Combinatorial Mathematics

This long-awaited textbook is the most comprehensive introduction to a broad swath of

combinatorial and discrete mathematics. The text covers enumeration, graphs, sets, and
methods, and it includes both classical results and more recent developments. Assuming
no prior exposure to combinatorics, it explains the basic material for graduate-level stu-
dents in mathematics and computer science. Optional more advanced material makes it
also valuable as a research reference.
Suitable for a one-year course or a one-semester introduction, this textbook prepares
students to move on to more advanced material. It is organized to emphasize connections
among the topics and facilitate instruction, self-study, and research, with more than 2200
exercises (many accompanied by hints) at various levels of difficulty. Consistent notation
and terminology are used throughout, allowing for a discussion of diverse topics in a uni-
fied language. The thorough bibliography, containing thousands of citations, makes this a
valuable source for students and researchers alike.

Douglas B. West is Professor of Mathematics at Zhejiang Normal University and Professor

Emeritus at the University of Illinois, where he won a campus-wide teaching award in
2002. Professor West has written more than 250 research articles on diverse topics in com-
binatorics and has advised 38 doctoral students. His earlier books include Introduction to
Graph Theory (1996, 2001), a popular textbook adopted around the world for courses. He
is Editor-in-Chief of Discrete Mathematics (since 2007) and Associate Editor of the Ameri-
can Mathematical Monthly (editing solutions for the Problems section since 1986). He also
maintains web pages at https://faculty.math.illinois.edu/∼west/ giving links to conferences
in discrete mathematics and advice on writing mathematics.
Combinatorial Mathematics

DOUGLAS B. WEST
Zhejiang Normal University, China and University of Illinois, Urbana-Champaign
 University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre, New Delhi – 110025, India
79 Anson Road, #06–04/06, Singapore 079906

Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781107058583
DOI: 10.1017/9781107415829
© Douglas B. West 2021
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2021
Printed in the United Kingdom by TJ International Ltd, Padstow Cornwall
A catalogue record for this publication is available from the British Library.
ISBN 978-1-107-05858-3 Hardback
Additional resources for this publication at www.cambridge.org/west
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
For my dear wife Ching
Contents

Preface xiii

Chapter 0 – Introduction. 1
Sets, Functions, and Relations: 2. Graphs: 3.
Discrete Probability: 6. Other Discrete Structures: 8. Complexity: 9.

Part I — Enumeration

Chapter 1 – Combinatorial Arguments. 13

1.1. Classical Models. 14
Elementary Principles: 14. Words, Sets, and Multisets: 16.
Exercises: 20.
1.2. Identities. 24
Lattice Paths and Pascal’s Triangle: 24. Delannoy Numbers: 28.
Exercises: 31.
1.3. Applications. 35
Graphs and Trees: 35. Multinomial Coefficients: 37.
The Ballot Problem: 39. Catalan Numbers: 41. Exercises: 46.

Chapter 2 – Recurrence Relations. 51

2.1. Obtaining Recurrences. 52
Classical Examples: 52. Variations: 56. Exercises: 59.
2.2. Elementary Solution Methods. 66
The Characteristic Equation Method: 67.
The Generating Function Method: 73. Exercises: 78.
2.3. Further Topics. 82
The Substitution Method: 82. Asymptotic Analysis: 84.
The WZ Method (optional): 87. Exercises: 91.

vii
viii Contents

Chapter 3 – Generating Functions. 93

3.1. Ordinary Generating Functions. 93
Modeling Counting Problems: 94. Permutation Statistics: 98.
Exercises: 103.
3.2. Coefficients and Applications. 107
Operations and Summations: 107. Snake Oil: 112. Exercises: 114.
3.3. Exponential Generating Functions. 118
Modeling Labeled Structures: 118. Stirling and Derangement
Applications: 121. The Exponential Formula: 124.
The Lagrange Inversion Formula (optional): 129. Exercises: 132.
3.4. Partitions of Integers. 137
Generating Function Methods: 137. Ferrers Diagrams: 141.
Bulgarian Solitaire (optional): 143. Distribution Models: 145.
Exercises: 148.

Chapter 4 – Further Topics. 153

4.1. The Inclusion-Exclusion Principle. 153
The Basic Principle: 153. Restricted Permutations: 159.
Signed Involutions: 163.
Determinants and Path Systems (optional): 165. Exercises: 171.
4.2. P ólya–Redfield Counting. 178
Burnside ’s Lemma: 179. The Pattern Inventory: 181.
Classical Cycle Indices: 185. Exercises: 187.
4.3. Permutations and Tableaux. 189
The Hook-Length Formula: 189. The RSK Correspondence: 193.
Switching P-Symbol and Q-Symbol: 198.
Jeu de Taquin (optional): 201. Exercises: 206.

Part II — Graphs

Chapter 5 – First Concepts for Graphs. 209

5.1. Definitions and Examples. 209
Graphs and Subgraphs: 209. Isomorphism: 211.
The Petersen Graph and Hypercubes: 213. Exercises: 216.
5.2. Vertex Degrees. 220
The Degree-Sum Formula: 220. Degree Lists: 221.
Extremality: 223. Directed Graphs: 224. Exercises: 226.
5.3. Connection and Decomposition. 229
Components and Walks: 229. Cycles and Cut-Edges: 231.
Eulerian Circuits: 233. Exercises: 235.
5.4. Trees and Distance. 239
Properties of Trees: 240. Distance and Diameter: 242.
Optimization on Weighted Graphs: 245. Exercises: 247.
Contents ix

Chapter 6 – Matchings. 253

6.1. Matching in Bipartite Graphs. 253
Hall’s Theorem: 254. Min-Max Relations: 258. Exercises: 260.
6.2. Matching in General Graphs. 264
Tutte ’s 1-Factor Theorem: 264. General Factors of Graphs: 268.
Exercises: 272.
6.3. Algorithmic Aspects. 276
Augmenting Paths: 277. Weighted Bipartite Matching: 279.
Fast Bipartite Matching (optional): 283.
Stable Matchings (optional): 284. Exercises: 286.

Chapter 7 – Connectivity and Cycles. 289

7.1. Connectivity Parameters. 289
Separating Sets: 289. Edge Cuts: 293. Blocks: 294. Exercises: 295.
7.2. Properties of ¾-Connected Graphs. 298
Menger ’s Theorem: 298. Applications of Menger ’s Theorem: 301.
2-Connected and 3-Connected Graphs: 304.
Highly Connected Orientations (optional): 307. Exercises: 311.
7.3. Spanning Cycles. 316
Properties of Hamiltonian Graphs: 317.
Sufficient Conditions: 319. Long Cycles (optional): 322.
Further Directions (optional): 325. Exercises: 328.

Chapter 8 – Coloring. 335

8.1. Vertex Coloring. 335
Upper Bounds: 336. Triangle-Free Graphs: 339. Exercises: 341.
8.2. Structural Aspects. 344
Color-Critical Graphs: 344. List Coloring: 346.
Forced Subgraphs (optional): 349. Exercises: 353.
8.3. Edge-Coloring and Perfection. 357
Special Classes: 357. Vizing ’s Theorem and Extensions: 359.
List Edge-Coloring: 362. Perfect Graphs: 366. Exercises: 370.

Chapter 9 – Planar Graphs. 377

9.1. Embeddings and Euler ’s Formula. 377
Drawings and Duals: 378. Euler ’s Formula: 383. Exercises: 385.
9.2. Structure of Planar Graphs. 390
Kuratowski’s Theorem: 390. The Separator Theorem (optional): 394.
Exercises: 397.
9.3. Coloring of Planar Graphs. 399
Edge-Colorings and Spanning Cycles: 400.
5-Colorable and 5-Choosable: 402. The Four Color Problem: 404.
Discharging and Light Edges: 407.
Other Aspects of Discharging (optional): 412. Exercises: 417.
x Contents

Part III — Sets

Chapter 10 – Ramsey Theory. 425

10.1. The Pigeonhole Principle. 425
Classical Applications: 426. Monotone Sublists: 430.
Pattern-Avoiding Permutations (optional): 431.
Large Girth and Chromatic Number (optional): 434.
Edge-Coloring of Hypergraphs (optional): 438. Exercises: 440.
10.2. Ramsey ’s Theorem. 443
The Main Theorem: 443. Applications: 445. Ramsey Numbers: 448.
Graph Ramsey Theory: 453. Exercises: 458.
10.3. Further Topics. 461
Van der Waerden’s Theorem: 461. Infinite Sets (optional): 468.
The Canonical Ramsey Theorem (optional): 470. Exercises: 473.

Chapter 11 – Extremal Problems. 475

11.1. Forced Subgraphs. 475
Turán’s Theorem: 475. Erdős–Stone Theorem: 478.
Linear Ramsey for Bounded Degree: 483. Roth’s Theorem: 484.
Proof of the Regularity Lemma (optional): 486. Exercises: 489.
11.2. Families of Sets. 493
The Kruskal–K atona Theorem: 493.
Antichains and Intersecting Families: 496. Chvátal’s Conjecture: 501.
Sunflowers (optional): 503. Entropy (optional): 504. Exercises: 510.
11.3. Matroids. 513
Hereditary Systems and Examples: 514. Axiomatics of Matroids: 519.
Duality and Minors: 523. The Span Function: 527.
Matroid Intersection: 529. Matroid Union: 533. Exercises: 536.

Chapter 12 – Partially Ordered Sets. 541

12.1. Structure of Posets. 541
Definitions and Examples: 541. Dilworth’s Theorem and Beyond: 546.
Exercises: 549.
12.2. Symmetric Chains and LYM Orders. 552
Graded Posets: 552. Symmetric Chain Decompositions: 553.
LYM and Sperner Properties: 558.
Products of LYM Orders (optional): 562. Exercises: 564.
12.3. Linear Extensions and Dimension. 568
Order Dimension: 569. Computation and Bounds: 572.
Bipartite Posets: 575. Exercises: 582.
12.4. Special Families of Posets. 585
Semiorders and Interval Orders: 585. Lattices: 588.
Distributive Lattices: 591. Correlational Inequalities: 595.
A Problem in Ramsey Theory (optional): 601. Exercises: 604.
Contents xi

Chapter 13 – Combinatorial Designs. 609

13.1. Arrangements. 609
Latin Squares: 609. Block Designs: 612. Symmetric Designs: 615.
Hadamard Matrices: 618. Exercises: 622.
13.2. Projective Planes. 624
Relation to Designs: 624. Applications to Extremal Problems: 628.
Difference Sets: 634. Exercises: 638.
13.3. Further Constructions. 640
Steiner Triple Systems: 641. Graphical Designs: 645.
Resolvable Designs and Other Tools: 648.
The Euler Conjecture (optional): 651. Exercises: 654.

Part IV — Methods

Chapter 14 – The Probabilistic Method. 657

14.1. Existence and Expectation. 657
The Union Bound: 658. Random Variables: 662. Exercises: 666.
14.2. Refinements of Basic Methods. 670
Deletions and Alterations: 670. The Symmetric Local Lemma: 674.
The General Local Lemma (optional): 679. Exercises: 684.
14.3. Moments and Thresholds. 686
“Almost Always”: 687. Threshold Functions: 690.
Convergence of Moments: 694. Graph Evolution: 698. Exercises: 702.
14.4. Concentration Inequalities. 706
Chebyshev and Chernoff Bounds: 706. Martingales: 712.
Bounded Differences (optional): 719. Exercises: 721.

Chapter 15 – Linear Algebra. 723

15.1. Dimension and Polynomials. 723
The Dimension Argument: 723. Restricted Intersections
of Sets (optional): 727. Combinatorial Nullstellensatz: 732.
The Alon–Tarsi Theorem: 738. Exercises: 744.
15.2. Matrices. 747
Determinants and Trees: 747. Cycle Space and Bond Space: 752.
Permanents and Planar Graphs: 754.
Möbius Inversion (optional): 757. Exercises: 762.
15.3. Eigenvalues. 766
Spectra of Graphs: 766. Eigenvalues and Graph Parameters: 768.
Regular and Strongly Regular Graphs: 772.
Laplacian Eigenvalues: 776. Exercises: 780.
xii Contents

Chapter 16 – Geometry and Topology. 783

16.1. Graph Drawings. 783
Embeddings on Grids: 783. Crossing Number: 790. Exercises: 797.
16.2. Combinatorial Topology. 798
Sperner ’s Lemma and Bandwidth: 799. Equivalent Topological
Lemmas: 802. The Borsuk–Ulam Theorem: 806. Kneser Conjecture
and Gale ’s Lemma: 811. Ham Sandwiches and Bisections: 815.
Borsuk’s Conjecture: 817. Exercises: 818.
16.3. Volumes and Containment. 821
Monotone Subsequences: 821. Balanced Comparisons: 822.
Containment Orders: 828. Exercises: 831.

Hints to Selected Exercises 833

References 849
Author Index 929
Glossary of Notation 943
Subject Index 949
Preface
Combinatorics is now a mature discipline. Although some see it as a mael-
strom of isolated problems, it has central themes, techniques, and results that
make it a surprisingly coherent subject. Meanwhile, it still rewards its students
with endless discovery and delight.
This book introduces the reader to a substantial portion of combinatorics.
It is not exhaustive in topics, results, or bibliography. However, it is thorough
enough to equip the reader with the tools needed to read or do research in com-
binatorics or to apply combinatorics in other areas of mathematics and computer
science. It assumes the maturity and sophistication of graduate students with-
out assuming prior exposure to combinatorics. It assumes basic undergraduate
mathematics, such as elementary set theory, induction, equivalence relations,
limits, elementary calculus, and some linear algebra.
More advanced or specialized material is planned to appear in The Art of
Combinatorics, a four-volume series of texts intended for researchers and for ad-
vanced graduate courses in combinatorics. Nevertheless, there is enough here to
reward substantial study and investigation.

History and Rationale

Despite its fundamental nature and its explosive growth in recent decades,
combinatorics still is not a standard part of mathematics instruction. Curricula
(and mathematicians) are slow to change.
Combinatorial ideas appear in courses on elementary discrete mathematics,
but such courses can be insubstantial. Serious undergraduate courses in combi-
natorics are seldom required for math majors. Graduate programs do not require
combinatorics. Nevertheless, it is an elegant and valuable subject.
In the mid-1980s, I began to teach graduate courses in combinatorics at the
University of Illinois. Excellent books existed for many topics, but every general
textbook omitted substantial areas. Gathering material for such a textbook, I
succumbed to the overabundance of riches before me. With so much beautiful ma-
terial in combinatorics, the project grew to become four rotating courses taught
from four books, now called The Art of Combinatorics.

xiii
xiv Preface

In 1996, I realized that this structure served only students already commit-
ted to focusing on combinatorics. For others, an overview of the subject could
have great value. An educated mathematical scientist should know some alge-
bra and analysis, and also such a person should be acquainted with fundamental
combinatorics and its relationships to other areas. Furthermore, disparities in
preparation of entering students make a core course worthwhile to establish a
common background before studying advanced material in combinatorics.
In 1997, I started a one-semester overview course to serve these goals. I ex-
tracted the fundamental material from The Art of Combinatorics and organized
it to emphasize connections among topics. This book is the result. However, with
so much beautiful combinatorics to choose from, I could not bring myself to cut
the book down to one semester. It can support a two-semester sequence, analo-
gous to fundamental two-semester sequences in classical areas of mathematics.
It can also support various one-semester courses, as discussed later.
Since the scope is large, I have also sought to make the book useful as a re-
search reference, rewarding further study after the courses are over. This leads to
a fair amount of optional material allowing the reader to probe farther into the
subject, plus remarks that provide statements and pointers to further results.
Nevertheless, I still aim to keep the material accessible to graduate students.

Organization

One can organize combinatorial mathematics in many ways: by structures

discussed, types of questions, methods used, etc. In a broad overview, the connec-
tions among topics are as important as the groupings within topics.
Most presentations of elementary combinatorics begin with enumeration or
with graph theory; the former is the more classical approach. Natural enumera-
tive questions arise in elementary graph theory, and many graph-theoretic argu-
ments use basic counting techniques, so each informs the other. Here the basic
notions of trees, cycles, and isomorphism are stated in Chapter 0 so that enumer-
ative problems about graphs can be used as examples in Part I.
Part I presents the basics of bijective arguments, recurrence relations, gen-
erating functions, and inclusion-exclusion, with enhancements. Young tableaux
and the elementary aspects of P ólya–Redfield counting appear here from a com-
binatorial point of view. Deeper algebraic aspects of enumeration are omitted.
Part II pursues central themes of elementary graph theory while reaching
important and classical results, particularly those having broad applications.
Graph theory is now a huge subject, so selecting fundamental core material is
difficult. Many large topics are mentioned here at most in passing or in exercises;
these include automorphism groups, Cayley graphs, graph representations, re-
construction, domination, decomposition, packings, genus, minors, nowhere-zero
flows, Tutte polynomials, graph labelings, and structured families of graphs.
Part III explores our most general structural object: families of sets, gener-
alizing graphs to hypergraphs. Four aspects of set systems are studied: Ramsey
theory, extremal set theory, the structure of partially ordered sets and matroids,
and combinatorial designs. Many aspects of posets and enumeration known as
algebraic combinatorics are omitted (but Möbius inversion is in Chapter 15).
Preface xv

Part IV develops methods from probability, algebra, and geometry/topology

and applies them to questions concerning graphs and sets. Also included are some
applications of combinatorics to geometric questions. When discussing methods
and connections, it helps to have the terminology and basic results of graph the-
ory and enumeration available. Thus the material in the latter half of the book
does depend on the earlier half.
Some topics omitted here are explored in The Art of Combinatorics or in my
earlier Introduction to Graph Theory, which is a more patient and less sophisti-
cated introduction to elementary graph theory. Some important topics in applied
discrete mathematics are largely omitted here, partly because they often already
have their own well-established courses; prominent among these are coding the-
ory and linear programming.
The nearly 2200 exercises here apply ideas from the text and/or explore fur-
ther concepts. Many have not appeared in texts before. More than 300 have hints
with the problem statements, and another 380 have hints in the back of the book.
Solutions are available to instructors via the book’s website at
http: / /www.cambridge.org /west .
I have tried to indicate difficulty by marking easier problems with (−) and harder
problems with (+). Problems of intermediate difficulty that are particularly in-
teresting or instructive are marked with (♦). There is much ambiguity (and taste)
in these designations, partly because the difficulty of finding a solution is not pro-
portional to its length or its complexity. Thus these labels should be taken lightly.

Usage

Most schools have few regular graduate courses in combinatorics. At such

schools this book is appropriate for a two-semester sequence to give a thorough
introduction. Instead of separating graph theory from other topics to make two
courses, this text integrates the topics into a coherent whole.
This approach enables students from other mathematical areas to acquire
the fundamental material about enumeration, graphs, and sets in the first
semester without continuing to the second. Also, topics that are best appreci-
ated after knowing the fundamentals in several areas of combinatorics are omit-
ted from courses that study only one part of combinatorics. Examples include the
existence (Chapter 14) or construction (Chapter 10) of graphs with large chro-
matic number and girth, powerful techniques like the Regularity Lemma and
entropy (Chapter 11), the application of projective planes to extremal graph prob-
lems (Chapter 13), the understanding of optimization via matroids (Chapter 11),
and combinatorial applications of linear algebra and topology (Chapters 15 and
16). With the approach here, such applications enliven the second semester.
Nevertheless, the text can also be used separately for courses in graph the-
ory and in “other ” combinatorics, as described later.
In a two-semester sequence where most students take the full year and prior
combinatorics is not expected, one can focus on Parts I and II in the first semester,
III and IV in the second. Parts I and II concisely present the basic material taught
in most undergraduate courses on combinatorics and graph theory, with a deeper
xvi Preface

point of view for graduate students. Students with prior exposure to the subject
also benefit from this discussion. Classical topics in graph theory reside in Part
II, but interactions between graphs and other topics and techniques appear in
other chapters. Later topics are more independent, but the order of presentation
here works well. Part III can be viewed as a third introductory area; it considers
basic questions about sets and order relations. The methods of Part IV then apply
to questions about the combinatorial contexts introduced in Parts I–III.
When the second semester is optional, with the first being the exposure that
students from other areas will have to combinatorics, the first semester should be
broader. Such a course has goals like those in a one-semester core course leading
to multiple advanced courses in combinatorics. With this in mind, I have desig-
nated some sections and subsections as “optional”, and in others the items marked
“*” are optional. Such material is more technical or advanced and can be skipped
at first reading without loss of continuity. In such a course, optional or more dif-
ficult topics should be skipped in Chapters 1–9 in order to present highlights of
early portions of Chapters 10 and 12–14. I used this approach in a one-semester
introduction at the University of Illinois that served as a departmental gradu-
ate exam course in combinatorics and prepared combinatorics students for four
independent advanced courses.
That was a fast-moving course. I spent 16 lectures on enumeration (this can
be less), 12 on graphs, and several each on Ramsey theory, posets, probabilistic
methods, and designs, aiming in the latter topics mostly to introduce the basic
ideas. Thus I was quite brief about signed involutions, the pattern inventory,
and the magical properties of Young tableaux. Most students already had some
acquaintance with graph theory (the clientele included many computer science
students), so from Chapter 5 I presented just a few highlights and left the rest
as background reading. In Chapters 6–9 I covered mostly fundamentals. Gradu-
ate students should see a bit extra, so in Chapter 6 I proved Plesnı́k’s Theorem
instead of Petersen’s Theorem, in Chapter 7 I gave the lower bound on connec-
tivity of graph products, in Chapter 8 I approached Brooks’ Theorem through
list coloring (an important theme in modern graph theory), and in Chapter 9 I
explained discharging. I skipped the subsections marked optional and also the
proof of the Perfect Graph Theorem. In Section 9.3, one can present just enough
about discharging to convey the idea.
In Chapter 10, I then presented a few choice pigeonhole applications and the
main Ramsey theorem with several applications. In Chapter 12, the main goals
were Dilworth’s Theorem and LYM Orders for their connections with graph the-
ory. Chapter 14 came before Chapter 13 to provide more time for homework prob-
lems, reaching the Local Lemma and threshold functions. In Chapter 13 the goal
was the connection between latin squares and projective planes and the applica-
tion of projective planes to extremal problems in graph theory.

Graph Theory vs. Combinatorics

Many institutions still have separate courses in graph theory and “other ”
parts of combinatorics, partly due to faculty interests.
A course in enumerative and set-theoretic combinatorics can use Part I and
Part III. From graph theory one needs only the definitions from Chapter 0 to
Preface xvii

present Cayley ’s Formula (Section 1.3) and count isomorphism classes (Section
4.2). The discussion of chromatic polynomials in Section 4.1 using inclusion-
exclusion can be skipped. Part II can be skipped completely. In Part III, one
can skip graph-theoretic applications of the Pigeonhole Principle in Section 10.1,
graph Ramsey Theory in 10.2, and all of 11.1. The material of Chapters 12 and
13 (except for symmetric chain decomposition of LYM orders and graph-theoretic
applications of projective planes) is mostly accessible without graph theory.
I have used Part II for a graph theory course for masters students at Zhe-
jiang Normal University in China. Moving more slowly to accommodate language
difference, I did not go much beyond Part II. One would cover Part II except for
some optional material. It does not require Part I except for binomial coefficients
and simple bijective arguments (counting two ways). In Part III one can use
graph-theoretic pigeonhole examples and graph Ramsey theory, possibly adding
Section 11.1. The material on matroids in Section 11.3 generalizes results in
graph theory, but it takes substantial time to develop the properties. It is more
beneficial to include the basic material of Chapter 14, since probabilistic tech-
niques are so effective and important and easily illustrated with graphs.
When a school has separate graduate courses in basic enumeration and basic
graph theory, most likely most of the material from Parts III and IV will not be
included, depending on the needs of the students and choices by the instructor. A
subsequent course requiring the two basic courses can then cover topics from the
last two Parts. Although the text often mentions connections between chapters,
the chapters after Part II are relatively independent except for the background
of language from the early parts.
It is worth noting that a two-semester sequence at U. Nebraska has for about
10 years used Part I combined with supplementary material on coding and infor-
mation theory in the first semester, and Part II combined with selections from
Chapters 10–15 in the second semester.

Highlights
Indeed, the connections between topics are among the features of this book.
One aspect is pedagogical: we solve several fundamental problems repeatedly to
show the usefulness of various techniques.
For example, Cayley ’s Formula to count labeled trees is obtained bijectively
in Chapter 1, inductively in exercises, via generating functions and Lagrange In-
version in Chapter 3, and in Chapter 15 via the Matrix Tree Theorem and via
eigenvalues. Derangements are counted via recurrence, generating functions,
and inclusion-exclusion; Catalan numbers are also obtained repeatedly, includ-
ing from Young tableaux. Turán’s Theorem on the maximum number of edges
in a graph containing no (r + 1)-vertex complete graph is proved inductively in
Chapter 5, via extremality in Chapter 11, algebraically in exercises in Chapter
11, and probabilistically in Chapter 14. Planar graphs are characterized induc-
tively in Chapter 9, via matroids in Chapter 11, and via dimension of partial
orders in Chapter 16. The K önig–Egerváry max/min relation of Chapter 6 in-
volving matchings and vertex covers in bipartite graphs is shown to be equiva-
lent to Dilworth’s Theorem on posets in Chapter 12 and is a special case of the
Matroid Intersection Theorem in Chapter 11.
xviii Preface

Other connections arise when techniques from one context are used to solve
problems from other contexts. For example, common systems of distinct represen-
tatives (characterized in Chapter 7 using Menger ’s Theorem) are used to obtain
symmetric chain decompositions of LYM orders in Chapter 12. Extremal prob-
lems for diameter in Chapter 5 and for graphs without 4-cycles in Chapter 11
are attacked using projective planes in Chapter 13. The problem in Chapter 8 of
finding small triangle-free graphs with large chromatic number is discussed us-
ing Ramsey theory in Chapter 10. Probabilistic methods from Chapter 14 are
used to obtained good bounds on crossing numbers in Chapter 16. Bounds on the
graph connectivity parameter from Chapter 7 are obtained via eigenvalues in
Chapter 15. Bounds on the list chromatic number, introduced for a richer study
of graph coloring in Chapter 8, are obtained using the Discharging Method in
Chapter 9 and the Combinatorial Nullstellensatz in Chapter 15. Pym’s Theorem
from Chapter 7 is used to prove the Planar Separator Theorem in Chapter 9.
Finally, since this is a graduate textbook, it covers the standard material of
an elementary introduction efficiently in order to go beyond and offer the reader
more. This permits the inclusion of many jewels that a standard elementary
introduction at the undergraduate level cannot reach. Some of these were men-
tioned earlier in describing my overview course. Here are more of them.
In enumeration, we become familiar with the Delannoy numbers and the
Eulerian numbers as additional basic counting models. We explain techniques
such as Wilf ’s Snake Oil technique for evaluating sums, the Exponential For-
mula for obtaining generating functions, and Lagrange Inversion for extracting
coefficients. We explore the combinatorial aspects of Young tableaux, obtaining
not only the Hook-Length Formula and the Robinson–Schensted–Knuth Corre-
spondence, but also Greene ’s Theorem about the largest union of ¾ increasing
subsequences in a permutation.
In graph theory, Chapter 5 provides unusual applications of the number of
vertices of odd degree being even. Orientations with small outdegree, a beauti-
ful application of Hall’s Theorem by Hakimi in Chapter 6, are later applied to
list coloring in Chapter 15. Optional material for advanced courses includes the
proof of Tutte ’s º -Factor Theorem and the fast matching algorithm of Hopcroft
and K arp in Chapter 6, and in Chapter 7 the Nash-Williams Orientation Theo-
rem characterizing ¾-connected orientations (generalizing Robbins’ Theorem on
strong orientations). Also in Chapter 7, the standard sufficient conditions for
spanning cycles in graphs are studied extended to long-cycle versions. List col-
oring provides a modern approach to coloring in Chapter 8, and the full form
of Vizing ’s Theorem for multigraphs is proved. Chapter 9 includes a thorough
introduction to the Discharging Method and an accessible proof of the Planar
Separator Theorem.
In Part III on sets, we present many beautiful results that are not easy to
find in general textbooks. Again some will typically be options for further read-
ing. Chapter 10 includes the proof of the Stanley–Wilf Conjecture on pattern-
avoiding permutations and the construction of graphs with large girth and chro-
matic number. The application of Ramsey ’s Theorem to table storage in computer
science is unusual and makes use of allowing many colors. In Chapter 11, the Reg-
ularity Lemma of Szemerédi is presented, applied, and proved. Also, the notion
of discrete entropy is developed to obtain extremal bounds on set counting prob-
Preface xix

lems. The basic structural aspects of posets are given in Chapter 12, along with
important results about poset dimension, and the treatment of lattices leads to a
rigorous discussion of correlational inequalities. Beyond the basic results about
designs and projective planes, Chapter 13 includes the Multiplier Theorem for
difference sets and the disproof of the Euler Conjecture.
Part IV on methods provides tools to attack many problems. Chapter 14
on the Probabilistic Method has a scope similar to the popular textbook of Alon
and Spencer, including the basic methods, Dependent Random Choice, the Lo-
cal Lemma, threshold functions, and concentration inequalities. The scope of
Chapter 15 is similar to the well-known notes of Babai and Frankl on Linear Al-
gebra Methods in Combinatorics; we also include K astelyn’s use of permanents
to count perfect matchings in planar graphs and a discussion of Möbius inver-
sion on posets. The Combinatorial Nullstellensatz is pursued as far as the re-
cent strengthening of Thomassen’s famous 5-choosability of planar graphs by
Grytczuk and Zhu, show that every planar graph contains a matching whose
edge-deletion yields a 4-choosable subgraph. Chapter 16 discusses geometric em-
beddings of graphs, applications of the Borsuk–Ulam Theorem and its relatives
from combinatorial topology, and geometric aspects of partially ordered sets.
I hope this brief sampling whets the appetite for the delights ahead.

Acknowledgments

When C.L. Liu heard in the mid-1980s that I was accumulating text mate-
rial on combinatorics, he showed me the lecture notes he had published as Topics
in Combinatorial Mathematics (Math. Assoc. of America, 1972). These came from
a summer seminar at Williams College in 1972 and were used in the combina-
torics graduate course at the University of Illinois that I inherited from him. He
proposed that we work them into a polished textbook; thus began The Art of Com-
binatorics. As described earlier, that project grew beyond the confines of a single
volume, and the present text is closer to what he had in mind (but still more than
twice as big). I thank him for the suggestion that started the process.
Also worthy of mention is Liu’s earlier book Introduction to Combinatorial
Mathematics (McGraw-Hill, 1968), which in 1972 introduced me to combina-
torics. This book established the overall shape and subject matter for modern
courses in combinatorics. Before it (at least in the U.S.) there was not much
more than a compilation of chapters from eminent researchers who delivered a
short course for engineers at UCLA (Applied Combinatorial Mathematics, 1964).
Courses in elementary graph theory were initially shaped by the seminal text-
books of Berge (1962), Harary (1969), and Bondy and Murty (1976).
This text has benefitted by comments from many users and reviewers. Those
who used pre-publication versions of the text several times include Garth Isaak
(Lehigh U.), Art Benjamin (Harvey Mudd), Stephen Hartke, Christine Kelley,
Xavier P érez-Giménez, and Jamie Radcliffe (Nebraska), Jozsef Balogh (Illinois),
and Mark K ayll and Cory Palmer (Montana). John Ganci and Leen Droogendijk
each gave the book an extremely thorough reading, catching many glitches.
Other reviewers contributing insightful comments on early versions in-
cluded Martin Aigner, Mike Albertson, Lowell Beineke, Miklós Bóna, Graham
xx Preface

Brightwell, Lynne Butler, Ira Gessel, Jay Goldman, Jerry Griggs, Mike Jacob-
son, Jenő Lehel, Herbert Maier, Michael Molloy, Chris Rodger, Bruce Rothschild,
László Szekely, and Wal Wallis. Comments on particular chapters in later ver-
sions came from Noga Alon, Louis DeBiasio, Stefan Felsner, David Gunderson,
Hemanshu K aul, Sasha Kostochka, Cory Palmer, Pawel Prałat, Joel Spencer,
Tom Trotter, Peter Winkler, and G ünter Ziegler.
I also thank generations of students who labored with slowly evolving iter-
ations of this material. Those who found numerous typos include Shivi Bansal,
Alfio Giarlotta, Farzad Hassanzadeh, Bill K innersley, Darren Narayan, Radhika
Ramamurthi, Michael Santana, Prasun Sinha, and Reza Zamani. I apologize to
many others I have forgottten to mention over the long years of development.
At Cambridge University Press I thank my editor David Tranah for his pa-
tience through years of delays as I slowly refined the text. He accurately concluded
that I view the book as a “work of art ”, which is part of why it took so long (25
or 35 years, depending on how you count). Clare Dennison shepherded the book
through production, and Sarah Routledge gave it an incredibly thorough proof-
reading. All remaining errors, which I am sure exist, are solely my responsibility,
especially since I continued to squeeze in exercises and make other refinements for
months after she finished her job.
This book has been typeset using TEX. The scientific community owes a vast
debt to its creator, Donald E. Knuth. With brilliance, foresight, and generosity,
he has provided a common language for the publication and communication of
technical material that is now used all over the world. Besides its versatility and
free availability, its incredible genius is that it runs amazingly fast.
Chris Hartman taught me perl, which I used to convert earlier groff files
to TEX. The “millenial” fonts were developed by Stephen Hartke, who helpfully
made it possible to use them with plain TEX instead of LATEX so that I could have
greater control over spacing and placement of material. The references and in-
dexes were assembled through herculean effort by Thomas Mahoney, who wrote
scripts to handle most of the processing, built an effective computing environment
for me to use in both China and the U.S., and patiently helped me resolve all my
internet difficulties. Finally, I thank my wife, Ching Muyot, for her assistance,
patience, and understanding with the crunch of each year ’s edition, no matter
how many times I declared the book “essentially finished”.

Feedback
I eagerly welcome comments on all aspects of this book. This includes se-
lection and presentation of topics, errors made in mathematics or attribution or
typography, items missing from the index, suggestions of additional hints, ma-
terial that should be added if there is ever another edition, etc. Please send com-
ments to [email protected]. Errata will be listed at
http: / /www.math.uiuc.edu/ ∼west /coerr.html .
Enjoy!
Douglas B. West
[email protected]
Urbana, IL, and Jinhua, China
Chapter 0

Introduction
Combinatorial problems and arguments have a long history in mathematics,
but only in the last half of the 20th century did they become a coherent sub-
ject. The discipline was long viewed as a collection of isolated tricks, but now the
methods are more systematic, and the connections and applications between com-
binatorics and other areas of mathematics (in both directions) are being studied.
In this book we explore some of these connections and many fundamental re-
sults of combinatorics. We do not assume any prior exposure to combinatorics,
but we assume mathematical maturity and basic undergraduate mathematics,
including elementary set theory, induction, equivalence relations, limits, calcu-
lus, linear algebra, etc.
One can classify mathematical problems by the type of question, the object
being studied, the method used, etc. These various aspects make it hopeless to
impose a linear order of development in the study of mathematics. We emphasize
different aspects at different times.
Our questions are of three general types. Given constraints specified for an
object, does it exist? If such objects exist, how many are there? With respect
to some criterion, which one is the best? These are the problems of Existence,
Enumeration, and Extremality. We emphasize enumerative problems in Part I
and problems of existence and extremality in most of the rest of the text.
We study objects that are discrete structures of various types. The simplest
is a set. More complicated structures arise by imposing constraints or relations
on sets or families of sets. We study various arrangements in Part I and graphs
in Part II. In Part III we study structures such as hypergraphs, partially ordered
sets, combinatorial designs, and matroids.
Finally, we also study methods of combinatorics. Many techniques arise in
conjunction with particular structures, but some are used in many contexts and
are worthy of study in their own right. Our focus on techniques is clearest in
Part IV, where we discuss the probabilistic method, algebraic methods, and con-
nections with geometry, but many other methods appear in earlier parts.
In this brief introduction, we review definitions from elementary mathemat-
ics, introduce elementary concepts about graphs for use in Part I, introduce ele-
mentary notions of probability as background for questions throughout the text,
describe some additional discrete structures, and mention the basic notions of
complexity. This is background material to be consulted as needed.

1
2 Chapter 0: Introduction

SETS, FUNCTIONS, AND REL ATIONS

Our most fundamental object is a set. We build other structures from sets and
relations. We use 0 for the set of nonnegative integers and for the set of pos-
itive integers, also called the natural numbers. We let [n] (pronounced “bracket
n”) denote the set {1 , . . . , n} of the first n natural numbers, with [0] = ∅. We take

as given the number systems , , , , (natural numbers, integers, rational
numbers, real numbers, complex numbers) and their elementary arithmetic and
order properties. Similarly we assume the elementary operations and notation
of sets, such as membership, containment, union, intersection, complement, and
difference. We write the difference of sets A and B as A − B, not A \ B.
A function from a set A to a set B, written : A → B, assigns each x ∈ A
an element (x) ∈ B. The set A is the domain; is defined on A. The image of
x ∈ A is (x), and { (x): x ∈ A} is the image of . The function is injective if
each y ∈ B is the image of at most one element of A. It is surjective if each y ∈ B
is the image of at least one element of A. It is a bijection if it is injective and
surjective, and then it provides a one-to-one correspondence between A and B.
A set S is finite if a bijection from S to [n] exists for some n ∈ 0 ; the value n
is then the size of S, written | S|. Counting a finite set means computing its size.
Two sets (finite or infinite) have the same cardinality if there is a bijection from
one to the other. For finite sets, “cardinality” is a synonym for “size”. A set with
the same cardinality as is countable.
A sequence is a function with domain (or 0); we write an for the im-
age of n, the nth term of the sequence. Usually one letter (such as ) denotes a
function; similarly we use ⟨a⟩ to denote the sequence with terms of the form an .
A list is a function defined on [n] for some n ∈ 0 ; this is the finite analogue
of a sequence. We write a list a of length n as an n-tuple (a1 , . . . , an). (Many au-
thors use “sequence” for an n-tuple; we try to use “sequence” only for functions
on .) A binary n-tuple or 0 , 1-list is a list with entries in {0 , 1}. Similarly,
a 0 , 1-matrix or binary matrix has entries in {0 , 1}. A ternary list has en-
tries in {0 , 1 , 2}. An n-ary list takes values from a specified set of size n. An
arithmetic progression is a list of equally spaced integers.
In contrast to a set, the order of elements in a list matters, and elements in
lists may repeat. A multiset differs from a set by allowing repeated elements,
but order remains unimportant. We can specify a multiset by specifying the set
of distinct elements and their multiplicities. Since the order is unimportant but
repetition is allowed, some authors refer to multisets as “unordered lists”.
A permutation of a finite set S is a bijection from S to itself. Since a func-
tion on [n] is a list, we may view a permutation  of [n] as a function from [n] to
[n] or as a listing of [n] in some order a1 , . . . , an , with ai denoting  (i). The latter
is the word form of the permutation. Both viewpoints will be useful.
We often discuss sets whose elements are also sets. To avoid confusion, we
use class and family as synonyms for “set ”. Instead of saying “a set in a set of
sets”, we say “a member of a family of sets”.
The cartesian product of sets S and T is the set S × T of ordered pairs
{(s , t): s ∈ S, t ∈ T}. A (binary) relation between S and T is a subset of the
cartesian product S × T . When S = T , we call this a relation on S. We say that
the pairs in a relation satisfy the relation.
Background Definitions 3

When S is a family of sets, the containment relation on S is the set of pairs

(A , B) ∈ S × S such that A ⊆ B. The pair (A , B) satisfies the disjointness rela-
tion when A ∩ B = ∅. As a relation, disjointness is a property of pairs. Hence “a
family of disjoint sets” technically has no meaning; nevertheless, by convention
the word “disjoint ” means “pairwise disjoint ” when applied to sets in a family.
An equivalence relation on S is a relation R with these properties:
1) reflexive — (x , x) ∈ R for all x ∈ S.
2) symmetric — (x , y) ∈ R if and only if (y , x) ∈ R.
3) transitive — if (x , y) , (y , ) ∈ R, then (x , ) ∈ R.
Containment is reflexive and transitive but not symmetric and hence not an equiv-
alence relation. “Having the same cardinality” defines an equivalence relation,
since the identity function is a bijection from a set to itself, the inverse of a bi-
jection is a bijection, and the composition of bijections is a bijection.
An equivalence class of an equivalence relation R on S is a maximal sub-
set T of S such that all pairs of elements of T satisfy R. Elements x and y are
in the same equivalence class if and only if (x , y) ∈ R, and each element of S be-
longs to exactly one such class. Hence the equivalence classes of an equivalence
relation on S form a partition of S, where a partition of a set S is a family of
disjoint nonempty sets whose union is S. The sets in a partition are its blocks.
Conversely, given a partition of S, putting (x , y) ∈ R if x and y are in the same
block yields an equivalence relation R on S. Hence partitions of S and equivalence
relations on S are essentially the same notion.

0.1. Example. Examples of equivalence relations.

Congruence modulo n. Integers x and y are congruent modulo n if x − y
is a multiple of n. The equivalence classes are the subsets of having a fixed
remainder upon division by n. These are the congruence classes modulo n, and
the family of congruence classes modulo n is denoted by n .
Orbits under a permutation. Viewing a permutation  of [n] as a bijection
from [n] to [n], we use  ¾ to denote the bijection obtained by applying  succes-
sively  times. The relation R on [n] that puts (i , j) ∈ R if  ¾ (i) = j for some
 ∈ 0 is reflexive and transitive. Since bijections have inverses, also R is sym-
metric. The equivalence class containing i in this equivalence relation is the set
of elements obtained as  is repeatedly applied, called its orbit.

0.2. Example. When S is a family of subsets of a “ground set ” X , the incidence

relation between X and S is the relation R consisting of the ordered pairs (x , A) ∈
X × S such that x ∈ A. When X = {x1 , . . . , x n} and S = {S1 , . . . , Sm}, we encode
R as a 0 , 1-matrix with position (i , j) being 1 if x i ∈ Sj and 0 otherwise. This is
the incidence matrix for R (with respect to the given indexing of X and S). The
jth column of the incidence matrix is the incidence vector of the set Sj ; this
0 , 1-vector records for each element of X whether it belongs to Sj .

GRAPHS

Many natural relations are symmetric and irreflexive, where irreflexive

means that for each x the pair (x , x) does not satisfy the relation (disjointness on
a family of nonempty sets, for example). Such relations are modeled by “graphs”.
4 Chapter 0: Introduction

A graph G is a pair consisting of a set V(G) of vertices and a set E(G) of

edges, where each edge is a set of two vertices. The order of G is | V(G)| , and its
size is | E(G)|. Vertices u and v forming an edge are adjacent, and the neighbors
of v are the vertices adjacent to it. Note that the adjacency relation is symmetric
and irreflexive. The two vertices in an edge are its endpoints, and we say that
the edge joins its endpoints.
An edge and its endpoints are incident. The incidence relation of a graph
G is the set of pairs (v , e) in V(G) × E(G) such that v and e are incident to each
other (v is an endpoint of e). When discussing graphs we drop set brackets for
edges and write the edge {u , v} as uv (or vu). To display a graph, we represent
each vertex by a point in the plane and each edge by a curve whose endpoints are
the points assigned to the vertices of the edge; this is a drawing of the graph.
Structural properties of graphs do not depend on the names of the vertices.
An isomorphism from a graph G to a graph H is a bijection : V(G) → V(H)
that preserves the adjacency relation: uv ∈ E(G) if and only if (u) (v) ∈ E(H).
The resulting isomorphism relation is an equivalence relation on any set of
graphs. Every graph isomorphic to G has the same structural properties as G,
so we treat isomorphic graphs as “equal” when not given explicit vertex names.
Containment and union for graphs follow those notions for sets. A graph H
is a subgraph of a graph G if V(H) ⊆ V(G) and E(H) ⊆ E(G). Similarly, the
union of graphs G and H is the graph G ∪ H with vertex set V(G) ∪ V(H) and
edge set E(G) ∪ E(H).

Many interesting enumerative questions can be asked about graphs. To per-

mit discussion of such enumerative questions in Part I, we develop some elemen-
tary examples and properties of graphs here.
Many useful graphs are defined by their structural properties. A path is a
graph whose vertices can be linearly ordered so that two vertices are adjacent if
and only if they are consecutive in the ordering; the endpoints of the path are
the first and last vertices in such an order. A cycle is a graph with an equal
number of vertices and edges whose vertices can be placed around a circle so that
two vertices are adjacent if and only if they appear consecutively along the circle.
A graph G is connected if for all u , v ∈ V(G), it contains a path with end-
points u and v. A component of G is a maximal connected subgraph of G, where
a maximal object of a given type is one that is not contained in another object of
that type (minimal is defined similarly). The graph below has three components.
u y
• •
r v w
• s• • • •
• •
t x
A u , v-path is a path with endpoints u and v. We say that u and v are con-
nected in G if G contains a u , v-path. The connection relation on V(G) is the set
of pairs (u , v) such that G has a u , v-path; it is reflexive and symmetric. Our first
proposition implies that it also is transitive and hence is an equivalence relation.
The point is that although a u , v-path and a v , w-path together need not form a
u , w-path, their union contains one. The equivalence classes of the connection
relation are the vertex sets of the components of G.
Background Definitions 5

The technique of extremality involves choosing an object that is extremal in

some respect (a maximal connected subgraph, for example). This often requires
finiteness, and our structures will be finite unless explicitly stated otherwise.

0.3. Proposition. If P is a u , v-path and P is a v , w-path, then P ∪ P contains

a u , w-path.
Proof: We use extremality. Let x be the first vertex along P from u to v that also
lies in P (x exists, since both sets contain v). The union of the u , x-path in P and
the x , w-path in P is a u , w-path, since these subgraphs share only x.

A spanning subgraph of G is a subgraph of G with vertex set V(G). A tree

is a graph that is connected and contains no cycles. A spanning tree in G is a
spanning subgraph of G that is a tree.

0.4. Proposition. Every connected graph contains a spanning tree.

Proof: Since every vertex is itself a tree, every connected graph G contains at
least one tree. Let T be a maximal tree contained in G.
If V(T)
= V(G), then a path from a vertex outside V(T) to a vertex in V(T)
yields an edge e with endpoints in V(T) and V(G) − V(T). Let T be the sub-
graph of G obtained from T by adding edge e and its endpoint v outside T . By
the transitivity of the connection relation, T is connected.
Also every cycle in T appears in T , since v is in only one edge in T . Thus T
has no cycle and is a tree. This contradicts our extremal choice of T . We conclude
that V(T) = V(G), and T is a spanning tree.

The degree of a vertex in a graph is the number of edges incident to it. A

leaf is a vertex of degree 1. A vertex of degree 0 is an isolated vertex.
A maximal path in a graph G is a path in G that is not a subgraph of another
path in G. Thus every vertex of G adjacent to an endpoint of a maximal path in
G must belong to the path.

0.5. Proposition. If all vertices in a finite graph G have degree at least 2, then
G contains a cycle.
Proof: Since V(G) is finite, G has a maximal path P. Let v be an endpoint of P.
Since d(v) ≥ 2, there is an edge vu not in P. Since P is maximal, u lies on P , and
vu completes a cycle with the u , v-path in P.

0.6. Proposition. Every tree with at least two vertices has at least two leaves.
Deleting a leaf from a tree yields a tree with one less vertex.
Proof: A connected graph with at least two vertices has no isolated vertices. A
tree has no cycle, so the endpoints of a maximal path in a tree with at least two
vertices have degree 1.
Given a leaf x in a tree G, obtain G from G by deleting x and its incident
edge. Since deleting a vertex creates no new subgraphs, G has no cycles. Hence
it suffices to show that G is connected. For distinct vertices u , v ∈ V(G ), there
a u , v-path P in G. Since internal vertices along a path have degree at least 2, P
does not contain x. Hence P is also a u , v-path in G .
6 Chapter 0: Introduction

0.7. Proposition. Every tree with n vertices has n − 1 edges. Furthermore, ev-
ery graph with n − 1 edges that arises from n isolated vertices by iteratively
adding an edge joining two components is a tree.
Proof: For the first statement, use induction on n. A 1-vertex tree has no edges.
For n > 1, Proposition 0.6 provides a leaf whose deletion yields a tree with n − 1
vertices. Since it has n − 2 edges, the original tree has n − 1 edges.
For the second statement, note first that each such edge addition creates no
cycles. Such a cycle would contain the new edge uv and another u , v-path from the
previous graph, which does not exist since u and v were in different components
in that graph. Also, each such edge addition reduces the number of components
by 1. Hence n − 1 additions reduce the number of components to 1. Thus the
resulting graph is acyclic and connected and is a tree.

We assume familiarity with these results in order to study counting problems

about trees in Part I. We discuss the structure of graphs more fully in Part II.

DISCRETE PROBABILITY

Many enumerative questions are easily motivated using discrete probability.

In such questions, there is a set U (the “universe”) of possible outcomes of some
process. These outcomes are assumed to be equally likely; this is the meaning of
phrases like “a random element ” and “chosen uniformly at random”. The proba-
bility of a desired property is then defined to be | A| / | U | , where A is the subset of
U consisting of all outcomes having the desired property.
In Part III we will also consider countable spaces, where an outcome may be
any natural number. Here we review the definitions of probability spaces.

0.8. Definition. A discrete probability space is a finite or countable set S

with a function  defined on the subsets of S (called events) such that
a) If A ⊆ S, then 0 ≤ (A) ≤ 1,
b) (S) = 1, and
c) If A1 , A2 , . . . are pairwise disjoint subsets of S, then
 (⋃ Ai) = ∑∞i=1 (Ai).
0.9. Remark. For a finite probability space, assuming (A ∪ B) = (A) + (B)
when A ∩ B = ∅ and applying induction on ¾ yields (A) = ∑i=1 (Bi) when
¾

B1 , . . . , B¾ is a partition of A. This follows from (c) above but does not imply it,
so we require the more general condition.
More general definitions of probability space allow the probability function to
be defined only on subsets of S with certain properties, but the simple definition
above suffices for our purposes. On the rare occasions where we mention continu-
ous probability spaces, we will be informal (that is, non-rigorous). For example,
when choosing a point at random from a region in the plane, we adopt the intu-
itive notion that the probability it lies in a particular subregion is proportional
to the area, with “regions” simple enough not to worry about measurability.
Background Definitions 7

Immediate consequences of Definition 0.8 include

a) (∅) = 0,
b) (A) = 1 − (A), and
c) (A ∪ B) = (A) + (B) − (A ∩ B).
Furthermore, (A) = ∑ a∈ A (a), writing (a) for ({a}) when a ∈ S. Elements
of a probability space S are sample points or outcomes (of an “experiment ”).

0.10. Definition. Events A and B in a probability space are independent if

(A ∩ B) = (A)(B). For events A and B with (B)
= 0, the conditional
probability of A given B, written (A | B), is defined to be (A ∩ B)/(B).

0.11. Remark. Saying that a probability space is generated by making choices

“independently” means that the space is a cartesian product. The probability of
an outcome is the product of the probabilities of its coordinates in the factors. For
example, when flipping a coin n times independently, each flip has outcome head
or tail, each with probability 1/2. The probability of a given list is 2−n .
Making a choice uniformly at random means that the possible outcomes are
equally likely. When flipping a coin, the flips are generated uniformly at random
(probability 1/2 of each outcome) and independently.
Studying conditional probability has the effect of normalizing or restricting
the space to the points within the event B. The intuitive idea is that the condi-
tional probability (given that B occurs) is the fraction of B where A also occurs.
The probability of a joint event can be computed as a product of conditional
probabilities. For example,
(⋂¾n=1 Ai) = (An | ⋂¾n=−11 Ai) (An−1 | ⋂¾n=−12 Ai) · · · (A2 | A1) (A1).

In the last half of the 20th century, advanced techniques for studying prob-
ability spaces found many applications to difficult combinatorial problems. Com-
binatorial techniques show that a “good enough” object exists by constructing it,
but probabilistic methods are nonconstructive. An object with a desired property
must exist when the probability of that property is nonzero in an appropriate
probability space. Similarly, one can show the existence of an object with a large
value of a parameter X by showing that the expected value of X is that large
when the objects are randomly generated.

0.12. Definition. A random variable on a discrete probability space S is a func-

tion X : S → . It is discrete when the range is finite or countable, often 0 . 

Let X = ¾ denote the event {a ∈ S: X(a) = ¾}, and write (X = ¾) for its prob-

ability. The expectation or expected value (X) of X is ∑ a∈S X(a) (a), 
when this sum converges. When X is a discrete random variable, we write
 ∞

this as (X) = ∑¾=0 ¾ · (X = ¾). The pigeonhole property of the expec-
tation is the statement that there is an element of the probability space for
which the value of X is as large as (or as small as) (X). 
Using the pigeonhole property requires a value or bound for (X). Often 
the computation applies the “linearity of expectation” to an expression for X as
a sum of simpler random variables. We restrict our attention to sums of finitely
8 Chapter 0: Introduction

many random variables on discrete probability spaces. Analogous results hold in

continuous probability spaces.

0.13. Lemma. (Linearity of Expectation) If X and X 1 , . . . , X ¾ are random

variables on the same space such that X = ∑ X i , then (X) = ∑ (X i). Also
(cX) = c (X) for any constant c.
Proof: In a discrete probability space, each sample point contributes the same
amount to each side of each of these equations.

One reason for the influence of probabilistic methods is that exact counts are
both too difficult and unnecessary in large structures. Somehow the probabilistic
methods capture the most important aspects or most dominant terms. Probabilis-
tic methods are especially effective for extremal problems when it turns out that
“most ” instances are near the optimum.

OTHER DISCRETE STRUCTURES

In addition to subsets, permutations, and graphs, many other structures are

used to model combinatorial problems. We briefly describe several studied in this
book in order to suggest the scope of the text. The reader should treat this section
lightly; precise definitions and examples will be given later.

0.14. Example. Digraphs and multigraphs. General binary relations are mod-
eled using directed graphs (digraphs); these differ from graphs in that the
edges are ordered pairs of vertices. In an edge from x to y, the first vertex is the
tail and the second is the head. We illustrate directed graphs by drawings that
use arrows (curves with direction) for the edges.
We sometimes allow edges in digraphs to be pairs of the form (x , x), called
loops. For example, the functional digraph of a function : A → A has vertex
set A; its edges are the pairs (x , (x)) for x ∈ A. The fixed points of become loops
in the functional digraph. A permutation is a bijection  : A → A; its functional
digraph consists of disjoint (directed) cycles corresponding to its orbits.
We may modify a graph or digraph both by allowing loops (one-vertex edges)
and by allowing more than one edge with the same endpoints. Here E(G) becomes
a multiset. The resulting model is a multigraph or multidigraph.

0.15. Example. Order relations. The containment relation on a family of sets

is a reflexive relation that is transitive but not symmetric and hence is not an
equivalence relation. It is a fundamental example of another important type of
relation. A relation R is antisymmetric if (x , y) ∈ R and (y , x) ∈ R imply x = y.
A relation is an order relation if it is reflexive, antisymmetric, and transitive.
Besides containment, other order relations include the divisibility relation
on the set of positive integers and the componentwise order on n , which contains
the pair (u , v) when ui ≤ vi for all i. Other examples arise in scheduling: events
occupy some interval in time, and we say that A “precedes” B if A ends before
B begins. This defines an order relation on any set of intervals. We study order
relations in Part III.
Background Definitions 9

0.16. Example. Hypergraphs. Another generalization of graphs allows edges of

arbitrary size. A hypergraph G consists of a set V(G) of vertices and a set E(G)
of edges, where an edge can be any subset of V(G). A ¾-uniform hypergraph
is a hypergraph whose edges all have size ¾ ; graphs are simply 2-uniform hyper-
graphs. A hypergraph is regular if every vertex is in the same number of edges.
Since the edge set of a hypergraph with vertex set X is a family of subsets of X , hy-
pergraphs can be studied from set-theoretic, order-theoretic, and graph-theoretic
viewpoints. They will be helpful in Parts III and IV.

0.17. Example. Designs and projective planes. In Chapter 13 we study a special

type of regular uniform hypergraph that has applications to design of experi-
ments and to extremal combinatorial problems. A block design is a regular uni-
form hypergraph in which any two points (vertices) appear together in the same
number of blocks (edges). Equivalently, any two rows in the incidence matrix for
the membership relation of elements in blocks have the same dot product.
When any two points lie in exactly one common block, we may request also
that any two blocks have exactly one common point. The resulting configurations
are projective planes, in which the blocks are called lines. Projective planes can
be obtained from finite fields and correspond to special families of latin squares. A
Latin square of order n is an arrangement of symbols of n types in an n-by-n ma-
trix so that each symbol appears exactly once in each row and in each column. A
classical application is the assignment of types of fertilizers and seeds to regions
in an agricultural plot to reduce the effect of soil differences.

0.18. Example. Matroids. In Chapter 11 we study another special structure.

Matroids can be viewed as special families of subsets of a set. This interprets
them as a special type of hypergraph. We postpone the precise definition and
observe merely that the matroid context permits common generalizations of fun-
damental results in graph theory, linear algebra, and the theory of ordered sets.
For example, matroids permit a natural generalization of the result that a span-
ning tree of minimum total weight in a connected graph with weighted edges
can be found by iteratively including the cheapest edge that does not form a cycle
with edges already chosen. Much of the elementary theory of dimension in linear
algebra also arises as a special case of matroid properties.

COMPLEXITY

The growth of combinatorics has been stimulated by computer science, which

studies the computational aspects of discrete mathematics. We will comment oc-
casionally on the computational complexity of problems. A simple measure of the
performance of an algorithm is its worst-case running time, as a function of the
size of the input. A problem is efficiently solved if it has a solution algorithm whose
running time is bounded by a polynomial in the size of the input.
The size of the input is its length in bits in some encoding of the problem. For
our purposes, natural parameters such as the order of a matrix or the number of
vertices suffice to measure size. A polynomial in n is bounded by a polynomial in
10 Chapter 0: Introduction

n2 or n3 , so the manner of encoding input is unimportant unless the problem has

exponentially large numbers as input data.
Complexity considers asymptotic growth rates. The set of functions whose
magnitude is bounded above by a constant multiple of (for sufficiently large ar-
guments) is called O( ). Several pertinent sets of functions arise when comparing
growth rates to , as listed below.
o( ) = { : |(x)|/| (x)| → 0}
O( ) = { : ∃c , a ∈ such that |(x)| ≤ |c (x)| for x > a}
( ) = { : ∃c , a ∈ such that |(x)| ≥ |c (x)| for x > a}
( ) = { : |(x)|/| (x)| → ∞}
( ) = O( ) ∩ ( )
Derbyshire [2003] attributed “Big Oh notation” to Landau [1909], but Landau
[1909, p. 883] borrowed it from Bachmann [1894].
Properly speaking, o( ), O( ), ( ), ( ), and ( ) are sets, and it is cor-
rect to write  ∈ O( ) when describing the growth rate of  . Mathematicians and
computer scientists routinely write (n) = O( (n)) to mean  ∈ O( ) (see Knuth
[1976a]). Some avoid this by writing “ (n) is O( (n))”, treating O( ) as an adjec-
tive. We can compute with members of O( ) somewhat as we do with congruence
classes, but we have no symbol like “ ≡ ” for doing this. Thus in this book we
sometimes use expressions like (n) = n2 + O(n3/2) (meaning (n) − n2 ∈ O(n3/2)),
but where the grammar is appropriate we use the membership symbol. The ex-
pression (n) ∼ (n) means that  is asymptotic to , which can be written as
(n) = 1 or 
limn→∞ º½(n) (n) = (n)(1 + o(1)).
Complexity classes are studied using decision problems that have yes/no
answers, such as “does the input graph have a spanning cycle?” Optimization
problems (such as “what is the maximum length of a cycle in this graph?”) can be
solved using successive decision problems, such as “does this graph have a cycle
of length at least  ?”, where  is part of the input. The class of decision problems
solvable by a worst-case polynomial-time algorithm (polynomial in the size of the
input) is called “P ”.
Many decision problems have no known polynomial-time solution algorithm
but have a polynomial-time algorithm for verifying a YES answer. For example,
existence of a spanning path can be verified by giving the order of vertices on such
a path and checking that successive vertices are adjacent. When checking all pos-
sible permutations in parallel, each computation path is short. It is verifying a
NO answer that is difficult.
A deterministic algorithm follows only one computation path on a given in-
put. A nondeterministic algorithm follows multiple computation paths simul-
taneously. In the example above, given an input graph, such an algorithm checks
all possible vertex orderings simultaneously to seek a spanning path; each order-
ing can be checked in polynomial time. A nondeterministic polynomial-time
algorithm follows one computation path for each way of specifying a polynomial-
length stream of bits, with each such computation running in polynomial time.
The bit stream is not the input to the problem; the bits specify an option to con-
sider, which in the example here is a vertex ordering.
A nondeterministic algorithm solves a decision problem if for every input I,
the answer to the problem on I is YES if and only if the algorithm applied to I has
Background Definitions 11

at least one computation path that returns YES. The class of decision problems
having nondeterministic polynomial-time solution algorithms is called “NP ”. Be-
cause a machine having the power to follow many computation paths in parallel
can also follow one, P ⊆ NP.
Most computer scientists believe P
= NP. In this context, a problem B
is as hard as a problem A if a polynomial-time algorithm for B would yield a
polynomial-time algorithm for A. This may involve using B as a subroutine in an
algorithm to solve A or providing a polynomial-time transformation that converts
an arbitrary instance of A into an instance of B such that the answer in B is YES
if and only if the answer in A was YES. In either case, an algorithm for B yields
an algorithm for A; we call this a reduction of A to B. A problem is NP-hard if
it is as hard as every problem in NP in this sense (every problem in NP reduces
to it). A problem belonging to NP is NP-complete if it is NP-hard.
Cook [1971] devised a generic transformation to reduce any problem in NP to
that of deciding whether an input logical formula is true under some truth assign-
ment for its variables. This problem is called SATISFIABILITY or SAT. Cook’s
result made SAT the first known NP-complete problem. To prove that a problem
B is NP-hard, we can reduce SAT to B. Every problem proved NP-complete can
then be used like SAT in this way. In practice, the known NP-complete problem in
most NP-completeness proofs is one of a few fundamental NP-complete problems,
such as those in K arp [1972].
A polynomial-time algorithm for any NP-complete problem could be used to
construct a polynomial-time algorithm for each problem in NP, yielding P =NP.
The conjecture that P
=NP is supported by the failure to find a polynomial-time
algorithm for any problem in the large class of NP-complete problems, despite
years of search.
Garey–Johnson [1979] gives a thorough and readable introduction to NP-
completeness. Nowadays, an NP-completeness proof is only a beginning. One
seeks to refine the boundary between P and NP by finding polynomial-time solu-
tion algorithms for large classes of inputs or by finding restricted classes of inputs
where the problem remains NP-hard.
Mostly we use NP-completeness as motivation. NP-completeness justifies the
study of heuristics that run quickly but do not obtain the optimal solution. It
also motivates extremal problems: we study bounds on optimization problems in
terms of various parameters of the input. A constructive proof of a bound for
graphs in a particular class yields an algorithm that may approximate the true
value of the parameter.

We hope that these few basic concepts suggest some of the underlying struc-
ture of this large subject. More thorough definitions and examples are given
where the concepts are explored in depth. With this as background, we are ready
to begin.
Chapter 1

Combinatorial Arguments
Enumerative questions are among the most natural in mathematics and have
been studied for eons. A simple way to count a set is to list its elements, but this
is inelegant and often impractical. We may view the study of enumeration as the
search for ways to avoid exhaustive listing.
In this chapter we use combinatorial arguments. We give no precise definition
of this term, but generally it refers to explicit counting arguments, “counting two
ways”, or the use of bijections to show that two sets have the same size.
Later chapters develop more sophisticated techniques. Combinatorial argu-
ments are elementary but can be hard to find. Studying them is worthwhile,
because elementary counting problems often serve as models or building blocks
in harder problems, and combinatorial arguments can yield more information.
Because we assume no prior experience in combinatorics, many results and
examples here appear also in undergraduate courses. Often we go farther, and
our treatment is more concise and assumes more mathematical maturity.
To avoid excessive formality, we may omit explicit declaration of variables
when the context is clear. The universe is the set of values where the resulting
objects make sense. This may happen with the set  of natural numbers or the

set 0 of nonnegative integers. Examples include “for all n-vertex graphs” and
“for all odd n”. A similar convention sometimes omits the limits on summations.

What does it mean to solve a counting problem? The problem may be ex-
pressed in terms of a parameter (variable), and we want to solve it simultaneously
for each value of the parameter. With one parameter n, this yields a sequence ⟨a⟩
consisting of the values {an}∞ n=0 . We may seek various expressions for the answer.
A formula for an expresses it as a function of the parameter(s). Sometimes
we accept a finite sum as the answer. This is common when we use the inclusion-
exclusion principle (Section 4.1), a technique that alternately overcounts and un-
dercounts the desired set until each desired element is counted exactly once. An
asymptotic formula (Section 2.3) for an is a formula ½(n) such that ½(n)/an → 1 as
n → ∞. Computational considerations make asymptotic analysis important, and
asymptotic formulas may be preferable when exact solutions are too complicated.
An effective and fast procedure for computing individual terms or asymptotic
behavior may be a good answer. A recurrence relation (Chapter 2) for ⟨a⟩ expresses
an as a function of a0 , . . . , an−1 ; appropriate initial values are needed to specify
⟨a⟩ completely. Solving a recurrence means obtaining a formula for an .

13
14 Chapter 1: Combinatorial Arguments

A generating function (Chapter 3) is a “formal” power series that encodes the

sequence ⟨a⟩ by using the terms as coefficients. The simplest form is the ordinary
∞
generating function ∑n=0 an x n . By “formal”, we mean that the powers of x are
placeholders for the terms in the sequence; we generally do not treat x as a num-
ber. Manipulating expressions for the generating function can lead to an explicit
or asymptotic formula for an .

1.1. Classical Models

In this section we use combinatorial arguments to develop elementary mod-
els used in solving many counting problems. First we present several elementary
techniques as “Principles”; the trivial examples here are just a warmup to illu-
minate a way of thinking.

ELEMENTARY PRINCIPLES

Our first two Principles are sometimes called “ad hoc” counting techniques.
Nothing sophisticated is involved; we merely organize the set to simplify counting
it, breaking the problem into smaller pieces.

1.1.1. Sum Principle: If a finite set A is partitioned into sets B1 , . . . , B¾ , then

| A| = ∑¾i=1 | Bi |.
Recall that a partition of A is a family of disjoint nonempty sets whose union is
A. This principle applies when a counting problem is broken into cases and is also
called Counting by Cases.

1.1.2. Product Principle: If elements of a set A are built by successive choices

and the number of options for the ith choice is independent of the outcomes of
earlier choices, then | A| is the product of the numbers of options for the choices.
This principle is also called Counting by Stages. The actual options at the ith
stage may depend on earlier choices, but the number of options does not. The for-
mula | S × T | = | S| | T | for counting a cartesian product is a simple application of
this principle.

n−1
1.1.3. Example. There are ∏i=0 (2n − 1 − 2i) ways to pair 2n people. The first
person can be paired with another in 2n − 1 ways. No matter how this choice
is made, we can take the least indexed unpaired person and choose a partner for
that person from the remaining people in 2n − 3 ways. Continuing through n
n−1
stages in this way produces ∏i=0 (2n − 1 − 2i) distinct pairings, and every pairing
is produced in this way. The 15 pairings for n = 3 appear below.

12 · 34 · 56 13 · 24 · 56 14 · 23 · 56 15 · 23 · 46 16 · 23 · 45
12 · 35 · 46 13 · 25 · 46 14 · 25 · 36 15 · 24 · 36 16 · 24 · 35
12 · 36 · 45 13 · 26 · 45 14 · 26 · 35 15 · 26 · 34 16 · 25 · 34
Section 1.1: Classical Models 15

1.1.4. Principle of Counting Two Ways: When two formulas count the same set,
their values are equal.
Hermann Weyl once described the property of getting the same answer no matter
how we count a finite set as one of the deepest theorems in mathematics. The
depth is not in its truth but rather in its myriad applications. The point is not
that we need to count a set twice; it is that we can prove equality of two formulas
by devising an appropriate set whose size is given by both formulas, counted in
different ways.

n−1
1.1.5. Example. ∑i=1 i = n(n − 1)/2. This formula is easily proved by induction
on n, but showing that both sides count the pairs of elements in [n] provides more
insight (recall from Chapter 0 that [n] = {1 , . . . , n}). The left side groups the
pairs by the larger element, applying the Sum Principle (i pairs have larger ele-
ment i + 1). The right side applies the Product Principle by picking two elements
successively and then dividing by 2 since order is irrelevant and we have counted
each pair exactly twice.
Interchanging the order of summation. This operation can be viewed as an
instance of counting two ways. Think of the value º (i , j) as the entry in row
i and column j of a matrix. The two sides of the identity ∑i=1 ∑ j =1 º (i , j) =
m n

∑ j =1 ∑i=1 º (i , j) sum the entries in the matrix, by rows or by columns.

n m

1.1.6. Bijection Principle: If there is a bijection from one set to another, then they
have the same size.
Recall that a bijection is a function º : A → B such that for all b ∈ B exactly
one x ∈ A satisfies º (x) = b. The Bijection Principle is just the definition of equal
size, but we use it as a counting technique. Establishing a bijection º from a set
A of unknown size to a set B of known size computes | A| .
We use bijections when modeling part of a new problem as an instance of a
classical problem. A bijection maps instances of the new problem into instances of
the previously solved problem. When there is a bijection from A to B, we say that
objects of A “correspond” to objects of B. A bijection is a function from one set to
another, but a one-to-one correspondence between two sets puts them in pairs and
can be followed in either direction. The latter term is less formal.

1.1.7. Example. Within the set [n], the number of subsets with ¾ elements equals
the number of subsets with n − ¾ elements. Complementation of subsets within [n]
provides a bijection from one family to the other.
The number of binary lists of length n equals the number of subsets of [n]. Map
any subset A of [n] to the list a defined by ai = 1 if i ∈ A and ai = 0 if i ∈ / A.
The list a is the incidence vector of A. Any binary list a is the image of exactly
one subset of [n], namely the set {i: ai = 1}. Hence the map is a bijection. This
bijection enables us to view subsets as incidence vectors and vice versa.

Proofs using counting two ways or bijections are examples of combinatorial

arguments. Asking for a combinatorial proof or bijective proof usually means
asking for a proof using one of these techniques.
16 Chapter 1: Combinatorial Arguments

1.1.8. Pigeonhole Principle: The maximum in a set of numbers is at least as

large as the average (and the minimum is at least as small).
In particular, placing more than ¾ n objects into n boxes puts more than ¾ ob-
jects into some box. This is not directly a counting technique, but it often helps
in applications of counting formulas. It is another simple statement with subtle
applications (for a precise proof, consider the contrapositive). We discuss appli-
cations and deep generalizations of the Pigeonhole Principle in Chapter 10, but
the simple form will often be useful before that.

1.1.9. Example. If A and B are finite and º : A → B and ½ : B → A are injections,

then | A| = | B|, and º and ½ are bijections. If | B| < | A| , then º puts more than
| B| objects into | B| boxes, but such functions are not injective. Hence | B| ≥ | A|.
Similarly, ½ yields | A| ≥ | B| , so | A| = | B| .
If some x ∈ B lies outside the image of º , then º has at most | A| − 1 elements
in its image, since we already have | B| = | A|. Again the Pigeonhole Principle
contradicts the injectivity of º . Thus º (and similarly ½) is surjective.

When A and B are infinite, and º : A → B and ½ : B → A are injections, it no

longer follows that º and ½ are bijections, but still A and B must have the same
cardinality (Exercise 47).

1.1.10. Polynomial Principle: If two polynomials in one variable are equal at

all positive integer values, then they are the same polynomial, meaning they are
equal for all values of the variables and have the same coefficients.
A polynomial of degree d is 0 at no more than d points. Hence polynomials of
degree at most d that agree at d + 1 points also agree everywhere (consider their
difference). A generalization for polynomials in more variables holds by induction
on the number of variables; see Exercise 33. As in the one-variable case, equality
at infinitely many points is not needed, but we usually invoke the principle when
giving an argument that proves equality at all choices of positive integers.
Our first application of this principle (using two variables) is Proposition
1.1.18; see also Exercises 34–35.

1.1.11. Remark. Counting problems may arise from questions of probability. In

a finite “sample space” U of equally likely outcomes, an event is a subset A of
U. The probability of A, written (A), is then defined as | A| / |U |. The prob-
ability of any property is the probability of the event consisting of all outcomes
with that property (see Chapter 0 for further probabilistic concepts). In Chapter
14 we consider more sophisticated applications of probability, but meanwhile we
occasionally use probabilistic questions to motivate counting problems.

WORDS, SETS, AND MULTISETS

In using bijective proofs to count sets, we need canonical sets of known sizes.
Modeling a piece of a counting problem as an instance of one of these classical
problems is a combinatorial argument.
Section 1.1: Classical Models 17

1.1.12. Definition. A ¾-word or word of length ¾ is a list of ¾ elements from

a given set (the alphabet). A simple word is a word whose letters are dis-
tinct. A ¾-set is a set with ¾ elements; a ¾-set in a set S is a subset of S with ¾
elements. We use (n¾ ), read “ n choose ¾ ”, to denote the number of ¾-sets in an
n-set. A multiset from a set S is a selection from S with repetition allowed.

Ways to have ¾ items from an n-set

no repetition repetition allowed
ordered simple ¾-words ¾-words/lists
(arrangements) (¾-permutations)
unordered subsets of size ¾ multisets of size ¾
(selections) (¾-combinations)

Words are simply lists; the term “word” is natural when we emphasize a par-
ticular finite alphabet from which they are formed. In contrast to sets, entries
in a list have specified positions. Also sets (and simple words) have no repeated
elements, while lists and multisets allow repetitions. The table above also lists
older equivalent terms parenthetically.
The easiest of these four counting problems is for ¾-words. A list of length ¾
can be viewed as a function º defined on [¾]. The function maps [¾] into the set
of entries allowed in the list, with º (i) being the entry in position i.

1.1.13. Proposition. There are n¾ words of length ¾ from an alphabet S of size

n. Equivalently, there are n¾ functions from [¾] to S.
Proof: For the entry in each position there are n choices, regardless of the earlier
choices made. The Product Principle applies.

1.1.14. Remark. Subsets as functions and lists. Subsets of [¾] correspond to func-
tions from [¾] to {0 , 1}; there are 2¾ . Given A ⊆ [¾], let º A(i) = 1 if i ∈ A and
º A(i) = 0 if i ∈/ A. The function º A is the characteristic function of A (or, infor-
mally, the membership function). Since the domain of º A is [¾], we can write
º A as a list of length ¾ with entries in {0 , 1}, as described in Example 1.1.7.
Before solving the counting problems with repetition forbidden, we introduce
notation for certain products.


1.1.15. Definition. For n , ¾ ∈ 0 , we define the following products:
n−1
n factorial: n! = ∏i=0 (n − i).
¾ −1
falling factorial: n(¾) = ∏i=0 (n − i).
¾ −1
rising factorial: n(¾) = ∏i=0 (n + i).

By convention, the result of an operation over an empty set is the identity

element for that operation, so 0! = 1 and n(0) = n(0) = 1. Since outer parenthe-
ses are not needed around a superscript or an exponent, their presence should
have a special meaning as in our definition here. Many conflicting notations
have been used for rising and falling factorials, as discussed at “MathWorld”
18 Chapter 1: Combinatorial Arguments

(http://mathworld.wolfram.com/) and “PlanetMath” (http://planetmath.org/).

Our notation was used long ago in published notes of a summer lecture course by
C. L. Liu [1972].
A simple word of length n from the set [n] writes the elements of [n] in some
order. Hence it is a permutation of [n]. Later we will discuss other ways of repre-
senting a permutation; this way is the word form. More generally, we may use
only ¾ elements.

1.1.16. Proposition. There are n(¾) simple ¾-words from an alphabet of size n.
Proof: We form a simple ¾-word from a set of size n one position at a time. Since
repetitions are forbidden, there are always n+1 − i ways to specify the ith element,
no matter how the earlier elements were specified. By the Product Principle, the
¾
word can be formed in ∏i=1 (n + 1 − i) ways.

For the number of ¾-subsets, notations used in the past instead of (¾n) include
C(n , ¾), Cn ,¾ , ¾ Cn , n C¾ , and Cn¾ ; these are obsolete and now have other meanings.

1.1.17. Proposition. The number (¾n) of ¾-sets in an n-set satisfies

n n! n( )
( )= = ¾ .
¾ ¾ !(n − ¾)! ¾!

Proof: In addition to the construction of simple ¾-words described in Proposition

1.1.16, we can also form a simple ¾-word by choosing a ¾-set and then placing the
elements in order. For each ¾-set, there are ¾ ! ways to permute the elements. By
counting the simple ¾-words in these two ways, we have ¾ !(¾n) = n(¾) , which yields
the claimed formula.

The numbers (¾n) are called binomial coefficients due to their role in the
Binomial Theorem. Our combinatorial definition of the symbol (¾n) makes it 0
unless 0 ≤ ¾ ≤ n. Hence we need not worry about limits on the summation in the
Binomial Theorem and many similar sums.

1.1.18. Proposition. (Binomial Theorem) For n ∈ 0 ,

n
(x + y)n = ∑ ( )x ¾ y n−¾ .
¾
¾

Proof 1: Expanding (x + y) · · · (x + y) yields a sum of many terms. Each corre-

sponds to a choice of x or y from each of the n factors. Terms of the form x¾ y n−¾
arise by choosing exactly ¾ of the n distinct factors to contribute x.
Proof 2: When x and y are integers, (x + y)n counts the n-words from an alphabet
S of size x + y (Proposition 1.1.13). We also count them by the number of positions
using the first x letters in S. There are (¾n)x¾ y n−¾ ways to choose ¾ such positions,
fill them, and fill the remaining positions (Product Principle). Summing over ¾
counts all the words (Sum Principle). Hence we have counted a set in two ways,
making the two sides equal whenever x and y are integers. By the Polynomial
Principle, the two polynomials are equal.
Section 1.1: Classical Models 19

Choosing ¾ objects from n types with no limit on repetitions produces a ¾-

element multiset from [n] (Definition 1.1.12). We care only how many objects are
chosen of each type; the objects have no “positions”. Thus each multiset corre-
sponds to a vector of multiplicities for each type of object. To count multisets, we
count these vectors of nonnegative integers.

1.1.19. Example. The integer-sum problem. Let x1 , . . . , x n be variables, with x i

representing the number of copies of i chosen in a ¾-element multiset from [n].
The number of such multisets is the same as the number of solutions to ∑i=1 x i =
n

¾ in nonnegative integers.
The solution vectors for (n , ¾) = (3 , 4) appear below.
(4,0,0) (3,1,0) (1,3,0) (2,2,0) (2,1,1)
(0,4,0) (0,3,1) (0,1,3) (0,2,2) (1,2,1)
(0,0,4) (1,0,3) (3,0,1) (2,0,2) (1,1,2)

1.1.20. Theorem. The number of ¾-element multisets from [n], or equivalently

the number of solutions to ∑i=1 x i = ¾ in nonnegative integers, is (¾+n−n−1 1 ),
n

which also equals (¾+n¾−1 ).

Proof: We model the integer-sum problem using subsets. For a nonnegative-
integer solution, form a list of ¾ dots and n − 1 bars, using x1 dots, then one
bar, then x2 dots, then one bar, etc. The integer variables can be retrieved from
such a list by letting x i count the dots between the (i − 1)th and the ith bars,
where we imagine a 0th bar before the list and an nth bar after it.

5 2 0 3
• • • • • | • • | | • ••

Since we can retrieve uniquely from each arrangement of dots and bars the
list of integers that yields it, we have established a bijection from the set of se-
lections with repetition to the set of lists of ¾ dots and n − 1 bars. The number of
solutions is thus (¾+n−n−1 1 ) (or (¾+n¾ −1 )), because the lists of dots and bars are specified
by choosing the positions for the bars (or the dots).

The alternative formulas (¾+n−n−1 1 ) and (¾+n¾−1 ) can be confusing, since the
model may arise with different (or reversed!) variables counting the objects and
the types. Thus it is more valuable to understand the argument than to memorize
the formula. The first formula emphasizes that the number of barriers between
types is one less than the number of types. The second emphasizes that ¾ is the
size of the multiset. What happens when we require x1 , . . . , x n to be positive?

1.1.21. Definition. A composition of an integer ¾ is a list of positive integers

summing to ¾ . The entries in the list are the parts of the composition.

The last column of the listing in Example 1.1.19 shows the three compositions
of 4 with three parts.
20 Chapter 1: Combinatorial Arguments

1.1.22. Corollary. There are (¾n−−11) compositions of ¾ with n parts.

Proof 1: We count the solutions of ∑ i=1 x i = ¾ in positive integers by reducing
n

it to a problem of counting nonnegative integer solutions. Solutions of ∑i=1 x i =

n

¾ in positive integers correspond to solutions of ∑i=1 yi = ¾ − n in nonnegative

n

integers. Thus the number of solutions is (¾−nn+ n−1 ¾ −1

−1 ), which simplifies to ( n−1).
(1,0,0) (0,1,0) (0,0,1)
(2,1,1) (1,2,1) (1,1,2)
• • | • |• • | • • |• • | • | • •

Proof 2: Consider the model of dots and bars. Bars split the row of ¾ dots into
segments, which become parts. To make each segment have at least one dot, no
two bars can be consecutive. Thus we select distinct places for the n − 1 bars from
the ¾ − 1 spaces between dots.

Discovering Proof 2 above after simplifying the computation in Proof 1 can

be considered our first illustration of the Erdős Principle: If there is a simple
answer, then there should be a simple proof!

EXERCISES 1.1
Problems marked “(−)” are easier or shorter than most (given the material of the text),
while those marked “(+)” are harder than most. Problems marked “(♦)” are particularly
valuable or instructive, not necessarily harder.
1.1.1. (−) When rolling n dice, what is the probability that the sum is even?
1.1.2. (−) Count all the rectangles with positive area formed by segments in a grid of m
horizontal lines and n vertical lines.
1.1.3. (−) The roman alphabet has 21 consonants and 5 vowels. How many strings can be
formed using r consonants and s vowels?
1.1.4. (−) Count the possible outcomes of an election with 30 voters and four candidates.
Count those in which no candidate receives more than half of the votes.
1.1.5. (−) Prove that (n5 − 5n3 + 4n)/120 is an integer for every n ∈ .
1.1.6. (−) Count the orderings of a standard 52-card deck such that the 13 cards in the
spade suit occur consecutively.
1.1.7. (−) Compute the probability that a random set of five cards from a standard 52-card
deck has at least three cards with the same rank.
1.1.8. (−) Count the sets of six cards from a standard deck of 52 cards that have at least
one card in every suit.
1.1.9. (−) Count the integers from 0 through 99,999 such that each digit occurs at most
twice; leading zeros are counted as appearances of 0.
1.1.10. (−) In how many distinguishable ways can the letters in “Mississippi” be ordered?
1.1.11. (−) Given a bag with many marbles in each of four colors, count the ways to select
12 marbles. Count the distinguishable ways to put 12 of them in a row.
Exercises for Section 1.1 21

1.1.12. (−) New York City has about 7 million residents; suppose that each has 100 coins
in a jar. The coins come in five types (pennies, nickels, dimes, quarters, half-dollars). Two
jars are equivalent if they have the same number of coins of each type. Is it possible that
no two people have equivalent jars?
1.1.13. (−) Count the compositions of ¾ in which every part is even.
1.1.14. (♦) Families of subsets.
(a) Count the subsets of [n] that contain at least one odd number.
(b) Count the ¾-sets in [n] having no two consecutive integers.
(c) Count the lists of subsets A0 , A1 . . . A n of [n] such that A0 ⊂ A1 ⊂ · · · ⊂ A n . Count
the lists such that A0 ⊆ A1 ⊆ · · · ⊆ A n .
1.1.15. (♦) Prove that the exponent on a prime p in the prime factorization of (2n n
) is the
number of powers p¾ of p such that ⌊ 2n/p¾ ⌋ is odd. Use this to determine which primes di-
vide (18
9
) and which divide (10
20
). For example, for n = 3, (63) = 20 and 20 = 22 · 5. The claim
holds here, because ⌊ 6/2⌋ and ⌊ 6/4⌋ are odd, ⌊ 6/3⌋ is even, ⌊ 6/5⌋ is odd, and ⌊ 6/p¾ ⌋ = 0
for all other prime powers p¾ .
1.1.16. Given positive integers a and b, let v(a , b) = ((b−a1) , (ab) , (b+a1 )) . Prove that v(c , d)
cannot be a multiple of v(a , b) when (a , b) and (c , d) are distinct pairs of positive integers.
1.1.17. (♦) Count the lists of m 1s and n 0s that have exactly ¾ runs of 1s, where a run is
a maximal set of consecutive entries with the same value.
1.1.18. (♦) For S ⊆ [n], a run in S is a maximal set of consecutive integers in S (corre-
sponding to a run of 1s in the incidence vector). For example, {1 , 2 , 4 , 5 , 6 , 9} is a subset
of [9] with three runs; every element is in one run.
(a) Count the subsets of [n] having ¾ runs.
(b) Count the t-element subsets of [n] having ¾ runs.
(c) Let s1 , . . . , sm be distinct positive integers. Among the t-element subsets of [n] hav-
ing ¾ runs, count those having exactly ri runs with length si , for 1 ≤ i ≤ m. (Comment:
Note that ¾ = ∑i=1 ri and t = ∑i=1 ri si .)
m m

1.1.19. Determine the number of binary strings of length n in which the number of copies
of 00 is the same as the number of copies of 11. For example, 00011011 has two copies of
each. (Deutsch [2009])
1.1.20. Give a summation for the number of elements of [3]n with ¾ odd entries that do
not have a 1 next to a 3.
1.1.21. The chords of a convex n-gon are the segments that join two corners. Count the
pairs of chords that cross inside the n-gon.
• • •
• • •
• •
• •
• • • • • • • •

1.1.22. Count all triangles (not only the empty triangular regions) in the picture formed
by drawing all (n2) chords of a convex n-gon, given that no three chords have a common
internal point. When n is 4, 5, or 6, the answers are 8, 35, and 111.
1.1.23. We have six dice. Each has three red faces, two green faces, and one blue face.
Determine the probability that three red faces, two green faces, and one blue face appear
when the six dice are rolled.
22 Chapter 1: Combinatorial Arguments

1.1.24. Given a standard deck of playing cards with 52 cards in four suits of 13 values,
which is more likely in a random set of five cards: having all five cards in the same suit , or
having cards of five consecutive values?
1.1.25. A trapezoid is a quadrilateral with at least one pair of parallel sides. For n ≥ 4,
how many sets of four distinct points forming the vertices of a trapezoid exist among the
vertices of a regular n-gon? (Nicoaescu [1986])
1.1.26. (♦) For a permutation of [n], the displacement of is ∑i=1 |i − (i)| . Prove
n

that the largest displacement of a permutation of [n] is ⌊ n2/2⌋ . (Comment: The extremal
permutation is not unique when n is odd and at least 3.)
1.1.27. (♦) Let A n be the set of permutations of [n]. Let Bn be the set of n-tuples (b1 , . . . , bn)
such that 1 ≤ bi ≤ i for each i ∈ [n]. Construct a bijection from A n to Bn . Below we illus-
trate a possible correspondence for n = 3.
A3 321 231 213 312 132 123
B3 111 112 113 121 122 123

1.1.28. Using interchanges of two students, a professor wants to rearrange the n students
sitting in a row so that no two students originally adjacent remain adjacent. Determine
the minimum number of interchanges needed, given n ≥ 6. (W. So)
1.1.29. Count the 0 , 1-matrices with n2 rows and n2 columns such that (1) each row and
column has exactly one 1, and (2) when the matrix is partitioned into n2 blocks of n con-
secutive rows and n consecutive columns, each block contains exactly one 1. For example,
when n = 2 the answer is 16. (Pratt [2011])
1.1.30. (♦) Count the permutations of [n] with (i + 1) ≤ (i) + 1 for 1 ≤ i ≤ n − 1. Give
both a proof by induction and a bijective proof. (Deutsch [2001a])
1.1.31. (+) A permutation is graceful if the absolute differences between successive ele-
ments are distinct. Prove that if the set of elements in even-indexed positions of a grace-
ful permutation of [2n] is [n], then the first and last elements differ by n. For example,
54637281 is such a permutation. (Comment: The converse also holds.) (Klove [1995])
1.1.32. A necklace is a circular arrangement that can rotate and flip without being con-
sidered different. Thus only one necklace can be made from three beads. A crown, in
contrast , can rotate but not flip.
(a) Count the necklaces with n beads that can be made from n distinct beads.
(b) Count the crowns with n beads that can be made from a supply of  types of beads,
given that n is prime (the  types need not all be used).
1.1.33. Polynomial Principle in several variables. Let p be a polynomial in x1 , . . . , x¾ . For
1 ≤ i ≤  , suppose that the degree of p as a polynomial in xi is at most d i , and let Si be a set
¾
of d i + 1 values. Prove that if p is 0 at all points in ∏i=1 Si (that is, on {(x1 , . . . , x¾): xi ∈
Si for 1 ≤ i ≤ }), then p is zero everywhere. Conclude that if two polynomials in several
variables are equal at all positive integer values, then they are equal at all values. (Hint:
Assume the well-known statement that a polynomial of degree at most d in one variable
is identically zero if it is zero at more than d values.)
1.1.34. (♦) Give a combinatorial proof that the identity below for falling factorials holds
for all x , y ∈ .

n
(x + y)(n) = ∑ ( ) x(¾) y(n−¾) .

¾
Exercises for Section 1.1 23

1.1.35. (♦) Flags on poles.

(a) Obtain a simple formula for the number of ways to put m distinct flags on a row
of r flagpoles. Poles may be empty, and changing the order of flags on a pole changes the
arrangement. The formula must only use one “ m” and one “r”. (The answer is 6 for m =
r = 2, as shown below.)
(b) Prove that the identity below for rising factorials holds for all x , y ∈ .

n
(x + y)(n) = ∑ ( ) x(¾) y(n−¾) .
¾
¾

2 1 2 1

1 2 1 2 2 1 1 2

A B A B A B A B A B A B

1.1.36. Given n distinct flags (n ≥ 1) and many identical flagpoles, we want to put all the
flags on equally spaced flagpoles on a rotating circular platform. We may use any number
of poles, but each pole used must have at least one flag, and order of flags on a pole mat-
ters. Arrangements that can be rotated into each other are not distinguishable. Count the
distinguishable ways to arrange the n flags. (Comment: There is a very short solution.)

1.1.37. For p prime and n ∈  , give a combinatorial proof that n divides (n+pp−1) − (np).
1.1.38. A spinner has a pointer that is equally likely to point to each of n regions num-
bered 1 , 2 , . . . , n. When we spin it three times, what is the probability that the sum of
the selected numbers is n?

1.1.39. Prove the identity below by showing that both sides count the same set of ternary
lists. (Karaivanov–Vassilev in Holland [2014]; identity (3.121) in Gould [1972])
⌊n/2⌋
n+1 s n− ¾
∑ (2s + 1)(¾ ) = ( ¾
)2 n−2¾
s=¾

1.1.40. (♦) Compositions of integers.

(a) Count the solutions in positive integers to ∑i=1 xi ≤ ¾ . Give a direct combinato-
n

rial argument for the resulting simple formula.

(b) Give a direct combinatorial argument to count all compositions of ¾ (with any num-
ber of parts).
(c) For ¾ ≥ 2, prove that the number of compositions of ¾ having an even number of
parts equals the number having an odd number of parts.
(d) For ¾ ≥ 2, prove that the number of compositions of ¾ with an even number of even
parts equals the number with an odd number of even parts.

1.1.41. (♦) Compositions of integers.

(a) Determine the total number of parts over all compositions of ¾ . (Hint: Consider
part (b) of Exercise 1.1.40.)
(b) For 1 ≤ m < ¾ , prove that all the compositions of ¾ together have (¾ − m + 3)2 ¾− m−2
parts equal to m. For example, 1 occurs twice in the composition 1 + 2 + 1 of 4, and all
the compositions of 4 together have twelve 1s, five 2s, and two 3s. (Hint: Relate the 1s in
compositions of ¾ to compositions of ¾ − 1.)
¾ −1
(c) Use parts (a) and (b) to evaluate 1 + ∑m=1 (¾ − m + 3)2 ¾− m−2 .
24 Chapter 1: Combinatorial Arguments

1.1.42. The Weights Problem. A balance scale tests whether the total weight put on the
left and right sides is the same. If one can specify ¾ known integer weights, what should
they be to maximize the value n such that every unknown object with weight in [n] can
be correctly weighed? What is this maximum n? (Example: when ¾ = 2 and the known
weights are {1 , 4}, one can weigh objects with weights in {1 , 3 , 4 , 5} but not 2. Using
{1 , 3}, one can balance every weight in [4].)
1.1.43. Consider a balance scale and positive integer weights w1 ≤ · · · ≤ w¾ . Let S0 = 0,
and let Sj = ∑i=1 wi for 1 ≤ j ≤ ¾ . Prove that it is possible to balance every integer weight
j

from 1 to S¾ if and only if wj ≤ 2Sj −1 + 1 for 1 ≤ j ≤ ¾ .

1.1.44. Choose points x1 , x2 , x3 on a circle. For each i, draw ¾i chords from xi to the arc
joining the other two points. What is the maximum number of regions inside the circle?
(Yaglom–Yaglom [1964])
1.1.45. Choose ¾ distinct numbers at random from [n], with all ¾-sets equally likely. Let
X be the sum of the chosen numbers. Prove that if n is divisible by r, and gcd(¾ , r) = 1,
then X is divisible by r with probability 1/r. (Kuczma [2000])

1.1.46. (+) For ¾ , m , n ∈ with ¾ ≤ n, prove the identity below by expressing the sum as
the probability of some event certain to happen. (Xiao [2012])
m
¾ (¾n)(mj)
∑ (¾ + j)(n+m) = 1 .
j =0 ¾+ j

1.1.47. Schr öder–Bernstein Theorem. Let A and B be sets. Prove that if º : A → B

and ½ : B → A are injections, then there is a bijection h: A → B. (Hint: Consider the
“chains” of elements formed by alternately using º to move from A to B and ½ to move
from B to A, and define h on these families.)

1.2. Identities
Using the Sum and Product Principles in a counting problem may produce
sums involving binomial coefficients. Standard formulas called identities can
help evaluate such sums.
Proofs of identities using induction or factorial formulas may involve tedious
manipulation. Combinatorial arguments can provide deeper understanding and
more information. Algebraic arguments (such as manipulating an identity) may
be easier to find; Graham–Knuth–Patashnik [1989] presents many. We focus
first on combinatorial arguments and introduce other techniques later.

L ATTICE PATHS AND PASCAL’S TRIANGLE

An argument by counting two ways shows that both sides of an identity count
the same set. The difficulty is devising an appropriate set. Arguments for sums
involving n¾ often use the set of n-ary ¾-tuples (words of length ¾ from an alpha-
bet of size n). There are several equivalent models for a set counted by (¾n); an
argument using one can be translated into the others. We have discussed models
involving subsets and binary lists; next we introduce the lattice path model.
Section 1.2: Identities 25

1.2.1. Definition. A lattice point is a vector with integer coordinates. A lat-

tice step changes one coordinate by 1. A lattice walk from a lattice point
(often the origin) moves by lattice steps. A lattice path is a lattice walk in
which each step increases one coordinate.

1.2.2. Example. Lattice paths and binary lists. A lattice path from (0 , 0) to
(¾ , n − ¾) has length n, with ¾ horizontal steps. Recording 1 for each horizon-
tal step and 0 for each vertical step produces a binary n-tuple with ¾ 1s. The
path is determined by where the horizontal steps occur in the list, so exactly one
path yields each such n-tuple. This establishes a bijection from the set of these
lattice paths to the set of binary n-tuples with ¾ 1s.

↔ 101101 ↔ {1 , 3 , 4 , 6} ⊆ [6]

Many authors use the term “lattice path” more generally, allowing steps of
other forms, but the restriction is convenient for us. A lattice path may end with
a horizontal step or a vertical step. This proves the formula (¾n) = (n−¾ 1) + (¾n−−11),
called Pascal’s Formula in honor of Blaise Pascal (1623–1662). The array of
numbers with (0n) , . . . , (nn) in the nth row is called Pascal’s Triangle, though it
was known to Chinese mathematicians much earlier.
1
11
1 2 1
1 3 3 1
1 4 6 4 1
1 5 10 10 5 1
1 6 15 20 15 6 1

Next we illustrate the use of various models in proving elementary identities.

1.2.3. Theorem. (Elementary Identities)

Elementary form Generalization
n n
(1) ( )=( ) (see Remark 1.3.11)
¾ n− ¾
n n−1 n−1
(2) ( )=( )+( ) (see Remark 1.3.11)
¾ ¾ ¾−1
n n−1 ¾ n n n− l
(3) ¾ ( ) = n( ) ( )( ) = ( )( )
¾ ¾−1 l ¾ l ¾−l
n n
(4) ∑ ( ) = 2n
¾ ∑ r¾ ( ) = (r + 1)n ¾
¾ ¾
n n
¾ n+1 m+ ¾ n−¾ m+ n+1
(5) ∑ (r) = (r + 1 ) ∑ (
r
)(
s
)=(
r+ s+1
)
¾ =0 ¾ =− m
2
n 2n m n m+ n
(6) ∑ ( ) = ( ) ∑ ( ¾ )(r − ¾) = ( )
¾ n r
¾ ¾
26 Chapter 1: Combinatorial Arguments

Proof: We give combinatorial proofs for these identities under the assumption
that the parameters are all nonnegative integers.
(1) Choosing ¾ positions for 1s or n − ¾ positions for 0s yields the same lists.
(2) Binary n-tuples with ¾ 1s end with a 0 or with a 1; equivalently, a lattice
path reaches (¾ , n − ¾) from (¾ − 1 , n − ¾) or (¾ , n − ¾ − 1).
(3) From n people, we form committees of size ¾ with a chair. On the left
we choose the committee and then select the chair; on the right the chair is cho-
sen first and then the rest of the committee. We call this the Committee-Chair
Identity. The generalization is the Subcommittee Identity: each side counts
committees of size ¾ with a subcommittee of size l. These are also the lists of
length n with l entries equal to A, ¾ − l equal to B, and n − ¾ equal to C. Be-
cause the positions for letters of each type can be chosen in various orders, there
are many products of two binomial coefficients that count this set.

¾−l
n l
n− ¾

(4) In the special case, both sides count all subsets of n elements. The gen-
eralization is the Binomial Theorem with x set to r and y set to 1; alternatively,
both sides count (r + 1)-ary lists of length n.
(5) For the special case, consider the (nr++11 ) binary (n + 1)-tuples with 1s in r + 1
positions. The left side counts them by the location of the rightmost 1; the ¾ th
term counts those where it is in position ¾ + 1. Alternatively, the left side counts
the lattice paths from (0 , 0) to (r + 1 , n − r) by the height at which the last hor-
izontal step occurs. We call this the Summation Identity. The generalization
counts binary (m + n + 1)-tuples with 1s in r + s + 1 positions by the position of
the (r + 1)th 1; the ¾ th term counts those where it is in position m + ¾ + 1. The
special case arises when m = s = 0.
(6) The general form is the Vandermonde Convolution; m = n = r yields
the special case. The left side counts r-sets from an (m + n)-set by how many el-
ements come from the first ¾ . Using lattice paths, the argument is that every
lattice path from (0 , 0) to (r, m + n − r) reaches the line x + y = m at some point
(¾ , m − ¾).

Writing (¾n) as ¾1! n(¾) treats it as a polynomial of degree ¾ in n. By the Poly-

nomial Principle, Identities (4)–(6) in Theorem 1.2.3 are valid as polynomials in
m and n. From this viewpoint, the general form of (6) is called Vandermonde’s
Theorem, in honor of Vandermonde [1772], but the convolution was known to
Chu Shih-Chieh (now spelled Zhu Shijie) in 1303. It will be helpful to study the
extension of binomial coefficients to non-integer arguments more formally.


1.2.4. Definition. For ¾ ∈ 0 , the extended binomial coefficient (u¾ ) is the
polynomial in u of degree ¾ defined by
¾ −1
u 1
( )=
¾ ¾! ∏(u − i) .
i=0
Section 1.2: Identities 27


For u ∈ 0 , the formula agrees with the combinatorial definition of (¾u) in
Definition 1.1.12; the value is 0 unless 0 ≤ ¾ ≤ u. The polynomial viewpoint
leads to an extension of the Binomial Theorem that was proved by Isaac Newton
(1643–1727) for rational exponents.

1.2.5. Theorem. (Extended Binomial Theorem) For u , x ∈ with | x| < 1,

u
(1 + x)u = ∑ ( )x ¾ .
¾
¾ ≥0

Proof: By Taylor ’s Theorem (of calculus), the coefficient of x ¾/¾ ! in the expansion
of (1 + x)u as a power series in x is the ¾ th derivative of (1 + x)u with respect to x,
evaluated at x = 0. This value is u(¾) .

1.2.6. Remark. Negative binomial and multisets. In Section 3.1 we will give a
combinatorial interpretation of Theorem 1.2.5 in the case where u is a negative
integer. Invoking Definition 1.2.4 as a polynomial yields
−n n+ ¾ −1
( ) = (−1)¾ ( ).
¾ ¾
Thus the expansion of (1 + x)−n is closely related to selection of ¾ elements (with
repetition) from n types of elements.

There are many methods for evaluating sums. Pascal’s Formula (Theorem
1.2.3(2)) may allow us to prove binomial coefficient identities by induction (as
in Exercise 11). However, induction does not work unless we know the answer
in advance; other methods may both discover and prove the formula. Bijective
proofs may give more “refined” results (see Theorem 1.3.26 and Example 1.3.27,
for example). Instead of devising new combinatorial arguments, we may also use
identities already proved; indeed, this is the motivation for proving identities.

1.2.7. Application. Initial values of a polynomial. For a polynomial p, we can

n
compute ∑ i=1 p(i) by writing the powers of i as integer combinations of binomial
coefficients and applying Theorem 1.2.3(5). The transformation uses identities
such as i2 = 2(2i ) + (1i ) and i3 = 6(3i ) + 6(2i ) + (1i ). To understand the identity
for i¾ , observe that the left side counts ¾-tuples with entries in [i] by filling the
positions independently, while the right side counts the same set according to the
number of distinct values used. Up to cubes, this argument yields the formulas
below. Exercise 8 requests the details.
n n n
n(n + 1) n(n + 1)(2n + 1) n2(n + 1)2
∑i = 2
, ∑ i2 = 6
, ∑ i3 = 4
.
i=1 i=1 i=1

The coefficient of (ri ) in the expression for i¾ counts the groupings of the po-
sitions in [¾] into r labeled boxes. We will obtain a general formula for this in
Section 3.3. For now, it suffices that such coefficients exist, because ( xj) is a poly-
nomial in x of degree j , and hence (0x) , . . . , (rx) is a basis for the vector space of
polynomials in x of degree at most r.
28 Chapter 1: Combinatorial Arguments

DEL ANNOY NUMBERS

Next we consider paths in the plane more general than lattice paths.

1.2.8. Definition. A Delannoy path is a path from (0 , 0) to (m , n) with each step

in {(1 , 0) , (0 , 1) , (1 , 1)}. The Delannoy number d m ,n is the number of such
paths. The numbers of the form d n ,n are the central Delannoy numbers.

1.2.9. Example. Below we show the seven paths from (0 , 0) to (2 , 2) that arise in
addition to the (42) lattice paths; altogether, d2 ,2 = 13.
The Delannoy numbers were introduced by Henri Auguste Delannoy (1833–
1915). The two formulas for d m ,n that we discuss below appear in Delannoy
[1889]; Banderier–Schwer [2005] discusses the history. Sulanke [2003] presents
29 problems for which the central Delannoy numbers provide the solution.

Delannoy paths are arrangements of three types of steps, and d m ,n counts

such arrangements. When proving that a sum counts a particular set, we cut the
set into pieces counted by the terms in the sum. A key step is to recognize what
the index of summation means in describing the pieces. Doing that can make it
easy to show that the summand counts the corresponding piece (see Exercise 49,
for example).

1.2.10. Proposition. The Delannoy number d m ,n is given by

m n+ m− j m n+ ¾
d m ,n = ∑ ( )( ) = ∑ ( )( ).
j m ¾ m
j ¾

Proof: Since we know there are (n+mm) paths from (0 , 0) to (m , n) using no diago-
nal steps, it makes sense to group the Delannoy paths by the number of diagonal
steps, j . To reach (m , n) there must also be m − j horizontal steps and n − j ver-
tical steps, in total n + m − j steps. Given the number of steps of each type, the
paths correspond to the orderings of the steps: the words with j copies of D, m − j
of H , and n − j of V . Forming the words in two stages yields (n+mm− j
)(mm− j ) as the
number of words.
Summing over the number j of diagonal steps completes the count. Grouping
instead by the number of horizontal steps, ¾ , yields the second expression (also
obtained from the first by substituting ¾ = m − j).

When no limits are given for the index parameter in a sum, as in the state-
ment of Proposition 1.2.10, all terms where the summand is nonzero are in-
cluded. When the summand involves binomial coefficients, the index values yield-
ing nonzero terms are usually clear.
Our next objects are closely related to Delannoy paths.
Section 1.2: Identities 29

1.2.11. Definition. The lattice ball of radius m in n dimensions is the set of

lattice points in n reachable from the origin by lattice walks with at most
m steps, shown below for m = n = 2.

•
• • •
• • • • •
• • •
•

A more combinatorial description for the lattice ball is the set of all inte-
ger n-tuples such that the absolute values of the entries sum to at most m. This
phrasing makes it easy to apply the integer equation model to count the points.

1.2.12. Proposition. (Golomb–Welch [1970], see also Vassilev–Atanassov [1994])

The size lm ,n of the lattice ball of radius m in n is given by
n m
lm ,n = ∑ ( )( )2¾ .
¾ ¾
¾

Proof: We count the integer n-tuples whose absolute values sum to at most m.
Group them by the number of nonzero coordinates, ¾ . We can pick these positions
in (¾n) ways and give signs to the nonzero entries in 2¾ ways. To finish building a
desired n-tuple, we give magnitudes to the ¾ nonzero entries by solving x1 + · · · +
x¾ ≤ m in positive integers.
Taking partial sums of such a list yields a ¾-element subset of [m], with the
j
jth element being ∑i=1 x i . To see that each ¾-set S in [m] arises exactly once,
write S as a1 , . . . , a¾ in increasing order and let x i = ai − ai−1 (with a0 = 0). Now
x1 , . . . , x ¾ is a solution to the inequality that yields S under the original map,
and it is the only solution that yields S. Hence the map is a bijection, and there
are (m ¾ ) solutions to the inequality.

Surprisingly, this formula for lm ,n is symmetric in m and n. The combina-

torial Definition 1.2.8 for d m ,n is also symmetric, but the formula in Proposition
1.2.10 is not. The real surprise, suggested by d2 ,2 = 13 = l2 ,2 , is that the two
formulas are equal! We present a bijection found in 2006 by a student (Pavithra
Prabhakar) after the second lecture of a combinatorics course; Sulanke [2003,
Example 20] used a similar idea.

1.2.13. Theorem. (Delannoy [1889]) For m , n ∈ 0 ,

m n+ ¾ n m
∑ ( ¾ )( ) = ∑ ( )( )2¾ .
m ¾ ¾
¾ ¾

Proof: Let A be the set of Delannoy paths to (m , n), and let B be the lattice ball
of radius m in n . By Propositions 1.2.10–1.2.12, the left side is | A| and the right
is | B| . We prove equality by a bijection º : A → B.
30 Chapter 1: Combinatorial Arguments

From a Delannoy path P to (m , n), we form an integer list (b1 , . . . , bn) with
∑i=1 |bi | ≤ m. Let |bi | be the number of horizontal or diagonal steps from the point
n

where P first reaches vertical coordinate i − 1 to the point where P first reaches
vertical coordinate i. The sign of bi is positive if the final step is vertical, negative
if it is diagonal. Let º (P) = b.
For every n-tuple b in the lattice ball, we show that º (P) = b for exactly one
Delannoy path P. Replace a nonnegative bi with bi horizontal steps and then one
vertical step. Replace a negative bi with −bi − 1 horizontal steps and then one
diagonal step. This produces a Delannoy path from (0 , 0) to (r, n), where r is the
sum of the absolute values in b. Complete P by adding m − r horizontal steps.
Since the portion of P between the lines y = i − 1 and y = i must yield bi , this is
the only way to form a Delannoy path that maps to b under º .

(0,5) (8,5)
0
−3
0
1
−2
(0,0) (8,0)

Using the Delannoy path and lattice ball models, Remark 2.1.11 gives an in-
ductive proof of Theorem 1.2.13 using recurrences. The equality can be proved
directly without these models by inventing a set counted by both sides. When
devising a set to be counted by sum of products, often one defines a set of pairs;
they can be counted by a two-step process, grouping the pairs according to one
coordinate. Exercise 39 requests a proof of Theorem 1.2.13 by this method.

1.2.14.* Remark. The bijection in Theorem 1.2.13 restricts to various state-

ments we have already proved. Forbidding diagonal steps in the path forbids
negative entries in the n-tuple. The n-tuples become the nonnegative integer so-
n+1
lutions to ∑i=1 x i ≤ m, or ∑i=1 x i = m. These correspond to multisets of size
n

m from n + 1 types; by Theorem 1.1.20, there are (m+nm+1−1 ). As expected, the

resulting value (mn+n) counts the lattice paths from (0 , 0) to (m , n).
Forbidding vertical steps in the path forbids nonnegative entries in the n-
tuple. Negating the n-tuple yields n positive integers summing to at most m.
Again adding x n+1 , we obtain compositions of m + 1 with n + 1 parts, of which
there are (m
n
) (Corollary 1.1.22). The corresponding paths arise by ordering n
diagonal steps and m − n horizontal steps in (m )
n ways.

We have given only a few identities; others appear in the exercises. H. Gould
[1972] listed more than 550 identities for binomial coefficients. Combinatorial
argument is not our only tool. Other techniques developed in the next several
chapters will enable us to prove identities or evaluate sums more systematically
(and maybe with less cleverness!).
Exercises for Section 1.2 31

EXERCISES 1.2
+ n m+¾
1.2.1. (−) Give combinatorial proofs of (¾+n1 ) = n−¾ n
¾ +1 ( ¾ ) and (m
m+¾ )( ¾ )
= (mm+n)(¾n).
1.2.2. (−) Evaluate ∑¾=0 ( m¾+¾) .
n

1.2.3. (−) Prove (¾n) − (n−¾ 2) = (¾n−−11 ) + (¾n−−12).

1.2.4. (−) Prove that Pascal’s Formula holds for the extended binomial coefficient.
1.2.5. (−) Use the Vandermonde Convolution to prove ∑¾ (m+r ¾ )( n−s ¾) = (mr++sn).

1.2.6. (−) Prove ∑¾=0 ( m+¾¾−1 ) = ∑¾=0 ( n+¾¾−1 ) and ∑¾ (n¾)(r+m¾ ) = ( nn++mr ).
n m

1.2.7. (−) Evaluate (−¾1 ).

1.2.8. (−) Complete the proofs of the formulas in Application 1.2.7 for ∑i=1 i, ∑i=1 i2 , and
n n

∑i=1 i3 , using the expressions

n
for powers of i in terms of binomial coefficients. Use these
formulas to evaluate ∑i=1
n
(2i 3
+ 3i2 − 5i).
1.2.9. (−) We put m white and n black marbles into a row of boxes. From the beginning of
the row, no box can be left empty until all the marbles are used, and each box has at most
one marble of each color. Prove that the number of ways to place the marbles in the boxes
is the Delannoy number d m ,n . (Schwer [2002])
1.2.10. Prove (¾n−−11 )(¾+n1 )( n+¾ 1 ) = (n−¾ 1 )(¾n++11 )(¾−n1) without expanding into factorials (use
known identities).
1.2.11. (♦) Use Pascal’s Formula to prove the following identities by induction.
(a) (¾n) = (c) (x + y)n = ∑¾=0 (¾n) x¾ y n−¾ .
n! n
¾ !(n−¾)! .
(b) ∑i=0 (¾i ) = (¾n++11 ) for ¾ , n ∈
n
0 . ¾
(d) ∑i=0(−1)i (¾n−i) = ( n−¾ 1 ).

1.2.12. (♦) When flipping 100 fair coins, is it more likely that the numbers of heads and
tails are equal or that they differ by two? Give an algebraic proof by manipulating formu-
las and a combinatorial proof by defining an injection.
1.2.13. (♦) Prove each identity below by counting a set in two ways. (One can count 3-tuples
with special properties or various geometric arrangements.) Then use the two identities
to obtain simple formulas that evaluate ∑i=1 ∑ j =1 min{i , j} and ∑i=1 ∑ j =1 max{i , j}.
n n n n

(a) ∑i=1 ∑ j =1 min{i , j} = ∑¾=1 ¾ 2 . (b) ∑¾=1 ¾ 2 = 2(n+3 1 ) + (n+2 1 ).

n n n n


1.2.14. For m ∈ , let º (m) = ∑ j =1 (m − j)2 j −1 . Obtain a simple formula for
m
º (m) induc-
tively. Also give a combinatorial proof by counting a set in two ways.
1.2.15. (♦) Prove each identity below by counting a set in two ways. In each case, give a
single direct argument without manipulating the formulas.
(a) (2n
n)
= 2(2n−1
n−1 )
. (c) ∑¾=1 q¾−1 =
n q n −1
q−1
for q , n ∈  with q
= 1.
¾
(b) ∑¾ ( l )(¾) = ( )2 (d) ∑ i(n − i) = ∑
n− l n n
n n
l
. i =1 i =1
( ).
i
2

1.2.16. (+) Prove Strehl’s Identity below (algebraically or combinatorially).

2 3
n 2¾ n
∑( ) (
¾ n
) = ∑( )
¾
(Strehl [1994, eq. 29])
¾ ¾
32 Chapter 1: Combinatorial Arguments

¾ −1
1.2.17. (♦) Use known identities to evaluate ∑¾=1 ∑i=1 (i − 1)(¾ − i − 1). Give a combina-
n

torial proof of the resulting identity.


1.2.18. (+) For m , r ∈ , use known identities to prove ∑¾≥1 ¾ (r+m¾++11 ) = ∑i=1 i2i−1 (mr−i) ,
m

and give a combinatorial proof by devising a set that both sides count.
1.2.19. Use known identities to evaluate the sums below.
n
1 n n 1
(a) ∑ ( ). (b) ∑(−1)¾ ( ) .
¾+1 ¾ ¾ n+ 1 − ¾
¾ ≥0 ¾ =0

1.2.20. (♦) Prove the following variation of Theorem 1.2.3(6):

n
⌊ n/2⌋ ⌈ n/2⌉ n+1
∑ (⌊ ¾/2⌋ )(⌈ ¾/2⌉ ) = (⌈ n/2⌉ ) .
¾ =0

1.2.21. Let Hn = ∑i=1

n 1
i
(so H0 0). = Suppose 0 ≤ ¾ ≤ n. (Based on Sondow [2003].)
(a) Prove ∑
n
j = ¾ +1
−
( 1) j −¾−1 nj j −1 ¾
() = (¾n)(Hn − H¾). (Hint: Use induction on n.)
¾ −1
(b) Prove ∑ j =0
−( 1) j −¾−1 nj ¾−1 j
() = (¾n)(Hn − Hn−¾).
1.2.22. (♦) Reciprocal powers in sums.
(a) Given bn = ∑¾=1 a¾ (¾n) for n ≥ 1, prove ∑¾=1 = ∑ ¾ =1
n n n
b¾
¾
a¾
¾ (¾n).
¾ −1
(b) Prove ∑ ¾ ···1 ¾m = ∑¾=1 (−1)
n
¾ m (¾ ) , where the sum on the left is over all ¾1 , . . . , ¾ m
n
( )
1
such that 1 ≤ ¾1 ≤ · · · ≤ ¾m ≤ n. (Dilcher [1995], ’t Woord [1999])

1.2.23. For n , m , j ∈ , prove the identities below. (Generalizes D íaz-Barrero [2005]; see
also Bang [1995] and ’t Woord [1999].)
(a) ∑¾= j (−1)¾ (¾n) = (−1) j (nj −−11 ) .
n

(b) ∑¾=1 (−1)¾+1 (¾n) ∑1 ≤i

n
1 ≤···≤ i m≤ ¾
1
i1 ··· i m
= 1
nm
.

1.2.24. Let Sn be the hexagonal arrangement consisting of n rings of dots, as illustrated

below for n ∈ {1 , 2 , 3}. Let an be the number of dots in Sn . Compute an . Compute ∑¾=1 a¾ .
n

The formula for ∑¾=1 a¾ is rather simple; is there a direct combinatorial proof?
n

• • •
• • • • • •
• • • • • • • • •
• • • • • •
• • •

1.2.25. (♦) Prove ∑¾=1 ¾ · ¾ ! = (n + 1)! − 1 by induction and by counting two ways.
n

1.2.26. By counting a set in two ways, prove ∑ A⊆[n] ∑ B⊆[n] | A ∩ B| = n4 n−1 .

1.2.27. (♦) Evaluate ∑S⊆[n] ∏i∈S 1

i
and ∑S⊆[n](−1)| S| ∏i∈S 1i .

1.2.28. Give both an algebraic proof (using known identities) and a combinatorial proof
(counting a set in two ways) for the identity (n − r)( n+rr−1 )( nr) = n(n+2rr−1 )(2rr).

1.2.29. (♦) Prove ∑ (¾n)((mn−−¾¾)/2 )2 ¾ = (2n

m
) , summed over ¾ with m − ¾ even. (Hint: Count
m-subsets of {x1 , . . . , x n , y1 , . . . , yn}.) Conclude ∑¾=0 (¾n)(⌊(mn−−¾¾)/2⌋) 2 ¾ = (2nm+1 ) . (R. Stanley)
n

1.2.30. Count the integer lists a− m , a− m+1 . . . an such that − p ≤ a− m ≤ a− m+1 ≤ · · · ≤ an ≤

q and a−1 ≤ 0 ≤ a1 . (Djokovic [1991])
Exercises for Section 1.2 33

1.2.31. (+) For the identity below,

(a) Give a combinatorial proof (construct a set that both sides count).
(b) Use the Binomial Theorem to prove that each side is the coefficient of x n in
(1 + 3x + x2)n . Explain the sense in which the two proofs are the same. (Beckwith [2008])
n 2¾ n 2¾
∑ (¾)( ¾ ) = ∑ (2¾)( ¾ )3n−2¾
¾ ≥0 ¾ ≥0

1.2.32. For the identity below, give a combinatorial proof, obtain the Summation Identity
(elementary form of Theorem 1.2.3(5)) as a special case, and apply a polynomial identity
from the text to express the general case as an instance of Vandermonde’s Theorem (the
polynomial version of Theorem 1.2.3(6)),
¾
a+ i−1 b+ ¾ − i−1 a+ b+ ¾ −1
∑( i
)(
¾−i
)=(
¾
).
i =0

1.2.33. (♦) For nonnegative integers m, n, r, and s, prove the identity below. (Ohtsuka–
Tauraso [2018])
s r
m+ r r + ¾ s m+ s s+ ¾ r
∑ ( n − ¾ )( ¾ )(¾) = ∑ ( n − ¾ )( ¾ )(¾)
¾ =0 ¾ =0

+s
1.2.34. (♦) For r ≥ m+ n, prove ∑¾ (mr+¾)(n+s ¾ ) = (r−rm+ n) , and use it to evaluate ∑¾ ¾ (¾ )(¾ ) .
a b

(Graham–Knuth–Patashnik [1989, p. 181])

1.2.35. Prove the identity below for n ∈ . (Sorel [2016])
n 2n+1 2
2n 2n + 1 2n + 1 2n 4n + 1 2n
∑ ( ¾ )( ¾
)+ ∑( ¾
)(
¾−1
)=(
2n
)+( )
n
¾ =0 ¾ = n+1

0 with n ≥ m, prove ∑¾m=0 (( )) = n−nm+1+1 . (Whitworth [1897])

m
1.2.36. For n , m ∈ ¾
n
¾

a(¾) −1
1.2.37. For a , b ∈ with b > a + 1, prove ∑¾≥0 b(¾)
= b−1
b−1 − a
. Use it to evaluate ∑¾≥0 ( n+n ¾) .
b−1 a(¾)
− ∑ ¾ =0
n
(Hint: First compute b−1 − a b(¾)
.)

1.2.38. Let cn = (⌊n/2⌋). Prove that ∑¾=0 (n¾)c¾ cn−¾ = cn cn+1 . (Bloom [2002]) (Comment: A
n n

combinatorial proof counts “balanced” teams of sizes n and n + 1 from n male/female pairs
and one extra male. An algebraic proof uses factorials and Exercise 1.2.20.)
n+¾
1.2.39. (♦) By Theorem 1.2.13, ∑¾ (m ¾ )( m ) = ∑j ( j )( j ) 2 . Ignoring that the two sides
n m j

count Delannoy paths and the lattice ball, give a direct combinatorial proof of the identity
by showing that both sides count the following set S. Let M and N be sets with sizes m
and n. Let S be the family of ordered pairs (A , B) such that A is a subset of M and B is an
m-subset of N ∪ A, as shown below.

M N

A B
34 Chapter 1: Combinatorial Arguments


1.2.40. (♦) For m , n ∈ 0 , generalize the argument of Exercise 1.2.39 to prove the follow-
ing polynomial identity by the Polynomial Principle:

m n+ ¾ ¾ m n
∑ ( ¾ )( m
) x = ∑ ( )( ) x m− j (1 + x) j .
j j
¾ j

1.2.41. Given 2m + 2n people named x1 , . . . , x m+ n and y1 , . . . , ym+ n , we want to form n

pairs using only pairs of the form {xi , yi } for 1 ≤ i ≤ m + n or {xi , xi+1 } for 1 ≤ i ≤ m + n − 1
(each xi can be involved in only one pair). Establish a bijection to show that the number of
ways to form n pairs is the Delannoy number d m ,n . (Deutsch; see Sulanke [2003])
1.2.42. (+) The augmented Aztec diamond of order n consists of 2n + 1 centered rows of
squares with lengths 2 , 4 . . . , 2n − 2 , 2n , 2n , 2n , 2n − 2 , . . . , 4 , 2, shown below for n = 3.
Prove that the number of ways to tile (partition) it using dominos (1-by-2 or 2-by-1 rectan-
gles) is the central Delannoy number d n ,n . (Sachs–Zernitz [1994]) (Hint: The figure below
suggests a way to obtain a path from a tiling. The easiest way to complete the proof is to
prove a more general result. Comment: The ordinary Aztec diamond lacks the central row
of length 2n. It has exactly 2 n(n+1)/2 tilings, a difficult result proved in Elkies–Kuperberg–
Larsen–Propp [1992], Kuperberg [2002], Kuo [2004], and Brualdi–Kirkland [2005].)

1.2.43. (♦) Let A = {a1 , . . . , an} ⊂ , with a1 < · · · < an . For i ∈ [n], let Si be the family
of i-element subsets of A, and let i = ∑ B∈S max(B). For  ∈ [n], prove
i
n
r−1
a¾ = ∑(−1)¾+r ( ) r. (Ash [2010])
−1
r =1

1.2.44. (+) Let n be the set of permutations of [n]. For positive real d1 , . . . , dn , prove
1 1
∑ d (1)(d (1) + d (2)) · · · (d (1) + · · · + d (n)) = d1 · · · d n
.
∈Ë n

Apply this to prove ∑v∈S(v1 · · · vn)−1 = ∑w∈ T (w1 · · · wn)−1 , where S = {v ∈ [N]n: ∑ vi ≤ N}
and T = {w ∈ [N]n: w1 , . . . , wn are distinct}. (Hill [2004], solution to Lubell [2003])

1.2.45. Given p , q ∈ , let d = gcd(p , q). By counting a geometric arrangement of points,
q−1
prove that 2 ∑i=1 ⌊ ip/q⌋ = (p − 1)(q − 1) + d − 1 .

1.2.46. Give a combinatorial proof that ∑i=0 ⌈ in ⌉ = ∑ j =0 ⌈ jmn ⌉ .

m n
m
(Hint: Use a geometric
arrangement of points, as in Exercise 1.2.45.)
1.2.47. (♦) Binomial coefficients as polynomials. Fix  ∈ . 
(a) Let (n) =  !(n) − n + (2) n−1 . Note that 2 is identically 0. For  ≥ 3, prove that
 is a polynomial of degree  − 2.
(b) Prove that ∑i=0 i is a polynomial in n with leading terms +1 1 n+1 + 12 n .
n
Section 1.3: Applications 35

1.2.48. (♦) Let I be the set of polynomials with rational coefficients that have integer val-
ues at all integers.
(a) Prove that every polynomial º of degree ¾ with rational coefficients can be ex-
¾
pressed in exactly one way as º (x) = ∑ j =0 bj ( xj) such that each bj is rational.
(b) Prove that º ∈ I if and only if the coefficients b0 , . . . , b¾ in part (a) are integers.
(Hint: Evaluate º at {0 , . . . , ¾}. Note that (00) = 1.)

1.2.49. By considering lists made from n − 2r copies of U (up) and n + 2r copies of D (down),
prove the identity below. (Hint: To explain the left side, view each list as a list of pairs of
successive elements; there are four types of pairs.)

n 2¾ 2n
∑ (2¾)(¾ − r)2n−2¾ = (n − 2r) . (Callan [2003a])
¾

(Comment: A similar argument yields ∑¾ (2¾n+1 )(2¾¾−+r1 )2 n−2¾−1 = (n−2n

2r−1 )
.)

1.2.50. (+) Prove the following identity by devising a set counted by both sides.

2m − ¾
n
2m + 1
∑ 2¾ ( m+ n
) = 4m − ∑ (
m+ j
). (Knuth [2007])
¾ ≥0 j =1

1.2.51. (+) Prove the following identity combinatorially. (Karaivanov–Vassilev [2017])

⌊n/2⌋
n+1 s n− ¾
∑ (2s + 1)(¾ ) = ( ¾ )2 n−2¾
s=¾

1.3. Applications
Proving identities provides plenty of opportunity for clever combinatorial ar-
guments, but often we want to count a new set rather than explain a given for-
mula. In this section we study several problems of this type.

GRAPHS AND TREES

A special case of Proposition 1.1.13 is that every ¾-element set has 2¾ subsets.
We apply this to count graphs (see Chapter 0 for basic definitions about graphs).

1.3.1. Corollary. There are 2(2) graphs with vertex set [n].
n

Proof: Each edge is an unordered pair of vertices, and a graph with a specified
vertex set is determined by choosing a family of vertex pairs as the edge set. There
are (n2) pairs of vertices to choose from.

We next count the trees (connected graphs without cycles) with vertex set [n].
Such trees have n − 1 edges and arise from a single vertex by iteratively adding a
new leaf with one old neighbor (see Chapter 0).
36 Chapter 1: Combinatorial Arguments

1.3.2. Example. Small trees. When n ∈ {1 , 2}, there is one tree with vertex set
[n]. With vertex set [3], there are three trees, pairwise isomorphic. They are
determined by choosing the central vertex, as shown below. With vertex set [4],
there are four “stars” and twelve paths; 16 trees in total. Case analysis yields
125 trees with vertex set [5].

3 1 2 1 2 3 1 3 2
• • • • • • • • •

Note that 1 , 3 , 16 , 125 fits the sequence ⟨a⟩ with an = nn−2 . To prove that
this is the formula, we seek a correspondence between the trees and the nn−2 lists
of length n − 2 with entries in [n]. We first view the lists as n-tuples by adding a
first entry 1 and last entry n. We then treat an n-tuple a with entries in [n] as a
function ºa : [n] → [n] defined by ºa(i) = ai for 1 ≤ i ≤ n. A function from a set to
itself has a useful graphical representation, which we use to define edges.

1.3.3. Definition. The functional digraph of a function º : S → S is the di-

rected graph with vertex set S having an edge from x to º (x) for each x ∈ S.
Each vertex is the tail of exactly one edge.

1.3.4. Theorem. (Cayley’s Formula; Cayley [1889]) There are nn−2 trees with
vertex set [n].
Proof: (Eğecioğlu–Remmel [1986]) There are nn functions º from [n] to [n].
Among these, nn−2 satisfy º (1) = 1 and º (n) = n. We establish a bijection from
this set to the set of trees with vertex set [n].
The functional digraph of º has vertex set [n] and n edges. Since each vertex
is the tail of exactly one edge, the digraph consists of cycles with directed trees
appended (iteration of º takes an element eventually to a cycle). Loops at 1 and
n are two of these cycles.
We modify the functional digraph of º to obtain a tree ( ). First place
the vertices of the cycles in  on a horizontal line, with the least label in each
cycle appearing first (followed in order by its successive images). List the cy-
cles in decreasing order of their least elements. When n = 15, the n-tuple
(1 , 5 , 4 , 3 , 10 , 12 , 5 , 15 , 12 , 7 , 1 , 7 , 15 , 3 , 15) yields the diagram on the left.
15 5 10 7 3 4 1 15 5 10 7 3 4 1
• • • • • • • • • • • • • •
• •• 12• • •11 → • •• 12• • •11
13 8 2 14 13 8 2 14
• • • •
6 9 6 9
For each cycle, cut the edge from the last vertex to the first. Replace each
such edge (except for the last cycle) with an edge to the first vertex of the next
cycle. This destroys all cycles, reduces the number of edges by one, and connects
the pieces. Ignoring the directions on the edges yields a connected graph with no
cycles: a tree ( ) with vertex set [n]; see the outcome on the right above.
The vertices of the cycles in the functional digraph form a path P in ( )
from n to 1 along the horizontal line. A vertex on P began a cycle in  if and only
if its label is less than all the labels that precede it on P. This is the key idea.
Section 1.3: Applications 37

For a tree T with vertex set [n], let P be the unique path from n to 1. Draw T
with P along a line. Orient edges not in P toward P , and orient P from n toward
1. For each vertex along P whose label is less than all preceding it, make it the
start of a new cycle by cutting the edge of P that reaches it and instead complete
the current cycle. Since n is first and 1 is last, these two become cycles of length
1. We obtain a functional digraph º of the desired form, so is surjective.
A vertex on the path from n to 1 in ( ) is the least element of a cycle in 
if and only if it is less than all earlier labels. Hence  is the only function with
( ) = T . Thus is also injective, and there are nn−2 trees with vertex set [n].

The idea used to arrange the cycles of this functional digraph will be used
again for the canonical cycle decomposition of a permutation in Section 3.1.

1.3.5.* Remark. Cayley ’s Formula has many proofs: Cayley [1889], K irch-
hoff [1847] (Exercise 15.2.4), Pr üfer [1918] (Exercise 12), P ólya [1937] (Theo-
rem 3.3.29), Moon [1967], Rényi [1970] (Theorem 15.2.9), Joyal [1981], Pitman
[1999], and others (Exercise 17). Equivalent results appeared earlier (Borchardt
[1860], Sylvester [1857]), but Cayley expressed the problem in terms of graph
theory and invented the term “tree”. Proofs also appear in Chapter 3 of Biggs–
Lloyd–Wilson [1976] and in Chapter 24 of Aigner–Ziegler [1999]. Moon [1970]
wrote a book on enumerating classes of trees.

MULTINOMIAL COEFFICIENTS

Cayley approached the tree-counting problem algebraically, using a generat-

ing function to enumerate labeled trees by their vertex degrees. The Eğecioğlu–
Remmel bijection also provides this information, using our next counting formula
(a special case was used in Proposition 1.2.10).
n
1.3.6. Proposition. There are m!/ ∏i=1 i ! words of length m having exactly i
letters of type i, where ∑ i=1 i = m.
n

Proof 1: We can produce the arrangement by choosing positions for the letters
of the first type, then choosing positions for the second type, and so on. The re-
sulting product simplifies to the desired formula.

m m − 1 m − 1 − 2 m! (m − 1)! m!
( )( )( )··· = ··· =
1 2 3 1 !(m − 1)! 2 !(m − 1 − 2)! Π i !

Proof 2: The symmetry in the formula suggests a more direct argument. We

temporarily assign subscripts to the letters of each type to distinguish them; now
there are m! arrangements. Each desired arrangement corresponds to exactly
∏ i ! of these, because there are ∏ i ! ways to assign subscripts once the posi-
tions of the letters are fixed.

1.3.7. Corollary. The number of trees with vertex set [n] in which vertices
1 , . . . , n have degrees d1 , . . . , d n , respectively, is (n − 2)!/ ∏i=1 (d i − 1)!.
n
38 Chapter 1: Combinatorial Arguments

Proof: In the functional digraph º for the list (1 , a2 , . . . , an−1 , n), one edge leaves
each i, and the number of edges entering i is the number of copies of i in the list.
For each i ∈ / {1 , n}, there is one incident edge in ( ) for each edge entering or
leaving i in  . However, turning the cycles into a path loses the edge entering n
and the edge leaving 1.
After deleting the first and last term in the list, it becomes true for all i that
the degree of i in the tree ( ) is one more than the number of copies of i in the list
a2 , . . . , an−1 . Therefore, we count trees with each i having degree d i by counting
lists of length n − 2 with d i − 1 copies of i for each i. By Proposition 1.3.6, the
formula is as claimed.

1.3.8. Example. Trees with fixed degrees. Consider 7-vertex trees whose vertices
(n−2)!
have degrees 3 , 1 , 2 , 1 , 3 , 1 , 1. Here Π(d i −1)!
= 30. Each tree is of one of the three
types shown below. There are six ways to complete the first tree (pick from the
remaining four vertices the two adjacent to vertex 1) and twelve ways to complete
each of the others (pick the neighbor of vertex 3 from the remaining four, and
then pick the neighbor of the central vertex from the remaining three).

1 3 5 1 5 3 5 1 3
• • • • • • • • • • • • •
• • • • • • • •

¾
n
The number  !/ ∏ i=1 i ! is written as (¾1 ,... )
,¾ n or P( ; 1 , . . . ,  n
). When n =
2, arranging the letters is the same as choosing positions for the first type, so
P(1 + 2 ; 1 , 2) = (¾1¾+1¾2 ). Due to the next result, the numbers (¾1 ,... ¾
)
,¾ n are called
multinomial coefficients.

1.3.9. Proposition. (Multinomial Theorem) For n ,  ∈ 0 ,

n ¾ n
⎛ ⎞ 
∑ xi = ∑ ( x ¾i ,
. . , n ∏ i
)
⎝ i=1 ⎠  1 , .
i=1

with the sum on the right over (1 , . . . , n) ∈ 0n with sum  .
Proof: In expanding (∑i=1 x i)¾ , the contributions to the coefficient of ∏i=1 x¾i i
n n

count the -letter words with i letters of type i for each i.

1.3.10. Application. (Fermat’s Little Theorem) If n ∈ and p is prime, then

np ≡ n (mod p). Furthermore, np−1 ≡ 1 (mod p) when p does not divide n.
Proof: (Leibniz) Multiplication with congruence classes is well defined, so we may
assume that n ∈ . We rewrite np as (∑i=1 1)p , which is obtained from (∑i=1 x i)p
n n

by setting each x i = 1.
By the Multinomial Theorem, the coefficient of ∏i=1 x¾i i in the expansion of
n
n n
(∑i=1 x i)p is p!/ ∏ i=1 i !. For each term of the form x i , the coefficient is 1; there
p

are n such terms. All other coefficients are divisible by p, since the denominator
in the formula for those multinomial coefficients has no divisor of p. Hence np =
(1 + · · · + 1)p ≡ n (mod p). (See Exercise 22 for another proof.)
Section 1.3: Applications 39

The second statement is the usual statement of Fermat ’s Little Theorem.

It follows from the first by the existence of multiplicative inverses modulo p for
numbers that are not multiples of p.

1.3.11. Remark. The multinomial coefficients provide extensions of identities (1)

¾
and (2) from Theorem 1.2.3. Since (¾1 ,... )
,¾ n counts words with specified multiplic-
ities of each letter, permuting ¾1 , . . . , ¾n does not change the value. The words
can also be interpreted as n-dimensional lattice paths. Grouping the paths by the
direction of the last step extends Pascal’s Formula:
n n−1 n−1
( )=( ) + ···+ ( ).
¾1 , . . . , ¾t ¾1 − 1 , ¾2 , . . . , ¾t ¾1 , . . . , ¾t−1 , ¾t − 1

THE BALLOT PROBLEM

A natural problem in combinatorial probability leads to a fundamental count-

ing sequence for special sets of lattice paths or binary lists.

1.3.12. Example. Bertrand’s Ballot Problem. Candidates A and B receive a

and b votes, respectively, with a ≥ b. When votes are counted in random order,
what is the probability that A never trails? The a + b votes can be removed from
the box in (a + b)! equally likely orders. When we record the votes as a list of As
and Bs, each possible list arises from a! b! orderings of the votes. This makes the
A , B-lists equally likely, and it suffices to determine the fraction of these lists in
which every initial segment has at least as many As as Bs.

1.3.13. Theorem. (Bertrand [1887]) Among the lists formed from a copies of A
and b copies of B, where a ≥ b, there are ( a+a b) − (aa++1b) lists in which every
initial segment has at least as many As as Bs.
Proof: Lists consisting of a copies of A and b copies of B correspond to lattice
paths from the origin to (a , b). The total number of paths is ( a+a b), and the desired
lists correspond to the paths that never move above the line y = x. We count these
by subtracting the others from the total.
A path steps above the diagonal if it reaches (¾ , ¾ + 1) for some ¾ . Consider
the first such occurrence. Modify the list after this by changing As to Bs and Bs
to As. Now there are b − ¾ − 1 additional As and a − ¾ additional Bs, so the total
for the modified list is (b − 1 , a + 1).
7
(4,7)
6
5 (6,5)

(2,3)

(2,2)

0
0 4 5 6
40 Chapter 1: Combinatorial Arguments

The switch reflects the part of the path after (¾ , ¾ + 1) through the line
y = x + 1. The new path ends at (b − 1 , a + 1). Since a + 1 > b − 1, each path
ending at (b − 1 , a + 1) rises above the line y = x. Reflecting the part of such a
path after it first visits a point of the form (¾ , ¾ + 1) generates a path to (a , b).
The second map inverts the first, so the bad paths to (a , b) correspond bijectively
to all paths reaching (b − 1 , a + 1). Hence there are (aa++1b) bad paths.

For the probability in Example 1.3.12, (a + 1)( aa++1b) = b(a+a b) yields

( a+a b) − (aa++1b) b a−b+1
=1− = .
( a+a b) a+1 a+1
The switching argument in Theorem 1.3.13 is called the reflection principle,
attributed to Antoine Désiré André [1887] (see Renault [2008] for what André
actually did). According to Feller [1968, p. 72], “ The reflection principle is used
frequently in various disguises, but without the geometrical interpretation it ap-
pears as an ingenious but incomprehensible trick.” That is, the argument works
as well using 0 , 1-lists; the reflection corresponds to complementing the final part
of the list. For the analogous problem with a > b and A being always ahead, the
reflection is through the line y = x and the formula is nicer (Exercise 5).
Theorem 1.3.13 leads to a short combinatorial proof of a beautiful identity
for the central binomial coefficients. It appears as early as a text by Whitworth
[1897] called Choice and Chance. In the 1930s P. Veress requested a combinato-
rial proof, found by G. Hajós (see Sved [1984], and K leitman [1975]). Example
3.2.13 presents another proof.

1.3.14. Lemma. The central binomial coefficient (2m )

m counts the following types
of lattice paths of length 2m that start at (0 , 0).
(A) Those ending at (m , m).
(B) Those never rising above the line y = x.
(C) Those never returning to the line y = x.
Proof: Type A is the familiar lattice path model. The number of paths of Type
B with length l (possibly l is odd) that end at (a , b) with a + b = l and a ≥ b is
(al ) − (a+l 1 ), by Theorem 1.3.13. Hence for the answer we sum over a and observe
that the sum telescopes. The computation is
l
l l l
∑ [( ) − (
a a+1
)] = (
⌈ l/2⌉
).
a=⌈l/2⌉

Paths of Type C first step to the right or up. If right, then after (1 , 0) the
rest has length 2m − 1 and does not rise above y = x − 1. Setting l = 2m − 1,
the number of these paths is (2mm−1). The same number first step up, so the total
number of Type C paths is 2(2mm−1 ). The Committee-Chair and Complementation
2m−1 2m−1
m = 2m( m−1 = 2m( m
Identities yield m(2m) ) ), so 2(2mm−1) = (2m
m
).

1.3.15. Theorem. For n ∈ 0 ,

n
2¾ 2n − 2 ¾
∑ ( ¾ )( n− ¾
) = 4n .
¾ =0
Section 1.3: Applications 41

Proof: There are 4n binary lists of length 2n; view them as lattice paths from
the origin. Since (2¾¾ ) is the number of lattice paths from the origin to (¾ , ¾), we
group the paths according to the point (¾ , ¾) where they last touch the line y = x.
Within such a path, the initial portion can combine with any path of 2n −
2 ¾ steps after (¾ , ¾) that does not return to y = x. By Lemma 1.3.14, there are
(2nn− 2¾
−¾ ) such paths. Hence the sum counts all paths with 2n steps, grouped by the
position of the last visit to the line y = x.

CATAL AN NUMBERS

In the special case a = b = n of Theorem 1.3.13, the number of good paths is

1
(
n+1 n
2n
). These numbers turn up surprisingly often.

1.3.16. Definition. The Catalan sequence is defined by Cn = n+1 1 (2n n ) for n ≥ 0.

The number Cn is called the nth Catalan number.
A ballot path of length 2n is a lattice path from (0 , 0) to (n , n) that never
rises above y = x. A ballot list of length 2n is a list of n 1s and n 0s such
that each initial segment has at least as many 1s as 0s.

Netto [1901] named Catalan numbers in honor of Eugène Charles Catalan

(1814–1894), (see Banderier–Schwer [2005]). Catalan [1838] studied them but
called them “Segner numbers” after Segner [1759], who obtained a recurrence.
For further discussion of the history, see I. Pak’s appendix to Stanley [2015a].
The correspondence between paths and lists in Example 1.2.2 restricts to
a correspondence between ballot paths and ballot lists. In Theorem 1.3.13, we
proved that the nth Catalan number Cn counts both the ballot paths of length 2n
and the ballot lists of length 2n. Our next proof of this yields more general results
(see Hilton–Pederson [1991] and Graham–Knuth–Patashnik [1989, 359–363]).

1.3.17. Theorem. If p and q are relatively prime, then the number of lattice
paths from (0 , 0) to (p , q) that do not rise above the line py = qx is p+1 q ( p+p q).
Proof: Exactly ( p+p q) words consist of p 1s and q 0s. Put them in groups; grouped
with A are all words obtained by reading A cyclically from some starting point.
If the rotations of A are not distinct, then A is periodic, making p and q both
divisible by the number of periods. Since gcd(p , q) = 1, each group has size p + q.
There are thus p+1 q ( p+p q) sets of words. Since the words correspond bijectively
to the lattice paths to (p , q), it suffices to show that each set contains exactly one
word that corresponds to a good path.

• • (2p , 2q)
• •
• • •
P
• • •
(p , q) = (5 , 3)
• •
• • •
P
(0 , 0) • •
42 Chapter 1: Combinatorial Arguments

Given a lattice path P from (0 , 0) to (p , q), let P be the translation of P start-

ing from (p , q). The subpaths of P ∪ P having length p + q that start at integer
points along P correspond to words in the same set.
Let L be the line {(x , y): py − qx = 0}. Since gcd(p , q) = 1, each line parallel
to L contains at most one integer point of P. (Otherwise, the difference (a , b)
satisfies a/b = p/q with a < p, but p/q is already in lowest terms.)
Each line parallel to L has equation py − qx = c for some c. The path P ∪ P
lies below the line parallel to L that has the highest value c among those hitting
P , and it does not lie below any other line in this family. This selects a unique
starting position among the cyclic rotations of P so that when the starting point
is translated to the origin the path stays below the line py = qx.

1.3.18. Definition. A q-ballot list is a list of n 0s and qn 1s where every initial

segment has at least q times as many 1s as 0s (q = 1 yields ballot lists).

1.3.19. Corollary. There are 1

qn+1
((q+n1)n) q-ballot lists of length (q + 1)n.
Proof: Such lists correspond to lattice paths from (0 , 0) to (qn , n) not rising above
qy = x, but gcd(qn , n)
= 1. Instead, consider paths from (−1 , 0) to (qn , n) not
rising above (qn + 1)y = n(x + 1). Such a path starts horizontal and then follows
a good path, since an integer point (a , b) above qy = x but not (qn + 1)y = n(x + 1)
requires bq ≥ a + 1 and b(qn + 1) ≤ n(a + 1), a contradiction.
We therefore count paths from (−1 , 0) to (qn , n) not rising above (qn + 1)y =
n(x + 1). Moved one step rightward, these correspond to paths from (0 , 0) to
(qn + 1 , n) not rising above (qn + 1)y = nx. Since gcd(qn + 1 , n) = 1, by Theo-
(q+1)n+1
+1 ( qn+1 ) such paths. The formula simplifies by the
1
rem 1.3.17 there are (q+1)n
Committee-Chair and Complementation Identities to the given expression.

The numbers qn1+1 ((q+n1)n) are the generalized Catalan numbers or Fuss–
Catalan numbers (first studied by Fuss [1791]). See Goulden–Serrano [2003]
for other instances, also Exercises 43–45.
A related argument yields the same generalizations and others. We state the
result here and leave the proof and applications to Exercises 37–39.

1.3.20. Theorem. (Cycle Lemma; Dvoretzky–Motzkin [1947]) For n , m , ¾ ∈ 0 

with m ≥ ¾ n, every arrangement of m 1s and n 0s in a circle has exactly m −
¾ n positions such that every clockwise segment starting there has more than
¾ times as many 1s as 0s.
0 0
1 1
1 1
1 0
1 1

The example above has (m , n , ¾) = (7 , 3 , 2), with the good position marked.
Theorem 1.3.20 provides a bijection from q-ballot lists (with an initial 1 added)
to cyclic arrangements with qn + 1 1s and n 0s. Since the number of such cyclic
arrangements is qn1+1 ((q+n1)n), Corollary 1.3.19 follows (see Peck [1989]).
Section 1.3: Applications 43

For the Catalan numbers themselves, Exercise 6.19 of Volume 2 of Enumer-

ative Combinatorics by Stanley [1999] lists 66 counting problems they solve. In
a later book, Stanley [2015a] expanded the list to 214 combinatorial interpreta-
tions of the Catalan numbers. See also www-math.mit.edu/ ∼rstan/ec/catadd.pdf.
Below are a few classical examples.

Ballot lists / ballot paths of length 2n (Corollary 1.3.19)

Binary trees of n + 1 leaves (grouping n + 1 terms) (Theorem 1.3.23)
Ordered trees with n edges (Example 1.3.25)
Triangulations of a convex (n + 2)-gon (Example 1.3.24)
Noncrossing pairings of 2n points on a circle (Exercise 31)
Noncrossing partitions of [n] (Exercise 41)
Stack-sortable permutations of [n] (Exercise 2.1.48)

Two of these examples use modified notions of trees.

1.3.21. Definition. A rooted graph has one vertex distinguished as a root. In

a tree with root r, the neighbor of a vertex v on the path from v to r is the
parent of v, and the other neighbors of v are its children. A leaf in a rooted
tree is a vertex with no children. An ordered tree is a rooted tree in which
the children of each vertex are given a fixed (left-to-right) linear order. A
binary tree is an ordered tree in which every vertex has zero or two children.

The root in a rooted tree has no parent; it is considered a leaf when it has
no children. Ordered trees have also been called rooted plane trees; the word
“plane” suggests the fixed ordering on the children when the tree is drawn in the
plane. In a binary tree, the left subtree and right subtree are the subgraphs
obtained by deleting the root r; they are rooted at the left and right children of r,
respectively. In some contexts in computer science, a vertex in a binary tree may
have exactly one child, designated “left ” or “right ”.

1.3.22. Example. A binary tree iteratively groups or “parenthesizes” its leaves.

Each matched pair of parentheses corresponds to a non-leaf non-root vertex of the
tree where the sets of leaves for two subtrees are “combined”. We omit the outer
parentheses corresponding to the combining operation at the root.
Below are the five binary trees with four leaves, the corresponding five ballot
lists of length 6 under the bijection in Theorem 1.3.23, and the expressions of the
trees as binary groupings of factors at the leaves.
A binary tree can be grown from the root by iteratively giving a leaf two
children. The numbers of leaves, non-leaves, and edges grow by 1, 1, and 2, re-
spectively. Thus a binary tree with n + 1 leaves has n non-leaf vertices and 2n
edges (the root is a leaf when n = 0).

((ab)c)d (a(bc))d (ab)(cd) a((bc)d) a(b(cd))

• • • • •
• • • • • • • • • •
• • • • • • • • • • • •
• • • • • • • •
111000 110100 110010 101100 101010
44 Chapter 1: Combinatorial Arguments

1.3.23. Theorem. The number of binary trees with n + 1 leaves is the nth Cata-
lan number, n+1 1 (2n
n ).
Proof: Using induction on n, we establish a bijection ºn from the set An of binary
trees with n + 1 leaves to the set Bn of ballot lists of length 2n (we know the latter
has size Cn). For n = 0, the 1-vertex tree corresponds to the empty ballot list.
Consider n > 0. A tree T in An has left and right subtrees T l and Tr with ¾
and n + 1 − ¾ leaves, for some ¾ ∈ [n]. Under the bijections º¾−1 and ºn−¾ provided
by the induction hypothesis, these subtrees yield ballot lists of lengths 2 ¾ − 2 and
2n − 2 ¾ . Let ºn(T) = (1 , º¾−1 (T l) , 0 , ºn−¾ (Tr)). The result is a ballot list of length
2n, since the sublists are ballot lists and the added 1 precedes the added 0.
To prove that ºn is a bijection, we argue that for a ballot list b of length 2n
exactly one binary tree T satisfies ºn(T) = b. This requires the initial 1 in b to
be paired with a 0 that occurs immediately after the ballot list corresponding to
T l and before the list corresponding to Tr . By the induction hypothesis, there is
exactly one choice for T l and one for Tr . Hence T is determined uniquely.

The key step for inductively producing the inverse of a bijection to the set of
ballot lists is to decompose the list using the unique 0 that “matches” the initial
1 by the first equality in the bits of each type (first return to the diagonal). In
Theorem 1.3.23, this discovers two smaller ballot lists to be used to find the in-
verse image (this is how one distinguishes the images of left and right subtrees
in a list like 10101010).
Bijections involving binary trees are especially easy to verify inductively.
This yields the next classical instance of the Catalan numbers.

1.3.24. Example. There are n+1 1 (2n

n ) triangulations of a convex
(n + 2)-gon. This is
the counting problem solved by Catalan. The result is surprising, but the bijec-
tion to a problem we already solved is easy to verify.
A parenthesized string of n + 1 letters corresponds to a binary tree with n + 1
leaves. With one root edge fixed, a triangulation of an (n + 2)-gon yields such a
tree in a natural way as suggested by the figure below. Exercise 30 requests the
proof that the resulting map is a bijection.
c
• •
b bc d

• de •
(bc)(de)
a e
• •
a((bc)(de))

An argument like that of Theorem 1.3.23 counts ordered trees with n edges.

1.3.25. Example. There are n+1 1 (2n )

n ordered trees with n edges. To define an in-
ductive bijection to the set of ballot lists, cut the edge in T from the root to its
leftmost child, forming ordered subtrees with ¾ − 1 and n − ¾ edges. By the induc-
tion hypothesis, these have corresponding ballot lists b and b of lengths 2 ¾ − 2
and 2n − 2 ¾ ; let º (T) = (1 , b , 0 , b ).
Section 1.3: Applications 45

This bijection motivates our claim that bijective arguments give more infor-
mation than inductive proofs. We record 1 when an edge is traveled to a subtree,
then the subtree is traversed, then a 0 is recorded when following the edge back
up, and finally we continue to the next subtree. Each edge contributes a 1 and
later a 0. A 1 is followed immediately by its matching 0 if and only if the sub-
tree has only one vertex and is a leaf. Thus the number of leaves in the ordered
tree is the number of times that 0 follows 1 in the corresponding ballot list. Our
bijection is “refined” by describing the images of subsets of interest.

In a binary list, a run is a maximal sublist of consecutive equal entries.

1.3.26.* Theorem. The number of ordered trees with n edges and ¾ leaves, called
the Narayana number Nn ,¾ , is 1¾ (¾−n1 )(¾n−−11).
Proof: As observed in Example 1.3.25, it suffices to count the ballot lists of length
2n that have ¾ runs of 1s (and ¾ runs of 0). Appending a 0 to the end of the list
yields a list whose corresponding path reaches (n , n + 1), still with ¾ runs of each
bit. As in Corollary 1.3.19, these are the paths reaching (n , n + 1) that do not
step above the line ny = (n + 1)x, and there is exactly one place to cut the cyclic
arrangement of any list with n 1s and n + 1 0s to obtain the list for such a path.
Hence it suffices to count the cyclic arrangements of these bits with ¾ runs
of each type. Consider starting at a run of 1s. The 1s form a composition of n
with ¾ parts, and the 0s form a composition of n + 1 with ¾ parts in the spaces
−1
following runs of 1s. By Corollary 1.1.22, there are (n¾− n
1 )(¾ −1 ) ways to form these
compositions. Starting at any one of the ¾ runs of 1s yields a cyclic rotation of the
same arrangement, so we divide by ¾ . Because n and n + 1 are relatively prime,
there is no periodicity, and each cyclic arrangement arises in ¾ ways.

The Narayana numbers are named after Narayana [1955], though they were
studied earlier (in more generality) by MacMahon [1916]; they arise again in Ex-
ercise 41 and in our last example.

1.3.27.* Example. Ballot paths have an alternative classical formulation. A

Dyck n-path is a path in the plane from (0 , 0) to (2n , 0) that takes n upsteps by
(1 , 1) and n downsteps by (1 , −1) and never falls below the axis y = 0. A ballot
path of length 2n becomes a Dyck n-path by turning horizontal steps (away from
the diagonal) into upsteps and vertical steps (toward the diagonal) into down-
steps. Each Dyck n-path thus arises from a unique ballot path, so there are Cn
Dyck n-paths. Dyck paths are named for Walther von Dyck (1856–1934).
The Dyck model is more natural for some concepts. A peak in a Dyck path is
an up-down subpath, while a valley is a down-up subpath. Left turns and right
turns in a ballot path become peaks and valleys in a Dyck path, which are natural
concepts. The Narayana number Nn ,¾ counts the Dyck n-paths having exactly ¾
peaks. The height of a peak or valley is the vertical coordinate of its midpoint.
See Exercise 32.

We have noted that the Catalan numbers arise in many counting problems,
such as in these exercises. Also, the formula for Cn can be found in many ways.
Indeed, a theme of this text is the flexibility of using various techniques to solve a
46 Chapter 1: Combinatorial Arguments

combinatorial problem. We therefore solve some problems repeatedly, such as de-

n
riving the Catalan formula (see Section 2.1 and Section 2.2), computing ∑i=1 i2 ,
studying the Delannoy numbers and the “derangement ” problem, and enumerat-
ing various types of permutations and partitions.

EXERCISES 1.3

1.3.1. (−) A box has ¾i letters of type i, for 1 ≤ i ≤ n. Count the words (using all letters in
the box) that have no two letters of type n adjacent.
1.3.2. (−) Given positive integers ¾1 , . . . , ¾m , how many ways are there to partition a set
of n distinct objects so that there are ¾i blocks of size i, for 1 ≤ i ≤ m?
1.3.3. (−) Prove bijectively that the Catalan number Cn counts these sets:
(a) Nondecreasing functions º : [n] → [n] such that º (i) ≤ i for all i.
(b) Nonnegative integer sequences of length 2n + 1 starting and ending at 0 such that
consecutive entries differ by 1.
(c) Arrangements of 2n people in two rows of n such that heights increase in each row
and in each column.
1.3.4. (−) Generate a random list b1 , . . . , bn in the following way. For 1 ≤ ¾ ≤ n, choose b¾
to be any of the ¾ values in [¾], equally likely. Obtain a simple formula for the probability
that the resulting list is nondecreasing.
1.3.5. (−) In Bertrand’s Ballot Problem (Example 1.3.12) with outcome (a , b) and votes
counted in random order, suppose that a > b. What is the probability that A is always
ahead of B (after the start)?
1.3.6. (−) A fair coin is flipped 2n times, consecutively. Compute the probability that the
lead changes, given that the final total is n heads and n tails.
1.3.7. (−) Prove the identities below by counting in two ways. (Whitworth [1897])
2 ¾ 2n − 2 ¾
n
1 2n + 1 1 2 ¾ − 2 2n − 2 ¾ + 1 2n
(a) ∑ ( )( )=( ) . (b) ∑ ( )( )=( ).
¾
¾+1 ¾ n−¾ n ¾ ¾−1 ¾ =1
n−¾ n−1

1.3.8. (♦) Prove bijectively that the number of graphs with vertex set [n] in which all
n−1
vertices have even degree is 2( 2 ) .
1.3.9. Among the trees with vertex set [n], use Corollary 1.3.7 to count those having n − 2
leaves and those having two leaves.
1.3.10. (♦) Let º (d1 , . . . , d n) be the number of trees with vertex set [n] in which, for each i,
the degree of i is d i . Use induction on n to prove directly that º (d1 , . . . , d n) = ( d −1 n,...
−2
,d n −1
).
1
n−2
Use this to obtain Cayley’s Formula n for the number of trees with vertex set [n].
1.3.11. Let G be a graph with n vertices and (n2) − 1 edges. Use Cayley’s Formula to prove
that G has exactly (n − 2)nn−3 spanning trees.
1.3.12. (♦) For a tree T with vertex set [n], the Pr üfer code ½(T) is formed by repeatedly
deleting the least leaf in the remaining tree and recording its neighbor, n − 2 times.
(a) Prove that ½ is a bijection from the set of trees with vertex set [n] to the set [n]n−2 .
(Comment: Pr üfer ’s argument appears in Biggs–Lloyd–Wilson [1976].)
(b) Prove that a tree T with vertex set [n] has {n − 1 , n} as an edge if and only if the
last entry in ½(T) is n − 1 or n. Use this to solve Exercise 1.3.11.
Exercises for Section 1.3 47

1.3.13. For n ≥ 4, construct a tree T with vertex set [n] whose Pr üfer code ½(T) (Exercise
1.3.12) differs from the list º (2) , . . . , º (n − 1) corresponding to it in Theorem 1.3.4.
n−1
1.3.14. (♦) Prove that t n = ∑¾=1 ¾ (¾n−−12) t¾ t n−¾ , where t n is the number of trees with vertex
set [n]. (Dziobek [1917])

1.3.15. (♦) Let t n be the number of trees with vertex set [n]. Explain the identity below
n−1
in terms of t n . Prove it. Conclude that t n = 2n ∑¾=1 (¾n−−12) t¾ t n−¾ . (Comment: The result
appears in Lovász [1979]; the combinatorial argument is due to L. Smiley.)
n−1
n
2(n − 1)nn−2 = ∑ ( )¾ ¾−1 (n − ¾)n−¾−1
¾
¾ =1

1.3.16. Let S be the set of pairs consisting of a rooted tree with vertex set [n] and a marked
vertex of that tree. Count S in two ways to prove that
n−1
n
nn = ∑ ( ) ¾ ¾(n − ¾)n−¾−1 . (Rey [1997])
¾
¾ =0

n−1
(Comment: Using this on forests, Smiley proved n(n + 1)n−1 = ∑¾=0 (n¾)(¾ + 1)¾ (n − ¾)n−¾−1 ,
which appears in Riordan [1968, p. 116].)

1.3.17. (+) Cayley’s Formula generalized. A forest is a graph whose components are trees.
Let (n , ¾) be the set of rooted forests with vertex set [n] in which the roots of the com-
ponents are the vertices 1 , . . . , ¾ , and let an ,¾ = | (n , ¾)|. This exercise develops three
distinct proofs that an ,¾ = ¾ nn−¾−1 . The formula appears in Cayley [1889].
(a) Use induction on n, with basis an ,n = 1. (Hint: Delete the roots.) (David [2007])
(b) Prove (¾ + 1)!an ,¾ = ¾ !n¾ an ,¾+1 and use an ,n = 1. (Hint: The operation obtaining
members of (n , ¾ + 1) from members of (n , ¾) will need to be symmetrized by permuting
[¾ + 1].) (Lovász [1979, problem 4.14]).
(c) Generalize Exercise 1.3.12(a). For F ∈ (n , ¾), form a list by iteratively deleting
the largest (nonroot) leaf and recording its neighbor, until only the roots remain. Prove
that this establishes a bijection from (n , ¾) to [n]n−¾−1 × [¾].

1.3.18. (♦) Let F be a forest with vertex set [n] and t components having n1 , . . . , nt ver-
tices. Prove that exactly nt−2 ∏i=1 ni trees with vertex set [n] contain F . (Moon [1967])
t

1.3.19. (+) Let M be the set of pairings of [2n], and let T be the set of “unordered” binary
trees with leaf set [n + 1]. Each tree has n unlabeled nonleaf vertices and 2n edges; nonleaf
vertices have exactly two children. “Unordered” means that the tree is unchanged when
we exchange left and right subtrees at a vertex.
(a) Show that |M| = (2n)!/(n!2 n).
(b) Prove bijectively that |T| = |M|. (Hint: First extend the labeling of the leaves of
such a tree to the full vertex set. When 1 , . . . , r have been used, assign r + 1 to the unla-
beled vertex that , among those with two labeled children, has the child with the smallest
label.) (Schröder [1870], Erdős–Székely [1989])

1.3.20. Let p be a prime, and let the p-ary expansions of positive integers n and ¾ be given
by n = ∑ ai pi and ¾ = ∑ bi pi .
(a) Use (x + 1)p ≡ (x p + 1) (mod p) to prove (¾n) ≡ ∏ ( abii ) (mod p). (Lucas [1878])
(b) Use part (a) to determine when (¾n) is odd. Determine also for which n the binomial
coefficients (n0) , . . . , ( nn) are all odd.
48 Chapter 1: Combinatorial Arguments

1.3.21. Let p be a prime. Part (a) of Exercise 1.3.20 implies ( pmpl

) ≡ (ml) (mod p).
Strengthen this by giving a combinatorial proof that ( pm
pl
) ≡ (ml) (mod p2). (Hint: Choose
points from an m-by-p grid.) (Stanley [1986, p. 53])

1.3.22. For a ∈ , use cyclic rotations of p-tuples to prove that p divides ap − a when p is
prime. (Comment: This yields a combinatorial proof of Fermat ’s Little Theorem.)
1.3.23. Let p be a prime, and let n1 , . . . , nm be integers summing to n. Consider their p-
¾ ¾
ary expansions: n = ∑ j =0 aj p j and ni = ∑ j =0 ai , j p j .
(a) Prove that the multinomial coefficient ( n ,...n ,nm) is not divisible by p if and only if
1
aj = ∑i=1 ai , j for 0 ≤ j ≤ ¾ . (Dickson [1902])
m

(b) Prove that the number of terms in the expansion of (x1 + x2 + · · · + x m)n whose
a + m−1
coefficients are not divisible by p is ∏j ( j m−1 ) . (Howard [1974])

1.3.24. By comparing sets of words, determine which coefficients in the expansion of

(x1 + · · · + x¾)n are the largest.
1.3.25. (♦) Prove bijectively that n pairwise-intersecting lines in 2 (no three at a point)
cut the plane into 1 + n + ( n2) regions. (Steiner [1826])

1.3.26. Generalize the formula of Exercise 1.3.25 by finding the number of regions in d
formed by n hyperplanes in general position. That is, no d + 1 of them have a common
point , and for ¾ ≤ d any ¾ of them intersect in a (d − ¾)-dimensional plane. (Schläfli [1852])
1.3.27. In an election where A receives a votes and B receives b votes, determine the prob-
ability that A never trails B by more than ¾ votes during the counting, assuming that all
orderings of the votes are equally likely.
1.3.28. (♦) Establish a direct bijection from the binary trees with n + 1 leaves to the or-
dered trees with n+ 1 vertices (Definition 1.3.21). Below are the instances of each for n = 2.
(Comment: By Theorem 1.3.23, the Catalan number Cn thus counts each set.) (Bernhart)
• • • •
• • • • • • •
• • • • •

1.3.29. (♦) The bijections in Exercise 1.3.28 and Example 1.3.25 show that there are
n+1 ( n )
1 2n
ordered trees with n + 1 vertices. Drawing them all yields a total of (2n
n)
vertices.
Prove that exactly half of these vertices are leaves in their trees. (Shapiro [1999])
1.3.30. Prove that the map suggested in Example 1.3.24 is a bijection from the set of tri-
angulations of a convex (n + 2)-gon to the set of binary trees with n + 1 leaves, thereby
showing that there are Cn such triangulations.
1.3.31. (♦) Place 2n points on a circle. Prove bijectively that the number of ways to pair
the points by drawing noncrossing chords equals the number of ballot lists of length 2n.
The possibilities for n = 3 appear below.
• • • • • • • • • •
• • • • • • • • • •
• • • • • • • • • •

1.3.32. (♦) Let En and On , respectively, be the number of Dyck n-paths (Example 1.3.27)
having an even or an odd number of peaks at even height. Compute En and On in terms of
the Catalan numbers. (Deutsch [2005])
Exercises for Section 1.3 49

1.3.33. (♦) Prove that the number of ballot lists of length 2n that are unchanged when
they are reversed and complemented is (⌈n/2⌉
n
).
1.3.34. (♦) Lattice walks in the plane. (See Definition 1.2.1; lattice walks can move in any
direction, while lattice paths only increase coordinates.)
(a) A positive lattice walk is a lattice walk that starts at the origin and never falls
below the horizontal axis. Prove that the number of positive lattice walks of length ¾ is
¾
∑ j =0 (¾j )(⌈ j/2⌉
j
)2¾− j . (Hint: Consider Lemma 1.3.14. Comment: This sum was shown com-
binatorially to equal (2¾¾+1 ) in Exercise 1.2.29 and will be evaluated by other means in both
Exercise 2.1.52 and Exercise 3.2.46.)
(b) In three dimensions, determine the number of walks of length n that start at the
origin and don’t fall below the horizontal plane. (Deutsch [2000])
1.3.35. (♦) Generalization of ballot paths.
(a) Prove that the number of lattice paths from (0 , 0) to (n , n + ¾) that never pass above
the line y = x + ¾ in the plane is n+¾+¾+1 1 (2nn+¾ ).
(b) A population grows from one individual of type ¾ in generation 0. For i ≥ 0, each
individual of type i in generation t produces one individual with each type 1 , . . . , i + 1 in
generation t + 1. For ¾ ≥ 0 and n ≥ 1, prove that the total number of individuals in gener-
ation n is the same as the answer in part (a). (Hint: Trace the ancestry of each individual.)
(See Kupka [1990], Beckwith [2006b])


1.3.36. For q ∈ , determine the number of lattice paths from (0 , 0) to (n , qn) not rising
above the line y = qx.
1.3.37. (♦) Prove the Cycle Lemma (Theorem 1.3.20). (Hint: Use induction.)
1.3.38. Apply the Cycle Lemma (Theorem 1.3.20) to solve Bertrand’s Ballot Problem (Ex-
ample 1.3.12).

1.3.39. For n ∈ , count the lists (a1 , . . . , an) of positive integers with a1 = 1 such
that ai − ai−1 is odd and at most 1 for i > 1. For example, for n = 5 the lists are
{12345 , 12343 , 12341 , 12323 , 12321 , 12123 , 12121}. (Deutsch [1999])

1.3.40. (+) Valid parenthesizations of x0 ÷ x1 ÷ · · · ÷ x n correspond to the Cn binary trees

with n + 1 leaves. Each such expression evaluates to a fraction with some variables in the
numerator (including x0) and the others (including x1) in the denominator.

(a) For n ∈ , determine which fractions occur most often.
(b) Prove that the number of occurrences of the most frequent fractions is the number
of nondecreasing nonnegative integer lists b0 , . . . , bn−1 such that bi ≤ i/2 for 0 ≤ i ≤ n − 1.
(c) Count the lists described in part (b). (Callan [2004])

1.3.41. (♦) Noncrossing partitions of [n]. A partition of [n] is noncrossing if there are no
a , b , c , d with a < b < c < d such that a and c are in one block and b and d are in another.
Thus (14|25|3) is crossing and (15|234) is noncrossing.
A B A
• • •
B
C •
• • • • • • • • • •
1 2 3 4 5 1 2 3 4 5
Establish a bijection from the set of noncrossing partitions of [n] to the set of ballot lists
of length 2n. From the bijection, argue that the number of noncrossing partitions of [n]
with ¾ blocks equals the number of ballot lists of length 2n with ¾ runs of 1. Thus the
value is the Narayana number Nn ,¾ of Theorem 1.3.26, counting the ordered trees with n
edges and ¾ leaves. (Prodinger [1983])
50 Chapter 1: Combinatorial Arguments


1.3.42. For n , t ∈ , count the nondecreasing integer lists (a1 , . . . , an) such that 1 ≤ ai ≤
ti for 1 ≤ i ≤ n. (Deutsch [2004a])
1.3.43. Establish a bijection from the set of q-ballot lists of length (q + 1)n to the set of
ordered (q + 1)-ary trees with qn + 1 leaves. The trees correspond to groupings of qn + 1
items, combining q + 1 at a time, shown below for q = n = 2. (Sands [1978])
• • •
• • •
• • • • • • • • • • • • • • •
a b c d e a b c d e a b c d e
((abc)de) ↔ abc| de| (a(bcd)e) ↔ abcd| e| (ab(cde)) ↔ abcde||

1.3.44. Consider an election where A receives ¾ n votes and B receives n votes. Given that
the vote order is random, let p be the probability that throughout the counting the number
of votes recorded for A is always at most ¾ times the number of votes recorded for B. Let q
be the analogous probability with “at most ” replaced by “at least ”. Prove p = q.
1.3.45. (♦) Counting q-ballot paths by the number of turns.
(a) Let r and s be relatively prime positive integers. Prove that the number of lattice
paths from (0 , 0) to (r, s) that do not rise above the line ry = sx and turn exactly 2 ¾ − 1
times is 1¾ (¾r−−11 )(¾s−−11) .

(b) For q ∈ , prove that the number of lattice paths from (0 , 0) to (n , qn) that do not
rise above the line y = qx and turn exactly 2 ¾ − 1 times is 1¾ (n¾−−11 )(¾qn
−1 ) .
(Cigler [1987]; see
also Krattenthaler [1997])
1.3.46. (♦) Let x1 , . . . , x n be a cyclic arrangement of integers with sum 1.
(a) Prove that for all r ∈ [n], there is exactly one place to break the cycle into a lin-
ear arrangement with exactly r positive partial sums. Let pj be the position from which
j partial sums are positive. Show that if r > s, then the positions ending positive partial
sums from position pr include all the positions ending positive partial sums from ps . For
example, for [x1 , . . . , x5 ] = [3 , −4 , 1 , 2 , −1], we have (p1 , p2 , p3 , p4 , p5) = (2 , 5 , 1 , 4 , 3).
For the ending positions of the positive partial sums from each pi , we have p1 = 2: {1};
p2 = 5: {1 , 4}; p3 = 1: {1 , 4 , 5}; p4 = 4: {1 , 3 , 4 , 5}; p5 = 3: {1 , 2 , 3 , 4 , 5}.
(b) Let all lists of n As and n Bs be equally likely. Let X be the random variable
counting the values i such that the ith A precedes the ith B. Apply part (a) to prove that
Prob(X = l) = 1/(n + 1) for each l ∈ {0 , . . . , n}. (Chung–Feller [1949])
(Comment: Many results similar or equivalent to parts of (a) have appeared: see
Spitzer [1956], Raney [1960], Kierstead–Trotter [1988], Mont ágh [1991], Snevily–West
[1998], and Chapter 3 of Mohanty [1979].)
1.3.47. (♦) Consider : [n] → [n]. Drivers 1 , . . . , n in order enter a street with parking
spots 1 , . . . , n. Driver i starts looking for parking at spot (i) and takes the first open spot.
If none is open from (i) to n, then Driver i fails. The functions such that all drivers
succeed are parking functions. For example, if ([9]) = (6 , 4 , 4 , 3 , 6 , 8 , 3 , 1 , 1) in order,
then all nine drivers succeed, parking in order in spots (6 , 4 , 5 , 3 , 7 , 8 , 9 , 1 , 2). Prove that
is a parking function if and only if there is a permutation  of [n] such that (i) ≤  (i)
for all i. (Riordan; see Knuth [1973, p. 545]) (Comment: More than 200 papers have been
written about parking functions and related topics; see Yan [2015] for a survey.)
1.3.48. (♦) Prove that the number of parking functions on [n] is (n + 1)n−1 (Pyke [1959],
Konheim–Weiss [1966]) (Hint: Add a position n + 1, which can be in the image of , and
move cars cyclically until they park; this proof is due to Pollak in the 1970s. Comment:
Foata–Riordan [1974] gave a correspondence with labeled trees.)
Chapter 2

Recurrence Relations
Often the solutions to a counting problem can be built from solutions to a
smaller problem of the same type. For example, let an count the 0 , 1-lists of length
n. A list of length n − 1 can extend by acquiring a 0 or a 1. The resulting lists are
distinct, and every list of length n arises in this way. Therefore, an = 2an−1 for
n ≥ 1. With a0 = 1 (there is one way to “do nothing ”), induction yields an = 2 n .
This example shows the steps in recursive solution of counting problems.
Combinatorial arguments express an in terms of earlier values and a function of
n. Along with initial values, this “recurrence” determines the sequence. “Solv-
ing ” the recurrence means obtaining an exact or asymptotic formula for an .
A solution formula for a recurrence can be verified by induction. When no
formula is apparent, solutions may arise by other techniques: the “characteristic
equation method” for special recurrences, a more general “generating function
method”, substitution techniques, special methods for asymptotic solutions, etc.
We first concentrate on recurrences with one (integer) parameter. We may
refer to the sequence of solutions to a counting problem as a counting sequence.

2.0.1. Definition. A recurrence relation or recurrence for a sequence a0 , a1 , . . .

is an expression of the form an = ½(n , a0 , . . . , an−1). It has order ¾ or degree
¾ if the formula for an depends only on n and an−1 . . . an−¾ . The recurrence
an = ½ 1 (n)an−1 + · · · + ½ ¾ (n)an−¾ + º (n) is linear if each ½ i (and º ) is indepen-
dent of ⟨a⟩. Here º (n) is the inhomogeneous term. If º is identically 0,
then the relation is homogeneous; otherwise, it is inhomogeneous.

2.0.2. Example. Below we illustrate these definitions using several recurrences

that we will derive for combinatorial problems.
an = an−1 + n inhomogeneous, linear, order 1
an = an−1 + an−2 homogeneous, linear, order 2
an = (n − 1)(an−1 + an−2) homogeneous, linear, order 2
an = ∑¾=1 a¾−1 an−¾
n
homogeneous, nonlinear, no finite order

2.0.3. Remark. A recurrence relation does not by itself specify a sequence; for
example, every constant sequence satisfies an = an−1 . A recursive definition must
include initial values that enable the recursive computation to proceed. A recur-
rence of order ¾ requires ¾ initial values to specify a sequence. We say that the

51
52 Chapter 2: Recurrence Relations

recursive formula is valid for larger values. Initial values correspond to the basis
step in inductive proofs; verifying that an explicit formula for an gives the initial
values is the basis step in an inductive proof that the formula is correct for all n.
In some contexts ½(a0 , . . . , an) = 0 is a more natural form for recurrences.
Here a recurrence is an operator, and one finds the sequences annihilated by it (its
¾
“nullspace”). Linear recurrences of order ¾ then become ∑i=0 ½ i(n)an−i = º (n);
the signs in the coefficients have changed. This convention is useful in the theory
of solving recurrences, but in this text we emphasize the combinatorial aspects
of describing a set using “earlier ” instances. Thus we choose the “ an = ” form.

2.0.4.* Remark. When studying integer sequences, one should know about the
On-Line Encyclopedia of Integer Sequences (OEIS), developed by Neil J. A. Sloane
and located at http://oeis.org. It contains enormous amounts of information
about nearly every counting sequence in this book, including the Fibonacci num-
bers (sequence A000045), Catalan numbers (A000108), Derangement numbers
(A000166), central Delannoy numbers (A001850), and many others. Readers can
consult the OEIS for more information about special sequences and contribute
interesting sequences not yet among the more than 100,000 sequences in the
database. The welcome page describes the database and gives instructions for
many ways of interacting with it.

2.1. Obtaining Recurrences

Obtaining a recurrence for a counting problem usually involves bijective ar-
guments. Some creativity may be needed; there is no algorithm. We give classical
examples to illustrate typical arguments.

CL ASSICAL EX AMPLES

In the simplest situation, our problem has one parameter, and we seek a sin-
gle recurrence for the resulting counting sequence.

2.1.1. Example. Regions in the plane.

an = an−1 + n for n ≥ 1, with a0 = 1.
Let an n-configuration be a set of n lines in the plane such that every two
lines have one common point but no three lines have a common point. Let an be
the number of regions formed by an n-configuration. It is not obvious that every
n-configuration forms the same number of regions; this follows inductively when
we establish a recurrence for an.
Section 2.1: Obtaining Recurrences 53

With no lines, a0 = 1. To prove an = an−1 + n for n ≥ 1 (and that all n-con-

figurations form the same number of regions), consider an n-configuration with
n ≥ 1, and let L be one line. The other lines form an (n − 1)-configuration; by the
induction hypothesis, they create an−1 regions. The intersections of L with the
other n − 1 lines cut L into n parts; each splits a region. Thus adding L increases
the number of regions by n, and all n-configurations have an−1 + n regions.

2.1.2. Example. Fibonacci numbers.

F n = F n−1 + F n−2 for n ≥ 2, with F0 = 0 and F1 = 1.
F̂ n = F̂ n−1 + F̂ n−2 for n ≥ 2, with F̂0 = F̂1 = 1 .
Imagine a linear parking lot with n spaces filled by two types of cars. Rab-
bits (sold by Volkswagen in the United States and Canada beginning in 1974) are
small cars taking one space. Cadillacs (manufactured since 1902 and by General
Motors since 1909) are large cars taking two spaces. Let F̂ n count the distin-
guishable ways to fill the parking lot. These correspond to the 1 , 2-lists with sum
n (also called compositions of n using parts in {1 , 2}); we emphasize this combina-
torial model. The resulting sequence solves many problems; for example, F̂ n also
counts the binary (n − 1)-tuples with no consecutive 1s (Exercise 14).

2
n
In a 1 , 2-list with sum n, the last element may be 1 or 2. There are F̂ n−1 ways
to fill the earlier part if the list is 1 and F̂ n−2 ways if it is 2. Thus F̂ n = F̂ n−1 + F̂ n−2
for n ≥ 2. Only the empty list has sum 0, and one list has sum 1, so F̂0 = F̂1 = 1.
Changing the initial conditions to F0 = 0 and F1 = 1 yields another sequence
⟨F⟩. Since F1 = F̂0 and F 2 = F̂1 , inductively F n = F̂ n−1 for n ≥ 1. The sequence
⟨F⟩ is the classical Fibonacci sequence introduced in 1202 by Leonardo of Pisa
(ca. 1170–1250), known as Fibonacci. The Fibonacci recurrence F n = F n−1 + F n−2
is a natural model for growth from two previous stages of a process and occurs fre-
quently in nature. The journal The Fibonacci Quarterly is devoted to the Fibonacci
numbers and related topics. Extensive material on their history, mathematics,
and applications appears in Koshy [2001]; we study them in Exercises 14–34.
Since F n = F̂ n−1 , both sequences use the same numbers, and we call ⟨ F̂⟩ the
adjusted Fibonacci numbers. Benjamin–Quinn [2003] also uses special nota-
tion to distinguish the adjusted and classical indexings, writing ºn for our F̂ n .
The initial conditions F0 = 0 and F1 = 1 are more common in the literature
and are appropriate in number theory due to divisibility properties (such as Ex-
ercise 22). Other initial conditions have been studied; the Lucas numbers arise
from the Fibonacci recurrence with initial conditions L1 = 1 and L2 = 3; see
Exercise 27. (The term “Fibonacci numbers” was popularized by Lucas.)

Many identities can be proved easily by induction, but combinatorial argu-

ments that present a set counted by both formulas can be more illuminating or
provide more detailed results (see Theorem 1.3.26, for example). We illustrate
this notion with Fibonacci numbers.
54 Chapter 2: Recurrence Relations

2.1.3. Application. Combinatorial arguments for Fibonacci identities. The ini-

tial conditions we use for ⟨ F̂⟩ facilitate combinatorial arguments using the model
of 1 , 2-lists with sum n. For example, consider the identity ∑i=0 F̂ i2 = F̂ n F̂ n+1 for
n

n ≥ 0. One can easily prove this by induction after replacing F̂ n+1 with F̂ n + F̂ n−1 .
Combinatorially, the product F̂ n F̂ n+1 counts the ways to choose two 1 , 2-lists,
with sums n and n + 1 (as in a parking lot with two rows). In a 1 , 2-list as a tiling
with segments of lengths 1 and 2, we view each partial sum as a “breakpoint ”. In
F̂ n2 of the pairs, both lists have a breakpoint at n; these are the pairs where the
second list ends with 1. Among those where the second list ends with 2, there are
F̂ n2−1 where both lists have a breakpoint at n − 1, and so on.
When the last common breakpoint is known, there is only one way to complete
the lists after that; it uses one 1 and the rest 2s. Thus there are exactly F̂ i2 pairs
where the last common breakpoint is at i, which proves the identity.

i n

Next we turn to the famous Hat-Check Problem. If n people check their

hats and retrieve them at random, what is the probability that each one gets
someone else ’s hat? The new assignment of hats gives a permutation of the in-
dex set [n], expressed as a bijection from [n] to [n]. In the functional digraph
(Definition 1.3.3), each number has one successor and one predecessor. Hence the
functional digraph consists of disjoint cycles. Applying the function repeatedly
sends an element x around its cycle. The successive elements form the orbit of x.

2.1.4. Definition. The cycles of a permutation are the orbits of elements under
iteration of the permutation; these are the (ordered) vertex sets of the cycles
in the functional digraph. A fixed point is a cycle of length 1. A derange-
ment is a permutation with no fixed points. The number of derangements of
an n-set is denoted by Dn .

2.1.5. Example. Viewing 641253798 as a permutation of [9], the cycles are (163),
(24), (89), (5), and (7), with 5 and 7 being fixed points.

In the Hat-Check Problem, all n! permutations are equally likely, and we

need to count those with no fixed points. The word “derangement ” is used be-
cause it is a variation on “arrangement ” and because “deranged” means “crazy”
or “mixed-up”: that is, nothing is in its place.

2.1.6. Example. Derangements of n objects.

Dn = n! − ∑¾=1 (¾n)Dn−¾ for n ≥ 1, with D0 = 1.
n

Dn = (n − 1)(Dn−1 + Dn−2) for n ≥ 2, with D0 = 1 and D1 = 0.

One way to form the n! permutations of [n] is to choose the fixed points and
then derange the other elements. If we fix all the points, then there is one way to
derange the remaining 0 elements, so it is sensible to set D0 = 1. Now counting
two ways yields n! = ∑¾=0 (¾n)Dn−¾ , which is the first recurrence above.
n
Section 2.1: Obtaining Recurrences 55

This recurrence needs more computation for each successive value; fixed or-
der is preferable. With D0 = 1 and D1 = 0 by inspection, we derive a second-order
recurrence valid for n ≥ 2. We count the ways to partition [n] into cycles of length
at least 2. There are two cases.
If the cycle containing the element n has length 2, then there are n − 1 ways
to pick its other element, and there are Dn−2 ways to derange the remaining el-
ements (that is, to put them into cycles of length at least 2). Hence there are
(n − 1)Dn−2 derangements of this type.
If the cycle containing n is longer, then skipping it in its cycle still leaves
cycles of length at least 2 and produces a derangement of [n − 1]. On the other
hand, every derangement of [n] with n in such a cycle arises from a derangement
of [n − 1] by inserting n immediately following some x ∈ [n − 1] on the cycle con-
taining x. Hence there are (n − 1)Dn−1 derangements of this type.

Derangements are studied in Exercises 40–46. The recurrences above are lin-
ear (Definition 2.0.1); our next recurrence is not.

2.1.7. Example. Catalan numbers and ballot paths.

Cn = ∑¾=1 C¾−1 Cn−¾ for n ≥ 1, with C0 = 1.
n

In Section 1.3 we counted the ballot paths: lattice paths from (0 , 0) to (n , n)

that don’t rise above the diagonal. We defined Cn to be the number of these paths
and proved that Cn = n+1 1 (2n
n
).

(n , n)

n− ¾
(¾ , ¾)
n−¾

¾−1

¾−1

Every ballot path to (n , n) has some first return to the diagonal, say at (¾ , ¾).
The part up to (¾ , ¾) begins rightward and then does not rise above y = x − 1 until
it reaches (¾ , ¾ − 1). Hence the possible initial portions of the path correspond
bijectively to the ballot paths of length 2 ¾ − 2. The portion from (¾ , ¾) to (n , n)
is a translation of a ballot path of length 2n − 2 ¾ . Hence exactly C¾−1 Cn−¾ ballot
paths of length 2n first return to the diagonal at (¾ , ¾). Summing over the choices
for ¾ yields Cn = ∑¾=1 C¾−1 Cn−¾ for n ≥ 1, with initial condition C0 = 1. This
n

is the Catalan recurrence. The sequence begins 1 , 1 , 2 , 5 , 14 , . . .. Only the

specification of C0 is needed as an initial condition.

2.1.8. Remark. Catalan numbers. Exercises 47–56 study the Catalan numbers.
In Section 1.3 we used bijections to show that this sequence solves many counting
problems. Another method, often easier, is to prove that a new problem satis-
fies the Catalan recurrence and initial condition; the counting sequence ⟨a⟩ must
56 Chapter 2: Recurrence Relations

then also be the Catalan sequence. As in combinatorial arguments for sums, the
key step in proving an = ∑¾=1 a¾−1 an−¾ is interpreting the index of summation so
n

that the resulting subset of objects is counted by the desired summand.

For example, let an count the triangulations of a convex (n + 2)-sided polygon
with vertices v0 , . . . , vn+1 in order. Consider n ≥ 1. In every triangulation, edge
vn+1 v0 belongs to some triangle; let v¾ be its third corner. To complete the triangu-
lation, we must triangulate the polygons formed by v0 , . . . , v¾ and by v¾ , . . . , vn+1 ,
with ¾ + 1 and n − ¾ + 2 sides, respectively. They can be triangulated in a¾−1 and
an−¾ ways. Summing over ¾ yields an = ∑¾=1 a¾−1 an−¾ for n ≥ 1, with a0 = 1 (a
n

2-sided polygon has one (empty) triangulation). Hence an = Cn for all n.

v¾
• •

• •

• •
v0 vn+1

Showing a0 = 1 and an = ∑¾=1 a¾−1 an−¾ for n > 0 gives more than an = Cn .
n

We also obtain a bijection from the given problem to any sequence of sets counted
by the Catalan numbers. The bijection applies recursively to the objects with
index ¾ − 1 and n− ¾ combined to form an object with index n. Thus it also matches
the sets for the ¾ th term in the recurrences for the two problems. For example, in
Theorem 1.3.23 we used the position of the 0 that matches the initial 1 in a ballot
list to recursively invert the bijection that maps binary trees to ballot lists.

VARIATIONS

Problems having more than one parameter may lead to recurrences with more
than one index. As with derangements and Fibonacci numbers, often one can de-
rive recurrences for problems involving distinct objects or positions by studying
how the last object or position is used.

2.1.9. Example. Arrangements with n distinct available objects.

C(n , ¾) = C(n − 1 , ¾) + C(n − 1 , ¾ − 1) C(n , ¾) = # ¾-subsets of [n]
P(n , ¾) = P(n − 1 , ¾) + ¾ P(n − 1 , ¾ − 1) P(n , ¾) = # simple ¾-words from [n]
S(n , ¾) = ¾ S(n − 1 , ¾) + S(n − 1 , ¾ − 1) S(n , ¾) = # partitions of [n] with ¾ blocks
c(n , ¾) = (n− 1)c(n− 1 , ¾) + c(n− 1 , ¾ − 1) c(n , ¾) = # permutations of [n] with ¾ cycles
C(n , ¾): When we select ¾ from n elements, we may use or not use the nth
element. This is Pascal’s Formula for the binomial coefficients.
P(n , ¾): When we list ¾ from n elements, we may use or not use the nth ele-
ment, and if we use it there are ¾ locations it can occupy.
S(n , ¾): When we partition [n] into ¾ blocks, the element n may or may not
form a block itself. There are S(n − 1 , ¾ − 1) partitions where it is alone. Other-
wise, it may join any block in a partition of [n − 1] into ¾ blocks; being nonempty,
these are distinguishable by their members.
Section 2.1: Obtaining Recurrences 57

c(n , ¾): When we permute [n] using ¾ cycles, the element n may or may not
form a cycle by itself. There are c(n − 1 , ¾ − 1) permutations where it is alone.
Otherwise, there are n − 1 elements it can follow to turn a permutation of [n − 1]
with ¾ cycles into a permutation of [n] with ¾ cycles. (Compare this with the
argument in Example 2.1.6.)
Each of these recurrences is valid for n ≥ 1 and computes the value for (n , ¾)
using values for (n − 1 , ¾) and (n − 1 , ¾ − 1). Hence it suffices to specify initial

values for {(0 , ¾): ¾ ∈ 0}. In each case, there is one construction when ¾ = n = 0
and none when ¾ > n = 0. Thus inductively the values are 0 whenever ¾ > n.

Next we find a recurrence for the size of the lattice ball (Definition 1.2.11).
n ¾
The solution ∑ (m ¾ )(¾ )2 was found bijectively in Proposition 1.2.12. The first
proof illustrates a useful technique: sometimes a summation can be eliminated
from a recurrence by taking the difference of two instances of the recurrence. The
second more directly obtains the desired recurrence.

2.1.10. Proposition. The number of lattice points within m lattice steps from
the origin in n satisfies the recurrence lm ,n = lm ,n−1 + lm−1 ,n−1 + lm−1 ,n for
m , n > 0, with lm ,0 = l0 ,n = 1.
Proof: Here l0 ,n counts the all-0 vector and lm ,0 counts the vector with no entries.

Let Bm ,n = {(a1 , . . . , an) ∈ n : ∑i=1 | ai | ≤ m}.
n

Proof 1 Group Bm ,n by the last coordinate. There are lm−|¾| ,n−1 points in Bm ,n
with an = ¾ and lm−|¾| ,n−1 with an = −¾ . Thus lm ,n = lm ,n−1 + 2 ∑¾=1 lm−¾ ,n−1 .
m

m−1
The same argument yields lm−1 ,n = lm−1 ,n−1 + 2 ∑¾=1 lm−1−¾ ,n−1 . Subtracting the
second equation from the first yields lm ,n − lm−1 ,n = lm ,n−1 + lm−1 ,n−1 .
Proof 2 Among the groups in Proof 1, there are lm ,n−1 vectors with an = 0
and lm−1 ,n−1 with an = 1. We claim that lm−1 ,n vectors remain. For each re-
maining vector, bringing its last coordinate one unit closer to 0 yields a vector
in Bm−1 ,n . This produces each vector in Bm−1 ,n exactly once; for example the vec-
tors in Bm−1 ,n with last coordinate 0 correspond to the vectors in Bm ,n with last
coordinate −1. Thus there are lm−1 ,n vectors remaining, and we obtain lm ,n =
lm−1 ,n + lm ,n−1 + lm−1 ,n−1 .

2.1.11. Remark. Delannoy paths and lattice ball. Recall that the Delannoy
number d m ,n is the number of paths from (0 , 0) to (m , n) that move by one of
{(1 , 0) , (0 , 1) , (1 , 1)} at each step (Definition 1.2.8). The analogue of Pascal’s For-
mula for Delannoy numbers is the recurrence d m ,n = d m ,n−1 + d m−1 ,n + d m−1 ,n−1 ,
valid for m , n ≥ 1. Also d m ,0 = d0 ,n = 1.
Proposition 2.1.10 obtains the same recurrence and initial conditions for lm ,n .
This gives a short proof (without finding a bijection) that the Delannoy numbers
also count the points in the lattice ball, or a short proof of Theorem 1.2.13 that
∑¾ (m¾ )(nm+¾) = ∑¾ (m¾ )(¾n)2¾ if we instead counted the two sets separately as in
Proposition 1.2.10 and Proposition 1.2.12.

When a single recurrence for a sequence is not apparent, it may help to derive
a system of recurrence relations for several sequences. Manipulating the system
may then yield a recurrence for the original problem.
58 Chapter 2: Recurrence Relations

2.1.12. Example. Systems of recurrences.

{an = bn + cn , bn = an−1 , cn = bn−1 } for n ≥ 1, with a0 = b0 = 1.

Let an count the binary n-tuples with no consecutive 1s. Let cn count those
ending in 1, and let bn count the rest. By definition, an = bn + cn . The lists ending
in 0 arise by appending 0 to a shorter good list, but those ending in 1 can append
1 only to lists ending in 0. This yields the expressions for bn and cn . Substituting
to eliminate b and c yields a single recurrence for a. It is the Fibonacci recurrence
an = an−1 + an−2 (obtained directly in Exercise 14).

Recurrences may be indexed by objects that are not numbers. Each object
must be evaluated in terms of objects evaluated earlier; finitely many computa-
tions must suffice to evaluate any object from the initial conditions. We consider
a recurrence on graphs that evaluates each in terms of graphs with fewer edges.

2.1.13.* Example. A spanning tree in a graph G is a subgraph that is a tree

and contains all the vertices; let Ì(G) be the number of spanning trees. Cayley ’s
Formula gives Ì(K n) = nn−2 (Theorem 1.3.4). For generalized graphs (loops and
repeated edges allowed), we develop a recurrence.
For e ∈ E(G), the graph G − e is obtained from G by deleting e from the edge
set. The graph G · e is obtained by deleting e and replacing its endpoints with a
single vertex incident to all edges formerly incident to e. This is called contract-
ing e; visually, e shrinks to a point. This may introduce loops and repeated edges
and reduces the number of edges by 1.
Ì(G) = Ì(G − e) + Ì(G · e) when e is not a loop in G,
Ì(G) = 0 when G is disconnected,
Ì(G) = 1 when G has one vertex.
The right side counts the spanning trees T by whether e is used. Those not
containing e are the spanning trees of G − e. If T contains e, then the edges of
T − e form a spanning tree in G · e. Conversely, adding e to the edge set of a
spanning tree T in G · e forms a spanning tree in G, since T cannot include a
path in G connecting the endpoints of e. Thus the number of spanning trees in
G that contain e is Ì(G · e).
Initial conditions are needed when all edges are loops. Here we have one span-
ning tree if G has one vertex and none if G has more vertices.
Below, the (multi)graphs on the right each have four spanning trees, and the
graph on the left has eight. The recurrence is useful in some special classes, but
generally it requires exponentially many computations (in the number of edges).
In Chapter 15 we develop a much faster method using determinants. Exercises
59–62 study Ì(G) for special graphs.

• • • • •
e + •
• • • • •
G G−e G·e
Exercises for Section 2.1 59

EXERCISES 2.1

2.1.1. (−) Obtain a recurrence relation for the number of ways to tile a 2-by-n checkerboard
with n identical dominoes.
2.1.2. (−) Obtain a recurrence for the number of pairings of 2n people.

2.1.3. (−) Fix r ∈ . Let an be the number of regions formed by n lines in the plane, given
that r of them are parallel, the other n − r lines each intersect n − 1 lines, and no three
lines have a common point. Obtain a recurrence for ⟨a⟩.
2.1.4. (−) Determine the number of binary n-tuples in which every run has odd length,
where a run is a maximal string of consecutive equal entries. (Beckwith [2014])
2.1.5. (−) Count the symmetric 1 , 2-lists with sum n (like 1221). (Alladi–Hoggatt [1975])
2.1.6. (−) A club grows as follows: at time 0 it has one man and no woman. At time t each
man selects one woman to join, and each woman selects one man and one woman to join.
What is the size of the club at time n, assuming that people never die?
2.1.7. (−) Let ⟨a⟩ satisfy an = an−1 + an−2 + an−3 for n > 3. Prove that an ≤ 2 n−2 for n ≥ 2
if ai = 1 for i ∈ {1 , 2 , 3}, and an < 2 n if ai = i for i ∈ {1 , 2 , 3}.
2.1.8. (−) Let an ,¾ = (2n
2¾
)/(¾n). Use the formula for the binomial coefficients to compute
an ,¾ recursively in terms of an ,¾−1 and in terms of an−1 ,¾ , including initial conditions.
2.1.9. (−) Let Dn be the number of derangements of [n], and let En be the number of per-
mutations of [n] having exactly one fixed point. Compute Dn − En as a function of n.
2.1.10. An arc diagram consists of a binary n-tuple plus arcs joining some pairs of posi-
tions. The arcs can only join positions with opposite values, and each position is the left
element of at most one arc. Below is an arc diagram of length 9. Determine the number
of arc diagrams of length n. (Callan [2008])

• • • • • • • • •
0 1 1 0 1 0 1 1 0


2.1.11. Fix r ∈ . Let an be the number of outcomes of n coin flips (in order) that contain
exactly r heads. Obtain a recurrence relation (in one variable!) for ⟨a⟩.
2.1.12. Let an count the binary n-tuples not having 0 , 1 , 1 consecutively. Find a system
of recurrences for an (see Example 2.1.12), and reduce it to a single recurrence for ⟨a⟩.
2.1.13. (♦) A Schröder n-path is a path from (0 , 0) to (2n , 0) in the first quadrant by steps
in {(1 , 1) , (2 , 0) ,(1 , −1)}. An uprun is a maximal upward segment; the example below has
two. One n-path has no uprun, and one has n. Let an and bn count the Schröder n-paths
having one uprun and n − 1 upruns, respectively (⟨a⟩ begins 0 , 1 , 4 , 11 , . . ., and ⟨b⟩ begins
0 , 1 , 4 , 9 , . . .). Derive first-order recurrence relations for ⟨a⟩ and for ⟨b⟩. (Deutsch [2004b])
•
• • • • •
• • •

2.1.14. For n ≥ 1, let an be the number of binary (n − 1)-tuples with no consecutive 1s.
Give two proofs that an = F̂ n , as follows.
(a) Show that ⟨a⟩ satisfies the same recurrence and initial conditions as F̂ n .
(b) Establish a bijection from the set of binary (n − 1)-tuples with no consecutive 1s to
the set of 1 , 2-lists with sum n.
60 Chapter 2: Recurrence Relations

2.1.15. Two adjacent panes of glass produce multiple images because some light reflects
each time a surface is encountered. Let an be the number of ways an image can reflect n
times from inner surfaces before emerging. As shown below, a0 = 1, a1 = 2, a2 = 3. Obtain
a recurrence relation for an . (Moser [1963])
0 reflections 1 reflection 2 reflections

2.1.16. For n ≥ 5, consider the digraph formed from n points around a circle by adding
an edge from each point to the next point and to the point after that. Count the cycles
(as subgraphs) in the digraph (cycles must follow arrows). In particular, when n = 5 the
answer is 12. (Juvan–Mohar–Škrekovski [1998b])

• •

• •

• •

• •

2.1.17. Prove the following by induction and by combinatorial argument.

n n
n− i
(a) F̂ n = ∑ ( ). (b) 1 + ∑ F̂ i = F̂ n+2 .
i
i =0 i =0

2.1.18. Give inductive and combinatorial proofs of (a). Use it to prove (b).
n 2n−1

(a) ∑ F̂2i = F̂2n+1 . (b) ∑(−1)i F̂2n−i = F̂2n−1 .

i =0 i =0

2.1.19. (♦) Prove that the classical Fibonacci number F n counts the compositions of n using
odd parts and the compositions of n + 1 using parts greater than 1.

2.1.20. (♦) Cassini’s Identity. Prove F̂ n2 = F̂ n−1 F̂ n+1 + (−1)n for n ≥ 1, both induc-
tively and combinatorially. (Comment: For classical Fibonacci numbers, this is F m 2
=
F m−1 F m+1 +(−1)m−1 , proved by Cassini in 1680). Use the identity to explain Lewis Carroll’s
“proof ” of 64 = 65 (and larger analogues). (Weaver [1938])


2.1.21. (♦) For m , n ∈ 0 with m ≥ n, prove F̂ n F̂ m − F̂ n−1 F̂ m+1 = (−1)n F̂ m− n inductively
and combinatorially. (Comment: This identity generalizes Cassini’s Identity. When rein-
dexed and written using classical Fibonacci numbers as F n+1 F m − F n F m+1 = (−1)n F m− n , it
is known as d’Ocagne’s Identity.)
Exercises for Section 2.1 61

2.1.22. (♦) Prove bijectively that F̂ m+ n = F̂ m F̂ n + F̂ m−1 F̂ n−1 . Conclude for ¾ ∈  that F̂ n−1
divides F̂ ¾n−1 , and hence F n divides F ¾n .
2.1.23. For classical Fibonacci numbers, Exercise 2.1.22 says F m+ n+1 = F m+1 F n+1 + F m F n .
Use this and Cassini’s Identity (Exercise 2.1.20) to prove Catalan’s Identity for a < b:
F b2 − F b− a F b+ a = (−1)b− a F a2 . (Comment: See Melham–Shannon [1995] and Rao [1953].)

2.1.24. (♦) Vajda’s Identity. Consider r, s , t ∈ with r < s.
(a) Give a combinatorial proof of F̂ r F̂ s−1 − F̂ r−1 F̂ s = (−1)r F̂ s−r−1 .
(b) Give a combinatorial proof of F̂ r+ t F̂ s − F̂ r F̂ s+ t = (−1)r−1 F̂ t−1 F̂ s−r−1 .
(c) For classical Fibonacci numbers, conclude Vajda’s Identity for i , j , n ∈ : 
F n+i F n+ j − F n F n+i+ j = (−1)n F i F j .
(Comment: This is a common generalization of d’Ocagne’s Identity and Catalan’s Iden-
tity; see Basin–Hoggatt [1963]. It was attributed by Dickson to Tagiuri [1900] and was
rediscovered in Everman–Danese–Venkannayah [1960].)
2.1.25. Extend classical Fibonacci numbers to negative indices via the recurrence. Prove

F− n = (−1)n−1 F n for n ∈ . (Graham–Knuth–Patashnik [1989, p. 279])
2.1.26. (♦) Prove that every positive integer has a unique expression as the sum of a set of
nonconsecutive Fibonacci numbers (1 and 2 cannot both be used). (Lekkerkerker [1952],
Zeckendorf [1972])
2.1.27. The Lucas numbers ⟨L⟩ satisfy the Fibonacci recurrence with L1 = 1 and L2 = 3.
Prove Ln = F n−1 + F n+1 . Conclude Ln = F 2n/F n .
2.1.28. Let an count the circular lists of n bits with no consecutive 1s (rotations are dis-
tinct , so a2 = 3 and a3 = 4). Compute an using Fibonacci numbers.

2.1.29. For n , d ∈ with d ≤ n, determine the number of partitions of [n] into arithmetic
progressions all having constant difference d and having length at least 1. Do the same
for length at least 2. (Getz–Jones [2003])

2.1.30. For n ∈ , let an = ∑¾=1 ¾ 21n−¾ , let bn = ∑¾=1 [¾ (¾n)]−1 , and let cn = 2− n+1 ∑¾ odd (¾n) 1¾ ,
n n

Note that a0 = b0 = c0 = 0. Prove an = bn = cn for all n, by showing that each sequence

satisfies 2nx n = nx n−1 + 2 for n ≥ 1 (see Galperin–Gauchman [2004]).
2.1.31. (♦) Let G be the square grid on n2 points in horizontal and vertical paths.
(a) Partition the edges of G into 2n − 2 zig-zag paths as shown for n = 4. Let an count
n−1
the edge sets with no two consecutive edges on any such path. Prove an = ∏ j =1 F̂ 22 j +1 .

(b) Let bn count the partitions of the vertices into up/right lattice paths (b2 = 9, shown
n−1
below). Prove bn = ∏ j =1 F̂ 22 j +1 . (Stanley [1992] gives a more general result.)
• • • • • • • • • • • • • • • • • •
• • • • • • • • • • • • • • • • • •

2.1.32. (+) Consider cards labeled 1 through n in some order. If the top card is m, we
reverse the order of the first m cards. The process stops only when card 1 is at the top.
Prove that the process must stop. Let an be the maximum (over all initial orderings) of the
number of steps in the process. Prove an < F̂ n . (Hint: Prove that if ¾ distinct cards appear
at the top during the process, then there are at most F̂ ¾ − 1 steps.) (Knuth)
62 Chapter 2: Recurrence Relations

2.1.33. Generalized Fibonacci. Let an+1 = an + an−1 for n ≥ 3, with a1 and a2 fixed.

(a) Determine T so that an+2 an + (−1)n T = a2n+1 for all n ∈ . (Comment: T is the
only x such that an+2 an + (−1)n x is always square.) (Horadam [1961], Kantrowitz [1986])
(b) Prove an+1 = F n−1 a1 + F n a2 for n ≥ 2, using the classical Fibonacci numbers.
(t)
2.1.34. The Fibonacci numbers of order t are defined by F n = 2 n−1 if 1 ≤ n ≤ t, and
(t) (t) (t)
F n = ∑ j =1 F n− j . The binomial coefficient of order t, Cn ,¾ , counts the n-tuples with entries
t

in {0 , . . . , t − 1} and total sum ¾ . Prove the statements below.

(a) F n(t) counts binary (n − 1)-tuples with no t consecutive 1s.
¾
(b) Cn(t),¾ = Cn(t),(t−1)n−¾ and Cn(t),¾ = ∑i=0 Cn(t)−1 ,¾−i , if n > 0 and 0 ≤ ¾ ≤ n(t − 1).
(c) Use the recurrences to prove F n(t) = ∑r≥0 Cn(t)−r,r .
(d) Give a combinatorial proof of (c).
2.1.35. (♦) Let an = 2an−1 + an−2 for n ≥ 2, with a0 = 1 and a1 = 2. Use Delannoy paths
to prove that an = ∑ (i+i! jj +!¾¾!)! , summed over nonnegative integer triples (i , j , ¾) such that
i + j + 2 ¾ = n. (Comment: This sequence is the Pell sequence.) (Deutsch [1998])

2.1.36. Given c1 , . . . , c¾ ∈  , define ⟨a⟩ by an = ∑¾i=1 an−c i for n > 0, with a0 = 1 and an =
¾ ¾
0 for n < 0. Prove an = ∑(∑i=1 mi)!/ ∏i=1 mi !, where the sum runs over all nonnegative
¾
m1 , . . . , m¾ such that ∑i=1 ci mi = n. (Barra [2000])

2.1.37. Fix ¾ ∈ and let bn , j = n j (¾j ). Prove bn , j = ∑i=0 (¾j −−ii)bn−1 ,i . Give both a combina-
j

torial proof and a proof using identities.

2.1.38. (♦) Let ¾ be a natural number. Create an array of ¾ rows with row 1 being the nat-
ural numbers. To obtain row j + 1 from row j , cross out every (¾ + 1 − j)th entry in row j
and take partial sums of the sequence that remains (leave blanks under the crossed-out en-
tries). Prove that row ¾ consists of the numbers n¾ . (Hint: Guess and prove a formula for
the numbers on the longest rising diagonals of the triangular wedges, using a recurrence.)
1 2 3 4 5 6 7 8 9 10 11 12
1 3 6 11 17 24 33 43 54
1 4 15 32 65 108
1 16 81
(Comment: The case ¾ = 2 states ∑i=1 2i − 1 = n2 . The procedure is called Moessner ’s
n

Process, after Moessner [1951]. Proofs appear in Perron [1951], Long [1966, 1982], and
Bender–Kochman–West [1990], generalized in Paasche [1956] and Long [1986].)
2.1.39. (+) Let a0 ,0 = 1, and let am ,n = 0 when m < 0 or n < 0. For m , n ∈  , let
am ,n−1 + am−1 ,n if m + n is even,
am ,n = {
am ,n−1 + 2am−1 ,n if m + n is odd.
By a combinatorial argument , obtain a summation for an ,n and prove that it equals the
central Delannoy number d n ,n . (Sulanke [2001])
2.1.40. Use the formula n! = ∑¾=0 (¾n) Dn−¾ of Example 2.1.6 to obtain the recurrence
n

Dn = ∑¾=1 (¾ − 1)(¾n) Dn−¾ for the derangement numbers.

n

2.1.41. (♦) Let d(n , ¾) count the derangements of [n] with ¾ cycles. Derive a recurrence.
2.1.42. Let Do(n) and De(n) be the numbers of derangements of [n] having an odd or an
even number of cycles, respectively. Compute Do(n) − De(n). (Stathopoulos [2012])
2.1.43. Determine the number of permutations of [n] that are derangements satisfying
the increase condition (i + 1) ≤ (i) + 1 for 1 ≤ i ≤ n − 1. (Deutsch [2001a])
Exercises for Section 2.1 63

2.1.44. (♦) Let A n be the set of permutations of [n] such that position i does not have ele-
ment i + 1, for 1 ≤ i ≤ n − 1. Let Bn be the set of those in which no element is followed imme-
diately by the next larger element. Thus A3 = {123 , 312 , 321} and B3 = {132 , 213 , 321}.
(a) Prove | A n| = Dn + Dn−1 for all n, where Dn is the number of derangements of [n].
(Hint: Consider the position of the element 1.)
(b) Obtain a second-order recurrence for | Bn | , and use it to conclude | Bn| = Dn + Dn−1
(Hint: Consider how the element n is used.)
(Comment: Thus | A n| = | Bn| ; a bijective proof is requested in Exercise 3.1.27.)

2.1.45. Let Dn ,¾ be the number of permutations of [n + ¾] having no fixed points in [n].

Generalize the argument of Example 2.1.6 to obtain a recurrence for Dn ,¾ .

2.1.46. Prove the identity below involving the derangement numbers.

¾ min{n ,¾}
¾ ¾ ¾ + n− j
∑ ( ) D¾+n− j = ¾ !
j ∑ ( )(
j ¾
) Dn− j (Callan [1998])
j =0 j =0

2.1.47. (♦) Prove recursively (no bijections needed) that in each case below, an is the nth
Catalan number. (Comment: In each case, one can also prove the claim bijectively by es-
tablishing a bijection from the set being counted to the set of ballot lists of length 2n.)
(a) Let an be the number of ordered trees with n + 1 vertices (see Definition 1.3.21).
The ordered trees with four vertices appear below.

• • • • •
• • • • • • • • •
• • • • •
•
(b) Let an be the number of noncrossing pairings of 2n points on a circle. The non-
crossing pairings of six points appear below.

• • • • • • • • • •
• • • • • • • • • •
• • • • • • • • • •

(c) Let an be the number of configurations of pennies on a base row of n pennies, where
pennies can be added so that each one not in the base rests on two in the row immediately
below it. An example for n = 6 appears below. (Propp)

2.1.48. (♦) A stack stores numbers with only the most recent number (the top element)
accessible. A stack processes an input list as follows. Let x be the next input number (if
any), and let y be the top element of the stack (if any). Move x to the stack when x exists
and y does not or when x < y; otherwise move y to the output. Eventually all input ele-
ments move input to the stack and later to the output. A list is stack-sortable if the out-
put is in increasing order. For example, (1 , 5 , 4 , 2 , 3) is stack-sortable, but the output for
(3 , 5 , 1 , 2 , 4) is (3 , 1 , 2 , 4 , 5). Let an be the number of stack-sortable permutations of [n].
(a) Prove recursively that an is the Catalan number Cn .
(b) Prove that a list x1 , . . . , x n of distinct numbers is stack-sortable if and only if there
do not exist i , j , ¾ with i < j < ¾ such that x¾ < xi < xj .
64 Chapter 2: Recurrence Relations

2.1.49. Prove that the Catalan numbers satisfy the recurrence below.
n−1
2n 1 2n − 2 ¾
Cn = (
n
)−
2 ∑ C¾ ( n− ¾
).
¾ =0

2.1.50. (♦) Noncrossing partitions of [n]. A partition of [n] is noncrossing if there are no
a , b , c , d with a < b < c < d such that a and c are in one block and b and d are in another.
Thus (14|25|3) is crossing and (15|234) is noncrossing. Let an be the number of these.

A B A
• • •
B
C •
• • • • • • • • • •
1 2 3 4 5 1 2 3 4 5
(a) Prove recursively that an is the Catalan number Cn .
(b) Use the recurrence to obtain a bijection from noncrossing partitions of [n] to or-
dered trees with n edges (Definition 1.3.21). (Prodinger [1983])
(c) From the bijection in part (b), observe that the number of noncrossing partitions of
[n] with l blocks equals the number of ordered trees with n edges and l leaves. (Comment:
The sizes of these subsets are the Narayana numbers; see Theorem 1.3.26.)

2.1.51. (♦) Let an count the partitions of 3n points on a circle into n nonintersecting tri-
angles. Let bn count the lists of 2n 1s and n 0s such that every prefix has at least twice as
many 1s as 0s (2-ballot lists; Definition 1.3.18). Prove an = bn by showing that the satisfy
the same recurrence. (Comment: By Corollary 1.3.19, bn = 2n1+1 (3n n)
.)

2.1.52. (♦) A positive lattice walk in n starts at the origin, at each step changes one
coordinate by 1, and visits no point with last coordinate negative. Vertices and edges may
repeat. Let an be the number of such walks of length n in 2 ending on the horizontal axis.
(a) Obtain a recurrence for ⟨a⟩. Show that an is the Catalan number Cn+1 .
(b) Let wn be the total number of positive lattice walks of length n in 2 . Use part (a)
to obtain a recurrence for ⟨w⟩, and conclude wn = (2nn+1 ). (Guy [2000])
(c) Use part (b) to show that the total number of positive lattice walks of length n in
three dimensions is ∑¾=0 (¾n)(2¾¾+1 )2 n−¾ .
n

(Comment: Exercise 1.3.34 obtains another formula for wn . It and part (b) provide a
¾
combinatorial proof of ∑¾=0 (¾n)(⌈¾/2⌉ )2 n−¾ = (2nn+1). For the corresponding problem in three
n

¾
dimensions, part (c) and Exercise 1.3.34 yield ∑¾=0 (n¾)(⌈¾/2⌉ )4n−¾ = ∑¾=0 (¾n)(2¾¾+1 )2 n−¾ .)
n n

2.1.53. (+) Let Sn count the paths from (0 , 0) to (n , n) that never rise above y = x and use
steps in {(0 , 1) , (1 , 0) , (1 , 1)}. Note that (S0 , S1 , S2 , S3) = (1 , 2 , 6 , 22).
(a) Prove that Sn is divisible by 3 when n is positive and even. (Shapiro–Rogers [1989])
(b) Prove an = 6an−1 − an−2 − 2Sn−1 for n ≥ 2, where an is the nth central Delannoy
number d n ,n (see Definition 1.2.8). (Peart–Woan [2002])
(Comment: Sn is the nth “large” Schröder number (there are also “small” ones),
named for Ernst Schröder (1841–1902) after Schröder [1870]. These numbers satisfy
(n + 1)Sn = 3(2n − 1)Sn−1 − (n − 2)Sn−2 for n ≥ 2; Sulanke [1998] gave a bijective proof.
Also nan = 3(2n − 1)an−1 − (n − 1)an−2 for the Delannoy numbers.)

2.1.54. (♦) Shapiro n-paths are lattice paths from (0 , 0) to (2n , 2n) avoiding the odd
points (2i − 1 , 2i − 1) on the diagonal; let Sn be the number of these. Prove Sn = C2n .

2.1.55. (♦) A Dyck n-path is a path from (0 , 0) to (2n , 0) using steps that move by (1 , 1)
or (1 , −1) without falling below the x-axis. Show that the number of Dyck (2n)-paths that
avoid {(4 ¾ , 0): 1 ≤ ¾ ≤ n − 1} is twice the number of Dyck (2n − 1)-paths. (Callan [2003b])
Exercises for Section 2.1 65

2.1.56. Two teams play until one wins n games. Team A has probability p of winning any
game, independently. Let q = 1 − p. Let an be the expected number of games played.
n−1
(a) Prove an = n ∑¾=0 (n+¾ ¾ )(pn q¾ + p¾ qn).
n−1 2¾
(b) Prove an = n ∑¾=0 C¾ (pq)¾ , where C¾ = ¾ +1 ( ¾ ) .
1
(Shapiro–Hamilton [1993])

2.1.57. (♦) An up-down permutation alternates ascents and descents, an ascent first.
Let an be the number of up-down permutations of [n]; note (a0 , . . . , a4) = (1 , 1 , 1 , 2 , 5).
Let bn be the number of equivalence classes of permutations of [n] under “low-flip” oper-
ations that reverse x1 , . . . , x¾ if ¾ = n or if x¾+1 > max{x1 , . . . , x¾ }; two permutations
are equivalent if low-flips can turn one into the other. Prove an = bn for n ≥ 0. (Knuth
[2009]) (Comment: These are called secant-and-tangent numbers; see Lehmer [1935],
Knuth–Buckholtz [1967], Riordan [1968], Comtet [1970, 1974], Gould [1972], etc.)

 n−1
2.1.58. For n ∈ , let T¾ (n) be the coefficient of (−1)¾ x n−¾ in the expansion of ∏i=0 (x − i).
(a) Give a combinatorial proof that T¾ (n) = T¾(n − 1) + (n − 1)T¾−1 (n − 1).
n−1
(b) Prove T¾(n) = ∑i=¾ iT¾−1 (i) for ¾ ≥ 1.
(c) Compute simple formulas for T¾(n) when ¾ ∈ {0 , 1 , 2}.
(d) Prove that T¾ is a polynomial of degree 2 ¾ .

2.1.59. Let the graph G n be the path ⟨x1 , . . . , x n ⟩ plus one vertex adjacent to all xi (shown
below for n = 5). Let an be the number of spanning trees in G n .
(a) Obtain a many-term recurrence for ⟨a⟩.
(b) Obtain a second-order recurrence for ⟨a⟩.
(c) Prove an = F 2n , where F ¾ is the ¾ th classical Fibonacci number.

• • • • •

(Comment: Relations for Fibonacci/Lucas numbers and counts of spanning trees in special
graphs have been studied in Sedláček [1970], Hilton [1972], Fielder [1974], Myers [1975],
Shannon [1978], Mikola [1980], and Exercises 60–61.)

2.1.60. A ¾-tree is a graph obtained from a complete graph with ¾ vertices by iteratively
adding a new vertex whose neighbors induce a complete graph with ¾ vertices. For ex-
ample, ordinary trees are 1-trees, and the graphs in Exercise 2.1.59 are 2-trees. A vertex
with degree ¾ in a ¾-tree is called a simplicial vertex. Prove that all n-vertex 2-trees hav-
ing exactly two simplicial vertices have the same number of spanning trees (this includes
the graphs in Exercise 2.1.59). (Comment: In fact , these 2-trees are the n-vertex 2-trees
with the most spanning trees.) (Xiao–Zhao [2013])

2.1.61. Form a graph G n from two n-vertex paths by making corresponding vertices adja-
cent as shown. Obtain a second-order recurrence for the number of spanning trees.

• • ··· • •

• • ··· • •

2.1.62. Let G be a graph with m spanning trees. Let G be the graph obtained by replac-
ing each edge of G with ¾ copies of that edge. Let G be the graph obtained by replacing
each edge uv ∈ E(G) with a u , v-path of length ¾ through ¾ − 1 new vertices. Determine
Ì(G ) and Ì(G ) in terms of m, ¾ , and Ì(G).
66 Chapter 2: Recurrence Relations

2.1.63. Let an be the number of domino tilings of a 4-by-n rectangle. Obtain a bounded-
order recurrence for ⟨a⟩. (Rymer [1979], Zhou [2005])

2.1.64. (♦) A Summation Identity for permutations. As in Example 2.1.9, let c(n , ¾) be the
number of permutations of [n] with ¾ cycles. For m ≥ ¾ , prove ∑ j =0 c(jj ,!¾) = c(m+m!
1 ,¾ +1)
m
.
(Hint: After multiplying by m!, there is both a direct bijective proof and an inductive proof
using a known identity.

2.1.65. Gambler’s Ruin. Player A starts with $r, B with $s. They flip a fair coin until one
goes broke. On a head, A pays B $1; on a tail, B pays A $1. (Kraitchik [1942])
(a) Prove that the probability of A winning is r+r s .
(b) Prove that the expected number of flips until the game ends is rs.

2.1.66. (+) A random walk starts at 0. On each step, it moves up or down by 1, each with
probability .5 (independent of earlier steps). Compute the expected number of steps until
the first time the walk is ¾ steps below the highest value it has reached. (Hint: Find a
recurrence for the expected number.) (Palacios [2012])

2.1.67. (+) The gambler and the devil play with n red balls and n + 1 blue balls. The
gambler starts with one unit of money (infinitely divisible). At each round, the gambler
bets some money (maybe 0), and the devil picks a remaining ball. When the ball is red,
the gambler loses the bet; when blue, the gambler gains that amount. The selected ball
is discarded. The gambler wants to maximize his money at the end, while the devil wants
to minimize it. If both play optimally, how much does the gambler end with? (Hint: Solve
the more general problem with r red balls and b blue balls.) (Pudaite [2000])

2.1.68. Compute det M , where M is the n-by-n matrix whose first row and column is all
1, and otherwise mi , j = mi−1 , j + mi , j −1 + xmi−1 , j −1 . (Bacher [2001])

2.1.69. Let m and n be nonnegative integers. Prove

⌊m/2⌋ ⌊n/2⌋ 2
i+ j m + n − 2i − 2 j ⌈ n+2m ⌉ !⌈ n+m2 +1 ⌉ !
∑ ∑( j
) (
n− 2j
)=
⌊ n/2⌋ !⌊ m/2⌋ !⌈ n/2⌉ !⌈ m/2⌉ !
.
i =0 j =0

(Hint: Writing each side as am ,n , compute am ,n − am ,n−1 − am−1 ,n .) (Andrews–Paule [1992])

2.1.70. Destructive testing. When dropped from the bth floor of a building, beanbags break.
From higher floors they also break, but from lower floors they survive and can be re-used.
Let º (t , ¾) be the largest value b that can be determined with certainty when ¾ beanbags
are available and up to t tests are permitted. Prove that º (t , 1) = t and that º (t , ¾) =
º (t − 1 , ¾ − 1) + º (t − 1 , ¾) + 1 for ¾ > 1. (G. Peterson)

2.2. Elementary Solution Methods

Finding a recurrence for a counting sequence can be a step toward a solution
formula. We next study methods for obtaining such formulas from recurrences.
A simple formula for an obtained from a recurrence may have an elegant direct
combinatorial proof. Although we may then discard the recursive approach in
presenting the final solution, solving the recurrence remains a useful tool.
Section 2.2: Elementary Solution Methods 67

THE CHARACTERISTIC EQUATION METHOD

Linear recurrence relations with constant coefficients can be solved by the

characteristic equation method or the more versatile generating function method
discussed later. When it applies, the first method is faster, abbreviating standard
computations in the second method. It also closely parallels standard methods for
elementary differential equations. In this discussion all recurrence relations are
linear with constant coefficients, though we may neglect to say so.

2.2.1. Definition. For the linear constant-coefficient recurrence relation

an = c1 an−1 + · · · + c¾ an−¾ + º (n),
the associated characteristic polynomial is the polynomial defined by (x) =
x¾ − c1 x¾−1 − · · · − c¾ x0 , the characteristic equation is the equation (x) = 0,
and its solutions are the characteristic roots.

2.2.2. Remark. Solutions to homogeneous recurrences. For the homogeneous re-

currence an =  an−1 , the characteristic equation is x =  . Successive substitu-
tion shows that the solution is  n a0 . In other words, the solution has the form
A n , and the constant A is determined by the initial condition. This behavior
generalizes to higher order.
Consider the homogeneous recurrence an = c1 an−1 + · · · + c¾ an−¾ . When  is a
characteristic root, substituting  in the characteristic equation and multiplying
by  n−¾ yields  n = c1  n−1 + · · · + c¾  n−¾ . Multiplying by a constant A does not
disturb the equality, so an = A n is a solution.
When both  and  are characteristic roots, the equations showing that A n
and B n satisfy the recurrence can be summed to show that setting an = A n +
B n also satisfies the recurrence.
A specific sequence is determined by the initial conditions. For  = 2 with
roots  and  , we choose values for A and B in A n + B n so that the formula for
an reduces to the values for a0 and a1 when n ∈ {0 , 1}.

2.2.3. Example. For the homogeneous recurrence an = 3an−1 − 2an−2 , the char-
acteristic equation is x 2 = 3x − 2, with roots 1 and 2. Hence every sequence of the
form an = A + B2 n satisfies the recurrence.
Setting n = 0 yields a0 = A + B, and n = 1 yields a1 = A + 2B. Given
the values for a0 and a1 as initial conditions, the system of two linear equations
determines A and B to complete the solution.

Irrational or complex numbers may arise as characteristic roots. Neverthe-

less, when the initial conditions and coefficients of the recurrence are integers,
each term in the resulting sequence must be an integer.

2.2.4. Example. The Fibonacci recurrence F̂ n = F̂ n−√1 + F̂ n−2 has the character-
√
istic equation x 2 − x − 1 = 0, with roots 1 = (1 + 5)/2 and 2 = (1 − 5)/2.
Requiring F̂0 = F̂1 = 1 in the general solution F̂ n = A1n + B2n yields 1 = A + B
and 1 = A1 + B2 . The solution is
68 Chapter 2: Recurrence Relations
√ n+1 √ n+1
1 ⎛1 + 5⎞ 1 ⎛1 − 5⎞
F̂ n = √ − √ .
5⎝ 2 ⎠ 5⎝ 2 ⎠

√ Fibonacci numbers (F0 = 0; F1 = 1), we obtain the solution

For the classical
F n = ( 1n − 2n)/ 5. This is Binet’s Formula, proved by de Moivre [1730] us-
ing generating functions√and rediscovered by Binet [1843] and Lamé [1844]. The
characteristic root (1 + 5)/2 is the Golden Ratio.

Since sums and constant multiples of sequences satisfying a homogeneous re-

currence also satisfy it, the solutions form a linear space. Since  initial values
are required to specify a sequence satisfying a recurrence of order  , the space of
solutions has dimension  . The next lemma makes this precise.

2.2.5. Lemma. The solutions to a homogeneous constant-coefficient linear recur-

rence of order  form a -dimensional subspace of the space of all sequences.
Proof: For completeness, consider sequences from the field of complex num-
bers; they form an infinite-dimensional vector space. Being closed under addi-
tion of sequences and under multiplication by constants, the set U of solutions to
a homogeneous linear recurrence with constant coefficients is a subspace.
Specify a solution sequence whose 0th through ( − 1)th terms are all 0 ex-
cept that the jth term is 1. These sequences are linearly independent. Also,
every solution sequence is a linear combination of these, using the values in the
initial conditions as coefficients. Thus the  special sequences form a basis for the
solution space, so it has dimension  .

Lemma 2.2.5 suggests that the method used above finds all the solutions,
when the characteristic roots are distinct.

2.2.6. Example. When an = 2an−1 − an−2 , the characteristic polynomial is

x2 − 2x + 1, which factors as (x − 1)2 . For any constants A1 and A2 , the expres-
sion A1 1 n + A2 1 n specifies a constant sequence. Nevertheless, the initial condi-
tions may have a0
= a1 . For example, when a0 = 1 and a1 = 2, it is easy to check
by induction that an = n + 1 is the solution.

The next theorem finds the missing solutions. “General Solution” means that
all sequences satisfying the recurrence are described; a specific sequence arises
by choosing coefficients in the polynomials Pi to satisfy the initial conditions.

2.2.7. Theorem. (General Solution) Let ⟨a⟩ satisfy a homogeneous linear

recurrence relation of order  with constant coefficients. If 1 , . . . , r
with multiplicities d1 , . . . , dr are the distinct characteristic roots, then
an = ∑i Pi(n) ni , where each Pi is a polynomial of degree less than d i .
Proof: The characteristic polynomial factors as (x) = ∏i=1 (x − i)d i . Multiplying
r

by x n−¾ yields
r
x n − c1 x n−1 − · · · − c¾ x n−¾ = x n−¾ ∏(x − i)
di
.
i=1
Section 2.2: Elementary Solution Methods 69

Using the product rule to differentiate the right side retains (x − i)d i −1 as a fac-
tor in each term. More generally, differentiating j times when j < d i retains
(x − i)d i − j as a factor in each term on the right; so the value at x = i remains 0.
Differentiating j times on the left, multiplying both sides by x j , and then setting
x = i yields
n−1 n−¾
n(j) n
i − c1 (n − 1)(j) i − · · · − c¾ (n − )(j) i =0.
Therefore, setting an = n(j) ni is a solution. Note that the falling factorial n(j) is a
polynomial in n of degree j . Every polynomial with degree less than d i is a linear
combination of {n(0) , . . . , n(di−1)}. Hence an = Pi(n) ni is a solution whenever Pi is
a polynomial of degree less than d i (Pi may have complex coefficients).
By linearity, every linear combination of such solutions from distinct roots is
also a solution. By Lemma 2.2.5, the set of sequences satisfying the recurrence
is a vector space of dimension  . In the expression for the solution, there are 
parameters (the coefficients in the polynomials). For a finite set of distinct pairs
( , r), Exercise 20 shows that the sequences defined by setting an = nr n are in-
dependent. Hence we have described a -dimensional space of solutions, which
consists of all solutions of the recurrence.

A direct proof that the sequences of the form nr n are independent involves
the Vandermonde determinant, which we have not derived. An alternative proof
using only the dimension of vector spaces will follow from Theorem 2.2.20.
Specifying  initial conditions determines one sequence and hence the coef-
ficients of the polynomials P1 , . . . , Pr in the general solution. To find it, we use
the initial conditions to impose  linear equations on the  unknown coefficients.
For 0 ≤ i <  , set n = i in the formula for an and equate this to the given value
ai . The system is uniquely solvable because the space of sequences (1) satisfying
the recurrence and (2) specified by the general solution are the same.

2.2.8. Example. Consider the recurrence an = 2an−1 − 2an−2 + 2an−3 − an−4 for
n ≥ 4, with (a0 , a1 , a2 , a3) = (1 , 3 , 5 , 5). The characteristic equation is x4 − 2x3 +
2x 2 − 2x + 1 = 0, which factors as (x2 + 1)(x − 1)2 = 0. The characteristic roots
are i , −i , 1 , 1. Thus the general solution is
an = A1 in + A2(−i)n + A3 1 n + A4 n1 n .
To determine these coefficients, we invoke the initial conditions:
n = 0: 1= A1 + A2 + A3
n = 1: 3 = i A1 − i A2 + A3 + A4
n = 2: 5 = − A1 − A2 + A3 + 2 A4
n = 3: 5 = −i A1 + i A2 + A3 + 3 A4
Summing the equations in pairs eliminates A1 , A2 to solve for A3 , A4 , and then
the first two yield A1 and A2 . The solution is (A1 , A2 , A3 , A4) = (−1/2 , −1/2 , 2 , 1),
which yields an = n + 2 − (in + (−i)n)/2.

Next we consider inhomogeneous linear recurrences. The method is moti-

vated by the observation that the solutions form an affine subspace of the space
of all sequences. The all-0 sequence is no longer a solution.
70 Chapter 2: Recurrence Relations

2.2.9. Example. The Tower of Hanoi problem.

an = 2an−1 + 1 for n ≥ 1, with a0 = 0 .

In this problem, n rings on a peg must move to another peg, with a third peg
available as workspace. No ring can be placed upon a smaller ring. Let an be the
minimum number of steps needed to move the pile. The bottom ring can move
to the desired peg only when we have moved the rings above it to the third peg,
and then after moving the bottom ring we must move the other rings on top of
it. Thus an = 2an−1 + 1. It takes no steps to move no rings. The value more than
doubles with each increase in n, so the growth is exponential. This is a linear
first-order recurrence.
Lucas popularized this problem using the legend of the Tower of Benares:
monks were moving a pile of 64 heavy golden disks according to these rules, and
the world would end when they finished. In fact, there was no such legend; Lucas
made it up! (Singmaster [1998, p. 581]). An early solution is in Ball [1892].
Dozens of papers have studied variations on this problem, such as starting
from an arbitrary configuration of the disks, moving disks only cyclically, finding
the minimum number of moves when p pegs are available, etc. A discussion of
work on the p-peg problem appears in K lavžar–Milutinović–Petr [2002].
In the 3-peg recurrence, each value roughly twice the previous value, the so-
lution should be close to 2 n . In fact, with a0 = 0, it follows by induction that
setting an = 2 n − 1 solves the recurrence. Proof by induction works only when we
already know the answer. If not given the formula 2 n − 1, how can we find it?

2.2.10. Remark. Solutions to inhomogeneous recurrences. To solve such a recur-

rence, we first seek a single sequence ⟨p⟩ satisfying the recurrence, without regard
to the initial conditions. We call this a particular solution. We then obtain the
general solution by adding ⟨p⟩ to the general solution of the homogeneous re-
currence obtained by dropping the inhomogeneous term. For each solution ⟨h⟩ to
the homogeneous recurrence, linearity (a special case of Exercise 10) implies that
setting an = pn + hn solves the original inhomogeneous recurrence.
¾
pn = ∑i=1 ci pn−i + º (n)
¾
hn = ∑i=1 ci hn−i
¾
pn + hn = ∑i=1 ci(pn−i + hn−i) + º (n)
As in Theorem 2.2.7, general solution means that all solutions of the recur-
rence have been described. The general solution again is parametrized by ¾ con-
stants; these are the constants in the general solution of the homogeneous part
of the recurrence. To complete the solution, we solve linear equations in these
constants to make the general formula satisfy the initial conditions.
Section 2.2: Elementary Solution Methods 71

When the inhomogeneous term is nonzero, many books suggest “guessing ”

the form of a particular solution. Usually it is not a guess at all; the next theorem
guarantees the form of a particular solution for certain kinds of recurrences.
For example, this theorem guarantees that the recurrence an = 2an−1 + 1
in Example 2.2.9 has a particular solution that is a constant sequence. If some
constant b works, then b must satisfy b = 2b + 1. This yields b = −1. The char-
acteristic root is 2, so the general solution is an = A2 n − 1. The initial condition
a0 = 0 requires A = 1, so an = 2 n − 1, with no inductive proof needed.

¾
2.2.11. Theorem. Let R denote the recurrence an = ( ∑i=1 ci an−i) + F(n) n for
n ≥  , where F is a polynomial of degree d. If has multiplicity r as a char-
acteristic root of the homogeneous recurrence obtained by setting F(n) = 0
(r may be 0), then the recurrence has a solution of the form an = P(n)nr n ,
where P is a polynomial of degree at most d.
Proof: Since R generates a sequence from any initial conditions, particular solu-
tions exist. The key is to find a particular solution to R among the solutions of a
homogeneous recurrence S of higher order.
Let H denote the homogeneous recurrence obtained from R by eliminating the
inhomogeneous term, and let  be the characteristic polynomial of H. Let S be the
homogeneous recurrence whose characteristic polynomial in x is (x − )d+1 (x).
Step 1: Every solution to R is a solution to S. From the expression for an given
by R, subtract times the expression for an−1 given by R:
¾
an − an−1 = ∑(ci an−i − ci an−i−1) + [F(n) − F(n − 1)] n
.
i=1

This new recurrence is still satisfied by ⟨a⟩, since we simply subtracted equalities
satisfied by ⟨a⟩. The recurrence has order  + 1, since it relates an , . . . , an−¾−1 .
Its characteristic equation is
¾
x ¾+1 − x¾ = ∑(ci x¾+1−i − ci x¾−i)
i=1

which simplifies to (x − )(x) = 0. Most importantly, F(n) − F(n − 1) is a polyno-

mial in n of degree d − 1.
Applying this operation d times, the resulting recurrence has characteristic
polynomial (x − )d (x) and inhomogeneous term G(n) n , where G is a polynomial
of degree 0; that is, G(n) is a constant. One more application of the differenc-
ing operation reduces the inhomogeneous term to 0 · n , so the recurrence is now
homogeneous and has characteristic polynomial (x − )d+1 (x). That is, it is the
recurrence S. Furthermore, throughout each step, ⟨a⟩ remains a solution.
Step 2: Choosing a particular solution. Since S is a homogeneous recurrence,
its solutions are as described in Theorem 2.2.7. Since is a root of multiplicity
d + r + 1 for (x − )d+1 (x), the solutions to S have the form Q(n) n + h(n), where
Q is a polynomial of degree at most d + r and h is a solution to H in terms of the
characteristic roots other than .
By Step 1, R has a solution of this form. The sum of this solution and any
solution to H is another solution to R, so we have one of the form Q(n) n . Fur-
72 Chapter 2: Recurrence Relations

thermore, n times any polynomial of degree less than r is a solution to H, so we

may also add such a sequence to our solution to cancel the coefficients of Q for pow-
ers less than r. We have now shown that there is a solution of the form P(n)nr n ,
where P is a polynomial of degree at most d.
Having proved that there is a particular solution of this form, we can use the
recurrence R to find one. Substitute P(n)nr n into R. With known, cancel n−¾
from both sides of R to obtain
¾
¾ ¾−i
P(n)nr = ∑ ci P(n − i)(n − i)r + F(n) ¾ .
i=1
We choose the coefficients in P(n) to make the polynomials in n obtained on both
sides be the same polynomial. Equating coefficients or evaluating at d + 1 values
of n yields d + 1 equations in d + 1 unknowns to determine the coefficients.

Theorem 2.2.11 can also be proved by the generating function method (Exer-
cise 19). As remarked in the proof, we can force two polynomials to be equal by
equating the coefficients of corresponding powers or by requiring the values to be
equal at more points than the degree. The next example illustrates both methods.

2.2.12. Example. Regions among lines, again (Example 2.1.1). Let an = an−1 + n
for n ≥ 1, with a0 = 1. Using iteration, a recurrence of the form an = an−1 +  (n)
has the solution an = a0 + ∑ i=1  (i). Since ∑i=1 i = (n+2 1 ), we obtain an = 1 + (n+2 1).
n n

Alternatively, we can use the characteristic equation method, which we do

to illustrate the method. Here (x) = x − 1, with root 1. The inhomogeneous
term is n · 1 n , which has the form F(n)cn with F having degree 1 and c being
a characteristic root of multiplicity 1. Theorem 2.2.11 guarantees a particular
solution of the form an = (B1 n + B2)n1 n .
Substituting the sequence B1 n2 + B2 n into the recurrence yields the polyno-
mial equation B1 n2 + B2 n = B1 (n − 1)2 + B2(n − 1) + n to determine B1 and B2 .
When equating corresponding coefficients, the coefficients of n2 yield B1 =
B1 . This gives no information, which is good, since corresponding powers give
three equations but we have only two parameters. The coefficients of n yield
B2 = −2B1 + B2 + 1, or B1 = 1/2. The constant term yields 0 = B1 − B2 , and
hence B2 = B1 = 1/2. The general solution is an = A1 n + (n+2 1 ). Satisfying the
initial condition a0 = 1 yields A = 1.
When evaluating at fixed values, n = 0 yields 0 = B1 − B2 , and n = 1 yields
B1 + B2 = 1. Again B1 = B2 = 1/2, and we obtain the same solution. This method
avoids the work of extracting the coefficients of powers of n.
A simple solution formula to a recurrence may suggest a direct counting ar-
gument. Exercise 13 requests it for this problem.

2.2.13. Remark. The form of the inhomogeneous term in Theorem 2.2.11 seems
quite special, but the method extends to handle sums of such terms. When the in-
homogeneous term is  (n) + (n), a particular solution can be found by summing
solutions obtained when the inhomogeneous term is  (n) or (n) alone. This prop-
erty is called the superposition principle (see Exercise 10 and Exercise 14).
Section 2.2: Elementary Solution Methods 73

THE GENERATING FUNCTION METHOD

Here we introduce generating functions as a tool to solve recurrences and to

explain the characteristic equation method. We postpone the combinatorial as-
pects of generating functions to Chapter 3.

2.2.14. Definition. The generating function for a sequence ⟨a⟩ of complex num-
∞
bers is the formal power series ∑n=0 an x n ; the term is also used for any expres-
∞
sion A(x) with power series expansion ∑n=0 an x n . The coefficient operator
[x¾ ] extracts the coefficient of x ¾ in a (formal) power series in x, so [x¾ ] A(x) =
∞
a¾ when A(x) = ∑ n=0 an x n .

The word “formal” in “formal power series” indicates that x n serves not as a
number but rather as a placeholder for the coefficient an . At present we ignore
the distinction between power series and formal power series, because a Taylor
series expansion of a function (around 0) will also be its formal power series (we
return to this issue in Section 3.2).

2.2.15. Algorithm. The generating function method uses the following steps
to solve a recurrence for ⟨a⟩ (see the example below).
(1) Sum the recurrence over its “region of validity” (the values of the param-
eter where the recurrence holds) to introduce the generating function A(x) and
obtain an equation that A(x) satisfies.
(2) Solve this equation to express A(x) in terms of x.
(3) Find the formal power series expansion and set an = [x n] A(x).

2.2.16. Example. Regions among lines, yet again. Let an = an−1 + n for n ≥ 1,
with a0 = 1. Multiplying by x n and summing over n ≥ 1 yields
∑ an x n = ∑ an−1 x n + ∑ nx n .
n≥1 n≥1 n≥1

With A(x) = ∑n≥0 an x n , the first sum is missing the first term in the generat-
ing function and becomes A(x) − 1. The second needs x factored out to make the
subscript and exponent agree; it is thus x A(x).
∞
Since the geometric series yields ∑ n=0 x n = (1 − x)−1 , termwise differentiation
expresses the third sum as x dx d (1
− x)−1 . Hence our equation for A(x) becomes
A(x) − 1 = x A(x) + (1− x)2 , which simplifies to
x

1 x
A(x) = + .
1 − x (1 − x)3
Now an = [x n] ((1 − x)−1 + x(1 − x)−3 ). The first term contributes 1. For the
second term, we seek the coefficient of x n−1 in the expansion of (1 − x)−3 . We will
see in Lemma 2.2.17 that (1 − x)−3 = ∑ n=0 (n+2 2)x n . Hence the answer is an =
∞

1 + (n+2 1 ), as in Example 2.2.12.

In Example 2.2.16, the generating function method automatically incorpo-

rates the inhomogeneous term and the initial conditions. It applies more gen-
erally than the characteristic equation method, but the latter is usually more
efficient when both apply.
74 Chapter 2: Recurrence Relations

As in Example 2.2.16, completing the generating function method often re-

quires the series for expressions such as (1 − cx)−¾ . We will give a combinatorial
proof avoiding Taylor ’s Theorem in Section 3.1. (When the summand is written
as (n+n¾−1 ), the expression is valid for all real ¾ .)

2.2.17. Lemma. For ¾ ∈ , the power series expansion of (1 − cx)−¾ is

∞
1 n+ ¾ −1 n n
= ∑( )c x .
(1 − cx)¾ ¾−1
n=0

Proof: In Theorem 1.2.5, Taylor ’s Theorem of calculus yields the power series
(1 + y)u = ∑n≥0 (un)y n , using the extended binomial coefficient. With y = −cx
and u = −¾ , we obtain (1 − cx)−¾ = ∑n≥0(−1)n(−n¾)cn x n , called the negative bi-
nomial expansion. By Definition 1.2.4, (−n¾ ) = 1
n! −¾)(n). As in Remark 1.2.6,
(
(−1)n (−n¾ ) = (¾+nn−1 ) = (n+¾−¾−1 1 ).

2.2.18. Application. We describe the general steps for applying the generating
function method to constant-coefficient linear recurrences of order ¾ ; it may be
helful to read the example below along with this.
¾
Suppose an = ∑i=1 ci an−i + (n) for n ≥  . To apply the generating function
method, we multiply by x n and sum over n with n ≥  . After factoring out ci x i ,
the term involving an−i becomes a multiple of A(x) with initial terms missing.
Moving multiples of A(x) to the left and the expressions for missing initial
¾
terms to the right yields Q(x)A(x) = P(x)+ R(x), where Q(x) = 1 − ∑ i=1 ci x¾ (a poly-
nomial), P is a polynomial arising from the initial terms, and R(x) is the formal
¾
power series for the inhomogeneous terms. Note that Q(x) = ∏i=1 (1 −  i x), where
1 , . . . , ¾ are the roots of the characteristic polynomial, since Q(x) = x¾ (1/x).
How do we find R(x)? Consider the inhomogeneous term F(n)cn of Theorem
2.2.11 (with deg(F) = d), so R(x) = ∑n≥¾ F(n)cn x n . Since (n+j −j −1 1) is a polynomial
in n of degree j − 1, we can express R(x) as a linear combination of expressions of
the form (1 − cx)− j for 1 ≤ j ≤ d + 1.
We have written A(x) as P̂(x)/Q̂(x), where P̂ and Q̂ are polynomials, with Q̂(x)
inheriting a higher power of (1 − cx) from the denominator of R(x). A ratio of two
polynomials is called a rational function.
To expand a generating function A(x) that is a rational function, we first use
partial fractions to write the expression for A(x) as a linear combination of ex-
pressions of the form (1 −  x)− j . We then extract [x n] A(x) by expanding these
expressions using Lemma 2.2.17.
The effect of the inhomogeneous term here explains (indeed, proves) the claim
of Theorem 2.2.11. The extra powers of (1 − cx) in the denominator produce the
same effect on the expansion (introducing cn times a higher-degree polynomial) as
the particular solution claimed in Theorem 2.2.11 (see Exercise 19).
The initial conditions determine P. The generating function method auto-
matically picks the one solution to the recurrence that satisfies the initial condi-
tions. To check for computational errors, it may be a good idea to show that the
resulting formula for an satisfies the recurrence.
Section 2.2: Elementary Solution Methods 75

The only part of the method we have not explained in detail is the partial
fraction expansion. This we illustrate before proving that it works.

2.2.19. Example. Partial fractions and the Heaviside method. Consider an =

2an−1 − an−2 + 2 n for n ≥ 2, with a0 = 1 and a1 = 3. Multiply by x n and sum
over n ≥ 2. The inhomogeneous term 2 n turns into (2x)2/(1 − 2x). The generating
function A(x) satisfies

A(x) − 1 − 3x = 2x[A(x) − 1] − x2 A(x) + ∑ n≥2(2x)n ,

which becomes [1 − 2x + x2 ] A(x) = −2x + 3x + 1 + (2x)2/(1 − 2x). Hence

1+x 4x 2 1 − x + 2x 2
A(x) = + = .
(1 − x)2 (1 − 2x)(1 − x)2 (1 − 2x)(1 − x)2

To obtain [x n] A(x), we need the numerators in the partial fraction expansion

A(x) = 1−B x + (1−Cx)2 + 1−D2x . Clearing fractions yields

1 − x + 2x 2 = B(1 − x)(1 − 2x) + C(1 − 2x) + D(1 − x)2 . (∗)

As in Example 2.2.12, there are two methods to compute the constants. Equating
coefficients yields three linear equations for {B , C, D}.
Alternatively, the Heaviside method (named for Oliver Heaviside) saves
work. It uses the fact that equal polynomials are equal at all values. To express
º (x)/∏(1 − i x) as ∑ bi/(1 − i x) when the roots are distinct, first clear fractions.
Letting x = 1/ i sets all but one term on the right to 0 and yields  (1/ i) =
bi ∏j
=i(1 − j/ i). This computes bi directly. With repeated roots, the method
still computes the coefficients of the highest-power terms, and then evaluating at
other appropriate values simplifies computation of the other coefficients.
For example, setting x = 1 and x = 1/2 in (∗) yields 2 = − C and 1 = D/4.
Setting x = 0 yields 1 = B + C + D, and hence B = −1. Finally,
−1 −2 4
an = [x n] ( + + ) = −1 − 2(n + 1) + 4 · 2 n .
1 − x (1 − x)2 1 − 2x
After such a computation, it is a good idea to check that the resulting formula
agrees with the initial conditions!

We have justified the generating function method for recurrences solved by

the characteristic equation method, except for proving that partial fraction ex-
pansion works. The next theorem does that, without saying how to find a partial
fraction expansion. It implies that the linear equations for partial fraction ex-
pansion always have a unique solution.
For simplicity, we state the theorem only in the homogeneous setting. It
extends to inhomogeneous terms of the form F(n)cn because the generating func-
tions for these are again rational functions, as discussed in Application 2.2.18.
The theorem then implies that the particular solutions arise as solutions to ho-
mogeneous recurrences of higher order.
76 Chapter 2: Recurrence Relations

2.2.20. Theorem. Let c1 , . . . , c¾ be complex numbers with c¾

= 0, and let Q(x) =
1 − c1 x − · · · − c¾ x¾ . If Q(x) = ∏i=1 (1 − i x)d i for distinct 1 , . . . , r , then the
r

following are equivalent for a sequence ⟨a⟩.

(A) ⟨a⟩ satisfies an = c1 an−1 + · · · + c¾ an−¾ for n ≥  .
(B) ⟨a⟩ has generating function P(x)/Q(x), where P is
a polynomial of degree less than  .
(C) ⟨a⟩ has generating function that is a linear combination of
expressions of the form (1 − i x)− j , where 1 ≤ j ≤ d i .
(D) an = ∑ri=1 Pi(n) ni for n ≥ 0, where each Pi is
a polynomial of degree less than d i .
Proof: (Stanley [1986, pp. 202–203]) Let V A , V B , VC , V D be the sets of sequences
satisfying conditions A , B, C, D above. These sets are closed under addition and
under multiplication by constants. Hence each is a linear subspace of the space of
sequences in . Each is specified by  independent parameters (initial conditions
a0 , . . . , a¾−1 , coefficients of P , coefficients of  functions of the form (1 − i x)− j ,
or d i coefficients in each Pi). Hence all have dimension at most  .
By Lemma 2.2.5, dim (V A) =  . When a vector space is contained in an-
other of the same finite dimension, they are equal. It therefore suffices to show
V A ⊆ V B ⊆ VC ⊆ V D . We show this by A ⇒ B ⇒ C ⇒ D (equivalent to justifying
the generating function method). The proof just reviews earlier discussion. Let
A(x) be the generating function for ⟨a⟩.
A ⇒ B. Multiply the recurrence by x n and sum over n at least  . The term
¾−i
ci an−i becomes ci x i [A(x) − G i(x)], where G i(x) = ∑ j =0 aj x j . Moving the terms with
A(x) to the left yields Q(x)A(x) = P(x) for a polynomial P of degree less than  .
B ⇒ C. This step is partial fraction expansion. Consider Q(x) =  (x)(x). If
 (x) and (x) have no common zeros, then there exist polynomials C and D such
that C(x) (x) + D(x)(x) = 1 (see Exercise 21). Multiplying by P(x)/Q(x) yields an
expansion; this generalizes for r distinct roots. With P having degree less than
 , in the resulting expansion ∑ri=1 pi(x)/(1 − i x)di the degree of pi is less than
d i . Expanding p(x)/(1 − ax)d into ∑ j =1 bj/(1 − ax) j follows inductively by clearing
d

fractions and matching coefficients.

C ⇒ D. By Lemma 2.2.17, (1− ji x) j = ∑n≥0 bj (n+j −j −1 1) ni x n . For fixed j , the
b

expression (n+j −j −1 1) is a polynomial in n of degree j − 1. A linear combination of

these expressions for 1 ≤ j ≤ d i (using coefficients b1 , . . . , bdi) is a polynomial of
degree less than d i . Doing this for all i yields an = ∑i=1 Pi(n) ni , where each Pi is
r

some polynomial of degree less than d i .

2.2.21. Remark. Stanley treated Theorem 2.2.20 as a statement in the theory

of rational generating functions, part of the calculus of finite differences, whose
origins have been ascribed to Taylor in 1717, Stirling in 1730, and Euler in 1755.
Theorem 2.2.20 also justifies the characteristic equation method, because
Q(x) = ∏i=1 (1 − i x)d i if and only if each i is a characteristic root of multiplicity
r

d i . This follows from Q(x) = x (1/x).

Finding the numerators in the partial fraction expansion of the generating
function is equivalent to finding coefficients to satisfy the initial conditions in the
characteristic equation method. This “trivial” step may take the most effort by
hand but is easily done by computer.
Section 2.2: Elementary Solution Methods 77

The generating function method also works on some nonlinear recurrences

where the equation for A(x) is quadratic or even differential. It may still be pos-
sible to extract coefficients and thus solve the recurrence.
We apply this to the Catalan recurrence, obtaining once again the Catalan
numbers. Here we use products of generating functions for the first time. When
multiplying formal power series, each term in the first is multiplied by each term
in the second, and the coefficient of xr in the product accumulates all contribu-
tions involving r factors of x. That is,
r
(∑ bm x m)(∑ cm x m) = ∑(∑ b¾ cr−¾ )x r .
m≥0 m≥0 r ≥0 ¾ =0

2.2.22. Theorem. If a0 = 1, and an = ∑¾=1 a¾−1 an−¾ for n ≥ 1, then

n

√
1− 1 − 4x 1 2n
∑ an x = n
2x
, and an = ( ).
n+1 n
n≥0

n−1
Proof: Shift the index to rewrite the recurrence as an = ∑¾=0 a¾ an−1−¾ for n ≥ 1.
To apply the generating function method, multiply by x n and sum over n ≥ 1.
The left side becomes A(x) − 1, where A(x) = ∑m≥0 am x m .
∞ n−1
The right side becomes x ∑n=1 ( ∑¾=0 a¾ an−1−¾ ) x n−1 . Letting r = n − 1 writes
this as x[A(x)]2 using the formula for products of power series.
equation for A(x) is A(x) − 1 = x[A(x)]2 . By the quadratic for-
The resulting √
mula, A(x) = (1 ± 1 − 4x)/(2x). We claim that only the minus sign gives a valid
solution. By definition, the coefficient of x−1 in A(x) must be 0, so the constant
terms in the √ numerator must cancel. By the Extended Binomial Theorem (The-
orem 1.2.5), 1 − 4x = ∑¾≥0 (1/2 ¾ )(−4x) . The term for ¾ = 0 in this sum is 1;
¾

hence we must choose the negative sign to make the constants cancel.
The coefficient of x−1 is the only term affected by the 1 in the numera-
tor. Thus to compute an for n ≥ 0, we seek [x n] −21 (1 − 4x)1/2 x−1 , which equals
−1 [x n+1 ](1 − 4x)1/2 . We now have the tools to compute
2

(−1)2 n+1
n
1 1/2 1
an = − ( )(−4)n+1 = (−2)n+1 ∏ ( − i)
2 n+1 2(n + 1)! 2
i=0

(−1)2 n!n n
∏
n
i=1 (2i) ∏
n
i=1 (2i − 1) 1 2n
(n + 1)! n! ∏
= (2i − 1) = = ( ).
(n + 1)n!n! n+1 n
i=0

The generating function method is valid also with multiple indices. Fixing
one index defines a sequence of generating functions, say An(x) = ∑¾≥0 an ,¾ x¾ .
The recurrence for the coefficients yields a recurrence for this sequence. The gen-
erating function method in one variable can then yield a generating function for
the sequence of generating functions, say as B(x , y) = ∑n≥0 An(x)y n .
78 Chapter 2: Recurrence Relations

2.2.23.* Example. Binomial coefficients. Let an ,¾ = an−1 ,¾−1 + an−1 ,¾ for n ≥ 1,

with an ,0 = 1 and a0 ,¾ = 0 for n , ¾ ≥ 1. Failing to notice that this is the recur-
rence given by Pascal’s Formula, we solve for an ,¾ .

For n ∈ 0 , let An(x) = ∑ an ,¾ x¾ . Multiply the original recurrence by x¾ and
sum over ¾ ≥ 1 to get An(x) − 1 = x An−1(x) + An−1 (x) − 1 for n ≥ 1. This simplifies
to An(x) = (x + 1)An−1 (x), with A0(x) = 1.
We now solve the recurrence for ⟨ A⟩. Failing to notice that An(x) = (x + 1)n , we
define B(x , y) = ∑ An(x)y n . Multiplying the recurrence for An by y n and summing
over n ≥ 1 yields B(x , y) − 1 = (1 + x)yB(x , y), and hence B(x , y) = [1 − (1 + x)y]−1 .
Now An(x), the coefficient of y n in the expansion of B(x , y), is (1 + x)n . Ex-
panding this in turn to obtain the coefficient of x ¾ yields (¾n) as the solution to the
original recurrence in two indices.

The generating function method also works on systems of recurrences. We

multiply the recurrences by x n and sum over n to obtain a system of equations for
the generating functions.

2.2.24.* Example. Systems of recurrence relations. Suppose that an = an−1 +

bn−1 , bn = bn−1 + cn−1 , and cn = an−1 − bn−1 , all valid for n ≥ 1. The resulting
system below can be solved for A, B, and C as functions of x.
A(x) − a0 = x A(x) + xB(x) (1 − x)A − xB = a0
B(x) − b0 = xB(x) + xC(x) or (1 − x)B − xC = b0
C(x) − c0 = x A(x) − xB(x) − x A + xB − C = c0

EXERCISES 2.2
To prove a formula by induction, one must already know the formula. Solutions found
non-inductively can be checked (for mistakes) using induction.
2.2.1. (−) Let ⟨a⟩ satisfy an = 2an−1 + 3an−2 for n ≥ 2.
(a) Prove that an is odd for all n ≥ 0 when a0 and a1 are odd.
(b) Solve for an in terms of a0 and a1 .
2.2.2. (−) Let ⟨a⟩ satisfy (n − 3)an = nan−1 − n2 − n for n ≥ 4, with a3 = 10. Prove an =
(n3) + 2(2n) + (1n) for n ≥ 3. Is there a simpler expression for an ?
2.2.3. (−) Let ⟨a⟩ satisfy an = an−4 for n ≥ 4. Use the characteristic equation method to
develop a single formula for an in terms of n and the initial conditions a0 , a1 , a2 , a3 .
2.2.4. (−) Let an = 3an−1 − 2an−2 + 1 for n ≥ 2, with a0 = 2 and a1 = 4. Obtain a simple
formula for an .
2.2.5. (−) At the start of each year, $100 is added to a savings account. At the end of the
year, 5% interest is added. Obtain and solve a recurrence for the account value at the end
of the nth year. The account is empty before the first year.
2.2.6. (−) A mortgage loan begins with $50,000 owed to the bank. During each year, 5%
of the unpaid loan amount is added as interest. At the end of each year, the borrower
pays $5,000. Obtain a recurrence for the unpaid amount at the end of the nth year. Use a
calculator to determine the number of years needed to pay off the mortgage. What happens
if the interest rate is 10% instead of 5%?
Exercises for Section 2.2 79

2.2.7. (−) Consider a set of n circles in the plane such that each intersects every other
(twice) and no three circles meet at a point. Obtain and solve a recurrence for the number
of regions into which the circles cut the plane.
2.2.8. (−) Let 1 , . . . , ¾ be integers, and let  1 , . . . ,  ¾ be real numbers. Let an =
¾
∑ i =1  i
i . Prove that if a0 , . . . , a¾ −1 are integers, then so is an for n ≥ 0.
n

2.2.9. (−) Let ⟨a⟩ satisfy an = c1 an−1 + c2 an−2 , with initial values a0 , a1 . Express the
generating function for a as a ratio of two polynomials.
2.2.10. (−) Let ⟨b⟩ and ⟨d⟩ satisfy inhomogeneous linear  th-order recurrences given by
¾ ¾
an = ( ∑i=1 hi(n)an−i) +  (n) and an = ( ∑i=1 hi(n)an−i) + (n), respectively. Prove that ⟨b⟩ +
¾
⟨d⟩ satisfies an = ( ∑i=1 hi(n)an−i) +  (n) + (n).
2.2.11. (♦) Let an be the number of n-tuples in [4]n that have at least one 1 and have no
2 appearing before the first 1 (note that ⟨a⟩ begins 0 , 1 , 6 , . . .). Obtain and solve a recur-
rence for ⟨a⟩. Give a direct counting argument (without using summations) to prove the
resulting simple formula.

2.2.12. For m , n ∈ , let am ,n be the number of squares (of all sizes, with positive area)
formed by a grid of m horizontal and n vertical lines, given that the two sets of lines are
equally spaced by the same amount. Obtain and solve a recurrence for am ,n . Give a short
alternative argument for a different solution formula.
2.2.13. Example 2.2.12 solves a recurrence to show that the number of regions formed by
n pairwise intersecting lines in the plane is 1 + n + ( 2n) . Give a bijective counting argument
to prove this formula.
2.2.14. Solve the recurrences below, given a0 = a1 = 1.
(a) an = 4an−1 − 4an−2 − n + 6.
(b) an = 5an−1 − 6an−2 + 2n − 1 + 2 n .
2.2.15. The chords of a convex n-gon are the segments joining pairs of vertices. Let an
denote the number of pairs of chords that cross.
(a) Compute an − an−1 .
(b) Solve the recurrence relation in part (a) to obtain a formula for an . (Comment:
Exercise 1.1.21 requests a direct combinatorial argument for the answer.)
2.2.16. (♦) Choose n points on a circle so that no point inside the circle lies on three of the
(n2) chords. Let an be the number of regions formed inside the circle by the chords.
(a) Obtain the recurrence relation an = an−1 +  (n) for n ≥ 1, where  (n) = n − 1 +
n−1
∑i=1 (i − 1)(n − 1 − i), with initial condition a0 = 1.
(b) Solve the recurrence of part (a) to obtain a formula for an as a sum of three bino-
mial coefficients. (Hint: Avoid the characteristic equation method.)
2.2.17. Let an = n3 . Find a constant-coefficient first-order linear recurrence relation sat-
isfied by ⟨a⟩. Does there exist a homogeneous constant-coefficient higher-order linear re-
currence relation satisfied by ⟨a⟩?
¾
2.2.18. (♦) Suppose that ⟨a⟩ satisfies an = ( ∑i=1 ci an−i) + F(n) n , where F is a polynomial
of degree d. Prove that ⟨a⟩ satisfies the homogeneous recurrence whose characteristic poly-
nomial is (x − )d+1 (x), where (x) is the characteristic polynomial of the recurrence that
arises by setting F(n) = 0.
2.2.19. Use the generating function method to prove Theorem 2.2.11. That is, given the
¾
recurrence an = ( ∑i=1 ci an−i) + F(n)cn , where F is a polynomial of degree d and c has mul-
tiplicity r as a characteristic root , prove that there is a solution of the form an = P(n)nr cn ,
where P is a polynomial of degree at most d.
80 Chapter 2: Recurrence Relations


2.2.20. Given (j , ) with j ∈ 0 and ∈ , let an = n j n . The proof of Theorem 2.2.7 can
be completed by showing that for any finite set of distinct such pairs (j , ), the resulting
sequences are linearly independent in the vector space of sequences over , meaning that
no linear combination of these sequences gives the all-0 sequence.
(a) Show that in an equation of dependence, attention can be restricted to the pairs
with | | maximized, and among these to the pairs with j maximized.
(b) For 1 , . . . , m , the Vandermonde matrix is the m-by-m matrix with rs−1 in po-
sition (r, s). Its determinant is ∏i< j (j − i). Use this to complete the proof.

2.2.21. Let [x] denote the family of polynomials in x with complex coefficients. A set in
[x] is an ideal if it is closed under addition and under multiplication by any polynomial.
(a) Prove that every ideal in [x] has the form {p : p ∈ [x]} for some polynomial  .
(b) Prove that if  and  have no common zeros, then there exist c , d ∈ [x] such
that c + d = 1. (Comment: This completes the proof of Theorem 2.2.20. Assume the
Fundamental Theorem of Algebra: every member of [x] is a product of linear factors.)

2.2.22. Prove that the Tower of Hanoi Problem with n rings (Example 2.2.9) is solved in
2 n − 1 moves as follows: On odd-indexed moves, move the smallest ring one peg rightward
(cyclically). On even-indexed moves, move a non-smallest ring. (Buneman–Levy [1980],
Walsh [1983])(Comment: The procedure is essentially unique, although there are other de-
scriptions, such as Hayes [1977], Wood [1981], Cull–Ecklund [1982], and Lavallée [1982].
In Walsh [1982], if we never move the same ring twice in a row, then it suffices to never
let two rings with indices of the same parity touch.)

2.2.23. (♦) For each recurrence below, valid for n ≥ 2 with a0 = a1 = 1, solve for an us-
ing both the characteristic equation method and the generating function method. What
changes if the initial conditions are a0 = 1 and a1 = 0?
(a) an = 3an−1 − 2an−2 + 2 n .
(b) an = 4(an−1 − an−2) + 2 n + 1.

2.2.24. Let an = 4an−1 − 5an−2 + 2an−3 + 2 n−2 for n ≥ 3, with (a0 , a1 , a2) = (1 , 2 , 7). Obtain
a simple formula for an .

2.2.25. For m ∈ 0 , let am = 2− m ∑¾=0 (2m2¾+1 )3¾ .

m
√ √ √ √
(a) Prove am = 12 (1 + 3)(2 + 3)m + 12 (1 − 3)(2 − 3)m .
(b) Use part (a) to prove am = 4am−1 − am−2 for m ≥ 2.
(c) Conclude that am is an odd integer for all m. (Alkan [1995])

2.2.26. (♦) Let an be the number of words of length n on the alphabet {0 , 1 , 2} such that
1 and 2 are never adjacent.
(a) Obtain a second-order recurrence relation for ⟨a⟩.
(b) Solve for an using both the characteristic equation method and the generating
function method.

2.2.27. The  -Fibonacci numbers. Fix  ∈ . Let F ¾ ,n =  F ¾ ,n−1 + F ¾ ,n−2 for n ≥ 2, with
F ¾ ,0 = 0 and F ¾ ,1 = 1.
(a) Solve the recurrence to obtain a formula for F ¾ ,n .
(b) Use the formula for F ¾ ,n to obtain a simple expression for ∑i=0 (2ni+1 ) F ¾ ,2n+1 −2i .
n

Note that (F1 ,0 , F1 ,1 , F1 ,2) = (1 , 5 , 25). (Plaza–Falcón [2016])

2.2.28. The Fibonacci recurrence can be encoded in matrices as ( FFnn+1 ) = (11 10)( FF n ) . Use
n−1 √
this expression and matrix eigenvalues to prove F n = √1 ( n
− n),
where 1 = (1 + 5)/2
√ 5 1 2

and 2 = (1 − 5)/2. (The initial conditions are F0 = 0 and F1 = 1.)

Exercises for Section 2.2 81

¾
2.2.29. (♦) Characteristic roots from eigenvalues. Let ⟨a⟩ satisfy an = ∑i=1 ci an−i .
(a) Express (an , . . . , an−¾+1 )T in terms of (an−1 , . . . , an−¾)T using a matrix A.
(b) Derive the relationship between the roots of the characteristic polynomial of the
recurrence and the eigenvalues of A (with proof).
(c) Use A to determine limn→∞ an/an−1 .
2.2.30. (♦) Given an = 2an−1 + an−2 for n ≥ 2, with a0 = 1 and a1 = 2, use the generating
function method to prove that an = ∑ (i+i! jj +!¾¾!)! , summed over all nonnegative integer triples
(i , j , ¾) such that i + j + 2 ¾ = n. (Comment: This is again the Pell sequence obtained
combinatorially in Exercise 2.1.35.)
2.2.31. Let an ,¾ be the number of compositions of n with ¾ parts. Derive a recurrence and
use the generating function method to obtain ∑ n ,¾≥0 an ,¾ x n y ¾ and compute an ,¾ .

2.2.32. (♦) Let bn ,¾ be the number of ¾-subsets of [n] having no two consecutive integers.
(a) Obtain a recurrence relation in two indices for these numbers.
(b) Use the generating function method in two variables to compute bn ,¾ . (Hint: Fol-
low Example 2.2.23. Let B¾(x) = ∑n≥0 bn ,¾ x n , and use the recurrence of part (a) to obtain
a recurrence for this sequence of generating functions.)
2.2.33. (♦) Let an ,¾ be the number of ways to partition [n] into ¾ sets so that no two con-
secutive numbers are in the same set. (“Partition” means that the ¾ sets are nonempty
and unlabeled; for example, an ,2 = 1.) Define bn ,¾ similarly, except that for bn ,¾ also n is
forbidden to be in the same set with 1; for example bn ,2 is 1 if n is even, but 0 if n is odd.
(a) Obtain a recurrence relation for an ,¾ .
∞
(b) Let A¾(x) = ∑n=0 an ,¾ x n . Use the generating function method to express A¾(x) as
a ratio of polynomials.
(c) Obtain a recurrence for bn ,¾ involving the number an ,¾ . Use this and part (b) to
∞
express ∑n=0 bn ,¾ x n as a ratio of polynomials. (Knuth [2005])
2.2.34. (♦) The Delannoy numbers satisfy d m ,n = d m ,n−1 + d m−1 ,n−1 + d m−1 ,n for m , n ≥
1, with d m ,0 = d0 ,n = 1 for m , n ≥ 0. Find the generating function ∑ m≥0 ∑n≥0 d m ,n x m y n .
Using a clever factorization of the denominator, prove that d m ,n = ∑¾ 2 ¾ (m
¾ )(¾ ).
n
(Hint: Let
(1 − x) and (1 − y) be factors of the denominator. Comment: The combination of Proposition
1.2.10 and Theorem 1.2.13 yields this formula combinatorially.) (Callan [2003c])
2.2.35. (♦) A row of n lightbulbs, initially all off, must be turned on. Bulb 1 can be turned
on or off at any time. For i > 1, bulb i can be turned on or off only when bulb i − 1 is on and
all earlier bulbs are off. Let an be the number of steps needed to turn all on; note that ⟨a⟩
begins (0 , 1 , 2 , 5 , . . .). Let bn be the number of steps to turn on bulb n for the first time.
(a) Find a recurrence for ⟨b⟩ and solve it.
(b) Use ⟨b⟩ to find a recurrence for ⟨a⟩. (Tucker)
(c) Solve the recurrence for ⟨a⟩.
2.2.36. Let an be the number of domino tilings of [0 , 2n] × [0 , 4] in which dominos do not
cross the n − 1 vertical cuts from (2i , 2) to (2i , 4) shown as heavy lines below.
(a) Obtain and solve a second-order recurrence for ⟨a⟩. (Hint: It may be easiest to first
obtain a system of recurrences for ⟨a⟩ and another sequence ⟨b⟩.)
(b) Let C(m , j) be the set of nonnegative integer vectors (x0 , . . . , x j ) with sum m. Prove
¾
an = ∑¾=0 2 n−¾ ∑ x∈ C(n−¾ ,¾) ∏i=0 F̂ 2xi . (Tauraso [2006])
n
82 Chapter 2: Recurrence Relations

2.2.37. (+) Let an count the ways to add noncrossing chords to a fixed convex (n + 2)-gon.
The empty set is allowed, so (a0 , a1 , a2 , a3) = (1 , 1 , 3 , 11). Find a recurrence and obtain
the generating function for ⟨a⟩. (Schröder [1870], Stanley [1971])
2.2.38. (♦) Touchard’s Formula (Shapiro [1976]). Given n points on a circle, let t n count
the ways to pair some by noncrossing chords and color the others red or blue. (Given n
politicians at a circular table, those not in the red party or blue party are shaking hands
in a noncrossing pairing; n need not be even.)
(a) Show that t n = ∑¾≥0 (2n¾)2 n−2¾ C¾ , where Cn = n+1 1 (2n
n)
.
(b) Find a recurrence for t n , and use it to obtain ∑ t n x n .
(c) From parts (a) and (b), prove that Cn+1 = ∑¾≥0 (2n¾ )2 n−2¾ C¾ .
2.2.39. Let L be the L-shaped region of area 3 obtained from [0 , 2] × [0 , 2] by deleting
[1 , 2] × [1 , 2]. Let an count the ways to cut a 4-by-n rectangle into unit squares and unro-
tated translations of L. Express ∑ an x n as a ratio of two polynomials. (Deutsch [2001])
√
2.2.40. (+) Let á = (1 + á2 = á + 1. Let º (1) = 1, and for n ≥ 1 let
5)/2; note that

º (n + 1) = { ºº (n) + 2 if º (º (n) − n + 1) = n
(n) + 1 otherwise
Prove º (n) = ⌊ án⌋ , and determine when º (º (n) − n + 1)
= n. (Doster [1990])

2.3. Further Topics

We consider asymptotic solution of recurrences and a technique using recur-
rences to evaluate summations.

THE SUBSTITUTION METHOD

The substitution method reduces a recurrence to a simpler recurrence for
a related sequence. In a2n = a2n−1 + 2, for example, we can substitute bn = a2n , solve
for bn , and take its square root. In general, we let bn be a suitable function of an
and rewrite the recurrence in terms of ⟨b⟩.
2.3.1. Example. an = nan−1 + n! for n ≥ 1, with a0 = 1. The recurrence is close to
a recurrence for n!. This suggests letting bn = an/n! to obtain
n!bn = n(n − 1)!bn−1 + n!,
which simplifies to bn = bn−1 + 1. Since a0 = 1, we have b0 = 1, and hence the
solution to the new recurrence is bn = n + 1. We obtain an = (n + 1)! as the solution
to the original recurrence.

2.3.2. Proposition. The number Dn of derangements of [n] is given by

n
(−1)¾
Dn = n! ∑ .
¾!
¾ =0

Proof: In Example 2.1.6 we showed Dn = (n − 1)(Dn−1 + Dn−2) for n ≥ 2, with

D0 = 1 and D1 = 0. (Even without knowing D0 , given only D1 and D2 , we can
define D0 to extend the region of validity of the recurrence.)
Section 2.3: Further Topics 83

We use the substitution method. To eliminate the factor n − 1, introduce an

auxiliary sequence ⟨a⟩ defined by an = Dn/n!. Note that an is then the probability
that a random permutation of [n] is a derangement. From D0 = 1 and D1 = 0 we
obtain a0 = 1 and a1 = 0. Substitution yields
n!an = (n − 1)(n − 1)!an−1 + (n − 1)!an−2 ,
and then dividing by n! yields an = (1 − 1/n)an−1 + (1/n)an−2 .
Now rewrite the recurrence for ⟨a⟩ as
an − an−1 = − 1n (an−1 − an−2).
The natural substitution bn = an − an−1 for n ≥ 2 leads to bn = − 1n bn−1 . This
substitution reduces the order of the recurrence, analogous to substitutions re-
ducing the order of differential equations. To recover ⟨a⟩ from ⟨b⟩, use ∑¾=1 b¾ =
n

∑¾=1 (a¾ − a¾−1) = an − a0 ; the sum “telescopes”.

n

The initial condition for ⟨b⟩ is b1 = a1 − a0 = −1. The recurrence then yields
bn = (−1)n/n!. Using the value of a0 and the solution for b¾ ,
n n
(−1)¾ (−1)¾
an = 1 + ∑ =∑ .
¾! ¾!
¾ =1 ¾ =0

Finally, Dn = n! ∑¾=0 (−1)¾/¾ !.

n

The sum ∑¾=0(−1)¾/¾ ! is a partial sum of the series ∑¾≥0 x¾/¾ ! at x = −1. As
n

n grows, the probability of avoiding fixed points rapidly approaches the constant
e−1 , where e denotes the base of the natural logarithm. Other derivations of the
formula appear in Exercise 4, Corollary 3.3.17, and Theorem 4.1.13.
Substitutions can involve some guesswork. The principle of reducing to a
simpler recurrence can suggest helpful substitutions.

2.3.3. Example. Define ⟨a⟩ by nan = (n + 1)an−1 , with a0 = 1. Multiplying the

left side by n + 1 and the right side by n would permit canceling n(n + 1). To
achieve this, let an = (n + 1)bn . (Also suggested by dividing the original recur-
rence by n(n + 1).) We obtain bn = bn−1 , with b0 = 1. Thus bn = 1 and an = n + 1,
which would be quite easy to check if we guessed it initially.

2.3.4. Example. Eliminating summations. Subtracting successive instances of a

n−1
recurrence may eliminate a summation. For example, consider an = ∑i=0 ai for
n−2
n ≥ 1, with a0 = 1. We also write an−1 = ∑i=0 ai . Subtraction yields an − an−1 =
an−1 , and hence an = 2an−1 . The solution to this first order recurrence is an = 2 n .
n−1
Unfortunately, this is wrong, since 2 n
= ∑i=0 2 i . Since our original recur-
n−2
rence is valid only for n ≥ 1, the expression an−1 = ∑ i=0 ai is valid only for n ≥ 2.
Thus we can subtract this expression to obtain an − an−1 = an−1 only when n ≥ 2.
We use the original recurrence to obtain a1 = 1, and now we have an = 2 n−1 for
n−1
n ≥ 1, with a0 = 1. This does check: 2 n−1 = 1 + ∑i=1 2 i−1 .
The subtraction trick can be simpler or quicker than eliminating summations
n−1
by substitution. In this example, substituting bn = ∑i=0 ai into an = ∑i=0 ai
n

yields bn − bn−1 = bn−1 for n ≥ 1, with b0 = 1. Thus bn = 2 n , and an = bn − bn−1 =

2 n−1 for n ≥ 1 as before.
84 Chapter 2: Recurrence Relations

2.3.5. Example. Binary search.

an = a⌈(n−1)/2⌉ + 1 with a0 = 0.
We seek a number x in a sorted list of length n. Let an be the number of
probes we may need. When we probe location j and find y, we compare y with x.
If y
= x, then we consider only the earlier or later part of the list, depending on
whether x < y or x > y. We may always be forced to search the larger part. We
minimize the larger size by probing the middle location; the resulting algorithm
is binary search. After the first probe, we may need to solve a problem of size
⌈ n−2 1 ⌉ . Hence an = 1 + a⌈(n−1)/2⌉ . Also a0 = 0, since no table of size 0 contains x.
For the growth rate of an , consider n when n + 1 is a power of 2. That is,
define a new sequence by b¾ = a2¾ −1 for ¾ ≥ 0. Substitution in the recurrence for
⟨a⟩ yields b¾ = b¾−1 + 1. Also a0 = 0 yields b0 = 0, and hence b¾ = ¾ for ¾ ≥ 0. Thus
an = log2(n + 1) when n + 1 is a power of 2. The full solution an = ⌈ log2(n + 1)⌉
follows by induction on n; it is the statement “ an = ¾ for 2¾−1 ≤ n < 2¾ ”.

2.3.6.* Remark. The technique of splitting a problem into smaller subproblems

and combining the solutions is called divide-and-conquer. Solutions to the re-
currences that result from this process can sometimes be found by substitutions.
Consider an = ca⌈n/d⌉ + e. To study this for powers of d, let b¾ = ad ¾ , so b¾ =
cb¾−1 + e. When c
= 1, we obtain b¾ = Ac¾ + 1−e c , or an = Anlog d c + 1−e c .
For a linear inhomogeneous term en, the behavior depends on the ratio c/d.
The substitution yields b¾ = cb¾−1 + ed¾ . When c
= d, we obtain b¾ = Ac¾ + ded ¾
−c d ,
yielding an = An log d c
+ d−c n (with A determined by the initial condition). When
ed

c = d, the solution is b¾ = Ad¾ + e ¾ d¾ , which becomes an = An + en logd n.

ASYMPTOTIC AN ALYSIS

Exact solutions may be unnecessary or too complicated. We may be content

with a simple formula that approximates or is asymptotic to an . For example,
in a linear recurrence with constant coefficients it may be enough to know the
exponential growth that arises from the characteristic root of largest magnitude.
The substitution method can provide asymptotic information. We “peel off
the controlling behavior ” by substitutions, yielding a simpler recurrence. If sub-
stitutions yield a constant sequence, then we have explained everything. Our
first example is easy to solve exactly.

2.3.7. Example. The Tower of Hanoi, again: an = 2an−1 + 1 with a0 = 0. We

expect an to be about 2 n , so we set an = 2 n bn (setting an = 2 n + bn does not help!).
Substituting and dividing by 2 n yields bn = bn−1 + 2−n , with b0 = 0. Iterating
yields bn = ∑i=1 2−i = 1 − 2−n , and hence an = 2 n − 1.
n

Asymptotic analysis first seeks the most extreme behavior. This may be facto-
rial growth (or worse). Next is exponential behavior and then perhaps polynomial.
Finding substitutions to extract the extreme behavior may take some cleverness.
To avoid guessing, we make substitutions with parameters and then determine
the values of the parameters that simplify the unexplained behavior.
Section 2.3: Further Topics 85

2.3.8. Application. Stirling’s Formula. Consider an = nan−1 for n ≥ 1, with

a0 = 1. The solution is an = n!. Computation with n! is difficult for large n, so
we seek an approximation using simpler functions. Since n < n! < nn for fixed
when n is large, we set bn = an/nn and study bn . Substituting into an = nan−1
yields bn = (1 − 1/n)n−1 bn−1 . Since (1 − 1/n)n−1 → e−1 , we have the “asymptotic
recurrence” bn ∼ e−1 bn−1 . This suggests that the behavior of ⟨b⟩ is exponential,
which is “simpler ”. (We could have tried the more general bn = an/n n , but then
we find that setting  = 1 eliminates the superexponential behavior from ⟨b⟩.)
We expect bn to be close to e−n , so we set cn = en bn . Substituting into the re-
currence for ⟨b⟩ yields cn = e(1 − 1/n)n−1 cn−1 . The ratio cn/cn−1 tends to 1, which
is necessary (but not sufficient) for the behavior of ⟨c⟩ to be “algebraic” (bounded
by a polynomial). A careful look at cn/cn−1 will lead us to the right exponent for
the leading term. Suppose cn ∼ n for some constants  and . This yields
cn/cn−1 ∼ n/(n − 1) = (1 − 1/n)− . By the Extended Binomial Theorem (Theorem
1.2.5), the expansion is 1 + n + (2n
+1)
2 + · · · ; only the linear term is accurate since
we parametrized only the leading behavior of cn .
To determine , we compare with our exact expression for the ratio: cn/cn−1 =
e(1 − 1/n)n−1 . Letting W = cn/cn−1 , we obtain
ln W = ln e + (n − 1) ln(1 − 1/n)
= 1 + (n − 1)(− 1n − 2n1 2 − 3n1 3 − 4n1 4 − · · ·)
= (1·12)n + (2·3)n
1
2 + (3·4)n3 + · · · .
1

Using the exponential series on eln W yields a series for W in powers of 1/n. We
obtain W = 1 + 2n1
+ 24n
7
2 + 16n3 +· · · . This yields
3
= 12 in the expansion for cn/cn−1 .
To further refine the analysis, define d n by cn = n1/2 d n . Substituting into
the recurrence for ⟨c⟩ yields the exact recurrence d n = e(1 − 1/n)n−1/2 d n−1 . Now
we hope that d n tends√to a constant C. If so, then undoing the substitutions so
far yields n! ∼ C(n/e)n n.
To prove this, let n = ln d n ; we show that n tends to a limit. By the Mono-
tone Convergence Theorem, it suffices to show that ⟨ ⟩ is decreasing and bounded
below. The recurrence for d n yields n = n−1 + 1 + (n − 1/2) ln(1 − 1/n). Note that
b1 = a1 = 1, d1 = c1 = e, and 1 = 1. Expanding the series for ln(1 − 1/n) shows
that the product (n − 1/2) ln(1 − 1/n) is less than −1; thus ⟨ ⟩ is decreasing, and
n < 1 for n > 1. To obtain a lower bound, we compute
n n
1 1 1 1 1 1
n = 1 + ∑ [1 + ( − ) ln (1 − )] = 1 + ∑ [1 − ( − ) ( + 2 + 3 +· · ·)]
 =2
2   =2
2  2  3
n n
1 1 t−1
= 1 − ∑∑( − )  −t = 1 − ∑  −t
t + 1 2t 2t(t + 1) ∑
=2 t≥2 t≥2  =2
#∞
t − 1 −1 −t+1 ####
∞
t−1 −t 1
>1−∑  d = 1 − ∑  ###
2t(t + 1) ∫1 2 t(t + 1) t − 1 ###=1
t≥2 t≥2
1 1 1 1 1 3
=1− ∑ = 1 − ∑( − )= .
2 t(t + 1) 2 t t+1 4
t≥2 t≥2
86 Chapter 2: Recurrence Relations

Since ⟨º ⟩ is decreasing and bounded below, it has a limit, by the Monotone

Convergence Theorem. Undoing the substitutions yields n! = an = rnne−n n1/2 ,
where e3/4 < r < e. Furthermore, we have proved that n!/(nn e−n n1/2) tends to a
constant C. This asymptotic approach cannot compute C, although √ we have con-
fined it between 2.1170 and 2.7183. We will in fact prove C = 2 ≈ 2.5066.
Continuing the asymptotic analysis expands d n in powers of 1/n. For exam-
ple, set d n = e n(1 + n). We choose  to make e n/e n−1 as close to 1 as possible.
Computing e n/e n−1 by substituting into d n/d n−1 yields e n/e n−1 = 1 + O(n−3) when
 = 12
1 (Exercise
12). The process continues forever, producing what is called an
asymptotic series; √ each substitution computes another coefficient. Using the
overall constant 2 computed below, the resulting asymptotic expansion is

√ n n 1 1 139
n! = 2 n ( ) [1 + + 2
− − · · ·] .
e 12n 288n 51840n3

√
2.3.9. Theorem. (Stirling’s Formula) n! ∼ nn e−n 2 n.
Proof: Let sn = nne−n!n n1/2 . The outcome of the asymptotic analysis in Application
2.3.8 is a proof that sn tends to a constant C. To compute the constant, we use the
∞
Wallis product, which expresses /2 as an infinite product: 2 = ∏=1 22− 1 22+ 1
√
(derived in Exercise 13). We combine the exact expression n! = sn(n/e)n n and
the fact that sn tends to a limit value C to compute
n
2 2 42n(n!)4
= lim∞ ∏ = lim∞
2 n→ 2 − 1 2 + 1 n→ (2n + 1)(2n)! 2
 =1
√
42n(sn(n/e)n n)4 s4n n2 C4 C2
= lim∞ √ = lim∞ = = .
n→ (2n + 1)(s2n(2n/e)2n 2n)2 n→ (2n + 1)s2 · 2n 4C 2 4
2n
√
From this we conclude C = 2 .

2.3.10. Example. Coin-flipping. When a fair coin is flipped n times (and n is

even), what is the probability of having the same number of heads and tails?
n
There are 2 n equally likely sequences, and exactly (n/2) have the same number of
heads and tails. Stirling ’s Formula yields
√
n n! nne−n 2 n 2n
( )= ≈ √ = √ .
n/2 (n/2)!2 [(n/2)n/2 e−n/2 2 (n/2)]2 n/2

Dividing√by 2 n tells us that the probability of equal numbers of heads and tails is
about c/ n, where c = (2/ )1/2 . For (nn), see Exercise 14.
Without knowing Stirling ’s Formula, we can still obtain the leading (expo-
n
nential) behavior of (n/2 ). We have observed that lim x n+1/x n = L yields exponen-
tial behavior of the form Ln . More precisely, lim x n+1/x n = L implies lim x1/n
n = L
(Exercise 15). Taking the ratio of successive middle binomial coefficients yields
n
2 n as the leading behavior of (⌊n/2⌋ ).
Section 2.3: Further Topics 87

2.3.11. Example. Inversion of functions. A natural question in many contexts is

how big n must be so that (n/2n
) ≥ t. Example 2.3.10 yields an approximate answer.
√
Roughly, we are given t = c2 n/ n, where c is a constant, and we want to invert
the expression of t in terms of n to obtain n in terms of t. The process is similar
to asymptotic analysis of recurrences: we approximate the leading behavior and
use substitutions to modify it to a more accurate solution.
We first take logarithms, where lg t denotes the base-2 logarithm: lg t ≈ n −
1
2 lg n, dropping the additive constant. Setting n to lg t leaves the right side too
small. Since the right side increases with n, we must enlarge n. With experience,
it is apparent that n should be about lg t + 12 lg lg t, but where does this come from?
From the leading behavior, substitute by setting n = (lg t) + x, with x a func-
tion of t. Now the right side is lg t + x − 12 lg(x + lg t). If we correctly found the lead-
ing behavior, then x ∈ o(lg t), and we rewrite the last term as 12 lg[(lg t)(1 + x/lg t)],
which equals 12 [lg lg t + lg(1 + x/lg t)]. When x/lg t → 0, our overall expression
“behaves like” lg t + x − 12 lg lg t, which simplifies to lg t when x = 12 lg lg t.
√ Essentially, the lower order term explains the effect of the lower order factor
n in the original function. The point is that when we set n to be lg t + 12 lg lg t,
the difference between the two sides is of lower order than when we set n to be
lg t. Thus we can write n = lg t + 12 lg lg t + o(lg lg t). With further substitutions,
we could reduce the error even further.

THE WZ METHOD (optional)

Many combinatorial identities evaluate sums. Combinatorial proofs are hard

to find, so mathematicians have sought automated methods to compute sums,
analogous to symbolic computation for integration. Here we describe a technique
that enables computers to evaluate many sums that arise in practice.
The technique relates instances of a formula over different values of a param-
eter, like a recurrence does. Also, recurrences are used in the method (albeit in a
step we will not explore in depth). This makes the method more an application of
recurrences rather than a derivation or solution of recurrences. Like other sub-
sections marked “optional”, this is a topic not used in the rest of the book and
may be skipped without loss of continuity.
Developed and popularized by H. Wilf and D. Zeilberger, this method is
known as the WZ method. It is sketched in Wilf [1990, pp. 120–126] and pre-
sented in detail in Petkovšek–Wilf–Zeilberger [1996], a book with the simple title
“ A = B ”.
Consider the identity ∑¾ U(n , ¾) = H(n) for n ≥ 0. Dividing by H(n) yields
an equivalent identity ∑¾ F(n , ¾) = 1. This transformation discards the bi-
jective approach; we are no longer summing integers. Nevertheless, since the
right side is now constant, it suffices to verify the equality at n = 0 and prove
∑¾ [F(n + 1 , ¾) − F(n , ¾)] = 0 for n ≥ 0.
The crux of the WZ method is finding an expression G(n , ¾) such that
F(n + 1 , ¾) − F(n , ¾) = G(n , ¾ + 1) − G(n , ¾). This turns the needed sum into a
telescoping sum.
88 Chapter 2: Recurrence Relations

2.3.12. Proposition. If F and G satisfy

F(n + 1 , ¾) − F(n , ¾) = G(n , ¾ + 1) − G(n , ¾)
and lim¾→±∞ G(n , ¾) = 0, then ∑¾ F(n , ¾) is independent of n.
Proof: Given such an expression, we write the difference in ∑¾ F(n , ¾) for two
successive values of n in terms of G:
B B

∑[F(n + 1 , ¾) − F(n , ¾)] = ∑[G(n , ¾ + 1) − G(n , ¾)] .

¾= A ¾= A

The sum over G telescopes to G(n , B + 1) − G(n , A). Summing over all ¾ yields
the desired result when lim B→∞ G(n , B) = lim A→−∞ G(n , A) = 0.

A pair (F , G) satisfying the conditions of Proposition 2.3.12 is a WZ pair and

certifies that ∑¾ F(n , ¾) is independent of n.

2.3.13. Example. ∑¾ (¾n) = 2 n. Of course we know this, but let us see how the
WZ method proves it. Let F(n , ¾) = (n¾ )/2 n . If we set G(n , ¾) = −(¾−n1 )/2 n+1 , then
the behavior of the binomial coefficient yields lim¾→±∞ G(n , ¾) = 0, and we need
only verify the difference condition. The check is

(n+¾ 1 ) 2(¾n) −(¾n) −(¾−n1 )

− = − n+1 .
2 n+1 2 n+1 2 n+1 2

By Proposition 2.3.12, ∑¾ F(n , ¾) is independent of n. To show that the value is

1, we set n = 0 in the sum, obtaining ∑¾ (0¾ )2−0 = 1.

+1 ,¾) ,¾ −1)
The difficulty is finding G(n , ¾). When F(n F(n ,¾)
and F(n
F(n ,¾)
both are rational
functions of n and ¾ , and ∑¾ F(n , ¾) = 1, usually there exists such a function G
of the form G(n , ¾) = R(n , ¾)F(n , ¾ − 1), where R(n , ¾) is a rational function of n
and ¾ . Moreover, there is an algorithm that finds such G when it exists or proves
that it does not exist. This is the advantage of computerizing the process, along
with the fact that computers can quckly perform manipulations that humans do
not even want to try.
The computation in the example above shows that R(n , ¾) = − 12 works when
F(n , ¾) = (¾n)/2 n . To prove ∑¾ (−1)¾ (¾n)(2¾¾ )4n−¾ = (2n n
), the magic expression is
2¾ −1 (see
R(n , ¾) = 2n +1 Exercise 19). The next one looks more complicated.

2.3.14. Example. Dixon’s Identity. The identity is

∞
a+b a+c b+ c (a + b + c)!
∑ (−1)¾ ( a + ¾ )(c + ¾)(b + ¾) = a!b!c!
.
¾ =−∞
(c+1−¾)(b+1−¾)
When we set a = n, the magic expression is R(n , ¾) = 2(n+¾)(n+b+ c+1)
.
Section 2.3: Further Topics 89

To seek the expression G(n , ¾) to mate with F(n , ¾), we treat n as a param-
eter, defining () = F(n + 1 , ) − F(n , ) for fixed n and seeking () such that
( + 1) − () = () to become G(n , ). Since summation is the inverse to differ-
¾
ence, we want () = ∑ j =0 (j) if the values are defined for  ≥ 0. However, we
also want () in “closed form”.

2.3.15. Definition. An expression  in terms of  is in closed form if ( +1)/()

is a rational function of  . Such an expression  is also called a hypergeo-
metric term. A hypergeometric series is a power series where the ratio
of successive coefficients is a rational function of the summation index.

In a geometric series, the term ratio (ratio of successive terms) is constant.

In a more general hypergeometric series, the term ratio is a ratio of polynomials.
Although  ! and most ratios of binomial coefficients are in closed form,  ¾ is not.
Closed form may involve factorials, polynomials, and exponentials in the index.
For () in closed form, an algorithm of Gosper [1978] tests whether there is
a hypergeometric term () such that () = ( + 1) − (), constructing it if so.
Graham–Knuth–Patashnik [1989] presents the algorithm (see also Petkovšek–
Wilf–Zeilberger [1996]); it is available in some symbolic computation software
packages. We sketch the steps.

2.3.16. Algorithm. (sketch of Gosper ’s Algorithm). Given a hypergeometric

term (), we seek  such that () = ( + 1) − ().
(1) Let r() = ( + 1)/ ().
¾) c(¾ +1)
(2) Express r() as a(b(¾) c(¾)
, where a and b are polynomials such that a() and
b( + j) are relatively prime for all j ∈ 0 . 
(3) Find a nonzero polynomial h solving a()h( + 1) − b( − 1)h() = c(), if
such a polynomial exists.
(4) If (3) fails, say  does not exist. Otherwise, let () = b(¾−c(1)h(
¾)
¾)
().

As discussed in Petkovšek–Wilf–Zeilberger [1996] (Section 5.3), every ratio-

nal function r has a factorization as requested in Step 2, and there is an algorithm
to find it. We then seek h in Step 3 by equating corresponding coefficients.
When has a particular form, we can state the factorization in Step 2 ex-
plicitly. If there are polynomials p and q and constants mi such that
p()(m1 + )!(m2 − )!
() = ¾ ,
q()(m3 + )!(m4 − )!
then first simplify by cancellation (for example, write (m1 + )!/(m3 + )! as (m1 +
)(m1 −m3) when m1 ≥ m3). Dropping factors that are then absent, we can choose
the polynomials a, b, and c roughly as below
a() = (m1 +  + 1)(m4 − )q() ,
b() = q( + 1)(m3 +  + 1)(m2 − ) ,
c() = p() .
90 Chapter 2: Recurrence Relations

2.3.17. Example. We apply Gosper ’s Algorithm to find the magic expression in

Example 2.3.13. We are given F(n , ¾) = (¾n)2−n , so

n+1 n 1 n 2¾ − n − 1
º (¾) = ( )2−n−1 − ( )2−n = ( ) .
¾ ¾ 2 n+1 ¾ n + 1 − ¾

Dividing º (¾ + 1) by º (¾) yields r(¾) = 2¾¾−+n1+1 2n¾+−1n−−¾1 . We set a(¾) = n + 1 − ¾ ,

b(¾) = ¾ + 1, and c(¾) = 2 ¾ − n − 1. For Step 3, we seek a solution to
(n + 1 − ¾)h(¾ + 1) − ¾ h(¾) = 2 ¾ − n − 1 .
Choosing h(¾) = −1 satisfies the equation. The algorithm now outputs
¾(−1) 1 n 2¾ − n − 1 −¾ n 1 −1 n
G(n , ¾) = ( ) = ( ) n+1 = ( ).
2¾ − n − 1 2 n+ 1 ¾ n+1 −¾ n+1 −¾ ¾ 2 2 ¾−1

With F(n , ¾) = (¾n)s−n , indeed G(n , ¾) = − 12 F(n , ¾ − 1), as in Example 2.3.13. In

general, as stated in Step 4, we are setting

b(¾ − 1)h(¾)
G(n , ¾) = (F(n + 1 , ¾) − F(n , ¾)) .
c(¾)

2.3.18. Proposition. If Gosper ’s Algorithm returns successful output ½(¾) from

input º (¾), then ½(¾ + 1) − ½(¾) = º (¾).
Proof: We compute

b(¾)h(¾ + 1) b(¾ − 1)h(¾)

½(¾ + 1) − ½(¾) = º (¾ + 1) − º (¾)
c(¾ + 1) c(¾)
b(¾)h(¾ + 1) b(¾ − 1)h(¾)
= º (¾) [ r(¾) − ]
c(¾ + 1) c(¾)
a(¾)h(¾ + 1) b(¾ − 1)h(¾)
= º (¾) [ − ]= º (¾) .
c(¾) c(¾)

Before concluding that the WZ method solves everything, note that our dis-
cussion of the method requires advance knowledge of the value of the sum. Oth-
erwise, we wouldn’t even know the expression F(n , ¾) that starts the process!
Fortunately, there are also methods to automate finding the sum. These stem
from an algorithm in the thesis of Sister Mary Celine Fasenmyer [1946]. The ba-
sic idea is to find a recurrence satisfied by the summand and then sum it to find
a recurrence satisfied by the sum. Fasenmyer ’s algorithm automates the search
for the first recurrence when the summand belongs to an appropriate class.
Zeilberger [1982, 1990, 1991] developed faster and more general algorithms.
Petkovšek [1991] automated the search for solutions of the resulting recurrences.
The algorithms find sums of hypergeometric terms or show that the sum has no
formula of a certain type. Again, see Petkovšek–Wilf–Zeilberger [1996].
Exercises for Section 2.3 91

EXERCISES 2.3

2.3.1. (−) Use the substitution method to solve the recurrences below.
(a) an = (1 − n+1 1 )an−1 for n ≥ 1, with a0 = 1.
(b) an = 23 (1 + 3 n2+1 )an−1 for n ≥ 1, with a0 = 1.

2.3.2. (−) Give the exact solution formula (for all n) for the recurrences below.
(a) an = a⌈n/2⌉ + 1 for n ≥ 2, with a1 = 0.
(b) an = a⌈(n−1)/2⌉ + 1 for n ≥ 2, with a1 = 1. (Comment: This is the recurrence for the
worst-case number of probes in binary search for an element in a sorted n-element list.)

2.3.3. (−) Let Hn = ∑¾=1 1¾ . Prove for n ∈

n
 that ∑in=−11 Hi = n(Hn − 1).
2.3.4. (−) Use the recurrence Dn = (n − 1)(Dn−1 + Dn−2) of Example 2.1.6 (with D0 = 1 and
D1 = 0) to prove Dn = nDn−1 +(−1)n . Use the new recurrence to prove Dn = n! ∑¾=0(−1)¾/¾ !.
n

2.3.5. Find the general solution to an = an/2 + 2an/4 + 3n/4 when n = 2 ¾ .

2.3.6. (♦) Solve the recurrence an = 2a2n−1 with a0 = 1.

2.3.7. (♦) Finding the max and min in a set S. When | S| = 1, the one element is both
answers. When | S| = 2, compare the two numbers and return the result. When | S| > 2,
split S into sets T and T of sizes ⌊ n/2⌋ and ⌈ n/2⌉ , apply the algorithm to T and T , and
compare the answers for the two subsets. Let an be the number of comparisons used when
| S| = n. Derive a recurrence for an . Use the substitution method to obtain a formula for
an when n is a power of 2.

2.3.8. (♦) In the game “High/low” , a number is chosen uniformly at random from [n],
where n is known. The player makes a guess, winning the prize if correct. Otherwise,
she learns whether the guess is high or low. When each guess is made randomly from
the remaining possible values, the expected numbers of guesses when n ∈ {0 , 1 , 2 , 3} are
0 , 1 , 3/2 , 17/9, respectively.
The harmonic number Hn is defined by Hn = ∑¾=1 1/¾ . Obtain a formula for the
n

expected number of guesses made to win the prize, using one harmonic number in the for-
mula. (Hint: Obtain a recurrence with many terms. Simplify via natural substitutions
and reduce to a first-order recurrence.) (Taylor [1991])

2.3.9. Let n(n − 3)an = (n − 1)(n2 − 3n + 1)an−1 − (n − 2)3 an−2 for n ≥ 4, with a2 = a3 = 1.
Use substitutions to obtain a simple formula for an . (Wilf [1997])

2.3.10. Let an = an−1 + an−2/(n − 1) for n ≥ 2, with a0 = and a1 =  . Find a formula for
an (one summation may appear in the formula).

2.3.11. Let an+1 = (2n + 3)an − 2nan−1 + 8n for n ≥ 1, with a0 = 1 and a1 = 3. Use
substitutions to obtain an asymptotic formula for an . (Doster [1994])

2.3.12. (♦) Continuation of Application 2.3.8 for Stirling’s Formula. Given the recurrence
d n = e(1 − 1/n)n−1/2 d n−1 , let d n = e n(1 + /n) and choose  so that e n/e n−1 = 1 + O(n−3).
/2
2.3.13. (♦) Let an = ∫ (sin x)n dx for n ∈ 0 .
0

(a) Use integration by parts to prove an = n−n1 an−2 for n ≥ 2.
∞
(b) Use the sequence ⟨a⟩ to prove the Wallis product: ∏n=0 2n 2n
2n−1 2n+1
= 2.
92 Chapter 2: Recurrence Relations

¾
2.3.14. Approximations for binomial coefficients. For ¾ fixed and n growing, (¾n) ∼ n¾ ! .
−n
n ( (1 − )1− )
(a) Using Stirling ’s Formula, prove ( )≈ √ .
n (1 − )2  n
(b) An ordinary deck of cards consists of 13 cards each in four suits. A bridge hand
distributes these randomly to four players around a table, each receiving 13 cards. Oppo-
site players are partners. To the nearest power of 10, approximate the probability that
some partnership will have all 13 cards of one suit.

2.3.15. Consider limits taken as n → ∞.

(a) Prove that if lim(x n+1 − x n) = L, then lim x n/n = L.
(b) Apply part (a) to prove that if lim(an+1/an) = M , then lim(a1/n
n ) = M.
1/n
(c) Conclude lim((⌊n/2⌋
n
) ) = 2.
2.3.16. Balls 1 , . . . , n will go in boxes 1 , . . . , n. Put ball 1 into a randomly chosen box.
Thereafter, for j from 2 to n, put ball j in box j if box j is empty; otherwise, put it in a
random empty box. Let n be the expected number of j such that ball j is in box j .
n
1 1
(a) Prove n = 1 + ∑ ( − 2 + n−+1 − ).
n n−  +1
 =2
(b) Conclude n = n − Hn−1 , where Hm = ∑ j =1 1/j . (Deshpande–Deshpande [2010])
m

−)!
2.3.17. Suppose an = 2an−1 + an−2 for n ≥ 2, with a0 = 1 and a1 = 2. Let cn ,i , = i!(n(n −i−2)!  !
when i,  , and n − i − 2  are all nonnegative; otherwise cn ,i , = 0.
(a) Prove cn ,i , = cn−1 ,i , + cn−1 ,i−1 , + cn−2 ,i ,−1 for n ≥ 2 and all i and  .
(b) Use part (a) to prove an = ∑ (i+i! jj +!!)! , summed over nonnegative integer (i , j , ) with
i + j + 2  = n. (Comment: This is again the Pell sequence.) (Stockmeyer [2000])

2.3.18. (+) An usher seats n people in a row of n narrow seats, choosing a seat for each
arrival in turn. In order to seat a new arrival, the people neighboring that seat in each
direction up to the first empty seat must stand so they can sit together at one time. For
example, if seat 4 is next filled when seats 1 , 3 , 5 , 6 , 8 are already filled, then those in seats
3 , 5 , 6 must stand so that 3 , 4 , 5 , 6 all sit together. The usher wants to minimize the total
number of seatings, counting 1 each time any person sits. Let  (n) be the optimal value;
for 1 ≤ n ≤ 5, the answers are 1 , 3 , 5 , 8 , 11. In general, filling the seats in order from one
end yields (n2 + n)/2 seatings (the worst!), but the optimum is asymptotic to n log 2 n.
(a) Let (n) = (n + 1) − 2  + 1, where  = ⌈ log 2(n + 1)⌉ . Prove (n) = (⌊ (n − 1)/2⌋ ) +
(⌈ (n − 1)/2⌉ ) + n.
(b) Prove that (n) − (n − 1) is a nondecreasing function of n.

(c) Prove  (n) = (n) for n ∈ . (Fredman–Knuth [1974], Vanden Eynden [1991])

2.3.19. Use the WZ method with R(n , ) = 22n−+11 to prove

n 2 2n
∑(−1) ()(  )4n− = ( n ) .


2.3.20. Use the WZ Method with G(n , ) = −n F(n , ) to prove that
t −1
1 t 1 t+n
∑(−1) ( )=
n+   n
(
t
) . (Wilf)
 =0

(Comment: Here F(n , ) is the summand divided by the right side; ∑ F(n , ) = 1 can also
be proved using Vandermonde’s Theorem and other familiar identities.)
Chapter 3

Generating Functions
∞
In Section 2.2, we called ∑n=0 an x n the generating function for a sequence
⟨a⟩ of complex numbers. There it was just a tool for solving recurrence relations,
but generating functions have deeper combinatorial aspects. They yield system-
atic solutions for many counting problems. Often we obtain an expression for the
generating function and a formula for an directly, without using recurrences.
There are many types of generating functions (see Stanley [1986]). We have
only mentioned the ordinary generating function (OGF), useful in studying multi-
sets and selection problems. A variation called the exponential generating function
(EGF) is useful for enumeration of “labeled” structures. We will also apply OGFs
to study partitions of integers.

3.1. Ordinary Generating Functions

Let an be the solution to a counting problem with n as a parameter. A problem
may have many parameters; at present we treat all but one as fixed and use the
varying parameter as the index of a sequence. We associate a generating function
in one variable with this counting sequence.
∞
3.1.1. Definition. A formal power series is an expression of the form ∑n=0 an x n .
The (ordinary) generating function (OGF) for a sequence ⟨a⟩ of complex
∞
numbers is the formal power series ∑ n=0 an x n or any expression having this
formal power series expansion (also called generating series in the litera-
ture on formal power series).
When objects in a universe U are assigned nonnegative integer values by
a weight function w, with an = |{u ∈ U: w(u) = n}| , we say that the generat-
ing function for ⟨a⟩ is the enumerator by w for U or is indexed by w.

3.1.2. Example. Let an be the number of binary lists of length n. The gener-
∞
ating function for ⟨a⟩ is ∑n=0 2 n x n ; it is the enumerator of binary lists, indexed
by length. We say “indexed by length” because the values of the weight function
index the terms in the sequence ⟨a⟩.
Next consider ¾-subsets of an n-set. We fix one parameter to form a generat-
ing function, writing a¾ = (¾n). In A(x) = ∑¾=0 (¾n)x¾ , the coefficient of x ¾ is the
n

93
94 Chapter 3: Generating Functions

number of ¾-subsets of [n]. Thus A(x), or more properly An(x), is the enumera-
tor of subsets of [n], indexed by size. Since the coefficient of x¾ in the expansion
of (1 + x)n is (¾n), we also say that (1 + x)n is the generating function for ⟨a⟩.
In a problem with several parameters, it may not be obvious which is most
useful as the index of summation in a generating function. We can also form a
∞ ∞
generating function in two variables: ∑ n=0 ∑¾=0 an ,¾ x n y ¾ .

3.1.3.* Remark. The word “function” in “generating function” indicates only

the notational form; x n is just a placeholder for the term an in the sequence, not a
number. In algebraic combinatorics the term generating series is used instead
of generating function; the series “generates” the sequence of coefficients.
Since x is not numerical, we call x a “formal variable”. Our first proof of
(1 + x)n = ∑¾n=0 (¾n)x¾ manipulated x as a symbol, not a number. This viewpoint
also excuses the notation “ A(x)” for a generating function. One should not write
“ º (x)” for a numerical function º , since º (x) is a number. We write “ A(x)” for a
generating function because “argument ” and “value” have no numerical mean-
ing; the notation indicates that we are encoding a sequence using powers of x.
The question of convergence when the formal variable takes a numerical
value is mostly irrelevant, but numerical convergence can be helpful in extract-
ing asymptotic formulas for coefficients (Section 3.4).

MODELING COUNTING PROBLEMS

The importance of generating functions arises from the way they model
counting arguments using the Sum and Product Principles, based on the behav-
ior of sum and product of formal power series. This allows generating functions
to model ad hoc counting for a sequence of counting problems all at once.
∞
3.1.4. Definition. The sum and product of formal power series ∑n=0 an x n and
∞
∑n=0 bn x n are defined as follows:
∞ ∞ n
sum : ∑(an + bn)xn product : ∑(∑ aj bn− j )x n .
n=0 n=0 j =0

The convolution of two sequences ⟨a⟩ and ⟨b⟩ is the sequence of coefficients
in the product of their generating functions. The term convolution is also
n
used for a sum of the form ∑ j =0 aj bn− j .

3.1.5. Example. Multisets from two types of objects. The formal power series
∞
∑¾=0 x¾ is the enumerator by size for multisets from one type of object. There is
exactly one multiset consisting of ¾ objects of the same type. Marbles of the same
color are indistinguishable, so there is only one way to have 10 white marbles.
With two types, such as black and white marbles, there are ¾ + 1 ways to form
a multiset of size ¾ , since the first type can contribute anywhere from 0 to ¾ ob-
jects. Thus the generating function is ∑¾=0(¾ + 1)x¾ . Alternatively, to reach size
∞

¾ , the first type contributes j and the second contributes ¾ − j , for some j . Hence
∞ 2
by the definition of product the generating function also equals ( ∑¾=0 x¾ ) .
Section 3.1: Ordinary Generating Functions 95

3.1.6. Example. Two types of coins. With ¾ coins on a table, each may show heads
or tails. The number of heads can be 0 through i, so there are i + 1 distinguishable
ways to have i coins. Thus the generating function for selections of coins, indexed
∞
by the number of coins, is ∑¾=0(¾ + 1)x¾ .
Now consider coins of two values, such as nickels and dimes. A selection with
¾ coins uses j nickels and ¾ − j dimes, for some j . If we also note how many of each
show heads or tails, then we have j + 1 ways to have j nickels and ¾ − j + 1 ways to
have ¾ − j dimes, and each choice for the nickels can be paired with each choice for
the dimes. Thus the generating function for distinguishable selections of nickels
¾
and dimes, indexed by the number of coins, is ∑¾≥0 [ ∑i=0(i + 1)(¾ − i + 1)] x¾ . By
∞ 2
the rule for multiplying formal power series, this is [ ∑¾=0(¾ + 1)x¾ ] .
We described this example in this way to illustrate that the coefficients in a
factor may be other than 0 or 1. In fact, we have just discussed the enumerator for
multisets from four types: head-nickels, tail-nickels, head-dimes, and tail-dimes.
∞ 4
The generating function is ( ∑¾=0 x¾ ) .

Products of generating functions model the enumeration of objects formed as

ordered pairs when the weight functions are “additive”.

3.1.7. Lemma. Let a¾ , b¾ , c¾ count the elements with weight ¾ in sets A , B , C,

respectively. The OGF for ⟨c⟩ is the product of the OGFs for ⟨a⟩ and ⟨b⟩ if
and only if, for all ¾ , the elements in C with weight ¾ correspond to the pairs
( , ) ∈ A × B such that  is the sum of the weights of and  .
¾
Proof: The correspondence specified is precisely the condition for c¾ = ∑ j =0 aj b¾− j ,
which by Definition 3.1.4 expresses the OGF for ⟨c⟩ as the product of the OGFs
for ⟨a⟩ and ⟨b⟩.

The product behavior generalizes to objects built in n stages. If objects in

C correspond to n-tuples in A1 × · · · × An , and the weight of an n-tuple is the
sum of the weights of the entries, then the OGFs satisfy C(x) = ∏i=1 Ai(x), using
n

Lemma 3.1.7 and induction on n.

3.1.8. Example. Multisets (selections with repetition). Let a¾ be the number of

multisets of size  from n types of objects, where n is fixed. Thus ∑¾≥0 a¾ x¾ is the
enumerator by size for multisets from n types of objects.
To specify a multiset, we list the multiplicity of each type of object; the size
is the sum of these multiplicities. Since the generating function for multisets
∞ ∞ n
from one type is ∑¾=0 x¾ , we obtain ∑¾≥0 a¾ x¾ = ( ∑¾=0 x¾ ) .
Each contribution to the expansion of the product is obtained by choosing a
term from each factor, independently. Within the jth factor, the different terms
represent options for the jth type in forming the compound object. The sum of
the exponents is the index (size) of the resulting compound object. This is why we
say that sum and product of generating functions model the use of the Sum and
Product Principles to solve a whole sequence of counting problems at once.
Consider red, white, and blue marbles. A multiset with three red, one white,
and two blue marbles corresponds to choosing x3 from the first factor, x from the
second, and x 2 from the third. This contributes 1 to the coefficient of x6 in the
96 Chapter 3: Generating Functions

product. Each choice of a term from each factor generates one multiset and con-
tributes 1 to the coefficient of x n , where n is the total size of the multiset.
In fact, we already know the coefficients in the generating function for mul-
tisets from n types. In Theorem 1.1.20, we proved a¾ = (¾+n−n−1 1 ). Equating the two
expressions for the OGF yields a combinatorial proof of an algebraic identity (for
each n) about formal power series:
∞
n ¾ + n−1
(1 + x + x + · · ·) = ∑ (
2
)x¾ .
n−1
¾ =0

Coefficients in factors need not always equal 1 (see Example 3.1.6).

Formal power series are specified by their coefficients; two formal power series
are equal if and only if corresponding coefficients are equal. We need convenient
notation for a specified coefficient. We repeat the definition from Chapter 2.

3.1.9. Definition. The coefficient operator [x¾ ], applied to a (formal) power

series in x, returns the coefficient of x¾ in the series. That is, if A(x) =
∞
∑n=0 an x n , then [x n] A(x) = an .
When the scope of the coefficient operator is clear, we usually do not place
parentheses around its argument.
The definition of sum (and multiplication by constants) makes the set of for-
mal power series in x an infinite-dimensional vector space. The definition of prod-
uct satisfies the additional axioms needed to make it an “algebra”. By the defini-
tion of product, the formal power series 1x0 + 0x1 + 0x 2 + · · · (or simply 1) is an
identity element for multiplication. We define A(x)−1 to be B(x) when A(x)B(x) =
1 (note that multiplication of formal power series is commutative). The condition
for A(x) having a multiplicative inverse is [x0 ] A(x)
= 0 (Exercise 6).
We obtained the power series for (1 − x)−n in Lemma 2.2.17. Here we obtain
the same expression as a formal power series expansion. We will consider why
these expressions are the same in the next section.

3.1.10. Theorem. For n ∈ , the formal power series expansion of the generating
function (1 − x)−n is
∞
¾ + n−1
(1 − x)−n = ∑ ( )x¾ .
n−1
¾ =0

∞ ∞ n
Proof: The product of (1 − x) and ∑ j =0 x j is 1, so (1 − x)n ( ∑ j =0 x j ) = 1 n = 1, and
therefore (1 − x)−n = ( ∑ j =0 x j ) . Since [x¾ ]( ∑ j =0 x j ) = (¾+n−n−1 1 ) (Example 3.1.8),
∞ n ∞ n

also [x ¾ ](1 − x)−n = (¾+n−n−1 1 ).

3.1.11. Example. Multisets with restricted multiplicities. When multiplicity is

unrestricted, the factor for each type of element in the OGF is 1 + x + x2 + · · · ,
which equals (1 − x)−1 . Restrictions on usage affect the choices listed in a factor.
When S is the set of multiplicities allowed for a type of object, the generating
function for that factor is ∑¾∈S x¾ .
Section 3.1: Ordinary Generating Functions 97

For ordinary subsets, the choice for each type is omission or inclusion (S =
{0 , 1}), and the generating function is (1 + x)n .
When each type must be used, we cannot select none. The enumerator is then
(x + x 2 + x3 + · · ·)n , which equals ( 1−x x )n . Thus

x n ¾ − n+ n−1 ¾−1
[ x¾ ] ( ) = [ x ¾−n ](1 − x)−n = ( )=( ),
1−x n−1 n−1

as in Corollary 1.1.22. The x in the numerator of each factor corresponds to re-

quiring a positive integer in the integer-sum model.
When a type must have even usage, its factor is 1 + x2 + x4 + · · · , which
equals (1 − x2)−1 . Any restriction on multiplicities can be encoded. For exam-
ple, in how many ways can we pick 20 coins that are pennies, nickels, or dimes,
with at least three nickels and at most four dimes (pennies unrestricted)? With
r +1
∑¾=0 x¾ = 1−1−x x , we compute
r

1 x3 1 − x 5 x3 − x8
[ x 20 ] = [ x20 ]
1−x1−x 1−x (1 − x)3

¾+3−1 ¾+3−1 19 14
= [ x17 ] ∑ ( )x¾ − [ x12 ] ∑ ( )x¾ = ( ) − ( ).
3−1 3−1 2 2
¾ ¾

Example 4.1.6 addresses restricted multiplicities in another way.

3.1.12. Remark. By definition, a generating function systematically solves a

sequence of problems. The automatic bookkeeping performed by modeling with
sums and products of generating functions avoids case analysis, and the expres-
sion for the full generating function is often simpler than more specific formulas.
For this reason, we do not artificially end factors at x t even when we seek only the
coefficient of x t .
∞
Also, an expression with formal power series expansion ∑¾=0 a¾ x¾ cannot de-
pend on ¾ , just as the value of ∑¾=0 º (¾) cannot depend on ¾ .
n

By “combinatorial argument ” to obtain a generating function, we mean us-

ing generating functions to model the sum (case) and product (stage) choices that
are made when building the objects being enumerated, in such a way that the
exponent on the formal variable accumulates the contributions to the index pa-
rameter. In Example 3.1.11, a multiset is formed by choosing some permitted
number of copies of each type, and the size of each resulting multiset is the sum
of the multiplicities chosen from each type. Accumulating the multiplicities in
the exponent models this additivity of size (the index parameter).

The notion of generating function extends naturally for problems with m pa-
rameters: the coefficient of the monomial x1¾1 · · · x¾mm is the number of objects in
which the value of the ith parameter is ¾i , for all i. For example, when flip-
ping nickels and dimes in Example 3.1.6, we could define a generating function
in which the coefficient of x r y s is the number of outcomes with r coins, of which
s show heads (see Exercise 16).
98 Chapter 3: Generating Functions

∞ ∞
3.1.13. Example. Let A(x , y) = ∑n=0 ∑¾=0 an ,¾ x¾ y n with an ,¾ = (¾n). Now
∞ ∞ ∞
n 1
A(x , y) = ∑ ∑ ( )x ¾ y n = ∑(1 + x)n y n = .
¾ 1 − y − xy
n=0 ¾ =0 n=0

Thus (1 − y − xy)A(x , y) = 1. For n ≥ 1, we read this as an ,¾ − an−1 ,¾ − an−1 ,¾−1 = 0,

which is the binomial recurrence (Pascal’s Formula).
This example reverses the process of Example 2.2.23; generating functions
can be used to obtain recurrence relations.

PERMUTATION STATISTICS

There are n! permutations of [n], but we can group them by the value of some
parameter. Such parameters are permutation statistics. Natural statistics for
permutations include number of fixed points, number of cycles, etc. These param-
eters have important roles in algebraic combinatorics and in the theory of special
functions, but here we study only the counting problems.
We use n for the set of permutations of [n]. This “blackboard” font is re-
served for important sets with special algebraic structure, such as  ,  ,  ,  for
number systems and  q for the finite field of order q. The permutations of [n] form
a “group” under composition (see Section 4.2), called the symmetric group (on
n elements). Its fundamental role makes a special font appropriate; we use for
“symmetric”. The historical notation is S in Fraktur font, which is hard to read.
It may be hard to count subsets of n explicitly. Instead, we seek a generat-
ing function indexed by the relevant parameter. Our first example illustrates the
combinatorial approach of finding a generating function by describing the choices
made in building the objects.

3.1.14. Definition. An inversion in ∈ n is a pair ( i , j) with i < j and i > j.

3.1.15. Proposition. The enumerator of n by number of inversions is

(1 + x)(1 + x + x2) · · · (1 + x + x2 + · · · + x n−1).

Proof: We build a permutation (viewed as an ordering of [n]) by successively

inserting 1 , . . . , n. Elements r and s with r < s form an inversion in if and
only if s is inserted leftward of r; their relative order cannot change later. The
number of inversions in which s is the higher element is the number of elements
in positions after s when s is inserted.
We have one factor in the generating function for each insertion. Inserting s
creates between 0 and s − 1 insertions. For each placement value, the choices for
the number of new inversions when inserting later values are the same.
After inserting all elements, we should contribute 1 to the coefficient of x t ,
where t is the total number of inversions. The factor (1 + x + · · · + x s−1) lists the
choices when we insert s.

Next we enumerate n by number of cycles.

Section 3.1: Ordinary Generating Functions 99

3.1.16. Definition. The 2-line form or tabular representation of a permu-

tation records the domain on top and the corresponding images below. The
word form of a permutation of [n] is simply the second row of the 2-line form,
listing the image of i in position i.
A cycle representation of a permutation is a list of its cycles, grouped
within parentheses. The canonical cycle representation puts the least el-
ement first in each cycle and puts cycles in decreasing order of least elements.

3.1.17. Example. The permutation with tabular representation (123456789 )

461752398 has
word form 461752398, and the cycles are (4731), (62), (89), and (5). The canonical
cycle representation is (89)(5)(26)(1473).
Dropping the parentheses in this representation yields 895261473, the word
form of another permutation. We show that this process defines a bijection.

3.1.18. Lemma. The canonical cycle representation º ( ) of a permutation

(without parentheses) defines a bijection from Ën (given as words) to itself.
Proof: In the canonical cycle representation, the next cycle begins when a new
left-to-right minimum is reached. By starting with the word form of a permuta-
tion  and inserting parentheses before the left-to-right minima, we obtain the
canonical cycle representation of a permutation such that  ( ) =  . Since a
surjection from a finite set to itself is a bijection,  is thus a bijection.

We used this idea in the enumeration of trees in Theorem 1.3.4. Knuth [1968]
described the bijection as “An unusual correspondence”; it also appeared in Rényi
[1962] and Foata–Schützenberger [1970]. It yields short proofs for many enumer-
ative and probabilistic problems; see Exercises 23–31. We give one example.

3.1.19. Proposition. When permutations of [n] are equally likely, the distribu-
tion of the length of the cycle containing a fixed element  is uniform, with
probability 1/n for each length.
Proof: By symmetry, it suffices to prove the claim when  = 1. The length of
the cycle containing 1 in a permutation  is n + 1 minus the position of 1 in the
canonical cycle representation of  . By Lemma 3.1.18, this position is uniformly
distributed over all n positions.

Next we use canonical cycle representation to enumerate Ën by number of

cycles. Note that setting x = 1 combines the n! permutations into one class.

3.1.20. Theorem. Let c(n , ) be the number of elements of Ën with  cycles. The
enumerator of Ën by number of cycles, Cn(x), is given by
n
Cn(x) = ∑ c(n , )x¾ = x(n) .
¾ =1

Proof: By definition, each permutation with  cycles contributes 1 to [x ¾ ]Cn(x).

Recall that x(n) = ∏i=1 (x + i − 1). To show that Cn(x) = x(n) , we interpret (x + i − 1)
n

as encoding the options when inserting element i to build the canonical cycle rep-
resentation of a permutation. We insert elements in increasing order as in Propo-
sition 3.1.15, but now the choices have different effects on the index parameter.
100 Chapter 3: Generating Functions

Again the factor for inserting i is independent of the choices made for earlier
elements, so the generating function is the product of these factors.
Choosing “ x” from (x + i − 1) means starting a new (leftmost) cycle in the
representation with i as least element. This increases the number of cycles, so
the objects built with this choice contribute to a term with higher exponent. The
number i − 1 counts the ways to make i follow a smaller number j in its cycle. For
example, the permutation of Example 3.1.17 is built as below. Here (i , x) indicates
an increase in the number of cycles by inserting i as the left, while (i , j) indicates
the insertion of i following the smaller number j .
(1 , x): (1) (4 , 1): (2)(143) (7 , 4): (5)(26)(1473)
(2 , x): (2)(1) (5 , x): (5)(2)(143) (8 , x): (8)(5)(26)(1473)
(3 , 1): (2)(13) (6 , 2): (5)(26)(143) (9 , 8): (89)(5)(26)(1473)

Since i is largest when inserted, it does not change left-to-right minima un-
less it is put first. When expanding the product, the exponent on x in a term is
the number of times the choice started a new cycle.
Each permutation is counted exactly once and with the correct exponent, be-
cause there is exactly one way to reconstruct the choices from the canonical cy-
cle representation of a permutation. To reconstruct the situation when i was
inserted, we discard all elements of the canonical cycle representation that are
larger than i; their insertion does not affect the cycle structure on the first i num-
bers. Now i is first if and only if the choice was made to start a cycle with i, and
otherwise i immediately follows the element it was chosen to follow.

Theorem 3.1.20 can be interpreted as a polynomial identity. Another tech-

nique for polynomial identities, different from building generating functions, is
the Polynomial Principle (Remark 1.1.10). To illustrate the difference, we reprove
Theorem 3.1.20 using the Polynomial Principle.

3.1.21. Remark. Enumerating Ën by cycles, again. (Stanton–White [1986, p. 79])

By the Polynomial Principle, it suffices to show for x ∈ 
that both sides of
∑¾=1 c(n , ¾)x¾ = x(n) count the same set. The right side counts placements of n
n

distinct flags on x flagpoles in a row; order on flagpoles matters (Exercise 1.1.35).

To show that the left side also counts these arrangements, begin with the cy-
cle representation of a permutation of [n]. For each cycle, assign the elements
to one of the x flagpoles. A permutation with ¾ cycles has x¾ such assignments.
Put all the flags in all the cycles assigned to a given pole on that pole, in the or-
der given by the canonical cycle representation. That is, each cycle has its least
element first, and the cycles are entered in decreasing order of least elements.
For each arrangement of flags, we retrieve the cycles (and the assignment of
cycles to poles). On each pole, a new cycle begins with each new minimum among
the flags listed so far. Hence the pairs consisting of a permutation of [n] with ¾
cycles and an x-ary word of length ¾ correspond bijectively to the arrangements
of n flags on x flagpoles.

Next we introduce another important permutation statistic and the corre-

sponding 2-parameter family of counting numbers.
Section 3.1: Ordinary Generating Functions 101

3.1.22. Definition. For ∈ Ën , a run is a maximal increasing string in the word

form. There is a descent or ascent at i if i > i+1 or i < i+1 , respectively.
The descent set D( ) is the subset of [n − 1] at which descents occur. The Eu-
lerian number A(n , ) is the number of permutations of [n] having  runs
(and  − 1 descents).

3.1.23. Example. The permutation 791368452 has four runs (79, 1368, 45, and
2), three descents (at 2, 6, and 8), and five ascents. For ∈ Ën , the ascent set of
is [n − 1] − D( ). When has  runs, it has  − 1 descents and n −  ascents, and
its reverse has  − 1 ascents and n −  descents.

The Eulerian numbers are the subject of the first chapter of Bóna [2004],
and there is an entire book about them (Petersen [2015]). Relevant surveys in-
clude Carlitz [1958–59], Foata–Schützenberger [1970], Knuth [1973], and Char-
alambides [2002]. The most common notation now is A(n , ), though An ,¾ is also
used. Knuth [1968] used ⟨ ¾n ⟩ for A(n , ), but this notation meant A(n ,  + 1) in
Graham–Knuth–Patashnik [1989], letting  count descents rather than runs.
The Eulerian numbers are studied in Exercises 34–43. Many of the exercises
use a natural recurrence.

3.1.24. Proposition. With A(0 , ) = 0 ,¾ , for n ,  ∈  the Eulerian numbers

satisfy A(n , ) =  A(n − 1 , ) + (n −  + 1)A(n − 1 ,  − 1).
Proof: When inserting n into a permutation of [n − 1], the number of runs is
unchanged if n goes at the end of a run; otherwise, it increases by 1. Hence per-
mutations with  runs arise by inserting n into permutations of [n − 1] with 
runs in  ways and by inserting n into permutations of [n − 1] with  − 1 runs in
(n − 1) − ( − 1) + 1 ways. Since every permutation arises by successive insertions
in a unique way, this counts each permutation of [n] with  runs exactly once.

We obtain a formula for A(n , ) from the next result by using an inversion
relationship for generating functions.

3.1.25. Theorem. (Worpitzky’s Identity; Worpitzky [1883]) For n ≥ 1, the

Eulerian numbers A(n , ) satisfy
n
x−+n
x = ∑ A(n , )(
n
).
n
¾ =1


Proof: (Gessel) When x ∈ , the left side counts functions from [n] to [x]. To turn
these into permutations, express them as distributions of [n] into x successive
bins, with each bin ended by a vertical bar. Write the elements of each bin in
increasing order. This produces a permutation of [n] with x bars inserted. There
is a bar at the end of each run and possibly in additional locations (empty bins
produce consecutive bars). Below we show two ways of distributing [9] into 6 bins
so that the resulting permutation is 791368452.
79 | 136 | 8 || 45 | 2 | and | 79 | 13 | 68 | 45 | 2 |
102 Chapter 3: Generating Functions

We have represented the x n distributions as barred permutations of [n]

with x bars; such an object consists of the word form of a permutation with x
bars inserted so that there is at least one bar immediately following every run
(including the last). We now count them by grouping them according to the num-
ber of runs in the permutation.
If the permutation has ¾ runs, then ¾ of the bars must appear at the ends
of the runs. (Note that ( x−¾n+n) = 0 if ¾ > x.) There remain x − ¾ bars to put in
n + 1 possible locations, with repetition allowed. Hence there are ( x−¾n+n) ways to
choose positions for the extra bars. (Bars are identical, so it doesn’t matter which
¾ bars were placed initially.)
Thus the identity counts barred permutations of [n] with x bars in two ways.
Since the argument is valid for every positive integer x, the result is an identity
in polynomials, by the Polynomial Principle.

Barred permutations may have appeared first in the thesis of Gessel [1977];
see also Gessel–Stanley [1978]. Bóna [2004, p. 6] gives another combinatorial
proof of Worpitzky ’s Identity. The identity is sometimes used to define the Eule-
rian numbers. It expresses the power x n in terms of the Eulerian numbers; our
goal now is to invert this relationship to obtain a formula for A(n , ¾).

3.1.26. Theorem. The Eulerian number A(n , ¾) (the number of permutations of

[n] with ¾ runs) is given by
¾
n+1
A(n , ¾) = ∑(−1)i ( )(¾ − i)n .
i
i=0

Proof: Change x to r in Worpitzky ’s Identity to treat it as a statement about

positive integers. Note that A(n , ¾) = 0 when ¾ = 0 or ¾ > n, and recall that every
run has a bar; that is, (r−¾n+n) = 0 when ¾ > r. Thus we change the summation
limits without changing the sum to obtain
r
r−¾+n
r n = ∑ A(n , ¾)( ).
n
¾ =0

Let A(x) = ∑¾=0 A(n , ¾)x¾ and B(x) = ∑¾≥0 (¾+n n)x¾ ; note that n is fixed. Wor-
n

pitzky ’s Identity says C(x) = A(x)B(x), where C(x) = ∑r≥0 r n xr . Thus A(x) =
1
B(x)
C(x). Since B(x) = (1− 1x)n+1 , we have A(x) = (1 − x)n+1 C(x). By the product rule,
¾
A(n , ¾) = [x¾ ] A(x) = ∑ i=0(−1)i (n+i 1)(¾ − i)n .

Many inversion formulas can be proved by substituting the original relation

into the desired expression and showing that unwanted contributions vanish.
This may require various identities and be rather painful! It also requires know-
ing the formula for the inverse in advance. Exercise 34 proves Theorem 3.1.26
this way. The proof by inverting Worpitzky ’s Formula in Theorem 3.1.26 is easy
and does not require knowing the formula for A(n , ¾).
The generating function ∑¾=0 A(n , ¾)x¾ is the nth Eulerian polynomial,
n

An(x). The proof above used An(x) = (1 − x)n+1 ∑¾≥0 ¾ n x¾ . Euler defined An(x) this
way. Curiously, this expression for a polynomial involves an infinite sum!
Exercises for Section 3.1 103

EXERCISES 3.1

3.1.1. (−) Use generating functions to count the distinguishable selections of six marbles
from a pile consisting of three red marbles, four white marbles, and five blue marbles.
Verify the answer by describing the selections explicitly.
3.1.2. (−) Let r1 , . . . , rn and s1 , . . . , sn be natural numbers with ri ≤ si for all i. Let a¾ be
the number of multisets of size ¾ from n types of objects such that the number ei of objects
of the ith type satisfies ri ≤ ei ≤ si . Express the generating function for ⟨a⟩ as (1 − x)− n
times a product of polynomial factors.
3.1.3. (−) A child wants to buy candy. Four types of candy have prices two cents, one cent ,
two cents, and five cents per piece, respectively. Build the enumerator by total cost for the
number of ways to buy candy.
3.1.4. (−) Let an be the number of nonnegative integer solutions to 4e1 + 2e2 + e3 + 2e4 = n.
Find the generating function for ⟨a⟩.
3.1.5. (−) Let an ,¾ be the number of permutations of [n] having ¾ inversions (Proposition
3.1.15). Obtain a recurrence relation for these numbers.
3.1.6. (−) From Worpitzky’s Identity (Theorem 3.1.25), prove
n
x+¾−1
x n = ∑ A(n , ¾)( ).
n
¾ =0

3.1.7. (−) Compute the Eulerian number A(4 , ¾) for all ¾ from the definition and from
Theorem 3.1.26. Compute the congruence class of A(p − 1 , ¾) modulo p when p is prime.
3.1.8. (♦) Let bn ,¾ be the number of ¾-subsets of [n] with no consecutive integers. Let
B¾(x) = ∑n≥0 bn ,¾ x n . Build B¾(x) using combinatorial arguments. Use known generating
function expansions to obtain a formula for bn ,¾ .
3.1.9. Sicherman dice. When rolling dice, on any die each face is equally likely to be the
outcome. Two 4-sided dice labeled (1 , 2 , 2 , 3) and (1 , 3 , 3 , 5) have the same distribution of
sums as two normal 4-sided dice labeled (1 , 2 , 3 , 4).
(a) Use generating functions to prove that no other pair of 4-sided dice in positive in-
tegers has the same sum distribution as the normal dice.
(b) (+) Determine the pairs of 6-sided dice with positive integer labels that have the
same sum distribution as two normal 6-sided dice. (Broline [1979], Gallian–Rusin [1979])
3.1.10. (♦) A coin is flipped 14 times, and three flips are tails. Use a generating function
to determine the probability that no five consecutive flips are heads.
3.1.11. In Example 3.1.6, the enumerator for distinguishable flippings of pennies, nickels,
∞
and dimes by number of coins is ( ∑¾=0(¾ + 1)x¾) . Give both an algebraic argument and a
3

direct combinatorial argument to show that the generating function equals (1 − x)−6 .
3.1.12. A meeting has five delegates from each of 100 countries. How many ways are there
to form a committee of 25 delegates using at most one person from each country? What is
the enumerator for committees with at most three people from each country, indexed by
committee size? (Note: People are distinguishable.)
3.1.13. Ten pairs of socks are washed; the two socks in a pair are indistinguishable. Some
are lost; use a generating function to count the distinguishable ways that exactly eight
socks survive. Explain why these outcomes are not equally likely when all socks are equally
likely to survive. (Comment: Extracting the desired coefficient from the generating func-
tion requires computation.)
104 Chapter 3: Generating Functions


3.1.14. (♦) Let an be the number of ways to choose r ∈ 0 , roll one six-sided die r times,
and obtain outcomes with sum n. Express the generating function for ⟨a⟩ as the ratio of
two polynomials that each have at most three terms. Obtain a recurrence for ⟨a⟩ from the
generating function. Give a direct combinatorial argument for the recurrence.
3.1.15. (♦) Given an alphabet consisting of a vowels and c consonants, let bn be the number
of words of length n not having consonants in consecutive positions. Obtain the generating
function B(x) = ∑ bn x n as the ratio of two polynomials.
3.1.16. Given t types of coins, let an ,¾ be the number of distinguishable ways to have a
multiset of n coins on a table with ¾ showing heads. Find a closed-form expression (without
summations) for the generating function ∑ n ,¾ an ,¾ x n y ¾ .
3.1.17. Let an ,¾ be the number of ways to tile a 3-by-n rectangle using unit squares and
¾ copies of L, the 2-by-2 square missing the top right 1-by-1 corner (no rotation allowed).
Build the generating function ∑n ,¾≥0 an ,¾ x n y ¾ .

3.1.18. Let am ,n ,¾ be the number of lattice paths from (0 , 0) to (m , n) using steps in

{(1 , 0) , (0 , 1) , (1 , 1)} such that ¾ diagonal steps are taken. Build the generating function
∑m ,n ,¾≥0 am ,n ,¾ xm y n ¾ .
3.1.19. (♦) Let F̂(x) = ∑n≥0 F̂ n x n , where ⟨ F̂⟩ is the adjusted Fibonacci number.
(a) Obtain F̂(x) from the Fibonacci recurrence.
(b) Obtain F̂(x) by building it combinatorially (without the recurrence!), using the
model that F̂ n is the number of 1,2-lists with sum n.
(c) Expand the generating function to prove F̂ n = ∑¾=0 ( n−¾ ¾) .
n

3.1.20. A -colored graph is a graph with its vertices partitioned into nonempty sets
V1 , . . . , V¾ such that adjacent vertices lie in different sets. Prove that the enumerator of
-colored graphs with vertex set [n], by number of edges, is

− ∑ n2i ) ,
∑ (n1 , . . . , n¾ )(1 + x) 2 (n
1 n 1 2

!

where the sum is over choices of nonzero n1 , . . . , n¾ with sum n.

3.1.21. Let an ,¾ be the number of directed graphs with vertex set [n] that have exactly 
sources (vertices of indegree 0) and no directed cycles. Note an ,n = 1. For n >  , prove
n−¾
n
an ,¾ = ( ) ∑ an−¾ , j (2 ¾ − 1) j 2 ¾(n−¾− j) .

j =1

Explain how to obtain from this the generating function indexed by number of edges.
3.1.22. Given a -set S ⊆ [n], let  (S) be the permutation consisting of S in increasing
order followed by [n] − S in increasing order. Letting the -sets be equally likely, determine
the expected number of inversions in  (S). (Comment: Also the variance is exactly (n + 1)/6
times the expectation, for each  .) (Deshpande–Laghate [2003])
3.1.23. (♦) Compute the expected number of cycles in a random permutation of [n],
(a) by using the generating function in Theorem 3.1.20, and
(b) by using Proposition 3.1.19 and linearity of expectation (the expectation of a sum
is the sum of the expectations). (Lovász [1979, p. 25])
3.1.24. (♦) Compute the probability that a random permutation of [n] has exactly two
cycles, in two ways:
(a) by using the generating function in Theorem 3.1.20, and
(b) by considering random canonical cycle representations (see Lemma 3.1.18).
Exercises for Section 3.1 105

3.1.25. Given S ⊆ [n], compute the probability that all elements of S are in the same cycle
in a random permutation of [n].

3.1.26. Let m0 = n. Starting with i = 0, iteratively choose xi ∈ [mi ] at random, subtract

xi from mi to obtain mi+1 , and continue (stopping if mi+1 = 0). Let Y¾ be the number of
times that ¾ is the randomly chosen integer. Prove that the distribution of the random vec-
tor (Y1 , . . . , Yn) is the same as the distribution of the random vector (X 1 , . . . , X n), where
X ¾ is the number of ¾-cycles in a random permutation of [n]. (Bach [1994])

3.1.27. (♦) For n ≥ 1, view Ën as the set of permutations of [n] in word form. Let A n =
{ ∈ Ën : i
= i + 1 for 1 ≤ i ≤ n − 1}. Let Bn = { ∈ Ën :  j +1
=  j + 1 for 1 ≤ j ≤ n − 1}. Use
the canonical cycle representation to prove bijectively that | A n| = | Bn |. (Comment: The
sizes of A n and Bn were computed in Exercise 2.1.44.)

3.1.28. (♦) The keys to n boxes are put randomly in the boxes, one per box. The boxes are
locked by closing them (a box may contain its own key). Find the probability that breaking
open  random boxes will allow unlocking the remaining boxes. (Hint: Consider canonical
cycle representations.) (Bognár–Mogyoródi–Prékopa–Rényi–Szász [1970, p. 56])

3.1.29. (♦) The reverse canonical representation of a permutation writes each cycle
with largest element first and puts the cycles in increasing order of those first elements.
Let n : Ën → Ën take each ∈ Ën to the permutation obtained by dropping the parentheses
in this representation of . Characterize and count the fixed points of n . (Deutsch [2008])

3.1.30. Put elements of Ë2n into E2n or O2n if their cycle lengths are all even or all odd,
respectively (| E4 | = |O4 | = 9). Prove | E2n | = |O2n| , bijectively. (Schmidt [1994])

3.1.31. (♦) Let Cn(x) be the enumerator of Ën by number of cycles, with c(n , ) = [x¾ ]Cn(x).
Theorem 3.1.20 proves combinatorially that Cn(x) = x(n) .
¾
(a) Use canonical cycle representations to prove c(n + 1 , m + 1) = ∑¾= m c(n , )( m
n
).
(b) Use part (a) to prove Theorem 3.1.20 by induction.
(c) By Example 2.1.9, c(n , ) = (n − 1)c(n − 1 , ) + c(n − 1 ,  − 1) for n ,  ≥ 1. Use this
and the generating function method to prove Theorem 3.1.20 again.

3.1.32. (♦) Create an array with row 0 being repeated copies of 1. To obtain row j + 1 from
row j , cross out the first element of row j , then the second subsequent entry, then the third
after that , and so on. Take partial sums of the remaining entries to form row j + 1, with
blanks under the deleted entries. Prove that the first entry in row j is j !. (Hint: The di-
agonal of the jth wedge sums to j!. Explain those values using Exercise 3.1.31. Comment:
Guy [1988] calls this an instance of Moessner ’s Process; see Exercise 2.1.38.)

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 2 3 4 5 6 7 8 9 10
2 6 11 18 26 35
6 24 50
24

3.1.33. Let F n() = ∑¾≥0 c( , n)¾/ !, where c( , n) is the number of permutations of []
with n cycles. Prove F n() = [− ln(1 − )]n/n!.

The remaining problems in this section involve the Eulerian numbers A(n , ), where
A(n , ) denotes the number of permutations of [n] with  runs.

3.1.34. Substitute the expression for ( − r)n given by using Worpitzky’s Identity with
¾
x =  − r into the sum ∑r=0(−1)r (n+r 1 )( − r)n to prove that the sum equals A(n , ).
106 Chapter 3: Generating Functions

¾
3.1.35. Prove A(n , ¾) = ∑r=0(−1)r (n+r 1 )(¾ − r)n from Proposition 3.1.24.

3.1.36. Eulerian numbers. There are A(n , ¾) permutations of [n] with ¾ runs.
(a) Show that ∑¾=1 A(n , ¾) = n! and A(n , ¾) = A(n , n + 1 − ¾).
n

(b) Prove ∑¾=1 ¾ A(n , ¾) = 12 (n + 1)! using the identities of part (a).
n

(c) Prove ∑¾=1 ¾ A(n , ¾) = 12 (n + 1)! by counting two ways.

n

(Comment: Thus the expected number of runs in a random permutation is n+2 1 .)


3.1.37. (♦) There are n! lists (b1 , . . . , bn) in n such that bi ≤ i for all i. Let B(n , ¾) be the
number that omit ¾ − 1 elements of [n]. Prove B(n , ¾) = A(n , ¾). (Bóna [2004, p. 31])

3.1.38. (♦) For n ∈  , prove that ∑¾≥0 ¾ n/2¾ is an integer. (Hint: Consider the proof of
Theorem 3.1.26.)

3.1.39. (♦) Two games begin with players A and B each having $1. On each play, A or
B wins $1. When A has $a and B has $b, the probability that A gets the next dollar is
a/(a + b) in Game 1, but in Game 2 it is b/(a + b). For 1 ≤ ¾ ≤ n, determine in each game
the probability that A has $ ¾ when the total is $(n + 1). (Comment: Game 1 has been
called P ólya’s Urn; see Johnson–Kotz [1977].)

3.1.40. (♦) An increasing tree is rooted with integer vertices, increasing along paths
from the root. Let T n be the set of increasing trees with vertices {0 , . . . , n}. For ∈ Ën ,
form by putting 0 before the word form of . For i ∈ [n], create the edge ji, where j is the
last element before i in that is less than i. Let  ( ) be the resulting graph on {0 , . . . , n}.
(a) Prove that  is a bijection from the set of permutations of [n] to T n .
(b) Prove that T n has c(n , ) elements whose root has degree  , where c(n , ) counts
the permutations of [n] with  cycles. (Hint: Consider canonical cycle representations.)
(c) Prove that T n has A(n , ) elements with  (non-root) leaves.
(Comment: Elements of T n correspond to inclusion arrangements of n square envelopes
with distinct sizes. There are n! arrangements, of which c(n , ) have  envelopes contained
in no others and A(n , ) have  envelopes containing no others.)

3.1.41. (♦) For ∈ Ën , there is an excedance at i if (i) > i. For example, the excedances
in 791368452 are at {1 , 2 , 5 , 6}. Given , let ˆ denote the permutation whose word form
is obtained by dropping the parentheses in the canonical cycle representation of . Prove
that has an excedance at i if and only if ˆ has an ascent at i. Thus the number of permu-
tations of [n] with  ascents is the number of permutations of [n] with  excedances, and
A(n , ) permutations of [n] have n −  excedances. (Foata–Schützenberger [1970])

3.1.42. (♦) For ∈ Ën , there is a weak excedance at i if (i) ≥ i. Use Exercise 3.1.41 to
prove that the number of permutations of [n] having  weak excedances is A(n , ). (Hint:
Compare with its reverse complement ∗ , defined by ∗(i) = n + 1 − (n + 1 − i).) (Bóna
[2004, p. 32])

3.1.43. (+) Smith College Diploma Problem. By tradition, the graduating students form a
circle and diplomas are distributed at random. Students having their own diplomas leave;
each remaining student passes the diploma she holds to the student on her right to com-
plete the round. Maurer [1973] asked for the probability that the process takes  rounds.
Don West [1974] used recurrence and induction; Gessel [2001] gave a bijective solution.
(a) For ∈ Ën , let s1 , . . . , sm be the non-fixed elements in increasing order. Let (si) =
(si+1 ) if i < m and (sm) = (s1 ). The fixed points of are also fixed points of . Prove
that if has  excedances (Exercise 3.1.41), then has  − 1 excedances. For example,
32514 → 52143 → 12345.
(b) Use part (a) and Exercise 3.1.41 to prove that the number of permutations of the
n diplomas that result in  rounds for sorting is A(n , ).
Section 3.2: Coefficients and Applications 107

3.2. Coefficients and Applications

Formal power series expansions of generating functions can be manipulated
much like ordinary power series in calculus, with the additional freedom of ig-
noring issues of convergence. (Niven [1969] presents a readable introduction to
the theory.) When a function has a power series expansion at 0, that is also its
formal power series expansion. In Lemma 2.2.17 we expanded (1 − x)−n using Tay-
lor ’s Theorem. In Theorem 3.1.10, the same expression emerged as the formal
power series expansion of (1 − x)−n using the combinatorial interpretation of for-
mal power series products. This is no coincidence.
Equality holds because operations on formal power series agree with corre-
sponding operations on power series that converge at numerical values. Since this
holds for sum and product, statements like A(x)m A(x)n = A(x)m+n have the same
proofs for formal power series as for power series.
To extend the correspondence, we define differentiation and “evaluation” for
formal power series to agree with these operations on power series. In calculus,
Taylor series are termwise differentiable, using uniform convergence of series of
functions. For formal power series, this is the definition of differentiation. We
will skip formal verifications, focusing instead on the use of generating functions.

3.2.1. Definition. When A(x) is the generating function with expansion A(x) =
∞
∑n=0 an x n , evaluation at 0∞ is defined by setting A(0) = a0 . The ∞derivative
of a formal power series ∑ n=0 an x n is the formal power series ∑n=1 nan x n−1 ;
we write A (x) for the derivative of a formal power series A(x) in x.

Since the termwise behavior holds by definition, the sum and product rules
for differentiation of formal power series are easy to verify (Exercise 30). Gen-
erally we have the following statement, which motivates writing A (x) for the
derivative of a formal power series A(x):
The derivative of the formal power series expansion of A(x)
is the formal power series expansion of the derivative of A(x).
The occurrence of power series as formal power series is another excuse for our
(ab)use of the word “function” in the term generating function. Some formal power
series converge in a neighborhood of the origin when viewed as power series, but
some do not. Evaluation of A(x) at numerical values for x (such as x = 1) is allowed
when the sum converges (such as when A(x) is a polynomial in x).

OPERATIONS AND SUMMATIONS

Since the formal power series expansion of (1 − x)−1 is also the power series for
(1 − x)−1 , the correspondence between operations allows us to derive our favorite
formal power series expansion by differentiation.
108 Chapter 3: Generating Functions

3.2.2. Example. (1 − x)−n = ∑¾≥0 (¾+n−n−1 1 )x¾ by formal differentiation. In contrast

to Example 3.1.8, we differentiate (n − 1 times) both sides of (1 − x)−1 = ∑¾≥0 x¾ .
(n−1)! ¾ − n+1 . Now divide by (n − 1)! and
− x)n = ∑¾≥ n−1 ¾ ¾ − 1) · · · ¾ − n + 2)x
This yields (1 ( (
“shift the index” by n − 1.

3.2.3. Remark. Shifting the index. The negative binomial expansion often arises
as ∑¾≥0 (¾r )x¾ . We find the generating function by
¾ ¾ ¾+r xr
∑ ( r )x¾ = ∑ ( r )x¾ = ∑ ( )x ¾+r = .
r (1 − x)r+1
¾ ≥0 ¾ ≥r ¾ ≥0

Here we discard terms that are 0, then shift the index, then use the known ex-
pansion of (1 − x)−n . Lowering the starting value of the index (by r) is done by
adding r to each appearance of the index in the summand. We often do this to
introduce or remove a power of the formal variable.

3.2.4. Remark. Extracting coefficients. We often study the generating function

A(x) to seek a formula for an . In principle, we can differentiate n times, divide by
n!, and evaluate at 0, since an = n!1 (n)
A (0). In practice, this approach is useless.
Usually we obtain coefficients from a known series, such as (1 + x)n , (1 − x)−n ,
1− x n+1 x
1− x , or e . We operate on the function and expansion simultaneously, maintain-
ing equality. This works with sums, products, derivatives, and partial fractions.
The other operations below are easy but useful manipulations. Property (3) was
used in Remark 3.2.3. Property (4) is a common application for differentiation,
where we also shift the index to restore equality of exponent and subscript.

3.2.5. Proposition. If A , B , C are the ordinary generating functions for ⟨a⟩,

⟨b⟩, and ⟨c⟩, respectively, then
(1) cn = an + bn for all n ⇔ C(x) = A(x) + B(x).
cn = ∑i=0 ai bn−i for all n ⇔ C(x) = A(x)B(x).
n
(2)
an−r for n ≥ r
(3) bn = { ⇔ B(x) = xr A(x).
0 for n < r
(4) bn = nan ⇔ B(x) = x A (x).
A(x)
cn = ∑i=0 ai for all n ⇔ C(x) =
n
(5) (special case of (2))
1−x
a for n even
(6a) bn = { n ⇔ B(x) = 12 [A(x) + A(− x)].
0 for n odd
an for n odd
(6b) bn = { ⇔ B(x) = 12 [A(x) − A(− x)].
0 for n even
an/m when m | n
(7) bn = { ⇔ B(x) = A(x m).
0 when m  n

Proof: (1,2) These are the definitions of arithmetic on formal power series (Defi-
nition 3.1.4). Convolution gives the coefficient of x n in a product of series.
(3) Multiplying by powers of the formal variable has the effect of shifting in-
dices in the sequence of coefficients.
(4) Differentiation of A(x) introduces a linear factor (Definition 3.2.1).
Section 3.2: Coefficients and Applications 109

(5) Multiplying A(x) by (1 − x)−1 produces a new generating function where

the coefficient of x n is the sum of the coefficients up to x n in A(x).
(6) Odd or even terms in the expansion of A(x) can be canceled out by forming
the sum [A(x) + A(− x)]/2 or the sum [A(x) − A(− x)]/2.
(7) Substituting y = x m spreads out the terms.

These manipulations are useful in several ways. We can start with the expan-
sion of a known generating function and manipulate it to obtain a desired sum,
operating simultaneously on the generating function. We may also recognize the
sum as a coefficient in a generating function. We call these algebraic proofs for
evaluating summations. As with recurrence relations, discovering the value of
the sum may suggest a short combinatorial proof that was not initially apparent.

3.2.6. Example. Summing the coefficients. The terms of a sequence can be

summed (if the sum converges) by setting the argument of the generating func-
tion to 1. For example, ∑ ¾ (¾n) arises from the binomial expansion ∑ (¾n)x¾ by
differentiating and then setting x = 1. Thus
n d ###
∑ ¾(¾) = dx (1 + x)n##### x=1 = n2n−1 .
¾ ≥0 #
Combinatorially, both sides count chaired committees from n people.

3.2.7. Example. Even coefficients. To sum the even coefficients in the expansion
of (1 + x)n , Proposition 3.2.5(6) first yields
n 1
∑ (2i)x2i = 2
[(1 + x)n + (1 − x)n].
i≥0

Setting x = 1 yields 2 n−1 when n > 0. When n = 0, both (1 + x)n and (1 − x)n equal
1, and setting x = 1 yields 1, not 1/2.
Combinatorially, the sum counts the even subsets of an n-set. The number
is 2 n−1 , since for each subset of [n − 1], including or omitting n yields one even
subset and one odd subset (when n > 0).

3.2.8. Example. Even minus Odd. Setting x = −1 in A(x) gives the sum of the
even coefficients minus the sum of the odd coefficients. If A(x) enumerates a set
having m0 objects with even index and m1 with odd index, then A(−1) = m0 − m1 .
For example, (1 + x)n enumerates subsets of [n] by size. Setting x = 1 yields
all 2 n subsets, and setting x = −1 proves equality between the numbers of subsets
with even size and odd size (if n > 0).
n−1 j
The generating function ∏ j =0 ∑i=0 x i enumerates permutations of [n] by
number of inversions (Proposition 3.1.15). Setting x = 1 yields all n! permuta-
tions, and setting x = −1 proves equality between the numbers of permutations
with even and odd numbers of inversions (if n > 1).
The generating function ∏ j =1 (x + j − 1) enumerates permutations of [n] by
n

number of cycles (Theorem 3.1.20). Setting x = 1 yields all n! permutations, and

setting x = −1 proves equality between the numbers of permutations with even
and odd numbers of cycles (if n > 1).
110 Chapter 3: Generating Functions

3.2.9. Example. Summation of polynomials, revisited. We saw in Applica-

tion 1.2.7 that every polynomial p of degree d in ¾ is a linear combination of
(d¾ ) , . . . , (¾0). This and ∑¾=0 (¾j ) = (nj ++11) allowed us to compute ∑¾=0 p(¾).
n n

Here instead we use Proposition 3.2.5(3–5). In ∑¾≥0 ¾ m x¾ , successive factors

of ¾ are introduced by differentiating and then multiplying by x to restore the
exponent. Thus ∑¾≥0 x¾ = 1−1 x and ∑¾≥0 ¾ x¾ = (1−xx)2 and ∑¾≥0 ¾ 2 x¾ = (1x+ x2
− x)3 . By
∑¾=0 ¾ 2 = [x n] (1x+− xx)4 . Summing [x n−1 ](1 − x)−4 and [x n−2](1 − x)−4 yields
n 2
(5),
n
n−1 +3 n− 2+3 (n + 1)n(2n + 1)
∑ ¾2 = ( 3
)+(
3
)=
6
.
¾ =0

Exercise 13 suggests another generating function approach.

Evaluation of sums by convolutions (Proposition 3.2.5(2)) deserves special at-

tention. We use the convenient abbreviation “OGF” for “ordinary” generating
function (in the next section we study a different type of generating function).
n
The value of ∑i=0 ai bn−i is the coefficient of x n in the product of the OGFs for ⟨a⟩
and ⟨b⟩. The coefficient operator conveniently encodes this phrase in notation.
Theorem 1.2.3 gave combinatorial arguments for the first two examples below.
r
3.2.10. Example. The Vandermonde Convolution ∑¾=0 (m n
¾ )(r −¾ ). Recognizing the
sum as a convolution with a¾ = (m
¾ ) and b¾ = (¾ ), we have
n

r
m n m n
∑ ( ¾ )(r − ¾) = [ xr ] ∑ ( ¾ )x¾ ∑ (¾ )x¾
¾ =0 ¾ ¾
m+ n
= [ x ](1 + x) (1 + x) = [ x ](1 + x)m+n = (
r m n r
).
r

3.2.11. Example. The convolution ∑¾=0 ¾(n − ¾). Let a¾ = b¾ = ¾ and cn =

n

∑¾=0 a¾ bn−¾ . The generating function ∑¾≥0 ¾ x¾ is x/(1 − x)2 (see Example 3.2.9).
n

We obtain cn = (n+3 1), because cn is the coefficient of x n in the computation below.

∞
x 2
x2 ¾ + 4 − 1 ¾ +2
∑ ¾ x¾ ∑ ¾ x¾ = ( (1 − x)2
) =
(1 − x)4 ∑
= (
4−1
)x .
¾ ≥0 ¾ ≥0 ¾ =0

3.2.12.* Example. Evaluation of ∑¾≥0(−1)¾ (n−¾ ¾ )(nm−−2¾¾) for m , n ∈ 0 . With so 

many uses of ¾ , it is unclear how to write this as a simple convolution. However,
when m + ¾ ≤ n, the Subcommittee Identity yields (n−¾ ¾)(nm−−2¾¾ ) = (nm
−¾ m
)( ¾ ), which
m n−¾ ¾
rewrites the sum as ∑¾ (−1) ( ¾ )( m ). Let a¾ = (−1) ( ¾ ) and b¾ = (m). The OGF
¾ ¾ m

for ⟨b⟩ from Remark 3.2.3 yields

n− ¾ n − 2¾ m n− ¾
∑(−1)¾ ( ¾
)(
m− ¾
) = ∑(−1)¾ ( )(
¾ m
) = ∑ a¾ bn−¾
¾ ≥0 ¾ ≥0 ¾ ≥0

if n ≥ m .
m
x
= [x n] (1 − x)m = {1
(1 − x)m+1 0 if n < m
Section 3.2: Coefficients and Applications 111

Next we apply generating functions to convolutions in a reverse way. Instead

of finding the unknown OGF for the convolution of sequences with known OGFs,
we use the known OGF for the convolution to find the OGF for a factor.

3.2.13. Example. Central binomial coefficients: ∑n≥0 (2n n

)x¾ = (1 − 4x)−1/2 .
Let A(x) be this generating function. The convolution ∑¾=0 (2¾¾ )(2nn− 2¾
n
−¾ ) equals
[x ] A(x) . Knowing the coefficients in A(x) is equivalent to knowing the sum.
n 2

In Theorem 1.3.15, combinatorial argument gave ∑¾=0 (2¾¾ )(2nn− 2¾

−¾ ) = 4 . Thus
n n

A(x)2 = (1 − 4x)−1 , so ∑ n≥0 (2n n )x =

n (1 − 4x)−1/2 .
If we did not know the sum, we could still find it and A(x) by using the Cata-
lan generating function. Termwise differentation yields A(x) = dx d

¾
√xC(x), where
C(x) = ∑¾≥0 ¾+1 ¾
1
( 2
)x ¾ . In Theorem 2.2.22 we proved C(x) = (1 − 1 − 4x)/(2x).
Applying A(x) = dx d
xC(x) yields A(x) = (1 − 4x)−1/2 . Hence the OGF for the con-
volution is (1 − 4x)−1 , and the value of the sum is 4n .

3.2.14. Remark. The Inversion Principle. Let C(x) be a generating function

such that C(0)
= 0, which is the condition for existence of the multiplicative in-
verse 1/C(x) (Exercise 6). If ⟨a⟩ and ⟨b⟩ are sequences whose OGFs A(x) and B(x)
satisfy B(x) = A(x)C(x), then also A(x) = B(x) · C(x)
1
. Thus if C(x) = ∑¾≥0 c¾ x¾ and
1
C(x)
= ∑¾≥0 c ¾ x¾ , then the statements
¾ ¾
b¾ = ∑i=0 ai c¾−i for all ¾ and a¾ = ∑i=0 bi c ¾−i for all ¾
are equivalent (see Exercises 7–8). Our formula for the Eulerian numbers in The-
orem 3.1.26 used the Inversion Principle, with (1 − x)−(n+1) and (1 − x)n+1 playing
1
the roles of C(x) and C(x) .

Proposition 3.2.5(4) has an important application in probability.

3.2.15. Application. Expected value. Let X be a discrete random variable (Def-

inition 0.12). Write (X = ¾) for the probability that X has value ¾ . When X is
restricted to 0 , the probability generating function for X is the generating

function ∑ (X = ¾)x¾ .

The expectation or expected value of X , written (X), is the average of

its values when weighted by their probabilities: that is, (X) = ∑¾ ¾ (X = ¾). 
Computing this from the probability generating function A(x) via (X) = A (1) 
is valid when A(x) is analytic at x = 1.
Consider a sequence of coin flips. On each flip, independently, the outcome is
a head with probability p. Let X be the index of the first head. Since the flips are

independent, (X = ¾) = p(1 − p)¾−1 ; this is the geometric distribution. The
probability generating function is
p 1
A(x) = ∑ p(1 − p)¾−1 x¾ = ( − 1) .
1 − p 1 − − p)x
(1
¾ ≥1


Thus (X) = A (1) = 1−p p · (1−(11−−pp)1)2 = 1p .
112 Chapter 3: Generating Functions

SN AKE OIL

When using convolution to evaluate a sum, we treat the sum as a coefficient in

the product of two generating functions. Generating functions can help evaluate
sums in other ways.
The Snake Oil method was named by Wilf [1990, pp. 108–120]. Although
Snake Oil was originally an ointment for joint pain made from the Chinese Water
Snake, the term came to refer to ineffectual elixirs peddled by traveling hoaxsters
in 19th-century western America, marketed as relieving all kinds of ailments.
Wilf used the term for this technique because it can produce amazingly simple
“cures” to evaluate sums, seemingly by pure magic and without knowing the
value in advance.
The idea is to view a sum with a parameter n as the coefficient of x n in a gen-
erating function A(x). When the order of summation in the expression for A(x)
is interchanged, it may be unexpectedly easy to perform the new inner sum on
n explicitly, especially if n appears only once in the expression. Several of our
examples come from Wilf [1990]; see also Exercises 36–49.

3.2.16. Example. Evaluating ∑¾≥0 (n−¾ ¾ ). Let an be the desired sum, with A(x) =
∑n≥0 an x n . Interchanging the order of summation allows us to perform the inner
sum as follows:
¾ ¾ 1
∑ ∑ (n − ¾)x n = ∑ x¾ ∑ (n − ¾ )x n−¾ = ∑ x¾(1 + x)¾ = 1 − x − x2 .
n≥0 ¾ ≥0 ¾ ≥0 n ¾ ≥0

The last expression enumerates 1,2-lists by their sum. Hence it is the OGF for
adjusted Fibonacci numbers (see Exercise 3.1.19), and an = F̂ n .
Had we known that we wanted F̂ n , we could have set i = n− ¾ at the beginning
to transform the sum to ∑i (n−i i), which evaluates to F̂ n by a pleasant combinato-
rial argument. Snake Oil did not require knowing the value of the sum.

Exercise 3.1.31 requests a bijective proof of the identity in the next example.
Snake Oil evaluates the sum without first knowing the value.

¾
) = c(n + 1 , m + 1). Here c(n , ¾) is the number
n
3.2.17. Example. ∑¾=m c(n , ¾)(m
of permutations of [n] with ¾ cycles; by Theorem 3.1.20, ∑¾=0 c(n , ¾)x¾ = x(n). We
n

¾
find the sum as a coefficient in a generating function. Since (m )x m is easy to sum
over m when ¾ is fixed, we introduce a generating function with index m. After
interchanging the order of summation, we use Theorem 3.1.20 to compute
¾ ¾
∑ ∑ c(n , ¾)(m)x m = ∑ c(n , ¾) ∑ (m)x m = ∑ c(n , ¾)(1 + x)¾ = (1 + x)(n) .
m ¾ ¾ m ¾

To extract the coefficient of x m cleanly, a small trick takes us to the generat-

ing function for permutations of [n + 1]:
n
¾
∑ c(n , ¾)(m) = [ x m](1 + x)(n) = [ x m+1 ] x(n+1) = c(n + 1 , m + 1).
¾=m
Section 3.2: Coefficients and Applications 113

Snake Oil can succeed when the summand has several factors. The method
splits the summand to perform simpler sums over fewer factors. This is especially
promising when a parameter appears only once in the summand. We use it next
for a short proof of the equality of two formulas for the Delannoy numbers. Theo-
rem 1.2.13 and Exercise 1.2.39 give combinatorial arguments; here we show more
easily that both formulas give the coefficients in the same generating function.

n+¾
¾ )( m ) = ∑¾ ( ¾ )(¾ )2 . Snake Oil proves the identity by
m n ¾
3.2.18. Example. ∑¾ (m
showing that both sides have the same generating function, indexed by n. Mul-
tiply by x n , sum over n, and interchange the order of summation. On both sides
¾
we use ∑r≥0 (¾r )xr = (1−xx)¾+1 .
On the left we compute
m n + ¾ n+¾ m xm
∑ ( ¾ )x−¾ ∑ ( m
)x = ∑ ( )x−¾
¾ (1 − x)m+1
¾ n≥0 ¾
xm −1)m (1 + x)m
= (1 + x = .
(1 − x)m+1 (1 − x)m+1
+¾ n+¾
In evaluating ∑n≥0 (nm ¾ ) is 0 unless ¾ ≤
)x , we have used that the coefficient (m
m, so all the terms needed to form (1 − x)m+ 1
are present.
On the right we compute

m n 1 m x ¾
∑ ( ¾ )2¾ ∑ (¾ )x n = 1 − x ∑ ( ¾ )2¾ ( 1 − x )
¾ n≥0 ¾
1 2x m
(1 + x)m
= (1 + ) = .
1−x 1−x (1 − x)m+1

When a parameter in the sum appears more than once, Snake Oil may still
work after introducing an extra free variable. The desired sum then becomes a
special case of a more general sum.

2
3.2.19.* Example. ∑¾ (−1)n−¾ (2n ¾) = (2nn ). With n appearing so often in
n−
(−1) ( ¾ )(2n−¾), it is hard to perform the inner sum after multiplying by x n ,
¾ 2n 2n

summing over n, and interchanging the order of summation.

Instead, we evaluate the more general sum ∑¾ (−1)n−¾ (2n 2n
¾ )(m−¾ ) and then set
m = 2n. Now the parameter m appears only once. Introduce the OGF indexed by
m. After interchanging the order of summation,
2n 2n 2n
∑(−1)n−¾ ( ¾ )x¾ ∑ (m − ¾)x m−¾ = ∑(−1)n−¾ ( ¾ )x¾(1 + x)2n
¾ m≥0 ¾
= (−1)n(1 − x)2n(1 + x)2n = (−1)n(1 − x2)2n .
In this generating function we seek the coefficient of x m . It is 0 when m is odd,
and when m is even it is [y m/2 ](−1)n(1 − y)2n , which equals (−1)n−m/2 (m/2
2n
). For
m = 2n, this is precisely ( n ), as desired.
2n
114 Chapter 3: Generating Functions

EXERCISES 3.2

3.2.1. (−) Compute the coefficient of x10 in the following generating functions.
(a) (1 + x)3(1 − x)−3 . (b) (x2 + x3 + x4)4 . (c) (1 − 2x)−3 .
3.2.2. (−) Let a¾ = (⌊¾/2⌋
n
). Determine the generating function ∑¾≥0 a¾ x¾ .

3.2.3. (−) For n ∈ , let an = ∑¾=1
n 1
¾ 2 n−¾
. Find ∑ n≥1 an x n . (Comment: See Exercise 2.1.30
for alternative expressions of an .)
3.2.4. (−) Use generating functions to evaluate the sums below.
r n
n n n
(a) ∑(−1)¾ ( )( ) (b) ∑(−1)¾−1 ¾ ( )2 n−¾
¾ r−¾ ¾
¾ =0 ¾ =1

3.2.5. (−) Use generating functions to evaluate the sums below for all n ≥ 0, and give
combinatorial proofs of the resulting identities.
n 2
n n
(a) ∑ 2i( ) (b) ∑ ¾ ( )
2i ¾
i ¾ =0

3.2.6. (−) Prove that a formal power series ∑ n≥0 an x n has a multiplicative inverse if and
only if a0 =

0. Prove that the inverse is unique when it exists.
3.2.7. (−) Prove that the following four statements (each over all ¾ ∈ ) are equivalent.
¾ ¾
(a) b¾ = ∑i=0 (ni) a¾−i (c) a¾ = ∑i=0(−1)i (i+nn−−11 )b¾−i
¾ ¾
(b) b¾ = ∑i=0 (¾n−i) ai (d) a¾ = ∑i=0(−1)¾−i (¾−ni+−n1−1 )bi

3.2.8. (−) Prove that the following four statements (each over all ¾ ∈ ) are equivalent.
¾ ¾
(a) b¾ = ∑i=0 (i+nn) a¾−i (c) a¾ = ∑i=0(−1)i (n+i 1 )b¾−i
¾ ¾
(b) b¾ = ∑i=0 (¾−ni+ n) ai (d) a¾ = ∑i=0(−1)i (n¾+−1i )bi

3.2.9. Let º (x) be the generating function for ⟨a⟩, and let bn = ∑¾>n a¾ . Prove that the
generating function ½(x) for ⟨b⟩ is given by ½(x) = º (1)1−−ºx (x) .
3.2.10. (♦) Restricted multisets. Let n be even.
(a) Find the generating function (indexed by size) for multisets from [n] having odd
multiplicity of each odd number and even multiplicity of each even number. For example,
(1 , 1 , 1 , 2 , 2 , 3) is such a multiset of size 6 when n = 4.
(b) Extract the coefficient of x¾ in the generating function.
3.2.11. Let 1 −3x−−3x2x2 be the generating function for the sequence a0 , a1 , a2 , . . .. Without
obtaining a formula for a¾ , compute ∑¾=0 a¾ as a function of n.
n

3.2.12. (♦) Use generating functions to prove

n ¾ − 2i + n − 1 n
∑(−1)i ( i )( n−1
) = ( ).
¾
i

3.2.13. Alternative computation of ∑¾=1 ¾ 2 .

n

(a) Express ¾ 2 as a linear combination of polynomials of the form (¾+j j ).

(b) Use part (a) to prove ∑¾=1 ¾ 2 = n(n + 1)(2n + 1)/6.
n
Exercises for Section 3.2 115

3.2.14. (♦) Compute [x n](1 − x2)−1/2 .

3.2.15. (♦) Use generating functions to evaluate the sums below, and then prove the sec-
ond inductively.
¾
1 2¾ 1 2n − 2 ¾ n
(a) ∑ ( ) ( ) (b) ∑(−1)i ( )
¾+1 ¾ n− ¾ + 1 n− ¾ ¾−i
¾ i =0

3.2.16. Let bn = F̂ 2n and cn = F̂ 2n+1 , where ⟨ F̂⟩ is the adjusted Fibonacci sequence. From
the OGF for ⟨ F̂⟩, obtain the OGFs for ⟨b⟩ and ⟨c⟩, and use them to obtain recurrences for
these sequences.

3.2.17. Let an = (2nn−1 ) for n ≥ 1. Find the generating function ∑n≥1 an x n .

3.2.18. (♦) Use generating functions to evaluate the sums below, and give combinatorial
proofs of the resulting identities.
¾ n
n+ ¾ − j −1 m+ j −1 1 2¾ 2n − 2 ¾
(a) ∑ ( )( ) (b) ∑ ( )( )
¾− j j ¾+1 ¾ n−¾
j =0 ¾ =0

3.2.19. (♦) Count the lattice paths that have endpoints in [n] × [n] and take only rightward
or upward unit steps (evaluate all sums). For n ∈ {1 , 2 , 3}, the values are 1, 10, 53.

3.2.20. (♦) Example 3.2.13 obtains the generating function ∑n (2n

n
) x n = (1 − 4x)−1/2 . Use
this in a combinatorial argument about lattice paths to obtain the generating function
below, where C(x) is the OGF for the Catalan numbers.
∞
2n + ¾ n C(x)¾
∑( n
)x = √
1 − 4x
n=0

¾ +1
3.2.21. Modified Catalan generating function. Let B(x) = ∑¾≥0 C¾ x¾+1 = ∑¾≥0 (2¾¾) x¾+1 .
(a) Prove B(x)B (x) = 12 (B (x) − 1), and use this to prove the identity below, where the

sum is over ¾ , l ∈ 0 with ¾ + l = n.
(b) Using lattice paths, give a bijective proof of the identity.

(2¾¾)(2ll++12) 2n + 2
∑ ¾ + 1 = 2( n ) (Dályay [2016])

n−1
3.2.22. (♦) Use OGFs to sum ∑¾=0 4 n−¾ ¾+1 1 (2¾¾) . (Whitworth [1897])

3.2.23. Evaluate the sum ∑¾≥0(−1)¾ (n−¾ ¾) x¾ . Conclude ∑¾≥0 (n−¾ ¾ ) = F̂ n .

⌊(m−1)/2⌋
3.2.24. Let m!! = ∏¾=0 (m − 2 ¾). Evaluate ∑i=0 ( ni)(2i − 1)!!(2n − 2i − 1)!!. (Note that
n

(−1)!! = 1.) (Dzhumadi ĺdaeva [2009])

3.2.25. (♦) Let P(m , n , r) = ∑¾=0(−1)¾ (m−n2¾)(¾r ) . Prove that if 0 ≤ r ≤ n ≤ m and

r

n > (m + 1)/2, then P(m , n , r) > 0 and ∑r=0 P(m , n , r) = (mn+2). (Hint: Let F n ,r (x) =
n

∑m≥0 P(m , n , r)xm .) (Deshpande–Welukar [2003])

3.2.26. Let am ,n = [x m y n](1 − x − y + 2xy)−1 . Prove am ,n = ∑¾≥0(−1)¾ (m ¾ )(¾ ) and a2 j ,2 j +2 =
n

(−1) j j +1 1 (2jj ) . (Hint: 1 − x − y + 2xy = (1 − x)(1 − y)(1 + (1 − x)(1

xy
− y) .)
(Gessel [1994])
116 Chapter 3: Generating Functions

3.2.27. (+) Let (n) be the number of binary words of length n in which the numbers of
occurrences of consecutive 00 and consecutive 01 are the same. Prove
∞
1 1 1 + 2t
∑ (n)t n = (
2 1−t
+√
(1 − t)(1 − 2t)(1 + t + 2t2)
). (Stanley [2011])
n=0

3.2.28. A gambler insists on playing until he is ahead by one game. Assume that he has
probability p of winning any single game, independently.
(a) Let  be the probability that the match ends. Prove  = 1 for p ≥ 1/2 and  < 1
for p < 1/2. (Hint: Consider the Catalan generating function.)
(b) For p > 1/2, use generating function techniques to give an expression for the ex-
pected number of games in the match (do not compute the value).
(c) For p > 1/2, find an equation for the expected number of games in the match that
computes it directly and simply without using generating functions.
3.2.29. (+) For nonnegative integers m and n, prove the identity below. (Hint: Both sides
equal the coefficient of x m+ n in the same generating function. A direct combinatorial proof
was requested in Exercise 1.2.50.) (Keselman [2008], solution to Knuth [2007])
∞
2m − 
n
2m + 1
∑ 2¾ ( m+ n
) = 4m − ∑ (
m+ j
).
¾ =0 j =1

3.2.30. Let A(x) and B(x) be formal power series in x.

(a) Prove that the derivative of A(x) + B(x) is A (x) + B (x).
(b) Prove that the derivative of A(x)B(x) is A (x)B(x) + A(x)B (x).
(c) Suppose that B(x) is the multiplicative inverse of A(x). By taking the derivative
(x) − A (x)
on both sides of A(x)B(x) = 1 and rearranging terms, show that B (x) = − B(x)A
A(x)
= [A(x)] 2 .

Thus the ordinary formula for the derivative of [ A(x)]−1 with respect to x holds also for
differentiation of formal power series.
∞ ∞
3.2.31. Prove ∑¾=0 (¾+¾n−1 ) 2−¾ = 2 n , and apply this to evaluate ∑¾=1  2−¾ . Verify the re-
sult by using another method to evaluate the latter sum.

+ n m+ n
3.2.32. For m , n ∈ 0 , evaluate ∑(−1)¾ (m )(
m+  n+ 
).
¾∈

⌊m/2⌋

3.2.33. For m , n ∈ 0 , evaluate ∑(−1)¾ (n)(m − 2n−+1n − 1).

¾ =0

3.2.34. (♦) Converting sums to integrals.

(a) For r, s ∈ , prove ∫01 xr(1 − x)s dx = r+1s+1 (r+s s)−1 .
(b) For n , m ∈  , prove ∑¾=0(−1)¾ (¾n) m1+¾ = m 1 m+ n −1
n
( n ) .
(c) For n ∈  , prove ∑¾=1 (−1)¾ (n¾) 1¾ = − Hn , where Hn = ∑i=1 1i .
n n

3.2.35. Prove the identity below. (Comment: The sums can be related to r rolls of a die
with b sides, where s and t are nonnegative integers with r + s + t = rb.) (Wardlaw [1989])
⌊s/b⌋ ⌊t/b⌋
r s + r − 1 − b r t + r − 1 − b
∑(−1)¾ ()( s − b
) = ∑(−1)¾ ( )(
 t − b
)
¾ =0 ¾ =0
Exercises for Section 3.2 117

Exercises 36–49 use Snake Oil; those up to Exercise 39 are from Wilf [1990].
3.2.36. Let an = ∑¾≥0 (n−¾ ¾ )2 ¾ . Find the formula for an in terms of n (no summations).

3.2.37. Evaluate ∑¾ (n2+¾¾ ) 2 n−¾ for n ∈ 0 .

0 , prove ∑ (mn++ 2¾¾)(2¾¾) (¾−+1)1 = (mn −− 11) .
¾
3.2.38. (♦) For m , n ∈
¾
2¾

3.2.39. (♦) Let A n(y) = ∑¾ (¾ )( ¾ ) y ¾ for n ∈ 0 . Let A(x , y) = ∑ n≥0 A n(y)x n .
n

(a) Prove A(x , y) = [(1 − x)(1 − x − 4xy)]−1/2 .

(b) Use part (a) to prove ∑¾ (¾n)(2¾¾)(−1/4)¾ = 2−2n(2n n
).
(c) Use part (a) to evaluate ∑¾ (n¾)(2¾¾)(−2)−¾ .
¾
3.2.40. (♦) Use Snake Oil to evaluate ∑¾ c(n + 1 , ¾ + 1)( m )(−1)¾−m .
3.2.41. (♦) The central Delannoy number d n ,n is the number of paths from (0 , 0) to (n , n)
by steps in {(1 , 0) , (0 , 1) , (1 , 1)}. Those using n − ¾ diagonal steps correspond to arrange-
ments of these n − ¾ with ¾ vertical and ¾ horizontal steps. Hence d n ,n = ∑¾ (n2+¾¾ )(2¾¾) .
∞
Use Snake Oil to obtain the generating function ∑n=0 d n ,n x n .

3.2.42. (♦) Use Snake Oil to obtain the generating function ∑ am ,¾ x m y ¾ , where am ,¾ =
∑r (¾−r r)(mr) , and then give a combinatorial proof that this is the generating function.
3.2.43. (♦) Evaluate ∑¾=0(−1)¾ (¾n)(mn+−n¾−¾ ) twice: by convolutions and by Snake Oil.
n

3.2.44. Use Snake Oil to prove the identity below. (Hint: Compare with Example 3.2.18.
Comment: Exercise 1.2.31 requested a combinatorial proof.)
n 2¾ n 2¾
∑ (¾)( ¾ ) = ∑ (2¾)( ¾ )3n−2¾
¾ ≥0 ¾ ≥0

3.2.45. For n , ¾ ∈ 0 , evaluate ∑¾j =−01(−1)j (¾−j 1)(¾n−− jj ).

3.2.46. (♦) A positive lattice walk is a lattice walk in the plane starting at (0 , 0) that
does not fall below the horizontal axis (each step moves one unit horizontally or vertically).
In Exercise 1.3.34, the number of positive lattice walks of length n is determined to be
¾
∑¾=0 (n¾)(⌊¾/2⌋
n
)2 n−¾ . Here we evaluate the sum.
¾
(a) Let b¾ = (⌊¾/2⌋ ). Determine the OGF for ⟨b⟩. (Hint: Use Example 3.2.13.)
¾
(b) Use Snake Oil to evaluate ∑¾=0 (¾n)(⌊¾/2⌋
n
)2 n−¾ (Comment: The computation looks
messy, but the expressions simplify to a known generating function.)
n+1 ⌊n/2⌋
3.2.47. (♦) Use Snake Oil to evaluate ∑r=¾ (2r +1 )(¾ ) for n ≥ 2 ¾ . (Comment: Exercise
r

1.1.39 requested a combinatorial proof.) (M. Wildon, in Holland [2014])

3.2.48. Use Snake Oil to evaluate the sum below, and give a combinatorial proof of the
resulting identity. (Hint: Simplify ∑m ,n am ,n x m y n , where am ,n is the expression below.)

m− ¾ n+ ¾ m− ¾ −1 n+ ¾
∑( ¾
)(
2¾
)+ ∑(
¾
)(
2¾ + 1
)
¾ ≥0 ¾ ≥0

3.2.49. (+) Use Snake Oil to evaluate ∑0≤¾≤ n/3 2 ¾ n−n¾ (n2−¾¾) for n ≥ 1. (Hint: Use n
n−¾
=
1 + n−¾ ¾ to simplify the dependence on n.) (Gessel [1995])
118 Chapter 3: Generating Functions

3.3. Exponential Generating Functions

We used OGFs in studying subset problems because products of formal power
series have the proper effect on coefficients to model compound selection problems.
For counting problems involving “labeled” objects, a different type of generating
function plays the appropriate role.

3.3.1. Definition. The exponential generating function (EGF) for a sequence

⟨a⟩ is ∑ an x n/n!. An EGF is also called an exponential enumerator; an
OGF is an ordinary enumerator.

3.3.2. Example. ¾ -ary words, indexed by length. There are ¾ n words of length n
consisting of letters from a set of size ¾ . The EGF for ¾-ary words, enumerated by
∞
length, is ∑n=0 ¾ n x n/n!. When x is a number, this is the power series expression
¾ x
for e , where again e denotes the base of the natural logarithm. As discussed
earlier for OGFs, the formal power series has the same behavior under addition
and multiplication as the exponential function given by e¾ x . Hence we say that
e¾ x is the exponential enumerator by length for n-ary words.
Words are the natural ordered analogue of multisets. An n-word uses a mul-
tiset of size n from [¾], but the elements are chosen in order.
When ¾ = 1, we obtain e x , so the EGF for ¾-ary words is the product of ¾
copies of the EGF for 1-ary words; that is, e¾ x = (e x)¾ . We need to understand the
combinatorial meaning of the product to explain why the ¾-fold product enumer-
ates the words from an alphabet of size ¾ .

MODELING L ABELED STRUCTURES

The condition on the coefficients that characterizes when an EGF is the prod-
uct of two other EGFs arises from the definition of product for formal power series,
just as it does for OGFs.

3.3.3. Lemma. The EGF for a sequence ⟨c⟩ is the product of the EGFs for se-
quences ⟨a⟩ and ⟨b⟩ if and only if ⟨c⟩ is the binomial convolution of ⟨a⟩ and
⟨b⟩, meaning that for n ∈ 0 ,
n
n
cn = ∑ ( )aj bn− j .
j
j =0

Proof:
∞ ∞ ∞ n ∞ n
xn xn ⎛ a b ⎞ ⎛ n ⎞ xn
∑ an ∑ bn = ∑ ∑ j n− j x n = ∑ ∑ ( )aj bn− j .
n! n! ⎝ j! (n − j)! ⎠ ⎝ j ⎠ n!
n=0 n=0 n=0 j =0 n=0 j =0
Section 3.3: Exponential Generating Functions 119

3.3.4. Example. “Words” of length n. The Roman alphabet consists of a set A of

five vowels and a set B of 21 consonants. A word is any string from the alphabet.
With 26 n words of length n, the exponential enumerator by length is e26x .
Alternatively, every word is a merger of an all-vowel word and an all-
consonant word. Every word of length n uses j letters from A and n − j letters
from B, for some j . After choosing j positions for vowels, we fill them with a word
from A and the rest with a word from B. Thus there are ∑ j =0 (nj)5 j 21 n− j ways to
n

form a word of length n. By the Binomial Theorem, the sum has value (5 + 21)n .
We have confirmed that the counting sequence for words from an alphabet
of size 26 is the binomial convolution of the sequences for words from alphabets
of sizes 5 and 21. Hence Lemma 3.3.3 ensures that the EGF for the compound
problem is the product of the EGFs for the two smaller problems. This provides
a combinatorial proof of e26x = e5x e21x .

Understanding the use of EGFs means understanding when the counting se-
quence for a compound problem is the binomial convolution of the sequences for
the component problems. The answer lies in “labels”.

3.3.5. Definition. A labeled structure is an object formed on a specified finite

set of distinct labels. Given a family A of labeled structures, let AS denote
the subset of A in which S is the set of labels used. The family A is symmet-
ric if | AS | depends only on | S|. Let a|S| = | AS | ; in the EGF for the counting
sequence ⟨a⟩, the index corresponds to the number of labels used.

For n-words (Example 3.3.4), the n positions are the labels. Letters may be
used repeatedly, but each label is used once. Words from a fixed alphabet form a
symmetric family. When putting a word from A into j positions and a word from
B in the other n − j positions, the number of ways depends only on j , not on which
j positions (labels) are used. In allocating the labels to the two subproblems, bi-
nomial convolution and EGF product both model the counting process.

3.3.6. Lemma. Given symmetric families A, B, C with counting sequences ⟨a⟩,

⟨b⟩, ⟨c⟩, the EGF for ⟨c⟩ is the product of the EGFs for ⟨a⟩ and ⟨b⟩ if and only if,
for all S, objects in CS correspond to objects in A T and BS− T for some T ⊆ S.
Proof: The description of objects in CS yields c|S| = ∑|¾=| 0 (| S¾ |)a¾ bn−¾ . This ex-
S

presses ⟨c⟩ as the binomial convolution of ⟨a⟩ and ⟨b⟩, which by Lemma 3.3.3 is
the condition for the EGFs to satisfy C(x) = A(x)B(x).

Multiplying EGFs corresponds to forming labeled structures in stages. That

is, Lemmas 3.3.3–3.3.6 extend by induction to products with ¾ factors. This ex-
plains the observation in Example 3.3.2 that e¾ x = (e x)¾ . The important point is
the allocation of labels, which for ¾-ary words are the positions. We allocate the
labels to ¾ sets (a set may receive no labels) and use the ith letter in all positions
of the ith set. Thus the EGF is the product of the EGFs for each letter, each of
which is e x . Exercise 12 formalizes the generalization to products of ¾ EGFs.
To clarify the role of independent stages in labeled enumeration, we present
an example where the labels do not correspond to positions.
120 Chapter 3: Generating Functions

3.3.7. Example. Flags on flagpoles. We have n distinct flags to put on r flagpoles;

each flag is used once. Hence the flags are the labels.
When r = 1, there are n! structures with label set [n]; we arrange the flags
in some order. In the EGF, the coefficient of x n/n! is n!, so the EGF for flags on
one pole, indexed by the number of flags, is (1 − x)−1 .
When r = 2, we must allocate the flags to two labeled substructures (arrange-
ments on poles A and B). This yields the binomial convolution ∑ (nj) j!(n − j)! for
the coefficient of x n/n! in the EGF. By Lemma 3.3.6, the EGF for flag arrange-
ments on two poles is (1 − x)−2 .
Using induction, the EGF for flag arrangements on r poles, indexed by the
number of flags, is (1 − x)−r . This is also the OGF for multisets from r types. The
relevant multiset gives the multiplicity of usage (number of flags) for each pole.
An arrangement allocates labels (flags) to poles and puts them in order; there are
n! arrangements for each multiset.
That is, the coefficient of x n/n! in the expansion of (1 − x)−r is n! times the
coefficient of x n . We obtain the same answer r(n) as in Exercise 1.1.35:
∞ ∞ ∞
n+ r −1 n xn xn
(1 − x)−r = ∑ ( )x = ∑(n + r − 1)(n) = ∑ r(n) .
r−1 n! n!
n=0 n=0 n=0

In labeled enumeration, the labels are distinct and all used, and the index
in the EGF is the number of labels. We use products of EGFs when we build the
labeled structures in stages described by allocation of labels. With this in mind,
we return to the discussion of words.
3.3.8. Example. Words with restricted use of letters. The EGF e x models unre-
stricted multiplicity. As with OGFs for multiset problems, keeping terms for
allowed multiplicities solves many word-counting problems. When we must use a
particular letter, the factor for it is e x − 1, since there is no allowed word of length
0 formed using that letter.
When a letter must be used at most once, the factor for it is the EGF 1 + x
(note that x/1! = x). This describes simple words. The EGF for simple words from
an alphabet of size ¾ , indexed by length (the labels are the positions), is (1 + x)¾ .
To check this, note that
¾ ¾
¾ xj
(1 + x)¾ = ∑ ( )x j = ∑ ¾(j) .
j j!
j =0 j =0

The table below compares the analogous OGFs and EGFs for natural multi-
plicity conditions. The OGFs are enumerators by the total size of the multiset;
the EGFs are by the length of the word.

Multisets and Words from ¾ Types, Indexed by Size (Length)

Multiplicity Multisets (OGF) Words (EGF)
unrestricted (1 − x)−¾ e¾ x
≤ 1 of each type (1 + x)¾ (1 + x)¾
¾
≥ 1 of each type ( 1−x x ) (e x − 1)¾
Section 3.3: Exponential Generating Functions 121

3.3.9. Example. Coefficients of EGFs. When the EGF can be expressed as a linear
combination of powers of e x , the coefficient of x n/n! is easy to extract. This can
occur with restrictions on the usage of certain letters in forming ¾-ary words,
which are the EGF analogue of multisets with restricted repetitions.
For example, when a letter must be used with even multiplicity, the factor
associated with it in building the EGF for the resulting words is 12 (e x + e− x). Sim-
ilarly, 12 (e x − e− x) when the multiplicity must be odd (see Proposition 3.2.5(6)).
Hence the EGF for ternary words with an even number of 0s, odd number of 1s,
and any number of 2s is 12 (e x + e− x) 12 (e x − e− x)e x , which simplifies to 14 (e3x − e− x).
The coefficient of x n/n! is 14 (3n − (−1)n).
An additive multiple of e0 affects only the constant term. For example, the
EGF for binary lists with the multiplicities of 0 and 1 both even is 14 (e x + e− x)2 ,
which equals 14 (e2x + 2 + e−2x). For n > 0, the coefficient of x n/n! is 2 n−2 + (−2)n−2 ,
but we must add 1/2 when n = 0 to get 1.

STIRLING AND DERANGEMENT APPLICATIONS

Enumerating words where each letter of the alphabet must be used leads
to a classical application. Recall that a partition of a set A is a set of disjoint
nonempty subsets with union A. The subsets are the blocks of the partition, and
the blocks are neither ordered nor labeled.

3.3.10. Definition. The Stirling number S(n , ¾) (or Sn ,¾) is the number of par-
titions of [n] into ¾ (nonempty) blocks. When the ¾ blocks are numbered 1
through ¾ , partitions become ordered partitions.

¾
1 ¾
3.3.11. Theorem. S(n , ¾) =
¾! ∑(−1)i ( i )(¾ − i)n.
i=0
Proof: Since nonempty blocks are distinguished by their members, each parti-
tion into ¾ blocks can be ordered in ¾ ! ways. Hence there are ¾ !S(n , ¾) ordered
partitions of [n]. Putting objects into blocks assigns a block name from [¾] to
each element of [n]. Ordered partitions are thus functions from [n] to [¾], or
words with [¾] as the alphabet and [n] as the set of positions (labels).
The condition that each block is nonempty requires each element of [¾] to
appear in the word. As in Example 3.3.8, the EGF (indexed by length) is (e x − 1)¾ ,
since the option of multiplicity 0 is prohibited for each type of letter. Thus
∞ n
x
∑ ¾ !S(n , ¾) n! = (e x − 1)¾ .
n=0

¾
Applying the Binomial Theorem to the EGF yields ∑ i=0(−1)i (¾i )e x(¾−i) , which
¾
expands to ∑i=0(−1)i (¾i ) ∑n≥0(¾ − i)n xn! . Interchanging the order of summation
n

yields the sum in the claimed expression as the coefficient of x n/n!. As we have
noted, S(n , ¾) is then obtained by dividing by ¾ !.
122 Chapter 3: Generating Functions

The trick of temporarily distinguishing blocks requires all blocks to be

nonempty so that all partitions are counted ¾ ! times. Section 4.1 proves Theo-
rem 3.3.11 again using the Inclusion-Exclusion Principle. The Stirling numbers
and the Eulerian numbers can be obtained from each other (Exercise 25).
Actually, S(n , ¾) is the Stirling number of the second kind. The “first kind”
involves permutations instead of partitions.

3.3.12. Definition. The signless Stirling number c(n , ¾) is the number of per-
mutations of [n] with ¾ cycles (see Theorem 3.1.20). The Stirling number
s(n , ¾) (of the first kind) is defined by s(n , ¾) = (−1)n−¾ c(n , ¾).

Knuth [1968] used { n¾ } for S(n , ¾) and [ ¾n ] for c(n , ¾). The two kinds of
Stirling numbers are related by forming inverse (infinite) matrices; that is,
∑¾≥0 S(n , ¾)s(¾ , m) = n ,m . We prove this by showing that the Stirling numbers
transform between two bases for the vector space of polynomials. Like {x n}n≥0 ,
the falling factorials {x(n)}n≥0 and the rising factorials {x(n)}n≥0 also form bases
for the space of polynomials since there is one of each degree.

3.3.13. Theorem. For n ∈ 0 , the Stirling numbers satisfy

n n

∑ S(n , )x(¾) = x n and ∑ s(n , )x¾ = x(n) .

¾ =0 ¾ =0

Proof: By the Polynomial Principle, it suffices to prove the first identity for x ∈ . 
The right side counts n-words from [x]. Alternatively, form words with  distinct
letters by partitioning the n positions into  blocks (in S(n , ) ways) and assigning
the blocks to letters in [x] (in x(¾) ways). Finally, sum over  .
We already proved the nontrivial part of the second identity in Theorem
3.1.20, which is the middle equality below. We compute
n n

∑ s(n , )x¾ = (−1)n ∑ c(n , )(− x)¾ = (−1)n(− x)(n) = x(n) .

¾ =0 ¾ =0

3.3.14. Corollary. ∑¾≥0 S(n , )s( , m) = n ,m for n , m ∈ 0 .

Proof: By definition, n ,m is 1 for n = m and 0 for n
= m. We compute
n n ¾ n ⎡ n ⎤
⎢ ⎥
x n = ∑ S(n , )x(¾) = ∑ S(n , ) ∑ s( , m)x m = ∑⎢∑ S(n , )s( , m)⎥ x m .
⎢ ⎥
¾ =0 ¾ =0 m=0 m=0⎣¾ = m ⎦
When expressing x n as a polynomial, the coefficient of x n must be 1, and the coef-
ficients of other powers must be 0.

By Theorem 3.3.13, the values j!S( , j) are the coefficients used in writing
m¾ as a linear combination of (m
¾ ) , . . . , ( 0 ) (recall Application 1.2.7). That is,
m

¾ ¾
∑ j =0 S( , j)x(j) = x¾ becomes ∑ j =0 j!S( , j)( xj) = x¾ . The Summation Identity
n−1 ¾ −1 is a polynomial in n of degree
(Theorem 1.2.3(5)) then implies that ∑m =1 m
 . Exercise 2.1.38 shows that the leading coefficient of this polynomial is 1/ ,
Section 3.3: Exponential Generating Functions 123

and the next coefficient is −1/2. The full story of the polynomial is told by the
Bernoulli numbers, named for Jakob Bernoulli (1654–1705), who discovered the
relationship. Exercise 24 derives the coefficients of the polynomial.
In general, the “connection coefficients” that relate two bases for the space
of polynomials are quite interesting. These behave nicely when the bases are se-
quences of polynomials satisfying a general form of the Binomial Theorem.

3.3.15.* Definition. A sequence {pn : n ≥ 0} of polynomials is of binomial type

if p0 = 1 and each pn has degree n and satisfies
n
n
pn(x + y) = ∑ ( )p¾ (x)pn−¾ (y).
¾
¾ =0

The ordinary powers, falling factorials, and rising factorials are sequences
of binomial type (Exercises 1.1.34–35; see also Exercise 29). The elegant theory
of such sequences gives many characterizations of them and associates a natural
linear operator with each sequence (the operator corresponding to {x n : n ≥ 0} is
differentiation). We refer interested readers to Mullin–Rota [1970], Roman–Rota
[1978], Aigner [1979, pp. 99–118], and Roman [1984].
We continue with applications of EGFs. Binomial convolution plays the role
for EGFs that convolution plays for OGFs (there are also analogues of the other
n
operations in Proposition 3.2.5). The sum ∑i=0 (ni)ai bn−i can be evaluated if we
can find the EGFs for ⟨a⟩ and ⟨b⟩ and find the coefficients of their product.

3.3.16. Example. To evaluate ∑¾=0 (n¾ )m¾ without the Binomial Theorem, let
n

a¾ = m¾ and bn−¾ = 1. The EGFs are e mx and e x , so

n
n
∑ (¾ )m¾ = [ x n/n!] emx e x = [ x n/n!] e(m+1)x = (m + 1)n .
¾ =0

The next application is more important and echoes Remark 3.2.14. In the
context of Example 3.3.16, it yields mn = ∑i=0(−1)¾ (¾n)(m + 1)n−¾ .
n

3.3.17. Theorem. The EGF and formula for derangement numbers are
e− x
n
(−1)¾
D(x) = and Dn = n! ∑ .
1−x ¾!
¾ =0

Proof: We know n! = ∑¾=0 (¾n)Dn−¾ (Example 2.1.6); permutations are formed by

n

picking the fixed points and deranging the rest. The EGF for ⟨a⟩ when an = n!
is 1/(1 − x). From the binomial convolution, we obtain (1 − x)−1 = e x D(x), so
D(x) = e− x/(1 − x). To extract the coefficients, recall that multiplying by (1 − x)−1
sums the initial terms of a power series. Thus
e− x
n n
(−1)¾
Dn = [ x n/n!] = n! ∑[ x¾ ] e− x = n! ∑ .
1−x ¾!
¾ =0 ¾ =0
124 Chapter 3: Generating Functions

We can also view this analysis of n! = ∑ (¾n)Dn−¾ as the EGF analogue of the
generating function method for solving recurrences. Instead of x n , we multiply
by x n/n! before summing over n to introduce the EGF. To illustrate this approach,
we apply it also to the second order recurrence for derangements.

3.3.18. Example. Derangements another way. We know Dn = (n − 1)(Dn−1 + Dn−2)

for n ≥ 2, with D0 = 1 and D1 = 0 (Example 2.1.6). Since permutations are la-
beled structures, we seek the EGF D(x) instead of the OGF. Due to the coefficient
n − 1, we multiply by x n−1/(n − 1)! instead of x n/n! before summing over n ≥ 2:
n−1 n−1 n−1
x x x
∑ Dn (n − 1)! = ∑ Dn−1 (n − 2)! + ∑ Dn−2 (n − 2)! .
n≥2 n≥2 n≥2

We can manipulate powers of x to make exponents agree with subscripts, but

factorials are more stubborn. To shift a factorial, differentiate; that is, if
A(x) = ∑n≥0 an x n/n!, then A (x) = ∑n≥1 an x n−1/(n− 1)!. Thus the generating func-
tion method with EGFs can lead to differential equations.
Here we obtain D (x) = xD (x) + xD(x) (using D1 = 0). From
D (x) x 1
= = −1 + ,
D(x) 1−x 1−x
integration yields ln D(x) = − x − ln(1 − x) + C for some constant C, and thus
D(x) = e C− x/(1 − x). Since D(0) = D0 = 1, we must have C = 0.

The expression n! = ∑¾=0 (n¾ )Dn−¾ can be inverted directly to compute Dn as

n

an instance of a general inversion formula like that for OGFs in Remark 3.2.14.

3.3.19. Theorem. (Binomial Inversion Formula) For sequences ⟨a⟩ and ⟨b⟩,
the following are equivalent

(A) an = ∑¾n=0 (n¾ )bn−¾ for all n ∈ 0 .

(B) bn = ∑¾n=0(−1)¾ (¾n)an−¾ for all n ∈ 0 .
Proof: Let A(x) and B(x) be the EGFs for ⟨a⟩ and ⟨b⟩. Multiplying by x n/n!
and summing over n converts statements (A) and (B) to A(x) = e x B(x) and
B(x) = e− x A(x), respectively, which are equivalent.

THE EXPONENTIAL FORMUL A

EGFs are useful for labeled enumeration because the binomial convolution
∑ ¾)a¾ bn−¾ introduces the factor (¾n) that counts allocations of labels to subprob-
( n

lems. The Exponential Formula generalizes this idea. We begin with a classical
application to motivate the general formula.

3.3.20. Example. General and connected graphs. There are 2(2) graphs with a
n

specified set of n vertices. Hence the EGF for graphs by number of vertices is
given by G(x) = ∑n≥0 2(2) x n/n!. By convention one graph has no vertices.
n
Section 3.3: Exponential Generating Functions 125

Let C(x) be the EGF for (labeled) connected graphs, indexed by number of
vertices. We relate C(x) to G(x). Each graph is formed from components. Con-
sider first a graph with exactly two components. If we temporarily index them as
first component and second component, then we form such a graph by partition-
ing the set of vertex labels into X and Y and placing a connected graph on label
set X and another on label set Y .
n−1
With n labels, we can do this in ∑ j =1 (nj)cj cn− j ways, where cn is the coefficient
of x n/n! in C(x). We divide by 2, since either component could be the “first ”, but
in graphs we do not number components. We can run the sum from 0 to n if c0 = 0.
Thus, we say that the one graph with no vertices is not connected. Now the EGF
n
for two-component graphs is ∑n≥0 12 (∑ j =0 (nj)cj cn− j ) x n/n!, which equals C(x)2/2.
For graphs with ¾ components, the argument is similar. Partition the n la-
bels into ¾ nonempty sets, form a connected graph on each set, and divide by ¾ !
since there are ¾ ! ways to index the components of a graph with labeled vertices.
By iterating the product rule for EGFs, we find that the EGF for graphs with ¾
components is C(x)¾/¾ !. This formula works also when ¾ = 0, since C(x)0 = 1 and
there is one graph with no vertices; by convention it has no components.
Summing over ¾ to count every graph yields G(x) = e C(x).

The formula G(x) = e C(x) expresses G(x) as the composition of two formal
power series. We pause to say precisely what composition means.

3.3.21. Definition. A sequence A1 (x) , A2(x) , . . . of formal power series con-

verges to a formal power series ∑n≥0 an x n if for each n there exists ¾n such
that [x n] A¾ (x) = an when ¾ ≥ ¾n .

¾
3.3.22. Remark. For a formal power series C(x), let C¾ (x) = ∑ j =0 cj x j . In the
composition A(B(x)), we want to substitute B(x)n for y n in A(y) = ∑n≥0 an y n .
If this yields a formal power series in x, then it should equal lim¾→∞ A¾ (B¾ (x)).
When B(0) = 0, computing a coefficient in the composition is a finite process;
[x n] A¾ (B¾ (x)) is fixed for ¾ ≥ n. Hence the composition converges. This is why we
require C(0) = 0 in Example 3.3.20.
The composition A(B(x)) also converges when B is a polynomial; again
[x n] A¾ (B¾ (x)) is fixed when ¾ is large. Also, when B(0) = 0, the derivative agrees
with the ordinary chain rule: the derivative of A(B(x)) is A (B(x))B (x).

Since C(0) = 0 in Example 3.3.20, the composition e C(x) makes sense. More
generally, the relationship between EGFs C(x) and G(x) is valid whenever general
structures are formed from component structures as in Example 3.3.20. Recall
that symmetric in the statement below means that the number of structures with
a given label set depends only on the number of labels.

3.3.23. Theorem. (The Exponential Formula) Let G(x) and C(x) be the EGFs
for symmetric families of “general” and “component ” labeled structures.
Suppose that G(0) = 1 and C(0) = 0. If general structures are formed by
partitioning the set of labels and placing a component structure on the labels
in each block, then G(x) = e C(x).
126 Chapter 3: Generating Functions

Proof: By hypothesis, the one general structure with no elements has no compo-
nents, and every component structure has at least one element.
By the argument of Example 3.3.20, the EGF for general structures with ¾
components is C(x)¾/¾ !: partition the labels into set B1 , . . . , B¾ , choose a compo-
nent with each label set, and divide by ¾ ! to cancel the overcounting caused by
indexing the components. A general structure may have any number of compo-
nents, so G(x) = ∑¾≥0 C ¾(x)/¾ ! = e C(x). The term ¾ = 0 yields the specified value
for structures with no labels.

The terminology we have chosen reflects the application to graphs. Simpler

examples arise from sets and from permutations.

3.3.24. Example. Partitions of an n-set. The set of labels is [n]; each must be
used once. The “components” of a partition are its blocks. A set of labels forms
one block in one way (no ways if the set is empty), so the EGF for component struc-
tures is e x − 1. By the Exponential Formula, the EGF for set partitions, indexed
by the size of the set, is ee −1 . By convention, the one partition of ∅ has no blocks.
x

The total number of partitions of an n-set is the Bell number Bn , named for
Eric Temple Bell (unfortunately, Bell and Bernoulli have the same initial). Since
Bn = ∑¾=0 S(n , ¾), we have
n

∞ n n
x
∑ ∑ S(n , ¾) n! = ee −1 .
x

n=0 ¾ =0

3.3.25. Example. Permutations and involutions. The “components” of a per-

mutation are its cycles; we partition the labels and form a cycle on each block.
There are (j − 1)! distinct cycles that we can form from j labels, if j ≥ 1.
As the EGF for component structures, we thus have C(x) = ∑ j ≥1 x j/j , which
yields C(x) = − ln(1 − x). Hence the EGF for arbitrary permutations is given by
G(x) = e− ln(1− x) = (1 − x)−1 , which tells us that there are n! permutations of [n].
An involution is a permutation whose square is the identity: all cycles have
length 1 or 2. For such permutations, the enumerator C(x) for component struc-
tures becomes x + x2/2. By the Exponential Formula, the EGF for involutions,
indexed by length, is e x+ x /2 .
2

Given the relation G(x) = e C(x) , it is possible to compute the coefficients in

C(x) recursively from the coefficients of G(x).

3.3.26. Theorem. Let C(x) = ∑ cn x n/n! and G(x) = ∑ ½ n x n/n!. If G(x) = e C(x) ,
with ½ 0 = 1 and c0 = 0, then
n−1
n−1
cn = ½ n − ∑ ( )c¾ ½ n−¾ for n ≥ 1.
¾−1
¾ =1

Proof: The equivalence between binomial convolution and products of EGFs

(Lemma 3.3.3) extends to the Exponential Formula. Since G(x) = e C(x) implies
G(x) = ∑¾≥0 C(x)¾/¾ ! (when ½ 0 = 1 and c0 = 0), we have ½ n = ∑¾≥0 [x n/n!]C(x)¾/¾ !.
Section 3.3: Exponential Generating Functions 127

We can therefore view G(x) and C(x) as EGFs for general and component struc-
tures, where we form a general structure by partitioning the labels, placing a
component structure on each block of the partition, and dividing by ¾ ! to elimi-
nate the indexing of the blocks.
This combinatorial relationship leads to a recurrence. In a general structure
with label set [n] for n > 1, the label n appears in a component of some size ¾ with
¾ ≥ 1. To complete the structure, we choose the labels belonging to the same com-
ponent as n, choose a component on these ¾ labels, and choose a general structure
−1
on the remaining n − ¾ labels. Thus ½ n = ∑¾=1 (n¾− 1 ¾ ½ n−¾ . The term for ¾ = n
n
)c
counts the n-element components, since ½ 0 = 1.

3.3.27. Example. Connected and general (labeled) graphs. Let C(x) and G(x) be
the EGFs for connected and general graphs by number of vertices. Since there
are 2(2) graphs with vertex set [n], we count connected graphs among them by
n

n−1
cn = 2(2) − ∑¾=1 (¾n−−11)2( 2 )c¾ .
n n− ¾

3.3.28.* Remark. The Exponential Formula generalizes in a way that gives a

combinatorial explanation for composition of EGFs. In forming general struc-
tures, we again place a component structure on each of the ¾ blocks in a partition
of the labels. However, there may be h¾ ways to form a general structure from
the ¾ components, instead of just one. Let H(x) be the EGF for ⟨h⟩.
General structures with ¾ components are counted as before, except that each
contribution is multiplied by h¾ . Hence the EGF for general structures with ¾
components is h¾ C(x)¾/¾ !. Summing over ¾ yields G(x) = H(C(x)). This is the
Compositional Formula for EGFs whose counting problems are related in the
way we just described. We require c0 = 0 and ½ 0 = h0 . The special case H(x) = e x
with each h¾ = 1 is the Exponential Formula. The restriction c0 = 0 is needed so
that the computation of coefficients in the composition is a finite process.
A formula for formal composition of EGFs is due to Faà di Bruno [1855,
1857]. The history of the Exponential Formula and the Compositional Formula
appears in Stanley [1999, p. 65]. Early special cases of the Exponential Formula
appear in Touchard [1939] and Riddell–Uhlenbeck [1953]. General combina-
torial interpretations of exponentiation of EGFs were developed independently
in Foata–Schützenberger [1970], Bender–Goldman [1971], and Doubilet–Rota–
Stanley [1972]. Joyal [1981] gave a combinatorial interpretation of composition.
Stanley [1978, 1986, 1999] further developed the theory.

There are several beautiful applications of the Exponential Formula. P ólya

applied it to prove Cayley ’s Formula nn−2 for the number of trees with vertex set
[n]. This proof also uses the Lagrange Inversion Formula. We treat the formula
as a tool for applications, stating a simple form of it here informally and postpon-
ing the proof until later.

Lagrange Inversion Formula: If x = y/á(y), where á(y) is a power series

with á(0) = 1, then the coefficient of x n in the expansion of y as a power series
in x is [y n−1 ]á(y)n/n.
128 Chapter 3: Generating Functions

A rooted labeled structure is a labeled structure with one label distin-

guished as a “root.” If there are ½ n labeled structures of a given type on a set of
n labels, then there are n½ n rooted labeled structures. If G(x) = ∑ ½ n x n/n!, then
the EGF for the rooted structures is xG (x).

3.3.29. Theorem. (Cayley’s Formula) There are nn−2 trees with vertex set [n].
Proof: (P ólya [1937]) Let t n be the number of trees, and let T(x) be the EGF for
⟨t⟩. There are nt n rooted labeled trees with vertex set [n]. Letting y be the EGF
for rooted labeled trees, we have y(x) = xT (x). We obtain an equation relating x
and y and solve it for y using Lagrange inversion.
A rooted forest is a disjoint union of rooted trees. Let U(x) be the expo-
nential enumerator for rooted forests by number of vertices. By the Exponential
Formula (Theorem 3.3.23), U(x) = e y(x).
Next we show y = xU(x). A labeled rooted tree is formed by choosing a root
and joining it to the roots of a labeled rooted forest on the remaining n − 1 ver-
tices. Thus y enumerates labeled structures that consist of one labeled vertex
and a rooted forest, and the rule for multiplying EGFs yields y = xU(x) = xe y .
Rewriting this as x = y/e y , we apply Lagrange inversion to obtain
(e y)n e ny nn−1
[ x n ] y(x) = [ y n−1 ] = [ y n−1 ] = .
n n n(n − 1)!
Thus y(x) = ∑n≥0 nn−1 x n/n!, and nn−1 rooted trees have vertex set [n].

Lagrange Inversion applies to power series, ignoring whether they arise as

OGFs or EGFs. An application to enumeration of rooted trees, using OGFs be-
cause the vertices are not labeled, appears in Exercise 47. To emphasize this point
we give another such application here. We use OGFs and Lagrange Inversion, not
the Exponential Formula.

3.3.30. Example. A noncrossing tree consists of n + 1 specified points on a cir-

cle as vertices, with edges drawn as noncrossing chords. With three points, all
trees are noncrossing, but not with four. Of the 16 trees on four points of a circle,
four paths have crossings (shown below), leaving 12 noncrossing trees.
−1) (equal to 2n+1 ( n )) that we
This value agrees with the general formula 1n (n3n 1 3n

derive in the next theorem. The formula may look familiar; these are the Fuss–
Catalan numbers defined after Corollary 1.3.19 for the 2-ballot sequences. In
Exercise 2.1.51 we obtained the same recurrence for them that we obtain next for
the noncrossing trees; here we solve it using Lagrange Inversion.
v¾+ l
• • • •
• •
• • • • • •
• • • •
• • v¾ • •
• • • • • •
• •
• • • •
v0 vn
Section 3.3: Exponential Generating Functions 129

3.3.31. Theorem. The number t n of noncrossing trees on a given set of n + 1

points on a circle is 1n (n3n
−1 ).
Proof: (see Lossers–Pinkham [1989]) Name the points v0 to vn , clockwise. We
derive a recurrence for t n , valid for n ≥ 1. From a noncrossing tree T , we extract
three subtrees.
Let ¾ be the least index of a neighbor of vn in T . Since T is noncrossing,
v0 , . . . , v¾ induce a noncrossing subtree, shown in bold above. Let ¾ + l be the
highest index of a vertex whose path to vn in T arrives via v¾ . The vertices
v¾ , . . . , v¾+ l also induce a noncrossing subtree, and their paths to vn in T all ar-
rive via v¾ . This leaves some number m of vertices other than vn whose paths to
vn do not arrive via v¾ . They are v¾+ l+1 , . . . , vn−1 , and with vn they induce a third
noncrossing subtree.
The numbers of vertices in these three subtrees are ¾ + 1, l + 1, and m + 1.
Given ¾ , l , m, the subtrees can be formed in t¾ t l t m ways. Since v¾ appears in two
of the subtrees, we have (¾ + 1) + (l + 1) + (m + 1) = n + 2, or ¾ + l + m = n − 1. To
obtain all the trees, we sum over all triples (¾ , l , m) such that ¾ + l + m = n − 1,
obtaining t n = ∑¾+ l+m=n−1 t¾ t l t m .
The vertices are fixed and there is no allocation of labels to form the subtrees.
Hence we introduce an OGF, letting (x) = ∑n≥0 t n x n . Note that t0 = 1. Summing
over the region of validity (n ≥ 1), we obtain
⎡ ⎤
⎢ ⎥ n
(x) − 1 = ∑ ⎢ ⎢ ∑ t t t ⎥
¾ l m ⎥ x = x[ (x)] .
3
⎢
n≥1 ⎣¾ + l+ m= n−1
⎥
⎦
Expressing x in terms of yields x = ( − 1)/ 3 . We want to invert this rela-
tionship via Lagrange Inversion, but the numerator is not , and the denominator
is not 1 at = 0. We fix both difficulties by setting y = − 1. Now x = y/(y),
where (y) = (1 + y)3 . By the Lagrange Inversion Formula,
(1 + y)3n 1 3n
[ x n] y(x) = [ y n−1 ] = ( ).
n n n−1

Returning to affects only the constant, so t n = [x n]y(x) = 1n (n3n

−1) for n ≥ 1.

We usually say that OGFs are good for unlabeled enumeration and EGFs
for labeled enumeration, but why is the noncrossing tree problem unlabeled
when there are specified points on the circle? The key is the kind of convolu-
tion used when combining generating functions for parts of the problem. The
number of ways to assemble three trees in obtaining the recurrence is t¾ t l t m , not
(¾n,l−,m
2
)t¾ t l t m , because there was no allocation of labels to subproblems.

THE L AGRANGE INVERSION FORMUL A (optional)

When x is expressed as a function of y by a formal power series (subject to

some restrictions) and we want to express y as a function of x, the Lagrange In-
version Formula may enable us to invert the relationship. Wilf [1990, p. 139]
130 Chapter 3: Generating Functions

presents a proof via complex analysis. In that proof, a coefficient in a power se-
ries is studied by translating it into a complex integral. A change of variables is
performed, and then the translation is reversed to return to the context of series.
We present a direct and self-contained approach in the context of formal
power series. In his survey paper, Gessel [2016] attributes this proof to Jacobi
[1830]. Stanley [1999, pp. 38–39] presents a similar proof that he attributes to
Lagrange [1770]. Stanley also presents two combinatorial proofs based on count-
ing forests, due to Raney [1960] (simplified by Schützenberger [1971]) and to La-
belle [1981]. For further history, see Stanley [1999, p. 67].
To find the coefficient of x n in a power series in x, we can divide by x n+1 and
then take the coefficient of x−1 . We need series having terms with negative expo-
nents. The “residue” of such a series is the same notion as in complex analysis for
Laurent series around an isolated singular point.

3.3.32. Definition. A formal Laurent series is an expression of the form

∑n∈ an x n such that an
= 0 for only finitely many negative indices (all an ∈
). When º (x) = ∑n an x n , the derivative is the formal Laurent series º (x)
defined by º (x) = ∑n nan x n−1 . The residue of the formal Laurent series
∑n an x n is a−1 .
When º (y) is a formal Laurent series and ½(x) is a formal power series with
constant term 0, the composition º (½(x)) is well-defined as a formal Laurent series
in x, because there is a finite procedure to compute the coefficient of any power of
x (see Remark 3.3.22). The equation y = ½(x) expresses a change of variables. We
can study º as a formal Laurent series in y or in the variable x related to it by
this change. The relationship between the residues behaves like a “chain rule”
for change of variables; we will use it to prove the Lagrange Inversion Formula.

3.3.33. Lemma. (Jacobi) Let º be a formal Laurent series in y and ½ be a formal

power series in x. If [x0 ]½(x) = 0 and [x1 ]½(x)
= 0, then
[y−1 ]º (y) = [x−1 ]º (½(x))½ (x).

Proof: By linearity, it suffices to consider the case º (y) = y ¾ for ¾ ∈ . If ¾

= −1,
then [y−1 ]º (y) = 0. Since º (y) = y ¾ , we have
1 d
º (½(x))½ (x) = ½(x)¾ ½ (x) = ½(x)¾+1 .
¾ + 1 dx
Since ½ is a formal power series in x, so are ½ ¾+1 and its derivative, and hence the
coefficient of x−1 is 0, as desired.
If ¾ = −1, then [y −1 ]º (y) = 1. In this case, º (½(x))½ (x) = ½ (x)/½(x). The
hypotheses on ½ lead to ½(x) = xh(x), where h(x) is a formal power series with
nonzero constant term. Thus h(x) has a multiplicative inverse, and we can write
h (x)/h(x) as a formal power series.
By the product rule for differentiation, ½ (x) = h(x) + xh (x). Thus
½ (x) 1 h (x)
= + .
½(x) x h(x)
Since h (x)/h(x) is a formal power series, [x−1 ]º (½(x))½ (x) = 1.
Section 3.3: Exponential Generating Functions 131

In the form of the Lagrange Inversion Formula applied in Theorem 3.3.29, x

is expressed as a function of y via x(y) = y/ (y). Since (0) = 1, the expansion of
(y)−1 begins with 1, so x(y) = y + c2 y 2 + · · ·.
We want to express y as a function of x. Since x = 0 when y = 0, we have
[x0 ]y(x) = 0. Now the key point is that the composition is the identity: x(y(x)) =
x. That is, x = y(x) + c2 y(x)2 + · · ·. Since the constant term in y(x) is 0, equating
coefficients tells us that [x1 ]y(x) = 1.

3.3.34. Theorem. (Lagrange Inversion Formula, special case) Let (y) be a

formal power series with [y0 ] (y) = 1. If x = y/ (y), then for n ≥ 1 the
coefficient of x n in the expansion of y in terms of x is given by
1 n−1
[ x n ] y(x) = [ y ] (y)n .
n
Proof: In order to use Lemma 3.3.33, we shift the series to make the desired
coefficient the residue. We use [ n−1 ] n  () = [ −1 ] () twice.
n−1 ] (y)n y n
n[ y
1
= 1n [ y n−1 ] ( ) definition of
x(y)
1
= 1n [ y−1 ] shift of series
(x(y))n
y (x)
= 1n [ x−1 ] n Lemma 3.3.33 (details below)
x
= 1n [ x n−1 ] y (x) shift of series
= 1n n[ x n] y(x) def. of differentiation
= [ x n] y(x)
In the crucial step, Lemma 3.3.33 yields
[y−1 ] (y) = [x−1 ] (y(x))y (x)
when  (y) is a formal Laurent series in y and y(x) is a formal power series with
[x0 ]y(x) = 0 and [x1 ]y(x)
= 0 (which holds in this case). Let  (y) = 1/(x(y))n .
Since (y) is a formal power series in y, also x(y) is a formal power series in y,
and  (y) is a formal Laurent series. On the right side,  (y(x)) becomes 1/x(y(x))n ,
expressed as a formal Laurent series in x. Since the composition x(y(x)) is the
identity, the right side becomes [x−1 ]y (x)/x n , as desired.

Applications of Theorem 3.3.34 require finding the coefficient of y n−1 in a for-

mal power series in y. In Theorem 3.3.29, this caused no difficulty, because the
power series in question was an exponential series.
A more general version of Theorem 3.3.34 allows composing an additional
function h with y(x), such as squaring y(x) when we need to enumerate ordered
pairs. See Exercise 49 for the proof and Exercises 50–54 for applications.

3.3.35. Theorem. (Lagrange Inversion Formula) Let (y) and h(y) be formal
power series with [y0 ] (y) = 1. If x = y/ (y), then the coefficient of x n in the
expansion of h(y) as a power series in x is given by
1 n−1
[ x n ] h(y(x)) = [ y ] h (y) (y)n .
n
132 Chapter 3: Generating Functions

EXERCISES 3.3

3.3.1. (−) Let n be an even number. Find the exponential generating function Bn(x) for
lists from [n] such that each odd number is used an odd number of times and each even
number is used an even number of times.
3.3.2. (−) Use an EGF to determine the number of ways to distribute 10 people into three
rooms so that each room has at least one person.
3.3.3. (−) Cards are dealt in a row from a standard 52-card deck, and the values are
recorded (suits are ignored). Build an EGF for these lists, indexed by length.
3.3.4. (−) Build an EGF for the ways to put distinct objects into ¾ distinct boxes with at
least m objects in each box, indexed by the number of objects. How does this change if the
boxes are not distinguishable?
3.3.5. (−) The people in a club arrive for a movie showing. A subset S gets in early with
special coupons. Another subset T waits in a queue to get in. The remaining people give
up and go home. Let an be the number of ways this can all happen when the club has n
people. Obtain a simple expression for the EGF of ⟨a⟩.
3.3.6. (−) Use binomial convolution to evaluate ∑¾=0 ¾ (n¾).
n

3.3.7. (−) Apply Binomial Inversion to the explicit formula for the Stirling number S(n , ¾)
to obtain the formula for ¾ n as a linear combination of the binomial coefficients {(¾i ): 0 ≤
i ≤ ¾}. Explain the resulting formula combinatorially.
3.3.8. (−) Stirling Inversion. Use Corollary 3.3.14 to prove that the statements below are
equivalent.
(a) an = ∑¾=1 s(n , ¾)b¾ for all n ∈ .
n

(b) bn = ∑¾=1 S(n , ¾)a¾ for all n ∈ .
n

3.3.9. (−) Let cn be the number of ways that n children can be arranged in teams, with a
x
captain for each team chosen from the team members. Prove that the EGF for ⟨c⟩ is e xe .
(Stanley [1978] states this problem using idempotent functions.)
3.3.10. (−) Use the generating function method with EGFs to solve the recurrence an =
nan−1 + n! for n ≥ 1, with a0 = 1.
3.3.11. (−) Compute [x3 ](1 + x) by composing the series for e y
and ln(1 + x).
1 n
3.3.12. Let A , . . . , A and C be symmetric families of labeled structures such that ob-
jects in C[] correspond bijectively to distributions of the label set [¾] into n sets S1 , . . . , Sn
# #
and choices of elements from AiSi for 1 ≤ i ≤ ¾ . Let ai = ### A[i] ### and c = #### C[] ####. Prove that
# #
the EGF for ⟨c⟩ is the product of the EGFs for ⟨a1 ⟩ , . . . , ⟨an ⟩.
3.3.13. (♦) Let an be the number of words of length n from the alphabet {w, x , y , } such
that x appears an even number of times and y appears an odd number of times. Build the
EGF for ⟨a⟩ and use it to obtain a formula for an . Give a direct combinatorial argument
to explain the resulting formula.
n−
3.3.14. (♦) Evaluate ∑=0 (n)( m
m
− ) using binomial convolution, and give a bijective proof
of the resulting identity.
3.3.15. (♦) Let br,n be the number of ways to place n distinct flags on r distinct flagpoles,
each pole having at least one flag (order of flags on poles matters). For example, b2 ,4 = 72.
(a) Use generating function arguments to build the EGF Br(x) = ∑ br,n x n/n!.
(b) Use the EGF in part (a) to obtain a simple formula for br,n .
(c) Give a direct combinatorial proof of the formula in part (b).
Exercises for Section 3.3 133

3.3.16. (♦) Let an be the number of involutions on an n-element set. Derive a recurrence
for ⟨a⟩, and use it to obtain the EGF, which appears in Example 3.3.25.
3.3.17. (♦) A ranking of candidates in an election allows ties. Let an be the number of
rankings of n distinct candidates. Note a2 = 3 and a3 = 13.
(a) Obtain the EGF for ⟨a⟩.
(b) Using (a), prove that ∑¾≥0 ¾ n/2 ¾ is an integer (also in Exercise 3.1.38).

3.3.18. Use the Stirling numbers of the second kind to count the trees with vertex set [n]
that have exactly ¾ leaves. (Rényi [1959])
3.3.19. (♦) Prove that S(n − 1 , ¾ − 1) is the number of ways to partition [n] into ¾ sets with
no two consecutive values in the same set. (Hint: See Exercise 2.2.33.)
3.3.20. (♦) For 0 ≤ ¾ < n, give three proofs to evaluate ∑i=0(−1)i ( ni) i¾ .
n

(a) Use induction. (Hint: Use an identity to reduce the exponent on i.)
(b) Use OGFs.
(c) Use the Stirling numbers of the second kind.
(Comment: This yields ∑i=0(−1)i (ni) p(i) whenever p is a polynomial of degree less than n.)
n

3.3.21. Let P(m , n) be the set of nonnegative integer vectors a = (a1 , . . . , am+1) such that
∑ ai = n and ∑ iai = m + n.
(a) Find a direct combinatorial argument for
m+1
n 1
S(m + n , n) = ∑ (n + 1)(n + 2) · · · (n + m)(
a1 · · · am+1 ∏ i! ai
) .
a∈ P(m ,n) i= 2

m+1 ⌊m/i⌋
(b) Prove that (n + 1)(m) divides ∏ i= 2
i S(m + n , n). (Knuth [1993])
3.3.22. Count the partitions of proper subsets of [n] into ¾ blocks.
3.3.23. (♦) Identities for Stirling numbers.
(a) Prove bijectively that S(n + 1 , m + 1) = ∑¾ (¾n) S(¾ , m).
(b) Apply part (a) to prove ( mn
) = ∑¾ S(n + 1 , ¾ + 1)s(¾ , m).
3.3.24. (♦) Define the Bernoulli number Bn by B0 = 1 and ∑ j =0 (m+j 1 ) Bj = 0 for m ≥ 1.
m

(a) Prove that the EGF B(x) for the Bernoulli numbers is x/(e x − 1).
(b) Evaluate the sum of the (¾ − 1)th powers of the first n − 1 positive integers in terms
of the Bernoulli numbers by proving
n−1 ¾ −1
1 ¾
∑ m¾−1 = ¾ ∑ ( j ) Bj n¾− j .
m=1 j =0

3.3.25. (♦) Stirling numbers and Eulerian numbers.

¾
(a) Prove bijectively that ¾ !S(n , ¾) = ∑i=0 (n¾−−ii) A(n , i). (Riordan [1964])
(b) Use part (a) to prove ∑¾=0 ¾ !S(n , ¾)x n−¾ = ∑i=0 A(n , i)(x + 1)n−i . Conclude that
n n

there are ∑ j =1 2 A(n , j) ordered partitions of [n] (it is 3 for n = 2 and 13 for n = 3).
n j −1

3.3.26. Another inversion formula. Prove that the two statements below are equivalent.
(Recall that (u0) is the constant polynomial 1 as a polynomial in the real variable u.)
¾
(a) b¾ = ∑i=0 (¾n−−ii) ai for 0 ≤ ¾ ≤ n.
¾
(b) a¾ = ∑i=0(−1)¾−i (¾n−−ii)bi for 0 ≤ ¾ ≤ n.
¾
Conclude A(n , ¾) = ∑i=0(−1)¾−i (¾n−−ii)i!S(n , i) from (a) of Exercise 3.3.25. (Bóna [2004, p. 13])
134 Chapter 3: Generating Functions

3.3.27. Given n ≥ m ≥ 2, prove nm S(n , m) ≥ mn (mn ). (Pit é [2017], Nikšić [2017])

3.3.28. For positive integers m , n , N , give a bijective proof of
min(n ,N) min(m ,N)
N N
∑ ( ) j !S(n , j)(N − j)m =
j ∑ ( )i!S(m , i)(N − i)n .
i
(Khan [1991])
j =1 i =1

3.3.29. (♦) Define a sequence of polynomials by letting pn(x) = ∑ j =0 S(n , j)x j for n ∈ 0 .
n

(These have been called both the Touchard polynomials and the exponential polyno-
mials; note p0(x) = 1.) Give a combinatorial proof of the polynomial identity
n
n
pn(x + y) = ∑ ( ) p¾(x)pn−¾ (y).
¾
¾ =0

3.3.30. (♦) A Stirling permutation linearly orders two copies of [n] so that for all i, all
entries between the two copies of i exceed i. Let a skyline be a Stirling permutation having
the additional property that no strictly increasing triple has its last two entries consec-
utive in the arrangement. For example, 122133 is a skyline. Let an be the number of
skylines of length 2n, and let A(x) = ∑n≥0 an x n/n!.
(a) Prove A (x) = e2x A(x), and conclude A(x) = e(e −1)/2 .
2x

(b) Prove an = ∑¾=0 2 S(n , ¾) by using part (a).

n−¾
n

(c) Obtain part (b) by establishing a correspondence between skylines of length 2n and
partitions of [n] with some elements marked. (Callan [2011])
3.3.31. (♦) The Bell number Bn is the total number of partitions of [n].
(a) Prove Bn+1 = ∑¾=0 (¾n) B¾ for n ≥ 0.
n

(b) Use part (a) to prove that the EGF for the Bell numbers is ee −1 .
x

3.3.32. (♦) Let Bn be the nth Bell number.

∞
(a) Prove ∑¾=1 ¾¾ ! x¾ = e x ∑¾=1 S(n , ¾)x¾ .
n n

(b) Conclude Dobiński’s Formula: Bn = e−1 ∑¾≥1 ¾ n/¾ !. (Dobi ński [1877])
(c) Conclude also that the difference between the number of partitions with an even
and an odd number of blocks equals e ∑¾≥1 (−1)¾ ¾ n/¾ !.
n(1 −1/ln n)
3.3.33. Bell numbers. For ¾ ∈ [n], prove Bn ≥ ¾ n−¾ . Conclude ( lnnn ) ≤ Bn ≤ nn .
3.3.34. Let S be a set of n marbles, consisting of two indistinguishable white marbles and
one each in n − 2 other colors. Prove that the number of distinguishable partitions of S is
(Bn + Bn−1 + Bn−2)/2, where Bm is the number of partitions of [m]. (When n = 3, there are
four: WWB , WW | B , W | BW, W | W | B.) (G. Beck)
3.3.35. A principal submatrix is a submatrix obtained by extracting the same set of
columns as rows. A symmetric matrix is positive semidefinite if all its principal subma-
trices have nonnegative determinant.
(a) A partial partition of a set X is a partition of a nonempty subset of X . Show that
the number of partial partitions of [n] is Bn+1 − 1.
(b) Prove that the number of positive semidefinite 0 , 1-matrices of order n is the Bell
number Bn+1 . (Schmidt [1995])
3.3.36. (+) Use the formula for the derangement numbers to give a combinatorial proof
of the identity below.
n 2l n
n (−1)m n
∑(−1)l ( l )(2l)! ∑ m! = ∑(−1)l ( l )2n−l(n + l)! (Yu [1997])
l =0 m=0 l =0
Exercises for Section 3.3 135

3.3.37. (♦) Derivation of the Eulerian polynomial from its EGF. The nth Eulerian polyno-
mial A n , written with t as the formal variable, is defined by A n(t) = ∑¾=1 A(n , ¾)t¾ , where
n

A(n , ¾ is the number of permutations of

) [n] with ¾ runs.
n−1 ¾ −1
(a) Prove A(n , ¾) = A(n−1 , ¾−1)+ ∑ m=1 ∑ j =1 (nm ) A(m , j)A(n−1−m , ¾− j) for n , ¾ ≥ 1.
(b) Use part (a) to obtain a recurrence for the Eulerian polynomials.
(c) Let A(x) = ∑n≥0 A n(t)x n/n!. Use part (b) to prove A (x) = (t − 1)A(x) + A2(x).
(d) Use part (c) to prove A(x) = 1 − te1 −(1t− t)x .
(e) Use part (d) to prove A n(t) = (1 − t)n+1 ∑i≥0 in ti . (See Theorem 3.1.26.)

3.3.38. Let G(x) be the EGF for graphs, indexed by number of vertices. A graph is even if
x
all vertices have even degree. Prove that 1 + ∫ G(t)dt is the EGF for even graphs, indexed
0
x
by number of vertices, and that ln(1 + ∫ G(t)dt) is the EGF for connected even graphs.
0

3.3.39. Let bn be the number of unordered binary trees with n labeled leaves (interchang-
ing the left and right subtrees of a vertex does not change the tree). The non-leaf vertices
have no labels (when n = 1 the root is a leaf). Set b0 = 0. Let B(x) be the EGF for ⟨b⟩.
(a) Prove B(x) = x + 12 B(x)2 .
n−1
(b) Use part (a) to prove bn = ∏i=1 (2i − 1).
(Comment: Equality of bn with the number of pairings of [2n − 2] is requested bijectively
in Exercise 1.3.19.) (Schröder [1870]; see also Erdős–Székely [1989])
3.3.40. (♦) Let cn be the number of ways to arrange n children in circles (holding hands)
with one child standing alone inside each circle. Each circle has at least one child (not
counting the one inside). Prove that the EGF for ⟨c⟩ is (1 − x)− x . (Stanley [1978])
3.3.41. Let an be the number of distinct matrices expressible as the sum of an n-by-n per-
mutation matrix and its inverse. Prove ∑ an x n/n! = (1 − x)−1/2 e x/2+ x /4 . (Stanley [1978])
2

3.3.42. (♦) Let G(x) be the EGF for graphs where every vertex has degree 2, indexed by
number of vertices. Let F(x) be the EGF for permutations with no fixed points or 2-cycles.
(a) Use the Exponential Formula to prove

e− x/2− x /4 e− x− x /2
2 2

G(x) = √ and F(x) = .

1−x 1−x

(b) Give a bijective proof of F(x) = G(x)2 . (Wilf [1990, p. 77])

3.3.43. (♦) Let an be the number of permutations of [n] whose cycles all have odd length.
Let bn be the number of such permutations having an even number of cycles. Let A(x) and
B(x) denote the EGFs for ⟨a⟩ and ⟨b⟩.
1/2
(a) Use the Exponential Formula to prove A(x) = ( 11 −+ xx ) and B(x) = (1 − x2)−1/2 .
(b) Let p be the probability that a random permutation of [n] has an even number of
cycles, all of odd length. Let q be the probability that a string of n coin tosses has exactly
n/2 heads. By expanding B(x), prove the startling fact that p = q. (Wilf [1990, p. 75])
3.3.44. (♦) A 2-colored bipartite graph is a bipartite graph (Definition 5.1.9) with the
first and second parts named. Let F(x) and G(x) respectively be the EGFs for 2-colored bi-
partite graphs and for bipartite graphs (parts unnamed),√ indexed by number of vertices.
(a) Use the Exponential Formula to prove G(x) = F(x).
(b) Give a bijective proof of G(x)2 = F(x). (Hint: Group the components of a 2-colored
bipartite graph by which part contains the least vertex.) (Wilf)
3.3.45. Let an and bn be the numbers of derangements of [n] having an even number and
an odd number of cycles, respectively. Use the Exponential Formula to determine | an − bn| .
(Comment: Solved using recurrences in Exercise 2.1.42.) (Stathopoulos [2012])
136 Chapter 3: Generating Functions

3.3.46. A block is a connected graph such that the subgraph obtained by deleting any
one vertex is connected. The blocks of a connected graph are the maximal such subgraphs
(Chapter 7). A 1-vertex graph is not a block, since we take the graph with no vertices to be
disconnected. Let bn and cn count the blocks and the connected graphs with vertex set [n].
Let B(x) = ∑¾≥2 bn x n/n! and C(x) = ∑¾≥1 cn x n/n!. A rooted graph marks one vertex.
(a) Let R¾ (x) be the EGF for rooted connected graphs whose root appears in ¾ blocks.
Prove R¾ (x)/x = (R1 (x)/x)¾/¾ !. (Hint: For what is R¾(x)/x the EGF?)
(b) Let R(x) be the EGF for rooted connected graphs, so R(x) = ∑¾≥0 R¾ (x). Prove
R1 (x)/x = ∑¾≥2 b¾ R(x)¾−1/(¾ − 1)!.
(c) Conclude B (xC (x)) = ln(C (x)), where is differentiation. (Riddell [1951], Ford–
Uhlenbeck [1956]; see Harary–Palmer [1973, pp. 9–11], Stanley [1999, pp. 119–120].)

3.3.47. (♦) Fix ¾ ∈ , and let an be the number of ¾-ary ordered trees with n vertices; ¾-
ary means each vertex has 0 or ¾ children (see Definition 1.3.21).
(a) Let y(x) be the OGF for ⟨a⟩. Obtain the functional equation y = x(1 + y ¾).
(b) Use Lagrange Inversion to obtain a simple expression for an . (Hint: Note that an
is nonzero only for some n.) (see Stanley [1999, p. 175])
(c) Use part (b) to prove the formula for the Catalan numbers (yet again).
3.3.48. Let º be a formal Laurent series in y, and let ½ be a formal power series in x such
that [x0 ]½(x) = · · · = [x m−1 ]½(x) = 0 and [x m]½(x)
= 0. Generalize Lemma 3.3.33:
1 −1
[ y−1 ] º (y) = [ x ] º (½(x))½ (x).
m

3.3.49. (♦) Prove the general Lagrange Inversion Formula (Theorem 3.3.35): If á(y) and
h(y) are formal power series in y with á(0) = 1, and x = y/á(y), then h(y) expands as a
power series in x as given below. (Hint: Mimic the proof of Theorem 3.3.34.)
1 n−1
[x n]h(y(x)) = [y ]h (y)á(y)n .
n

3.3.50. Find the power series expansion of eay in powers of x, given y = xe y .

3.3.51. (♦) Generalization of Cayley’s Formula, again. Let an ,¾ be the number of rooted
forests with vertex set [n] and root set [¾]. Apply Theorem 3.3.35 to prove an ,¾ = ¾ nn−¾−1 .
(Hint: Express an ,¾ in terms of a coefficient in y ¾ , where y(x) is the EGF for rooted trees
(see Theorem 3.3.29).)
3.3.52. (♦) The Catalan recurrence yields the equation C(x) − 1 = xC(x)2 for the generat-
ing function C(x). Letting y = xC(x), use the general form of Lagrange Inversion (Theorem
3.3.35) to prove the identity below. (Comment: Setting ¾ = 1 yields another derivation of
the formula for Cn .)
¾ 2n + ¾
[ x n] C(x)¾ = ( )
2n + ¾ n
√
3.3.53. Apply Lagrange Inversion (Theorem 3.3.35) to the equation y = x 1 + y to prove
√ ¾ ∞
4 + x2 + x ¾ 12 (n + ¾) n
( ) =∑ ( )x .
2 n+ ¾ n
n=0

3.3.54. (+) Given an = ∑¾ ( n−¾ ¾)b¾ for n ≥ 0, use Theorem 3.3.35 for n ≥ 1 to prove

1 2n − ¾ − 1
) ¾ a¾ .
n∑
bn = (−1)n−¾ ( (Wilf [1990, pp. 140–141])
n−1
¾
Section 3.4: Partitions of Integers 137

3.4. Partitions of Integers

Corollary 1.1.22 counts lists of n positive integers with sum ¾ ; the order of
the integers matters. Sometimes, we do not care about the order.

3.4.1. Example. Distributions of bridge hands. A card deck has 13 cards each in
“spades”, “hearts”, “diamonds”, and “clubs”. A bridge hand consists of 13 cards.
Strategy depends heavily on the distribution of a player ’s cards among suits. A
hand having 5 spades, 4 hearts, 4 diamonds, and no clubs lies in an equivalence
class with one having 4 spades, no hearts, 5 diamonds, and 4 clubs. A “5440 distri-
bution” denotes a hand with five cards in one suit and four in each of two others.

3.4.2. Definition. A partition of an integer n is a multiset of positive inte-

gers with sum n. A composition of n places those integers in a fixed order.
The numbers used in a partition or composition are its parts (parts can be
repeated). Since the order of parts in a partition of n is irrelevant, we canon-
ically write them in nonincreasing order.

The possible bridge distributions are the partitions of the integer 13 into at
most four parts (one or more suits may have no cards). Exercise 6 compares the
probabilities of bridge distributions.

GENERATING FUNCTION METHODS

To enumerate partitions with specified properties, indexed by the sum, we

seek an OGF where the coefficient of x n is the number of such partitions of n
(the OGF is appropriate because we are partitioning identical units). One way
to specify a partition is by the number of times each integer is used as a part.
This suggests building the generating function by using a factor for each type of
number allowed to be a part.

3.4.3. Theorem. The OGFs for partitions using parts in {1 , . . . , ¾}, partitions
with largest part ¾ , and all partitions are, respectively,
¾ ¾ ∞
1 1 1
∏1−x , x¾ ∏ , ∏ 1 − xi .
i 1−x i
i=1 i=1 i=1

Proof: Let e i be the number of copies of i in a partition. Partitions of n using

parts in [¾] correspond to nonnegative integer solutions of 1e1 + 2e2 + · · · + ¾ e¾ =
n. The index in the OGF is the sum; when j copies of i are used, the contribution
to the exponent is j · i. Hence the factor listing the options for i is 1 + x i + x2i + · · · ,
which equals (1 − x i)−1 .
Deleting x0 from the factor for copies of ¾ ensures that at least one ¾ is used.
The infinite product allows all positive integers as parts.
138 Chapter 3: Generating Functions

The third generating function of Theorem 3.4.3 is the “master ” generating

function for partitions; it includes all partitions we might want to count. Re-
strictions on usage of particular parts can be incorporated into the correspond-
ing factors. For example, the generating function for partitions of n in which all
parts are odd, 1 is used at least three times, and 3 is used at most three times is

x3(1 − x)−1(1 + x3 + x6 + x9)(1 − x5)−1 (1 − x7)−1 · · · .

Generating function methods can sometimes be used to find the asymptotic

behavior of a sequence of coefficients. Let an ,¾ be the number of partitions of n
using parts in [¾]. For small ¾ , we can extract the exact formula for an ,¾ . Note
that an ,1 = 1 and an ,2 = 1 + ⌊ n/2⌋ .

3.4.4. Proposition. The number of partitions of n using parts in {1 , 2 , 3} is the

1
nearest integer to 12 (n + 3)2 .
Proof: By Theorem 3.4.3, the generating function for partitions using parts in
[3] is (1 − x)−1 (1 − x2)−1(1 − x3)−1 . Note that

3
1 1 1 1 A
∏ 1 − x i = (1 − x)3 1 + x 1 + x + x2 = (1 − x)3 + other terms.
i=1

Multiplying by the denominators leaves at least one factor of (1 − x) in every

term except the first. Hence setting x = 1 (Heaviside method, Example 2.2.19)
yields 1 = 6 A, so A = 1/6. Since [x n](1 − x)−3 = (n+2 2), the leading behavior will
be n2/12 if the contributions from the other terms grow at most linearly in n.
To be more precise, we factor 1 + x + x2 over the complex numbers. With
= e2 i/3 , the denominator is (1 − x)3(1 + x)(1 − x)(1 − 2 x) in linear factors.
Expanding by partial fractions yields
3
1 1/6 1/4 17/72 1/8 1/9 1/9
∏ 1 − x i = (1 − x)3 + (1 − x)2 + 1 − x + 1 + x + 1 − x + 1 − 2 x .
i=1

Summing the coefficients of x n in these terms yields

1 n+2 1 17 (−1)n 1
an ,3 = ( ) + (n + 1) + + + ( n
+ ).
2n
6 2 4 72 8 9

Note that 16 (n+2 2) + 14 (n + 1) = 12

1
(n + 3)2 − 13 . Moving 12
1
(n + 3)2 to the left side and
applying the triangle inequality on the right yields, as desired,
### ##
### an ,3 − 1 (n + 3)2 #### ≤ 7 + 1 + 2 < 1 .
### 12 ### 72 8 9 2

Similar reasoning applies to an , for any fixed  . The dominant contribution

comes from (1 − x)− . We obtain the leading coefficient.
Section 3.4: Partitions of Integers 139

3.4.5. Theorem. The number of partitions of n using parts in [¾] is asymptotic

to ¾ !(¾1−1)! n¾−1 .
Proof: The generating function is the first formula in Theorem 3.4.3. Since
1 − xr = (1 − x)(1 + x + · · · xr−1), we have
¾
1 1 1 1 1
∏ 1 − xr = (1 − x)¾ 1+x 1+x+x 2
···
1 + x + · · · + x ¾ −1
.
i=1

To obtain the partial fraction expansion and complete the analysis, we must fur-
ther factor the denominator into complex linear factors, as in Proposition 3.4.4.
Note the use there of the primitive cube roots of unity. In general, 1 − xr is the
product of 1 − x over all primitive dth roots of unity over all d that divide r. (For
¾
example, 1 − x4 = (1 − x)(1 + x)(1 − ix)(1 + ix).) When we factor ∏r=1 (1 − xr), we
obtain 1 − x as a factor ⌊ /d⌋ times if is a primitive dth root.
In the partial fraction expansion, we therefore have constant multiples of
(1− x) j for 1 ≤ j ≤ /d⌋ . The contribution of a fixed dth-root
1
⌊ to the coefficient
of x in the expansion of the generating function is thus p (n) n , where p is a
n

polynomial of degree at most ⌊ /d⌋ − 1.

The only one of these polynomials having a term with degree as large as
 − 1 is the one associated with (1 − x) . To find its coefficient, we multiply

by ∏r=1 (1 − xr) (the product of the denominators) and set x = 1 (again using
the Heaviside method). For this operation we use the initial factorization given
above, where setting x = 1 yields 1 = A ·  !, so A = 1/ !. We obtain
A 1 n+  −1 1
[x n] = ( )∼ n−1 .
(1 − x) !  − 1  !( − 1)!
Furthermore, the magnitude of each root is 1. Hence the magnitude of the
leading behavior of the coefficient of x n is bounded by  !(1−1)! n−1 plus the sum of
polynomials in n with degree less than  − 1. We conclude an , ∼ 1  −1 .
 !( −1)! n

The asymptotic behavior of the total number of partitions of n is more com-

plicated. We will obtain an upper bound. In order to judge how good it is, we
first build many partitions. Let p(n) denote the total number of partitions of n,
so ∑n≥0 p(n)x n = ∏i≥1 (1 − x i)−1 .

3.4.6. Remark. A lower bound. For fixed  , the number of partitions of n us-
ing parts in [] is asymptotic to n−1/( !( − 1)!) as n → ∞. When  grows
√ slowly
enough with n, the formula should still
√ √
be fairly
√
accurate. With  = n,
√
by Stir-
ling ’s Formula it becomes roughly n n/( n/e)2 n , which simplifies to e2 n .
This lower bound seems reasonable; not many partitions can have very large
parts. Partitions of n with largest part  correspond to partitions of n −  into
parts no bigger than  . When  is large,
√ p(n − ) is much smaller than p(n), but
there are many terms between  = n and  = n, so the outcome is unclear.

The
√ intuition of Remark 3.4.6 is not bad; the actual upper bound is less than
2.565 n
e , so the order of growth of the logarithm is correct. For asymptotic anal-
ysis, we treat the generating function as a power series and worry about conver-
gence. The short proof of this upper bound is due to van Lint [1974].
140 Chapter 3: Generating Functions

3.4.7. Theorem. The number p(n) of partitions of n satisfies

√
p(n + 1) < √ e2 n/6
.
6n

Proof: Let P(x) = ∑n≥0 p(n)x n = ∏≥1 (1 − x)−1 . For 0 < x < 1, we obtain a
numerical bound on ln P(x) and then use it to bound ln p(n + 1). The idea behind
this comparison is the following, using that ⟨p⟩ is increasing:
x n+1
P(x) = ∑ p()x > ∑ p()x > p(n + 1) ∑ x = p(n + 1) . (∗)
1−x
 ≥0 >n >n

Since P(x) is a product, it is convenient to study ln P(x). The inner sum below
is bounded by an absolutely convergent geometric series, so we can interchange
the order of summation and evaluate the new inner geometric series ∑≥1 (x j ) .
∞ ∞ ∞ ∞
x j 1 xj
ln P(x) = − ∑ ln(1 − x) = ∑ ∑ =∑ .
j j 1 − xj
 =1  =1 j =1 j =1

Next we bound 1−x x j . Since jx j −1 < 1 + x + · · · + x j −1 = 11−−xx when 0 < x < 1,

j j

xj x jx j −1 1 x
= < .
1 − xj j 1 − xj j1−x
∞
Since 1−x x is independent of j , we have ln P(x) < 1−x x ∑ j =1 j −2 = 6 1−x x .
2

Now combine (∗) with the bound on ln P(x) (also set u = 1−x x , so x−1 = 1 + u):
P(x)(1 − x) 2
x 1 1−x
ln p(n + 1) < ln < · + n ln + ln
xn · x 6 1−x x x
2
1
< · + nu + ln u.
6 u
√
Here we have also used ln(1 + u)√< u. Choosing u = / 6n to make the bound
small, we obtain ln p(n + 1) < 2 n/6 + ln u, as desired.

The upper bound in Theorem 3.4.7 is within a constant

√ factor
√ of the true
behavior. Hardy–Ramanujan [1918] showed p(n) ∼ (4n 3)−1 e2 n/6 . Rademacher
[1937] gave more detail; see Chandrasekharan [1970] or Andrews [1976, Chap. 5].

3.4.8.* Remark. Analytic methods. Various methods from complex analysis shed
light on the coefficients. If a formal power series A(x) converges absolutely when
the argument is set to some complex number in a neighborhood of the origin,
then the function is analytic in that neighborhood. Now [x n] A(x) is the residue
of A(x)/x n+1 , and the Cauchy Integral Formula yields an = 21 i ∫ C xA(x)
n+1 dx , where

C is a simple closed curve around the origin. The partition function is analytic
for | x| < 1 and has singularities at every complex root of unity on the unit cir-
cle. This leads to the asymptotic analysis of p(n). These notions are explored in
Andrews [1976, Chapter 6] and Wilf [1990, Chapter 5].
Section 3.4: Partitions of Integers 141

FERRERS DIAGRAMS

We have an OGF for the partitions of n with largest part ¾ . What about the
partitions of n with ¾ parts? In fact, these sets have the same size. This and
many combinatorial identities arise from a geometric view of partitions.

3.4.9. Definition. The Ferrers diagram of a partition ë of n is an array of n

dots, with ë i dots in row i, where ë i is the ith largest part. The rows are
left-justified, each at least as long as the row below it.
• • • • •
• • •
•
•

When drawn with boxes instead of dots, Ferrers diagrams are called Young
diagrams. We can cut and reassemble a diagram to turn a partition of one type
into a partition of another type. When this yields a bijection, the two types of
partitions are equinumerous. For example, we can count dots by columns instead
of by rows. This maps the partition 5,3,1,1 shown above into 4,2,2,1,1.

3.4.10. Definition. The conjugate ë∗ of a partition ë is the partition whose

Ferrers diagram is the transpose of the Ferrers diagram of ë.

3.4.11. Proposition. The number of partitions of n with largest part ¾ equals

the number of partitions of n into ¾ parts.
Proof: Taking the conjugate converts every partition with ¾ parts into a parti-
tion with largest part ¾ , and conjugation is a bijection.

By Proposition 3.4.11, the OGFs for partitions with largest part ¾ and par-
titions into ¾ parts are equal. Similarly, partitions with parts in [¾] correspond
to partitions with at most ¾ parts.
Euler [1748] proved the following identity algebraically (Exercise 13):
∞ ∞

∏(1 + xi) = ∏(1 − x2i−1)−1 .

i=1 i=1

The left side is the OGF for partitions into distinct parts, and the right side is
the OGF for partitions into odd parts. The natural bijective proof presented next
shows that corresponding coefficients are equal. See Sylvester–Franklin [1882],
Pak [2006], and Exercises 14–15 for extensions.

3.4.12. Proposition. The number of partitions of n into distinct parts equals

the number of partitions of n into odd parts.
142 Chapter 3: Generating Functions

Proof: (Glaisher [1883]) Map an odd-parts partition into a distinct-parts parti-

tion by iteratively combining two identical parts until no identical parts remain.
Each part in the resulting partition is a power of 2 times an odd number, say
2¾ (2 j + 1), obtained by combining 2¾ copies of 2 j + 1 from the original odd-parts
partition. Since every positive integer is expressible as 2¾ (2 j + 1) in a unique
way, for each distinct-parts partition there is exactly one odd-parts partition that
maps to it in this way. Hence the map is a bijection.

Next consider congruence classes of integer triangles; triangles whose sides

have integer lengths. Two triangles are congruent if their side-lengths form the
same integer partition. For perimeter m, we partition m into three parts a , b , c
with a ≥ b ≥ c, but they must satisfy the strict triangle inequality b + c > a.
The methods we have discussed do not constrain one part in terms of oth-
ers, so the triangle inequality is hard to enforce directly. Exercise 18 develops
an OGF for the congruence classes of triangles, indexed by perimeter. Although
somewhat longer than the proof here, that approach does not require prior knowl-
edge of the answer. Here we use Ferrers diagrams to give a short proof.

3.4.13. Theorem. The OGF for congruence classes of triangles with integer-
length sides, indexed by perimeter, is
x3
.
(1 − x2)(1 − x3)(1 − x4)
Proof: We show that 3-part partitions satisfying the strict triangle inequality
correspond to partitions with parts in {2 , 3 , 4} and at least one 3.
For a partition of the first type, the triangle inequality ensures that the first
row of the diagram exceeds the second row by less than the length of the third.
Therefore, moving the excess to a fourth row yields a new Ferrers diagram. The
new fourth row is strictly shorter than the third, so the conjugate partition has
a 3. Since we moved all the excess, the conjugate has no 1; it uses only parts in
{2 , 3 , 4}, with at least one 3.
3
7 • • • • • • •
5 • • • • •
3 • • •

Given a partition with conjugate of the second type, this process reverses to
yield the unique partition of the first type that maps to it. Since the conjugate
has at least one 3, the excess of the largest resulting part over the second part is
less than the third part, and hence the three parts satisfy the triangle inequality.
Thus we have a bijection, and the two generating functions are the same.

Many other bijections for partitions can be found in the survey by Pak [2006];
see also Sylvester–Franklin [1882]. We close with a classical such argument to
prove another famous identity that Euler proved algebraically.
Section 3.4: Partitions of Integers 143

3.4.14. Theorem. (Euler ’s Identity)

∞ ∞
x¾
2
2i−1
∏(1 + x )=1+∑
(1 − x2)(1 − x4) · · · (1 − x2¾)
i=1 ¾ =1

Proof: (possibly due to Durfee) We prove that both expressions enumerate self-
conjugate partitions (those unchanged by transposing the diagram). To do this,
we cut the diagram of a self-conjugate partition in two ways.
The coefficient of x n in ∏(1 + x2i−1) counts partitions of n into distinct odd
parts. In a self-conjugate partition, grouping dots by the smaller coordinate cuts
the Ferrers diagram into L-shapes as on the left below. Corresponding rows and
columns have the same length and share a diagonal dot, so each piece has odd size.
Also, the sizes are distinct. Conversely, a partition into distinct odd parts yields
the self-conjugate partition that generates it by bending each part in the middle.
Hence ∏(1 + x 2i−1) is the OGF for self-conjugate partitions.

• • • • • • • • • • • • • • • •
• • • • • • • • • • • •
• • • • • • • • • • • •
• • • • • • • • • •
• • • • • • • •
• • • • • •
• •
• •

Each self-conjugate partition has some number ¾ of dots on the diagonal. The
largest square in the diagram (the “Durfee square”) has ¾ 2 dots. In the example
above, ¾ = 4. The number of dots at any distance to the right of the Durfee square
is the same as the number below it, so grouping the remaining dots by distance
from the square partitions them into even parts. These parts do not reach the
diagonal, so they are at most 2 ¾ . Thus the OGF for self-conjugate partitions with
Durfee square of size ¾ is the term for ¾ on the right side of the identity (¾ = 0
yields the one partition of 0). Summing over ¾ completes the proof.

BULG ARIAN SOLITAIRE (optional)

Ferrers diagrams facilitate solution of a curious problem that occurred to

many people independently over the years. Martin Gardner [1983] gave the prob-
lem its name in his column in Scientific American.

3.4.15. Problem. (Bulgarian Solitaire) Given n identical coins distributed into

piles, a move takes one coin from each pile and puts them into a new pile. What
happens under repeated moves?
Gardner phrased the problem using cards (hence “solitaire”), with n = 45.
Since the coins are identical and the piles are not distinguished, each configu-
ration is a partition of n, and the specified move defines a function B (for “Bul-
144 Chapter 3: Generating Functions

garian”) from the set of partitions of n to itself. Since the set is finite, repeated
moves must eventually reach a cycle.
When n is a triangular number (having the form (¾+2 1 )), it is easy to see that
the partition ¾ , ¾ − 1 , . . . , 1 is fixed under B, and such partitions are the only
fixed points (Exercise 48). More surprising is that when n = (¾+2 1 ) this is the only
cycle. Brandt [1982] proved this and described all the cycles for all values of n (we
will count them in Application 4.2.22). Akin–Davis [1985] discussed the proof in
terms of dynamical systems.
Similar processes studied in Akin–Davis [1985], Cannings–Haigh [1992],
Yeh [1995], Servedio–Yeh [1995], Broline–Loeb [1995], and Griggs–Ho [1998]
were called Austrian solitaire, Montreal solitaire, Carolina solitaire, etc.

3.4.16.* Remark. Let d(ë) be the number of iterations of B needed to reach a

partition on the final cycle when starting from ë. Brandt [1982] conjectured that
d(ë) ≤ ¾(¾ − 1) when n = (¾+2 1 ). This was proved by Igusa [1985], Bentz [1987],
and Etienne [1991] (the latter submitted in 1984), and the bound is sharp (Ex-
ercise 50). Igusa and Etienne proved more generally that d(ë) ≤ ¾(¾ − 1) when
n = (¾2) + r with 0 < r ≤ ¾ .
Griggs–Ho [1998] proved d(ë) ≤ (¾ − 1)2 when (¾2) < n < (¾+2 1 ), which is sharp
when n = (¾+2 1 ) − 1 with ¾ ≥ 4. They also conjectured the form of partitions max-
imizing d(ë) over partitions of n and found necessary conditions (not sufficient)
for partitions achieving the maximum. Partitions achieving the maximum are
not unique, even when n = (¾+2 1 ).

Study of B is aided by Ferrers diagrams. Given the diagram T for a parti-

tion ë, the diagram for B(ë) is obtained by removing the first column of T and
inserting it instead as a new row.

3.4.17. Example. The key to understanding B is to form the diagram for B(ë) in
another way. A slant in the diagram is a set of positions (i , j) having the same
value of i + j (the top left position is (1 , 1)). The ¾ th slant, counting from the
upper left, has ¾ possible positions.
Consider B(7 , 6 , 3 , 3) = (6 , 5 , 4 , 2 , 2). When T has s rows, B(T) gains s as a
part. We can produce the new diagram by putting the dots taken from the start
of each row as a new top row. We then move each subsequent column up by one
row (see the middle diagram below), and s appears as a part in the correct position
(indicated by the circled dot). Since dots moved from the left to the top, all dots
in the first s columns that were in the ¾ th slant remain in the ¾ th slant, and dots
in the later columns move to the (¾ − 1)th slant. One can equivalently say that
every slant rotates down by one position, except that when this leaves dots with
a gap above them they move up into the previous slant.

• • • •
• • • • • • • • • • • • • • • • • • •
• • • • • • • • • • • • • • • •
• • • → • •  • • • •
• • • • • • •
• •
Section 3.4: Partitions of Integers 145

The ¾ th slant is complete if it contains ¾ dots, incomplete if it is nonempty

but has fewer than ¾ dots.

3.4.18. Theorem. (Brandt [1982]) A partition is in a cycle under Bulgarian soli-

taire if and only if its Ferrers diagram has at most one incomplete slant.
Proof: (Griggs–Ho [1998]) Under the description of B in Example 3.4.17, com-
plete slants are preserved (dots rotate by one position). When T has at most one
incomplete slant, B(T) is obtained by one cyclic shift on the incomplete slant. No
dots move upward, because the first column becomes the full top row.
For the converse, dots remain on the same slants except for those in the right-
most columns (as described in Example 3.4.17), which move to a shorter slant.
Dots never move to a longer slant. Hence for a partition on a cycle, iteration of B
can never move a dot to a shorter slant.
If T has more than one incomplete slant, then there is a value r such that T
has a gap in slant r and a dot in slant r + 1. If T is on a cycle, then these slants
shift cyclically with each application of B. However, since the sizes of slant r and
slant r + 1 are relatively prime, the gap in slant r catches up to the dot in slant
r + 1, reaching the same column. This array is not a Ferrers diagram, and hence
such a diagram cannot be on a cycle under B.

DISTRIBUTION MODELS (summary)

Many of the classical counting problems can be phrased in the context of dis-
tributions of objects into boxes. There are conditions on the objects, the boxes,
and the allowed distributions. We use a uniform notation with ¾ being the num-
ber of objects and n being the number of boxes.
A distribution of ¾ distinct objects into n distinct boxes is a function º : [¾] →
[n], assigning a box to each object. We use all ¾ objects, but boxes may be empty.
It suffices to list º (1) , . . . , º (¾) in order. Thus the distributions correspond to
words of length ¾ from n letters.
When the order of letters is unimportant, words using the same multiset of
letters are equivalent, and we count the equivalence classes (the multisets). The
word model is now less natural than the distribution model using ¾ indistinguish-
able (or “identical”) objects.
Making boxes indistinguishable yields another equivalence relation: a distri-
bution groups the objects without labels on the groups. This yields partitions of
the set [¾] into (at most) n blocks. When the objects also are identical, a distri-
bution just groups ¾ identical dots and forms a partition of the integer ¾ into (at
most) n parts. The number of blocks or parts is exactly n when we restrict to dis-
tributions that use all the boxes. We can also restrict distributions by allowing
only one object in each box.
We obtain twelve distribution problems, in an array called “ The Twelvefold
Way”. Indexing the rows, distributions may be Unrestricted, Injective (at most
one per box), or Surjective (no box empty). Indexing the columns, objects may be
Distinct or Identical, and boxes may be Distinct or Indistinguishable, Stanley
[1986] publicized the table, attributing the idea to G.-C. Rota and the term to
Joel Spencer, suggested by the Eightfold Way of Buddhism. Where possible, we
146 Chapter 3: Generating Functions

give a classical model, a formula, a generating function, and a recurrence (several

of the recurrences appear in Example 2.1.9). We use subscripts (for example, S¾ ,n
instead of S(¾ , n)) for clarity. We break the table into two parts.

¾ Distinct Objects ¾ Identical Objects

n Distinct Boxes n Distinct Boxes

Words of length ¾ Multisets of size ¾

from n types (repetition okay) from n types (repetition okay)
P n ,¾ = n¾ P n ,¾ = nP n ,¾−1 C n ,¾ = (¾+n−n−1 1 ) = (−1)¾ (−¾n)
U ∑¾≥0 P n ,¾ x¾ = 1−1nx
∑¾≥0 C n ,¾ x¾ = ( 1−1 x )n
∑¾≥0 P n ,¾ x¾/¾ ! = enx
P n ,¾ = P n−1 ,¾ + ∑i (ni) P n−1 ,¾−i C n ,¾ = C n−1 ,¾ + C n ,¾−1

Simple words of length ¾ Subsets of size ¾

from n types (no repetition) from n elements (no repetition)
I P n ,¾ = n!
(n−¾)! Cn ,¾ = (¾n) = n!
¾ !(n−¾)!

≤1 ∑¾≥0 Pn ,¾ x¾/¾ ! = (1 + x)n ∑¾≥0 Cn ,¾ x¾ = (1 + x)n

Pn ,¾ = Pn−1 ,¾ + ¾ Pn−1 ,¾−1 Cn ,¾ = Cn−1 ,¾ + Cn−1 ,¾−1
Pn ,¾ = nPn−1 ,¾−1

Partitions of ¾ objects Compositions of integer ¾

into n labeled blocks into n positive parts
S¾ ,n = n!S¾ ,n (solns to ∑i=1 ei = ¾ ; ei ≥ 1)
n
S
≥1 S¾ ,n = ∑i=0(−1)i ( ni)(n − i)¾ qn ,¾ = (¾n−−11 )
n

∑¾≥0 S¾ ,n x¾/¾ ! = (ex − 1)n ∑¾≥0 qn ,¾ x¾ = ( 1−x x )n

S¾ ,n = n(S¾−1 ,n−1 + S¾−1 ,n) qn ,¾ = qn ,¾−1 + qn−1 ,¾−1

References

UDD: Proposition 1.1.13, UID: Theorem 1.1.20

Example 3.3.8 Example 3.1.8, Theorem 3.1.10
IDD: Proposition 1.1.16 IID: Proposition 1.1.17
SDD: Theorem 3.3.11 SID: Corollary 1.1.22

In the second part of the table (below), our model here of distributing ¾ ob-
jects into n boxes exchanges the meanings of n and ¾ that were used in the text
for surjective distributions or partitions of integers. For Stirling numbers of the
second kind, we thus write S¾ ,n here; written in second position, n is the num-
ber of blocks. For partitions of an integer with a restricted number of parts, the
first argument is always the integer being partitioned.
Section 3.4: Partitions of Integers 147

¾ Distinct Objects ¾ Identical Objects

n Indistinguishable Boxes n Indistinguishable Boxes

Partitions of ¾-set Partitions of integer ¾

into at most n blocks into at most n parts
∑i=0 S¾ ,i p¾ ,n = ∑i=0 p¾ ,i
n n

U For n = ¾ , is the Bell number B¾

∑¾≥0 B¾ x¾/¾ ! = ee − 1
x 1
∑¾≥0 p¾ ,n x¾ = Πin=1 (1 − xi)
¾ −1
B¾ = ∑i=0 (¾−i 1) Bi p¾ ,n = p¾− n ,n + p¾ ,n−1

I 1 way if ¾ ≤ n 1 way if ¾ ≤ n
≤1 0 ways if ¾ > n 0 ways if ¾ > n

Partitions of ¾-set Partitions of integer ¾

into n blocks into n parts
S¾ ,n = ∑i=0(−1)i (ni)(n − i)¾
1 n
S n!

≥1 ∑¾≥0 S¾ ,n x¾/¾ ! = 1
n!
(e x − 1)n
xn
x n
∑¾≥0 p¾ ,n x¾ = Πin=1 (1 − xi)
∑¾≥0 S¾ ,n x¾ = Πn (1 − ix)
i =1
S¾ ,n = S¾−1 ,n−1 + nS¾−1 ,n p¾ ,n = p¾− n ,n + p¾−1 ,n−1

References
UDI: Exercise 3.3.31 UII Theorem 3.4.3
IDI: trivial III: trivial
SDI: Theorem 3.3.11 SII: Theorem 3.4.3
Bogart [1990] expanded the table to 18 cells by also considering the order of
distinct elements within a box. Arrangements of ¾ flags on n flagpoles, counted
by the rising factorial n(¾) , are distributions of ¾ distinct objects into n distinct
boxes (cell UDD), where in addition order of objects within the boxes matters.
Similarly, permutations of [¾] with n cycles arise by putting ¾ distinct ob-
jects surjectively into n identical boxes (cell SDI) and then cyclically arranging
within each box. Hence c(¾ , n) is not in our table.
Permutations of [¾] with n runs (Theorem 3.1.26) can be viewed as surjec-
tive distributions of ¾ objects into n boxes, but the largest object in one box must
exceed the smallest object in the next. When constraints involve more than one
box, the distribution language is not very useful.
Proctor [2006] suggested increasing the number of cells to 24 and then 30.
His first augmentation allowed an arbitrary number of boxes, thus introducing
the Bell numbers and the total number of partitions of an integer. His other
augmentation, suggested by Brylawski, specified multiplicities within boxes, in-
troducing multinomial coefficients.
A summary of Proctor ’s table with links to the relevant counting sequences
appears in the index of Sloane ’s On-Line Encyclopedia of Integer Sequences at
http://oeis.org. We consider only the Twelvefold Way because we seek only an
overview of basic results to show some of the relationships among them.
148 Chapter 3: Generating Functions

EXERCISES 3.4

3.4.1. (−) Count the partitions of 30 into 1s and 3s using an odd number of 3s. Express
the answer using generating functions. Verify the answer by describing these partitions
explicitly.
3.4.2. (−) In how many ways can a roll of five ordinary six-sided dice of different colors
sum to 20? How many distinguishable ways if the dice are identical? (It suffices to give
answers using generating functions.)
3.4.3. (−) Build a generating function to count the positive integer solutions to ∑i=1 ei = ¾
n

such that e1 ≥ · · · ≥ e n . (Hint: Choose the index appropriately.)

3.4.4. (−) Use generating functions to show that every nonnegative integer has a unique
binary expansion.
3.4.5. (−) Prove that the number of partitions of n + 1 having no 1 is at most the number
of partitions of n in which the number of copies of 1 is positive and is less than the size of
the smallest other part (if such a part exists). For example, when n = 3 the partitions of
the first type are 4 and 22, while those of the second type are 111 and 21. (Merca [2014b])
3.4.6. (−) List the possible distributions of bridge hands with no void suit (Example 3.4.1;
a void is a suit with no cards). Which distribution is more likely, 4333 or 5431?
3.4.7. (♦) Let an count the distinguishable ways to toss n indistinguishable dice and obtain
an even sum. Prove ∑ an x n = 12 (1 − x)−3 [(1 − x)−3 + (1 + x)−3 ].
3.4.8. (♦) The left and right pillars of an arch must have the same total height. They will
be built using blocks of height 1 or 2, but blocks of height 2 may not sit on blocks of height
1. Let an be the number of ways to do this using a total of n blocks. Note that a0 = 1,
a1 = 0, a2 = 2, a3 = 2, etc.
(a) Build the OGF for ⟨a⟩ by showing that an equals the number of partitions of n into
red 2s, blue 2s, red 3s, and blue 3s that use at most one color of 3s.
(b) Determine the asymptotic behavior of an . (Beckwith [2005])
3.4.9. Let pn ,¾ be the number of partitions of n into ¾ parts. Give bijective proofs for the
following facts.
(a) p2r+¾ ,r+¾ is independent of ¾ .
(b) pr+¾ ,¾ counts the partitions of r into parts of size ≤ ¾ .
(c) pr+¾ ,¾ counts the partitions of r + ¾(¾ + 1)/2 into ¾ distinct parts.
3.4.10. Let pn ,¾ be the number of partitions of n into ¾ parts, and A¾(x) = ∑ n pn ,¾ x n .
(a) Prove bijectively that pn ,¾ = pn−1 ,¾−1 + pn−¾ ,¾ . When is the equation valid?
(b) Obtain a recurrence for A¾ . Solve it and explain the result combinatorially.
3.4.11. Let pj ,¾(n) be the number of partitions of n that have j parts and have ¾ as the
∞
largest part. Compute ∑n=0 pj ,¾(n).
3.4.12. Prove that the number of partitions (of any integer) whose Ferrers diagram fits in
an i-by-(n − i) rectangle is one more than the number of permutations of [n] having exactly
one descent , at position i.
3.4.13. Prove Proposition 3.4.12 by direct algebraic manipulation of the generating func-
∞ ∞
tions. That is, give an algebraic proof of ∏i=1 (1 + xi) = ∏i=1 (1 − x2i−1)−1 . (Euler [1748])

3.4.14. (♦) Generalization of Proposition 3.4.12. For n , ¾ ∈ 0 , use generating functions to
prove that the number of partitions of n having ¾ even parts is the same as the number
of partitions of n in which ¾ is the largest repeated part (¾ = 0 when there is no repeated
part). (Comment: Proposition 3.4.12 is the case ¾ = 0.) (Andrews–Deutsch [2016])
Exercises for Section 3.4 149

3.4.15. (♦) Further generalization of Proposition 3.4.12.

(a) (Glaisher ’s Theorem) Prove for d ∈ 
that the number of partitions of n with
no part divisible by d equals the number of partitions of n with no part occuring at least
d times. (Comment: Proposition 3.4.12 is the case d = 2.) (Glaisher [1883])
(b) Prove that the number of partitions of n having exactly ¾ parts divisible by d is
the same as the number of partitions of n in which ¾ is the largest part that occurs at
least d times. For example, 6 has one partition with three even parts (2 + 2 + 2) and one
in which the largest repeated part is 3 (3 + 3). (Comment: Setting ¾ = 0 yields part (a);
setting d = 2 yields Exercise 3.4.14.) (Smoot [2018])
3.4.16. Prove that for n ≥ 2, exactly half of the partitions of n into powers of 2 have an
even number of parts. For example, when n = 2 the partitions are 2 and 11, when n = 3
they are 21 and 111, and when n = 4 they are 4, 22, 211, 1111.
3.4.17. Let an be the number of partitions of n in which each part is at least as big as
the sum of all subsequent parts. Let bn be the number of partitions of n into powers of 2.
Prove an = bn for n ≥ 1. (Beckwith [2009])

3.4.18. (♦) For n ∈ 0 , let an be the number of congruence classes of triangles with integer-
length sides and perimeter n. Equivalently, an is the number of partitions of n into three
parts less than n/2. This reproves Theorem 3.4.13, which is forbidden from use here.

(a) For ¾ ∈ 0 , prove that a2¾ counts the partitions of ¾ into three parts.
(b) For ¾ ≥ 2, prove a2¾−3 = a2¾ .
(c) Use parts (a) and (b) and manipulation of generating functions to prove that the
OGF for ⟨a⟩ is x3/[(1 − x2)(1 − x3)(1 − x4)].
3.4.19. (♦) Let an be the number of partitions of n into three parts less than n/2, corre-
sponding to congruence classes of integer triangles.
(a) For n ≥ 3, prove
0 if n is even,
an = an−3 + {
⌊ n+4 1 ⌋ if n is odd.
√
(b) For n ≥ 3, prove an = an−3 + 1 −(−8 1) (n + in+1 ), where i = −1.
n

(c) Obtain the generating function for ⟨a⟩ from the recurrence.

3.4.20. With an as in Exercise 3.4.19, prove an = an−2 + ⌊ n/3⌋ − ⌊ n/4⌋ for n ≥ 2. Find the
generating function ∑n≥0 ⌊ n/¾ ⌋ x n , and use it to find the generating function ∑ n≥0 an x n .

3.4.21. Prove the identity below. (MacMahon [1916], Beckwith [2000])

∞ ∞
x n(n+1)/2 xn
∑ 1 − xn
=∑
1 − x2n
n=1 n=1

∞ ∞
3.4.22. Let A(x) = ∑n=0 an x n/n! and B(x) = ∑n=1 bn x n/n!. Given a partition ë of the
integer n, let ei be the number of parts equal to i in ë. Prove
∞
n! ∏ bë j ⎞ x n
n
⎛
A(B(x)) = a0 + ∑ ∑ a¾ · ∑ ,
n=1
⎝ ¾ =1 ∏i(i!)ei (ei !) ⎠ n!
where the inner sum is over partitions ë of n with ¾ parts.
3.4.23. (♦) Let an be the total number of 2s over all partitions of n. Let bn be the total
number of nonrepeated parts over all partitions of n − 1. For n ≥ 1, both sequences begin
(0 , 1 , 1 , 3 , 4 , . . .). Prove an = bn for all n by showing that for each sequence the OGF is
∞
1 − x 2 ∏ i =1 1 − x i
x2 1
. (Deutsch [2006])
150 Chapter 3: Generating Functions

∞ ¾ + 1 −1 ∞
3.4.24. (♦) Use partitions to prove ∑¾=0 x2 ∏ j =¾+1 (1 + x2 −1) =
j x
1− x
. (GCHQ Problem
Solving Group [2018])
3.4.25. (♦) Prove ∑ ∏ j ≥1 ( ëë j ) = 2 n−1 , where the sum is over all partitions ë of n (when
j +1
ë has ¾ parts, by convention ë ¾+1 = 0). (Beckwith [2011])
3.4.26. For a partition ë, let m(ë) denote the total number of parts, and let mi(ë) denote
the number of parts equal to i. Prove
m(ë)!
∑∏ mi ( ë)!
= 2n − Fn ,
i

where the sum is over all partitions of n + 1 having at least one 1 and F n is the nth classical
Fibonacci number. (Merca [2014a])

3.4.27. (♦) Let p(n , m) be the number of partitions of n with m parts, and let H(x , t) =
∑m ,n p(n , m)xn tm .
∞
(a) Prove H(x , t) = ∏¾=1 1 −1x¾ t .
(b) Let º (n) be the sum, over all partitions ë of n, of the number of parts in ë. Let
¾
F(x) = ∑n≥0 º (n)x n . Prove F(x) = ∑¾≥1 1 −x x¾ P(x), where P(x) = ∑n≥0 p(n)x n with p(n)
being the number of partitions of n. (Stanley [2014])
3.4.28. (♦) A double-partition of the integer n consists of two partitions ë and Þ such
that min{ ë i , Þi } ≥ max{ ë i+1 , Þi+1 } for all i (trailing zeros added as needed) and ∑ ë i +
∑ Þi = n. Note that the numbers of nonzero parts in ë and Þ must differ by at most 1.
Interchanging ë and Þ yields a different double-partition when ë and Þ are distinct. Let
a¾ ,n be the number of double-partitions of n such that ë and Þ each have at most ¾ parts.
∞ ¾ 2¾
Let A¾(x) = ∑n=0 a¾ ,n x n . Prove A¾(x) = ∏i=1 (1 + x2i−1)/ ∏i=1 (1 − xi). (Hint: Use induction
on ¾ .) (Corteel–Savage)
3.4.29. (♦) Let ë and Þ be integer partitions, and let ë∗ and Þ∗ be their conjugates. By
counting a set in two ways, prove ∑i , j min{ ë i , Þj } = ∑¾ ë∗¾ Þ∗¾ .

3.4.30. (+) Let n be a positive integer. For each positive integer ¾ , let q¾ = ⌊ n/¾⌋ , let
r¾ = n − ¾ q¾ , and let º (¾) = ( q2¾ )¾ + r¾ q¾ . Prove combinatorially that º (¾) ≥ º (¾ + 1) for all

¾ ∈ . (Hint: For each ¾ , create a Ferrers diagram for a partition of n in which º (¾) counts
something. Transform the diagram for ¾ into the diagram for ¾ + 1 by small changes that
don’t increase º .) (Griggs)

3.4.31. Prove the variation of Euler ’s Identity given below.

∞
x¾(¾+1)
∏(1 + x2i) = 1 + ∑ (1 − x2)(1 − x4) · · · (1 − x2¾)
i =1 ¾ ≥1

3.4.32. (♦) Prove the variation of Euler ’s Identity for all partitions.
∞
x¾
2
1
∏ 1 − xi =1+∑
(1 − x)2(1 − x2)2 · · · (1 − x¾)2
i =1 ¾ ≥1

3.4.33. (♦) Prove the variation of Euler ’s Identity for partitions into distinct parts.
∞
x¾(¾+1)/2 y ¾
∏(1 + xi y) = 1 + ∑ (1 − x)(1 − x2) · · · (1 − x¾)
i =1 ¾ ≥1
Exercises for Section 3.4 151

3.4.34. With mi being the number of copies of i in a partition , let  ( ) = 1 / ∏i≥1 mi !imi .

Prove for n ∈ that the sum of  ( ) over partitions of n equals 1. (Stanley [1986])

3.4.35. Rogers–Ramanujan partitions are partitions into distinct parts that do not
use two consecutive integers as parts. For n > 1, prove that at least half of the Rogers–
Ramanujan partitions of n do not use 1 as a part. (Andrews [1997a])

 n−1
3.4.36. (♦) For n ∈ , let P(n) = ∑i=0 p(i), where p(i) counts all partitions of i.
(a) Consider a set S of marbles consisting of one black marble and n − 1 white marbles.
White marbles are indistinguishable from one another. Prove that P(n) is the number of
distinguishable partitions of S.
(b) Prove that P(n) is the sum, over all partitions of n, of the number of distinct parts
in the partition. The count is one for 4, 2 + 2, and 1 + 1 + 1 + 1, and it is two for 3 + 1 and
2 + 1 + 1, so the sum is 7 when n = 4. (G. Beck)

3.4.37. (+) In a Dyck n-path (Example 1.3.27), a peak is a consecutive pair UD; its height
is the y-coordinate reached between its U and D. Counting the peak heights with multi-
plicity yields the peak multiset: for example, the peak multiset of UDUDUD is {1 , 1 , 1}.
Prove that the number of distinct multisets that occur as peak multisets of Dyck n-paths
n−1
is 2 n − 1 − ∑¾=1 p(), where p() is the number of partitions of  . (Callan–Deutsch [2012])

3.4.38. (♦) Euler ’s Pentagonal Number Theorem.

(a) Let pe(n) and po(n) count the partitions of n into an even number and an odd num-
ber of distinct parts, respectively. For example, pe(9) counts {8 + 1 , 7 + 2 , 6 + 3 , 5 + 4},
while po(9) counts {9 , 6 + 2 + 1 , 5 + 3 + 1 , 4 + 3 + 2}. Use the left diagram below to show
m−1
that pe(n) equals po(n) except when n has the form ∑i=1 (m + i) or ∑i=0 (m + i), in which
m

case they differ by 1. (Franklin [1881])

• • • • • • • • •
• •
• • • • • • • • • • •
• • • •
• • • • • • • • • • •
• • • • •
• • • • • •

(b) Let (m) = (3m2 − m)/2. Use part (a) to prove

∏(1 − x¾) = 1 + ∑(−1)m ( x (m) + x (−m)) . (Euler [1748])

¾ ≥1 m≥1

m−1
(Comment: “Pentagonal Number Theorem” arises from writing (m) as ∑ j =0 (3 j + 1). The
sum counts dots arranged pentagonally as on the right above, so (m) is a pentagonal num-
m−1 m−1
ber. Similarly, ∑ j =0 (j + 1) is a triangular number and ∑ j =0 (2 j + 1) is a square number.)

3.4.39. Pentagonal numbers, continued.

(a) With p(n) being the number of partitions of n and (m) = (3m2 − m)/2, use Exercise
3.4.38b to prove the identity below for n ≥ 1.

p(n) = ∑(−1)m−1 [ p( n − (m)) + p (n − (− m))] (Euler [1748])

m≥1

(Hint: For n ≥ 0, let an be p(n) minus the right side of the identity. Prove ∑n≥0 an x n = 1.)
(b) Prove ∑ j ≥1 (−1) j qj (n − (j)) = 0, where qj () is the number of partitions of  not
having j or 2 j as a part. (For n = 9, we compute − q1(9) + q2(6) − q2(0) = −4 + 5 − 1 = 0.)
(Andrews [1997b])
152 Chapter 3: Generating Functions
n+ j
3.4.40. (♦) The Gaussian polynomial [mm+ n] q is defined by [ mm+ n]q = ∏ j =1 11−−qq j .
m

(a) Let pm ,n; ¾ be the number of partitions of ¾ into at most m parts of size at most n.
Prove pm ,n; ¾ = pm−1 ,n; ¾ + pm ,n−1; ¾− m .
(b) For fixed m and n, prove ∑¾ pm ,n; ¾ q¾ = [mm+ n] . Conclude that [ mm+ n] is a polynomial
q q
and is symmetric in m and n.
3.4.41. Letting mi be the number of parts equal to i in a partition , prove for n ∈  that
∏ ë (∏¾≥1  m¾ ) = ∏ ë (∏ j ≥1 mj !) , where each outer product is over all partitions of n. For
example, the values of both sides for 1 ≤ n ≤ 4 are 1 , 2 , 6 , 96, respectively. (Hint: Show
that the logarithms of both sides are equal.) (Kirdar–Skyrme [1982], Hoare [1986])
3.4.42. (♦) A corner of a partition is a dot in its Ferrers diagram that is the last dot in its
row and in its column. Let ( ) denote the number of corners in a partition , and let (n)
denote the set of all partitions of n.
(a) Prove ∑ ë∈(n) ( ) = |(n − 1)| + ∑ ë∈(n−1) ( ).
(b) Conclude ∑n≥0 ∑ ë∈(n) ( )x n = 1 −x x ∏i≥1 1 −1xi . (Pak [2006])

3.4.43. (♦) Let and  be conjugate partitions, with trailing zeros added. Prove that the
infinite multisets { i + i: i ≥ 1} and { i + i: i ≥ 1} are the same. (Knuth [2018])
3.4.44. (+) Prove that the number of partitions of n equals the number of partitions (of all
integers) into distinct parts whose odd-indexed parts sum to n. For example, the partitions
of 4 are 4, 3 + 1, 2 + 2, 2 + 1 + 1, and 1 + 1 + 1 + 1, and the partitions into distinct parts
whose odd-indexed parts sum to 4 are 4, 4 + 3, 4 + 2, 4 + 1, and 3 + 2 + 1. (Schmidt [1997])
3.4.45. (+) Let t(n) be the number of unordered factorizations of n into divisors greater
∞
than 1 (for example, t(12) = 4). Prove ∑n=2 t(n)/n2 = 1. (Beckwith [1998])
3.4.46. Let A be an m-by-n integer matrix. Given 1 , . . . , m with n ≥ 1 ≥ · · · ≥ m ≥ 0,
ë
let  A( ) = ∑i=1 ∑ j =i 1 ai , j . A move subtracts 1 from some entry and adds 1 to the next
m

entry rightward or below or to no entry. Prove that some sequence of moves reduces A to
the zero matrix if and only if  A( ) ≥ 0 for all such . (Bennett [1992])
3.4.47. Let  be a function defined on the subsets of [n] by letting  (S) = S ∪ {1} if 1 ∈
/S
and  (S) = [n] − { − 1:  ∈ S} if 1 ∈ S. Show that  2 has a unique fixed point S∗ reached
by iterating  from any initial set S. (Huang–Scully [2003])
3.4.48. (♦) Determine all the partitions that are fixed points in Bulgarian solitaire.
3.4.49. (♦) Determine the number of steps in Bulgarian solitaire needed to turn the posi-
tion consisting of a single pile of size n into a position on a cycle. For example, 5 → 41 →
32 → 221 → 311 → 32, so the answer is 2 when n = 5. (Cranston–West [2013])
3.4.50. For  ≥ 3, let be the partition of (¾+2 1 ) whose parts are two copies of  − 1, two
copies of 1 , and one copy each of 2 through  − 2. Prove that Bulgarian solitaire takes
( − 1) steps to reach the “staircase” partition from . (Griggs–Ho [1998])
3.4.51. For a partition , let d( ) be the number of moves of Bulgarian solitaire needed to
reach a partition that lies on a cycle when starting from .
(a) Prove that if the Ferrers diagram for is contained in the Ferrers diagram for  ,
then d( ) ≤ d(). (Akin–Davis [1985])
(b) As stated in Remark 3.4.16, d( ) ≤ ( − 1) when is a partition of (¾+2 1 ) (Exercise
3.4.50 shows that this is sharp). Conclude that d( ) ≤ ( − 1) when is a partition of n
with (¾2) < n < (¾+2 1 ). (Igusa [1985], Bentz [1987], Etienne [1991], Griggs–Ho [1998])
Chapter 4

Further Topics
The topics in this chapter exhibit common themes of indirect counting
techniques and applications to permutations and lattice paths. The Inclusion-
Exclusion Principle gives a combinatorial explanation for many alternating sums,
while the P ólya–Redfield method counts equivalence classes of objects made indis-
tinguishable by symmetry operators. Young tableaux, which arise in the theory
of representations of symmetric groups, have surprising connections to permuta-
tions and sorting.

4.1. The Inclusion-Exclusion Principle

“Sieve methods” are counting methods that within a universe allow only a de-
sired set to survive a process of overcounting and undercounting. The Inclusion-
Exclusion Principle is the most common sieve and is our focus. Knuth attributed
it to Doctrine of Chances by de Moivre [1718]. Riordan [1958] wrote
The logical identity on which it rests is very old; Dickson’s History
of the Theory of Numbers (vol. I, p. 119) mentions its appearance in a
work by Daniel da Silva in 1854, but Montmort ’s solution in 1713 of a
famous problem, known generally by its French name, “le problème des
rencontres” . . . effectively uses it and it may have been known to the
Bernoullis.
Montmort corresponded with Nicolaus Bernoulli on such topics during 1710-12.

THE BASIC PRINCIPLE

Many elementary texts describe inclusion-exclusion as a way to count the el-

ements in a union of sets. Summing the sizes of two sets counts the intersection
twice, so | A1 ∪ A2 | = | A1 | + | A2 | − | A1 ∩ A2 |. For three sets, the computation is
+ | A1 | − | A1 ∩ A2 |
| A1 ∪ A2 ∪ A3 | = + | A2 | − | A2 ∩ A3 | + | A1 ∩ A2 ∩ A3 | .
+ | A3 | − | A3 ∩ A1 |
153
154 Chapter 4: Further Topics

The first column overcounts each element in more than one Ai . An element in any
two sets is cancelled by subtracting the intersection, but this completely cancels
elements in all three sets. We correct by adding back the number of elements in
all three sets. Now every element of the union is counted exactly once.

4.1.1. Example. The classical model. In most generalizations and applications

with n sets, we want instead to count the elements of an appropriate universe
U that are outside the union of specified sets. We can do this by subtracting the
formula for the size of the union from |U |.
## #
### A1 ∪ A2 #### = |U | − | A1 | − | A2 | + | A1 ∩ A2 |
− | A1 | + | A1 ∩ A2 |
### A ∪ A ∪ A ### = U
## 1 2 3 ## | | − | A2 | + | A2 ∩ A3 | − | A1 ∩ A2 ∩ A3 | .
− | A3 | + | A3 ∩ A1 |
This model is natural for counting intersections instead of unions, since
⋃ Ai = ⋂i=1 Ai . Imagine n constraints on objects in U. If Ai is the set of
n n
i=1
elements violating the ith constraint, then the desired objects (satisfying all the
constraints) are those in ⋂i=1 Ai .
n

For example, the derangements of [n] are the permutations of [n] with no
fixed points. If Ai is the set of permutations that fix element i, then Dn =
### n A ###. Similarly, surjective distributions of objects into boxes are the distri-
##⋂i=1 i ##
butions with no box empty, and the numbers in [m] relatively prime to m are the
numbers not divisible by any prime factor of m. We will discuss all of these.
Consider the Venn diagram of sets A1 , . . . , An in a universe U. In addition
to counting the elements in the region outside all the circles, we may want to
count the elements in any given cell. The cell corresponding to the index set T
for T ⊆ [n] consists of the elements that are in Ai for i ∈ T and are not in Ai for
i∈/ T . We introduce precise notation for these concepts.

4.1.2. Definition. Let A1 , . . . , An be subsets of a universe U. For x ∈ U , let P(x)

be the property set of x, defined by P(x) = {i ∈ [n]: x ∈ Ai}. For S, T ⊆ [n],
# #
let º (T) = |{x ∈ U: P(x) = T}| and ½(S) = ####⋂i∈S Ai ####.

U
A3 º (∅)
º (3)
º (13) º (23)
º (123)
A1 A2
º (1) º (12) º (2)

The notation here echoes the meaning. Often A1 , . . . , An are defined by spec-
ified properties for the elements, so P(x) indexes the Properties satisfied by x.
Also, º (T) counts the elements with a º ixed property set T , while ½(S) counts
those whose property set is ½ reater than or equal to S. We drop set brackets and
commas in the arguments of º and ½ when specifying a particular subset of [n].
Section 4.1: The Inclusion-Exclusion Principle 155

The relationship between ½ and is (S) = ∑ T ⊇S (T). That is, ⋂i∈S Ai con-
sists of the elements of U belonging to Ai for i ∈ S and possibly to other sets
also. Thus (1) = | A1 | = (1) + (12) + (13) + (123) and (12) = | A1 ∩ A2 | =
(12) + (123) when n = 3 and the sets are A1 , A2 , A3 . We use (S) and ####⋂i∈S Ai ####
# #
interchangeably.
In many settings, #####⋂i∈S Ai ##### is easy to compute directly but (T) is not. For
# #
example, in the derangements problem ####⋂i∈S Ai #### = (n − | S|)!, because after fixing
the elements indexed by S we can permute the remainder arbitrarily, not caring
whether other elements also become fixed points.
When (S) is easy to compute, we want to invert the relationship (S) =
∑ T ⊇S (T) to compute (T) using values of (S). This is like inversion princi-
ples in Chapter 3 for sequences ⟨a⟩ and ⟨b⟩; now values are indexed by subsets of

[n] instead of by 0 . (Notation such as N(S) and N ∗(S) or N=(S) and N≥(S) has
been used, but and  better suggest the more general context of inversion.)
The next theorem provides a formula for the inversion. The proof uses the
observation that a nonempty finite set has the same number of subsets of even
size and odd size.

4.1.3. Theorem. (Inclusion-Exclusion Principle — PIE). Let A1 , . . . , An be

sets in a universe U. For T ⊆ [n], let (T) be the number of elements x ∈ U
such that x ∈ Ai if and only if i ∈ T . The inclusion-exclusion formula for is
# #
(T) = ∑(−1)| S|−| T | ####⋂ Ai #### .
## ##
S⊇ T i∈ S

Proof: We prove that each element of U contributes equally to both sides. An

element x contributes to the term for S in the sum if and only if x appears in
⋂i∈S Ai . Since S ⊇ T , only elements in ⋂i∈ T Ai contribute.
Recall that P(x) = {i ∈ [n]: x ∈ Ai}. Thus x appears in ⋂i∈S Ai precisely when
T ⊆ S ⊆ P(x). When | S| − | T | is even, the contribution is positive; when | S| − | T |
is odd, it is negative. Thus x contributes +1 for each even subset of P(x) − T and
−1 for each odd subset of P(x) − T .
Nonempty sets have equal numbers of even subsets and odd subsets, so the
net count from x is 0 when P(x)
= T , as desired. When P(x) = T , only the term
S = T contributes, and x counts +1 on both sides.

Most applications need (∅), given by (∅) = ∑S⊆[n](−1)| S| #####⋂i∈S Ai #####. We start
# #
with | U | and next subtract the sizes of the sets Ai , letting ####⋂i∈∅ Ai #### = |U |. Treat-
ing ⋂i∈∅ Ai as U makes sense because intersection over no sets should be the “in-
tersective identity” U ; every element is in all of “none of the sets”.
One of the earliest applications counts the elements of [m] relatively prime
to m (the number 1 is included). The resulting value (m) is the Euler totient;
the function is studied extensively in number theory (see Exercises 15–16).

4.1.4. Proposition. Let p1 , . . . , pn be the distinct prime factors of m. The num-

ber (m) of elements of [m] relatively prime to m is given by
m
(m) = ∑ (−1)|S| .
S ⊆[n]
∏i∈S pi
156 Chapter 4: Further Topics

Proof: Within the universe [m], let Ai be the set of multiples of pi . The numbers
relatively prime to m are those in none of A1 , . . . , An . To apply the inclusion-
exclusion formula, we compute #####⋂i∈S Ai #####. The numbers divisible by all of {pi : i ∈
# #
S} are those divisible by ∏i∈S pi , and thus ####⋂i∈S Ai #### = m/∏i∈S pi .

Note that (1) = 1. As another example, 60 = 22 · 3 · 5, so

(60) = 60 − 60
2 − 60
3 − 5 + 6 + 10 +
60 60 60 60
15 − 60
30 = 16.

4.1.5. Remark. (1) In the PIE formula, the sets with a given size all contribute
with the same sign. Letting h¾ be the sum of #####⋂i∈S Ai ##### over all S with size  , we
can write the formula as  (∅) = ∑¾=0 (−1)¾ h¾ .
n

(2) With h¾ defined in this way, the value h¾ does not equal the number of el-
ements lying in at least  of the sets. The sum overcounts elements lying in more
than  sets.
(3) In many applications (not Proposition 4.1.4), ####⋂i∈S Ai #### depends only on | S|.
# #
The sets of size  all make the same contribution, and the sum with 2 n terms
becomes a sum with n + 1 terms involving binomial coefficients. Letting  ¾ be the
# #
common value of ####⋂i∈S Ai #### when | S| =  reduces the formula to
n
n
 (∅) = ∑(−1)¾ ( ) ¾ .
¾ =0


4.1.6. Example. Multisets with restricted usage. A girl has red, white, and blue
marbles and wants to bring 20 to school. In how many ways can she do this? When
20 marbles of each color are available, the multiset formula yields (202+2 ) ways. If
she has only seven of each, then we must discard multisets having at least eight
of the same color.
Let am ,n ,r be the number of multisets of size m from n types of elements using
fewer than r of each type. These correspond to the integer solutions to x1 + · · · +
x n = m such that 0 ≤ x i < r for each i.
Let U be the set of all multisets of size m from n types: | U | = (mn+−n1−1). For 1 ≤
i ≤ n, let Ai be the subset of U where type i violates its limit. These correspond to
x i ≥ r in the integer-sum model. Since the units being distributed are identical,
we can count such solutions by reserving r units for x i and distributing the rest
arbitrarily. Thus | Ai | = (m−nr−+1n−1). This may include distributions where other
constraints also are violated.
Violating  constraints leaves m − r units to distribute. Thus ####⋂i∈S Ai #### =
# #
(m−|Sn|−r+1 n−1), which depends only on | S| . By inclusion-exclusion,

n m − r + n − 1
n
am ,n ,r = ∑(−1)¾ ( )( ).
¾ =0
 n−1

When m = n(r − 1) − 1, the answer must be n, since we just choose a marble of

one type to leave home. Thus inclusion-exclusion provides a combinatorial proof
that ∑¾=0(−1)¾ (¾n)((n−n¾−)r1−2) = n. If m = n(r − 1) + 1, then we have proved that
n

∑¾=0(−1)¾ (¾n)( n−−¾1 ) = 0.

n (n )r
Section 4.1: The Inclusion-Exclusion Principle 157

4.1.7. Theorem. The formula for S(n , ¾), the number of partitions of an n-
element set into ¾ blocks, is
¾
1 ¾
S(n , ¾) =
¾! ∑(−1)i ( i )(¾ − i)n .
i=0

Proof: In each partition, the blocks are distinguished by their elements, so there
are ¾ ! distinguishable ways to index the blocks. Hence we can compute ¾ !S(n , ¾)
by counting surjective distributions of [n] into ¾ labeled blocks.
Altogether there are ¾ n distributions of [n] into ¾ labeled blocks with blocks
allowed to be empty. Let A j be the subset in which block j is empty. Given a set
S of blocks required to be empty, we distribute [n] arbitrarily into the remaining
# #
blocks. Hence ####⋂ j ∈S A j #### = (¾ − i)n whenever | S| = i. Combining the terms for sets
# #
of the same size yields, as desired,
¾
¾
¾ !S(n , ¾) = º (∅) = ∑(−1)i ( )(¾ − i)n .
i
i=0

We proved Theorem 4.1.7 using generating functions in Theorem 3.3.11. Stir-

ling numbers appear also in Exercises 47–50.
We have used PIE as an inversion formula: given sets A1 , . . . , An within
a universe U , we count each ⋂i∈S Ai and obtain º (∅), the number of elements
outside all the sets. Next we use PIE to evaluate sums by interpreting them as
inclusion-exclusion computations. This has the same flavor as evaluating convo-
lutions using generating functions.

4.1.8. Example. PIE evaluation of sums. Consider the sum ∑¾=0(−1)¾ (¾n)c¾ . If
n

the sum is the inclusion-exclusion formula for the number of elements outside
some sets A1 , . . . , An in a universe U , then counting those elements in another
way gives the value of the sum.
To fit this model, we need |U | = c0 and | Ai | = c1 ; we seek natural sets with
these sizes. Some creativity may be needed. The sets must be chosen so that also
c¾ = ####⋂i∈S Ai #### whenever | S| = ¾ .
# #
For example, when c¾ = 2 n−¾ , we seek a universe of size 2 n , such as the subsets
of [n]. To obtain | Ai | = 2 n−1 for 1 ≤ i ≤ n, let Ai consist of the subsets containing
# #
i. Now ####⋂i∈S Ai #### = 2 n−|S| , as desired. Hence the sum counts the subsets of [n] out-
side all of A1 , . . . , An . These are precisely the subsets containing no elements.
The only such set is the empty set, so ∑¾=0(−1)¾ (¾n)2 n−¾ = 1. (In the language of
n

the Binomial Theorem, we have just proved that (−1 + 2)n = 1.)
Similarly, in ∑¾=0(−1)¾ (n¾ ), each c¾ is 1. The sum counts by inclusion-
n

exclusion the items in a universe U of size 1 that belong to none of n sets each
equal to U. There is no such item if n > 0 and one such if n = 0, so the sum is 0 if
n > 0 and 1 if n = 0. We have counted a set both by PIE and directly.

4.1.9. Example. Evaluation of ∑¾=0(−1)¾ (¾n)(m+ n−¾

). Here c0 = (m+
n n
r −¾ r
); we want a
universe of that size. Let U be the family of r-sets in [m + n]. To model this as
an inclusion-exclusion computation, we need n sets of size (mr+−n1−1). Let Ai consist
of the r-sets in [m + n] that use element i. When | S| = ¾ , a member of ⋂i∈S Ai
158 Chapter 4: Further Topics

has ¾ required elements, so such members are completed by choosing r − ¾ of the

# #
remaining m + n − ¾ elements. Thus ####⋂i∈S Ai #### = (m+ n−¾ m+ n−¾
r −¾ ). Since c¾ = ( r −¾ ), we
conclude that the sum is the inclusion-exclusion sum to count (∅).
The r-sets in none of A1 , . . . , An are the r-sets in [m + n] that use none of the
first n elements. Thus the value of the sum is (m r ). We did not need to know the
value in advance to compute the sum.
This example generalizes the sum in Exercise 3.2.42.

Using inclusion-exclusion in this way is an argument by “counting two ways”

(see also Exercises 29–34). Next we return to the setting of finding a formula to
count a known set. We rederive the formula for the Eulerian numbers.

4.1.10.* Theorem. The Eulerian number A(n , ) is given by

¾
n+1
A(n , ) = ∑(−1)i ( )( − i)n .
i
i=0

Proof: (Stanley; see Bóna [2004, p. 8–9]) Let U be the universe of barred permu-
tations of [n] with  bars, defined in the proof of Theorem 3.1.25 to be permu-
tations of [n] in word form with  bars inserted, having at least one bar at the
end of each run. We showed there that |U | =  n ; the elements correspond to dis-
tributions of [n] into  boxes, since the numbers placed immediately before any
bar must be written in increasing order. The A(n , ) permutations with exactly
 runs are the elements of U having one bar at the end of each run and no others.
A barred permutation with  bars but fewer than  runs may have excess
bars in any of the n + 1 positions. For 1 ≤ j ≤ n + 1, let A j be the set of elements
of U having at least one excess bar in the jth position. Since (∅) = A(n , ), it
suffices to show that the sum is the inclusion-exclusion computation for (∅).
For S ⊆ [n + 1], we count the barred permutations having at least one excess
bar in each position indexed by S. When | S| = i, form a barred permutation with
 − i bars and then add one bar at each position indexed by S. Bars are identical,
so there is no distinction between the bars distributed in the two steps.
# #
For | S| ≤  , we have ####⋂ j ∈S A j #### = ( − | S|)n , since there are ( − i)n barred
# #
permutations of [n] with  − i bars. When | S| >  , the value is 0, because no
barred permutation with  bars has excess bars in more than  positions. Hence
the sum stops at  even though we avoid n + 1 sets.

Whitney [1932b] applied PIE to count proper colorings of graphs. A color-

ing of a graph is a function that assigns each vertex a label from a set of colors.
A proper coloring gives distinct labels to adjacent vertices. Typically, we use
[] for the set of colors. We are placing the n distinguishable vertices into  dis-
tinguishable boxes, with adjacent vertices forbidden to go in the same box.

4.1.11.* Application. For a set S of edges in a graph G, let c(S) be the number of
components in the subgraph with vertex set V(G) and edge set S. For the number
 G() of proper colorings of G using colors in [],
 G() = ∑ (−1)| S|  c(S) .
S ⊆ E(G)
Section 4.1: The Inclusion-Exclusion Principle 159

Proof: The universe U of unrestricted colorings is the set of all functions from
V(G) to [¾]; this has size ¾ n , where n = | V(G)|. Let Ai denote the set of colorings
in which the vertices of the ith edge have the same color. We want to count the
colorings belonging to none of these sets.
# #
To apply inclusion-exclusion, we compute ####⋂i∈S Ai ####. Let H be the subgraph of
G with vertex set V(G) and edge set S. A coloring in ⋂i∈S Ai makes each edge of H
monochromatic. If x and y are the ends of a path in H , then by transitivity x and
y have the same color. Thus all vertices of a component of H (a maximal connected
subgraph) have the same color. The color can be chosen in ¾ ways, independently,
# #
on each component of H. Thus ####⋂i∈S Ai #### = ¾ c(S) , as claimed.

In a graph G with n vertices and m edges, subgraphs with vertex set V(G)
are spanning subgraphs. Spanning subgraphs with one edge have n − 1 com-
ponents; those with more edges have fewer components. Thus Application 4.1.11
shows that G() is a polynomial in  with leading terms  n − m n−1 ; it is called
the chromatic polynomial of G. Exercises 36–38 concern chromatic polynomi-
als and their properties.

4.1.12.* Example. A chromatic polynomial. The computations in Application

4.1.11 begin  n − m n−1 + (m ) n−2 for | S| ≤ 2. When | S| = 3, the number of
2 
components is again n − 2 if the three edges form a triangle; otherwise it is n − 3.
The graph below has two triangles, and the other eight triples of edges yield one
component each. All subgraphs with four or five edges have only one component.
Hence the computation is
G ()=  4 − 5 3 + 10 2 − (2 2 + 8 1) + 5 −  =  4 − 5 3 + 8 2 − 4 .
By ad hoc counting, one can also see that G() = ( − 1)( − 2)( − 2). This is 0
when  is 1 or 2, but G(3) = 6.
• •

• •

RESTRICTED PERMUTATIONS

For selection problems, we considered restricted multiplicities. For arrange-

ment problems, we consider restricted positions. The first example is the derange-
ment problem; derangements are permutations without fixed points. We counted
these in Proposition 2.3.2 and Corollary 3.3.17; inclusion-exclusion yields the
same formula almost immediately.

4.1.13. Theorem. The formula for Dn , the number of derangements of [n] (per-
mutations with no fixed points), is
n
(−1)¾
Dn = n! ∑ .
!
¾ =0
160 Chapter 4: Further Topics

Proof: Let U = Ën . For i ∈ [n], let Ai be the set of permutations that fix i; we

seek º (∅). If | S| = , then the elements of ⋂i∈S Ai have specified fixed points
and permute the remaining elements of [n] arbitrarily. Thus #####⋂i∈S Ai ##### = (n − )!.
Collecting terms with | S| = yields
n n
n ¾ (−1)¾
 (∅) = ∑(−1) ( )(n − )! = n! ∑ .
!
¾ =0 ¾ =0

4.1.14. Example. Permutations with restrictions. The derangement problem gen-

eralizes to forbid any values from any positions. The permutation matrix for
a permutation  is the 0 , 1-matrix with 1 in entry (i , j) if and only if  (i) = j .
Forbidding value j from position i of the word form forbids 1 from entry (i , j) of
the matrix. We view the forbidden locations for 1s in the matrix as squares on a
chessboard, forming a board of forbidden positions. Below is a board with five
forbidden positions for permutations of [4]; any value may appear in position 4.

The universe is the set of n-by-n permutation matrices. Let A s consist of

those matrices with 1 in square s of the forbidden board B. We count the desired
permutations by inclusion-exclusion. There exist permutations with 1s in all po-
sitions indexed by S if and only if S has at most one position in every row and
column; such a set of positions in a matrix (or on a board) is independent.
# #
If the positions in S are independent in B, then ####⋂i∈S Ai #### = (n − | S|)!. Since
the universe has only permutation matrices, the number of such terms with | S| =
is not (¾n); it is the number of ways to choose independent positions on B. Let
r¾ (B) denote this number. For the board above, (r0 , r1 , r2 , r3 , r4) = (1 , 5 , 7 , 2 , 0).
PIE then counts the permutation matrices with no 1s in forbidden entries.

4.1.15. Proposition. Given a board B of forbidden positions for the 1s in an n-by-

n permutation matrix, the number of permutations of [n] having no elements
in forbidden positions is ∑¾=0(−1)¾ r¾ (B)(n − )!, where r¾ (B) is the number of
n

-element subsets of B having no two elements in the same row or column.

Proof: The sum is the PIE computation for  (∅) within the universe of permu-
tation matrices, where Ai consists of the permutation matrices with 1 in the ith
# #
square. We have ####⋂i∈S Ai #### = (n − | S|)! when S has no two positions in the same
row or column, and otherwise ####⋂i∈S Ai #### = 0.
# #
In the derangement problem, the forbidden positions are those along the di-
agonal, and every subset of these is independent. In this case r¾ (B) = (¾n), and
the computation reduces to Theorem 4.1.13. In general, we model independent
positions by pairwise non-attacking rooks on a chessboard, since a rook is a piece
that can attack (or move to) any position in its row or column (if there are no
intervening pieces). This is why we call the set of forbidden positions a board.
Section 4.1: The Inclusion-Exclusion Principle 161

4.1.16. Definition. For a board B of positions in a grid, the rook polynomial

R B(x) is the generating function ∑ r¾ (B)x¾ , where r¾ (B) is the number of
placements of ¾ pairwise non-attacking rooks on B.

4.1.17. Remark. The rook polynomial factors when a board B decomposes into
sub-boards B1 and B2 occupying distinct rows and distinct columns. Since ¾ rooks
on B must consist of j on B1 and ¾ − j on B2 , and rooks on the two sub-boards cannot
attack each other, the rule for multiplying OGFs yields RB(x) = RB1 (x)RB2(x). For
example, we compute RB(x) for the board in Example 4.1.14 as (1 + 3x + x2)(1 + 2x),
which equals 1 + 5x + 7x 2 + 2x3 .
To compute r¾ (B) in general, group rook placements by whether they occupy
a fixed square s in B. Deleting s from B yields a board B − s, and deleting the
entire row and column containing s yields a board we write as B · s. Counting the
placements of ¾ rooks by whether they use s yields r¾ (B) = r¾ (B − s) + r¾−1 (B · s).
Note that r0(B) = 1 for every board and r¾ (∅) = 0 for ¾ > 0. Multiplying by x¾
and summing where the recurrence is valid (¾ ≥ 1) yields RB(x) − 1 = RB− s(x) −
1 + xRB· s(x). Thus
RB(x) = RB− s(x) + xRB· s(x).
An extensive classical treatment of rook polynomials appears in Riordan
[1958]. See also Exercises 50–54.

How many permutations have exactly ¾ elements in forbidden positions? For

inclusion-exclusion in general, we develop a generating function to enumerate ob-
jects by “number of properties”. Having property i is the same as being in set Ai
in the formulation of PIE using subsets.

4.1.18. Theorem. For sets A1 , . . . , An in a universe U , let h¾ be the sum of

## #
###⋂i∈S Ai #### over all sets S of size ¾ . If ºp is the number of elements of U in
exactly p of the sets A1 , . . . , An , then
n n

∑ ºp x = ∑(x − 1)¾ h¾ .
p

p=0 ¾ =0

Proof: As in Theorem 4.1.3, we show that each element contributes the same
to both sides. For a ∈ U , the contribution of a to the left side is x p , where p =
|{i ∈ [n]: a ∈ Ai}|. Since the number of ¾-subsets of { A1 , . . . , An} whose intersec-
tion contains a is (¾p), element a contributes (¾p) times to h¾ . Hence a contributes
∑¾=0(x − 1)¾ (¾p) to the right side. Since (¾p) = 0 for ¾ > p, the sum is x p by the
n

Binomial Theorem.

4.1.19.* Remark. Using the Binomial Theorem to expand (x − 1)¾ in Theorem

¾
4.1.18 yields ∑¾=0 (∑ p=0 (¾p)(−1)¾− p x p h¾ ) on the right. Interchanging the order
n

of summation and extracting the coefficient of x p yields ºp = ∑¾= p(−1)¾− p (¾p)h¾ .

n

Some authors express Theorem 4.1.18 in this form.

Setting x = 0 reduces to the usual inclusion-exclusion formula º (∅) =
∑¾=0(−1)¾ h¾ . The general formula of Theorem 4.1.18 can also be proved by sum-
n

ming the inclusion-exclusion formulas for each º (T) (Exercise 25).

162 Chapter 4: Further Topics

To apply Theorem 4.1.18 to permutations with restricted positions given by a

board B, recall that h¾ = r¾ (B)(n − ¾)!. With ºp being the number of permutations
having p elements in forbidden positions,
n n

∑ ºp x p = ∑(x − 1)¾ r¾(B)(n − ¾)!.

p=0 ¾ =0

4.1.20. Example. Expected number of fixed points (Montmort [1708]). Consider

an urn with balls labeled 1 through n. We extract the balls one by one, randomly.
What is the expected number of times that the label on a ball equals the step on
which it is drawn? These occurrences were called “rencontres”, and this question
was the “problème des rencontres”.
In the language of forbidden positions, we are seeking the expected num-
ber of fixed points in a random permutation of [n]. Let N(x) be the generating
function for permutations by the number of positions with forbidden elements:
N(x) = ∑ ºp x p . As in Application 3.2.15, we find the expected weight of a ran-
dom object in a finite universe U by differentiating the weight enumerator: the
expected weight is N (1)/ | U |.
To enumerate permutations by fixed points, we use the diagonal board of for-
bidden positions used to count derangements. We obtain
n n
n (x − 1)¾
N(x) = ∑(x − 1)¾ ( )(n − ¾)! = n! ∑ .
¾ =0
¾ ¾!
¾ =0

The answer is N (1)/n!, which equals 1, a remarkably simple answer.

When a fancy method yields a simple answer, we seek a simpler explanation
(we called this the “Erdős Principle”). Here, the number X of fixed points is ∑ X i ,
where X i is 1 if i is a fixed point and 0 otherwise. By the linearity of expectation,
(X) = ∑ni=1 (X i = 1) (see Chapter 14 for further explanation and use of this
concept). Since each element has probability 1/n of being fixed, (X) = 1.

We consider another classical example.

4.1.21. Example. Problème des ménages (Lucas [1891]). The hosts of a party with
n male/female couples want to seat men and women alternately at a circular ta-
ble with no spouses adjacent. Index the couples 1 , . . . , n by the circular ordering
of sex A. Let position i be the position between Ai and Ai+1 . Neither Bi nor Bi+1
is allowed to sit there. Acceptable seatings of B1 , . . . , Bn now correspond to per-
mutations of [n] not having i or i + 1 in position i for any i (modulo n).

•
• 1 •

•4 2 •

• 3 • s
•
Section 4.1: The Inclusion-Exclusion Principle 163

To count the allowed seatings, compute r¾ (B) with this board of forbidden
positions. Applying the recurrence to the square s in the lower left yields two
zig-zag boards; we must place ¾ rooks on a board with 2n − 1 positions or ¾ − 1
rooks on a board with 2n − 3 positions.
Instead of continuing the recurrence, note that l non-attacking rooks on a zig-
zag board with m positions correspond to a 0 , 1-string of length m with l copies
of 1, no two consecutive. These strings arise from an all-0 string of length m − l
by choosing l distinct places to insert a 1. There are (m−ll+1) such choices.
With (m , l) set to (2n − 1 , ¾) or (2n − 3 , ¾ − 1), the computation yields r¾ (B) =
(2n¾−¾ ) + (2n¾−−¾1−1 ) = 2n2n−¾ (2n¾−¾ ). Having seated one sex, the enumerator of arrange-
ments by number of consecutively-seated couples is
n
2n 2n − ¾
∑ 2n − ¾ ( )(n − ¾)!(x − 1)¾ .
¾
¾ =0

Setting x = 0 counts the arrangements with no couples consecutive.

SIGNED INVOLUTIONS

The basic Inclusion-Exclusion Principle has many generalizations. It gener-

alizes to the classical Möbius Inversion Formula of number theory and further to
the setting of partially ordered sets (see Section 15.2).
Here we discuss a simple but flexible generalization. Inclusion-exclusion
counts a set within a larger universe X by canceling unwanted elements. If we
weight the elements ±1 so that unwanted elements come in plus/minus pairs,
and elements of the desired set F have positive weight, then | F | will be the sum
of the weights over X . We consider a more general setting.

4.1.22. Definition. An involution is a permutation whose square is the iden-

tity; the cycles have length 1 or 2. With respect to a partition of X into a
positive part X + and a negative part X − , a signed involution on X is an in-
volution such that every 2-cycle pairs a positive element with a negative
element. Let F Ì and G Ì denote the sets of fixed points within X + and X −
under a signed involution .

FÌ • • • • X+

GÌ • • • • X−

4.1.23. Example. Parity of permutations. A permutation  is even [odd] when

the number of inversions in  is even [odd] (see Proposition 3.1.15). A more com-
mon definition is the parity of the number of transpositions needed to turn  into
the identity permutation; these are the same because every exchange of two ele-
ments in the word form changes the number of inversions by an odd amount.
even permutations: 123 231 312
odd permutations: 213 321 132
164 Chapter 4: Further Topics

Interchanging the first two elements in the word form changes the parity.
Repeating the transposition changes it back. Thus we have a signed involution
with X = Ën , where X + and X − are the sets of even and odd permutations, re-
spectively. There are no fixed points unless n = 1.

4.1.24. Proposition. If Ì is a signed involution on a set X weighted by w(x) = 1

for x ∈ X + and w(x) = −1 for x ∈ X − , then
∑ x∈ X w(x) = | F Ì | − | G Ì |.
Proof: The contributions to the sum from 2-cycles in Ì cancel, leaving those from
fixed points: | F Ì | with weight +1, | G Ì | with weight −1.

The inclusion-exclusion formula for º (∅) is a special case of Proposition

4.1.24. We embed U in a still larger set X and show that the alternating sum
in the formula for º (∅) is the sum of the weights for a signed involution on X
having º (∅) fixed points, all in X + .

4.1.25. Example. PIE by signed involution. We have a universe U and sets

A1 , . . . , An . We seek º (∅). Recall the “property set ”: P(x) = {i ∈ [n]: x ∈ Ai}. In
the inclusion-exclusion formula, each element x of U appears in ⋂i∈S Ai for each
S such that S ⊆ P(x). Grouping the contributions to reflect this yields
# #
º (∅) = ∑ (−1)| S|#####⋂ Ai ##### = ∑ ∑ (−1)|S| = ∑ (−1)| S| ,
# i∈S # x∈U
S ⊆[n] S ⊆ P(x) (x ,S)∈ X

where X is the set of ordered pairs (x , S) such that x ∈ U and S ⊆ P(x). The
contribution is +1 for even | S| and −1 for odd | S| . Partition X into X + and X − by
the parity of | S| . To apply Proposition 4.1.24, we need a signed involution with
respect to (X + , X −) whose fixed points lie in X + and correspond to the elements
of U outside all the sets (those with P(x) = ∅). Proposition 4.1.24 then proves the
inclusion-exclusion formula.
Given a pair (x , S) ∈ X , view the elements of S as marked elements of P(x).
Let Ì operate on elements of X by switching whether the largest element of P(x)
is marked, if any element of P(x) exists. That is,
⎧
⎪ (x , S) if P(x) = ∅ ,
Ì(x , S) = ⎪⎨⎪ (x , S − i)if i = max P(x) ∈ S,
⎪
⎩ (x , S ∪ i) if i = max P(x) ∈
/ S.

Applying Ì again acts on the same element of P(x) and returns us to (x , S).
Hence Ì is an involution. It is a signed involution relative to (X + , X −), be-
cause Ì changes the parity of | S| when it makes a change. The fixed points are
{(x , ∅): P(x) = ∅}. These lie in X + and correspond bijectively to the elements of
U outside all of A1 , . . . , An .

Many applications of signed involution follow this pattern. We embed the

desired set in a larger set X . For elements in X , we define a “switch” opera-
tion such that applying it twice retrieves the original element. The desired fixed
points have no candidates for switching, so the map does not change them (see
Exercise 57, for example).
Section 4.1: The Inclusion-Exclusion Principle 165

By taking advantage of cancellation, signed involutions can make it possible

to find short proofs that avoid needless detailed counting.

4.1.26. Proposition. (Deutsch [2005]) For a fixed convex n-gon, if e n and on count
the dissections by nonintersecting diagonals into an even or an odd number
of bounded regions, then e n − on = (−1)n .
Proof: Let the parity of a dissection be the parity of the number of regions. If a
signed involution Ì pairs dissections with opposite parity, then e n − on is the num-
ber of even fixed points minus the number of odd fixed points. Adding or deleting
one diagonal changes the parity; we want Ì to do this and repeating Ì to undo it
by changing the same diagonal.
Let v1 , . . . , vn be the vertices in order around the n-gon. If the diagonals
incident to vn are not all present, then let ¾ be the index before the first non-
neighbor of vn (if any exists), and let l be the index of the next actual neighbor
of vn . (For example, when no diagonals are present, there is only one region and
(¾ , l) = (1 , n − 1).) Define Ì by inserting v¾ vl if it is missing and removing v¾ vl if
it is present. When Ì inserts an edge, no existing edge crosses it, so the image is
indeed a dissection.
The dissection is unchanged by Ì if and only if ¾ and l are undefined, meaning
that vn is adjacent to all the other corners and there are n − 2 regions. When Ì
produces a change, the parity changes. Because Ì does not change the set of edges
incident to vn , the values ¾ and l for a dissection are the same as for its image
under Ì . Thus repeating Ì changes the presence of the same diagonal, restoring
the original dissection.
We have shown that Ì pairs even and odd dissections except for the one fixed
point, whose parity is the same as the parity of n.
vn v1
• •
• • v2

vl • • v¾
• •

Exercises 56–67 concern signed involutions.

DETERMIN ANTS AND PATH SYSTEMS (optional)

Many counting problems about lattice paths are solved by determinants.

Gessel–Viennot [1985] explained this phenomenon by using signed involutions.
Lattice paths can be viewed as paths in a directed graph. Every determinant
can be interpreted as a signed counting problem in a weighted bipartite digraph.
The common generalization is a theorem by Lindström [1973] that includes the
results on lattice paths. Aigner [2001] gave an elegant exposition of the topic
with many striking applications. We begin with a simple example as motivation.
166 Chapter 4: Further Topics

4.1.27. Example. We know how to count lattice paths using binomial coefficients.
Consider disjoint pairs of lattice paths with specified endpoints. For example,
with initial points defined by x1 = (0 , 1) and x2 = (1 , 0), and terminal points
defined by y1 = (1 , 2) and y2 = (3 , 2), there are eight such pairs, shown below.
An x , y-path is a path from x to y. By computing the familiar binomial co-
efficients, the numbers of x i , y j -paths, arranged in a matrix for i , j ∈ {1 , 2}, are
(12 46). As it happens, the determinant of this matrix also equals 8. We will see
that this is no coincidence.

We defined the relevant concepts for directed graphs in the Introduction. Re-
call that each edge in a directed graph (digraph) G with vertex set V(G) and
edge set E(G) is an ordered pair (u , v) of vertices, abbreviated to uv. An edge is
directed from its first vertex (the tail) to its second vertex (the head). An x , y-
path consists of vertices that can be ordered as v0 , . . . , v¾ with v0 = x and v¾ = y
and edges {vi vi+1 : 0 ≤ i ≤ ¾ − 1}.
For example, we used functional digraphs in Chapter 1; this is the special
case where each vertex is the tail of exactly one edge. A path in the functional
digraph of follows repeated iteration of .

4.1.28. Example. Lattice paths can be modeled as paths in a digraph with grid
points as the vertices and allowed transitions as the edges. The ballot paths of
length 6 are the x , y-paths in the digraph on the left below.

•y y1 yj yn
• ··· • ··· •
• •
a1 ,1 an ,n
• • • ai , j
• ··· • ··· •
x• • • • x1 xi xn

We introduce a weight w(e) on each edge e. In the digraph G on the right

above, the edges are {x i yj : i , j ∈ [n]}. Let w(x i yj ) = ai , j , where the numbers ai , j
are the entries in an n-by-n matrix A. Using sets of paths in G, we will interpret
the computation of det A combinatorially.

4.1.29. Definition. In a weighted digraph G, the weight of a subgraph is the

product of the weights of its edges. Given vertex subsets X , Y ⊆ V(G), with
X = {x1 , . . . , x n} and Y = {y1 , . . . , yn}, the X , Y -path matrix has as its
(i , j)-entry the sum of the weights of all paths from x i to yj . An X , Y -path
system  consists of a permutation   of [n] and paths P1 , . . . , Pn such that
Pi is a path from x i to y (i). It is a disjoint-path system if the paths are
pairwise disjoint.
Section 4.1: The Inclusion-Exclusion Principle 167

4.1.30. Example. Determinants and path systems. For the weighted digraph on
the right in Example 4.1.28, the path matrix is the matrix A of order n that
provided the weights; there is exactly one x i , yj -path for each pair (i , j), and its
weight is ai , j . All the path systems are disjoint-path systems, and there is one
such system  for each permutation .
In terms of permutations, det A = ∑ ∈Ën(sign ) ∏i=1 ai , (i) , where the sign
n

of a permutation is +1 or −1 depending on whether the number of inversions

is even or odd (see Example 4.1.23). Since w( ) in this example is precisely
∏i=1 ai , (i) , we obtain ∑ ∈Ën(sign )w( ) = det A.
n

Our aim is to extend this formula to all disjoint-path systems in all digraphs.
Aigner wrote, “It is precisely this step to arbitrary graphs that makes the [The-
orem] so widely applicable — and what ’s more, the proof is stupendously simple
and elegant.”

To keep the sums finite, we consider only finite digraphs without cycles.

4.1.31. Theorem. (Lindström [1973], Gessel–Viennot [1985]) Let X and Y be n-

sets of vertices in a finite acyclic digraph G with edges weighted by w. If A
is the X , Y -path matrix, and P is the set of disjoint- X , Y-path systems, then

∑(sign )w() = det A

∈P

Proof: Let Q be the set of all X , Y-path systems, disjoint or not. We seek a
signed involution on Q whose fixed points are the disjoint-path systems and whose
systems paired in 2-cycles have the same weight and have associated permuta-
tions with opposite sign. This is a signed involution with X + and X − being the
sets of X , Y-path systems whose associated permutations are even or odd, re-
spectively. The paired contributions cancel in ∑∈Q(sign )w(), leaving just
∑∈P(sign )w(). The final step is to interpret the sum over all of Q as a deter-
minant, proving

∑(sign )w() = ∑(sign )w() = det A.

∈P ∈Q
Let Pi denote the x i , y (i)-path in a path system  with permutation . Let i∗
be the least i such that Pi intersects another path in  (if such an index exists).
Let  be the first vertex of Pi∗ shared with another path, and let j ∗ be the smallest
index other than i∗ such that P j ∗ also contains . Let () be the path system
obtained from  by switching the portions of Pi∗ and P j ∗ after . Let be the
permutation associated with ().
•y (3) •y (1)

x1 • • • •y (1) x1 • • • •y (3)

←→
x 2• • •y (2) x 2• • •y (2)

x 3• x 3•
168 Chapter 4: Further Topics

The switch transposes (i∗) and (j ∗) in the word form of to obtain . Trans-
positions reverse sign, so sign Ì() = −sign  . Also (()) =  , since paths re-
main unchanged before intersections, so  selects the same intersection point and
pair of paths in () and undoes the switch.
For a disjoint-path system, no intersection is found and  makes no change.
Always w() = w(()), since both systems use the same edges with the same
multiplicities. Thus  is a signed involution having the desired properties, so
∑∈Q(sign )w() = ∑∈P(sign )w().
For the remaining step, recall that det A = ∑ (sign )a1 , (1) · · · an , (n) , where
ai , j is the sum of the weights of all x i , yj -paths. Each path system associated
with chooses one x i , y (i)-path for each i, and its weight is the product of their
n
weights. Hence ∏i=1 ai , (i) is the sum of the weights of all the path systems asso-
ciated with . Summing over yields ∑∈Q(sign )w(), as desired.

Aigner noted that basic properties of determinants follow easily from The-
orem 4.1.31. For example, det A T = det A follows by reversing the edges in the
digraph of Example 4.1.28. The path system associated with is associated with
−1 instead, but they have the same sign, so the computation of the determinant
is the same. Exercise 62 uses Theorem 4.1.31 to prove that a matrix with linearly
dependent rows has determinant 0.
Another application generalizes the product formula for determinants to
products of non-square matrices. We will apply this result in Chapter 15. This
proof is very easy; others require tedious manipulation or clever linear algebra.

4.1.32. Proposition. (Cauchy–Binet Formula) If A and B are n-by-p and p-by-

n matrices, respectively, with n ≤ p, then
det(AB) = ∑(det AS)(det BS) ,
S ⊆S

where S is the family of all n-subsets of [p], AS consists of the columns of A

indexed by S, and BS consists of the rows of B indexed by S.
Proof: (Aigner [2001]) Form a digraph with vertex set X ∪ Y ∪ Z, where X =
{x1 , . . . , x n}, Y = {y1 , . . . , yn}, and Z = { 1 , . . . , p}. Let the edge set be
(X × Z) ∪ (Z × Y), with w(x i j ) = ai , j and w(j y) = bj , . Entry (i , ) in AB is
∑j w(x i j)w(j y), so AB is the X , Y -path matrix. By Theorem 4.1.31, det(AB)
sums the weights of nonintersecting X , Y -path systems.
Such a path system  selects for each i a vertex  (i) on a path from x i to y (i) ;
let S be the set of indices selected. Since w() is the product of all edge weights,
w() is the product of weights of path systems in the digraph computations for
det AS and det BS . Also, sign is the product of the signs of the permutations
of [n] for those path systems. Therefore, the right side gives the sum of weights
over all nonintersecting X , Y -path systems, grouped by the choice of S.

When all disjoint- X , Y-path systems have the same associated permutation
, a determinant counts them. For this discussion, we return to the context of lat-
tice paths from Example 4.1.27 and name the initial points p1 , . . . , pn and termi-
nal points q1 , . . . , qn . The digraphs for lattice paths have edges oriented upward
and rightward as in Example 4.1.28.
Section 4.1: The Inclusion-Exclusion Principle 169

4.1.33. Corollary. (Gessel–Viennot [1985]) Let h((a , b) , (c , d)) = (d−cb−+ac− a). If ev-
ery lattice disjoint-path system from {p1 , . . . , pn} to {q1 , . . . , qn} matches pi
to qi for all i, then the number of such disjoint-path systems is the determi-
nant of the matrix with (i , j)-entry h(pi , qj ).
Proof: Form the digraph on 2 with edges for upward and rightward steps. The
number of paths from point p to point q is h(p , q). By Theorem 4.1.31, the spec-
ified determinant sums the signed weights of all disjoint-path systems from the
initial points to the terminal points. Under the given conditions, the only such
systems use the identity permutation, which has positive sign. The weight of ev-
ery path system is 1, since all the edge weights are 1, so the sum of the weights
is the number of disjoint-path systems.

4.1.34. Example. Example 4.1.27, continued. The hypothesis of Corollary 4.1.33

holds when (p1 , . . . , pn) and (q1 , . . . , qn) are nondecreasing in first coordinates
and nonincreasing in second coordinates. This property holds in Example 4.1.27.
By Corollary 4.1.33, the number of disjoint paths of paths from p1 and p2 to q1
and q2 is det #### 12 46 #### , which is 8.
# #
We can also enumerate the disjoint-path systems in more detail by generat-
ing functions. When each edge on the line y = s has weight x s , for each s, the
coefficient of ∏ x ess counts the disjoint- X , Y-path systems having exactly e s hori-
zontal steps on y = s. This allows us to enhance the definition of h(p , q) via

h((a , b) , (c , d)) = ∑ ∏ x¾ .
i
b≤¾1 ≤···≤¾c− a ≤ d

Setting all x i = 1 yields the previous value (d−cb−+ac− a) that counts each path once.
In Example 4.1.27, the resulting generating function is

x12 x2 + x1 x22 + x0 x22 + 2x0 x1 x2 + x0 x12 + x02 x2 + x02 x1 .

Exercise 61 applies Corollary 4.1.33 to count paths avoiding a specified set of

lattice points. Other applications of Theorem 4.1.31 appear in Lindström [1973]
and Gessel–Viennot [1989].
As a striking application of the method, we present Aigner ’s proof of MacMa-
hon’s formula for the number of rhombic tilings of a regular hexagon of side-
length n. Cut the hexagon into 6n2 equilateral triangles; each rhombus consists
of two triangles with side-length 1 sharing an edge. Below we show the triangles
in a hexagon of length 3 and one of its rhombic tilings.
170 Chapter 4: Further Topics

4.1.35. Theorem. (MacMahon [1916]) The number of rhombic tilings of a

hexagon with side-length n is the determinant of the matrix with (n+2n
i− j ) in
(i
position , j).
Proof: (Aigner [2001]) Let L and R be the lower left and upper right edges of the
diagram. We interpret rhombic tilings as disjoint-path systems joining points on
L and R. Rhombi in a tiling have three possible orientations. Shade the rhombi
whose short diagonal is parallel to L and R. Each unshaded rhombus has two
edges parallel to L and R. The segment joining the centers of those edges can
be viewed as a step in a path from L to R. These paths naturally partition the
unshaded rhombi; the paths can be seen as climbing 3-dimensional piles of boxes
fixed in one coordinate.

R • y3 y2
• • y1
• • •
• • •
• • •
• • •
• • •
• • •
• • •
x3• • • •
•
L x2 •
x1

These edges belong to the graph on the right above; orient them from lower
left to upper right to form a digraph G. The edge crossing an unshaded rhombus
continues a path that enters it at the lower left, so the chosen edges form paths
from L to R. Each rhombus has only one entrance and one exit, so the paths are
disjoint. Thus the unshaded rhombi yield a disjoint-path system from x1 , . . . , x n
to y1 , . . . , yn in G.
The map is a bijection; each edge in a disjoint-path system bisects one rhom-
bus, which is the unshaded rhombus that generates it in the corresponding tiling.
(The path system has n2 steps up and n2 steps rightward, so every tiling has n2
rhombi with each orientation.)
Hence it suffices to count the disjoint-path systems in G. They all consist of
paths from x i to yi for i ∈ [n] and have positive sign. By Theorem 4.1.31, the num-
ber of disjoint-path systems is the determinant of the matrix whose (i , j)-entry
is the number of paths from x i to yj . This is (n+2ni− j
), the ordinary count of lattice
paths from (0 , 0) to (n + i − j , n + j − i).

4.1.36. Remark. Evaluation of combinatorial determinants. The reader may won-

der whether the determinant in Theorem 4.1.35 evaluates nicely. In fact, it does:
n−1 (2n+ i)(n)
i− j ), then det A = ∏ i=0 (n+ i)(n) , where as usual the
if the (i , j)-entry of A is (n+2n
parenthesized subscript indicates the falling factorial. When n = 3, the number
of tilings evaluates to 980.
Many methods for evaluating determinants arise combinatorially; Kratten-
thaler [1999, 2005] is a thorough two-part survey. Most of these methods are
beyond the scope of this text, but a simple method called the Condensation
Method works for some determinants including those in this application.
Exercises for Section 4.1 171

The method is often attributed to Dodgson [1866] (the clergyman known as

Lewis Carroll), but it may have been known earlier. The identity relates several
determinants, computing det A in terms of determinants of submatrices. For a
square matrix A of order n, with I , J ⊆ [n], let  I: J denote the matrix obtained
by deleting from A the rows indexed by I and the columns indexed by J. The
identity is
det A det Â1 ,n:1 ,n = det Ân:n det Â1:1 − det Ân:1 det Â1:n .
This yields inductive proofs of some determinant formulas. For n , a , b ∈ , 
let Mn(a , b) denote the n-by-n matrix whose (i , j)-entry is ( aa−+i+b j ). The result is
det Mn(a , b) = ∏ j =1 ∏¾=1 n+j +j +¾−¾−1 1 (see Exercise 69).
a b

Further applications of Theorem 4.1.31 appear in Exercises 61–65.

EXERCISES 4.1

4.1.1. (−) How many ways are there to place 10 distinct people within three distinct rooms?
How many ways are there to place 10 distinct people within three distinct rooms so that
every room receives at least one person? How many decimal n-tuples contain at least one
each of {1 , 2 , 3} (leading 0s allowed)?
4.1.2. (−) Two distinct dice are rolled (together) n successive times. What is the probabil-
ity that each roll of doubles (two 1s , . . ., two 6s) appears?
4.1.3. (−) 13 cards are chosen at random from a standard deck of 52 cards. Compute the
probability of each event below.
(a) At least one card in each suit.
(b) No cards in some suit.
(c) Some four cards with the same value.
4.1.4. (−) Given the special case º (∅) = ∑ S⊆[n](−1)| S| ½(S) of Theorem 4.1.3, prove the gen-
eral formula for º (T) by reducing it to the special case for º (∅).
4.1.5. (−) How many natural numbers in [252] are relatively prime to 252?
4.1.6. (−) How many natural numbers in [200] have no divisor in {6 , 10 , 15}?
4.1.7. (−) Given pennies, nickels, dimes, quarters, and half-dollars, how many ways can
one select n coins so that no type of coin is selected more than four times?
4.1.8. (−) How many permutations of [n] leave no odd number fixed?
4.1.9. (−) Use inclusion-exclusion to prove that ∑(−1)¾ (¾n)(n − ¾)n = n!.
4.1.10. (−) A 6-sided die is rolled repeatedly until the numbers 1 through 5 have appeared
at least once each. What is the probability that this happens in the first n rolls? Use a
calculator to find the least n where the probability exceeds .5.
4.1.11. (−) Compute the rook polynomial of the board below.
172 Chapter 4: Further Topics

4.1.12. (−) Compute the rook polynomials for all boards with at most four squares.
4.1.13. (−) Using Theorem 4.1.31 and the example of lattice paths with initial points
{(1 , 0) , (0 , 1)} and terminal points {(2 , 2) , (3 , 3)}, explain the role of the assumption in
Corollary 4.1.33 that disjoint-path systems occur only with the identity permutation. In
general, what happens when that condition fails?
4.1.14. Given fixed positive integers s1 , . . . , sn , let an ,¾ be the number of integer solutions
to 1 + · · · + n =  such that 0 ≤ i < ri for each i.
(a) Obtain the generating function ∑ an ,¾ x¾ (as a ratio of polynomials).
(b) Without using (a), compute an ,¾ as a finite sum.
(c) Explain why the answers to (a) and (b) are equivalent.
4.1.15. Prove that ∑¾|n () = n, where  is the Euler totient function. (Gauss)
4.1.16. (♦) Let  denote the Euler totient function.

(a) For m ∈ , prove that (m) = m ∏ p∈ P(m)(1 − 1/p), where P(m) is the set of distinct
prime factors of m.

(b) For n ∈ with n > min{m/2 , m − (m)}, prove that ∑i=1 (−1)i (ni) im is divisible by
n

m. (Hint: What does the sum count?) (Popescu [1999])

4.1.17. Let M be a multiset consisting of two each of n types of letters. When all arrange-
ments of M into a row are equally likely, what is the probability that no two consecutive
letters are the same?
4.1.18. Given n boys and n girls, use inclusion-exclusion to obtain sums for the number of
ways to pair the 2n people as lab partners under the following criteria. (No simple closed
formulas are available.)
(a) For each i, the ith tallest boy is not matched to the ith tallest girl.
(b) Same condition as (a) and also each pair has one person of each sex.
4.1.19. (♦) A mathematics department has n professors and 2n courses, each professor
teaching two each semester. How many ways are there to assign the courses in the fall
semester? How many ways are there to assign them in the spring so that no professor
teaches the same two courses in the spring as in the fall?
4.1.20. (♦) At a circular table are n students taking an exam. The exam has four versions.
Given that no two neighboring students have the same version and n > 1, how many ways
are there to assign the exams? Do not leave the answer as a sum.
4.1.21. (♦) Let an count the permutations of [n] with no cycles of length 2.
(a) Use PIE to obtain a formula for an (a summation).
(b) Use the Exponential Formula to obtain the same answer as in part (a).
(c) What is the asymptotic probability that a permutation has no 2-cycles?
4.1.22. (♦) Count the distinguishable ways to seat the people in n married couples at a
rotating round table with 2n seats so that no person sits next to his or her spouse. The
sexes need not alternate. (Leave the answer as a summation.)
4.1.23. (♦) Let S = {a1 , . . . , a¾ } ⊆ [n − 1], with a1 ≤ · · · ≤ a¾ . Count the permutations of
[n] whose descent set is contained in S. Use the resulting formula to count the permuta-
tions of [n] whose descent set equals S.
4.1.24. Prove Theorem 4.1.3 by substituting ####⋂i∈S Ai #### = ∑ Q⊇ S  (Q) into the formula and
simplifying. Explain how this proof relates to that in the text.
4.1.25. Prove Theorem 4.1.18 from Theorem 4.1.3 by algebraic computation.
4.1.26. Determine the number of graphs with vertex set [n] that have exactly m edges
and exactly  isolated vertices.
Exercises for Section 4.1 173

4.1.27. A number n ∈ 
is a power if it equals am for some a , m ∈ 
with m ≥ 2. For

S ⊆ , let l(S) be the least common multiple of the elements of S. For x ≥ 4, prove that
the number of powers in the interval [1 , x] is ∑S(−1)| S|+1 ⌊ x1/l(S) ⌋ , where the sum is over
all nonempty subsets of the integer interval [2 , ⌊ log 2 x⌋ ]. (Nyblom [2004])
4.1.28. (♦) Let t n be the number of triangles in the drawing of the equilateral triangular
grid with side length n, shown below for n = 3. Note that t1 = 1, t2 = 5, and t3 = 13.
Use PIE to derive a third-order recurrence for ⟨t⟩. Solve it by the characteristic equation
method. (Hint: Express the inhomogeneous term using exponentials.)


4.1.29. (♦) Given m , n , r ∈ 0 , evaluate each sum below using generating functions, and
give inclusion-exclusion proofs of the resulting identities. In part (c), m ≥ n.
n n
n n m− ¾
(a) ∑(−1)¾ ( )(n − ¾) (c) ∑(−1)¾ ( )( )
¾ ¾ r
¾ =0 ¾ =0
m n
n− ¾
n n r + n− ¾ −1
(b) ∑(−1)¾ ( )( ) (d) ∑(−1)¾ ( )( )
¾ m− ¾ ¾ r
¾ =0 ¾ =0

4.1.30. For n ≥ t, apply Theorem 4.1.3 directly to prove the identity below. (Hint: Con-
sider the universe of n-sets in [2n], and do not shift the index of summation. Comment:
Shifting the index and additional substitutions can turn this identity into part (c) of Ex-
ercise 4.1.29; instead we seek a direct proof.)
n
n− t 2n − r n
∑(−1)r−t ( r − t )( n − r ) = (n − t)
r= t


4.1.31. (♦) For n ∈ , use inclusion-exclusion to give a combinatorial proof of the identity
below. (Beckwith [2006a])
n
n 2n − 2 ¾
∑(−1)¾ (¾ )( n−1
)=0
¾ =0


4.1.32. (♦) For n , ¾ ∈ , evaluate the sum below, both by using generating functions and
by using inclusion-exclusion. (Cox–Thieu [2015])
¾
¾ ¾ n − in
∑(−1)i ( i )( ¾+1
)
i =0

4.1.33. Evaluate the sum below. Use inclusion-exclusion to prove the resulting identity.
r −1
r−1 n− ¾
∑(−1)¾ ( ¾
)(
r−¾
)
¾ =0

4.1.34. Let b , r, s , t be nonnegative integers such that r + s + t = rb > 0. Prove the identity
below using constrained compositions of integers. (Wardlaw [1989])
174 Chapter 4: Further Topics

⌊s/b⌋ ⌊t/b⌋
r s + r − 1 − b¾ r t + r − 1 − b¾
∑(−1)¾ (¾)( s − b¾ ) = ∑(−1)¾ (¾)( t − b¾ )
¾ =0 ¾ =0

# #
4.1.35. Let A1 , . . . , A¾ be finite sets. For J ⊆ [¾], let NJ = ###⋃j ∈ J Aj ### , and let Sm =
# #
∑| J|=m NJ . Let Tm be the number of elements that belong to exactly m of the sets
¾
A1 , . . . , A¾ . Prove T m = ∑i=1 (−1)¾+i+ m+1 (¾−i m) Si . (Sofair [2014])
4.1.36. Every tree with n vertices has n − 1 edges, and the cycle with n vertices has n
edges. Use the inclusion-exclusion formula to derive the chromatic polynomials of trees
with n vertices and cycles with n vertices.
4.1.37. In a proper coloring of a graph G, the vertices receiving a particular color must
form an independent set (a set of pairwise nonadjacent vertices). Hence another formula
for the chromatic polynomial is ∑r=1 ¾(r) pr(G), where pr(G) denotes the number of parti-
n

tions of V(G) into r independent sets.

(a) Let T be an n-vertex tree. Prove that pr(T) = S(n − 1 , r − 1), where S(m , ¾) is the
Stirling number of the second kind.
(b) Compute the chromatic polynomial of T from the information above. (Hint: Use a
result about Stirling numbers from Chapter 3.) (Voloshin [2002])
4.1.38. Prove that a graph G with n vertices and m edges has at most 13 (m 2)
cycles with
three edges. Conclude that the coefficient of ¾ n−2 in G () is positive unless | E(G)| ≤ 1.
4.1.39. (♦) Let an be the number of permutations of [n] such that no number i is fol-
lowed immediately by i + 1. Use PIE to derive a formula for an , and from it conclude
an = Dn + Dn−1 , where Dn counts the derangements of [n]. Explain how the first compu-
tation changes when we also forbid n to be followed by 1. (Comment: an = Dn + Dn−1 was
proved using recurrences in Exercise 2.1.44.) (Brualdi [2010, Section 6.5])
4.1.40. (♦) Let an be the number of permutations of [n] such that position i does not con-
tain element i + 1, for 1 ≤ i ≤ n − 1. Use Proposition 4.1.15 to prove an = Dn + Dn−1 , where
Dn counts the derangements of [n]. Use rook placements to give a direct proof. (Comment:
Exercise 3.1.27 compares this set and that in Exercise 4.1.39, having the same size.)
4.1.41. (♦) Without using derangement numbers, use Theorem 4.1.18 to count the permu-
tations of [n] having an even number of fixed points. From the resulting formula, find the
approximate probability (when n is large) that a random permutation of [n] has an even
number of fixed points. Is the number of fixed points more likely to be even or odd?
4.1.42. (♦) For n ≥ 2, let n be the polynomial defined below. Interpret [xr ]n(x) combina-
torially; that is, find a set that it counts. Use this to determine the degree of n . (Hint:
n−1
Consider the combinatorial meaning of ∏ j =0 (x + j).) (Caro–Pohoata [2011])

n i −1
n
n(x) = ∑ ( )(− x)n−i ∏(x + j)
i
i =0 j =0

4.1.43. (♦) Count the -subsets of [n − 1] having exactly r runs of consecutive elements.
Use this to obtain a double summation for the number of permutations of [n] in which no
two numbers differing by 1 are adjacent (in word form).
4.1.44. (+) Squares on an n-by-n chessboard alternate black and white. How many sets of
 white squares and n −  black squares have no two in a row or column? Interpret the
result in terms of permutations. (Andrews–Wang [1988])
Exercises for Section 4.1 175

4.1.45. (+) Let (j) denote the number of positions having 1 in the binary expansion of j .
2 n −1
Prove that ∑j =0 (−1)n− (j) j r is 0 when r < n and is n!2 n(n−1)/2 when r = n. (Callan [1999])

4.1.46. (♦) Prove the identity below, where s(m , ) is the signed Stirling number of the
first kind (recall that s(m , ) = (−1)n− c(n , ), where c(n , ) is the number of permutations
of [m] having  cycles). (Kauers [2011])
n n
2n
∑(−1) ( n+ 
) s(n +  , ) = ∏(2i − 1)
 =0 i =1

4.1.47. (♦) Stirling numbers and derangements.

(a) Use known formulas or EGFs to prove the identity below.
(b) Prove it also by counting a set in two ways. (Leuck [1987])
n n
n
∑  r () Dn− = n! ∑ S(r, m)
 =0 m=0

4.1.48. Prove the identities below.

n n−1
1 (n − j)n 1 (n − 1 − j)n n
(a) ∑(−1) j =1 (b) ∑(−1) j =( )
j ! (n − j)! j ! (n − 1 − j)! 2
j =0 j =0

4.1.49. (+) For n ∈  , prove ∑n=1 (−1) ¾ −1

(n)(n − )n = nn( 12 + 13 + · · · + 1n ). (Ungar [1984])
4.1.50. (♦) Let T be the triangular board with n rows of lengths 1 , . . . , n. For the
coefficients of the rook polynomial, prove both bijectively and recursively that r(T) =
S(n + 1 , n + 1 − ), where S(m , j) is the number of partitions of [m] into j blocks. (Hint:
Encode the positions of T as {(i , j): 1 ≤ i < j ≤ n + 1}.)
4.1.51. Let Rm ,n(x) be the rook polynomial for a full m-by-n board. Give combinatorial
arguments (bijections) to prove
(a) Rm ,n(x) = Rm ,n−1 (x) + mxRm−1 ,n−1 (x).
(b) dx
d
Rm ,n(x) = mnRm−1 ,n−1 (x).
4.1.52. (♦) Let B be a board whose squares form the Ferrers diagram of a partition  with
at most n parts (append 0s as needed). Prove ∑=0 r (B)x(n−) = ∏i=1 (x +  n+1 −i − i + 1). For
n n

example, with  = (3 , 1) and n = 3, both sides equal x2(x + 1). (Comment: The resulting
polynomial is called the factorial polynomial of B.) (Goldman–Joichi–White [1975])
4.1.53. (♦) Two boards are equivalent if their rook polynomials are equal. They are com-
plementary if they can be expressed as a partition of a rectangular board.
(a) For complementary boards B and B in an m-by-n rectangle, prove

n− j m− j
r (B ) = ∑(−1) j rj (B)( )( )( − j)! .
− j − j
j =0

(b) Prove that two boards are equivalent if they have equivalent complements in some
rectangular board. Conclude that the boards below are equivalent. Using the rook polyno-
mial, find a third board equivalent to them that cannot be obtained from them by rotations
and/or permutations of rows and/or columns.
176 Chapter 4: Further Topics

4.1.54. (+) For a board B, let R∗B(x) = ∑¾ r¾(B)x(n−¾) , where x(n−¾) denotes the falling
factorial. Let B and B be complementary boards in an n-by-n square. Prove R∗B (x) =

(−1)n R∗B(− x − 1). (Hint: For x ∈ 0 , add x extra rows to the square.) (Chow [1996])

4.1.55. (♦) Give three proofs of the identity below: by identities, by Snake Oil, and by
Theorem 4.1.18.
n
¾ n n
∑(−1)¾−p( p)(¾)2n−¾ = ( p) .
¾= p

4.1.56. A mathematics department has m students, of which n are male. All m students
want to take the combinatorics course, limited to l students and graded pass/fail. Assume
that l , n ≤ m. Define a signed involution on the set of all possible grade reports to prove
the identity below. (Comment: There are also other proofs.) (West–Wiedemann [1988])
l i
m− i n m− n m− n
∑ ∑(−1)j (m − l)( j )( i − j ) = 2 l( l
).
i =0 j =0

4.1.57. (♦) Let X be the set of all partitions of n. For i ∈ {0 , 1}, let pi(n) be the number
of elements of X whose number of even parts is congruent to i modulo 2. Prove that the
number of partitions of n into distinct odd parts equals p0(n) − p1(n). (Gupta [1976])

4.1.58. (♦) For a permutation of [n], the displacement of is ∑i=1 |i − (i)|.

n

(a) Prove that every permutation has even displacement.

(b) Let  (n , ) be the number of even permutations of [n] with displacement 2  minus
the number of odd permutations of [n] with displacement 2  . Prove  (n , ) = (−1)¾ (n−¾ 1 ) .
(Hint: Obtain a recurrence for  (n , ) from a signed involution on the set of permutations
of [n] with displacement 2  .) (Olmsted [1986])

4.1.59. (♦) Colored balls 1 through m are put in boxes 1 through  so that no box is empty
and no box has two balls with the same color. The coloring of the balls is fixed, and they
do not all have the same color. Let Pm() be the number of such arrangements. Prove
¾
∑¾=1 −¾1) Pm() = 0. (Schmuland [2013])
m (

4.1.60. (+) Given m , n , q ∈  such that 0 ≤ n − m < q ≤ n, prove

( + q)/2 − 1 n − ( + q)/2 n
∑ (
−1
)(
m− 
) ≡ ( ) (mod 2).
m
¾ ≡ q (mod 2)

(Hint: Define an involution on the set of arrangements of m unit squares and n − m domi-
noes in a row of length 2n − m.) (Knuth [1996])

4.1.61. (♦) Let T be a set of integer lattice points along a single lattice path from (0 , 0) to
(r, s), with (0 , 0) , (r, s) ∈
/ T . Let r,s(T) be the number of lattice paths from (0 , 0) to (r, s)
that avoid all of T .
(a) Use Corollary 4.1.33 to compute r,s(T).
(b) Use inclusion-exclusion to compute r,s(T).
(Comment: Together, (a) and (b) provide a combinatorial proof of an identity.)

4.1.62. (♦) Use Theorem 4.1.31 to prove that det A = 0 when a matrix A has linearly de-
n−1
pendent rows u1 , . . . , un . (Hint: Given un = ∑i=1 ci ui , form a digraph with vertex set
{x1 , . . . , x n} ∪ {y1 , . . . , yn} having edges xi yj for (i , j) ∈ [n − 1] × [n] and edges x n xi for
i ∈ [n − 1]. Define edge weights appropriately.) (Aigner [2001])
Exercises for Section 4.1 177

4.1.63. A path diagram is a Ferrers diagram with dots replaced by numbers such that the
number in position (i , j) is the number of paths from the bottom of column j to the right
end of row i, moving one box up or right at each step, always staying within the diagram.
Let M be the largest square matrix containing the upper left box (in bold below).

22 10 3 2 1 1
6 3 1 1 1
3 2 1
1 1 1

(a) Prove that for every path diagram, det M = 1. (Hint: Use Theorem 4.1.31.)
(Carlitz–Roselle–Scoville [1971]; Aigner [2001])
(b) Prove that the Catalan numbers form the only sequence ⟨a⟩ of real numbers such

that for all n ∈ 0 ,

⎛ a0 a1 · · · an ⎞ ⎛ a1 a2 · · · an+1 ⎞
⎜ a1 a2 · · · an+1 ⎟ ⎜ a2 a3 · · · an+2 ⎟
det ⎜ . .. .. .. ⎟ = det ⎜ .. . .. .. ⎟ = 1.
⎜ .. . . . ⎟ ⎜ . .. . . ⎟
⎝a an+1 · · · a2n ⎠ ⎝a an+2 · · · a2n+1 ⎠
n n+1

(Hint: Apply part (a). Comment: Matrices of this form are called Hankel matrices. Re-
lated work appears in Kellogg [1997] and Radoux [1997].)
4.1.64. (♦) For m ≥ 0 and n ≥ 1, prove the determinant identity below, where Sn ,¾ is the
Stirling number of the second kind. (Hint: Use a weighted digraph like that shown on
the right for the case (m , n) = (3 , 4). The diagonal edges have weight 1 and the horizontal
edges at height ¾ have weight ¾ .) (Aigner [2001])
• • • y4 = yn
•
· · · Sm+1 ,n ⎞ • • • • y3
•
⎛ Sm+1 ,1 Sm+1 ,2
⎜ Sm+2 ,1 Sm+2 ,2 · · · Sm+2 ,n ⎟ • • • • • y2
•
det ⎜ .. ⎟ = (n!)m
⎜ ... ..
.
..
. . ⎟ • • • • • • y1
•
⎝S · · · Sm+n ,n ⎠
m+ n ,1 Sm+ n ,2 • • • • • • • •
x n = x4 x3 x 2 x1 0 1 2 3=m


4.1.65. (+) Let S = {a1 , . . . , an} ⊆ . The GCD matrix A of S has (i , j)-entry gcd(ai , aj ).
Prove that if S is closed under taking divisors, then det A = ∏i=1 (ai), where is the
n

Euler totient function. (Hint: Form a digraph using three copies of S. For elements q , r ∈
S with q | r, add edges from r1 to q2 and from q2 to r3 , with w(r1 q2) = (q) and w(q2 r3) = 1.
Apply Theorem 4.1.31. Comment: The formula is due to Smith [1876]; this proof is from
Altinisik–Sagan–Tuglu [2005], generalized in Exercise 15.2.35.)

4.1.66. (♦) Monotone trees and alternating permutations. Let Mn be the set of n-vertex or-
dered trees with vertex labels [n] such that each vertex has at most two children and labels
increase along paths from the root. For T ∈ Mn with left subtree T1 , right subtree T2 ,
and root label a, recursively define  (T) to be the concatenation  (T1) , a ,  (T2).
(a) Prove that  is a bijection from Mn to the set of permutations of [n].
(b) When  = (a1 , . . . , an), the element ai is a climb, slide, peak, or valley if the values
move up-up, down-down, up-down, or down-up, respectively, in the triple (ai−1 , ai , ai+1).
Characterize each type in terms of whether ai has left and/or right children in  −1().
(Set a0 = an+1 = 0, so a1 is a climb or a peak, and an is a slide or a peak.) (Stanley [1986])
(c) Let n be odd. In an alternating permutation of [n], odd positions are peaks and even
178 Chapter 4: Further Topics

positions are valleys. Characterize the elements of Mn mapped to alternating permuta-

tions by º . Prove that [n] has exactly | d n − e n | alternating permutations, where [n] has
d n permutations with an odd number of ascents and e n with an even number (d3 − e3 =
2, as shown below). (Hint: Using part (b), define an appropriate involution.) (Foata–
Schützenberger [1970, 83–84])
permutation 123 132 213 231 312 321
# of ascents 2 1 1 1 1 0
alternating? no no yes no yes no

4.1.67. (♦) Let G be the set of graphs with vertex set [n] and m edges. A graph is even if
every vertex has even degree. Let wn ,m be the number of even graphs in G. Let X be the
set of pairs (G , S) such that G ∈ G and S ⊆ V(G).
(a) Define a signed involution on X for which the set of fixed points consists of all pairs
(G , S) in which G is an even graph.
(b) Let a¾ ,l be the number of graphs in G such that l of the m edges join [¾] to [n] − [¾].
Use the involution in part (a) to prove wn ,m 2 n = ∑¾=0 (n¾) ∑ l=0(−1)l a¾ ,l .
n m

(c) Obtain the generating function for even graphs on [n] by number of edges:
n ¾(n−¾)
n 1−x
∑ wn ,m x + x)(2) ∑ ( ) (
n
− n(1
m
=2 ) .
¾ 1+x
¾ =0

4.1.68. Let A be a square matrix in which every entry not in the first row or column is
the sum of the two entries immediately above it and to its left. Prove that if every entry
in the first column of A is 1, then det(A) = 1. (Comment: The special case for binomial
coefficients is in Rupp [1930].) (Ionin [2014])
4.1.69. (+) The Condensation Method for determinants (Remark 4.1.36).
(a) For an n-by-n matrix A, let  I: J denote the matrix obtained by deleting from A
the rows indexed by I and the columns indexed by J. Prove
det A det Â1 ,n:1 ,n = det  n:n det Â1:1 − det  n:1 det Â1:n . (Dodgson [1866])
(Comment: Zeilberger [1997] gave a bijective proof.)

(b) For n , a , b ∈ , let Mn(a , b) be the n-by-n matrix with (i , j)-entry (aa−+i+b j ). Use part
n+ j +¾ −1
(a) to prove det Mn(a , b) = ∏ j =1 ∏¾=1
a b
j + ¾ −1
(see Zeilberger [1996]). (MacMahon [1916])

4.2. P ólya–Redfield Counting

When we count “distinguishable” objects, the objects are grouped into classes
by an equivalence relation, and we count these classes. For example, we can form
a string of n beads drawn from ¾ types in ¾ n ways, but when we join the ends to
make a necklace, the necklaces obtained from it by cyclic rotations and flips are
indistinguishable from it.
A classical application counts the isomorphism classes of n-vertex graphs.
There are 2(2) graphs with vertex set [n], but two graphs are isomorphic when a
n

vertex permutation turns one into the other. Since bijections can be composed or
inverted, isomorphism is an equivalence relation. To view this as a “distinguish-
able colorings” problem, we view vertex pairs as colored by “edge” or “non-edge”.
Two simple examples introduce the technique.
Section 4.2: P ólya–Redfield Counting 179

4.2.1. Example. Painted batons. A baton is partitioned into n bands of equal

length that can be painted with any of ¾ colors; there seem to be ¾ n paintings.
However, the baton can be flipped without changing it, so most equivalence classes
of colorings are listed twice. Symmetric colorings appear only once. Therefore,
before dividing by 2, we add the number of symmetric colorings to make each
equivalence class be counted exactly twice. The number of distinguishable paint-
ings is thus (¾ n + ¾ ⌈n/2⌉)/2.

4.2.2. Example. Tri-cornered hats. A clown with a three-cornered hat wants

to pin a flower to each corner. Given ¾ types of flowers, how many distinguish-
able patterns exist? There are ¾ 3 colorings of the fixed hat, but most equivalence
classes appear three times in the list, since the hat can rotate to turn one pattern
into another. The ¾ monochromatic patterns appear only once. We add 2 ¾ so that
those classes also are counted thrice. The number of patterns is (¾ 3 + 2 ¾)/3.

In these examples, the equivalence classes have different sizes, so we cannot

just divide the total number of objects by the size of the classes. The idea in gen-
eral is to adjust the counting so that each equivalence class contributes the same
amount and then divide by that amount.

BURNSIDE’S LEMMA

Equivalence classes of colorings result from “symmetry operators” that per-

mute the things being colored. When we color pieces of a physical object, a sym-
metry of the object is a rigid motion that leaves it occupying the same position.
Typically, a symmetry permutes the pieces of the object being colored. For ex-
ample, a regular n-gon has 2n symmetries in space; n rotations and n reflections
(flips). These operations permute the vertices or edges.
Symmetries can be inverted and can be composed. Under composition, the
symmetries of an object form a group.

4.2.3. Definition. A group is a set G together with a composition law satisfying

the four properties below. Here we write composition as multiplication.
(1) identity: there exists e ∈ G such that e = = e for all ∈ G.
(2) inverse: for ∈ G, there is a unique −1 ∈ G with −1 = e = −1 , where
e is the identity element.
(3) closure:  ∈ G for all ,  ∈ G.
(4) associativity: (  ) = () for all ,  ,  ∈ G.

Two colorings of a set X are equivalent if we can move the underlying object
to turn one coloring into the other. Such a motion permutes X .

4.2.4. Definition. Given a set Y of colors, a coloring of X is a function  : X →

Y . Given a group G of permutations of X , colorings  and  are indistin-
guishable or equivalent if there exists ∈ G such that  ( (x)) =  (x) for
all x ∈ X . We then say that turns  into  .
180 Chapter 4: Further Topics

4.2.5. Remark. Note that “ turns  into  ” means that  assigns to (x) the
same color that  assigns to x, for all x ∈ X . Using identity, inverse, and closure
for G, indistinguishability is reflexive, symmetric, and transitive. Thus it is an
equivalence relation on the colorings; we want to count the equivalence classes.
Symmetries of an object permute the set X being colored and thereby per-
mute the set C of colorings. The group structure is the same: the permutation
of C induced by composing and  is the composition of the permutations of C
induced by and  . Thus we may view ∈ G as a symmetry operator, a permu-
tation of X , or a permutation of C. We use the same notation in each context.
When G is a group of permutations of a set C, and u ∈ C, the orbit of u is
{ (u): ∈ G}. The defining properties of a group imply that the orbits partition
C, and “in the same orbit ” is an equivalence relation. In our model, the orbit of
a coloring is the set of colorings equivalent to it.
When all elements of G map a coloring  to distinct colorings, the orbit of 
has size | G | , and dividing by | G | counts it as one class. The orbit is smaller when
two elements of G take  to the same  . Since every element of G maps  to
some element of its orbit, summing the number taking  to each element of its
orbit again counts the orbit | G | times, and then dividing the total by | G | counts
it once. A classical lemma of group theory will simplify the computation.

4.2.6. Lemma. If G is a group of permutations of C, and u , v ∈ C are in the same

orbit, then |{ ∈ G: (u) = v}| = |{ ∈ G: (v) = v}|.
Proof: Let A = { ∈ G: (u) = v} and B = { ∈ G: (v) = v}. We establish
injections in both directions. Since u , v are in the same orbit, there is some  ∈ G
such that  (u) = v; we use this fixed  to define the injections. For each ∈ B, the
composition  takes u to v. For each ∈ A, the composition  −1 takes v to v.
This defines functions from B to A and from A to B. Both are injective, because
for  fixed we can multiply 1  = 2  on the right by  −1 to obtain 1 = 2 .

4.2.7. Lemma. (Burnside’s Lemma). Under the action of a group G of per-

mutations of C, the number of orbits of C is | G1 | ∑ ∈ G ( ), where ( ) is the
number of elements of C left fixed by the action of on C.
Proof: We count 1 for each orbit (equivalence class) by counting | G | for each orbit
and dividing the total by | G |. Choose a distinguished element uE from each orbit
E. Apply each in G to uE , and group the elements of G according to the image
(uE) ∈ E. Thus | G | = ∑v∈ E |{ : (uE) = v}|. Using Lemma 4.2.6, the number of
operators in G taking uE to any element v equals the number of operators fixing
v, which we write as (v). The computation of the number of orbits becomes
1 1 1 1
∑ 1 = | G| ∑ | G|= | G| ∑∑ |{ : (uE) = v}| =
| G| ∑
(v) =
| G| ∑
( ).
E E E v∈ E v∈ C ∈G

The sum ∑v∈ C (v) counts all pairs (v , ) such that (v) = v, grouped by the color-
ings. The sum ∑ ∈ G ( ) counts the same set, grouped by the operators. We just
interchanged the order of summation.
Section 4.2: P ólya–Redfield Counting 181

The last step in Lemma 4.2.7 is crucial for computations. We study questions
where the set X and group G acting on X are fixed, but the color set Y and thus
C varies. The computations for the baton and hat in Examples 4.2.1–4.2.2 were
those of the final step in Burnside ’s Lemma.
Although Lemma 4.2.7 is generally known as Burnside ’s Lemma, Wikipedia
says: “ William Burnside stated and proved this lemma, attributing it to Frobe-
nius [1887] in his 1897 book on finite groups. But even prior to Frobenius, the
formula was known to Cauchy in 1845. In fact, the lemma was apparently so well
known that Burnside simply omitted to attribute it to Cauchy. Consequently,
this lemma is sometimes referred to as the lemma that is not Burnside ’s.”
When counting colorings left fixed by , we use the observation that fixes
a coloring if and only if for all x ∈ X it assigns the same color to x and  x. We
impose this requirement one cycle at a time.

4.2.8. Lemma. In the set of all colorings of X from a set of  colors, the number
() of colorings left fixed by the action of  on X is  t , where t is the number
of cycles created by  on X .
Proof: To be fixed by  , a coloring must give the same color to all elements in a
cycle produced by  on X . There are  choices for that color, and the choice on
each cycle is made independently of the other choices.

4.2.9. Example. 4-bead necklaces. We model the problem as coloring X (the cor-
ners of a square) from a set of  colors. The square has eight symmetries; four
rotations and four reflections. Rotations of 0◦ , 90◦ , 180◦ , and 270◦ have 4 , 1 , 2 , 1
cycles, respectively, as permutations of the corners. Non-diagonal flips have two
cycles of length 2; diagonal flips have one cycle of length 2 and two of length 1.
Among the eight permutations, 1 , 2 , 3 , 2 have 4 , 3 , 2 , 1 cycles, respectively.
By Burnside ’s Lemma, there are 18 ( 4 + 2  3 + 3 2 + 2 ) distinguishable colorings
of X . When  = 2, the answer is 6. There are two monochromatic necklaces, two
where one color occurs once, and two with two beads of each color.
We can add constraints by applying Burnside ’s Lemma to restricted sets of
colorings. If adjacent beads must have distinct colors, then the only operators
that fix legal colorings are the identity, the 180◦ rotation, and the diagonal flips,
fixing two legal colorings each. Burnside ’s Lemma then yields 18 (2 + 2 + 2 + 2) as
the number of legal 4-bead necklaces.

Dividing by | G | when using Burnside ’s Lemma provides a computational

check. The number of equivalence classes is an integer. If the computation pro-
duces a non-integer, then mostly likely some operator or some colorings fixed by
an operator have been overlooked.

THE PATTERN INVENTORY

By recording how many times each color is used, we can count the distin-
guishable patterns using each color a specified number of times. For this we need
the lengths of the cycles in the permutations of X . We represent a cycle of length
j by the variable x j .
182 Chapter 4: Further Topics

e
4.2.10. Definition. The cycle structure of a permutation on a set is ∏ x j j ,
where has e j cycles of length j . The cycle index ZG(x1 , x2 , · · ·) of a group
G of permutations is | G |−1 times the generating function that enumerates G
e
by cycle structure, meaning that the coefficient of ∏ x j j counts the elements
ej
of G having cycle structure ∏ x j as permutations of the set being colored.

4.2.11. Example. We have applied Burnside ’s Lemma to count colorings of the

vertices of a rotating triangle and a square in space (Examples 4.2.2–4.2.9). For
the rotating triangle, the cycle index is 13 (x31 + 2x13). For 4-bead necklaces, it is
8 1 + 2x4 + 3x 2 + 2x1 x 2
1
(x4 2 2 ).
We just record, for each element of the group, the lengths
of the cycles in the corresponding permutation of the objects being colored.
In the cycle structure ∏ x j j for a permutation of X , note that ∑ je j = | X | ,
e

because each element of X appears in one cycle.

For cycle indices, Burnside ’s Lemma states the following:

4.2.12. Corollary. If G is a group of permutations of X , then the number of

distinguishable -colorings of X under G is ZG( ,  ,  , · · ·).
Proof: Setting x j =  for each j accumulates ( ) for each ∈ G; we then divide
by | G | to count one for each class.

We now want to record how many times each color is used.

4.2.13. Definition. Let Y = {y1 , . . . , y¾ } be a set of colors. The pattern in-

ventory of colorings of X is the OGF in which the coefficient of ∏ yiei is the
number of equivalence classes of colorings that for each i have color yi on e i
elements of X .

The pattern inventory records the options instead of just the number of op-
tions. Instead of setting x j =  in the cycle index, we let x j list the  options
“ j blacks or j reds or j greens . . .” To do this, we set x j = y1j + y2j + y3j + · · · ,
in the usual way of building generating functions to model the listing of alter-
natives. The choice from the expression for x j records a contribution of j to the
exponent of the chosen color. This represents using that color on all j vertices of
the corresponding cycle.

4.2.14. Example. Pattern inventory for 4-bead necklaces. Setting x j = b j + r j in

the cycle index of Example 4.2.11 yields the pattern inventory
1 [(b
8 + r)4 + 2(b4 + r4) + 3(b2 + r 2)2 + 2(b + r)2(b2 + r 2)]
for distinguishable necklaces with blue and red beads. This simplifies to b4 + b3 r +
2b2 r 2 + br3 + r4 . With two beads of each color we have two patterns: alternating
or non-alternating.

4.2.15. Example. A cube has 24 “symmetries”, meaning rigid rotations that

leave it occupying the same space. Other than the identity, there are three rota-
tions about the axis through a pair of opposite faces, one about the axis through
Section 4.2: P ólya–Redfield Counting 183

a pair of opposite edges, and two about the axis through a pair of opposite ver-
tices. The groups we consider when coloring the vertices, faces, or edges of a cube
are isomorphic (each corresponds to the rigid motions of the cube), but the cycle
structures are different, because the set X being colored is different. The cycle
structures are listed below, with “multiplicity” being the number of motions of
the type indicated.
Axis Degrees Mult. Vertices Faces Edges
anywhere 0 1 x81 x61 x12
1
opp. faces 90,270 6 x42 x12 x4 x34
opp. faces 180 3 x42 x12 x22 x62
opp. verts. 120,240 8 x12 x32 x32 x43
opp. edges 180 6 x42 x32 x12 x52
The resulting cycle indices are
1
(x8
24 1 + 6x42 + 9x42 + 8x12 x32) for coloring vertices,
1 (x6
24 1 + 6x12 x4 + 3x12 x22 + 8x32 + 6x32) for coloring faces,
1
(x12
24 1 + 6x34 + 3x62 + 8x43 + 6x12 x52) for coloring edges.
Setting each x j to ¾ counts the equivalence classes of colorings when ¾ colors
¾
are available. Setting x j = ∑ i=1 yij forms the pattern inventory listing equiva-
lence classes by how many times each color is used.

Using Burnside ’s Lemma, we can count isomorphism classes of graphs with n

vertices. With the pattern inventory, we can count them by the number of edges.

4.2.16. Application. Isomorphism classes of n-vertex graphs. Graphs G and H

with the same vertices are isomorphic if some permutation of the vertices turns
G into H. That is, uv ∈ E(G) if and only if (u) and (v) are adjacent in H.
A graph G is formed by deciding for each pair of vertices whether they are
adjacent or not. This yields two colors: “edge” and “non-edge”. A permutation of
the vertices can turn one coloring into another by inducing a permutation of the
vertex pairs. The cycle index for the counting of classes is always for a group of
permutations of the objects being colored.
The group of permutations of the vertices is Ën . These permutations induce
n! permutations of the set X of vertex pairs. The resulting group is the pair
group Ën . In Example 4.2.25 we describe it for general n. Meanwhile, the table
(2)

below shows the computation of ZË(2) .

4

cycle form of # of this Induced edge perm. Induced

vertex perm. type 12 13 14 23 24 34 cycle struc.
(1)(2)(3)(4) 1 12 13 14 23 24 34 x61
(1234) 6 23 24 21 34 31 41 x4 x 2
(123)(4) 8 23 21 24 31 34 14 x32
(12)(34) 3 21 24 23 14 13 43 x12 x22
(12)(3)(4) 6 21 23 24 13 14 34 x12 x22
184 Chapter 4: Further Topics

The cycle index of Ë4 is 24

(2)
1 + 6x 2 x4 + 8x3 + 9x1 x 2 Setting each x i = 2
1 (x6 2 2 2).

yields 11 as the number of isomorphism classes with four vertices.

• • • • • • • • • •
• • • • • • • • • • • •
• • • • • • • • • • • •
• • • • • • • • • •
Setting x j to 1 + y models the choices in constructing graphs invariant under
j

so that the number of edges accumulates in the exponent. We can choose no

edges (y0) or j edges on a cycle consisting of j pairs. There is such a factor for
each cycle of the permutation, and multiplying them yields a contribution of 1 to
the coefficient of y ¾ for each graph with  edges that is unchanged by .
Thus ( ) is a generating function for the graphs left fixed by , indexed by
the number of edges. Summing over all to count each isomorphism class once
accumulates those with  edges in the coefficient of y ¾ .
For example, there are three isomorphism classes with three edges:

[y3 ] 24
1
((1 + y)6 + 6(1 + y 2)(1 + y4) + 8(1 + y3)2 + 9(1 + y)2(1 + y 2)2 )
= 1
24 ((63) + 6 · 0 + 8(12) + 9(12)(12)) = 3.

The argument of Application 4.2.16 works for all n.

4.2.17. Corollary. (Redfield [1927], P ólya (see Harary [1955b])) The number of
isomorphism classes of n-vertex graphs with m edges is the coefficient of y m
in the OGF obtained by letting x j = 1 + y j for each x j in the cycle index ZË(2)
n
of the pair group on n vertices.

Application 4.2.16 uses one color of weight 1 (edge) and one color of weight 0
(non-edge). We may have several colors of each weight.

4.2.18. Definition. Given ai options with weight i, the generating function

∑i≥0 ai y i is the weight enumerator w(y). When options have various
weights in r colors, use w(y1 , . . . , yr): the coefficient on each monomial counts
the options where the weights on each color are given by the exponents.

Weight enumerators yield a common generalization of the pattern inventory

arguments we have given for necklaces and for isomorphism classes of graphs.

4.2.19. Example. Given three types of paint that cost one dollar per use and an-
other that costs two dollars, the weight enumerator (by cost) is w(y) = 3y + y 2 .
To enumerate distinguishable colorings by total cost, set x j = w(y j ) in the cycle
index for each j , because using a color on a j-cycle uses it j times.
Similarly, with two types of red paint and two types of blue paint, costing $2
or $4 per face, let w(r, b) = r 2 + r4 + b2 + b4 . The coefficient of r i b j in the pattern
inventory counts the distinguishable paintings using $i of red paint and $ j of
blue paint. To compute it, set x j = w(b j , r j ) in the cycle index.
Section 4.2: P ólya–Redfield Counting 185

We summarize the general discussion in the following theorem. We have es-

sentially explained the proof; it uses Burnside ’s Lemma to argue that each class
of colorings has been counted and assigned to the correct weight.

4.2.20. Theorem. (P ólya’s Theorem; P ólya [1937], Redfield [1927]) Let G be a

group of permutations of X , and let w be a weight enumerator for colors. The
OGF for equivalence classes of colorings of X under action of G, by weight
in each color, is obtained from the cycle index ZG by setting each x j to the
expression obtained from w by replacing each variable by its jth power.

CL ASSICAL CYCLE INDICES

Various classical groups arise in using the P ólya–Redfield Theorem.

4.2.21. Example. The cyclic group. Counting distinguishable jeweled crowns

uses the cycle index of the cyclic group Cn , the group of rotations of an n-gon.
With elements n , the jth element j takes i to (i + j) (mod n) and has d cycles
of length n/d, where d = gcd(j , n). Its cycle structure is x dn/d . The values j with
gcd(j , n) = d are multiples of d; we have gcd(id , n) = d when gcd(i , n/d) = 1.
Thus there are (n/d) permutations with cycle structure x dn/d , where () counts
the numbers in [] that are relatively prime to  (the Euler totient—Proposition
4.1.4). For ease of application, we exchange d and n/d in the sum to write
1
n∑
ZCn = (d)x n/d
d .
d|n

4.2.22. Application. Crowns with r missing jewels. A crown with  places for
diamonds is missing r of them; how many distinguishable ways can this happen?
We are asking for the number of distinguishable cyclic arrangements of r 1s and
 − r 0s under rotational symmetry.
The answer also counts the cycles in Bulgarian solitaire on partitions of n
when n = (¾2) + r (Theorem 3.4.18). The “move” in Bulgarian solitaire maps the
set of Ferrers diagrams with n dots to itself. We showed that the Ferrers diagrams
lying on cycles in the resulting functional digraph consist of  − 1 full slants plus
r dots in the slant of size  . Each move shifts these dots one step down the slant,
cyclically. Thus the number of cycles is again the number of cyclic arrangements
of r 1s and  − r 0s.
Applying Example 4.2.21, our choices for x d in the pattern inventory are y0
d
or y to accumulate contributions by the number of 1s. Since the cycle structures
of permutations in Cn have the form x n/dd , we obtain contributions to the desired
term only when d divides both  and r. Setting x d = (1 + y d)¾/d for all d, contri-
butions to the coefficient of y r arise when we take r/d of the /d cycles to have 1s.
Using the Binomial Theorem, the number of cyclic arrangements is
1 /d
 ∑ (d)(
r/d
).
d|gcd(¾ ,r)
186 Chapter 4: Further Topics

4.2.23.* Example. The dihedral group. To count distinguishable necklaces, we

need the cycle index for the dihedral group Dn , the group of symmetry opera-
tors on an n-gon in space. It has the same rotational operators as Cn plus n “flips.”
We divide the cycle index for Cn by 2 and add the contributions for the flips. If n
is odd, then every flip uses an axis between a vertex and the opposite edge, so each
consists of a fixed point and (n − 1)/2 pairs forming 2-cycles. If n is even, then the
n/2 flips around an axis through opposite edges consist solely of 2-cycles, while
the n/2 flips around an axis through opposite vertices have two fixed points. Thus
ZCn + 12 x1 x 2 −1)/2
1 (n
when n is odd
ZDn = { 12 1 2 n/2−1
2 ZCn + 4 x 2 + 4 x1 x 2
1 n/2
when n is even.

4.2.24.* Example. Symmetric group. The symmetric group Ën is the group of

all permutations of [n]. A cycle structure ∏¾ x¾e¾ corresponds to the partition of
n having e¾ parts of size  . To count the permutations with this cycle structure,
view the cycles linearly as
(− − − − − − − − −)(− − − − −)(−−)(−−),
in decreasing order of length, waiting to be filled. There are n! ways to put [n]
in the blanks. Each permutation with this structure arises in ∏  e¾ e¾ ! ways, be-
cause (1) a cycle of length  has  different starting points, and (2) cycles of the
same length can appear in any order. Hence n!/ ∏  e¾ e¾ ! permutations have cycle
structure ∏¾ x¾e¾ . Letting  n indicate a sum over all partitions of n,
x¾e¾
ZËn = ∑ ∏ .
ën ¾
 e¾ e¾ !

4.2.25.* Example. The pair group. The pair group Ën is the group of permu-
(2)

tations of the pairs in [n] induced by Ën . We used this in Application 4.2.16 and
Corollary 4.2.17. The cycle structure of the permutation in Ën induced by  ∈ Ën
(2)

depends only on the cycle structure for  . Pairs are of two types: those contained
in one cycle of  and those consisting of elements from two cycles in  .
For two elements on the same cycle,  shifts each by one along the cycle, leav-
ing them separated by the same distance along the cycle. Thus a cycle of length
 in  generates ⌊( − 1)/2⌋ cycles of length  among the pairs contained in it.
When  is even it also generates one cycle of length /2; pairs with displacement
/2 return to themselves after /2 steps.
A pair of elements from distinct cycles of lengths r and s repeats after
lcm (r, s) iterations of  . The induced permutation has gcd(r, s) cycles of length
lcm (r, s) on the rs pairs chosen from the two cycles.
Thus when  has cycle structure ∏¾ x¾e¾ , the induced permutation in Ën has
(2)

cycle structure
e ¾ (¾ −1)/2 (¾/2−1) e ¾ (e2¾)¾
∏ x¾ ∏ x¾e¾ ∏ x¾ ∏ xlcm (r,s)
e r e s gcd(r,s)
x¾/2 .
odd ¾ even ¾ ¾ r
= s
Exercises for Section 4.2 187

EXERCISES 4.2

4.2.1. (−) A game is played on a table with 15 balls in a triangular array. The array can
slide on the table and rotate. There is a supply of three types of balls. How many distin-
guishable ways are there to arrange balls in the array?
4.2.2. (−) The clown of Example 4.2.2 wants to put flowers also on the edges of the tri-
cornered hat; the same ¾ types are available, and we count distinguishable patterns under
rotation. The number of patterns on edges alone is the same as the number on vertices
alone. Why is the number of patterns with flowers on both the edges and the vertices not
equal to the square of this?
4.2.3. (−) The n bands of a baton (Example 4.2.1) are to be painted under the additional
requirement that adjacent bands must not have the same color. With ¾ colors available,
how many distinguishable ways are there to paint the baton?
4.2.4. (−) Given that n is prime, count the distinguishable n-bead necklaces that can be
formed when ¾ colors of beads are available.
4.2.5. (−) For n = 6, n = 9, and n = 10, count the n-bead necklaces that can be formed
using three colors of beads.
4.2.6. (−) Determine the cycle indices for coloring the vertices, edges, or faces of a tetra-
hedron that can rotate in space.
4.2.7. (−) Determine the pattern inventory for colorings of the squares of a rotating 4-by-4
chessboard with two colors available.
4.2.8. (−) Let c(x) be the OGF for isomorphism classes of connected graphs, indexed by
number of vertices. Find the OGF for isomorphism classes of graphs with two components.
4.2.9. (−) A double domino is a solid 2-sided chip of wood such that each side consists
of two squares sharing an edge. Each of the four squares is marked with a number of dots
between 0 and ¾ − 1, inclusive. Determine the number of distinguishable double dominoes.
4.2.10. The squares of a 4-by-4 chessboard are to be painted (front only), with four colors
available for each square. Determine the number of distinguishable paintings that have
four squares of each color.
4.2.11. Determine the cycle index for coloring the vertices and the edges of an n-gon that
can rotate or flip (movement in space around any axis).
4.2.12. (♦) Caterpillars are trees in which the deletion of all leaves leaves a path. Prove
that the number of isomorphism classes of n-vertex caterpillars is 2 n−4 + 2⌊n/2⌋−2 if n ≥ 3.
(Harary–Schwenk [1973], Kimble–Schwenk [1981])
4.2.13. Lemma 4.2.6 states that when a group G acts on a set S, the number of permuta-
tions that map x to y is the same for all y in the orbit of x. Use this to count the ways to
pair up 2n people. (Hint: Let S be the set of pairings.)
4.2.14. (♦) Partitions of an integer n are obtained from compositions of n by ignoring the
order of the parts. Use Burnside’s Lemma and the symmetric group Ë3 to count the par-
titions of 9 into three parts. List the partitions explicitly to check the computation.
4.2.15. (♦) Let p , n , l be positive integers with p prime. Use Burnside’s Lemma to prove
l −1
that np − np is divisible by pl . (Chappell–Hartman)
l

4.2.16. With four colors of beads available, how many distinguishable 6-bead necklaces
are there in which no two adjacent beads have the same color?
188 Chapter 4: Further Topics

4.2.17. (♦) Consider distinguishable n-bead necklaces formed from three colors of beads.
Use the methods of this section to answer the following questions.
(a) For n = 4, how many have no two consecutive beads of the same color?
(b) For n = 6, how many have no two of the three colors used equally often?
(c) For n = 9, how many use all three colors three times each?
4.2.18. With three colors of beads available, how many distinguishable n-bead necklaces
are there with each color used at least once, for n = 4 and n = 7?
4.2.19. Determine the number of isomorphism classes of 4-vertex loopless multigraphs in
which each pair of vertices occurs as the set of endpoints of at most two edges. Use the
pattern inventory to compute the number of these classes that have six edges. Draw them
explicitly to check the computation.
4.2.20. (♦) A financial company is designing a new symbol: a regular tetrahedron built
from six metal bars of equal length. Given that the bars can be copper, silver, or gold,
determine the number of distinguishable ways there are to do this (the tetrahedron can
rotate in space but not reflect). How many ways are there using two bars of each type?
4.2.21. (♦) Pattern inventories for cubes.
(a) Count the distinguishable 17-dollar cubes that can be built using (as edges) $1 lead
bars, $2 silver bars, and $3 gold bars.
(b) Determine the number of distinguishable ways to paint the faces of a cube with
three colors so that each color is used twice. Describe them geometrically.
4.2.22. Use P ólya’s Theorem to count isomorphism classes of spanning subgraphs of K3 ,3 .
4.2.23. Use P ólya’s Theorem to count the isomorphism classes of 5-vertex graphs with
four edges.
4.2.24. Determine the number of terms in the pattern inventory for n-bead necklaces
with ¾ colors of beads available.
4.2.25. (♦) A rotating square table has a pocket at each corner. Each pocket may have
1, 2, or 3 balls. Compute the total number of distinguishable arrangements. Use P ólya’s
Theorem to compute the number of distinguishable arrangements with a total of 7 balls.
Describe how the computation would change if we have two colors of balls, but the balls
within a pocket must all have the same color.
4.2.26. An even graph is a graph whose vertex degrees are all even.
(a) Compute the weight enumerator for isomorphism classes of connected even graphs
with at most five vertices, indexed by number of vertices.
(b) Use part (a) and P ólya’s Theorem to enumerate, by number of vertices, the even
graphs with three components in which every component has at most five vertices. Check
your work by explicitly listing those with nine vertices.
4.2.27. Let e n and on count the isomorphism classes of graphs on n vertices whose number
of edges is even or odd, respectively. Prove that e n ≥ on for all n and that equality holds if
and only if n is congruent to 2 or 3 modulo 4. (Schmidt [1993])
4.2.28. (♦) The Ordinary Exponential Formula.
(a) By comparing coefficients, prove ∑n≥0 ZËn (x1 , x2 , · · ·) = e ¾≥1 x ¾/¾ .
¾
(b) Prove ½(x) = e ¾≥1 c(x )/¾
, where ½(x) and c(x) are the OGFs by number of vertices for
isomorphism classes of graphs and isomorphism classes of connected graphs. (Comment:
This unlabeled version of the Exponential Formula applies whenever a general labeled
structure consists of component structures and the symmetric group acts on the set of la-
bels to produce isomorphisms.)
¾ n−1
(c) For OGFs ½(x) and c(x) related by ½(x) = e ¾≥1 c(x )/¾ , prove cn = ½ n − 1n ∑¾=1 ¾c¾ ½ n−¾ .
Section 4.3: Permutations and Tableaux 189

4.2.29. Let r(x) be the OGF by number of vertices for isomorphism classes of rooted trees.
¾
(a) Use Exercise 4.2.28 to prove r(x) = xe ¾≥1 r(x )/¾ .
(b) Obtain a recurrence for the coefficients of r(x).
4.2.30. Let Tr(x) be the OGF by number of vertices for isomorphism classes of rooted trees
such that every vertex is reachable from the root by a path with at most r edges.
(a) Find a formula for Tr in terms of Tr−1 .
(b) In terms of these functions, find an expression for the OGF by number of vertices
for isomorphism classes of unrooted trees in which the longest path has d edges.

4.3. Permutations and Tableaux

The subject of Young tableaux expands upon our discussions of permutations,
partitions of integers, and bijections. Young tableaux correspond to representa-
tions of the symmetric group and are central in the study of symmetric functions
(see Sagan [1991], Fulton [1997], and Stanley [1999]). We confine our attention
to their enumeration and their connections to monotone sublists in permutations.

THE HOOK-LENGTH FORMUL A

A generalization of Bertrand ’s Ballot Problem considers m candidates in-

stead of two. Candidate i receives ë i votes, with ë 1 ≥ · · · ≥ ë m and ∑ ë i = n.
What is the probability that the partial vote totals always maintain the same
order while the votes are being counted?
Record j in the row for the candidate receiving the jth vote. At the end the
elements of [n] occupy the positions in the Ferrers diagram for a partition ë. By
construction, the numbers increase in each row. Partial votes totals satisfy the
stated condition if and only if each candidate reaches ¾ votes before the next can-
didate does; thus the numbers must also increase in each column, as shown below.
To solve our problem, we divide the number of such tableaux for ë by the multi-
nomial coefficient ( ë 1 ,...n , ë m ), which counts all vote sequences.

1357
248
69

4.3.1. Definition. A reverse plane partition consists of positive integers in the

positions of a Ferrers diagram with rows and columns nondecreasing. It is a
standard Young tableau if the entries are 1 , . . . , n, distinct. A general-
ized Young tableau or column-strict tableau is a reverse plane partition
with strictly increasing columns. Its shape is the list of row-lengths.

The word “partition” indicates that the entries form a partition of their total,
“plane” suggests the geometric arrangement, and “reverse” refers to nondecreas-
ing instead of nonincreasing rows and columns. (Standard) Young tableaux were
introduced in Young [1901] to study matrix representations of permutations but
190 Chapter 4: Further Topics

actually were used earlier by Frobenius [1900]. They have many connections to
algebraic combinatorics and to combinatorial aspects of permutations. Much of
the subject (and the terminology) was developed by French mathematicians.
Let º ( ) be the number of Young tableaux of shape (the value needed in
the generalized ballot problem). The first formula for  ( ) was found by Frobe-
nius [1900] using group representations. The modern formula uses the sizes of
particular subsets of the diagram.

4.3.2. Definition. In a tableau with shape , the hook Hi , j consists of all posi-
tions (i , s) and (r, j) with s ≥ j or r ≥ i. The hook length hi , j is the size of
Hi , j . A corner is a position in the Ferrers diagram with no item to its right
or below (an (i , i) with i > i+1).

4.3.3. Lemma. Let position ( , ) be a corner of shape . If i ≤  and j ≤  , then

(hi − 1) + (h j − 1) = hi , j − 1.
Proof: By shifting the sets labeled x and u in the figure, we conclude that
(Hi − {( , )}) ∪ (H j − {( , )}) has the same size as Hi , j − {(i , j)}.

i, j ←u→ i, ← v →

↑ ↑
x x
↓ ↓
, j ←u→ ,
↑
y
↓

A tableau puts [n] into a shape in one of n! ways. The result is a standard
Young tableau if and only if for each (i , j) the entry in position (i , j) is the least
entry in Hi , j . When Hi , j is filled randomly, position (i , j) is least with probability
1/hi , j . Knuth [1973] joked that we would like to multiply these values to obtain
1/ ∏ hi , j as the probability of being a Young tableau, but the events are far from
independent (consider h2 ,1 and h1 ,2 when = (2 , 2)).
Nevertheless, this formula is correct! The original proof used a determinant
for  ( ) known to Frobenius [1900] and Young [1902] (Exercise 13). Another uses
the fundamental theorem of algebra (Bandlow [2008]). Ours is probabilistic.

4.3.4. Theorem. (Hook-Length Formula; Frame–Robinson–Thrall [1954])

For a partition of n, the number  ( ) of standard Young tableaux of shape
is n!/∏ hi , j , where the product is over all positions (i , j).
Proof: (Greene–Nijenhuis–Wilf [1979]). Let F( ) = n!/ ∏ hi , j ; we want to prove
 ( ) = F( ). We extend the domains of F and  by setting F( ) =  ( ) = 0 when
is a list that is not nonincreasing.
When a list is in fact a partition (that is, non-increasing), let ˆ  denote
the list of numbers obtained from by subtracting 1 from  . Since the ele-
Section 4.3: Permutations and Tableaux 191

ment n must occupy the last position in its row, we have the recurrence º( ) =
∑¾=1  ( ˆ ¾), where m is the number of rows in the shape.
m

Since F and  agree when  ( ) = 0 and when n = 1, it suffices to prove F( ) =

∑¾=1 F( ˆ ¾), or equivalently 1 = ∑ F( ˆ ¾)/F( ), which we interpret probabilisti-
m

cally. We define an experiment whose outcomes correspond to rows in the shape.

It suffices to prove that the probability of the outcome in row  is F( ˆ )/F( ).
Each trial of the experiment is a path in the tableau. Begin at a random cell,
each with probability 1/n. Thereafter, when at position (i , j), move to another
position in Hi , j , each with probability 1/(hi , j − 1). This move increases i or j ;
repeat until a corner square is reached. The outcome of the trial is the corner
square ( , ) that is reached.
To find the probability of outcome ( , ), we compute conditional probabili-
ties from each starting location. We then sum over starting locations and multi-
ply by 1/n to remove the conditioning. In computing the conditional probability,
we do not consider each path from (a , b) to ( , ) separately. We combine paths by
recording only the list A of rows and the list B of columns used in the path. For
example, from (a , b) the paths (a , b) , (a , ) , ( , ) and (a , b) , ( , b) , ( , ) both ap-
pear in the conditional event A , B with A = (a , ) and B = (b , ). The grouped
probabilities will correspond to terms in the formula F( ˆ )/F( ).
Let p(A , B) be the probability that the path visits row set A and column set
B, given the start at (a , b). The probability is 0 if A or B is empty; it is 1 if
(a , b) is a corner and A = {a}, B = {b}. Otherwise, let A = {a , a . . . } and B =
{b , b . . . } in order. We can move to (a , b ) and thereafter have A , (B − {b}), or
we can move to (a , b) and thereafter have (A − {a}) , B. Since the probability of
moving from (a , b) to any location in the hook is 1/(ha ,b − 1), the recurrence is
1 1
p(A , B) = p(A − {a} , B) + p(A , B − {b})
ha ,b − 1 ha ,b − 1
We claim that p(A , B) = Q(A , B), where
1 1
Q(A , B) = ∏ hi ,  − 1 ∏ h ,j −1
.
i∈ A− j ∈ B− 

Observe that p(A , B) = 1 = Q(A , B) when | A| = | B| = 1; we proceed by induction.

Since Q(A − {a} , B) = (ha ,  − 1)Q(A , B) and Q(A , B − {b}) = (h ,b − 1)Q(A , B),
the recurrence and the induction hypothesis yield
1
p(A , B) = [(ha ,  − 1)Q(A , B) + (h − 1)Q(A , B)] .
ha ,b − 1
,b

Lemma 4.3.3 yields ha ,  − 1 + h ,b − 1 = ha ,b − 1, and thus p(A , B) = Q(A , B).

The probability computation finishes with ( , ) = (1/n) ∑ p(A , B), where the
summation runs over all A , B such that  ∈ A and  ∈ B.
Now, consider the ratio of F( ˆ ) and F( ). Each has a factorial and a product
of hook lengths. When the corner position  is deleted, the factor h  = 1 disap-
pears, and the hooks in row  and column  decrease by one. Since (n − 1)!/n! =
1/n, this yields
F( ˆ ) 1 h h ,j 1 1 1
= ∏ i , ∏ = ∏(1 + ) ∏(1 + ).
F( ) n hi ,  − 1 h ,j − 1 n hi ,  − 1 h ,j − 1
i< j<  i< j< 
192 Chapter 4: Further Topics

In expanding this product, every term corresponds to choosing a set A of row

indices less than and a set B of column indices less than  . Let A = A ∪ { },
and let B = B ∪ {}. The corresponding term in the product is precisely Q(A , B).
The terms together generate all A , B that contain ,  . As desired, we obtain

F(ˆ ) 1
= ∑ Q(A , B) = ∑ p(A , B) = ( , ).
1
F() n n

In this experiment, the probability of each outcome is the fraction of tableaux

whose largest element is in that corner. The proof is really a counting argument;
probabilistic language does the bookkeeping.

4.3.5. Remark. The weighting of corner squares in this experiment lets us gen-
erate Young tableaux of a given shape with equal probability. First, run a trial
and put n in the corner location specified by the outcome. Delete this position,
run the experiment for the smaller shape, and place n − 1 in the resulting lo-
cation. Repeat until the entire tableau is filled. At each step, the weights for
placing  are as they should be, given the placement of  + 1 , . . . , n.
Also, since it is the size of a set, n!/ ∏ hi , j is always an integer.

To count column-strict tableaux with shape  and entries in [m], see Exer-
cise 14. Thrall [1952] introduced a variation called shifted tableaux, where the
shape  has distinct parts and the rows start at positions (i , i) instead of (i , 1);
there is a counting formula using generalized hooks (Exercise 15).
Researchers long sought a simple bijective argument for the Hook-Length
Formula. There are n! ways to fill the shape  with [n]. One would like to prove the
formula by grouping them into sets of size ∏ hi , j , each containing one standard
Young tableau. Complicated bijective proofs were found by Franzblau–Zeilberger
[1982], Remmel [1982], Zeilberger [1984], and Krattenthaler [1995]. Novelli–
Pak–Stoyanovskii [1997] found a bijection that maps each permutation into a
pair consisting of a standard Young tableaux of shape  and an element of a set
of size ∏ hi , j . Generating a random permutation and applying this bijection pro-
vides another way to generate a random standard tableau with shape .

4.3.6.* Remark. Bijection for Hook-Length Formula (Novelli–Pak–Stoyanovskii

[1997]). Number the cells of the shape  in reverse lexicographic (RL) order, first
the last column from bottom to top, then the next, and so on, ending with n in the
upper left corner. Enter the given permutation into the cells in this “RL” order.
The bijection involves moving elements to reach a standard tableau and
recording data about the process. A hook function  assigns an integer to each
position in the shape, with −(∗j − i) ≤  (i , j) ≤  i − j , where ∗ is the conjugate
of . Given , there are ∏ hi , j hook functions.
Now consider an entry a with entries b to the right and c below (let entries
outside the shape be infinite). Leave a in place if a < min{b , c}. Otherwise,
switch a with the smaller of b and c and continue from the new position of a. We
call this movement “bubbling ” by analogy with one-dimensional “bubble sort ”.
Bubbling a until it stops produces a new tableau.
Section 4.3: Permutations and Tableaux 193

If cells 1 through l (in RL order) are increasing in rows and columns and we
bubble the entry in cell l + 1, then cells 1 through l + 1 of the resulting tableau
will be increasing in rows and columns. Hence successively bubbling the elements
in cells 1 through n produces a standard Young tableau.
During this process, we develop a hook function, starting with º (i , j) = 0 for
all (i , j). When a bubbles from position (i , j) in a tableau with hook function º ,
it comes to rest in a new position (i , j ) with i ≥ i and j ≥ j . We modify the
hook function, changing values only in column j . The value in row i of column
j becomes j − j . The values that were in rows i + 1 through i (if any) move one
position higher and decrease by 1. Since the higher cell has one more cell below
it than the lower one, and position (i , j) has at least j − j positions to its right,
the result is again a hook function.
Below we apply the bijection to the permutation 123456 in shape (2 , 2 , 2).
Bubbling 1 from the corner produces no change. In each figure, we underscore
the element next to bubble. To complete the proof, we must prove that every pair
consisting of a standard tableau and a hook function with that shape arises from
exactly one permutation by this map. The probabilistic proof remains shorter,
but the bijection is useful.

6 3 0 0 6 3 0 0 6 1 0 -2
→ 5 2 0 0 → 5 1 0 -1 → 5 2 0 -1
4 1 0 0 4 2 0 0 4 3 0 0

6 1 0 -2 6 1 0 -2 1 4 -1 -2
→ 5 2 0 -1 → 2 4 0 -1 → 2 5 0 -1
3 4 1 0 3 5 1 0 3 6 1 0

THE RSK CORRESPONDENCE

We next study a remarkable correspondence between permutations and pairs

of same-shape Young tableaux. It enables us to count the n-element Young
tableaux and to compute the maximum size of the union of increasing sequences
in a given permutation. Its properties are so astonishing that Knuth [1973]
wrote, “ The unusual nature of these coincidences might lead us to suspect that
some sort of witchcraft is operating behind the scenes!” We call it the RSK Cor-
respondence to honor Robinson [1938], Schensted [1961], and Knuth [1970].
Algebraists discovered using group representations that there are n! pairs of
same-shape Young tableaux with n elements. The RSK Correspondence proves
this combinatorially. To count the tableaux with n elements, it then suffices to
count permutations of [n] whose images under the RSK Correspondence are pairs
of identical tableaux.
Counting these permutations uses a remarkable property of the RSK Cor-
respondence: if it maps  to the pair (P, Q), then it maps  −1 to (Q , P). Hence
the permutations corresponding to pairs of identical tableaux are the involutions
194 Chapter 4: Further Topics

(permutations consisting of fixed points and 2-cycles). Involutions with i 2-cycles

correspond to partitions of [n] into n − i blocks with i blocks of size 2. Summing
over i yields the formula below to count both involutions and standard Young
tableaux with n elements.
⌊n/2⌋
n!
∑ 2i i!(n − 2i)!
i=0

4.3.7. Example. For n = 4, the sum yields 10 involutions. They appear be-
low, using the two-line notation for permutations. Also shown are the 10 Young
tableaux, using the correspondence we will develop.
1234 1234 1234 1234 1234
1234 2134 3214 4231 1324
1234 134 14 13 124
2 2 2 3
3 4

1234 1234 1234 1234 1234

1432 1 243 2143 341 2 4321
12 123 13 12 1
3 4 24 34 2
4 3
4
A permutation : [n] → [n] can be expressed as a set of pairs {(i , i)}. Knuth
generalized the correspondence to allow the input to be any multiset of n inte-
ger pairs {(qi , pi)}. The output is then a pair P, Q of same-shape column-strict
tableaux. We describe the insertion of a new pair recursively.

4.3.8. Algorithm. INSERT(P , Q , p , q) If p is as large as everything in row 1 of

P , return the new pair P , Q obtained by putting p at the end of row 1 of P and
putting q at the end of row 1 of Q.
Otherwise, p bumps the first value p in row 1 that exceeds p. Replace p
with p in row 1 of P and leave row 1 of Q unchanged; these become the top rows of
the new pair P , Q . For the remaining rows of P , Q , use INSERT(P , Q , p , q),
where T denotes the tableau consisting of all but the top row of T .

4.3.9. Example. When processing a permutation, the pairs are inserted in in-
creasing order of the values of qi . If = 617258934, then the situation when we
insert the next-to-last pair (8 , 3) is
P Q P Q
12589 13567 1 2389 13567
67 24 → 57 24
6 8
Here 3 bumps 5, 5 bumps 6, and 6 is inserted at the end of row 3. Hence Q ac-
quires 8 in a new position at the end of row 3. To complete P( ) , Q( ), we insert
(9 , 4). Here 4 bumps 8 to row 2, and the ninth new position is at the end of row 2.
Section 4.3: Permutations and Tableaux 195

In the generalized input allowing any multiset of pairs as input, the insertion
order is lexicographic on the pairs (qi , pi). For example, with (q , p) = (211246133
133444555
)
(written with qi above pi), the last insertion is
P Q P Q
11136 13444 11133 13444
224 355 → 2246 3555

This algorithm grows the common shape in one location; the end of the row
where insertion succeeds without bumping. If (q , p) was the pair being inserted,
then we add q to the current tableau Q in the new location to obtain the new Q .
This new position receiving q is the end of a row. Since we want nondecreasing
rows, we INSERT input pairs {(qi , pi)} in increasing order of qi . For permuta-
tions, this reduces to inserting as pi the values 1 . . . n in order. We write qi first
in generalized input pairs to agree with the 2-line notation for permutations.

4.3.10. Definition. The RSK Correspondence is the map that takes a gener-
alized input to the pair (P( ) , Q( )) of tableaux generated by successively
INSERTing input pairs (qi , pi) in lexicographic order (nondecreasing order in
q, and nondecreasing in p among pairs with the same q). In the output, P( )
is the P-symbol and Q( ) is the Q-symbol of the input .

The RSK Correspondence is also called the Robinson–Schensted Corre-

spondence, Schensted’s Algorithm, or the bumping procedure. It was im-
plicit in Robinson [1938], explicit in Schensted [1961] (independently), and gen-
eralized beyond permutations in Knuth [1970].
The proof that P( ) and Q( ) are column-strict uses a technical lemma about
the positions where tableaux successively grow.

4.3.11. Lemma. Let pi and pi+1 be values successively INSERTed into a column-
strict tableau P , creating positions A and B, respectively.
(a) The successive values bumped when inserting pi form a strictly in-
creasing sequence. The successive column indices of the bumped positions
(and the new position) form a nonincreasing sequence.
(b) B is in a column to the right of A if and only if pi+1 ≥ pi .
Proof: (a) An element bumps only a larger element, so the values strictly in-
crease. If in row i the element x in column j is bumped, then x cannot enter a
column past j in row i + 1, because the columns of P are strictly increasing.
(b) We use induction on the number of rows. Whether it bumped or not, pi
is in row 1 when we start to INSERT pi+1 . If pi+1 enters without bumping, then
pi+1 ≥ pi , a new rightmost position is created, and B is farther right than A. If
pi does not bump but pi+1 bumps, then pi+1 < pi and B is not farther right than
A, since no later row reaches the column of A.
In the remaining case, pi bumps some p i and pi+1 bumps some p i+1 . Now
pi+1 ≥ pi if and only if p i+1 ≥ p i . Since A and B are the positions created when
p i and p i+1 are inserted in the tableau P consisting of the remaining rows, the
induction hypothesis completes the proof.
196 Chapter 4: Further Topics

4.3.12. Lemma. The tableaux P( ) and Q( ) produced by the RSK Correspon-

dence are column-strict.
Proof: The processing order makes the rows of Q( ) nondecreasing, and the bump-
ing procedure makes the rows of P( ) nondecreasing. Since dropping (qn , pn) from
the lexicographic input order leaves the rest in lexicographic order, it suffices to
show that the outputs of INSERT(P , Q , p , q) are column-strict when P and Q are
column-strict and produced by pairs lexicographically earlier than (q , p).
The value q is as large as all entries in Q and appears at the bottom of its
column. All copies of q appear as (q , pj) , . . . , (q , pn) at the end of the input, with
p j ≤ · · · ≤ pn . By Lemma 4.3.11b, the columns of the new locations for these
insertions are strictly increasing and thus distinct.
For P , let t be the column where p enters row 1. Since P is column-strict,
all entries in column t of P are larger then p. Suppose that p bumps an element
p . By Lemma 4.3.11a, p cannot survive past column t when inserted in row 2.
If p enters row 2 under element p of row 1, then p ≤ p < p . Continuing the
INSERTion leaves every element in row 2 larger than the element above it. The
result of inserting p into P is also column-strict, since this is the P-symbol for
the input consisting of elements bumped from the first row (in the same order).

Next we show that the RSK Correspondence is a bijection. Since we insert

the pairs in a canonical order, the input is specified by the multiplicities of dis-
tinct pairs. Encode the input as a matrix A whose (r, s)-entry is the multiplicity
of (r, s) in the input. From A, we can recover the pairs and INSERT them in
lexicographic order. If ∈ Ën , then A is an n-by-n permutation matrix.

4.3.13. Theorem. The RSK Correspondence maps nonnegative integer matrices

with total sum n bijectively into pairs of column-strict tableaux with n cells.
It restricts to a bijection from Ën to the set of pairs of same-shape standard
Young tableaux with entries [n].
Proof: It suffices to invert the procedure. Given same-shape tableaux P and Q
on n elements, we specify an inverse for the last INSERT operation.
DELETE(P, Q). Let q be the rightmost copy of the largest element in Q,
and let p be the entry in that position in P ; say that they are in row ¾ .
Delete q from Q to obtain the new Q ; also delete p from P. If ¾ = 1,
then return the new tableau pair and the deleted elements. If ¾ > 1,
then let p be the rightmost value in row ¾ − 1 that is strictly less than
the “unbumped element ” p. Replace p with p in P and let p become the
unbumped element. Decrease ¾ and repeat.
Since P and Q are column-strict, the deleted position was a corner. Column-
strictness also makes the next unbumped element p well-defined. The col-
umn indices of the “unbumps” are nondecreasing, and the values unbumped
are strictly decreasing. The resulting P and Q are column-strict tableaux of
the same shape. The modifications undo in reverse the changes performed by
INSERT(P , p , Q , q) when (q , p ) is the last pair in a lexicographically ordered
sequence and the pair (P , Q ) arises by applying RSK to the preceding portion of
the sequence. Induction on the input length completes the argument.
Section 4.3: Permutations and Tableaux 197

Using Lemma 4.3.11, we can obtain information about a permutation in

terms of the positions of elements in P( ) and Q( ).

4.3.14. Definition. A permutation of [n] (in word form) has an ascent at posi-
tion i if (i + 1) > (i); otherwise there is a descent at i. It has an advance
at value x if value x + 1 is rightward of x; otherwise there is a retreat at x.

Reading the word form from left to right and writing + or − for ascents or
descents produces the up-down sequence. Following the values from low to high
and writing + or − for advances or retreats produces the right-left sequence.

4.3.15. Example. Let = 617258934. This permutation has advances at values

{1 , 2 , 3 , 6 , 7 , 8}, retreats at values {4 , 5}, ascents at positions {2 , 4 , 5 , 6 , 8}, and
descents at positions {1 , 3 , 7}. Hence the right-left sequence is + + + − − + + +
and the up-down sequence is − + − + + + − +. These sequences can also be
obtained by tracing the numbers through P( ) and Q( ), respectively.
P Q
12349 13567
578 249
6 8

4.3.16. Theorem. (Schützenberger [1963], Foulkes [1976]). A permutation

has an ascent at position c if and only if c + 1 is rightward from c in Q( ). It
has an advance at c if and only if c + 1 is rightward from c in P( ).
Proof: The row and column inequalities of tableaux imply that c+1 cannot appear
in a position with both indices smaller or both indices larger than the position of
c. Hence the only possible movements are rightward (larger column index) and
downward (larger row index). By Lemma 4.3.11b, the bumping procedure places
c + 1 in a column to the right of c in Q( ) if and only if (i + 1) ≥ (i).
Since the up-down and right-left sequences interchange when we replace
with −1 , the proof of the second statement will follow from P( ) = Q( −1).

A tool used to prove P( ) = Q( −1) will allow us to obtain the maximum length
of an increasing sublist of .

4.3.17. Definition. In the 2-line input (q , p), the pair (qi , pi) belongs to class t
if pi initially enters P in the tth position of the first row. We refer to pi or qi
as a class t element or to (qi , pi) as a class t pair.

4.3.18. Example. Nondecreasing sublist. In our example = 617258934, element

pi is in class t if and only if the longest nondecreasing sublist up to pi has t terms.

q 123456789
p 617258934
t 112234534
198 Chapter 4: Further Topics

4.3.19. Lemma. In the class t pairs (qi1 , pi1 ) , . . . , (qim , pim), the second coordi-
nates are strictly decreasing, and the first coordinates are strictly increas-
ing. Also, these pairs are those for which the longest nondecreasing sublist
of p ending at this position has t elements.
Proof: Each pi in class t bumps the previous element in class t, and thus in order
they are strictly decreasing. The lexicographic input order then implies that the
corresponding qi ’s are strictly increasing.
The top left element is always the least yet seen; thus the longest nondecreas-
ing sublist ending at a class 1 element is just that element. By induction on t and
the number of elements seen, the current element in position t of row 1 is always
the least element yet seen that ends a nondecreasing sublist of length t.

4.3.20. Corollary. (Schensted’s Theorem; Schensted [1961]) The maximum

length of an increasing sublist of a permutation is the length of the first
row of P( ). For a decreasing sublist it is the length of the first column.

Schensted noted that the second statement follows from the first by revers-
ing the permutation. Later we prove much stronger statements than Corollary
4.3.20. Note that the elements of the first row in P( ) need not themselves form
an increasing subsequence of ; in Example 4.3.18, the first row of P( ) is 12349,
but the only increasing sequence of length 5 in is 12589.

SWITCHING P-SYMBOL AND Q-SYMBOL

To finish counting n-element standard Young tableaux, we still need P( −1) =
Q( ), equivalent to Q( −1) = P( ). For the generalized input matrix A, inter-
changing q and p corresponds to transposing A. The next lemma and theorem
obtain P(A T ) = Q(A) and Q(A T ) = P(A) directly from class t pairs. We then
interpret the RSK Correspondence geometrically for an alternative proof.

4.3.21. Lemma. The pair (q¾ , p¾) belongs to class t of (q , p) if and only if the pair
(p¾ , q¾) belongs to class t of (p , q).
Proof: If (q¾ , p¾) is in class t of (q , p), then by Lemma 4.3.19 some nondecreas-
ing subsequence pi1 , . . . , pit ends with p¾ . From the lexicographic input order,
qi1 , . . . , qit is also nondecreasing. When p and q are interchanged and re-ordered
lexicographically, the indices of these pairs remain in the same order. Thus q¾
will appear at the end of a nondecreasing subsequence of length at least t. By
Lemma 4.3.19, (p¾ , q¾) belongs to class t or higher in the (p , q) input. The same
argument applies when the switch is repeated, so the pairs (q¾ , p¾) from (q , p)
and (p¾ , q¾) from (p , q) belong to the same class in their respective processing.

4.3.22. Theorem. (Schützenberger [1963], Knuth [1970]). Given a finite non-

negative integer matrix A, the P-symbols and Q-symbols satisfy P(A T ) =
Q(A) and Q(A T ) = P(A). In particular, if is a permutation, then P( −1) =
Q( ) and Q( −1) = P( ).
Proof: The algorithm ends with the last class t element in position t of row 1 in
P. The value in position t of row 1 of Q was put there when the position was first
Section 4.3: Permutations and Tableaux 199

created; it is the qi from the first pair in class t. Lemma 4.3.21 says that the pairs
in class t are the same (with coordinates reversed) when we transpose the input
A; Lemma 4.3.19 says that they are processed in opposite order. Together these
facts imply that transposing A interchanges the first rows of P and Q.
We have proved the basis step for induction on the number of rows. The rows
below row 1 are formed by applying the same algorithm to a modified input. The
modified input consists of the pairs (qi , pj) such that pi bumps pj from row 1. We
need only show that the modified input creating the pair of tableaux below the
first row for (p , q) is the transpose of the modified input that does so for (q , p).
If pi bumps pj from position t of row 1, then (qj , pj ) is the class t pair before
(qi , pi). By Lemmas 4.3.19–4.3.21, (pi , qi) and (pj , qj) are successive class t pairs
for A T , and qj bumps qi from position t when we insert (pj , qj ). Thus (pj , qi) is in
the modified input for A T whenever (qi , pj ) is in the modified input for A.

qi1 qi2 ··· qi¾−1 qi¾

pi1 pi2 ··· pi¾−1 pi¾

Since the P-symbol and Q-symbol are formed so differently, it is surprising

that inverting a permutation exchanges them. The result becomes natural under
a geometric interpretation called planarization developed by Viennot [1977]. It
encodes bumping in a special diagram that treats P( ) and Q( ) symmetrically.
Given input with corresponding matrix A, put Ar,s points at position (r, s)
in the plane (the pair (r, s) occurs times in  ). Let T( ) be the resulting arrange-
ment of points with multiplicities. On this multiset we build what Viennot called
outstanding lines. The tth outstanding line corresponds to the class t pairs in the
bumping procedure. These lines yield row 1 of P(A) and Q(A) in a symmetric
manner. They also yield the input for building the remaining rows, inductively.

4.3.23. Example. Planarization, outstanding lines, and skeletons. We illustrate

the construction with the sample permutation  = (123456789 )
617258934 used by Viennot.
Its planarization appears below. In placing this diagram in the plane, we have
temporarily adopted the “French notation”, in which everything is indexed from
the bottom left. (Indexing from the upper left is the “English convention”.)
The first line is determined by the elements of the input T( ) that are min-
imal in the component-wise dominance ordering, defined by (i , j) ≤ (r, s) if and
only if i ≤ r and j ≤ s. The minimal points of T( ) are those that dominate no
other points; here they are (1 , 6) and (2 , 1). (If there are repeated points in the
input, then we just use one copy of each minimal point, but here for simplicity we
only discuss permutations.)
Let M denote the set {(qi1 , pi1 ) . . .(qi¾ , pi¾ )} of minimal points of T( ). Since
none dominates any other, we may index them so that qi1 < · · · < qi¾ and pi1 >
· · · > pi¾ . From M we form the outstanding line consisting of the vertical and
horizontal zig-zag path linking the following points in order: (qi1 , ∞), (qi1 , pi1 ),
(qi2 , pi1 ), (qi2 , pi2) . . .(qi¾ , pi¾−1 ), (qi¾ , pi¾ ), (∞ , pi¾ ).
200 Chapter 4: Further Topics

1 2 1 2 1 1 1 3 2
9 • 1
8 • ◦ 2
7 • ◦ 2
6 • ◦ × 3
p 5 • ◦ 2
4 • 1
3 • 1
2 • 1
1 • 1
1 2 3 4 5 6 7 8 9
q

Geometrically, consider light shining from the left and below. The outstand-
ing line is the outline of the shadow of M. Alternatively, it is the outline of the
set of points in the plane that dominate points of M.
The outstanding line changes direction at 2 ¾ − 1 points. Of these, ¾ form M ;
the other ¾ − 1 we place in the skeleton of T( ), denoted S( ). The full skele-
ton consists of these points plus the skeleton of T( ) − M. Thus we find the first
outstanding line, put the inside turning points in the skeleton, delete the mini-
mal points M from T( ), and iterate until no points remain. If T( ) has n points,
then S( ) has n −  1 points, where  1 is the number of outstanding lines.

In the diagram of Example 4.3.23, we have iterated the process of finding the
skeleton. The solid points are T( ), with solid outstanding lines. The open points
are the skeleton S( ), with dashed outstanding lines. The skeleton of the skeleton
has only the one point denoted × and one outstanding (dotted) line. An outstand-
ing line receives label  where it exits the diagram if it arises when forming the
skeleton the  th time. The creation of the skeleton has the same effect as bump-
ing, and the labels of the lines give the rows of the elements in P( ) and Q( )!

4.3.24. Theorem. (Viennot [1977]). Given the permutation as input to the

RSK Correspondence,
(1) The skeleton S( ) is the set of pairs produced by the bumping proce-
dure for computation of P( ) and Q( ) below row 1.
(2) The  th row of P( ) consists of the rows labeled  in the planarization.
The  th row of Q( ) consists of the columns labeled  in the planarization.
(3) P(A T ) = Q(A) and Q(A T ) = P(A).
Proof: The point (qi , pi) is a minimal point of in the plane if and only if pi is a
left-to-right minimum in p, making it a class 1 pair for (q , p). Since the class t
pairs are the class 1 pairs for the input in which the earlier classes are deleted,
the points of on the tth outstanding line of are the class t pairs for (q , p).
(1) If (qj , pj ) and (qi , pi) are consecutive in class t, then (qi , pj) belongs to the
skeleton. Also, pi bumps pj when inserted (Lemma 4.3.19), and hence (qi , pj ) be-
longs to the input below row 1.
Section 4.3: Permutations and Tableaux 201

(2) By (1), it suffices to prove this for ¾ = 1 and apply induction on the number
of rows. As observed in proving Theorem 4.3.22, the value in position t of row
1 of P is the p-coordinate of the last class t pair in the input, and the value in
position t of the first row of Q is the q-coordinate of the first class t pair, placed
when the position is created. Since the tth outstanding line visits the class t pairs
in order, these two values are the horizontal and vertical lines on which the tth
outstanding line exits the planarization diagram. Being an outstanding line of
T( ), the label is 1, correctly placing these elements in row 1 of P( ) and Q( ).
When the input pairs are not restricted to be distinct, several outstanding lines
may exit at the same value.
(3) Transposing the input means reflecting the planarization diagram through
the line x = y, which interchanges rows and columns.

JEU DE TAQUIN (optional)

By Schensted ’s Theorem, the length of the longest nondecreasing subse-

quence in a list is the length of the first row in its P-symbol. More generally, how
can we find the largest size of the union of  nondecreasing sequences? Greene
proved that this is the sum of the lengths of the first  rows in the P-symbol.
We need several ideas to prove this. The first generalizes the notion of “non-
decreasing sequence” to planar arrangements and incorporates the properties of
advances while ignoring consecutivity among the values.

4.3.25. Definition. A partial tableau consists of numbers in some positions

in rows and columns, strictly increasing in columns and nondecreasing in
rows. A nondecreasing sequence in a partial tableau T is a nondecreasing
sequence of elements in positions with increasing column indices and nonin-
creasing row indices. A -sequence in T is a union of  nondecreasing se-
quences in T . The Greene number G ¾ ,l(T) is the largest size of a -sequence
in the first l positions in T under lexicographic order on (value, column).

When the entries are distinct, the “first l positions” are those of the l smallest
values. In the partial tableau T( ) obtained by writing the elements of a permu-
tation diagonally from lower left to upper right, nondecreasing sequences are
the same as increasing sequences in ; this motivates Definition 4.3.25. To prove
Greene ’s Theorem, we need
(1) T( ) and P( ) have the same Greene numbers, and
(2) in P( ) the first  rows form a largest -sequence.
We will move from T( ) to P( ) using a class of partial tableaux containing both.

4.3.26. Example. Three partial tableaux and their Greene numbers. Let =
617258934. The diagonal tableau T( ) puts the values in order of the word form.
As noted above, G ¾ ,n(T( )) is the maximum size of the union of  increasing se-
quences in . The other tableaux below are obtained by sliding elements; all three
have the same Greene numbers.
202 Chapter 4: Further Topics

In R (in fact, it is P( )), the first two rows form a largest 2-sequence, as do
1234 and 6789. Also T( ) has 1234 and 6789 as a 2-sequence of size 8, but it
cannot be found greedily; the only increasing sequence of length 5 is 12589, and
deleting it leaves no increasing triple.
We will slide elements to move from T( ) to P( ).
T( ) strip shape R = P( )
4
3
9
8 34 12349
5 2589 578
2 17 6
7 6
1
6
l
G ¾ ,l 1 23456789
1 1 23444445
2 1 23455678
3 1 23456789

The process of moving from T( ) to the “reduction” R is called Jeu de

Taquin. In French, the term means “Game of Teasing ”, or brain-teaser. It is
the French name for the English 15-puzzle. In this puzzle, tiles numbered 1
through 15 slide in a 4-by-4 grid with one empty position; the aim is to put the
numbers in order. We need an operation to slide numbers in a partial tableau.

4.3.27. Definition. Given two partitions  and  such that  i ≥ i for all i, a
skew tableau of shape ( , ) is a partial tableau filling the positions in the
diagram of  that are outside the diagram of . A hole in such a tableau is
an unfilled position within the skew tableau shape: outside  but inside .

The tableau T( ) is a skew tableau, where  is a triangular shape and  is

the next smaller triangle. The strip shape is a more complicated skew tableau,
and in P( ) the partition  is empty. Jeu de Taquin reduces the vacant portion 
to an empty shape by successively filling holes.

4.3.28. Definition. Given a hole in a skew tableau, with elements immediately

below and to the right, a SHIFT modifies the tableau by sliding into the
hole the smaller of those two elements (from below if they are equal); posi-
tions outside  are assumed to hold +∞.
Given a skew tableau of shape ( , ), a FILL operation deletes a corner
of  to form  , which leaves a skew tableau of shape ( ,  ) with one hole,
and then successively applies SHIFT to flush the hole out of the shape, leav-
ing a skew tableau of shape ( ,  ), where  is obtained from  by deleting
the corner reached by the hole.
Section 4.3: Permutations and Tableaux 203

In Example 4.3.26, the strip shape arises from T( ) by FILL operations that
compare consecutive elements of . This leaves i+1 rightward of i if i+1 ≥ i
(by a vertical SHIFT) and i+1 above i if i+1 < i (by a horizontal SHIFT). After
comparing i and i+1 , a horizontal SHIFT is followed by SHIFTing the later ele-
ments left (by FILLing each later corner of the current ), while a vertical SHIFT
is followed by SHIFTing the earlier elements up (by FILLING each earlier corner
of the current . The resulting strip shape encodes the up-down sequence of  .
Further FILLing transforms the strip shape to a tableau R, which in Exam-
ple 4.3.26 we see is P( ) from Example 4.3.15. Jeu de Taquin is the process of
successively FILLing corners of , in some order. We will show Jeu de Taquin
always produces a column-strict tableau from T( ) and that it is always P( ).
Jeu de Taquin was invented by Schützenberger [1977]. Our motivation here
is the purely combinatorial result of Greene. Viennot called the tableau R( )
resulting from Jeu de Taquin the redressé of the original, meaning “straight-
ened”; we use the English word reduction.

The plan. We show first that Jeu de Taquin can reduce the diagonal tableau
T( ) to P( ) by modeling the bumping procedure with SHIFTs. Next, in a column-
strict tableau with n elements, G ¾ ,n is just the sum of the lengths of the first 
rows. The fact that every reduction R( ) equals P( ) follows because the Greene
numbers determine exactly one column-strict tableau and each SHIFT in Jeu de
Taquin preserves the Greene numbers. Greene ’s Theorem is then a corollary.

4.3.29. Lemma. For an input  , some instance of Jeu de Taquin reduces the
diagonal tableau T( ) to the P-symbol P( ).
Proof: We model the bumping procedure, by induction on the length of  . No
SHIFT is needed when  has length 1. Otherwise, form  from  by ignoring the
last element x and inductively transform T( ) into P( ) using Jeu de Taquin.
FILLing corners now can move P( ) up from below and x in from the right to
obtain a partial tableau where consists of one part equal to the length of the
first row in P , with x immediately to its right and P( ) below it.
We now model the bumping procedure to insert x into P( ). Apply SHIFT to
the left of x until x slides left no farther. Now we have P( ) plus x sitting above
the element x that x bumps when inserted into P( ). SHIFTs into the other
positions of row 1 (in decreasing order) now produce the first row of P( ), with x
above x . Now x sits alone in row 2, ready to enter the tableau below the first
row of P( ). In particular, since the list of column indices of bumped positions is
nonincreasing (Lemma 4.3.11a), the newly bumped element y is no farther left
than where it should enter the next row. By the induction hypothesis, we can
perform the rest of the bumping procedure with further SHIFTs.
We have not made these SHIFTs in the order of Jeu de Taquin, but they can
be reordered to express the process (after the induction hypothesis) as FILLing
row 1 from right to left. As x moves in, the rest of each FILL pulls up an element
from row 2 behind it. After x reaches its resting place, the first SHIFT in each
fill to its left pulls up the element from row 2. Instead of performing these imme-
diately, we can postpone them to subsequent FILLs, since they are not disturbed
by FILLING the hole to the left of x. The induction hypothesis similarly permits
reordering the other SHIFTs to complete each FILL in the first row.
204 Chapter 4: Further Topics

4.3.30. Lemma. If T is a column-strict tableau (without holes), then the first ¾

rows of T form a largest ¾-sequence in T .
¾
Proof: For a tableau T of shape ë, let º¾ (T) = ∑i=1 ë i . Since the first ¾ rows
form a ¾-sequence, it suffices to show that no ¾-sequence is larger than º¾ (T). We
use induction on the number of elements; the claim is trivial when T has one
element. Let x in position (i , j) be the rightmost copy of the largest entry in T ;
delete x to obtain T . Every ¾-sequence in T that omits x is a ¾-sequence in T ;
by the induction hypothesis, we need only consider a ¾-sequence S containing x.
Since S − {x} is a ¾-sequence in T , we have | S| ≤ º¾ (T ) + 1. If i ≤ ¾ , then
º¾ (T) = º¾ (T ) + 1 and the bound holds. Hence we may assume i > ¾ . Now every
nondecreasing sequence in T containing x lies completely below the first ¾ rows.
Also it has size at most j , since the elements of a nondecreasing sequence occupy
distinct columns. The remainder of S is a (¾ − 1)-sequence in T . Also j ≤ ë ¾ ,
since i > ¾ . Thus | S| ≤ º¾−1 (T) + ë ¾ = º¾ (T).

4.3.31. Lemma. The shape of a column-strict tableau and the placement of its
multiset of entries is determined by the Greene numbers.
Proof: Let T and T be column-strict tableaux with the same elements. If T
=
T , then let l be the least integer such that the subtableaux T l and T l formed by
the first l values in T and in T differ (repeats of a single value are taken in order
from left to right to form the first l values).
Note that T l and T l differ only in where the last element is, say in row ¾ in
T l and row ¾ in T l . We may assume ¾ < ¾ by symmetry. By Lemma 4.3.30,
G ¾ ,l(T) = º¾ (T l) > º¾ (T l ) = G ¾ ,l(T ). Thus distinct tableaux with the same en-
tries cannot have the same Greene numbers.

The proofs of Lemmas 4.3.30–4.3.31 yield an algorithm to construct the

column-strict tableau determined by the Greene numbers. It suffices to find for
each l the row for the lth element; this is the least ¾ such that G ¾ ,l > G ¾ ,l−1 .
Finally, we show that each SHIFT in a FILL preserves the Greene numbers.

4.3.32. Lemma. If T is a skew tableau with one hole, and T is obtained from
T by a SHIFT into the hole, then their Greene numbers are the same:
G ¾ ,l(T ) = G ¾ ,l(T) for all ¾ and l.
Proof: A horizontal SHIFT does not change the sets of entries forming ¾-
sequences, so here G ¾ ,l(T ) = G ¾ ,l(T) holds trivially. Vertical SHIFTs can change
¾-sequences. We prove G ¾ ,l(T ) ≥ G ¾ ,l(T) for this case and leave the similar in-
equality G ¾ ,l(T) ≥ G ¾ ,l(T ) as Exercise 17.
Let S of size G ¾ ,l(T) be a largest ¾-sequence among the first l elements in T .
Suppose that T arises by SHIFTing y from row j to row j − 1. The set S is still
a ¾-sequence unless y lies in a sequence A in S with elements to its right in row
j . Let A = aybc, where b is the part after y in row j .
Since T has only one hole, all positions in row j − 1 above elements of b are
filled; let b be this list above b in row j − 1. Since max b < max b, we have
max b < min c; also y ≤ min b since y SHIFTs up. If no element of b appears in
S, then we can modify A by replacing b with b to obtain G ¾ ,l(T ) ≥ G ¾ ,l(T).
Section 4.3: Permutations and Tableaux 205

Otherwise, let be the leftmost element of b in S, and let B be the nonde-

creasing sequence in S containing . Write B = a dd , where d is the part of B
in row j after y (d begins with ). If d is empty, then y can move from A into B.
If d
= ∅, then break b into b1 and b2 , where b1 is the part to the left of the
right end of d. Let b 1 be the part of b above b1 . Replace b1 by b 1 in A and switch
the latter parts of A and B to form A and B :
A = aybc A = ayb 1 d
B = a dd B = a db2 c
# #
Note that A ∩ B = ∅ and ####b 1 #### +|b2 | = |b| , so | A |+| B | = | A|+| B| . The elements
of b 1 are available; by the choice of , none of them are in S.
The new sequences move strictly rightward in columns and weakly upward in
rows. It suffices to check inequalities at the splices. If b 1
= ∅, then y ≤ min b 1 ≤
max b 1 < max b1 ≤ max d ≤ min d . If b 1 = ∅, then the assumption B ∩ b
= ∅
guarantees y ≤ min d , since d cannot avoid row j − 1. For B , it suffices that db2
lies entirely in row j . New elements are bounded by others used, so all elements
are still among the l smallest.
d c
b 1 d b 2 c
a y b1 d b2 b2
a a
a

Various cases in the proof of Lemma 4.3.32 change lists. Although G ¾ ,l does
not change, the previous largest -sequence may no longer be a -sequence. We
already saw that the top line of P( ) need not be a nondecreasing sequence in  .

4.3.33. Theorem. The reduction R of a skew tableau T is independent of the

order of FILLs. Also, the sum ¾ (R) of the first  row-lengths in R is the
maximum size of a -sequence in T .
Proof: Successive FILLs at corners of  produce a column-strict tableau. By
Lemma 4.3.32, the Greene numbers remain unchanged. By Lemma 4.3.31, the
shape and placement of entries in a column-strict tableau with fixed entries
and Greene numbers is fixed. Since also Lemma 4.3.29 shows that P( ) can be
reached, the reduction R is independent of the sequence of FILLs, and ¾ (P( ))
is the maximum size of a -sequence in T( ), which corresponds to the union of 
increasing sequences in  .

As a corollary, we obtain Greene ’s Theorem. In order to state it for both

increasing and decreasing sequences, we need one more observation.

4.3.34. Corollary. If  ∗ denotes the reverse (word form) of a permutation  , then

P( ∗) is the transpose of P( ).
Proof: Initially, T( ∗) is the transpose of T( ). If a tableau T has distinct entries
and T is the transpose of T , then the result of SHIFT to fill position (j , i) in
T is the transpose of the result of SHIFTing to fill (i , j) in T . Hence for any
sequence of SHIFTs reducing T( ), there is a corresponding sequence of SHIFTs
that reduces T( ∗) and maintains the transpose relationship at every step.
206 Chapter 4: Further Topics

4.3.35. Corollary. (Greene’s Theorem; Greene [1974]) The maximum size of

the union of ¾ increasing [decreasing] subsequences of a permutation is the
total size of the first  rows [columns] of P( ).
Proof: Since Lemma 4.3.29 implies that the reduction is P( ) when T is the diag-
onal tableau of , the claim for increasing sequences is a special case of Theorem
4.3.33. For decreasing subsequences, combine this with Corollary 4.3.34.

EXERCISES 4.3

4.3.1. (−) Find the P-symbol and Q-symbol for the permutation 596134278. Find the two-
line generalized input such that the RSK Correspondence yields the results below.
P( ) = 11122 Q( ) = 11233
233 223

4.3.2. (−) Use Young tableaux to prove the Erdős–Szekeres Theorem: every list of n2 + 1
distinct numbers has a monotone subsequence of length n + 1.
4.3.3. (−) How many permutations of [n] have element 2 in the first row of P( )?
4.3.4. The Non-Messing-Up Theorem. Consider a matrix such that in each row the en-
tries appear in increasing order. Prove that after sorting the entries of each column into
increasing order, the entries of the rows are still in increasing order.
4.3.5. (♦) Find the minimum and maximum of the number of comparisons between ele-
ments when applying the RSK Correspondence to a permutation of [n].
4.3.6. A riffle shuffle of a permutation cuts into initial and final segments and then
merges them. Let c( ) be the minimum number of riffle shuffles that can sort into order.
Let d( ) be the number of descents in . Prove c( ) = ⌈ log 2(1 + d( ))⌉ . (Schwenk [1986])
4.3.7. Show that the number of ways to label the vertices of an n-vertex ordered tree (Def-
inition 1.3.21) with [n] so that each label is less than those on its descendants is n! divided
by the numbers of vertices in each subtree rooted at a single vertex. (Knuth)
n−1
4.3.8. Prove combinatorially that (n2)! is divisible by nn ∏i=1 [i(2n − i)]i .

4.3.9. A tableau with shape  is a drawing with ∑  i squares in the plane. The corners
of the squares have degrees 2, 3, or 4 in the drawing. In terms of  1 ,  , and the number
of distinct parts in , count the corner points with odd degree. (Stanley [2014])
4.3.10. Obtain a simple formula for the number of permutations of [rs] having no increas-
ing sublist with s + 1 terms and no decreasing sublist with r + 1 terms. (Schensted [1961])
4.3.11. (♦) Permutations with forbidden patterns.
(a) Establish a one-to-one correspondence between tableaux consisting of two rows of
length n each and pairs of n-element same-shape tableaux with at most two rows.
(b) Conclude that the number of permutations of [n] having no decreasing sublist of
length 3 is the Catalan number Cn .
(c) Exercise 2.1.48 shows that Cn also counts the permutations of [n] having no triple
i < j <  such that () < (i) < (j). Conclude that for each of the six orderings of three
elements, the set of permutations of [n] whose triples avoid that pattern has size Cn .
4.3.12. (♦) Show that the number of fixed points in an involution is the number of columns
having odd length in its P-symbol. (Schützenberger [1963])
Exercises for Section 4.3 207

4.3.13. (♦) Alternative proof of Hook-Length Formula. Let be a partition of n with m

parts, and define  ( ) and F( ) as in Theorem 4.3.4.
(a) Group terms in the Hook-Length Formula F( ) to prove

F( ) ∏i< j (hi ,1 − hj ,1) ∏i< j ( i − i − j + j)

= =
n! ∏i hi ,1 ! ∏i( i − i + m)!
(b) With Ëm being the set of permutations of m, prove

n
( ) = ∑ (sign )( 1 +  (1) − 1 , . . . , m +  (m) − m
),
∈Ë m

(c) Prove the Hook-Length Formula  ( ) = F( ) by proving that the expressions in

parts (a) and (b) are equal. (Hint: Let bi = i + m − i. For any values b1 , . . . , bm , the
product ∏i< j (bi − bj ) is the Vandermonde determinant for the matrix with entry bim− j
in position (i , j). Comment: This formula for  ( ) was known to Frobenius [1900] and
Young [1902]; MacMahon [1916] found another proof. This proof is from Linial [1982].)
4.3.14. (+) Given a partition with r parts, and m ≥ r, let  ( ; m) count the column-strict
tableaux with shape and elements in [m]. Let M be the m-by-m matrix with (i , j)-entry
(  j m+m −j
−i ) , where j = 0 for r < j ≤ m.
(a) Prove det M = V( 1 + m − 1 , 2 + m − 2 , . . . , m)/ ∏i=1 (m − i)!, where V(b1 , . . . , bm)
m

is the Vandermonde determinant with (i , j)-entry bjm−i , equal to ∏i< j (bi − bj ).

(b) Prove  ( ; m) = det M. (Littlewood [1940]; see also Bender–Knuth [1972])

4.3.15. (+) For a partition with distinct parts, the shifted shape is the set of locations
{(i , j): i ≤ j ≤ i + i − 1}; row i starts with the diagonal position (i , i). With distinct parts,
no row extends past the previous row. A standard shifted tableau fills the positions in
a shifted shape of size n with the elements of [n], increasing in both rows and columns.
Thrall [1952] proved that the number of standard shifted tableaux with shape is

∏ i< j ( i − j )
n! ,
∏i i ! ∏i< j ( i + j)

(a) For a position (i , j) in the shifted shape, define the shifted hook Hi∗, j by

Hi∗, j = {(i , s): s ≥ j} ∪ {(r, j): r ≥ i} ∪ { j + 1 , s): s ≥ j + 1}.

Below we show the hook H1∗ ,2 in the shape (5 , 3 , 2). Use part (a) of Exercise 4.3.13 to show
# #
that Thrall’s formula equals n!/ ∏i , j ### Hi∗, j ### (for (5 , 3 , 2), both formulas yield 54).
# #
(b) Prove Thrall’s formula for the number of shifted tableaux. (Hint: Prove that the
formula and the number of shifted tableaux satisfy the same recurrence. Comment: Other
proofs appear in Stanley [1972] and Sagan [1980]; the latter is probabilistic.)

4.3.16. Consider a zig-zag skew shape with n boxes, two in each row (except one in the first
row if n is odd), as shown below. Let an be the number of skew tableaux of this shape with
entries [n] (these correspond to permutations with up-down sequence + − + − · · · , called
alternating permutations; see Stanley [2010] for a survey).
208 Chapter 4: Further Topics

(a) Find a (nonlinear) recurrence for ⟨a⟩.

(b) Let (x) = ∑ an x n/n! and h(x) = 12 [ (x) − (− x)]. Prove h(x) (x) = (x) − 1.
(c) Solve the differential equation in (b) to obtain (x) = sec x + tan x. (André [1881])
4.3.17. Let T be a skew tableau with one hole, and let T be a partial tableau obtained
from T by a vertical SHIFT filling the hole. Prove G ¾ ,l(T) ≤ G ¾ ,l(T ) for all  , l. (Com-
ment: This completes the proof of Greene’s Theorem.)
4.3.18. The row-canonical permutation of a standard Young tableau is obtained by
reading the tableau row by row (left to right) starting at the bottom. Its column-
canonical permutation reads it column by column (bottom to top) starting at the left.
Prove that the P-symbol of a row-canonical or column-canonical permutation is the Young
tableau from which it was obtained. Conclude also that the reverse of a row-canonical
permutation is the column-canonical permutation for the transposed tableau.
4.3.19. (♦) In a permutation  (in word form), a transposition of two consecutive values is
admissible if some value between their values appears in a position next to one of them.
(Comment: Such transpositions are also called Knuth transformations, with permuta-
tions being Knuth-equivalent if one can be obtained from the other via such operations.
The definition and results extend also to generalize inputs (q , p).) (Knuth [1970])
(a) Prove P( ) = P( ) when  is obtained from  by an admissible transposition.
(Comment: There is a proof by induction, but using Jeu de Taquin is much simpler.)
(b) Prove that every permutation transforms to a row-canonical permutation using
admissible transpositions. (Hint: Model the bumping procedure.)
(c) Conclude from (b) that every permutation can be transformed to a column-
canonical permutation using admissible transpositions.
4.3.20. For a tableau T , the deletion operator produces a tableau d(T) by emptying the
upper left corner and filling the hole by Jeu de Taquin. Define  from  by deleting the
first occurrence of the least element. Prove P() = d(P( )). (Schützenberger [1963])
4.3.21. (♦) A balanced tableau of shape  is a tableau in which for every position (i , j),
the hook Hi , j (Definition 4.3.2) has ∗j − i numbers less than entry (i , j) and  i − j num-
bers greater than entry (i , j), where ∗ is the conjugate of . For example, there are three
balanced tableaux of shape (3 , 1) and 16 of shape (3 , 2 , 1). An n-staircase tableau is a
tableau with shape  such that  i = n − i for 1 ≤ i ≤ n − 1. For the n-staircase shape, let
B be the set of balanced tableaux, and let Y be the set of standard Young tableaux.
(a) Prove that an n-staircase tableau T is balanced if and only if for every position
(i , j) and all  with i <  ≤ n − j , exactly one of the entries ( , j) and (i , n + 1 − ) (in the
hook Hi , j ) is less than entry (i , j).
(b) One can pass from the identity permutation to its reverse via (2n) steps of trans-
posing two adjacent elements to increase the number of inversions by 1. Let P be the set of
such paths. Suppose that a path P ∈ P transposes adjacent elements a and b on step t. Let
(P) be an n-staircase tableau defined by putting t in position (n + 1 − b , a), for 1 ≤ t ≤ (2n) .
Prove that  is a bijection from P to B.
(Comment: Edelman–Greene [1987] proved this and other results about balanced
tableaux, including the amazing result that for every shape  there are the same num-
ber of balanced tableaux and Young tableaux. The combination of those results provided
a combinatorial proof of the algebraic result by Stanley [1984] that |P | and | B| are equal.)
Chapter 5

First Concepts for Graphs

The usefulness of graphs stems from their flexibility to model diverse situa-
tions: for example, in Theorem 4.1.35 we used paths in a directed graph to model
tilings of hexagons by rhombi. In Part II we discuss graphs for their own sake,
emphasizing proof techniques and classical topics. With this foundation, later we
explore further connections between graph theory and other areas.

5.1. Definitions and Examples

We introduced some definitions about graphs and directed graphs in Chap-
ter 0 to facilitate discussion of enumerative questions. We repeat these for conve-
nience and provide others to be consulted as needed when reading later material.
Although graph theory begins with an onslaught of definitions, most are natural
and visual, and the terms become familiar with use.

GRAPHS AND SUBGRAPHS

5.1.1. Definition. A graph G is an ordered pair consisting of a vertex set V(G)

and an edge set E(G), where each edge (element of E(G)) is a set of two ver-
tices (elements of V(G)). The vertices of an edge are its endpoints. We write
xy for an edge with endpoints x and y. The order of a graph G is | V(G)| ; its
size is | E(G)|. A graph with order n is an n-vertex graph.

A graph is defined “on” its vertex set by specifying the edges. It is finite if its
vertex set and edge set are finite. Our graphs are finite unless stated otherwise.

5.1.2. Definition. A path with n vertices is a graph whose vertices can be named
v1 , . . . , vn so that the edges are {vi vi+1 : 1 ≤ i ≤ n − 1}. We write ⟨v1 , . . . , vn ⟩
to specify a path having vertices v1 , . . . , vn in order.
A cycle with n vertices is a graph whose vertices can be named v1 , . . . , vn
so that the edge set is {vi vi+1 : 1 ≤ i ≤ n − 1} ∪ {vn v1 }. We write [v1 , . . . , vn]
to specify this cycle. The length of a path or cycle is the number of edges. A
cycle is an odd cycle [even cycle] if its length is odd [even].

209
210 Chapter 5: First Concepts for Graphs

The paths ⟨v1 , . . . , vn⟩ and ⟨vn , . . . , v1 ⟩ are the same graph. The square brack-
ets used to denote a cycle suggest “closing ” it. Since the starting point and direc-
tion do not matter, Definition 5.1.2 gives 2n ways to specify any n-vertex cycle.

5.1.3. Definition. A subgraph of a graph G is a graph H such that V(H) ⊆ V(G)

and E(H) ⊆ E(G). A spanning subgraph of G is a subgraph H such that
V(H) = V(G). An induced subgraph of G, induced by vertex set S ⊆ V(G),
is the subgraph H with vertex set S such that E(H) = {uv ∈ E(G): u , v ∈ S};
we write this as H = G[S]. A graph contains its subgraphs.

5.1.4. Example. Every cycle of length n contains n paths with n vertices; these
are not induced subgraphs. However, a cycle of length n does contain (n − 1)-
vertex paths as induced subgraphs. The graph below has 49 subgraphs (we list
three); 16 are induced subgraphs. Its only cycle is an odd cycle.
The only spanning induced subgraph of a graph G is G itself.

•
w• •
x
•y
vertex edge spanning induced
set set subgraph? subgraph?
w, x , y, xy , y yes no
x , y, xy , y no no
x , y, xy , y , x no yes

A graph defines relations on its vertex set.

5.1.5. Definition. Vertices x and y in a graph G are adjacent (or neighbors)

if xy ∈ E(G). We may write x ↔ y for “ x is adjacent to y ” and x  y
for “ x is not adjacent to y ”, denoting the adjacency and nonadjacency
relations on V(G). The neighborhood NG(x) or N(x) of a vertex x is
{y ∈ V(G): xy ∈ E(G)}. An isolated vertex is a vertex x with N(x) = ∅.
The degree of a vertex v in a graph G, written d G(v) or d(v), is the num-
ber of edges containing v. The minimum vertex degree is (G); the maximum
is (G). A graph G is regular if (G) = (G); -regular if that degree is  .
Graphs that are 3-regular are also called cubic graphs.

5.1.6. Definition. A complete graph is a graph whose vertices are pairwise ad-
jacent. A set of pairwise adjacent vertices is a clique; it is a -clique when
it has size  . A set of pairwise nonadjacent vertices is an independent set,
also called a stable set. The complement of a graph G is the graph G with
vertex set V(G) defined by uv ∈ E(G) if and only if uv ∈ / E(G).

A complete graph with n vertices is (n − 1)-regular; a cycle is 2-regular. The

vertex set of a complete subgraph is a clique. In Example 5.1.4, the largest clique
has three vertices, and the largest stable set has two. In the complement, ⟨y , w , ⟩
is a path and x is isolated, having no neighbor. The complement of an n-vertex
graph G has (2n) − | E(G)| edges, and the complement of a regular graph is regular.
Section 5.1: Definitions and Examples 211

5.1.7. Definition. A graph G is bipartite if V(G) is the union of two independent

sets called its parts. An X , Y -bigraph is a bipartite graph G with parts X
and Y , and (X , Y) is then a bipartition of G. An X , Y -bigraph with all of
X adjacent to all of Y is a complete bipartite graph. A ¾-partite graph
is a graph whose vertex set is the union of ¾ independent sets.

5.1.8. Example. A graph is 1-partite if and only if it has no edges, such as the
complement of a complete graph. Every graph is ¾-partite for some ¾ , since indi-
vidual vertices are independent sets. When l ≥ ¾ , every ¾-partite graph is also
l-partite, since independent sets may be empty. The complement of a complete
bipartite graph consists of two disjoint complete graphs.
The vertices of a path or cycle in a bipartite graph must alternate between
the parts. Hence every cycle in a bipartite graph is an even cycle. Below are three
drawings of a complete bipartite graph with three vertices in each part. The
fourth drawing also has six vertices, all of degree 3, but it cannot be a drawing of
a bipartite graph because it has an odd cycle.

• • • • • • • • •
• • • • • •
• • • • • • • • •

We draw a graph on paper by assigning each edge a curve joining points as-
signed to the endpoints of the edge. We require that curves for distinct edges have
only finitely many common points. Since the incidence relation associating each
curve with its endpoints is the incidence relation of the graph, we think of the
drawing as the graph. More precisely, it is a member of the isomorphism class of
the graph, which brings us to the next topic.

ISOMORPHISM

To specify a graph, we must specify the vertices and edges. We can list edges
or neighborhoods. Also matrix representations can be useful.

5.1.9. Definition. For an ordering v1 , . . . , vn of V(G), the adjacency matrix

A(G) is the 0 , 1-matrix having Ai , j = 1 if and only if vi ↔ vj .
Vertex v and edge e are incident if v is an endpoint of e. Also edges with
a common endpoint are incident. Given a vertex ordering v1 , . . . , vn and
an edge ordering e1 , . . . , e m , the incidence matrix M(G) is the 0 , 1-matrix
defined by Mi , j = 1 if and only if vi ∈ e j .

5.1.10. Example. Below are two adjacency matrices and an incidence matrix for
the 4-vertex path ⟨w , x , y , ⟩ with edges e1 , e2 , e3 in order.
w x y w y x e1 e 2 e3
w⎛ 0 1 0 0⎞ w⎛ 0 0 0 1⎞ w⎛ 1 0 0 ⎞
x ⎜ 1 0 1 0⎟ y ⎜ 0 0 1 1⎟ x ⎜1 1 0⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
y ⎜0 1 0 1⎟ ⎜ 0 1 0 0⎟ y ⎜0 1 1⎟
⎝ 0 0 1 0⎠ x ⎝ 1 1 0 0⎠ ⎝0 0 1⎠
212 Chapter 5: First Concepts for Graphs

The adjacency matrix depends on the vertex ordering. However, structural

properties do not depend on this ordering or on how we draw a graph or name its
vertices. Isomorphism captures this notion.

5.1.11. Definition. An isomorphism from a graph G to a graph H is a bijection

º : V(G) → V(H) such that uv ∈ E(G) if and only if º (u)º (v) ∈ E(H). When
an isomorphism exists, we write G ∼
= H and say “ G is isomorphic to H ”. The
isomorphism relation is the set of pairs (G , H) such that G ∼= H.
5.1.12. Remark. Isomorphism classes and “unlabeled” graphs. Isomorphisms
can be composed and inverted. Hence isomorphism is an equivalence relation on
graphs, and the equivalence classes are called isomorphism classes. A drawing
of a graph specifies a member of its isomorphism class, so we often omit vertex
and edge labels in drawings.
The informal term unlabeled graph indicates an isomorphism class. When
two graphs are isomorphic, we often use the same name for them, because the ob-
ject of importance is the isomorphism class, and all members of it have the same
structural properties. Thus we often write G = H instead of G ∼ = H ; asking
whether a given graph is G means asking whether it is isomorphic to G. We say
“graph” in discussing an isomorphism class to refer to each member of the class.
Exercises 15–27 study isomorphism. To prove G ∼ = H , one shows that a partic-
ular vertex bijection is an isomorphism or that G and H have the same structural
description. To prove G ∼
H , one shows that some aspect of their structure that
=
does not depend on vertex names differs. For example, a bipartite graph cannot
be isomorphic to a graph having an odd cycle.

5.1.13. Definition. Special isomorphism classes. We use Cn , Pn , K n for the cycle,

path, and complete graph of order n, respectively, viewed as isomorphism
classes. We use K r,s for the complete bipartite graph with parts of sizes r and
s. A star is a graph of the form K 1 ,t . The claw is K 1 ,3 . The triangle is K 3 .

We use “ H is a subgraph of G ” to mean that G has a subgraph isomorphic

to H. Thus we say that C¾ is a subgraph of G whenever G has a cycle with ¾
vertices. A subgraph of G isomorphic to H is a copy of H in G.

5.1.14. Definition. An automorphism of G is an isomorphism from G to itself;

G is vertex-transitive if for any u , v ∈ V(G), some automorphism maps u to
v. It is edge-transitive if for any e , º ∈ E(G), some automorphism maps e
(as a set of two vertices) to º .

The identity is an automorphism, composition of automorphisms is an auto-

morphism, and they are invertible, so they form a group under composition.

5.1.15. Example. The cycle Cn has 2n automorphisms (n rotations and n reflec-

tions), corresponding to the 2n ways we can specify the cycle by listing its vertices
(compare to Example 4.2.23).
The path Pn has two automorphisms, and K n has n!. Although K r,s has r!s!
automorphisms when r
= s, it has 2r!s! when r = s. Both Cn and K n are vertex-
transitive and edge-transitive; Pn is neither. Although K r,s is edge-transitive, it
is not vertex-transitive when r
= s.
Section 5.1: Definitions and Examples 213

THE PETERSEN GRAPH AND HYPERCUBES

The complete graph K n has n! automorphisms, but its structure is not very
interesting, and it has too many edges for practical applications. The cycle Cn
has only n edges, but also only 2n automorphisms. We next discuss a 10-vertex
graph that is almost as sparse as C10 but has many more automorphisms and rich
structure. It is an important example in many contexts, and there are many fam-
ilies in which this is the smallest or the only graph. An entire book was devoted
to topics involving it (Holton–Sheehan [1993]).

5.1.16. Example. The Petersen graph. The graphs below are isomorphic (more
precisely, any two of them are isomorphic). We refer to any graph isomorphic
to these as the Petersen graph. A structural description of the Petersen graph
names its vertices as the 2-element subsets of [5] (we write the vertex names with-
out set brackets). Two vertices form an edge in this graph if and only if as 2-sets
from [5] they are disjoint. Although the graph is named for Petersen [1891], it
appeared earlier (using the rightmost drawing) in Kempe [1886, p. 11].
12
• • • • •
• • • •
45• 13 •35 34
• • • • • • • • • • • • •
41• •
52
• • •
• •
24
• • • •
23• •51 • • •

We derive several important properties of the Petersen graph.

5.1.17. Definition. A ¾-cycle is a cycle of length ¾ . The girth of a graph with a

cycle is the length of a shortest cycle in the graph (with no cycle, the girth is
infinite). A chord of a cycle C is an edge not in C whose endpoints are in C.

5.1.18. Proposition. The Petersen graph G is 3-regular, and any two nonadja-
cent vertices in it have exactly one common neighbor. Also it has girth 5 but
no spanning cycle.
Proof: View the vertices as 2-sets in [5], with adjacency being disjointness. There
are three ways to select two elements of [5] − {a , b}, so every vertex has degree 3.
If ab and cd are nonadjacent, then they have a common element. Their union
omits one pair in [5], which is their unique common neighbor.
A 3-cycle requires three pairwise disjoint pairs in [5], which don’t exist. A 4-
cycle thus requires nonadjacent vertices with two common neighbors, also forbid-
den. Hence every cycle has length at least 5, and [12 , 34 , 51 , 23 ,45] is a 5-cycle.
Suppose that G contains a spanning cycle C. The remaining edges must be
chords of C. Since G has girth 5, chords can only join vertices that are opposite
or nearly-opposite on C. Making two consecutive vertices of C adjacent to the
vertices opposite them creates a 4-cycle. By symmetry, we may therefore assume
v0 v4 ∈ E(G), where C = [v0 , . . . , v9 ]. Now every way of giving v9 another neighbor
creates a cycle of length at most 4.
214 Chapter 5: First Concepts for Graphs

In fact, the Petersen graph is the only 3-regular graph with 10 vertices that
has girth at least 5 and has no spanning cycle (Exercise 32). The structural de-
scription immediately makes it vertex-transitive and edge-transitive. Every per-
mutation of [5] permutes the pairs in [5] and preserves the disjointness relation;
this generates 120 automorphisms of the Petersen graph. In fact, these are all
the automorphisms (Exercise 33).
Our next family of graphs has many applications in computer science.

5.1.19. Example. The ¾-dimensional cube or hypercube is the graph Q¾ with

vertex set {0 , 1}¾ such that two ¾-tuples are adjacent if and only if they differ in
exactly one position. Below we draw Q3 . The number of ways to form two vertices
differing in one coordinate is exactly ¾ 2¾−1 , so | E(Q¾)| = ¾ 2¾−1 . Each edge joins
vertices whose names have an even and an odd number of 1s, so Q¾ is bipartite.
Permuting coordinates and/or switching 0 and 1 within fixed coordinates gen-
erates ¾ !2¾ automorphisms of Q¾ ; these are all the automorphisms (Exercise 42).
For ¾ ≥ 1, the graph Q¾ arises naturally from two copies of Q¾−1 . Append 0
to the name of each vertex in the first copy, and append 1 to the name of each ver-
tex in the second copy. Adding one edge from each vertex of the first copy to the
corresponding vertex of the second copy completes Q¾ . Below we show Q3 gener-
ated from Q2 in this way, with the first copy outside and the second copy inside.
The inductive description facilitates many proofs about Q¾ by induction on ¾ .

010 110
• •

011 • • 111

001 • • 101

• •
000 100

The inductive construction of Q¾ is a special case of a general operation for

combining two graphs.

5.1.20. Definition. The cartesian product of graphs G and H , written G H ,

is the graph with vertex set V(G) × V(H) such that (u , v) is adjacent to (u , v )
if and only if (1) u = u and vv ∈ E(H), or (2) v = v and uu ∈ E(G).

5.1.21. Example. The cartesian product is symmetric; G H = ∼ H G. As exam-

ples, C3 K 2 appears below, and Q¾ = Q¾−1 K 2 if ¾ ≥ 1. In general, G H is an
edge-disjoint union of copies of H for each vertex of G plus copies of G for each
vertex of H. Thus d G H (u , v) = d G(u) + d H (v).
b• b1• •b2
1 2
•c • • = c1 • •c2
a• a1• •a2
Section 5.1: Definitions and Examples 215

We use instead of × to avoid confusion with other product operations, re-

serving × for the cartesian product of vertex sets. The symbol visually repre-
sents the observation that K 2 K 2 is a cycle with four vertices. This notation was
popularized by Nešet řil.

We mention several additive combining operations on graphs.

5.1.22. Definition. The union of graphs G and H is the graph G ∪ H with vertex
set V(G) ∪ V(H) and edge set E(G) ∪ E(H). We use G + H for the disjoint
union when V(G) and V(H) are disjoint; ¾ G is the disjoint union of ¾ copies
of G. The join of G and H , denoted G H , is obtained from G + H by adding
edges to make all of V(G) adjacent to all of V(H).

5.1.23. Example. Union and join. Using the natural vertex names, G ∪ G is
a complete graph. The notation G + H for disjoint union (usually for unlabeled
graphs) is convenient for making “arithmetic of graphs” consistent with ordinary
arithmetic: mG = G + · · · + G. The symbol “ ” illustrates the graph K 2 K 2 ,
with the interior “ + ” reflecting the additivity of the sizes of the vertex sets. This
is consistent with the notation “ ” for cartesian product.
Using G = K 1 ,2 and H = K 3 , we illustrate below their disjoint union, their
join, and their union in the case where the vertices are named so that the end-
points of G are two of the vertices of H.
We can express K r,s using K r,s = K r + K s or K r,s = (rK 1) (sK 1). More gener-
ally, G + H = G H. Using the word “ join” in “u and v are joined by an edge” is
consistent with the definition of the join of two graphs.

• • • • •
• • • • • •
• • • • •
G+H G H G∪H

5.1.24.* Remark. There is no standard notation for the order and size as graph
parameters. Textbook authors have tried: (G) , (G) (Bondy–Murty [1976]);
v(G) , e(G) (Bondy–Murty [2008]); | G | , e(G) (Bollobás [1978, 1998]); n(G) , e(G)
(West [1996, 2001]); | G | ,  G  (Diestel [1997, etc.]). Conflicts arise with us-
ing v and e as a specific vertex and edge and using n as an integer variable.
Reserving special characters has problems when there are several graphs un-
der discussino: p , q (Harary [1969], Gould [1988]); v , e (Wallis [2000]); n , m
(Chartrand–Lesniak [1986, etc.], Volkmann [1991, 1996], Buckley–Lewinter
[2002]). Although | G | ,  G  avoids these problems, internet respondents disliked
them. V. Strehl suggested #V(G) , #E(G); this avoids the difficulties but is not in
common use.
We therefore introduce no notation for these operators, but we often say “let
G be an n-vertex graph” or “let n = | V(G)| ”. Thus we encourage using n for the
order of a particular graph, with no global definition (this is common in computer
science). We similarly often let m = | E(G)| , reserving “ e” for a specific edge.
216 Chapter 5: First Concepts for Graphs

EXERCISES 5.1

5.1.1. (−) Count the automorphisms of Pn , Cn , and K n .

5.1.2. (−) Using the disjointness definition of the Petersen graph, prove that the graphs
of Example 5.1.16 are isomorphic.
5.1.3. (−) Prove by induction that | E(Q¾)| = ¾ 2 ¾−1 .
5.1.4. (−) Prove that each graph below is isomorphic to C3 C3 .
• •
• •
• •
•
• •
• •
• • • •
• • •

5.1.5. (−) Let G be a graph with rs vertices that decomposes into rK s and sK r . Prove
G∼ = K r K s . Construct infinitely many examples to prove that in general a graph with rs
vertices may decompose into rH and sF , where r = | V(F)| and s = | V(H)| , without being
isomorphic to F H .
5.1.6. (−) Determine which pairs of graphs below are isomorphic.
d
• •c v• • • •

• •
h• • u• •y
e• • t• •x
• •
a
• •b s• •w • •

5.1.7. (−) Prove that the graph below is isomorphic to Q3 .

• •
• •
• •
• •

5.1.8. (−) Let G be a graph with adjacency matrix A and incidence matrix M . What do
the entries in position (i , j) of A2 and MM T say about G?
5.1.9. (−) In terms of a known enumerative problem, find the number of isomorphism
classes of X , Y -bigraphs such that | X | = m and | Y | = n and every vertex of X has degree 1.
5.1.10. (−) In terms of the numbers of vertices and edges in a graph G, count the vertices
and edges in the cartesian product G · · · G with  factors all isomorphic to G.
5.1.11. (−) Prove that if (G) ≥ 3, then some cycle in G has a chord. (Czipzer [1963])
5.1.12. Let G n be the graph whose vertices are the permutations of [n], with two permu-
tations a1 , . . . , an and b1 , . . . , bn adjacent if they differ by switching two entries. Prove
that G n is bipartite.
Exercises for Section 5.1 217

5.1.13. Determine the least number of vertices in a graph in which the minimum length
of an even cycle is r and the minimum length of an odd cycle is s. (Harary–Kovacs [1982])

5.1.14. (+) Determine all r such that Cr K 2 can be expressed as the union of two cycles.

5.1.15. Determine whether the graphs below are isomorphic.

• • • •
• • • •

• • • •
• • • •

5.1.16. Prove that the two graphs below are isomorphic.

• • •
• •
• • • • •
• •
•• • •
• •
• •
• • • •

5.1.17. Determine which pairs of graphs below are isomorphic.

• • • • •
• • • • • • •
• • • • • •
• • ••
• • • • • •
• • • •
• • • •
• • • • • •

5.1.18. Determine which pairs of graphs below are isomorphic.

• • • • • • •
• • • • • •
• • • • • •
• •
• • • • • •
• •
• • • • • •
• • • • • •
• • • • • • •

5.1.19. Determine which pairs of graphs below are isomorphic.

• • • • • • • •
• • • • •
• •
• • • • • • •
• • • • • • • •
218 Chapter 5: First Concepts for Graphs

5.1.20. (♦) Determine which pairs of graphs below are isomorphic.

• • • • • •
• • • • • • • •
• • • • • • • • • • •
• • • • • • • • • •

5.1.21. (♦) Determine which pairs of graphs below are isomorphic, presenting the proof
by testing the fewest possible pairs.

• • • • • • • • • • • • • • • •
• •• • • •• • • •
• • • • • • • • • • • • • •
• • • • • • • • • • • • • •
• •• • • •• • • •
• • • • • • • • • • • • • • • •

5.1.22. For each ¾ at least 2, determine the smallest n such that

(a) There is a ¾-regular graph with n vertices.
(b) There exist nonisomorphic ¾-regular graphs with n vertices.
5.1.23. For ¾ ≥ 2, determine the smallest n such that there exist non-isomorphic ¾-regular
bipartite graphs with n vertices.
5.1.24. Determine the number of isomorphism classes of 3-regular graphs with six ver-
tices and with eight vertices.
5.1.25. (♦) A graph isomorphic to its complement is self-complementary. Prove that
there is a self-complementary graph with n vertices if and only if n or n − 1 is divisible by
4. Determine for which n there is a regular self-complementary n-vertex graph.
5.1.26. Let G be a self-complementary graph. Let be a permutation of V(G) that is an
isomorphism from G to G. Prove that every cycle of other than fixed points has length
divisible by 4. (Ringel [1963])
∼ G and n ≡ 1 (mod 4). Prove that G has at
5.1.27. Let G be an n-vertex graph with G =
least one vertex of degree (n − 1)/2.
5.1.28. Suppose that n is congruent to 0 or 1 modulo 4. Construct an n-vertex graph G
with 12 (2n) edges such that (G) − (G) ≤ 1.
5.1.29. Construct a 6-vertex graph having no nontrivial automorphism. Construct a
graph having exactly three automorphisms.
5.1.30. Edge-transitive versus vertex-transitive.
(a) For n ≥ 4, form a graph G by replacing each edge of K n with a path of length two
(the new vertices have degree 2). Prove that G is edge-transitive but not vertex-transitive.
(b) Prove that every edge-transitive graph that is not vertex-transitive is bipartite.
5.1.31. Prove that the Petersen graph has no cycle of length 7.
5.1.32. Prove that the Petersen graph is the only 3-regular graph with 10 vertices that
has girth at least 5.
5.1.33. View the Petersen graph as the graph whose adjacency relation is the disjointness
relation on the 2-element subsets of [5]. Prove that every automorphism is determined by
a permutation of [5], with vertex i j mapped to the vertex (i) (j) for all i , j ∈ [5].
Exercises for Section 5.1 219

5.1.34. Count the independent sets of size 4 and the independent sets of size 3 in the Pe-
tersen graph. Find all smallest sets whose deletion eliminates all independent 4-sets.
5.1.35. Let e and e be non-incident edges in the Petersen graph. Prove that e and e lie
in a 5-cycle or in a 6-cycle. Determine the number of pairs of nonincident edges that lie in
a 5-cycle and the number that do not.
5.1.36. The Odd graph O¾ is the disjointness graph on the ¾-subsets of [2 ¾ + 1] (O2 is the
Petersen graph). For ¾ > 2, determine the girth of O¾ .
5.1.37. For the Heawood graph below, determine all cycle lengths and find a second
spanning cycle.
• •
• •
• •
• •
• •
• •
• •

5.1.38. For even n, determine the minimum number of distinct cycle lengths in an n-
vertex graph with minimum degree n/2.
5.1.39. Determine the smallest bipartite graph that is not a subgraph of the ¾-dimensional
cube Q¾ for any ¾ .
5.1.40. Define a graph Q ¾ with vertex set {0 , 1}¾ by uv ∈ E(Q¾ ) if and only if u and v agree
in exactly one coordinate. Prove that Q ¾ ∼
= Q¾ if and only if ¾ is even. (D.G. Hoffman)

5.1.41. For ¾ ∈ , let G be the subgraph of Q2¾+1 induced by the vertices in which the num-
ber of ones and zeros differs by 1. Prove that G is vertex-transitive and edge-transitive,
and compute the order, size, and girth of G.
5.1.42. Prove that Q¾ has exactly ¾ !2 ¾ automorphisms.
5.1.43. A subcube of dimension r in Q¾ is a subgraph induced by 2 r vertices that agree
on some ¾ − r coordinates.
(a) Prove that every cycle of length 2r in Q¾ lies in a subcube of dimension at most r.
Prove that this subcube is unique when r = 2 or r = 3 but need not be unique when r = 4.
(b) Count the 4-cycles and 6-cycles in Q¾ .
5.1.44. (♦) Regular graphs with no short cycle.
(a) Prove that every ¾-regular graph with girth at least 4 has at least 2 ¾ vertices.
Determine all such graphs where equality holds.
(b) Prove that every ¾-regular graph with girth at least 5 has at least ¾ 2 + 1 vertices.
Find examples where equality holds when ¾ is 2 or 3.
5.1.45. (♦) Prove that every ¾-regular graph with girth at least 6 has at least 2 ¾ 2 − 2 ¾ + 2
vertices. (Comment: The Heawood graph of Exercise 5.1.37 achieves equality when ¾ = 3.)

5.1.46. (+) Fix n , ¾ ∈ with 1 < ¾ < n − 1. Let G be an n-vertex graph such that every ¾-
vertex induced subgraph of G has m edges. For l ≥ ¾ , prove that every induced subgraph
of G with l vertices has m(¾l )/(¾l−−22) edges. Use this to prove G = K n or G = K n .
5.1.47. (+) A (¾ , ½)-cage is a smallest ¾-regular graph with girth ½ . For ¾ ≥ 2 and ½ ≥ 3,
prove that there exists a ¾-regular graph with girth ½ . (Hint: For ½ ≥ 5, construct such
a graph from a (¾ − 1)-regular graph H with girth ½ and a graph with girth ⌈ ½/2⌉ that is
regular with degree | V(H)|.) (Erdős–Sachs [1963])
220 Chapter 5: First Concepts for Graphs

5.2. Vertex Degrees

Vertex degrees give more information than the numbers of vertices and
edges. Here we briefly consider enumerative, existential, and extremal problems
about degrees. We also introduce corresponding terminology for directed graphs.

THE DEGREE-SUM FORMUL A

A simple counting argument yields a fundamental tool counting edges in
terms of vertex degrees. It has been called the “First Theorem of Graph The-
ory” and the Handshaking Lemma (modeling handshakes as edges in a graph).

5.2.1. Proposition. (Degree-Sum Formula) For a graph G with m edges,

1
2 ∑
m= d(v).
v∈ V(G)

Proof: Grouped by edges, there are 2m pairs (v , e) such that e is an edge incident
to vertex v. Grouped by vertices, there are ∑v∈V(G) d(v) pairs.

Thus a ¾-regular n-vertex graph has ¾ n/2 edges (Q¾ has ¾ 2¾−1 edges).

5.2.2. Corollary. Every graph has an even number of vertices of odd degree, and
thus no regular graph of odd degree has an odd number of vertices.

We expressed the argument for the Degree-Sum Formula as “counting two

ways” for the set of incidences of edges with vertices. More simply, summing the
degrees counts every edge twice. Here is another example of multiple counting.

5.2.3. Example. The Petersen graph has twelve 5-cycles. The Petersen graph has
15 edges (it is 3-regular with ten vertices). We claim that every edge lies in four
5-cycles. If so, then 4 · 15 counts every edge in every 5-cycle; since a 5-cycle has
five edges, there must thus be 12 of them.
A 5-cycle through edge e contains a copy of P4 with central edge e. The end-
points of such a path are nonadjacent, since the girth is 5. Nonadjacent vertices
have one common neighbor, so each copy of P4 extends to one 5-cycle. Since e has
two incident edges at each end, and there are no triangles, e is the central edge
of exactly four copies of P4 . Thus e lies in four 5-cycles.

The simple statements in Corollary 5.2.2 have nontrivial applications. If a

graph has an odd-degree vertex, then it must have another. We present one ap-
plication; see also Exercise 12, Exercise 13, Theorem 5.4.7, and Theorem 16.2.2.

5.2.4. Proposition. If a rectangular region R is decomposed into nonoverlapping

rectangular subregions R1 , . . . , R¾ , and each Ri has a side of integer length,
then the full region R has a side of integer length.
Section 5.2: Vertex Degrees 221

Proof: Translate one corner of R to the origin, as illustrated below. Form a bi-
partite graph G whose two parts are the subrectangles and the integer lattice
points, with Ri adjacent to (r, s) if (r, s) is a corner of Ri .

R2
2• • R1 R2 R3 R4 R5
R4 • • • • •
R3
1 R1 • •
• • • • • •
R5 (0,0) (0,2) (2,2) (2,1) (3,1) (3,0)
0• •
0 1 2 3
Since each subrectangle has a side of integer length, it has an even number
of integer corners; hence each vertex Ri has even degree in G. An integer point
is the corner of an even number of subrectangles if it is internal or on the side
of R, but it is a corner of only one subrectangle if it is a corner of R. Hence any
vertex of odd degree in G is an integer corner of R. Since the origin is one such
corner, Corollary 5.2.2 yields another. Hence R has a side of integer length.

Wagon [1987] published 14 proofs of Proposition 5.2.4, some using complex

analysis. Another proof appeared in Winkler [2004, pp. 6–7].

DEGREE LISTS

We may ask whether some graph in a given class has specified values of cer-
tain parameters? For example, consider the vertex degrees. We can ask this about
the class of all n-vertex graphs.

5.2.5. Definition. The degree list or degree sequence of a graph G is the list
of its vertex degrees, usually written in nonincreasing order. A list of inte-
gers is graphic if it is the degree list of some graph G; then G realizes it.
When v ∈ V(G), vertex deletion yields the graph denoted G − v by delet-
ing v and all its incident edges from G.

For example, (3 , 3 , 1 , 1) is not graphic; when two vertices in a 4-vertex graph

have degree 3, the other two vertices must have degree at least 2. Degree “se-
quence” is the traditional term, but a sequence is a function defined on 0 , so the 
term “list ” is more correct. We next determine which integer lists are graphic.

5.2.6. Theorem. (Havel [1955], Hakimi [1962]) For n > 1, a nonnegative integer
list d of size n is graphic if and only if d is graphic, where d is the list of
size n − 1 obtained from d by deleting its largest element ¾ and subtracting
1 from its ¾ next largest elements.
Proof: If d = 3322211, for example, then d = 211211. Given a list d and a graph
G with degree list d , we add a new vertex adjacent to vertices in G having de-
grees that were reduced in forming d from d. The resulting graph G realizes d
and proves that the condition is sufficient.
222 Chapter 5: First Concepts for Graphs

For necessity, we begin with a graph G realizing d and find G realizing d .

Let ¾ = (G), and let w be a vertex of degree ¾ in G. Let S be a set of ¾ other
vertices in G with largest degrees. If N(w) = S, then we let G = G − w.
Otherwise, some vertex x of S is missing from N(w), and w has a neighbor
outside S (since d(w) = | S|). Note that d(x) ≥ d( ), by the choice of S. Since
w ∈ N( ) − N(x), some vertex y outside {x , , w} must satisfy y ∈ N(x) − N( ), as
illustrated below. Replacing {w , xy} with {wx , y } in E(G) yields a realization
of d such that w has more neighbors in S. Since | N(w) ∩ S| can increase at most
 times, repeating this procedure converts G into a graph G ∗ that realizes d and
has N(w) = S. Now G ∗ − w is the desired graph G realizing d .

• S
x
•
w• • •y
•

The operation we used has further applications.

5.2.7. Definition. Given xy , w ∈ E(G) and y , wx ∈ E(G), a 2-switch is the

replacement of {xy , w} with {wx , y } in G.

5.2.8. Theorem. (Fulkerson–Hoffman–McAndrew [1965], Berge [1970]) If G

and H are graphs with vertex set V , then 2-switches can turn G into H if
and only if d G(v) = d H (v) for every v ∈ V .
Proof: Every 2-switch preserves vertex degrees, so the condition is necessary.
For the converse, we use induction on | V | . When | V | ≤ 3, each specification of
degrees is realized by at most one graph.
For | V | ≥ 4, let  = (G), and let w be a vertex of degree  . Let S be a fixed set
of  vertices other than w with the highest degrees. As in the proof of Theorem
5.2.6, we use 2-switches to turn G into a graph G ∗ with NG∗ (w) = S and turn H
into a graph H ∗ with NH ∗ (w) = S.
Since NG∗ (w) = NH ∗ (w), deleting w leaves graphs G = G ∗ − w and H = H ∗ − w
with d G (v) = d H (v) for every vertex v. By the induction hypothesis, G trans-
forms to H by 2-switches. Since these do not involve w, and w has the same
neighbors in G ∗ and H ∗ , the same 2-switches transform G ∗ to H ∗ . Hence we can
transform G to H by transforming G to G ∗ , then G ∗ to H ∗ , then (in reverse order)
the transformation of H to H ∗ (the inverse of a 2-switch is a 2-switch).

w w w w
• • • •
• •••• •••• •••• • ••••
• S → S → S ← S
• G H
•
•
G G∗ H∗ H
Section 5.2: Vertex Degrees 223

EXTREMALITY

An important technique in discrete mathematics is the selection of an ex-

tremal instance of some structure. The absence of a more extreme structure gives
extra leverage to prove the desired conclusion. We give two examples here.

5.2.9. Theorem. Every graph G with m edges has a bipartite subgraph with at
least m/2 edges.
Proof: Among all bipartite subgraphs of G, let H be one having the most edges.
We may assume that H has bipartition X , Y with X ∪ Y = V(G), and every edge
of G having endpoints in both X and Y belongs to H.

• •

X • • • Y

v• •

If some vertex v contributes fewer than half of its incident edges to H , then
v has more neighbors in its own part of the bipartition than in the opposite part.
Moving v to the other part puts more edges into H than it deletes from H , yield-
ing a larger bipartite subgraph.
Since H is a bipartite subgraph with the most edges, we conclude that
d H (v) ≥ d G(v)/2 for every v ∈ V(G). Summing over all vertices and applying
the Degree-Sum Formula yields | E(H)| ≥ | E(G)| /2.

We could also start with a spanning bipartite subgraph and move vertices to
gain edges until we reach the desired degree property. Choosing a largest bipar-
tite subgraph skips directly to that point. Many proofs by extremality have equiv-
alent inductive or algorithmic versions; using extremality is usually shorter. We
next use extremality by choosing a vertex of largest degree.

5.2.10. Definition. A graph is triangle-free if it does not contain K 3 . More

generally, a graph G is H-free if G has no induced subgraph isomorphic to
H. The Tur án graph T n ,r is the n-vertex complete r-partite graph whose
parts differ in size by at most 1.

5.2.11. Theorem. (Tur án’s Theorem; Turán [1941]) The largest n-vertex
K r+1 -free graph is T n ,r . For the triangle-free case r = 2, it has ⌊ n2/4⌋ edges.
Proof: (Erdős [1970]) Every r-partite graph is K r+1 -free, since any r + 1 vertices
have two in the same part, which are not adjacent. A largest r-partite graph is a
complete r-partite graph. If two parts differ in size by more than 1, then moving
a vertex from the larger part to the smaller part gains more edges than it loses.
Hence the graph described is the unique largest r-partite graph with n vertices.
It thus suffices to prove that when G is K r+1 -free, there is an r-partite graph
H with V(H) = V(G) such that d H (v) ≥ d G(v) for all v ∈ V(G) (and therefore
| E(H)| ≥ | E(G)|). We prove this by induction on r, trivial for r = 1.
224 Chapter 5: First Concepts for Graphs

Suppose r > 1. Choose x ∈ V(G) with d G(x) = (G). Since K r+1

⊆ G, we
have K r
⊆ G , where G = G[N(x)]. By the induction hypothesis, there is an
(r − 1)-partite graph H with d H (v) ≥ d G (v) for all v ∈ N(x).
Let S = V(G) − N(x). Form H from H by making every vertex of S adjacent
to every vertex of N(x), leaving S an independent set. Note that H is r-partite.
Now d H (v) = d G(x) = (G) for v ∈ S. Hence d H (v) ≥ d G(v) for all v ∈ S, and H has
all the required properties.
In the special case r = 2, the triangle-free graph with the most edges has
parts of size ⌊ n/2⌋ and ⌈ n/2⌉ , and the size is ⌊ n2/4⌋ .
Uniqueness of the extremal graph also follows inductively.
x• •
N(x) x N(x)
• → S ••
• G H
• •
• •

Mantel [1907] (jointly with others!) proved the case r = 2. We will revisit
Turán’s Theorem in Chapters 11 and 14 to study other techniques.

DIRECTED GRAPHS

Edges in a graph are unordered pairs of distinct vertices, so graphs model

binary relations that are symmetric and irreflexive. A general relation on a set
V is a subset of the cartesian product V × V . We can model general relations
using ordered pairs of vertices as edges.

5.2.12. Definition. A directed graph or digraph G is an ordered pair consist-

ing of a vertex set V(G) and an edge set E(G), with E(G) ⊆ V(G) × V(G).
An edge exits from its first vertex (the tail) and enters its second vertex (the
head). We write xy instead of (x , y) for an edge from x to y. An edge from x
to x is a loop. Given a digraph G, we may write x → y to mean xy ∈ E(G).

The term digraph was introduced by Harary [1955a] at the request of P ólya
(as told in Harary [1989]). It is natural to allow loops in digraphs, such as in
the functional digraphs of permutations. For the analogue of vertex degrees, we
count entering and exiting edges separately.

5.2.13. Definition. For a vertex x in a digraph G, the indegree d−G(x) and out-
degree d+G(x) are the number of edges entering x or leaving x, respectively.
The out-neighborhood NG+(x) is {y: x → y}; the in-neighborhood NG−(x)
is {w: w → x}. Subscripts may be dropped. A vertex with indegree 0 is a
source; one with outdegree 0 is a sink.
Given a vertex ordering v1 , . . . , vn , the adjacency matrix A of a di-
graph is the 0 , 1-matrix defined by Ai , j = 1 if and only if vi → vj . With edges
e1 , . . . , e m , the incidence matrix M of a loopless digraph has Mi , j = 1 if vi
is the tail of e j , Mi , j = −1 if vi is the head of e j , and Mi , j = 0 otherwise.
Section 5.2: Vertex Degrees 225

5.2.14. Example. Compare the matrices for the digraph below with those for P4
in Example 5.1.10.

w x y a b c d
x
w⎛ 0 1 0 0⎞ • c • w ⎛ +1 0 0 0 ⎞
x ⎜0 0 1 0⎟ x ⎜ −1 +1 −1 0 ⎟
⎜ ⎟ a d ⎜ ⎟
y ⎜0 1 0 0⎟ y ⎜ 0 −1 +1 −1 ⎟
⎝0 0 1 0⎠ • b •y ⎝ 0 0 0 +1 ⎠
w
A(G) G M(G)

A graph can be viewed as a special kind of digraph.

5.2.15. Definition. A digraph is symmetric if its adjacency matrix is symmet-

ric (a graph is a symmetric loopless digraph). A digraph D is antisymmetric
if xy and yx are not both edges when x , y ∈ V(D). To form the underlying
graph of an antisymmetric digraph, make the edges unordered. An orien-
tation of a graph G is an antisymmetric digraph whose underlying graph is
G. An oriented graph is a digraph that is an orientation of a graph.

In drawing a digraph, we view the curve for an edge as an “arrow” from its
tail to its head. Basic properties of digraphs are like those of graphs, so we focus
primarily on graphs. For example, definitions of sub(di)graphs and isomorphism
are the same as those for graphs. Nevertheless, interesting problems for digraphs
arise when we consider orientations of graphs.

5.2.16. Definition. A tournament is an orientation of a complete graph. In the

directed sense, a path is a digraph whose vertices can be ordered as v1 , . . . , v¾
such that vi → vj if and only if j = i + 1. A cycle consists of a path ⟨v1 , . . . , v¾ ⟩
plus the edge v¾ v1 .

Thus, a path or cycle in a digraph is a “consistent ” orientation of an undi-

rected path or cycle. A tournament is a natural model for a competition in which
each team plays every other, with one winner in each game.

5.2.17. Definition. In a digraph, a king is a vertex from which every vertex is

reachable by a path of length at most 2. A successor of v is a vertex w such
that vw is an edge. A predecessor of v is a vertex u such that uv is an edge.

5.2.18. Proposition. (Landau [1953]) Every tournament has a king.

Proof: Let x be a vertex of maximum outdegree in a tournament T . If x is not
a king, then some vertex y is not reachable from x in at most two steps. Hence
no successor of x is a predecessor of y. Since T is an orientation of a complete
graph, every successor of x is thus a successor of y, as is x. Hence d+(y) > d+(x),
contradicting the choice of x.

An algorithmic proof starts with an arbitrary vertex and moves to a vertex

of larger degree if it is not a king. The extremal choice of a vertex with maximum
degree goes directly to the key point. Exercise 43 generalizes Proposition 5.2.18.
226 Chapter 5: First Concepts for Graphs

EXERCISES 5.2

5.2.1. (−) Prove or disprove:

(a) Deleting a vertex of maximum degree cannot increase the average degree.
(b) Deleting a vertex of minimum degree cannot reduce the average degree.
5.2.2. (−) Count the copies of C2¾ in K r,s .
5.2.3. (−) Let G be a graph with maximum degree 3. Prove that V(G) can be partitioned
into two sets that each induce a subgraph consisting of isolated edges and vertices.

5.2.4. (−) For ¾ ∈ , prove that every graph G has a ¾-partite subgraph H (Definition
5.1.7) with at least (1 − 1/¾) | E(G)| edges.
5.2.5. (−) Determine the minimum number of edges in an n-vertex graph having no inde-
pendent set of size 3.
5.2.6. (−) Prove that some n-vertex tournament satisfies d+(v) = d−(v) for every vertex v
if and only if n is odd.
5.2.7. (−) An oriented graph G is transitive if xy , y ∈ E(G) implies x ∈ E(G). Prove
that K n has exactly n! transitive orientations.
5.2.8. Count the cycles of lengths 6, 8, and 9 in the Petersen graph.
5.2.9. Use graphs and bijections (not algebra!) to prove the following:
(a) (n2) = (¾2) + (n − ) + (n−2 ¾ ) for 0 ≤  ≤ n.

(b) If ∑ ni = n, then ∑ (n2i ) ≤ (n2).

5.2.10. (♦) Let G be a graph with average vertex degree a.
(a) In terms of d(x), determine when G − x has average degree at least a. Conclude
that if a > 0, then G has a subgraph with minimum degree greater than a/2.
(b) For  > 0, show that a subgraph with minimum degree greater than (a/2) + 
cannot be guaranteed.
5.2.11. Let G be a graph with positive average degree a. Let t(v) denote the average of the
degrees of the neighbors of v. Prove that t(v) ≥ a for some v ∈ V(G). Construct infinitely
many graphs G such that t(v) > a for all v ∈ V(G). (Ajtai–Komlós–Szemerédi [1980])
5.2.12. Let F be a family of subsets of [n], with [n] ∈
/ F . Prove that the number of sets
S ⊆ [n] such that S contains an odd number of sets in F is even. (J. Edmonds)
5.2.13. (♦) A mountain range is the curve of a piecewise-linear function from (a , 0) to (b , 0)
in the upper half-plane. Travelers A and B start at (a , 0) and (b , 0), respectively. Give a
graph-theoretic proof (without induction) that they can meet by traveling on the mountain
range so that at all times their heights above the horizontal axis are the same.

5.2.14. (♦) For n ≥ 2, determine whether an n-vertex graph can have n distinct vertex
degrees. In an n-vertex graph with n − 1 distinct vertex degrees, what are the possibilities
for the repeated and missing degrees?
5.2.15. (+) Let G be a graph such that | N(x) ∩ N(y)| is odd for all x , y ∈ V(G). Prove that
| V(G)| is odd. (Engel [1998], Royle [2010], Silwal [2018])
Exercises for Section 5.2 227

5.2.16. (♦) Let d be a list of n nonnegative integers with even sum that differ by at most
1 and all are at most n − 1. Prove that d is graphic.
5.2.17. Given a nonincreasing list d of nonnegative integers, let d be obtained by deleting
d ¾ and subtracting 1 from the d ¾ largest elements remaining in the list. Prove that d is
graphic if and only if d is graphic. (Wang–Kleitman [1973])
5.2.18. Let a1 , . . . , a¾ be integers such that 0 < a1 < · · · < a¾ .
(a) Construct a graph with a¾ + 1 vertices whose set of distinct vertex degrees is
a1 , . . . , a¾ . (Kapoor–Polimeni–Wall [1977])
(b) Prove that when a1 = 1, the construction is unique if and only if there is a unique
graph on fewer than a¾ vertices with degree set {a2 − 1 , . . . , a¾−1 − 1}.
5.2.19. Let G and H be graphs with V(G) = V(H) and d G(v) = d H (v) for all v ∈ V(G).
Use induction on number of edges belonging to exactly one of G and H to prove that G
can be turned into H by using 2-switches. (Comment: This yields an alternative proof of
Theorem 5.2.8.)
5.2.20. Let p and q be nonnegative integer lists, with p = (p1 , . . . , pr) and q = (q1 , . . . , qs).
The pair (p , q) is bigraphic if some bipartite graph has p1 , . . . , pr as the vertex degrees in
one part and q1 , . . . , qs as the degrees in the other. When ∑ pi > 0, prove that (p , q) is bi-
graphic if and only if (p , q ) is bigraphic, where (p , q ) is obtained from (p , q) by deleting
the largest element ¾ from p and subtracting 1 from the ¾ largest elements of q.
5.2.21. Let A and B be r-by-s matrices with entries in {0 , 1}, having the same vector of
row sums and the same vector of column sums. Prove that A can be transformed into B
via steps in which the 0s and 1s are interchanged in a 2-by-2 permutation submatrix (that
is, submatrices (01
10
) and (10
01
) can be substituted for each other). (Ryser [1957])
5.2.22. An expansion of a graph replaces two edges with paths of length 2 through two
new vertices joined by an edge. Erasure is the inverse. Prove that a 2-switch can be per-
formed using expansions and erasures. Use this to show that every 3-regular graph arises
from K4 by expansions and erasures.

• • • •
expansion
• • • •
erasure
• • • •
5.2.23. Given an X , Y -bigraph G and edges x1 y1 and x2 y2 with x1 , x2 ∈ X and y1 , y2 ∈ Y ,
say that a bridging operation deletes x1 y1 and x2 y2 and adds two new vertices x3 and y3
with N(x3) = {y1 , y2 , y3 } and N(y3) = {x1 , x2 , x3 }. Prove that every 3-regular bipartite
graph can be obtained from a graph whose components are isomorphic to K3 ,3 by repeated
bridging operations. (Kotzig [1966a])
5.2.24. For m > 0, prove that every graph with m edges has a bipartite subgraph with
strictly more than m/2 edges. Construct a sequence of graphs (with m > 0) whose largest
bipartite subgraphs tend to half the number of edges.
5.2.25. Prove that the maximum number of edges in a bipartite subgraph of the Petersen
graph is 12. (Comment: Bondy–Locke [1986] proved that every triangle-free graph with
maximum degree 3 has a bipartite subgraph with at least 4/5 of its edges.)
5.2.26. Prove or disprove: Every graph has a vertex partition into two nonempty sets such
that each vertex has at least half of its neighbors in its own set.
5.2.27. (♦) For a graph G with no isolated vertices, prove that V(G) can be partitioned
into sets of size at least 2 whose induced subgraphs each have a spanning star. (Winkler)
228 Chapter 5: First Concepts for Graphs

5.2.28. A directed graph is unipathic if for all vertices x and y there is at most one (di-
rected) x , y-path. Let T n be the tournament with vertex set [n] such that i → j if and only
if i < j . What is the maximum number of edges in a unipathic subgraph of T n ? How many
unipathic subgraphs have that size? (Maurer–Rabinovitch–Trotter [1980])
5.2.29. (♦) Two inductive proofs of Mantel’s Theorem (the case r = 2 of Theorem 5.2.11).
(a) Given adjacent vertices x and y in an n-vertex graph, prove that xy lies in at least
d(x) + d(y) − n triangles. Use this to prove by induction on n that every n-vertex graph
with more than n2/4 edges contains a triangle.
(b) Prove that an n-vertex graph G with more than n2/4 edges has a vertex v such that
G − v has more than (n − 1)2/4 edges, again yielding Mantel’s Theorem by induction.
5.2.30. (♦) Strengthening Mantel’s Theorem (the case r = 2 of Theorem 5.2.11).
(a) For a vertex v in an n-vertex graph G, let º (v) be the maximum size of an inde-
pendent set contained in N(v). Prove ∑v∈V(G) º (v) ≤ n2/2, and determine which graphs
achieve equality. (Hint: Consider a largest independent set in G.)
(b) Use part (a) to obtain the case r = 2 of Theorem 5.2.11. (Galvin [1999])
5.2.31. (♦) Let G be an n-vertex graph having an orientation in which every triangle is ori-
ented cyclically. Prove that | E(G)| ≤ n2/3 and that this is sharp. (Brown–Harary [1970])
5.2.32. The game “bridge” has two teams of two partners each. Consider a club in which
four players cannot play if two of them have previously been partners that night. One
night 15 members arrive, but one wants to study graph theory. The other 14 play until
each has partnered with four others. After six more games (12 partnerships), they cannot
find four players with no pair of previous partners. Prove that adding the graph theorist
allows at least one more game to be played. (adapted from Bondy–Murty [1976, p. 111])
5.2.33. (♦) Let tr(n) be the number of edges in the n-vertex r-partite Turán graph, and let
G be an n-vertex graph not containing K r+1 . Prove that if tr(n) − | E(G)| ≤ s, then G can
be made into an r-partite graph by deleting at most s edges. (Hint: Consider the proof of
Turán’s Theorem (Theorem 5.2.11).) (Füredi [2015])

5.2.34. (♦) Let G be a graph such that ||V(H)

E(H)|
|
≤ ¾ for all H ⊆ G. By improving a bad orien-
tation, prove that G has an orientation in which every vertex has outdegree at most ¾ .
5.2.35. Determine the least n such that some two nonisomorphic n-vertex tournaments
have the same list of outdegrees.
5.2.36. (♦) Let T be an n-vertex tournament.
+ −
(a) Prove that ∑v∈V(T) ( d 2(v)) = ∑v∈V(T) (d 2(v)) by counting a set in two ways.
+
(b) Prove that T has (n3) − ∑v∈V(T) (d 2(v)) 3-cycles. (Kendall–Smith [1940])
(c) When n is odd, determine the maximum possible number of 3-cycles in T .
5.2.37. Prove that every tournament T with no source has at least three kings. For n ≥ 3,
construct an n-vertex tournament with no source and only three kings.
5.2.38. Prove that there is no 4-vertex tournament in which every vertex is a king. For
n ≥ ¾ ≥ 1, construct an n-vertex tournament with exactly ¾ kings, except possibly when
¾ = 2 and when n = ¾ = 4. (Exercise 5.2.37 excludes ¾ = 2.) (Maurer [1980])

5.2.39. (♦) Prove that p1 , . . . , pn ∈ 0 with p1 ≤ · · · ≤ pn are the outdegrees of some
¾
tournament if and only if ∑i=1 pi ≥ (¾2) for 1 ≤ ¾ < n and ∑i=1 pi = (n2). (Landau [1953])
n

5.2.40. (♦) Let G and H be tournaments on a vertex set V . Prove that d+G(v) = d+H (v) for
all v ∈ V if and only if G can be turned into H using direction-reversals on cycles of length
3. (Hint: Consider the graph of edges oriented differently in G and H .) (Ryser [1964])
Section 5.3: Connection and Decomposition 229

5.2.41. Given an ordering v1 , . . . , vn of the vertices of a tournament G, let S denote the

sum of j − i over edges vj vi such that j > i. Prove that every vertex ordering of G that
minimizes S puts the vertices in nonincreasing order of outdegree.
5.2.42. (♦) In an ordering L0 of the vertices of a tournament G, a reverse edge is an edge
yx ∈ E(G) such that y immediately follows x in L0 . Switching the order of such x and y
may create more reverse edges. Form L0 , L1 , · · · by iteratively switching vertices of re-
verse edges. Prove that this always reaches a list with no reverse edges. What is the
maximum number of steps to termination? (Comment: When the vertices are numbers,
and each edge points to the higher number, the result implies that iteratively switching
any two wrongly ordered adjacent numbers always sorts a list.) (Locke [1995])
5.2.43. (♦) Prove that every loopless digraph D has an independent set S such that every
vertex outside S is reached from S by a path of length at most 2. (Hint: Use induction on
| V(D)|. Comment: This generalizes Proposition 5.2.18.) (Chvátal–Lovász [1974])

5.3. Connection and Decomposition

We have defined paths and cycles in graphs and digraphs. Here we discuss
their use in connection and traversal. We also introduce multigraphs and the
notion of graph decomposition.

COMPONENTS AND WALKS

5.3.1. Definition. A u , v-path is a path with first and last vertices u and v, called
its endpoints; other vertices are internal vertices. A graph G is con-
nected if it contains a u , v-path for all u , v ∈ V(G). The components of G
are its maximal connected subgraphs.
The connection relation on the vertex set of a graph G is the set of
pairs u , v ∈ V(G) such that G has a u , v-path; we say “u is connected to v”
or “u and v are connected by a path”. A connected set of vertices is a set S
such that the induced subgraph G[S] is connected.

To state the stronger property that u and v are adjacent, we avoid ambigu-
ity by saying “u and v are joined by an edge”, not “u and v are connected”. For
digraphs, we use a stronger notion of connection.

5.3.2. Definition. In a digraph, a path ⟨v1 , . . . , v¾ ⟩ is a subdigraph such that

vi → vj if and only if j = i + 1. A digraph is strongly connected or strong
if it has a u , v-path for each ordered pair (u , v) of vertices. The strong com-
ponents of a digraph are its maximal strongly connected subgraphs.

For example, the digraph in Example 5.2.14 has three strong components,
with vertex sets {w}, {x , y}, and { }.
We need more general notions of movement. The definitions and results hold
for both graphs and digraphs, with edges in digraphs being ordered pairs.
230 Chapter 5: First Concepts for Graphs

5.3.3. Definition. A walk is a vertex list ⟨v0 , . . . , v¾ ⟩ such that each vi vi+1 is an
edge. The length of ⟨v0 , . . . , v¾ ⟩ is ¾ , the number of edges “traversed”. A
walk or cycle is odd or even when its length is odd or even. A walk with first
vertex u and last vertex v is a u , v-walk; these are its endpoints. A walk is
closed if its endpoints are the same.

5.3.4. Lemma. Given distinct vertices u and v in a graph or digraph, every

u , v-walk contains the vertices of a u , v-path as a sublist.
Proof: “Contains” refers to the vertex list, not the subgraph relation. The ver-
tices of the desired u , v-path must appear in order in the walk, but they need not
appear consecutively.
If no vertex repeats, then a u , v-walk W just traverses a u , v-path. If some ver-
tex x repeats, then deleting the portion of W after the first x up to and including
its next appearance leaves a shorter u , v-walk contained in W . Hence a shortest
u , v-walk contained in W has no repeated vertex and traverses a u , v-path.

u• • • • •v → u• • • • •v

5.3.5. Corollary. The union of a u , v-path and a v , w-path contains a u , w-path.

In particular, the connection relation is transitive.
Proof: Concatenating a u , v-path and a v , w-path yields a u , w-walk, and Lemma
5.3.4 applies.

5.3.6. Remark. Proving connectedness for a graph can be nontrivial. For exam-
ple, given numbers d1 , . . . , d n , let S be the set of graphs with vertices v1 , . . . , vn
in which the degree of vi is d i , for each i. Define a graph with vertex set S by
making two vertices adjacent if one can be obtained from the other by a 2-switch.
Theorem 5.2.8 proves that every such graph is connected. Exercise 20 is another
problem of this type. For undirected graphs, Corollary 5.3.5 implies that for
connectedness it suffices to show that there are paths from a single vertex to all
others. Connected graphs are those having one component.

5.3.7. Remark. By Corollary 5.3.5, the connection relation for a graph is an

equivalence relation on its vertex set. This motivates the term “component ” for a
maximal connected subgraph. The equivalence classes of the connection relation
are the vertex sets of the components; they are the maximal connected sets.
In a directed graph, the connection relation need not be symmetric (consider
a path, for example). We can define vertices u and v to be strongly connected if
both a u , v-path and a v , u-path exist. By Corollary 5.3.5 for digraphs, this is an
equivalence relation, and its equivalence classes are the vertex sets of the strong
components.

Lemma 5.3.4 has an analogue for closed walks. Note that a single vertex
forms a closed walk of length 0, but this is not a cycle.

5.3.8. Lemma. Every odd closed walk contains the vertices of an odd cycle (in
order) as a sublist.
Section 5.3: Connection and Decomposition 231

Proof: Given an odd closed walk W , let W be a minimal odd closed walk contained
in W (viewing W cyclically, without a fixed beginning vertex). If some vertex re-
peats in W , then it splits W into portions of odd and even positive lengths, each
of which is closed. The odd portion is a shorter odd closed walk contained in W .
Hence no vertex repeats, and the vertices of W in order traverse a cycle.

• • • • •
v
• • •
• • • • • •

Lemma 5.3.8 yields an important characterization of bipartite graphs.

5.3.9. Theorem. (K önig) A graph is bipartite if and only if it has no odd cycle.
Proof: A walk in a bipartite graph must alternate between the parts. Hence
every closed walk (including every cycle) has even length.
Conversely, when G has no odd cycle, we construct a bipartition of each com-
ponent H of G. Choose any v ∈ V(H). If for some vertex x there is both an even
v , x-walk and an odd v , x-walk, then they combine to form a closed odd walk, which
by Lemma 5.3.8 contains an odd cycle. Hence we can label each x ∈ V(H) as “odd”
or “even” by the parity of the v , x-walks in H. This is a bipartition of H , because
appending the edge xy to a v , x-walk shows that x and y have opposite parity.

Theorem 5.3.9 guarantees short proofs of whether a graph is bipartite (ex-

hibit a bipartition) or is not bipartite (exhibit an odd cycle).

CYCLES AND CUT-EDGES

Deleting parts of a graph can increase the number of components. Recall that
G − v denotes the subgraph of G obtained by deleting a vertex v. We denote edge
deletion similarly; we need to know whether we are deleting a vertex or an edge.

5.3.10. Definition. The subgraph of G obtained by deleting an edge e is denoted

G − e. A cut-vertex or cut-edge of G is a vertex or edge whose deletion
leaves a subgraph with more components than G.

Cut-edges (also called bridges) can be characterized using cycles.

5.3.11. Proposition. An edge is a cut-edge if and only if it belongs to no cycle.

Adding an edge to a graph combines the vertex sets of two components into
one or leaves the vertex sets of components unchanged.
Proof: Consider an edge uv in a graph G. If u and v are in the same component in
G − uv, then G and G − uv have the same number of components. Hence the follow-
ing are equivalent: (1) uv is a cut-edge of G, (2) u and v are in distinct components
of G − uv, (3) G − uv has no u , v-path, (4) G has no cycle containing uv.
Since deleting an edge can only increase the number of components by 1,
adding an edge can only reduce the number of components by 1.
232 Chapter 5: First Concepts for Graphs

5.3.12. Corollary. Every graph with n vertices and ¾ edges has at least n − ¾
components. Every connected n-vertex graph has at least n − 1 edges, which
is sharp for paths.
Proof: With n vertices and no edges, there are n components. Each added edge
reduces the number of components by at most 1.

Finite graphs have maximal (nonextendible) paths; choosing one is another

use of extremality. A nontrivial graph is a graph with at least one edge.

5.3.13. Proposition. Every nontrivial graph has at least two vertices that are
not cut-vertices.
Proof: Let u be an endpoint of a maximal (nonextendible) path P. Since P − u
is connected and N(u) ⊆ V(P), the neighbors of u lie in one component of G − u.
Hence u is not a cut-vertex.

• • •u

5.3.14. Lemma. A graph with minimum degree at least 2 has a cycle.

Proof: Again consider an endpoint u of a maximal path P. Since d(u) ≥ 2, there
is an edge incident to u other than its incident edge on P. By the maximality of
P the other endpoint of this edge is also in V(P), thereby completing a cycle.

We first apply this lemma to “decomposition” of graphs.

5.3.15. Definition. A decomposition of a graph G is a family of subgraphs that

partition E(G). An even graph is a graph where every vertex degree is even.

5.3.16. Theorem. Every even graph decomposes into cycles.

Proof: We use induction on the number of edges, m. A graph with no edges has a
decomposition into 0 cycles. If m > 0, then G has a nontrivial component H , and
(H) ≥ 2. By Lemma 5.3.14, H has a cycle C. Deleting the edges of C reduces the
degree of its vertices by 2, leaving an even graph G . By the induction hypothesis,
G decomposes into cycles, which with C complete a decomposition of G.

The reader deserves an apology about “even” and “odd”. An odd cycle is an
even graph, and a graph that is not an even graph is not called an odd graph.
Given a family , we can ask whether an input graph G decomposes into
graphs in . Decomposing a complete graph into copies of a smaller complete
graph is the central topic of Design Theory (Chapter 13). A complete graph de-
composes into two copies of a graph F if and only if F is self-complementary.
Hajós conjectured that every n-vertex even graph decomposes into at most
⌊ n/2⌋ cycles, and Gallai conjectured that every connected n-vertex graph decom-
poses into at most ⌈ n/2⌉ paths. These have been open for many years; see Exercise
48 for complete graphs. Lovász [1968a] proved that every n-vertex graph decom-
poses into at most ⌊ n/2⌋ subgraphs when both paths and cycles are allowed.
Section 5.3: Connection and Decomposition 233

The exercises pose many decomposition problems. For example, does the Pe-
tersen graph decompose into copies of P4 ? Does it decompose into copies of K 1 ,3 ?
The center of a star with at least two edges is its vertex of maximum degree.

5.3.17. Proposition. A ¾-regular graph G decomposes into copies of the star K 1 ,¾

if and only if G is bipartite.
Proof: Sufficiency. For an X , Y -bigraph G, form a decomposition using the stars
with centers at vertices of X . They are copies of K 1 ,¾ since G is ¾-regular, and
they form a decomposition because every edge has exactly one endpoint in X .
Necessity. The claim is trivial for ¾ = 1, so we may assume ¾ ≥ 2. Let X be
the set of centers of the stars in the decomposition. Since G is ¾-regular, each
star uses all the edges at its center; hence X is an independent set. Since the
stars decomposes G, no edges join vertices outside X .

EULERIAN CIRCUITS

5.3.18. Example. The Königsberg Bridge Problem. It is said that graph theory
began in K önigsberg. Located on the Pregel river at the island of Kneiphopf,
K önigsberg had seven bridges as shown on the left below. The citizens wondered
whether it was possible to take a stroll crossing each bridge exactly once.
x
•
X e1 e6
1 2 e2
6 e5
W 5 Y w• •y
7
3 4 e4
e3 e7
Z
•

We model this problem by shrinking the land masses to single points to ob-
tain the diagram of connections shown on the right above. We want to traverse
the edges once each. However, the diagram has multiple edges joining vertices.
This suggests a model more general than graphs.

5.3.19. Definition. A multigraph G is an ordered pair consisting of a vertex

set V(G) and an edge multiset E(G) consisting of subsets of V(G) with size 1
or 2. A loop is an edge consisting of one vertex. Multiedges are repeated
copies of edges with the same endpoints. The multiplicity of an edge is its
multiplicity as an element of E(G). The degree of a vertex is the sum of the
multiplicities of its incident edges, except that each loop contributes twice.

The convention that a loop contributes twice to the degree of its vertex ex-
tends the Degree-Sum Formula (and its consequences). A loop would arise in Ex-
ample 5.3.18 from a bridge whose endpoints lie in the same land mass. We treat
loops and paired multiedges in a multigraph as cycles of lengths 1 and 2.
234 Chapter 5: First Concepts for Graphs

5.3.20. Remark. Graphs vs. multigraphs. Many definitions and results for
graphs hold also for multigraphs (with the same proofs). Exceptions are state-
ments that need distinctness of the other endpoints of the edges incident to a
given vertex. The underlying graph of a loopless multigraph is the graph hav-
ing the same edges, but with multiplicity 1.
In the adjacency matrix of a multigraph, entry Ai , j counts the edges with
endpoints vi and vj . A loopless multigraph can be viewed as a complete graph
with a nonnegative integer weight function giving edge multiplicities. An iso-
morphism is a vertex bijection that preserves edge multiplicities.
When moving from vertex to vertex in a multigraph, there may be a choice
of edges. For technical precision, one should thus define a walk as an alternating
list v0 , e1 , v2 , . . . , e¾ , v¾ of vertices and edges. In the absence of multiedges, list-
ing only the vertices is precise. Extending elementary concepts to multigraphs is
straightforward, so we will not be formal about notation for walks.
Graphs have no loops or multiedges. We may emphasize this by calling a
graph a simple graph, particularly when also discussing multigraphs. When
an object under discussion is a multigraph, the term “subgraph” refers to any
multigraph contained in it; we avoid the awkward term “submultigraph”.
Of course, the digraph model also extends to allow multiedges.

In the K önigsberg problem, the people wanted a special walk in the multi-
graph of Example 5.3.18. Such walks are named in honor of Euler ’s paper about
the problem and its generalizations.

5.3.21. Definition. A trail is a walk that traverses each edge at most once (ver-
tices may repeat). A circuit is a closed trail, viewed cyclically regardless of
the starting vertex (it is an equivalence class of closed trails with the same
cyclic ordering of edges). An Eulerian trail in a multigraph G is a trail
traversing all edges of G. An Eulerian circuit is a closed Eulerian trail.

The difference between a trail and the subgraph consisting of its edges is that
the trail specifies an order on the edges.
The K önigsberg multigraph has no Eulerian trail. Each visit of a trail to a
vertex contributes twice to the degree by its entrance and exit. Only at the start
and end can the degree be odd. Thus a graph with an Eulerian trail has at most
two vertices of odd degree, and only even graphs can have Eulerian circuits. Eu-
ler [1736] observed that this condition is also sufficient for graphs with only one
nontrivial component, but the first published full proof was by Hierholzer [1873].

5.3.22. Theorem. A multigraph G has an Eulerian circuit if and only if it has

at most one nontrivial component and all vertex degrees are even.
Proof: Necessity. Each visit of a circuit to a vertex uses two incident edges.
Sufficiency. Since G is an even graph, it decomposes into cycles, by Theorem
5.3.16. Traversing any cycle is following a closed trail. Hence a decomposition of
G into cycles is a set of edge-disjoint circuits partitioning E(G).
Since such a set of circuits exists, we may take a smallest one, S. We claim
that S consists of a single Eulerian circuit. If S has two distinct circuits, and no
two circuits have a common vertex, then the circuits traverse distinct nontrivial
Exercises for Section 5.3 235

components of G. This contradicts the hypothesis, so there must be two circuits

sharing a vertex v. We can traverse one, starting and ending at v, and then tra-
verse the other, thereby forming a single circuit and partitioning E(G) into fewer
circuits than S does.
• • • •
v
• • • • •
• • • •

An Eulerian [multi] [di] graph is a [multi][di]graph having an Eulerian

circuit. For digraphs, the corresponding necessary degree condition is d+(v) =
d−(v) for all v ∈ V(G) (Exercise 58).
With the characterization of Eulerian multigraphs, it is easy to solve the
problem of optimal decomposition into trails, where “decomposition” means that
each edge is traversed by exactly one of the trails. The transformation by adding
edges to pair vertices of odd degree is a common technique that takes advantage
of the generality of multigraphs.

5.3.23. Corollary. If G is a connected multigraph with 2 ¾ vertices of odd degree,

where ¾ > 0, then G decomposes into ¾ trails and no fewer.
Proof: A trail contributes odd degree only to its endpoints, so a partition into
trails has a trail ending at each odd-degree vertex. To prove that ¾ trails suffice,
add ¾ edges joining pairs of odd-degree vertices. The resulting multigraph G is
Eulerian. Deleting the ¾ added edges from an Eulerian circuit of G cuts it into
¾ edge-disjoint trails decomposing G.

EXERCISES 5.3

5.3.1. (−) Prove that the complement of any disconnected graph is connected.
5.3.2. (−) Let ¾ , l , n be nonnegative integers with ¾ + l = n. Find necessary and sufficient
conditions on (¾ , l , n) such that there exists a connected n-vertex graph with ¾ vertices of
even degree and l vertices of odd degree.
5.3.3. (−) Prove or disprove the following statements about graphs.
(a) Every union of two distinct u , v-walks contains a cycle.
(b) Every union of two distinct u , v-paths contains a cycle.
(c) A circuit contains a cycle through each of its vertices.
(d) If (G) ≥ 2, then every vertex of G belongs to some cycle.
5.3.4. (−) Prove that every finite digraph with no vertex of outdegree 0 contains a cycle.
5.3.5. (−) Prove that the number of vi , vj -walks of length  in a graph is entry (i , j) in A¾ ,
where A¾ is the  th power (multiplication) of the adjacency matrix.
5.3.6. (−) Prove that a multigraph G is bipartite if and only if every subgraph H of G has
an independent set consisting of at least half of V(H).
5.3.7. (−) The Möbius ladder M¾ is the graph formed by adding to a 2 -cycle a set of
 edges joining opposite vertices on the cycle. Determine all  such that M¾ is bipartite.
(Comment: Why is this graph called a “Möbius ladder ”?)
236 Chapter 5: First Concepts for Graphs

5.3.8. (−) Let D1 , . . . , D¾ be the strong components of a digraph D. Let D∗ be the loopless
digraph with vertices v1 , . . . , v¾ such that vi → vj if and only if i
= j and D has an edge
from Di to Dj . Prove that D∗ has no cycle. Conclude that in every digraph, some strong
component has no entering edges, and some strong component has no exiting edges.
• • •

D • • • D∗

• • •

5.3.9. (−) Determine all ¾ such that the hypercube Q¾ decomposes into 4-cycles.
5.3.10. (−) Prove that in a connected graph that is not a complete graph, every vertex
belongs to an induced copy of P3 . Does the same conclusion hold for every edge?
5.3.11. (−) Decompose Q3 into copies of K1 ,3 and into copies of P4 . Does the Petersen graph
decompose into copies of P4 ? Does it decompose into copies of P6 ?
5.3.12. (−) Prove or disprove: A 3-regular graph with more than six vertices cannot de-
compose into three paths.
5.3.13. (−) Characterize the pairs of positive integers r, s such that K r,s decomposes into
two isomorphic subgraphs.
5.3.14. (−) Let G be a disconnected graph with n vertices. Find the maximum possible
value of (G). When (G) =  , find the maximum possible value of (G).
5.3.15. (−) Let G and H be two disconnected graphs with the same vertex set. Prove that
some two vertices are in different components in both G and H .
5.3.16. (−) Prove or disprove the following statements.
(a) No connected Eulerian graph has odd size and even order.
(b) If e and  are incident edges in an Eulerian graph G, then e and  appear consec-
utively in some Eulerian circuit of G.
5.3.17. (−) Two Eulerian circuits are equivalent if they have the same pairs of consecutive
edges, viewed cyclically. For example, a cycle has only one equivalence class of Eulerian
circuits. How many equivalence classes of Eulerian circuits are there in K 2 ,r ?
5.3.18. (−) For  , r ∈  , define G on the set of integer -tuples by uv ∈ E(G) if and only if
∑ |ui − vi | = r. Prove that G is disconnected when r is even and bipartite when r is odd.
5.3.19. Let G be the graph defined on all binary -tuples by making x and y adjacent if
they differ in exactly two positions. Determine the number of components of G.
5.3.20. Let G n be the graph whose vertices are the permutations of [n] in word form,
with permutations a1 , . . . , an and b1 , . . . , bn adjacent if they differ by interchanging two
consecutive entries. Prove that G n is connected.
5.3.21. Let G be a connected graph with no induced subgraphs that are paths or cycles of
length 3. Prove that G is a complete bipartite graph.
5.3.22. For n ≥ 6, prove that the minimum number of edges that must be deleted from
K n to obtain a graph having a decomposition into two disconnected subgraphs without iso-
lated vertices is 4. (Hint: Use induction. Comment: This result strengthens the easy fact
that the complement of a disconnected graph is connected.) (Caro–Roditty [2004])
5.3.23. Given vertices u and v in a connected graph G, prove that there is a unique parti-
tion of V(G) into sets A and B inducing connected subgraphs such that u ∈ A, v ∈ B, and
every edge joining A and B is incident to v. (Liu [1985])
Exercises for Section 5.3 237

5.3.24. (♦) In a connected graph G, let P and Q be two paths chosen to maximize the sum
of their lengths. Prove that P and Q have a common vertex.

5.3.25. (♦) Prove that a connected graph G with at least three vertices has two vertices x
and y such that (1) G − {x , y} is connected and (2) x and y are adjacent or have a common
neighbor. (Hint: Consider a longest path.)

5.3.26. (♦) For integer weights on E(K n), not all even, prove the following equivalent.
(A) Every cycle has even total weight.
(B) Every triangle has even total weight.
(C) The edges with odd weight form a spanning complete bipartite subgraph.

5.3.27. A signed graph assigns +1 or −1 to each edge. A signed graph is balanced if

every cycle has an even number of negative edges. Prove that a signed graph G is balanced
if and only if V(G) can be expressed as the disjoint union of sets A and B such that an edge
is negative if and only if it joins A and B. (Harary [1953], Cartwright–Harary [1956])
(Comment: Zaslavsky [1998] and Kaiser–Lukot ’ka–Rollová [2017] survey signed graphs.)

5.3.28. (♦) Let D be a strongly connected orientation of a graph G. Prove that if G has
an odd cycle, then D has an odd cycle.

5.3.29. (+) For a strong digraph D, let º (D) be the length of the shortest closed walk
visiting every vertex. Prove that the maximum of º (D) over all n-vertex strong digraphs
is ⌊ (n + 1)2/4⌋ if n ≥ 2. (Vizing–Goldberg [1969], Cull [1980])

5.3.30. Let G be an oriented graph with | V(G)| ≥ 3. Prove that edges can be added to
extend G to a strongly connected oriented graph if and only if V(G) contains no nonempty
proper subset S such that xy ∈ E(G) for all x ∈ S and y ∈ S. (Boesch–Tindell [1980],
Farzad–Mahdian–Mahmoodian–Saberi–Sadri [2006])

5.3.31. (+) For a graph G, let c(G) denote the least ¾ such that every edge lies in a cycle
of length at most ¾ (infinite when G has a cut-edge). Prove√ for n ≥ 3 that the minimum of
| E(G)| + c(G), over n-vertex connected graphs, is n + ⌈ 2 n − 1 ⌉ . (Butler–Mao [2007])
5.3.32. Let W be a nontrivial closed walk that does not contain a cycle. Prove that some
edge of W occurs twice in succession (once in each direction).

5.3.33. Let C be a closed walk in a graph G, and let H be the subgraph of G whose edges
are those used an odd number of times in C. Prove that H is an even graph.

5.3.34. (♦) A parity subgraph of a graph G is a spanning subgraph H such that d G(v) ≡
d H (v) (mod 2) for all v ∈ V(G). Prove that a parity subgraph contains every cut-edge of G.

5.3.35. (♦) Cut-edges.

(a) Prove that a ¾-regular bipartite graph has no cut-edge when ¾ ≥ 2.
(b) Prove that every even graph has no cut-edge. For each odd ¾ with ¾ ≥ 3, construct
a ¾-regular connected graph having a cut-edge.

5.3.36. (+) Determine the maximum number of cut-edges in a connected 3-regular graph
with n vertices. (O–West [2010] determined this for (2r + 1)-regular n-vertex graphs.)

5.3.37. (♦) Let d1 , . . . , d n be the vertex degrees of a graph G, with d1 ≤ · · · ≤ d n . Prove

that G is connected if d j ≥ j whenever j ≤ n − 1 − (G).

5.3.38. (♦) Let G be a graph with minimum degree ¾ and girth ½ .

(a) For ¾ ≥ 2, prove that G has a cycle of length at least ¾ + 1 and that this is sharp.
(b) For ¾ ≥ 3, prove that G has an even cycle.
(c) For ¾ ≥ 2, prove that G has a cycle of length at least (½ − 2)(¾ − 1) + 2.
238 Chapter 5: First Concepts for Graphs

5.3.39. (♦) Non-bipartite graphs have short odd cycles. Let G be a non-bipartite triangle-
free n-vertex graph, and let ¾ = (G). Let C be a shortest odd cycle in G, with length l.
(a) Prove that every vertex not in V(C) has at most two neighbors in V(C).
(b) Prove that l ≤ 2n/ . (Campbell–Staton [1991])
(c) Prove that the inequality of part (b) is best possible when  is even.
5.3.40. (♦) Prove that some loopless multigraph has degree list d1 , . . . , d n if and only if
∑ di is even and no value is more than half the total. (Hakimi [1962])
5.3.41. Prove that there is a connected multigraph whose vertex degrees are the positive
integers d1 , . . . , d n if and only if ∑ d i is even and at least 2n − 2.
5.3.42. Prove that the degree list of any loopless multigraph has a realization whose un-
derlying graph has no cycle or has one cycle, with length 3. (Will–Hulett [2004])
5.3.43. (+) Let G be a graph with 2r + 1 vertices in which any r vertices have a common
neighbor. Prove that some vertex is adjacent to all others. (Burungale [2006])
5.3.44. (+) Suppose that n wires run from the top to the bottom of a tall building through
a tube. The correspondence between the tops and the bottoms is unknown. The inspector
has a battery and a lamp. When a circuit is completed using battery, lamp, and wires, the
lamp lights. Prove that the inspector can match up the ends using only one trip up and
down the building. The inspector can link some pairs of wires at the bottom, go up and
make similar links and tests, and finally return to the bottom to complete the analysis.
5.3.45. (♦) Prove that K n decomposes into three isomorphic subgraphs if and only if n + 1
is not divisible by 3.
5.3.46. Decompose the Petersen graph into three isomorphic connected subgraphs.
5.3.47. Let G be a -regular graph, where  is odd. Prove that in any decomposition of G
into paths, the average length of the paths is at most  .
5.3.48. (♦) Prove that K n decomposes into ⌈ n/2⌉ paths. Prove that K n decomposes into
⌊ n/2⌋ cycles when n is odd. (Walecki, in Lucas [1892])
5.3.49. (♦) Prove that every connected graph has an orientation with at most one vertex of
odd outdegree. Conclude that a connected graph with an even number of edges decomposes
into paths with two edges. (Rotman [1991])
5.3.50. Use induction on  to prove that a connected graph with 2  edges decomposes into
paths with two edges. Is this true without the hypothesis of connectedness?
5.3.51. (♦) Nash-Williams [1970] conjectured that when n is sufficiently large, an n-vertex
even graph G with | E(G)| divisible by 3 decomposes into triangles if (G) ≥ 43 n. Use the
graph Ht to prove that the degree threshold is sharp, where Ht is formed from C4 by ex-
panding each vertex into K t and each edge into K t ,t .

5.3.52. Let G be an n-vertex graph with (G) ≥ n/4, no cut-edge, and | E(G)| divisible by
3. Barát–Thomassen [2006] proved that G decomposes into copies of K1 ,3 . Prove that this
is sharp using a graph formed from 4K n/4 by one edge for each pair of components.
Section 5.4: Trees and Distance 239

5.3.53. Prove that the edges of every n-vertex graph can be covered using at most ⌊ n2/4⌋
edges and triangles, with equality needed only for K⌊n/2⌋ ,⌈n/2⌉ . (Erdős–Goodman–P ósa
[1966]) (Comment: Bollobás [1976a] generalized to coverings by edges and copies of K r .)
5.3.54. (+) The broom Bn ,r is the n-vertex graph formed from the union of the path Pn−r
and the star K1 ,r by letting the center of the star be an endpoint of the path. Prove that if
K n decomposes into copies of Bn ,r , then n is even and n ≥ 4r − 2. (Kovář–Kubesa–Meszka
[2012]) (Comment: the condition is also sufficient , except for (n , r) = (6 , 2).)
¾
5.3.55. (+) For a graph G, let (G) = max{r: K r ⊆ G}. Prove ∑i=1 (G i) ≤ n + (¾2) for any
decomposition (G1 , . . . , G ¾) of K n . Provide a construction to show that the bound is best
possible when n ≥ (¾2) . (Füredi–Kostochka–Škrekovski–Stiebitz–West [2005])
5.3.56. (♦) Prove or disprove each statement below.
(a) Every graph G has an orientation D such that for every v ∈ V(G), the numbers of
edges entering v and leaving v in D differ by at most 1.
(b) Every graph G has an orientation D such that for every S ⊆ V(G), the numbers
of edges entering S and leaving S in D differ by at most 1.
5.3.57. (♦) Let G be a connected graph having | E(G)| even and an even number of vertices
of each odd degree. Prove that G decomposes into two spanning subgraphs with the same
degree list. Show that the last condition is needed. (Choi–Ozkahya–West [2010])
5.3.58. Eulerian digraphs. Let D be a digraph with no isolated vertex.
(a) Prove that D has an Eulerian circuit if and only if d+(v) = d−(v) for each vertex v
and the underlying graph is connected.
(b) Suppose that d−(v) = d+(v) for every vertex v, except that d+(x) − d−(x) =  =
d−(y) − d+(y). Prove that D contains  edge-disjoint x , y-paths.

5.3.59. (♦) de Bruijn cycles. For  , l ∈ , form a digraph G with V(G) = [] l (called the
de Bruijn graph) as follows. For u , v ∈ [] l , create an edge uv if the last l − 1 positions
in u form the same (l − 1)-tuple as the first l − 1 positions in v.
(a) Prove that the de Bruijn graph is an Eulerian digraph. (de Bruijn [1946])
(b) Prove that there is a cyclic arrangement of  l+1 characters chosen from [] such
that the  l+1 strings of length l + 1 are distinct. (Good [1946], Rees [1946])
5.3.60. (+) Alternative characterization of Eulerian graphs.
(a) For uv ∈ E(G) when G is Eulerian, prove that an odd number of u , v-trails in G − uv
reach v only at the end, and that an even number of these are not paths. (Toida [1973])
(b) Let v be a vertex of odd degree in a graph G. Prove that for some edge incident to
v, the number of cycles through it is even.
(c) Use parts (a) and (b) to show that a nontrivial connected graph is Eulerian if and
only if every edge belongs to an odd number of cycles. (McKee [1984])

5.4. Trees and Distance

When a graph models a communication network, its spanning trees are the
minimal subgraphs that permit communication among all vertices. We restate
the definitions from the Introduction.

5.4.1. Definition. An acyclic graph is a graph with no cycles. A tree is a con-

nected acyclic graph. A graph whose components are all trees is a forest. A
spanning tree in a graph is a spanning subgraph that is a tree.
240 Chapter 5: First Concepts for Graphs

We studied enumeration of trees in Part I, counting those with a fixed vertex

set in Chapter 1, isomorphism classes in Chapter 4, and spanning trees of a given
graph (recursively) in Chapter 2 (see also Theorem 15.2.5). Here we consider
graph-theoretic properties and applications of trees. We also study distances in
graphs and their relation to trees.

PROPERTIES OF TREES

We expand on properties proved earlier. A leaf is a vertex of degree 1.

5.4.2. Proposition. Every nontrivial tree has at least two leaves. Deleting a
leaf from a tree yields a tree.
Proof: Nontrivial trees have no isolated vertices. Since trees are acyclic, the
endpoints of every maximal path are leaves.
Deleting a vertex cannot create a cycle. A leaf v in a graph G cannot be a cut-
vertex of G, since v can have a neighbor only in one component of G − v. Thus if
v is a leaf of a tree G, then G − v is acyclic and connected and hence is a tree.

Proposition 5.4.2 implies that each tree with n + 1 vertices arises from a tree
with n vertices by adding an edge to a new vertex. This justifies inductive proofs
about trees that perform the induction step by “growing a new leaf ” from an ar-
bitrary vertex of an arbitrary smaller tree.
The fact that the number of edges in a tree is one less than the number of
vertices is one of the most fundamental properties of a tree.

5.4.3. Proposition. For a graph G with n vertices, any two of the following three
properties implies the third (and hence characterize trees).
(a) G is acyclic.
(b) G is connected (that is, G has one component).
(c) G has n − 1 edges.
Proof: Adding an edge to a graph reduces the number of components (by 1) if
and only if it joins vertices from distinct components and hence creates no cycle.
Starting with no edges, we thus conclude that a graph with ¾ edges has exactly
n − ¾ components if and only if it has no cycles.
Therefore, an acyclic graph has one component if and only if it has n− 1 edges,
and a graph with n − 1 edges and one component must be acyclic.

Various other properties also characterize the graphs that are trees. It is
efficient to prove a cycle of implications.

5.4.4. Proposition. For each property below, a graph G is a tree if and only if it
satisfies that property.
(a) G is connected and every edge is a cut-edge.
(b) G contains exactly one u , v-path whenever u , v ∈ V(G).
(c) G has no cycle, and adding any edge creates exactly one cycle.
Section 5.4: Trees and Distance 241

Proof: (G is a tree) ⇒ (a): By Proposition 5.3.11, an edge is a cut-edge if and only

if it is in no cycle. Since a tree is acyclic, every edge is a cut-edge.
(a) ⇒ (b): Since G is connected, it contains a u , v-path. If G has more than
one u , v-path, then let xy be an edge belonging to one but not all of them. By
Corollary 5.3.5, G − xy contains an x , y-path, contradicting xy being a cut-edge.
(b) ⇒ (c): Two vertices on a cycle are connected by two paths along the cycle,
so (b) implies that G has no cycle. For u , v ∈ V(G), adding the edge uv creates
exactly one cycle since G contains exactly one u , v-path.
(c) ⇒ (G is a tree): Adding a missing edge uv creates a cycle, so G contains a
u , v-path. Hence G is connected. With lack of cycles also given, G is a tree.

5.4.5. Proposition. Every connected graph contains a spanning tree.

Proof: Since deleting an edge of a cycle does not disconnect a graph, iteratively
deleting edges from cycles in a connected graph leads to a connected acyclic span-
ning subgraph of the original graph.

The properties of trees allow us to switch edges in spanning trees, moving

from one spanning tree to another (see also Exercises 64–65).

5.4.6. Theorem. If T, T are spanning trees of a graph G and e ∈ E(T) − E(T ),

then there exists e ∈ E(T ) − E(T) such that both T − e + e and T + e − e
are spanning trees.
Proof: Since e is a cut-edge of T , we may let U and U be the vertex sets of the
two components of T − e (in the 8-vertex graph below, T is bold and T is solid).
Let e = uu , with u ∈ U and u ∈ U . Because T is a spanning tree, T contains
a unique u , u -path. This path has an edge from U to U ; let e be such an edge.
The only edge of T joining U and U is e, so e ∈ E(T ) − E(T). Since e joins
the components of T − e, the subgraph T − e + e is a spanning tree. Since e is
in the u , u -path in T , it is in the unique cycle formed by adding e to T . Hence
T + e − e also is a spanning tree.

e
• •
U • • e • • U
• •

5.4.7.* Theorem. (Berman [1986]) Let T be a spanning tree in a graph G with

at least three vertices. If d G(v) − d T (v) is odd for every vertex that is not a
leaf of T , then G has an even number of spanning trees T (including T) such
that d T (v) = d T (v) for all v ∈ V(G).
Proof: (Cameron–Edmonds [1999]) We define an auxiliary graph H such that the
vertices of odd degree in H correspond to the desired trees. Corollary 5.2.2 then
implies that the number of them is even.
Specify one non-leaf vertex w in T . The vertices of H are the spanning trees
in G whose vertex degrees agree with T (Type 1) or agree with T except for having
degree d T (w) + 1 at w and degree d T (u) − 1 at some other vertex u (Type 2).
242 Chapter 5: First Concepts for Graphs

w w w w
• • • •
u
• • • •
• • • • • • • •
• • • •
u
T
• Type 1
• • Type 2
•
Spanning trees S and S that are vertices of H are adjacent in H if and only
if each tree has exactly one edge not belonging to the other. Thus a neighbor S
of S is obtained from S by adding one edge of S and deleting one edge of S.
One endpoint of the added edge increases in degree; call it v. If S is Type 1,
then v = w. If S is Type 2, then v must be the vertex u such that d S(u) = d T (u)− 1.
Since u is not isolated in S, in both cases v is not a leaf in T . Hence the number of
choices for the added edge is odd when S is Type 1 and is even when S is Type 2.
Let vx be the added edge, where x ∈ NG(v)− NS(v). Adding vx creates a unique
cycle with edges of S. The degree of x must not increase, because already degree
is increasing at v. Hence the edge deleted to form S must be the edge reaching
x along the v , x-path in S. The resulting tree S has vertex degrees putting it in
V(H); it is Type 1 if y = w and Type 2 otherwise.
We have proved d H (S) = d G(v) − d S(v) and shown that the degree is odd when
S is Type 1 and even otherwise. Hence the number of Type 1 trees is even.
v=w • • • y v=u • • • y (= w?)
• S • • S •
S ← S S ← S
• •
Type 1 x x Type 2

A famous result of Smith (see Tutte [1946]) states that in a 3-regular graph
the number of spanning cycles containing a particular edge is even. A stronger
version of this follows easily from Theorem 5.4.7.

5.4.8.* Corollary. (Thomason [1978]) If every vertex of G except possibly the

endpoints of a particular edge e has odd degree, then the number of spanning
cycles containing e is even.
Proof: (Cameron–Edmonds [1999]) Deleting e from a spanning cycle of G leaves
a spanning tree T in which every vertex other than the endpoints of e has degree
2. Hence those vertices have odd degree outside T , and Theorem 5.4.7 applies.
Each of the resulting trees becomes a spanning cycle by adding e.

DISTANCE AND DIAMETER

In a connected graph, each vertex can be reached from every other; how many
steps are needed? The concept is valid for graphs or digraphs.

5.4.9. Definition. If x is connected to y in G, then the length of the shortest

x , y-path is the distance from x to y, written d G(x , y) or simply d(x , y). If
G has no x , y-path, then d(x , y) = ∞.
Section 5.4: Trees and Distance 243

The distance function of a graph is a metric on the vertex set: it is nonnega-

tive, symmetric, positive for distinct vertices, and satisfies the triangle inequal-
ity d(x , y) + d(y , ) ≥ d(x , ) (Exercise 38). The distance function for a digraph is
not symmetric unless the digraph is symmetric.
One measure of the performance of a communication network is the worst
time it might take to reach one point from another.

5.4.10. Definition. The diameter of a graph G is max u ,v∈V(G) d(u , v). The ec-
centricity (u) of a vertex u is max v∈V(G) d(u , v). The radius of G equals
minu∈V(G) (u). The center of G is the set of vertices with least eccentricity.

The diameter is the largest vertex eccentricity. The term “diameter ” comes
from geometry, where diameter is the greatest distance between two vertices in
a set. When G is not connected, the diameter, the radius, and each eccentricity
are infinite. The graph below has three vertices in its center; we have labeled
each vertex with its eccentricity.

5• •5 4• •5
5•
3
• • • •4
4 3
5• 3•

The uniqueness of paths in trees makes it easy to compute distances. Since

every path is the shortest (and only!) path between its endpoints, the maximum
of the vertex distances is the length of the longest path.
The next observation, about centers of trees, was used in Section 4.2.

5.4.11. Theorem. (Jordan [1869]) The center of a tree consists of one vertex or
two adjacent vertices.
Proof: Let d be the diameter of a tree T . Let S be the central vertex or edge on a
path P of length d in T . The vertices of S have eccentricity ⌈ d/2⌉ , since a vertex
at distance greater than ⌈ d/2⌉ from v ∈ S would combine with half of P to form
a path longer than P.
All vertices outside S have eccentricity larger than ⌈ d/2⌉ , by the uniqueness
of paths in trees. From a vertex u, follow a path to P , and then add at least ⌈ d/2⌉
edges to reach the farther end of P , with equality in the last step only if the first
vertex reached on P is in S.

Theorem 5.4.11 also has an easy inductive proof (Exercise 44) and a proof us-
ing the Pigeonhole Principle (Exercise 10.1.24).
Beyond trees, small diameter remains desirable for efficiency of communi-
cation. A complete graph has diameter 1, while the star K 1 ,n−1 has diameter 2
with only n − 1 edges, but physical constraints may limit the maximum degree.
The problem of finding the smallest diameter among n-vertex graphs with maxi-
mum degree  can be solved by instead determining for all  and d the maximum
number of vertices in a graph with diameter d and maximum degree at most  .
244 Chapter 5: First Concepts for Graphs

5.4.12. Proposition. (Moore bound) Graphs with maximum degree ¾ and di-
ameter d have at most 1 + ¾ (¾−¾1) −1
d

−2 vertices.
Proof: Let G be a graph with maximum degree ¾ and diameter d. From a vertex
v, at most ¾(¾ − 1)i−1 vertices have distance i in G, and all vertices are within
distance d. Thus G has at most ¾ ∑ i=1 (¾ − 1)i−1 vertices other than v. Evaluating
d

the geometric sum yields the desired bound.

• •
• •
• • • • •
•
• • • • • • • •
• • • • • • • • • • • • • •

When d(v) < ¾ for some vertex v, starting the search at v multiplies the bound
by the fraction d(v)/¾ . Thus graphs within a fraction 1/¾ of the bound must be
¾-regular. Equality requires adding edges among the leaves of the search tree to
establish distance at most d for all vertex pairs. As shown above, the Petersen
graph accomplishes this when d = 2 and ¾ = 3 (further motivating study of the
Petersen graph!).
Equality in the Moore bound requires the full tree grown from a vertex and
hence girth at least 2d + 1. This is best possible; every graph of diameter d that
is not a tree has girth at most 2d + 1 (Exercise 9).

5.4.13. Example. Moore graphs. A Moore graph is a ¾-regular graph (for ¾ ≥ 3)

that has diameter d, girth 2d + 1, and order 1 + ¾ (¾−¾1) −1
d

−2 (any two of these three

restrictions imply the third; see Exercise 59).
Damerell [1973] and Bannai–Ito [1973] independently proved that no Moore
graph exists with diameter at least 3. For diameter 2, Hoffman–Singleton [1960]
proved that Moore graphs exist only for ¾ = 3, ¾ = 7, and possibly ¾ = 57 (The-
orem 15.3.26). For ¾ = 3, we have the Petersen graph. For ¾ = 7, the 50-vertex
Hoffman–Singleton graph contains many copies of the Petersen graph (Exercise
59). It remains unknown whether a 57-regular Moore graph exists.

For diameter 2, the least number of edges in an n-vertex graph with maxi-
mum degree ¾ is known √ exactly when ¾ ≥ (n + 1)/2 (Erdős–Rényi–Sós [1966]) and
asymptotically
√ when n − 1 ≤ ¾ ≤ n/2 (Bollobás [1971]).
Why n − 1? The Moore bound requires √ n ≤ 1 + ¾ 2 for diameter 2. Graphs
with diameter 2 and maximum degree near n don’t look much like trees. We
can get close with a cartesian product.
√
5.4.14. Example. Diameter 2 with degree as small as 2 n. Given n = m2 , let
G m = K m K m . Thus V(G m) = {(i , j): 1 ≤ i , j ≤ m}, with (i , j) and (¾ , l) adjacent
if i = ¾ or j = l. Each vertex has degree 2m − 2, and non-adjacent vertices (i , j)
and (¾ , l) have common neighbors (i , l) and (¾ , j). √
√
However, G m is regular of degree 2 n − 2, not n − 1. Reducing the degree
√
to approximately n − 1 requires finite projective planes, which can be obtained
Section 5.4: Trees and Distance 245

from finite fields; see Chapter 13. For values of n having the form q2 + q + 1 (where
q is a power of a prime), the result is a√nearly-regular graph of diameter 2 with
maximum degree q + 1, which equals ⌈ n − 1⌉ (some vertices have degree q).

• • • •

• • • •

• • • •

• • • •

OPTIMIZATION ON WEIGHTED GRAPHS

The notion of length suggests a more general graph model: a weighted

graph has real-valued edge weights. In a weighted graph, the length of a walk is
the sum of the weights of its edges; the unweighted case is when each edge weight
is 1. Classical problems in this setting are computing distances and finding a
spanning tree of minimum total weight.
Algorithms to find minimum-weight spanning trees in connected weighted
graphs rest on Theorem 5.4.6 or similar observations. Kruskal’s Algorithm is a
well-known example; it is a “greedy” algorithm. In Chapter 11 we study a more
general setting where this algorithm works.

5.4.15. Theorem. (Kruskal’s Algorithm; Kruskal [1956]) Given a connected

n-vertex graph with edge weights, a minimum weight spanning tree is con-
structed by n − 1 repetitions of selecting a cheapest edge that completes no
cycle with edges already selected.
Proof: Always there are edges that join components in the spanning subgraph of
chosen edges, so the algorithm produces a spanning tree.
Let T be a tree found by the algorithm, and let T ∗ be a minimum-weight
spanning tree sharing the most edges with T . If T ∗
= T , then because they
have the same size there exists e ∈ E(T) − E(T ∗). By Theorem 5.4.6, there exists
e ∈ E(T ∗) − E(T) such that both T − e + e and T ∗ + e − e are spanning trees.
Since T ∗ + e − e shares more edges with T than does T ∗ , the choice of T ∗
implies that e has larger weight than e . On the other hand, since adding e to
T − e creates no cycles, both e and e were eligible when the algorithm choose e,
and thus e cannot have larger weight than e . The contradiction yields T ∗ = T .

5.4.16. Example. Choices in Kruskal’s Algorithm depend only on the order of

weights, not the values. Below we have used positive integers as weights to indi-
cate the order of examination of edges. The four cheapest edges are selected, but
then we cannot take the edges of weight 5 or 6. We can take the edge of weight 7,
but then not those of weight 8 or 9.
246 Chapter 5: First Concepts for Graphs

• 9 •
1 8 2 4
• 7 • 6 •
11 10 5 3
• •
12

Other early algorithms were Bor u̇vka [1926], Jarnı́k [1930], and Prim [1957]
(Exercise 67). Modern improvements use clever data structures to merge compo-
nents quickly. Fast versions appear in Tarjan [1984] when the edges are pre-
sorted by cost and in Gabow–Galil–Spencer–Tarjan [1986] when they are not.
For further discussion and references, see Ahuja–Magnanti–Orlin [1993, Chap-
ter 13]. More recent developments appear in K arger–K lein–Tarjan [1995].

Next we consider finding shortest paths. Dijkstra’s Algorithm (Dijkstra

[1959] and Whiting–Hillier [1960]) solves this problem quickly, using the obser-
vation that the u , v-portion of a shortest u , -path is a shortest u , v-path. It finds
shortest paths from u to all other vertices in increasing order of d(u , ), creating
a spanning tree from u. The distance d(u , ) in a weighted graph is the minimum
sum of the edge weights along a u , -path (we assume nonnegative weights).

5.4.17. Algorithm. (Dijkstra’s Algorithm for distances from one vertex.)

Given a graph (or digraph) G with nonnegative edge weights and a starting ver-
tex u, with weight w(xy) on edge xy (and w(xy) = ∞ if xy ∈ / E(G)), we maintain a
set S of vertices to which a shortest path from u is known and a tentative distance
t( ) from u to each ∈ / S, being the length of the shortest u , -path yet found.
The initial state is S = {u}, t(u) = 0, and t( ) = w(u ) for
= u. At each it-
eration, select v outside S such that t(v) = min ∈/S t( ). Add v to S. Explore edges
from v to the rest of V(G) − S to update tentative distances: for each edge v with
∈
/ S, set t( ) to min{t( ) , t(v) + w(v )}.
The iteration continues until S = V(G) or until t( ) = ∞ for every ∈ / S. At
the end, set d(u , v) = t(v) for all v.

5.4.18. Example. In the example below, shortest paths from u are found to the
other vertices in the order a , b , c , d , e, with distances 1 , 3 , 5 , 6 , 8, respectively.
To obtain the paths, we only need the edge on which each reaches its destina-
tion, because the earlier portion of a shortest u , -path that reaches along v is
a shortest u , v-path. Here the final edges on the paths to a , b , c , d , e generated
by the algorithm are ua , ub , ac , ad , de, respectively; these are the edges of the
spanning tree generated from u.

a• 5 •d
1 4 2
u• •e
4
3 6
b• 5
•c
Exercises for Section 5.4 247

With the phrasing given in Algorithm 5.4.17, Dijkstra’s Algorithm works

also for digraphs, generating an out-tree rooted at u if every vertex is reachable
from u. The proof works for graphs and for digraphs.

5.4.19. Theorem. Given a [di]graph G and a vertex u ∈ V(G), Dijkstra’s Algo-

rithm computes d(u , ) for every ∈ V(G).
Proof: We prove the stronger statement that at each step, (1) t( ) = d(u , ) if
∈ S, and (2) if ∈ / S, then t( ) is the minimum length of a u , -path reaching
directly from S.
We use induction on | S|. Basis step: | S| = 1. From the initialization, S = {u},
d(u , u) = t(u) = 0, and the minimum length of a u , -path reaching directly from
S is w(u ), which is the initial value for t( ) and is infinite when u ∈ / E(G).
Induction step: Suppose that (1) and (2) are true when | S| = . Let v be a
vertex among  ∈ / S such that t() is smallest. Let S = S ∪ {v}. We first show
d(u , v) = t(v). A shortest u , v-path must exit S before reaching v. By the induc-
tion hypothesis, the length of the shortest path going directly to v from S is t(v).
The induction hypothesis and choice of v imply that a path reaching v via any
vertex outside S has length at least t(v). Hence d(u , v) = t(v), and (1) holds for S .
To prove (2) for S , let  be a vertex outside S other than v. By hypothesis,
the shortest u , -path reaching  directly from S has length t() (∞ if there is
no such path). When we add v to S, we also consider paths reaching  from v.
Since we have now computed d(u , v) = t(v), the shortest such path has length
t(v) + w(v ), which we compare with the previous value of t() to find the shortest
path reaching  directly from S .
Thus (1) and (2) hold for the new set S of size + 1.

• •v •

u• • S •
t(v) ≤ t()
• •

By maintaining d(u , x) ≤ t() for all x ∈ S and  ∈

/ S, the algorithm computes
distances from u in nondecreasing order. It yields d(u , v) = ∞ when v is unreach-
able from u. The special case for unweighted graphs is Breadth-First Search
from u; here both the algorithm and the proof (Exercise 70) are simpler.

EXERCISES 5.4

5.4.1. (−) Prove that each property below characterizes the class of forests.
(a) Every induced subgraph has a vertex of degree at most 1.
(b) Every connected subgraph is an induced subgraph.
(c) If two paths have a common vertex, then their intersection is a path.
5.4.2. (−) Prove that a tree with 2 ¾ vertices of odd degree decomposes into ¾ paths.
248 Chapter 5: First Concepts for Graphs

5.4.3. (−) Prove that a tree with maximum degree ¾ has at least ¾ leaves. For n > ¾ > 1,
construct an n-vertex tree having maximum degree ¾ and exactly ¾ leaves.
5.4.4. (−) Determine the number of vertices in a tree with average degree a.
5.4.5. (−) Determine the possible numbers of vertices in a tree with each degree in {1 , ¾}.
5.4.6. (−) Prove that every nontrivial tree has at least two maximal independent sets,
with equality only for stars. (Note: maximal
= maximum.)
5.4.7. (−) Count the isomorphism classes of n-vertex trees with diameter 3.
5.4.8. (−) Decompose the graph below into two isomorphic spanning trees.
•
• •
• • •
• •
•
5.4.9. (−) Prove that a connected graph G with a cycle has girth at most 2 diam(G) + 1.

For ¾ ∈ , exhibit a graph with diameter ¾ and girth 2 ¾ + 1.
5.4.10. (−) Prove or disprove: If T is a tree in which every vertex adjacent to a leaf has
degree at least 3, then T has two leaves with a common neighbor.
5.4.11. (−) Given a graph G, let G be the graph with vertex set V(G) such that xy ∈ E(G )
if and only if d G(x , y) ∈ {1 , 2}. Prove diam (G ) = ⌈ diam(G)/2⌉ .
5.4.12. (−) Find the least radius among spanning trees of the ¾-dimensional hypercube.
5.4.13. (−) In the graph K1 C4 , assign weights (1 , 1 , 2 , 2 , 3 , 3 , 4 , 4) to the edges in two
ways: one way so that the minimum-weight spanning tree is unique, and another way so
that the minimum-weight spanning tree is not unique.
5.4.14. Let G be an n-vertex graph such that deleting any one vertex of G yields a tree.
Determine the number of edges in G, and use this to determine G itself.
5.4.15. For ¾ > 2, prove that if the vertices of a ¾-regular graph can be partitioned into
sets A and B that each induce a tree, then | A| = | B|. (X. Lv)
5.4.16. (♦) Let d1 , . . . , d n and c1 , . . . , cn be positive integers.
(a) Prove that some n-vertex tree has degrees d1 , . . . , d n if and only if ∑ d i = 2n − 2.
(b) Prove that some 2n-vertex tree has degrees d1 , . . . , d n in one part of the biparti-
tion and c1 , . . . , cn in the other if and only if ∑ d i = ∑ ci = 2n − 1. (Hollingsworth [2013])
5.4.17. (♦) Let T be a tree with ¾ edges. Prove that if G is a graph with minimum degree
at least ¾ , then T ⊆ G. (Comment: By Exercise 5.2.10, the same conclusion follows when
G has average degree at least 2 ¾ . The Erdős-Sós Conjecture states that it suffices to have
average degree more than ¾ − 1, or equivalently more than n(¾ − 1)/2 edges.)
5.4.18. Prove for n > ¾ that every tree with ¾ edges is a subgraph of every n-vertex graph
with more than n(¾ − 2) − (¾2) edges. (Comment: The Erdős-Sós Conjecture implies this.)

5.4.19. (♦) Two n-vertex graphs G and H pack if H ⊆ G.

(a) Prove that if | E(G)| | E(H)| < (2n) , then G and H pack. This is sharp, since K n does
not pack with K 2 + (n − 2)K1 . For even n, find another sharpness example. (Comment:
Sauer–Spencer [1978] proved this and that G and H pack if 2 (G) (H) < n, which also is
sharp. See also Kierstead–Kostochka [2009].)
(b) Prove that if T is an n-vertex tree, and G is an n-vertex graph with #### E(G)#### < n/2,
# #
then T ⊆ G. Is the bound best possible? (Zhou [1984])
Exercises for Section 5.4 249

5.4.20. (+) Let G be a graph such that the degrees of any two vertices that are distance 2
apart sum to at least 2 ¾ (and there is at least one such pair). Prove that G contains every
tree with ¾ edges. (T. Jiang)
5.4.21. (♦) Prove that if K n decomposes into ¾ spanning connected subgraphs, then n ≥ 2 ¾ .
Show that this bound is sharp by decomposing K 2¾ into ¾ spanning double-stars, where a
double-star is a tree of diameter 3. (Hint: Use double-stars whose central edges share no
vertices.) (Lovász [1966], Palumb íny [1973])
5.4.22. (♦) For 2 ≤ d < n, determine the maximum number of edges in an n-vertex graph
having diameter d, and find all the extremal graphs. (Ore [1968], Qiao–Zhan [2019])
5.4.23. Let G be a tree. Prove that V(G) can be split into two nonempty sets such that
each vertex has at least half its neighbors in its own set if and only if G is not a star.
5.4.24. Let U be the set of nonleaf vertices in a forest G with ¾ nontrivial components.
Prove that G has 2 ¾ + ∑u∈U [d(u) − 2] leaves. In terms of r alone, determine the number
of vertices in a tree whose nonleaf vertices are one of degree i for each i with 2 ≤ i ≤ r.
5.4.25. Let e be an edge in a connected multigraph G. Prove that e is a cut-edge of G if
and only if e belongs to every spanning tree of G. Prove that e is a loop if and only if e
belongs to no spanning tree of G.
5.4.26. (♦) Prove that a connected n-vertex multigraph has exactly one cycle if and only
if it has exactly n edges.
5.4.27. (♦) Let G be an n-vertex graph with m edges, where n ≥ 4 and m ≥ 2n − 3.
√ the same length. (Comment: Chen–Lehel–Jacobson–Shreve
Prove that G has two cycles of
[1998] proved that m ≥ n + 2n − 4 − 1 suffices and is asymptotically sharp.)
5.4.28. Let T be a tree of even order. Prove that T has exactly one spanning subgraph in
which every vertex has odd degree.
5.4.29. (♦) A parity subgraph of a graph G is a graph H contained in G such that
d H (v) ≡ d G(v) (mod 2) for all v ∈ V(G).
(a) Prove that every spanning tree of a connected graph G contains exactly one parity
subgraph of G. (Itai–Rodeh [1978])
(b) Hajós conjectured that every even graph with n vertices decomposes into at most
⌊ n/2⌋ cycles. Prove that if this conjecture is true, then every n-vertex graph decomposes
into fewer than 3n/2 cycles and edges. (Pyber [1984, 1992], Dean [1987])
5.4.30. (♦) Let e1 , . . . , e¾ be distinct edges in a graph G having ¾ edge-disjoint spanning
trees. Prove that G has edge-disjoint spanning trees T1 , . . . , T¾ with ei ∈ Ti for 1 ≤ i ≤ ¾ .
5.4.31. (♦) Let G be a tree with ¾ leaves. Prove that G is the union of paths P1 , . . . , P⌈¾/2⌉
such that Pi ∩ Pj
= ∅ for all i
= j . (Ando–Kaneko–Gervacio [1996])
5.4.32. Let G be an n-vertex graph having a decomposition into ¾ spanning trees. Suppose
also that (G) = (G) + 1. For 2  ≥ n, show that this is impossible. For 2  < n, determine
the degree list of G in terms of n and  .
5.4.33. (+) Spanning trees with fixed degrees.
(a) Let G be a graph decomposing into trees S and T . Prove that the number of de-
compositions of G into trees S and T with (d S (v) , d T (v)) = (d S(v) , d T (v)) for all v ∈ V(G)
is even. (Berman [1986], Cameron–Edmonds [1999])
(b) Let G be 4-regular, with e , e ∈ E(G). Prove that G has an even number of decom-
positions into spanning cycles with e in the first and e in the second. (Thomason [1978])
5.4.34. Let x and y be the endpoints of a maximal path P in a graph G, and let l be the
length of P. Prove that d(x , y) ≤ max{2 + l − d(x) − d(y) , 2}. (Tracy [2000])
250 Chapter 5: First Concepts for Graphs

5.4.35. Decomposition into trees. (Chung [1978a])

(a) Let G be a connected graph with at least three vertices. Prove that G has two ver-
tices x , y such that G − {x , y} is connected and d(x , y) ≤ 2.
(b) (+) Prove that every connected n-vertex graph G decomposes into ⌈ n/2⌉ trees.
(Hint: Prove that G has a tree decomposition T1 , . . . , T⌈n/2⌉ and V(G) has a partition into
pairs S1 , . . . , S⌈n/2⌉ such that Si ⊆ V(Ti) for all i (except that | S1 | = 1 when n is odd).
(c) Let G be a graph with n vertices and ¾ components. Prove that G decomposes into
at most ⌊ (n + ¾)/2⌋ trees.
5.4.36. (♦) Let (n , ) be the maximum possible girth of an n-vertex graph with n + 
edges. Trivially (n , 0) = n. Prove (n , 1) = ⌊ (2n + 2)/3⌋ and (n , 2) = ⌊ (n + 2)/2⌋ . (Com-
ment: Bollobás–Thomason [1997] proved (n , ) ≤ ⌊ n+¾ ¾ ⌋ log 2(2 ), and Bollobás–Szemerédi
[2002] improved this to (n , ) ≤ 23 (n + )(log 2  + log 2 log 2  + 4).)
5.4.37. (♦) Diameter and complements. Let G be a graph.
(a) Prove (diam G ≥ 3 ⇒ diam G ≤ 3) and (diam G ≥ 4 ⇒ diam G ≤ 2).
(b) Prove that if G is regular, then diam G ≥ 3 implies diam G ≤ 2.
5.4.38. (♦) Diameter and radius. Let G be a graph.
(a) Prove the triangle inequality for distance: d G(u , v) + d G(v , w) ≥ d G(u , w).
(b) Use part (a) to prove that diam G ≤ 2radG.
(c) Given positive integers r, d with r ≤ d ≤ 2r, construct a graph with radius r and
diameter d. (Hint: Build a suitable graph with one cycle.)
5.4.39. Let S be a set of  values in n that is unchanged under negation. Let G be the -
regular graph with vertex set n whose vertices are adjacent if their difference modulo n
is in S. Prove that diam G > 3 requires  < ⌊ n/2⌋ . (A. Pasotti)
5.4.40. Let G be a connected Eulerian graph with at least three vertices. A vertex v in
G is extendible if every trail beginning at v can be extended to form an Eulerian circuit of
G. For example, in the graphs below only the marked vertices are extendible. Prove the
following statements about G (adapted from Chartrand–Lesniak [1986, p. 61]).
a) A vertex v ∈ V(G) is extendible if and only if G − v is a forest. (Ore [1951])
b) If v is extendible, then d(v) = (G). (Bäbler [1953])
c) All vertices of G are extendible if and only if G is a cycle.
d) If G is not a cycle, then G has at most two extendible vertices.
• • •
•
• • •
•
• • •
5.4.41. (♦) Prove that the isomorphism class of a tree is determined by the distances among
its leaves. That is, a tree S with leaves {x1 , . . . , x¾ } is determined by knowing for each i
and j the value of d S(xi , xj ). (Smolenskii [1962])
5.4.42. Prove that every automorphism of a tree T with | V(T)| odd has a fixed point.
5.4.43. Pathological behavior of eccentricity.
a) Prove that the center of a graph can be disconnected and can have components ar-
bitrarily far apart , by constructing a graph where the center consists of two vertices and
the distance between these two vertices is  .
b) Prove that if G is a tree, then every vertex x outside the center of G has a neigh-
bor with eccentricity (x) − 1.
c) For r ≥ 3, construct a graph with radius r in which some vertex x has eccentricity
r + 2 and has no neighbor with eccentricity r + 1.
Exercises for Section 5.4 251

5.4.44. Theorem 5.4.11 states that the center of a tree consists of one vertex or two adja-
cent vertices. Give two more proofs of this fact , as follows:
(a) By induction on the number of vertices.
(b) By showing directly that any two vertices in the center of a tree must be adjacent.
5.4.45. (♦) Let the weight of a vertex v in a tree T be the maximum order of a component
of T − v. A centroid is a vertex of smallest weight. Prove that a tree has one centroid or
two adjacent centroids, the latter only when each has weight | V(T)| /2. (Jordan [1869])
5.4.46. Prove that every tree T has a vertex v such that for all e ∈ E(T), the component
of T − e containing v has at least ⌈ | V(T)| /2⌉ vertices. Prove that v is unique or there are
two adjacent such vertices.
5.4.47. (♦) For v ∈ V(G), let s(v) = ∑w∈V(G) d(v, w). The set of vertices minimizing s(v) is
called the barycenter of G.
(a) Prove that if G is a tree with edges xy and y , then 2s(y) < s(x) + s( ). Use this to
prove that the barycenter of a tree consists of one vertex or two adjacent vertices.
(b) Let G be a tree of diameter d. Determine the maximum possible distance in G
between the center and the barycenter.
5.4.48. Let T be a tree with n vertices,  leaves, and maximum degree  . Determine the
maximum and minimum possible values of diam T .
5.4.49. (♦) The Kneser graph K(n , ) (Kneser [1955]) is defined on the -element subsets
of [n], with two -sets adjacent if and only if they are disjoint. For example, the Petersen
graph is K(5 , 2). Prove diam(K(n , )) = 2 for n ≥ 3 − 1. Prove diam(K(2  + 1 , )) =  .
(Comment: In general, diam(K(n , )) =⌈ n¾−−21¾ ⌉ + 1; see Valencia-Pabon & Vera [2005].)

5.4.50. (♦) For a connected n-vertex graph, the average distance is the average of all (n2)
distances between pairs of vertices. Determine the average distance in Pn , and prove that
every connected n-vertex graph other than Pn has smaller average distance.
5.4.51. Determine the average distance between points in the -dimensional hypercube
Q¾ (averaged over all pairs of points).
5.4.52. Let G be a connected graph of order n and minimum degree  , with 2 ≤  ≤ n − 3.
Prove diam G ≤ 3 ¾n−+12 − 1. Prove that equality can hold whenever  ≥ 2 and ¾n+−12 is an
integer greater than 1. (Moon [1965a])
| E(H)|
5.4.53. Let F1 , . . . , F ¾ be forests whose union is G. Prove that  ≥ max H ⊆ G ⌈ | V(H) |−1 ⌉
.
(Comment: Chapter 11 has a proof using matroids that equality holds for some decompo-
sition into forests.) (Nash-Williams [1964], Edmonds [1965a])
5.4.54. Let G be a graph having  edge-disjoint spanning trees. Prove that for any par-
tition of V(G) into r parts, there are at least (r − 1) edges of G whose endpoints are in
different parts of the partition. (Comment: Chapter 11 has a proof using matroids that
the condition is also sufficient.) (Tutte [1961a], Nash-Williams [1961], Edmonds [1965b].)
5.4.55. (♦) Prove that a graph G is a forest if and only if for every set of pairwise inter-
secting paths in G, some vertex belongs to all the paths in the set.
5.4.56. (+) Prove that every n-vertex tree other than K1 ,n−1 is contained in its comple-
ment. (Burns–Schuster [1978])
5.4.57. Let S be an n-element set , and let A1 , . . . , A n be n distinct subsets of S. Prove that
S has an element x such that the sets A1 ∪ {x} , . . . , A n ∪ {x} are distinct. (Hint: Define a
graph G with vertices a1 , . . . , an such that ai aj ∈ E(G) if and only if Ai and Aj differ by
one element. Use that element as a label on the edge. Prove that some forest contains all
the labels that occur on edges, and use it to obtain the desired element x.) (Bondy [1972a])
252 Chapter 5: First Concepts for Graphs

5.4.58. (+) Let G be a graph with fewest vertices among ¾-regular graphs with girth at
least ½ , where ¾ ≥ 2 and ½ ≥ 3. (By Exercise 5.1.47, such graphs exist.) Prove that G has
diameter at most ½ . (Hint: If d G(x , y) > ½ , then obtain from G a smaller ¾-regular graph
with girth at least ½ .) (Erdős–Sachs [1963])
5.4.59. (♦) Moore graphs.
(a) For ¾-regular connected graphs (with ¾ ≥ 3), prove that any two of the three prop-
erties {diameter d, girth 2d + 1, order 1 + ¾[(¾−¾−1)2 −1] } imply the third.
d

(b) The Hoffman–Singleton graph with 50 vertices is the graph G formed by adding
to ten disjoint 5-cycles Q0 , . . . , Q4 and R0 , . . . , R4 a 5-regular bipartite graph as follows.
Let the vertex sets of the 5-cycles be copies of 5 in order. Make i ∈ V(Qj ) and l ∈ V(R¾)
adjacent if and only if 2l ≡ i + 2 j ¾ (mod 5). Use part (a) to prove that G has diameter 2.
5.4.60. For odd d, prove that the maximum number of vertices in a tree with diameter d
(d+1)/2 (d+1)/2
and maximum degree ¾ is ∑i=1 (¾ − 1)i−1 + ∑i=1 (¾ − 1)i−1 .
5.4.61. Let G be a graph with order n, size m, maximum degree ¾ , and diameter d. Prove
m ≥ (n2)/ ∑i=1 (¾ − 1)i−1 . (Hint: Count paths starting at each edge.) (Erdős–Rényi [1962])
d

5.4.62. Let G be a connected graph with distinct edge weights. Without Kruskal’s Algo-
rithm, prove that G has only one minimum-weight spanning tree.
5.4.63. Suppose that in the hypercube Q¾ , each edge whose endpoints differ in coordinate
i is given weight 2 i . Compute the minimum weight of a spanning tree.
5.4.64. (♦) Let G be a connected n-vertex graph. Define a graph G on the set of spanning
trees of G by putting TT ∈ E(G ) if and only if T and T have exactly n − 2 common edges.
Prove that G is connected. Obtain a formula for d G (T , T ).
5.4.65. (♦) Let T be a spanning tree in a graph G with minimum-weight spanning tree
T . Prove that T can transform into T by steps exchanging one edge of T for one edge of
T so that the edge set is always a spanning tree and the total weight never increases.
5.4.66. A minimax or bottleneck spanning tree is a spanning tree in which the maxi-
mum of the edge weights is as small as possible. Prove that every minimum-weight span-
ning tree is a minimax spanning tree.
5.4.67. (♦) Prim’s Algorithm grows a spanning tree from a given vertex of a connected
weighted graph, iteratively adding the cheapest edge between a vertex already absorbed
and a vertex not yet absorbed, finishing when all the vertices have been absorbed. (Ties are
broken arbitrarily.) Prove that Prim’s Algorithm produces a minimum-weight spanning
tree. (Jarnı́k [1930], Prim [1957], Dijkstra [1959], independently).
5.4.68. (♦) In a connected weighted graph, iteratively delete a heaviest non-cut-edge until
no cycle remains. Prove that what remains is a minimum-weight spanning tree.
5.4.69. A greedy algorithm for minimum-weight spanning path iteratively selects the
edge of least weight so that the edges chosen so far form a disjoint union of paths. After n −
1 steps, a spanning path is obtained. Prove that this algorithm always gives a minimum-
weight spanning path, or give an infinite family of counterexamples where it fails.
5.4.70. Breadth-First Search for distances from a vertex u0 in an unweighted graph. Place
the initial vertex u0 on a queue, and declare its distance from u0 to be 0. While there re-
mains a vertex in the queue, remove the oldest vertex v from the queue, declare its neigh-
bors that have not yet been reached to have distance d(u0 , v) + 1 from u0 , and add them
to the queue. When the queue is empty, say that any unreached vertices have infinite
distance from u0 . Prove that this algorithm computes distances from u0 to all vertices.
Chapter 6

Matchings
Can we fill jobs using qualified applicants from an applicant pool? Can we
pair students as roommates with all pairs compatible? How do we find a strategy
for playing Tic-Tac-Toe? These are all questions about matchings in graphs.

6.0.1. Definition. A matching in a graph G is a set of pairwise non-incident

edges, also called independent edges. A matching covers (or saturates)
the vertices in its edges. A matching that covers all of V(G) is a perfect
matching in G. The matching number (G) is the maximum size (number
of edges) of a matching in G.

Study of matchings involves existence (conditions for a perfect matching),

enumeration (how many perfect matchings), and optimization (finding a maximum-
sized matching). The traditional term “saturates” is from optimization; it indi-
cates that the usage of edges at a vertex has reached its bound. We use the simpler
word “covers”. A vertex not covered by a matching M is “missed” by M. For the
history of matching theory, see Mulder [1992] and Plummer [1992].

6.1. Matching in Bipartite Graphs

Filling jobs with qualified applicants is a matching problem in a bipartite
graph. The parts are the jobs and the applicants; the edges record which appli-
cants can do which jobs. We ask whether some matching fills all the jobs.

6.1.1. Remark. We can restate the job-filling problem using sets. For the ith job,
there is a set Ai of qualified applicants. The problem is to find a system of dis-
tinct representatives (SDR) of the sets A1 , . . . , An , meaning distinct elements
1 , . . . , n such that i ∈ Ai for all i.
The two problems are equivalent; an SDR of a family of sets corresponds to
a matching covering X in a natural bipartite graph. The incidence graph of a
family A1 , . . . , An is the X , Y -bigraph with X = { A1 , . . . , An} and Y = ⋃i∈[n] Ai
such that Ai ∈ X is adjacent to y ∈ Y if and only if y ∈ Ai . An SDR with i
representing Ai for all i corresponds to the matching { Ai i : i ∈ [n]}; we use this
correspondence freely.

253
254 Chapter 6: Matchings

In discussing matchings covering X , we have used n for | X | . When there are

n sets and n elements, the matrix of the membership relation between the sets
and the elements is a (0 , 1)-matrix of order n. An SDR then corresponds to a per-
mutation of [n] that maps indices in one part into indices of the matched vertices
in the other part. Hence in the context of bipartite matching with parts of equal
size, it is convenient to let the total number of vertices be 2n.

HALL’S THEOREM

Philip Hall [1935] characterized when an SDR exists. There is an obvious

necessary condition. For J ⊆ [n], an SDR for a family A1 , . . . , An must have
distinct elements representing the sets indexed by J , and thus ####⋃i∈ J Ai #### ≥ | J |
# #
is necessary. In terms of matchings in bipartite graphs, the condition is stated
using neighborhoods.

6.1.2. Definition. The neighborhood of a set of vertices is the union of the

neighborhoods of its members. In notation, N(S) = ⋃v∈S N(v).

If an X , Y-bigraph has a matching of size | X | , then the elements of any S ⊆ X

have distinct neighbors in the matching, so | N(S)| ≥ | S|. Hall’s Condition is the
# #
statement that | N(S)| ≥ | S| for all S ⊆ X (or, in the SDR setting, ####⋃i∈ J Ai #### ≥ | J |
for all J ⊆ [n]). Hall proved that this necessary condition for a matching that
covers X is also sufficient.

6.1.3. Theorem. (Hall’s Theorem; P. Hall [1935]) In an X , Y-bigraph, some

matching covers X if and only if | N(S)| ≥ | S| for all S ⊆ X .
Proof: (M. Hall [1948], Halmos–Vaughan [1950], Mann–Ryser [1953]). As re-
marked above, existence of such a matching requires Hall’s Condition. We prove
sufficiency by induction on | X | for such a graph G. When | X | = 1, the claim is
immediate. For | X | > 1, we consider two cases.
If | N(S)| > | S| for every nonempty proper subset S of X , then we fix a vertex
x ∈ X and choose y ∈ N(x), which exists since Hall’s Condition prohibits isolated
vertices in X . The graph G − {x , y} satisfies Hall’s Condition, because each sub-
set of X − {x} loses at most one neighbor, y. The induction hypothesis yields a
matching covering X − {x} in G − {x , y}, and adding the edge xy completes the
desired matching in G.
In the remaining case (see figure below), there is a nonempty proper subset S
of X such that | N(S)| = | S|. Let G 1 be the subgraph induced by S ∪ N(S), and let
G 2 = G − (S ∪ N(S)). Hall’s Condition holds for G 1 , since N(T) ⊆ N(S) for T ⊆ S.
Hall’s Condition also holds for G 2 as follows: since NG 2 (T) = NG(T ∪ S) − NG(S)
for T ⊆ X − S, we have
| NG2 (T)| = | NG(T ∪ S)| − | NG(S)| ≥ | T ∪ S| − | S| = | T |.
By the induction hypothesis, G 1 has a matching covering S and G 2 has a match-
ing covering X − S. Their union is the desired matching in G.
Section 6.1: Matching in Bipartite Graphs 255

X −S T S
• • • • • • • • X

G2 G1
• • • • • • • • Y
Y− N(S) NG 2(T) N(S)

6.1.4.* Remark. A closer look at the argument of Theorem 6.1.3 yields a quan-
titative result (M. Hall [1948]): if G is an X , Y -bigraph satisfying Hall’s Condi-
min{ ,n}
tion, then at least ∏i=1 ( + 1 − i) matchings in G cover X , where n = | X | and
= min x∈ X d(x) (Exercise 34). That is, when Hall’s Condition holds there tend to
be many matchings covering X . For example, K r,s has ∏i=1 (s + 1 − i) matchings
r

covering the part of size r.

Hall’s Theorem states that an obvious necessary condition for existence of

a matching that covers X is also sufficient. The British mathematician Crispin
St. James Alva Nash-Williams coined an acronym for such results: TONCAS (The
Obvious Necessary Conditions are Also Sufficient). Such results usually guaran-
tee a short verification of the answer. Here, one can prove that some matching
covers X by exhibiting one. One can prove that none exists by exhibiting a set
S ⊆ X with | N(S)| < | S|. Always exactly one of these alternatives holds.
It should be noted, however, that the proof given here does not lead to a fast
algorithm for checking Hall’s Condition or showing that the guaranteed match-
ing exists. The techniques in Section 6.3 are useful for that.
Hall’s Theorem has many TONCAS-type applications. We seek an auxiliary
X , Y -bigraph H such that the claimed sufficient condition implies Hall’s Condi-
tion for H , and a matching in H covering X yields the desired conclusion. We
present such an application; others appear in Exercises 19–23. A more detailed
version of this result is in Exercise 18.


6.1.5. Corollary. (Hakimi [1965]) For d ∈ , a graph G has an orientation in
which each vertex has outdegree at most d if and only if every subgraph H
of G has at most d |V(H)| edges.
Proof: Necessity: If some subgraph H has more than d |V(H)| edges, then in any
orientation some vertex of H is the tail of more than d edges.
Sufficiency: Form an X , Y -bigraph G in which X = E(G) and Y consists of
d copies of V(G). For each v ∈ V(G), we have vertices v1 , . . . , vd ∈ Y . For each
uv ∈ X , make uv adjacent to all copies of u and v in Y .

• • • • copies of u
•
xuv • • • •
•
• • • • copies of v
•
• • • •
X • Y
• • • •
256 Chapter 6: Matchings

A matching M in G that covers X selects an endpoint of each edge uv, by

matching uv to a copy of u or a copy of v. Since there are d copies of each vertex,
each vertex of G is selected at most d times. Therefore, orienting each edge so
that the endpoint selected by M is the tail produces the desired orientation of G.
It thus suffices to show that G has a matching that covers X . To do this, we
verify Hall’s Condition. For each S ⊆ X , there is a corresponding subgraph H of
G whose edge set is S. Let U be the set of vertices incident to the edges of S in G.
By the hypothesis, | S| ≤ d |U |. However, d |U | = | NG (S)| , since the neighbors of S
in Y are precisely the d copies of the vertices incident to edges of S. Thus Hall’s
Condition holds.

The application of Hall’s Theorem to regular bipartite graphs is called the

“Marriage Theorem”. Multiedges can be important to achieve regularity. Our
statements about existence (but not enumeration) of matchings in bipartite graphs
extend to bipartite multigraphs. (For ¾-regular bipartite multigraphs, a sharp
lower bound on the number of perfect matchings is n!(¾/n)n , weaker than Re-
mark 6.1.4. The bound follows from the famous van der Waerden Conjecture,
proved independently by Egorychev [1981] and Falikman [1981]; see Exercises
15.2.30-31.)

6.1.6. Corollary. (Marriage Theorem; K önig [1916], Frobenius [1917]) Every

nontrivial regular bipartite multigraph has a perfect matching.
Proof: In a ¾-regular bipartite multigraph with parts X and Y , consider any
S ⊆ X . Counting edges joining S and N(S) yields ¾ | S| ≤ ¾ | N(S)| , since N(S)
receives all the edges from S and maybe more. “Nontrivial” implies ¾ ≥ 1, so we
can divide by ¾ , and Hall’s Condition holds.

S X

¾ | S|

N(S) Y

The biadjacency matrix or reduced adjacency matrix of an X , Y-bigraph

G is the submatrix of A(G) induced by the rows for X and the columns for Y . For
a ¾-regular bipartite multigraph, this matrix is a nonnegative integer matrix
whose rows and columns all sum to ¾ . The positions for a perfect matching are
those of the 1s in a permutation matrix: one 1 in each row and each column.
Deleting a perfect matching leaves a (¾ − 1)-regular subgraph and reduces each
row and column by 1. By induction on ¾ , every nonnegative integer matrix whose
rows and columns all sum to ¾ is a sum of permutation matrices.

6.1.7.* Definition. A doubly stochastic matrix is a nonnegative matrix whose

rows and columns all have sum 1. A convex combination is a linear combi-
nation using nonnegative coefficients that sum to 1.

If the entries in a doubly stochastic matrix M are rational, then we can mul-
tiply M by the least common multiple of the denominators (call it ¾) to obtain
Section 6.1: Matching in Bipartite Graphs 257

a nonnegative integer matrix whose rows and columns all have sum ¾ . As we
have just observed, this is the biadjacency matix of a ¾-regular bipartite multi-
graph G, and the matrix can be expressed as a sum of permutation matrices. Let
P1 , . . . , Pr be the permutation matrices used, and let ci be the number of times Pi
is used. By letting the coefficient i of Pi be ci/ , we have M = ∑ i=1 i Pi . This ex-
r

presses M as a convex combination of permutation matrices, proving the special

case of the next theorem in which the entries are all rational.

6.1.8.* Theorem. (Birkhoff–von Neumann Theorem; Birkhoff [1946], von

Neumann [1953]) Every doubly stochastic matrix is a convex combination of
permutation matrices.

In fact, the full strength of the theorem (for any real entries) can be proved
by induction on the number of nonzero entries using Hall’s Theorem (Exercise
25). One may wonder about the size of the coefficients. Is it possible to find a
permutation matrix whose 1s correspond to entries in M that are not too small?

6.1.9.* Example. An island of area n is home to n married couples, each consist-

ing of a farmer and a rancher. The Ministry of Agriculture splits the island into
n farms of area 1. The Ministry of Meat independently splits it into n ranches
of area 1. The Ministry of Marriage requires that each couple ’s farm and ranch
overlap. Theorem 6.1.8 guarantees that such an assignment exists. Can we match
the farms and ranches so that each couple ’s intersection is large?
In the example below, the ranches (1 , 2 , 3) are horizontal strips, and the
farms (a , b , c) extend vertically. The matrix records areas of intersection. Since
1 and 2 share area .5 only with a, one of them must match to a farm with area
only .25. This example is the worst; when n = 3 there always are independent
positions with values at least .25. We next find the worst case in general.

1 b a c a b c
1 ⎛ .5 .25 .25 ⎞
2 b a c
2 ⎜ .5 .25 .25 ⎟
3⎝ 0 .5 .5 ⎠
3 b c

6.1.10.* Theorem. (Marcus–Ree [1959], Floyd [1990]) Every doubly stochastic

matrix of order n has n independent entries all with value at least  , where
 = (n+41)2 for odd n and  = n(n4+2) for even n. Furthermore, no larger mini-
mum can be guaranteed.
Proof: We use the language of Example 6.1.9. We first generalize the construc-
tion. Partition a square of area n into n equal strips as ranches. Let T be a set
of t ranches. Let t − 1 farms intersect each ranch in T with area 1t . Split the re-
maining 1t for each ranch in T equally among the remaining n + 1 − t farms. Some
ranch in T must match to a farm intersecting it in area n+11−t 1t . Set t = ⌈ n2 ⌉ .
To prove the guarantee of finding a matching with common area at least 
for each pair, define a bipartite graph G by making ranches and farms adjacent
258 Chapter 6: Matchings

when they have common area at least . It suffices to prove that G satisfies Hall’s
Condition. Let (A) denote the area of A.
For Hall’s Condition, consider a set S of  ranches, and let R be their union.
We need  farms that each intersect some ranch in S with area at least . If
(F ∩ R) ≥ n+11−¾ for some farm F , then F intersects some ranch in S with area
at least ¾(n+11−¾) and hence at least . We show that there are at least  such farms.
Index the farms F1 , . . . , F n so that (F1 ∩ R) ≥ · · · ≥ (F n ∩ R). Let
a = (F ¾ ∩ R). Since (F i ∩ R) ≤ 1 for i <  and (F i ∩ R) ≤ a for i ≥  ,
n
 − 1 + a(n + 1 − ) ≥ ∑ (F i ∩ R) = (R) =  .
i=1

Simplifying yields a ≥ 1
n+1−¾ . We conclude that N(S) contains F1 , . . . , F ¾ , and
Hall’s Condition holds.

MIN-MA X REL ATIONS

Hall’s Theorem leads to a formula for the maximum size of a matching in a

bipartite graph. It turns out to equal a trivial upper bound.

6.1.11. Definition. Let  (G) denote the maximum size of a matching in a graph
G. For an X , Y -bigraph G, the defect df(S) of a set S ⊆ X is | S| − | N(S)|.

One would like to call df(S) the “deficiency” and write def(S), but we use that
term and notation later to discuss matching in general graphs.

6.1.12. Corollary. (Ore’s Defect Formula; Ore [1955]). In an X , Y-bigraph

G, a formula for the maximum size of a matching is given by
 (G) = min{| X | − df(S)} .
S⊆ X

Proof: Every matching misses at least df(S) vertices of S, so  (G) ≤ | X | − df(S)

for every S. We construct a matching of size | X | − d, where d = max S⊆ X df(S).
Form G by adding d vertices to Y , each adjacent to all of X . Now G satisfies
Hall’s Condition, since each subset of X has received d more neighbors. Hence G
has a matching of size | X | . By the construction, at most d edges of this matching
are in G − G, so the remaining edges form a matching of the desired size in G.

The size of a matching is bounded using a natural “dual” concept.

6.1.13. Definition. A vertex cover is a set of vertices having at least one end-
point of every edge; the stars centered at these vertices cover the edges. The
minimum size of a vertex cover in a graph G is denoted (G).

6.1.14. Theorem. (K önig–Egerváry Theorem; K önig [1931], Egerváry [1931]).

If G is bipartite, then max |matching| = min |vertex cover|. That is,
 (G) = (G).
Section 6.1: Matching in Bipartite Graphs 259

Proof: No vertex can cover more than one edge of a matching, so (G) ≥  (G). We
seek a vertex cover of size  (G). Let G be an X , Y -bigraph. By Corollary 6.1.12,
there is a set T ⊆ X such that  (G) = | X | − | T | + | N(T)|. Since N(T) ∪ (X − T) is
a vertex cover, (G) ≤ | N(T)| + | X − T | =  (G).

X−T T

Y − N(T) N(T)

A statement of equality for the optima of minimization and maximization

problems over a class of inputs (like Theorem 6.1.14) is a min-max relation. The-
orem 6.1.14 ensures that in a bipartite graph we can always prove optimality of
a maximum matching by finding a vertex cover of the same size. Exercise 42 re-
quests another proof.
Matching and vertex cover are called dual problems because every vertex
cover must have size at least as big as every matching, making (G) ≥  (G) triv-
ial for every graph G. A similar pair of optimization problems involves indepen-
dent sets of vertices. The independence number of a graph is the maximum
size of an independent set of vertices. An X , Y -bigraph may have independent
sets larger than max{| X | , | Y |}, as shown below.

• • •

• • •

Just as no vertex covers two edges of a matching, no edge contains two ver-
tices of an independent set. Again we have a dual covering problem.

6.1.15. Definition. An edge cover of G is a set of edges that cover the vertices of
G (only graphs without isolated vertices have edge covers). We list notation
for related optimization parameters:
maximum size of independent set (G)
maximum size of matching  (G)
minimum size of vertex cover (G)
minimum size of edge cover (G)

We use (G) for minimum vertex cover due to its interaction with maximum
matching. We put the “prime” on (G) rather than on (G) because (G) counts
a set of vertices and (G) counts a set of edges.
Always (G) ≥ (G). We now prove the min-max relation.

6.1.16. Theorem. (K önig’s Other Theorem, K önig [1916]) If G is a bipartite

graph with no isolated vertices, then (G) = (G).
Proof: Let Q be a minimum vertex cover in an X , Y -bigraph G. Let R = X ∩ Q
and S = Y ∩ Q. Since  (G) = (G) by the K önig–Egerváry Theorem, G has a
260 Chapter 6: Matchings

matching M of size | Q|. Since Q covers every edge, M matches R into Y − S and
S into X − R, and each edge of M covers one vertex of (Y − S) ∪ (X − R). Adding
one edge to cover each remaining vertex of (Y − S) ∪ (X − R) completes an edge
cover of size | Y − S| + | X − R|. Furthermore (Y − S) ∪ (X − R) is an independent
set, so we have produced an edge cover and an independent set of the same size;
they are a smallest edge cover and largest independent set.

6.1.17. Remark. Every n-vertex graph G satisfies (G) + (G) = n, since the
complement of any independent set of vertices is a vertex cover. Surprisingly, the
edge parameters satisfy the same equality: Gallai’s Theorem (Gallai [1959])
states that (G) +  (G) = n for every n-vertex graph G without isolated vertices.
For bipartite graphs, this follows from the proof of Theorem 6.1.16, where
we produced an edge cover of size n − (G). It was a smallest edge cover since we
produced an independent set of the same size. For non-bipartite graphs this is
not enough: Exercise 50 requests a proof of the full statement.

EXERCISES 6.1

6.1.1. (−) Prove (G) ≥ (G)/2 for every graph G. Prove (G) ≥ (G) for every bipartite
graph G. In each case, give infinitely many examples where equality holds.
6.1.2. (−) A cycle-factor of a digraph is a set of (directed) cycles such that each vertex
lies on exactly one of the cycles. Prove that a digraph D has a cycle-factor if and only if
| N +(S)| ≥ | S| for all S ⊆ V(D). (Ore [1962])
6.1.3. (−) Let T be an n-vertex tree with independence number  . Determine (T).

6.1.4. (−) Prove that a graph G is bipartite if and only if (H) = (H) for all H ⊆ G.
6.1.5. (−) Prove that (G) ≥ m/ (G) when G is a bipartite graph with m edges.
6.1.6. (−) Prove that (G) ≤ n/2 when G is a nontrivial regular n-vertex graph. Prove
that equality holds if and only if G is bipartite.
6.1.7. (−) Prove n
(G)+1 ≤ (G) ≤ n − (G) for every n-vertex graph G.
6.1.8. Prove that a bipartite graph G has a perfect matching if and only if | N(S)| ≥ | S| for
all S ⊆ V(G), and present an infinite class of examples to prove that this characterization
does not hold for all graphs.
6.1.9. Let G be a bipartite graph having a perfect matching. For r < | V(G)|/2, prove that
if every matching of size r extends to a perfect matching, then also every matching of size
r − 1 extends to a perfect matching.
6.1.10. Given n red and n blue points in the plane, with no three on a line, prove that
there is a matching of all red points to blue points by straight line segments so that no two
of the segments cross.
6.1.11. Let M be a matching in the hypercube Q¾ such that the vertices covered by M
induce no edges other than M. Prove that M is contained in a perfect matching of Q¾ .
(Vandenbussche–West [2009])
6.1.12. The hypercube Q¾ has vertex set {0 , 1}¾ ; the weight of a vertex is the number of
nonzero positions. Let M be a perfect matching in Q¾ . For each i, compute the number of
edges of M whose endpoints have weights i and i + 1.
Exercises for Section 6.1 261

¾− 2
6.1.13. For ¾ ≥ 2, prove that Q¾ has at least 2 2 perfect matchings.
6.1.14. Let A and B be disjoint independent sets in a graph G, with | A| = (G). Prove
that G[ A ∪ B] contains a matching of size | B|.
6.1.15. (♦) Let G be an X , Y -bigraph with no isolated vertices and with | X | ≤ | Y | .
Prove that Hall’s Condition (| N(S)| ≥ | S| for all S ⊆ X) is equivalent to the statement
“ | T | − | N(T)| ≤ | Y | − | X | for all T ⊆ Y ”.
6.1.16. (♦) Let G be a connected regular bipartite graph. Prove that any graph obtained
from G by deleting one vertex from each part has a perfect matching. Conclude that every
edge in G appears in some perfect matching. For  ≥ 2, construct a -regular bipartite
graph having two nonincident edges that do not appear together in a perfect matching.
6.1.17. Let G be an X , Y -bigraph. Prove that there exists x ∈ X such that every edge
incident to x belongs to some maximum matching.
6.1.18. (♦) Prove that a multigraph G with m edges has an orientation with specified out-
degree d i at each vertex vi if and only if ∑vi ∈V(G) d i = m and ∑vi ∈U d i ≥ | E(G[U])| for every
U ⊆ V(G). (Hakimi [1965]) (Comment: The special case G = K n was proved in Landau
[1953], characterizing “score sequences” of tournaments (Exercise 5.2.39). Hall’s Theo-
rem was used to prove that result in Bang–Sharp [1979].)
6.1.19. (♦) Consider a deck of mn cards with n suits and m values; each value appears on
one card in each suit. The cards are placed randomly in a grid with n rows and m columns.
(a) Prove that it is possible to choose one card from each of the m columns so that the
values are all distinct.
(b) Use part (a) to prove that it is possible to iteratively exchange positions of two cards
of equal value so that eventually each suit appears in each column. (Enchev [1994])
6.1.20. (♦) Let T be a set of  permutations of [n]. Prove that if  ≤ n/2, then some per-
mutation of [n] disagrees in every position with every member of T . Prove that if  > n/2,
then there may be no such permutation. (K ézdy–Snevily; see Cameron–Wanless [2005])
6.1.21. A travel club is planning vacations. Trips t1 , . . . , t n are available, but trip ti has
capacity ci . Each person likes some trips and will take at most one. In terms of which
people like which trips, derive a necessary and sufficient condition for being able to fill all
trips (to capacity) with people who like them.
6.1.22. Let G be an X , Y -bigraph having a matching that covers X . Given S, T ⊆ X such
that | N(S)| = | S| and | N(T)| = | T | , prove that | N(S ∩ T)| = | S ∩ T | .
6.1.23. (♦) Let G be an X , Y -bigraph with no isolated vertices, with d(x) ≥ d(y) when-
ever xy ∈ E(G) with x ∈ X and y ∈ Y . Prove that G has a matching covering X . (Hint:
Consider a smallest set violating Hall’s Condition.) (F. Galvin)
6.1.24. A positional game consists of a set X of positions and a family W1 , . . . , Wm of
winning sets of positions (Tic-Tac-Toe has nine positions and eight winning sets). Two
players alternately choose positions; the player who first occupies the positions of a win-
ning set wins. Let a be the minimum size of a winning set , and let b be the maximum
number of winning sets containing a given position. Prove that Player 2 can force a draw
if a ≥ 2b. (Comment: More generally, Player 2 can force a draw in d-dimensional Tic-Tac-
Toe when the sides are long enough, given the appropriate definition of winning sets. For
fixed side-length, Player 1 can win if the dimension is large enough; see Chapter 10.)
6.1.25. Prove the Birkhoff–von Neumann Theorem (Theorem 6.1.8) that every doubly
stochastic matrix can be written as a convex combination of permutation matrices. (Hint:
Use induction on the number of nonzero entries.)
262 Chapter 6: Matchings

6.1.26. Let M be a nonnegative integer matrix of order n in which every row and column
has sum t. Determine the least value r such that if t ≥ r, then M must contain a set of n
entries, with one in each row and each column, that all exceed 2. (Stanley [2015b])

6.1.27. (+) Let G be an n-vertex graph with (G) ≥  .

(a) Let F and F be forests with the same number of components. Prove that if F ⊆ F
and | V(F)| ≤  + 1, then every copy of F in G is contained in a copy of F in G.
(b) Let F be a forest with  edges and at most n vertices. Prove that F ⊆ G. (Hint:
When F has no isolated vertices, deleting a leaf from each component yields a -vertex
forest. Comment: The case where F is a tree is Exercise 5.4.17.) (Brandt [1994])

6.1.28. Use the K önig–Egerváry Theorem to prove both the Marriage Theorem and Hall’s
Theorem.

6.1.29. (♦) For an X , Y -bigraph G with | X | = | Y | = n, prove  (G) ≥ min{n , y (G)}:

(a) Using Ore’s Defect Formula.
(b) Using the K önig–Egerváry Theorem.

6.1.30. Let G be a connected X , Y -bigraph with vertex degrees d1 , . . . , d n in nonincreas-

ing order. Prove  (G) ≥ max¾ min{| X | −  , | Y | −  , 2d n−¾−1 } (this strengthens Exercise
6.1.29). (Jahanbekam–West [2013])

6.1.31. A multigraph is r-semiregular if every vertex degree is r or r − 1.

(a) For 0 ≤ r ≤ d, prove that every loopless d-semiregular multigraph has a spanning
r-semiregular submultigraph. (Thomassen [1981c])
(b) Prove that every graph with maximum degree at most s + t − 1 is the union of
graphs G1 and G 2 with (G1) ≤ s and (G 2) ≤ t. (Lovász [1966])

6.1.32. Let A1 , . . . , A m be -sets such that each element of the union lies in at most  sets.
Prove that there exist disjoint B1 , . . . , Bm of size ⌊ /⌋ such Bi ⊆ Ai for all i. (T. Jiang)

6.1.33. (♦) Let G be an X , Y -bigraph without (m + 1)K 2 as an induced subgraph.

(a) Prove that if G has no isolated vertices, then N(S) = Y for some S ⊆ X with
| S| ≤ m. (Liu–Zhou [1997])
(b) Prove that G has at most m( (G))2 edges. (Faudree–Gyárfás–Schelp–Tuza [1989])

6.1.34. (♦) Let G be an X , Y -bigraph with m = | X | and = min x∈ X d(x). Prove that if
G satisfies Hall’s Condition for X , then at least ∏i=1
min{ ,m}
( + 1 − i) matchings in G cover
X . (Hint: Follow the method of Theorem 6.1.3, proving this lower bound instead of mere
existence. Multiedges are forbidden.) (M. Hall [1948])

6.1.35. Exhibit a perfect matching in the graph drawn below or give a short proof that it
has none. (Lovász–Plummer [1986, p. 7])

• •
• • • •
• • • • • • •
• • • • • • • •
• • • • • • • •
• • • • • • •
• • • •
• •

6.1.36. Determine the maximum number of edges in a bipartite graph that has no match-
ing with  edges and contains no star with l edges. (Isaak)
Exercises for Section 6.1 263

6.1.37. (♦) Prove that every subgraph of K n ,n with more than (¾ − 1)n edges has a matching
of size at least ¾ .
6.1.38. Prove that in a bipartite graph G, a vertex belongs to some smallest vertex cover
if and only if it is covered by every maximum matching.
6.1.39. For ¾ ≤ r ≤ s, characterize the maximal subgraphs of K r,s that have no matching
of size ¾ , and determine which have the most edges.
6.1.40. (♦) Let G be an X , Y -bigraph with (G) = | X | = r. Prove that at most (2r ) edges
of G lie in no maximum matching. Show that this is sharp for every r.
6.1.41. (♦) Determine the largest number b such that every maximal matching in every
3-regular graph G has size at least b | V(G)|.
6.1.42. (+) Let G be a bipartite graph, and let G be a minimal spanning subgraph of G
such that (G ) = (G). Prove that G has no two incident edges, and use this to prove the
K önig–Egerváry Theorem. (Lovász [1975])
6.1.43. (♦) Let G be an n-vertex graph with m edges. Prove (G) ≤ n/2 + m/6, with equal-
ity only when every component of G is K3 or K4 . (Hint: Consider cases depending on (G).)
(Henning–Yeo [2013a])
6.1.44. Let G be a nonbipartite graph having exactly one cycle, C. Prove that (G) ≥
(n − 1)/2, with equality if and only if G − V(C) has a perfect matching.
6.1.45. (♦) Let G be a graph with vertex cover number  .
(a) Build a vertex cover R greedily as follows: iteratively select a vertex of maximum
degree in the remaining graph, add it to R, and delete it , until the remaining graph has
no edges. Prove that the final set R may be as large as  ln  − O(). (Hint: Construct an
¾
X , Y -bigraph with | X | =  and | Y | = ∑i=2 ⌊ /i⌋ so that the algorithm chooses all of Y .)
(b) Another algorithm adds to R the vertices of a remaining edge and deletes them.
Prove that this always produces a vertex cover of size at most 2  .
6.1.46. (♦) Build an independent set S greedily as follows: iteratively select a vertex v of
minimum degree in the remaining graph, add it to S, and delete {v} ∪ N(v). Prove that
the final independent set S has size at least ∑u∈V(G) d (u)
1
+1 .
(Caro [1979], Wei [1981])
G

6.1.47. For a graph G, order V(G) as v1 , . . . , vn by iteratively deleting a vertex of max-

imum degree in G − {v1 , . . . , vi−1 } and calling it vi . Let  = ⌈ ∑i=1 1 + d1 (vi) ⌉ . Prove that
n
G
{vn−¾+1 , . . . , vn} is an independent set in G. (Caro–Tuza [1991])
6.1.48. (♦) Let G be an n-vertex graph not having K1 ,3 as an induced subgraph. Prove
(G) ≤ (G)
2n
+2 .

6.1.49. Let G be an n-vertex graph with no isolated vertex.

(a) Prove (G) ≥ 1 + n(G) . (Hint: Use induction on | E(G)| .) (Weinstein [1963])
n (G)
(b) Prove  (G) ≤ 1 + (G))
, and show that both bounds are sharp infinitely often for
each maximum degree.
6.1.50. (♦) Prove Gallai’s Theorem: (G) +  (G) = n whenever G is an n-vertex graph
with no isolated vertices. (Comment: Since G need not be bipartite, the K önig–Egerváry
Theorem and K önig ’s Other Theorem are not relevant.) (Gallai [1959])
6.1.51. Consequences of Gallai’s Theorem (Exercise 6.1.50). Let M be a maximal matching
and L be a minimal edge cover in a graph G having no isolated vertices.
(a) Prove that | M | = (G) if and only if M is contained in a smallest edge cover.
(b) Prove that | L| =  (G) if and only if L contains a largest matching. (Norman–
Rabin [1959], Gallai [1959])
264 Chapter 6: Matchings

6.1.52. An edge e of a graph G is -critical if (G − e) > (G). Let xy and x  be -critical

edges in G. Prove that some induced subgraph of G is an odd cycle containing xy and x .
(Hint: Let Y and Z be largest independent sets in G − xy and G − x  , respectively. Let
H = G[Y  Z]. Prove that every component of H has the same number of vertices from Y
and from Z. Conclude that y and  lie in the same component of H .) (Berge [1970])
6.1.53. -critical edges.
(a) Prove that in a bipartite graph, every -critical edge intersects every largest in-
dependent set of vertices.
(b) Prove that in a tree, the -critical edges form a matching. (Zito [1991])
(c) Prove that both properties above may fail in general graphs.

6.2. Matching in General Graphs

The matching theorems that we proved for bipartite graphs do not hold for
general graphs.

6.2.1. Example. Odd cycles. Note that (C2¾+1) = (C2¾+1) =  , but (C2¾+1) =
 (C2¾+1) =  + 1. That is, the min-max relations fail. Also Hall’s Theorem fails:
although C2¾+1 satisfies “ | N(S)| ≥ | S| for all S ⊆ V(G)” (see Exercise 6.1.8), it has
no perfect matching.

Nevertheless, we still seek a necessary and sufficient condition for the exis-
tence of a perfect matching in a general graph.


6.2.2. Definition. For  ∈ , a -factor in a graph G is a -regular spanning
subgraph of G. In the context of factors, let an odd component of a graph
be a component of odd order (odd number of vertices). We use o(H) to denote
the number of odd components of a graph H.

A perfect matching is a set of edges; a 1-factor is a subgraph whose edge set

is a perfect matching. The terms are almost interchangeable.

TUTTE’S 1-FACTOR THEOREM

As with Hall’s Theorem, we start with an obvious necessary condition. If

a graph G has a 1-factor, and S ⊆ V(G), then G − S must have at most | S| odd
components, since each odd component of G − S must have a vertex matched into
S by the 1-factor. Tutte proved that this condition is also sufficient for a 1-factor.

odd odd

S
even even
Section 6.2: Matching in General Graphs 265

6.2.3. Theorem. (Tutte’s 1-Factor Theorem; Tutte [1947]) A graph G has a

1-factor if and only if o(G − S) ≤ | S| for all S ⊆ V .

Tutte ’s Theorem has many proofs, including Gallai [1950, 1963, 1964],
Maunsell [1952], Kotzig [1959–60], Edmonds [1965c], Anderson [1971], Lovász
[1975], Mader [1973]. Like Hall’s Theorem, Tutte ’s Theorem leads to a min-max
relation for the maximum size of a matching. Every matching must leave an un-
covered vertex in at least o(G − S) − | S| odd components of G − S. This yields
an upper bound on (G); the Berge–Tutte Formula is the statement that equal-
ity holds. This can be proved from Tutte ’s Theorem by a proof like our proof
of Ore ’s Defect Formula from Hall’s Theorem (see Exercise 4). The forced uncov-
ered amount max{o(G − S) −|S|} is analogous to the notion of “defect ” in bipartite
matching, and again we have a special term for it.

6.2.4. Definition. The deficiency def(S) of a vertex set S in a graph G is

o(G − S) − | S|. Tutte’s Condition is the statement “ o(G − S) ≤ | S| for all
S ⊆ V(G)”. A set with positive deficiency is a Tutte set.

graph class characterization min-max relation

Hall s Condition Ore s Defect Formula
bipartite (G)
∀S ⊆ X , | N(S)| ≥ | S| = min[| X | − df(S)]
S⊆ X
Tutte s Condition Berge − Tutte Formula
general ∀S ⊆ V(G) , o(G − S) ≤ | S| (G)
= min 12 [n − def(S)]
S ⊆ V(G)

In fact, the min-max relation and the characterization are equivalent. We

have remarked that Tutte ’s Theorem implies the min-max relation (Exercise 4).
Conversely, Tutte ’s Theorem is a special case of the Berge–Tutte Formula. Just
as Hall’s Condition is the statement “max df(S) = 0”, so Tutte ’s Condition is the
statement “max def(S) = 0”. Thus Hall’s Theorem and Tutte ’s Theorem follow
as special cases of the min-max relations.
To emphasize this relationship, we prove the full statement of the Berge–
Tutte Formula directly. We begin with two lemmas.

6.2.5. Lemma. (Parity Lemma) If S is a set of vertices in an n-vertex graph G,

then o(G − S) − | S| ≡ n (mod 2).
Proof: Counting all vertices shows that o(G − S) + | S| ≡ n (mod 2).

6.2.6. Remark. The Parity Lemma aids in proving Tutte ’s Condition when
|V(G)| is even. To prove o(G − S) ≤ | S| , it then suffices to prove o(G − S) < |S| + 2,
since two integers with the same parity cannot differ by 1.

Many results about maximum matchings can be proved by studying vertex

sets with maximum deficiency. When T is a maximal such set (contained in no
other), the components of G − T are very well behaved.

6.2.7. Lemma. Let T be a maximal subset of V(G) among those having largest
deficiency. If u is a vertex of an odd component C of G − T , then C − u satisfies
Tutte ’s Condition. Also, G − T has no even components.
266 Chapter 6: Matchings

Proof: Subscripts on def denote the relevant graph. For S ⊆ V(C − u), the odd
components of C − u − S are in C − u. Comparing o(G − T − u − S) to o(G − T), we
lose one odd component (C) and gain odd components in C − u − S. Thus
def G(T ∪ {u} ∪ S) = o(G − T − u − S) − (| T | + 1 + | S|)
= o(G − T) − 1 + o(C − u − S) − | T | − 1 − |S|
= def G(T) − 2 + def C−u(S).
Since T is a maximal set of maximum deficiency and T ∪ {u} ∪ S contains T ,
we have def G(T ∪ {u} ∪ S) < def(T). Thus def C−u(S) < 2. Since C − u has even
order, def C−u(S) is even, by the Parity Lemma. We conclude that def C−u(S) ≤ 0
for all S ⊆ V(C − u), which is Tutte ’s Condition.
If G − T has an even component C, then let T = T ∪ {v}, where v is a leaf of
a spanning tree of C. We have | T | = | T | + 1 and o(G − T ) = o(G − T) + 1. Hence
def G(T ) = def G(T), which contradicts the choice of T .

6.2.8. Theorem. (Berge–Tutte Formula; Berge [1958]). In an n-vertex graph

G, the maximum number of vertices covered by a matching is n − d, where
d = max S⊆V(G) def(S). That is,
(G)
= min 12 (n − def(S)).
S ⊆ V(G)

Proof: We know (G) ≤ 12 (n − def(S)) for all S ⊆ V(G). Hence it suffices to con-
struct a matching leaving only d vertices uncovered. We use induction on n; the
claim is trivial for n = 0. For n > 0, let T be a maximal set with deficiency d.
By Lemma 6.2.7, G − T has no even components, and C − u satisfies Tutte ’s
Condition when u is a vertex in an odd component C of G − T . Since C − u has
fewer vertices than G, the induction hypothesis yields a perfect matching in C − u.
Since this holds for any vertex u in any component C of G − T , and there are | T |+ d
such components, it suffices to cover T using edges to | T | distinct components of
G − T . The resulting matching leaves only d uncovered vertices.
Let Y be the set of components of G − T . To match T into Y , we define an
auxiliary T, Y-bigraph H. Let H have an edge ty for t ∈ T and y ∈ Y if and only
if t is adjacent in G to some vertex of the component of G − T corresponding to y.
To obtain a matching that covers T , we apply Hall’s Theorem to H.
For S ⊆ T , the vertices of Y outside NH (S) remain as odd components when
only T − S is deleted. Using also def(T − S) ≤ d and | Y | = | T | + d, we compute
| NH (S)| ≥ | Y | − o(G − (T − S))
= | Y | − (def(T − S) + | T − S|) ≥ | Y | − d − | T | + |S| = |S| .
Since this is true for every subset S of T , Hall’s Condition holds, and Hall’s
Theorem guarantees the needed matching in H.

• •
• Y
T S •
• NH (S)
• •
Section 6.2: Matching in General Graphs 267

Like many other characterization theorems, Tutte ’s 1-Factor Theorem guar-

antees short proofs of both yes and no answers. To prove that G has a 1-factor,
present one. To prove that G has no 1-factor, present a set with positive deficiency
(a “ Tutte set ”). Always one of these two proofs is available. Note that if S is a
Tutte set in a graph of even order, then o(G − S) − | S| ≥ 2, by the Parity Lemma.
Applications of Tutte ’s Theorem usually involve showing that some other con-
dition implies Tutte ’s Condition and hence guarantees existence of a 1-factor.
There are 3-regular graphs without 1-factors; below we show the smallest. Nev-
ertheless, long before Tutte ’s Theorem, Petersen proved a sufficient condition
that now follows easily from Tutte ’s Theorem.

•
• • •
• • • • • •
• • • • • •

6.2.9. Corollary. (Petersen [1891]) Every 3-regular graph with no cut-edge has
a 1-factor.
Proof: Given any vertex subset S, let m be the number of edges joining S to G − S.
Since G is 3-regular, m ≤ 3 | S| . Since any odd component H of G − S must have
even degree-sum, H is joined to S by an odd number of edges. This number is
at least 3 since G has no cut-edge. Hence m ≥ 3o(G − S), so o(G − S) ≤ | S| , and
Tutte ’s Theorem applies.

A more careful look at the argument yields various strengthenings. For ex-
ample, every edge lies in some 1-factor (Schönberger [1934]). Extensions to ¾-
regular graphs were proved by Bäbler [1938] and Berge [1973] (see also Exercise
25 and Exercise 32). Most of these follow from Theorem 6.2.10 (even the hypoth-
esis here can be weakened; see Exercise 29). The condition of even order in the
statement is included to cover the case of even ¾ .

6.2.10. Theorem. (Plesnı́k [1972]) Let G be a ¾-regular loopless multigraph of

even order that remains connected when at most ¾ − 2 edges are deleted. If
G arises from G by deleting at most ¾ − 1 edges, then G has a 1-factor.
Proof: We prove that G has no Tutte set. Consider S ⊆ V(G ), and let m be the
number of edges of G joining S to odd components of G − S. Since G ⊆ G and
G is ¾-regular, m ≤ ¾ | S|. If we can prove m ≥ ¾o(G − S) − 2(¾ − 1), then dividing
by ¾ will yield | S| > o(G − S) − 2. Since | V(G )| is even, the Parity Lemma then
requires | S| ≥ o(G − S), as desired.
For the lower bound on m, consider an odd component H of G − S. Let l
be the number of edges in G leaving V(H); by hypothesis, l ≥ ¾ − 1. Note that
∑v∈V(H) d H(v) = ¾ |V(H)| − l. Since H is a graph, the sum is even, but |V(H)| is
odd, so ¾ and l have the same parity. Thus l ≥ ¾ .
Over all odd components of G − S, the total is at least ¾o(G − S), but this
counts edges of G. Up to ¾ − 1 such edges may be missing in G , counted from both
ends if they join odd components of G − S. Thus we subtract 2(¾ − 1) to ensure
having a lower bound on m. That is, m ≥ ¾o(G − S) − 2(¾ − 1), as desired.
268 Chapter 6: Matchings

6.2.11. Corollary. (Berge [1973]) Let G be a ¾-regular loopless multigraph of

even order. If deleting at most ¾ − 2 edges leaves a connected subgraph, then
every edge of G appears in some perfect matching.
Proof: Delete ¾ − 1 of the ¾ edges incident to one endpoint of the edge desired in
the 1-factor, and apply Theorem 6.2.10.

Lemma 6.2.5 and Lemma 6.2.7 also lead to the Gallai–Edmonds Structure
Theorem describing the structure of maximum matchings (Exercise 40).

GENERAL FACTORS OF GRAPHS

A factor is a spanning subgraph of G; we consider the vertex degrees. For

surveys, see Akiyama–K ano [1985], Volkmann [1995], and Plummer [2007].


6.2.12. Definition. Given a graph G and º : V(G) → 0 , an º -factor of G is a
spanning subgraph H such that d H (v) = º (v) for all v ∈ V(G).


Recall that for ¾ ∈ , a ¾-factor of G is a ¾-regular factor; this is the special
case of º -factor when º (v) = ¾ for all v. Petersen found a sufficient condition for
2-factors, which we prove using only Eulerian circuits and bipartite matching.
We observed in Chapter 5 that even graphs decompose into edge-disjoint cycles.
For regular graphs, these cycles can be grouped into 2-factors.

6.2.13. Theorem. (Petersen’s 2-Factor Theorem; Petersen [1891]) Every reg-

ular multigraph of positive even degree has a 2-factor.
Proof: Let G be a 2 ¾-regular graph with vertices v1 , . . . , vn . By considering
each component, we may assume that G is connected. Hence G is Eulerian; let
C be an Eulerian circuit. Define a U, W-bigraph H with U = {u1 , . . . , un} and
W = {w1 , . . . , wn} by making ui and wj adjacent if vj immediately follows vi some-
where on C. Since C leaves and enters each vertex ¾ times, H is ¾-regular.
Hence the Marriage Theorem yields a 1-factor in H. It picks one edge “leav-
ing ” vi (incident to ui in H) and one edge “entering ” vi (incident to wi in H). Thus
the edges of a 1-factor in H form a 2-factor in G.

6.2.14. Example. Construction of a 2-factor. Consider the Eulerian circuit in K 5

that successively visits 1231425435. The corresponding bipartite graph H is on
the right. For the 1-factor whose u , w-pairs are 12, 25, 54, 43, 31, the resulting
2-factor is the cycle [1 , 2 , 5 , 4 , 3]. The remaining edges form another 1-factor
that yields the remaining cycle [1 , 4 , 2 , 3 , 5].

1
• U W
1• •1
5• •2 2• •2
3• •3
4• •4
4• •3 5• •5
Section 6.2: Matching in General Graphs 269

The general º -factor problem specifies the degree at each vertex. Multiedges
are irrelevant for 1-factors but becomes important for º -factors. Tutte [1952]
found a necessary and sufficient condition for a multigraph to have an º -factor
(proved earlier by Belck [1950] for regular factors). Tutte ’s first proof was quite
difficult; later [1954] he reduced the problem to checking for a 1-factor in a related
graph. We describe this construction. It is a beautiful example of transforming
a graph problem into a previously solved problem.

6.2.15. Example. A graph transformation (Tutte [1954]). When º satisfies the

obvious necessary condition that º (w) ≤ d(w) for all w, we construct a graph H
that has a 1-factor if and only if G has an º -factor. Let e(w) = d(w) − º (w); this
is the excess degree at w and is nonnegative.
To construct H , replace each vertex v with a copy of K d(v) ,e(v) having parts
A(v) and B(v) of sizes d(v) and e(v), respectively. For each vw ∈ E(G), add an edge
joining one vertex of A(v) to one vertex of A(w). Each vertex of A(v) participates
in one such edge.
The figure below shows a multigraph G, vertex labels given by º , and the
resulting graph H. The bold edges in H form a 1-factor that corresponds to an
º -factor of G. The º -factor is not unique.

3 1 • • ••
• • • •
• •
• A(v)•
G 3•
v
•2 • •• • •
• • H
• • B(v)
• •
2
• •1 • • •
• • • •

6.2.16. Theorem. A multigraph G has an º -factor if and only if the blow-up

graph H constructed from G and º in Example 6.2.15 has a 1-factor.
Proof: Necessity. Start with the edges in H corresponding to an º -factor in G.
They leave e(v) vertices of A(v) unmatched, and these match into B(v) (for each v)
to complete a 1-factor of H.
Sufficiency. From a 1-factor of H , deleting B(v) and the vertices of A(v)
matched into B(v) leaves º (v) vertices of degree 1 corresponding to v. Doing this
for each v and merging the remaining º (v) vertices of each A(v) yields a subgraph
of G with degree º (v) at each vertex v.

By Theorem 6.2.16, any algorithm to test for 1-factors can be used to test for
an º -factor. Nevertheless, we still seek a necessary and sufficient condition in
terms of the structure of G. Applications of the resulting Tutte º -Factor The-
orem include a proof of the Erdős–Gallai [1960] characterization of degree lists
of graphs (Exercise 43). The theorem can be proved by translating the statement
of Tutte ’s 1-Factor Condition for the blow-up into a structural condition for G.
First we show necessity of the condition directly.
270 Chapter 6: Matchings

6.2.17. Definition. Given disjoint sets Q , T ⊆ V(G), let  Q , T  denote the num-
ber of edges in G with one endpoint each in Q and T . When º : V(G) → 0 , 
let º (S) = ∑v∈S º (v). Similarly, d H (S) = ∑v∈S d H (v).

6.2.18. Example. The º -factor condition. Given disjoint vertex sets S and T
in G, the condition states a bound on the contributions to º (T) in a hypothetical
º -factor F . Thus º (T) cannot exceed that bound. The necessary condition is
º (T) ≤ º (S) + d G−S(T) − q(S, T) ,
where q(S, T) is the number of components of G − S − T such that º (Q) +  Q , T 
is odd, where Q denotes the vertex set of the component.
Consider how the edges of an º -factor can contribute to º (T). At most º (S)
can join S to T , each contributing once to º (T). The remaining contribution at
each vertex v ∈ T is bounded by d G−S(v), so º (T) ≤ º (S) + d G−S(T). However, sat-
isfying this is not sufficient (Chvátal gave a 13-vertex example of insufficiency).
We can tighten the bound further by subtracting 1 for some components G[Q]
of G − S − T . If the º -factor uses any edge from Q to S, then that edge reduces
the contribution from º (S) in the bound on º (T) by 1. If no such edge is used,
and the º -factor uses all edges joining T to Q, then these edges plus the edges
used within Q form a subgraph H of F . Since no edge from Q to S is used, the
quantity º (Q) +  Q , T  counts each edge of H twice, and hence it must be even.
If it is odd, then the edges joining T to Q cannot all be used, which reduces the
contribution to the bound from d G−S(T). Hence when º (Q) +  Q , T  is odd we get
one way or another a reduction of at least 1 in the bound on º (T).

6.2.19. Theorem. (Tutte’s º -Factor Theorem; Tutte [1952, 1954]). If G is a


multigraph and º : V(G) → 0 , then G has an º -factor if and only if
º (T) ≤ º (S) + d G−S(T) − q(S, T) (∗)
for all disjoint S, T ⊆ V(G), where q(S, T) counts the sets Q such that G[Q]
is a component of G − S − T and º (Q) +  Q , T  is odd.
Proof*: Necessity was shown in Example 6.2.18; consider the converse. With
S = ∅ and T = {w}, (∗) yields º (w) ≤ d(w) for every vertex w; hence the blow-
up graph H from G and º is well-defined. By Theorem 6.2.16, it suffices to prove
that (∗) guarantees Tutte ’s 1-Factor Condition for H.
When Tutte ’s Condition fails, there is a Tutte set C (with o(H − C) > | C|). We
use a minimal Tutte set to construct a partition of V(G) (with R = V(G) − S − T)
violating (∗). Let a parasite in a Tutte set C be a vertex with neighbors in at most
two components of H − C. If x is a parasite in C, then
o(H − (C − x)) ≥ o(H − C) − 1 > | C| − 1 = | C − x| ,
so a minimal Tutte set C has no parasite.
Using C, we prove that each v ∈ V(G) is in one of three types:
Type 1 Type 2 Type 3
Bv ⊆ C Bv ∩ C = ∅ Bv ∩ C = ∅
Av ∩ C = ∅ Av ⊆ C Av ∩ C = ∅
Below we show edges of H − C in bold. If Bv has a vertex x in C and a vertex
y outside C (shown on the left), then the vertices of Av not in C all lie in one
Section 6.2: Matching in General Graphs 271

component of H − C, and x is a parasite. Hence Bv ⊆ C or Bv ∩ C = ∅ (on the

right). If Bv ⊆ C, then any x ∈ Av ∩ C has at most one neighbor in H − C and is a
parasite, so Av ∩ C = ∅ (Type 1); this applies also when Bv = ∅ (that is, e(v) = 0).
For Bv
= ∅, if Bv ∩ C = ∅ and there exist ∈ Av − C and x ∈ Av ∩ C, then Bv
lies in one component of H − C. Now x has neighbors in at most two components
of H − C and is a parasite. Hence C contains all or none of Av (Type 2 or 3).

• x•∈ C
y C
Bv • • • •
Av • • • • • • • • •
x x∈C
Given a minimal Tutte set C, let T, S, R be the subsets of V(G) consisting of
the Type 1, Type 2, and Type 3 vertices, respectively. For v of each type, Bv and
Av are distributed to C and V(H) − C as listed below.
size v∈T v∈S v∈ R
Bv e(v) C V(H) − C V(H) − C
Av d(v) V(H) − C C V(H) − C
We claim that o(H − C) > | C| violates (∗) for this choice of R , S, T . By con-
struction, | C| = ∑v∈S d(v) + ∑v∈ T e(v) = d G(S) + d G(T) − (T). In the figure below,
edges of H − C are in bold. Let R be the family of vertex sets of components of
G[R]. The components of H − C are of four types:
Description of component Order Number with odd order
y ∈ Bv with v ∈ S 1 ∑v∈S e(v)
y  with y ∈ Au ,  ∈ Av , uv ∈ E(G[T]) 2 none
y ∈ Av for each edge uv ∈ [S, T] 1  T, S
one for each Q ∈ R large q(S, T)
• •R

• • •
• •
R• T
• •
C
• • • • •
C
• • •
S T
Consider the large component of H − C corresponding to some connection class
Q ∈ R. It has ∑v∈ Q(d(v) + e(v)) + Q , T  vertices, since it gains one vertex for each
edge from Q to T . Since e(v) = d(v) − (v), this quantity has the same parity
as (Q) +  Q , T . By definition, the number of components of G[R] for which
(Q) +  Q , T  is odd is q(S, T).
Now we rewrite o(H − C) > | C| as
q(S, T) + ∑v∈S [d(v) − (v)] +  T, S > d G(S) + d G(T) − (T).
When we cancel d G(S) and rearrange what remains, we have
q(S, T) + (T) > (S) + (d G(T) −  T, S).
The final quantity in parentheses equals d G−S(T), so (∗) fails.
272 Chapter 6: Matchings

EXERCISES 6.2

6.2.1. (−) Let M be a perfect matching in a ¾-regular graph G, where ¾ is odd. Prove that
M contains every cut-edge of G.
6.2.2. (−) Find maximum matchings in the graph G below and in the graph G obtained
by adding the edge xy. For each prove optimality using “duality”.
•
• • •
•
x y
• • • • • •
•
•
• • •
•

6.2.3. (−) Let G be a (2r − 1)-regular graph. Prove that G decomposes into r factors whose
components are paths or cycles. For example, the Petersen graph decomposes into a perfect
matching and 2C5 .
6.2.4. Prove the Berge–Tutte Formula (Theorem 6.2.8) from Tutte’s 1-Factor Theorem
(Theorem 6.2.3).

6.2.5. (♦) Prove that a tree T has a perfect matching if and only if o(T − v) = 1 for every
v ∈ V(T). (Chungphaisan)
6.2.6. (♦) Minimal Tutte sets.
(a) Let S be a minimal Tutte set in a graph G. Prove that every vertex in S has neigh-
bors in at least three odd components of G − S.
(b) Let G be a connected graph with | V(G)| even. Prove that if G does not contain K1 ,3
as an induced subgraph, then G has a 1-factor. (Sumner [1974], Las Vergnas [1975])
(c) Use part (b) to prove that every connected graph with | E(G)| even decomposes into
3-vertex paths. (Chartrand–Polimeni–Stewart [1973])
6.2.7. Example 5.2.3 shows that every edge in the Petersen graph lies in four 5-cycles.
Use this to count the perfect matchings.
6.2.8. Derive a formula for the number of 1-factors in K n ,n that do not match xi to yi for
any i, and for the number of 1-factors in K 2n that do not match x2i−1 to x2i for any i. No
simple closed formulas are known.
6.2.9. For each ¾ ∈  except ¾ = 1, construct a ¾-regular graph with no 1-factor.

6.2.10. For ¾ ∈ with ¾ ≥ 2, construct a 2 ¾-regular connected graph having no spannning
subgraph in which at most ¾ − 2 vertices have degree 2 ¾ and the rest have degree 2 ¾ − 1.
6.2.11. (♦) Let G be a graph with (G) =  .
(a) Determine the maximum possible value of (G).
(b) Determine the minimum possible size of a maximal matching in G.

6.2.12. Let G be a graph with (G) =  . Prove that G has at most ( − 1) edges that
belong to no maximum matching. Construct examples to show that this bound is best pos-
sible for every  . (F. Galvin)
Exercises for Section 6.2 273

6.2.13. Obtain sufficiency in Hall’s Theorem (Theorem 6.1.3) from Tutte’s Theorem.
(Hint: Transform an X , Y -bigraph G into a graph G that has a 1-factor if and only if
G has a matching that covers X .)

6.2.14. Let T be a set of vertices in a graph G. Prove that G has a matching that covers
T if and only if for all S ⊆ V(G), the number of odd components of G − S contained in G[T]
is at most | S|. (Lovász [1965], extended to a min-max relation in Bollobás [1978])

6.2.15. (♦) Let G be a graph of even order not having K1 ,r+1 as an induced subgraph
(K1 ,r+1 ⊆ G is allowed). Prove that if G − S is connected whenever S ⊆ V(G) with | S| < r,
then G has a 1-factor. (Sumner [1976])

6.2.16. Given graphs G and H , the tensor product G ∗ H is the graph with vertex set
V(G) × V(H) having (u , v) adjacent to (u , v ) if and only if uu ∈ E(G) and vv ∈ E(H).
Prove that the following conditions are equivalent for G (George [1991]):
(A) For every graph H , the graph G ∗ H has no 1-factor.
(B) G ∗ K 2 has no 1-factor.
(C) There exists S ⊆ V(G) such that G − S has more than | S| isolated vertices.

6.2.17. (♦) For a 3-regular graph G with n vertices and c cut-edges, prove (G) ≥ n/2 − c/3.
(Chartrand–Kapoor–Lesniak–Schuster [1984])

6.2.18. (+) Let i(H) be the number of isolated vertices in a graph H ; always i(H) ≤ o(H).
For m ≥ 2, prove that a graph G has a factor whose components are nontrivial stars with
at most m edges if and only if i(G − S) ≤ m| S| for all S ⊆ V(G). (Amahashi–Kano [1982])

6.2.19. (♦) Extension of König–Egerváry to general graphs. Given a graph G, let S1 , . . . , S¾

and T be subsets of V(G) such that each Si has odd size. These sets form a generalized
cover of G if every edge of G has at least one endpoint in T or has both endpoints in some
Si . The weight of a generalized cover is | T | + ∑(| Si | − 1)/2. Let ∗(G) be the minimum
weight of a generalized cover. Prove (G) = ∗(G). (Hint: Use the Berge–Tutte Formula,
Theorem 6.2.8. Comment: Since every vertex cover is a generalized cover, ∗(G) ≤ (G).)

6.2.20. (♦) Use induction on n to determine the maximum number of edges in a 2n-vertex
graph having no 1-factor.

6.2.21. For even n, prove that the maximum number of edges in an n-vertex graph having
a Tutte set of size  is (¾2) + (n − ) + (n−22¾−1 ). Use this to find the maximum number of
edges in an n-vertex graph with no 1-factor. (Erdős–Gallai [1961])

6.2.22. (♦) Prove (G) ≥ min{(G) , ⌊ n/2⌋ } when G has n vertices. (Erdős–P ósa [1962])

6.2.23. Let G have vertex degrees d1 , . . . , d n in nonincreasing order. Strengthen Exercise

6.2.22 by proving (G) ≥ max¾ min{d n−¾ , ⌊ (n − )/2⌋ }. (Jahanbekam–West [2013])

6.2.24. Prove that a 3-regular graph G has a 1-factor when all its cut-edges lie on one
path (Petersen [1891]). Conclude that the minimum number of cut-edges in a 3-regular
graph with no 1-factor is 3. (Comment: More generally, every (2r + 1)-regular graph with
at most 2r cut-edges has a 2-factor, and this is sharp (Hanson–Loten–Toft [1998]).)

6.2.25. (♦) Sharpness of the connectivity threshold for 1-factors in regular graphs. Let  be
an integer at least 3, and let l be the least odd number greater than  .
(a) Construct a -regular graph G of even order that has no 1-factor and remains con-
nected whenever fewer than  − 2 edges are deleted. (Hint: Use (l + 1) − 2 vertices.)
(b) Prove that every answer to part (a) has at least (l + 1) − 2 vertices. (Hint: For a
Tutte set S, treat the large odd components and small odd components of G − S differently.)
(See Niessen–Randerath [1998] for a more general result.)
274 Chapter 6: Matchings

6.2.26. (♦) Let G be a 3-regular graph. Prove that G has a 1-factor if and only if G decom-
poses into copies of P4 . (Comment: Favaron–Genest–Kouider [2010] conjectured that ev-
ery (2 ¾ + 1)-regular graph with a perfect matching decomposes into paths of length 2 ¾ + 1.)
(Kotzig [1957], Bouchet–Fouquet [1983])
6.2.27. The graph on the left below is the Heawood graph. Prove that every 2-factor in
this graph is a spanning cycle. Is this true also for the other graph?

•
• •
• • • • •
• •
• • • •
• • • •
• • • •
• •
• • • • • •
• •

6.2.28. Prove Corollary 6.2.11 directly from Tutte’s 1-Factor Theorem. That is, if xy
is an edge in a ¾-regular graph G of even order such that G cannot be disconnected by
deleting fewer than ¾ − 1 edges, then G − {x , y} has no Tutte set.
6.2.29. Weaken the hypothesis of Theorem 6.2.10 by requiring only that vertex sets of
odd size are not disconnected from the rest of G by deleting fewer than ¾ − 2 edges. Prove
that any G obtained from G by deleting at most ¾ − 1 edges has a 1-factor. (Cruse [1977])
6.2.30. Petersen’s proof of Petersen’s 2-Factor Theorem (Theorem 6.2.13).
(a) Prove that every 4-regular multigraph decomposes into two 2-factors.
(b) For ¾ > 2, prove by induction on | E(G)| that every multigraph G with (G) ≤ 2 ¾
decomposes into ¾ factors with maximum degree at most 2. (Petersen [1891])
6.2.31. (♦) A d-regular multigraph is primitive if it does not decompose into two regular
factors with degree less than d. By Petersen’s Theorem, no primitive multigraphs have d
even and d > 2. Let G be a d-regular n-vertex loopless multigraph with d odd.
(a) Construct an example where G is primitive with n = 3d + 1. (J.J. Sylvester)
(b) Prove (G) ≥ n/3.
(c) For (G) = n − s, prove that G has an odd-regular factor with degree at most 2s + 1.
Conclude that G cannot be primitive if n ≤ 3d − 3. (Petersen [1891]; see Mulder [1992])
6.2.32. (♦) A balloon in a graph G is a maximal subgraph without cut-edges that is inci-
dent to exactly one cut-edge of G; let b(G) be the number of balloons in G (Petersen [1891]
called such subgraphs “leaves”). Let G be a connected n-vertex 3-regular graph.
(a) Prove b(G) ≤ (n + 2)/6, with equality achievable when n ≡ 4 (mod 6).
(b) Prove (G) ≥ 2n − b(G)
3
. (Hint: Use the Berge–Tutte Formula.)
(b) Conclude (G) ≥ (4n − 1)/9. Show that equality can hold when n ≡ 16 (mod 18).

(Comment: For r ∈ , Henning–Yeo [2007] proved (G) ≥ n2 − 2r (2r+(2r −1)n+2
1)(2r 2 +2r−1)
for any
(2r + 1)-regular graph G with n vertices. O–West [2010] proved that equality holds if and
only if G arises from a tree with degrees 1 and 2r + 1 having all leaves on one side of the
bipartition by attaching a copy of P3 + rK 2 at each leaf; below is an example.)

• • •

• • • • • • •
• • • • • • • • • • • •
• • • • • • • • • • • •
Exercises for Section 6.2 275

6.2.33. (+) For ¾ ≥ 2 and ½ ≥ 2 with ½ even, construct a ¾-regular bipartite (multi)graph
with girth ½ . (Hint: Construct it from a (¾ − 1)-regular bipartite graph H with girth ½ and
a bipartite graph with girth ⌈ ½/2⌉ that is | V(H)|-regular. Comment: The requirement of
being bipartite strengthens the result of Erdős–Sachs [1963]; see Exercise 5.1.47.)
6.2.34. (+) Let G be a 2m-regular graph with girth ½ , and let T be a tree with m edges
and diameter d. Prove that if d < ½ , then G decomposes into copies of T . (Hint: Using
Theorem 6.2.13, produce inductively such a decomposition in which each vertex of G is
used once as an image of each vertex of T .) (Häggkvist)
6.2.35. (+) Let G be a 2n-vertex graph. Let P(c) be the statement “ | N(S)| ≥ c | S| whenever
S is a set of at most 2n/c vertices.” (Anderson [1973])
(a) Prove that P(4/3) implies that G has a 1-factor. (Hint: In applying Tutte’s
1-Factor Theorem, consider the cases | S| ≥ n/2 and | S| < n/2 separately.)
(b) Prove that P(c) does not imply that G has a 1-factor when 1 ≤ c < 4/3.
6.2.36. (♦) Let v be a vertex in a graph G. Prove def(G − v) ≤ def(G) + 1.
6.2.37. (♦) A graph is factor-critical if every subgraph obtained by deleting one vertex
has a 1-factor. Prove the following statements.
(a) Every factor-critical graph has odd order, is connected, and is not bipartite.
(b) A connected graph is factor-critical if and only if ∅ is its only Tutte set.
(c) A connected graph is factor-critical if and only if no vertex is in every maximum
matching. (Hint: Consider a Tutte set of maximum deficiency.) (Gallai [1963b])
6.2.38. Let v be a vertex in a factor-critical graph G. Let G be a connected graph obtained
from G by splitting v into two vertices whose neighborhoods form a nontrivial partition of
NG(v). Prove that G has a 1-factor.
6.2.39. Let G be a connected graph such that every edge is in a triangle and all cycles are
triangles. Suppose also that every vertex has degree 2 or 4.
(a) Prove that G is factor-critical.
(b) Prove that for every odd-sized set S of vertices having degree 2, G − S has a unique
perfect matching. (Lovász–Plummer [1986, p. 89]; see also Došlić–Rautenbach [2015])
6.2.40. (♦) In a graph G, let B be the set of vertices covered by every maximum match-
ing. Let A be the set of vertices in B having a neighbor outside B. Let C = B − A and
D = V(G) − B (see figure below). Let G1 , . . . , G ¾ be the components of G[D]. The Gallai–
Edmonds Structure Theorem (Gallai [1963b], Edmonds [1965c]) asserts the claims be-
low about every maximum matching M in G; prove them. (Hint: Let T be a maximal set
of maximum deficiency in G. Use Lemma 6.2.7 to find the sets A , C, D; see West [2011].)
(a) M covers C and matches A into distinct components of G[D].
(b) Each G i is factor-critical and has a near-perfect matching in M.
(c) If ∅
= S ⊆ A, then N(S) intersects at least | S| + 1 of G1 , . . . , G ¾ .
(d) def(A) = def(G) = ¾ − | A|.
G4 •
• • •
G3
• • • •
G2 • • •
• •
G1 • • • •
• •
D A C
276 Chapter 6: Matchings

6.2.41. (+) Let G be a 2n-vertex graph having exactly one 1-factor, M.

(a) Prove that if G has no cut-vertex, then for every two vertices x , y ∈ V(G), the
graph G − {x , y} has a 1-factor. (Lovász [1972c])
(b) Prove that G has a cut-edge that belongs to M. (Kotzig [1959])
(c) Prove (G) ≤ ⌊ log 2(n + 1)⌋ and show that this bound is sharp.
(d) Prove that G has at most n2 edges and that this bound is sharp.
(Comment: Part (c) is Hetyei’s Theorem and appears in Lovász [1972c]. Generalizing
(d), one can ask for the maximum number of edges in a 2n-vertex graph having exactly p
perfect matchings; see Dudek–Schmitt [2012] and Hartke–Stolee–West–Yancey [2013].)

6.2.42. (+) Let G be an even multigraph, and let  assign an odd integer to each vertex.
Prove that G has an even number of  -factors. (Cameron–Edmonds [1999])

6.2.43. (+)  -Factor Theorem and graphic lists. The  -Factor Theorem of Tutte [1952,
1954] (Theorem 6.2.19) states that a graph G has an  -factor if and only if

 (T) ≤  (S) + d G−S(T) − q(S, T)

for all disjoint sets S, T ⊆ V(G), where q(S, T) counts the components G[Q] of G − S − T
such that  (Q) +  Q , T  is odd.
(a) The Parity Lemma. Let (S, T) =  (T) −  (S) − d G − S(T) + q(S, T). Prove that
(S, T) has the same parity as  (V(G)) for disjoint sets S, T ⊆ V(G).
(b) Let  (vi) = d i , where ∑ d i is even and d1 ≥ · · · ≥ d n . Use the  -factor condition
¾
and part (a) to prove that K n has an  -factor if and only if ∑i=1 d i ≤ (n − 1 − s) + ∑i= n+1 − s d i
n

for all  , s with  + s ≤ n.

(c) Show that d1 , . . . , d n is graphic when d1 ≥ · · · ≥ d n if and only if ∑ d i is even and
¾
∑i=1 di ≤ ( − 1) + ∑i=¾+1 min{ , di} for 1 ≤  ≤ n. (Erdős–Gallai [1960])
n

6.2.44. Belck [1950] proved the special case of Tutte’s  -Factor Theorem for regular fac-
tors, restricting  to assign the same integer to all vertices. Use this result to prove that
if j ,  , l are three odd integers with j <  < l and G has a j-factor and an l-factor, then G
also has a -factor. (Katerinis [1985])


6.2.45. For a graph G and a function  : V(G) → 0 , say that G is  -soluble if there ex-

ists w: E(G) → 0 such that ∑uv∈ E(G) w(uv) =  (v) for all v ∈ V(G).
(a) Prove that G has an  -factor if and only if the graph H obtained from G by sub-
dividing each edge twice and defining  to be 1 on the new vertices is  -soluble. (This
reduces testing for an  -factor to testing  -solubility.)

(b) Given G and an  : E(G) → 0 , construct a graph H (with proof) such that G is
 -soluble if and only if H has a 1-factor. (Tutte [1954])

6.3. Algorithmic Aspects

Min-max relations guarantee short proofs of optimality by exhibiting solu-

tions to dual problems. However, we still must find optimal solutions. The needed
tool for maximum matching is augmenting paths. We will use them to give a fast
algorithm for bipartite matching and then consider more general weighted ver-
sions of the dual matching and vertex cover problems.
Section 6.3: Algorithmic Aspects 277

AUGMENTING PATHS

Maximal matchings need not be maximum(-sized) matchings.

6.3.1. Example. Maximal

= maximum. In the drawing of P6 below, the solid
edges form a maximal matching; we can add no other. However, the three dashed
edges form a larger matching.

• • •
• • •

6.3.2. Definition. Given a matching M , an M-alternating path is a path that

alternates between edges in M and edges not in M. An M-alternating path
whose endpoints are missed by M is an M-augmenting path.

If P is an M-augmenting path, then replacing M ∩ E(P) with E(P) − M pro-

duces a new matching M with one more edge than M. Thus a maximum matching
admits no augmenting path. Furthermore, this characterizes maximum match-
ings. Petersen [1891] stated the characterization for the special case of regular
graphs, without giving a proof; Berge [1957] provided a proof in general.

6.3.3. Definition. The symmetric difference A B of sets A and B is (A ∪ B) −

(A ∩ B). The symmetric difference G H of graphs G and H has vertex set
V(G) ∪ V(H) and edge set E(G) E(H).

6.3.4. Theorem. (Berge [1957]) A matching M in a graph G is a maximum

matching in G if and only if G has no M-augmenting path.
Proof: We have noted that an M-augmenting path produces a larger match-
ing. Conversely, when G has a matching M larger than M , we construct an
M-augmenting path. Let F = M M . Since M and M are matchings, every ver-
tex has at most one incident edge from each of them. Hence (F) ≤ 2 when F is
viewed as a subgraph of G.
Since (F) ≤ 2, the components of F are paths and cycles (as illustrated be-
low). Since each component alternates between M and M , each cycle in F has
even length. Since | M | > | M | , some component has more edges of M than of M.
Such a component must be a path that starts and ends with an edge of M and
hence is an M-augmenting path.

• • • •
• • • •
• • • •

Theorem 6.3.4 is useful in studying the properties of maximum matchings.

For example, the thorough three-volume text by Schrijver [2003] on combinato-
rial optimization presents a direct proof of the Berge–Tutte Formula using in-
duction and augmenting paths.
278 Chapter 6: Matchings

Most algorithms for finding maximum matchings are based on augmenting

paths. A simple such algorithm for bipartite graphs also provides a constructive
proof of the K önig–Egerváry Theorem.

6.3.5. Algorithm. Augmenting path algorithm for an X , Y -bigraph G. Given a

matching M , let U be the set of vertices in X not covered by M. Grow a forest
rooted at U by exploring M-alternating paths, letting S and T be the sets of ver-
tices reached in X and Y , respectively. To do this, explore from vertices of S along
edges not in M ; when a vertex of Y covered by M is reached, add it to T , and add
its neighbor via M to S (see the figure below).
Reaching a vertex of Y missed by M completes an M-augmenting path; aug-
ment the matching. Because the search always reaches X along edges of M and
reaches Y along edges not in M , every edge is explored at most once, and record-
ing the edge used to reach each vertex permits retrieval of the augmenting path.
If no vertex of Y missed by M is reached, then when all vertices in S have
been explored the set (X − S) ∪ T forms a vertex cover of size | M |. This vertex
cover proves that M is a maximum matching in G.

U S X −S
X • • • • • • • • •

Y • • • • • • • • •
T

The augmenting path characterization of maximum matchings (Theorem

6.3.4) holds also for nonbipartite graphs, but odd cycles make the exploration of
alternating paths more difficult. Edmonds [1965c] overcame this difficulty using
the “Blossom Algorithm” in his famous paper “Paths, Trees, and Flowers”.

6.3.6. Example. An M-alternating path can travel in either direction around

an odd cycle in order to continue onward from any of its vertices, as suggested
below with M in bold. A search for shortest M-augmenting paths from u reaches
x via the edge ax outside M. If we do not also consider a longer path reaching x
via dx in M , then we miss the augmenting path u , v, a , b , c , d , x , y. To postpone
deciding which route to take, the algorithm shrinks such an odd cycle to a single
vertex before continuing to seek an augmenting path.

y x d y
• • • •
• •
• • •a → • • •
u v u v A
•w •w
• •
b c

The running time of the algorithm is cubic in the number of vertices. With
another approach, describing a polyhedron whose vertices correspond to match-
ings, Edmonds [1965d] extended the result to weighted graphs.
Section 6.3: Algorithmic Aspects 279

WEIGHTED BIPARTITE MATCHING

Like the shortest path problem (Section 5.4), maximum matching generalizes
to weighted graphs. We assign weights to edges and seek the matching with max-
imum total weight. The maximum matching problem is the special case where
edges of G have weight 1 and edges not in G have weight 0. The dual notion of
vertex cover also extends. For bipartite graphs, we present a common solution to
the weighted matching and cover problems using augmenting paths.
When | X | = | Y | = n, the problem reduces to seeking a maximum-weight per-
fect matching in K n ,n with nonnegative edge weights. If G
= K n ,n , then we can
insert the missing edges with weight 0 without changing the answer; hence we
may assume G = K n ,n . We may also assume that all weights are nonnegative,
by changing negative weights to 0, solving the resulting problem, and deleting
chosen edges with weight 0.

6.3.7. Example. Weighted bipartite matching and its dual. A farming company
has n farms and n stores. Each farm can produce corn to supply one store. The
profit from sending the output of farm i to store j is wi , j . This yields a weighted
X , Y -bigraph with X = {x1 , . . . , x n} and Y = {y1 , . . . , yn}; the weight on edge
x i yj is wi , j . The company seeks the matching with maximum total weight.
The government offers to pay subsidies to reduce corn production. The gov-
ernment will pay ui if the company doesn’t use farm i and vj if it doesn’t use store
j . If ui + vj < wi , j , then the company makes more by using the edge x i yj than by
accepting the payments ui and vj . In order to stop all production, the government
must offer amounts such that ui + vj ≥ wi , j for all (i , j). The government wants
to find such values for {ui} and {vj } to minimize the cost ∑ ui + ∑ vj .

6.3.8. Definition. A transversal of an n-by-n matrix consists of n positions,

one in each row and each column. Finding a transversal with maximum sum
is the Assignment Problem. This is the matrix formulation of Maximum
Weighted Matching, where nonnegative weight wi , j is assigned to edge x i yj
of K n ,n and we seek a perfect matching M to maximize the total weight w(M).
With these weights, a (weighted) cover is a choice of labels u1 , . . . , un
and v1 , . . . , vn such that ui + vj ≥ wi , j for all i , j . The cost c(u , v) of a cover
(u , v) is ∑ ui + ∑ vj . The minimum weighted cover problem is that of find-
ing a cover of minimum cost.

A minimization and a maximization problem are dual problems when the

value of every solution to the minimization problem is at least as large as the
value of every solution to the maximization problem. Thus a solution to either
problem establishes a bound for the other. (See Section 8.3 and Section 11.3 for
further discussion of duality.) The next lemma shows that weighted matching
and weighted cover are dual problems.

6.3.9. Lemma. (Weighted matching /cover duality) If M is a perfect match-

ing in a weighted bipartite graph G and (u , v) is a cover, then c(u , v) ≥ w(M).
Also, c(u , v) = w(M) if and only if M consists of edges x i yj with ui + vj = wi , j ,
and then M and (u , v) are optimal.
280 Chapter 6: Matchings

Proof: Since M is a perfect matching, summing the constraints ui + vj ≥ wi , j that

arise from its edges yields c(u , v) ≥ w(M) for every cover (u , v). Furthermore, if
c(u , v) = w(M), then equality must hold in each of the n inequalities summed.
Finally, since c(u , v) ≥ w(M) for every matching and every cover, c(u , v) = w(M)
implies that there is no matching with weight greater than c(u , v) and no cover
with cost less than w(M).

The observation that a matching and cover can have equal weight and cost
only by using edges covered with equality leads to an algorithm.

6.3.10. Definition. The equality subgraph G u ,v for the cover (u , v) is the span-
ning subgraph of K n ,n whose edges are {x i yj : ui + vj = wi , j }. In the cover
(u , v), the excess e i , j for the pair (i , j) is ui + vj − wi , j .

If G u ,v has a perfect matching, then its weight is ∑ ui + ∑ vj , and by Lemma

6.3.9 we have the optimal matching and cover. Otherwise, we use the equality
subgraph and matrix of excesses to change the cover (u , v). The resulting algo-
rithm was named the “Hungarian Algorithm” by Kuhn in honor of the work of
K önig and Egerváry on which it is based.

6.3.11. Algorithm. (Hungarian Algorithm; Kuhn [1955], Munkres [1957]).

Let X and Y be the parts in the bipartition of K n ,n . Given a matrix of weights on
E(K n ,n) and a weighted cover (u , v) as in Definition 6.3.8, we first find a match-
ing M and a vertex cover Q of the same size in the equality subgraph G u ,v (by any
method, such as the Augmenting Path Algorithm). Let R = Q ∩ X and T = Q ∩ Y .
Our matching of size | Q| consists of | R| edges from R to Y − T and | T | edges from
T to X − R, as shown below.
−
U S R
X • • • • • • • • •

Y • • • • • • • • •
T
+
To seek an equality subgraph having a larger matching, we modify (u , v) to
introduce an edge from X − R to Y − T while maintaining equality on all edges of
M. Edges joining X − R and Y − T are not in G u ,v and thus have positive excess.
Let be the minimum among these; that is,
= min{e i , j : x i ∈ X − R , yj ∈ Y − T}.
Reducing ui by for all x i ∈ X − R maintains the cover condition for these edges
while bringing at least one into the equality subgraph. To maintain the cover
condition for the edges from X − R to T , increase vj by for y j ∈ T , as shown in
the figure above. Recompute the equality subgraph.
The algorithm proceeds by iteratively adjusting the cover to obtain a cover
whose equality subgraph contains a perfect matching.
Section 6.3: Algorithmic Aspects 281

We explained the adjustment step using the equality subgraph, but redraw-
ing a changing equality subgraph is awkward. More efficient is using the matrix
of excesses, called the excess matrix for the cover.

6.3.12. Example. Solving the Assignment Problem. The first matrix below lists
the weights (wi , j ). The others display a cover (u , v) and the corresponding excess
matrix. The 0s in an excess matrix correspond to edges in the equality subgraph
G u ,v ; an independent family of 0s (no two in a row or column) corresponds to a
matching in G u ,v . We underscore independent 0s to mark a maximum matching,
shown in bold in the equality subgraphs for the first two excess matrices.
A set of rows and columns covering the 0s in the excess matrix is a covering
set; these form a vertex cover in G u ,v . A covering set of size less than n yields
progress, since the next weighted cover costs less. For the 0s in the excess matrix,
we find an independent family and a covering set of the same size. In a small
matrix, we can do this by inspection.
0 0 0 0 0
⎛6 3 4 3 4⎞ 6 ⎛0 3 2 3 2⎞ R
8 ⎜3 0 2 2⎟ X • • • • •
⎜5 8 6 4 6⎟ 4
⎜ ⎟ ⎜ ⎟
⎜6
⎜ 2 3 4 1⎟
⎟
→ 6 ⎜0 4 3
⎜
2 5⎟
⎟
⎜3 7 6 5 0⎟ 7 ⎜4 0 1 2 7⎟
Y • • • • •
⎝4 5 8 2 3⎠ 8 ⎝4 3 0 6 5⎠R T T
T T
We use “ R” and “ T ” to label the rows and columns of the covering set and the
corresponding vertex cover in G u ,v . At each iteration, we compute the minimum
excess on the positions that are not in a covered row or column (in rows of X − R
and columns of Y − T). These uncovered positions have positive excess; that is,
the corresponding edges are not in the equality subgraph. As suggested earlier,
let be the minimum of these excesses. We reduce the label ui by on rows not
in R and increase the label vj by on columns in T . The outlined positions in the
diagram are the ones that change to reach the next matrix.
1 1 0 0 0 2 2 1 0 0
5⎛0 3 1 2 1⎞ 4⎛0 3 1 1 0⎞
7⎜3 0 1 3 1⎟ X• • • • • 6⎜3 0 1 2 0⎟
⎜ ⎟ ⎜ ⎟
→ 5⎜
⎜
0 4 2 1 4⎟
⎟ → 4⎜
⎜
0 4 2 0 3⎟
⎟
6⎜4 0 0 1 6⎟ 5⎜4 0 0 0 5⎟
Y • • • • •
8⎝5 4 0 6 5⎠ T T T 7⎝5 4 0 5 4⎠
T T T
In the example given, the first iteration reduces the cost of the cover but does
not increase the size of the maximum matching in the equality subgraph. The
second iteration produces a perfect matching. Using the first three columns as a
covering set in the first iteration would augment the matching immediately.
The transversal of 0s at the end identifies a perfect matching whose total
weight equals the cost of the final cover. Its edges have weights 6 , 8 , 8 , 5 , 4 in the
original data; they sum to 31. The labels 4 , 6 , 4 , 5 , 7 and 2 , 2 , 1 , 0 , 0 in the final
cover satisfy these edges exactly and also sum to 31. The value of the optimal so-
lution is unique, but the solution itself is not; this example has many maximum
weight matchings and many minimum cost covers, all summing to 31.
282 Chapter 6: Matchings

6.3.13. Theorem. The Hungarian Algorithm finds a maximum weight matching

and a minimum cost cover.
Proof: The algorithm begins with a cover. It can terminate only when the equal-
ity subgraph has a perfect matching, which guarantees equal value for the cur-
rent matching and cover. Suppose that (u , v) is the current cover and that the
equality subgraph has no perfect matching. Let (u , v ) denote the new lists of
numbers assigned to the vertices. Because is the minimum of a nonempty finite
set of positive numbers, > 0.
We verify first that (u , v ) is a cover. The change of labels on vertices of X − R
and T yields u i + v j = ui + vj for edges x i yj from X − R to T or from R to Y − T .
If x i ∈ R and yj ∈ T , then u i + v j = ui + vj + , and the weight remains covered. If
x i ∈ X − R and yj ∈ Y − T , then u i + v j equals ui + vj − , which by the choice of
is at least wi , j .
The algorithm terminates only when the equality subgraph has a perfect
matching, so it suffices to prove termination. If the weights wi , j are rational,
then multiplying by their least common denominator yields an equivalent prob-
lem with integer weights, where the changes made in labels are always integers.
By the K önig–Egerváry Theorem, | T | < | X − R| when the equality subgraph has
no perfect matching. The changes in labels reduce the cost of the cover by a mul-
tiple of | X − R| − | T | , which is an integer. Since the cost is bounded below by the
weight of a perfect matching, after finitely many iterations we have equality.
For real-valued weights, see the next remark.

6.3.14. Remark. The algorithm also works with real-valued weights if we obtain
covers more carefully. Because M remains in the equality subgraph, the size of
the matching never decreases. Since it can increase at most n times, it suffices to
show that it must increase within n iterations.
If we find M using the Augmenting Path Algorithm, then the last iteration
produces a vertex cover by exploring M-alternating paths from the subset U of
uncovered vertices in X . With S and T being the sets of vertices reached in X
and Y , the vertex cover is R ∪ T , where R = X − S.
A step of the Hungarian Algorithm using this cover R ∪ T maintains equality
on M and all edges in M-alternating paths from U. Edges from T to R disappear
from the equality subgraph, but they don’t appear in M-alternating paths from
U. Introducing an edge from S to Y − T either creates an M-augmenting path or
increases T while leaving U unchanged. Since T can increase at most n times,
we obtain a larger matching in the equality subgraph within n iterations.

When running the Hungarian Algorithm with all inputs wi , j ∈ {0 , 1}, with
initial cover assigning 1 for each ui and 0 for each vj , the excesses always remain
0 or 1. Hence the values in the weighted cover are all integers. The algorithm
produces a minimum weighted cover, but it does not use integers larger than 1
(using nonnegative cover values). Thus in the special case of maximum matching,
the dual cover problem reduces to finding Î(G).
We have not discussed weighted matching in general graphs. Edmonds
[1965d] found an algorithm for this, which was implemented in time O(n3) by
Gabow [1975] and by Lawler [1976]. Faster algorithms appear in Gabow [1990]
and in Gabow–Tarjan [1989].
Section 6.3: Algorithmic Aspects 283

Weighted bipartite matching is an important special case of the general prob-

lem of optimizing a linear function of several variables under linear constraints
and requiring variables to have integer values (like “use an edge or don’t ”). Lin-
ear optimization problems always have natural dual problems, and when the inte-
grality constraints are dropped the optimal values of the two problems are always
equal. We refer the reader to the many textbooks on Linear Programming.

FAST BIPARTITE MATCHING (optional)

In a bipartite graph with n vertices and m edges, exploring all edges for each
augmentation could take time O(nm) to find a maximum matching, since (G)
may be linear in n. By looking first for√short augmenting paths, Hopcroft–K arp
[1973] reduced the running time to O( nm).
The idea is to explore M-alternating paths simultaneously from all M-
uncovered vertices of X , seeking disjoint M-augmenting paths of the same length
with one examination of the edge set. Subsequent augmentations must use longer
paths, so the procedure splits into phases finding paths of the same length. The
key is that not many phases are needed. The first remark follows from the proof
of Theorem 6.3.4.

6.3.15. Remark. If M and M are matchings, with | M | = s > r = | M | , then

M M ∗ contains at least s − r disjoint M-augmenting paths.

We next show that successive shortest augmenting paths get longer.

6.3.16. Lemma. If P is a shortest M-augmenting path and P is M P-augmenting,

then | P | ≥ | P | + 2 | P ∩ P | (treating paths as edge sets).
Proof: The edge set M P is the matching obtained by using P to augment M.
Let N be the matching given by using P to augment M P. Since | N | = | M | +
2, Remark 6.3.15 guarantees disjoint M-augmenting paths P1 and P2 in M N .
Each is at least as long as P , since P is a shortest M-augmentating path.
Since N arises from M by switching on P and then switching on P , an edge
belongs to exactly one of M and N if and only if it belongs to exactly one of P and
P . That is, P P = M N . Hence | P P | = | M N | ≥ | P1 | + | P2 | ≥ 2 | P | , and
2 | P | ≤ | P P | = | P | + | P | − 2 | P ∩ P | .
We conclude | P | ≥ | P | + 2 | P ∩ P |.

6.3.17. Lemma. Augmenting paths of the same length in a list of successive

shortest augmentations share no vertices.
Proof: Let P1 , . . . , Pt be successive shortest augmenting paths starting from
some matching, in which P¾ and P l are two closest paths of the same length that
are not disjoint. Let M be the matching that results from using the augmenting
paths P1 , . . . , P¾ . Thus every vertex of P¾ is covered by an edge of M in P¾ .
By the choice of P¾ and P l , no two of P¾+1 , . . . , P l share a vertex. Hence P l
is an M-augmenting path. Since every vertex of P¾ belongs to an edge of P¾ ∩ M ,
284 Chapter 6: Matchings

the M-augmenting path P l cannot end at a vertex of P¾ , and it cannot have an

internal vertex on P¾ without sharing an edge of M with P¾ . However, by Lemma
6.3.16 P¾ and P l share no edge.
The contradiction implies that there is no such pair (¾ , l).

6.3.18.√Theorem. (Hopcroft–K arp [1973]) Maximum matching can run in

O( nm) time on bipartite graphs with n vertices and m edges.
Proof: Lemmas 6.3.16–6.3.17 imply that searching simultaneously for shortest
M-augmenting paths from all vertices of X not covered by M yields disjoint paths,
and further augmenting paths are longer. The augmentations of one length are
thus found by exploring the edges at most once.
List the resulting paths as P1 , . . . , P s in order by length, with s = (G).
Each Pi+1 is an augmenting path for the matching M √i formed by using P1 , . . . , Pi .
It suffices to prove that these paths have at most 2 s distinct lengths.
√
Let r = s − ⌈ s ⌉ . Because | Mr | = r and the maximum matching has size
s, Remark 6.3.15 yields s − r disjoint Mr -augmenting paths. The shortest of
these uses at most s−r r edges from Mr . Hence Pr+1 has at most 2 s−r r + 1 edges,
and each of P1 , . . . , Pr has no more than that. Since augmenting paths have √ odd
length, P1 , . . . , Pr+1 have at most s−r r + 1 lengths. Even if the remaining ⌈ s ⌉ − 1
√
paths have distinct lengths, altogether we use at most s−r r + ⌈ s ⌉ phases. Since
√ √
s− r < ⌈ s ⌉ ≤ s, we have the desired upper bound of at most 2⌈ s ⌉ phases.
r √s

√ in time O(n ). There

3
On general graphs, Edmonds’ Blossom Algorithm runs
is a complicated faster algorithm with running time O( nm) (Micali–Vazirani
[1980]); the proof of correctness is in Vazirani [1994].

STABLE MATCHINGS (optional)

Instead of total weight, we may consider other criteria for matchings. Given
n men and n women, we want to establish n “stable” marriages. If man x and
woman a are paired with other partners, but x prefers a to his current partner
and a prefers x to her current partner, then they might leave their current part-
ners and switch to each other. In this situation we say that the unmatched pair
(x , a) is an unstable pair.

6.3.19. Definition. A perfect matching is a stable matching if it yields no un-

stable unmatched pair.

6.3.20. Example. Given men x , y ,  , w, women a , b , c , d, and preferences listed

below, the matching {xa , yb , d , wc} is stable.
Men {x , y ,  , w} Women {a , b , c , d}
x: a>b>c>d a: > x> y>w
y : a>c>b>d b: y>w> x> 
: c>d>a>b c: w> x> y> 
w: c>b>a>d d: x> y> > w
Section 6.3: Algorithmic Aspects 285

In their paper “College admissions and the stability of marriage”, Gale and
Shapley proved that a stable matching always exists and can be found using a
relatively simple algorithm.

6.3.21. Algorithm. (Gale–Shapley Proposal Algorithm) Given preference

rankings by each of n men and n women, iteratively perform the following. Each
man proposes to the highest woman on his preference list who has not previously
rejected him. If each woman receives exactly one proposal, then stop and use the
resulting matching. Otherwise, every woman receiving more than one proposal
rejects all except the one that is highest on her preference list. Every woman
receiving a proposal says “maybe” to the most preferred proposal received.

6.3.22. Theorem. (Gale–Shapley [1962]) The Proposal Algorithm produces a

stable matching.
Proof: Key Observation: the sequence of proposals made by each man is non-
increasing in his preference list, and the sequence of men to whom a woman
says “maybe” is nondecreasing in her preference list, culminating in the man
assigned. This holds because men propose repeatedly to the same woman until
rejected, and women say “maybe” to the same man until a better offer arrives.
Once a woman has a proposal, she keeps receiving proposals and giving one
“maybe” answer. Hence if n − 1 women have rejected man x, then they are all
holding proposals from n − 1 distinct men who propose to them again on the next
round. When x proposes to the nth woman, a matching is completed.
Hence the algorithm can only terminate with a matching. It does terminate,
since at each iteration the total length of the remaining proposal lists decreases.
If the result is not stable, then some unstable unmatched pair (x , a) has x
matched to b and y matched to a. By the key observation, x never proposed to
a, since a received a mate less desirable than x. The key observation also im-
plies that x would not have proposed to b without earlier proposing to a. This
contradiction confirms the stability of the result.

The proposal algorithm is asymmetric; which sex is happier? When the first
choices of the men are distinct, they all get their first choice, and the women re-
ceive whoever proposed. In Example 6.3.20, when women propose the algorithm
immediately yields the matching {xd , yb , a , wc}; all women receive their first
choices. In fact, among all stable matchings, every man is happiest in the one
produced by the male-proposal algorithm, and every woman is happiest under the
female-proposal algorithm (Exercise 24). Societal conventions thus favor men.
The algorithm is used in another setting. Each year, the graduates of medical
schools submit preference lists of hospitals where they want to be residents. The
hospitals also have preferences; a hospital with multiple openings acts as several
hospitals with the same preference list. Chaos in the market for residents (then
called interns) forced hospitals to devise and implement the algorithm ten years
before the Gale–Shapley paper defined and solved the problem! The result was
the National Resident Matching Program, a non-profit corporation established
in 1952 to provide a uniform appointment date and matching procedure.
Since the medical organizations ran the algorithm, it is not surprising that
initially they did the proposing and were happier with the outcome. This is even
286 Chapter 6: Matchings

clearer in another setting; students applying for jobs have preferences, but the
employers make the proposals, called “ job offers”. Unhappiness with the NRMP
caused the system to be changed in 1998 to a student-proposing algorithm. In
1998 the system processed 35,823 applicants for 22,451 positions. Additional de-
tails are at nrmp.aamc.org /nrmp/mainguid/ on the World Wide Web.
There may be stable matchings other than those found by the two versions
of the proposal algorithm. To seek a “fair ” stable matching, we could give each
person a number of points with which to rate preferences. The weight for the pair
xa is then the sum of the points that x gives to a and a gives to x. The Hungarian
Algorithm would yield a matching of maximum total weight, but this might not
be a stable matching (Exercise 23). Other approaches appear in the books Knuth
[1976b] and Gusfield–Irving [1989] on stable marriages and related topics.

EXERCISES 6.3

6.3.1. (−) Given a graph G, suppose that S ⊆ V(G) is covered by some matching. Prove or
disprove each statement below:
(a) S is covered by some maximum matching.
(b) S is covered by every maximum matching.

6.3.2. (−) Use the symmetric difference operation to give an alternative proof that if x
and y are vertices in a tree G, then G has exactly one x , y-path.
6.3.3. (−) Using nonnegative edge weights, construct a 4-vertex weighted graph in which
the matching of maximum weight is not a matching of maximum size.
6.3.4. (−) Show how to use the Hungarian Algorithm to test for the existence of a perfect
matching in a bipartite graph.
6.3.5. (−) Give an example of the stable matching problem with two men and two women
in which there is more than one stable matching.
6.3.6. (−) Determine the stable matchings resulting from the Proposal Algorithm with
men proposing and with women proposing, given the preference lists below.
Men {u , v , w , x , y , } Women {a , b , c , d , e ,  }
u: a>b>d>c>>e a: >x> y>u>v>w
v: a>b>c>> e> d b: y> >w>x>v>u
w: c>b>d>a>>e c: v> x>w> y>u>
x: c>a>d>b>e> d: w> y>u> x> >v
y : c> d>a>b>> e e: u>v> x>w> y>
: d>e>>c>b>a  : u>w> x>v> > y

6.3.7. (♦) Let G be a graph with m edges and maximum degree  .

(a) Let M be a maximum matching in G. For  ≥ 3, prove that the number of edges
joining vertices covered by M to vertices not covered by M is at most ( − 1) | M |. Use this
to prove  (G) ≥ 2m/(3 − 1). (Feng [2009])
(b) Let M be a maximal matching in G. Prove that M has size at least m/(2  − 1). For
each  , construct infinitely many -regular graphs having some maximal matching with
m/(2  − 1) edges (thereby proving that the inequality is sharp). (Biedl–Demaine–Duncan–
Fleischer–Kobourov [2004])
Exercises for Section 6.3 287

6.3.8. (♦) Two people play a game on a graph G by alternately selecting vertices v1 , v2 , · · ·
such that for i > 1, vertex vi is adjacent to vi−1 and has not previously been chosen. The last
player able to select a vertex wins. Prove that the second player has a winning strategy
if G has a perfect matching and that the first player has a winning strategy if G has no
perfect matching.

6.3.9. Prove that every nontrivial tree has a maximum matching that covers every non-
leaf vertex plus at least one leaf. Determine which trees have a matching covering at least
two leaves. (Lih–Lin–Tong [2006])

6.3.10. Let G be an X , Y -bigraph with | X | ≤ | Y | . Prove that G has a maximal path with
an endpoint in Y . (K ündgen–Ramamurthi [2002])

6.3.11. Let M and M be matchings in an X , Y -bigraph G such that S ⊆ X is covered by

M and T ⊆ Y is covered by M . Prove that G contains a matching that covers S ∪ T .

6.3.12. Let G be an X , Y -bigraph, not necessarily finite. Prove that if G has a matching
covering X and a matching covering Y , then G has a perfect matching. (J. Zaks)


6.3.13. (♦) Given a function º : V(G) → 0 on a graph G, an º -matching in G is a subset
M of E(G) having at most º (v) edges incident to each vertex v. An M-augmenting trail is
a trail that alternates between edges outside and in M , beginning and ending with edges
outside M at vertices having “excess capacity”. Use the transformation in Example 6.2.15
to prove that an º -matching M in G is a maximum-sized º -matching if and only if G has
no M-augmenting trail. (Gondran–Minoux [1984])

6.3.14. (+) Let G be a connected X , Y -bigraph with girth at least 2s. Prove that if the
distance between any two vertices of X having degree 1 is at least 2s, then every subset of
X with size at most s is covered by some matching in G. (Hint: Consider a smallest subset
of X that cannot be covered.) (Horák–Tuza)

6.3.15. Let u be a vertex missed by a matching M . Prove that if no M-augmenting path

starts at u, then u is missed by some maximum matching.

6.3.16. (♦) Proof of König–Egerváry Theorem by augmenting paths. Let M be a matching

in an X , Y -bigraph G. Let U be the subset of X missed by M. Let W be the set of vertices
reachable from U by M-alternating paths (this includes U). Let S = W ∩ X and T = W ∩ Y .
Prove that if G has no M-augmenting path, then T ∪ (X − S) is a vertex cover of size | M |.

6.3.17. (♦) Find a transversal of maximum total sum (weight) in each matrix below. Prove
that there is no larger weight transversal by exhibiting a solution to the dual problem.
Explain why this proves that there is no larger transversal.

(a) (b) (c)

44436 78987 1 2345
11434 87676 67872
14535 96546 13445
56479 85764 36287
53683 76555 41354

6.3.18. Given a connected graph G with nonnegative edge weights, the Chinese Post-
man Problem (posed in Guan [1962]) is the problem of finding the shortest closed walk
that includes all the edges. Reduce this problem to a problem of finding a perfect matching
of minimum total weight in a complete graph with weights on the edges.
288 Chapter 6: Matchings

6.3.19. Find a minimum-weight transversal in the matrix below, and use duality to prove
that the solution is optimal. (Hint: Use a transformation of the problem.)

⎛4 5 8 10 11 ⎞
⎜7 6 5 7 4 ⎟
⎜8 6 ⎟
⎜ 5 12 9 ⎟
⎜6 6 13 10 7 ⎟
⎝4 5 7 9 8 ⎠

6.3.20. (♦) The Bus Driver Problem. Suppose that bus drivers are paid overtime for the
time by which their routes in a day exceed t. Let there be n bus drivers, n morning routes
with durations x1 , . . . , x n , and n afternoon routes with durations y1 , . . . , yn . The objec-
tive is to assign one morning route and one afternoon route to each driver to minimize the
total overtime. Express this as a weighted matching problem. Prove that overtime is min-
imized by giving the ith longest morning route and ith shortest afternoon route to the
same driver, for each i. (R.B. Potts)
6.3.21. Let the entries in matrix A have the form wi , j = ai bj , where a1 , . . . , an are num-
bers associated with the rows and b1 , . . . , bn are numbers associated with the columns.
Determine the maximum weight of a transversal of A. What happens when wi , j = ai + bj ?
(Hint: In each case, guess the general pattern by examining the solution when n = 2.)
6.3.22. (♦) A mathematics department offers ¾ seminars in different topics to its n stu-
dents. Each student will take one seminar; the ith seminar will have ¾i students, where
∑ ¾i = n. Each student submits a preference list ranking the ¾ seminars. An assignment
of the students to seminars is stable if no two students can both obtain more preferable
seminars by switching their assignments. Show how to find a stable assignment using
weighted bipartite matching. (Isaak)
6.3.23. Consider n men and n women, each assigning n − i points to the ith person in his or
her preference list. Let the weight of a pair be the sum of the points assigned by those two
people. Construct an example where no maximum weight matching is a stable matching.
6.3.24. (♦) Prove that if man x is paired with woman a in some stable matching, then
a never rejects x in the Proposal Algorithm with men proposing. Conclude that among
all stable matchings, every man is happiest in the matching produced by this algorithm.
(Hint: Consider the first occurrence of such a rejection.)
6.3.25. In the Stable Roommates Problem, each of 2n people has a preference ordering on
the other 2n − 1. A stable matching is a perfect matching such that no unmatched pair
prefers each other to their current roommates. Prove that the preferences below do not
permit a stable matching. (Gale–Shapley [1962])
a: b > c > d b: c > a > d c: a > b > d d: a > b > c

6.3.26. In the Stable Roommates Problem, let each individual declare a top portion of
the preference list as “acceptable”. Define the acceptability graph to be the graph whose
vertices are the people and whose edges are the pairs who rank each other as acceptable.
Prove that all sets of rankings with acceptability graph G lead to a stable matching if and
only if G is bipartite. (Abeledo–Isaak [1991]).
Chapter 7

Connectivity and Cycles

Every connected graph contains a spanning tree, but a tree is just barely con-
nected; deleting any edge or non-leaf vertex disconnects it. In a communication
network, we want to preserve service by ensuring that the graph (or digraph) of
possible transmissions remains connected even when some vertices or edges fail.

7.1. Connectivity Parameters

How difficult is it to disconnect a graph? We consider first deletion of vertices
and then deletion of edges.

SEPARATING SETS

7.1.1. Definition. A separating set or vertex cut of a graph G is a set S ⊆ V(G)

such that G − S has more than one component. A graph G is ¾-connected if
it has more than ¾ vertices and every vertex cut has size at least ¾ . The con-
nectivity of G, written (G), is the maximum  such that G is -connected.

A non-complete graph G has a separating set, and (G) is the minimum size
of such a set. In this case, the neighborhood of a vertex of minimum degree is
a separating set, and hence (G) ≤ (G). Also, if a graph G is -connected, and
 > 0, then G is also ( − 1)-connected, by the definition.
7.1.2. Example. Connectivity of K n and K r,s . A complete graph has no separat-
ing set. By requiring -connected graphs to have more than  vertices, we obtain
(K n) = n − 1. This allows general connectivity results (such as (G) ≤ (G)) to
hold also for K n . A graph with more than two vertices has connectivity 1 if it is
connected and has a cut-vertex. A graph with more than one vertex has connec-
tivity 0 if and only if it is disconnected. Unfortunately, K 1 is an anomaly; it is
connected but has connectivity 0.
Every induced subgraph of K r,s having at least one vertex from each part is
connected. Hence every vertex cut contains a full part, and (K r,s) = min{r, s}
(the convention for K 2 is consistent with this).

289
290 Chapter 7: Connectivity and Cycles

Since (G) is an upper bound on (G), it is natural to ask when equality holds.
Such graphs have been called maximally connected. Our first family with this
property also solves the extremal problem for the minimum number of edges in a
-connected graph G with n vertices. Since (G) ≥  , the Degree-Sum Formula
requires at least ⌈  n/2⌉ edges; this is achieveable whenever n >  .

7.1.3. Example. The  -connected Harary graph H¾ ,n . Given 2 ≤  < n, place

vertices 1 , . . . , n around a circle, equally spaced. Make each vertex adjacent to
the nearest ⌊ /2⌋ vertices in each direction around the circle. If  is odd, then
also add the edges from i to i + ⌊ n/2⌋ for 1 ≤ i ≤ ⌈ n/2⌉ .
When  n is even, H¾ ,n is -regular. When  n is odd, vertex (n+ 1)/2 has degree
 + 1 and the rest have degree  . See H4 ,8 , H5 ,8 , and H5 ,9 below.

5 5
• • 4 •
• • 4
• •6 • •6
3• •7
• • 3• •7
• •8
• • • •8 2
• 2
• •
1
•9
1

The first general method for proving a graph -connected is the definition:
consider a separating set S and prove | S| ≥  . For the Harary graphs, we will use
another tool for the last case; we postpone it until after the main proof.

7.1.4. Theorem. (Harary [1962a]) (H¾ ,n) =  , and hence the minimum number
of edges in a -connected graph on n vertices is ⌈  n/2⌉ .
Proof: Let r = ⌊ /2⌋ and G = H¾ ,n . Since (G) =  , it suffices to prove (G) ≥  .
Let S be a separating set; we prove | S| ≥  .
Choose x , y ∈ V(G) − S. Deleting x and y from the circular arrangement of
vertices leaves two maximal segments A and B of consecutive vertices. In G − S,
we have the potential for traveling from x to y in a clockwise or a counterclockwise
direction, through A or B.
Since each vertex is adjacent to the next r vertices in each direction, there is
an x , y-path in G − S via A or B unless S contains r consecutive vertices both in
A and in B. Thus | S| ≥  unless  is odd and S consists exactly of r consecutive
vertices in A and r consecutive vertices in B.
In this case, when n is even we find an x , y-path in G − S using the diagonal
edge at x or y. Let x and y be the neighbors of x and y along these edges, respec-
tively. Label A and B so | A| ≥ | B| ; now x , y ∈ A (see figure below). Since S has
r vertices between x and y in B, there are also r vertices between x and y in A.
Therefore, deleting r consecutive vertices in A leaves intact the x , y-path in A
or the x , y -path in A; adding xx or yy completes an x , y-path in G − S.
For odd n, the lack of rotational symmetry leads to annoying technical details
in such arguments. Instead, note that H¾ ,n is obtained from H¾ ,n−1 by using the
operation in Definition 7.1.5, which by Lemma 7.1.6 preserves -connectedness.
Section 7.1: Connectivity Parameters 291

The new vertices 1 and n each have degree ¾ , and each vertex that was adjacent
to the split vertex now has a neighbor in {1 , n}.
y
•
•x

•
x •
y
7.1.5. Definition. A vertex ¾-split forms a graph H from a graph G by delet-
ing one vertex x and replacing it with adjacent vertices x1 and x2 such that
NH (x1) ∪ NH (x2) = NG(x) ∪ {x1 , x2 } and d H (x i) ≥ ¾ .

7.1.6. Lemma. If a graph H arises from a ¾-connected graph G by a vertex ¾-

split, then H is ¾-connected.
Proof: Form H by splitting x into x1 and x 2 as in Definition 7.1.5. Let S be a
separating set of H , and let X = {x1 , x2 }. If S ∩ X = ∅, then also S separates G,
so | S| ≥ ¾ . If X ⊆ S, then H − S = G − T , where T = (S − X) ∪ {x}, so | S| > | T | ≥ ¾ .
The remaining case is | S ∩ X | = 1, say x1 ∈ S. Let T = S − {x1 } ∪ {x}. If
T separates G, then | S| = | T | ≥ ¾ , so assume G − T is connected. Now H − S
is obtained from G − T by adding x 2 and the edges joining x 2 to its neighbors.
Hence H − S is connected unless S contains NH (x2), again yielding | S| ≥ ¾ .

• •x1
G • •x H
• •x2

Theorem 7.1.4 implies that ⌈ ¾ n/2⌉ edges and minimum degree ¾ are the
smallest values that allow an n-vertex graph to be ¾-connected. The smallest
value of the minimum degree that forces an n-vertex graph to be ¾-connected (for
1 ≤ ¾ < n) is much larger: ⌈(n + ¾ − 2)/2⌉ (Exercise 17).
Next we show that hypercubes are also maximally connected.

7.1.7. Example. (Q¾) =  . The hypercube Q¾ is -regular, so (Q¾) ≤  . We

prove the lower bound by induction on  . Note that (Q¾) =  for  ≤ 1, since Q0
and Q1 are complete graphs.
For  ≥ 2, express Q¾ as two copies of Q¾−1 (called G 1 and G 2) plus a match-
ing joining corresponding vertices (each vertex of Q¾−1 becomes two vertices in Q¾
joined by an edge). By the induction hypothesis, G 1 and G 2 are ( − 1)-connected.

G1 G2
292 Chapter 7: Connectivity and Cycles

Let S be a separating set in Q¾ . If S disconnects G 1 or G 2 , then S must have

at least ¾ − 1 vertices in that copy of Q¾−1 . Also S must have at least one vertex
in the other copy, since otherwise any two vertices of G − S would be connected
by a path through the other copy. On the other hand, if G 1 − S and G 2 − S both
are connected, then S must contain at least one vertex from every edge joining
G 1 and G 2 to avoid G − S being connected; this yields | S| ≥ | V(G 1)| = 2¾−1 ≥ ¾ .

The argument in Example 7.1.7 also proves that whenever G is ¾-connected,

the product G K 2 is (¾ + 1)-connected. The result generalizes further.

7.1.8. Theorem. (Niu–Zhu [1994], Chiue–Shieh [1999]) If G and H are con-

nected graphs, then (G H) ≥ (G) + (H).
Proof: The cartesian product G H (Definition 5.1.20) decomposes into a copy G v
of G for each v ∈ V(H) and a copy Hu of H for each u ∈ V(G). Let nG = | V(G)| and
nH = | V(H)| , and let S be a separating set of G H.
Case 1: S separates some G v or Hu . By symmetry, let G v − S be disconnected,
with (u , v) and (u , v) in different components. Thus | S ∩ V(G v)| ≥ (G). If Hu − S
or Hu − S is disconnected, then S also has (H) vertices in Hu or Hu , yielding
|S| ≥ (G) + (H). Otherwise, with Hu − S and Hu − S connected, we claim that S
contains a vertex of each G w with w
= v. If not, then we can travel from (u , v) to
(u , w) in Hu − S, from (u , w) to (u , w) in G w − S (since S ∩ V(G w) = ∅), and from
(u , w) to (u , v) in Hu − S, completing a walk from (u , v) to (u , v). We conclude
|S| ≥ (G) + nH − 1 ≥ (G) + (H).
Gw Gv G v Gw Gv

Hu • • Hu • •

Hu • • Hu • •

Case 1 Case 2

Case 2: S separates no G v or Hu . Choose (u , v) and (u , v ) in different com-

ponents of (G H) − S; note that u
= u and v
= v . For any w ∈ V(H) with
w∈ / {v , v }, we must have (u , w) or (u , w) in S, since otherwise we can travel from
(u , v) to (u , w) in Hu − S, from (u , w) to (u , w) in G w − S, and from (u , w) to (u , v )
in Hu − S, since S does not separate any copy of G or H. Similarly, S contains
(x , v) or (x , v ) for x ∈
/ {u , u }. Finally, S contains (u , v) and (u , v ). Thus
|S| ≥ (nG − 2) + (nH − 2) + 2 = (nG − 1) + (nH − 1) ≥ (G) + (H).

Equality holds when (G) = (G) and (H) = (G), since (G H) ≤ (G H) =
(G) + (H). However, (G H) can be large when (G) = (H) = 1 (Exercise
27). In fact, always (G H) = min{(G) + (H) , |V(G)| (H) , | V(H)| (G)}, stated
without proof by Liouville [1978] and proved by Špacapan [2008] (Exercise 28).
Section 7.1: Connectivity Parameters 293

EDGE CUTS

Edge deletion may be more relevant than vertex deletion. For example, per-
haps transmitters (vertices) are secure, but communication links (edges) can be
disrupted. In this setting, the redundancy of multiedges can be valuable.

7.1.9. Definition. A disconnecting set of edges in a multigraph G is a set F ⊆

E(G) such that G − F is disconnected; G is ¾-edge-connected if every discon-
necting set has size at least ¾ . The edge-connectivity (G) is max{ : G is
-edge-connected}. For S, T ⊆ V(G), write [S, T] for the set of edges joining S
to T . An edge cut is a set of the form [S, S], where ∅
= S ⊂ V(G). A minimal
set of edges whose deletion increases the number of components is a bond.

Deleting any edge cut of a graph G disconnects it: G − [S, S] has no path from
S to S. In fact, all minimal disconnecting sets of edges have this form.

7.1.10. Proposition. For a graph G, every minimal disconnecting set is an edge

cut. If G is connected, then an edge cut [S, S] is a bond if and only if the
induced subgraphs G[S] and G[S] are both connected.
Proof: If G − F is disconnected, then let T be the vertex set of a component of
G − F . The edge cut [T , T] is contained in F , so F is not a minimal disconnecting
set unless F = [T, T].
If G is connected, F is the edge cut [S, S], and both G[S] and G[S] are con-
nected, then no proper subset of [S, S] disconnects G, so F is a bond. On the other
hand, if G[S] is not connected, and T is the vertex set of a component of G[S],
then [T, T] is a proper subset of [S, S], as shown below, so F is not a bond.

T
S S

7.1.11. Example. Edge-connectivity of K n. Every edge cut in K n has size (n − )

for some  with 1 ≤  ≤ n − 1. The minimum occurs at  = 1, so (K n) = n − 1.

Deleting one endpoint of each edge in an edge cut F deletes every edge of
F . Hence we expect that (G) ≤ (G), but we must avoid leaving a connected
subgraph by deleting all isolated vertices of G − F .

7.1.12. Theorem. (Whitney [1932a]) (G) ≤ (G) ≤ (G).

Proof: The edges incident to a vertex v of minimum degree disconnect G; hence
(G) ≤ (G). To prove (G) ≤ (G), let  = (G), and let [S, S] be a smallest
edge cut. If all of S is adjacent to all of S, then  = | S| ####S#### ≥ | V(G)| − 1 ≥ (G).
# #
Otherwise, there exist x ∈ S and y ∈ S with xy ∈ / E(G). For each edge of
[S, S] incident to x, put the other endpoint into T . For other edges of [S, S],
put the endpoints in S into T , as shown below. Now | T | ≤  , and x and y lie in
different components of G − T . Hence (G) ≤ (G).
294 Chapter 7: Connectivity and Cycles

x•
•T
•
•T
S T• S
•
T•
•y
T•

When (G) < (G), no smallest edge cut isolates a vertex; in fact, both sides
of a smallest edge cut [S, S] must then be larger than (G). This follows from a
simple expression for the size of an edge cut.

7.1.13. Proposition. If S ⊆ V(G), then ####[S, S]#### = [∑v∈S d(v)] − 2 | E(G[S])|.

# #
Proof: The sum ∑v∈S d(v) counts edges in G[S] twice and edges in [S, S] once.
# #
7.1.14. Corollary. If G is a graph and ####[S, S]#### < (G) for some nonempty proper
subset S of V(G), then | S| > (G).
Proof: By Proposition 7.1.13, (G) > ∑v∈S d(v) − 2 | E(G[S])|. Since d(v) ≥ (G)
and 2 | E(G[S])| ≤ | S| (| S|− 1), we have (G) > | S| (G) −|S| (| S|− 1). This inequality
requires | S| > 1, so rearranging and canceling | S| − 1 yields | S| > (G).

We give a sufficient condition for equality in the trivial upper bound. Note
that diameter 2 does not imply (G) = (G); consider K 1 2K r .

7.1.15. Theorem. (Plesnı́k [1975]) If diam G = 2, then (G) = (G).

# # # #
Proof: If #### [S, S]#### < (G), then Corollary 7.1.14 yields | S| > (G) and #### S#### > (G).
If each vertex in S has a neighbor in S, then #### [S, S]#### ≥ | S| > (G). Otherwise,
# #
some x ∈ S has no neighbor in S, and similarly some vertex y ∈ S has no neighbor
in S. Now d(x , y) > 2, contradicting diam G = 2.

BLOCKS

Decomposition into components can reduce a problem to connected graphs.

There is a similar decomposition into connected pieces without cut-vertices.
7.1.16. Definition. A block of a graph G is a maximal connected subgraph of G
that has no cut-vertex.

• •
• •
• • •
• •
• •

7.1.17. Example. If H is a block of G, then H as a graph has no cut-vertex, but

H may contain cut-vertices of G. The graph above has five blocks, circled.
Exercises for Section 7.1 295

7.1.18. Remark. Properties of blocks. An edge in a cycle is not a block, since it

lies in a larger subgraph with no cut-vertex. An edge (with its endpoints) is a
block if and only if it is a cut-edge; the blocks of a tree are its edges. Any block
with more than two vertices is 2-connected. The blocks of a graph are its isolated
vertices, its cut-edges, and its maximal 2-connected subgraphs.
If blocks B and B share two vertices, then their union cannot be disconnected
by deleting one vertex. Thus B and B fail the maximality condition in Definition
7.1.16. Hence any two blocks in a graph share at most one vertex.
This implies that the blocks of a graph form a decomposition of it. A vertex
shared by two blocks must be a cut-vertex of G. The interaction between blocks
and cut-vertices is described by a special graph. The block-cutpoint graph of a
graph G is the bipartite graph H whose parts are the cut-vertices and the blocks
of G, with cut-vertex v adjacent to block B in H if and only if v ∈ B.
The block-cutpoint graph is a forest with a component for each component of
G (Exercise 34). The leaves of this graph correspond to blocks of G containing
only one cut-vertex of G; such blocks are leaf blocks.

EXERCISES 7.1

7.1.1. (−) Prove that if G has more than ¾ vertices and is not ¾-connected, then G has a
separating set of size exactly ¾ − 1.
7.1.2. (−) Prove that G is ¾-connected if and only if G K r is (¾ + r)-connected.
7.1.3. (−) For ¾ , l , m ∈
(G
 with ¾ ≤ l ≤ m, construct a graph G ¾ ,l ,m with (G ¾ ,l ,m) =  ,
¾ ,l ,m) = l, and (G ¾ ,l ,m = m. (Chartrand–Harary [1968])
)
7.1.4. (−) Let M be a matching of size r in K r,s , where r ≤ s. Prove that the graph K r,s − M
is (r − 1)-connected unless (r, s) = (2 , 2).
7.1.5. (−) A cactus is a connected graph whose blocks are all edges or cycles. Prove that
a connected graph is a cactus if and only if no two vertices are joined by three internally
disjoint paths. Conclude that a connected graph with no even cycle is a cactus.
7.1.6. (−) Determine the smallest 3-regular graph with connectivity 1.
7.1.7. (−) Construct a graph with degree list 5543333 that is 3-connected. Prove that it is
3-connected by using vertex splits.

7.1.8. (−) For n ,  ∈ with n −  odd and at least 3, show that the list with  − 1 copies
of n − 1 and n −  + 1 copies of  is graphic but has no -connected realization. (F. Jao)
7.1.9. (−) Prove that every even graph has even edge-connectivity.
7.1.10. Determine (G), (G), and (G) for each graph G below. (Hint: For the graph on
the left , use Proposition 7.1.13 to establish the edge-connectivity.)
• • • •
• • • •
• • • • • •
• • • •
• • • •
296 Chapter 7: Connectivity and Cycles

7.1.11. A graph G is ¾ -expansive if for every set S ⊆ V(G) with | S| ≤ n− ¾ , there are at least
| S| + ¾ vertices of G whose neighborhood intersects S. Prove that every ¾-expansive graph
is ¾-connected. Provide infinitely many ¾-connected graphs that are not ¾-expansive.
7.1.12. For ¾ ≥ 2, prove that the only separating sets of size at most ¾ in the hypercube
Q¾ are vertex neighborhoods. (Ramras [2004])
7.1.13. Let G be an n-vertex graph such that (G) = (G) < n − 1. Prove that every
smallest edge cut consists of one edge incident to each vertex of a smallest separating set.
7.1.14. Prove that a connected graph is -edge-connected if and only if each of its blocks
is -edge-connected.
7.1.15. Use Proposition 7.1.13 to prove that the Petersen graph has an edge cut of size 
if and only if 3 ≤  ≤ 12 (thus its edge-connectivity is 3).
7.1.16. Let G be an n-vertex triangle-free graph, with (G) ≥ 3. Prove that if n ≤ 11 ,
then G is 3-edge-connected. Prove sharpness by finding a 3-regular bipartite graph with
12 vertices that is not 3-edge-connected. (F. Galvin)
7.1.17. (♦) Let G be an n-vertex graph. Prove that if (G) ≥ (n +  − 2)/2, where
1 ≤  ≤ n − 1, then G is -connected. Prove sharpness by constructing an n-vertex graph
with minimum degree ⌊ (n +  − 3)/2⌋ that is not -connected.

7.1.18. (+) Let G be an n-vertex graph with n ≥  + l and (G) ≥ n+ll(+¾1−2) . Prove that if
G − S has more than l components, then | S| ≥  . Prove that the hypothesis on (G) is
sharp for n ≥  + l by constructing an appropriate n-vertex graph with minimum degree
⌊ n+ l(l¾+−12)−1 ⌋ . (Comment: This generalizes Exercise 7.1.17.)
7.1.19. (♦) Generalization of Exercise 5.3.37 to ( + 1)-connected graphs. (Bondy [1969])
(a) Let G have degrees d1 , . . . , d n in nondecreasing order, and suppose 0 ≤  ≤ n − 2.
Prove that if d j ≥ j +  whenever j ≤ n − 1 − d n−¾ , then G is ( + 1)-connected.
(b) Prove sharpness by constructing an example for each  to show that requiring
d j ≥ j +  whenever j < n − 1 − d n−¾ does not imply (G) >  .
7.1.20. (♦) Let G be an n-vertex graph with m edges. For  ≥ 2, prove that if n ≥ 2  − 1 > 1
and m > (2  − 3)(n −  + 1), then G has a -connected subgraph. Conclude that average
vertex degree a guarantees a subgraph with connectivity at least a/4. (Mader [1972])
7.1.21. Let G be an n-vertex graph.
(a) Prove that if (G) ≥ n − 2, then (G) = (G).
(b) For n ≥ 4, prove that part (a) is sharp by constructing an n-vertex graph with
minimum degree n − 3 and connectivity less than n − 3.
(c) Prove that if (G) = n − 3 and G contains no 4-cycle, then (G) = (G).

7.1.22. (♦) Let G be an n-vertex graph such that d(x) + d(y) ≥ n − 1 whenever xy ∈
/ E(G).
Prove that (G) = (G). As a function of n, determine the least  such that (G) ≥ 
implies (G) = (G).
7.1.23. (♦) Prove that if G is a bipartite graph with diameter 3, then (G) = (G). (Hint:
Enhance the argument of Theorem 7.1.15.) (Plesnı́k–Znám [1989])
7.1.24. (♦) Let G be a graph with diameter 2.
(a) Prove that (G) = (G) when G has girth at least 5.
(b) Prove that part (a) is sharp by constructing an example with girth 4, diameter
2, and (G) < (G). (Hint: Create a 4-vertex cut S such that G − S = 2 K3 ,3 . Comment:
Soneoka–Nakada–Imase–Peyrat [1987] proved (G) = (G) for G with girth  and diame-
ter at most 2⌈ /2⌉ − 3. Diameter at most 2⌈ /2⌉ − 2 guarantees (G) = (G).)
Exercises for Section 7.1 297

7.1.25. (♦) Let G be an n-vertex bipartite graph. Prove the following implications and
show that they are sharp.
(a) If (G) > n/4, then  (G) = (G). (Volkmann [1988])
(b) If (G) ≥ n/3, then (G) = (G). For sharpness, construct for each odd integer  a
bipartite graph G with (G) =  and | V(G)| = 3 + 1 that is not -connected. (Kostochka)
7.1.26. (+) For  ≥ 2, prove that every graphic list of integers with smallest value at least
 is the degree list of some -edge-connected graph. (Hint: Consider a realization whose
smallest edge cut has size h. If h <  , then find a 2-switch that produces a graph with
fewer edge cuts of size at most h.) (Edmonds [1964])
7.1.27. Theorem 7.1.8 states (G H) ≥ (G) + (H) for connected graphs G and H .
(a) Prove that the bound can be weak by computing (G H) when G = H = K1 2 K r .
(b) Prove that  (G H) ≥  (G) +  (H) when G and H are connected.
7.1.28. (+) Strengthen Theorem 7.1.8 by proving, for nontrivial graphs G and H ,
(G H) = min{ (G) + (H) , | V(G)| (H) , | V(H)| (G)}.
(Hint: Prove that if S is a separating set with | S| < min{| V(G)| (H) , | V(H)| (G)}, then S
has at least (G) + 1 vertices in some two copies of G and (H) + 1 vertices in some two
copies of H .) (Špacapan [2008]; also proved in Govorčin–Škrekovski [2014])
7.1.29. (♦) Let H be a spanning subgraph of a connected graph G. Prove that H is a span-
ning tree if and only if G − E(H) is a maximal subgraph containing no bond of G. (Com-
ment: Note that H is a spanning tree if and only if H is a maximal subgraph containing
no cycle. See Chapter 11 to relate the two statements.)
7.1.30. (♦) Prove that the edges of any edge cut can be partitioned into bonds.
7.1.31. (♦) Edge cuts and cycles.
(a) Let F be a nonempty set of edges in a graph G. Prove that F is an edge cut if and
only if F has an even number of edges in every cycle in G. (Hint: For sufficiency, find an
appropriate bipartition of the components of G − F .)
(b) A signing of a graph is a map  : E(G) → {+1 , −1}. A signing is positive if each
cycle has an even number of negative edges. Prove that the number of positive signings of
any connected n-vertex graph is 2 n−1 . (See also Exercise 5.3.27 about signed graphs.)
7.1.32. For n ≥  , prove that an n-vertex graph with no -connected subgraph has at most
(n2) − (n−¾+31) −1 edges. (Matula [1983])
2

7.1.33. Prove that G is an even graph if and only if every block of G is Eulerian.
7.1.34. (♦) Let H be the block-cutpoint graph (Remark 7.1.18) of a graph G that has a cut-
vertex. (Harary–Prins [1966])
(a) Prove that H is a forest.
(b) Prove that at least two blocks of G contain only one cut-vertex of G.
(c) Prove that G has exactly  + ∑v∈V(G)(b(v) − 1) blocks, where  is the number of
components of G and b(v) is the number of blocks containing v.
(d) Prove that every graph has fewer cut-vertices than blocks.
7.1.35. Prove that a graph has no connected induced subgraph with three leaf blocks if
and only if it is claw-free and net-free, where the claw is K1 ,3 and the net is the graph
formed from K3 and K 3 by adding a matching joining them. (Kelmans [2006])
7.1.36. The cyclic edge-connectivity of a graph is the least number of edges whose dele-
tion leaves a disconnected graph with a cycle in each component. For m ≥ 6, prove that
Cm−2 2 K1 is 4-connected and has cyclic edge-connectivity m. (Plummer [1972])
298 Chapter 7: Connectivity and Cycles

7.2. Properties of ¾-Connected Graphs

Being ¾-connected is more restrictive than being connected. When G is con-
nected and x , y ∈ V(G), there is an x , y-path. When G has ¾ such paths sharing
no internal vertices, separating y from x requires deleting at least ¾ vertices.
We will show that this obviously sufficient condition for being ¾-connected is also
necessary. When phrased appropriately, these concepts apply also to digraphs.

7.2.1. Definition. A digraph D is strongly connected (or strong) if it contains

a path from each vertex to every other. It is ¾-connected if it has more than
¾ vertices and D − S is strong for any set S ⊆ V(D) with | S| < ¾ . The connec-
tivity (D) is the maximum  such that D is -connected. For ∅
= S ⊂ V(D),
the edge cut [S, S] is the set of edges from S to S. A digraph D is -edge-
connected if every edge cut has size at least  . The edge-connectivity
(D) is the minimum size of an edge cut.

Note that a graph or digraph G is -edge-connected if and only if for every

nonempty proper vertex subset S, at least  edges leave S.

MENGER’S THEOREM

When studying paths from one specified vertex to another, there is a natural
local version of separation, valid for both graphs and digraphs.

7.2.2. Definition. For xy ∈ / E(G), an x , y-separating set is a subset S of

V(G) − {x , y} such that G − S has no x , y-path. We write (x , y) for the min-
imum size of such a set. Paths from x to y are independent if they share no
internal vertex. We write (x , y) for the maximum size of a set of pairwise
independent x , y-paths.

Every x , y-separating set is at least as large as every family of independent

x , y-paths, so always (x , y) ≥ (x , y). Thus like vertex cover and matching in
Chapter 6, separation and path-packing are dual optimization problems (see also
Section 8.3 for further discussion of duality). Again, a guarantee of equality for
the optima of a pair of dual problems is called a min-max relation, guaranteeing
that an optimal solution to an instance of one problem can be proved optimal by
exhibiting a solution to the other problem. Min-max relations also often lead to
polynomial-time algorithms to find optimal solutions.
Menger [1927] proved (x , y) = (x , y) for nonadjacent vertices x and y in
a graph G. Separating G requires separating some two vertices, so (G) equals
min{ (x , y): xy ∈ / E(G)}, which yields a global duality. The global version and
analogues for edge-connectivity and digraphs were observed by others, but all
eight variants are called Menger ’s Theorem. We start with a similar result.
Section 7.2: Properties of ¾-Connected Graphs 299

7.2.3. Definition. For X , Y ⊆ V(G), an X , Y -path is a path from X to Y visiting

X and Y only at its endpoints. An X , Y-barrier is a vertex set Z such that
G − Z has no X , Y -path. An X , Y -link is a set of pairwise disjoint X , Y -paths.

We use “barrier ” instead of “cut ” since an X , Y-barrier may intersect X and

Y ; in particular, X and Y themselves are X , Y-barriers.

7.2.4. Theorem. (Pym’s Theorem; Pym [1969]) In a graph or digraph G, the

minimum size of an X , Y-barrier is the maximum size of an X , Y-link.
Proof: We use induction on | V(G)| + | E(G)| ; the claim is trivial for small graphs.
Let ¾ be the minimum size of an X , Y-barrier in G; we may assume ¾ ≥ 1. For
the induction step, we consider two cases (illustrated on the left and right below).
Case 1. G has an X , Y -barrier Z of size ¾ other than X or Y . Since G − Z has
no X , Y -path, every X , Y-path visits Z. Hence no X , Z-path and Z, Y-path share
a vertex outside Z. Let G 1 and G 2 be the subgraphs of G induced by the vertices
on X , Z-paths and on Z, Y-paths, respectively; we have V(G 1) ∩ V(G 2) = Z. Since
Z is not X or Y and ¾ ≥ 1, both G 1 and G 2 are smaller than G.
Since every X , Y-path visits Z, every X , Z-barrier is an X , Y -barrier. Hence
G 1 has no X , Z-barrier of size ¾ − 1. By the induction hypothesis, in G 1 there is
an X , Z-link of size ¾ . Similarly, G 2 has a Z, Y-link of size ¾ . They meet at Z to
form an X , Y -link of size ¾ .

• •
• • •
• • • •
• • • •
x w
• • • •
X Z Y X Y

Case 2. G has no X , Y -barrier of size ¾ other than X and/or Y . By symmetry,

we may assume | X | = ¾ . If X ⊆ Y , then the link consists of paths of length 0.
Otherwise, there exists x ∈ X − Y . Since X is a smallest X , Y -barrier, the graph
G − (X − {x}) has an X , Y -path; it begins along some edge xw with w ∈ / X . Let Z
be a smallest X , Y -barrier in G − xw.
If G − xw has an X , Y -link of size ¾ , then it is an X , Y -link of size ¾ in G.
Otherwise, by the induction hypothesis, | Z | < ¾ . Now Z is not an X , Y -barrier
in G, so G − Z has an X , Y -path. Every such path must use the edge xw. Hence
Z ∪ {x} and Z ∪ {w} are X , Y-barriers in G. Since these sets have size at most ¾ ,
the hypothesis for Case 2 yields Z ∪ {x} = X and Z ∪ {w} = Y . Now the 1-vertex
paths in Z together with the x , w-path of length 1 form the desired X , Y -link.

The analogous duality holds for edge-connectivity. Given ¾ pairwise edge-

disjoint x , y-paths, we must delete at least ¾ edges to make y unreachable from
x. Appropriate notation and the “line graph” operation yields the min-max rela-
tions for edges from those for vertices as corollaries of Pym’s Theorem.
300 Chapter 7: Connectivity and Cycles

7.2.5. Definition. Let (x , y) be the minimum number of edges whose deletion

makes y unreachable from x, and let  (x , y) be the maximum size of a set of
pairwise edge-disjoint x , y-paths.
The line graph of a (multi)graph G, written L(G), is the graph whose
vertices are the edges of G, with e  ∈ E(L(G)) when e ,  ∈ E(G) with e = uv
and  = vw. The line digraph of a digraph is defined in the same way with
edges being ordered pairs.
• •
•
• •
  
•e• h e• •h •e• h • j• e• •h •j
 i
• i•
• •
G L(G) H L(H)

The line graph operation motivates our use of the “prime” for related param-
eters involving vertices or edges. We have  (G) = (L(G)), and the next proof
shows that (G) is closely related to (L(G)).

7.2.6. Theorem. (Menger [1927], Ford–Fulkerson [1956]) If x and y are vertices

in a (multi)graph or digraph G, then (x , y) = (x , y) when xy ∈
/ E(G), and
(x , y) =  (x , y) always.

Proof: Pym’s Theorem yields the first statement, as follows. For graphs, use
X = N(x) and Y = N(y), and add edges from x to X and from Y to y to complete
the desired paths. For digraphs, use X = N +(x) and Y = N −(y).
For the second statement, apply Pym’s Theorem to the line graph of G. Let
X be the set of edges incident to [leaving] x, and let Y be the set of edges incident
to [entering] y. The X , Y -link of size (x , y) in L(G) transforms into  pairwise
edge-disjoint x , y-paths in G.
These edge statements extend without change to multi[di]graphs. Under the
line graph operation, multiedges become distinct vertices, and the application of
Pym’s Theorem is unchanged.

The resulting characterization of -connected graphs (the “global” version of

Menger ’s Theorem) was observed by Whitney [1932a]. The global versions for
edges and digraphs appeared in Ford–Fulkerson [1956]. For vertices we need a
lemma (valid also for digraphs; see Exercise 8).

7.2.7. Lemma. Deletion of an edge reduces connectivity by at most 1.

Proof: If S is a smallest separating set of G − xy, then S ∪ {x} or S ∪ {y} separates
G (yielding (G) ≤ (G − xy) + 1), unless x and y are the only vertices of G − S.
In that case | S| = | V(G)| − 2 ≥ (G) − 1, so again (G) ≤ (G − xy) + 1.

• x X

• y Y
Section 7.2: Properties of ¾-Connected Graphs 301

7.2.8. Theorem. (Menger ’s Theorem; Whitney [1932a], Ford–Fulkerson [1956])

For a graph or a digraph G,
(1) G is ¾-connected if and only if (x , y) ≥  for all x , y ∈ V(G), and
(2) G is -edge-connected if and only if (x , y) ≥  for all x , y ∈ V(G).
Proof: Since  (G) = min x ,y∈V(G)  (x , y), Theorem 7.2.6 yields (2).
For (G), Theorem 7.2.6 yields (x , y) = (x , y) for xy ∈
/ E(G), and (G) is the
least of these values. We need only show that (x , y) ≥ (G) for xy ∈ E(G). Since
xy forms an x , y-path and lies in no other x , y-path,

G (x , y) = 1 + G − xy (x , y) = 1 +  G− xy(x , y) ≥ 1 + (G − xy) ≥ (G).

APPLICATIONS OF MENGER’S THEOREM

Our first application is a short proof of a useful fact.

7.2.9. Corollary. (G) =  (G) when (G) = 3.

Proof: Always (G) ≤  (G),
so it suffices to show for x , y ∈ V(G) that G has
 (G) independent x , y-paths. By Menger ’s Theorem, G has  (G) edge-disjoint
x , y-paths. Since (G) = 3, two such paths share no internal vertices, so the
 (G) paths are independent.

The next lemma gives another way (besides vertex -splits) to enlarge a graph
while preserving -connectedness.

7.2.10. Lemma. (Expansion Lemma) If G is formed from a -connected graph

G by adding a vertex y with at least  neighbors in G, then G is -connected.
Proof: If | S| <  , then G − S is connected, and a neighbor of y remains, so by
transitivity G − S is connected.

Analogous conditions yield expansion lemmas for digraphs and for edge-
connectivity of multigraphs; see Exercise 23.

7.2.11. Definition. For x ∈ V(G) and U ⊆ V(G), an x , U-fan of size  is a set of 

paths from x to U that pairwise share only x and reach U only at their ends.

7.2.12. Lemma. (Fan Lemma; Dirac [1960]) A graph with more than  vertices
is -connected if and only if it has an x , U-fan of size  for each choice of x
and U with |U | ≥  and x ∈
/ U.
Proof: Let G be -connected, and construct G from G by adding a new vertex y
adjacent to all of U. Since G is -connected, the Expansion Lemma implies that
G is -connected, and by Menger ’s Theorem there is in G a set of  independent
x , y-paths. Stopping these paths where they reach U produces an x , U-fan of size
 in G, as illustrated below.
If G is not -connected, then G has a separating ( − 1)-set S (Exercise 7.1.1).
Choose x and y in distinct components of G − S, and let U = S ∪ {y}. There is no
x , U-fan, since every x , y-path intersects S.
302 Chapter 7: Connectivity and Cycles

G •
U
x• • •y
•

Exercises 28–29 generalize the Fan Lemma. We give an application.

7.2.13. Theorem. (Dirac [1960]) When ¾ ≥ 2, in any ¾-connected graph G each

set of ¾ vertices lies on some cycle.
Proof: We use induction on ¾ . When ¾ = 2, two independent x , y-paths guaran-
teed by Menger ’s Theorem form a cycle containing {x , y}.
For ¾ > 2, observe that G is also (¾ − 1)-connected, and choose x ∈ S. By the
induction hypothesis, G has a cycle C through S − {x}. If | V(C)| = ¾ − 1, then
consider an x , V(C)-fan of size ¾ − 1. Using two successive such paths to form a
detour through x completes the desired cycle.
If | V(C)| ≥ ¾ and x ∈
/ V(C), then use an x , V(C)-fan of size ¾ . Partition V(C)
into ¾ − 1 segments, each starting at an element of S − {x}. Since the fan has size
¾ , two paths reach C in the same segment (by the Pigeonhole Principle). Detour
from C along them to visit x between two vertices of S − {x}.
u
v1
C •
x ¾=4
u •
• •
v3 v2

When ¾ ≥ 3, the converse does not hold, as shown when G is a cycle. Bondy–
Lovász [1981] proved the stronger result that for ¾ ≥ 3 in a ¾-connected graph
every set of ¾ vertices lies on an even cycle.
To apply Menger ’s Theorem, model a problem by defining a graph or digraph
whose paths yield the desired objects. Exercise 31 requests such proofs of Hall’s
Theorem and the K önig–Egerváry Theorem. (One can also prove Menger ’s Theo-
rem from those results, as K önig did to fill a gap in Menger ’s original proof.)
Ford and Fulkerson solved a more general problem. Recall that an SDR for a
family A1 , . . . , Am of sets consists of distinct elements 1 , . . . , m with i ∈ Ai .

7.2.14. Definition. Let A and B be two families of sets, each of size m. A com-
mon system of distinct representatives (CSDR) is a set of m elements
that form an SDR for A and also for B.

7.2.15. Theorem. (Ford–Fulkerson [1958]) Let A and B be families of m sets:

A = { A1 , . . . , Am} and B = {B1 , . . . , Bm}. Using the notation A(I) = ⋃i∈ I Ai
and B(J) = ⋃ j ∈ J Bj , families A and B have a (CSDR) if and only if
### A(I) ∩ B(J)### ≥ I + J − m for each pair I , J ⊆ [m].
## ## | | | |
Section 7.2: Properties of ¾-Connected Graphs 303

Proof: We prove that the condition of satisfying the equivalent inequality

### A(I) ∩ B(J)### + (m − I ) + (m − J ) ≥ m (∗)
## ## | | | |
for all I , J ⊆ [m] is both necessary and sufficient for the existence of a CSDR.
Create a digraph G with special vertices s and t and vertex sets A , B , and
X , where A = {ai : Ai ∈ A}, B = {bj : Bj ∈ B}, and X = ( ⋃ Ai ) ∪ ( ⋃ Bj ) . As
illustrated below, the edges are
{sai : Ai ∈ A}, {ai x: x ∈ Ai},
{bj t: Bj ∈ B}, {xbj : x ∈ Bj }.

An s , t-path is ⟨s , ai , x , bj , t⟩ for x ∈ Ai ∩ Bj . Thus A and B have a CSDR if and

only if G has m independent s , t-paths. By Menger ’s Theorem, m independent
s , t-paths exist if and only if every s , t-separating set has size at least m.
Let R be a minimal s , t-separating set. Let I = {i ∈ [m]: ai ∈ / R} and J = { j ∈
[m]: bj ∈/ R}. Since the vertices of A and B indexed by I and J are not being
deleted, R will be an s , t-cut if and only if A(I) ∩ B(J) ⊆ R, and equality will hold
when R is minimal. In that case,

| R| = ##### A(I) ∩ B(J)##### + (m − | I |) + (m − | J |).

Therefore, requiring (∗) for all I , J ⊆ [m] is both necessary and sufficient for each
s , t-separating set to have size at least m.

4
•
a3 b3
I • • J
3
•
s• • • •t
a2 2 b2
R • R
a1 b1
• •
1
•
A X B

7.2.16. Example. Digraph for CSDR. In the figure above, the elements are
{1 ,2 ,3 ,4}, with A = {12 , 23 , 31} and B = {14 , 24 , 1234}. When R contains
{a1 , a2 , b1 , b2 } but not a3 or b3 , we set I = J = {3}. Now R is an s , t-cut if
and only if it also contains {1 , 3}, which is ( ⋃i∈ I Ai ) ∩ ( ⋃ j ∈ J Bj ) .

Like many min-max relations, Menger ’s Theorem follows from the Max-Flow
Min-Cut Theorem of network flow theory. Conversely, Menger ’s Theorem implies
both the flow theorem (for rational capacities) and most of its combinatorial ap-
plications. For example, Theorem 7.2.15 above is usually proved in the language
of network flows, but Menger ’s Theorem suffices. We will apply Theorem 7.2.15
in Chapter 12.
304 Chapter 7: Connectivity and Cycles

2-CONNECTED AND 3-CONNECTED GRAPHS

Menger ’s Theorem characterizes ¾-connected graphs but does not say how to
construct them. Construction or decomposition procedures for a class of graphs fa-
cilitate the development of iterative algorithms and inductive proofs. For graphs
that are 2-connected or 3-connected, such procedures exist using vertex splits.
First we introduce an inverse operation.

7.2.17. Definition. Given an edge e in a graph G, the contraction G · e is the

graph obtained from G by replacing the endpoints of e with a single vertex
whose neighbors are their neighbors in G outside e. We say that G · e is ob-
tained from G by contracting e. A ¾-contractible edge is an edge whose
contraction leaves a ¾-connected graph.

The endpoints of an edge e may have common neighbors. When discussing

multigraphs, we may let these yield multiedges with the new vertex in G · e. Here
we consider only graphs and do not introduce multiedges. Thus contracting the
edge x1 x2 introduced by splitting vertex x inverts the splitting operation.
An easy construction procedure for 3-connected graphs results from finding
3-contractible edges.

7.2.18. Lemma. (Contraction Lemma; Tutte [1961]) Every 3-connected graph

other than K 4 has a 3-contractible edge.
Proof: (Thomassen [1980a]) Let G be a 3-connected graph with at least five ver-
tices. If G has no contractible edge, then for each edge e in G the graph G · e has
a separating 2-set. This 2-set must consist of the vertex formed by contracting e
and some other vertex , which we call a companion of e. If e = xy, then {x , y , }
is a separating triple in G.
Among all the edges of G, choose xy and a companion of xy so that the result-
ing disconnected graph G − {x , y , } has a component H with the largest possible
number of vertices. Let H be another component of G − {x , y , }, as sketched be-
low. Since {x , y , } is a minimal separating set, each element of {x , y , } has a
neighbor in both H and H .
Let u be a neighbor of in H . Since u ∈ E(G), we are guaranteed a separat-
ing set { , u , v} of G. The subgraph of G induced by V(H) ∪ {x , y} is connected.
Deleting v from this subgraph, if it occurs there, cannot disconnect it, since { , v}
would be a separating set of G. Therefore G[V(H) ∪ {x , y}] − v lies in a compo-
nent of G − { , u , v} with more vertices than H , which contradicts the choice of
{x , y , }. Hence G does have a contractible edge.

• • •u
x
• • •

H • • • H
y

We use Lemma 7.2.18 in characterizing planar graphs in Chapter 9.

Section 7.2: Properties of ¾-Connected Graphs 305

7.2.19. Theorem. A graph is 3-connected if and only if it can be obtained from

K 4 by a sequence of vertex 3-splits.
Proof: By Lemma 7.1.6, every graph so constructed is 3-connected. By Lemma
7.2.18, every 3-connected graph can be obtained.

For ¾ > 3, finding ¾-contractible edges is harder. In fact, when ¾ is even,

the ¾-connected graph H¾ ,n of Example 7.1.3 has no ¾-contractible edge (Exercise
33). Thomassen [1981b] proved that every triangle-free ¾-connected graph has a
¾-contractible edge (Exercise 34).
A more traditional and harder characterization of 3-connected graphs is due
to Tutte [1961]. We use only disjoint 3-splits, where a disjoint ¾-split requires
the neighborhoods of the new vertices to be disjoint. Inverting a disjoint 3-split
means finding an edge not on a triangle to contract. Tutte proved that the 3-
connected graphs are the graphs that arise by disjoint 3-splits and edge additions
from graphs of the form K 1 Cn−1 , called wheels.
Similarly, the 2-connected graphs can be obtained from C3 by disjoint 2-splits
and edge additions, but there is a more useful way to describe the construction
procedure. We start with a special case of disjoint 2-split.

7.2.20. Definition. In a graph G, subdivision of an edge uv is the operation of

replacing uv with a path u , w , v through a new vertex w.

• • → • ••
u v u w v

If d(v) ≥ 2, then subdivision of uv is equivalent to a vertex 2-split at v where

one new vertex inherits only u and the other inherits all other neighbors of v. By
Lemma 7.1.6, subdivision therefore preserves 2-connectedness. We also want to
apply it to 2-edge-connected graphs.

7.2.21. Lemma. Edge-subdivision preserves 2-edge-connectedness.

Proof: Let G be a 2-edge-connected graph, and let H be obtained from G by
adding w to subdivide uv. Since cut-edges are those in no cycle, every edge of
G lies in a cycle. The same edge in H lies in the corresponding cycle, lengthened
by the replacement of uv with uw and wv if it used uv. Similarly, uw and wv lie
in a cycle corresponding to a cycle through uv.
Also, lengthening paths of G (if needed) shows that H is connected. Since H
is connected and has no cut-edge, it is 2-edge-connected.

The construction procedure for 2-connected or 2-edge-connected graphs as-

sembles them from a decomposition into cycles and paths.

7.2.22. Definition. An ear in a graph G is a path whose internal vertices have

degree 2 in G and whose endpoints have degree greater than 2. An ear de-
composition of G is a decomposition Q¾ , . . . , Q0 such that Q0 is a cycle and
Qi for i ≥ 1 is an ear in Q0 ∪ · · · ∪ Qi .
A closed ear in a graph G is a cycle whose vertices have degree 2 in G
306 Chapter 7: Connectivity and Cycles

except for one of degree at least 4. A weak ear decomposition of G is a de-

composition Q¾ , . . . , Q0 such that Q0 is a cycle and Qi for i ≥ 1 is either an
ear or a closed ear in Q0 ∪ · · · ∪ Qi .
•
Q3
•
• •
•Q Q1
•
2

• Q0 •
Q4
• •
•
Weak ear decompositions are also called “closed-ear decompositions”. We use
“weak” to emphasize that the requirements for the decomposition are relaxed;
each added Qi may now be an ordinary ear or a closed ear.

7.2.23. Theorem. (Whitney [1932a]) A graph is 2-connected if and only if it has

an ear decomposition. A multigraph is 2-edge-connected if and only if it has a
weak ear decomposition. Furthermore, when such decompositions exist they
can start with any cycle.
Proof: Sufficiency. Cycles are 2-connected, so it suffices to show that adding an ear
or closed ear preserves the desired properties. Ear addition can be accomplished
by adding an edge (or a new vertex with two neighbors) and then possibly subdi-
viding. These operations preserve 2-connectedness and 2-edge-connectedness.
When adding a closed ear to a 2-edge-connected graph, the properties of being
connected and having every edge on a cycle are maintained.
Necessity. Given a 2-connected graph G, we build an ear decomposition of G
from any cycle C. Let G 0 = C. Consider a subgraph G i built by adding ears. If
G i
= G, then we may select an edge uv ∈ E(G) − E(G i) with u ∈ V(G i) (since G
is connected). Since G is 2-connected, there is a path P in G − u from v to V(G i)
(it has length 0 if v ∈ V(G i)). Now {uv} ∪ P is an ear that can be added to G i .
Repeating the argument absorbs all of G.
For a weak ear decomposition of a 2-edge-connected (multi)graph, the proof
is the same, except that we find a path in G − uv instead of G − u. If the resulting
path ends at u, then we obtain a closed ear addition; otherwise we have an ear
addition as before.

Weak ear decompositions yield a short solution of the “one-way street ” prob-
lem. When can the streets of a road network all be made one-way and still permit
each location to reach every other? A strong orientation of a graph is an orien-
tation that is a strongly connected digraph (Definition 7.2.1).

7.2.24. Corollary. (Robbins [1939]) A multigraph has a strong orientation if

and only if it is 2-edge-connected.
Proof: One cannot travel both directions across an edge cut having fewer than
two edges. Hence the condition is necessary.
Section 7.2: Properties of ¾-Connected Graphs 307

For sufficiency, use a weak ear decomposition. Orient the initial cycle consis-
tently. Orient each subsequent ear or closed ear Q from one end (u) to the other
(v). By transitivity, the new digraph is strong: old vertices can reach u, which
can reach all new vertices, and new vertices can reach v, which can reach all old
vertices. After all of G is added, we have a strong orientation.

Similarly, an obvious necessary condition for a ¾-edge-connected orientation

of G is (G) ≥ 2  . We next generalize Corollary 7.2.24 to prove that this is suffi-
cient. The proof of the generalization is much more difficult.

HIGHLY CONNECTED ORIENTATIONS (optional)

In a -edge-connected orientation of G, every edge cut needs at least  edges

in each direction, so G must be 2 -edge-connected. Nash-Williams [1960] gener-
alized Robbins’ Theorem by proving this obvious necessary condition sufficient.
It is easy for Eulerian multigraphs. Orienting consistently along an Eulerian
circuit crosses each cut the same number of times in each direction. If each edge
cut has size at least 2  , the orientation is then -edge-connected. For the gen-
eral result, Nash-Williams showed that the vertices of odd degree can be paired
by paths so that an Eulerian circuit of the rest gives the desired orientation.
Lovász simplified the proof that 2 -edge-connected multigraphs have -edge-
connected orientations. Our presentation follows the survey by Frank [1993].

7.2.25. Definition. (Lovász) In a multigraph or digraph G, with X ⊆ V(G), let

 G(X) or (X) denote the number of edges of G leaving X , so (X) = ##### [X , X ]##### .

Note that (X) < ∑v∈ X d(v) when X is not an independent set. The (X) no-
tation makes it easy to compare the sizes of related edge cuts.

7.2.26. Proposition. Given a graph G and X , Y ⊆ V(G),

(a) (X) + (Y) = (X ∩ Y) + (X ∪ Y) + 2#### [X − Y, Y − X]#### .
# #
(b) (X) + (Y) = (X − Y) + (Y − X) + 2#### [X ∩ Y, X ∪ Y]#### .
# #
Proof: Each equation counts a set in two ways. In terms of the multiplicities
shown in the figure below, the equations are
(b + e +  + c) + (a + e +  + d) = (a + b + e) + (c + e + d) + 2 
(b + e +  + c) + (a + e +  + d) = (a + c +  ) + (b + d +  ) + 2e

X
a
X∩Y X−Y
e
Y b c

Y−X X∪Y
d
308 Chapter 7: Connectivity and Cycles

A function r on sets is submodular if r(X ∩ Y) + r(X ∪ Y) ≤ r(X) + r(Y) for all

sets X and Y (see Chapter 11). Proposition 7.2.26a implies that is submodular.
Note that G is -edge-connected if and only if (X) ≥ whenever ∅
= X ⊂ V(G).

7.2.27. Definition. A multigraph G is -edge-connected relative to a vertex

 if (X) ≥ for all ∅
= X ⊂ V(G) − { }. If G has edges u  and v, then the
u , v-shortcut of  is the graph G = G − u  − v + uv.
 
• •
→
• • • •
u v u v

A u , v-shortcut of  may have uv as a multiedge. Multiedges cause no trouble

when discussing edge-connectivity (we ignore loops). The content of “relative to”
in this definition is to ignore the condition for the cut between  and V(G) − { }.
Lovász proved that when  has even degree there is a shortcut of  that pre-
serves the edge-connectivity relative to . Mader [1978, 1982] proved stronger
shortcut results. The proof of Lovász’s weaker version is more difficult when is
odd. Fortunately, we only need the even case.

7.2.28. Lemma. (Shortcut Lemma; Lovász [1974, 1979]) If  is a vertex of even

degree in a multigraph G that is -edge-connected relative to  , then for all
u ∈ N() there exists v ∈ N() such that the u , v-shortcut of  is also -edge-
connected relative to .
Proof: (for even only). Let u be a fixed neighbor of  , and let V = V(G) − { }.
Let X be a nonempty proper subset of V , and consider v ∈ N() − {u}. Among the
edges leaving X , the u , v-shortcut of  destroys u  or v  if X contains u or v, and
it adds uv if X contains exactly one of {u , v}. Thus the shortcut violates (X) ≥
only if u , v ∈ X and (X) ≤ + 1.

•
•
• • v
v u V
X Y

We call X dangerous if X is a nonempty proper subset of V that contains u

and satisfies (X) ≤ + 1. A u , v-shortcut of  is -edge-connected relative to  if
and only if every dangerous set omits v. Letting M be the union of all dangerous
sets, it thus suffices to show that
M is dangerous and omits some neighbor of .
To show that M is dangerous, it suffices to show that the union of any two
dangerous sets X and Y is dangerous. We may assume that X − Y and Y − X
are nonempty. We prove first that (X ∪ Y) is odd and then use this to prove that
X ∪ Y is a proper subset of V and satisfies (X ∪ Y) ≤ + 1.
Section 7.2: Properties of ¾-Connected Graphs 309

Since X and Y are dangerous, (X) , (Y) ≤ + 1. By the definition of dan-

gerous, u ∈ X ∩ Y , so u  ∈ [X ∩ Y, X ∪ Y]. The hypothesis on relative edge-
connectivity yields (X − Y) ≥ and (Y − X) ≥ . We use Proposition 7.2.26b to
rewrite (X) + (Y) in terms of X − Y and Y − X :
# #
2 + 2 ≥ (X) + (Y) = (X − Y) + (Y − X) + 2#### [X ∩ Y, X ∪ Y ]#### ≥ 2 + 2.
Equality must hold throughout, and hence (X) = (Y) = + 1 and (X − Y) =
(Y − X) = . Now, since (Y) + (X − Y) ≡ (X ∪ Y) (mod 2), we conclude that
(X ∪ Y) is odd. Also, (X ∩ Y) + (X − Y) ≡ (X) (mod 2), so (X ∩ Y) is odd.
Since (X ∪ Y) is odd and (V ) = d(), evenness of d() yields X ∪ Y
= V .
Now relative edge-connectivity, grouping of the edges leaving X and leaving Y ,
and the dangerousness of X and Y yield
+ ≤ (X ∩ Y) + (X ∪ Y) ≤ (X) + (Y) ≤ 2 + 2.
Since (X ∩ Y) and (X ∪ Y) are odd, they must both equal + 1, and hence X ∪ Y
is dangerous.
We have proved that M is dangerous. If there is no dangerous set, then every
u , v-shortcut of  works, so we may assume M
= ∅. Since M is a nonempty proper
subset of V and G is -edge-connected relative to  , we have (V − M) ≥ .
If M contains all neighbors of  , then (M ∪ { }) = (M) − d(). Since d() is
a positive even number and M is dangerous, we then have
≤ (V − M) = (M ∪ { }) = (M) − d() ≤ (M) − 2 ≤ − 1 .
The contradiction implies that M omits some v ∈ N(), and the u , v-shortcut of 
is -edge-connected relative to .

To prove the Orientation Theorem from the Shortcut Lemma inductively, we

need a vertex of even degree. With = 2t and seeking a t-edge-connected orien-
tation, we will discard edges from our -edge-connected multigraph to obtain a
minimal -edge-connected multigraph, one where the deletion of any edge de-
stroys -edge-connectedness. Submodularity of the degree function now makes it
fairly easy to obtain a vertex with degree .

7.2.29. Lemma. (Mader [1971]) Every minimal -edge-connected multigraph

has a vertex of degree .
Proof: If (X) > for all ∅
= X ⊂ V(G), then deleting any edge leaves a -edge-
connected multigraph. Thus (X) = for some set X .
Suppose that G[X] has an edge xy. Since G − xy is not -edge-connected,
there is a set Z ⊂ V(G), with Z containing exactly one of the vertices x and y,
such that − 1 ≥  G− xy(Z) = (Z) − 1. Since (Z) ≥ , equality holds.
Now -edge-connectedness of G and submodularity of  yield
+ ≤ (X ∩ Z) + (X ∪ Z) ≤ (X) + (Z) = + .
Since G is -edge-connected, we obtain (X ∩ Z) = . Since Z contains exactly
one of {x , y}, the set X ∩ Z is smaller than X .
Hence a minimal set X such that (X) = must be an independent set. Since
each vertex of X has at least incident edges leaving X , we have | X | = 1, and
this is the desired vertex of degree .
310 Chapter 7: Connectivity and Cycles

7.2.30. Theorem. (Orientation Theorem; Nash-Williams [1960]) An n-vertex

multigraph has a t-edge-connected orientation if and only if it is 2t-edge-
connected.
Proof: (Lovász [1974, 1979], see Frank [1993]). Necessity. We need t edges each
way across each cut. Sufficiency. We use induction on n. For n = 2, the two ver-
tices are joined by at least 2t edges.
For n > 2, consider a 2t-edge-connected multigraph G. We discard edges to
obtain a minimal 2t-edge-connected multigraph; we later orient the deleted edges
arbitrarily. By Lemma 7.2.29, what remains has a vertex of degree 2t. Lovász’s
Shortcut Lemma iteratively finds shortcuts of until we reduce the degree of
to 0. This process maintains 2t-edge-connectedness relative to . At the end,
deleting yields a 2t-edge-connected multigraph G with n − 1 vertices.
By the induction hypothesis, G has a t-edge-connected orientation. Orient
G by replacing each shortcut edge uv with {u , v} or {v , u}, agreeing with the
orientation of uv. For X
= { }, lifting uv preserves (X) ≥ t in the orientation;
the only edge lost is uv, and if uv leaves X , then u  or v is a new edge leaving
X , depending on whether  ∈ X . After t lifts, also ({ }) = t, so now (X) ≥ t
whenever ∅
= X ⊂ V(G).

Mader [1978] used his Shortcut Lemma to give another proof of the Strong
Orientation Theorem. Other proofs of the Orientation Theorem use polyhedral
combinatorics (see Frank [1980a], Frank–Tardos [1984]).
We close this section with the analogue of Lemma 7.2.29 for vertices. A min-
imal -connected graph is a graph from which the deletion of any edge destroys
-connectedness. Halin [1969, 1971] proved that every such graph has a vertex
of degree  . Mader strengthened Halin’s result by showing that every cycle con-
tains a vertex of degree  . We need a technical lemma.

7.2.31. Lemma. (Mader [1972]). Let ax and ay be edges in a minimal -connected

graph G, with d G(a) ≥  + 1 ≥ 3. If S is a ( − 1)-set separating G − ax, and T
is a ( − 1)-set separating G − ay, then the component of G − ay − T containing
y has fewer vertices than the component of G − ax − S containing a.
Proof: The graph G − ax − S has two components, with vertex sets A containing
a and X containing x. Also G − ay − T has two components, with vertex sets B
containing a and Y containing y (see figure below). Note that x ∈ / Y and y ∈ / X.
The claim is | Y | < | A|. Since Y is partitioned by its intersections with the
sets A , S, X , and since A ∩ B is nonempty (it contains a), it suffices to show (1)
| Y ∩ S| ≤ | A ∩ T | , and (2) Y ∩ X = ∅.
(1) Let U = (S − Y) ∪ (T − X). If | A ∩ T | < | Y ∩ S| , then |U | < | S| =  − 1.
Since d(a) ≥  + 1, the vertex a has a neighbor  outside U ∪ {x , y}. All neighbors
of a not in U ∪ {x , y} lie in A ∩ B. Since a is the only vertex of A ∩ B having a
neighbor (in G) in X ∪ Y , the set U ∪ {a} separates  from X ∪ Y . This contradicts
the -connectedness of G, since | U ∪ {a}| ≤  − 1, and hence | Y ∩ S| ≤ | A ∩ T |.
(2) Let D = X ∩ Y , and suppose D
= ∅. Let W be the set of vertices outside
D having neighbors in D. Since x ∈ / Y and y ∈ / X , no vertex of W is in A or in
B. Hence W ⊆ (S ∩ Y) ∪ (T − A). From (1), this yields | W | ≤ | T | =  − 1. This is
impossible, since W is a separating set and G is -connected, so D = ∅.
Exercises for Section 7.2 311

B T Y

A
•
a• •y

S U

X
• D
x

7.2.32. Theorem. (Mader [1972]). Every cycle in a minimal ¾-connected graph

contains a vertex of degree ¾ .
Proof: Let [a1 , . . . , a l] be a cycle with no vertex of degree ¾ , and let Si be a sep-
arating (¾ − 1)-set of G − ai−1 ai , where a0 denotes a l . Let ni be the order of the
component of G − ai−1 ai − Si containing ai . By applying Lemma 7.2.31 with ai =
a, Si = S, and Si+1 = T , we obtain ni > ni+1 . Iteration yields the contradiction
n1 > n2 > · · · > nl = n0 > n1 .

Via a simple counting argument, Theorem 7.2.32 implies that almost half the
vertices of a minimal ¾-connected graph have degree ¾ .

7.2.33. Corollary. (Bollobás [1978, p. 25]). Every minimal ¾-connected graph

G with n vertices has at least (¾−2¾1)n+2
−1 vertices of degree ¾ .
Proof: Let S be the set of vertices having degree ¾ . By the Degree-Sum Formula,
G has at least 12 (¾ n + n − | S|) edges. By Theorem 7.2.32, G − S is a forest, and
hence G − S has at most n − | S| − 1 edges. Since deleting S removes at most ¾ | S|
edges, 12 (¾ n + n − | S|) − ¾ | S| ≤ n − | S| − 1. This yields the desired inequality.

Corollary 7.2.33 is almost the best possible bound. Mader [1979] (see Mader
[1996], a survey) proved that a minimal ¾-connected graph with n vertices has at
least (¾−21)n +2 ¾
¾ −1 vertices of degree ¾ . Equality holds for infinitely many examples,
such as those in Exercise 59.

EXERCISES 7.2

7.2.1. (−) Prove or disprove: If P is a u , v-path in a 2-connected graph G, then G has a

u , v-path Q sharing no internal vertices with P.
7.2.2. (−) Let x and y be nonadjacent vertices in a 2-connected graph G. Prove that if
G − x − y is connected, then y lies on a cycle in G − x. Is this true when xy ∈ E(G)?
7.2.3. (−) From a connected graph G, form G by adding the edge xy whenever d G(x , y) = 2.
Prove that G is 2-connected.
7.2.4. (−) Prove that a graph G is 2-connected if and only if G can be obtained from C3 by
edge additions and edge subdivisions.
312 Chapter 7: Connectivity and Cycles

7.2.5. (−) A thread in a graph is a path that is maximal subject to the condition that the
internal vertices have degree 2. Prove or disprove: The last ear added in an ear decompo-
sition of a 2-connected graph G that is not a cycle can be any thread in G.

7.2.6. (−) Let G be a graph with at least three vertices, none isolated. Prove that G is
2-connected if and only if any two edges appear in a common cycle.

7.2.7. (−) Prove that a graph with at least three vertices is 2-connected if and only if for
every 3-tuple (x , y , ) of vertices, the graph has an x , -path through y. (Chein [1968])

7.2.8. For an edge e in a digraph D, proved (D − e) ≥ (D) − 1.

7.2.9. (−) Prove that a graph G with at least four vertices is 2-connected if and only if for
every pair X , Y of disjoint vertex subsets with | X | , | Y | ≥ 2, there are two disjoint paths
from X to Y in G that have no internal vertex in X or Y .

7.2.10. (−) Let v be a vertex in a 3-connected 3-regular graph G. Form G by expanding

v into a triangle whose vertices are joined to the original neighbors of v by a matching.
Thus G is a 3-regular graph. Prove that G is 3-connected.

7.2.11. (−) Show that Theorem 7.2.13 is best possible by constructing, for all  ∈  with
 ≥ 3, a -connected graph having  + 1 vertices that do not lie on a cycle.

7.2.12. (−) For  ≥ 2, prove that a graph G with at least  + 1 vertices is -connected if
and only if for disjoint S, T ⊆ V(G) with | S| =  − 2 and | T | = 2, there is a cycle in G that
contains T and avoids S. (Lick [1973])

7.2.13. (−) Prove that a connected graph G with more than  vertices is -connected if
and only if  G(x , y) ≥ 2 whenever d(x , y) = 2. (Li [1994], Naatz [2000])

7.2.14. (−) Use Theorem 7.2.19 to prove that the Petersen graph is 3-connected.

7.2.15. (−) Let H be a -connected graph obtained by contracting the edges of a matching
in a graph H with (H) ≥  . Prove that H is -connected. (Savage–Zhang [1998])

7.2.16. (−) Let G be a 2-connected n-vertex graph with m edges. Prove that in every ear
decomposition of G, the number of ears after the initial cycle is m − n.

7.2.17. (−) Let G be a 2-connected graph. Prove that if G has an ear decomposition with
initial cycle length at least l where each added ear has length at least l − 1, then G has
girth at least l. Prove that the converse is not true (provide a counterexample). (Kelmans)

7.2.18. (−) Győri [1978] and Lovász [1977] proved that a graph G with more than  ver-
tices is -connected if and only if for all distinct v1 , . . . , v¾ ∈ V(G) and n1 , . . . , n¾ ∈
summing to | V(G)| , there is a partition V1 , . . . , V¾ of V(G) such that each Vi has size ni ,
contains vi , and induces a connected subgraph. Show that the condition is sufficient.

7.2.19. Prove or disprove: If G is a 2-connected graph that is not a cycle, then G contains
a cycle C such that G − V(C) is connected.

7.2.20. For a graph G with x , u , v ∈ V(G), prove  G(u , v) ≥ min{  G(x , u) ,  G(x , v)}. Does
the analogous inequality hold for  G(u , v)?

7.2.21. (♦) The pinch operation subdivides  edges of a digraph (each becomes a path of
length 2 through a new vertex) and merges the  new vertices into a single vertex. Prove
that pinching a -edge-connected digraph always yields a -edge-connected digraph.

7.2.22. (♦) Prove that if G is an Eulerian graph, then  G(x , y) is even for all x , y ∈ V(G).
(Comment: M. Ghorbani conjectured that the converse is true; this remains open.)
Exercises for Section 7.2 313

7.2.23. Expansion Lemmas.

(a) Let G be obtained from a ¾-edge-connected multigraph G by adding a new vertex
w with degree at least ¾ (multiedges are allowed). Prove that G is ¾-edge-connected.
(b) Prove analogues of part (a) and the Expansion Lemma (Lemma 7.2.10) for digraphs.
7.2.24. (♦) Let G be a ¾-edge-connected multigraph, with ¾ ≥ 2. Prove that G has a closed
trail through any ¾ specified vertices.

7.2.25. Fix j , ¾ ∈ with j even and j ≤ ¾ . Form a ¾-regular graph G from 2K ¾+1 by delet-
ing a matching of size j/2 from each component and adding a matching of size j joining
the components. Prove (G) = j . (Kostochka)

• •
• •
• •
• •

7.2.26. A u , v-necklace is a graph consisting of cycles C1 , . . . , C¾ with u ∈ C1 and v ∈ C¾

such that consecutive cycles share one vertex and non-consecutive cycles are disjoint. Prove
that a graph G is -edge-connected if and only if for all u , v ∈ V(G) there are  pairwise
edge-disjoint paths such that the union of any two is a u , v-necklace. (T. Jiang)

u• • •v
• •

7.2.27. (♦) Let G be an n-vertex graph with connectivity  and diameter d. Prove both
n ≥ (d − 1) + 2 and (G) ≥ ⌈ (1 + d)/2⌉ . For  , d ≥ 2, construct an example such that
equality holds in both bounds. (Watkins [1967])
7.2.28. Let G be a -connected graph. For sets S, T ⊂ V(G) with size at least  , prove
that G has  pairwise disjoint S, T-paths.
7.2.29. (♦) Let X and Y be disjoint sets of vertices in a -connected graph G. Let w be a
positive integer function on X ∪ Y such that ∑ x∈ X w(x) = ∑ y∈ Y w(y) =  . Prove that there
are  independent X , Y -paths such that for each vertex v ∈ X ∪ Y , the number of these
paths having an endpoint at v is w(v).
7.2.30. Assuming that (x , y) =  (x , y) and (x , y) = (x , y) hold for digraphs (the latter
when xy is not an edge), derive the same statements for graphs.
7.2.31. (♦) Prove the K önig–Egerváry Theorem and Hall’s Theorem from Menger ’s The-
orem. Prove Hall’s Theorem from the Ford–Fulkerson CSDR Theorem.
7.2.32. (♦) Fix  ≥ 2, and let G be a -connected graph with n vertices.
(a) For n ≥ 2  , prove that G has a cycle of length at least 2  .
(b) For n ≥ 3 , prove that G has a cycle of length at least 3 if C4
⊆ G. (Kostochka)
7.2.33.  -contractible edges (see Martinov [1982]).
(a) Prove that a -connected -regular graph in which every edge lies in a triangle has
no -contractible edge. (Comment: This includes H¾ ,n of Example 7.1.3 for even  .)
(b) The Harary graph H2 ,n is also denoted Cn2 . For n ≥ 7, prove that contracting some
two edges in Cn2 leaves Cn2−2 (multiedges after contraction are turned into simple edges).
7.2.34. (+) Prove that every triangle-free -connected graph has a -contractible edge.
(Thomassen [1981b])(Comment: Thomassen used this to prove that every ( + 3)-connected
graph has a cycle such that deleting its vertices leaves a -connected graph.)
314 Chapter 7: Connectivity and Cycles

7.2.35. Let S be a set of vertices in a graph G such that (x , y) ≥  for all x , y ∈ S. Prove
that if | S| ≤  + 1, then the vertices in S lie on a single path in G. Show this cannot be
guaranteed when | S| =  + 2.
7.2.36. (♦) A graph is -linked if for every choice of distinct vertices s1 , . . . , s¾ and
t1 , . . . , t¾ , there exist disjoint paths P1 , . . . , P¾ such that Pi is an si , ti -path. Prove that
every -linked graph is (2  − 1)-connected. The graph below is 5-connected; prove that it
is not 2-linked (Watkins [1968]). For general  , construct a (3 − 3)-connected graph that
is not -linked. (Comment: Thus  () ≥ 3 − 2, where  () is the least j such that every
j-connected graph is -linked. Thomassen [1980b] and Seymour [1980] proved  (2) = 6.
The best general result is  () ≤ 10  ; in fact , 2 -connected graphs with average degree at
least 10  are -linked (Thomas–Wollan [2005]).)
•

• • •
• •
• • • •
• •
• • •

•
7.2.37. Let G be a -edge-connected graph. Prove that the line graph L(G) is -connected
and (2  − 2)-edge-connected.
7.2.38. Use Tutte’s 1-Factor Theorem to prove that every connected line graph with an
even number of vertices has a perfect matching. Conclude that every connected graph
with an even number of edges decomposes into paths of length 2.
7.2.39. (♦) Let A1 , . . . , A m and B1 , . . . , Bm be two partitions of a set E such that all the
sets Ai and Bj have the same size. Prove that the two set systems have a common system
of distinct representatives. (Ryser [1963, p. 51])
7.2.40. Let G1 and G 2 be disjoint -connected graphs with  ≥ 2. For i ∈ {1 , 2}, choose
vi ∈ V(G i) and let X i = NG i (vi). Let B be an X 1 , X 2 -bigraph that has no isolated vertex
and has a matching of size at least  . Prove that (G1 − v1) ∪ (G 2 − v2) ∪ B is -connected.
7.2.41. (♦) Let (j) be the least  such that any two vertices in a -connected graph are
joined by a path P such that G − V(P) is j-connected. Prove (1) = 3. (Tutte [1961])
(Comment: Lovász conjectured that (j) exists; Chen–Gould–Yu [2003] proved (2) ≤ 5.)
7.2.42. Let u be a vertex in a 3-connected graph G. Prove that G − u contains a cycle C
such that G − V(C) is connected.
7.2.43. Prove that applying the expansion operation of Exercise 5.2.22 to a 3-connected
graph yields a 3-connected graph. Obtain the Petersen graph from K4 by expansions.
(Comment: Tutte [1966] proved that all 3-regular 3-connected graphs arise this way.)
7.2.44. Prove that if every edge in a connected graph G is the unique common edge in some
pair of cycles, then G is 3-edge-connected. Prove this property for the Petersen graph.
7.2.45. Use induction on the distance between vertices to prove that a graph with at least
three vertices is 2-connected if and only if for any two vertices x and y there exist two
independent x , y-paths. (Whitney [1932a])
7.2.46. Let G be a graph having no induced subgraph that is C¾ for  > 3 or K4− (five
edges). Prove that every block of G is a complete graph. (Lehel)
Exercises for Section 7.2 315

7.2.47. (♦) For a connected graph G with | V(G)| ≥ 3, prove the following equivalent.
(A) G is 2-edge-connected.
(B) Every edge of G appears in a cycle.
(C) G has a closed trail containing any specified pair of edges.
(D) G has a closed trail containing any specified pair of vertices.

7.2.48. Let v be a vertex of a 2-connected graph G. Prove that v has a neighbor u such
that G − u − v is connected. (Chartrand–Lesniak [1986, p. 51])

7.2.49. Let G be an n-vertex connected graph having a cut-vertex. In terms of n, de-

termine the maximum number of edges that may need to be added to change G into a
2-connected graph. Obtain all n-vertex graphs achieving the maximum.

7.2.50. (♦) Let C be a cycle in a 2-connected graph G that is not a cycle. Prove that C
contains an ear of G whose deletion leaves a 2-connected subgraph. That is, removal of
ears in an ear decomposition of G can start by breaking any cycle.

7.2.51. (♦) Let s and t be vertices in a 2-connected graph G. Prove that the vertices of
G can be linearly ordered so that each vertex outside {s , t} has a neighbor that is ear-
lier in the order and a neighbor that is later in the order. (Hint: Use ear decompositions.
Comment: This is called an s , t-numbering of G.)

7.2.52. (♦) Let G be a 2-connected graph, and fix r ∈ V(G). Prove that G has two spanning
trees such that for every v ∈ V(G), the r, v-paths in the two trees are independent. (Hint:
Use ear decomposition to prove the stronger statement for each r that G has two trees and
a labeling of V(G) − {r} by real numbers such that in one tree the labels increase along
paths from r and in the other they decrease along paths from r.)
(Comment: Itai and Rodeh conjectured a ¾-connected graph always has ¾ such trees.
This was proved for ¾ = 3 in Itai–Zehavi [1989] and for ¾ = 4 in Curran–Lee–Yu [2006].)

7.2.53. (+) Alice and Bob play a game on a 2-connected n-vertex graph G. Alice picks ver-
tices u and v. Next Bob orients up to º (n) of the edges. Alice then orients the remaining
edges and selects an edge e, which may have been oriented by her or by Bob. If the orien-
tation contains a u , v-path through e, then Bob wins; otherwise, Alice wins. Prove that
the least º (n) such that Bob always has a winning strategy is 2n − 3. (Kerimov [2009])

7.2.54. Prove or disprove: If G is a ¾-edge-connected graph with minimum degree greater

than ¾ , then some two vertices in G are joined by more than ¾ edge-disjoint paths.

7.2.55. Let G be a minimal 2-connected n-vertex graph other than Cn .

(a) Prove that if P is the last ear in an ear decomposition of G, then the graph G
formed before adding P is a minimal 2-connected graph. (Dirac [1967], Plummer [1968])
(b) For n ≥ 4, prove that G has at most 2n − 4 edges, with equality holding only when
G = K 2 ,n−2 . (Comment: Mader [1972] proved that for n ≥ 3¾ − 2, a minimal ¾-connected
n-vertex graph has at most ¾(n − ¾) edges, achieved only by K ¾ ,n−¾ when n ≥ 3¾ − 1.)

7.2.56. (♦) Minimal 2-connected graphs. (Dirac [1967], Plummer [1968])

(a) For a 2-connected graph G, prove that G − xy is 2-connected if and only if x and y
lie on a cycle in G − xy. Conclude that a 2-connected graph is a minimal 2-connected graph
if and only if no cycle has a chord.
(b) Vertices x and y are compatible in a graph G if no x , y-path in G has a chord. For
t ≥ 2, let each of G1 , . . . , G t be an edge or a minimal 2-connected graph such that xi and
yi are compatible vertices in G i . Form G by merging yi with xi+1 for 1 ≤ i < t and y¾ with
x1 . Prove that G is a minimal 2-connected graph, except when ¾ = 2 and K 2 ∈ {G1 , G 2 }.
Conversely, prove that every minimal 2-connected graph can be constructed from edges
and smaller 2-connected graphs in this way. (Hint: Use part (a).)
316 Chapter 7: Connectivity and Cycles

7.2.57. Given vertices x and y in a minimal 2-connected graph G such that no x , y-path
has a chord, prove that every x , y-path has an internal vertex with degree 2 in G. Use
this to improve Corollary 7.2.33 by showing that every minimal 2-connected graph with n
vertices has at least (n + 4)/3 vertices of degree 2. (Dirac [1967], Plummer [1968])
7.2.58. (♦) Applications of Mader’s Theorem (Theorem 7.2.32).
(a) Prove: a minimal ¾-connected graph has at least ¾ vertices of degree ¾ .
(b) Let S be any ¾-set in a minimal ¾-connected graph G, and let C be a component of
G − S. Prove that C contains a vertex of degree ¾ in G (see Mader [1972]).
(c) Prove that a minimal ¾-connected graph G has at least ¾ + 1 vertices of degree ¾ .
(Comment: This improves Corollary 7.2.33 when | V(G)| ≤ 2 ¾ .)
7.2.59. Construct a graph G ¾ ,m as follows. Begin with mK ¾−1 ,¾ , having m(¾ − 1) vertices
yr, j and m¾ vertices xr,i such that yr, j and xr,i are adjacent for i ∈ [¾] and j , r ∈ [m]. Also
make xr,i and xr+1 ,i adjacent for i ∈ [¾] and r ∈ [m − 1]. Finally, add two vertices and of
degree  having neighbors {x1 ,i } and {x m ,i }, respectively. Prove that G ¾ ,m is a minimal -
connected graph with (¾−21)n +2¾
¾ −1 vertices of degree  . (Bollobás [1978, p. 25], Mader [1979])

7.3. Spanning Cycles

A ¾-connected graph has a cycle through any ¾ vertices (Theorem 7.2.13), but
cycles through larger sets are not guaranteed (Exercise 7.2.11). However, when
the minimum degree is large enough in terms of the number of vertices, there is
a cycle through all the vertices. We first consider conditions for spanning cycles,
and later we discuss “long cycle” versions of these results.
Spanning cycles were introduced by K irkman in 1855 in a paper on polyhedral
graphs. Nevertheless, they are called Hamiltonian cycles to honor Sir William
Rowan Hamilton (1805–1865). In studying non-commutative algebra, he devised
a game on the dodecahedron graph where one player would specify a 5-vertex path
and the other would extend it to a spanning cycle. Hamilton sold “ The Icosian
Game” to a dealer in puzzles, for £25. In another version, “ The Traveller ’s Do-
decahedron”, the vertices were named for 20 important cities. Neither was com-
mercially successful, but Hamilton’s name was associated with the concept.
•
•
• • • •
• • • •
• • • • •
• • •
• •
A graph with a spanning cycle is a Hamiltonian graph. Until the 1970s,
interest in spanning cycles was related to the Four Color Problem (Chapter 9).
Later, the Traveling Salesman Problem (TSP) became an important compu-
tational problem: it asks for the minimum weight of a spanning cycle in an edge-
weighted input graph. Testing for a spanning cycle is a special case and is NP-
complete, even for 3-regular triangle-free planar graphs or for line graphs.
Section 7.3: Spanning Cycles 317

Survey papers about Hamiltonian graphs and related topics include Bondy
[1978a, 1995], Gould [1991, 2003, 2014], Faudree [1996, 2001], Lesniak [1996],
K awarabayashi [2001], Yamashita [2004], Li [2013], K ühn–Osthus [2014], and
many others. We will not discuss analogues for directed graphs, surveyed in
Bermond–Thomassen [1981] and Zhang–Song [1991].

PROPERTIES OF HAMILTONIAN GRAPHS

Every Hamiltonian graph is 2-connected, since deleting any vertex leaves a

spanning path. Bipartite graphs suggest an extension of this condition.
7.3.1. Example. Bipartite graphs. A spanning cycle in a bipartite graph visits
the two parts alternately, so such a cycle requires the parts to have the same size.
In particular, K r,s is not Hamiltonian when r < s. Note that deleting the part of
size r leaves more than r components.
7.3.2. Proposition. If G has a Hamiltonian cycle, then #(G − S) ≤ | S| whenever
∅
= S ⊆ V(G), where #(H) denotes the number of components of H.
Proof: Follow a Hamiltonian cycle C. When C leaves a component of G − S, it
can go only to S. Arrivals in S must be at distinct vertices of S, so S must have
as many vertices as G − S has components.

• • • •
•
• S •
• • •
• •

The necessary condition #(G − S) ≤ | S| for all nonempty S sounds like Tutte ’s
1-Factor Condition, but it is not sufficient for a spanning cycle.
7.3.3. Example. Necessary but not sufficient. The graph on the left below fails
the condition of Proposition 7.3.2, even though it is bipartite with parts of equal
size. Hence it is not Hamiltonian. (Tutte [1971] conjectured that all 3-connected
3-regular bipartite graphs are Hamiltonian; counterexamples were found with
96 vertices (Horton [1982]) and eventually only 50 vertices (Georges [1989]).
• • •
• •
• •
• •
• • • • •••
• •
• • •
• • • •
• •
In the graph on the right, the necessary condition holds, but there is no span-
ning cycle; both edges incident to any vertex with degree 2 must be used, which
forces three edges at the central vertex. Exercise 28 generalizes this example.
318 Chapter 7: Connectivity and Cycles

7.3.4.* Remark. Perhaps strengthening the necessary condition by requiring

|S| ≥ t · #(G − S) for some t guarantees a spanning cycle. A connected graph G is
t-tough if | S| ≥ t · #(G − S) for every separating set S, and the toughness of G
is the maximum such t. By Proposition 7.3.2, Hamiltonian graphs are 1-tough.
For example, the Petersen graph G is 4/3-tough (hence also 1-tough), since
#(G − S) = 2 ⇒ | S| ≥ 3, #(G − S) = 3 ⇒ | S| ≥ 4, #(G − S) = 4 ⇒ | S| = 6,
and no larger value of #(G − S) is possible since (G) = 4. The Petersen graph is
not Hamiltonian (Proposition 5.1.18), so 4/3-tough is not sufficient.
Chvátal [1973] conjectured that some t suffices. Many thought t = 2 would be
enough, but non-Hamiltonian graphs are now known with toughness approach-
ing 9/4 (Bauer–Broersma–Veldman [2000]). For graphs having no induced cycle
of length at least 4, Chen–Jacobson–K ézdy–Lehel [1998] showed that toughness
18 suffices, and K abela–K aiser [2017] improved that to 10. Enomoto–Jackson–
K aterinis–Saito [1985] proved that every ¾-tough n-vertex graph has a ¾-factor
(Chapter 6) when n > ¾ and n¾ is even. Broersma [2002] and Bauer–Broersma–
Schmeichel [2006] surveyed results on toughness.

For regular graphs with even degree, one can study decompositions into span-
ning cycles, called Hamiltonian decompositions (we write “H-decomposition”).
The complete graph K 2r+1 has H-decompositions (Exercise 5.3.48). Sloane [1969]
asked whether every 4-regular graph with an H-decomposition has another.

7.3.5.* Theorem. (Thomason [1978]) The number of Hamiltonian decomposi-

tions of any loopless 4-regular multigraph G with at least three vertices is
even. Furthermore, for any two edges e and e in G, the number of Hamilto-
nian decompositions having e and e in the same cycle is even.
Proof: Let R(G) be the set of H-decompositions of G. For e , e ∈ E(G), let P G(e , e )
and Q G(e , e ) denote the sets of H-decompositions having e and e in the same or
different cycles, respectively. Thus P G(e , e ) ∪ Q G(e , e ) = R(G). Let the sizes of
these sets be pG(e , e ), qG (e , e ), and r(G), respectively.
Let v be an endpoint of e. In every H-decomposition, the cycle through e con-
tains another edge incident to v. Hence r(G) is even if pG(e , e ) is even whenever e
and e are incident edges. With r(G) even, the stronger statement that pG(e , e ) is
even for all e , e ∈ E(G) is equivalent to having qG(e , e ) even for all e , e ∈ E(G).
We prove the claims by induction on | V(G)|. In the only loopless 4-regular 3-
vertex multigraph, each vertex pair forms a double edge. When e and e share
both endpoints, pG(e , e ) = 0; when they share one endpoint, pG(e , e ) = 2.
Now consider larger G. We prove first that pG(e , e ) is even when e and e
have a common endpoint v. Let º and º be the other two edges incident to v. If e
and e share both endpoints or º and º share both endpoints, then pG(e , e ) = 0.
Otherwise, obtain Ĝ from G by deleting v, adding an edge ê joining the other
endpoints of e and e , and adding an edge ºˆ joining the other endpoints of º and
º . Note that Ĝ is a loopless 4-regular multigraph, smaller than G.
The H-decompositions of G in P G (e , e ) put º and º into the cycle not con-
taining e and e ; hence they correspond to the H-decompositions of Ĝ in Q Ĝ(ê , ºˆ).
By the induction hypothesis, R(Ĝ) and Q Ĝ(ê , ºˆ) have even size. Hence pG(e , e ) is
even, since it equals qĜ(ê , ºˆ). This holds for all incident e and e , so r(G) is even.
Section 7.3: Spanning Cycles 319

Now consider any e , e ∈ E(G). If R(G)

= ∅, then G is connected, and some
path P contains e and e . We show by induction on the length of P that pG(e , e )
is even. We have shown it when E(P) = {e , e }, the base case.
When e and e are not incident, let ê be the edge before e on P. Note
that Q G(e , e ) = P G (e , ê) P G (ê , e ), since e and e are in different cycles in a H-
decomposition if and only if ê is in exactly one of those cycles. Since we have al-
ready proved that P G (ê , e ) and P G (e , ê)(shorter path) have even size, also qG(e , e )
is even. Since we already proved that r(G) is even, also pG(e , e ) is even.

Applications of Theorem 7.3.5 appear in Exercises 32–33. For the number of

Hamiltonian cycles, Corollary 5.4.8 proves the result of Smith that a 3-regular
graph has an even number of Hamiltonian cycles through any edge. This does not
imply that a 3-regular graph has an even number of spanning cycles! The com-
plete graph K 4 and the cartesian product C3 K 2 each have exactly three spanning
cycles; every edge lies in two of them. Kotzig proved (see Bosák [1967]) that every
bipartite 3-regular graph has an even number of spanning cycles.

SUFFICIENT CONDITIONS

The number of edges needed to force a spanning cycle in an n-vertex graph is

(n−2 1 ) + 2 (Exercise 36). Conditions that “spread out ” the edges permit a smaller
threshold. The idea underlying such results is the edge-switching technique ex-
tracted by Ore from an earlier argument of Dirac.

7.3.6. Lemma. (Ore’s Lemma; Ore [1960]) Let x and y be distinct nonadjacent
vertices of an n-vertex graph G. If d(x) + d(y) ≥ n, then G is Hamiltonian if
and only if G + xy is Hamiltonian.
Proof: If G is Hamiltonian, then so is G + xy. If G + xy is Hamiltonian but G is
not, then xy lies in each spanning cycle in G + xy. Index the vertices from v1 to
vn along a spanning x , y-path in G.
If some neighbor of x immediately follows a neighbor of y on the path,
say vi+1 ∈ N(x) and vi ∈ N(y), then omitting vi vi+1 yields the spanning cycle
[x , vi+1 , vi+2 , . . . , y , vi , vi−1 , . . . , v2 ] in G shown below.

• • • •
x vi vi+1 y
To guarantee such a cycle, we prove that S and T have a common element,
where S = {i: vi+1 ∈ N(x)} and T = {i: vi ∈ N(y)}. Summing their sizes,
|S ∪ T | + |S ∩ T | = | S| + | T | = d(x) + d(y) ≥ n.
Neither S nor T contains the index n. Thus | S ∪ T | < n, and hence | S ∩ T | ≥ 1.
We conclude that G has a spanning cycle.

7.3.7. Corollary. (Ore’s Theorem; Ore [1960]) Let G be an n-vertex graph. If

n ≥ 3 and d(x) + d(y) ≥ n whenever xy ∈
/ E(G), then G is Hamiltonian. (Let-
ting 2(G) = min{d(x) + d(y): xy ∈
/ E(G)}, Ore’s Condition is 2(G) ≥ n.)
320 Chapter 7: Connectivity and Cycles

Proof: Since K 2 satisfies Ore ’s Condition but is not Hamiltonian, we need n ≥ 3.

# #
For n ≥ 3, we use induction on #### E(G)#### ; the basis step is K n . When xy ∈ E(G),

adding xy to G yields a graph G that also satisfies Ore ’s Condition. By the in-
duction hypothesis, G is Hamiltonian, and then by Lemma 7.3.6 so is G.

Ore ’s Condition is immediately implied by Dirac’s Condition, which is the

stronger hypothesis (G) ≥ n/2. Thus the sufficiency of Dirac’s Condition is a
weaker result than the sufficiency of Ore ’s Condition.

7.3.8. Corollary. (Dirac’s Theorem; Dirac [1952b]) For n ≥ 3, an n-vertex

graph G with (G) ≥ n/2 is Hamiltonian, and this threshold is sharp.
Proof: Again n ≥ 3 is needed, because K 2 satisfies the condition but is not Hamil-
tonian. For n ≥ 3, the non-Hamiltonian graph K ⌊(n−1)/2⌋ ,⌈(n+1)/2⌉ shows that the
bound is sharp. For sufficiency, when (G) ≥ n/2 we have d(x) + d(y) ≥ n when-
ever x , y ∈ V(G), and Corollary 7.3.7 applies.

Using Lemma 7.3.6 to add edges, we can test whether G is Hamiltonian by

testing whether a larger graph is Hamiltonian.

7.3.9. Definition. A (Hamiltonian) closure of an n-vertex graph G, denoted

C(G), is obtained from G by iteratively adding edges joining nonadjacent
vertices whose degree sum is at least n, until no such pair remains. We say
G is closed when C(G) = G.
• • • •
• • • • • • • •
→ → →
• • • • • • • •
• • • •

7.3.10. Theorem. (Bondy–Chvátal [1976]) The closure of a graph is unique (does

not depend on the order of edge additions), and a graph is Hamiltonian if and
only if its closure is Hamiltonian.
Proof: From G, form closures G 1 by adding the edges e1 , . . . , er and G 2 by adding
the edges º1 , . . . , ºs . If the degree-sum of two nonadjacent vertices reaches
|V(G)| , then they must be made adjacent. Thus if º1 , . . . , ºi−1 ∈ E(G 1), then
ºi becomes addable and must belong to G 1 . Hence neither list contains a first
edge omitted by the other, so G 1 ⊆ G 2 and G 2 ⊆ G 1 .
The second statement follows immediately from Lemma 7.3.6.

Bondy–Chvátal [1976] considered many related concepts and closures by

varying the threshold for d(x) + d(y). Material on these and other closure con-
cepts includes Ryj áček [1997] and Broersma–Ryjáček–Schiermeyer [2000].
Theorem 7.3.10 characterizes Hamiltonian graphs but does not provide a good
algorithm to test whether a graph is Hamiltonian. Nevertheless, it provides
a method for proving sufficient conditions. A condition that forces C(G) to be
Hamiltonian (as when C(G) = K n) forces a Hamiltonian cycle in G. Chvátal used
this method to find the best possible sufficient condition using only vertex degrees.
The condition on the degree list in this theorem is called Chvátal’s Condition.
Section 7.3: Spanning Cycles 321

7.3.11. Theorem. (Chvátal’s Theorem; Chvátal [1972]) For n ≥ 3, let G be

a graph with vertex degrees d1 , . . . , d n in nondecreasing order. If d i > i or
d n−i ≥ n − i whenever i < n/2, then G is Hamiltonian.
Proof: By Theorem 7.3.10, it suffices to prove that if G satisfies Chvátal’s Con-
dition, then C(G) = K n , since a complete graph is Hamiltonian. Adding edges
reduces no entries in the degree list, so Chvátal’s Condition is preserved by tak-
ing the closure. Thus it suffices to prove the implication in the case where G is
already closed. We prove the contrapositive: if G is closed and G
= K n , then
Chvátal’s Condition fails. Violation means that for some i less than n/2, at least
i vertices have degree at most i and at least n − i have degree less than n − i.
Since G
= K n , we may pick a pair {u , v} of nonadjacent vertices with largest
degree sum. Because G is closed, uv ∈ / E(G) requires d(u) + d(v) < n; label u and
v so that d(u) ≤ d(v). Since d(u) + d(v) < n, we obtain d(u) < n/2. Let i = d(u).
Since we chose {u , v} with largest degree sum, every nonneighbor of v has
degree at most d(u), and there are n − 1 − d(v) of them. Since n − 1 − d(v) ≥ d(u) =
i, at least i vertices have degree at most i.
Similarly, every nonneighbor of u has degree at most d(v), and for u there are
n − 1 − d(u) nonneighbors. Since d(u) ≤ d(v), also u itself has degree at most d(v).
Since d(v) < n − d(u) = n − i, these n − i vertices have degree less than n − i.
Hence we have proved that d i ≤ i and d n−i < n − i when i = d(u), so Chvátal’s
Condition does not hold.

• u N(v) N(u) •v

7.3.12. Example. Non-Hamiltonian graphs with “large” vertex degrees. Theorem

7.3.11 characterizes the degree lists of graphs that force Hamiltonian cycles. If G
fails Chvátal’s Condition at the value i, then the largest we can make the terms
in d1 , . . . , d n is d j = i for j ≤ i, d j = n − i − 1 for i + 1 ≤ j ≤ n − i, and d j = n − 1 for
j > n − i. The unique graph with this degree list is not Hamiltonian. The graph
is (K i + K n−2i) K i ; it is not Hamiltonian because deleting the i vertices of degree
n − 1 leaves a subgraph with i + 1 components.

iK 1 Ki K n−2i

7.3.13. Remark. A Hamiltonian path is a spanning path. Results for spanning

cycles yield analogous results for paths by a standard transformation: G has a
spanning path if and only if G K 1 has a spanning cycle. For example, 2(G) ≥
n − 1 yields a spanning path (Ore [1961]); see also Exercise 3 and Exercise 38.

Our next sufficient condition involves connectivity and independence. The

proof yields a good algorithm to find a spanning cycle or contradict the hypoth-
esis. The theorem strengthens the degree results in the probabilistic sense (see
Chapter 14); the Chvat ál–Erdős Condition almost always holds, while the degree
conditions do not. Surveys of material related to the Chvátal–Erdős Theorem
include Jackson–Ordaz [1990] and Saito [2008].
322 Chapter 7: Connectivity and Cycles

7.3.14. Theorem. (Chvátal–Erdős Theorem; Chvátal–Erdős [1972]) If (G) ≥

(G), then G has a Hamiltonian cycle (unless G = K 2).
Proof: Let  = (G) ≥ (G). With G
= K 2 , the conditions require (G) > 1, so
there is a longest cycle C in G. Since (G) ≥ (G), and since every graph with
(G) ≥ 2 has a cycle of length at least (G) + 1 (Exercise 5.3.38), the length of C
is at least  + 1. Let H be a component of G − V(C). Since (G) =  , at least 
vertices of C have neighbors in H.
Let u1 , . . . , u¾ be vertices of C with neighbors in H , indexed in order along
C. For each i, let ai be the vertex following ui along C. If ai and aj are adjacent,
then we construct a longer cycle by replacing ui ai and uj aj with ai aj and a ui , uj -
path through H (see illustration). Similarly, ai has no neighbor in H. Hence
{a1 , . . . , a l} plus a vertex of H forms an independent set of size greater than  .
This contradiction implies that C is a Hamiltonian cycle.
ui
• •
• ai
•
• • H •
•

• • •uj
aj

The graph K r,r+1 shows that (G) ≥ (G) − 1 is not sufficient.

7.3.15.* Remark. Lu [1994] strengthened the Chvátal–Erdős Theorem by prov-
ing that an n-vertex graph G is Hamiltonian when the following inequality holds
for every nonempty proper subset S of V(G):
### #
## N(S) ∩ S#### (G)
## ## ≥ .
### S### n
To obtain the Chvátal–Erdős Theorem from Lu’s Theorem, we view the latter as
lowering the threshold for the size of a separating set. Let S be the vertex set of a
component of G − T , where T is a smallest separating set. Lu’s Condition requires
## # ## ##
### N(S) ∩ S#### ≥ (G)
n ### S ###. In fact, N(S) ∩ S = T , since T is a minimal separating set.
Thus | T | ≥ (G)(1 − | Sn | ). Since | T | = (G), the Chvátal–Erdős Condition implies
this, so Lu’s Theorem implies the Chvátal–Erdős Theorem. This greatly reduces
the required size of a separating set T when the components of G − T are large.
Cuts that isolate single vertices still must have size at least (G).

LONG CYCLES (optional)

When a sufficient condition for spanning cycles fails slightly, we still expect
a long cycle. The maximum length of a cycle in G is its circumference, c(G).
We begin with the number of edges to force c(G) > q. We also consider long-cycle
versions of conditions for spanning cycles.
Section 7.3: Spanning Cycles 323

7.3.16. Theorem. (Erdős–Gallai [1959]) For q ≥ 2, if an n-vertex graph G has

more than q(n − 1)/2 edges, then c(G) > q.
Proof: (Woodall [1972]) Fix q; we use induction on n. When n = q + 1, fewer
than (n − 1)/2 edges are missing, so (G) ≥ n/2 and G is Hamiltonian. Consider
n > q + 1 and c(G) ≤ q. If d(x) ≤ q/2, then | E(G − x)| > q(n − 2)/2. Applying the
induction hypothesis to G − x yields c(G − x) > q. Hence we may assume (G) >
q/2. Using the induction hypothesis we may also assume that G is connected.
Among all longest paths in G, choose P to maximize the degree of the first
vertex. Let P = ⟨v1 , . . . , vl ⟩ and d = d G(v1). Since G is connected, v1 vl ∈ / E(G)
(otherwise an edge from V(P) to V(G) − V(P) would yield a longer path). Let
W = {vi : vi+1 ∈ N(v1)}. All neighbors of v1 lie on P , so | W | = d.
For v¾ ∈ W , let Q be the path obtained from P by replacing v¾ v¾+1 with v1 v¾+1
(see figure below). Since Q has the same length (and vertex set) as P , we have
N(v¾) ⊆ V(P), and the choice of P yields d G(v¾) ≤ d. Avoiding longer paths and
cycles with length greater than q yields N(v¾) ⊆ Z, where Z = {v1 , . . . , vmin{l ,q} }.
We bound w, the number of edges incident to W . Since N(v) ⊆ Z for v ∈ W ,
# # # #
we have 2w = ∑v∈W d G(v) + #### [W, Z − W]#### . Since #### [W, Z − W]#### ≤ | W | | Z − W | and
d G(v) ≤ d = | W | , we obtain w ≤ 2 (d + d(| Z | − d)) ≤ 2 dq.
1 2 1

Therefore, G − W has n − d vertices and more than q(n − d − 1)/2 edges. Using
the induction hypothesis, c(G) ≥ c(G − W) > q.
v1 v¾ v¾+1 vl
• • • •

Theorem 7.3.16 is sometimes sharp; Kopylov [1977] gave the complete answer.
Next, the long-cycle result of Dirac [1952b] will follow from our later results.

7.3.17. Theorem. For G a 2-connected n-vertex graph, c(G) ≥ min{n , 2 (G)}.

Theorem 7.3.17 improves the observation that (G) = ¾ guarantees a cycle of

length at least ¾ + 1 (Exercise 5.3.38). Requiring 2-connectedness eliminates the
example K 1 2K ¾+1 , which has circumference ¾ + 1.
The long-cycle version of Ore ’s Theorem (Corollary 7.3.7) came much later.
It is implicit in Bondy [1971b] and was explicit in Bermond [1976] and in Linial
[1976]. The fundamental argument used in this and many other long-cycle re-
sults appears in Bondy [1971b]. Called Bondy’s Lemma, it strengthens the
Ore/Dirac switching argument (Lemma 7.3.6) by considering “gaps”.

7.3.18. Lemma. (Bondy [1971b]) If G is a 2-connected n-vertex graph, then

c(G) ≥ min{n , d(x) + d(y)}, where x and y are the ends of a longest path P.
Proof: Suppose c(G) < n. Let P = ⟨v1 , . . . , vl ⟩, where v1 = x and vl = y. For
i , j ∈ [l], let Pi , j denote the vi , vj -path in P. Since G is connected, l-cycles are
forbidden (they yield longer paths or a spanning cycle). In particular, xy ∈ / E(G).
Case 1: P has a crossover: edges xvj and vi y with i < j. Adding xvj and vi y to
P1 ,i ∪ P j ,l yields a cycle with length l − (j − i − 1).

• • • •
x = v1 vi vj y = vl
324 Chapter 7: Connectivity and Cycles

Given a crossover {xvj , vi y} with j − i smallest, x and y have no neighbors

between vi and vj on P. Also N(y) contains no predecessor on P of a neigh-
bor of x, since l-cycles are forbidden. Hence N(y) lies in V(P) − {y} but avoids
{vi+1 , . . . , v j −2 } and {vr−1 : vr ∈ N(x)}. Thus d(y) ≤ (l − 1) − (j − 2 − i) − d(x), which
simplifies to l − (j − i − 1) ≥ d(x) + d(y), so the crossover cycle is long enough.
Case 2: P has no crossover. We construct a cycle containing x and y and all
their neighbors. Since avoiding crossovers requires | N(x) ∩ N(y)| ≤ 1, such a cy-
cle has length exceeding d(x) + d(y).
Let t0 = max{ j: vj ∈ N(x)} and u = min{ j: vj ∈ N(y)}. We define paths
Q1 , Q2 , . . .. Given t i−1 , choose integers si and t i , with si < t i−1 < t i and t i as large
as possible, so that G has a path Qi from vsi to vti that has no internal vertex on
P. Such a path exists when t i−1 < l since G − vti−1 is connected. If Qi and Qj with
i < j share an internal vertex, then Qi would follow Qj after the intersection to
reach vtj , farther than vti . Hence the paths are disjoint. Similarly, si+1 ≥ t i−1 ,
since otherwise Qi+1 would be chosen as Qi .
Let r be the least index such that tr > u (r = 5 in the figure below). Set
a = min{ j: vj ∈ N(x) and j > s1 } , b = max{ j: vj ∈ N(y) and j < tr }.
Since s1 < t0 and tr > u, the indices a and b are well-defined. We use the paths
Qi for even i to build one x , y-path; we build another using the paths Qi for odd
i. When r is odd, the two paths are as follows, where Q : Q denotes the union of
paths Q and Q when the end of Q is the beginning of Q .
xva : P a ,s2 : Q2 : Pt2 ,s4 : Q4 : · · · : Ptr−1 ,b : vb y

P1 ,s1 : Q1 : Pt1 ,s3 : Q3 : Pt3 ,s5 : · · · : Qr : Ptr ,l

Q2 Q r −1
a t0 s2 t2 s4 t r −1 u b
1 s1 t1 s3 t3 sr tr l
Q1 Q3 Qr
When r is even, the path starting with xva reaches tr and ends with Ptr ,l , while
the other path reaches vb and ends with vb y.
Since si+1 ≥ t i−1 ,
s 1 < a ≤ t 0 ≤ s 2 < t 1 ≤ s3 < t 2 · · · < t r −1 ≤ u ≤ b < t r
Thus the two concatenations are x , y-paths forming a cycle. The definitions of a
and b yield N(x) ⊆ V(P1,s1 ∪ P a ,t0 ) and N(y) ⊆ V(Pu ,b ∪ Ptr ,l), so the cycle includes
x and y and their neighbors and has length exceeding d(x) + d(y).

Bondy ’s Lemma implies the long-cycle version of Ore ’s Theorem, which

strengthens the long-cycle version of Dirac’s Theorem.

7.3.19. Theorem. If G is a 2-connected n-vertex graph and d(u) + d(v) ≥ q for

every nonadjacent pair u , v ∈ V(G), then c(G) ≥ min{n , q}.
Proof: Ore ’s Theorem provides a spanning cycle if q ≥ n, so we may assume q < n.
The endpoints of a longest path are nonadjacent, since G is connected. By hypoth-
esis, their degree sum is at least q, so Bondy ’s Lemma completes the proof.
Section 7.3: Spanning Cycles 325

Fan [1984] strengthened Theorem 7.3.19 by weakening the degree condition

and by requiring it only for some nonadjacent pairs. As shown next, Fan’s The-
orem (Theorem 7.3.21) yields a sufficient condition for Hamiltonian cycles that
does not force the closure to be complete (see Exercise 50 for another).

7.3.20. Example. A Hamiltonian graph. With n = 4r, let G 1 = K 2r and

G 2 = rK 2 , and form G by adding a matching joining V(G 1) and V(G 2). The
Hamiltonian closure of G is G itself, so our previous sufficient conditions do
not apply. Although G has n/2 vertices of degree 2, Fan’s Theorem only needs
max{d(x) , d(y)} ≥ n/2 when d G(x , y) = 2 and implies that G is Hamiltonian.

• • • • • •
• • • • • •

Tian [1988] shortened the proof by applying Bondy ’s Lemma. Later, Tian
[2004] shortened it yet again, obtaining the needed longest path more directly.
This improvement uses the switching idea in the proof of Theorem 7.3.16.

7.3.21. Theorem. (Fan’s Theorem; Fan [1984]) Let G be a 2-connected n-vertex

graph. If d G(u , v) = 2 implies max{d(u) , d(v)} ≥ q/2, then c(G) ≥ min{n , q}.
Proof: (Tian [2004]) Let U = {v ∈ V(G): d(v) ≥ q/2}. Among all longest paths in
G, let Q be one with the most endpoints in U. If both ends are in U , then Bondy ’s
Lemma yields c(G) ≥ min{n , q}. Hence we may assume that Q starts outside U.
Let x be the last vertex of Q. Among all longest paths in G, let P be the set of
those ending at x. By the choice of Q, all paths in P start outside U. Let ¾ be the
maximum distance along a path in P from the first vertex to its last neighbor; let
P be a path achieving this. Thus P has vertices v0 , . . . , vl in order, with v¾ being
the last neighbor of v0 in the list. Since P ∈ P, we have v0 ∈ / U.
Also N(v0) ⊆ W , where W = {v1 , . . . , v¾ }. We claim W ⊆ N(v0). Otherwise,
let vt be the last vertex of W missing from N(v0). Note that vt , . . . , v0 , vt+1 , . . . , vl
in order form a path in P beginning at vt , so vt ∈ / U. This contradicts the hypoth-
esis, since d G(v0 , vt) = 2 but v0 , vt ∈
/ U. Thus N(v0) = {v1 , . . . , v¾ }.
For 1 ≤ j ≤ ¾ − 1, the path v j , . . . , v0 , v j +1 , . . . , vl also lies in P. It starts
with v0 , . . . , v¾ in some order. Hence the choice of P in P to maximize the index
of the last neighbor of the first vertex yields N(vj) ⊆ {v0 , . . . , v¾ }.
We have shown that v¾ is the only vertex outside {v0 , . . . , v¾−1 } having neigh-
bors in this set. Hence G − v¾ is disconnected, which is a contradiction since G is
2-connected. We conclude that some longest path has both endpoints in U.

FURTHER DIRECTIONS (optional)

The results of Dirac, Ore, and Chvátal are sharp but only begin the study
of cycles. Besides generalizing to long cycles as discussed above, we can consider
restricting G or modifying the hypotheses or conclusions in various ways.
326 Chapter 7: Connectivity and Cycles

7.3.22. Remark. Restricted classes. When we restrict attention to spanning sub-

graphs of K n/2 ,n/2 , the degree threshold cannot be larger; indeed, we hope for
an easier threshold. In fact, in this case minimum degree (n + 1)/4 suffices for a
spanning cycle (Moon–Moser [1963]; Exercise 39), and (G) = n/4 is not enough.
Bagga–Varma [1999] provides a survey.
We can also restrict attention to regular graphs. Every -regular n-vertex
graph with n ≤ 3 is Hamiltonian (Jackson [1980]); thus regularity drops the
degree threshold from n/2 to n/3. Only the Petersen graph prevents lowering it
to (n − 1)/3 (Zhu–Liu–Yu [1985], partly simplified in Bondy–Kouider [1988]). The
example K r,r+1 shows how crucial regularity is.
Spanning cycles are also studied in special classes of graphs. A famous conjec-
ture of Matthews–Sumner [1984] asserts that every 4-connected graph with no
four vertices inducing K 1 ,3 is Hamiltonian. Connectivity at least 7 is sufficient
(Jackson [1989, unpublished], Zhan [1991]). Another proof of this by Ryjaček
[1997] reduced the conjecture to the case of line graphs of triangle-free graphs.
The prism over a graph G is the cartesian product graph G K 2 , and G
is prism-Hamiltonian when G K 2 has a spanning cycle. Every graph having
a spanning path is prism-Hamiltonian, but spanning paths are not necessary.
The concept is interesting because a spanning cycle in G K 2 collapses to a 2-
walk in G, which is a spanning closed walk that visits each vertex at most twice
(a weaker version of a spanning cycle). Many classes of graphs are known to be
prism-Hamiltonian, including all 3-connected 3-regular graphs (Paulraja [1993]).
K aiser–Ryjáček–Král’–Rosenfeld–Voss [2007] contains a survey.
For graphs on surfaces, see Ellingham [1996] and Holton–Aldred [1999].

7.3.23. Remark. Stronger conclusions, weaker hypotheses. A graph with cycles of

all lengths (3 to | V(G)|) is pancyclic. A “meta-conjecture” of Bondy [1971a] as-
serts that sufficient conditions for spanning cycles usually imply pancyclicity with
few exceptions. Bondy [1971a] showed that n-vertex graphs satisfying Ore ’s Con-
dition are pancyclic except for K n/2 ,n/2 (see Exercise 25). With suitable exceptions,
this also holds for Chvátal’s Condition (Schmeichel–Hakimi [1974]) and Fan’s
Condition (Tian–Shi [1986], Benhocine–Wojda [1987]); see Mitchem–Schmeichel
[1985] for a survey. Schmeichel–Hakimi [1988] showed that if some spanning cy-
cle has consecutive x and y with d(x) + d(y) ≥ n, then the graph is pancyclic or
bipartite or lacks only cycle length n − 1 (pancyclic if d(x) + d(y) ≥ n + 1). Using
this, Bauer–Schmeichel [1990] gave unified proofs for pancyclicity (with excep-
tions) under the conditions of Chvátal [1972], Fan [1984], and Bondy [1980].
Strengthening a sufficient condition for spanning cycles often forces more
structure. P ósa [1963/4] proved that if the degree sum of nonadjacent vertices is
always at least n + , then any set of edges forming a disjoint union of paths can
be extended to a Hamiltonian cycle (Exercise 49). Berge conjectured the same con-
clusion when (G) ≥ (G) + , proved in Häggkvist–Thomassen [1982]. Zamani–
West [2012] proved the analogue for bipartite graphs. For a survey on disjoint
cycles (and paths), see Chiba–Yamashita [2018].
A graph G is Hamiltonian-connected if it has a spanning u , v-path when-
ever u , v ∈ V(G). Often a slight strengthening of a sufficient condition for Hamil-
tonian graphs implies that a graph is Hamiltonian-connected (Exercises 44–45).
Section 7.3: Spanning Cycles 327

Ore ’s Theorem (Corollary 7.3.7) has been strengthened by considering inde-

pendent sets of size greater than 2 and by considering neighborhood unions in-
stead of degree-sums(see Bondy [1981], Faudree–Gould–Jacobson–Schelp [1989],
Faudree–Gould–Jacobson–Lesniak [1992]).

7.3.24. Remark. Weakening the conditions for spanning paths. Spanning paths
are usually guaranteed by slight weakenings of sufficient conditions for spanning
cycles, since G has a spanning path if and only if G K 1 has a spanning cycle (Re-
mark 7.3.13). Dirac’s Condition, Ore ’s Condition, and the Chvátal–Erdős Con-
dition can be modified in this way to guarantee a spanning path in an n-vertex
graph G when (G) ≥ (n − 1)/2, or 2(G) ≥ n − 1, or (G) ≤ (G) + 1.
Viewing a spanning path as a restricted spanning tree leads to many gener-
alizations. A spanning path has maximum degree 2, no branch vertices, only two
leaves, etc. Weakening the conditions for a spanning path may guarantee sim-
ilarly weakened spanning trees. Ozeki–Yamashita [2011] provides a thorough
survey about such problems, including what we mention here (see also Kouider–
Vestergaard [2005]). Say that a spanning tree is -bounded if it has maximum
degree at most  , -branched if it has at most  branch vertices (vertices with
degree at least 3), and -ended if it has at most  leaves.
For example, Gargano–Hammar–Hell–Stacho–Vaccaro [2004] proved that
 (G) ≥ (n − 1)/3 guarantees a 1-branched spanning tree. Strengthening ear-
lier conjectures, Ozeki–Yamashita [2011] conjectured that (G) ≥ (n − )/( + 3)
guarantees a -branched spanning tree, which is sharp (Exercise 60); DeBiasio–
Lo [2017+] proved the conjecture asymptotically (that is, (G) ≥ ( ¾+1 3 + o(1))n is
sufficient). In fact, the conjecture asserts sufficiency of the Ore-type condition
that the sum of the degrees of any  + 3 independent vertices is at least n −  .
Recall from Remark 7.3.4 on toughness that #(G) denotes the number of com-
ponents of G. A -bounded spanning tree requires #(G − S) ≤ ( − 1) | S| + 1 for all
S ⊆ V(G) (Exercise 61). This necessary condition is almost sufficient: Win [1989]
proved that if #(G − S) ≤ ( − 2) |S| + 2 for all S ⊆ V(G), then G has a -bounded
spanning tree (Ellingham–Zha [2000] gave a short proof).
In terms of connectivity, (G) ≤ ( − 1)(G) + 1 guarantees a -bounded span-
ning tree (Neumann-Lara & Rivera-Campo [1991]). Win [1975] (proved below)
obtained an Ore-type condition: it suffices that any  independent vertices have
degree-sum at least n − 1 (implied by (G) ≥ (n − 1)/). A “long-cycle” analogue
for -bounded subtrees appears in Caro–Krasikov–Roditty [1991].
For -ended spanning trees, a necessary condition is #(G − S) ≤ | S| +  − 1
for all S ⊆ V(G) (Exercise 62). Also the Chvátal–Erdős and Ore conditions were
generalized: Win [1979] showed that (G) ≤ (G) +  − 1 suffices, and Broersma–
Tuinstra [1998] proved that 2(G) ≥ (n −  + 1)/2 suffices.
The notion of Hamiltonian-connected also generalizes to the setting of span-
ning trees. A graph G with more than  vertices is -leaf-connected if for every
set S of  vertices, there is a spanning tree whose set of leaves is S. A graph is
Hamiltonian-connected if and only if it is 2-leaf-connected. Gurgel–Wakabayashi
[1986] proved that G is -leaf-connected when 2(G) ≥ n +  − 1, and Egawa–
Matsuda–Yamashita–Yoshimoto [2008] improved this: 2(G) ≥ n+ 1 suffices when
G is ( + 1)-connected.
328 Chapter 7: Connectivity and Cycles

7.3.25. Theorem. (Win [1975]) For ¾ ≥ 2, if every independent set of ¾ vertices

in a connected n-vertex graph G has degree sum at least n − 1, then G has a
¾-bounded spanning tree (maximum degree at most ¾), and this is sharp.
Proof: Sharpness is shown by K ¾+1 ,¾ 2 +1 . In this graph, any ¾ independent ver-
tices have degree sum at least ¾ 2 + ¾ , which equals n − 2. However, a spanning
tree has ¾ 2 + ¾ + 1 edges, and in a ¾-bounded spanning tree only ¾ 2 + ¾ edges can
be incident to the small part.
Now assume the degree condition, and let T be a largest ¾-bounded tree con-
tained in G. If T is not spanning, then some vertex x ∈ V(T) has a neighbor
w outside V(T), since G is connected. Since T cannot be enlarged, d T (x) = ¾ .
Let T1 , . . . , T¾ be the components of T − x, with t i the neighbor of x in T i and
ni = | V(T i)|. Select a leaf pi in each subtree T i (with pi
= t i unless ni = 1).
The vertices p1 , . . . , p¾ form an independent set S in G, since an edge pi pj
would yield a larger ¾-bounded subtree by replacing xt i with pi pj and xw.
Let Yj be the set of vertices in T j having a neighbor in S − pj . If y ∈ Yj , then
d T (y) = ¾ , since otherwise replacing xt j with xw and an edge ypi enlarges T .
For y ∈ Yj , let y be the neighbor of y on the y , pj -path in T j . If there exists
∈ NT (y) − {y } such that pj ∈ E(G), then replacing y and xt j with all of xw,
pj , and some ypi enlarges T . (If y = t j and = x, then just replace xy with xw
and ypi .) Hence y has no such neighbor in T .
If is a common neighbor in T of two vertices y1 , y2 ∈ Yj , then ∈ {y1 , y2 };
otherwise, adding the edges y1 and y2 to the y1 , pj -path and y2 , pj -path com-
pletes a cycle in T j . Thus the sets NT (y) − {y } are pairwise disjoint for all y ∈ Yj .
For v ∈ V(G), let d j (v) = | NG(v) ∩ V(T j )|. From the definition of Yj ,
1
| Yj | ≥ max d j (pi) ≥
i∈[¾]−{ j} −1
∑ d j (pi).
i∈[¾]−{ j}

We further consider d j (pi). The sets NT (y) − {y } for y ∈ Yj are pairwise dis-
joint and have size  − 1. These sets, which have no neighbor of pi , cannot contain
pj , which is another nonneighbor of pi . However, x is in one of those sets if t j ∈ Yj .
Let  = 1 if t j ∈ Yj ; otherwise,  = 0. We obtain
d j (pi) ≤ nj − 1 − ( − 1) | Yj | +  ≤ nj − 1 +  − ∑ d j (pr).
r ∈[¾]−{ j}
¾
Thus ∑r=1 d j (pr) ≤ nj − 1 + . Since all neighbors of pr lie in T , summing
this inequality over j ∈ [] provides an upper bound on D, the sum of the degrees
of the  independent vertices in S. If  = 1 in the inequality for j , then all pi
for i
= j are not adjacent to x in G, and D ≤ ∑ nj = | V(T)| − 1. If  = 0 in each
inequality, then vertices of S may be adjacent to x in G; now D ≤ ∑(nj − 1) +  =
|V(T)| − 1. Since D ≥ n − 1 is given, T must be a spanning tree.

EXERCISES 7.3

7.3.1. (−) Prove that K n ,n has (n − 1)!n!/2 Hamiltonian cycles when n ≥ 2.

7.3.2. (−) Prove that every 5-vertex path in the dodecahedron graph extends to a Hamil-
tonian cycle.
Exercises for Section 7.3 329

7.3.3. (−) Prove that G has a Hamiltonian path only if for every S ⊆ V(G), the number of
components of G − S is at most | S| + 1.

7.3.4. (−) Prove that every ¾-regular ¾-edge-connected graph is 1-tough. (Comment:
Hence every such graph also has a 1-factor.)

7.3.5. (−) Prove or disprove: If an n-vertex graph G has an edge xy with the property that
d(x) + d(y) ≥ n + 2, then G is Hamiltonian if and only if G − xy is Hamiltonian.

7.3.6. (−) Let G be a 3-regular graph whose edge set has exactly one partition into three
perfect matchings. Prove that G is Hamiltonian. (Greenwell–Kronk [1973])

7.3.7. (−) Prove or disprove: If G is an n-vertex graph with n ≥ 3, and G has at least (G)
vertices of degree n − 1, then G is Hamiltonian.

7.3.8. A mouse eats its way through a 3 × 3 × 3 cube of cheese by eating all the 1 × 1 × 1
subcubes. If it starts at a corner subcube and always moves on to an adjacent subcube
(sharing a face of area 1), can it do this and eat the center subcube last? Give a method or
prove impossible. (Ignore gravity.)

7.3.9. (♦) Let x and y be vertices in a Hamiltonian bipartite graph G. Prove that G − x − y
has a perfect matching if and only if x and y lie in opposite sets of the bipartition. Use this
to conclude that an 8-by-8 chessboard with two missing unit squares can be partitioned
into 31 one-by-two rectangles if and only if the two missing squares have opposite colors.

7.3.10. Place n vertices around a circle. Let G n be the 4-regular graph obtained by making
each vertex adjacent to the two nearest vertices in each direction. For n ≥ 5, prove that
G n is the union of two Hamiltonian cycles.

7.3.11. For  ≥ 3, let G ¾ be the graph obtained from two disjoint copies of K ¾ ,¾−2 by adding
a matching joining the independent -sets in the two copies. Determine all values of  such
that G ¾ is Hamiltonian.
• •
• •
• • • •
• • • •
• • • •
• •
• •

7.3.12. (♦) For n ≥ 10 with n even, construct a 3-regular 3-connected graph that is not
Hamiltonian. (Hint: Start with the Petersen graph and expand vertices.)

7.3.13. (♦) Prove that the line graph of a graph G is Hamiltonian if and only if G has a
closed trail that contains at least one endpoint of each edge (this holds for the Petersen
graph). Use the Petersen graph to construct a 3-regular 3-connected graph whose line
graph is not Hamiltonian. (Hint: Make an appropriate substitution for each vertex of the
Petersen graph.) (Harary–Nash-Williams [1965])

7.3.14. Prove that if n is even, then K n ,n does not contain three perfect matchings such
that any two together form a spanning cycle. (Hint: Interpret each matching as a permu-
tation.) (Basavaraju–Chandran–Kummini [2010])

7.3.15. (♦) Prove that the cartesian product of two nontrivial graphs with spanning paths
fails to have a spanning cycle if and only if both graphs are bipartite and have odd order,
in which case the product has a spanning path. (Behzad–Mahmoodian [1969])
330 Chapter 7: Connectivity and Cycles

7.3.16. (♦) On a chessboard, a knight can move from one square to another that differs
by 1 in one coordinate and by 2 in the other coordinate (some such moves are shown be-
low). A knight ’s tour is a traversal of a rectangular board by knight ’s moves that visits
each square once and returns to the start. Prove that a 4-by-n chessboard cannot have a
knight ’s tour.

• • • • • •
• • • • • •
• • • • • •
• • • • • •

7.3.17. Let G be a claw-free graph (that is, G has no induced K1 ,3). Prove that the maxi-
mum t such that G is t-tough is (G)/2. (Matthews–Sumner [1984])

7.3.18. (♦) Prove that every 2-connected, claw-free, paw-free graph is Hamiltonian. (The
paw is obtained from the claw K1 ,3 by adding one edge.) (Goodman–Hedetniemi [1974])

7.3.19. (♦) Prove that a graph obtained by deleting one vertex from each part of the bi-
partite graph C2m K 2 is Hamiltonian (Stong [2003]). Conclude that deleting one vertex
of each parity from the hypercube Q¾ with  ≥ 3 leaves a Hamiltonian graph. (Comment:
Locke [2001] asked whether deleting up to  − 2 vertices of each parity from Q¾ always
leaves a Hamiltonian graph. Stong proved that asymptotically /2 of each can be deleted.)

7.3.20. (+) In the -dimensional hypercube Q¾ with  ≥ 2, prove that every perfect match-
ing extends to a spanning cycle. (Hint: Prove the stronger statement that every perfect
matching in the complete graph with vertex set V(Q¾) extends to a spanning cycle by adding
a perfect matching in Q¾ .) (Fink [2007, 2009])

7.3.21. Prove that the -dimensional hypercube Q¾ has a spanning path whose endpoints
are complementary if and only if  is odd. Prove that it has a spanning path with endpoints
differing in all but one position if and only if  is even.

7.3.22. Let G be an n-vertex Hamiltonian graph. Prove that G K1 ,m is Hamiltonian

if and only if m ≤ n. (Comment: Batagelj–Pisanski [1982] proved that the statement
remains true with any tree of maximum degree m in place of K1 ,m .)

7.3.23. (+) For each odd  , construct a ( − 1)-connected -regular bipartite graph that
is not Hamiltonian. (Comment: Nash-Williams conjectured that 4-connected 4-regular
graphs are Hamiltonian, but the 70-vertex Meredith graph disproves this (duplicate a
matching in the Petersen graph to obtain a 4-regular multigraph G, and then form the
blow-up graph H as in Example 6.2.15 that seeks a 1-factor in G). Tutte [1971 /2] con-
jectured that every 3-connected 3-regular bipartite graph is Hamiltonian; the first coun-
terexample was the Horton graph with 96 vertices (Bondy–Murty [1976, p. 240]). The
smallest known counterexample has 54 vertices (Ellingham–Horton [1983]).)

7.3.24. (♦) Determine which complete multipartite graphs are pancyclic. (Recall that an
n-vertex graph is pancyclic if it has cycles of all lengths from 3 to n.)

7.3.25. (+) Prove that every n-vertex Hamiltonian graph G with more than n2/4 edges is
pancyclic. (Bondy [1971a]) (Hint: Use induction on n (Thomassen). Consider separately
the cases when G does or does not contain a cycle of length n − 1. Comment: Bondy also
proved that n2/4 edges are sufficient unless G = K n/2 ,n/2 . From this it follows easily that
Dirac’s Condition and Ore’s Condition imply that G is pancyclic unless G = K n/2 ,n/2 .)
Exercises for Section 7.3 331

7.3.26. There is a subgraph of K n ,n with n(n − 1) + 1 edges that has no spanning cycle.
Prove that every subgraph of K n ,n having at least n(n − 1) + 2 edges has cycles of all even
lengths from 4 through 2n. (Entringer–Schmeichel [1988])
7.3.27. Determine whether the graph below has a spanning path.
• •
• • • •
• • • •
• • • • • • • •

• • • • • • • •
• • • •
• • • •
• •

7.3.28. (♦) The ¾ th power of a graph G is the graph G ¾ with vertex set V(G) and edge
set {uv: d G(u , v) ≤ ¾}.
(a) Prove that if a vertex x in G has exactly one neighbor in at least three nontrivial
components of G − x, then G 2 is not Hamiltonian.
(b) Prove that if G is connected and has at least three vertices, then G3 is Hamilto-
nian. (Sekanina [1960], Karaganis [1968]) (Comment: Fleischner [1974] proved that the
square of each 2-connected graph is Hamiltonian.)
7.3.29. Let G(¾ , t) be the class of connected ¾-partite graphs in which each part has size
t and each subgraph induced by two parts is a matching of size t.
(a) Show that every graph in G(3 , t) is a cycle and hence is Hamiltonian. (Comment:
Also every graph in G(¾ , 3) is Hamiltonian.)
(b) For ¾ ≥ 4 and t ≥ 4, construct a non-Hamiltonian graph in G(¾ , t). (Hint: Start
with a cycle spanning three parts, and make the other parts into cliques. Finish the con-
struction so that deleting some three vertices leaves four components.) (Ayel [1982])
7.3.30. (♦) The lollipop graph. (Thomason [1978])
(a) Let P be a u , -path in a multigraph G, and let S = V(P) − { }. Prove that G has
an even number of spanning paths that start along P and end at vertices whose degree
in G − S is even. (Hint: Define a graph H on the set of spanning paths beginning along
P , making a u , x-path Q and a u , y-path R in V(H) adjacent if and only if xv ∈ E(G) and
R = Q + xv − vy, where v is just before y on Q. Which vertices in H have odd degree?)
x•
u
• • v•
y•

(b) Let u and v be vertices in a multigraph G such that every vertex outside {u , v} has
odd degree. Prove that the number of spanning u , v-paths is even. Conclude that when
every vertex has odd degree, the number of spanning cycles through any edge is even.
(Comment: This was proved also in Corollary 5.4.8.)
7.3.31. Hamiltonian cycles through two incident edges. (Thomason [1978])
(a) Let uw and wv be edges in a multigraph G with | V(G)| ≥ 4 such that all vertices
outside {u , v , w} have odd degree. Use Exercise 7.3.30(a) to prove that if G − u is not Hamil-
tonian, then the number of Hamiltonian cycles in G using both uw and wv is even.
(b) For odd  with  > 1, prove that every -regular bipartite graph has an even num-
ber of Hamiltonian cycles. (For  = 3, this was proved by Kotzig, as noted in Bosák [1967].)
332 Chapter 7: Connectivity and Cycles

7.3.32. Let G be the union of two spanning cycles C and C . Say that G is rearrangeable if
G can be expressed as the union of two other spanning cycles with the same intersection.
(a) Use Theorem 7.3.5 to show that G is rearrangeable if | E(C) ∩ E(C )| = 1.
(b) For n ≥ 5, construct an n-vertex example to show that G may not be rearrangeable
if | E(C) ∩ E(C )| = 2. (West [1978])
7.3.33. Let G be the union of m edge-disjoint spanning cycles. Use Theorem 7.3.5 to prove
that every edge of G lies in at least 3m − 2 spanning cycles. Conclude that G has at least
m(3m − 2) spanning cycles. (Thomason [1978])
7.3.34. (♦) Let G be a graph that is not a forest and has girth at least 5. Use Ore’s Theo-
rem to prove that G is Hamiltonian. (N. Graham)
7.3.35. (♦) Let G be an X , Y -bigraph with | X | = | Y | = n/2 > 1. (Moon–Moser [1963])
(a) Given x ∈ X and y ∈ Y with xy ∈ / E(G) and d(x) + d(y) > n/2, prove that G is
Hamiltonian if and only if G + xy is Hamiltonian (see Lemma 7.3.6).
(b) Prove that (G) > n/4 guarantees a spanning cycle in G and that this is sharp.
7.3.36. (♦) For n ≥ 2, give two proofs that the maximum number of edges in a non-
Hamiltonian n-vertex graph is ( n−2 1 ) + 1.
(a) Use induction on n. (Ore [1959])
(b) Use Chvátal’s Theorem. (Bondy [1972b])
7.3.37. Edge threshold for spanning cycles in terms of minimum degree.
(a) Let  (i) = 2i2 − i + (n − i)(n − i − 1), and suppose n ≥ 6  . Prove that among i such
that  ≤ i ≤ n/2, the maximum value of  (i) is  ().
(b) Let G be an n-vertex graph with minimum degree  . Use part (a) and Chvátal’s
condition to prove that if G has at least 6  vertices and has more than ( n−2 ¾) +  2 edges,
then G is Hamiltonian. Is this always sharp? (Erdős [1962a])
7.3.38. Let G be a graph with vertex degrees d1 , . . . , d n such that d1 ≤ · · · ≤ d n .
(a) Prove that if d i ≥ i or d n+1 −i ≥ n − i whenever i < (n + 1)/2, then G has a spanning
path. (Hint: Use Theorem 7.3.11.) Construct an example based on Example 7.3.12 to show
that this result is best possible. (Chvátal [1972])
(b) Let d 1 , . . . , d n be the vertex degrees in G, indexed in nondecreasing order. Use
part (a) to prove that if d i ≥ d i for i ≤ n/2, then G has a spanning path. (Hence self-
complementary graphs have spanning paths.) (Clapham [1974])
7.3.39. (+) Let G be an X , Y -bigraph having | X | = | Y | = n/2 > 1 and vertex degrees
d1 , . . . , d n in nondecreasing order. Form G by adding edges to turn Y into a clique.
(a) Prove that G is Hamiltonian if and only if G is Hamiltonian, and describe the re-
lationship between d and the degree list of G .
(b) Prove that if d ¾ >  or d n/2 > n/2 −  whenever  ≤ n/4, then G is Hamiltonian.
(Hint: Prove that Chvátal’s Condition holds for G .) (Chvátal [1972])
7.3.40. Prove that Ore’s Condition implies Chvátal’s Condition. Conclude that Theorem
7.3.11 (Chvátal’s Condition is sufficient for a spanning cycle) implies Corollary 7.3.7 (Ore’s
Condition is sufficient for a spanning cycle).
7.3.41. (♦) Use Ore’s Lemma to prove that a graph G and its closure C(G) have the same
circumference.
7.3.42. For n ≥ 3, prove that Chvátal’s Condition (Theorem 7.3.11) cannot hold for both
an n-vertex graph and its complement. (Kostochka–West [2006])
7.3.43. (+) By showing that their hypotheses imply the Chvátal–Erdős Condition, prove
that the Chvátal–Erdős Theorem implies each of the following. (Bondy [1978b])
(a) Ore’s Theorem (Corollary 7.3.7).

(b) For  ∈ , every -regular graph with 2  + 1 vertices is Hamiltonian.
Exercises for Section 7.3 333

7.3.44. Let G be an n-vertex graph such that G has m edges. Prove that G is Hamiltonian
if m ≤ n − 3 and G is Hamiltonian-connected if m ≤ n − 4. (Ore [1963])
7.3.45. Let G be an n-vertex graph with m edges, where n ≥ 4.
(a) Prove that if G is Hamiltonian-connected, then m ≥ ⌈ 3n/2⌉ . Use C¾ K 2 to show
that the result is sharp when ¾ is odd. (Moon [1965b])
(b) Prove that G is Hamiltonian-connected if 2(G) ≥ n + 1. For  > 1, construct a
2 -vertex graph with minimum degree  that is not Hamiltonian-connected. (Ore [1963])
7.3.46. For a graph G with degrees d1 , . . . , d n in nondecreasing order, prove that G is
Hamiltonian-connected if d i > i + 1 or d n−i−1 ≥ n − i whenever i < (n − 1)/2. (Berge [1973])
7.3.47. For even n, prove that n − 1 is the least t such that 2 (G) ≥ t forces an n-vertex
graph G to have a perfect matching.
7.3.48. For even n, prove that if G has vertex degrees d1 , . . . , d n in nondecreasing or-
der, and d i ≥ i + t − 1 for 1 ≤ i ≤ n/2, then G has t edge-disjoint perfect matchings.
(Jahanbekam–West [2016])
7.3.49. (♦) Let G be an n-vertex graph with 2 (G) ≥ n +  . Prove that any set of  edges
forming a disjoint union of paths in G is contained in a Hamiltonian cycle. Construct an
example to show that 2(G) ≥ n +  − 1 is not sufficient. (P ósa [1963/4])
7.3.50. (+) Las Vergnas’ Condition. Let G be a graph having a vertex ordering v1 , . . . , vn
for which there is no nonadjacent pair vi , vj such that i < j , d(vi) ≤ i, d(vj ) < j ,
d(vi) + d(vj ) < n, and i + j ≥ n. Las Vergnas [1970] proved that the closure of G is complete
(and hence G is Hamiltonian).
(a) Prove that Chvátal’s Condition for Hamiltonian cycles implies Las Vergnas’ Con-
dition. (Thus Las Vergnas’ Theorem strengthens Chvátal’s Theorem.)
(b) Prove that each of the graphs below fails Chvátal’s Condition but has a complete
graph as its Hamiltonian closure. Prove that the smaller graph satisfies Las Vergnas’
Condition but the larger one does not.
•
• • • • • •

• • • • • •
•

7.3.51. (♦) Let G be a connected n-vertex graph with 2 ≤  = (G) < n/2.
(a) Let P be a maximal path in G (not a subgraph of any longer path). Prove that if
P has at most 2  vertices, then the induced subgraph G[V(P)] has a spanning cycle (this
cycle need not have its vertices in the same order as P).
(b) Use part (a) to prove that G has a path with at least 2  + 1 vertices. Give an
example for each  to show that G may have no cycle with more than  + 1 vertices.
7.3.52. Use Theorem 7.3.17 that c(G) ≥ min{n , 2 (G)} when G is 2-connected to prove
that every -regular graph with 2  + 1 vertices is Hamiltonian. (Nash-Williams)
7.3.53. (♦) For r,  ≥ 2, let G be a -connected graph with girth at least 2r − 1. Prove that
if G has at least r vertices, then G has a cycle of length at least r . (Kostochka)
7.3.54. (♦) Let G be an n-vertex
√ graph not having K1 ,t+1 as an induced subgraph, where

t ∈ . Prove that if (G) ≥ tn, then G is Hamiltonian. (Faudree)
7.3.55. Let G be a 4-regular multigraph that is a union of two spanning cycles. Form G
by subdividing one edge from each cycle and adding a double-edge joining the new vertices.
Show that G is also a union of two spanning cycles if | V(G)| ≤ 3. Conclude for 2 ≤  ≤ 4
that two longest cycles in a -connected graph have at least  common vertices. (S. Smith)
334 Chapter 7: Connectivity and Cycles

7.3.56. (♦) Prove that if (G) ≥ 3 − 1, or if G is triangle-free and (G) ≥ 2  , then G con-
tains  disjoint cycles. Show also that lower degree does not suffice. (Thomassen [1988])
7.3.57. (♦) Use Theorem 7.3.16 to prove that in an n-vertex graph with many edges,
one can cover at least half the edges using at most n/2 cycles. Conclude that every n-
vertex graph decomposes into at most O(n log n) cycles and edges. (Erdős–Gallai [1959])
(Comment: Erdős and Gallai conjectured that O(n) cycles and edges suffice. Conlon–Fox–
Sudakov [2014] reduced the bound to O(n log log n).)
7.3.58. Fouquet–Jolivet [1978] conjectured a long-cycle version of the Chvátal–Erdős
Theroem: for  < a, every n-vertex -connected graph with independence number a has a
cycle of length at least ¾a (n + a − ). For a ≥  ≥ 2 and r ≥ 1, construct a -connected graph
with  + ar vertices that has no longer cycle, thereby proving sharpness. (Comment: The
conjecture was proved in O–West–Wu [2011].)
7.3.59. The prism over a graph G is the cartesian product G K 2 . Prove that the prism
over K ¾ ,2¾+1 has no spanning cycle. (Comment: Ellingham–Salehi Nowbandegani [2018]
proved that if (G) ≤ 2 (G), then the prism over G is Hamiltonian.)
7.3.60. Ozeki–Yamashita [2011] conjectured that every n-vertex connected graph G with
(G) ≥ (n − )/( + 3) has a spanning tree with at most  branch vertices. Prove sharpness
using a connected graph G formed from ( − 1)K t+1 + 2 K t ,t+1 by adding one path of length
 . (Gargano–Hammar–Hell–Stacho–Vaccaro [2004], DeBiasio–Lo [2017+])
7.3.61. Prove that if G has a -bounded spanning tree (maximum degree at most ), then
#(G − S) ≤ ( − 1) | S| + 1 for all S ⊆ V(G). (Ozeki–Yamashita [2011])
7.3.62. (♦) Prove that if G has a spanning tree with no more than  leaves, then
#(G − S) ≤ | S| +  − 1 for all S ⊆ V(G). (Salamon–Wiener [2007])
7.3.63. (♦) Prove that every tournament has a Hamiltonian path and that every strong
tournament has a Hamiltonian cycle. (Rédei [1934], Camion [1959])
7.3.64. (♦) Let T be a strong n-vertex tournament. Suppose 3 ≤  ≤ n.
(a) For each u ∈ V(T), prove that u belongs to a cycle of length  in T . (Hint: Use
induction on  .)
(b) Prove that T has at least n −  + 1 cycles of length  . Construct an example for
each n to show that the bound is sharp for all  , simultaneously. (Moon [1966])
7.3.65. Let T be a 7-vertex tournament in which every vertex has outdegree 3. Prove that
T has two disjoint cycles. (Hint: Use Exercise 7.3.64. Comment: Thomassen [1983] proved
that every digraph with minimum outdegree at least 3 has two disjoint cycles.)
7.3.66. (♦) A bipartite tournament is an orientation of a complete bipartite graph.
Prove that a bipartite tournament has a spanning path if and only if it has a spanning
subgraph whose components are all cycles except for possibly one path. (Comment: Gutin
[1993] proved this for all complete multipartite graphs.)
7.3.67. Let D be an n-vertex digraph having no loops and having at most one copy of each
ordered pair as an edge; such a digraph is strict.
(a) Prove that if min{ −(D) , +(D)} ≥ n/2, then D has a spanning cycle. (Hint: Con-
sider a longest cycle in D.) (Ghouilà-Houri [1960])
(b) Construct for each even n an n-vertex digraph D with loops that is not Hamiltonian
even though it satisfies all other conditions above.
Chapter 8

Coloring
When the edges of a graph represent vertex conflicts, we often seek the min-
imum number of conflict-free classes needed to partition the vertices. For exam-
ple, when scheduling examinations or committee meetings into the fewest time
slots, each course or committee is a set of people, and intersecting sets must have
different times. The time slots are “colors”, and the problem is to use the fewest
colors on the vertices of a graph so that adjacent vertices have different colors.
Because labeling to avoid conflicts is such a fundamental concept, coloring
theory is enormously broad. Texts and surveys with extensive material on color-
ing include Berge [1973], Bollobás [1978], Chartrand–Lesniak [1986, etc.], Toft
[1995], Jensen–Toft [1995], Molloy–Reed [2002], and Kubale [2004].

8.1. Vertex Coloring

We call vertex labels “colors” because (1) the problem originated with the
coloring of regions on maps, and (2) the labels need not be numbers.

8.1.1. Definition. A ¾-coloring of G is a vertex labeling º : V(G) → S, where

|S| = ¾ . A ¾-coloring º is proper if º (x)
= º (y) whenever xy ∈ E(G), and G
is ¾-colorable if it has a proper ¾-coloring. The chromatic number Ò(G) is
the least ¾ such that G is ¾-colorable. If Ò(G) = ¾ , then G is ¾-chromatic.

8.1.2. Example. In a proper coloring, each color class is an independent set. Thus
G is ¾-colorable if and only if G is ¾-partite. Thus G is 2-colorable if and only if
G has no odd cycle. The 5-cycle and the Petersen graph (below) are 3-chromatic.

a b
• a• •
•c
c• •b c• •a
•a
b• •b
b• •a •
a
•
c

335
336 Chapter 8: Coloring

Breadth-first search (Exercise 5.4.70) makes 2-colorability easy to test; a con-

nected graph is bipartite if and only if the sets of vertices at odd and even distance
from some vertex x are independent sets. No easily testable characterization of
3-colorable graphs is known; indeed, testing 3-colorability is NP-complete.

We use (G) for the independence number and (G) for the maximum size
of a clique in G, called the clique number. As the first and last letters of the
Greek alphabet, they suggest the beginning and end of growing a graph.

8.1.3. Remark. Easy bounds. When G has n vertices, using distinct colors yields
(G) ≤ n. Since each color class is an independent set, (G) ≥ n/(G). Since
adjacent vertices require distinct colors, (G) ≥ (G). More generally, a proper
coloring of G must properly color its subgraphs, so (G) ≥ (H) when H ⊆ G.
Equality holds for K n in these bounds. Each can be very bad. Despite having
n vertices, K r,n−r is 2-colorable. Exercise 40(b) shows that (G) − n/(G) can be
large. Theorem 8.1.17 shows that (G) can be large when (G) = 2.

8.1.4. Example. Additivity under joins (Dirac). For odd cycles, (C2r+1) = 3 and
(C2r+1) = 2. This yields graphs with larger difference via the join operation (Def-
inition 5.1.22). Recall that the join G H is the graph obtained from the disjoint
union G + H by making every vertex of G adjacent to every vertex of H.
The edges that join V(G) and V(H) in G H force disjoint sets of colors; hence
(G H) = (G) + (H). Also, combining cliques from G and H yields (G H) =
(G) + (H). Thus, (G) = 3 and (G) = 2  when G is the join of  odd cycles.

• •
• • • •
C5 C5
• • • •

A bound that is rarely sharp begs for improvement. Only complete graphs
achieve equality in (G) ≤ | V(G)| , so we focus first on improving this bound.

UPPER BOUNDS

Most upper bounds on chromatic number arise from coloring algorithms. An

n-vertex graph can be properly n-colored without even looking. A natural im-
provement is to color the vertices in some order and use the first available color.

8.1.5. Algorithm. For a graph G, the greedy coloring (or first-fit coloring)
with respect to a vertex ordering v1 , . . . , vn colors vertices in the order v1 , . . . , vn ,
giving vi the least-indexed color not used on its neighbors colored earlier.

8.1.6. Proposition. (G) ≤ (G) + 1.

Proof: In a vertex ordering, each vertex has at most (G) earlier neighbors, so
the greedy coloring will not use more than (G) + 1 colors.

Although (G) + 1 ≤ | V(G)| , still Proposition 8.1.6 is rarely sharp.

Section 8.1: Vertex Coloring 337

8.1.7. Theorem. (Brooks’ Theorem; Brooks [1941]) (G) = (G) + 1 only if G

has K (G)+1 (or an odd cycle when (G) = 2) as a component.

Cranston–Rabern [2015] surveyed proofs of Brooks’ Theorem and its exten-

sions. We will prove it as Corollary 8.2.14 from a stronger result. Meanwhile,
since (G) ≤ (G) + 1 using any ordering, maybe clever orderings give better
greedy colorings. We give an example that yields optimal colorings.

8.1.8. Definition. The intersection graph of sets A1 , . . . , Am is the graph with

vertices {v1 , . . . , vm} where vi vj is an edge if and only if Ai ∩ A j
= ∅.

Proper colorings of intersection graphs can model allocation of scarce re-

sources where intersecting sets represent conflicting demands.

8.1.9. Example. An interval graph is the intersection graph of a family of

intervals on the real line. Computations in a computer program put values of
variables into expensive “registers”. If two variables are never simultaneously in
use, then we can assign them to the same register. Each variable occupies a time
interval from its first to its last use. The number of registers needed is the chro-
matic number of the resulting interval graph. (Other applications include DNA
analysis (Benzer [1959]), timing of traffic lights (Roberts [1978]), etc.)
Given intervals representing an interval graph, order the vertices by the left
endpoints of the intervals. Under greedy coloring, when x is to be colored, it re-
ceives color  only if the left endpoint of the interval for x belongs to intervals
already having colors 1 through  − 1. If x receives the color  of maximum in-
dex, then the neighbors of x already colored at that time induce a -clique with
x. Since (G) ≥ (G), the coloring is therefore optimal. For the vertex ordering
a , b , c , d , e ,  ,  of the interval graph below, greedy coloring assigns the colors
1 , 2 , 3 , 1 , 1 , 2 , 3, respectively, which is optimal.

•d b 
b• •c •e • a d e
•a 
•
c

Replacing (G) with a quantity that is never larger strengthens Proposition

8.1.6. For any vertex ordering, a closer look at the vertex degrees produces the
improvement below. This bound is best when we dispose of high-degree vertices
first, coloring the vertices in decreasing order of degree.

8.1.10. Proposition. (Welsh–Powell [1967]) If a graph G has vertex degrees

d1 , . . . , d n , then (G) ≤ 1 + max i min{d i , i − 1}.
Proof: When we color the ith vertex by the greedy algorithm, the number of its
neighbors that have already been colored is bounded by d i and by i − 1, so the
label used is at most one more than min{d i , i − 1}.

Instead of coloring high-degree vertices first, we can color low-degree vertices

last. Applied iteratively to produce an ordering, this generally works better.
338 Chapter 8: Coloring

8.1.11. Definition. A graph G is ¾-degenerate if every subgraph has a vertex

of degree at most ¾ . The degeneracy of G is max H ⊆ G (H); that is, the mini-
mum  such that G is -degenerate. A smallest-last ordering of an n-vertex
graph G is constructed iteratively from index n to 1 by letting vi be a vertex
of minimum degree in G − {vi+1 , . . . , vn}.

The 1-degenerate graphs are precisely the forests. Some authors call the de-
generacy plus 1 the “coloring number ” of G; this causes some confusion since it
generally differs from the chromatic number but provides an upper bound. “Col-
oring bound” would be a less-confusing term.

8.1.12. Proposition. (Szekeres–Wilf [1968]) If G is a -degenerate graph, then

G is ( + 1)-colorable. In particular, (G) ≤ 1 + max H ⊆ G (H).
Proof: If G is -degenerate, then a smallest-last ordering gives each vertex at
most  neighbors among the earlier vertices. Hence the greedy coloring for such
a vertex ordering is a proper ( + 1)-coloring. For the second statement, the max-
imum of (H) is the least  such that G is -degenerate.

A smallest-last ordering gives the smallest possible greedy coloring bound on

(G) in terms of degrees alone, because in every ordering u1 , . . . , un some vertex
ui has at least max H ⊆ B (H) earlier neighbors (Finck–Sachs [1969]).
Greedy coloring runs fast and constructs a proper coloring even when the
graph is shown only one vertex at a time, requiring an immediate and permanent
choice of color. With a random vertex ordering on a random graph, the greedy al-
gorithm almost always uses only about twice as many colors as the minimum,
although with a bad ordering it may use many colors on a tree (Exercise 23).
Other results bound (G) using orientations of G.

8.1.13. Theorem. (Minty’s Theorem; Minty [1962]) In a connected graph G

that is not a forest, let C be the set of all cycles, listed in both directions. For
an orientation D of G, let r(D) = max C∈C ⌈ a/b⌉ , where a and b are the num-
bers of edges of D followed forward and backward by C, respectively. Always
(G) ≤ 1 + min D r(D), with equality for some orientation.
Proof: When D has a cycle, r(D) is infinite, and the bound holds. Hence we may
assume D is acyclic. We produce a proper coloring of G with 1 + r(D) colors. We
may assume that G is connected.
Fix a vertex x ∈ V(G). For a walk W starting at x, let (W) = a − b · r(D),
where a and b are the numbers of steps along W that are forward or backward in
D, respectively. For y ∈ V(G), let (y) = max{(W): W is an x , y-walk}. By the
definition of r(D), traversing a cycle of G (or traversing an edge in both directions)
makes no positive contribution to (W), so (y) is the maximum over x , y-paths.
If uv ∈ E(D), then an x , u-walk attaining (u) extends along uv to yield (v) ≥
(u) + 1. An x , v-walk attaining (v) extends along uv to yield (u) ≥ (v) − r(D).
Hence (u) + 1 ≤ (v) ≤ (u) + r(D). We conclude that coloring each vertex y by
the congruence class of (y) modulo 1 + r(D) produces a proper r(D)-coloring.
•u
x•
•v
Section 8.1: Vertex Coloring 339

For sharpness, consider a proper coloring using colors 1 , . . . , (G). Form D∗

by orienting each edge from lower color to higher color. Colors increase along ev-
ery path, so every path has length at most (G) − 1. A cycle thus follows at most
(G) − 1 forward edges before a backward edge. Hence a/b ≤ (G) − 1 for each
cycle. Since (G) is an integer, we obtain (G) ≥ 1 + r(D∗) ≥ 1 + min D r(D).

Minty ’s Theorem implies the subsequent better-known Gallai–Roy Theorem,

proved in Gallai [1968], Roy [1967], Vitaver [1962], and Hasse [1964/5] (in En-
glish, French, German, and Russian, respectively).

8.1.14. Corollary. (Gallai–Roy Theorem) (G) ≤ 1 + l(D) for every orientation

D of G, where l(D) is the length of a longest path in D. Equality holds for
some orientation.
Proof: Suppose first that D is acyclic. Since (G) ≤ 1 + r(D) by Minty ’s Theorem,
it suffices to prove r(D) ≤ l(D). Since r(D) = max C ⌈ a/b⌉ , along a cycle C achieving
the maximum we have a/b > r(D) − 1. By the Pigeonhole Principle, C has at least
r(D) forward edges between some two backward edges. Hence l(D) ≥ r(D).
When D is not acyclic, r(D) is infinite. We instead find an acyclic orientation
D̂ with l(D)
̂ ≤ l(D), so (G) ≤ 1 + r(D) ̂ ≤ 1 + l(D)
̂ ≤ 1 + l(D).

Let D be a maximal acyclic subdigraph of D. Obtain D ̂ from D by reversing
̂
all edges of D not in D . If D contains a cycle C , using reversed edges e1 , . . . , e¾
that create cycles C1 , . . . , C¾ in D when added to D without reversal, then re-
placing each e i in C with Ci − e i yields a closed walk in D . A closed walk in a
digraph contains a cycle, but D is acyclic. Hence D ̂ is acyclic.
̂ ̂
To prove l(D) ≤ l(D), let P be a u , v-path in D. If xy ∈ E(P) was reversed in
forming D,̂ then D has an x , y-path. Replacing all such edges of P with paths in
D yields a u , v-walk in D at least as long as P ; it is a path since D is acyclic.

Equality again is achieved by directing edges from low label to high label
with respect to an optimal coloring of G.

The Gallai–Roy and Minty bounds give the chromatic number exactly on any
graph IF we guess the right orientation, as does greedy coloring if we guess the
right vertex ordering (Exercise 23).
Tuza [1992] strengthened Theorem 8.1.13 by showing that only some cycles
need be considered: if the ratio a/b in the orientation D is at most  − 1 for every
cycle in G whose length is 1 modulo  , then (G) ≤  . Thus if G is not -colorable,
then G has a cycle whose length is congruent to 1 modulo  (Exercise 31).
Many open problems remain for bounds on (G). One of the most famous is
Reed ’s Conjecture (Reed [1998]) that (G) ≤⌈ (G)+2 (G)+1 ⌉ for every graph G, im-
proving the trivial bound (G) ≤ (G) + 1.

TRIANGLE-FREE GRAPHS

What forces chromatic number to be large? We know (G) ≥ (G), but almost
all graphs have large chromatic number but no large clique: when n is large, the
values of (G), (G), and (G) for almost all n-vertex graphs are approximately
340 Chapter 8: Coloring

2 log2 n, 2 log2 n, and n/(2 log2 n), respectively (see Chapter 14). Hence (G) is
usually a weak lower bound on (G), while n/(G) is usually a good lower bound.
There are many constructions of triangle-free graphs with large chromatic
number (see Sachs [1969] for a survey). Our first is perhaps the easiest to describe
and best known (see also Exercise 21).

8.1.15. Definition. Given a graph G, Mycielski’s Construction produces a

new graph G as follows. Letting V(G) = {v1 , . . . , vn}, add a vertex w and
an independent set U consisting of vertices u1 , . . . , un . For each i, make ui
adjacent to all of NG(vi). Finally, let NG (w) = U.

8.1.16. Example. Mycielski [1955] actually considered only the sequence gener-
ated from K 2 using Definition 8.1.15, but the construction can be applied to any
graph. From K 2 , one application yields the 3-chromatic C5 . A second yields the
4-chromatic Gr ötzsch graph, drawn in two ways below.

• • •
•
• •
• • • • • • •w
•
w • •
• • • •
• • G U

8.1.17. Theorem. Given a graph G, let G be the graph generated from G by

Mycielski’s construction. If G is triangle-free, then G is triangle-free. If G
is -chromatic, then G is ( + 1)-chromatic.
Proof: Name the vertices as in Definition 8.1.15, with V(G) = {v1 , . . . , vn} and
V(G ) = V(G) ∪ U ∪ {w}. Since U is independent in G , the other vertices of any
triangle containing a vertex ui ∈ U belong to V(G) and are neighbors of vi , which
would yield a triangle in G. Therefore, Mycielski’s Construction preserves the
absence of triangles.
A proper -coloring  of G extends to a proper ( + 1)-coloring of G by defining
 (ui) =  (vi) for each i and  (w) =  + 1; hence (G ) ≤ (G) + 1.
To prove that the chromatic number increases, we prove (G) < (G ). If G
has a proper -coloring  , then we may assume (w) =  , which restricts  to
{1 , . . . ,  − 1} on U. Let A = {vi : (vi) = }; we change colors on A to obtain a
proper ( − 1)-coloring of G. For each vi ∈ A, change the color of vi to (ui).
Since A lies in one color class of  , it is independent. Thus we need only
check colors on {vi , v } with vi ∈ A and v ∈ V(G) − A. If v vi ∈ E(G), then also
v ui ∈ E(G ), which yields (v )
= (ui). Hence our alteration does not violate
edges within G. We now delete U ∪ {w} and have a proper ( − 1)-coloring of G.

Iterating Mycielski’s Construction from G 2 = K 2 yields the smallest -

chromatic graphs for  ≤ 4, but then the graphs grow rapidly: | V(G ¾)| =
2 | V(G ¾−1)| + 1 yields the exponential growth | V(G ¾)| = 3 · 2¾−2 − 1. How many
vertices are needed?
Exercises for Section 8.1 341
√
8.1.18. Proposition. If G is a triangle-free n-vertex graph, then Ò(G) ≤ 2 n.
Equivalently, n ≥ ¾ 2/4 when G is triangle-free and ¾-chromatic.
Proof: The neighborhood of any vertex in G is an independent set; we can use one
color
√ on it. This suggests an algorithm. As long as G has a vertex with at least
⌊ n⌋ neighbors not yet colored, use one new color on those vertices.
√
Because G has only n vertices, this phase uses at most n colors. Afterwards,
the
√ subgraph induced by the remaining vertices has maximum degree less than
⌊ n⌋ . Thus greedy coloring with any ordering properly colors the remaining sub-
√
graph using at most n additional colors.
√
Finally, Ò(G) ≤ 2 n if and only if n ≥ Ò(G)2/4.

The minimum number of vertices needed is surprisingly close to ¾ 2/4, but

building such small ¾-chromatic triangle-free graphs is not easy.

8.1.19. Remark. Let º (¾) be the smallest order of a triangle-free ¾-chromatic

graph. Using probabilistic (non-constructive) methods (see Theorem 14.2.22),
Erdős [1961] proved º (¾) ≤ c(¾ log ¾)2 . We now know c1 ¾ 2 log ¾ ≤ º (¾) ≤ c2 ¾ 2 log ¾
for some constants c1 and c2 (see Chung–Graham [1998, p. 61]). Explicit construc-
tions of relatively small triangle-free ¾-chromatic graphs appear in Lubotzky–
Phillips–Sarnak [1988] and Kriz [1989].
In fact, we can forbid all cycles up to any fixed length and still have large
chromatic number. Methods to prove this involve hypergraphs or probabilistic
constructions, so we postpone proofs to Chapters 10 and 14.

EXERCISES 8.1

8.1.1. (−) Compute (G) and (G) for the graph G below.

• •

• • • •

• • • •

• •

8.1.2. (−) Let the blocks of G be G1 , . . . , G ¾ . Prove that (G) = maxi (G i).
8.1.3. (−) Find an optimal coloring of C5 C5 where sizes of color classes differ by at most 1.
8.1.4. (−) For each  ∈ 
with  ≥ 2, construct a -chromatic graph having no optimal
coloring using a color class of size (G).
8.1.5. (−) Prove or disprove: (G) ≤ 1 + a(G) for every connected graph G, where a(G) is
the average vertex degree in G.
8.1.6. (−) Prove or disprove: If G is an n-vertex graph, and n is sufficiently large, then
(G) ≤ (G) + n/ (G).
342 Chapter 8: Coloring

8.1.7. (−) Prove that (G) ≥ n− n(G) for every n-vertex graph G. Prove that there are in-
finitely many graphs (other than complete graphs) where equality holds.
8.1.8. (−) Use the Gallai–Roy Theorem (Corollary 8.1.14) to prove that every tournament
has a spanning path.
8.1.9. Let G be a graph without K1 ,t+1 as an induced subgraph. Prove (G) ≥ 1 + (G)/t.

For  , t ∈ , construct a -chromatic example where equality holds. (Zaker [2011a])
8.1.10. Determine the largest possible minimum degree of an n-vertex -colorable graph.
8.1.11. (♦) Let G be a graph in which any two odd cycles have a common vertex. Prove
(G) ≤ 5. Construct a graph to show that the bound cannot be improved.
8.1.12. Suppose that every edge of a graph G appears in at most one cycle. Prove that
every block of G is an edge or a cycle. Use this to prove (G) ≤ 3.
8.1.13. (♦) Let G be the unit-distance graph in the plane; the vertices are all the points
in 2 , adjacent when their Euclidean distance is exactly 1. Prove 4 ≤ (G) ≤ 7. (Comment:
de Grey [2018] and Heule [2018] proved (G) ≥ 5; see Soifer [2008, 2019].)
8.1.14. (♦) Given a set of lines in the plane with no three meeting at a point , form a graph
G whose vertices are the intersections of the lines, with two vertices adjacent if they ap-
pear consecutively on one of the lines. Prove (G) ≤ 3. (Comment: This can fail when
three lines may meet at a point.) (H. Sachs)

• •
→ • •
• •

8.1.15. Given a spanning tree T in a graph G, let (G; T) denote the least  such that G
has a proper coloring using colors in [] such that vertices adjacent in T have colors differ-
ing by at least 2. The backbone chromatic number of G is the minimum of (G; T) over
all spanning trees T . Prove that every nonbipartite graph has backbone chromatic num-
ber at least 4. (The concept originated in Broersma–Fomin–Golovach–Woeginger [2007].)
8.1.16. The integer simplex with dimension d and side-length m is the graph T md whose
vertices are the nonnegative integer (d + 1)-tuples summing to m, with two vertices adja-
cent when they differ by 1 in two positions and agree elsewhere. Determine (T md ).

8.1.17. (♦) The Kneser graph K(n , ) has vertex set ([n]
 ) , with two vertices are adjacent
when they are disjoint -sets (the Petersen graph is K(5 , 2)). Prove (K(n , )) ≤ n − 2  + 2.
Prove that this is optimal when n = 2  + 1. (Comment: Lovász [1978] proved Kneser ’s
conjecture that always (K(n , )) = n − 2  + 2; see Theorem 16.2.29.)
8.1.18. Let G be the intersection graph of a family of equal-sized squares in the plane,
all having horizontal and vertical sides. Prove that G has a vertex whose neighborhood is
covered by two cliques. Conclude that (G) ≤ 2(G) − 1.
8.1.19. For a graph G containing no 4-cycle, prove (G) ≤  (G) + 2. (Zaker [2011b])
8.1.20. Let G be a 3-regular K4 -free graph with m edges. Prove that G has a bipartite
subgraph with at least 97 m edges. (Hint: Use Brooks’ Theorem.)
8.1.21. The achromatic number of a graph G is the maximum number of colors in a
proper coloring of G such that all pairs of colors appear on adjacent vertices.
(a) Show that the achromatic number is well defined for every graph G.
(b) For n ≥ 3, determine the achromatic number of the cycle Cn .
Exercises for Section 8.1 343

8.1.22. Consider the coloring algorithm that iteratively chooses a maximal independent

set and gives it the next color. For ¾ ∈ , construct a 2-colorable graph on which this
algorithm may use ¾ colors.
8.1.23. (♦) Greedy coloring. Let G be a graph.
(a) Prove that G has a vertex ordering where greedy coloring uses (G) colors.

(b) For  ∈ , construct a tree T¾ with maximum degree  and an ordering  of V(T¾)
such that greedy coloring with respect to  uses  + 1 colors. (Comment: Thus the perfor-
mance ratio of greedy coloring to optimal coloring can be ( (G) + 1)/2.) (Bean [1976])
8.1.24. Prove that 11 is the least number of vertices in a triangle-free 4-chromatic graph.
8.1.25. Prove that an n-vertex graph with chromatic number  has at most  n−¾ vertex
partitions into  independent sets, with equality only for K ¾ + (n − )K1 .
8.1.26. The graph below has a proper 3-coloring with each color used twice. Prove that it
cannot arise from greedy coloring. (Fon-Der-Flaass)
•
•
• •
• •
r −1
8.1.27. Let G be an n-vertex graph that does not contain K r+1 . Prove (G) ≤ 2 r−1 n1 −1/2 .
(Comment: This generalizes Proposition 8.1.18.)
(G)+1
8.1.28. (♦) Prove that if G has no induced 2K 2 , then (G) ≤ ( 2
). (Hint: Cover the
vertices using ( (G)
2 )
+ (G) independent sets.) (Wagon [1980])
8.1.29. (♦) Let G be a -degenerate graph, and let a and b be positive integers such that
a + b =  − 1. Prove that V(G) can be partitioned into sets A and B such that G[ A]
is a-degenerate and G[B] is b-degenerate. Prove that the vertex set of any graph G can
be partitioned into max H ⊆ G ⌈ (H)+1
d+1
⌉ sets inducing d-degenerate subgraphs. (Comment:
Proposition 8.1.12 is the case d = 0. The case d = 1 is in Chartrand–Kronk [1969].)
8.1.30. Let G be an n-vertex graph. Prove that if (G) = n −  , then (G) ≥ n − 2  . For
even  and n ≥ 5 /2, construct an n-vertex graph with chromatic number n −  and clique
number at most n − 3/2.
8.1.31. (♦) Chromatic number and cycle lengths. Let G be a connected graph.
(a) Fix v ∈ V(G). Choose a spanning tree T to maximize ∑u∈V(G) d T (u , v). Prove that
the endpoints of any edge in G lie along a path in T that starts at v.
(b) Prove that if (G) >  ≥ 2, then G has a cycle with length congruent to 1 modulo
 . (Hint: Define a -coloring using the tree T of part (a). Comment: This generalizes the
characterization of bipartite graphs.) (Tuza [1992])
(c) Prove that a graph having no odd cycle of length more than 2 j − 1 is 2 j-colorable.
(Erdős–Hajnal [1966])
8.1.32. (♦) Prove that a graph is -colorable if every vertex lies in fewer than () odd cy-
2
cles. (Hint: Use induction on | V(G)|.) (Stong [2006])
8.1.33. Prove that (H) = (H) when H is bipartite.
8.1.34. Permutation graphs. Let G be the intersection graph of segments having end-
points on two parallel lines. Prove (G) = (G). (Pnueli–Lempel–Even [1971])
8.1.35. (♦) Let G and H be graphs. Prove (G H) = max{ (G) , (H)}. (Sabidussi [1957],
Vizing [1963], Aberth [1964], Behzad–Mahmoodian [1969])
344 Chapter 8: Coloring

8.1.36. (♦) Prove that an n-vertex graph G is r-colorable if and only if the cartesian prod-
uct G K r has an independent set of size n. (Plesnevič–Vizing [1965])
8.1.37. (♦) Prove that a graph G is 2 ¾ -colorable if and only if G is a union of ¾ bipartite
graphs. (Harary–Hsu–Miller [1977])
8.1.38. (♦) Prove that every ¾-chromatic graph has at least (¾2) edges. Conclude that if
√
G is the union of t copies of K t , then (G) < 1 + t t − 1. (Comment: This bound is near
tight , but see Conjecture 10.1.33.) (Horák–Tuza [1990])
√
8.1.39. Let G be an n-vertex graph. Prove (G) · (G) ≥ n and (G) + (G) ≥ 2 n. Prove
√
that equality can hold when n is an integer. (Nordhaus–Gaddum [1956], Finck [1968])
8.1.40. (♦) Let G be an n-vertex graph.
a) Prove (G) + (G) ≤ n + 1. (Nordhaus–Gaddum [1956])
(n+1)2
b) Using part (a), prove (G) · (G) ≤ (n + 1)2/4. From this, prove (G) ≤ 4 (G)
. For
(n+1)2
each odd n, construct G such that (G) = 4 (G)
. (Nordhaus–Gaddum [1956], Finck [1968])

8.1.41. For edge-disjoint graphs G and H on the same set of n vertices, let  (G , H) =
(G)+ (H)− (G ∪ H). Determine max  (G , H) and limn→∞ min n(G ,H) . (Hint: Use Exercise
8.1.40(a).) (Bloome–Johnson–Saritzky [2012])
8.1.42. (♦) Use the Gallai–Roy Theorem (Corollary 8.1.14) to prove that every n-vertex
digraph D has a path with at least n/(D) vertices. Conclude the following theorem of
Erdős–Szekeres [1935]: every list of rs + 1 distinct numbers has an increasing sublist of
size r + 1 or a decreasing sublist of size s + 1. (Schmeichel)
8.1.43. Paths and chromatic number in digraphs.
(a) Let G be the union of graphs F and H . Prove (G) ≤ (F) (H).
(b) Let D be an orientation of a graph G with (G) > rs. Assign each v ∈ V(D) a real
number  (v). Use (a) and the Gallai–Roy Theorem to prove that D has a path ⟨u0 , . . . , ur ⟩
with  (u0) ≤· · ·≤  (ur) or a path ⟨v0 , . . . , vs ⟩ with  (v0) >· · ·>  (vs).
(c) Use part (b) to conclude the Erdős–Szekeres Theorem (Exercise 8.1.42).
8.1.44. Prove that every digraph decomposes into two acyclic digraphs.

8.2. Structural Aspects

Since computing the chromatic number is very difficult (testing Ò(G) ≤ ¾ is
NP-complete when ¾ is a fixed integer at least 3), we study various structural
aspects of ¾-chromatic graphs to understand what makes them hard to color.

COLOR-CRITICAL GRAPHS

8.2.1. Definition. A graph G is color-critical if Ò(H) < Ò(G) for every proper
subgraph H of G. If also Ò(G) = ¾ , then G is ¾-critical.

Every ¾-chromatic graph contains a ¾-critical subgraph. The only 1-critical

and 2-critical graphs are K 1 and K 2 . The 3-critical graphs are the odd cycles.
Some structural properties follow easily.
Section 8.2: Structural Aspects 345

8.2.2. Proposition. If G is a ¾-critical graph, then (G) ≥  − 1.

Proof: If some vertex x has smaller degree, then a proper ( − 1)-coloring of G − x
(guaranteed by -criticality) extends to G by giving x a color not used on N(x).

Proposition 8.2.2 can be used instead of greedy coloring to prove Propositions

8.1.10–8.1.12 (Exercise 10). Using greedy coloring underscores the algorithmic
nature of upper bounds. Properties of -critical graphs don’t help to find optimal
colorings, because we have no good algorithm to find -critical subgraphs. Nev-
ertheless, these properties are useful for proving bounds and for understanding
the structure of graphs with large chromatic number.
The technique of proof in Proposition 8.2.2 is typical for proving properties
of -critical graphs. If the desired property of -critical graphs fails for such a
graph G, then we obtain a proper ( − 1)-coloring of an appropriate subgraph of
G and use the failure of the desired property to produce a proper ( − 1)-coloring
of G. This contradiction implies that the desired property does hold.

8.2.3. Remark. A nontrivial graph G is color-critical if and only if

(1) G has no isolated vertex, and
(2) (G − e) < (G) for every e ∈ E(G).
Hence to prove that a connected graph is color-critical, we need only consider sub-
graphs obtained by deleting a single edge.

8.2.4. Proposition. If v ∈ V(G) and (G − v) < (G) =  , then G has a proper -
coloring  having  (v) only on v and the other  − 1 colors all on N(v).
If e ∈ E(G) and (G − e) < (G) =  , then every proper ( − 1)-coloring of
G − e gives the same color to the endpoints of e.
Proof: Let  be a proper ( − 1)-coloring of G − v. If any color is not used in
N(v), then we can use it on v to complete a proper ( − 1)-coloring of G. This
contradicts (G) =  , so the colors all appear on N(v), and extending  to G by
letting  (v) =  completes the desired coloring.
If some proper ( − 1)-coloring of G − e gave distinct colors to the endpoints
of e, then it would also be a proper ( − 1)-coloring of G.

Remark 8.2.3 and Proposition 8.2.4 can be used to show that if G is color-
critical, then applying Mycielski’s Construction yields a color-critical graph with
larger chromatic number (Exercise 16). For -critical graphs, we strengthen
Proposition 8.2.2 to a lower bound on connectivity instead of minimum degree.

8.2.5. Theorem. (Dirac [1953]) Every -critical graph is ( − 1)-edge-connected.

Proof: (Dirac–Sorensen–Toft [1974]) Let G be -critical, and let [X , Y ] be an edge
cut. Since G is -critical, G[X] and G[Y ] are ( − 1)-colorable. Let X 1 , . . . , X ¾−1
and Y1 , . . . , Y¾−1 be the color classes in proper ( − 1)-colorings of G[X] and G[Y].

X2 X1 Y1 Y2
X Y
X3 X4 Y4 Y3
346 Chapter 8: Coloring

By induction on r, we show that if ##### [X , Y ]##### < r with X and Y partitioned

into X 1 , . . . , X r and Y1 , . . . , Yr , then Y1 , . . . , Yr can be reindexed so that [X i , Yi]
is empty for all i. This is trivial for r = 1. For r > 1, some X i is incident to no
edge of [X , Y ]. If some e ∈ [X , Y ] exists, then let Yi be the subset of Y incident to
e; otherwise any pairing works. The induction hypothesis completes the pairing.
We can thus pair the colors in the proper (¾ − 1)-colorings of G[X] and G[Y]
to produce a proper (¾ − 1)-coloring of G, which contradicts the hypothesis.

A ¾-critical graph need not be (¾ − 1)-connected (Exercise 8 and Exercise 24).

Theorem 8.2.5 holds also for graphs where deleting any vertex decreases the chro-
matic number. This is a larger class (Exercise 5).

LIST COLORING

In graph coloring problems, some colors may be forbidden from certain ver-
tices, representing unavailable resources. Also, when extending partial color-
ings, the color on a colored vertex is forbidden from use on its neighbors. Such
applications lead to the more general model of “list coloring ”, introduced in Vizing
[1976] and Erdős–Rubin–Taylor [1979] and surveyed in Stiebitz–Voigt [2015].

8.2.6. Definition. For a graph G, a list assignment L assigns to each vertex

v ∈ V(G) a set L(v) of colors allowed at v. An L-coloring is a proper color-
ing á of G such that á(v) ∈ L(v) for all v. A graph G is ¾-choosable or list
¾-colorable if it has an L-coloring whenever | L(v)| ≥ ¾ for all v. The list
chromatic number or choice number or choosability Ò l(G) is the mini-
mum ¾ such that G is ¾-choosable.

Here the “lists” are actually sets; there is no order or multiplicity for the
colors. Nevertheless, the term “list ” is thoroughly entrenched for this model, and
“set coloring ” has other meanings. Exercises 30–39 concern list coloring.
One option for L is assigning all vertices the same list S, in which case an L-
coloring exists if and only if | S| ≥ Ò(G). Hence Ò l(G) ≥ Ò(G) for every graph G.
However, Ò l(G) may be much larger than Ò(G), even when G is bipartite.

8.2.7. Example. Choice numbers of C2m and K m ,m.

Even cycles are 2-choosable. We may assume that all lists have size 2. If all
the lists are the same, then alternate the colors along the cycle. Otherwise, find
adjacent vertices x and y such that L(x) has a color c not in L(y). Use c on x, and
then color the other vertices in order from x to y. At each new vertex, choose
a color from its list that was not used on the previous vertex. Such a choice is
always available, and the coloring is proper since the colors on x and y also differ.
As observed in Erdős–Rubin–Taylor [1979], Ò l(K m ,m) > ¾ when m = (2¾¾−1 ).
In each part, let L assign all ¾-subsets of [2 ¾ − 1] as lists. In an L-coloring, at
least ¾ colors must be used on each part, since otherwise some vertex has no color
chosen from the ¾ in its list. Since altogether the lists have only 2 ¾ − 1 colors,
some color is used on each part. Since K m ,m is complete bipartite, we now have
adjacent vertices with the same color, so the coloring is not proper. Since K m ,m is
not L-colorable for this assignment of lists of size ¾ , it is not ¾-choosable.
Section 8.2: Structural Aspects 347

Some bounds on chromatic number are strengthened by proving that they

hold also for l , such as the Szekeres–Wilf (degeneracy) bound.

8.2.8. Proposition. Every -degenerate graph is ( + 1)-choosable. Thus also

l(G) ≤ 1 + max H ⊆ G (H) ≤ 1 + (G).
Proof: When G is -degenerate, iteratively delete a vertex with at most  neigh-
bors in what remains. In the reverse ordering,  , each vertex has at most  earlier
neighbors. When all lists have size at least  + 1, coloring the vertices in the or-
der of  guarantees that each vertex, when reached, has a color in its list not used
on earlier neighbors.

We will prove l(G) ≤ (G) when G is connected and not a complete graph or
an odd cycle. This is stronger than Brooks’ Theorem, since l(G) ≥ (G).

8.2.9. Lemma. (Vizing [1976]) Given a connected graph G, let L be a list assign-
ment such that | L(v)| ≥ d(v) for all v.
(a) If | L(y)| > d(y) for some vertex y, then G is L-colorable.
(b) If G is 2-connected and some two lists differ, then G is L-colorable.
Proof: (a) Order the vertices by iteratively deleting leaves of a spanning tree
rooted at y; each vertex before y has a later neighbor. Choose colors for ver-
tices in this order. At a vertex v earlier than y, we have colored fewer than d(v)
neighbors (since v has a later neighbor), so a color remains available in L(v) for v.
Although y has no later neighbor, | L(y)| > d(y), so also at y a color is available.
(b) Since the lists are not identical and G is connected, we can find adjacent x
and y such that L(x) − L(y)
= ∅. Choose c ∈ L(x) − L(y). Define L on V(G − x) by
L (v) = L(v) if v ∈
/ N(x) and L (v) = L(v) − c if v ∈ N(x). We have | L (v)| ≥ d G− x(v)
for all v ∈ V(G − x), and | L (y)| > d G− x(y) (since c ∈
/ L(y)). Hence part (a) yields an

L -coloring of G − x, which extends to an L-coloring of G by using c on x.

8.2.10. Definition. A graph G is  -choosable if it is L-colorable whenever


| L(v)| ≥  (v) for each vertex v, where  : V(G) → . The graph is degree-
choosable if it is L-colorable whenever | L(v)| ≥ d(v) for each vertex v.

8.2.11. Lemma. If a connected graph G has a degree-choosable induced subgraph

H , then G is degree-choosable.
Proof: Let L be a list assignment with | L(u)| ≥ d(u) for each u ∈ V(G). Using the
same lists, every component of G − V(H) has a vertex y with d G−V(H)(y) < | L(y)| ,
since G is connected. By Lemma 8.2.9a, G − V(H) has an L-coloring  .
Form a list assignment L for H by deleting from L(v) the colors used by 
on neighbors of v. We lose at most one color for each neighbor not in V(H). Since
| L(v)| ≥ d G(v), we have | L (v)| ≥ d H (v) for v ∈ V(H). Since H is degree-choosable,
H has an L -coloring  . Together,  and  form an L-coloring of G.

•
y• •v
G H • •
• •
•
348 Chapter 8: Coloring

We need an elementary structural result, called Rubin’s Block Theorem

in Entringer [1985]. The proofs by Entringer and Hladký–Král’–Schauz [2010]
are shorter than the original, as is the proof here.

8.2.12. Lemma. (Erdős–Rubin–Taylor [1979]) Every 2-connected graph G that

is not a complete graph or odd cycle has an even cycle with at most one chord.
Proof: Since G is 2-connected, G contains a cycle. If G contains a triangle, then
let Q be a largest complete subgraph of G. Since G is not complete, G has a vertex
outside Q. Since G is 2-connected, we may pick a shortest path P joining some
vertices u and v of Q via edges not in Q. If P has length at least 3, then its inter-
nal vertices have no neighbors in Q, and P forms the desired even cycle with uv
or with a u , v-path of length 2 in Q. If P has length 2, then its internal vertex
has some nonneighbor w in V(Q), and the desired cycle is [u , , v , w].
Hence G has no triangle. A shortest cycle C in G has no chord, so we may
assume that C has odd length. Since G is not an odd cycle and is 2-connected,
some path joins vertices of C via edges not in C, forming an even cycle with part
of C. Hence G has a shortest even cycle C . Chords of C create two odd cycles
with C . Any two such chords yield a shorter even cycle (whether they cross or
not), since G has no triangle. Thus C is an even cycle with at most one chord.

The chromatic number of a complete graph or odd cycle exceeds its degree, so
these graphs are not degree-choosable.

8.2.13. Theorem. (Borodin [1977], Erdős–Rubin–Taylor [1979]) If graph G is not

degree-choosable, then every block of G is a complete graph or an odd cycle.
Proof: By Lemma 8.2.11, it suffices to show that any block B other than a com-
plete graph or odd cycle is degree-choosable. A single edge is a complete graph,
so we may assume B is 2-connected. By Lemma 8.2.12, B has a subgraph that is
an even cycle with at most one chord. By Lemma 8.2.11, it suffices to show that
such a graph H is degree-choosable.
Let L be a list assignment with | L(v)| ≥ d H (v) for all v. If the lists are not
identical, then Lemma 8.2.9b applies (H is 2-connected). If they are identical,
then ordinary proper coloring suffices, using two colors when H is an even cycle,
three colors when H is a cycle with one chord.

Exercise 30 establishes the converse of Theorem 8.2.13; degree-choosability

fails when every block is a complete graph or odd cycle.

8.2.14. Corollary. (List Extension of Brooks’ Theorem) If a connected graph

G is not a complete graph or an odd cycle, then l(G) ≤ (G).
Proof: We prove the contrapositive. Suppose l(G) > (G).
By Proposition 8.2.8, G is not ( (G) − 1)-degenerate. Hence (H) ≥ (G) for
some subgraph H. Since G is connected and vertices with degree (G) in H have
no neighbors outside H , we have H = G.
Theorem 8.2.13 now implies that every block of G is a complete graph or an
odd cycle. Since G is regular, G must therefore have only one block, since the cut-
vertex in a leaf block would have higher degree than the other vertices. Hence G
is a complete graph or an odd cycle.
Section 8.2: Structural Aspects 349

FORCED SUBGRAPHS (optional)

What subgraphs must appear in a ¾-chromatic graph? We have seen that

the chromatic number can be arbitrarily large even when triangles are forbidden.
Thus a ¾-chromatic graph need not contain K ¾ , but perhaps it must contain some
weaker version of a ¾-clique.

8.2.15. Remark. For a graph F , an F-subdivision is a graph obtained from F

by successive edge subdivisions. Equivalently, each edge of F is replaced by a path
of length at least 1 whose internal vertices are new vertices.
The Hajós Conjecture asserts that every ¾-chromatic graph contains a K ¾ -
subdivision. The thesis of Dirac [1951, p. 5] attributes this question to Hajós in
the 1940s. The claim is trivial for ¾ ≤ 3, since every 3-chromatic graph contains
a cycle. As we will show next, Dirac [1952b] proved a stronger version of the con-
jecture for ¾ = 4. Catlin [1979] found counterexamples for ¾ ≥ 7 (see Exercise 42
for ¾ ∈ {7 , 8}). For ¾ = 5 and ¾ = 6, the question is open.
The Hadwiger Conjecture (Hadwiger [1943]) is weaker: every ¾-chromatic
graph contains a subgraph reducible to K ¾ by edge contractions. This is weaker
because a K ¾ -subdivision is a graph reducible to K ¾ by edge contractions. The
Hadwiger Conjecture remains open for ¾ ≥ 7. For ¾ = 4, it is equivalent to
Hajós’ Conjecture (and proved by Hadwiger [1943]). For ¾ = 5, it is equivalent to
the Four Color Theorem described in the next chapter. For ¾ = 6, it was proved
by Robertson–Seymour–Thomas [1993] using the Four Color Theorem. Seymour
[2016] gave a survey of results on Hadwiger ’s Conjecture.
In fact, Hajós’ Conjecture fails for almost all graphs, while Hadwiger ’s Con-
jecture holds for almost all graphs, as discussed in Section 14.4.

Every 4-chromatic graph has a 4-critical subgraph, which has minimum

degree at least 3. Hence our next theorem is stronger than guaranteeing K 4-
subdivisions in 4-chromatic graphs.

8.2.16. Definition. For S ⊆ V(G), an S-lobe of G is a subgraph of G induced by

the union of S and the vertices of a component of G − S.

8.2.17. Theorem. (Dirac [1952a]) Every graph with minimum degree at least 3
contains a K 4-subdivision.
Proof: Say that a vertex is weak if it has degree less than 3. By induction on
n, we prove the stronger result that every nontrivial n-vertex graph G with at
most one weak vertex contains a K 4-subdivision. We need this stronger statement
because we will apply the induction hypothesis to graphs having a weak vertex.
For the base step, the only nontrivial graph with at most four vertices that has
at most one weak vertex is K 4 itself, which trivially contains a K 4-subdivision.
For S ⊆ V(G), only one S-lobe of G can contain a weak vertex of G not in S,
since G has only one weak vertex. All other S-lobes are good S-lobes. Let S be
a smallest vertex cut. When | S| ≤ 1, let G be a good S-lobe. The only possible
weak vertex in G is the vertex of S (when | S| = 1). By the induction hypothesis,
G contains a K 4-subdivision, which is also contained in G.
350 Chapter 8: Coloring

When | S| = 2, let S = {x , y}, and obtain G from a good S-lobe of G by adding

/ E(G) (see the figure below). In G only x and y can be weak. If at most
xy if xy ∈
one is weak, then G contains a K 4-subdivision H. Replacing xy (if xy ∈ E(H))
with a path through another S-lobe of G yields a K 4-subdivision in G.

x
• • • •
S H
• • • •
y

If both x and y are weak in G , then each has exactly one neighbor outside S
in H. These neighbors are distinct, since (G) ≥ 2. Form G from G by contract-
ing the edge xy. The only weak vertex in G is the new vertex  , so G contains
a K 4-subdivision H. Subdividing an edge incident to  in H (if  ∈ V(H)) yields a
K 4-subdivision in G . Again G contains a K 4-subdivision by replacing xy with a
path through another component of G − S, if xy is used.
Hence we may assume that G is 3-connected. Select x ∈ V(G). Since G − x is
2-connected, it has a cycle C of length at least 3. Since G is 3-connected, the Fan
Lemma (Lemma 7.2.12) yields an x , V(C)-fan of size 3 in G. These three paths
combine with C to form a K 4-subdivision in G.

Many forced structures in -chromatic graphs are in fact forced by minimum

degree  − 1, which occurs in a -critical subgraph. For example, large chro-
matic number forces a subdivision of K t because large minimum degree is enough.
Mader [1967] proved that (G) ≥ 2(2) yields a K t-subdivision in G. Thomassen
t

[1988] extended the argument to F-subdivision for any graph F .

8.2.18. Lemma. (Mader [1967], Thomassen [1988]) If (G) ≥ 2  , then G con-

tains disjoint subgraphs G and H such that (1) (G ) ≥  , (2) each vertex of
G has a neighbor in H , and (3) H is connected.
Proof: We may assume that G is connected. When we contract the edges of a
connected subgraph H , the set V(H ) becomes a single vertex; let G · H denote
the resulting graph. Consider all connected induced subgraphs H of G such that
G · H has at least  | V(G · H )| edges. Since (G) ≥ 2  , every 1-vertex subgraph
has this property, so we may let H be a maximal such subgraph. Let G be the
subgraph induced by the vertices of G − V(H) that have neighbors in H.
It remains only to show (G ) ≥  . Each x ∈ V(G ) has a neighbor y in V(H).
In G · (H ∪ xy), the edges incident to x in G collapse onto edges from V(G ) to
the single vertex representing H in G · H. The edge xy also contracts. Other
edges of G · H remain in G · (H ∪ xy), so contracting the edge xy loses d G (x) + 1
additional edges. The maximality of H implies that contracting xy loses at least
 + 1 additional edges. Thus d G (x) ≥  . This holds for each x, so (G ) ≥  .

•
G x • •y H
•
Section 8.2: Structural Aspects 351

8.2.19. Theorem. (Mader [1967], Thomassen [1988]) If F has m edges and no

isolated vertex, and (G) ≥ 2 m , then G has an F-subdivision.
Proof: We use induction on m; the claim is trivial for m = 1. Consider m ≥ 1.
By Lemma 8.2.18, there are disjoint subgraphs H and G in G such that H is
connected, (G ) ≥ 2 m−1 , and each vertex of G has a neighbor in H.
If (F − e) ≥ 1 for some e ∈ E(F), then the induction hypothesis yields an
(F − e)-subdivision J in G . Add a path through H connecting the vertices of J
that represent the endpoints of e.
If (F − e) = 0 for all e ∈ E(F), then every edge of F has a leaf endpoint, and
F is a forest of stars. If F = K 1 ,m , then F ⊆ G, since (G) ≥ 2 m ≥ m. Otherwise,
deleting the vertex set S of a copy of K 1 ,r in G leaves (G − S) ≥ 2 m − r − 1 ≥ 2 m−r
for r < m, and the remaining stars can be found iteratively in G − S.

8.2.20.* Remark. Better bounds can be proved with more detailed arguments.
When F is a tree with m edges, (G) ≥ m suffices (Exercise 5.4.17). For F = K ¾ ,
let () be the smallest (G) that forces a K ¾ -subdivision in G; Theorem 8.2.19
yields () ≤ 2(2). In fact, 18  2 < () < c 2 . Komlós–Szemerédi [1996] and Bol-
¾

lobás–Thomason [1998] proved the upper bound (the latter shows c ≤ 256); the
lower bound follows from K r,r with r about (¾/2 2 )
(Exercise 41).
Theorem 8.2.17 yields (4) = 3. Furthermore, (5) = 6. The icosahedron
(pictured in Exercise 9.3.32) yields (5) ≥ 6, since this graph is 5-regular and has
no K 5-subdivision (being planar). Equality holds because every n-vertex graph
with at least 3n − 5 edges contains a K 5-subdivision (conjectured by Dirac [1964]
and proved by Mader [1998]).

The chromatic number is unbounded not only for triangle-free graphs, but
in fact for the family graphs with any fixed girth, as shown in Chapter 10 and in
Chapter 14. Thus the only fixed graphs that must appear in graphs with large
chromatic number are forests.
Since G contains every -vertex tree when (G) ≥  − 1 (Exercise 5.4.17),
each -vertex tree appears in every -critical graph. To strengthen this, we seek
a given forest F as an induced subgraph. Equivalently, we want to show that
forbidding large cliques and induced copies of F bounds the chromatic number.

8.2.21. Definition. A family G of graphs is -bounded if there is a function

such that (G) ≤ ((G)) for G ∈ G. A family is weakly -bounded if there
is a bound on the chromatic number of its triangle-free members. A graph is
F-free if it does not contain the graph F as an induced subgraph. The family
of F-free graphs is Forb(F).

If Forb(F) is -bounded and F is an induced subgraph of F , then Forb(F )

is also -bounded, since Forb(F ) ⊆ Forb(F). We have observed that if Forb(F) is
-bounded, then F must be a forest.
8.2.22. Conjecture. (Gyárf ás [1975], Sumner [1981]) For every forest F ,
Forb(F) is -bounded.

For example, Forb(2K 2) is -bounded (Exercise 8.1.28): G ∈ Forb(2K 2) im-

+1
plies (G) ≤ ( (G)
2
). In studying Conjecture 8.2.22, it suffices to consider trees,
352 Chapter 8: Coloring

since forests are induced subgraphs of trees. The broom Br,s is the tree with r + s
vertices consisting of the path Pr plus s edges attached to one end. Gyárf ás [1985]
proved that Forb(Br,s) is -bounded. We present only a weaker earlier result.

8.2.23. Theorem. (Gyárf ás–Szemerédi–Tuza [1980]) The class Forb(Br,s) is

weakly -bounded. In particular, triangle-free Br,s-free graphs have chro-
matic number less than r + s.
Proof: It suffices to show that every vertex-color-critical triangle-free graph with
chromatic number r + s has Br,s as an induced subgraph. We prove a stronger
statement. Let the end of the broom Br,s be the vertex farthest from the vertex
of degree s + 1. Let v be a vertex in a triangle-free connected graph G such that
(G) ≥ r + s. We prove that if all vertices other than v have degree at least r + s − 1,
then G has an induced copy of Br ,s with its end at v, for some r with r ≥ r.
We use induction on r. For r = 2, the desired broom is a star, and we can take
any neighbor of v as the center, since G is triangle-free.
Consider r > 2. Let G = G − N[v]. From G , iteratively delete vertices that
have degree less than r + s − 2 in the graph that remains. Let W be the final set of
vertices; we claim (G[W]) ≥ r + s − 1. Otherwise, iteratively replace the deleted
vertices and color them properly from [r + s − 2] to obtain a proper (r + s − 2)-
coloring of G , add N(v) with color r + s − 1, and assign any color in [r + s − 2] to
v. This yields a proper (r + s − 1)-coloring of G, a contradiction.
Hence (H) ≥ r + s − 1 for some component H of G[W]. Let v be the last
vertex before V(H) on a shortest path P from v to V(H). Let H be the subgraph
of G induced by V(H) ∪ {v }. We have (H) ≥ r + s − 2, since vertices with smaller
degree would be deleted. By the induction hypothesis, H has an induced copy of
Br ,s with end at v for some r ≥ r − 1. Adding the rest of P completes the desired
induced broom, since on P no vertex before v has a neighbor in H.

H
• • • • G
v v
N(v)

8.2.24.* Remark. Other results toward Conjecture 8.2.22 are harder. Gyárf ás
[1985] proved that Forb(Bm ,n) is -bounded. K ierstead–Penrice [1994] proved
that Forb(T) is -bounded when T has radius 2 (Gyárf ás–Szemerédi–Tuza [1980]
had proved it weakly -bounded). The smallest non-broom tree not having radius
2 is the 7-vertex tree Q obtained by attaching an edge to a central vertex of P6 .
K ierstead–Penrice [1990] proved that Forb(Q) is weakly -bounded ( (G) ≤ 4 for
every triangle-free graph in Forb(Q)). K ierstead [1992] proved that Forb(T) is
weakly -bounded when T is the union of copies of P4 with a common endpoint.
K ierstead–Zhu [2004] further strengthened these results by proving that if
T is a tree of radius 3 obtained from a tree of radius 2 by subdividing every edge
incident to a central vertex, then Forb(T) is -bounded. In another direction,
Scott [1997] proved that for each tree T , there is a function T such that every
K ¾ -free graph with (G) >  () contains a subdivision of T as an induced sub-
graph. Thus Forb(T) is -bounded when T is a subdivision of a star.
Exercises for Section 8.2 353

EXERCISES 8.2

8.2.1. (−) Let º be a proper -coloring of a -chromatic graph G. Prove that each color
class under  has a vertex adjacent to vertices of all other colors.
8.2.2. (−) Suppose that (G − x − y) = (G) − 2 for all pairs x , y of distinct vertices. Prove
that G is a complete graph.
8.2.3. (−) Prove that the Petersen graph can be 2-colored so that the subgraph induced
by each color class has no vertex of degree exceeding 1.
8.2.4. (−) Prove (G) ≤ 1 + max H ⊆ G  (H) for every graph G. (Matula [1969])
8.2.5. (−) A graph G is vertex-color-critical if (G − v) < (G) for all v ∈ V(G). Prove
that every 3-chromatic vertex-color-critical graph is an odd cycle (and thus color-critical).
Prove that the graph on the left below is vertex-color-critical but not color-critical.

•
•
• • •
• • • •
•
• • • • •
• •

8.2.6. (−) Prove that the graph on the right above is not 2-choosable.
8.2.7. (−) Given that minimum degree 3 forces a K4 -subdivision (Theorem 8.2.17), prove
that the maximum number of edges in an n-vertex graph with no K4 -subdivision is 2n − 3.
(Dirac [1964] conjectured and Mader [1998] proved that every n-vertex graph with at least
3n − 5 edges contains a subdivision of K 5 .)
8.2.8. Prove that the graphs below are 4-critical, thereby confirming that -critical
graphs need not be ( − 1)-connected.
• • • •
• • •
• • •
• •
• • •
• • • •

8.2.9. The smallest -critical graph other than K ¾ is C5 K ¾−3 .

(a) Prove that N(x) ⊆ N(y) is impossible when x and y are vertices in a -critical graph
G. Conclude that no -critical graph has + 1 vertices.
(b) Prove that (G H) = (G) + (H) and that G H is color-critical if and only if
both G and H are color-critical. Conclude that C5 K ¾−3 is -critical. (Comment: In fact ,
it is the only -critical graph with + 2 vertices.)
8.2.10. Use Proposition 8.2.2 to prove Propositions 8.1.6, 8.1.10, and 8.1.12.
8.2.11. Vertex cuts in color-critical graphs. Let G be a -critical graph with a vertex cut
S. Prove | S| ≥ 2. Prove that if S = {x , y}, then xy ∈
/ E(G). Prove also that G has exactly
two S-lobes (Definition 8.2.16), which can be named G1 and G 2 so that G1 + xy and G 2 · xy
(add xy and then contract it) are -critical.
8.2.12. Determine the least number of edges in an n-vertex -chromatic connected graph.
(Eršov–Kožuhin [1962]; see Bhasker–Samad–West [1994] for higher connectivity.)
354 Chapter 8: Coloring

8.2.13. For j < ¾ < n − j , determine the least number of edges in an n-vertex ¾-chromatic
graph with minimum degree at least j .
8.2.14. For the graph G on the left below, obtain (G) and a (G)-critical subgraph.
• • • • • •
• • • • •
• • • • • •
• • • • •
• • • • • •
A X Y B

8.2.15. Dense 4-critical graphs. For odd m, form G m with V(G) = A ∪ X ∪ Y ∪ B as fol-
lows. Let G[ A] = G[B] = Cm , let G[X ∪ Y ] = K m ,m (an X , Y -bigraph), and add matchings
joining A to X and B to Y (G 5 appears above). Prove that G m is 4-critical. (Toft [1970])
8.2.16. (♦) Let G be a color-critical graph. Prove that the graph G generated from G by
Mycielski’s Construction (Definition 8.1.15) is also color-critical.
8.2.17. (♦) A special case of the Erdős–Lovász Tihany Conjecture (see Erdős [1968]) is that
if (G − x − y) = (G) − 2 for all xy ∈ E(G) and G is connected, then G is a complete graph.
Prove this for (G) ≤ 4. (Comment: For (G) = 5, see Stiebitz [1987] and Mozhan [1987].)
8.2.18. Let G be an n-vertex graph with m edges.
(a) Prove that if (G) ≤ r, then m ≤ (1 − 1/r)n2/2, and hence (G) ≥ n2/(n2 − 2m).
(Hint: Consider the neighborhood of a vertex of maximum degree.) (Myers–Liu [1972])
(b) Use part (a) to prove (G) ≥ ⌈ n/(d + 1)⌉ , where d is the average vertex degree of
G. (Erdős–Gallai [1961])
8.2.19. (♦) Let G be a ( + 1)-critical graph. Prove that every edge of G lies in at least
( − 1)! cycles whose lengths are congruent to 1 modulo  . (Moore–West [2019])

8.2.20. (♦) Let G be the 12-vertex graph on the left below. Prove (G) = 4. Determine
whether G is 4-critical. (Hint: Consider the independent sets. Comment: G is the smallest
triangle-free 4-regular 4-chromatic graph.) (Chvátal [1970])

• G1 G2 G3
• • • • • • •
• • • • •
• •
• •
• • • • • • • • • • • • T
•

8.2.21. Let G1 = K1 . For  > 1, construct G ¾ by adding to G1 + · · · + G ¾−1 a set T of

¾ −1
∏i=1 |V(Gi)| additional vertices, such that for each (v1 , . . . , v¾−1) ∈ V(G1) × · · · × V(G¾−1),
one vertex of T has neighborhood {v1 , . . . , v¾−1 }. The graph on the right above suggests
G4 . Prove that G ¾ is -critical and triangle-free. (Zykov [1949], Schäuble [1969])
8.2.22. Let {X , Y } be a partition of V(G). Prove that if G[X ] and G[Y ] are -colorable,
then (G) ≤  + m/ , where m = | [X , Y ]| . (Hakimi–Schmeichel [2004])
Exercises for Section 8.2 355

8.2.23. Let G be a graph, and let t = ⌈ ( (G) + 1)/j ⌉ . Prove that V(G) has a partition
V1 , . . . , Vt such that each G[Vi ] has no j-edge-connected subgraph. Prove that ⌈ (G)/j ⌉
classes do not suffice in the following cases: (1) G is a j-regular j-edge-connected graph,
(2) G = K n with n ≡ 1 (mod j), (3) j = 1 and G is an odd cycle. (Comment: Matula [1973]
generalized Brooks’ Theorem by showing that ⌈ (G)/j ⌉ suffices except in these cases.)

8.2.24. Hajós construction. For ¾ ≥ 3, let G and H be ¾-critical graphs sharing only
vertex v, with vu ∈ E(G) and vw ∈ E(H). Prove that (G − vu) ∪ (H − vw) ∪ uw is ¾-critical.
Use this to construct 4-critical n-vertex graphs for n ≥ 6. (Hajós [1961])

8.2.25. (♦) Generalized Hajós construction. Let G1 and G 2 be disjoint ¾-critical graphs.
Choose w ∈ v(G1 ) with d G 1 (w) = ¾ − 1, and partition NG 1 (w) into nonempty sets X and Y .
Given uv ∈ G 2 , form G from (G1 − w) + (G 2 − uv) by adding X to N(u) and Y to N(v).
(a) Prove that G is ¾-critical.
(b) For ¾ ≥ 3, apply this operation once to construct a ¾-critical n-vertex graph with
degree-sum n(¾ − 1) + ¾ − 3, where n = 2 ¾ − 1. (Comment: Every ¾-critical n-vertex non-
complete graph has degree-sum at least n(¾ − 1) + ¾ − 3 (Dirac [1957]). Kostochka–Yancey
[2014] proved a conjecture of Gallai [1963a] and almost a conjecture of Ore [1967] by show-
ing that every ¾-critical n-vertex graph has at least ⌈ (¾+1)(¾2(−¾2)n−¾(¾ −3)
−1) ⌉ edges.)
8.2.26. Let G be a connected graph that is not a complete graph or a cycle whose length
is an odd multiple of 3. Prove that every proper (G)-coloring of G has two vertices of the
same color with a common neighbor. (Tomescu [1990])

8.2.27. (♦) Extension of Brooks’ Theorem. Let G be a graph.


(a) Given D1 , . . . , Dt ∈ 0 such that ∑ Di > (G), prove that there is a partition
V1 , . . . , Vt of V(G) such that (G[Vi ]) < Di for each i. (Lovász [1966])
(b) Prove that if (G) = r, where 3 ≤ r ≤ (G), then (G) ≤ ⌈ r+r 1 ( (G) + 1)⌉ .
(Borodin–Kostochka [1977], Catlin [1978], Lawrence [1978])

8.2.28. (+) Given a graph G, let 1 , . . . , ¾ be nonnegative integer functions on V(G) such
that ∑i i(v) > d G(v) for all v ∈ V(G). Prove that G has a coloring c: V(G) → [] such that
for all v ∈ V(G), fewer than c(v)(v) neighbors of v have color c(v). (Bernardi [1987])

8.2.29. Gallai’s Theorem: If G is a ( + 1)-critical graph, then every block of the subgraph
of G induced by the vertices of degree  is an odd cycle or a complete graph. (Gallai [1963a])
(a) Use Theorem 8.2.13 to prove Gallai ’s Theorem.
(b) Gallai ’s Theorem can be proved without list colorings. Use Gallai ’s Theorem to
prove Brooks’ Theorem for chromatic number: Every connected graph G that is not (G)-
colorable is a complete graph or an odd cycle.

8.2.30. Prove that if every block of a graph G is a complete graph or an odd cycle, then G
is not degree-choosable. (Comment: This is the converse of Theorem 8.2.13.)

8.2.31. (♦) Let G be an n-vertex graph. Prove max H ⊆ G (H) + max H ⊆ G (H ) ≤ n − 1.
Conclude l(G) + l(G) ≤ n + 1. (Comment: This strengthens Exercise 8.1.40(a).)

8.2.32. (♦) Prove that K ¾ ,r is -choosable if and only if r <  ¾ . (Vizing [1976])

8.2.33. (+) Small Pot Lemma. Let L(X) = ⋃ v∈ X L(v) for X ⊆ V(G).
(a) Prove that if G is not -choosable, then G is not L-colorable for some -uniform
lists L with | L(V(G))| < | V(G)|. (Galvin [1998], Kierstead [2000], Reed–Sudakov [2002])
(b) Let G be a complete -partite graph. Prove that if one part has size 4, the others
have size 2, and  is even, then l(G) >  (Enomoto–Ohba–Ota–Sakamoto [2002]) (Com-
ment: This graph is a sharpness example for the conjecture of Ohba [2002], proved in Noel–
Reed–Wu [2015], that -colorable graphs with at most 2  + 1 vertices are -choosable.)
356 Chapter 8: Coloring

8.2.34. (♦) Let G be the complete ¾-partite graph obtained by deleting a perfect matching
from K 2¾ . Prove l(G) =  . (Erdős–Rubin–Taylor [1979])

8.2.35. Prove l(K3 ,3 ,1) = 3. (Comment: More generally, Gravier–Maffray [1998] proved
l(G) = (G) when G = K 3 ,3 ,2 ,... ,2 .)

8.2.36. (♦) Let K m∗¾ denote the complete -partite graph with m vertices in each part.
(a) Prove l(K m∗¾ ) ≥ ⌈ (m−1)(2m¾−1)+1 ⌉ . (Comment: The exact value is ⌈ 4¾3−1 ⌉ when m = 3
(Kierstead [2000]) and ⌈ 3¾2−1 ⌉ when m = 4 (Kierstead–Salmon–Wang [2014]).)
(b) Prove l(K m∗¾) > ( − 1) log m
log ¾
. (Noel–West–Wu–Zhu [2015]) (Hint: Consider m of
the form ((¾¾−j −1)j
1
). Comment: Alon [1992] proved l(K m∗¾ ) ∈ ( log m).)
8.2.37. For a graph G and list assignment L, let S = ⋃v∈V(G) L(v). Let H be the graph
with vertex set V(G) ∪ S consisting of G, a complete graph with vertex set S, and the edges
{vc: v ∈ V(G) , c ∈ S − L(v)}. Prove that G is L-colorable if and only if (H) ≤ | S|.

8.2.38. (♦) Reduction of L-colorability to independence number. For a graph G and list as-
signment L, let H be the graph with vertex set {vc : v ∈ V(G) , c ∈ L(v)} such that uc and
vc are adjacent if u = v or if uv ∈ E(G) and c = c . Prove that G is L-colorable if and only
if (H) = | V(G)|. (Vizing [1976])(Comment: This generalizes Exercise 8.1.36. A more gen-
eral model in which H allows any matching joining the cliques assigned to adjacent vertices
u and v in G was introduced in Dvořák–Postle [2018]; it is now called DP-coloring.)

8.2.39. (+) Characterization of 2-choosable graphs.

(a) For n ≥ 2, prove that Pn is not  -choosable when  is 1 at the endpoints and 2 else-
where. Prove that Cn is not  -choosable when  is 1 at one vertex and 2 elsewhere.
(b) Let L be a 2-uniform list assignment on a path with lists not all identical. Prove
that L-colorings exist having at least three of the four possibilities at the endpoints.
(c) The graph (l1 , ... , l¾) with branch vertices u and v is the union of  pairwise
internally-disjoint u , v-paths with lengths l1 , ... , l¾ . Prove that ( , l , m) with  ≤ l ≤ m
is 2-choosable if and only if ( , l , m) = (2 , 2 , 2q) for some q ∈ . 
(d) Prove that a connected bipartite graph is 2-choosable if and only if it has at most
one cycle or the edges in its cycles form (2 , 2 , 2q) for some q. (Erdős–Rubin–Taylor [1979])

8.2.40. (♦) Let G be a -colorable graph, and let P be a set of vertices in G such that
d(x , y) ≥ 4 whenever x , y ∈ P. Prove that every coloring of P with colors from [ + 1]
extends to a proper ( + 1)-coloring of G. (Albertson [1998])

8.2.41. (♦) Prove that K r,r contains a subdivision of K 2¾ if and only if r ≥ (¾+2 1 ) . Determine
the threshold on r for the existence of a K 2¾+1 -subdivision in K r,r .

8.2.42. Let G7 and G8 denote the graphs on the left and right below. Heavy edges in-
dicate that every vertex inside one circle is adjacent to every vertex inside another. For
 ∈ {7 , 8}, prove that (G ¾) =  and that G ¾ has no K¾ -subdivision. (Comment: Thus
Hajós’ Conjecture is false for  ∈ {7 , 8}.) (Catlin [1979])

• •
• • • •
• •
• • • • • • • •
G7 G8

• • • •
• • • • • • • •
Section 8.3: Edge-Coloring and Perfection 357

8.2.43. (+) Let F be a forest with m edges. Let G be a graph with at least | V(F)| vertices
such that (G) ≥ m. Prove F ⊆ G. (Brandt [1994])
8.2.44. (♦) Let G be a -chromatic graph with girth at least 5. Prove that G contains
every -vertex tree as an induced subgraph. (Hint: Reduce to the case (G) ≥  − 1 and
then use induction on  .) (Gyárfás–Szemerédi–Tuza [1980])
8.2.45. (+) Forcing  -colored subtrees in  -chromatic graphs. Let  be a proper -coloring
of a graph G, and let T be a tree with vertex set {w1 , . . . , w¾ }. Prove that there is a map
: V(T) → V(G) such that the images of adjacent vertices in T are adjacent in G and
 ((wi)) = i for all i. (Gyárfás–Szemerédi–Tuza [1980])
8.2.46. (+) Let G be a graph having a proper coloring in which no color class has size 1.
Prove that G has a proper (G)-coloring in which no color class has size 1. (Gallai [1963c])

8.3. Edge-Coloring and Perfection

Just as we partition V(G) into independent sets to avoid conflicts between
adjacent vertices, we can also partition E(G) into matchings to avoid conflicts of
incident edges. For example, when the vertices are people and the edges are pairs
that must meet, we cannot schedule two meetings for one person at the same time.
In a sense, this is a better-behaved special case of vertex coloring. Subsequently,
we put the results into perspective by discussing the more-better-behaved family
of “perfect graphs”, where the subject of min-max relations has a natural home.
8.3.1. Definition. A ¾-edge-coloring of a graph is a labeling of its edges from
a set of ¾ colors; it is proper if incident edges receive distinct colors. A
graph is ¾-edge-colorable if it has a proper ¾-edge-coloring, and the edge-
chromatic number or chromatic index (G) of G is the minimum  such
that G is -edge-colorable.
These definitions for edge-coloring extend without change to loopless multi-
graphs. Although multiedges are irrevelant for vertex colorings, they can greatly
affect edge-chromatic number. Loops cannot be properly colored.
8.3.2. Remark. Comparing (G) with (G) is not merely analogy; always
(G) = (L(G)). This usage of parallels  (G) = (L(G)).
Since the edges at a vertex must have distinct colors, (G) ≥ (G). Applying
greedy coloring to any ordering of the edges of a multigraph (or of the vertices of
its line graph) yields (G) ≤ 2 (G) − 1. (That is, (L(G)) ≤ 2( (G) − 1).)
A clique in the line graph L(G) corresponds to pairwise-incident edges in G.
In a graph (no loops or multiedges), such edges can have one common endpoint
or can form a triangle. Thus for graphs with maximum degree at least 4, the
statements (G) ≥ (G) and (L(G)) ≥ (L(G)) have the same content.

SPECIAL CL ASSES

We have the greedy upper bound (G) ≤ 2 (G) − 1. For multigraphs, the
bound cannot be reduced below 3 (G)/2.
358 Chapter 8: Coloring

8.3.3. Example. The fat triangle. In the multigraph below, each edge is incident
to all others, so all edges must receive distinct colors, and Ò (G) = 3 (G)/2.

• •

Shannon [1949] proved that Example 8.3.3 is the worst in terms of (G) (Ex-
ercise 26). Vizing [1964] and Gupta [1966] both proved the more detailed bound
Ò (G) ≤ (G) + Þ(G), where Þ(G) is the largest edge-multiplicity. For a graph
G, this reduces to Ò (G) ≤ (G) + 1 (graph restricts edge-multiplicities to {0 , 1}).
There is standard terminology for the two alternatives when G is a graph.

8.3.4. Definition. A graph G is Class 1 when Ò (G) = (G); Class 2 otherwise.

A decomposition of a multigraph into perfect matchings is a 1-factorization.

To be Class 1, a regular graph must be colored with the edges of each color
forming a perfect matching. Thus regular graphs of odd order are Class 2.

8.3.5. Example. Edge-coloring of K 2n . In a league with 2n teams, we want to

schedule games between all pairs, but each team plays at most once a week. Since
each team must play 2n − 1 others, the season lasts at least 2n − 1 weeks. The
games of each week must form a matching. We can schedule the season in 2n − 1
weeks if and only if K 2n has a 1-factorization.
The figure below suggests a solution. Arrange 2n − 1 vertices cyclically. Let
the length of an edge be the number of steps between its endpoints along the cir-
cle. There are 2n − 1 edges of each length 1 , . . . , n − 1. In the figure, the bold
matching has one edge of each length, plus an edge from the central vertex to one
vertex on the circle. Rotating the picture yields others matchings, again with one
edge of each length. The 2n− 1 rotations of the figure yield the desired matchings.
The 1-factorization of K 2n is related to design theory (see Chapter 13).
•
• •
• •
•
• •
• •

8.3.6. Example. The cube and the Petersen graph. The d-dimensional cube is
Class 1. A 1-factorization is given by letting the ith 1-factor consist of the edges
joining the d-tuples that differ in the ith coordinate.
The Petersen graph is Class 2. Deleting a 1-factor always leaves a 2-factor,
and the Petersen graph has no 10-cycle (Proposition 5.1.18). Thus the 2-factor
that remains must be 2C5 , and this has no 1-factor.
Section 8.3: Edge-Coloring and Perfection 359

There are many proofs that bipartite graphs are Class 1 (see Exercise 18).
Perhaps the easiest uses the Marriage Theorem. The proof is also valid when
multiedges are allowed; indeed, that simplifies the proof.

8.3.7. Theorem. (K önig [1916]) Ò (G) = (G) when G is a bipartite multigraph.

Proof: Let G be an X , Y-bigraph, and let ¾ = (G). Constructing a ¾-regular bi-
partite supergraph H of G reduces the problem to the regular case, because dis-
carding edges from a proper ¾-edge-coloring of H yields a proper ¾-edge-coloring
of G. To construct H , add vertices to the smaller of X and Y until both have the
same size. Then edges joining vertices of X and Y with degree less than ¾ until
all vertices have degree (G); multiedges may arise.
Hence we may assume G is ¾-regular. By the Marriage Theorem (Corollary
6.1.6), a ¾-regular bipartite multigraph has a 1-factor, and hence by induction
on ¾ it has a 1-factorization.

Although (G) ≤ Ò (G) ≤ (G) + 1 for every graph G, deciding between the
two options is NP-complete (Holyer [1981]). Thus we seek conditions for Class 1
or Class 2. For example, regular graphs with cut-vertices are Class 2 (Exercise
13). For cartesian products, G H is Class 1 if G or H is Class 1 or if G and H
both have 1-factors (Exercise 25).

8.3.8.* Remark. An overfull subgraph of a graph G is a subgraph H such

that | E(H)| > (G)⌊ | V(H)| /2⌋ . An overfull subgraph forces Ò (G) > (G) and
can only occur when | V(H)| is odd. When discussing also multigraphs, we add the
restriction (H) = (G), which holds implicitly for simple graphs.
The Overfull Conjecture (Chetwynd–Hilton [1984, 1986]; see also Hilton
[1989]) states that an n-vertex graph G with (G) > n/3 is Class 1 if and only
if G has no overfull subgraph H. Deleting one vertex from the Petersen graph
shows that the condition is not sufficient when (G) = n/3 (Exercise 21).
The Overfull Conjecture implies the following 1-Factorization Conjec-
ture: If ¾ ≥ m, or if ¾ ≥ m − 1 when m is even, then every ¾-regular graph
of order 2m is Class 1 (Exercise 22). This is sharp (Exercise 23).
The conclusions of both conjectures hold when the maximum degree is a large
enough fraction of the number of vertices (see Chetwynd–Hilton [1989], Niessen–
Volkmann [1990], Perkovic–Reed [1997], Plantholt [2004]).

VIZING’S THEOREM AND EXTENSIONS

Vizing [1964, 1965] and Gupta [1966] proved Ò (G) ≤ (G) + Þ(G) for every
loopless multigraph G. Commonly known as Vizing’s Theorem, this has many
proofs and many extensions to stronger results.
For example, let Þ(x , y) denote the multiplicity of xy as an edge. In addition,
let Þ(v) = max x∈ N(v) Þ(x , v). Ore [1968] proved
Ò (G) ≤ max {d(v) + Þ(v)} ≤ (G) + Þ(G).
v∈ V(G)
360 Chapter 8: Coloring

Ore [1967] also proved

(G) d(x) + d(y) + d()
≤ max { (G) , max },
⟨x ,y , ⟩∈ P 2
where P is the set of all 3-vertex paths. This implies the bound of Shannon [1949]:
(G) ≤ 3 (G). Both of Ore ’s bounds follow immediately from our main result:
2

8.3.9. Theorem. (Andersen [1977], Goldberg [1977, 1984]) If G is a loopless

multigraph, and P is the set of 3-vertex paths in G, then
(G) d(x) + (x , y) + (y , ) + d()
≤ max { (G) , max ⌊ ⌋} . (∗)
⟨x , y , ⟩∈P 2

The proof of the Anderson–Goldberg Theorem, like many proofs of edge-

coloring bounds, yields a polynomial-time algorithm to augment a partial proper
edge-coloring of G until all edges have been colored using the specified number of
colors. The actual value of (G) may be less than the bound being proved.
The proof here also yields other useful results such as Vizing ’s Adjacency
Lemma (Corollary 8.3.11). The tool that permits extension of partial edge-
colorings is the following technical statement. The proofs can be followed for the
special case of graphs, where all multiplicities equal 1. In the general case, “uv”
indicates a particular edge with endpoints u and v.


8.3.10. Theorem. For q ∈ , let G be a multigraph having an edge yw such that
d(y) , d(w) ≤ q and G − yw is q-edge-colorable. If
d(x) + (x , y) + (y , ) + d() ≤ 2q + 1
whenever x and  are distinct neighbors of y, then G is q-edge-colorable.

Proof of Theorem 8.3.9: We use induction on | E(G)| , with trivial basis. For
the induction step, let q be the bound computed in (∗), and let wy be an edge in
G. Since deleting edges cannot increase the computation in (∗), the induction
hypothesis implies that G − wy is q-edge-colorable. Also d(y) , d(w) ≤ (G) ≤ q.
When x and  are distinct neighbors of y, we have ⟨x , y , ⟩ ∈ P. Also
d(x)+ (x , y)+ (y , )+ d() ≤ 2q + 1, since otherwise this path ⟨x , y , ⟩ contradicts
the computation of q in (∗). Now Theorem 8.3.10 applies and yields (G) ≤ q.

Proof of Theorem 8.3.10: (Kostochka [2014]) Let G = G − yw. Let  be a

proper q-edge-coloring of G . For v ∈ V(G), let O(v) be the set of colors omitted at
v (used on no edges at v). If w is the only neighbor of y, then any color in O(w) also
lies in O(y), and using it on yw extends  to G. Hence we may assume | NG(y)| ≥ 2.
A list v0 , . . . , v¾ of neighbors of y forms a color fan if v0 = w and (vi y) ∈
O(vi−1) for i ≥ 1, where vi y is some edge joining vi and y (see w , x , v ,  below). If a
color  is missing at both y and v¾ , then moving color (vi y) to vi−1 y (for 1 ≤ i ≤ )
and using color  on v¾ y produces a proper q-edge-coloring of G. Hence (G) > q
requires O(y) ∩ O() = ∅ when  is reached by some color fan.
Let X be the set of vertices on color fans. Form an auxiliary digraph H with
vertex set X having an edge from u to v for each edge in G with endpoints v and
y whose color under  is in O(u). Color fans correspond to paths from w in H.
Section 8.3: Edge-Coloring and Perfection 361

H
,  ∈ O(w)  ,  ∈ O(x)  ∈ O(v)  ∈/ O()
• • • •
w x v
  
G

•
y  ∈ O(y)
Suppose  (G) >q. We will show that the sets O(v) are disjoint for v ∈ X ∪ {y}.
We first show how this yields bounds on | E(H)| leading to a contradiction.
Counting argument: When the sets O(v) are disjoint, each color lies in O(x)
for at most one x ∈ X , so each edge of G with endpoints v and y generates at most
one edge of H entering v. Thus counting E(H) by heads yields
| E(H)| = ∑v∈ X d−H (v) ≤ (∑v∈ X (v , y)) − 1,
where we lose 1 from the sum because the edge yw is not in G .
Next we count by tails. For v ∈ X , each color in O(v) appears on one edge y 
at y, which yields an edge v  in H. Thus d+H (v) = |O(v)| = q − d G (v) for v ∈ X .
Again adjusting by 1 since d G (w) = d G(w) − 1, we have
| E(H)| = ∑v∈ X d+H (v) ≥ 1 + ∑v∈ X (q − d G(v)).
Combining the two bounds on | E(H)| and grouping by vertices yields
∑v∈ X (d G(v) + (v , y)) ≥ q | X | + 2.
Note that q ≥ d G(w) yields d+H (w) ≥ 1 and | X | ≥ 2. By the given hypothesis, the
two largest terms on the left sum to at most 2q + 1, with one of them being at
most q. Thus the other | X | − 2 terms are also bounded by q, and the total sum is
at most 2 | X | + 1. This contradiction yields  (G) ≤ q if the sets O(v) are disjoint.
Disjointness argument: The definition of color fans yields O(y) ∩ O() = ∅
for  ∈ X . If d() = (G) = q, then O() is empty and disjoint from all sets, so we
may consider  ∈ O(y) and ∈ O(). Since  appears at  and does not, we can
follow a maximal path P from  in colors  and . Let u be the other end of P.
We claim u = y. If u
= y, then switching  and on P changes the omitted
set only at  and u. Let Q be a w , -path in H. The new coloring has a color fan
reaching a vertex where  does not appear: u if u ∈ V(Q) (since  ∈
/ O(u)), or  if
u∈/ V(Q). Again changing colors along the fan extends the q-edge-coloring to G.
If ∈ O(v) ∩ O(v ), then the paths from v and v via  and both reach y. The
vertex where they meet has two edges of the same color, a contradiction.

u• • •
 
w• x• •v •
∈ O()

y•  ∈ O(y)
362 Chapter 8: Coloring

For edge-coloring of simple graphs, a critical graph is a graph G with

(G) = (G) + 1 such that (H) < (G) for every proper subgraph H. Delet-
ing an edge e would still leave a Class 2 graph if all maximum-degree vertices of
G are incident to e. In fact, this cannot happen; every Class 2 graph has at least
three maximum-degree vertices. This follows immediately from our next result,
given the name Vizing’s Adjacency Lemma in Fiorini–Wilson [1977], a book
that explores early results on edge-coloring. We will apply it in Exercise 9.3.64.

8.3.11.* Corollary. (Vizing’s Adjacency Lemma; Vizing [1965]) If w and y

are adjacent in a critical graph G, then y has at least max{ (G) − d(w) + 1 , 2}
neighbors with degree (G).
Proof: Since G is critical, (G) > (G − yw) = q with q = (G) ≥ max{d(y) ,d(w)}.
Since (G) > q, Theorem 8.3.10 yields distinct x ,  ∈ N(y) having the property
that d(x)+ (x , y)+ (y , )+ d() > 2q+ 1. Since each multiplicity is 1, we conclude
that y has distinct neighbors with degree q.
In addition, since the counting bounds on E(H) in the proof of Theorem
8.3.10 used only the hypothesis (G) > q and not the assumed inequalities
d(x) + (x , y) + (y , ) + d() ≤ 2q + 1, the argument and computation producing
∑v∈ X (v , y) ≥ 2 + ∑v∈ X (q − d G(v) is also valid here, where X ⊆ N(y).
With multiplicity always 1, the inequality reduces to | X | ≥ 2 + ∑ v∈ X (q − d(v)).
Letting  be the number of vertices in X − {w} having degree (G), we have
 terms equal to 0 in the sum on the right, one term (G) − d(w) for w itself,
and | X | − 1 −  terms contributing at least 1 each. Thus the inequality becomes
| X | − 1 ≥ 1 + (G) − d(w) + | X | − 1 −  , which yields  ≥ (G) − d(w) + 1.

LIST EDGE-COLORING

We next consider the list version of edge-coloring, analogous to list coloring.

Just as the chromatic number behaves better when restricted to line graphs,
where it is equivalent to the edge-chromatic number of the original graph, also
the list chromatic number behaves better when restricted to line graphs.

8.3.12. Definition. A list assignment L for the edges of a graph G speci-

fies a list L(e) of available colors for each edge e. An L-edge-coloring is a
proper edge-coloring  with  (e) chosen from L(e) for each e. The list edge-
chromatic number or edge-choosability l(G) is the least  such that G is
L-edge-colorable whenever | L(e)| ≥  for all e ∈ E(G). Note that l(G) equals
l(L(G)), where L(G) is the line graph of G.

Note that (L(G)) ≤ 2 (G) − 2. Hence l(G) ≤ 2 (G) − 1 by greedy coloring,

as in Proposition 8.2.8. Thus l(G) < 2 (G), since (G) ≥ (G). As in ordinary
coloring, the edge version of choosability behaves much better than the vertex
version. Even so, the conjectured bound for edge-choosability is surprising. The
List Color Conjecture was posed independently by many researchers. It was pub-
lished in Bollobás–Harris [1985] but was independently formulated earlier by
Albertson and Collins in 1981 and by Vizing as early as 1975 (both unpublished).
Section 8.3: Edge-Coloring and Perfection 363

8.3.13. Conjecture. (List Color Conjecture) (G) = (G) for any graph G.

8.3.14.* Remark. Bollobás–Harris [1985] proved l(G) < c (G) for each con-
stant c exceeding 11/6, when (G) is sufficiently large. This and subsequent
improvements used probabilistic methods. K ahn [1996] proved the conjecture
asymptotically: l(G) ≤ (1 + o(1)) (G). Häggkvist–Janssen [1997] sharpened the
√
error term, proving l(G) ≤ D + 23D2/3 log D), where D = (G). Molloy–Reed
[2000] further sharpened it to l(G) ≤ D + O(D1/2(log D)4).
The special case of the List Color Conjecture for G = K n ,n was posed by Dinitz
in 1979 (Janssen [1993] proved the LCC for K n ,n−1). The Dinitz Conjecture be-
came popular in its matrix formulation: If each position of an n-by-n grid contains
a set of size n, then one can choose one element from each set so that those chosen
in each row are distinct and those chosen in each column are distinct.

Galvin [1995] proved the List Color Conjecture for bipartite multigraphs
(generalizing the Dinitz Conjecture). Galvin wrote “ The proof is very simple,
and uses no new ideas”. We start with a fundamental tool for list coloring.

8.3.15. Definition. A kernel in a digraph is an independent set S containing

a successor of every vertex outside S. A digraph is kernel-perfect if every
induced subdigraph has a kernel.

In a directed even cycle, either maximum independent set is a kernel. Indeed,

a directed even cycle is kernel-perfect, so the next result gives another proof that
even cycles are 2-choosable. Recall the definition of  -choosable from Definition
8.2.10. The next proof uses the same idea as Lemma 8.2.11.

8.3.16. Lemma. (Bondy–Boppana–Siegel) If D is a kernel-perfect orientation of

G and  (x) ≥ 1 + d+D(x) for all x ∈ V(G), then G is  -choosable.
Proof: Let n = | V(G)|. We use induction on n, with trivial basis n = 1. For n > 1,
consider a list assignment L with | L(x)| =  (x) for each x. Choose a color  in
some list. Let U = {v:  ∈ L(v)}. Let S be a kernel of the induced subdigraph
D[U], and let D = D − S. Assign color  to all of the independent set S.
Let L (v) = L(v) − {} and  (v) =  (v) − 1 for each v ∈ U − S, but L (v) = L(v)
and  (v) =  (v) for v ∈
/ U. The assignment L on V(D ) satisfies  (x) ≥ 1 + d+D (x),
since vertices of U − S that lost a color from their lists also lost a successor in S.
By the induction hypothesis, D is  -choosable. We complete an L-coloring for G
by combining an L -coloring of D with color  on S.

•
•
U−S S
•
•
V(D) − U

We apply Lemma 8.3.16 to orientations of line graphs of bipartite graphs.

364 Chapter 8: Coloring

8.3.17. Lemma. (Galvin [1995]) Let G be a bipartite multigraph with parts X

and Y . Let c be a proper edge-coloring of G by integers. For uv , vw ∈ E(G)
with c(uv) < c(vw), orient the edge in L(G) as uv → vw if v ∈ X and vw → uv
if v ∈ Y . The resulting orientation Dc of L(G) is kernel-perfect.
Proof: Note that multiedges in G become pairs of oppositely-directed edges in Dc .
It suffices to prove that the full digraph Dc has a kernel, since the same argument
applies to induced subdigraphs, which are obtained by discarding edges of G.
We use induction on | E(G)| , with trivial basis for E(G) = ∅. When E(G)
= ∅,
let S be the set of edges in G consisting of the highest-colored edge incident to
each vertex of X . If S is a matching in G, then S is independent in L(G) and is a
kernel in Dc , since every edge e ∈ E(G) − S shares an endpoint in X with some
edge e ∈ S, and c(e) > c(e ), so e is a successor of e in Dc .
If S is not a matching, then let e and e be edges in S having a common end-
point y ∈ Y ; by symmetry, we may assume c(e ) < c(e). Let G = G − e. By the
induction hypothesis, the orientation of L(G ) obtained by deleting vertex e from
Dc has a kernel S . We claim that S is also a kernel in Dc ; we need only show
that e has a successor in S (see figure on the left below).
Either e ∈ S , or e has a successor ê in S . By the choice of S, any successor
ê of e in Dc shares its endpoint y and has a color less than c(e ). Since c(e) > c(e ),
in either case e shares its endpoint y with a member of S having smaller color
than e, and hence e has a successor in S .

i−1 x
• • • • • • •
ê →← e
e 1 i ¾
←←
• • • • • • •
y y ¾−i

8.3.18. Theorem. (Galvin [1995]) l(G) = (G) for bipartite multigraphs G.

Proof: Let  = (G); we prove that G is -edge-choosable. Let c be a fixed proper
-edge-coloring of G using color set []. Define Dc as in Lemma 8.3.17.
If c(xy) = i, then in Dc the vertex xy has at most i − 1 successors sharing
endpoint y and at most  − i successors sharing endpoint x (see figure on the right
above). Hence d+Dc (e) ≤  − 1 for e ∈ V(Dc). Since Lemma 8.3.17 implies that Dc
is kernel-perfect, Lemma 8.3.16 implies that L(G) is -choosable.

Slivnik [1996] gave another proof. Borodin–Kostochka–Woodall [1997] then

strengthened Galvin’s result, showing that smaller lists suffice on edges whose
endpoints have small degree. Their corollary yielding the best upper bound on
(G) in terms of (G) does not follow directly from Galvin’s result.
l

8.3.19.* Lemma. Let G be an X , Y -bigraph G with | X | ≥ | Y | and (G) ≥ 1. For

some nonempty S ⊆ X , there is a matching with vertex set S ∪ N(S).
Proof: Since | X | ≥ | Y | , we may choose S to be a minimal nonempty subset of X
such that | S| ≥ | N(S)|. For x ∈ S, the choice of S yields
Section 8.3: Edge-Coloring and Perfection 365

|S| ≥ | N(S)| ≥ | N(S − x)| ≥ |S − x| + 1 = | S|.

Hence | S| = | N(S)|. Now | N(T)| > | T | for every nonempty proper subset T of S, by
the choice of S. Hence the subgraph induced by S ∪ N(S) satisfies Hall’s Condition
and has a perfect matching.

8.3.20.* Theorem. (Borodin–Kostochka–Woodall [1997]) In a bipartite multi-

graph G, if º (xy) = max{d(x) , d(y)} for xy ∈ E(G), then L(G) is º -choosable.
Proof: Let G be an X , Y-bigraph, named so that | X | ≥ | Y | after discarding iso-
lated vertices. By Lemmas 8.3.16–8.3.17, it suffices to produce a proper edge-
coloring c (using integers) such that Dc satisfies d+Dc (e) ≤ max{d G(x) , d G(y)} − 1
for each edge e, where e has endpoints x and y. We produce c by induction on
| E(G)| ; when G is just a matching, we use just one color.
For the induction step, let M be a matching (of S to N(S)) as given by Lemma
8.3.19. Let G = G − M , and let c be the edge-coloring of G guaranteed by the in-
duction hypothesis. Extend c to a proper edge-coloring c of G by giving one color
to M , larger than the colors in c . Edges not incident to S receive no additional
successors in moving from Dc to Dc , so they satisfy the desired bound on d+Dc (e).
Consider an edge incident to x ∈ S. If x ∈ M , then d+Dc (x ) = d G( ) − 1 ≤
max{d G(x) , d G( )} − 1. If xy ∈ E(G ), then restoring M increases the outde-
gree: d+Dc (xy) = d+Dc (xy) + 1, since M is a matching. Nevertheless, both x and
y have larger degree in G than in G , since M covers all of N(S). Thus d Dc (xy) ≤
max{d G (x) , d G (y)} − 1 yields d+Dc (xy) ≤ max{d G(x) , d G(y)} − 1, as desired.

• • •x • S

• •
y

Theorem 8.3.20 implies that the degree bound of Shannon [1949] for edge-
chromatic number holds also for edge-choosability. Since the bound is sharp for
edge-chromatic number (Example 8.3.3), it is also sharp here. This direct proof
of Shannon’s bound through list edge-coloring does not require proving Vizing ’s
Theorem or any of its extensions.

8.3.21.* Corollary. (Borodin–Kostochka–Woodall [1997]) If G is a multigraph,

then  l(G) ≤ 32 (G).
Proof: Let L assign lists of size ⌊ 32 (G)⌋ . Let H be a largest bipartite subgraph
of G, so d H (v) ≥ 12 d G(v) for v ∈ V(G) (Theorem 5.2.9). Let H = G − E(H). Since
(H ) ≤ 12 (G), we have (L(H )) ≤ (G) − 2. Since 32 (G) ≥ (G) − 1 when
(G) ≥ 2, there is an L-edge-coloring for H .
For xy ∈ L(H), form L (xy) by deleting from L(xy) the colors now on edges at
x and y in H . Thus | L (xy)| ≥ 23 (G) − d H (x) − d H (y). Since d H (y) ≤ 12 (G),
we obtain | L (xy)| ≥ (G) − d H (x) ≥ d H (x). Similarly, d H (x) ≤ 12 (G) yields
| L (xy)| ≥ d H (y). By Theorem 8.3.20, H has an L -edge-coloring that combines
with the edge-coloring of H to complete an L-coloring of G.
366 Chapter 8: Coloring

PERFECT GRAPHS

Two optimization problems (one maximization and one minimization) are

dual if every way to meet the constraints for the minimization yields value at
least as large as every way to meet the constraints for the maximization. Exam-
ples include minimum coloring and maximum clique, minimum edge-coloring and
maximum degree, minimum vertex cover and maximum matching, etc.
A min-max relation holds for dual integer optimization problems on a fam-
ily of instances when they have solutions of equal value on those instances. We
have proved such relations for coloring and cliques on interval graphs, for edge-
coloring and maximum degree on bipartite graphs, etc. Min-max relations are
among the most elegant and valuable results in combinatorics. Such a theorem
guarantees short proofs of optimality for both optimization problems. Exhibiting
a coloring and a clique of the same size proves that neither can be improved, and
a min-max relation states that such a coloring and clique can be found.
The general topics of duality and min-max relations belong to the subject
of “integer linear programming ”, which we do not study in this text. Instead,
we stay in graph theory and interpret several of the min-max relations we have
proved in the context of a family of graphs called “perfect graphs”.

8.3.22. Definition. A graph G is perfect if (H) = (H) for every induced sub-
graph H of G.

The induced-subgraph requirement prevents bad local behavior from being

hidden by adding a large clique. Claude Berge introduced perfect graphs in 1959;
the notion appears first in abstracts of his talks (Berge [1960, 1961]). The concept
arose from several families found to be perfect. Many such families are closed
under taking induced subgraphs.

8.3.23. Definition. A graph class G is hereditary if every induced subgraph of

a graph in G is also in G.

8.3.24. Remark. To prove that graphs in a hereditary class G are perfect, it

suffices to prove (G) = (G) for all G ∈ G, since the same argument applies to
induced subgraphs of G. Note that interval graphs form a hereditary family.

8.3.25. Definition. A hole or chordless cycle in a graph is an induced sub-

graph that is a cycle of length at least 4 (recall that a chord of a cycle C is an
edge not in C with endpoints in C). A chordal graph is a graph with no hole.

From the definition, the family of chordal graphs is hereditary (deleting a

vertex cannot create a chordless cycle). Using an interval representation, it is
easy to see that every interval graph is a chordal graph (see also Exercise 44).
Chordal graphs were first called triangulated graphs or rigid circuit
graphs. They helped suggest the notion of perfect graphs. Chordal graphs are
used in problems such as fast numerical inversion of large sparse symmetric ma-
trices (see Golumbic [1978], Andreae [1988]). They have many alternative char-
acterizations; we prove one used in many inductive proofs about them.
Section 8.3: Edge-Coloring and Perfection 367

8.3.26. Definition. A vertex is simplicial if its neighborhood is a clique. A sim-

plicial elimination ordering of an n-vertex graph G (also called perfect
elimination ordering) is a vertex ordering vn , . . . , v1 such that each vertex
vi is simplicial in the subgraph induced by {v1 , . . . , vi}.

8.3.27. Theorem. (Dirac [1961]) A graph G is chordal if and only if it has a

simplicial elimination ordering.
Proof: Sufficiency: If G has a chordless cycle C, then the first vertex deleted
from C in an elimination ordering has nonadjacent remaining neighbors. Hence
a graph with a chordless cycle has no simplicial elimination ordering.
Necessity: Since the family of chordal graphs is hereditary, it suffices to prove
by induction that every chordal graph has a simplicial vertex. To facilitate the in-
duction, we prove the stronger statement that if a chordal graph is not a complete
graph, then it has two nonadjacent simplicial vertices.
If G is not complete, then it has a minimal separating set S. By the mini-
mality of S, vertices u and v of S both have neighbors in any components A and
B of G − S. The union of shortest u , v-paths through A and through B is a cycle
of length at least 4 (as on the left below). By the choice of the paths and the ab-
sence of edges from A to B, this cycle has no chord other than uv. Since G has no
chordless cycle, uv ∈ E(G). Since u , v ∈ S were chosen arbitrarily, S is a clique.
Now let G 1 and G 2 be the S-lobes of G containing A and B (Definition 8.2.16),
as on the right below. Note that G i a chordal graph. If G i is complete, then any
vertex of G i outside S is simplicial. Otherwise, G i has two nonadjacent simplicial
vertices, and since S is a clique at least one of them is outside S. Since a vertex
outside S has no neighbors in G outside its S-lobe, vertices outside S that are
simplicial in their S-lobes are nonadjacent and simplicial in G.

u
•
A S B x1 • G 1 S G 2 • x2
•
v

An elimination ordering iteratively deletes vertices. Its reverse is a construc-

tion ordering: Every chordal graph can be grown from K 1 by iteratively adding a
new vertex whose neighborhood among the existing vertices is a clique.

8.3.28. Theorem. Chordal graphs are perfect.

Proof: (Berge [1960]) Since the chordal graphs form a hereditary family, we need
only prove that (G) = (G) when G is chordal.
We color the vertices of G in the reverse order to a simplicial elimination or-
dering. If vi has  earlier neighbors, then a color in [ + 1] is available to use on
vi . Since vi forms a clique with its earlier neighbors, there is a clique whose size
is at least the number of the color we use on vi . After coloring all the vertices,
(G) ≥ (G). Since always (G) ≥ (G), equality holds.
Chordal graphs are also the intersection graphs of families of subtrees of a
tree (Exercise 40). This again implies that all interval graphs are chordal graphs.
368 Chapter 8: Coloring

Many special families of perfect graphs are known. They may have a struc-
tural definition, a characterization by a decomposition or construction procedure,
a characterization using intersection graphs, fast algorithms for recognition and
for optimal coloring, etc. We present some of these families in the exercises.
Meanwhile, we return to the interpretation of our earlier min-max relations
in terms of perfection. In discussing independence and coloring, four optimiza-
tion parameters are natural. A short synonym for “independent set ” often used
in this context is stable set.

8.3.29. Definition. A clique covering of a graph G is a partition of V(G) into

cliques (a proper coloring of G is a partition of V(G) into stable sets). We
consider four parameters:
independence number (G) max size of a stable set
clique number (G) max size of a clique
chromatic number (G) min size of a coloring
clique covering number (G) min size of a clique covering

A clique and a stable set share at most one vertex, so always (G) ≥ (G) and
(G) ≥ (G); covering V(G) requires covering the biggest object of the dual type.
Since (G) = (G) and (G) = (G), the statement that (H) = (H) for every
induced subgraph H of G is just the statement that G is a perfect graph.

8.3.30. Example. Bipartite graphs and their line graphs. Bipartite graphs form
a hereditary class, and (G) = 2 = (G) for every nontrivial bipartite graph.
Hence bipartite graphs trivially are perfect.
Line graphs of bipartite graphs form another hereditary class (if G = L(G),
then every induced subgraph of G is the line graph of a subgraph of G ). A sta-
ble set in L(G) is a matching in G, and a clique in L(G) is a set of edges in G with
a common vertex (or a triangle in G). Hence (L(G)) = (G) (if G has no trian-
gles), and perfection for line graphs of bipartite graphs is just K önig ’s Theorem
(Theorem 8.3.7) that  (G) = (G) when G is bipartite.
Consider the K önig–Egerváry Theorem ( (G) = (G) when G is bipartite).
We have  (G) = (L(G)), and cliques in L(G) come from stars in G, so a vertex
cover in G provides a clique covering in L(G). Thus the K önig–Egerváry Theorem
says (L(G)) = (L(G)) when G is bipartite. This is another hereditary family, so
complements of line graphs of bipartite graphs are perfect.
Since (G) = 2, we have (G) =  (G) when G is bipartite with no isolated
vertex. Thus K önig ’s Other Theorem ((G) =  (G), Theorem 6.1.16) states that
complements of bipartite graphs are perfect.

Bipartite graphs and their line graphs are very special, but these and other
families suggest that the complement of every perfect graph is also perfect. Berge
conjectured this around 1960; many tried to prove it. Lovász [1972a,b] found the
first proof at the age of 22 (Berge–Ramírez-Alfonsín [2001] presents the history).
It is now known as the “Perfect Graph Theorem” (PGT). Fulkerson reduced it to
a claim he thought was too strong to be true. When Berge told him that Lovász
had proved the theorem, within hours he proved the claim, illustrating that a
theorem becomes easier to prove when known to be true.
Section 8.3: Edge-Coloring and Perfection 369

8.3.31. Remark. We proved perfection for each family in Example 8.3.30, but
the PGT makes this unnecessary. By the PGT, K önig ’s Other Theorem follows
from observing that bipartite graphs are perfect. Similarly, the K önig–Egerváry
Theorem and K önig ’s Theorem (bipartite graphs are Class 1) imply each other.

Lovász proved the PGT twice. Since (H) ≥ | V(H)|

(H) , perfection of G
(or G) re-
quires (H)(H) ≥ | V(H)| for each induced subgraph H. Lovász [1972b] proved
the converse, making perfection of G and G both equivalent to this and hence
to each other. Later, Gasparyan [1996] captured the idea in a short proof using
linear algebra, which we present. (See Perz–Rolewicz [1990] for another proof.)

8.3.32. Definition. A minimal imperfect or p-critical graph is an imperfect

graph whose proper induced subgraphs are all perfect.

8.3.33. Lemma. If S is a stable set in a p-critical graph G, then (G) = (G − S).
Proof: Since G is p-critical, (G) > (G) and (G − S) = (G − S). Since G can
be colored by adding S as one class to an optimal coloring of G − S,
(G − S) + 1 ≥ (G) > (G) ≥ (G − S) = (G − S).
The two ends of the display differ by 1, so (G) − 1 = (G) = (G − S).

8.3.34. Theorem. (Lovász [1972b]) A graph G is perfect if and only if the in-
equality (H)(H) ≥ | V(H)| holds for every induced subgraph H.
Proof: (Gasparyan [1996]) Since (H) ≥ | V(H) |
(H) , having
(H) = (H) requires
(H)(H) ≥ | V(H)|. Since imperfect graphs have p-critical induced subgraphs,
for the converse it suffices to show (G)(G) < n when G is p-critical with n ver-
tices. Let a = (G) and w = (G).
Let S0 be a maximum stable set in G, with S0 = {x1 , . . . , x a}. By Lemma
8.3.33, each G − xr is w-colorable. Let {S(r−1)w+1 , . . . , Srw} be the stable sets in an
optimal coloring of G − xr . For 0 ≤ j ≤ aw, Lemma 8.3.33 yields (G − Sj) = w,
so a maximum clique Qj in G − Sj is a w-clique in G that avoids Sj .
We claim that Qj intersects Si when i
= j . The sets S(r−1)w+1 , . . . , Srw parti-
tion V(G − xr) into w stable sets, and Qj contains at most one vertex of each (since
Q j is a clique). If xr ∈/ Qj , then Qj must have exactly one vertex from each. If
xr ∈ Qj , then Qj has only w − 1 vertices other than xr and hence misses exactly
one of the sets in the coloring of G − xr . Since Qj can only have one vertex in S0 ,
at most one stable set in the entire list is missed by Qj . On the other hand, we
know that Qj misses Sj , since Qj was chosen as a maximum clique in G − Sj .
Let A and B be the n-by-(aw + 1) incidence matrices for the set families
{S0 , . . . , Saw} and {Q0 , . . . , Qaw}, respectively. Since | Si ∩ Qj | = 1 when i
= j ,
and | Sj ∩ Qj | = 0, we obtain A T B = J − I, where J is the all-1 square matrix of
order aw + 1. Since J − I is nonsingular, both A and B must have rank at least
aw + 1, and hence n ≥ aw + 1. We conclude (G)(G) < n.

8.3.35. Corollary. (Perfect Graph Theorem (PGT) ; Lovász [1972a]) A graph

G is perfect if and only if G is perfect.
Proof: The condition in Theorem 8.3.34 is the same for both G and G.
370 Chapter 8: Coloring

This leads to the question of which graphs are p-critical.

8.3.36. Example. Odd cycles of length at least 5 are imperfect, since (C2¾+1) =
3 > 2 = (C2¾+1). They are p-critical, since proper induced subgraphs are bipar-
tite. By the PGT, C2¾+1 is also p-critical when  ≥ 2.
Berge [1963] conjectured both the Perfect Graph Theorem and the Strong
Perfect Graph Theorem (SPGT): The odd cycles and their complements are the
only minimal imperfect graphs. Since the condition is self-complementary, the
SPGT implies the PGT.
The Strong Perfect Graph Conjecture remained open for more than 40 years.
It was proved in Chudnovsky–Robertson–Seymour–Thomas [2006] (announced in
2002). The paper is 178 pages long; we will not attempt to describe the proof.

Many survey articles and at least four books are devoted to perfect graphs
and related topics. Golumbic [1980, 2003] emphasizes algorithmic aspects and
special classes. Berge–Chvátal [1984] collects fundamental early papers that de-
veloped the theory of perfect graphs. Brandst ädt–Le–Spinrad [1999] provides a
thorough catalogue of properties of and relationships among nearly 200 classes
of perfect graphs. Ramírez-Alfonsín–Reed [2001] discusses classical and more
recent aspects, emphasizing the structure of perfect graphs. More recent dis-
cussions that emphasize the proof of the SPGT include Roussel–Rusu–Thuillier
[2009] and Trotignon [2015].

EXERCISES 8.3

8.3.1. (−) Prove that L(K r,s) = K r K s and that L(K 5) is the Petersen graph.
8.3.2. (−) Prove that if L(G) is connected and regular, then G is regular or is bipartite
with equal degree at vertices of the same part. (Ray-Chaudhuri [1967])
8.3.3. (−) Give an explicit edge-coloring to prove that K r,s is Class 1.
8.3.4. (−) Prove that the smallest graph with maximum degree 3 and edge-chromatic
number 4 has five vertices.
8.3.5. (−) Determine whether it is possible for a graph G to have more than (G) pairwise
disjoint maximum matchings. Determine whether it is possible to have more than (G)
pairwise disjoint maximal matchings.
8.3.6. (−) Let D be a directed multigraph (loops allowed) such that each indegree and each
outdegree is at most d. Prove that D has an edge-coloring with d colors such that edges
with a common head or a common tail have distinct colors.
8.3.7. (−) Let G be a Class 1 regular graph of even degree. Prove that | E(G)| is even.
8.3.8. (−) Given integers r and d with 0 ≤ r ≤ d, use Vizing ’s Theorem to prove that
every d-regular graph has a spanning subgraph in which every vertex has degree r or r + 1.
(Comment: This is generalized in Tutte [1978].)
8.3.9. (−) Compute (G) and (G) for the complement of an odd cycle.
8.3.10. (−) Show that an n-vertex chordal graph has at most n maximal cliques, with
equality if and only if it has no edges. (Fulkerson–Gross [1965])
Exercises for Section 8.3 371

8.3.11. (−) Clique identification preserves perfection. Prove that G ∪ H is perfect when G
and H are perfect graphs and G ∩ H is a complete graph.
8.3.12. Let G be a graph.
(a) Prove that the number of edges in the line graph L(G) is ∑v∈V(G) (d(v)
2
).
(b) Prove that G is isomorphic to L(G) if and only if G is 2-regular.
(c) Determine the graphs G such that | E(L(G))| < | E(G)| .

8.3.13. (♦) Prove that every regular graph having a cut-vertex is Class 2.
8.3.14. Use Brooks’ Theorem to prove that every graph with maximum degree 3 is 4-edge-
colorable. (This is a special case of Vizing ’s Theorem; do not use Vizing ’s Theorem.)
8.3.15. (♦) Prove that a 3-regular graph is 3-edge-colorable if and only if it is the union of
two even subgraphs.
8.3.16. Prove that every bipartite graph G has a (G)-regular bipartite supergraph (no
multiedges).
8.3.17. (♦) Greedy edge-coloring of a graph G produces a proper edge-coloring using at

most 2 (G) − 1 colors. For ¾ ∈ , construct a tree with maximum degree ¾ and an ordering
of its edges such that greedy edge-coloring uses 2 ¾ − 1 colors.
8.3.18. Use induction on | E(G)| to prove K önig ’s Theorem that Ò (G) = (G) for every
bipartite multigraph G. (K önig [1916])
8.3.19. The d-dimensional integer simplex of length m is the graph T md whose vertices
are the nonnegative integer (d + 1)-tuples summing to m, adjacent when they differ by 1
in two places and agree elsewhere. Compute Ò (T md ) for m > d. (Ma–West [2013])
8.3.20. Density Conditions for Class 2. Prove that if G is a graph with 2 ¾ + 1 vertices,

where ¾ ∈ , then Ò (G) > (G) under each condition below (with r ≥ 2).
(a) G has more than ¾ · (G) edges.
(b) G arises from an r-regular graph by deleting fewer than r/2 edges.
(c) G arises from an r-regular graph with 2 ¾ vertices by subdividing one edge.

8.3.21. Prove that the graph obtained by deleting one edge from the Petersen graph is
Class 2 but has no overfull subgraph with maximum degree 3.
8.3.22. (♦) Overfull Conjecture ⇒ 1-Factorization Conjecture (Remark 8.3.8).
(a) Prove that in a regular graph of even order, an induced subgraph is overfull if and
only if the subgraph induced by the other vertices is overfull.
(b) Let G be a ¾-regular graph of order 2m having an overfull subgraph. Prove ¾ < m
for odd m and ¾ < m − 1 for even m. Conclude that the Overfull Conjecture implies the
1-Factorization Conjecture.
8.3.23. Let G be the (m − 1)-regular connected graph formed from 2K m by applying a 2-
switch (Definition 5.2.7). Prove that G has no 1-factorization if m is odd and greater than
3. (Comment: This shows that the 1-Factorization Conjecture in Remark 8.3.8 is sharp.)

• •

• •
Km − e Km − e

8.3.24. (♦) For nontrivial graphs G and H , prove Ò (G H) = (G H) when H satisfies

Ò (H) = (H). For example, Ò (K3 K2) = 3. (Mahmoodian [1981])
372 Chapter 8: Coloring

8.3.25. Kotzig’s Theorem for cartesian product of graphs.

(a) Use Vizing ’s Theorem to prove (G K 2) = (G K 2).
(b) Let G1 and G 2 be edge-disjoint graphs on V , and let H1 and H2 be edge-disjoint
graphs on W . Prove (G1 ∪ G 2) (H1 ∪ H2) = (G1 H2) ∪ (G 2 H1).
(c) Use parts (a) and (b) to prove (G H) = (G H) when both G and H have 1-
factors. (Kotzig [1979], George [1991])

8.3.26. Shannon’s Theorem. The edge-coloring theorem of Shannon [1949] states that
(G) ≤ 3 (G) for every loopless multigraph G.
2
(a) Prove Shannon’s Theorem from Vizing ’s Theorem (G) ≤ (G) + (G).
(b) Prove Shannon’s Theorem without Vizing ’s Theorem by showing that with 32 (G)
colors available, a partial edge-coloring can be augmented.
(c) Prove the same bound for even (G) from Petersen’s 2-Factor Theorem.

8.3.27. (♦) Equitable edge-colorings.

(a) Prove that every graph G with maximum degree  − 1 has a proper -edge-coloring
with each color used ⌈ | E(G)| /⌉ or ⌊ | E(G)| / ⌋ times. (de Werra [1971], McDiarmid [1972])
(b) Prove that every bipartite graph with maximum degree at least  has a -edge-
coloring in which at each vertex v, each color appears ⌈ d(v)/⌉ or ⌊ d(v)/⌋ times. (Com-
ment: In particular, bipartite graphs are Class 1. Gupta [1966] notes the case  = (G),
which yields a (G)-edge-coloring such that every vertex has incident edges of all colors.)

8.3.28. (+) Given an edge-coloring of a graph G, let c(v) be the number of distinct colors
on edges at v. An optimal coloring maximizes ∑v∈V(G) c(v) among -edge-colorings of G.
(a) Suppose that no component of G is an odd cycle. Prove that G has a 2-edge-coloring
in which both colors appear at each vertex of degree at least 2.
(b) Let  be an optimal -edge-coloring of G in which color a appears at least twice at
u ∈ V(G) and color b does not appear at u. Let H be the subgraph of G consisting of edges
colored a or b. Prove that the component of H containing u is an odd cycle.
(c) Use part (b) to prove Vizing ’s Theorem for graphs. (Fournier [1973])

8.3.29. Let G ¾ be the graph consisting of three “parallel” -cycles {xi }, {yi }, { i } and 
additional vertices {wi } such that N(wi) = {xi , yi , i } for each i. Obtain H¾ from G ¾ by
deleting x¾ x1 and y¾ y1 and adding x¾ y1 and y¾ x1 . (Comment: The graphs {H2 j + : j ≥ 2}
are called flower snarks. A snark is a 2-edge-connected 3-regular Class 2 graph that has
girth at least 5 and has no edge cut of size 3 that does not isolate a vertex.)
(a) Prove that G ¾ is Class 1.
(b) Prove that H¾ is Class 2 if  is odd. (Isaacs [1975])

• •
•
•
• •
• •• •• •

• •
•• ••
• •

8.3.30. Let M be a maximal -edge-colorable subgraph of a multigraph G. For x ∈ V(G),

let (x) be the maximum multiplicity of an edge incident to x. Let S = {x ∈ V(G): d M (x) ≤
 − (x)}. Puleo [2017] proved that if x ∈ S, then d G[S](x) ≤ d M (x). Prove that Puleo’s
result implies Ore’s strengthening of Vizing ’s Theorem: (G) ≤ maxv∈V(G)(d G(v) + (v)).
Exercises for Section 8.3 373

8.3.31. The generalized Petersen graph P(n , ¾) is the graph with vertices {u1 , . . . , un}
and {v1 , . . . , vn} and edges {ui ui+1 }, {ui vi }, and {vi vi+¾ }, where addition is modulo n. The
usual Petersen graph is P(5 , 2), with = 4.
(a) Prove that the subgraph of P(n , 2) induced by  consecutive pairs {ui , vi } has a
spanning cycle if  ≡ 1 (mod 3) and  ≥ 4. (Comment: The figure below is a proper sub-
graph of P(n , 2) for n ≥ 10.)
(b) Use part (a) to prove (P(n , 2)) = 3 for n ≥ 6.

v1 v10
• • • • • • • • • •
• • • • • • • • • •
u1 u10

8.3.32. (♦) Given an edge-coloring of a graph G, for v ∈ V(G) let S(v) be the set of colors
used on edges incident to v. Let ¾(G) be the minimum number of colors in a proper edge-
coloring of G such that | S(x) ∩ S(y)| ≤  whenever xy ∈ E(G). Prove ¾(K r,s) ≥ ⌈ rs/ ⌉ .
(Comment: Borozan et al. [2015] proved ¾ (K r,s) = ⌈ rs/ ⌉ , plus 2(G) ≤ 6 for (G) ≤ 3 and
bounds for d-degenerate and complete graphs. Strong edge-coloring is the case  = 1;
Erdős and Nešet řil conjectured 1 (G) ≤ 45 (G)2 , improving to 14 (5 (G)2 − 2 (G) + 1) when
(G) is odd, both achieved by expanding vertices of C5 into independent sets. The sharp
bound of 10 for (G) = 3 was proved in Andersen [1992] and Horák–He–Trotter [1993];
Cranston [2006] proved 1 (G) ≤ 22 when (G) = 4. See also Molloy–Reed [1997].)
8.3.33. A total coloring of G assigns a color to each vertex and each edge so that objects
adjacent or incident have different colors. Let (G) denote the number of colors needed.
(a) Prove (G) ≤ l(G) + 2 for every graph G.
(b) Prove (G) ≤ (G) + 2 when G is bipartite. (Rosenfeld [1971])
(c) Given the List Color Conjecture, prove (G) ≤ (G) + 3. (Bollobás–Harris [1985])
(Comment: The Total Coloring Conjecture is (G) ≤ (G)+2 for all G (Behzad [1965]).
It holds for (G) ≤ 3 (Rosenfeld [1971], Vijayaditya [1971]), (G) = 4 (Kostochka [1977]),
(G) = 5 (Kostochka [1996]), and complete multipartite graphs (Yap [1989]); see Exercise
13.3.5 for (K n). Hind–Molloy–Reed [1999] proved (G) ≤ (G) + 8 log8 (G) when (G)
is large. Geetha–Narayanan–Somasundaram [2018] surveys the topic.)
8.3.34. (♦) Prove the Total Coloring Conjecture (above) for 4-colorable Class 1 graphs.
8.3.35. (+) An interval coloring of G is a proper edge-coloring of G by integers such that
at each vertex, the incident colors form an interval of integers.
(a) Let G be an X , Y -bigraph in which every vertex of X has degree 2 and every ver-

tex of Y has degree 2  for some  ∈ . Prove that G has an interval coloring. (Comment:
The statement also holds when vertices of Y have degree 2  + 1, using the result that a
(2  + 1)-regular graph with at most  cut-edges has a 2-factor.) (Hansen [1992])
(b) Prove that K r,s has an interval coloring using r + s − gcd(r, s) colors. (Hint: Com-
bine solutions for the cases r = s and gcd(r, s) = 1. Comment: In fact , fewer colors cannot
suffice, but this is harder.) (Hanson–Loten–Toft [1998])
8.3.36. (♦) A parity edge-coloring of a graph G is an edge-coloring such that on every
path, some color appears an odd number of times. Prove that G is a subgraph of the hyper-
cube Q¾ if and only if G has a parity -edge-coloring such that on every cycle, every color
appears an even number of times. (Havel–Morávek [1972])
8.3.37. With parity edge-coloring defined as above, let p(G) denote the minimum number
of colors in a parity edge-coloring of G. Prove p(K 2 ,3) = 4 and p(K 5) = 7.
8.3.38. (♦) Give fast algorithms for computing (G), (G), (G), and (G) when supplied
with a simplicial elimination ordering of G.
374 Chapter 8: Coloring

8.3.39. (♦) Prove that l(G) = (G) when G is a chordal graph.

8.3.40. (♦) Intersection representation of chordal graphs.
(a) Prove that pairwise intersecting subtrees of a tree have a common vertex.
(b) Prove that a graph G is chordal if and only if it is an intersection graph of some
subtrees of a tree. (Walter [1972, 1978], Gavril [1974], Buneman [1974])
(c) Conclude that a non-complete chordal graph has nonadjacent simplicial vertices.
8.3.41. Show that every intersection graph of subtrees of a tree is an intersection graph
of subtrees of a tree with maximum degree 3.
8.3.42. A clique tree of a graph G is a tree T whose vertices are the maximal cliques of
G, such that for v ∈ V(G) the cliques containing v induce a subtree of T .
(a) Prove that every tree with fewest vertices in which G has a subtree representa-
tion (Exercise 8.3.40) is a clique tree of G.
(b) Let M(G) be the weighted graph whose vertex set is the set {Q1 , . . . , Q¾ } of max-
imal cliques in an n-vertex graph G, with weight | Qi ∩ Qj | on the edge Qi Qj . Prove that if
T is a spanning tree of M(G), then w(T) ≤ ∑ | Qi | − n, with equality if and only if T is a
clique tree of G. (Acharya–Las Vergnas [1982], McKee [1993])
8.3.43. (♦) A graph G is a comparability graph (see Chapter 12) if and only if it has a
transitive orientation, which is an orientation F of G such that xy , y  ∈ E(F) implies
x  ∈ E(F). Prove that comparability graphs are perfect.
8.3.44. (+) This exercise proves that a graph G is an interval graph if and only if G is a
chordal graph and G is a comparability graph. (Gilmore–Hoffman [1964])
(a) Prove that the condition is necessary.
For sufficiency, henceforth let G be a chordal graph such that G is a comparability graph.
Let Q1 , . . . , Q¾ be the maximal cliques in G.
(b) Let F be a transitive orientation of G. Prove that all edges of G joining Qi and Qj
point in the same direction under F .
(c) Define a tournament T on q1 , . . . , q¾ by putting qi qj ∈ E(T) in T when all edges of
F joining Qi and Qj point from Qi to Qj . Prove that T is a transitive tournament.
(d) Prove that Q1 , . . . , Q¾ can be linearly ordered so that the cliques containing any
one vertex form a consecutive portion of the list.
(e) Prove that G has an interval representation.
8.3.45. Prove that a connected interval graph G that is not complete has an interval rep-
resentation whose interval starting leftmost represents a vertex with degree less than
(G). (Comment: Thus G cannot be regular.) (Kostochka–Pelsmajer–West [2003])
8.3.46. Prove that G is an interval graph if and only if G has a vertex ordering v1 , . . . , vn
such that vi v¾ ∈ E(G) guarantees vj v¾ ∈ E(G) whenever i < j <  . (Ramalingam–
PanduRangan [1988], Jacobson–McMorris–Mulder [1991])
8.3.47. Let Q be a maximal clique in a chordal graph G. Prove that if Q contains no
simplicial vertex of G, then G − Q is disconnected. (Voloshin–Gorgos [1982])
8.3.48. The chromatic polynomial  G () (Application 4.1.11) counts the proper colorings
 : V(G) → []. Prove that G is chordal if and only if for every induced subgraph H ,  H ()
has the form  ∏i( − j i) with all j i ∈  (Voloshin [2002]). (Comment: Subdividing one
edge of K6 yields a non-chordal G such that  G so factors (Read [1975], Dmitriev [1980]).)
8.3.49. Let v be a vertex in a chordal graph G. From the definition (no chordless cycle),
prove that there is a simplicial vertex of G among the vertices having maximum distance
from v in G. (Voloshin [1982], Farber–Jamison [1986])
(G)
8.3.50. Prove ∑r=1 (−1)r−1 r = 1 when r counts the r-cliques in a connected chordal G.
Exercises for Section 8.3 375

8.3.51. (♦) An x , y-separator in a graph G is a set S of vertices such that x and y lie in
distinct components of G − S. A minimal vertex separator is a minimal x , y-separator
for some pair {x , y}. Every minimal separating set is a minimal vertex separator. In the
graph below, S is a minimal vertex separator (for {x , y}) but is not a minimal separating
set. Prove that a graph is chordal if and only if every minimal vertex separator is a clique.
•
•
x• •y
•
S

8.3.52. A -tree is a graph grown from K ¾ by iteratively adding a vertex whose neigh-
borhood is a -clique. Prove that a connected graph G is a -tree if and only if (1) G has a
-clique but no ( + 2)-clique and (2) Every minimal vertex separator of G is a -clique.
8.3.53. Let G be an n-vertex chordal graph with clique number r. Prove that G has at
most ( rj ) + ( rj −−11 )(n − r) cliques of size j , with equality only when G is an (r − 1)-tree.
8.3.54. A Meyniel graph is a graph in which every odd cycle has at least two chords;
such graphs are perfect (Meyniel [1976]).
(a) Prove that every chordal graph is a Meyniel graph.
(b) Prove that every P4 -free graph is a Meyniel graph.
(c) Find a comparability graph that is not a Meyniel graph.
8.3.55. (♦) Let G be a P4 -free graph, meaning that no induced subgraph of G is a 4-vertex
path. Prove that every maximal independent set in G intersects every maximal clique in
G. Use this to prove that G is perfect.
8.3.56. Let G be a nontrivial connected P4 -free graph. Prove that if Q is a maximal clique
in G, then Q contains a neighbor of every vertex of G. Conclude that every connected
triangle-free P4 -free graph is a complete bipartite graph.
8.3.57. Let G be a P4 -free graph. Prove that for every vertex ordering  , the greedy
coloring algorithm with respect to  uses (G) colors (such graphs are called perfectly or-
derable). (Hint: When greedy coloring with respect to  uses  colors, let i be the least
integer such that G has a clique whose vertices receive colors i through  . Prove i = 1.)
8.3.58. (♦) A complement reducible graph (or cograph), is a graph reducible to a triv-
ial graph by successive complementation within components.
(a) Prove that a graph G is P4 -free if and only if it is a cograph.
(b) Use (a) and the PGT to prove P4 -free graphs are perfect. (Seinsche [1974])
(c) Prove that if G is P4 -free, then G and G are comparability graphs.
8.3.59. (♦) (Star-Cutset Lemma) A star-cutset of a graph G is a separating set S con-
taining a vertex x adjacent to all of S − {x}.
(a) Prove that if G has no stable set intersecting every maximum clique, and every
proper induced subgraph of G is (G)-colorable, then G has no star-cutset.
(b) Prove that no p-critical graph has a star-cutset. (Chvátal [1985b])
8.3.60. Find an imperfect graph G having a star-cutset S such that the S-lobes of G are
perfect graphs. (Comment: Thus identification at star-cutsets does not preserve perfec-
tion, although no p-critical graph has a star-cutset.) (T. Shermer)
8.3.61. A graph G is weakly chordal if neither G nor G has an induced cycle of length
at least 5. Prove that every chordal graph is weakly chordal. (Comment: Hayward [1985]
proved that if G and G are non-complete weakly chordal graphs, then G or G has a star-
cutset. Thus weakly chordal graphs are perfect , by the Star-Cutset Lemma.)
376 Chapter 8: Coloring

8.3.62. (♦) (Substitution Lemma) If V(G) = {v1 , . . . , vn}, and H1 , . . . , Hn are pairwise
disjoint graphs, then the composition G[H1 , . . . , Hn] is the graph H1 + · · · + Hn plus the
edges {xy: x ∈ V(Hi), y ∈ V(Hj ), vi vj ∈ E(G)}. Prove that every composition of perfect
graphs is perfect. (Hint: Use the Star-Cutset Lemma, Exercise 8.3.59) (Lovász [1972a])
v1 • •
•
• •
v2 • • v4 •
• • •
v3 •••

8.3.63. Vertex multiplication is the special case G[K h1 , . . . , K hn ] of composition (Exer-

cise 8.3.62), denoted G ◦ h (the ith vertex of G “multiplies” into an independent hi-set).
Without using the PGT or Substitution Lemma, prove directly that vertex multiplication
preserves perfection of G and perfection of G.
8.3.64. Proof of the PGT by vertex multiplication. The goal: prove that G is perfect when
G is perfect , using vertex multiplication (Exercise 8.3.63). A minimal counterexample G
must be p-critical. Lemma 8.3.33 then implies that every maximal stable set S in G misses
some maximum clique Q(S). Let S1 , . . . , S¾ be the list of maximal stable sets in G. Let
hj = |{i: vj ∈ Q(Si)}|. Let H = G ◦ h.
(a) By counting vertices of H , prove (H) ≥  .
(b) By counting contributions from cliques, prove (H) <  .
(c) Obtain a contradiction, using Exercise 8.3.63. (Lovász [1972a])
8.3.65. Let G be a cartesian product of complete graphs. Prove (G) = (G). Prove that
K 2 K 2 K3 is not perfect.
8.3.66. A graph G is a split graph if V(G) is the union of a clique and an independent set.
(a) Prove that if G is a split graph, then G and G are chordal.
(b) Prove that if G and G are chordal, then G has no induced C4 , 2K 2 , or C5 .
(c) Prove that a graph having no induced C4 , 2K 2 , or C5 is a split graph.
8.3.67. Let d1 , . . . , d n be the vertex degrees in a graph G, in nonincreasing order. Let m
be the largest index  such that d ¾ ≥  − 1. Prove that G is a split graph if and only if
∑i=1 di = m(m − 1) + ∑i=m+1 min{m , di} (Erdős–Gallai; see Exercise 6.2.43).
m n

8.3.68. (♦) A graph G is a threshold graph if there is a nonnegative weight function

w and a threshold t such that S ⊆ V is a stable set if and only if ∑ x∈S w(x) ≤ t; this is
a threshold weighting. An edge-threshold weighting imposes the condition only for
| S| = 2. Prove A ⇒ B ⇒ C ⇒ D ⇒ E ⇒ F ⇒ A:
(A) G has a threshold weighting.
(B) G has an edge-threshold weighting.
(C) G has no x1 , x2 , x3 , x4 such that x1 x2 , x3 x4 ∈ E(G) and x2 x3 , x4 x1 ∈ E(G).
(D) G has no 4-vertex induced subgraph isomorphic to C4 , 2K 2 , or P4 .
(E) G is a P4 -free split graph.
(F) G arises from K1 by successively adding vertices made adjacent either to every
vertex already present or to no vertex already present.
8.3.69. A graph G is strongly perfect if every induced subgraph H has a stable set that
intersects all maximal cliques of H . Note that every bipartite graph is strongly perfect.
(a) Prove that every strongly perfect graph is perfect (justifying the name).
(b) Construct an example to show that not every perfect graph is strongly perfect.
Chapter 9

Planar Graphs
Topological graph theory studies layouts of graphs on surfaces. In the plane
this was stimulated by the famous Four Color Problem, asking whether the (con-
nected) regions of any map on the globe can be colored from four colors so that
regions sharing a nontrivial boundary have different colors. More recent motiva-
tion comes from circuit layouts on silicon chips. Crossing wires cause problems
in layouts, so we want to know which circuits can be laid out without crossings.
Many general texts treat planar graphs and topological graph theory at some
length, such as Chartrand–Lesniak [1986], Diestel [1997], and Gross–Yellen
[1999] (and more recent editions of all three). Gross–Tucker [1987] and Mohar–
Thomassen [2001] are devoted completely to topological graph theory.

9.1. Embeddings and Euler ’s Formula

A famous puzzle asks for noncrossing paths to enable three hermits to access
three utilities. This is equivalent to embedding K 3 ,3 in the plane. Postponing
definitions, we first explain informally why it cannot be done.

9.1.1. Proposition. K 5 and K 3 ,3 are not planar graphs.

Proof: Consider a drawing of K 5 or K 3 ,3 in the plane. Let C be a spanning cycle.
If the drawing does not have crossing edges, then C is drawn as a closed curve.
Chords of C must be drawn inside or outside this curve. Two chords conflict if
their endpoints on C occur in alternating order. When two chords conflict, we
can draw only one inside C and one outside C.
A 6-cycle in K 3 ,3 has three pairwise conflicting chords. We can draw at most
one inside and one outside, so the embedding cannot be completed. When C is a
5-cycle in K 5 , at most two chords can go inside or outside. There are five chords,
so again the embedding cannot be completed. Hence neither graph is planar.
•
• •
• •
• •

• • • •
377
378 Chapter 9: Planar Graphs

DRAWINGS AND DUALS

We use standard concepts about curves and regions. We treat the formal def-
initions lightly, since the terms mean what one expects them to mean.

9.1.2. Definition. A curve in the plane is the image of a continuous function

from the interval [0 , 1] into 2 . An open set is a set U ⊆ 2 such that for
p ∈ U , all points within some small distance from p belong to U. A region
is an open set U such that every two points in U lie on a curve contained in
U. A curve in the plane is closed if its first and last points are the same; it
is simple if it does not otherwise intersect itself.

Many aspects of planarity need the generality of multigraphs. A drawing

of a multigraph on a surface maps the vertices into points and the edges into
curves, preserving the incidence relation. Edges now have a physical geometric
meaning. Since incidence and adjacency are unchanged, we can view these points
and curves as the vertices and edges themselves.
By moving edges slightly, we may assume that no three edges share an in-
ternal point, that no edge has any vertex as an internal point, and that no two
edges are “tangent ”. We may also assume that no two incident edges cross and
that no edge crosses itself, since the change shown below eliminates the crossing.
We consider only drawings satisfying these properties.

• •
• •
• •
9.1.3. Definition. In a drawing of a multigraph, a crossing of two edges is a com-
mon internal point. An embedding is a drawing without crossings. A multi-
graph that embeds in the plane is planar. A particular planar embedding
of a multigraph is a plane multigraph. A subgraph of a plane multigraph
inherits its embedding. The faces of a plane multigraph are the maximal
regions disjoint from the edges.

Every finite plane multigraph has one unbounded face. Computation with
faces uses the Jordan Curve Theorem: a simple closed curve cuts the plane into
two maximal regions (“inside” and “outside”). In topology this is a deep notion.
For finite graphs we may restrict drawings to use only polygonal curves (unions
of finitely many line segments). In this model the proof is not difficult, and it
provides a test for whether a point is inside or outside.

9.1.4. Theorem. (Restricted Jordan Curve Theorem) A simple closed polygo-

nal curve C partitions the plane into two faces, each with boundary C.
Proof*: (see Tverberg [1980]) Because C has finitely many segments, noninter-
secting segments on C cannot be arbitrarily close. Thus as we follow C, the nearby
points on one side all lie in the same face, and similarly for the points on the other
side. Every point not on C lies in the same face with at least one of these two sets.
Section 9.1: Embeddings and Euler ’s Formula 379

To distinguish the two sets, consider rays in the plane. For a ray Ô from a
point p not on C, let º (Ô) be the number of segments of C that Ô intersects. As
the direction changes, º (Ô) changes only when Ô visits an endpoint of a segment
of C, but the parity is the same before and after that direction. Let the parity of
p be the parity of º (Ô) for the rays not containing endpoints of segments in C.

Points x and y in the same face of C are joined by a polygonal x , y-curve P

avoiding C. A segment of P lies along a ray. Since the segment does not intersect
C, its endpoints have the same parity. Hence any two points in the same face
have the same parity. However, the endpoints of a short segment crossing C have
opposite parity. Hence there are two faces. The even points and odd points form
the outside face and the inside face, respectively.

A map on the plane or the sphere can be viewed as a plane multigraph. The
“boundary-sharing ” relation on the faces yields a natural “dual” multigraph.

9.1.5. Definition. The dual G ∗ of a plane multigraph G is a plane multigraph

whose vertices correspond to the faces of G and whose edges correspond to
the edges of G. The edge e∗ ∈ E(G ∗) that corresponds to e ∈ E(G) joins the
vertices corresponding to the faces of G whose boundary contains e.
A canonical embedding of G ∗ puts the dual vertex x for each face X of G
inside X . For each appearance of an edge e in the boundary of X , put a curve
within X from x to the midpoint of e; these curves are noncrossing. Each
such curve meets another from the other side of e to form the dual edge e∗ .

By construction, the dual G ∗ is a connected plane multigraph, each face of

G contains one vertex of G ∗ , and each edge of G ∗ is embedded to cross the corre-
sponding edge of G and no other. With this definition, (G ∗)∗ is isomorphic to G if
and only if G is connected (Exercise 15); this motivates using the word “dual”.
A plane graph may have loops and multiedges in its dual. A cut-edge of G
becomes a loop in G ∗ ; the faces on both sides of it are the same. Multiedges arise
in the dual when distinct faces of G have more than one common boundary edge.

9.1.6. Example. Below in bold we show a plane graph with four vertices, four
edges, and two faces. The two faces have three common boundary edges, and the
cut-edge has the same face on both sides. Hence the dual has a triple-edge and a
loop. We show it in solid edges, with four faces, four edges, and two vertices.

•
• • • •
•
380 Chapter 9: Planar Graphs

Every planar embedding of K 4 has four faces, and every two of them have a
common boundary edge. The dual is another copy of K 4 .
Every planar embedding of the cube Q3 has eight vertices, 12 edges, and six
faces. “Opposite” faces have no common boundary; the dual is an embedding of
K 2 ,2 ,2 , with six vertices, 12 edges, and eight faces.

9.1.7. Example. Two embeddings of a planar graph may have nonisomorphic du-
als. Each embedding shown below has three faces, so in each case the dual has
three vertices. In the embedding on the right, the dual vertex corresponding to
the unbounded face has degree 4. In the embedding on the left, no dual vertex
has degree 4, so the duals are not isomorphic.
In contrast, every 3-connected planar graph G has essentially only one em-
bedding; the duals of any two drawings of G are isomorphic. This follows from the
2-Isomorphism Theorem of Whitney [1933], which describes when planar graphs
have isomorphic duals.

• • • •
•
• • • • •
• • • •

In a connected plane multigraph, the boundary of each face is a closed walk.

Otherwise, face boundaries may consist of more than one closed walk.

9.1.8. Definition. The length of a face in a plane multigraph G is the total

length of the closed walk(s) in G bounding the face.

9.1.9. Example. A cut-edge belongs to the boundary of only one face, and it con-
tributes twice to its length. Each graph in Example 9.1.7 has three faces. In the
embedding on the left the lengths are 3 , 6 , 7; on the right they are 3 , 4 , 9. The
sum of the lengths is 16 in each case, which is twice the number of edges.

9.1.10. Proposition. (Dual Degree-Sum Formula) If l(ºi) denotes the length

of face ºi in a plane multigraph G with m edges, then 2m = ∑ l(ºi).
Proof: The face lengths are the degrees of the dual vertices. Since G ∗ and G have
the same number of edges, the statement 2m = ∑ l(ºi) is just the Degree-Sum
Formula for G ∗ . (Both sums count each edge twice.)

Statements about a connected plane multigraph become statements about its

dual when we interchange the roles of vertices and faces. Vertex degrees become
face lengths and vice versa, since edges incident to a vertex are exchanged with
edges bounding a face.
Geometrically, the Jordan Curve Theorem states that a simple closed curve
cuts its interior from its exterior. In plane multigraphs, the duality between
curve and cut becomes a duality between cycles and bonds.

9.1.11. Theorem. Edges in a plane multigraph G form a cycle in G if and only if

the corresponding dual edges form a bond in G ∗ .
Section 9.1: Embeddings and Euler ’s Formula 381

Proof: Consider D ⊆ E(G), and let D∗ denote the set of edges in G ∗ corresponding
to the edges in D. Since a bond is a minimal edge cut, it suffices to show that D∗
contains an edge cut if and only if D contains a cycle.
If D is the edge set of a cycle in G, with S being the set of dual vertices corre-
sponding to faces inside the cycle, then the edge cut [S, S] consists of all the edges
of G ∗ that cross edges of D; this is the set D∗ (by the Jordan Curve Theorem, S
and S are nonempty).
If D contains no cycle in G, then D encloses no region. The unbounded face of
G is reachable from anywhere without crossing D. Hence G ∗ − D∗ is connected,
and D∗ contains no edge cut. Hence when D is a cycle the dual cut is minimal.

• • • •
D w∗ D w∗
• • • • • • • • • • • •
• v∗ • • • • v∗ • • •
• •
• • • •

9.1.12. Remark. If a plane graph is not a forest, then every face boundary con-
tains a cycle.

9.1.13. Remark. Deleting a non-cut edge of G has the effect of contracting its
dual edge in G ∗ , since two faces of G combine into one. Contracting a non-loop
edge of G has the effect of deleting its dual edge in G ∗ . Letting G be the central
solid graph below, we have G − e on the left and G · e on the right.
To maintain this duality, when discussing planarity we often keep multiedges
and loops that arise from edge contraction.

• •
e
• • • ← • • • • → • • • • •
• •

Face boundaries allow us to characterize bipartite planar multigraphs. The

result can also be proved by induction (Exercise 17).

9.1.14. Theorem. Equivalent for a plane multigraph G are

(A) G is bipartite.
(B) Every face of G has even length.
(C) The dual G ∗ is Eulerian.
Proof: A ⇒ B. A face boundary consists of closed walks. Every odd closed
walk contains an odd cycle. Therefore, in a bipartite plane multigraph every
face boundary has even length.
B ⇒ A. Let C be a cycle in G. Since G has no crossings, C becomes a simple
closed curve in the drawing. Every face of G is wholly inside C or wholly outside
C. The total of the face lengths for the faces inside C is even, since each face
length is even. The sum counts each edge of C once. It also counts each edge in
382 Chapter 9: Planar Graphs

the region inside C twice, since each such edge belongs twice to faces inside C.
Hence the length of C and the full sum have the same parity, which is even.
B ⇔ C. The dual G ∗ is connected, and degrees in G ∗ are face lengths in G.

• • •
C

• • • •

• • •

Many questions we consider for general planar multigraphs can be answered

rather easily for graphs in a special class.

9.1.15. Definition. A multigraph is outerplanar if it has an embedding with

every vertex on the boundary of the unbounded face. An outerplane multi-
graph is such an embedding. A maximal outerplanar graph is a (sim-
ple) outerplanar graph that is not a proper subgraph of another outerplanar
graph on the same vertices.

By another drawing, the graph in Example 9.1.7 is outerplanar.

9.1.16. Proposition. The boundary of the outer face of a 2-connected outerplane

multigraph is a spanning cycle.
Proof: This boundary contains all the vertices. If it is not a cycle, then it visits
some vertex more than once; this is a cut-vertex.

9.1.17. Proposition. K 4 and K 2 ,3 are planar but not outerplanar.

Proof: The figure below shows that K 4 and K 2 ,3 are planar.
• •
• • •
•
• • •
Since they are 2-connected, an outerplane embedding requires a spanning cycle.
There is no spanning cycle in K 2 ,3 (a bipartite graph has no 5-cycle). There is
a spanning cycle in K 4 , but the endpoints of the remaining two edges alternate
along it. Hence these chords conflict and cannot both be drawn inside.

Some facts about outerplane graphs have elegant proofs using duals.

9.1.18. Definition. The weak dual of a plane multigraph G is obtained from

the dual G ∗ by deleting the vertex for the unbounded face of G.

9.1.19. Proposition. A plane graph is an outerplane graph if and only if its weak
dual is a forest.
Section 9.1: Embeddings and Euler ’s Formula 383

Proof: Since components of a plane graph G that are trees do not affect whether
G is an outerplane graph, we may assume that all components have cycles. Now
components of the weak dual H correspond to components of G. A component of
H has a cycle if and only if G has a vertex v surrounded by bounded faces of G,
which holds if and only if v is not incident to the unbounded face of G.

9.1.20. Proposition. If G is an outerplanar graph, then (G) ≤ 2. If also

|V(G)| = n ≥ 2, then G has at most 2n − 3 edges.
Proof: Let G be an outerplane graph. By Proposition 9.1.19, the weak dual H of
G is a forest. If H has no vertices, then G is a forest. If H has an isolated vertex,
then G has a component with at most one cycle. In both cases, (G) ≤ 2.
Otherwise, H has a vertex of degree 1. The corresponding face º in G shares
only one edge with a bounded face of G. Since G is simple, º has length at least
3, and the internal vertices of the path along º whose edges are not shared with
a bounded face have degree 2 in G.
An outerplanar graph with two vertices has at most one edge. For larger n,
deleting a vertex of minimum degree yields the result by induction.

EULER’S FORMUL A

Euler ’s Formula is the basic counting tool relating the numbers of vertices,
edges, and faces of multigraphs drawn in the plane.

9.1.21. Theorem. (Euler ’s Formula; Euler [1758]) If a connected plane multi-

graph G has n vertices, m edges, and r faces, then
n − m + r = 2.

Proof: We use induction on n. Basis step (n = 1): G is a “bouquet ” of loops, each a

closed curve in the embedding. If m = 0, then r = 1, and the formula holds. Each
added loop passes through a face and cuts it into two faces (by the Jordan Curve
Theorem). This augments the edge count and the face count each by 1. Thus the
formula holds for n = 1.
Induction step (n > 1): Since G is connected, it has an edge that is not a
loop. Contracting it yields a plane multigraph G with n vertices, m edges, and
r faces. Note that n = n − 1 and m = m − 1. The number of faces does not
change (we merely shorten boundaries but may create loops and multiple edges),
so r = r. By the induction hypothesis,
n − m + r = n + 1 − (m + 1) + r = n − m + r = 2.

• • •
• • • →
• •
n=1 n>1

9.1.22. Remark. (1) By Euler ’s Formula, all planar embeddings of a connected

multigraph G have the same number of faces. The dual G ∗ may depend on the
embedding of G, but the number of vertices in G ∗ does not.
384 Chapter 9: Planar Graphs

(2) Euler ’s Formula as stated fails for disconnected multigraphs. If a plane

multigraph G has ¾ components, then adding ¾ − 1 edges to G yields a connected
plane multigraph without changing the number of faces. This yields the gener-
alization n − m + r = ¾ + 1.
(3) Euler ’s Formula is familiar for polyhedra as V − E + F = 2, where V, E , F
are the numbers of vertices, edges, and faces. However, in graph theory V and
E are operators returning the sets of vertices and edges, and F is usually a set
of edges or a graph. Since we often use n and m for the numbers of vertices and
edges, using r to count the faces is consistent and frees º for individual faces. We
use r because it is the number of maximal regions, that is, faces.

We will see various applications of Euler ’s Formula.

9.1.23. Theorem. Let G be a planar n-vertex graph with m edges. If n ≥ 3, then

m ≤ 3n − 6. If also G is triangle-free, then m ≤ 2n − 4.
Proof: By adding edges, it suffices to consider connected graphs. Let r be the
number of faces in some embedding of G, Euler ’s Formula will relate n and m if
we can dispose of r. Proposition 9.1.10 provides an inequality for m and r.
For n ≥ 3, every face boundary in a connected plane graph has length at
least 3. Summing the face lengths yields 2m ≥ 3r. Substituting r ≤ 2m/3 into
n − m + r = 2 yields m ≤ 3n − 6.
When G is triangle-free, all faces have length at least 4. In this case 2m ≥ 4r
and r ≤ m/2, and we obtain m ≤ 2n − 4.

9.1.24. Example. Nonplanarity of K 5 and K 3 ,3 follows again from Theorem

9.1.23. A planar graph with five vertices has at most nine edges (3n − 6), but
K 5 has 10 edges. Although K 3 ,3 has only nine edges, it is triangle-free, and a
triangle-free graph with six vertices has at most eight edges (2n − 4). In both
cases, there are too many edges to be planar.

9.1.25. Definition. A maximal planar graph is a planar graph that is not a

spanning subgraph of another planar graph. A triangulation is a plane
multigraph where every face boundary is a 3-cycle.

9.1.26. Proposition. For an n-vertex plane graph G, equivalent are:

(A) G has 3n − 6 edges.
(B) G is a triangulation.
(C) G is a maximal plane graph.
Proof: A ⇔ B. For an n-vertex plane graph, the proof of Theorem 9.1.23 shows
that having 3n − 6 edges is equivalent to 2m = 3r, which occurs if and only if
every face has length 3.
B ⇔ C. Some face has length more than 3 if and only if there is a way to add
an edge to the drawing and obtain a larger plane graph.

9.1.27. Remark. “Maximal plane multigraph” makes no sense, since edges may
have high multiplicity, but an n-vertex triangulation still has 3n − 6 edges. To
obtain a triangulation with multiedges from a maximal plane graph, an edge xy
bounding faces A and B can be “widened” into two edges and two vertices inserted
Exercises for Section 9.1 385

as shown below, adding six edges and four faces. Every face still has length 3. In
the dual, the edge AB is replaced using four new vertices as shown.
x x
• • •

A B → A • • B
• •
• •
• •
y y
•
• •
• •
A• • B → A• • • • B
•

9.1.28. Remark. A graph embeds in the plane if and only if it embeds on a

sphere. Given an embedding on a sphere, puncture the sphere inside a face and
project the embedding onto a plane tangent to the opposite point. This yields
a planar embedding in which the punctured face on the sphere becomes the un-
bounded face in the plane. The process is reversible.

9.1.29. Application. A regular polygon is a closed simple polygonal curve hav-

ing sides of equal length and equal angles between successive sides. A regular
polyhedron is a solid whose faces are isomorphic regular polygons, with the same
number of faces meeting at each vertex. When we expand the polyhedron to a
sphere and then move the drawing to the plane as in Remark 9.1.28, we obtain a
regular plane graph with faces of the same length. Hence the dual also is regular.
Let G be a plane graph with n vertices, m edges, and r faces. If G is ¾-regular
and G ∗ is l-regular, then ¾ n = 2m = lr, using the Degree-Sum Formulas for G
and G ∗ . Substituting for n and r in Euler ’s Formula yields m( ¾2 − 1 + 2l ) = 2.
Since m and 2 are positive, the other factor must also be positive, which yields
(2/¾) + (2/l) > 1 and hence 2l + 2 ¾ > ¾ l. We rewrite this as (¾ − 2)(l − 2) < 4.
Because the dual of a 2-regular graph is not simple, we require ¾ , l ≥ 3. The
pairs solving (¾ − 2)(l − 2) < 4 are now (3 , 3), (3 , 4), (3 , 5), (4 , 3), (5 , 3).
Given ¾ and l, there is only one way to lay out the plane graph when we start
with any face. Hence the regular polyhedra are only the five Platonic solids listed
below, one for each pair (¾ , l) that satisfies the requirements.
¾ l (¾ − 2)(l − 2) m n r name
3 3 1 6 4 4 tetrahedron
3 4 2 12 8 6 cube
4 3 2 12 6 8 octahedron
3 5 3 30 20 12 dodecahedron
5 3 3 30 12 20 icosahedron

EXERCISES 9.1

9.1.1. (−) Let G be a plane graph with dual G∗ . What is the effect on the dual when an
edge of G is subdivided? What is the effect when an edge of G is duplicated (that is, when
another edge with the same endpoints is added)?
386 Chapter 9: Planar Graphs

9.1.2. (−) Count the isomorphism classes of planar duals of the graph below.
• •

• • •
• •

9.1.3. (−) Prove or disprove: If G is a 2-connected plane graph with minimum degree 3 (no
multiedges), then its dual G∗ has no multiedges.
9.1.4. (−) Let G be a maximal plane graph with | V(G)| ≥ 3. Prove or disprove each below:
(a) G must be a chordal graph.
(b) G must have an even number of faces.
(c) G∗ must be 2-edge-connected and 3-regular.
9.1.5. (−) Let G be a plane graph with an odd number of vertices. Prove or disprove: If
G∗ is Eulerian, then G cannot be Hamiltonian.
9.1.6. (−) Prove that a 2-edge-connected 3-regular plane graph is 3-connected if and only
if no two faces share at least two boundary edges.
9.1.7. (−) Use edge deletion to prove Euler ’s Formula by induction on the number of edges,
with trees in the basis step.
9.1.8. (−) Let S be a set of n points in the plane with no two points in S closer together
than 1 inch. Prove that at most 3n − 6 pairs are exactly 1 inch apart.
9.1.9. (−) Obtain the number of faces in a ¾-regular plane graph with n vertices.
9.1.10. (−) Prove that every plane graph with minimum degree at least 4 has a face of
length 3.
9.1.11. (−) Prove that every plane graph with minimum degree at least 3 has a face with
length at most 5.
9.1.12. (−) Determine the number of faces in a plane graph with n vertices, m edges, and
¾ components.
9.1.13. (−) For n ≥ 5, construct an n-vertex maximal planar graph that does not have
three disjoint cycles.
9.1.14. (−) Let G be a maximal planar graph with at least five vertices. Prove that G does
not have adjacent vertices of degree 3.
9.1.15. Let G be a plane multigraph. Prove that (G∗)∗ ∼
= G if and only if G is connected.
9.1.16. Contracting an edge of a plane multigraph deletes the corresponding edge from the
dual: (G · e)∗ = G∗ − e∗ (Remark 9.1.22(2)). Use this to prove Theorem 9.1.11 inductively:
a set D ⊆ E(G) is a cycle in G if and only if the dual set D∗ is a bond in G∗ .
9.1.17. Prove by induction on the number of faces that a plane multigraph is bipartite if
and only if every face has even length.
9.1.18. Prove that every plane n-vertex multigraph isomorphic to its dual has 2n− 2 edges.
For n ≥ 4, construct an example with no loops or multiedges.
9.1.19. For ¾ ≥ 3, determine all maximal planar graphs with ¾ + 2 vertices such that all
vertices have degree 4 except for two vertices of degree ¾ .
9.1.20. Construct a vertex-transitive 3-regular planar graph of diameter 3 with 12 ver-
tices. (Comment: T. Barcume proved the conjecture of Erdős that no 3-regular planar
graph with diameter 3 has more than 12 vertices.)
Exercises for Section 9.1 387

9.1.21. Construct a 16-vertex 5-regular planar graph with diameter 3. (Hint: The unique
such graph can be drawn with 4-fold rotational symmetry and contains a 4-regular span-
ning subgraph with diameter 3. Comment: Preen [2012] proved that this graph is the
only 5-regular planar graph with diameter 3 other than the icosahedron.)
9.1.22. Let S be a separating 3-set of the dual of a maximal planar graph G. Prove that
G∗ − S has two components.
9.1.23. (♦) Prove that every maximal planar graph having at least four vertices is 3-
connected. (Ore [1967, p. 6])
9.1.24. (♦) A fullerene is a 3-regular 3-connected plane graph whose faces have length 6

except for twelve of length 5. For ¾ ∈ 0 , construct a fullerene with 5 ¾ faces of length 6 and
another with 6 ¾ + 2 faces of length 6. (Comment: Gr ünbaum–Motzkin [1963] constructed

fullerenes with r faces of length 6 for all r ∈ with r > 1.)
9.1.25. (♦) Prove that edges in a connected plane multigraph G form a spanning tree if
and only if the duals of the other edges form a spanning tree in G∗ . (von Staudt [1847])
9.1.26. (♦) Let G be a connected plane multigraph. Prove that G is Hamiltonian if and
only if V(G∗) can be partitioned into two sets S and T such that G∗ [S] and G∗ [T] are both
trees. (Stein [1970] proved this for triangulations.)
9.1.27. Let G be an Eulerian plane multigraph. Prove that G cannot satisfy either prop-
erty below. (Comment: These facts can be used in Exercise 9.3.28.)
(a) One face of G has length 2 or 4 and the rest have length 3.
(b) Every face of G has length 3 and G has a loop.
9.1.28. Use Euler ’s Formula to prove that an n-vertex maximal outerplanar graph has
2n − 3 edges.
9.1.29. Alternative proof of Proposition 9.1.20. Prove inductively that every outerplanar
graph with at least four vertices has two nonadjacent vertices with degree at most 2.
9.1.30. Let G be an outerplane graph other than K3 whose bounded faces all have length
3. Let t be the number of faces in G sharing no edge with the unbounded face. Prove that
G has exactly t + 2 vertices of degree 2.
9.1.31. Let G0 = K3 . For ¾ ≥ 1, let G ¾ be the plane graph obtained from G ¾−1 by adding a
new vertex ve for each edge e on the unbounded face of G ¾−1 , adjacent only to the endpoints
of e. Note that G ¾ is outerplanar.
(a) Determine the degree list of G ¾ .
(b) Prove that every outerplanar graph is a subgraph of G ¾ for some ¾ .
9.1.32. (♦) A triangulation of a convex n-gon is a maximal outerplanar graph obtained
by adding chords joining corners of the n-gon. Flipping a chord replaces it with the other
chord that cuts the union of the two triangles bounding it , as shown below. Flipping a
chord yields another triangulation. Prove that any triangulation can be turned into any
other using at most 2n − 6 flips, and prove that at most 2n − 10 suffice when n > 12.
(Comment: Sleator–Tarjan–Thurston [1988] used hyperbolic geometry to prove that the
bound 2n − 10 is best possible when n is sufficiently large. Pournin [2014] found a shorter
combinatorial proof that 2n − 10 is sharp when n > 12.)

• • • •

• • ↔ • •

• • • •
388 Chapter 9: Planar Graphs

9.1.33. For G ∈ {K6 , K7 }, determine whether G decomposes into two outerplanar graphs.
9.1.34. (♦) Let G be a nontrivial maximal outerplanar graph with vertex degrees all even.
Prove that | E(G)| is divisible by 3. (Jonsson–Propp [2007])
9.1.35. Let l be the length of a longest cycle in a planar triangulation G. Prove that G
has cycles of all lengths from 3 through l. (Balister)
9.1.36. (♦) Euler’s Formula using linear algebra. Let A be an n-by-m matrix over a field ;
its nullspace is {x ∈ m : Ax = 0}. The Rank–Nullity Theorem of linear algebra states
that the rank of A plus the dimension of its nullspace equals the number of columns. Apply
this to the incidence matrix (over the field 2 of size 2) to prove Euler ’s Formula.
9.1.37. Find the maximum number of edges in a planar subgraph of the hypercube Q¾ .
9.1.38. Let G be a 3-regular connected plane graph in which every vertex is incident to
one face of length 4, one face of length 6, and one face of length 8. Without drawing G,
use Euler ’s Formula to count the faces of G.
9.1.39. The rhombicosidodecahedron is a polyhedron in which every vertex is incident
to one triangular face, one pentagonal face, and two (opposite) quadrilateral faces. Deter-
mine the number of faces in the rhombicosidodecahedron. (Comment: The toy construction
system “ Zometool” is based on this polyhedron.)
9.1.40. (♦) Use Euler ’s Formula to count the regions formed by n lines in the plane, as-
suming that no two are parallel and no three have a common point.
9.1.41. (♦) Consider a convex n-gon such that no three segments joining corners have a com-
mon internal point. Use Euler ’s Formula (not induction!) to count the bounded regions in
the drawing after adding all (n2) segments joining corners of the n-gon. The answers for
3 ≤ n ≤ 6 are 1, 4, 11, and 25.
9.1.42. Let R be a convex region in the plane. Suppose that q chords with distinct end-
points on the boundary of R are drawn and form p points of intersection, with no three
chords having a common point. In terms of p and q, compute the number of regions into
which the chords cut R. (Alexanderson–Wetzel [1977])
9.1.43. (♦) Let º be a proper vertex coloring of an n-vertex triangulation using the colors
{a , b , c , d}. Let t be the number of edges joining colors a and b plus the number of edges
joining colors c and d. Determine t. (X. Lv)
9.1.44. Prove that if G is a planar graph with at least 11 vertices, then G is nonplanar.
Construct a self-complementary planar graph with eight vertices.
9.1.45. (♦) Let G be a connected n-vertex planar graph with m edges. Prove that
m ≤ ½ −½ 2 (n − 2) when G has girth ½ .
9.1.46. Prove that every n-vertex planar graph with n + ¾ edges has a cycle of length
+¾) 2(n+¾)
at most 2(n
¾ +2 . To prove that the bound is sharp, construct an example with girth ¾ +2
whenever n − 2 is divisible by ¾ + 2.
9.1.47. Unusual graphs.
(a) For a planar graph with minimum degree at least 3 and ni vertices of degree i,
prove 3n3 + 2n4 + n5 ≥ 12.
(b) Let G be a 3-connected plane graph in which no three faces have the same length.
Prove that G has two faces each of lengths 3, 4, and 5 and no face at all of length at least
7. There are four such graphs (Jorza [2001]); construct one.
9.1.48. (♦) Prove that every n-vertex plane graph decomposes into at most 2n − 4 edges
and facial triangles. Determine the graphs for which 2n − 4 such subgraphs are needed.
Exercises for Section 9.1 389

9.1.49. (♦) Let G be a connected even plane graph. An Eulerian circuit is noncrossing if
it does not cross itself when viewed as a closed curve. That is, in the embedding around a
vertex v, the two edges used on a visit through v are consecutive. Prove the following.
(a) G has a noncrossing Eulerian circuit. (Abrham–Kotzig [1979], Singmaster [1981])
(b) If also every bounded face is a triangle, then | E(G)| is divisible by 3.
(c) If also G is a maximal outerplanar graph, then | V(G)| is divisible by 3.
• • •
• • •
• •


9.1.50. For ¾ ∈ , prove that the following exist: an Eulerian planar graph with min-
imum degree 4 where the distance between degree-4 vertices is at least ¾ , and a planar
graph with minimum degree 5 where the distance between degree-5 vertices is at least ¾ .

9.1.51. Prove that every 4-regular plane graph has at least eight triangular faces. For

each ¾ ∈ , construct a 4-regular plane graph with exactly eight triangular faces and
such that any path connecting two of these faces has length at least ¾ .

9.1.52. Determine all n such that there exists a 4-regular planar graph with n vertices.
(Chvat ál [1969], Owens [1971])

9.1.53. For ¾ ∈  , construct a connected 5-regular planar graph with 12¾ vertices.
9.1.54. (♦) For n ≥ 4, prove that every n-vertex planar graph has at least four vertices
with degree less than 6. Prove that equality may hold when n is even and at least 8.
(Gr ünbaum–Motzkin [1963])

9.1.55. (♦) Directed plane graphs. Let G be a plane graph, and let D be an orientation of
G. The dual D∗ of D is an orientation of G∗ such that when an edge of D is viewed from
tail to head, the dual edge in D∗ crosses it from right to left. For example, if the solid
edges below are in D, then the dashed edges are in D∗ .

• • •

• • •
Prove that if D is strongly connected, then D∗ has no directed cycle, so −(D∗) = +(D∗) =
0. Conclude that if D is strong, then D has a face whose boundary is a clockwise cycle and
another face whose boundary is a counterclockwise cycle.

9.1.56. Prove that every 2-connected plane graph has an ear decomposition in which ears
are successively removed from the boundary of the external face of the remaining graph.

9.1.57. (♦) Pick’s Theorem. A lattice polygon is a polygon whose corners are at integer
lattice points in the plane.
(a) Let G be a 2-connected plane graph whose bounded faces are triangles, with B
vertices on the unbounded face and I other vertices. In terms of B and I, determine the
numbers of edges and faces in G.
(b) Let T be a lattice triangle. Prove that if no other lattice points lie on the bound-
ary or in the interior of T , then T has area 1/2. (Hint: Use induction on the area of the
smallest lattice rectangle containing T .)
(c) Let P be a lattice polygon with B lattice points on the perimeter and I lattice
points inside. Prove Pick’s Theorem: the area of the region bounded by P is I + B/2 − 1.
(DeTemple–Robertson [1974], Gaskell–Klamkin–Watson [1976])
390 Chapter 9: Planar Graphs

9.2. Structure of Planar Graphs

Which graphs embed in the plane? We know that K 5 and K 3 ,3 do not. These
two graphs are critical for Kuratowski’s characterization of planar graphs. After
the publication of Frank Harary ’s textbook, K asimir Kuratowski asked Harary
about the origin of the notation for K 5 and K 3 ,3 . Harary replied, “ The K in K 5
stands for K asimir, and the K in K 3 ,3 stands for Kuratowski!”
In this section we also discuss the Planar Separator Theorem, a tool for solv-
ing computational problems quickly on planar graphs.

KURATOWSKI’S THEOREM

We generalize the notion of subdividing an edge of a graph (Definition 7.2.20).

9.2.1. Definition. A subdivision of F or F-subdivision is a graph obtained

from a graph F via edge subdivisions; edges become paths through new ver-
tices. When (F) ≥ 3 and H is an F-subdivision, the vertices with degree at
least 3 in H are the branch vertices of H ; these were the vertices of F .
• • • •
• •

• •
•
• •
•
• • •
a K 3 ,3-subdivision

9.2.2. Proposition. A graph containing a subdivision of K 5 or K 3 ,3 is nonplanar.

Proof: Subgraphs of planar graphs are planar, so it suffices to show that subdi-
visions of K 5 and K 3 ,3 are nonplanar. By Proposition 9.1.1, K 5 and K 3 ,3 are not
planar. Subdividing edges does not affect planarity; the curves in an embedding
of a subdivision of G can be used to obtain an embedding of G, and vice versa.

By Proposition 9.2.2, avoiding subdivisions of K 5 and K 3 ,3 is a necessary con-

dition for being a planar graph. It also is sufficient:

9.2.3. Theorem. (Kuratowski [1930]) A graph is planar if and only if it does not
contain a subdivision of K 5 or K 3 ,3 .

Thus K 5 and K 3 ,3 are the only topologically minimal nonplanar graphs (not
a subdivision of another nonplanar graph). Wagner [1937] proved another char-
acterization. Deletions and contractions preserve planarity, so we can seek the
Section 9.2: Structure of Planar Graphs 391

minimal nonplanar graphs under these operations. Wagner proved that G is pla-
nar if and only if it has no subgraph contractible to K 5 or K 3 ,3 . This follows easily
from Kuratowski’s Theorem (see Exercise 13), the proof of which is our next goal.

9.2.4. Definition. A Kuratowski subgraph of G is a subgraph of G that is

a subdivision of K 5 or K 3 ,3 . A minimal nonplanar graph is a nonplanar
graph such that every proper subgraph is planar.

We will show that a smallest nonplanar graph not having a Kuratowski sub-
graph must be 3-connected. To prove Kuratowski’s Theorem, it then suffices to
show that 3-connected graphs without Kuratowski subgraphs are planar.

9.2.5. Lemma. If F is the edge set of a face in an embedding of G, then G has

an embedding where F is the edge set of the unbounded face.
Proof: Project the embedding onto the sphere, where the edge sets of regions
remain the same and all regions are bounded, and then return to the plane by
projecting from inside the face bounded by F .

9.2.6. Lemma. Every minimal nonplanar graph is 2-connected.

Proof: Let G be a minimal nonplanar graph. If G is disconnected, then we embed
one component of G inside one face of an embedding of the rest.
If G has a cut-vertex v, then let G 1 , . . . , G ¾ be the {v}-lobes of G. By the
minimality of G, each G i is planar. By Lemma 9.2.5, we can embed each G i with
v on the outside face. We fit each embedding into an angle smaller than 360/¾
degrees at v and then combine them to obtain an embedding of G.

9.2.7. Lemma. Let {x , y} be a separating 2-set of G, let G 1 , . . . , G ¾ be the {x , y}-

lobes of G, and let Hi = G i ∪ xy. If G is nonplanar, then some Hi is nonplanar.
Proof: If Hi is planar, then by Lemma 9.2.5 it embeds with xy on the outside
face. Combine such embeddings, with each successive one embedded in a face of
the current graph having xy on the boundary. Now delete xy if it is not in E(G).
The result is a planar embedding of G.

We next reduce Kuratowski’s Theorem to the 3-connected case. The hypoth-

esized graph doesn’t exist, but if it did, it would be 3-connected.

9.2.8. Lemma. If G is a graph with fewest edges among all nonplanar graphs
without Kuratowski subgraphs, then G is 3-connected.
Proof: Deleting an edge of G cannot create a Kuratowski subgraph in G. Thus
deleting one edge produces a planar subgraph, and hence G is a minimal nonpla-
nar graph. By Lemma 9.2.6, G is 2-connected.
Suppose that G has a separating 2-set S, with S = {x , y}. Since G is nonpla-
nar, the union of xy with some S-lobe is nonplanar (Lemma 9.2.7); let H be such
a graph. Since H has fewer edges than G, minimality of G forces H to have a
Kuratowski subgraph F . All of F appears in G except possibly the edge xy.
Since S is a minimal vertex cut, both x and y have neighbors in every S-lobe.
Thus we can replace xy in F with an x , y-path through another S-lobe to obtain a
392 Chapter 9: Planar Graphs

Kuratowski subgraph of G. This contradicts the prohibition of Kuratowski sub-

graphs, so G has no separating 2-set.

x
• • • •
F • H
• • • •
y

To prove that 3-connected graphs without Kuratowski subgraphs are planar,

we use induction. For the induction step, we obtain from such a graph a smaller
3-connected graph without Kuratowski subgraphs, by contracting an appropri-
ate edge. We show first that no edge contraction can introduce a Kuratowski
subgraph if none was present before.

9.2.9. Lemma. If G · xy has a Kuratowski subgraph, then G has a Kuratowski

subgraph.
Proof: Let H be a Kuratowski subgraph of G · xy, and let be the vertex of G ·
xy obtained by contracting xy. If is not in H , then H itself is a Kuratowski
subgraph of G. If ∈ V(H) but is not a branch vertex of H , then we obtain a
Kuratowski subgraph of G from H by replacing with x or y or with the edge xy.
Similarly, if is a branch vertex in H and at most one edge at in H is in-
cident to x in G, then expanding into xy lengthens that path, and y becomes a
branch vertex for a Kuratowski subgraph in G.
In the remaining case (shown below), H is a K 5-subdivision, is a branch
vertex of H , and each of x and y is incident in G to two of the four edges incident
to in H. Let u1 and u2 be the branch vertices of H reached via the paths leaving
on edges incident in G to x. Similarly let v1 and v2 be the branch vertices of H
reached by the paths leaving on edges incident in G to y. Deleting the u1 , u2-
path and v1 , v2-path from H yields a K 3 ,3-subdivision in G having y , u1 , u2 as the
branch vertices for one part and x , v1 , v2 as the branch vertices for the other.

u1 v1 u1 v1
• • • •
y x
• → • •
• • • •
v2 u2 v2 u2
H in G in G

Now we can prove Kuratowski’s Theorem. In fact, we prove a stronger prop-

erty for 3-connected graphs without Kuratowski subgraphs.

9.2.10. Definition. A convex embedding of a planar graph is an embedding in

which each face boundary is a convex polygon (including the unbounded face).

9.2.11. Theorem. (Tutte [1960, 1963]) If G is a 3-connected graph containing no

subdivision of K 5 or K 3 ,3 , then G has a convex embedding in the plane with
no three vertices on a line.
Section 9.2: Structure of Planar Graphs 393

Proof: (Thomassen [1980a, 1981a]) We use induction on n, the order of G. The

only 3-connected graph with at most four vertices is K 4 , which has such an em-
bedding. For the induction step, consider n ≥ 5.
Let e with endpoints x and y be an edge such that G · e is 3-connected, guar-
anteed by Lemma 7.2.18. Let be the vertex obtained by contracting e, and let
H = G · e. By Lemma 9.2.9, H has no Kuratowski subgraph. By the induction
hypothesis, we obtain a convex embedding of H with no three vertices on a line.
In this embedding, the subgraph obtained by deleting the edges incident to
has a face containing (perhaps unbounded). Since H − is 2-connected, the
boundary of this face is a cycle C. All neighbors of lie on C; they may be neigh-
bors in G of x or y or both.
The embedding of H has segments from to its neighbors. Let x1 , . . . , x ¾ be
the neighbors of x in cyclic order on C. If all neighbors of y lie in the portion of
C from x i to x i+1 , then we obtain a convex embedding of G by putting x at in H
and putting y at a point close to in the wedge formed by xx i and xx i+1 (slightly
moving the segments from to NG(y)), as shown in Case 0 below.
If this does not occur for any i, then either (1) y shares three neighbors u , v , w
with x, or (2) y has neighbors u and v that alternate on C with neighbors x i and
x i+1 of x. If (1), then the union of C with xy and the edges from {x , y} to {u , v , x}
is a K 5-subdivision. If (2), then the union of C with the paths uyv, x i xx i+1 , and
xy is a K 3 ,3-subdivision. Since our graph has no Kuratowski subgraph, only Case
0 occurs.
• • w
• •
x i+1
• •
• • •
x x• x•
xi • • • • x i+1 • • x • • u v• •u
•y y •y •y
• •
• • • v• • •
xi
•
x i+1 xi
Case 0 Case 1 Case 2

9.2.12.* Remark. Thomassen [1984] proved that the regions in Theorem 9.2.11
can be required to be convex hexagons, but not convex pentagons. Other exten-
sions guarantee special Kuratowski subgraphs in nonplanar graphs. Kelmans
[1981] conjectured that every 3-connected nonplanar graph with at least six ver-
tices has a cycle with three pairwise crossing chords. This was proved by Kelmans
[1984] and by Thomassen [1984].
A 2-connected planar graph need not have a convex embedding (consider
K 2 ,4), but every planar graph has an embedding where all edges are straight
line segments. Proved by Wagner [1936], Fáry [1948], and Stein [1951], this is
known as Fáry’s Theorem (Exercise 9).

9.2.13.* Remark. Kuratowski’s Theorem does not directly yield fast planarity
testing: too many subgraphs to check. Demoucron–Malgrange–Pertuiset [1964]
obtained a quadratic-time algorithm, and it produces an embedding. The proof of
Kuratowski’s Theorem by K lotz [1989] provides another algorthm, and it finds a
Kuratowski subgraph when the graph is not planar.
394 Chapter 9: Planar Graphs

Linear-time planarity testing algorithms appear in Hopcroft–Tarjan [1974]

and Booth–Luecker [1976]; these are complicated (Gould [1988, pp. 177–185] de-
scribes the Hopcroft–Tarjan algorithm). Later linear-time algorithms appear in
Boyer–Myrvold [1999] and Shih–Hsu [1999].
For computer display, one seeks a straight-line embedding with the ver-
tices at grid points. Schnyder [1990] proved that a planar n-vertex graph has
a straight-line embedding at points of [n − 1] × [n − 1], and it can be found in
linear time from any embedding of G. We discuss this theorem in Chapter 16.

9.2.14.* Remark. Planar graphs have various geometric representations. The

tangency graph of a family of geometric objects has a vertex for each object,
with vertices adjacent when the objects are tangent. Every planar graph is a
tangency graph of rectangles in the plane (Thomassen [1986]) and also of circles
in the plane (Koebe [1936], Brightwell–Scheinerman [1993], Thurston [1997]).
Many computational problems involve visibility in the plane. The famous
Art Gallery Problem (Exercises 9.3.25–26) seeks the minimum number of
guards needed to watch all locations of interest (see O ’Rourke [1987]). Visibil-
ity can also be in restricted directions. A graph is a bar visibility graph if each
vertex can be assigned a bar (a horizontal segment) in the plane so that vertices
are adjacent if and only if their bars can “see” each other along an unblocked ver-
tical channel. Wismath [1985] and Tamassia–Tollis [1986] proved that a graph
G is a bar visibility graph if and only if G has a planar embedding with all cut-
vertices on the unbounded face. Later Hutchinson [2002] gave a short proof.

THE SEPARATOR THEOREM (optional)

In “divide-and-conquer ” algorithms, we solve a problem recursively by split-

ting it into smaller pieces. This works well for planar graphs.

9.2.15. Definition. An (m , )-separation of an n-vertex graph G splits V(G)

into sets A , B , C such that | A| , | B| ≤ n, and | C| ≤ m, and no edges join A
and B. A graph family F has an ( , )-separator if every graph G ∈ F has
an ( (| V(G)|) , )-separation.

An ( , )-separator may yield good divide-and-conquer algorithms on a hered-

itary class F. For G ∈ F with n vertices, the approach is
(1) find an ( (n) , )-separation of G, with vertex partition A , B , C,
(2) solve the specified problem on G[A] and G[B], and
(3) combine these answers with C to solve the problem on G.
If (n) is the worst-case cost for doing steps (1) and (3), then the worst-case cost
h(n) for n-vertex graphs in F satisfies h(n)√ ≤ 2h( n) + (n). If  = 1/2 and (n) =
an, then h(n) ∈ O(n log n). When (n) = a n with  = 1/2, the running time is
linear (Exercise 20). With  = 2/3, the complexity results are similar.
Trees have a (1 , 2/3)-separation (Exercise 21). For n-vertex
√ subgraphs of a
cartesian product of two paths (grid graphs), there is a ( n , 2/3)-separation.
General planar graphs behave almost as well; the Separator √ Theorem of Lipton–
Tarjan [1979] provides an ( , )-separator with (n) = 2 2n and  = 2/3.
Section 9.2: Structure of Planar Graphs 395

Planar graphs are sparse, with average degree less than 6, but sparseness
alone does not guarantee easy separation. Erdős–Graham–Szemerédi [1976]
proved that for > 0 there is a positive constant c such that almost all graphs
with (2 + )¾ vertices and c ¾ edges retain a component with at least ¾ vertices
when any ¾ vertices are deleted. (See Chapter 14 for discussion of “almost all”.)
Thus the Separator Theorem truly needs the properties of planarity. The
original proof (Exercise 24) used weighted vertices and obtained separations from
spanning trees of small diameter. We present a later short proof based on connec-
tivity and path-lengths. A near-triangulation is a 2-connected plane graph in
which every bounded face has length 3.

9.2.16. Lemma. Let G be a near-triangulation with vertices colored red or blue.

If the outer cycle C has red points u and v, then G has a red u , v-path or a
blue path joining the components of C − {u , v}.
Proof: Let A be the subset of V(G) reachable from u by paths in red. If G has
no red u , v-path, then the edge cut [A , A] contains a bond B that breaks all u , v-
paths. By Theorem 9.1.11, the edges corresponding to B in the dual graph G ∗
form a cycle B∗ with one of {u , v} inside and the other outside. The bounded faces
in G corresponding to vertices of B∗ are triangles having a red vertex and a blue
vertex. Deleting A from the union of these triangles (as on the left below) leaves
a path through blue vertices joining the two components of C − {x , y}.

•u • v0
• • • • • • • •

• • • • • • • • • •
Y S X
◦ ◦ ◦ ◦ ◦ • • • • •
•◦ ◦ ◦• ◦ •◦ • • •
•◦ v •v


9.2.17. Lemma. Let G be a near-triangulation; outer cycle C = [v0 , . . . , v2−1 ].

If d G(v0 , v) ≥ ¾ , then G has disjoint paths linking vi to v2−i for 0 ≤ i ≤ ¾ .
If also d G(vi , vj ) = d C(vi , vj) for all vi , vj ∈ V(C), then | V(G)| ≥ (¾ + 1)2/2.
Proof: Let X = {v0 , . . . , v } and Y = {v , . . . , v2 }, where v2 = v0 . Let S be a
smallest X , Y-barrier. Color S red, and color the remaining vertices blue. Since
S is an X , Y-barrier, there is no blue X , Y-path. By Lemma 9.2.16, there is a red
v0 , v -path. By hypothesis, such paths have length at least ¾ , so | S| ≥ ¾ + 1.
By Pym’s Theorem, G has an X , Y -link of size | S| (shown above). It has ¾ + 1
disjoint X , Y -paths. They must link vi to v2−i for all i; otherwise edges cross.
If G has no shorter vi , v2−i-path than the one along C, then the vi , v2−i-path
in the set we have obtained has at least 1 + min{2i , 2 ¾ − 2i} vertices. Hence the
total number of vertices in these ¾ + 1 paths is at least 1 + 3 + 5 + · · · + 5 + 3 + 1,
with ¾ + 1 terms. The sum of the first j positive odd integers is j 2 , so we compute
⌈( +1)/2⌉ ⌊( +1)/2⌋
¾+1 2
¾+1 2
(¾ + 1)2
∑ (2i − 1) + ∑ (2i − 1) = ⌈
2
⌉ +⌊
2
⌋ ≥
2
.
i=1 i=1
396 Chapter 9: Planar Graphs

9.2.18. Theorem. (Planar Separator √ Theorem; Lipton–Tarjan [1979]). Ev-

ery n-vertex planar graph G has a (2 2n , 2/3)-separation.
Proof: (Alon–Seymour–Thomas [1994]) Since separators of graphs are separa-
tors of√their spanning subgraphs,
√ we may assume that G is a triangulation. Let
¾ = ⌊ 2n⌋ ; since ¾ > 2n − 1, we have n < (¾ + 1)2/2. We seek a (2 ¾ , 2/3)-
separation using a cycle as the separating set, separating the inside from the
outside. We guarantee this by showing that if a well-chosen cycle fails, then it
will meet the conditions of Lemma 9.2.17, yielding | V(G)| ≥ (¾ + 1)2/2.
Given a cycle C in G, let c− and c+ denote the numbers of vertices in the in-
terior and exterior of C, respectively. Among all cycles of length at most 2 ¾ such
that c+ < 2n/3, let C be one minimizing c− − c+ . Such a cycle exists, since the
external triangle
√ has fewer than 2 ¾ vertices and has no vertices outside.
The (2 2n , 2/3)-separation exists if c− < 2n/3; suppose not. Let D be the
subgraph on and inside C. For u , v ∈ V(C), let c(u , v) and d(u , v) be the lengths
of shortest u , v-paths in C and D; C ⊆ D implies d(u , v) ≤ c(u , v).
It now suffices to prove d(u , v) = c(u , v) for u , v ∈ V(C) and also | V(C)| = 2 ¾ .
With d(v0 , v¾) = ¾ , Lemma 9.2.17 then applies to yield n ≤ | V(D)| ≥ (¾ + 1)2/2.
Claim 1: d(u , v) = c(u , v) for u , v ∈ V(C). If not, then among the pairs with
d(u , v) < c(u , v), choose {u , v} with d(u , v) smallest. Let P be a shortest u , v-path
in D. By the choice of {u , v}, no internal vertex of P lies on C. Thus P combines
with the two u , v-paths on C to form cycles C1 and C2 , indexed so c−1 ≥ c−2 .

• ••
C2 • c−
2 •P c−1 C1
• •

We claim that C1 contradicts the minimality of C. Since d(u , v) < c(u , v), the
cycle C1 is shorter than C and hence has length at most 2 ¾ . We have c−1 ≤ c− and
c+1 > c+ , so c−1 − c+1 < c− − c+ . To show c+1 < 2n/3, use c−1 ≥ c−2 to compute
n − c+1 = c−1 + | V(C1)| > 12 c− ≥ 13 n.
Thus C1 is an eligible cycle and would be preferred to C, a contradiction.
Claim 2: C has 2 ¾ vertices. By Claim 1, C has no chords; hence every edge
of C is one edge of a triangular face with a vertex inside C, since c− ≥ 2n/3. If
|V(C)| < 2¾ , then obtain C from C by detouring to absorb one such interior ver-
tex. Since c+ vertices remain outside C , and fewer than c− vertices are inside C ,
the cycle C contradicts the choice of C. Hence we must have | V(C)| = 2 ¾ .

Iterating the Separator Theorem reduces to 1/2 at the expense of increas-

ing the constant in  (n).

9.2.19. Corollary.
√
(Lipton–Tarjan [1979]). Every n-vertex planar graph G has
2 √2n
a( , 2 )-separation.
1
1− 2/3
Proof: We build A , B , C gradually; the vertices not yet placed lie in D. Initially
A0 = B0 = C0 = ∅ and D0 = V(G). Step i − 1 produces disjoint sets Ai−1 , Bi−1 ,
and Di−1 such that no edges join any two of them and | Ai−1 | ≤ | Bi−1 | ≤ n/2.
Exercises for Section 9.2 397

At step i, apply Theorem 9.2.18 to the remaining “large” piece Di−1 . This
produces C∗ separating A∗ and B∗ within Di−1 , labeled so that | A∗ | ≤ | B∗ |. Com-
bine A∗ with Ai−1 . Since | A∗ | ≤ | B∗ | and | Ai−1 | ≤ | Bi−1 | , we have | Ai−1 ∪ A∗ | ≤ n/2.
Also | Bi−1 | ≤ n/2 is given. Let Ai be the smaller of Ai−1 ∪ A∗ and Bi−1 , and let Bi
be the other. Now | Ai | ≤ | Bi | ≤ n/2. Let Ci = Ci−1 ∪ C∗ .
We put A∗ into the final sets, and B∗ becomes Di to split at the next stage.
No edges join Di to Ai or Bi , so any vertices of Di can be added to Bi or to Ai .
Since | Di | ≤ 32 | Di−1 | , the bounds on the contributions to | C| form a geometric
√ √
sum. When D¾ = ∅, we have | A¾ | , | B¾ | ≤ n
and | C¾ | ≤ ∑i 2 2n(2/3)i < 2 √2n
.
2 1− 2/3

For = 2/3, √ the coefficient in Theorem √ 9.2.18

√ was reduced successively in
five papers from 2 2 (about 2.828) to reach 2/3 + 4/3 (about 1.971) in Djidjev–
Venkatesan [1997]. For example, Alon–Seymour–Thomas [1994] applied √ their
technique for proving Theorem 9.2.18 more carefully to reduce it to 2 2 (about
3

2.121). Along the way were also Djidjev [1982, 1987] and Gazit–Miller
√ [1990].
Djidjev [1982] showed that the coefficient cannot be less than 23 ( 3)1/2 (about
1.555). Thus trees of separations of planar graphs down to pieces of constant size
need logarithmic depth.
Proofs of the Planar Separator Theorem produce separations in linear time.
Due to this, the Planar Separator Theorem yields fast approximation algorithms
for computational problems that are difficult on planar graphs. Exercise 23 con-
siders the example of finding a large independent set.

EXERCISES 9.2

9.2.1. (−) Prove that the complement of the 3-dimensional cube Q3 is nonplanar.
9.2.2. (−) Prove or disprove: The union of any two paths is a planar graph.
9.2.3. (−) For each graph below, prove nonplanarity or give a convex embedding.
• •
• • •
• •
•
• •
•
• •
• •
• • • •
• •

9.2.4. (−) Let M be a matching in G as given below. Determine whether G − M is planar.

(a) G = K4 ,4 and | M | = 4. c) G = K6 and | M | = 3.
(b) G = K4 ,4 and | M | = 3. d) G = K6 and | M | = 2.
9.2.5. (−) Give two proofs that the Petersen graph is nonplanar: one by using Kura-
towski’s Theorem, and one by using Euler ’s Formula.
9.2.6. (−) Prove or disprove: Every bipartite graph with minimum degree at least 4 con-
tains K3 ,3 as a subgraph.
398 Chapter 9: Planar Graphs

9.2.7. When G has degree list (4 , 4 , 4 , 4 , 3 , 3 , 3 , 3), also G has this degree list. Show that
such G and G can be both planar, both nonplanar, or one of each type.
9.2.8. (♦) For n ≥ 5, let G be the graph whose vertices are n points on a circle, with each
vertex adjacent to the four nearest others. Prove that G is planar if and only if n is even.
9.2.9. (♦) Fáry’s Theorem. Let R be a planar region bounded by a simple polygon with
at most five sides (the edges do not cross). Prove that some point x inside R “sees” all of
R, meaning that the segment from x to any point of R does not cross the boundary of R.
Conclude inductively that every planar graph has a straight-line embedding.
9.2.10. Find a convex embedding in the plane for the graph below.
• •

• • •
• •
• • •

• •
9.2.11. (♦) Outerplanarity. (Chartrand–Harary [1967])
(a) Use Kuratowski’s Theorem to prove that G is outerplanar if and only if G does
not contain a subdivision of K4 or K 2 ,3 . (Hint: To apply Kuratowski’s Theorem, modify G
appropriately. This is much easier than following the steps of Kuratowski’s Theorem.)
(b) Use part (a) to prove that G K 2 is planar if and only if G is outerplanar.
9.2.12. For G1 and G 2 connected, with | V(G i)| ≥ 3, prove that G1 G 2 is planar if and
only if one is a path and the other is a cycle or a path. (Behzad–Mahmoodian [1969])
9.2.13. Wagner [1937] proved that a graph G is planar if and only if neither K 5 nor K3 ,3
can be obtained from G by deletions and contractions of edges.
(a) Show that deletion and contraction of edges preserve planarity. Conclude from
this that Wagner ’s condition is necessary.
(b) Use Kuratowski’s Theorem to prove that Wagner ’s condition is sufficient.
9.2.14. (♦) The Hanani–Tutte Theorem: a planarity criterion.
(a) Prove that in every drawing of K 5 or K3 ,3 in the plane, some two nonincident edges
cross an odd number of times. (Hint: Prove that the number of such pairs of edges is al-
ways odd.) (Hanani [1934], as Chojnacki)
(b) Prove that a graph is planar if it has a drawing in the plane in which every two
nonincident edges cross an even number of times. (Tutte [1970])
9.2.15. (♦) Prove that every 3-connected graph with at least six vertices that contains a
subdivision of K 5 also contains a subdivision of K3 ,3 .
9.2.16. (♦) For a cycle C in a graph G, a C-fragment is a component of G − V(C) together
with its edges and vertices of attachment to C. Two C-fragments conflict if they have
three common vertices of attachment or four alternating vertices of attachment on C. The
conflict graph of C has a vertex for each C-fragment , adjacent when they conflict. Prove
that G is planar if and only if the conflict graph of each cycle is bipartite. (Tutte [1958])
9.2.17. Let x and y be vertices in a planar graph G. Prove that G has a planar embedding
with x and y on the same face if and only if G − x − y has no cycle C with x and y in
conflicting C-fragments in G. (Hint: Use Kuratowski ’s Theorem. Comment: Tutte proved
this without Kuratowski’s Theorem and used it to prove Kuratowski’s Theorem.)
Section 9.3: Coloring of Planar Graphs 399

9.2.18. Let C be a cycle in a 3-connected plane graph G. Prove that C is a face-boundary

in G if and only if G has exactly one C-fragment. (Comment: Tutte [1963] used this to
prove the result of Whitney [1933] that 3-connected planar graphs have only one planar
embedding.)
9.2.19. (+) Let C be a cycle in a graph G, and let H1 and H2 be V(C)-lobes. Say that
H1 and H2 are skew if C has vertices v1 , v2 , v3 , v4 in order such that v1 , v3 ∈ V(H1) and
v2 , v4 ∈ V(H2). The skew-overlap graph of C has a vertex for each C-component and an
edge joining every skew pair. Use Kuratowski’s Theorem (without Kelman’s Conjecture)
to prove that G is nonplanar if and only if G contains a cycle whose skew-overlap graph
has a cycle of length 3 or 5. (Hint: Use induction on the number of edges.)
√
9.2.20. Solve the recurrence º (n) = 2 º (n/2) + a n, where n is a power of 2, with º (1) = 0.
9.2.21. (♦) Separator theorems.
(a) Prove that every tree has a (1 , 2/3)-separation.
(b) Prove that every outerplanar graph has a (2 , 2/3)-separation.
9.2.22. The naive algorithm to check for K4 as a subgraph of an n-vertex graph runs in
time O(n4). Use the planar separator theorem to design an algorithm that solves this prob-
lem in time O(n log n) on planar graphs. Assume that the separation guaranteed by the
separator theorem can be found in linear time.
9.2.23. Let be a function of n satisfying (log log n) ≤ ≤ O(log n). Prove that there is
¾
√ graphs that runs in time O(n2 ) and produces√an indepen-
an algorithm on n-vertex planar
dent set of size within O(n/ ) of the maximum. (Hint: Find a set C of O(n/ ) vertices
such that every component of G − C has at most vertices.) (Lipton–Tarjan [1980])
9.2.24. (+) Lipton–Tarjan [1979] proof of Theorem 9.2.18. Let G be a planar graph with
nonnegative weight c(v) at each vertex v, with ∑ c(v) ≤ 1. An (m , )-separation is a parti-
tion {A , B , C} of V(G) where [ A , B] = ∅ , with c(A) , c(B) ≤  c(V(G)) and | C| ≤ m.
(a) Prove that a planar graph with a spanning tree of diameter 2r has a (2r + 1 , 32 )-
separation, with | C| = 2r + 1 in such a separation only if C contains the center of T .
(b) Given a rooted tree T , let T l denote the set of vertices at distance l from the root.
Let T be a tree grown from vertex x in a planar graph G so that v ∈ T l if and only if
d G(x , v) = l. Let r and s be integers with r < s such that ∑i<r c(Ti) ≤ 32 and ∑i> s c(Ti) ≤ 23 .
Prove that G has an (m , 23 )-separation with m ≤ | Tr | + | T s | + 2(s − r − 1).
√
(c) Prove that every weighted n-vertex planar graph has a ( 8n , 23 )-separation.

9.3. Coloring of Planar Graphs

We come now to the notorious Four Color Problem: can the regions of any map
be colored from four colors so that adjacent regions have different colors? Equiva-
lently, is every planar graph 4-colorable? The question was in a letter of October
23, 1852 from Augustus de Morgan to Sir William Hamilton. It was asked by
de Morgan’s student Frederick Guthrie, who heard it from his brother Francis.
Cayley presented the problem to the London Mathematical Society in 1878.
Its notoriety stems from its ease of statement (especially the map-coloring form)
and from many published faulty proofs. The history is discussed in Ore [1967],
Saaty–K ainen [1977, 1986], Aigner [1984, 1987], and Fritsch–Fritsch [1998].
Kempe [1879] published the first “solution”, refuted by Heawood [1890]. Tait
[1880] published another; we consider both arguments.
400 Chapter 9: Planar Graphs

EDGE-COLORINGS AND SPANNING CYCLES

Tait reduced the Four Color Problem to edge-coloring of 3-regular planar

graphs. It suffices to prove that maximal planar graphs are 4-colorable. It there-
fore suffices to prove that their duals are 4-face-colorable. The dual of a maximal
planar graph is 3-regular and 2-edge-connected.

9.3.1. Theorem. (Tait’s Theorem; Tait [1880]) A 2-edge-connected 3-regular

plane graph is 4-face-colorable if and only if it is 3-edge-colorable.
Proof: Let G be such a graph. Suppose first that G is 4-face-colorable using four
colors named as follows: c0 = 00, c1 = 01, c2 = 10, c3 = 11. We define an edge-
coloring. To the edge between faces with colors ci and cj we assign the color that
is the binary coordinate sum of ci and cj . Since G is 2-edge-connected, ci
= cj , and
hence 00 is not obtained. If two edges incident to v receive the same color c, then
subtracting the color on their common face from c leaves adjacent faces with the
same color, a contradiction. Hence we have produced a proper 3-edge-coloring.

• • • • •
01
11 10 • • • 01
11 00 10
01 10 01 11 • • • • 00 • 11 •
• • 01
10 01 01 11 • 01 •
• 11 • • 00 • c b
a
00
• •

From a proper 3-edge-coloring of G using colors a , b , c on the subgraphs

Ea , Eb , Ec , we construct a 4-face-coloring using the four colors as above. Since G is
3-regular, each color appears at every vertex, and the union of two of {Ea , Eb , Ec}
is a 2-factor of G. Each face in such a subgraph is a union of faces of G. Let
H1 = Ea ∪ Eb and H2 = Eb ∪ Ec . To a face º , assign the color whose ith coordinate
is the parity of the number of cycles in Hi that contain º (0 for even, 1 for odd).
We show that the resulting 4-face-coloring is proper. The faces º and º
bounded by an edge e are distinct, since G is 2-edge-connected. Suppose that e
lies on a cycle C in Hi . By the Jordan Curve Theorem, C separates º and º , and
no other cycle in Hi separates º and º . Hence the numbers of cycles enclosing º
and º in Hi have different parity, and the colors we have given to º and º differ
in coordinate i (in both coordinates when e has color b).

Due to Theorem 9.3.1, a proper 3-edge-coloring of a 3-regular graph is called a

Tait coloring. Proving that all 2-edge-connected 3-regular planar graphs have
Tait colorings reduces to the 3-connected case (Exercise 9), and every Hamilto-
nian 3-regular graph has a Tait coloring. Thus the Four Color Theorem follows
if every 3-regular 3-connected planar graph is Hamiltonian. It was noticed that
Tait didn’t prove that assumption, but no counterexample was found until 1946.
Grinberg [1968] found a simple necessary condition for planar graphs to
be Hamiltonian. It yields many 3-regular 3-connected non-Hamiltonian planar
graphs, including Tutte ’s 1946 example and the Grinberg graph of Exercise 13.
Section 9.3: Coloring of Planar Graphs 401

9.3.2. Theorem. (Grinberg [1968]) If G is a plane multigraph with a Hamilto-

nian cycle C, and G has
i faces of length i inside C and
i faces of length i
outside C, then ∑i(i − 2)(
i −
i) = 0.
Proof: It suffices to prove ∑i(i − 2)
i = n − 2, since the same argument applies to
the regions outside C. Consider the outerplanar graph G formed by C and the
chords inside it, having n vertices, m edges, and r faces.
Note that ∑i i
i counts the n edges of C once and chords twice, while ∑i 2
i
counts bounded faces twice. Using Euler ’s Formula,

∑(i − 2)
i = (2m − n) − 2(r − 1) = 2 − 2(n − m + r) + n = n − 2.

i

9.3.3. Example. The Tutte graph. Tutte [1946] found the 3-connected 3-regu-
lar non-Hamiltonian plane graph on the left below. Let H denote each component
obtained by deleting the central vertex and the three long edges. Since a Hamil-
tonian cycle must visit the central vertex, it must traverse one copy of H along a
Hamiltonian path joining the other entrances to H , which we call x and y.
We therefore study a graph that has a Hamiltonian cycle if and only if H has
a spanning x , y-path. Such a graph H (on the right below) is obtained by adding
an x , y-path of length 2 through a new vertex.
• • • •
• • •• • 4
••• • • •
••• 5 9
• 5 • • •
G • • • • 5 4 H
• •y •
• •
• • •• •• • • • •
•• • • •• •
• •• ••
5 5
•H
• • x• • •y
4
• •x •
A face of length j is a j-face. Considering the faces of each length, Grin-
berg ’s condition for H becomes 2a4 + 3a5 + 7a9 = 0, where ai =
i −
i . Since the
unbounded face is outside, the equation reduces mod 3 to 2a4 ≡ 7 (mod 3). Since

4 +
4 = 3, we have a4 ∈ {±3 , ±1}. The only choice satisfying 2a4 ≡ 7 (mod 3)
is a4 = −1, which requires that two 4-faces lie outside the Hamiltonian cycle.
However, those having a vertex of degree 2 cannot lie outside the cycle.
A faster contradiction arises by subdividing one edge incident to each vertex
of degree 2. This does not change the existence of a spanning cycle. The result-
ing graph has seven 5-faces, one 4-face, and one 11-face. The required equation
becomes 2 · (±1) = 9 − 3a5 , but the left side is not a multiple of 3.

Tutte [1956] proved that all 4-connected planar graphs are Hamiltonian.
Thomassen [1983] proved the stronger result (conjectured by Plummer [1975])
that every 4-connected planar graph is Hamiltonian-connected (every vertex pair
occurs as the endpoints of a spanning path). By proving that 4-connected trian-
gulations are Hamiltonian, Whitney [1931] reduced the Four Color Problem to
the case of Hamiltonian planar graphs (Exercise 10).
402 Chapter 9: Planar Graphs

5-COLORABLE AND 5-CHOOSABLE

We proved from Euler ’s Formula that planar n-vertex graphs have at most
3n − 6 edges (Theorem 9.1.23). Thus planar graphs are 5-degenerate, which im-
plies that they are 6-colorable (recall Proposition 8.1.12). Heawood improved the
upper bound from 6 to 5.

9.3.4. Theorem. (Five Color Theorem; Heawood [1890]) Every planar graph
G is 5-colorable.
Proof: Since subgraphs of G are planar, it suffices to forbid 6-critical planar
graphs. If G is 6-critical, then (G) ≥ 5, so with planarity we may let v be a
vertex of degree 5. Let º be a proper 5-coloring of G − v. All five colors must ap-
pear on N(v). Let v1 , v2 , v3 , v4 , v5 be the neighbors of v in clockwise order around
v, and name the colors so that º (vi) = i.
Let G i , j be the subgraph of G − v induced by the vertices of colors i and j .
Switching the colors on a component of G i , j yields another proper 5-coloring of
G − v. If the component of G i , j containing vi does not contain vj , then switch-
ing the colors on it removes color i from N(v). Now giving color i to v produces a
proper 5-coloring of G. Thus G is 5-colorable unless, for each choice of i and j , the
component of G i , j containing vi also contains vj . Let Pi , j be a path in G i , j from vi
to vj , shown below for (i , j) = (1 , 3).
3 1
1
• • •
4 2
• •
5• •
• 2 •4 •3
v
•
2
• • •1
4 3
The cycle C completed with P1 ,3 by v separates v2 from v4 . By the Jordan
Curve Theorem, the path P2 ,4 crosses C. Since G is planar, paths cross only at
shared vertices. The vertices of P1 ,3 all have color 1 or 3, and those of P2 ,4 all
have color 2 or 4, so they have no common vertex.
This contradiction implies that every planar graph is 5-colorable.

After 86 more years, the upper bound was lowered from 5 to 4 in Appel–
Haken [1976]. Before we explain the approach to the proof, we describe another
way to strengthen the Five Color Theorem.
Vizing asked in 1975 whether every planar graph is 5-choosable; we prove
this below. Voigt [1993] found a planar graph with 238 vertices that is not 4-
choosable; Mirzakhani [1996] (Exercise 35) reduced this to 63 vertices (both exam-
ples generalize to infinite families). Gutner [1996] (Exercise 34) and Voigt–Wirth
[1997] showed that in fact some 3-colorable planar graphs are not 4-choosable.
As is common in inductive arguments about plane graphs, the external ver-
tices (vertices of the unbounded face) play a special role. Recall that a near-
triangulation is a 2-connected plane graph in which every bounded face is a
triangle. The boundary of the outer face is a cycle.
Section 9.3: Coloring of Planar Graphs 403

9.3.5. Theorem. (Thomassen [1994b]) Every planar graph G is 5-choosable.

Proof: Adding edges cannot reduce Ò l , so we may assume that every bounded face
is a triangle. By induction on | V(G)| , we prove the stronger result that G has an
L-coloring even when two adjacent external vertices have distinct lists of size 1
and all other external vertices have lists of size 3. The basis step is a triangle; a
color remains available for the third vertex. For | V(G)| > 3, let the external cycle
C be [v1 , . . . , vp], with consecutive vertices vp and v1 having fixed colors.
Case 1: C has a chord vi vj with 1 ≤ i < j − 1 < p. We apply the induction
hypothesis to the subgraph induced by the cycle [v1 , . . . , vi , v j , . . . , vp] and the
vertices inside it. This selects a proper coloring giving vi and vj some fixed col-
ors. Now we apply the induction hypothesis to the subgraph induced by the cycle
[vi , vi+1 , . . . , v j ] and the vertices inside it to complete the L-coloring of G.
Case 2: C has no chord. Let v1 , u1 , . . . , u¾ , v3 be the neighbors of v2 in order.
Since bounded faces are triangles, ⟨v1 , u1 , . . . , u¾ , v3 ⟩ is a path. Since C is chord-
less, u1 , . . . , u¾ are internal vertices, and the outer face of G − v2 is bounded by
a cycle C in which ⟨v1 , u1 , . . . , u¾ , v3 ⟩ replaces v1 , v2 , v3 . Let G = G − v2 .
Let c be the color assigned to v1 . Since | L(v2)| ≥ 3, there are distinct colors
x , y ∈ L(v2) − {c}. We reserve x and y for v2 by letting L (ui) = L(ui) − {x , y}. Since
| L(ui)| ≥ 5, we have | L (ui)| ≥ 3. Hence we can apply the induction hypothesis to
G , with u1 , . . . , u¾ having lists of size at least 3 and other lists being the same
as in G. Note that v1 and u1 , . . . , u¾ receive colors outside {x , y}. We extend the
coloring to G by choosing for v2 a color in {x , y} not used on v3 .
vi v2
v1 • v1 • v
• • •3
vp • vp • • • •
u1 u¾

•
vj
Case 1 Case 2

Building on this method, Grytczuk–Zhu [2020] proved that every planar

graph contains a matching whose deletion leaves a 4-choosable graph (Theorem
15.1.45). Bipartite planar graphs are 3-choosable (Exercise 15.1.32). Thomassen
[1995a] proved that also planar graphs with girth at least 5 are 3-choosable.
To better understand choosability, a more refined notion was introduced.

9.3.6.* Definition. (Kratochvı́l–Tuza–Voigt [1998]) Graph G is (¾ ,d)-choosable

if G is L-colorable for every list assignment L such that | L(v)| ≥ ¾ for all
v ∈ V(G) and | L(x) ∩ L(y)| ≤ d for all xy ∈ E(G).

Note that (¾ , ¾)-choosable means ¾-choosable, and for ¾ ≥ 1 every graph is

(¾ , 0)-choosable. Thus when 1 ≤ ¾ < Ò l(G) there is a threshold d such that G
is (¾ , d)-choosable but not (¾ , d + 1)-choosable. It is unknown whether all planar
graphs are (4 , 2)-choosable; they are not all (4 , 3)-choosable (Exercises 34–35).

9.3.7.* Theorem. (Kratochvı́l–Tuza–Voigt) Planar graphs are (4 , 1)-choosable.

404 Chapter 9: Planar Graphs

Proof: If | E(H)| ≤ d|V(H)| for every subgraph H of a graph G, where d ∈ , 

then G has an orientation where every vertex has outdegree at most d (Corollary
6.1.5, applying Hall’s Theorem). A planar n-vertex graph G has at most 3n − 6
edges, so G has an orientation D with maximum outdegree at most 3.
Delete from each list L(v) any color occurring in the list of an outneighbor of v
in D. This deletes at most one color for each outneighbor, so a color remains. Now
adjacent vertices have disjoint lists, so any coloring chosen from them is proper.

THE FOUR COLOR PROBLEM

In proving the Five Color Theorem, we argued that a smallest non-5-colorable

planar graph cannot contain a vertex of degree at most 5. This suggests an ap-
proach to the Four Color Problem. We need only consider triangulations, since
every subgraph of a 4-colorable graph is 4-colorable.
9.3.8. Definition. In a graph, a configuration is a substructure, often a type
of subgraph. For the Four Color Problem, a configuration in a triangulation
is a separating cycle C (called the ring) plus the portion of the graph inside
C. A set S of configurations is unavoidable for a problem if every instance
contains a member of S. A configuration is reducible for a property if it
cannot occur in a minimal structure lacking that property.
Thus the goal is to find an unavoidable set of reducible configurations.
For the Four Color Problem, this means showing that every triangulation con-
tains a configuration that is reducible for the property of being 4-colorable.
9.3.9. Example. An unavoidable set. We have remarked that (G) ≤ 5 for every
simple planar graph. In a triangulation with at least four vertices, every vertex
has degree at least 3. Thus the set of configurations below is unavoidable.

• • • •
• • • •
•
• • • • • •
The edges inward from the ring are dashed because a configuration in a trian-
gulation is determined by the interior subgraph and the degrees of its vertices
(Exercise 37). Thus we write these as “ • 3”, “ • 4”, and “ • 5”, respectively.
Saying that a configuration C is reducible for 4-colorability of triangulations
means that if C appears in G, then G can be changed to a smaller triangulation
G such that every proper 4-coloring of G leads to a proper 4-coloring of G.
The idea Heawood used in Theorem 9.3.4 came from Kempe ’s failed proof and
is important in proving reducibility of configurations. A path whose colors alter-
nate between two colors is a Kempe chain.
9.3.10. Remark. Kempe’s proof. Let us try to prove the Four Color Theorem
by induction using the unavoidable set {• 3 , • 4 , • 5}. The approach is similar to
Theorem 9.3.4. We can extend a 4-coloring of G − v to complete a 4-coloring of G
unless all four colors appear on N(v). Thus “ • 3” is reducible. If d(v) = 4, then
the Kempe-chain argument works as in Theorem 9.3.4, and “ • 4” is reducible.
Section 9.3: Coloring of Planar Graphs 405

Now consider “ • 5”. When d(v) = 5, the repeated color on N(v) in a proper
4-coloring of G − v appears on nonconsecutive neighbors of v, since G is a triangu-
lation. Let v1 , v2 , v3 , v4 , v5 be the neighbors of v in order. In the 4-coloring º of
G − v, we may assume by symmetry that º (v5) = 2 and that º (vi) = i for 1 ≤ i ≤ 4.
Define G i , j and Pi , j as in Theorem 9.3.4. We can eliminate color 1 from N(v)
unless chains P1 ,3 and P1 ,4 exist from v1 to v3 and v4 , respectively, as on the left
below. The component H of G 2 ,4 containing v2 is separated from v4 and v5 by the
cycle through P1 ,3 and v. The component H of G 2 ,3 containing v5 is separated
from v2 and v3 by the cycle through P1 ,4 and v. Switching colors 2 and 4 in H and
colors 2 and 3 in H eliminates color 2 from N(v). This was Kempe ’s final case.
1
4• • •3
4 3
• • P1 ,4 3• •4 P1 ,3
P1 ,4 1 P1 ,3 1
• •
1• 2• •2 •1 1• 2• •2 •1
• •
v v
• • • •
4 3 4 3

Unfortunately, P1 ,3 and P1 ,4 can intertwine, crossing at a vertex with color

1 as on the right above. Performing the switches in both H and H now produces
a pair of adjacent vertices with color 2. The resulting coloring is not proper.

To replace “ • 5”, we need to find other reducible configurations.

9.3.11.* Example. All configurations with ring size 3 or 4 are reducible (Exercise
38). Equivalently, no minimal 5-chromatic triangulation has a separating cycle
of length at most 4. Birkhoff [1913] proved that configurations with ring size 5
other than • 5 are reducible.
Birkhoff also proved reducibility of the configuration below with ring size 6,
called the Birkhoff diamond. One must consider all proper 4-colorings of the
ring. Some cases extend to proper colorings of the interior. In others, Kempe
chains are needed to change the coloring into one that extends (Exercise 39).
• • 5
• •
• • • • = 5• •5
• •
5
• •
We have barely begun; an enormous amount of detail remains. From 1913 to
1950, enough reducible configurations were found to prove that all planar graphs
with at most 36 vertices are 4-colorable. This was slow progress. In the 1960s,
Heesch focused attention on the size of the ring, gave heuristics for finding re-
ducible configurations, and developed methods for generating unavoidable sets.
Appel and Haken, working with Koch, improved the heuristics to consider
only “promising ” configurations. Using 1000 hours of computer time on three
406 Chapter 9: Planar Graphs

computers in 1976, they found an unavoidable set of 1936 reducible configura-

tions with ring size at most 14. A ring of size 13 already has 66430 distinguish-
able 4-colorings. Proving reducibility requires showing that each leads to a proper
4-coloring of the full graph. Kempe chains and other arguments were used.

9.3.12. Theorem. (Four Color Theorem; Appel–Haken–Koch [1977]) Every

planar graph is 4-colorable.

By 1983, refinements led to an unavoidable set of 1258 reducible configura-

tions. Robertson–Sanders–Seymour–Thomas [1997] revisited the proof. Their
simplifications yielded an unavoidable set of 633 reducible configurations. They
posted their code on the Internet; in 1997, it would prove the Four Color Theorem
on a desktop workstation in about three hours. Now it takes just minutes.
We have explained the meaning of reducibility, but how to generate the un-
avoidable set? Concise terminology for vertices by degrees is very helpful.

9.3.13. Definition. A j-vertex, j + -vertex, or j − -vertex is a vertex with degree

equal to j , at least j , or at most j , respectively. Similarly, a j-neighbor of v
is a j-vertex adjacent to v. Analogously, j-face, j + -face, and j − -face refer to
the face length, which is the degree in the dual. Let d(G) denote the average
of the vertex degrees in G.

Every planar graph has a 5−-vertex, but a 5-vertex is not reducible for 4-
colorability. We need configurations with small degree on more vertices, such as
in Example 9.3.11. Wernicke [1904] guaranteed a 5-vertex having at least one
6−-neighbor. We prove a stronger version to introduce “discharging ”.

9.3.14. Proposition. (Franklin [1922]) In every planar graph G with minimum

degree 5, some 5-vertex has at least two 6−-neighbors.
Proof: We may assume that G is a triangulation, since added edges only increase
degrees. Give each vertex initial “charge” equal to its degree. If no 5-vertex has at
least two 6−-neighbors, then we modify the charges so that each vertex ends with
charge at least 6, without changing the total charge. This contradicts d(G) < 6,
which implies that the desired configuration must occur at some 5-vertex.
Each 5-vertex needs charge; let it take 1/4 from each 7+-neighbor. When
a 5-vertex does not have two 6−-neighbors, it receives charge from at least four
neighbors and reaches charge 6. A 6-vertex neither gains nor loses charge.
A 7+-vertex loses charge 1/4 to each 5-neighbor. Therefore an 8+-vertex keeps
enough charge, since j − j/4 ≥ 6 when j ≥ 8. Hence it remains only to show
that every 7-vertex ends with charge at least 6. To fall below charge 6, a 7-vertex
v must give 1/4 to at least five 5-neighbors. Since G is a triangulation, G has
a cycle through N(v). With at least five 5-vertices, this 7-cycle must have three
consecutive 5-vertices, yielding a 5-vertex with two 5-neighbors. Hence when the
desired configuration does not occur, every vertex ends with charge at least 6.

The condition d(G) < 6 alone does not guarantee the conclusion. The graph
K 5 ,7 has no 5-vertex with a 6−-neighbor, despite (K 5 ,7) = 5 and d(K 5 ,7) < 6. We
needed planarity to find a cycle through the neighbors of a 7-vertex.
Section 9.3: Coloring of Planar Graphs 407

DISCHARGING AND LIGHT EDGES

Discharging arguments yield sparse local structures from global sparseness

conditions. For example, d(G) < b guarantees some vertex with degree less than
b. Adding planarity or stricter bounds on average degree can guarantee more,
as in Proposition 9.3.14. Cranston–West [2017] discussed discharging via the ap-
proach here, Borodin [2013] surveyed applications to coloring planar graphs, and
Jendrol’–Voss [2013] discussed guaranteeing subgraphs with small degree-sum.
In order to explain the technique, we first assume only a bound on d(G).
Most discharging arguments are motivated by applications to coloring or de-
composition problems. We want the configurations forced by discharging to be
reducible for the desired property. Thus discharging applications often start by
finding reducible configurations and later study what sort of sparseness is needed
force their appearance. Combining the reducibility arguments and the discharg-
ing arguments to give an inductive proof is called the Discharging Method.
To illustrate the method, consider edge-coloring. Just as a vertex with degree
less than ¾ is reducible for ¾-coloring, an edge incident to fewer than ¾ others is
reducible for ¾-edge-coloring. In fact, since we will color this edge after coloring
the rest of the graph, the configuration is also reducible for ¾-edge-choosability.

9.3.15. Definition. When e = uv ∈ E(G), the weight of e is d G(u) + d G(v). More

generally, the weight of a subgraph H in a graph G is ∑v∈V(H) d G(v).

9.3.16. Lemma. A minimal graph G such that Ò (G) > ¾ has no edge of weight at
most ¾ + 1. That is, an edge of weight at most ¾ + 1 is reducible for Ò (G) ≤ ¾ .
Proof: Let L assign lists of size ¾ to the edges of G. If uv is an edge of weight
at most ¾ + 1, then let G = G − uv. If Ò (G ) ≤ ¾ , then G has an L-coloring á .
Since uv is incident to at most d G(u) + d G(v) − 2 edges, some color in L(uv) is not
used by á on any edge incident to uv. Using it on uv extends the coloring to G.

Always Ò l(G) ≥ (G). For equality, we seek a sparseness condition that guar-
antees occurrence of our reducible configuration. We measure sparseness by small
density, where density is usually defined to be | E(G)| / | V(G)|. Equivalently, and
more pertinently for discharging, we consider graphs with small average degree.

9.3.17. Definition. Degree charging assigns to each vertex in a graph an ini-

tial charge equal to its degree. Discharging rules modify charges assigned
to elements of a graph in ways that do not change the total charge.

9.3.18. Remark. Degree charging and average degree. We want to show that some
set of configurations is unavoidable when d(G) is less than some value b. Discharg-
ing brings extra charge to vertices with initial charge less than b. Vertices with
higher initial charge can give charge but must keep at least b. As in Proposi-
tion 9.3.14, the goal is to show that if no desired configuration occurs, then every
vertex ends with charge at least b, which requires d(G) ≥ b and violates the hy-
pothesis. The argument proves the contrapositive of the structural claim.
An edge whose weight is at most some desired value is a light edge. More
generally, we can speak of light triangles, etc.
408 Chapter 9: Planar Graphs

9.3.19. Lemma. If (G) ≥ 2 and d(G) < ¾4+¾2 , then G has an edge of weight at most
+ 1 (a light edge). Furthermore, d(G) ≤ ¾4+¾2 is not sufficient. (In particular,
for ∈ {3 , 4 , 5 , 6}, respectively, d(G) < b suffices, where b ∈ { 12 8 20
5 , 3 , 7 , 3}.)

Proof: The conditions require ≥ 3, since d(G) ≥ (G). We show that if G has
no light edge, then d(G) ≥ b, where b = ¾4+¾2 . We use degree charging and move
charge so that every vertex ends with charge at least b. Here we give the proof
for 3 ≤ ≤ 6, leaving > 6 to Exercise 43.
When ≤ 6, we have b ≤ 3, so only 2-vertices need charge (when > 6, also 3-
vertices need charge). A 2-vertex v needs b− 2; let it take half from each neighbor.
Since G has no light edge, no neighbor of v needs charge, so v ends with charge b.
We also must show that all vertices still have charge at least b. With no
light edges, a 2-vertex has only +-neighbors, so only +-vertices lose charge. A
j-vertex loses b−2 2 to each 2-neighbor. Hence its final charge is at least j − j b−2 2 .
With j ≥ , it suffices to show − b−2 2 ≥ b, which holds when b ≤ ¾4+¾2 .
To prove sharpness, we construct a graph having average degree ¾4+¾2 and no
light edge. Form G by subdividing each edge in a -regular multigraph H. If
n = | V(H)| , then G has n + n/2 vertices and n edges (twice as many as H), for
average degree exactly ¾4+¾2 . However, every edge has weight + 2.

In applying Lemma 9.3.19 to prove -edge-choosability, a light edge e allows

an L-coloring  of E(G − e) to extend, if  exists. Thus an inductive proof needs
the same sparseness condition on subgraphs. This suggests a natural parameter.

9.3.20. Definition. The maximum average degree of a graph G, written

| E(H)|
Mad(G), is max H ⊆ G d(H); it equals max H ⊆ G 2|V(H) | .

4 (G)
9.3.21. Theorem. If Mad(G) < (G)+2 , then  l(G) = (G).
Proof: Always  l(G) ≥ (G); we prove the upper bound. Deleting edges can re-
duce the degree, so to permit an inductive proof we prove more generally that
Mad(G) < ¾4+¾2 and (G) ≤ together imply  l(G) ≤ . Since every subgraph
of a graph satisfying these hypotheses also satisfies them, it suffices to show that
every graph in this family contains a configuration that is reducible for  l(G) ≤ .
We may ignore isolated vertices. Since (G) ≤ , the edge at a 1-vertex has
weight at most + 1. If (G) ≥ 2, then Lemma 9.3.19 yields an edge of weight at
most + 1. By Lemma 9.3.16, this configuration is reducible for  l(G) ≤ .

9.3.22.* Remark. Lemma 9.3.19 is easy to prove when ≤ 6: only 2-vertices

need charge, and they take what they need from their neighbors. With light
edges forbidden, -vertices are the most in danger of losing too much charge.
When 2-vertices take b−2 2 from each neighbor to reach b, the final charge will be
at least b at each vertex if and only if − b−2 2 ≥ b. Hence the proof works if and
only if b ≤ ¾4+¾2 . We find the statement of the lemma after we have the proof!
The bound on d(G) in the structural result of Lemma 9.3.19 is sharp, but
that does not make the coloring application sharp. Our sharpness examples were
bipartite, so by Galvin’s Theorem (Theorem 8.3.18) they satisfy  l(G) = (G).
We have not given graphs with average degree ¾4+¾2 and  l(G) > (G) = .
Section 9.3: Coloring of Planar Graphs 409

9.3.23. Remark. Mad(G) and girth of planar graphs. By Euler ’s Formula, an n-

vertex plane graph G with all faces of length at least ½ has at most ½ −½ 2 (n − 2)
edges (Exercise 9.1.45). This holds also for subgraphs, so Mad(G) < ½2−½2 .
We often get stronger results for planar graphs than for general graphs. If
2½ 4¾
½ −2 ≤ ¾ +2 , then Lemma 9.3.19 guarantees a light edge (weight at most ¾ + 1)
in a planar graph of girth ½ with minimum degree 2. However, planar graphs
with smaller girth, which allow larger average degree, may also force such a light
edge. For various ¾ , below we list the bound b on average degree that guarantees
a light edge and often find a larger bound on average degree for planar graphs
with at least the specified girth where the light edge is also guaranteed.
2½
¾ weight b = ¾4+¾2 girth ½ ½ −2 reference
6 7 3 5 10/3 Lemma 9.3.27
5 6 20/7 no gain Exercise 8
4 5 8/3 7 14/5 Lemma 9.3.28
3 4 12/5 ? ? Exercise 44

This brings us to the special role of discharging on planar graphs. The Dis-
charging Method is particularly effective for planar graphs because the dual is
also planar and hence also has bounded average degree. Thus we may also use
charge on faces. Euler ’s Formula yields several ways to exploit this interaction.

9.3.24. Proposition. The following formulas hold for a connected plane graph
G, where F(G) denotes the set of faces and l(º ) is the length of a face º .

∑ (d(v) − 6) + ∑ (2l(º ) − 6) = −12

v∈ V(G) º ∈ F(G)

∑ (2d(v) − 6) + ∑ (l(º ) − 6) = −12

v∈ V(G) º ∈ F(G)

∑ (d(v) − 4) + ∑ (l(º ) − 4) = −8
v∈ V(G) º ∈ F(G)

Proof: Multiply Euler ’s Formula by 6 or 4 and split the contribution from the
number of edges, obtaining the three formulas below.
6n − 2m − 4m + 6p = 12; 6n − 4m − 2m + 6p = 12; 4n − 2m − 2m + 4p = 8.
Multiply each equation by −1 and substitute 12 ∑v∈V(G) d(v) for the first occur-
rence of m in each and 12 ∑º ∈ F(G) l(º ) for the second. Collecting the contributions
by vertices and by faces completes the proofs.

9.3.25. Definition. Three ways to assign charge on planar graphs.

charge on v ∈ V(G) charge on º ∈ F(G)
vertex charging d(v) − 6 2l(º ) − 6
face charging 2d(v) − 6 l(º ) − 6
balanced charging d(v) − 4 l(º ) − 4

9.3.26. Remark. For triangulations, vertex charging puts charge 0 on all faces,
just a translation of degree charging. Dually, face charging puts charge 0 on all
410 Chapter 9: Planar Graphs

vertices in a 3-regular plane graph and can be useful for plane graphs with large
girth. Balanced charging treats vertices and faces symmetrically.
In each case, the total charge is negative. When a hypothesis allows moving
charge so that every vertex and face ends with nonnegative charge, the hypothesis
must be false. For example, there is no 4-regular bipartite plane graph, since un-
der balanced charging every vertex and face would start with nonnegative charge.
In discharging arguments, we call an element “happy” when it has enough
charge. For degree charging, usually “enough charge” means at least the bound
on the average degree. For the methods on planar graphs in Definition 9.3.25, it
means nonnegative charge. When every vertex and face is happy, the contradic-
tion disproves the assumption that permits the discharging.

Next we illustrate balanced charging and face charging by proving two of the
results mentioned in Remark 9.3.23. Later we illustrate vertex charging.

9.3.27. Lemma. Every planar graph G with girth at least 5 and (G) ≥ 2 has a
2-vertex with a 5− neighbor or an edge joining two 3-vertices.
Proof: Suppose that G has no such light edge. Use balanced charging (put
charge d(v) − 4 on each vertex v and l( ) − 4 on each face ; all faces have excess
charge). The total charge is −8. We obtain a contradiction by showing that each
element ends happy (nonnegative charge) under the following discharging rule:
(R1) Each vertex v takes 4−d(v)
d(v)
from each incident face.

The amount taken is negative when d(v) ≥ 5, which means that such vertices
give charge to faces. The rule immediately makes each vertex happy; they all end
with charge 0, having taken or sent charge equally from or to the incident faces.
It remains only to check that each face ends happy. Since light edges are
forbidden, a j-face loses charge to at most ⌊ 2j ⌋ vertices. Charge is lost only to
3−-vertices; a 3-vertex takes only 13 , but a 2-vertex takes 1.
When a j-face loses charge to a 2-vertex, the next vertex on the face is a 6+-
vertex since G has no light edge. The net loss to these two vertices is only 23 . For
a 3-vertex and its successor, which is a 4+-vertex, the loss is only 13 .
Hence a j-face ends with charge at least j − 4 − 23 ⌊ 2j ⌋ . This is nonnegative
when j ≥ 6. When j = 5 it is − 13 , but a 5-face with two incident 2-vertices has
three incident 6+-vertices, which provides the extra 13 to make it happy.

In principle, any method in Proposition 9.3.24 (or others) can be used in a dis-
charging argument on plane graphs, but one method may require more movement
of charge (and work) than another.

9.3.28. Lemma. Every planar graph G with girth at least 7 and (G) ≥ 2 has an
edge of weight at most 5.
Proof: Assume that G has no light edge. Use face charging (put initial charge
2d(v) − 6 on each vertex v and l( ) − 6 on each face ). Since G has girth at least
7, the only objects with negative initial charge are 2-vertices. Discharging:
(R1) Each 2-vertex takes 1
2 from each neighboring vertex and each incident face.
Section 9.3: Coloring of Planar Graphs 411

To complete the proof, we check that all vertices and faces end happy. The
discharging ensures that 2-vertices end with charge 0. Since 3-vertices have no
2-neighbors, their charge remains 0. For j ≥ 4, a j-vertex may lose 12 along each
edge and ends with charge at least 2 j − 6 − 2j , which is nonnegative.
A j-face has at most ⌊ 2j ⌋ incident 2-vertices, since 2-vertices are not adja-
cent. Hence a j-face has final charge at least j − 6 − 12 ⌊ 2j ⌋ , which is nonnegative
whenever j ≥ 8. To help the 7-faces, we add another discharging rule.
(R2) When an edge e joins 4+-vertices, redirect the charge 12 that each has avail-
able for a 2-neighbor so that instead the two faces bounded by e each receive 12 .
Now when a 7-face gives away 32 to three 2-vertices, it recovers 12 from the adjacent
4+-vertices on its boundary and ends with charge 0.

•
• •
• • •
• •
• •

9.3.29. Remark. Lemma 9.3.28 strengthens Lemma 9.3.16 for planar graphs; it
can also be proved by balanced charging (Exercise 67). The proof here shows both
“redirection” of transmitted charge and the possible need for extra discharging
to help elements that lose too much charge.
When d(G) < 125 and
(G) = 2, already Lemma 9.3.19 guarantees an edge of
weight 4, meaning adjacent 2-vertices. One may wonder what additional sparse-
ness is guaranteed as the average degree decreases toward 2. Instead of lighter
edges, we obtain longer threads, where an -thread is a path in G with inter-
nal vertices having degree 2 in G (see Exercise 53). Long threads are reducible
for various properties of colorability or decomposition.

We next use vertex charging to prove a classical result about light edges in a
planar graph G. When (G) = 5, Franklin’s result (Proposition 9.3.14) guaran-
tees a 5-vertex with two 6−-neighbors and thus an edge of weight at most 11.
When (G) = 2, we cannot guarantee a light edge: every edge in K 2 ,n−2 has
weight n. The fundamental result of Kotzig [1955] (known as Kotzig’s Theo-
rem) is that every 3-connected planar graph has an edge of weight at most 13.
We prove a stronger version due to Borodin [1989b]. A normal plane map is a
plane multigraph where every vertex degree and face length is at least 3.

9.3.30.* Lemma. (Borodin [1989b]) Every normal plane map G has an edge of
weight at most 11 or a 4-cycle through two 3-vertices and a 10−-vertex (and
hence an edge of weight at most 13 at a 3-vertex).
Proof: (Jendrol’ [1999], Cranston–West [2017]) Suppose that G has no edge of
weight at most 11. If any face  has length more than 3, then adding a chord
joining the neighbors on  of a lowest-degree vertex of  does not create one of the
desired configurations. Hence any desired subgraph in a triangulation obtained
from G must have occurred in G.
412 Chapter 9: Planar Graphs

We may therefore assume that every face has length 3. Assign charge by
vertex charging and use this discharging rule (vertices take what they need):
(R1) Every 5− vertex v takes charge 6−d(v)
d(v)
from each neighbor.
Faces start and end with charge 0. Since we assume that G has no edge of
weight at most 11, only 7+-vertices lose charge. Hence 5−-vertices become happy
and 6-vertices remain happy.
Since G is a triangulation, the neighbors of a vertex v form a cycle C. With
d(v) = j and no light edges, at most ⌊ j/2⌋ vertices along C take charge from v. A
7-vertex loses charge only to 5-vertices, so it loses at most 3( 15 ). An 8-vertex loses
charge only to 4+-vertices and hence loses at most 4( 12 ). A 9+-vertex v loses charge
at most 1 to at most ⌊ d(v)/2⌋ neighbors; this leaves charge at least ⌈ d(v)/2⌉ − 6,
which is nonnegative when d(v) ≥ 11. Thus in these cases v remains happy.
This leaves d(v) ∈ {9 , 10}. If v has no 3-neighbor, then v loses at most ⌊ d(v)
2 ⌋ 2,
1

leaving nonnegative charge. Hence already we have proved that G must have an
edge of weight at most 13 at a 3-vertex.
In addition, any three consecutive vertices on C form a 4-cycle with v. Thus
avoiding the specified 4-cycle separates 3-vertices by at least three edges along C,
so v gives charge to at most ⌊ d(v)/3⌋ such vertices. If a 9-vertex v has at least
three 3-neighbors, then it has exactly three and loses no other charge. Hence a
9-vertex loses at most max{3 · 1 , 2 · 1 + 2 · 12 } and ends happy. A 10-vertex has
at most five 5−-neighbors, and with three 3-neighbors it can only have one other
5−-neighbor. It loses at most max{4 · 1 , 2 · 1 + 3 · 12 } and also ends happy.

9.3.31.* Remark. For sharpness of the guarantee on light edges in Lemma

9.3.30, see Exercise 49. We also mention an application. Proper edge-coloring
is decomposition into subgraphs with maximum degree 1. The arboricity of a
graph is the number of forests needed to decompose it (planar graphs have ar-
boricity at most 3; Exercise 41). A linear forest is a disjoint union of paths.
The linear arboricity, la(G), is the minimum number of linear forests
needed to decompose G. Since 2-regular graphs contain cycles, la(G) ≥ (2r + 1)/2
when G is 2r-regular. Akiyama–Exoo–Harary [1980, 1981] conjectured that al-
ways la(G) ≤ ⌈( (G) + 1)/2⌉ . Their papers plus Enomoto–P éroche [1984] and Gul-
dan [1986] together proved this for (G) ≤ 6 and (G) ∈ {8 , 10}. For > 0 and
(G) large, Alon [1988] proved la(G) ≤ ( 12 + ) (G).
The conjecture has been proved for planar graphs, with (G) ≥ 9 in Wu
[1999] and (G) = 7 in Wu–Wu [2008]. The proof for planar G with (G) ≥ 9 (Ex-
ercise 51) uses the fact provided by Lemma 9.3.30 that G has a 4-cycle through
two 3-vertices when it has no edge with weight at most 11.

OTHER ASPECTS OF DISCHARGING (optional)

Although the remainder of this section is optional, the proofs illustrate im-
portant aspects of discharging arguments, especially Theorem 9.3.32.
We return to the List Color Conjecture  l(G) =  (G). Thinking that possibly
(G) = (G) + 1 for some graph G with  (G) = (G), Vizing conjectured  (G) ≤

(G) + 1. Juvan–Mohar–Skrekovski [1998a, 1999] proved this for (G) ≤ 4.
Section 9.3: Coloring of Planar Graphs 413

For planar graphs, Borodin [1990] proved (G) ≤ (G) + 1 when (G) ≥ 9.
We use balanced charging to prove Borodin’s result. An interesting tool is a reser-
voir or “pot ” of charge that can flow to or from vertices or faces without regard
to location; this enables charge to move long distances. In this proof, the pot al-
lows charge to move from maximum-degree vertices to 3-vertices; we need not
name specific recipients. Notions analogous to the pot of charge for long-distance
transfer of charge appear in Havet–Sereni [2006] and Borodin–Ivanova [2009]; a
general term for such methods is global discharging.

9.3.32. Theorem. (Borodin [1990]) (G) ≤ (G) + 1 for planar G with (G) ≥ 9.
Proof: (Cohen–Havet [2010]) To ensure that the family is closed under taking
subgraphs, we prove more generally that if  ≥ 9 and G is planar with (G) ≤  ,
then (G) ≤  + 1. Let G be a minimal counterexample, with L assigning lists of
size  + 1 to edges so that G has no L-edge-coloring. By Lemma 9.3.16, edges with
weight at most  + 2 are reducible. Hence we may assume that (G) ≥ 3 and that
every edge has weight least  + 3. Since  ≥ 9, edges have weight at least 12.
We use balanced charging and a pot of charge. Initially, the pot has charge 0.
Discharging that makes each vertex and face happy and maintains nonnegative
charge in the pot contradicts negative total charge. We use two discharging rules.
(R1) The pot takes 12 from every -vertex and gives 1 to every 3-vertex.
(R2) Each 3-face takes 12 from each incident 8+-vertex and j −j 4 from each incident
j-vertex with j ∈ {5 , 6 , 7}.
To ensure positive charge in the pot, we prove n¾ > 2n3 , where nj is the num-
ber of j-vertices in G. The edges incident to 3-vertices form a bipartite graph H
with the 3-vertices as one part and the -vertices as the other. Any cycle C in
H has even length, since H is bipartite. By minimality of the counterexample,
G − E(C) has an L-edge-coloring . Each edge of C has (G) − 1 incident edges
colored by  , so it still has at least two available colors. Since even cycles are 2-
edge-choosable (Example 8.2.7), the L-edge-coloring extends to G. Thus G being
a counterexample requires H to be acyclic, with fewer than n3 + n¾ edges. Since
| E(H)| = 3n3 , we have 3n3 < n3 + n¾ , as desired.
3  3 
• • • •

For vertices, (R1) makes 3-vertices happy. A j-vertex v with j ∈ {4 , 5 , 6 , 7}

loses altogether at most j − 4, its initial charge. Since  ≥ 9, an 8-vertex gives
nothing to the pot and loses at most 4. For j ≥ 9, possibly sending 12 to the pot, a
j-vertex loses at most j +2 1 and is happy.
For faces, the 4+ -faces lose no charge and remain happy; we must show that
each 3-face  gains at least 1. Let j be the least degree among vertices incident to
 . If j ≤ 4, then  has at least two incident 8+-vertices, giving 12 each. If j = 5,
then  has at least two incident 7+-vertices giving at least 37 each, plus 15 for the
5-vertex. If j ≥ 6, then each vertex incident to  gives at least 13 to  .
Here again discharging produces an unavoidable set of reducible configura-
tions. The configurations are light edges (weight at most (G) + 2) and cycles
alternating between 3-vertices and (G)-vertices.
414 Chapter 9: Planar Graphs

9.3.33. Remark. Continuing the study of edge-choosability in planar graphs,

Bonamy [2015] extended Theorem 9.3.32 to (G) ≥ 8 by discharging, but the de-
tails are much longer. Cranston [2009] showed that (G) ≥ 6 suffices when no two
3-faces share an edge (Exercise 59). For (G) = 5, we have Ò l(G) ≤ (G) + 1 when
G is planar with no 3-cycle (Zhang–Wu [2004]), with no 4-cycle (Cranston [2009]),
or with no 5-cycle (Wang–Lih [2002]). Higher maximum degree reduces the up-
per bound; Borodin [1990] proved Ò l(G) = (G) for planar graphs with (G) ≥ 14
(Exercise 63), which Borodin–Kostochka–Woodall [1997] reduced to (G) ≥ 12.

9.3.34. Remark. Even without planarity, we can study the effect of Mad(G) on
edge-coloring. Using the Vizing Adjacency Lemma (Corollary 8.3.11) and degree
charging, one can show that when (G) ≥ 8 the condition Mad(G) < 6 implies
Ò (G) = (G) (Exercise 64). In fact, Mad(G) < 6.5 suffices (Miao–Sun [2010]).
If Mad(G) < 2, then G is a forest, and Ò l(G) = Ò (G) = (G). Hence for both
Ò and Ò l there is a threshold on Mad(G) (in terms of (G)) to guarantee equaling

(G). Vizing [1968] conjectured that Mad(G) ≤ (G) − 1 implies Ò (G) = (G).
In fact, he conjectured 2 | E(G)| ≥ | V(G)| ( (G) − 1) + 3 for critical graphs. Fiorini
[1975] proved that Mad(G) < 14 (G) implies Ò (G) = (G).
Sanders–Zhao [2002] proved that Mad(G) < 12 (G) implies Ò (G) = (G), and
√
in fact 2 | E(G)| ≥ 12 | V(G)| ( (G) + 2 (G) − 1) when G is critical. This result
pioneered the interaction between Mad(G) and discharging; they described it as
“the first time that the discharging method is applied to a graph theory problem
in which Euler ’s formula is not used and embeddings of graphs in surfaces are
not mentioned.” Later, Woodall [2007] proved that Mad(G) < 23 ( (G) + 1) suffices.
Woodall [2010] conjectured that Mad(G) < (G)−1 also implies Ò l(G) = (G),
but here the results are much weaker. Using an “iterated discharging ” argument
as presented√ in Woodall [2010], Borodin–Kostochka–Woodall [1997] proved that
Mad(G) < 2 (G) implies Ò l(G) = (G).

For ordinary proper coloring, the Four Color Theorem begs the question of
which planar graphs are 3-colorable. Grötzsch [1959] proved that planar triangle-
free graphs are 3-colorable; other proofs (all using discharging) are by Thomassen
[1994a], Dvořák–K awarabayashi–Thomas [2011] (with a linear-time algorithm
to find a proper 3-coloring), and Kostochka–Yancey [2014] (see Exercise 31).

9.3.35. Remark. Steinberg’s 3-Color Conjecture from 1975 (see Aksionov–

Melnikov [1978]) asserts that planar graphs with no 4-cycle or 5-cycle are 3-
colorable; triangles are allowed. Erdős asked whether it suffices to forbid cycle
lengths 4 through ¾ for some ¾ . This was proved for ¾ = 11 (Abbott–Zhou [1991])
and ¾ = 9 (Sanders–Zhao [1995] and Borodin [1996a]; see below). With more
discharging, Borodin–Glebov–Raspaud–Salavatipour [2005] proved that forbid-
ding j-faces for 4 ≤ j ≤ 7 is enough. Many others proved 3-colorability under
various restrictions on cycle-lengths. Eventually, Cohen-Addad–Hebdige–Kr ál’–
Li–Salgado [2017] showed that Steinberg ’s Conjecture is false. Their smallest
example has 85 vertices.
For choosability, Thomassen [2003] proved Ò l(G) ≤ 3 when G is planar with
girth at least 5. Dvořák–Postle [2018] (on arXiv in 2015) proved that planar
graphs having no cycles of lengths 4 through 8 are 3-choosable.
Section 9.3: Coloring of Planar Graphs 415

The traditional proof of the lemma for the result of Sanders–Zhao [1995] and
Borodin [1996a] uses balanced charging, but by face charging it is a bit simpler.

9.3.36. Lemma. (Borodin [1996a]) Every 2-connected plane graph G such that
(G) ≥ 3 has two 3-faces with a common edge, or a j-face with 4 ≤ j ≤ 9, or a
10-face whose vertices all have degree 3.
Proof: Let G be a plane graph with (G) ≥ 3 and no such configuration. Use face
charging (2d(v) − 6 on each vertex v and l( ) − 6 on each face ). The total charge
is −12. With no 4-faces or 5-faces, only triangles start with negative charge, −3.
(R1) Each 3-face takes 1 from each neighboring face.
(R2) Each non-triangular face takes 1 from each incident 4+-vertex lying on a
triangle sharing an edge with (to help faces that lose too much).
Here (R1) makes 3-faces happy (since no two 3-faces share an edge), and 3-
vertices remain at charge 0. For j ≥ 4, a j-vertex loses charge at most ⌊ 2j ⌋ (since
3-faces do not share edges) and ends with at least ⌈ 32j ⌉ − 6, which is nonnegative.
For j ≥ 10, a j-face loses 1 for every path along its boundary such that the
neighboring faces are triangles and the endpoints have degree 3, as shown on the
left below. Face gives 1 to each of those 3-faces but takes 1 from each interven-
ing vertex, since forbidding 3-faces with shared edges requires the intervening
vertices to have degree at least 4. If an endpoint of a maximal such path has de-
gree at least 4, then there is no net loss. Hence the net loss for is at most ⌊ 2j ⌋ ,
and the final charge is at least ⌈ 2j ⌉ − 6, which is nonnegative when j ≥ 11.
Hence negative charge occurs only at 10-faces. Losing more than 4 requires
losing 1 through five paths. The paths must be single edges sharing no vertices,
and all the vertices incident to must have degree 3, as on the right below.

•
• • • •
• •
↓ ↑→← ↑ →←↑ ↓ • • • •
→ • ↑ • ↑ • ←
• ↑ ↓ ↓ ↓
↑ • • •
• • • •
• •

9.3.37. Theorem. (Borodin [1996a], Sanders–Zhao [1995]) Every plane graph

having no j-cycle with 4 ≤ j ≤ 9 is 3-colorable.
Proof: Since the family described is hereditary, a smallest counterexample G
must be 4-critical, so (G) ≥ 3 and (G) ≥ 2. Since there is no 4-cycle, no two 3-
faces share an edge. By Lemma 9.3.36, we may thus assume that G is embedded
with at least one 10-face C, whose vertices all have degree 3.
Let  be a proper 3-coloring of G − V(C). If each vertex on C has exactly one
neighbor outside C, then two colors remain available at each vertex of C. Other-
wise, C has a chord (drawn outside C), and at the endpoints of a chord three colors
remain available. Since even cycles with chords are degree-choosable (Theorem
8.2.13), the coloring can be completed.
416 Chapter 9: Planar Graphs

Discharging can also guarantee many desired configurations. For example,

the amount by which total charge is negative can force many light edges.

9.3.38. Theorem. (Borodin–Sanders [1994]) For a plane graph G with (G) = 5,

2e5 ,5 + e5 ,6 + 27 e5 ,7 ≥ 60,
where e i , j is the number of edges joining vertices of degrees i and j . Also,
the coefficients in this inequality are sharp.
Proof: Add edges to obtain a triangulation H ; still (H) = 5. Since (G) = 5,
no edges at 5-vertices of H were added. Hence e5 , j (H) ≤ e5 , j (G), and it suffices to
prove the result for triangulations. We use vertex charging.
−6
Each vertex distributes its charge equally to all incident edges, giving d(v) d(v)
to
each. Thus each vertex ends with 0 and only edges end with nonzero charge. Only
5-vertices take charge ( 15 from each edge), so only their incident edges can have
negative charge. An 8+-vertex gives each incident edge at least 14 , which exceeds
1
5 . Thus negative charge remains only on edges from 5-vertices to vertices of de-
grees 5, 6, or 7, with charges −52 , −51 , and −352 , respectively. Since the total charge is
−12 and there is no negative charge elsewhere, we obtain −52 e5 ,5 + −51 e5 ,6 + −352 e5 ,7 ≤
−12. Multiplying by −5 completes the proof.
Equality forbids edges with positive charge, since that would force more edges
with negative charge. Hence every sharpness example is a triangulation with
maximum degree at most 7. The icosahedron has e5 ,5 = 30 and no other edges;
thus the coefficient of e5 ,5 cannot be reduced. The graph obtained from the dodec-
ahedron by putting in each face a 5-vertex adjacent to its corners has e5 ,6 = 60
with all other edges joining 6-vertices; thus the coefficient of e5 ,6 cannot be re-
duced. The graph below (three edges wrapping from left to right), with 2e5 ,5 =
e5 ,6 = 28 and e5 ,7 = 14, shows that the coefficient of e5 ,7 cannot be reduced.

•
• • • • • • •
a a
b • • • • • • • b
c c
• • • • • • •
•

9.3.39. Remark. To explain the method, we have presented discharging argu-

ments with fairly simple details. Many applications are quite intricate. Robert-
son–Sanders–Seymour–Thomas [1997] used 32 discharging rules to produce
their unavoidable set of 633 reducible configurations for the Four Color Theorem.
Discharging was also used for Tutte’s 3-Edge-Coloring Conjecture: 2-
edge-connected 3-regular graphs containing no subdivision of the Petersen graph
are 3-edge-colorable. Robertson–Seymour–Thomas [1997, 2019] reduced this
to 3-edge-colorability of two types of nearly planar 3-regular graphs: “double-
cross graphs” (Edwards–Sanders–Seymour–Thomas [2016]) and “apex graphs”
(Sanders–Thomas, unpublished). Exercise 16.1.13 is one step in the proof.
Exercises for Section 9.3 417

Borodin [1984, 1995] used discharging (without computers) to prove Ringel’s

Six Color Conjecture from 1965: the vertices and faces of a planar graph can be
6-colored so that neighboring objects have different colors. The original discharg-
ing argument produced 35 reducible configurations, later cut by about half.
Inductive proofs using discharging to produce reducible configurations also
have algorithmic implications. The proof of the Four Color Theorem in Robertson–
Sanders–Seymour–Thomas [1997] yields a quadratic-time algorithm that finds a
proper 4-coloring; the Appel–Haken proof provides a quartic algorithm. Hea-
wood ’s proof produces proper 5-colorings in quadratic-time, but Thomassen’s
proof of 5-choosability (without discharging) yields a linear-time algorithm.

EXERCISES 9.3

9.3.1. (−) Prove that every maximal plane graph other than K4 is 3-face-colorable.
9.3.2. (−) Exhibit 3-regular graphs with these properties:
(a) planar but not 3-edge-colorable.
(b) 2-connected but not 3-edge-colorable.
(c) planar with connectivity 2, but not Hamiltonian.
9.3.3. (−) Prove that a 2-edge-connected plane graph is 2-face-colorable if and only if it is
Eulerian.
9.3.4. (−) Use the Four Color Theorem to prove that every planar graph decomposes into
two bipartite graphs. (Hedetniemi [1969], Mabry [1995])
9.3.5. (−) Prove that every graph with maximum average degree at most d is d-
degenerate. Prove that every d-degenerate graph has maximum average degree at most
2d. Give infinitely many examples to show that the second statement is nearly sharp.
9.3.6. (−) Prove or disprove: Every triangle-free planar graph is 4-choosable.
9.3.7. (−) Prove or disprove: For ¾ sufficiently large, a graph G with (G) =  and d(G) <
 + 1 has adjacent -vertices.
9.3.8. (−) Let G be a planar graph with girth at least 6 and (G) ≥ 2. By Lemma 9.3.19,
G has an edge with weight at most 7. Prove sharpness by constructing a planar graph
with girth 6 and minimum degree 2 having no edge with weight at most 6.
9.3.9. Reduction of Four Color Problem to Tait’s conjecture. Let G be a 3-regular graph
with edge-connectivity 2. (Recall that (G) =  (G) when G is 3-regular.)
(a) Prove that there exist subgraphs G1 , G 2 ⊆ G and vertices u1 , v1 ∈ V(G1) and
u2 , v2 ∈ V(G 2) such that u1 v1 , u2 v2 ∈
/ E(G) and G consists of G1 , G 2 , and a “ladder ” (of
some length) joining them at {u1 , v1 , u2 , v2 } as illustrated below.
(b) Prove that  (G1 + u1 v1) =  (G 2 + u2 v2) = 3 implies  (G) = 3.
(c) Use Tait ’s Theorem (Theorem 9.3.1) to reduce the Four Color Theorem to Tait ’s
Conjecture: “every simple 3-regular 3-connected planar graph is 3-edge-colorable.”

v1 • • • • v2
G1 G2

u1 • • • • u2
418 Chapter 9: Planar Graphs

9.3.10. Reduction of Four Color Problem to Hamiltonian graphs.

(a) Prove that every triangulation with at least four vertices is 3-connected and that
a separating triple in such a graph induces a 3-cycle.
(b) Whitney [1931] proved that every 4-connected planar triangulation is Hamilto-
nian. Use this to reduce the Four Color Problem to the problem of proving that every
Hamiltonian planar graph is 4-colorable. (See also Saaty–Kainen [1977])
9.3.11. Prove Grinberg ’s Theorem by induction on the number of edges.
9.3.12. (♦) Use Grinberg ’s Theorem to prove that each graph below is not Hamiltonian.

• • • • • •
• • • • •
• • • • • •
• • • • • • • •
• • • • • • •
•
• • • • • •

9.3.13. (♦) Use Grinberg ’s Theorem to prove that the Grinberg graph (on the left below)
is not Hamiltonian.

•
• •
• • • • •
• • • • • • •
• • •• •• • • • • • • • • •
• • • • • • •
• • • • • • • • •
• • • • •
• • • • • • •
• • •
• • •
• •

9.3.14. Give a short proof that the graph on the right above is not Hamiltonian. Show
that Grinberg ’s Theorem cannot be used directly to prove this. Show that Grinberg ’s
Theorem does apply to a modification of the graph.
9.3.15. The smallest known 3-regular 3-connected planar graph that is not Hamiltonian
has 38 vertices and appears on the left below. Prove that this graph is not Hamiltonian.
(Lederberg [1966], Bosák [1967], Barnette)
• • •• • • • • •
• • A B A A
• • •
• • • • • • • • • •
• • • • • • • • • • ••• • B B B A
• •• • • • • • • •
• • B A B A
• • • • • • • •

9.3.16. (♦) Let G be the grid Pr P s . Let Q be a Hamiltonian path from the upper left cor-
ner vertex to the lower right corner vertex, such as that shown on the right above in bold.
Note that Q partitions the grid into regions, some opening to the left or downward and
others opening to the right or upward. Prove that the total area of the up-right regions
(B) equals the total area of the down-left regions (A). (Fisher–Collins–Krompart [1994])
Exercises for Section 9.3 419

9.3.17. Prove that an outerplanar graph can be 2-colored so that the subgraph induced
by each color class is a disjoint union of paths. (Mihók [1983], Akiyama–Era–Gervacio–
Watanabe [1989], Goddard [1991])
9.3.18. Prove that a planar graph can be 2-colored so each color class induces an outerpla-
nar graph. (Chartrand–Geller–Hedetniemi [1971], Buršteı̆n [1974], Penaud [1975])
9.3.19. Prove that every outerplanar graph can be properly 3-colored so that the union of
any two color classes induces a forest. Conclude that the vertices of an outerplanar graph
can be partitioned into an independent set and a set inducing a forest so that the indepen-
dent set has at least 1/3 of the vertices.
9.3.20. (♦) Short proof of the Five Color Theorem. Let v be a 5-vertex in a graph G. Given
that x and y are nonadjacent neighbors of v, let G be the graph obtained from G by con-
tracting the edges vx and vy. Prove that if G is 5-colorable, then G is 5-colorable. Use
this to prove by induction that every planar graph is 5-colorable. (Kainen [1974])
9.3.21. Albertson–Berman [1979] and Akiyama–Watanabe [1987] conjectured that ev-
ery n-vertex planar graph has an induced subgraph with at least ⌈ n/2⌉ vertices that is a
forest. This would yield an independent set of at least ⌈ n/4⌉ vertices without using the
Four Color Theorem. Show that the conjecture is sharp by constructing an n-vertex pla-
nar graph having no induced forest with more than ⌈ n/2⌉ vertices. Akiyama–Watanabe
[1987] also conjectured that every n-vertex bipartite planar graph has an induced forest
with at least ⌈ 5n/8⌉ vertices; show that this conjecture also is sharp.
9.3.22. Without using the Four Color Theorem, prove that every Hamiltonian plane
graph is 4-face-colorable (nothing is assumed about the vertex degrees).
9.3.23. (♦) Let G be a maximal plane graph. Prove that G∗ has a 2-factor and use it to
show that V(G) can be 2-colored so that both colors appear on each face. Obtain the same
conclusion from the Four Color Theorem. (Buršteı̆n [1974], Penaud [1975])
9.3.24. (♦) Thomassen [1995b] proved that the vertices of any planar graph can be parti-
tioned into sets inducing a forest and a 3-degenerate graph. To generalize this, prove for
 ¾
d1 , . . . , d ¾ ∈ 0 and d ≤ ¾ − 1 + ∑i=1 d i that the vertices of any d-degenerate graph can
be partitioned into sets V1 , . . . , V¾ such that G[Vi ] is d i-degenerate for 1 ≤ i ≤ ¾ .
9.3.25. (♦) The Art Gallery Theorem. (Chvátal [1975], Fisk [1978])
(a) Prove that outerplanar graphs are 3-colorable (without the Four Color Theorem).
(b) Prove that an art gallery laid out as an n-gon can be protected by ⌊ n/3⌋ guards so
that every point of the interior is visible to some guard.
(c) For each n, construct an art gallery with n outer segments that cannot be watched
by fewer than ⌊ n/3⌋ guards.
9.3.26. (♦) An art gallery with walls is a polygon plus nonintersecting chords called “walls”
that join vertices. Each interior wall has a tiny “doorway”. A guard in a doorway can see
along both sides of the wall containing it , but no guard can see through a wall. Determine
the least t such that for every walled art gallery with n vertices, one can place t guards so
that every interior point is visible to some guard. (Hutchinson [1995], K ündgen [1999])
9.3.27. (♦) Heawood [1898] stated that a maximal planar graph is 3-colorable if and only
if it is Eulerian. Golovina–Yaglom [1963] proved it (Hutchinson [2001] generalized it).
(a) Prove that every 3-colorable triangulation (including multigraphs) is Eulerian.
(b) Let G be an even plane graph with all bounded faces having length 3. By Exercise
9.1.49, G has a noncrossing Eulerian circuit. Prove that every subcircuit has length di-
visible by 3. Conclude that G is 3-colorable. (Tsai–West [2011])
(c) Let G be as in part (b). Prove that G is 3-colorable using the fact that the dual is
bipartite. (Hint: Use induction on the number of bounded faces.) (Lovász [1993])
420 Chapter 9: Planar Graphs

9.3.28. (♦) Prove that every 3-colorable planar graph is a subgraph of some 3-colorable
triangulation. (Comment: C5 is not a spanning subgraph of a 3-colorable triangulation;
hence it may be necessary to add vertices.) (Król [1972])
9.3.29. Use the Four Color Theorem to prove that every 2-edge-connected 3-regular graph
that can be drawn in the plane with only one edge-crossing is 3-edge-colorable.
9.3.30. Gr ötzsch’s Theorem states that every triangle-free n-vertex planar graph G is
3-colorable, so (G) ≥ n/3. Tovey and Steinberg showed (G) > n/3. Prove sharpness using
the following graphs. Let G1 be the 5-cycle, with vertices a , x0 , x1 , y1 , 1 in order. For  >
1, obtain G ¾ from G ¾−1 by adding three vertices x¾ , y¾ , ¾ , the path ⟨x¾−1 , x¾ , y¾ , ¾ , y¾−1 ⟩
and the edge ¾ x¾−2 . The graph G3 appears below. (Fraughnaugh [1985])
x0 x1 x2 x3
• • • •

• • • • • • •
a 1 y1 2 y2 3 y3
9.3.31. (♦) Kostochka–Yancey [2014] proved (using some discharging)that every 4-critical
n-vertex graph has at least ⌈ (5n − 2)/3⌉ edges (conjectured by Ore [1967]). Use their result
to prove Grötzsch’s Theorem: Every triangle-free planar graph is 3-colorable.
9.3.32. It has been conjectured that every planar triangulation has edge-chromatic num-
ber (G), and this has been proved for sufficiently large (G). Show that  (G) = (G) for
the icosahedron, shown on the left below.
• • • •
•
• • •
•
• • • • •
• • • • • •
• • • •
•
• • •
9.3.33. An equitable coloring of a graph is a proper coloring whose color classes differ in
size by at most 1. Let G n be the graph formed from n concentric 4-cycles triangulating the
region between any two consecutive 4-cycles (each vertex has two consecutive neighbors on
the other 4-cycle); G4 appears above on the right.) Prove that for G n with n even and for
the icosahedron (on the left above), every proper 4-coloring is equitable. (M. Albertson)
9.3.34. (♦) Non-4-choosable planar graph with 75 vertices.
(a) Prove that the graph below cannot be properly colored from the given lists; the
vertices of degree 5 have lists of size 1 and the others have lists of size 4.
(b) Use part (a) to construct a 3-colorable planar graph with 75 vertices that is not
4-choosable. (Gutner [1996])
x
•

xa12
• •xb34
xy12• • xyab • xy34
• •yb34
ya12

•
y
Exercises for Section 9.3 421

9.3.35. (♦) Non-4-choosable planar graph with 63 vertices.

(a) In the list assignments for the graph on the left below, S denotes [4] and i denotes
S − {i}. Prove that this graph has no proper coloring chosen from these lists.
4• •1 4• •1 1• •5 5• •1 1• •5
S 5 4 2 3
• • • • •
1 5 1
4• 2• •3 •1 4• 2• •3 • 3• •2 • 4• •3 • 2• •4 •5
S S S 5 5 5 4 4 4 2 2 2 3 3 3
• • • • • • • • • • • • • • •
3• • • •2 3• • • • • • • • • • • • •2
1 S 4 1 5 4 2 5 4 1 3 1 2 5 4 5 3 1
• • • • •
3• •2 3• •2 2• •3 3• •4 4• •2
(b) In the list assignments for the graph G on the right above, i denotes [5] − {i};
each list has size 4. Let G be the graph obtained from G by adding one vertex with list
[5] − {1} adjacent to all vertices on the outside face of this drawing of G. Prove that G
has no proper coloring chosen from these lists. (Mirzakhani [1996])
9.3.36. Without the Four Color Theorem, prove that every planar graph with at most 12
vertices is 4-colorable. Conclude that planar graphs with at most 32 edges are 4-colorable.
9.3.37. Let H be a configuration in a planar triangulation(Definition 9.3.8). Let H be ob-
tained by labeling the neighbors of the ring vertices with their degrees and then deleting
the ring vertices. Prove that H can be retrieved from H when H has no cut-vertex.
9.3.38. Without using anything not proved in this section, show that a smallest 5-
chromatic triangulation has no separating cycle of length 3 or 4. That is, show that any
configuration with a ring of length at most 4 is reducible. (Birkhoff [1913])
9.3.39. (+) Prove that the Birkhoff diamond (the configuration on the left below) is re-
ducible for 4-colorability of plane triangulations. (Hint: Prove first that it suffices to
consider colorings of the ring obtained by replacing the Birkhoff diamond with the con-
figuration on the right. For each such coloring, extend it to a 4-coloring of the interior
diamond, possibly after modifying it via Kempe chain arguments.) (Birkhoff [1913])
w1 w6 w1
• • •
v1
•
v2 • w6
w2 • • v4 • w5 w2 • • • w5
w4
•
v3
• • •
w3 w4 w3

9.3.40. Sharpness of Steinberg’s Conjecture. Construct two plane graphs with chromatic
number 4, one having no 5-cycles and the other having no 4-cycles.
9.3.41. (+) Arboricity of planar graphs. Let u be a 5-vertex in a plane triangulation G. Let
N(u) = {v, w , x , y , }, ordered so [v , w, x , y , ] is a cycle. Form G from G − u by adding
the edges vx and vy. Prove that if G decomposes into three forests, then also G decom-
poses into three forests. Conclude that every planar graph decomposes into three forests.
(Comment: Balogh–Kochol–Pluhár–Yu [2005] proved that one of the three forests can be
required to have maximum degree at most 8. Gonçalves [2009] reduced 8 to 4.)
9.3.42. Every graph G with Mad(G) <  + 1 is -degenerate. Disprove the converse: for

each  ∈ with  ≥ 2 construct a -degenerate graph with average degree at least  + 1.
9.3.43. Let G be a graph with (G) ≥ 2 and d(G) < ¾4+¾2 . Complete the proof of Lemma
9.3.19 by proving when  > 6 that G has an edge with weight at most  + 1.
422 Chapter 9: Planar Graphs

9.3.44. (♦) Adjacent vertices of minimum degree.

(a) Determine the largest b such that if (G) =  and d(G) < b, then G must have two
adjacent vertices of degree  .
(b) Determine the least  such that every planar graph G with girth at least  and
(G) ≥ 2 has an edge of weight 4. Explain how this relates to part (a).
9.3.45. Let G be a graph with (G) =  . For 0 ≤ j <  , determine the largest  such that
d(G) <  +  guarantees in G a -vertex with more than j neighbors of degree  .
9.3.46. Prove that if d(G) < 4, then (G) ≤ 2 or G has a 3-vertex with a 5−-neighbor. Show
that when (G) = 2 and d(G) < 4 there is no upper bound on the smallest edge weight.
9.3.47. Let G be a graph with (G) = 3 and d(G) < 103
. Prove that G has a 3-vertex whose
neighbors have degree-sum at most 10. (Comment: G. Tardos constructed such a graph
with 98 vertices and d(G) < 10
3
having no 3-vertex with three 3-neighbors.)
9.3.48. (♦) Let G be a connected plane graph such that (G) ≥ 3 and (G∗) ≥ 3.
(a) Use balanced charging to prove that G has a 3-vertex on a 5−-face or a 5− -vertex
on a 3-face. (Comment: There are five such combinations of degrees and face-lengths; the
Platonic solids show that none can be excluded.)
(b) Strengthen part (a) by proving that G contains a 3-vertex on a 5−-face, a 4-vertex
on a 3-face, or a 5-vertex with at least four incident 3-faces. (Comment: (Lebesgue [1940]
proved this for 3-connected plane graphs.)
9.3.49. Construct a triangulation where all lightest edges join 5-vertices and 6-vertices
and another where all lightest edges join 3-vertices and 10-vertices (see Lemma 9.3.30).
9.3.50. Determine whether every planar graph with girth at least 4 and minimum de-
gree 3 has a 3-vertex with a 4− -neighbor. Construct a plane graph G ¾ with girth 4 and
minimum degree 3 where the distance between 3-vertices is at least  . Construct a planar
graph H¾ with minimum degree 5 where the distance between 5-vertices is at least  .
9.3.51. (+) For t ≥ 5, prove that every planar graph G with (G) < 2t decomposes into t
linear forests. (Wu [1999])
9.3.52. Vizing’s Planar Graph Conjecture states that planar graphs with (G) ≥ 6
are Class 1. Prove necessity by finding a graph with maximum degree 6 that is not 6-
edge-colorable and a planar graph with maximum degree 5 that is not 5-edge-colorable.
(Comment: Vizing [1965] proved the conjecture for (G) ≥ 8. Sanders–Zhao [2001] and
Zhang [2000] proved it for (G) = 7, but for (G) = 6 it is only known in special classes.)
9.3.53. (♦) An -thread in a graph G is a path with  internal vertices, all with degree 2 in
G. Prove that if d(G) < 2 + 3t1−2 and G has no 2-regular component , then G has a 1 − -vertex
or a (2t − 1)-thread. Construct infinitely many examples with average degree 2 + 3t1−2 but
no (2t − 1)-thread. (Comment: (2t − 1)-threads are reducible for C2t+1 -coloring, where an
H-coloring of G is a map : V(G) → V(H) such that uv ∈ E(G) implies (u)(v) ∈ E(H).)
9.3.54. (♦) An I , F-partition of a graph G is a partition of V(G) into sets I and F such
that G[F] is a forest and the distance between any two vertices of I is at least 3. A star
-coloring of G is a proper -coloring such that no 4-vertex path is 2-colored.
(a) Prove that if d(G) < 2 + 2t1−1 and G has no 2-regular component , then G contains
(1) a 1 − -vertex, or (2) a 3-vertex with at least 4t − 3 vertices of degree 2 on its maximal in-
cident threads, or (3) a 4+ -vertex incident to a (2t − 1)-thread.
(b) Prove that if Mad(G) < 73 , then G has an I , F-partition. (Timmons [2008])
(c) Prove that if G has an I , F-partition, then G has a star 4-coloring.
(d) Construct infinitely many graphs with average degree 25 having no I , F-partition.
(Comment: Brandt–Ferrara–Kumbhat–Loeb–Stolee–Yancey [2016] proved that d(G) < 52
implies an I , F-partition of G. Perhaps higher Mad(G) still guarantees a star 4-coloring.)
Exercises for Section 9.3 423

9.3.55. Prove that if d(G) < 25 and G is connected, then G contains a 3− -vertex with
a 1-neighbor, a 4− -vertex with two 2−-neighbors, or a 5+ -vertex v having at least d(v)2−1
neighbors of degree at most 2. (Cranston–Jahanbekam–West [2014])
9.3.56. Let G be a connected graph with at least four vertices. Prove that if d(G) < 52 and
(G) ≥ 2, then G contains a 2-thread or a 3-vertex having three 2-neighbors, one of which
has a second 3-neighbor. (Cranston–Kim–Yu [2010])
9.3.57. (♦) Prove that every planar graph with minimum degree 5 contains two 3-faces
sharing an edge with weight at most 11. (Borodin [1996b])
9.3.58. Prove that every plane triangulation with minimum degree 5 has two 3-faces shar-
ing an edge such that the non-shared vertices have degree-sum at most 11. (Comment:
With this and a 4− -vertex as an unavoidable set , Albertson [1976] proved that (G) ≥ 2n/9
when G is an n-vertex planar graph, without using the Four Color Theorem.)
9.3.59. (♦) Say that a plane graph is clean if no two 3-faces share an edge. For  ≥ 7, prove
that a clean plane graph G with (G) ≤  has an edge with weight at most  + 2. Conclude
that  l(G) ≤ (G) + 1 when G is a clean plane graph with (G) ≥ 7. (Comment: Cranston
[2009] proved the same conclusion for (G) = 6, strengthening several earlier results.)
9.3.60. (♦) For (G) ≥ 2, prove that (G) ≤ 6 and d(G) < 27 guarantee in G an edge with
weight at most 7 or a cycle alternating between 2-vertices and 6-vertices. Conclude that
(G) ≤ 6 and Mad(G) < 27 imply  ç(G) ≤ 6. (Hint: Use degree charging and a pot of charge.
Comment: This result strengthens Theorem 9.3.21.)
9.3.61. (+) Light triangles. Let G be a planar graph with minimum degree 5. Proposition
9.3.14 guarantees weight at most 17 on some three-vertex path.
(a) Prove that G contains a 3-face with weight at most 17 (Borodin [1989a])
(b) Provide a single example to show that smaller weight cannot be guaranteed. Show
that (G) = 5 is needed by constructing for each  ∈ a planar graph with minimum
degree 4 having no 3-face with weight at most  .
9.3.62. Let G be a plane graph having no 4-cycle and no face length in {4 , . . . , }. Prove
that the average face length is at least 6 − ¾18
+4 . Conclude Mad(G) < 3 + 2¾ −1 . In particular,
9

Mad(G) < 4 when G is a plane graph with no 4-face or 5-face.

9.3.63. (+) Planar graphs with high maximum degree are Class 1.
(a) Let G be a plane graph with (G) ≥ 2. Prove that G contains an edge with weight
at most 15 or a cycle that alternates between 2-vertices and higher-degree vertices.
(b) Prove that  (G) = (G) when G is planar and (G) ≥ 14. (Borodin [1990])
9.3.64. Recall that a critical graph is one with  (G) = (G) + 1 such that deleting any
edge reduces the edge-chromatic number. The Vizing Adjacency Lemma (VAL, Corollary
8.3.11) says that for an edge yw in a critical graph, y has at least max{ (G) − d(w) + 1 , 2}
neighbors with degree (G). Use VAL to prove that if (G) ≥ 8 and Mad(G) < 6, then
 (G) = (G). (Vizing [1968], Luo–Zhang [2004])
9.3.65. (♦) Let G be a planar graph with minimum degree 5. As in Theorem 9.3.38, let
ei , j be the number of edges in G with endpoints of degrees i and j . Prove that e5 ,6 ≥ 60
when e5 ,5 = 0, and show that equality can hold. (Gr ünbaum [1973]) (Comment: More
generally, Borodin–Sanders [1994] proved 37 e5 ,5 + e5 ,6 ≥ 60.)
9.3.66. An acyclic coloring is a proper coloring in which the union of any two color
classes induces a forest. A graph is acyclically -choosable if an acyclic coloring can be
chosen from any lists of size  at the vertices. Prove that if Mad(G) < 3, then G is acycli-
cally 6-choosable. (Comment: Cranston–Yu [2011] proved this also for planar graphs with
girth at least 5: the added reducible configuration is a 5-face whose incident vertices are
four 3-vertices and a 5−-vertex.)
424 Chapter 9: Planar Graphs

9.3.67. Use balanced charging to prove that every planar graph with girth at least 7 and
minimum degree at least 2 has a 2-vertex adjacent to a 3− -vertex. Prove that the conclu-
sion does not always hold when Mad(G) < 14
5
(thus planarity is needed).

9.3.68. A dynamic coloring is a proper coloring with each vertex v adjacent to vertices
of at least min{2 , d(v)} colors (Montgomery [2001]). A graph is dynamically ¾-choosable
if a dynamic coloring can be chosen from any vertex lists of size ¾ (Kim–Park [2011]).
(a) Prove that every cycle of length at least 6 is dynamically 4-choosable.
(b) Prove that except for C5 , a minimal graph among those that are not dynamically
4-choosable cannot contain an edge of weight at most 5.
(c) Prove that every planar graph with girth at least 7 is dynamically 4-choosable, but
some planar graphs with girth 6 are not dynamically 4-choosable.

9.3.69. (♦) Prove that every planar graph has a 5− -vertex with at most two 12+-neighbors.
(Comment: Balogh–Kochol–Pluhár–Yu [2005] strengthened 12 to 11; see Exercise 9.3.41.
van den Heuvel–McGuinness [2003] applies this to coloring squares of planar graphs.)

9.3.70. (♦) Choosability of squares of subcubic graphs. Let G be a graph.

(a) Prove that if (G) = 2 and (G) ≤ 3 and d(G) < 14 5
, then G has one of the con-
figurations shown below: adjacent 2-vertices, a 3-vertex with two 2-neighbors, adjacent
3-vertices with 2-neighbors, or a 3-vertex with three 3-neighbors having 2-neighbors. (In
the last two configurations, the specified 2-vertices need not be distinct.)
(b) Prove that if (G) ≤ 3 and Mad(G) < 14 5
, then (G 2) ≤ 7. What does this imply
for planar graphs? (Cranston–Kim [2008])
• •
• •
• •
• • •
• • •
• •
• •

9.3.71. Choosability of squares. Let G be a graph.

(a) Prove that if 2 ≤ (G) ≤ (G) ≤ 4 and d(G) < 18 7
, then G has adjacent 2-vertices, or
a 3-vertex with three 2-neighbors, or a 4-vertex path alternating 2-vertices and 3-vertices.
(b) Prove that if (G) ≤ 4 and Mad(G) < 18 7
, then (G 2) ≤ 7. What does this imply
for planar graphs? (Cranston–Erman–Škrekovski [2014])

9.3.72. Choosability of denser squares. Let G be a graph.

(a) Prove that for 2 ≤ (G) ≤ (G) ≤ 4 and d(G) < 10 3
, some 2-vertex has a 3− -neighbor,
or a 3-vertex has two 3-neighbors, or a 4-vertex has a 2-neighbor and a 3− -neighbor.
(b) Prove that the result in part (a) is sharp by presenting infinitely many graphs with
d(G) = 103
and (G) = 4 that contain no such configuration.
(c) Prove that if (G) ≤ 4 and Mad(G) < 10 3
, then (G 2) ≤ 12. What does this imply
for planar graphs? (Cranston–Erman–Škrekovski [2014])

9.3.73. An injective coloring gives distinct colors to vertices with a common neighbor,
and the injective chromatic number, written i(G), is the minimum number of colors
in an injective coloring of G. Injective colorings need not be proper colorings.
(a) Determine the injective chromatic numbers of C5 and the Petersen graph.
(b) Let G be a graph with (G) = 2 and (G) = 3. Prove that if d(G) < 13 36
and G has
no 2-vertex in a triangle, then G has vertices x and y of degree 2 such that d(x , y) ≤ 3.
(Hint: 2-vertices may need more charge than their neighbors can provide.)
(c) Given (G) ≤ 3 and Mad(G) < 13 36
, prove i(G) ≤ 5. (Cranston–Kim–Yu [2010])
(d) Use the graph obtained by deleting one vertex from the Heawood graph (Exercise
6.2.27) to prove sharpness for parts (b) and (c).
Chapter 10

Ramsey Theory
Ramsey theory, named for Ramsey ’s Theorem, is the study of patterns in
partitions of large structures. A simple special case of Ramsey ’s Theorem is the
intuitively obvious Pigeonhole Principle. Partitioning theorems in general state
that some class in the partition must have special properties.
Motzkin described this informally by saying “Complete disorder is impossi-
ble”. Stargazers have always found interesting patterns of stars and given them
names; Ramsey Theory implies that any given pattern must occur (scaled) if there
are enough stars. Similarly, any sufficiently large amount of data (baseball statis-
tics, letters in the Bible, etc.) must have patterns with special structure.
Books on the subject include Graham–Rothschild–Spencer [1980, 1990],
Nešet řil–Rödl [1990], Soifer [2011], Prömel [2013], and Landman–Robertson
[2014], and recent developments are surveyed in Sudakov [2010] and Conlon–
Fox–Sudakov [2015]. Graham–Butler [2015] provides an introduction.

10.1. The Pigeonhole Principle

The Pigeonhole Principle (also called the Dirichlet Drawer Principle) is a sim-
ple idea with subtle applications. The subtlety often lies in recognizing when it
can be used. Its use can eliminate lengthy case analysis.
The Pigeonhole Principle implies that n + 1 shoes from a closet containing n
pairs of shoes must include a matched pair; they cannot all come from different
pairs. We have already used the Pigeonhole Principle in proving the Cycle Lemma
(Section 1.3), the existence of a cycle through any ¾ vertices in a ¾-connected
graph (Section 7.2), lower bounds on coloring (Section 8.1), Vizing ’s Theorem (Sec-
tion 8.3), etc. The basic idea is that every set of numbers has an element at least
as large as the average (also one at least as small).

10.1.1. Proposition. (Pigeonhole Principle) Placing more than ¾ n objects into ¾

classes puts more than n objects into some class.
Proof: With at most n objects per class, there are at most ¾ n objects.

425
426 Chapter 10: Ramsey Theory

A more general form, which generalizes to Ramsey ’s Theorem in Section

10.2, allows distinct thresholds (quotas) in different classes. In this section, we
only need the symmetric form above.

10.1.2. Theorem. (Pigeonhole Principle) If ( ∑ pi ) − ¾ + 1 objects are put into ¾

classes with quotas {pi}, then some class meets its quota.
Proof: If not, then at most ∑(pi − 1) objects can be accommodated.

CL ASSICAL APPLICATIONS

Rarely is the relevance of the Pigeonhole Principle apparent when a problem

is first stated. It may emerge only after the structure of the problem leads to a
notion of classes and objects so that having many (or few) of these objects in one
class yields a desired property. Often “many” means “more than one”.

10.1.3. Example. The (symmetric) acquaintance relation. In every set of at least

two people, some two people have the same number of acquaintances. Equiva-
lently, in every graph G some two vertices have the same degree. If | V(G)| = n,
then the degrees lie in {0 , . . . , n − 1}. Degrees n − 1 and 0 cannot both occur, so
at most n − 1 distinct degrees occur, and hence two must be equal.

10.1.4. Example. Midpoints between lattice points. Given five integer points in
the plane, the midpoint of the segment joining some pair also has integer coor-
dinates. The midpoint between (a , b) and (c , d) is ( a+ c b+ d
2 , 2
). This is an integer
point if and only if a and c have the same parity and b and d have the same parity.
Hence we group integer points by parity of coordinates: (odd/even, odd/even).
With five points, two must be in the same class and yield a lattice midpoint. With
four points, we may have one in each class and no lattice midpoint.

10.1.5. Example. Covering K n with bipartite graphs. We prove that the minimum
number of bipartite graphs with union K n is ⌈ log2 n⌉ . Equivalently, a complete
graph covered by ¾ bipartite subgraphs has at most 2¾ vertices. (Exercise 8.1.37
presents a generalization; the Pigeonhole Principle is not actually needed.)
Suppose that K n is covered by bipartite subgraphs G 1 , . . . , G ¾ , with G i hav-
ing bipartition {X i , Yi}. We may assume that each G i contains all the vertices,
since adding isolated vertices does not introduce odd cycles.
For each vertex v, define a binary ¾-tuple a by setting ai = 0 if v ∈ X i and
ai = 1 if v ∈ Yi . There are only 2¾ binary ¾-tuples. If n > 2¾ , then two vertices
have the same code. They lie in the same part in each G i , so the edge between
them is not covered. Thus n ≤ 2¾ .
To show that n = 2¾ is achievable, assign distinct binary ¾-tuples to the ver-
tices. The ith bipartite subgraph consists of all edges joining vertices with ith
coordinate 0 to vertices with ith coordinate 1. Since the ¾-tuples are distinct, ev-
ery edge is covered by some such subgraph. Indeed, edges joining complementary
codes appear in each subgraph.
Section 10.1: The Pigeonhole Principle 427

10.1.6. Example. Forcing divisible pairs. Erdős [1935] observed that any set of
n + 1 numbers in [2n] contains a pair such that one divides the other. This is
sharp: the n largest numbers in [2n] contain no such pair. To apply the Pigeon-
hole Principle, we partition [2n] into n classes such that for every two numbers
in the same class, one divides the other.
Every natural number is uniquely expressed as an odd number times a power

of 2. For fixed ¾ , the set {(2 ¾ − 1)2 j −1 : j ∈ } has the desired property; the smaller
of any two divides the larger. Since only n odd numbers are less than 2n, we have
n such classes. The ¾ th class is {m ∈ [2n]: m = (2 ¾ − 1)2 j −1 and j ∈ }. 
Like Example 10.1.5, Example 10.1.6 can be viewed as an extremal problem.
How many elements can be chosen from [2n] such that none divides another? The
Pigeonhole Principle may establish a bound, and a construction can show that the
bound is best possible. Nevertheless, presenting a construction and showing that
no element can be added does not prohibit larger configurations. For example,
we can avoid divisible pairs in Example 10.1.6 by first selecting primes. When
n = 5, this yields {2 , 3 , 5 , 7}, but now we cannot add any more elements. Our set
is maximal but not maximum-sized: {6 , 7 , 8 , 9 , 10} is a larger example.

10.1.7. Example. Inverses modulo n. If a and n are relatively prime, then ab ≡

1 (mod n) for some b ∈ {1 , . . . , n − 1}. Otherwise, a , 2a , . . . , (n − 1)a lie in n − 2
congruence classes other than 1. By the Pigeonhole Principle, two lie in the same
class. If they are ia and ja, then n | (i − j)a. Since a and n are relatively prime,
n | (i − j), a contradiction. Hence some multiple of a is congruent to 1 modulo n.
Similarly, an element of a group permutes the group elements via composi-
tion. Otherwise, it maps two elements into one, contradicting the uniqueness of
compositional inverses. We used this in Section 4.2.

10.1.8. Example. Domino tilings. A 6-by-6 checkerboard can be partitioned into

18 dominoes consisting of two squares each; this is a tiling by dominoes. We prove
that every such tiling can be cut between two adjacent rows or adjacent columns
without cutting any dominoes. In the example below, the tiling can be cut along
the middle horizontal line.

In a tiling, every domino cuts one line between adjacent rows or between ad-
jacent columns. There are 18 dominoes and 10 possible lines to be cut, so the
average number of cuts per line is 1.8. Since some line cuts at most the average,
some line is cut at most once, but perhaps still every line is cut at least once.
428 Chapter 10: Ramsey Theory

It suffices to show that every line is cut by an even number of dominoes, since
then a line cut by at most one is not cut at all. Since the number of squares on
one side of a line is even, and dominoes pair an even number of those, the number
covered by dominoes crossing the line must also be even.

The next theorem implies the result of Graham–Harary [1992] that every
spanning tree of Q¾ has diameter at least 2 ¾ − 1.

10.1.9. Theorem. (Graham–Entringer–Székely [1994]) If T is a spanning tree

of the ¾-dimensional cube Q¾ , then there is an edge of Q¾ outside T whose
addition to T creates a cycle of length at least 2 ¾ .
Proof: For each vertex v of Q¾ , let v denote the complementary vertex at distance
¾ from v. There is a unique v , v -path in T ; orient its first edge toward v . Since
T has more vertices than edges, the Pigeonhole Principle implies that some edge
receives orientations from both endpoints.
Since this edge uv receives an orientation from u and from v, we have v on
the u , u -path and u on the v , v -path in T . Hence the u , v -path and v , u -path in
T are disjoint. Each has length at least ¾ − 1, since the distance in Q¾ between
a vertex and its complement is ¾ . Finally, uv ∈ E(Q¾) implies also u v ∈ E(Q¾),
which completes a cycle of length at least 2 ¾ .

v u
• •

• •
u v

10.1.10.* Theorem. (Two Squares Theorem; Fermat) A natural number is

the sum of two squares if and only if all primes congruent to 3 modulo 4 occur
with even exponent in its prime factorization.
Proof: Necessity. Suppose that n = x2 + y 2 , and let p be a prime congruent to 3
modulo 4 that divides n. We will prove that p divides both x and y, which yields
p2 | n. Dividing the equation by p2 then expresses n/p2 as the sum of two squares,
and we can use induction on the sum of the exponents in the prime factorization.
By symmetry, it suffices to show that p divides x. If not, then x has a mul-
tiplicative inverse modulo p (Example 10.1.7). Since n ≡ 0 (mod p), multiplying
x 2 + y 2 = n by a number congruent to x−2 yields 1 + (x−1 y)2 ≡ 0 (mod p). How-
ever, when p = 4¾ + 3, applying Fermat ’s Little Theorem (Application 1.3.10) to
1 + a2 ≡ 0 (mod p) produces
1 ≡ a p−1 ≡ a4¾+2 ≡ 1 ¾ a2 ≡ −1 (mod p).
The contradiction yields p | x.
Sufficiency (Thue). If n1 = x12 + y12 and n2 = x22 + y22 , then n1 n2 = (x1 x2 +
y1 y2)2 + (x1 y2 − x2 y1)2 , so we need only consider n = 2, primes congruent to 1
modulo 4, and squares of other primes. Since 2 = 1 2 + 1 2 and p2 = p2 + 02 , it
suffices to decompose primes p such that p ≡ 1 (mod 4).
Section 10.1: The Pigeonhole Principle 429

For prime p with p ≡ 1 (mod 4), there is an integer √ a such that a2 + 1 ≡

0 (mod p) (see Exercise 22). Consider {ai + j: 0 ≤ i , j ≤ ⌊ p⌋}. The number of
√ 2
pairs (i , j) indexing this set is ( ⌊ p⌋ + 1) , which exceeds p.
By the Pigeonhole Principle, two of these pairs yield congruent numbers,
ai + j ≡ ai + j (mod p). Letting
√ y = j − j and x = i −2 i, we obtain y = ax with
x , y
≡ 0 (mod p) and | x| , | y| ≤ ⌊ p⌋ . The latter yields x + y 2 < 2p. We compute
x 2 + y 2 ≡ x2 + (ax)2 ≡ x2(a2 + 1) ≡ 0 (mod p).
Since 0 < x2 + y 2 < 2p, we conclude that x 2 + y 2 = p.

10.1.11. Example. The Chess Player Problem. A chess player wants to practice
for a match for 11 weeks. She wants to play at least one game per day but at most
132 games altogether. No matter how she schedules the games, there is a period
of consecutive days (among the 77 days) on which she plays exactly 22 games.
We can study totals over consecutive days by considering partial sums. Let ai
be the number of games played on days 1 through i, and set a0 = 0. The number of
games played on days i + 1 through j is aj − ai . We seek i , j such that ai + 22 = aj .
This suggests considering both {aj : 1 ≤ j ≤ 77} and {ai + 22: 0 ≤ i ≤ 76}. With at
least one game each day, each of these sets consists of 77 distinct numbers. Hence
repetition among these 154 numbers implies the desired result. Since a77 ≤ 132,
and a76 + 22 ≤ 153, we have 154 numbers in [153], and some number repeats.
Because a76 + 23 can be as large as 154, this argument does not force a period
of consecutive days with exactly ¾ games when ¾ ≥ 23.

With a total of at most b games on d consecutive days, the partial sum tech-
nique in Example 10.1.11 guarantees consecutive days with a total of ¾ games
when ¾ ≤ 2d − b. Another trick for ¾ ∈ {23 , 24 , 25} leads to a general solution.
For a circular analogue of the problem, see Clark–Lewis [1989, 1993].

10.1.12.* Theorem. (Hutchinson–Trow [1980]) For d , b ∈ 

with b > d, let ¾
be forced if every list of d positive integers summing to at most b has a con-
secutive segment summing to ¾ . A positive integer ¾ is forced if and only if
r < 2d − b, where d = m¾ + r with 0 ≤ r < ¾ .
Proof: Call a list of positive integers with sum at most b a game list, and call its
list of partial sums (increasing and bounded by b) a sum list. If a game list has a
consecutive segment summing to ¾ (so its sum list has two elements differing by
¾), then we say that it has property ¾ .
A game list alternating ¾ − 1 copies of 1 with one copy of ¾ + 1 does not have
property ¾ . Let a(¾ , d) be the corresponding sum list of length d. If d = m¾ + r,
then each segment of ¾ consecutive terms sums to 2 ¾ , and the final sum 2m¾ + r
equals 2d − r. If r ≥ 2d − b, then the list has sum at most b and lacks property ¾ .
entry 1 ··· 1 ¾+1 1 ··· 1 ¾+1 1 ··· 1
position 1 ··· ¾ − 1 ¾ ¾ + 1 · · · 2¾ − 1 2¾ 2¾ + 1 · · · d
To prove the converse, we show that 2m¾ + r is the least sum of a list of length
d lacking property ¾ . We prove this by induction on d, with d = 0 as a trivial
basis. If d > 0 and d is not a multiple of ¾ , then the sum must exceed the sum of
the initial segment of length d − 1, which also does not have property ¾ .
430 Chapter 10: Ramsey Theory

Hence we may assume that d = m¾ , so r = 0. Let S = [2m¾ − 1] − {¾}. If a

list with length d and sum less than 2m¾ lacks property ¾ , then its partial sums
consist of m¾ numbers among the 2m¾ − 2 numbers in S. Partition S into m¾ − 1
pairs by pairing x and x + ¾ whenever ⌊ x/¾ ⌋ is even, as shown below.
1 · · · ¾ − 1 2¾ · · · 3¾ − 1 4¾ · · · 2m¾ − ¾ − 1
¾ + 1 · · · 2 ¾ − 1 3¾ · · · 4 ¾ − 1 5 ¾ · · · 2m¾ − 1
By the Pigeonhole Principle, some pair contains two of the partial sums,
which implies that the list has property ¾ .

We have seen much variety in applications of the Pigeonhole Principle. The

classes may have different sizes. Classes are defined so that having many (or few)
objects in a class yields the desired outcome. Partial sums may help with numer-
ical problems involving order or sums. An extremal example can suggest classes
and objects for applying the Pigeonhole Principle to prove optimality. The prin-
ciple can be combined with proof by contradiction or with auxiliary results.

MONOTONE SUBLISTS

A classical use of the Pigeonhole Principle guarantees a long monotone sub-

list in a list of numbers. The result follows from Schensted ’s Theorem on Young
tableaux (Section 4.3), from Dilworth’s Theorem on partial orders (Section 12.1),
and from other results, but the Pigeonhole Principle gives a short direct proof.

10.1.13. Theorem. (Erdős–Szekeres Theorem; Erdős–Szekeres [1935]) Every

list of n2 + 1 distinct real numbers has a monotone sublist with n+ 1 numbers.
This is sharp; length n2 does not suffice.
Proof: (Seidenberg [1959]) For a given list a1 , . . . , an2 +1 , let x¾ and y¾ be the
lengths of a longest increasing and longest decreasing sublist ending at a¾ , re-
spectively (see example below). If no monotone sublist has length n + 1, then x¾
and y¾ never exceed n, and there are only n2 possible pairs (x¾ , y¾).
¾ 1 2 3 4 5 6 7 8 9 10
a¾ 7 4 1 8 5 2 9 6 3 10
x¾ 1 1 1 2 2 2 3 3 3 4
y¾ 1 2 3 1 2 3 1 2 3 1

By the Pigeonhole Principle, some two pairs are the same. However, when
i < j and ai < aj we can add aj to a longest increasing sublist ending at ai . When
i < j and ai ≥ aj , we can add aj to a longest decreasing sublist ending at ai .
Exercise 28 requests a list proving that the result is sharp.

Steele [1995] surveyed seven proofs of Theorem 10.1.13. An extension to

paths in digraphs is in Exercise 8.1.43; another generalization is Dilworth’s The-
orem (Exercise 12.1.15). For a higher-dimensional generalization, we allow repe-
titions in lists and extend “monotone” to strictly increasing or weakly decreasing:
a1 < · · · < a l or a1 ≥ · · · ≥ a l . The proof of Theorem 10.1.13 remains valid.
Section 10.1: The Pigeonhole Principle 431

10.1.14. Definition. A list of d-dimensional vectors is monotone if in each coor-

dinate it is monotone.

10.1.15. Theorem. (de Bruijn, unpublished; see Kruskal [1953]) A list of more
d
than l2 d-dimensional vectors has a monotone sublist with more than l vec-
d
tors. Length l2 does not suffice.
Proof: The Erdős–Szekeres Theorem is the case d = 1; we use induction on d.
d−1
For d > 1, the given list A has a sublist B of length more than l2 in which the
last coordinate is monotone. By the induction hypothesis, the (d − 1)-dimensional
list B formed by the first d − 1 coordinates of the vectors in B has a monotone
(d − 1)-dimensional sublist C of length more than l. Since the last coordinate is
monotone over all of B, the positions in B corresponding to C form a d-dimensional
monotone sublist of A.
Exercise 28 requests a list proving that the result is sharp.

By Theorem 10.1.13, every list of n2 numbers has a monotone sublist of length

at least n. A similar result holds for graphs with an ordering on the edges.
Graham–K leitman [1973] considered the case where the graph is K n . The idea
of the “Pedestrian Algorithm” in the general proof is due to Friedgut, as cited in
Winkler [2008].

10.1.16. Theorem. In any n-vertex graph with m edges having distinct numeri-
cal labels on the edges, there is an increasing trail of length at least 2m/n.
Proof: Consider n pedestrians on the graph, starting one at each vertex. Process
each edge, in order of the labels, by switching the pedestrians on its endpoints
and adding 1 to the number of times each of those two have moved. Always each
vertex has one pedestrian. Each pedestrian follows an increasing trail. The total
movement is 2m. By the Pigeonhole Principle, some pedestrian follows a trail of
length at least 2m/n.

PATTERN-AVOIDING PERMUTATIONS (optional)

An increasing sublist of fixed length in a permutation is an example of a

“pattern”. The Pigeonhole Principle was applied to settle a long-standing open
problem about the number of permutations of [n] avoiding a given pattern.

10.1.17. Definition. A permutation of [ ] is a pattern in a permutation  of

[n] if the word form of  has a -element sublist  such that the element in
position i of  is the  i th smallest element in  . A permutation of [n] is  -
avoiding if it does not contain  as a pattern. Let Sn( ) denote the number
of  -avoiding permutations of [n].

For example, the pattern 4132 occurs in 18256734 at positions 2 , 3 , 6 , 8; the

sublist 8274 has this pattern. An increasing sublist is an occurrence of the iden-
tity pattern. The Erdős–Szekeres Theorem states that no permutation of length
2
+ 1 avoids both the identity pattern of length + 1 and its reverse.
432 Chapter 10: Ramsey Theory

10.1.18. Example. An inversion is an occurrence of the pattern 21; only iden-

tity permutations avoid this. Proposition 3.1.15 gives the generating function
for permutations of [n] by the number of occurrences of the pattern 21.
Exercise 2.1.48 shows that Sn(231) is the Catalan number Cn . By combin-
ing Schensted ’s Theorem (Corollary 4.3.20), the Hook-Length Formula (Theorem
4.3.4), and Exercise 4.3.11, it follows that also Sn(321) = Cn .
Given a permutation of [n], form by subtracting each entry from n + 1. A
-element pattern in is obtained by subtracting a -element pattern in from
 + 1. Hence Sn(213) = Sn(231). Also the reverse of a  -avoiding permutation
avoids the reverse of  . We conclude that Sn( ) = Cn for every 3-element  .

The equalities of Example 10.1.18 fail for longer patterns. Bóna [1997] gave
the first example of  and  of equal length such that Sn( )
= Sn( ), proving
Sn(1234) < Sn(1324). (For this and other pattern results, see Bóna [2004].)
One property of 3-element patterns does generalize. Note that n+1 1 (2n n
) ∼
cn − 4 for a constant c (Example 2.3.10). Stanley and Wilf conjectured around
3/2 n

1990 that Sn( ) is always bounded by an exponential function of n whose ra-

tio to n! tends to 0. (They conjectured the seemingly stronger statement that
limn→∞(Sn( ))1/n exists, but by Arratia [1999] the two are equivalent.)
Among various partial results, Bóna [1999] proved the conjecture for permu-
tations where each run consists of consecutive elements, such as 456312. Finally,
Marcus–Tardos [2004] used the Pigeonhole Principle to prove another conjecture
that implies the Stanley–Wilf Conjecture.

10.1.19. Definition. Let A and P be 0 , 1-matrices of orders n and  with n ≥  .

The matrix A covers P if A has a submatrix B such that Bi , j = 1 whenever
Pi , j = 1 (rows and columns from A are indexed in the same order in B). If A
does not cover P , then A avoids P.

10.1.20. Conjecture. (Stanley–Wilf Conjecture) For every pattern  , the

number Sn( ) of  -avoiding permutations of [n] is at most exponential in n.
(Füredi–Hajnal Conjecture; Füredi–Hajnal [1992]) For every permu-
tation matrix P , there is a constant cP such that the number of 1s in any
P-avoiding 0 , 1-matrix of order n is at most cP n.

10.1.21. Lemma. (K lazar [2000]) The Füredi–Hajnal Conjecture implies the

Stanley–Wilf Conjecture.
Proof: A permutation avoids a pattern  if and only if the matrix of avoids
the matrix of  . Given a permutation matrix P , let T n be the set of P-avoiding
0 , 1-matrices of order n. It suffices to show that | T n| ≤ cn for some constant c,
since the permutation matrices avoiding P lie in T n .
Consider A ∈ T2n , so A avoids P. Partition A into 2-by-2 blocks. Form an n-
by-n matrix A by replacing each all-0 block with a 0 and each nonzero block with
a 1. If A covers P , then since P only has one 1 in each row and column, we can
take any 1 from the corresponding block in A to play the role of the 1 in A that
covers a 1 in P ; hence A covers P. By this contradiction, A avoids P.
There are 15 2-by-2 matrices that can collapse to a 1 in A and only one that
can collapse to a 0. Therefore, each P-avoiding matrix A arises by collapsing at
Section 10.1: The Pigeonhole Principle 433

most 15º (n) P-avoiding matrices of order 2n, where º (n) is the maximum number
of 1s in a P-avoiding matrix of order n.
We have proved | T2n | ≤ | T n| 15º (n) . The Füredi–Hajnal Conjecture is º (n) ≤
¾ −1
cP n. For n = 2¾ , iterating the recurrence yields | T n | ≤ | T1 | 15cP ¾=1 2 ≤ (15cP )n ,
n

since | T1 | = 2. When n is not a power of 2, let m be the next larger power of 2;

note that m < 2n. Hence | T n| ≤ | T m| ≤ (15cP )2n . With | T n| ≤ (15cP )2n for all n, we
obtain the Stanley–Wilf Conjecture with the constant 152cP .

10.1.22. Theorem. (Marcus–Tardos [2004]) Let º (n) be the maximum number

of 1s in a P-avoiding 0 , 1-matrix of order n. If P is a permutation matrix of
order ¾ , then º (n) ≤ 4¾ 4 (¾¾ )n, so the Füredi–Hajnal Conjecture is true.
2

Proof: Suppose first that n is a multiple of ¾ 2 . Let A be a P-avoiding 0 , 1-matrix

of order n. Again we use the collapsing technique to obtain a recursive bound.
Partition A into blocks of order ¾ 2 by splitting the matrix into n/¾ 2 sets of con-
secutive rows and n/¾ 2 sets of consecutive columns.
Let A be the matrix of order n/¾ 2 obtained by collapsing blocks of A into
single entries with a block becoming a 0 if and only if it is an all-0 block. The
matrix A avoids P , since otherwise taking any 1 from each block corresponding
to a position in A that represents a 1 in P yields a submatrix of A that covers P.
Let a block be wide [tall] if it has 1s in at least ¾ columns [rows]. We claim
that each column of blocks has at most ¾ (¾¾ ) wide blocks (and each row has at most
2

¾ (¾¾ ) tall blocks). Each wide block has ¾ nonzero columns among its ¾ 2 columns,
2

and there are (¾¾ ) possible such sets. With ¾ (¾¾ ) wide blocks in a column of blocks,
2 2

the some ¾ wide blocks are nonzero in the same set S of ¾ columns. We use the
ith highest among these ¾ blocks to cover the 1 in the ith row of P , using the
appropriate column in S. That is, in this column of S the block has a 1 in some
row, and we use that row from this block. This contradicts that A avoids P.
We thus have at most ¾n2 ¾ (¾¾ ) wide blocks and the same bound on the number
2

of tall blocks. Since A avoids P , at most º (n/¾ 2) nonzero blocks are neither wide
nor tall. Such a block has at most (¾ − 1)2 nonzero entries, and we use the trivial
bound of ¾ 4 on the nonzero entries in wide and tall blocks. We now have
n ¾2 n
º (n) ≤ 2 ¾ 4 ¾ ( ) + (¾ − 1)2 º ( 2 ) .
¾ 2 ¾ ¾
When n is not a multiple of ¾ 2 , consider a submatrix whose order is the next
smaller multiple of ¾ 2 ; call it n . Fewer than 2 ¾ 2 n positions are lost; at worst
they could all be 1. We now prove the desired bound by induction. We compute
n ¾2
º (n) < º (n ) + 2 ¾ 2 n ≤ (¾ − 1)2 º ( 2 ) + 2 ¾ 3 ( )n + 2 ¾ 2 n
¾ ¾
¾ 2
n ¾ 2
≤ (¾ − 1)2 [2¾ 4 ( ) ] + 2 ¾ 3 ( )n + 2 ¾ 2 n
¾ ¾ 2 ¾
¾ 2
¾2
≤ 2¾ 2 ((¾ − 1)2 + ¾ + 1) ( )n ≤ 2¾ 4 ( )n ,
¾ ¾

where the last inequality holds for ¾ ≥ 2.

434 Chapter 10: Ramsey Theory

10.1.23. Corollary. If is a permutation of [ ], then Sn( ) ≤ cn , where c =

4 ¾2
152¾ ( ¾ ). That is, the Stanley–Wilf Conjecture is true.
The bound on c in Corollary 10.1.23 is quite large; c = 4 is adequate when
= 3. The best constant is ( − 1)2 when  is the identity permutation (Regev
[1981]). Arratia [1999] conjectured that c = ( − 1)2 always suffices, but this was
disproved at = 4 in Albert et al. [2006] (for  = 4231).
Wilf later asked also how many different patterns can occur in a permutation
of [n]. A trivial upper bound is 2 n . Wilf provided a sequence of examples where
the number of patterns grows as fast as the Fibonacci numbers, and Coleman
[2004] improved this to 2(1−o(1))n (Exercise 37).

L ARGE GIRTH AND CHROMATIC NUMBER (optional)

Erdős proved by probabilistic arguments that graphs exist with large girth
and large chromatic number. Explicit constructions came later. The construction
we present intimately involves the Pigeonhole Principle, applying it iteratively
in a way that generates very large structures to ensure the existence of a special
substructure. (The proof of Ramsey ’s Theorem in Section 10.2 will also do this.)
To motivate the hypergraph-based construction, we first construct graphs
with girth 6 and large chromatic number. This may be the earliest construction of
triangle-free graphs with large chromatic number. It appeared in an undergrad-
uate magazine at Cambridge University. “Blanche Descartes” was a pseudonym
used by R.L. Brooks, C.A.B. Smith, A.H. Stone, and W.T. Tutte when under-
graduates together at Cambridge; all became mathematicians.

10.1.24. Example. Graphs with girth 6 and high chromatic number (Blanche
Descartes [1947, 1954], Kelly–Kelly [1954]). Like Mycielski’s construction, this
is inductive; we start with C6 for = 2. The construction increases chromatic
number but preserves girth 6. Let G be a ( − 1)-chromatic graph with girth 6.
To construct G , let r = | V(G)| and N = (r − 1)( − 1) + 1. Begin with a set
S of N isolated vertices. Add ( Nr ) copies of G, one for each r-subset of S. For
1 ≤ i ≤ ( Nr ), add a matching from the ith copy of G to the ith r-set in S. The
resulting graph is G . A cycle visiting more than one copy of G has at least one
interior edge and two departing edges from each copy it visits. Thus cycles in G
not in one copy of G have at least six edges.
A proper ( − 1)-coloring of G can be extended to a -coloring of G by using
that ( − 1)-coloring on each copy of G and using a new color on S. In any ( − 1)-
coloring of G , by the Pigeonhole Principle some r vertices of S have the same
color. This color is forbidden from the copy of G corresponding to this r-set. Since
this copy of G cannot be properly ( − 2)-colored, we conclude that (G ) = .

G G G G
•••• •••• ( Nr ) copies

•••••• S
Section 10.1: The Pigeonhole Principle 435

In an early triumph for probabilistic arguments, Erdős [1959] showed that

for appropriate sparseness, almost every graph has chromatic number and girth
above fixed constants (Theorem 14.2.3). By extending the question to hyper-
graphs, Lovász [1968b] constructed such graphs explicitly. We present a sim-
plification from Nešet řil–Rödl [1979]. Other constructions appear in Lubotzky–
Phillips–Sarnak [1988] and Kriz [1989]; see also Nešet řil [2013] and Alon–
Kostochka–Reiniger–West–Zhu [2016] (Exercise 40).

10.1.25. Definition. A hypergraph H consists of a vertex set V(H) and an edge

set E(H) such that each edge is a subset of V(H). It is r-uniform if all edges
have size r. In a complete r-uniform hypergraph, all r-sets are edges.
A proper ¾-coloring of a hypergraph H labels V(H) from a set of ¾ colors
so that no edge with size at least 2 is monochromatic; the least ¾ needed is the
chromatic number (H). A hypergraph H is -chromatic when (H) =  .

Proper coloring of 2-uniform hypergraphs is just proper coloring of graphs.

Since the input structures are more general, hypergraph -coloring is at least as
hard as graph -coloring; it is NP-complete to decide whether a general hyper-
graph has a proper 2-coloring.

10.1.26. Example. Let F be the 3-uniform 7-vertex hypergraph with edge set
{124, 235, 346, 457, 561, 672, 713}. Three consecutive vertices contain no edge,
so (F) ≤ 3. To show (F) = 3, consider a 2-coloring; some four vertices have the
same color. Any four vertices not having three consecutive include three vertices
of the form i , i + 1 , i + 3, forming an edge. A 4-set containing i , i + 1 , i + 2 avoids
completing an edge only by adding i + 5, but then the three vertices with the
other color form an edge. Hence every 2-coloring has a monochromatic edge. (This
hypergraph is known as the Fano plane; see Chapter 13).

10.1.27. Example. Complete r-uniform N-vertex hypergraph. An independent

set in a hypergraph, as in a graph, is a set containing no edge of size at least
2. If (H) denotes the maximum size of an independent set in H , then (H) ≥
|V(H)| /(H), as for graphs. In the complete r-uniform hypergraph, every r-set is
an edge, so (H) = min{r − 1 , N} and (H) = ⌈ N/(r − 1)⌉ .

10.1.28. Definition. For l ≥ 2, a cycle of length l in a hypergraph is a cyclic list

[x1 , e1 , . . . , x l , e l] of l distinct vertices and l distinct edges such that x i and
x i+1 lie in e i (indices modulo l). The girth of H is the length of a shortest
cycle. A hypergraph is N-partite if its vertices can be partitioned into N
disjoint classes such that each edge has at most one vertex in each class.

Cycles in 2-uniform hypergraphs can be viewed as cycles in graphs. However,

when naming a cycle in a graph we normally specify only the edges, because in a
2-uniform hypergraph two vertices can determine only one edge. Hence graphs
have no cycles of length 2, though hypergraphs may. Cycles as defined above are
often called Berge cycles. This is the most general notion of cycle in hyper-
graphs; there are more restrictive definitions. In multihypergraphs, there may
be distinct edges with the same set of vertices; thus three edges that have the
same set of vertices in a 3-uniform multihypergraph form a Berge cycle.
436 Chapter 10: Ramsey Theory

We will build an r-uniform ¾-chromatic hypergraph with girth at least l,

using induction on l. Setting r = 2 yields the desired graphs. The inductive con-
struction involving N-partite hypergraphs uses hypergraphs with large edges.

10.1.29. Example. A bigger hypergraph. From an N-partite hypergraph H and

another hypergraph Y , we build a large N-partite hypergraph ºj (H, Y), where
j ∈ [N]. We require Y to be s-uniform, where s is the size of the jth class in H.
Let m be the number of edges in Y . We use m copies of H. For i ∈ [m], let H i
be a copy of H using the vertices of the ith edge in Y as its jth class. The copies
of H are disjoint except for the jth class. Below we sketch the construction for
N = 3 and m = 5, with A1 , . . . , A5 denoting the edges of Y . Each solid ellipse
represents a copy of H ; the three regions inside represent the three classes.
An edge of Y is not an edge in ºj (H, Y); it is just the jth class in a copy of
H contained in ºj (H, Y). The vertex set V(Y) becomes the jth class in ºj (H, Y),
and for ¾
= j the ¾ th class in ºj (H, Y) is the union of the (disjoint) ¾ th classes in
H 1 , . . . , H m . Thus the ¾ th class in ºj (H, Y) is large. If we use the construction
again with ¾ in place of j and ºj (H, Y) in place of H , then the new auxiliary Y
will need large edges.

H2

H1 A2 H3
A1 A3

A5 A4
H5 H4
V(Y)

10.1.30. Lemma. (Nešet řil–Rödl [1979]) If H is an N-partite hypergraph with

girth at least l and jth class of size s, and Y is an s-uniform hypergraph
with girth at least l/2, then the hypergraph ºj (H, Y) constructed in Example
10.1.29 is N-partite with girth at least l.
Proof: Let the vertex classes of H i be X 1i , . . . , X ji−1 , Ai , X ji+1 , . . . , X N
i
. Since H is
N-partite, no edge of ºj (H, Y) has two vertices in V(Y), so ºj (H, Y) is N-partite
with classes ⋃i X 1i , . . . , ⋃i X ij −1 , V(Y), and ⋃i X ij +1 , . . . , ⋃i X N i
. Each edge of
ºj (H, Y) has the same size as the corresponding edge in H.
Let C be a cycle in ºj (H, Y). If C is contained in one copy of H , then C has
length at least l (with equality for some such C). Suppose that C visits more than
one copy of H. Since no edge of ºj (H, Y) contains two elements of V(Y), the cycle
C cannot have consecutive vertices in V(Y). Also, C can pass from V(Hi) − V(Y)
to V(Hj ) − V(Y) only through a vertex of Ai ∩ A j . Hence the vertices of V(Y) used
by C form a cycle in Y . Since C contains at least one vertex not in V(Y) between
any two successive vertices in V(Y), the length of each such cycle is at least l.
Section 10.1: The Pigeonhole Principle 437

10.1.31. Theorem. (Lovász [1968b]) For integers r, ¾ , l ≥ 2, there is an r-

uniform ¾-chromatic hypergraph with girth at least l.
Proof: (Nešet řil–Rödl [1979]) We construct an r-uniform hypergraph H(r, ¾ , l)
with girth at least l and chromatic number at least ¾ . Reducing chromatic num-
ber to ¾ by deleting edges does not decrease the girth.
We use induction on l. For r, ¾ ≥ 2, let N = (r − 1)(¾ − 1) + 1. Let H(r, ¾ , 2)
be the complete r-uniform hypergraph with N vertices. When ¾ = 2, there is one
edge and no cycle, which is treated as infinite girth, and the chromatic number
is 2. For l > 2, we assume that H(t , ¾ , u) has already been constructed for all
{t , ¾ , u} with t ≥ 2 and 2 ≤ u < l.
We construct hypergraphs M1 , . . . , MN +1 and let H(r, ¾ , l) = MN +1 . Let M1
be the N-partite r-uniform hypergraph having vertex classes of size ( Nr−−11 ) and ( Nr )
pairwise disjoint edges. There is one edge for each choice of r classes, consisting
of one vertex from each chosen class. The edges are disjoint, so there is no cycle:
again infinite girth.
Each successive hypergraph Mi will be r-uniform, N-partite, and have girth
at least l, constructed by applying Lemma 10.1.30 with H = Mi−1 . We will ar-
gue that every (¾ − 1)-coloring of MN +1 contains, among its astronomically many
copies of M1 , some copy of M1 in which every class has a single color on its ver-
tices. Since M1 has (r − 1)(¾ − 1) + 1 classes, some r of those classes have the same
color. By construction, M1 has some edge consisting of a vertex from each of those
r classes. This edge is monochromatic, and hence MN +1 is not (¾ − 1)-colorable.
Having constructed Mi , let si be the size of the ith class in Mi . Let Yi =
H(si , ¾ , l − 1). Since Yi is si-uniform, we may define Mi+1 = ºi(Mi , Yi) as in
Lemma 10.1.30. Note that si+1 = si | E(Yi)| , so the construction needs smaller-
girth hypergraphs with huge edges. By Lemma 10.1.30, Mi+1 is r-uniform and
N-partite and has girth l.
Let º be a (¾ − 1)-coloring of MN +1 . Note that MN +1 has a copy of MN for each
edge A of YN , in which A serves as the Nth class of MN . Since YN is not (¾ − 1)-
colorable and º uses only ¾ − 1 colors on V(YN), the vertices of some edge A of
YN (it is not an edge in MN +1) receive the same color under º . The copy of MN
corresponding to A is thus a copy of MN where º uses one color on the Nth class.
Within this copy L N of MN , we apply the argument inductively. In L N , the
(N − 1)th class is a copy of V(YN −1), and L N contains a copy of MN −1 for each edge
of YN −1 . The Nth class of L N contains the Nth class of each copy of MN −1 used
in forming it. By the choice of L N , the Nth class of each copy of MN −1 in L N is
monochromatic under º . Since YN −1 is not (¾ − 1)-colorable, L N contains a copy
of MN −1 in which the (N − 1)th class is the vertex set of an edge of YN −1 that is
monochromatic under º .
At stage i, working back, we have a copy Li of Mi in which classes i through
N are monochromatic under º . Hence they are also monochromatic in each copy
of Mi−1 contained in Li . Since Yi−1 is not (¾ − 1)-colorable, in one of these copies
the (i − 1)th part also is monochromatic under º . Iterating through i = 2, we
obtain a copy of M1 in which every class is monochromatic under º , yielding a
monochromatic edge as discussed earlier. Hence every (¾ − 1)-coloring of MN +1
contains a monochromatic edge, and we set H(r, ¾ , l) = MN +1 .
438 Chapter 10: Ramsey Theory

EDGE-COLORING OF HYPERGRAPHS (optional)

Like vertex coloring, proper edge-coloring also generalizes naturally to hyper-

graphs; the edges containing a given vertex must have distinct colors. We seek
an analogue of Vizing ’s Theorem.

10.1.32. Definition. A hypergraph is linear if no two edges share more than

one vertex. A loop is an edge of size 1. A simple hypergraph is a linear
hypergraph without loops.

Bounds for edge-coloring of simple hypergraphs have been hard to find. We

consider a long-standing conjecture in this area that Erdős described as one of
his favorite problems. It originated at a party in 1972 in Boulder, Colorado. The
proposers agreed to meet the next day to write out the solution, but they failed.
Erdős [1976] offered 50 British pounds for the solution, later raised to $500.

10.1.33. Conjecture. (Erdős–Faber–Lovász Conjecture) Any union of n edge-

disjoint copies of K n has chromatic number n.

10.1.34. Conjecture. Every simple hypergraph with n vertices is properly n-

edge-colorable.

The first conjecture sounds a bit artificial, but they are equivalent, and both
are known as the Erdős–Faber–Lovász Conjecture. Conjecture 10.1.33 is a way
of stating Conjecture 10.1.34 in the language of ordinary graph coloring. The
transformation uses hypergraph duality.

10.1.35. Definition. The dual of a hypergraph H is the hypergraph H ∗ whose

incidence matrix is the transpose of the incidence matrix of H.

In forming H ∗ from H , vertices and edges exchange roles, but the incidences
remain the same (multiedges arise in H ∗ when vertices in H have the same in-
cident edges). Proper edge-coloring of H is more restrictive than proper vertex
coloring of H ∗ ; it is equivalent to coloring V(H ∗) so that in each edge of H ∗ the
vertices all have distinct colors.

10.1.36. Theorem. Conjectures 10.1.33–10.1.34 are equivalent.

Proof: Let G be a union of n pairwise edge-disjoint copies of K n with vertex sets
Q1 , . . . , Qn . Each Qi has an element in no other Qj , since each Qj has at most one
vertex of Qi . Let Q i be the proper subset of Qi obtained by deleting the elements
in no other Qj . It suffices to show that the union G of complete graphs on the sets
Q1 , . . . , Q n is n-colorable, since the coloring extends to the missing vertices.
View Q1 , . . . , Q n as edges of a hypergraph H. Vertices in an edge are adja-
cent in the original graph, so we want them to have distinct colors. Hence the
vertex coloring problem for G becomes the edge-coloring problem for H ∗ . The n
original cliques in G become the n vertices of H ∗ .
The original edge-disjointness condition becomes the condition that two edges
of H share at most one vertex. Thus H is linear. A hypergraph is linear if and
Section 10.1: The Pigeonhole Principle 439

only if its dual is linear, since both mean forbidding a 2-by-2 all-1 submatrix in
the incidence matrix. By discarding elements of V that lie in only one clique in
G, we have ensured that the dual hypergraph H ∗ has no edge of size 1. Thus H ∗
is a simple hypergraph, and G is n-colorable if and only if H ∗ is n-edge-colorable.
The transformation from G to H ∗ is reversible. A simple hypergraph H ∗ with
n vertices has each vertex in at most n − 1 edges, since no other vertex is in two
such edges. Hence the dual hypergraph H is a simple hypergraph with n edges of
size at most n − 1. Augmenting these edges gives Q1 , . . . , Qn and G.

A simple hypergraph whose edges all have size 2 is a graph. An n-vertex

graph has maximum degree at most n − 1, and thus Vizing ’s Theorem yields n-
edge-colorability in that case. Hindman [1981] proved Conjecture 10.1.34 in the
case where the union of the edges with at least three vertices has size at most 10.
K ahn [1997] proved that Conjecture 10.1.34 holds asymptotically. That is,
there is a proper n(1 + o(1))-edge-coloring (here o(1) represents some function of n
that tends to 0 as n tends to ∞).
An upper bound of 2n − 3 is easy. Color the edges in decreasing order of size,
greedily using the least available color. An edge of size ¾ intersects at most ¾n−−1¾
edges of size at least ¾ at each vertex, so the number of previously-colored edges
intersecting the current edge is at most (n¾−−¾1)¾ . For simple hypergraphs, this is
largest at ¾ = 2 and always at most 2n − 4. Chang and Lawler extended this
argument to reduce the upper bound.

10.1.37. Theorem. (Chang–Lawler [1988]) The edge-chromatic number of a sim-

ple hypergraph on n vertices is at most ⌈ 3n/2⌉ − 2.
Proof: Color the edges in non-increasing size order. For an edge of size ¾ , at most
(n−¾)¾
¾ −1 edges intersecting it have already been colored. For ¾ ≥ 3 this is bounded
by (3n − 9)/2, which is less than ⌈ 3n/2⌉ − 3, so a color is available for the edge.
For an edge uv of size 2, if some color avoids both vertices (c avoids v if it has not
already been used on an edge containing v), then we can color the edge.
Otherwise, we recolor an edge to free a color for uv. Suppose that a avoiding v
and b avoiding u appear on edges uw and vw of size 2, for some vertex w (as shown
below). Since each vertex is in at most n − 1 edges and uv is uncolored, at most
n − 2, n − 2, and n − 1 colors appear at u, v, or w, respectively. Hence at least
n/2, n/2, and n/2 − 1 colors avoid u, v, and w, respectively. With only ⌈ 3n/2⌉ − 2
colors, some color avoids two of these vertices. No color avoids both u and v, so we
may assume that c avoids w and u. We change the color of uw to c and use a on uv.
u v
• •
no b no a
a b
•
w
Hence it suffices to find such a vertex w when no color is available for uv. Let
A be the set of colors used on edges of size 2 at u and not on larger edges containing
v. Similarly, let B be the set of colors used on edges of size 2 at v and not on larger
edges at u. Colors in A − B do not appear at v, so edges at u with these colors are
440 Chapter 10: Ramsey Theory

candidates for the edge uw. Similarly, colors in B − A do not appear at u. Since
colors in A − B specify distinct edges containing u and colors in B − A specify
distinct edges containing v, it suffices to prove | A − B| + | B − A| ≥ n − 1.
We count the ⌈ 3n/2⌉ − 2 colors present at u and/or v. A color not in A ∪ B
appears on an edge of size at least 3 containing u or v. Since each color in A elim-
inates a potential neighbor of u from such edges and we also exclude {u , v}, there
are at most (n − | A| − 2)/2 such edges containing u. Similarly, there are at most
(n − | B| − 2)/2 such edges containing v. Having counted all the colors, we have
| A ∪ B| + (n − | A| − 2)/2 + (n − | B| − 2)/2 ≥ ⌈ 3n/2⌉ − 2 ≥ (3n − 5)/2.
This simplifies to 2 | A ∪ B| − (| A| + | B|) ≥ n − 1. Since the left side equals | A − B| +
| B − A| , the proof is complete.

EXERCISES 10.1

10.1.1. (−) Prove that every graph with at least three vertices, all of even degree, has at
least three vertices with the same degree.
10.1.2. (−) Prove that every set of 2 n + 1 integer lattice points in n
contains a pair of
points whose centroid (mean vector) is also an integer lattice point.
10.1.3. (−) Prove that any√set of five points in a square of area 1 contains two points sepa-
rated by distance at most 2/2 and that no closer distance can be forced.
10.1.4. (−) Determine the least ¾ such that every ¾-element subset of [2n] contains two
relatively prime elements.
10.1.5. (−) Determine the least ¾ such that every ¾-element subset of [3n] contains three
consecutive numbers.

10.1.6. Given n , ¾ ∈ , determine the maximum size of a subset of [n] that has no two
numbers differing by ¾ .
10.1.7. (♦) Determine the least n such that every set of n integers contains two elements
whose sum or difference is a multiple of ¾ .
10.1.8. Prove that every set of 2 ¾ integers has elements of opposite parity differing by at
least 2 ¾ − 1 or elements of equal parity differing by at least 4 ¾ − 2.
10.1.9. Use the Pigeonhole Principle to prove that for every rational number, the decimal
expansion is eventually repeating.
10.1.10. Generalization of Example 10.1.6. Show that every set of (2 m − 1)n + 1 distinct
integers chosen from {1 , . . . , 2 m n} contains m + 1 distinct integers a0 , . . . , am such that
ai−1 divides ai for 1 ≤ i ≤ m. (Khare [1989])
10.1.11. Consider the grid [n] × [n] of n2 points in the plane. Each point is colored black or
white. How large must n be so that every such coloring yields a rectangle with horizontal
and vertical sides whose corners have the same color?
10.1.12. An exam consists of ¾ true/false questions. Determine the least n such that
whenever n students take the exam, some two students agree in their answers to at least
one-third of the questions.
10.1.13. Let T be a complete ternary tree with ¾ levels whose 3¾ leaves are 2-colored. Prove
that in some complete binary subtree (2 ¾ leaves), all leaves have the same color. (Milans)
Exercises for Section 10.1 441

10.1.14. In a chess tournament , each person plays every other, scoring 1 for a win, 0 for
a draw, and −1 for a loss. Prove that if at least 3/4 of the games are draws, then some two
players score the same total. Show that this is asymptotically sharp. (Bóna [2002, p. 50])
10.1.15. In the chess tournament of Exercise 10.1.14, suppose that the players come from
two countries (each person plays all others). Prove that some player scores at least as many
points against players from the same country as against players from the other country.
10.1.16. (♦) Determine the least n such that every set of n integers has a nonempty subset
whose sum is a multiple of ¾ . (Comment: A more general and stronger statement valid for
any group (here ¾) appears in Moser [1948]. See also Vince [1965] for the special case.)
10.1.17. Let x be a real number. Show that some number in {x , 2x , . . . , (n − 1)x} differs
by at most 1/n from an integer.
10.1.18. (♦) Let º (n) be the least ¾ such that every set of ¾ elements in [n] has two disjoint
subsets with the same sum. Prove 1 + ⌊ log 2 n⌋ < º (n) ≤ ⌈ 1 + log 2 n + log 2 log 2 n⌉ for n ≥ 3.
10.1.19. For n < 2 ¾ , prove that any multiset of ¾ positive integers with sum n has a subset
with sum j for all j ∈ [n], and show that this fails when n = 2 ¾ . (Brown [1961], Ganter–
Teirlinck [1977])
10.1.20. Let S1 , . . . , Sm be a partition of [n] with at least two sets having size more than
m. Prove that there are distinct indices ¾ , l ∈ [m] such that some difference between two
elements of S¾ is also a difference between two elements of Sl . (LeSaulnier [2006])
10.1.21. Use Example 10.1.7 to prove the Chinese Remainder Theorem: If n1 , . . . , nr
are pairwise relatively prime natural numbers, and a1 , . . . , ar are integers, then one
congruence class modulo ∏ ni is congruent to ai modulo ni for each i. (Hint: Consider
x = ∑i=1 aj Nj yj , where Nj = (∏ ni)/nj and yj is chosen using multiplicative inverses.)
r

10.1.22. Given a prime p, let p be the field of congruence classes modulo p.

(a) The order of a nonzero element a ∈ p is the least ¾ such that a¾ ≡ 1 (mod p). Us-
ing multiplication by a on p , conclude that ap−1 ≡ 1 (mod p). This statement is Fermat ’s
Little Theorem, also proved in Application 1.3.10.
(b) Prove that if a2 ≡ 1 (mod p), then a ≡ ±1 (mod p). Use this and multiplicative in-
verses to prove Wilson’s Theorem: (p − 1)! ≡ −1 (mod p).
(c) Let r = (p − 1)/2. Use Fermat ’s Theorem to show that x2 ≡ −1 (mod p) has no solu-
tion when p ≡ 3 (mod 4). Use Wilson’s Theorem to show that x2 ≡ −1 (mod p) has ±(r!) as
solutions when p ≡ 1 (mod 4). (Hint: Modify (p − 1)! to express it in terms of r.)
10.1.23. (♦) For m ≥ 2n, let S be a set of m points on a circle, no two on a diameter. Say
that x ∈ S is “free” if fewer than n points of S − x lie in the semicircle clockwise from x.
Prove that S has at most n free points. (Knuth [1991])
10.1.24. (♦) Use the technique of Theorem 10.1.9 (not induction!) to prove the following
two statements about trees.
(a) The center of a tree T consists of one vertex or two adjacent vertices (see Defini-
tion 5.4.10). (Graham–Entringer–Székely [1994])
(b) If T1 , . . . , T¾ are subtrees of T that pairwise intersect , then some vertex of T
belongs to all of T1 , . . . , T¾ . (Lehel)
10.1.25. (♦) A club has 90 rooms and 100 members. Keys must be distributed so that any
90 members can fill the 90 rooms without sharing or trading keys. Prove that 990 keys are
enough and 989 keys are not enough. (Members may have more than one key, and keys to
a room may be given to more than one member.)
10.1.26. Let x1 , x2 , · · · be a sequence drawn from a finite alphabet. Prove that there is a
positive integer ¾ such that x n+¾ = x n for infinitely many n.
442 Chapter 10: Ramsey Theory


10.1.27. For n ∈ , prove that every list of more than n3 numbers has a sublist with more
than n entries that is strictly increasing or strictly decreasing or constant. Construct an
example to show that the result is sharp.
10.1.28. Construct a list of length n2 having no monotone sublist of length more than n.
d
Extend this by induction to construct a list of l 2 vectors in d having no monotone sublist
of length more than l. (Alon–Füredi–Katchalski [1985])
10.1.29. (♦) Let Cn¾ be the 2 ¾-regular graph obtained from Cn by making any two vertices
adjacent if the distance between them in Cn is at most ¾ . For ¾ ≥ 2, prove that (Cn¾) >  + 2
if n = ( + 1) − 1 and (Cn¾) ≤  + 2 if n ≥ ( + 1).
10.1.30. (♦) A function  : [n] → [n] is contractive if  (i) ≤ i for all i. A monotone  -list
for  is a strictly increasing list a1 , . . . , a¾ from [n] such that  (a1) ≤ · · · ≤  (a¾). Prove
that 2 ¾−1 is the least n such that for every contractive mapping on [n] there is a monotone
-list. (Hint: For how many a in [n] can the longest monotone list ending with a have j
elements? Comment: This result is generalized in West–Trotter–Peck–Shor [1984].)
10.1.31. (+) Prove that Theorem 10.1.16 is sharp for n ≥ 6 by giving an ordering of E(K n)
such that the maximum length of an increasing trail is n − 1. (Comment: The construction
is easy when n is even.) (Graham–Kleitman [1973])
10.1.32. Prove that at most ( − 1)2n permutations of [n] avoid the identity permutation
of length  . (Bóna [2004]) (Comment: Although this bound seems loose, in fact there are
asympotically c ((¾¾2−−1)2¾)/2 such permutations, where c is a known constant.)
2n

10.1.33. Determine the maximum number of 1s in a 0 , 1-matrix of order n such that no 1

has a 1 somewhere to its right and somewhere below (see Füredi–Hajnal [1992]).
10.1.34. Prove that the number of permutations of [n] avoiding all three of the patterns
123, 132, and 213 is the adjusted Fibonacci number F̂ n . (Simion–Schmidt [1985])
10.1.35. (♦) Establish a bijection from the 123-avoiding permutations to the 132-avoiding
permutations. (Hint: Leave all left-to-right minima fixed.) (Simion–Schmidt [1985])
10.1.36. Generalize Exercise 10.1.35 by proving Sn(123 · · ·  ) = Sn(123 · · ·  ), where
 =  − 1 (see Definition 10.1.17). This compares the identity pattern with that switch-
ing the last two elements. (Comment: Backelin–West–Xin [2007] proved Sn() = Sn( )
whenever  is a permutation of [ j],  = (j + 1 , . . . , ) in order, and  is the reverse of  .)
10.1.37. Permutations with many patterns.
(a) Prove that 1 , n , 2 , n − 1 , . . . , ⌈ n + 1 ⌉ 2 contains at least F̂ n patterns. (Wilf)
(b) Let ¾ be the permutation of  2 whose runs have size  and step by  . For example,
3 = 369258147. Prove that ¾ contains more than 2(¾−1) patterns. (Coleman [2004])
2

10.1.38. (+) A deck of n cards is a permutation of [n]; view it as a word. A shuffle splits
a word into an initial block and a final block (one may be empty) and merges them, main-
taining order within each block. For example, one shuffle can produce 3142 from 1234.
Consider permutations of [n] produced from the identity permutation by one shuffle.
a) Prove that these are the permutations avoiding 321, 2143, and 2413. Count them.
b) Show that among these, the permutations that can be returned to the identity by
one more shuffle are those that split [n] into four segments A , B , C, D with B , C
= ∅ and
reorder them as A , C, B , D. Count them. Characterize them as the permutations avoiding
321, 2143, 2413, and 3142. (DeSario [2002])
10.1.39. (♦) Let G be a ( − 1)-chromatic n-vertex graph with girth l. Let H be a -
chromatic n-uniform hypergraph with girth at least l/3, edges e1 , . . . , e m , and vertices U.
Form G from U and disjoint copies G1 , . . . , G m of G by adding a matching of size n from
V(G i) to ei for each i. Prove that G has girth l and chromatic number  . (Nešet řil–Rödl)
Section 10.2: Ramsey’s Theorem 443

10.1.40. (+) Graphs with large girth and chromatic number. A (d , r, ½)-graph is a graph
with girth at least ½ formed from a complete d-ary tree by giving each leaf r neighbors
(besides its parent) on its path to the root. (A complete d-ary tree of height h is a rooted
tree where all non-leaves have d children and all leaves have distance h from the root.)
(a) Assume that (d , r, ½)-graphs exist. Use a (¾ , ¾ + 1 , 2½)-graph to construct a graph
with girth at least ½ that is not ¾-colorable. (Hint: Use the canonical edge-coloring of the
underlying tree, which assigns color i to the edge from a non-leaf vertex to its ith child.)
(b) Let m(d , r, ½) be the least height of the underlying tree in a (d , r, ½)-graph, if one
exists. Prove that a (d , r, ½)-graph exists for even ½ by showing (1) m(d , r, 4) = 2r + 1, (2)
m(d , r, ½ + 2) ≤ 2 + m(d , d 2 , ½), and (3) m(d , r, ½) ≤ m1 + m2 − 1, where m1 = m(d , 1 , ½)
and m2 = m(d m1 , r − 1 , ½). (Alon–Kostochka–Reiniger–West–Zhu [2016]) (Comment: By
setting m1 = 2⌊ m(d , 1 , ½)/2⌋ + 1 in (3), one can make bipartite (d , r, ½)-graphs, which
yields additional applications to list coloring.)

10.2. Ramsey’s Theorem

The Pigeonhole Principle guarantees that partitioning many objects into
classes yields some class with many objects. The famous theorem of F. P. Ram-
sey [1930] makes a similar statement about partitioning the r-element subsets of
the objects. The study of Ramsey Theory often starts with a classic brain-teaser,
which appeared on the Putnam examination in 1952.

10.2.1. Example. Among any six people, there are three mutual acquaintances or
three mutual strangers. Focus on one person, x. Among the remaining five people,
x must have at least three acquaintances or at least three non-acquaintances. By
symmetry, we may assume that x has at least three acquaintances. If any two of
these are acquainted, then they form the desired set with x; if they are pairwise
non-acquainted, then we have three mutual strangers.
x
•

• • • • •

Example 10.2.1 can be modeled as 2-coloring E(K 6). Edges are sets of size 2;
more generally, Ramsey ’s Theorem considers sets of size r. We partition the r-
subsets of a set S into ¾ classes and look for p elements of S whose r-sets all lie in
the same class. Example 10.2.1 states that when | S| = 6, there is always a triple
whose 2-sets lie in the same class.

THE MAIN THEOREM

By definition, blocks of a partition are nonempty. The language of coloring

allows classes to be empty.
444 Chapter 10: Ramsey Theory

10.2.2. Definition. A ¾-coloring is a function that labels each domain element

with one of ¾ colors. Let (Sr) denote the family of r-subsets of a set S. When
coloring (Sr), a set T ⊆ S is homogeneous if its r-sets all have the same color.

Roughly speaking, Ramsey ’s Theorem says that, for ¾ , r, p ∈ , every ¾- 

coloring of the r-sets of a large enough set has a homogeneous p-set. The precise
statement is more detailed, allowing different thresholds in different colors; it
facilitates an inductive proof.

10.2.3. Definition. In a ¾-coloring of (Sr), a homogeneous set whose r-sets have


color i is i-homogeneous. For quotas p1 , . . . , p¾ ∈ , the Ramsey number

R(p1 , . . . , p¾ ; r) is the least N ∈ such that every ¾-coloring of ([N]
r
) has an
i-homogeneous set of size pi for some i.

Ramsey ’s Theorem states that R(p1 , . . . , p¾ ; r) exists for every choice of the
quotas p1 , . . . , p¾ . The case r = 1 is the quota version of the Pigeonhole Principle
(Theorem 10.1.2). To suggest the approach, we first consider the case r = ¾ = 2,
which includes and extends Example 10.2.1.

10.2.4. Example. When r = 2, a partition of (S2 ) is just an edge-coloring of the

complete graph with vertex set S. When ¾ = 2, by time-honored tradition we
make color 1 red (solid) and color 2 blue (bold). We prove
R(p1 , p2 ; 2) ≤ R(p1 − 1 , p2 ; 2) + R(p1 , p2 − 1; 2).
Choose a vertex x. The quota version of the Pigeonhole Principle implies that if
there are N + M − 1 vertices other than x, then x must have N incident red edges
or M incident blue edges. Let
N = R(p1 − 1 , p2 ; 2) , M = R(p1 , p2 − 1; 2) ,
and consider a 2-coloring of E(K N + M). By symmetry, we may assume that x has
at least N incident red edges. By the definition of N , the clique induced by
the neighbors of x along these edges has a blue p2 -clique or a red (p1 − 1)-clique
(which combines with x to form a red p1 -clique). In either case, we obtain an
i-homogeneous set of size pi for some i.
The inequality inductively proves the existence of R(p1 , p2 ; 2). We postpone
discussion of the resulting bound on R(p1 , p2 ; 2).
x
•

• T • •U • •

| T | ≥ R(p1 , p2 − 1; 2) or |U | ≥ R(p1 − 1 , p2 ; 2)

In Example 10.2.4, we could define an auxiliary coloring on S − x by giving

each vertex the color of its edge to x. We would then apply the Pigeonhole Prin-
ciple to this coloring of the 1-subsets of S − x. This is what the induction step of
the general theorem reduces to when r = 2.
Section 10.2: Ramsey’s Theorem 445

10.2.5. Theorem. (Ramsey’s Theorem; Ramsey [1930]) For ¾ , r, p1 , . . . , p¾ ∈

, the Ramsey number R(p1 , . . . , p¾ ; r) exists.
Proof: We use induction on r, proving the induction step by induction on ∑ pi .
As basis steps, we need the existence of R(p1 , . . . , p¾ ; r) for r = 1, and we need
the existence of R(p1 , . . . , p¾ ; r) for r > 1 when ∑ pi is small. The case r = 1 is
the threshold version of the Pigeonhole Principle (Theorem 10.1.2). When r > 1
and some quota pi is less than r, a set of pi objects has no r-subsets, so vacu-
ously its r-sets all belong to class i. Hence R(p1 , . . . , p¾ ; r) = min{p1 , . . . , p¾ } if
min{p1 , . . . , p¾ } < r.
For clarity, we describe the induction step only for ¾ = 2; the general argu-
ment is similar (Exercise 12). Write (p , q) for (p1 , p2). Let
p = R(p − 1 , q; r) , q = R(p , q − 1; r) , N = 1 + R(p , q ; r − 1).
When we consider the triple (p , q , r), the induction hypothesis (induction on p+ q)
for fixed r implies that p and q are well-defined. The induction hypothesis for
the induction on r then implies that also N is well-defined. Let S be a set of N
elements, and choose x ∈ S. Consider a 2-coloring º of (Sr). We need to show that
there is a red-homogeneous p-set or a blue-homogeneous q-set.

T
x• r−1 S
p = R(p − 1 , q; r)

Let S = S − {x}. We use º to define a 2-coloring º of (rS−1). For A ∈ (rS−1),
let º (A) = º (A ∪ {x}); thus º is a 2-coloring of the (r − 1)-sets in S . Since
|S | = R(p , q ; r − 1), the induction hypothesis implies that S contains a red-
homogeneous set of size p or a blue-homogeneous set of size q under º (when
r = 2, this step is the invocation of the Pigeonhole Principle). By symmetry, we
may assume that the red quota is met. Let T be a p -element subset of S whose
(r − 1)-sets are red under º .
We return to the original coloring º on ( Tr ). Since | T | = p = R(p − 1 , q; r),
within T there is a (p − 1)-set that is red-homogeneous under º or a q-set that
is blue-homogeneous under º . If there is a blue-homogeneous q-set, then we are
done. If there is a red-homogeneous (p − 1)-set P , then consider P ∪ {x}. From
the definition of T , the (r − 1)-sets of P are all red under º , and so their unions
with x are red under º . Hence P ∪ {x} is a red-homogeneous p-set under º .

The structure of this proof is called “double induction”; the proof of the in-
duction step is itself a proof by induction.

APPLICATIONS

Like the Pigeonhole Principle, Ramsey ’s Theorem has subtle applications

and provides elegant existence proofs. Unlike the Pigeonhole Principle, the
bounds from Ramsey ’s Theorem usually are horribly loose.
446 Chapter 10: Ramsey Theory

Our first application was a problem posed to some talented mathematics stu-
dents in 1932 by their friend Eszther K lein. Paul Erdős, then 19, was in the
group. Erdős and Szekeres gave several solutions, including one requiring re-
discovery of Ramsey ’s Theorem, which was unknown to them. Four years later,
K lein married Szekeres, so Erdős named this the Happy End Theorem.


10.2.6. Theorem. (Erdős–Szekeres [1935]). For m ∈ , there is a (least) inte-
ger N(m) such that every set of N(m) points in the plane (no three collinear)
contains an m-subset forming a convex m-gon.
Proof: We need two facts. (1) Among any five points in the plane, four determine a
convex quadrilateral (if no three are collinear). Construct the convex hull of the
five points. When it is a pentagon or a quadrilateral, the claim is immediate.
When it is a triangle, the other two points lie inside. By the Pigeonhole Princi-
ple(!), two corners of the triangle are on one side of the line through the points
inside. These plus the points inside form a convex quadrilateral.

• • • •
• •
• • • •
• • • •
• • • •
• • • •

(2) If every 4-subset of m points in the plane forms a convex quadrilateral, then
the m points form a convex m-gon. If not, then the convex hull of the m points
consists of some t points, where t < m. The remaining points lie inside the t-gon.
When we triangulate the t-gon, as shown on the right above, a point inside the t-
gon lies in one of the triangles. With the vertices of that triangle, it forms a 4-set
that does not determine a convex quadrilateral.
In order to combine these two results into a proof of the theorem, let N =
R(m , 5; 4). Given N points in a plane, color each 4-set by convexity: red if it de-
termines a convex quadrilateral, blue if it does not. By fact (1), there cannot be 5
points whose 4-subsets are all blue. By Ramsey ’s Theorem, there must then be m
points whose 4-subsets are all red. By fact (2), they form a convex m-gon. Hence
N(m) exists and is at most R(m , 5; 4).

The upper bound R(m , 5; 4) is exact for m = 4 by fact (1), but already it is
−4
very loose at m = 5 (Exercise 6 shows N(5) = 9). In general, N(m) ≤ (2m m−2 + 1
)
(Erdős–Szekeres [1935]) and N(m) > 2 m−2 (Erdős–Szekeres [1960]). In fact, they
conjectured N(m) = 2 m−2 + 1 (Exercises 7–9). Erdős offered $500 for a proof of
the conjecture, raised to $1000 by Graham. After more than 60 years, Chung–
Graham [1998] improved the upper bound by 1. About ten papers gave further
small improvements, but none improved the leading behavior of 4 m . Finally, Suk
[2016] combined various ways of finding a convex polygon to solve the problem
asymptotically, proving N(m) ≤ 2 m+o(m). Other extensions and generalizations
are surveyed in Morris–Soltan [2000, 2016].
In the next application, again we need an auxiliary structural lemma and
obtain a bound that likely is much larger than needed.
Section 10.2: Ramsey’s Theorem 447

10.2.7. Example. Table storage and search. From a large universe of numbers
(called “keys”), n numbers arrive to be stored in a table of size n. If they are stored
in order, then binary search can test the presence of any key by probing at most
⌈ lg(n + 1)⌉ locations, where lg denotes log2 .
Can another strategy always test membership in fewer probes? It depends
on the size of the key space M. Yao [1981] proved that sorted storage and binary
search is optimal when M is large.
A storage strategy T assigns each n-set A of keys a storage permutation:
T(A) = puts the jth smallest element of A in location (j), for 1 ≤ j ≤ n.
A query asks whether a key x is present in the table. A search strategy S probes
table locations, based on x, T, and the outcome of earlier probes. The answer to a
probe is the key stored in that location. The (worst-case) cost of a strategy (S , T)
is the maximum, over all table contents A and queries x, of the number of probes
used by S to determine whether x ∈ A when T is used to store the contents of the
table. For m = | M | , let (m , n) be the minimum cost over all strategies.
When m = n, every key is present, and (n , n) = 0. When m = n + 1, sorted
storage has worst-case search cost 2, since the jth element x may be stored at j
or j − 1, depending on whether the missing element is higher or lower than x. On
the other hand, probing location 1 suffices if we store in location 1 the key that
cyclically follows the missing key. Hence (n + 1 , n) = 1; sorting is not optimal
when m is small. Yao [1981] proved that (m , n) ≤ 1 if and only if m ≤ 2n − 2.

The next lemma generalizes the argument that binary search in sorted tables
needs lg(n + 1) probes in the worst case.

10.2.8. Lemma. (Yao [1981]) Let T be a storage strategy for n-sets from a uni-
verse M with | M | ≥ 2. A set P ⊆ M is stored consistently if each n-set in P
is stored according to the same permutation. If some set of size 2n − 1 in M
is stored consistently under T, then any strategy using T for storage costs at
least ⌈ lg(n + 1)⌉ probes in the worst case.
Proof: Let P be a set of size 2n − 1 stored consistently under T; there is a permu-
tation  such that T stores any n-set from P by putting its jth smallest element
in position  (j), for 1 ≤ j ≤ n. Let P = {x1 , . . . , x 2n−1 } with x1 < · · · < x2n−1 . By
reindexing the table locations in the order  (1) , . . . ,  (n), we may assume that 
is the identity permutation.
Using induction on n, we prove that an adversary can force the algorithm to
use at least ⌈ lg(n + 1)⌉ probes to test whether the “central element ” x n of P is
present. For n = 1, a probe is needed, since | M | ≥ 2.
For n ≥ 2, let n = ⌊ n/2⌋ . The adversary responds to the first probe to leave
a problem of the same type, with size n and with x n still the central element.
Define P by eliminating the top ⌈ n/2⌉ and bottom ⌈ n/2⌉ elements of P. Since
2n − 1 = 2n − 1 − 2⌈ n/2⌉ , the size of P is 2n − 1, and the central element is x n .
If the algorithm first probes location j with j ≤ ⌈ n/2⌉ , then the adversary
responds that key x j is there. The adversary also grants that keys x1 , . . . , x⌈n/2⌉
are present in locations 1 , . . . , ⌈ n/2⌉ (in order, since P is consistent) and keys
{x2n−⌈n/2⌉ , . . . , x 2n−1 } are not in the table (see exception below for n ≤ 3). Now all
is known except which n -subset of P appears in locations ⌈ n/2⌉ + 1 , . . . , n.
448 Chapter 10: Ramsey Theory

probe
1 j ⌈n/2⌉ ⌈n/2⌉ + 1 n
x1 · · · x j · · · x⌈n/2⌉
n -set from P x2n−⌈n/2⌉ · · · x2n−1
present smaller problem absent

When n = 2 or n = 3, the adversary instead guarantees the absence of only

the top ⌈ n/2⌉ − 1 elements. It thus remains uncertain whether x n is in the last
unknown location, and one more probe will be needed.
For n ≥ 4, consistency of T on P implies that for n-sets of P containing
x1 , . . . , x⌈n/2⌉ and omitting x 2n−⌈n/2⌉ , . . . , x 2n−1 , the elements of P in locations
⌈ n/2⌉ + 1 , . . . , n are in order. Thus the set P of size 2n − 1 with central element
x n is stored consistently under a strategy with n locations. By the induction hy-
pothesis, the adversary can force at least ⌈ lg(n + 1)⌉ further probes to test for the
presence of x n . This completes the proof, using 1 + ⌈ lg(⌊ n/2⌋ + 1)⌉ = ⌈ lg(n + 1)⌉ .
Symmetrically, if the first probe is after ⌈ n/2⌉ , then the adversary puts the
highest ⌈ n/2⌉ elements in the highest ⌈ n/2⌉ positions and says that the lowest
⌈ n/2⌉ keys in P are absent. Now the induction hypothesis applies with an n -
subset of P in locations 1 , . . . , n .

10.2.9. Theorem. (Yao [1981]) Let º (m , n) be the complexity of membership

testing when n-sets from a space of m keys are stored in a table of size n. If
m is sufficiently large, then º (m , n) = ⌈ lg(n + 1)⌉ .
Proof: By Lemma 10.2.8, we need only show that for large m, under every stor-
age strategy T there is a set P of size at least 2n − 1 that is stored consistently.
This is a job for Ramsey ’s Theorem. We need P such that every n-subset of P is
stored using the same permutation , so we color the n-subsets of M using n! col-
ors, with the color on a set being the permutation used to store it. That is, assign
color to A if T stores A by putting x i in location (i) for all i.
If m ≥ R(2n − 1 , . . . , 2n − 1; n) (all n! quotas equal 2n − 1), then Ramsey ’s
Theorem guarantees a color such that some set of 2n − 1 keys is consistently
stored by . By Lemma 10.2.8, ⌈ lg(n + 1)⌉ probes are required in the worst case
to test the presence of some key in that set.

RAMSEY NUMBERS

Ramsey ’s Theorem defines the Ramsey numbers R(p1 , . . . , p¾ ; r). No formula

is known, and few values have been computed. To prove R(p1 , . . . , p¾ ; r) = N , one
must guarantee a -coloring of the r-sets on N − 1 points that meets no quota and
show that every coloring on N points meets some quota.
In principle, a computer can check all -colorings of ([n]
r
) for various n to find
the first n such that every coloring meets some quota. Even when r =  = 2, the
2(2) colorings rapidly become too many to handle. Erdős joked that if an evil spirit
n

threatens to destroy the human race unless we tell it the value of R(5 , 5; 2), then
we should have all computers in the world look for the answer. If instead it wants
R(6 , 6; 2), then he advised destroying the evil spirit before it destroys us.
Section 10.2: Ramsey’s Theorem 449

To understand the joke, consider trying to prove R(5 , 5; 2) > 43. Without
making use of symmetry or other reductions of the search, the number of graphs
with vertex set [43] is 2( 2 ) , which equals 2903 . We need one having no 5-clique or
43

independent 5-set. The number of particles in the universe is thought to be at

most 1088 . The age of the universe is thought to be at most 20 billion years, and
there are fewer than 32 million seconds per year. Even if every particle in the
universe had checked one graph every second since the beginning of time, fewer
than 2350 graphs would have been checked by now. Hence Erdős was an optimist.
For often-studied subcases, we abbreviate the notation.

10.2.10. Definition. When r = 2, we write R(p1 , . . . , p¾ ; r) as R(p1 , . . . , p¾).

When p = p1 = · · · = p¾ , we write R(p1 , . . . , p¾ ; r) as R¾ (p; r) and call it a
diagonal Ramsey number. In particular, R¾ (3; 2) is the least n such that
every ¾-coloring of E(K n) has a monochromatic triangle.

For r > 2, little is known other than R(4 , 4; 3) = 13 (McK ay–Radziszowski

[1991]). Even when r = 2, only one Ramsey number is known exactly with ¾ > 2:
R(3 , 3 , 3) = 17 (Greenwood–Gleason [1955]). The table below contains the known
values of R(p , q) and the best known bounds for some other values as of August
2009. The current bounds are maintained in Radziszowski [1994–2017] (periodi-
cally updated). Computer searches have been slowly increasing the lower bounds.
3 4 5 6 7 8 9
3 6 9 14 18 23 28 36
4 18 25 36/41 49/61 56/84 73/115
5 43/48 58/87 80/143 101 /216 133/316
6 102/165 115/298 134/495 183/780
The values of R(3 , 9) (Grinstead–Roberts [1982]), R(3 , 8) (McK ay–Zhang
[1992]), and R(4 , 5) (McK ay–Radziszowski [1995]) are the most recent. The
others were found much earlier (due primarily to Greenwood–Gleason [1955],
K albfleisch [1967], and Graver–Yackel [1968]).
When r = ¾ = 2, we can view the colors on the edges of K n as “in” and “out ”.
Ramsey ’s Theorem for this case then becomes: “ There exists a smallest integer
R(p , q) such that every graph with R(p , q) vertices has a clique of size p or an
independent set of size q.”
The proof of Ramsey ’s Theorem yields a recursive upper bound on R(p , q; r).
It is very large but gives useful values for small thresholds when r = 2. For r = 2,
it repeats the discussion in Example 10.2.4.

10.2.11. Theorem. R(p , q) ≤ R(p − 1 , q) + R(p , q − 1). If both summands on the

right are even, then the inequality is strict.
Proof: If a vertex in an arbitrary graph has R(p − 1 , q) neighbors or R(p , q − 1)
nonneighbors, then the graph contains K p or K q . With R(p − 1 , q) + R(p , q − 1)
vertices altogether in the graph, the Pigeonhole Principle guarantees that one of
these possibilities occurs.
Equality requires a graph with R(p − 1 , q) + R(p , q − 1) − 1 vertices hav-
ing no clique of size p or independent set of size q. If some vertex has at least
R(p − 1 , q) neighbors or at least R(p , q − 1) nonneighbors, then there is a p-clique
450 Chapter 10: Ramsey Theory

or an independent q-set. Otherwise, with exactly R(p − 1 , q) + R(p , q − 1) − 2 other

vertices, every vertex must have degree exactly R(p − 1 , q) − 1. When R(p − 1 , q)
and R(p , q − 1) are both even, this requires a regular graph of odd degree with an
odd number of vertices, which does not exist.

10.2.12. Corollary. R(p , q) ≤ ( p+p−q−1 2 ).

Proof: A 2-edge-coloring of K p has an edge of the second color or doesn’t, so
R(p , 2) = p. Hence the upper bound holds with equality when q (or p) is 2. This
provides the basis for induction on p + q. By Theorem 10.2.11, the induction
hypothesis, and Pascal’s Formula,
R(p , q) ≤ R(p − 1 , q) + R(p , q − 1) ≤ ( p+p−q−1 3) + ( p+p−q−2 3) = ( p+p−q−1 2).

10.2.13. Example. R(3 , 3) = 6 and R(3 , 4) = 9.

For R(3 , 3), Example 10.2.1 yields 6 as the upper bound. Equality holds be-
cause the 5-cycle and its complement have no triangles.
For R(3 , 4), the inequality in Theorem 10.2.11 is strict. With R(2 , 4) = 4 and
R(3 , 3) = 6, we have R(3 , 4) < 10 since both are even. The graph on the right
below proves R(3 , 4) ≥ 9. Four independent vertices on an 8-cycle must include
opposite vertices, but in this graph such vertices are adjacent.

• • •
• • • •

• •
• • • •

√
Corollary 10.2.12 yields R(p + 1 , p + 1) ≤ (2p p) ∼ 4 /
p
p. Conlon [2009a]
improved the upper bound by putting a superpolynomial function of p in the de-
nominator: R(p + 1 , p + 1) ≤ (2p
p
)p−c log p/log log p for some constant c.
The constructive lower bounds in Exercises 17–18 are polynomials in p. The
best constructive lower bound known grows faster than every polynomial but
slower than every exponential (Exercise 19). A survey of constructive bounds ap-
pears in Frankl [1990]. An exponential lower bound can be proved by counting.

10.2.14. Theorem. (Erdős [1947]) R(p , p) > √1 p2 p/2(1 − o(1)).

e 2

Proof: Consider the graphs with vertex set [n]. Each possible p-clique occurs in
2(2)−(2) of these 2(2) graphs. Similarly, each set of p vertices occurs as an inde-
n p n

pendent set in 2(2)−(2) of the graphs. Discarding these leaves only graphs with no
n p

p-clique or independent p-set. Since there are (np) ways to choose p vertices, the
inequality 2(n)2(2)−(2) < 2(2) implies R(p , p) > n. We seek n such that (n) < 2(2)−1 .
n p n p

p p
p
To obtain the desired bound, we use (np) < ( ne
p
) , proved in Section 14.1. It
ne p (p
)−
thus suffices to have ( ) < 2 2 , which simplifies to n < (1 − o(1)) √p 2 p/2 .
1
p e 2
Section 10.2: Ramsey’s Theorem 451

We revisit this argument in probabilistic language in Theorem 14.1.3. Deter-

mining whether the limit of R(p , p)1/p exists (and its value if it exists) is the fore-
most open problem in√ Ramsey theory. By Theorem 10.2.14 and Corollary 10.2.12,
the value is between 2 and 4.

10.2.15.* Remark. Fixing one quota. In contrast to the diagonal case, tighter
bounds are known when q is fixed and p is large. Here R(p , q) ≤ cpq−1 logloglogp p
(Graver–Yackel [1968], Chung–Grinstead [1983]). For q = 3, the answer is
known within a constant factor:
c p2
cp2
log p ≤ R(p , 3) ≤ log p

The upper bound is due to Ajtai–Komlós–Szemerédi [1980], the lower to K im

[1995]. Both use probabilistic methods (see Chapter 14).
This result relates to triangle-free ¾-chromatic graphs (Remark 8.1.19). Let
n = R(p , 3) − 1. By the meaning of R(p , 3), some n-vertex graph has no trian-
gle and no independent p-set. Its chromatic number is thus at least n/p. With
¾ = n/p, there is a ¾-chromatic triangle-free graph with at most n vertices.
√
We solve
√ for n in terms of ¾ . With n ≥ cp2
/log p, we have p ≤ O( n log n).
Now ¾ ≈ n/log n. Thus there are ¾-chromatic triangle-free graphs with at most
O(¾ 2 log ¾) vertices. This yields the upper bound in Remark 8.1.19.

Before generalizing the Ramsey problem for r = 2, we derive the classical

bounds on the diagonal 2-color hypergraph Ramsey numbers. We write simply
R(p; r) instead of R2(p; r). For further discussion of hypergraph Ramsey num-
bers, see Conlon–Fox–Sudakov [2010, 2013] and Mubayi–Suk [2017].
To introduce the main idea for the upper bound, we consider first an easy
explicit upper bound for R(p; 2), a bit weaker than Corollary 10.2.12.

10.2.16. Proposition. R(p; 2) ≤ 22p .

Proof: Consider the elements of [22p] in order as vertices, and let º be a red/blue
coloring of the pairs. Initially, all vertices are “live”. When x is the first remain-
ing live vertex, we label x red or blue to agree with the color given by º to the
majority of the edges from x to later live vertices. A later live vertex y then re-
mains alive only if º (xy) agrees with the label of x.
After doing this R(p − 1; 1) times, the Pigeonhole Principle yields p − 1 live
vertices that have the same label, red or blue. Together with one more live ver-
tex, we have a set U of p vertices that were all kept because their edges to earlier
vertices in U have that same color.
Since we lose at most a factor of 2 each time we process the first unlabeled
live vertex, and R(p − 1; 1) = 2p − 1, it suffices to start with 1 + 22p−1 vertices.

For larger r, we generalize the idea, considering extensions of (r − 1)-sets by

adding one later vertex.
r −1 ]
10.2.17. Theorem. (Erdős–Rado [1952]) R(p; r) ≤ 2[R(p−1;r−1) .
r −1
Proof: Initially, all vertices are live. Let N = 2[R(p−1;r−1) ] , consider the ele-
ments of [N] in order, and let º be a red/blue coloring of ([N]
r
). Let S be the set
452 Chapter 10: Ramsey Theory

of the first r − 1 vertices in [N]. Label S with the color given by º to the major-
ity of the r-sets obtained by adding a later vertex to S. Keep alive only the later
live vertices y such that º (S ∪ {y}) is the same as the label of S.
Consider the situation when x is the first remaining vertex. Iteratively la-
bel each (r − 1)-set S consisting of x and r − 2 earlier vertices that have been
kept alive. The label of S is the majority color of the extensions of S to r-sets
by adding a later remaining live vertex y. Keep y alive only if º (S ∪ {y}) is the
same as the label on S .
Each time we label an (r − 1)-set, we lose at most half of the remaining live
vertices. We want to wind up with R(p − 1; r − 1) vertices whose (r − 1)-sets have
all been labeled, plus another live vertex. Since ( R(p−r−1;r
1
−1)
) sets of size r − 1 are
to be labeled, it suffices to start out with N vertices. By the definition of the
Ramsey number, we then have a set U of p vertices in which every (r − 1)-set not
containing the last vertex has been labeled with the same color, meaning that the
r-sets obtained by adding a later vertex of U all have that color.

10.2.18. Corollary. R(p; r) ≤ tow r−1 (cr p), where tow 0(x) = x, tow h(x) =
2tow h−1(x) for h > 0, and cr is a constant that depends on r but is indepen-
dent of p.
Proof: By Proposition 10.2.16, R(p; 2) ≤ 22p , so c2 = 2 suffices. For larger r,
r −1
Theorem 10.2.17 yields R(p; r) ≤ 2[tow r−2(cr−1(p−1))] ≤ tow r−1 (cr p). The presence
of p − 1 in the recurrence allows cr to remain relatively small.

A straightforward extension of the counting argument in Theorem 10.2.14

(counting 3-uniform hypergraphs instead of graphs) yields R(p; 3) > 2 cp . Induc-
2

tion on r leads to a lower bound on R(p; r) that is also a tower, but having height
lower by 1. The key is the next lemma, whose proof is due to Erdős and Hajnal
but seems to be published only in Graham–Rothschild–Spencer [1980].

10.2.19. Lemma. (Stepping-Up Lemma) If r ≥ 3, then R(p; r) > n implies

R(2p + r − 4; r + 1) > 2 n .
Proof*: Let º be a red/blue coloring of ([n] )
r avoiding homogeneous p-sets. Letting
T = [2 n], we form a red/blue coloring of (r+T1 ). For x ∈ T , let (x1 , . . . , x n) be the
binary expansion of x − 1, so that x = 1 + ∑ i=1 x i 2 i−1 .
n

For x , x ∈ T , let (x , x ) = max{i: x i

= x i }. If x < x and (x , x ) = j , then
x j = 0 and x j = 1. Therefore, x < x < x implies (x , x )
= (x , x ); that is, in a
monotone list the derived list of high-bit changes has no immediate repeats.
Given x1 , . . . , x t ∈ T , indexed in increasing order,
(x1 , x t) = max1≤i<t { (x i , x i+1)}. (∗)
The operator extends to monotone lists of any length by letting (x1 , . . . , x t) =
( (x1 , x2) , . . . , (x t−1 , x t).
Now we define a red/blue coloring  of (r+T1 ), “stepping up” from  . Index
any (r + 1)-set Y in increasing order, forming an (r + 1)-tuple. If the r-tuple (Y)
is strictly monotone, then let  (Y) =  ( (Y)). If (Y) has a peak at the second
position (value larger than the neighboring values), then let  (Y) be red. If (Y)
Section 10.2: Ramsey’s Theorem 453

has a valley at the second position (value smaller than the neighboring values),
then let º (Y) be blue. Otherwise, (Y) is chosen arbitrarily.
Now let Z be a (2p + r − 4)-set in T ; we show that Z cannot be homogeneous.
Suppose that Z is red-homogeneous (an analogous argument applies for blue).
With Z indexed in increasing order, no consecutive values in (Z) are equal.
Case 1: For some t, the positions t through t + p − 1 in (Z) form a strictly
monotone p-tuple Ẑ. Suppose first that Ẑ is decreasing, and let s = (Zs , Zs+1),
so Ẑ = ( t , . . . , t+ p−1). Since  admits no homogeneous p-tuple, some r entries in
Ẑ form a set W that is blue under  . Let i1 , . . . , ir be the corresponding indices
of these positions in Z, and let Y = (Zi1 , . . . , Zir , Zir +1).
We claim that  (Y) is blue, contradicting that Z is red-homogeneous. By
(∗), the value (Yj , Yj +1) is the largest (Zs , Zs+1) among s such that i j ≤ s < i j +1 ,
where ir+1 = ir + 1. Since (Z) is monotone decreasing, this largest value is Wj .
Thus (Y) = W . Since W is a sublist of the monotone decreasing list (Z), also W
is monotone decreasing. Hence by definition  (Y) =  (W), which is blue.
If (Z) is strictly increasing, then an analogous argument applies, with Y
consisting of the positions in Z just after all those corresponding to the positions
occupied by W , plus the first position occupied by W .
Case 2: No consecutive p-tuple in (Z) is monotone. Consider the first 2p − 3
positions in (Z). None after the first can be a valley, because then Z would con-
tain an (r + 1)-tuple assigned blue. Hence (Z) has at most one peak. If it has no
peak, then (Z) is monotone. If it has a peak before or after the middle, then the
portion on the opposite side of the middle of (Z) is a monotone list of length at
least p. In each of these possibilities, Case 1 applies.

10.2.20. Corollary. R(p; r) > tow r−2(c r p2), where c r = c/42r−6 and R(p; 3) > 2 cp .
2

Proof: Since R(q; r) > n implies R(2q + r − 4; r + 1) > 2 n by the Stepping-Up

Lemma, and we may assume q > r, also R(q; r) > n implies R(4q; r + 1) > 2 n . With
p = 4q, we have R(p; r) > 2 R(p/4;r−1) . Thus each iteration increases the height of
the tower and decreases the constant by a factor of 42 .

Thus tow r−2(c r p2) < R(p; r) ≤ tow r−1 (cr p); the heights differ by 1. Erdős con-
jectured R(p; 3) > 22 , yielding R(p; r) > tow r−1 (c r p). Erdős and Hajnal proved
cp

that R4(p; 3) is doubly-exponential, so for four colors the towers above and below
have the same height. Conlon–Fox–Sudakov [2013] proved R3(p; 3) > 2 p .
c log p

GRAPH RAMSEY THEORY

By Ramsey ’s Theorem, -coloring the edges of a large enough complete graph

forces a monochromatic complete subgraph of order p. It thus also forces a
monochromatic copy of every p-vertex graph G. When we seek G and don’t need
K p , we may not need R(p , p) vertices. For example, every 2-coloring of the edges
of K 3 has a monochromatic P3 , but K 6 is needed to force a monochromatic K 3 .

10.2.21. Definition. For graphs G 1 , . . . ,G ¾ , the Ramsey number R(G 1 , . . . ,G ¾)

is the least n such that every -coloring of E(K n) contains a copy of G i in color
i for some i. When G i = G for all i, we write R(G 1 , . . . ,G ¾) as R¾ (G).
454 Chapter 10: Ramsey Theory

Sometimes R(G 1 , G 2) has a simple formula. Again we color red and blue.

10.2.22. Theorem. (Chvátal [1977]). If T is an m-vertex tree, then R(T, K n) =

(m − 1)(n − 1) + 1.
Proof: For the lower bound, color K(m−1)(n−1) by letting the red graph consist of
(n − 1)K m−1 . Red components of order m − 1 have no red m-vertex tree. The blue
graph is (n − 1)-partite and hence does not contain K n .

K m−1 K m−1 K m−1 K m−1

We prove the upper bound by induction on n, using a property of trees proved

by induction on m. The basis step is n = 1; no edges are needed to obtain K 1 .
For n > 1, consider a 2-coloring of E(K(m−1)(n−1)+1). If some vertex x has more
than (m − 1)(n − 2) incident blue edges, then the induction hypothesis gives a red
T or a blue K n−1 in N(x), and thus a red T or a blue K n (with x) in the full graph.
Otherwise, every vertex has at most (m − 1)(n − 2) incident blue edges and
at least m − 1 incident red edges. Now there is a red T , because every graph G
having minimum degree at least m − 1 contains T .
This claim for trees (Exercise 5.4.17) holds by induction on m; it is trivial for
m = 1. For m > 1, let u be a leaf, with neighbor v, and let T = T − u. By the
induction hypothesis, G contains T . The vertex serving as v has at least m − 1
neighbors in G, at most m − 2 of which appear in the copy of T . Thus v has
another neighbor in G to serve as u.

The lower bound holds more generally, with the same coloring. If the largest
component of G has m vertices, and H has chromatic number n, then R(G , H) ≥
(m − 1)(n − 1) + 1 (Chvátal–Harary [1972]). Burr–Erdős [1983] conjectured that
R(G , K n) = (m − 1)(n − 1) + 1 when m is sufficiently large relative to n and
max F ⊆ G ||V(F)
E(F)|
|
, but Brandt [1996] disproved this. Nevertheless, Nikiforov [2005]
proved R(Cm , K n) = (m − 1)(n − 1) + 1 for m ≥ 4n + 2, partially settling a conjecture
of Erdős, Rousseau, and Schelp (see Faudree–Schelp [1978]).
The upper bound for Theorem 10.2.22 depends on color classes in H being sin-
gle vertices. Otherwise, the lower bound can be weak. For example, the Chvátal–
Harary result yields R(mK 3 , mK 3) ≥ (3 − 1)(3 − 1) + 1 = 5, but the truth is 5m.

10.2.23. Theorem. (Burr–Erdős–Spencer [1975]). R(mK 3 , mK 3) = 5m for m ≥ 2.

Proof: Let the red graph be K 3m−1 + K 1 ,2m−1 , as shown on the left. Every red
triangle uses three vertices from the (3m − 1)-clique, which is not big enough to
permit m pairwise disjoint triangles. The complementary blue graph on the right
is (K 2m−1 + K 1) K 3m−1 . Every blue triangle has at least 2 vertices in the (2m − 1)-
clique, which again is not big enough to permit m pairwise disjoint triangles.

x• K 2m−1 K 3m−1 K 2m−1 K 3m−1 •x

Section 10.2: Ramsey’s Theorem 455

For the upper bound, we use induction on m. The basis step m = 2 requires
case analysis (Exercise 28). If m ≥ 3, then 5m > R(3 , 3) = 6, so every 2-coloring
of E(K 5m) contains a monochromatic triangle. Discarding three vertices, we find
more such triangles until fewer than 6 vertices remain. Since 5m − 3m ≥ 6 for
m ≥ 3, we find m disjoint monochromatic triangles. If these have the same color,
then we are done; otherwise, we have a triangle in each color.
Let {a , b , c} and {d , e , º } be the vertex sets of disjoint red and blue triangles.
Of the nine edges joining them, we may assume by symmetry that at least five
are red. Some two of these have a common endpoint in {d , e , º }. This yields a red
triangle and a blue triangle with a common vertex. If m > 2, then the induction
hypothesis yields (m − 1)K 3 in one color among the remaining 5m − 5 vertices. We
then we add the appropriately colored triangle from the five special vertices.

Using R(3 , 3) = 6 as a basis instead yields R(mK 3 , mK 3) ≤ 5m + 1. A sim-

ilar but easier result is R(mK 2 , mK 2) = 3m − 1 (Exercise 27). The main induc-
tion argument generalizes to an upper bound on R(m1 G 1 , . . . , m¾ G ¾) in terms of
R(G 1 , . . . , G ¾) (Exercise 29).
Ramsey numbers of graphs (such as K n) may be exponential in the number
of vertices. However, Chvátal–Rödl–Szemerédi–Trotter [1983] proved that for
graphs with fixed maximum degree d, the Ramsey number grows only linearly in
the number of vertices! That is, R(G , G) ≤ c | V(G)| , where c is a constant that
grows quickly with d but does not depend on | V(G)|. We prove this in Section 11.1.
Meanwhile, we mention other directions in graph Ramsey theory. Having
generalized the target monochromatic subgraph from complete graphs to any
graph, we can also generalize the host graph.

10.2.24. Definition. The notation H → (G 1 , . . . , G ¾) means that every -

coloring of E(H) has a copy of G i in color i for some i. When G 1 = · · · = G ¾ ,
we write H →¾ G, or just H → G when = 2.

Note that K n → (G 1 , . . . , G ¾) when n ≥ R(G 1 , . . . , G ¾), so some such H ex-

ists. Classical graph Ramsey numbers minimize | V(H)| such that H → G, since
K n → G whenever some n-vertex graph H suffices. More generally, we can mini-
mize any parameter over the graphs forcing G.

10.2.25. Definition. For a graph parameter  , the -Ramsey number of a

graph G, written R (G), is min{(H): H → G}. More generally, R (G; s) =
min{(H): H → s G} and R (G 1 , . . . , G s ; s) = min{(H): H → (G 1 , . . . , G s)}.

We have so far studied vertex Ramsey numbers. There are also the size Ram-
sey number ((H) = | E(H)|), the clique Ramsey number ( = ), the chro-
matic Ramsey number ( = ), and the degree Ramsey number ( = ). The
initial result on parameter Ramsey numbers was R(G) = (G) for every graph
G (Folkman [1970]), extended to R(G; s) = (G) by Nešet řil–Rödl [1976].
The size Ramsey number of G, defined in Erdős–Faudree–Rousseau–Schelp
[1978], is obviously at most ( R(G)
2 ); they proved equality when G is a complete
graph. Solving a problem for which Erdős offered 100 euros, Beck [1983] proved
that the size Ramsey number of Pn is linear in n (improved by Dudek–Prałat
456 Chapter 10: Ramsey Theory

[2015]). The size Ramsey number also grows linearly in n for cycles (Haxell–
Kohayakawa–Łuczak [1995], but not for graphs with maximum degree 3 (Rödl–
Szemerédi [2000]). The result of Chvátal–Rödl–Szemerédi–Trotter [1983] (Sec-
tion 11.1) implies a quadratic upper bound in n for graphs with maximum de-
gree at most ¾ , and this was improved to O(n2( logn n)1/¾) by Kohayakawa–Rödl–
Szemerédi–Schacht [2011]). Faudree–Schelp [2002] surveyed the early results
on the topic, but there are many additional papers on the size Ramsey numbers
of special graphs or special families of graphs.
Here we study only the chromatic Ramsey number. This is the primary fo-
cus of Burr–Erdős–Lovász [1976], where parameter Ramsey numbers were first
proposed. We begin with an old result implying R(G) = 2 when G is bipartite.

10.2.26. Theorem. (Erdős–Rado [1956]) Given p , q , s, there exist m , n such that

every s-edge-coloring of K m ,n has a monochromatic K p ,q .
Proof: Let m = s(p − 1) + 1 and n = sm(q − 1) + 1. Consider an arbitrary s-coloring
of K m ,n with bipartition X , Y . Let the vertices in X be x1 , . . . , x m . Label each
vertex yj in Y by a s-ary vector whose ith coordinate is the color of the edge x i yj .
Now we apply the Pigeonhole Principle twice. Since there are sm such distinct
vectors, there must be q vertices in Y with the same vector v as label. Since only
s colors are available as entries for the m coordinates of v, v must use some color
at least p times. The p vertices of X in those positions and the q vertices of Y
with label v induce a monochromatic K p ,q .

To study chromatic Ramsey numbers of non-bipartite graphs, we apply The-

orem 10.2.26 to obtain specially colored complete multipartite subgraphs.

10.2.27. Theorem. (Burr–Erdős–Lovász [1976]) Let K r [t] denote the complete r-


partite graph with parts of size t. For n , r, s ∈ , there exists N ∈ such
that every s-coloring of E(K r [N]) contains a copy of K r [n] where any two parts
induce a monochromatic copy of K n ,n .
Proof: By Theorem 10.2.26, sufficiently large M for fixed m guarantees in parts
i and j a monochromatic copy of K m ,m . We want to do this with m large enough
so that, even when we keep only m vertices in each part, enough vertices remain
to ensure monochromatic copies of K n ,n joining all other pairs.
To make this precise, let b(m) be the least M such that s-colorings of E(K M,M)
force a monochromatic copy of K m ,m . Let N0 = n, and for i > 0 let Ni = b(Ni−1).
Let N = N(r). We claim that K r [N] has the desired property.
2

List the pairs of parts in some order. For i from 0 to (2r ), we show that N
is large enough so that for every s-coloring of E(K r [N]), some copy of K r [N(r)−i]
2
has the property that the edges of the jth pair of parts are colored monochro-
matically, for each j with 1 ≤ j ≤ i. For i = 0, there is nothing to prove; no
requirement is imposed on the full graph K r [N].
After such a copy of K r [N(r)−i+1 ] has been found for the first i−1 pairs of parts,
2
consider the ith pair. The vertices retained in these two parts induce K M,M , where
M = N(r)−i+1 . By the definition of N(r)−i+1 , the inherited coloring of E(K M,M) con-
2 2
tains a monochromatic copy of K m ,m with m = N(r)−i . Keep these m vertices in
2
Section 10.2: Ramsey’s Theorem 457

these two parts and any m vertices in each remaining part. The resulting copy
of K r [N(r)−i] has the desired property for the first i parts. After reaching i = (2r ),
2
we have the desired copy of K r [n], since N0 = n.

In Theorem 10.2.27, the colors on monochromatic subgraphs induced by pairs

of parts in the resulting K r [n] may differ. We need two more notions.

10.2.28. Definition. For a family of graphs, let R( ; s) be the minimum

number of vertices in a graph H such that every s-coloring of E(H) has a
monochromatic copy of some graph in . Write R( ) for R( ; 2). A homo-
morphism from a graph G to a graph H is a map
: V(G) → V(H) such that
adjacent vertices in G are mapped to adjacent vertices in H.
A proper ¾-coloring is the same thing as a homomorphism into K ¾ . Indeed,
the notion of homomorphism generalizes graph coloring, and a homomorphism
mapping G to H is called an H-coloring of G. Graph homomorphism is a vast
subject (the book Hell–Nešet řil [2004] is devoted to it), but here we introduce it
only to describe the chromatic Ramsey numbers of graphs (see also Exercise 41).
10.2.29. Theorem. (Burr–Erdős–Lovász [1976]) For every graph G, the s-color
chromatic Ramsey number equals R( ; s), where is the family of all homo-
morphic images of G.
Proof: Let n = | V(G)| , let r = R( ; s), and let H = K r [N], where N is the thresh-
old guaranteed for n , r, s by Theorem 10.2.27. We show that H → s G, which
completes the proof since H has the desired chromatic number.
Consider an s-coloring of E(H). By Theorem 10.2.27, the coloring of some
copy H of K r [n] is constant on the edges joining any pair of parts. Collapsing
each part to a single vertex thus yields an s-coloring of E(K r). By the definition
of R( ; s), this s-coloring contains a monochromatic copy of some homomorphic
image G of G, say with color c.
Let á be a homomorphism from G to G . The set mapped to any vertex v in G
by á is an independent set Sv in G, and | Sv | ≤ n. Hence we can return to H and
map each Sv injectively into the part in H that collapsed to v in G . If uv ∈ E(G ),
then all of Su is adjacent in H to all of Sv by edges of color c. Hence H contains
a monochromatic copy of G within H .
There are many variations. In on-line Ramsey theory, a Builder presents
edges one by one to force Painter (who must color each edge immediately) to make
a monochromatic G (see K ierstead–Konjevod [2009], Conlon [2009b/2010], But-
terfield et al. [2011], etc.). Ordered Ramsey theory was introduced in Fox–
Pach–Sudakov–Suk [2012]; with vertex orders on G and H , a monochromatic
G respecting the ordering is sought in every coloring of E(H). Theorem 12.4.45
treats such a problem; more recent papers include Balko–Cibulka–Král’–Kynčl
[2015], Conlon–Fox–Lee–Sudakov [2017], and Mubayi [2017].
Anti-Ramsey theory asks how many colors are needed on E(H) to force a
copy of G with distinct colors on its edges. Erdős–Sós–Simonovits [1975] conjec-
tured that when H = K n and G is a p-cycle this is n( p−2 2 + p−1 1 ) + O(1), proved
by Montellano-Ballesteros & Neumann-Lara [2005]. K ano–Li [2008] and Fujita–
Magnant–Ozeki [2010] are early surveys. The area has more than 200 papers.
458 Chapter 10: Ramsey Theory

EXERCISES 10.2
10.2.1. (−) Let Dn be the digraph with vertex set [n] whose edges are the ordered pairs
of distinct vertices. A tournament with a linear order on its vertices is monotone if ev-
ery edge points toward its lower endpoint or every edge points toward its higher endpoint.
Given ¾ , prove that if n is sufficiently large, then every spanning subdigraph of Dn con-
tains ¾ vertices that induce no edges or a monotone tournament or a copy of D¾ .
10.2.2. (−) Let n = R(¾ , l). Prove that the edges of the graph obtained from K n by deleting
one edge can be 2-colored with no red K ¾ or blue K l . (Chvatal [1974])
10.2.3. (−) Use the graph below to prove R(3 , 5) = 14.

• • •
• •
• •

• •
• •
• •

10.2.4. (−) Prove that 2-colorings of E(K6) have at least two monochromatic triangles.
10.2.5. (♦) For ¾ ≥ 2, say that a graph G is ¾ -balanced if every induced subgraph with 2 ¾
vertices has independence number ¾ .
(a) By explicit construction, prove R(3 , ¾) > 3¾ − 4.
(b) Prove that every ¾-balanced graph has at most R(3 , ¾ + 1)+ 2 vertices. (A. Brieden)
(c) Construct a ¾-balanced graph with 2 ¾ + 2 vertices. (Comment: Q. Zhu [1989]
showed for ¾ ≥ 4 that no ¾-balanced graph has more than 2 ¾ + 2 vertices.)
10.2.6. Prove N(5) = 9. That is, any nine points in 2 with no three collinear determine
some convex 5-gon, and some set of eight points does not. (Comment: Due first to Makai
and Turán in 1935 (unpublished), later proofs include Kalbfleisch–Kalbfleisch–Stanton
[1970] and Bonnice [1974]. By computer, Szekeres–Peters [2006] showed N(6) = 17.)
10.2.7. (+) In a tournament with weights on the edges, a nondecreasing [decreasing]
path is a path whose edge weights are successively nondecreasing [decreasing]. Prove that
the maximum number of vertices in an edge-weighted tournament having no nondecreas-
ing path of length p + 1 and no decreasing path of length q + 1 is ( p+p q). (Hint: Let (cx y , d x y)
denote the maximum length of (nondecreasing, decreasing) paths ending with edge xy.
Associate with each vertex the set of maximal labels on entering edges. Show that at most
(p+pq) distinct sets of labels can occur at vertices.) (Chvátal–Komlós [1971])
10.2.8. A list of points v1 , . . . , vn , where vi = (xi , yi) with x1 < · · · < x n , is convex [con-
cave] if whenever i < j the segment vi vj passes above [below] {vi+1 , . . . , vj −1 }. Use Exercise
10.2.7 to prove that the maximum number of points in the plane (with no three collinear)
containing no convex sequence of r points or concave sequence of s points is (r+r−s−2 4) . Con-
−4
clude that (2m
m−2
) + 1 points force a convex m-gon.
10.2.9. (+) Find 2 m−2 points in the plane (no three collinear) that form no convex m-gon.
(Hint: Use sets Ti for 0 ≤ i ≤ m − 2 (from Exercise 10.2.8) that have ( m−i 2) points but
contain no concave (i + 2)-sequence or convex (m − i)-sequence.) (Erdős–Szekeres [1960])
10.2.10. (♦) Let S be a set of R(m , m; 3) points in the plane with no three collinear. Prove
that S contains m points that form a convex m-gon. (Tarsi)
Exercises for Section 10.2 459

10.2.11. Prove that every graph with 2 ¾ vertices has a clique Q and an independent set S
such that | Q| + | S| = ¾ + 1. Conclude from this that R(¾ , ¾) ≤ 4¾−1 .
10.2.12. Ramsey numbers for r = 2 and multiple colors.
(a) Let p = (p1 , . . . , p¾), and obtain qi from p by subtracting one from pi but leaving
¾
the other coordinates unchanged. Prove R(p) ≤ ∑i=1 R(qi) − ¾ + 2.
(b) Prove R(p1 + 1 , . . . , p¾ + 1) ≤ (p1p+···+ p¾ )!
!··· p !
.
1 ¾

10.2.13. Prove that every graph G with p vertices, q edges, and automorphism group of
size s satisfies R¾(G) > (s¾ q−1)1/p .
10.2.14. (♦) Let r¾ = R¾ (3; 2) (see Definition 10.2.10).
(a) Prove r¾ ≤ ¾(r¾−1 − 1) + 2.
(b) Prove ⌊ t!e⌋ = t! ∑i=0 i!1 for t ∈ .
t

(c) Prove r¾ ≤ ⌊ ¾ !e⌋ + 1 for ¾ ≥ 2. (Thus r3 ≤ 17; in fact , equality holds).
10.2.15. (♦) For the diagonal ¾-color Ramsey number (Definition 10.2.10), prove the bound
(q−1)¾ −1 j
R¾(q; 2) ≤ 1 + ∑j =0 ¾ . (Comment: This implies the simpler-looking but weaker bound
R¾(q; 2) < ¾ , generalized in the next exercise.)
q¾

¾
10.2.16. (♦) For ¾ ≥ 2, let M¾(p; r) = max{R(q1 , . . . , q¾ ; r): ∑i=1 qi = p}.
(a) Prove M¾(p; r) ≤ 1 + M¾(¾ M¾(p − 1; r); r − 1).
(b) Prove M¾(p; 2) < ¾ p .
(c) Prove R¾(q; 3) ≤ t(¾ q + 1 , ¾ + 1), where t(1 , b) = b and t(h , b) = bt(h−1 ,b) (thus t(h , b)
t(p−1 , ¾+1)
is a tower of h copies of b). (Hint: Prove the stronger bound M¾(p; 3) < ¾ ¾ . Com-
ment: Conlon–Fox–Sudakov [2010] improved the upper bound; in particular, R¾(6; 3) ≤
(4+ o(1))¾ log ¾ c¾
22 . The technique of Lemma 10.2.19 yields R¾(6; 3) > 2 2 .)
10.2.17. Give a constructive proof of R(p + 1 , p + 1) > (3p) by considering the graph G
whose vertex set is ([p]
3
) with xy ∈ E(G) if and only if | x ∩ y| = 1. (Nagy [1972])
10.2.18. (♦) The composition or lexicographic product of graphs G and H is the graph
G[H] on V(G) × V(H) defined by making (u , v) and (u , v ) adjacent if and only if (1) uu is
an edge of G, or (2) u = u and vv is an edge of H .
(a) Prove (G[H]) ≤ (G) (H).
(b) Prove that the complement of G[H] is G[H].
(c) Use (a) and (b) to prove by construction that
R(pq + 1 , pq + 1) − 1 ≥ [R(p + 1 , p + 1) − 1] · [R(q + 1 , q + 1) − 1] .
(d) Deduce that R(2 n + 1 , 2 n + 1) ≥ 5 n + 1 for n ≥ 0, and compare this to the non-
constructive lower bound for R( , ) in Theorem 10.2.14. (Abbott [1972])
10.2.19. Frankl–Wilson [1981] √
constructed explicit n-vertex graphs with clique and inde-
pendence numbers at most 2 c log n log log n , where c is a constant. Obtain from this a lower
bound for R(p , p) larger than every polynomial in p but smaller than any exponential in p.
10.2.20. (+) Prove pe  M1 < R¾ (p; r + 1) ≤ r +  M2 , where M1 = 1
r +1
( p−r 1) and M2 = ( R¾ (p;r)
r
).
10.2.21. (♦) Determine R(K1 ,r1 , . . . , K1 ,r¾ ) (in all cases). (Burr–Roberts [1973])
10.2.22. (♦) Let T be a tree with m edges, and let n be a multiple of m. Determine the
Ramsey number R(T , K1 ,n+1 ). (Burr [1974])
10.2.23. (♦) Let T be a tree with m vertices. Prove that only one 2-coloring of E(K(m−1)(n−1))
(up to isomorphism) has no red T or blue K n . Conclude that if G is formed from K(m−1)(n−1)
by adding one vertex adjacent to s vertices of the clique, then all red/blue colorings of G
have a red T or a blue K n if and only if s > (m − 1)(n − 2). (Hook–Isaak [2011])
460 Chapter 10: Ramsey Theory

10.2.24. For m-vertex trees T , show R(T, K n1 ,...,K n¾ ) = (m− 1)(R(n1 ,...,n¾)− 1)+ 1. (Burr)
10.2.25. (♦) Let G and H be graphs with n = | V(G)| and ¾ = Ò(H). Let s be the least size
of a color class in any proper ¾-coloring of H . When s ≤ n, prove R(G , H) ≥ (n − 1)(¾ − 1) + s.
(Rousseau–Sheehan [1978], Burr [1981])
10.2.26. Prove R(C4 , C4) = 6. (Comment: There are many proofs.)
10.2.27. (♦) Prove R(mK 2 , mK 2) = 3m − 1.
10.2.28. Complete the proof that R(2 K3 , 2K3) = 10.
10.2.29. (♦) For 1 ≤ i ≤ ¾ , let G i be a graph with pi vertices, and mi ∈ . Prove
R(m1 G1 , . . . , m¾ G ¾) ≤ ∑(mi − 1)pi + R(G1 , . . . , G ¾).
10.2.30. (♦) Determine R(P3 , G) for every n-vertex graph G, as a function only of n and
the maximum size of a matching in G.
10.2.31. Let G be an n-vertex graph with m edges. Build a graph H with n + m vertices
such that in every red/blue coloring of E(H), the red subgraph has P3 as an induced sub-
graph or the blue subgraph has G as an induced subgraph. (Comment: The bound n + m
cannot be reduced when the components of G are complete.) (Kostochka–Sheikh [2006])
10.2.32. (♦) Prove that every ¾-coloring of E(K s ,t) has a monochromatic connected sub-
graph with at least (s + t)/¾ vertices. Conclude that every ¾-coloring of E(K n) has such a
subgraph with at least n/(¾ − 1) vertices. For ¾ = 3, show that one cannot guarantee more
than n/2. (Gerencsér–Gyárfás [1967], Gyárfás [1977]) (Comment: Mubayi [2002] and
Liu–Morris–Prince [2009] guaranteed double-stars as such subgraphs of K s ,t . One can
use at least 1/¾ of the vertices in each part when ¾ ≤ 3 (Bucić–Letzter–Sudakov [2018]),
but not when ¾ > 3 (DeBiasio et al. [2018]). For surveys, see Gyárfás [2011, 2016].)
10.2.33. Let G be an n-vertex graph with m edges.
(a) Prove that if ∑v∈V(G) ( d(v)) > (2n) , then G contains a 4-cycle.
2
√
(b) Prove that if m > 4n (1 + 4n − 3), then G contains a 4-cycle.
(c) Prove R¾(C4) ≤ ¾ 2 + ¾ + 2. (Comment: Using difference sets (Section 13.2), there
is a lower bound of ¾ 2 − ¾ + 2.) (Chung–Graham [1975])
10.2.34. (♦) Prove that for n sufficiently large, every 2-coloring of E(K n ,n) has a monochro-
matic copy of Kp ,p . Show that the analogue fails for tripartite graphs: 2-colorings of
E(K n ,n ,n) need not have monochromatic triangles no matter how large n is.
10.2.35. (♦) Bondy [1971a] proved that if xy ∈
/ E(G) implies d(x) + d(y) ≥ n for an n-vertex
graph G, then G = K t ,t or G has a cycle of each length from 3 to n. Use this to prove
R(Cm , K1 ,r) = max{m , 2r+1}, except maybe when m is even and m ≤ 2r. (Lawrence [1973]).
10.2.36. (♦) Ramsey numbers of connected n-vertex graphs.
(a) Prove that if G is a connected bipartite graph with parts of sizes r and s, with
r ≥ s, then R(G , G) ≥ max{2r − 1 , r + 2s − 1}. (Burr [1974])
(b) Let G be a connected n-vertex graph. Prove R(G , G) ≥ ⌊ (4n − 1)/3⌋ .
(c) Prove R(H , H) ≥ max{2r − 1 , r + 2s − 1}, where H = K1 ,r−1 + K1 ,s and r > s.
(d) The formula of part (c) is in fact an upper bound (Rosta). Use this to prove
that parts (a) and (b) are sharp by computing R(G , G) when G is the tree obtained from
K1 ,r−1 + K1 ,s−1 by adding an edge joining leaves of the two components. (Burr–Erdős [1976])
(Comment: Burr and Erdős conjectured R(G , G) ≤ 2n − 2 when G is an n-vertex tree.)
10.2.37. (+) For m ≥ n ≥ 2, prove R(Pm , Pn) = m + ⌊ n/2⌋ − 1. (Gerencsér–Gyárfás [1967])
10.2.38. For n ≥ 3, prove that every 2-coloring of E(K n) has a spanning cycle that consists
of at most two monochromatic paths. (Gerencsér–Gyárfás [1967]) (Comment: A version
for digraphs due to Raynaud [1973] was proved also in Gyárfás [1983].)
Section 10.3: Further Topics 461

10.2.39. (+) Ramsey numbers for cycles. Let º be a 2-coloring of E(K n).
(a) Prove that if º contains a monochromatic copy of C2¾+1 for some ¾ ≥ 3, then º also
contains a monochromatic copy of C2¾ .
(b) Prove that if º contains a monochromatic C2¾ for some ¾ ≥ 3, then º also contains
a monochromatic C2¾−1 or 2K ¾ .
(c) Prove R(Cm , Cm) ≤ 2m − 1 for m ≥ 5 (m = 4 is Exercise 10.2.26). (Hint: Use the
Erdős–Gallai [1959] result that an n-vertex graph with more than (m − 1)(n − 1)/2 edges
has a cycle of length at least m (Theorem 7.3.16). There is still one difficult case.)
10.2.40. 3-color Ramsey numbers for cycles.
(a) Prove R(Cn , Cn , Cn) ≥ 2n when n is even.
(b) Prove R(Cn , Cn , Cn) ≥ 4n − 3 when n is odd.
(Comment: Equality holds in both statements when n is sufficiently large, proved
by Benevides–Skokan [2009] for even n. Jenssen–Skokan [2016] proved for odd n that
R¾(Cn) = 2 ¾−1(n − 1) + 1. Both papers use the Regularity Method (Section 11.1).)
10.2.41. (♦) Chromatic Ramsey number.
(a) Prove the general lower bound RÒ(G; s) > (Ò(G) − 1)s . (Comment: Burr–Erdős–
Lovász [1976] conjectured that this lower bound is sharp; this was proved by Zhu [2011].)
(b) Let G be a 3-chromatic graph. Prove that 5 ≤ RÒ(G; 2) ≤ 6, with equality in the
lower bound if and only if there is a homomorphism from G into C5 .
10.2.42. (♦) In a family of axis-parallel squares in 2 , a set is homogeneous if it is pairwise
intersecting √
or pairwise disjoint. Prove that n squares always have a homogeneous set of
size at least n/4.√Prove for infinitely many n that some family has no homogeneous set
of size more than 4n/5. (Hagelstein–Herden–Young [2017], Hagelstein–Herden [2018])
(Comment: See Larman–Matoušek–Pach–Törőcsik [1994] for related problems.)
10.2.43. (♦) Given ¾ , l ∈ , prove that there exists N¾ ,l such that for d ≥ N¾ ,l , every
¾-coloring of the edges of the d-dimensional cube Qd contains a monochromatic path of
length l that is a shortest path joining its endpoints. (Hint: Given the value N¾ ,l , obtain
an upper bound on N¾ ,2l to prove that it exists.) (Stong [2018])

10.2.44. The zero-sum Ramsey number R(G , ¾) of a graph G with | E(G)| divisible by
¾ is the least n such that every coloring of E(K n) by integers has a copy of G on which the
sum of the colors is divisible by ¾ . (Caro [1996] surveys such problems.)

(a) Prove that R(G , ¾) is well-defined.

(b) Prove R(K3 , 3) = 11.

10.3. Further Topics

Ramsey Theory is also called “partition calculus”. The idea is that every
partition of a large configuration contains a subconfiguration with more struc-
ture. We consider coloring integers to obtain arithmetic structure, the infinite
version of Ramsey ’s Theorem and its application to the finite version, and finally
the Canonical Ramsey Theorem, which allows infinitely many colors.

VAN DER WAERDEN’S THEOREM

Ramsey ’s Theorem was not the first result in the partition calculus, but it
is the best known and most thoroughly studied. The first result about monochro-
462 Chapter 10: Ramsey Theory

matic structures in colorings of natural numbers was due to Hilbert [1892] (Ex-
ercise 3). We begin with a later result that still pre-dates Ramsey ’s Theorem and
arose from an attempt to prove Fermat ’s Last Theorem.

10.3.1. Theorem. (Schur ’s Theorem; Schur [1916]) Given ¾ > 0, there is an

integer s¾ such that every ¾-coloring of {1 , . . . , s¾ } yields monochromatic (not
necessarily distinct) x , y , solving x + y = .
Proof: Let r¾ = R¾ (3; 2). We show that s¾ < r¾ by showing that every -coloring
 of [r¾ − 1] has a monochromatic solution to x + y = . From  , we define a -
coloring  of E(K r¾ ). Let V(K r¾ ) = [r¾ ]. Let the color of edge i j in  be  (| i − j |).
By the choice of r¾ ,  yields a monochromatic triangle with some vertices
a , b , c. We may assume a < b < c. Let x = b − a, y = c − b, = c − a; we have
 (x) =  (y) =  ( ). By construction, x + y = .

10.3.2. Example. Bounds for Schur numbers. Like Ramsey numbers, few Schur
numbers are known exactly. We know s1 = 2, s2 = 5, s3 = 14, and s5 = 45 (by
computer). The upper bound of r¾ − 1 yields s¾ ≤ ⌊  !e⌋ (see Exercise 10.2.14).
Schur proved s¾ ≥ (3¾ + 1)/2. Given a -coloring  of [n] with no monochro-
matic solution to x + y = , define a ( + 1)-coloring  of [3n + 1] by

⎧ (i) if i ≤ n,
⎪
⎪
 (i) = ⎨  (i − 2n − 1) if i ≥ 2n + 2,
⎪
⎪
⎩ +1 if n + 1 ≤ i ≤ 2n + 1.

Checking cases shows that  produces no monochromatic solution to x + y =

(Exercise 1). The resulting recurrence s¾ ≥ 3s¾−1 − 1 yields s¾ ≥ (3¾ + 1)/2.

Schur ’s Theorem generalizes considerably. For a homogeneous linear equa-

tion c1 x1 + · · · + ct x t = 0, we may ask whether -colorings of [n] must contain
a monochromatic solution when n is large. Rado (Schur ’s student) proved that
this holds if and only if ∑i∈ I ci = 0 for some nonempty index set I ⊆ [t]. More
generally, he determined the systems of homogeneous linear equations such that
-colorings eventually force monochromatic solutions.

10.3.3. Definition. Let Cx = 0 be a homogeneous system of linear equations with

rational coefficients, where no column of C is 0. The system is regular if for
all  there is an integer n such that every -coloring of [n] yields a monochro-
matic solution to Cx = 0. The matrix C satisfies the Columns Condition
if there is a partition C1 , . . . , Ct of its columns such that the columns in the
first block sum to 0 and in each subsequent block the sum of the columns is a
rational linear combination of columns in the preceding blocks.

When C has just one row (with no 0), the Columns Condition reduces to hav-
ing a nonempty index set I such that ∑i∈ I ci = 0. These form the first block, and
each later coefficient is a multiple of one in the first block.

10.3.4. Theorem. (Rado’s Theorem; Rado [1933]) A linear system Cx = 0 is

regular if and only if C satisfies the Columns Condition.
Section 10.3: Further Topics 463

Rado ’s Theorem was extended by Deuber [1973]. We will prove a weaker ver-
sion of Rado ’s Theorem for coefficient matrices of the type below.

⎛1 −1 0 0 0 −1 ⎞
⎜0
⎜
1 −1 0 0 −1 ⎟
⎟
⎜0 0 1 −1 0 −1 ⎟
⎝0 0 0 1 −1 −1 ⎠
The columns before the last sum to zero, and the last is a linear combination of
the others by assigning weight i to column i. Thus the matrix does satisfy the
Columns Condition.
To interpret the equations corresponding to the matrix, name the variables
x1 , . . . , x l and d. The equations then become x i+1 = x i + d for 1 ≤ i < l. In other
words, x1 , . . . , x l must form an l-term arithmetic progression. Under Rado ’s The-
orem, we are guaranteed a monochromatic arithmetic progression where the con-
stant difference d also has the same color. We will prove the weaker statement
that does not require d to have the same color as the elements of the progression.

10.3.5. Theorem. (van der Waerden’s Theorem; van der Waerden [1927])
Given positive integers l and ¾ , there is an integer w(l , ¾) such that every
¾-coloring of [w(l , ¾)] has a monochromatic l-term arithmetic progression.

We present the proof by Graham–Rothschild [1974]. Mills [1983] rephrased

it (and that of the subsequent Hales–Jewett Theorem) more compactly. Our pre-
sentation follows that of Graham–Rothschild–Spencer [1980]. The details of a
small case suggest the ideas in the proof.

10.3.6. Example. We prove w(3 , 2) ≤ 325. This is quite loose, since w(3 , 2) = 9
(Exercise 8). Like the inductive proof of Ramsey ’s Theorem, this inductive proof
gives a ridiculously large upper bound for w(l , ¾).
Consider a 2-coloring of [325]. Partition [325] into 65 blocks of 5 consecutive
integers. The coloring gives some red-blue pattern to each block; there are 32
possible patterns. By using 65 blocks, some pattern is repeated by the time we
reach the middle block. Thus we can find three equally-spaced blocks such that
the first two have the same pattern. Similarly, we use 5-element blocks so that we
find a duplicated color by the middle of any block. Thus we can find three equally
spaced elements in the block such that the first two have the same color.
For example, when blocks 3 and 30 have the same pattern RBRBB, con-
sider block 57 with elements 281-285. We obtain either a blue progression as
{15 , 150 , 285} or a red progression as {11 , 148 , 285}.

Of the 32 color patterns on blocks, 19 contain 3-term monochromatic progres-

sions. Hence we need only 27 blocks, not 65. We also ignored other ways to obtain
3-term progressions. By considering all possible progressions, Exercise 8 cuts the
bound from 325 to the actual minimum, 9.
Nevertheless, block structures lead to a proof of van der Waerden’s Theorem
by induction on l without seeking a strong bound on w(l , ¾). The bounds become
enormous because when we want a progression of l − 1 blocks having the same
pattern, the exponentially many patterns behave as distinct colors. With larger
464 Chapter 10: Ramsey Theory

¾ , we will also need to consider more potential monochromatic progressions. To

manage this, we introduce a way to embed a multidimensional grid in [n].

10.3.7. Definition. Let [0 , l]m be the set of integer m-tuples (x1 , . . . , x m) such
that 0 ≤ x i ≤ l. The jth critical class Cj is defined by
Cj = {x ∈ [0 , l]m : x i = l ⇔ i ≤ j}.
An m-dimensional block structure (of length l) in [n] is a choice of posi-
tive integers a , d1 , . . . , d n that map [0 , l]m to [n] injectively by sending x to
a + ∑ x i d i . A block structure is layered for a coloring º of [n] if º is constant
on the image of each critical class.
x2
0 ··· l
0
.. C0
x1 .

l C1 C2

The set [0 , l]m is an m-dimensional grid of lattice points with l + 1 points on

each side, sketched above for m = 2. The critical class Cj consists of the m-tuples
whose first j entries equal l and whose last m − j entries are less than l. There
are m + 1 critical classes, with | Cj | = l m− j .
Van der Waerden’s Theorem says roughly that for large [n] every ¾-coloring
contains a layered 1-dimensional block structure of length l. The image of C0 is
then the monochromatic arithmetic progression {a + rd1 : 0 ≤ r ≤ l − 1}.
To prove this by induction on l we will need higher-dimensional layered block
structures. In a layered block structure on [0 , l]m using m colors, two of the m + 1
critical classes must have the same color. For any two critical classes, there is an
l-term arithmetic progression in the image of the larger one that extends to an
(l + 1)-term progression using an element of the smaller one, as suggested by the
figure above. To guarantee progressions of length l + 1 we will need to guarantee
layered block structures of length l for all dimensions. Hence we again have a
double induction, like in Ramsey ’s Theorem.

10.3.8. Theorem. (van der Waerden [1927], Graham–Rothschild [1974]) For


l , m , t ∈ , there exists w(l , m , t) such that every t-coloring of [w(l , m , t)]
contains a layered block structure on [0 , l]m .
Proof: We use induction on l. For each l, we use induction on m. For each l we
prove a basis step for m = 1 (using the hypothesis for shorter lengths) and an
induction step for m > 1. Each argument holds for all t.
For the basis step m = 1, we bound w(l + 1 , 1 , t) in terms of w(l , m , t ) for all

m , t (focusing the classes in a layered structure). For the induction step m > 1,
we bound w(l , m + 1 , t) in terms of w(l , ¾ , t ) for all ¾ ≤ m and all t (creating
the high-dimensional block structure). Since we have w(1 , 1 , t) = 2 as the overall
basis, we do the latter step first.
Section 10.3: Further Topics 465

Step 1: w(l , m+ 1 , t) ≤ MM , where M = w(l , m , t) and M = w(l , 1 , t M). Con-

sider a t-coloring º : [MM ] → [t]. Partition [MM ] into M consecutive blocks of
length M. Induce a coloring º : [M ] → [t M ] by the color patterns used on the
blocks. There are t M possible color patterns for a block of M elements; we have M
such blocks. Since M = w(l , 1 , t M), there are l + 1 equally spaced special blocks
with the same pattern on the first l of them. Let a be the index of the first special
block, and let d be the difference between indices of successive special blocks.
Since M = w(l , m , t), the pattern (t-coloring) in the first special block con-
tains a layered block structure on [0 , l]m . Stacking the l + 1 special blocks in-
creases the dimension of the block structure for the full coloring. Since the pat-
tern is the same for the first l special blocks, the corresponding critical classes
have the same color and stack to form a monochromatic critical class l times as
large in the (m + 1)-dimensional block structure.
To be precise, let the layered block structure in the first special block be
a , d1 , . . . , d m when viewing its first position as 1 instead of 1 + (a − 1)M. For
the block structure in the full coloring, let d m+1 = d M and a = (a − 1)M + a .
The repetition of pattern in the first l special blocks yields monochromatic criti-
cal classes l times as large as before. The (m + 1)th critical class, corresponding to
x = (l , . . . , l), is the last element of the (l + 1)th special block. No other element
of this block belongs to any critical class, so the colors in this block are irrelevant.
Step 2: w(l + 1 , 1 , t) ≤ ⌈ w(l , t , t) l+l 1 ⌉ . Let n = ⌈ w(l , t , t) l+l 1 ⌉ , and con-
sider º : [n] → [t]. We force the 1-dimensional arithmetic progression by find-
ing a t-dimensional layered block structure. By the choice of n, º is constant on
critical classes of some block structure on [0 , l]t in [w(l , t , t)] with parameters
a , d1 , . . . , d t . Two of the t + 1 monochromatic critical classes must have the same
color, say classes r and s with r < s. To obtain the desired 1-dimensional block
structure, we use special elements in [0 , l]t . Let x(i) have the first r coordinates
be l, the next s − r coordinates be i, and the remaining coordinates be 0.
The elements x(0) , . . . , x(l−1) belong to class r, and x l belongs to class s, but
their images have the same color. Set a = a + l ∑i=1 d i and d = ∑i=r+1 d i .
r s

These l + 1 images form a monochromatic arithmetic progression with constant

difference d ; this is the 1-dimensional critical class. Finally, add the element
a + (l + 1)d as a singleton critical class. This element belongs to [n] due to the
factor (1 + 1/l) in the definition of n.


10.3.9.* Remark. When we color with finitely many colors, van der Waerden’s
Theorem guarantees arbitrarily long arithmetic progressions. “Arbitrarily long ”

does not mean infinite! Consider 2-coloring by the parity of ⌈ lg n⌉ . An arith-
metic progression in one color has a fixed difference d. No matter where it starts,
later there are more than d consecutive integers in the other color. Hence there
is no infinite monochromatic arithemetic progression.
We know even less about van der Waerden numbers than about Ramsey num-
bers: w(2 , 2) = 3, w(3 , 2) = 9, w(4 , 2) = 35, w(3 , 3) = 27, w(3 , 4) = 76, w(5 , 2) =
178, w(6 , 2) = 1132 (Kouril–Paul [2008]), and w(4 , 3) = 293 (Kouril [2012]).
The proof of Theorem 10.3.8 yields what Graham–Rothschild–Spencer [1980]
called “Eeeeenormous Upper Bounds.” For diagonal Ramsey numbers, the induc-
r −1
tion yields roughly R(q; r) ≤ 2 R(q;r−1) ; this builds a “tower ” of r exponentiations.
The bound from Theorem 10.3.8 grows even faster. Let wl(¾) be the result-
466 Chapter 10: Ramsey Theory

ing bound on the number required to force monochromatic l-progressions in a

¾-coloring, viewed as a function of ¾ for each l. Let vl(¾) be defined by letting
v2(¾) = 2 ¾ and for l ≥ 3 letting vl(¾) be the result of ¾ iterations of vl−1 starting
with argument 1. Thus v3(¾) = 2¾ , v4(¾) is an exponential tower of ¾ 2s, and v5(¾)
is already difficult to describe. Even vl(2), in terms of l, grows unimaginably fast.
Shelah [1988] found a new proof that improves the upper bound. His bound
is what logicians call “primitive recursive”, whereas the bound from Theorem
10.3.8 is not. Gowers [2001] further improved the bounds.
Another short proof of van der Waerden’s Theorem uses the language and
properties of “ultrafilters”. An ultrafilter on a set X is a nonempty family of
subsets of X that is upwardly closed, is closed under intersection, omits ∅, and
contains at least one of A and A for all A ⊆ X . Hindman [1979] proved that
results in Ramsey theory that guarantee homogeneous sets correspond to the ex-
istence of an ultrafilter on the set being colored. Many results in Ramsey theory
have been proved by this method. A thorough survey of the subject, focusing on
van der Waerden’s Theorem, can be found in Johannson [2007].

Van der Waerden’s Theorem implies Rado ’s Theorem for systems with one
equation, but this still takes some effort. We present an easier application. For a
finite set S of integers, let P(S) be the set of sums of nonempty subsets of S. Folk-
man and others proved that in every ¾-coloring of there are arbitrarily large
sets S such that P(S) is monochromatic. This follows easily from Rado ’s Theorem
(Exercise 19). To prove it from van der Waerden’s Theorem, we need a lemma.

10.3.10.* Lemma. A set S ⊆  with sum at most n is max-governed under a

coloring of [n] if the color of the sum of a subset of S depends only on its max-

imum element. Given l , ¾ ∈ , there exists M(l , ¾) ∈  such that under
every ¾-coloring of [M(l , ¾)] some l-set is max-governed.
Proof: We use induction on l; note that M(1 , ¾) = 1. Existence of M(2 , ¾) follows
using the Schur number s¾ . For l > 2, let m = M(l − 1 , ¾). Set n = 2w(m + 1 , ¾),
and consider a ¾-coloring º of [n]. By van der Waerden’s Theorem, the top half
of [n] has a monochromatic arithmetic progression T with m + 1 terms. Let a l
be its first term and d be its constant difference. Since T fits in an interval of
length n/2, we have dm ≤ n/2.
Define a coloring º : [m] → [¾] by º (j) = º (d j). The induction hypothesis
yields an (l − 1)-set S with sum at most m that is max-governed under º . Let
S = (dS ) ∪ {a l}. Since dm ≤ n/2 < a l , the largest element of S is a l . The sum
of any subset containing a l exceeds a l by a multiple of d that is at most dm, so it
belongs to T and has the same color as a l .

10.3.11.* Theorem. (Folkman’s Theorem; Folkman [1970], Sanders [1969],


Rado [1970]) For ¾ , t ∈ , there exists N ∈ 
such that every ¾-coloring of
[N] yields a t-set with sum at most N whose nonempty subsets have the same
color as their sums.
Proof: In Lemma 10.3.10, let N = M(¾(t − 1) + 1 , ¾). Given a ¾-coloring º of [N],
we obtain a1 , . . . , a¾(t−1)+1 such that the sum of every nonempty subset has the
same color as its largest element. The Pigeonhole Principle yields t of them in
the same color. These form the desired set.
Section 10.3: Further Topics 467

The block structures in the proof of van der Waerden’s Theorem suggest that
it concerns sets and arrangements more than integers. The Hales–Jewett Theo-
rem captures the combinatorial essence. The proof is analogous, so we just state
the result. The point is that when we color a grid in sufficiently many dimensions,
we obtain a monochromatic line. We must define “line” carefully.
Let S = {a0 , . . . , a l−1 }, and let S m be the set of m-tuples drawn from S. We
designate elements of S m by the vector of subscripts. Thus S m can be viewed as
Clm , the m-dimensional grid with l points on each edge. Only the order of ele-
ments in S matters, not their values.

10.3.12. Definition. A line in Clm is a set of l distinct vectors x(0) , . . . , x(l−1) de-
termined by a partition B0 , B1 of [m] and constants cj for j ∈ B0 by setting
(i) (i)
x j = cj if j ∈ B0 and x j = i if j ∈ B1 (B0 may be empty).

In this notion of line, coordinate values vary in the same order on coordinates
in B1 , and they are fixed on each coordinate in B0 . Not every geometric line is a
line here; {210 , 111 , 012} is not a line in C33 .

10.3.13. Theorem. (Hales–Jewett Theorem; Hales–Jewett [1963]) Given pos-

itive integers l , t, there is an integer HJ(l , t) such that every t-coloring of Clm
with m ≥ HJ(l , t) has a monochromatic line.

The Hales–Jewett Theorem includes van der Waerden’s Theorem and implies
Folkman’s Theorem. We present a geometric application. Two sets V, W ∈ m
are homothetic if one arises from the other by a scale change and a transla-
tion, meaning a correspondence between {vi} and {wi} such that wi = cvi + with
c ∈ − {0} and ∈ m .

10.3.14. Theorem. (Gallai–Witt Theorem; see Rado [1933], Witt [1952]) If the
points of m are colored with finitely many colors, then for every finite set
V ⊂ m there exists a monochromatic W that is homothetic to V .
Proof: Let : m → [t] be the coloring. Let l = | V | and N = HJ(l , t). View the
elements of the cube ClN as lists (x1 , . . . , x N) with x i ∈ V . Define : ClN → m
by (x1 , . . . , x N) = ∑ i=1 i x i for constants 1 , . . . ,  N chosen to make  injective.
N

We need only avoid finitely many equalities of the form ∑ i(x i − x i) = 0 and can
choose 1 , . . . ,  N to be integers.
Once  is injective, define : ClN → [t] by (x) = ((x)). The choice of
N guarantees a monochromatic line; let B0 , B1 be its coordinate partition. The
image of the line under  is monochromatic under . To see that it is homothetic
to V , observe that (x(j)) = cvj +  , where c = ∑i∈ B1 i and  = ∑i∈ B0 i x i ; both c
(j)

and  are constants.

If the elements of V are positive integer vectors, then we find a monochro-

matic set homothetic to V in a sufficiently large (i.e., finite) grid. If we choose V
to be a cube itself, then Gallai’s Theorem guarantees a higher-dimensional ver-
sion of a monochromatic arithmetic progression.
468 Chapter 10: Ramsey Theory

INFINITE SETS (optional)

Most statements of Ramsey theory can be phrased concisely using the vague-
sounding “sufficiently large”. A statement about infinite sets would be simpler.

That is, when we ¾-color the r-subsets of , must there be an infinite homogeneous
set? This seems too much to ask but would avoid quotas and Ramsey numbers and
would imply the finite version.
Extensions to infinite sets can be dangerous, leading to unsupported leaps of
faith and false statements. For example, injections from A to B and B to A must
be bijections when A and B are finite, but this fails when A and B are infinite
(doubling and negation are injections from to , but doubling is not a bijection).
In contrast, Ramsey ’s Theorem behaves as desired. Like the proof of the fi-
nite version, this proof invokes the induction hypothesis for both r − 1 and 1.
However, we ignore quotas and “Ramsey numbers” and don’t care which color
provides the infinite homogeneous set.

10.3.15. Theorem. (Infinite Ramsey Theorem; Ramsey [1930]) For finite r

and ¾ , every ¾-coloring of (r ) has an infinite homogeneous set.
Proof: We use induction on r. The case r = 1 is the Pigeonhole Principle: every
finite partition of an infinite set has an infinite class. For r > 1, consider a ¾-
coloring º of (r ). We produce an infinite sequence of vertices x∗1 , x∗2 , · · · forming
a homogeneous set under º .
First we construct sequences ⟨x⟩ , ⟨c⟩ , ⟨D⟩. The elements of ⟨x⟩ are natural
numbers (candidates for ⟨x∗ ⟩), those of ⟨c⟩ are colors, and those of ⟨D⟩ are infinite
subsets of . Begin with D0 = . Having constructed Dn−1 , construct x n , cn ,
and Dn as follows. Let x n be the least element of Dn−1 . Induce a ¾-coloring º of
the (r − 1)-subsets S of Dn−1 − {x n} by setting º (S) = º (S ∪ {x n}). Since Dn−1
is infinite, the induction hypothesis guarantees an infinite subset of Dn−1 that is
homogeneous under º in some color c. Let this set be Dn and this color be cn .
Since Dn is infinite, the construction proceeds and constructs sequences ⟨x⟩
and ⟨c⟩. Since there are finitely many colors, some value c∗ occurs infinitely often
in ⟨c⟩, by the Pigeonhole Principle (r = 1). Let {ni : i ∈ } be the set of indices
such that cni = c∗ . Define ⟨x∗ ⟩ by x∗i = x ni .
We claim that º (S) = c∗ whenever S consists of r terms in ⟨x∗ ⟩. Let the least
element of S be x n . All other elements of S belong to Dn . By the construction,
the union of these with the element x n has color c∗ under º . We conclude that
⟨x∗ ⟩ forms the desired infinite set.

Ramsey [1930] proved both versions separately. The infinite version can be
used to prove the finite version (but not the other way!) via the “Compactness
Principle” (also called Rado Selection Principle), which was discovered repeat-
edly (see Rado [1949], Erdős [1950], Gottschalk [1951], de Bruijn–Erdős [1951]).
The Compactness Principle relates colorings of an infinite hypergraph to col-
orings of finite subhypergraphs. If A is a subset of the vertex set of H , then the
induced subhypergraph H A is the hypergraph with vertex set A whose edges
are the edges of H contained in A.
Section 10.3: Further Topics 469

10.3.16. Theorem. (Compactness Principle) Let H be a hypergraph with a

countable vertex set and finite edges. If Ò(H A) ≤ ¾ for every finite set of ver-
tices A, then Ò(H) ≤ ¾ .
Proof: Given Ò(H A) ≤ ¾ for all finite A, we construct a proper ¾-coloring of H.

Rename the vertices so V(H) = . We develop the coloring iteratively, defining
each º ∗(j) so that the full coloring º ∗ :  → [¾] properly colors H. Let ºn be a
proper ¾-coloring given for H[n] by hypothesis.
Having defined º ∗(1) , . . . , º ∗(j − 1), let Sj −1 = {n ≥ j: ºn(i) = º ∗(i) for
1 ≤ i ≤ j − 1}. That is, Sj −1 is the set of future values n such that ºn agrees
with the choices made so far for º ∗ . Given that Sj −1 is infinite, we will choose
º ∗(j) so that Sj is also infinite.

Begin with S0 = . Suppose that º ∗(1) , . . . , º ∗(j − 1) have been defined so
that Sj −1 is infinite. For n ∈ Sj −1 , partition Sj −1 into ¾ classes according to ºn(j).
Since Sj −1 is infinite, some such class is infinite; call it Sj , and define º ∗(j) to
agree with ºn(j) for all n in Sj .
This defines a coloring º ∗ ; we check that º ∗ is a proper ¾-coloring. Let E be
an edge of H , and let m be its highest vertex. Since Sm is infinite, there exists
n ∈ Sm such that º ∗ agrees with ºn on [m]. Since ºn properly colors E, also º ∗
properly colors E.

The Compactness Principle yields results on finite structures from results on

infinite structures. This can yield shorter existence proofs of theorems on finite
sets but gives no information on bounds. In some sense this is not such a great
deficiency, since the bounds from inductive proofs are so bad.
The contrapositive is the usual form for applications: If Ò(H) > ¾ , then there
is a finite A ⊆ V(H) such that Ò(H A) > ¾ . We demonstrate the use of the Com-
pactness Principle on Ramsey ’s Theorem itself. We will need to restate Ramsey ’s
Theorem in terms of hypergraph coloring.


10.3.17. Corollary. (Finite Ramsey Theorem) For ¾ , r, q ∈ , there is n ∈ 
such that every ¾-coloring of ([n]
r
) has a homogeneous q-set.
Proof: Let H q be the hypergraph with vertex set (r ) that has an edge (Sr) for each

q-set S ⊆ ; note that every edge is finite (with size (qr)). By Ramsey ’s Theorem
(Theorem 10.3.15) every ¾-coloring of (r ) has an infinite homogeneous set; hence
it also has a homogeneous set of size q.
We conclude that H q is not ¾-colorable. Every edge of H q is finite, so the Com-

pactness Principle guarantees a finite set A ⊆ such that H Aq is not ¾-colorable.
Letting n be the largest element of A, we have H Aq ⊆ H[n]q
. We conclude that the
desired n exists and is at most | A|.

If we finitely color the r-subsets of an uncountable set, will there be an un-

countable homogeneous subset? The answer is yes when r = 1, but when r ≥ 2
we have trouble seeking a set with an order-preserving bijection to . The Well-
Ordering Principle states that every set S has a well ordering, meaning a linear
ordering with respect to which every nonempty subset of S has a least element.
The Well-Ordering Principle is equivalent to the Axiom of Choice.
470 Chapter 10: Ramsey Theory

10.3.18. Theorem. Given the Well-Ordering Principle, there is a 2-coloring of

the unordered pairs in such that no subset that corresponds to under the
usual ordering has its pairs monochromatic.
Proof: Let ≺ be a well ordering of , and let < be the usual ordering. For a , b ∈
, color {a , b} red if a and b have the same order under < and ≺ ; otherwise color it
blue. If some red-homogeneous subset corresponds bijectively to under <, then
< agrees with the well ordering ≺ on those points, but is not well-ordered by <.
Similarly, such a blue-homogeneous set yields a well ordering of under >.

THE CANONICAL RAMSEY THEOREM (optional)

What happens for infinite sets of colors? As usual, we color (r ). When r = 1,
the situation is simple. If we use infinitely many colors, then we have an infinite
set receiving distinct colors. Otherwise, we use only finitely many colors, and
then the Pigeonhole Principle guarantees an infinite monochromatic set. When
r ≥ 2, we can no longer guarantee an infinite set whose r-sets all receive distinct
colors or all receive the same color, but we can guarantee a more general object.


10.3.19. Definition. For S ⊂ , a coloring º of (Sr) is canonical if there is a set
of indices I ⊆ [r] such that º (A) = º (B) if and only if for i ∈ I the ith largest
elements of A and B are equal. A colored set is rainbow (or polychromatic
or heterochromatic) if its elements have distinct colors.

The condition for º (A) = º (B) is not that the ith largest elements of A and B
are equal if and only if i ∈ I; additional equalities between A and B are allowed.
For example, if I = ∅, then the condition holds for all {A , B}, so in a canonically
colored S with I = ∅ all r-sets have the same color. On the other hand, if I =
[r], then r-tuples get the same color if and only if they agree completely, so (Sr) is
rainbow colored when I = [r], using distinct colors on all r-sets.
When r = 2, there are two other possibilities: a canonically-colored S with
I = {1} has º (A) = º (B) if and only if min(A) = min(B), and with I = {2} the
requirement is max(A) = max(B). We state the infinite and finite versions of the
general theorem (see Graham–Rothschild–Spencer [1980, 1990, Chapter 5]).

10.3.20. Theorem. (Canonical Ramsey Theorem; Erdős–Rado [1950]) Every

coloring of (r ) is canonical on some infinite set. For every q, there exists n
such that every coloring of ([n]r
) is canonical on some q-set.

For the finite version when r = 1, see Exercise 11. We prove just the infinite
version for r = 2. The four possibilities for a canonical coloring are monochro-
matic, min-coloring, max-coloring, and rainbow.

10.3.21. Theorem. Every coloring º of (2 ) is canonical on some infinite set.

Proof: Write a 4-tuple A in  as {a1 , a2 , a3 , a4} with a1 < a2 < a3 < a4 . The
behavior of º on A induces a partition of the edges of the complete graph on
{1 , 2 , 3 , 4} by putting i j and ¾ l in the same class if º ({ai , aj }) = º ({a¾ , a l}).
Section 10.3: Further Topics 471

With this fixed vertex set {1 , 2 , 3 , 4}, there are 203 possible such patterns, cor-
responding to the 203 distinct partitions of [6] (that is, ∑¾=1 S(6 , ¾) = 203).
6

By the Infinite Ramsey Theorem (Theorem 10.3.15) with 203 colors, under º
all 4-sets in some infinite set S are colored by the same pattern. In fact, as soon
as S has at least seven elements, only a few patterns can hold for all 4-sets in S.
Consider the equalities below, dropping the set brackets within º ({i , j}).
º (a1 , a2) = º (a1 , a3) º (a2 , a4) = º (a3 , a4)
º (a1 , a2) = º (a1 , a4) º (a1 , a4) = º (a3 , a4)
º (a1 , a3) = º (a1 , a4) º (a1 , a4) = º (a2 , a4)
º (a2 , a3) = º (a2 , a4) º (a1 , a3) = º (a2 , a3)
If the pattern on {1 , 2 , 3 , 4} satisfies the first column of equalities whenever
{a1 , a2 , a3 , a4} ⊂ S with a1 < a2 < a3 < a4 , then the coloring is min-determined.
This is distinct from min-coloring in that min-determined does not imply that
º (i , j)
= º (¾ , l) when min{i , j}
= min{¾ , l}. Similarly, a pattern satisfying the
second column is max-determined.
Without loss of generality, let [7] be the first seven elements of S, in order.
Since the pattern is constant over (S4 ), a single equality yields the four require-
ments in either column when applied to various 4-sets in [7]:
if º (2 , 4) = º (2 , 6), if º (2 , 6) = º (4 , 6),
then min-determined then max-determined
2 4 6 7 º (a1 , a2) = º (a1 , a3) 1 2 4 6 º (a2 , a4) = º (a3 , a4)
2 4 5 6 º (a1 , a2) = º (a1 , a4) 2 3 4 6 º (a1 , a4) = º (a3 , a4)
2 3 4 6 º (a1 , a3) = º (a1 , a4) 2 4 5 6 º (a1 , a4) = º (a2 , a4)
1 2 4 6 º (a2 , a3) = º (a2 , a4) 2 4 6 7 º (a1 , a3) = º (a2 , a3)
If º (2 , 4) = º (2 , 6) = º (4 , 6), then the pattern puts all edges of K 4 in the
same class, and º is monochromatic on (S2 ). To complete the proof, we need to
consider the consequences of º (2 , 6)
= º (4 , 6) and º (2 , 4)
= º (2 , 6). Again we
list these in two columns
if º (2 , 6)
= º (4 , 6) if º (2 , 4)
= º (2 , 6)
2 4 6 7 º (a1 , a3)
= º (a2 , a3) 1 2 4 6 º (a2 , a3)
= º (a2 , a4)
2 4 5 6 º (a1 , a4)
= º (a2 , a4) 2 3 4 6 º (a1 , a3)
= º (a1 , a4)
2 3 4 6 º (a1 , a4)
= º (a3 , a4) 2 4 5 6 º (a1 , a2)
= º (a1 , a4)
1 2 4 6 º (a2 , a4)
= º (a3 , a4) 2 4 6 7 º (a1 , a2)
= º (a1 , a3)
If º (2 , 4) = º (2 , 6)
= º (4 , 6), then both columns on the left hold. From the
first, the coloring is min-determined. The second then ensures that the pattern
is a min-coloring. Similarly, º (2 , 4)
= º (4 , 6) = º (2 , 6) forces max-coloring.
Finally, when º (2 , 4)
= º (2 , 6)
= º (4 , 6), we must show that the pattern has
no two edges with the same color. The two columns of inequalities establish dis-
tinctness for eight pairs among the six edges; there are seven more to check. For
example, if º (a1 , a2) = º (a2 , a3), then {1 , 2 , 4 , 5} and {1 , 2 , 6 , 7} yield º (2 , 4) =
º (1 , 2) = º (2 , 6), a contradiction. We leave the remaining cases as Exercise 12.
For the finite version with r = 2 and q ≥ 7, it suffices to let n be the Ramsey
number such that every 203-coloring of ([n]
2 ) has a homogeneous q-set S. The proof
S
above then guarantees that ( 2 ) is canonically colored.
472 Chapter 10: Ramsey Theory

The Canonical Ramsey Theorem has interesting applications to graphs. In a

graph with vertex set [n], an increasing path is a path through successively in-
creasing vertices. The Canonical Ramsey Theorem implies that for large enough
n, every coloring of E(K n+1) has a monochromatic increasing path of length l or
a rainbow increasing path of length ¾ . Let º (l , ¾) be the least such n. Lefmann–
Rödl–Thomas [1992] proved constructively that º (l , ¾) ≥ l¾−1 and conjectured
that equality holds. (Hence also the value of n needed to force a canonical q-set
for r = 2 is at least (q + 1)q .) Jiang–Mubayi [2000] provided support for √ the con-
jecture by proving that º , ¾ is bounded by + o(1))l
(l ) (1 ¾ − 1
when ¾ ∈ o( l).
Another variation ignores the names of colors. For example, when we want a
monochromatic copy of G we care not which colors are used. More generally, we
may seek certain patterns of contrast among colors.

10.3.22. Definition. An edge-coloring of a graph is lexical if the vertices can be

linearly ordered so that edges have the same color if and only if they have the
same lower endpoint. A pattern is a graph G plus a partition of E(G) into
color classes. A pattern G contains a pattern H if H occurs in G with edges
partitioned in G the same as in the given classes in H.

The formally weaker version of the Canonical Ramsey Theorem for r = 2 is

 
that for q ∈ , there exists n ∈ such that every edge-coloring of K n contains a
monochromatic, a rainbow, or a lexically colored K q .

10.3.23. Definition. A family F of patterns is a Ramsey family if for some in-

teger n every edge-coloring of K n contains a pattern in F. The smallest such
n is the pattern Ramsey number R(F).

10.3.24. Theorem. (Jamison–West [2004]) A family of color patterns is a Ram-

sey family if and only if it has a monochromatic pattern, a rainbow pattern,
and a lexical pattern (not necessarily distinct).
Proof: Let F be a Ramsey family with pattern Ramsey number n. Every sub-
graph of a pattern that is monochromatic, rainbow, or lexical inherits that prop-
erty. By considering monochromatic, rainbow, and lexical colorings of K n , we see
that F must contain a pattern of each type.
Conversely, when F contains a monochromatic pattern M , a rainbow pattern
R, and a lexical pattern L, let p be the maximum number of vertices among M , R,
and L. By the Canonical Ramsey Theorem, there exists n such that every edge-
coloring of K n contains a canonically colored K p. If an edge-coloring of K n contains
a monochromatic or rainbow K p , then it contains M or R.
More care is needed when K n contains a lexical K p. The lexical coloring is
relative to some linear order ≺1 on V(K p). Also, the lexical coloring of L assumes
a linear order ≺2 on V(L). Having L as a subpattern requires an order-preserving
injection of V(L) into V(K p). Since both orders are linear and L has at most p
vertices, such an injection exists.

Jamison [2011] showed that lexical colorings can be recognized in time that
is linear in the number of edges. The same is true for monochromatic and rainbow
patterns, so Ramsey families of color patterns can be recognized in linear time.
Exercises for Section 10.3 473

When F consists of a monochromatic G and a rainbow K 1 ,t , the pattern Ram-

sey number is the smallest n such that every edge-coloring of K n with fewer than t
colors at each vertex contains a monochromatic G. Introduced in Gyárf ás–Lehel–
Nešet řil–Rödl–Schelp–Tuza [1987], this is the local Ramsey number of G.
When F consists of a monochromatic and a rainbow copy of one tree T , Chen–
Schelp–Wei [2001] showed that the pattern Ramsey number is at most quartic in
the order of T , and Jamison–Jiang–Ling [2003] obtained a cubic upper bound. It
is conjectured that the truth is quadratic.

EXERCISES 10.3
10.3.1. (−) Confirm that under the (¾ + 1)-coloring in Example 10.3.2, there is no
monochromatic solution to x + y = .
10.3.2. (−) For edge-colorings of graphs (see Definition 10.3.22), prove that the only pat-
terns that are monochromatic and lexical are monochromatic stars, and prove that the only
patterns that are rainbow and lexical are rainbow forests. (Jamison–Jiang–Ling [2003])

10.3.3. (♦) Hilbert ’s Cube Lemma (Hilbert [1892]). Given a ∈ 0 and d1 , . . . , d m ∈ , 

the affine m-cube H(a , d1 , . . . , d m) is the set {a + ∑i∈ I d i : I ⊆ [m]}. For m ,  ∈ , prove
that there is a least integer hm ,¾ such that every -coloring of [hm ,¾ ] contains a monochro-
matic affine m-cube. (Hint: Prove hm ,¾ ≤  n + n, where n = hm−1 ,¾ . Comment: Brown–
Erdős–Chung–Graham [1985] proved h2 ,¾ = (1 + o(1)) 2 and showed  c1 m ≤ hm ,¾ ≤  c2 for
m

constants c1 and c2 . Gunderson–Rödl [1998] gave further refinements.)

10.3.4. (♦) Show that for large n, every -coloring of the nonempty subsets of [n] gives the
same color to some two disjoint sets and their union.
10.3.5. (♦) Consider the equation ∑i=1 xi = x0 , for m ≥ 1.
m

(a) Prove that m2 + m − 1 is the least n such that every 2-coloring of [n] contains a
monochromatic solution (numbers may repeated). (Beutelspacher–Brestovansky [1982])
(b) For each n, determine the maximum size of a subset of [n] containing no solution.

(Comment: Given a1 , . . . , am ∈ , Guo–Sun [2008] found the least n such that all
2-colorings of [n] have monochromatic solutions of ∑i=1 ai xi = x0 . It is rs2 + s − r, where
m

m−1
r = min{a1 , . . . , am} and s = ∑i=1 ai . Schaal [1993] studied c + ∑i=1 xi = x m . No n suffices
m

when c is odd and m is even, but otherwise the value is known. When c ≥ −(m − 2), it is
m2 + (c − 1)(m + 1) if m is odd or c is even. For smaller c it is more complicated.)
10.3.6. (♦) Given a homogeneous linear equation with integer coefficients, let () be the
least n such that every -coloring of [n] produces a monochromatic solution to the equation.
For each equation below, (1) use Rado’s Theorem (Theorem 10.3.4) to determine whether
() exists for all  , and (2) determine (2).
(a) x + y = 2 . (b) x + y = 3 . (c) w + x + y = .

10.3.7. Given m, prove that if p is prime and sufficiently large, then x m + y m = m has a
solution in integers modulo p. (Dickson [1909], Schur [1916])
10.3.8. Prove w(3 , 2) = 9. That is, every 2-coloring of [9] has a 3-term monochromatic
arithmetic progression, but some 2-coloring of [8] does not. Generalize the latter to prove
w(l , 2) > 2(l − 1)2 . (Hint: On [9], consider possible colorings of {4 , 5 , 6} or {1 , 5 , 9}.)

10.3.9. (♦) Let ⟨a⟩ and ⟨b⟩ be sequences of distinct positive integers, and fix  ∈ . Prove
that if ⟨a⟩ contains arbitrarily long arithmetic progressions, and |bn − an | ≤  for all n,
then ⟨b⟩ contains arbitrarily long arithmetic progressions. (Rubel [1988])
474 Chapter 10: Ramsey Theory

10.3.10. Prove that every infinite sequence has an infinite nondecreasing subsequence or
an infinite decreasing subsequence.
10.3.11. Prove that every finite colored set of at least (¾ − 1)2 + 1 integers has a canonically
colored ¾-set. (Note that the colored objects are 1-sets.)
10.3.12. Finish the proof of the Canonical Ramsey Theorem for r = 2 (Theorem 10.3.21).
10.3.13. Prove that for every 2-coloring of the points of m , there is a set of n integer
lattice points with the same color whose centroid (mean vector) is an integer lattice point
also having that color. (Bòna [1991])
10.3.14. (♦) Prove that every coloring of E(K n) has a rainbow triangle or a monochromatic
spanning tree. (Thus a graph or its complement is connected.) (Galvin [1975])
10.3.15. (♦) A Gallai coloring is an edge-coloring of a complete graph with no rainbow
triangle. (Gyárfás–Simonyi [2004])
(a) Prove that every Gallai coloring with at least three colors has a color class con-
taining no spanning tree. (Hint: Show that the edges joining two components formed by
a color class have the same color.)
(b) Prove that every Gallai coloring arises from a 2-colored complete graph H by ex-
panding each vertex into a Gallai-colored complete graph. That is, each vertex vi expands
into a vertex set Vi so that each Vi induces a Gallai coloring, and each edge xy with x ∈ Vi
and y ∈ Vj inherits the color originally on vi vj in H .
(c) Prove that the pattern Ramsey number of the family consisting of a monochromatic
K1 ,m and a rainbow K3 is the least n such that ⌈ 2n/5⌉ ≥ m.
10.3.16. (♦) Product Ramsey Theorem. For r = (r1 , . . . , rd) ∈ d , an r-grid is a set
S1 × · · · × Sd , where Si is a set of ri natural numbers. Given r, ¾ , and vector thresholds
p1 , . . . , p¾ ∈ d , prove that there exists a vector (n1 , . . . , nd) ∈ d such that any ¾-coloring
of the r-grids of [n1 ] × · · · × [nd ] yields for some i a pi -grid whose r-grids all have color i.
10.3.17. Product van der Waerden Theorem. An arithmetic m-grid of size l is a set

of points in m of the form {(a1 + i1 d1 , . . . , am + im d m): 0 ≤ ij < l}, where d1 , . . . , d m are
fixed positive integers.
(a) Using van der Waerden’s Theorem (not the Hales–Jewett Theorem), prove for
r, l , m ∈ that there is a number W such that every r-coloring of [W ]m contains a
monochromatic arithmetic m-grid of size l.
(b) For r, n ∈ 
with n > 2, conclude that every r-coloring of 2 contains n points
with the same color spanning a convex polytope with nonzero 2-dimensional volume whose

centroid (mean vector) also has that color. (Comment: This generalizes to m with m ≥ 2.)
10.3.18. Determine the least n such that for every ¾-coloring of [n] there is a pair S of
elements for which the color of the sum of every nonempty subset of S is determined by its
maximum element (see Lemma 10.3.10).
10.3.19. (♦) Prove Folkman’s Theorem from Rado’s Theorem. In particular, show that
the system yT = ∑i∈ T xi for nonempty sets T ⊆ [t] satisfies the Columns Condition.
10.3.20. Consider a positional game with positions x1 , . . . , x n and winning sets {S1 , . . . , Sm}.
(Tic-Tac-Toe has nine positions and eight winning sets.)
(a) Prove that Player 1 can win Tic-Tac-Toe on the m-dimensional cube Clm of side-
length l if l is fixed and m is sufficiently large. (Hales–Jewett [1963])
(b) Suppose that each winning set has size at least a and each position appears in at
most b winning sets. Prove that Player 2 can force a draw if a ≥ 2b.
(c) In Tic-Tac-Toe on Clm , prove that Player 2 can force a draw when l ≥ 3m − 1.
Chapter 11

Extremal Problems
In this chapter, we study extremal problems for families of sets and subsets.
We seek the largest or smallest structure with certain properties, or extreme val-
ues of a parameter over a class of structures.

11.1. Forced Subgraphs

In Ramsey ’s Theorem, we ¾-color ([n]
t ) for some large n and find that some
q-set is homogeneous. That is, we find a monochromatic copy of the complete t-
(t)
uniform hypergraph with q vertices, written K q . Given n and q with n ≥ q,
(t)
we can also force a copy of K q by having enough edges in a t-uniform hypergraph
(t)
on n vertices; the maximum number of edges avoiding K q is Tur án’s Problem.
For t ≥ 3, the problem is notoriously difficult; Füredi [1991] and Keevash
[2011] give extensive surveys. We do not even know the asymptotic answer for
(3)
K 4 (see Exercise 30). The “flag algebra” method pioneered by Razborov [2007,
2010] has improved many of the bounds; Falgas-Ravry–Vaughan [2013] summa-
rizes the results. Hence we focus on t = 2, where the central question started the
field of extremal graph theory.

TUR ÁN’S THEOREM

The traditional phrasing asks for the maximum number of edges in an n-

vertex graph having no (r + 1)-clique. This was solved for r = 2 by Mantel [1907]
and for general r by Turán [1941]. We already proved the theorem as Theorem
5.2.11; here we take a more quantitative look and give another short proof.

11.1.1. Definition. The Tur án graph T n ,r is the complete r-partite graph with
n vertices having b parts of size a + 1 and r − b parts of size a, where a = ⌊ n/r⌋
and b = n − ra. Let tr (n) = | E(T n ,r)|.
Turán proved that T n ,r is the unique largest n-vertex graph not containing
K r+1 . The proof by Erdős in Theorem 5.2.11 uses induction on r and the Degree-

475
476 Chapter 11: Extremal Problems

Sum Formula, showing that for every graph G not containing K r+1 , there is an
r-partite graph H with d H (v) ≥ d G(v) for all v ∈ V(G). It also uses that T n ,r is
the unique largest r-partite n-vertex graph (if part-sizes differ by more than 1,
shifting a vertex from a larger part to a smallest part gains edges). Here we note
the asymptotic value of the number of edges.

11.1.2. Proposition. tr (n) = (1 − 1r ) n2

2 − O(n).
Proof: With part-sizes x1 , . . . , x r , there are ∑i< j x i x j edges. With ∑ x i = n fixed
and all x i real, the sum is maximized when all x i equal n/r (averaging any two
increases the value). The maximum is thus (2r ) nr 2 , which proves tr (n) ≤ (1 − 1r ) n2 .
2 2

For the asymptotic optimality of the bound, note that each part-size exceeds
n/r − 1. Counting the edges by pairs of parts thus yields
r(r − 1) n − r 2 1 1 1 n2
tr (n) > ( ) = (1 − ) ( n2 − 2nr + r 2 ) = (1 − ) − O(n).
2 r 2 r r 2

Turán’s Theorem has many proofs. Turán’s original proof used induction on
n (Exercise 7), different from the Erdős proof. The proof in Exercise 8 uses a con-
tinuous optimization problem. Theorem 14.1.15 gives a probabilistic proof. Six
proofs appear in Aigner [1995], and the five proofs in Aigner–Ziegler [1999] in-
clude the proof we present here.

11.1.3. Theorem. (Tur án’s Theorem; Turán [1941]) Among n-vertex graphs
with no (r + 1)-clique, the unique largest graph is T n ,r .
Proof: (Zykov [1949]) Since T n ,r is the unique largest r-partite graph, it suffices
to show that any largest graph G not containing K r+1 is a complete multipartite
graph, which means it does not have K 2 + K 1 as an induced subgraph. If G has
such a subgraph, with vertex set {u , v , w} and vw ∈ E(G), then we find a larger
graph not containing K r+1 .
If d(u) < d(v), then we replace u with a new vertex v having the same neigh-
bors as v, as shown below. Since v and v are not adjacent, we did not create an
(r + 1)-clique. Since d(u) < d(v) = d(v ), the new graph has more edges.
By symmetry in v and w, we may assume d(u) ≥ max{d(v) , d(w)}. Now we
replace both v and w by two new copies of u, as shown below. Again we have create
no larger clique. We lose only d(v) + d(w) − 1 edges, since vw is counted twice, and
we gain 2d(u) edges, so again the new graph is larger.

w• w• •
v•• v •u v•
•u •u •u
• • • • • • • • •

In fact, for 2 ≤ ¾ ≤ r, the Turán graph T n ,r is the unique n-vertex graph

without K r+1 that has the most ¾-cliques (Zykov [1949]). This generalization was
rediscovered by Erdős [1962b] and by Sauer [1971].
An n-vertex graph with more than n2 rr−
2
2
−1 edges contains K r . More generally,
how many r-cliques are forced by m edges on n vertices?
Section 11.1: Forced Subgraphs 477

11.1.4. Definition. Let ¾r(G) denote the number of r-cliques in G.

Let G be an n-vertex graph with m edges. Moon–Moser [1962a] gave a lower

bound for ¾r (G) in terms of n and m. Goodman [1959] minimized ¾3(G) + ¾3(G)
in terms of n. Lovász [1972d] obtained both results via inclusion-exclusion. One
would expect a random n-vertex graph and its complement together to have ap-
proximately 2 · 18 (n3) triangles; surprisingly, the minimum asymptotically equals
the average. The proof uses a counting formula that has other applications.

11.1.5. Lemma. If a graph G has n vertices and m edges, then ¾3(G) + ¾3(G)
equals each formula below, where d(v) is the degree of v in G.
(a) (n3) − 12 ∑v∈V(G) d(v)[n − 1 − d(v)] (Sauvé [1961])
(b) (n3) − (n − 2)m + ∑v∈V(G) (d(v)
2
). (Goodman [1959])

Proof: (Sauvé) View G and G as a 2-coloring of E(K n). Edges of G have one color
and edges of G the other. We count monochromatic triangles.
Assign each pair of incident edges in K n weight 2 if they have the same color,
weight −1 if not. A vertex triple inducing a monochromatic triangle contributes
6 to the total weight; other triples contribute 0. Hence the sum of the weights of
all pairs of incident edges is 6 times the number of monochromatic triangles.
Over the pairs of edges incident to v, the weights sum to
n−1− d(v)
2 ) + 2(
2(d(v) 2 ) − d(v)[n − 1 − d(v)].

Using (¾2) + (m2−¾ ) = (m

2) − ¾
(m − ¾), the contribution from v becomes 2(n−2 1 ) −
3d(v)[n − 1 − d(v)]. Summing over v and dividing by 6 yields (a).
To obtain (b) from (a), replace n − 1 with n − 2 + 1, apply the Degree-sum
Formula, and collect terms.

11.1.6. Theorem. (Goodman [1959]) An n-vertex graph and its complement to-
gether have at least n(n − 1)(n − 5)/24 triangles, sharp when n ≡ 1 (mod 4).
Proof: (Sauvé [1961]) To minimize formula (a) of Lemma 11.1.5, we maximize
the subtracted terms. This is achieved by setting each d(v) to (n − 1)/2. The for-
mula becomes (3n) − 2n (n−41) , which simplifies to the claimed lower bound.
2

By Lemma 11.1.5, ¾3(G) + ¾3(G) depends only on the vertex degrees, not the
choice of edges. Equality in the bound holds if and only if every vertex has degree
(n − 1)/2. This can happen only for n odd and (n − 1)/2 even, so n = 4¾ + 1 and G is
2 ¾-regular. When n ≡ 1 (mod 4), this is achieved by a regular self-complementary
graph, such as the graph obtained by adding one vertex adjacent to the low-degree
vertices in the near-regular self-complementary graph P4 [K ¾ , K ¾ , K ¾ , K ¾ ]. For
other congruence classes, the bound can be improved slightly (Exercise 27).

We return to the counting of cliques in the graph G itself. For r = 3, Lemma

11.1.5 provides a lower bound.

11.1.7. Corollary. (Moon–Moser [1962a]) A graph G with n vertices and m edges

m
has at least 3n (4m − n2) triangles.
478 Chapter 11: Extremal Problems

Proof: Write the formula of Lemma 11.1.5b in terms of G, letting m = ##### E(G)#####

and d (v) = d G(v). Note that 3¾3(G) ≤ ∑ (d 2(v)), since incident edges lie in at most
one triangle. Subtracting this upper bound on ¾3(G) from Lemma 11.1.5b yields

¾3(G) ≥ (n3) − (n − 2)m + 23 ∑ (d 2(v)).

Replace ∑ (d 2(v)) with the lower bound n(2m2/n), replace m with (n2) − m, and sim-
plify (Exercise 1) to obtain ¾3(G) ≥ 3n
m
(4m − n2).

11.1.8.* Remark. Lower bounds. For q + 1 ≥ r > p, the Turán graph T n ,q has the
fewest r-cliques among n-vertex graphs G with ¾p(G) = ¾p(T n ,q). Bollobás [1976b]
proved that linear interpolation yields lower bounds on ¾r(G) for intermediate
values of ¾p(G) (see Bollobás [1978, pp. 297–301]). When p = 2 and r = 3 and m =
| E(G)| , interpolation improves Corollary 11.1.7 for m ∈ [ n4 , n3 ]. Since K n/2 ,n/2 has
2 2

n2/4 edges and no triangles, while K n/3 ,n/3 ,n/3 has n2/3 edges and n3/27 triangles,
interpolation yields at least 9n (4m − n2) triangles. This improves Corollary 11.1.7,
since 9n > 3n m
when m < n2/3.
Just above tr(n) (with n2/4 + 1 edges), Turán’s Theorem guarantees one trian-
gle, the Moon–Moser bound guarantees n/3, and interpolation guarantees 4n/9.
In fact, G has at least ⌊ n/2⌋ triangles (Rademacher; see Erdős [1955], in Hebrew).
This is sharp, by adding one edge to T n ,2 .
Adding q edges inside one part of T n ,2 (without forming triangles) creates
only q⌊ n/2⌋ triangles. Erdős [1962b] proved that this minimizes ¾3(G) when q <
cn, for some constant c. When q = n/2, adding a (¾ + 1)-cycle to K ¾+1 ,¾−1 produces
only (n/2)(n/2)− 1 triangles. For large n, Lovász–Simonovits [1976] proved Erdős’
conjecture that c = 1/2 is best, plus similar results for larger complete graphs.
Mubayi [2010] greatly generalized the results. For an (r + 1)-chromatic graph
F having an edge e such that (F − e) = r, there is a constant c F such that for
1 ≤ q ≤ c F n, every n-vertex graph with tr (n) + q edges has at least qs copies of
F , where s is the minimum number of copies of F formed by adding one edge to
T n ,r . This is sharp for odd cycles and asymptotically sharp in general. The tool is
the Graph Removal Lemma, generalizing the Triangle Removal Lemma (Lemma
11.1.22) obtained from the Szemerédi Regularity Lemma (Theorem 11.1.13).
The problem was also studied for larger numbers of edges: we want to min-
imize r(G) when G is an n-vertex graph with at least  n2 edges, where  >
1
2
(1 − r−1 1 ). Lovász–Simonovits [1983] conjectured that the minimizing graph is
a complete r-partite graph where all the parts (except one smaller) have the same
size. This was proved for r = 3 by Razborov [2008], for r = 4 by Nikiforov [2011],
and in general by Reiher [2016].

ERD ŐS–STONE THEOREM

Just as the Ramsey problem extends by seeking a monochromatic copy of any

target graph, so the Turán problem extends by asking how many edges force a
given graph. When K q is forced, every q-vertex graph is forced, but fewer edges
may force sparser subgraphs.
Section 11.1: Forced Subgraphs 479

11.1.9. Definition. The Tur án number of F , written ex (n; F), is the maximum
number of edges in an n-vertex graph not containing F . Let G[s] denote the
graph obtained from G by replacing each vertex v with an independent set
Sv of size s, with vertices in Su and Sv adjacent if and only if uv ∈ E(G).

Since any edge beyond a maximal triangle-free subgraph must lie in a trian-
gle, n-vertex graphs with t2(n)+ cn2 edges have many triangles when c is a positive
constant. Similarly, tr(n) + cn2 edges force many (r + 1)-cliques. Erdős and Stone
proved that when n is large enough, tr(n) + cn2 edges force copies of K r+1 whose
union is a complete (r + 1)-partite graph with large vertex classes.

11.1.10. Theorem. (Erdős–Stone Theorem; Erdős–Stone [1946]) Fix s ∈ 

and a positive constant c. If n is sufficiently large, then every n-vertex graph
with tr (n) + cn2 edges contains K r+1 [s].

This theorem yields the asymptotics of ex (n; F) when Ò(F) > 2.

11.1.11. Theorem. (Erdős–Simonovits [1966]) If F is an (r + 1)-chromatic graph,

then limn→∞ ex (n; F)n−2 = 12 (1 − 1r ).
Proof: Since T n ,r is r-partite, F
⊆ T n ,r when Ò(F) > r. Thus ex (n; F) ≥ tr(n) =
1 (1
2 − 1r )n2 − O(n). For the upper bound, let s be the maximum size of a color class
in some proper (r + 1)-coloring of F . In an n-vertex graph with 12 (1 − 1r + c)n2
edges, where n is sufficiently large in terms of c, the Erdős–Stone Theorem guar-
antees the appearance of K r+1 [s] and thus also F . Thus the ratio ex (n; F)/n2 can
be brought down as close to 12 (1 − 1r ) as desired by making n sufficiently large.

Few exact results are known. For odd cycles, in fact ex (n; C2¾+1) = ⌊ n2/4⌋
when n ≥ 4¾ − 2 (see Füredi–Gunderson [2015], including all n).
For bipartite graphs, Theorem 11.1.11 gives only ex (n; F) ∈ o(n2). Counting
arguments yield better bounds; in fact, ex (n; C4) ∈ O(n3/2). The constructions
involve designs and projective planes, so we postpone discussion of ex (n; C4) to
Chapter 13. Simonovits [1968] pioneered a method for studying Turán numbers
(also called extremal numbers). These have been studied for various bipartite
graphs and families of graphs, surveyed in Füredi–Simonovits [2013].
Of most interest is ex (n; C2¾). Bondy–Simonovits [1974] proved ex (n; C2¾) <
90¾ n1+1/¾ (see Exercise 14.2.11 for a lower bound). The constant was improved by
Verstraëte [2000] and √then Pikhurko [2012]. More recently, Bukh–Jiang [2017]
proved ex (n; C2¾) ≤ 80 ¾ log ¾ n1+1/¾ + 10¾ 2 n when ¾ ≥ 2 and n ≥ (2 ¾)8¾ .
2

For the Erdős–Stone Theorem, Exercise 31 requests a direct proof. Here we

use an enormously important tool for proving asymptotic results in combinatorics.
It is the Szemerédi Regularity Lemma, developed originally to prove that sets of
integers with positive density contain arbitrarily long arithmetic progressions.
Roughly speaking, the Regularity Lemma states that every sufficiently large
graph contains within it a multipartite subgraph with large parts such that al-
most all the bipartite subgraphs induced by pairs of parts look fairly random.
This is useful when seeking a particular subgraph, because in randomly gen-
erated graphs with a fixed density of edges (see Chapter 14) the number of oc-
curences of a particular subgraph is almost always close to its expected value.
480 Chapter 11: Extremal Problems

The induced subgraphs of fixed size in a “fairly random” bipartite graph have
about the same edge density, if they are not too small. Hence we restrict the edge
density only for subgraphs with at least a fraction of the vertices from each
part. Making smaller is more restrictive, yielding more “regular ” behavior.

11.1.12. Definition. Given disjoint vertex sets A and B in a graph G, the

subgraph generated by (A , B) is the A , B-bigraph with edge set [A , B].
Its density, denoted (A , B), is | | AA||,B] |
[
B| . The pair (A , B) is -regular if
| (A , B ) − (A , B)| <  whenever A ⊆ A and B ⊆ B with | A | ≥  | A| and

| B | ≥  | B|. An equipartition of G is a partition of V(G) into parts whose
sizes differ by at most 1.

When we partition V(G) into singleton sets, every pair is -regular, but ap-
plications require larger parts. We need equipartitions with not too many parts
such that almost all the pairs of parts are -regular. The Regularity Lemma guar-
antees such a suitable partition in a precise way.

11.1.13. Theorem. (Regularity Lemma; Szemerédi [1978]). Given  , l > 0,

there exist constants M, N ∈  such that every graph with at least N ver-
tices has an equipartition with  parts for some  ∈ [l , M] such that fewer
than  2 pairs of parts fail to be -regular.

Though Szemerédi [1978] is the seminal paper, already the idea was used in
Szemerédi [1975]. Szemerédi [2015] describes a number of variants of the lemma.
We will prove the one stated below.

11.1.14. Definition. For an n-vertex graph G, an  , -partition is a partition

V0 , . . . , V¾ of V(G) such that | V0 | ≤  n (V0 may be empty) and | V1 | = · · · =
|V¾ | ; call V0 the exceptional part. An  , -partition is an -regular partition
if fewer than  2 pairs of nonexceptional parts fail to be -regular.

11.1.15. Theorem. (alternate Regularity Lemma). For  , l > 0, there are con-
stants M, N ∈  such that every graph with at least N vertices has an -
regular  , -partition for some  with l ≤  ≤ M.

Before proving the Regularity Lemma, we use it to prove the Erdős–Stone

Theorem and striking results in graph Ramsey theory and additive combina-
torics. Applications of the Regularity Lemma avoid technical detail by capturing
the standard technical argument in an “Embedding Lemma”. We form a reduced
graph R from the partition of G provided by the Regularity Lemma. The ver-
tices of R are the (nonexceptional) parts of the equipartition, adjacent when they
form an -regular pair with density more than some parameter d. With d >  ,
the Embedding Lemma implies that any specified subgraph H of R[s] that has
small enough maximum degree occurs as a subgraph of G, for suitable s.
For the Erdős–Stone Theorem, the Regularity Lemma provides a reduced
graph with enough edges so that Turán’s Theorem forces K r+1 in R, and then the
Embedding Lemma yields K r+1 [s] ⊆ G. That is, we apply the Embedding Lemma
with R = K r+1 and H = R[s].
Section 11.1: Forced Subgraphs 481

11.1.16. Theorem. (Embedding Lemma) Let R be a graph, and fix m , s ∈ . 

Given d , with d > > 0, let G be a subgraph of R[m] in which each pair of
parts corresponding to an edge of R is an -regular pair with density at least
d. Let H be a subgraph of R[s] with n vertices and maximum degree D. If
≤ and s − 1 ≤ m, where = (d − )D/(D + 2), then H ⊆ G. In fact, G
contains more than ( m)n copies of H (as labeled subgraphs).
Proof: For each vertex v ∈ V(R), let Rv be the corresponding s-set in R[s], and
let Av be the corresponding m-set in G. Let V(H) = x1 , . . . , x n . In V(G) we will
find representatives y1 , . . . , yn of x1 , . . . , x n such that x i x j ∈ E(H) implies yi yj ∈
E(G). Furthermore, if x t ∈ Rv , then yt ∈ Av .
We pick y1 , . . . , yn in order. At time t, we will choose yt and update (for j > t)
the set Bj of candidates for yj . Initially Bj = Av , where x j ∈ Rv . The set Bj
shrinks when we choose a vertex yt that y j needs as a neighbor. For each j with
j ≥ t, let Yj = {yi : i < t and x i x j ∈ E(H)}; these already-chosen vertices are
required to be neighbors of the vertex we choose as yj . Since we want yj ∈ Av ,
the set Bj of candidates for yj is {y ∈ Av : Yj ⊆ NG(y)}. Letting = d −  , we will
make our choices to guarantee that when we are ready to pick yt ,

| Bj | ≥ | Yj | m for j ≥ t. (∗)
When we are choosing yt , we do not know which vertices will be chosen to
represent the neighbors of x j in {x t+1 , . . . , x j −1 }. The vertex eventually chosen
as yj must be adjacent to all of them. For this reason, we initially preserve many
candidates for y j . As more neighbors of x j are chosen, the number of candidates
we keep decreases. When | Yj | = D, one candidate suffices.
If x j ∈
/ NH (x t), then Yj and Bj do not change when yt is chosen. To maintain
(∗), it suffices to choose yt from Bt so that yt has at least | Bj | neighbors in Bj for
each j with j > t and x j ∈ NH (x t).
For such j , define u and v by x t ∈ Ru and x j ∈ Rv . The pair (Au , Av) is -
regular in G, by definition. We have Bt ⊆ Au and Bj ⊆ Av . Since the hypotheses
guarantee D ≥  , by (∗) we have | Bj | ≥  | Av |.
Let B be the set of vertices in Bt that do not have at least | Bj | neighbors
in Bj . Thus (B , Bj ) < . If | B| ≥  m, then -regularity of (Au , Av) guarantees
(B , Bj ) ≥ d −  = , a contradiction.
Hence all but  m vertices of Bt have at least | Bj | neighbors in Bj . Since
there are at most D values of j to consider, we discard at most D m vertices of Bt
in this way. We also discard vertices of Bt already chosen to represent vertices of
H in Rv ; there are at most s − 1 of these. There remain at least ( D − D)m − (s − 1)
vertices in Bt ; choose one to be yt .
Since at each t there are at least ( D − D)m − (s − 1) choices for yt , there are
at least [( D − D)m − (s − 1)]n labeled copies of H in G. From  = D/(D + 2), we
have D = 2  + D ≥ 2  + D. Thus 2  ≤ D − D. Since also s − 1 ≤  m, we
have ( D − D)m − (s − 1) ≥  m.

The Embedding Lemma is used to find small graphs in a large graph G.

(Komlós–Sárközy–Szemerédi [1997] proved a more difficult extension called the
Blow-up Lemma to find spanning subgraphs.) To complete the proof of the
Erdős–Stone Theorem, it now suffices to find a copy of K r+1 in the reduced graph
that comes from the partition of G provided by the Regularity Lemma.
482 Chapter 11: Extremal Problems

Proof of Erdős–Stone Theorem (Theorem 11.1.10) : Fix r, s ∈ and a positive 

constant c. We want to prove that if n is sufficiently large, then an n-vertex graph
with more than tr (n) + cn2 edges contains K r+1 [s]. We may assume c < 12 .
Choose l with l > max{r, 2c 1
}. Choose d with 0 < d2 < c − 2l 1
< 12 . Choose
(d− )rs
small enough so that ≤ rs+2 and < 3 (c − 2 − 2l ). Let M and N be the
1 d 1

constants in terms of and l needed to apply the Regularity Lemma. Choose n so

that n > max{N , M(s − 1)/ }, where = (drs−+)2 .
rs

Let G be an n-vertex graph with more than tr(n) + cn2 edges. By the alterna-
tive form of the Regularity Lemma, G has an -regular , -partition V0 , . . . , V ,
where l ≤ ≤ M. Let m = | V1 | = · · · = | V | .
The reduced graph R with vertex set v1 , . . . , vn has vi vj ∈ E(R) if and only
if the pair (Vi , Vj) is -regular with density at least d. Showing that | E(R)| >
tr( ) forces K r+1 ⊆ R. The Embedding Lemma then implies K r+1 [s] ⊆ G. Since
(K r+1 [s]) = rs, applying it requires  ≤  and s − 1 ≤  m, which were guaran-
teed by the choice of  and then n.
The “bad” edges of G are those incident to V0 , within Vi for i ≥ 1, or joining
Vi and Vj when (Vi , Vj ) is not -regular or is -regular with density less than d.
To show that R has enough edges, we show that the number of bad edges in G
is small enough that more than tr ( ) pairs with density at least d are needed to
accommodate the good edges. Note that n −  n ≤ m yields n ≤ (1 + ) m, and
hence n2 < 2 2 m2 , since  < 1/4. Of course, n2 > 2 m2 .
Since | V0 | ≤  n, the number of edges incident to V0 is less than  n2 , which is
less than 2  2 m2 . Since at most  2 pairs are not -regular, at most  2 m2 edges
of G lie in such pairs. Fewer than 12 2 pairs have density less than d, and hence at
most 12 2 dm2 edges lie in such pairs. Each G[Vi] has at most 12 m2 edges, totaling
at most 12 m2 edges. Finally, each edge of R arises from a pair with density at
least d; even when the density is 1, it contributes at most m2 edges to G. Thus
| E(G)| ≤ 3 2 m2 + 12 d 2 m2 + 12 m2 + | E(R)| m2 .
Solving for | E(R)| , further substitution yields the desired bound.
| E(G)| − 3 2 m2 − 12 d 2 m2 − 12 m2
| E(R)| ≥
m2
tr(n) + cn
2
d 1
≥ 2
( 2 m2
− 3 − − )
2 2
tr(n) d 1
≥ 2
( + c − 3 − − )
n2 2 2l
> 1
2
2(1
− 1/r) ≥ tr( ).
Hence K r+1 ⊆ R, and the Embedding Lemma gives K r+1 [s] ⊆ G.

11.1.17.* Remark. As we have noted, the Erdős–Stone Theorem has proofs avoid-
ing these tools. The theorem also can be made more precise. We fixed c and s and
showed that K r+1 [s] is forced by tr (n) + cn2 edges when n is large enough. Instead,
we can fix c and n and ask how large s can be guaranteed as a function of n when
forcing K r+1 [s]. If it grows with n, then the Erdős–Stone Theorem follows. A
lower bound on the growth of s is a quantitative strengthening.
Section 11.1: Forced Subgraphs 483

Fixing r and c, one can guarantee s that is logarithmic in n (Bollobás–Erdős

[1973]). Several improvements culminated in the result of Ishigami [2002] (via
Regularity) that every n-vertex graph with more than (1 − 1/r + c)n2/2 edges con-
tains K r+1 [s] with s ≥ ⌊ 11r
log 1/c n
⌋ . Superlogarithmic s is not forced (Bollobás–Erdős–
Simonovits [1976]). Bollobás–Kohayakawa [1994] guaranteed that one part has
close to linear size. Nikiforov [2008] proved a more general theorem yielding both
logarithmic growth of s and near-linear size of one part.

LINEAR RAMSEY FOR BOUNDED DEGREE

Our second application of the Regularity Lemma is from graph Ramsey the-
ory. The Ramsey number R(G , G) may grow exponentially in the number of ver-
tices of G, such as when G = K n . In some families of relatively sparse graphs, the
Ramsey number grows at most linearly in the number of vertices! Most famously,
one such family is the family of graphs with bounded maximum degree.

11.1.18. Theorem. (Chvátal–Rödl–Szemerédi–Trotter [1983]) For d ∈ , there

is a constant cd such that R(G , G) ≤ cd | V(G)| whenever (G) = d.
Proof: Let l = 3 max{3d , R(d + 1 , d + 1)}, and let = l−1 . The Regularity Lemma

provides constants M, N ∈ such that every graph with at least N vertices has
an -regular , -partition for some with l ≤ ≤ M. Let c = max{N , M/ },
where  = (1/2 − )d/(d + 2). Now fix a graph G with maximum degree d and
vertices x1 , . . . , x n ; we prove that R(G , G) ≤ cn.
Consider a red/blue coloring of E(K cn). Let H and H be the subgraphs in red
and blue, respectively. Via Steps 1 and 2 below, we find d + 1 large sets such that
every pair of them is -regular and has high density in the same color. In Step 3,
the Embedding Lemma allows us to find a copy of G among the edges of that color.
Step 1: Some -regular partition of V(H) has at least l/3 parts with all pairs -
regular in H and H. For a pair (A , B) of vertex subsets, the densities in H and H
sum to 1. Hence (A , B) is -regular in H if and only if it is -regular in H , so we
consider only H. By the choice of c, there is an -regular  , -partition A0 , . . . , A¾
of H , where l ≤ ≤ M. All but  2 pairs are -regular. Let H ∗ be the graph with
vertex set [ ] putting i and j adjacent if and only if (Ai , A j ) is -regular in H.
By Turán’s Theorem, a -vertex graph with no (t + 1)-clique has fewer than
(1 − 1t ) ¾2 edges. Since | E(H ∗)| ≥ (¾2) −  2 , we find a clique of size at least 31 , equal
2

to 3l . We thus have 3l parts in { A1 , . . . , A } among which every pair is -regular.

Step 2: This -regular partition has d + 1 parts with all pairs -regular of den-
sity at least 12 in the same color (H or H). We color the edges of the 3l -clique in H ∗
found in Step 1, using red if  H (Ai , A j) ≥ 12 and blue if  H (Ai , A j) < 12 . Since
3 ≥ R(d + 1 , d + 1), Ramsey ’s Theorem yields a monochromatic copy of K d+1 in
l

H ∗ ; by symmetry, we may assume it is in red. In the original coloring of E(K cn),

each pair among these parts is -regular with density at least 12 in the red graph.
Step 3: The given 2-coloring of E(K cn) has a monochromatic copy of G. We
now have d + 1 vertex sets in K cn , all of size at least n(1 − )/ , such that each
pair is -regular with density at least 1/2 in the red graph, H. Let G be the
subgraph of H induced by these parts. In the language of the Embedding Lemma
484 Chapter 11: Extremal Problems

(Theorem 11.1.16), the reduced graph R of G is K d+1 . Since (G) ≤ (G) + 1 ≤

d + 1, the graph G is a subgraph of R[n]. If the Embedding Lemma applies, then
G contains G, as desired. It is important that the requirements to apply the
Embedding Lemma are only in terms of (G), not the number of vertices. Since
the size of the parts is at most cn/ , we need  ≤  and n − 1 ≤  cn/ , where
 = ( 12 − )d/(d + 2). Since we define  to be less than (1/3)d and defined c so that
 c/ ≥  c/M ≥ 1, the hypotheses hold and the proof is complete.
11.1.19.* Remark. Unfortunately, the threshold for N in the Regularity Lemma
applies is huge. As a result, the constant cd from the proof of Theorem 11.1.18
grows like an exponential tower with height d. Using a variant of the Regularity
O(d)
Lemma, Eaton [1998] reduced the constant to doubly-exponential growth (22 ).
Graham–Rödl–Ruciński [2000] avoided the Regularity Lemma and reduced
cd to 2O(d(log d) ) (Conlon–Fox–Sudakov [2012] then reduced it to 2O(d log d)). For bi-
2

partite graphs, Graham–Rödl–Ruciński [2001] further improved the bound to

8(8d)d and showed that cd is at least exponential in d. Like Ramsey ’s Theorem,
the Regularity Lemma is a powerful tool to prove the existence of bounds, but
accurate bounds generally require more detailed direct arguments.
Theorem 11.1.18 was conjectured by Burr–Erdős [1975], who also conjectured
the same claim (with a larger constant) for the larger family of d-degenerate
graphs. Kostochka–Rödl [2001, 2004] proved bounds that are quadratic in the
number of vertices. Finally, C. Lee [2017] proved the conjecture, showing the ex-
istence of a constant c such that every d-degenerate graph H with (H) =  and
2 c¾ c¾
|V(H)| ≥ 2d 2 has Ramsey number at most 2d2 |V(H)|.

ROTH’S THEOREM

We give one more application of the Regularity Lemma. Szemerédi devel-

oped the Regularity Lemma to prove the long-standing conjecture of Erdős–Turán
[1936] that subsets of the integers with positive density contain long arithmetic
progressions. More precisely, subsets of [n] containing no -term arithmetic pro-
gression have size at most o(n) (strengthening van der Waerden’s Theorem).
Roth [1953] proved this for  = 3; Szemerédi [1970] proved it for  = 4. Sze-
mer édi’s Theorem (Szemerédi [1975]) used a bipartite version of the Regularity
Lemma to prove it for all  . A later proof by Gowers [2001] using Fourier analy-
sis avoids regularity and gives much better numerical bounds. We use regularity
to prove the case  = 3, following Palmer [2015]; see also Szemerédi [2015]. Roth
proved that subsets of [n] with size at most O(n/log log√ n) contain a 3-term pro-
gression; Behrend [1946] constructed a set of size n/e c log n that does not.
We begin with a standard computation, useful also in some of the exercises.

11.1.20. Lemma. (Degree-Density Lemma) In a graph G, let (A , B) be an -

regular pair with density d. Given a set Y ⊆ B with | Y | ≥  | B| , the number
of vertices in A having at most (d − ) | Y | neighbors in Y is less than  | A|.
Proof: Let X be the set of vertices in A having at most (d − ) | Y | neighbors in
Y . By direct counting, (X , Y) ≤ d − . If | X | ≥  | A| , then -regularity of (A , B)
requires (X , Y) > d − . Hence | X | <  | A|.
Section 11.1: Forced Subgraphs 485

11.1.21. Lemma. (Triangle-Counting Lemma) Let A , B , C be a partition of

the vertices of a graph G such that each pair is -regular. Let ,  ,  be the
densities of the pairs (A , B), (B , C), (C, A), respectively. If ,  ,  ≥ 2  , then
the number of triangles having a vertex in each class is at least
(1 − 2 )( − )( − )( − ) | A| | B| | C|.

Proof: By Lemma 11.1.20, fewer than  | A| vertices in A have at most ( − ) | B|

neighbors in B, and fewer than  | A| vertices in A have at most ( − ) | C| neigh-
bors in C. Deleting both such sets from A leaves at least (1 − 2 ) | A| vertices in A
having more than ( − ) | B| neighbors in B and ( − ) | C| neighbors in C.
For such a vertex x, let B = N(x) ∩ B and C = N(x) ∩ C. Since  ≥ 2  , by
-regularity we have (B , C ) ≥  −  , so [B , C ] ≥ ( − ) | B | | C |. With at least
(1 − 2 ) | A| choices for x and at least ( − )( − )( − ) | B| | C| desired triangles for
each such x, the claim follows.

The Triangle-Counting Lemma is a special case of the Embedding Lemma,

finding many copies of a fixed subgraph given a partition with all pairs -regular.
We next show that a graph with few triangles can be made triangle-free by re-
moving few edges. A more recent proof and extension by Fox [2011] gives better
bounds and avoids the Regularity Lemma (see also Conlon–Fox [2013]). The more
general Graph Removal Lemma states that when F is a -vertex graph, all copies
of F in any n-vertex graph H containing o(n¾) copies of F can be destroyed by
deleting o(n2) edges of H. Rödl–Schacht [2009] generalizes to hypergraphs.

11.1.22. Lemma. (Triangle-Removal Lemma; Ruzsa–Szemerédi [1978]) For

 > 0, there exists a positive number  such that for sufficiently large n, any
n-vertex graph G in which at least  n2 edges must be deleted to break all
triangles has at least  n3 triangles.
Proof: Since a complete graph has only (n2) edges, we may assume  < 1/2. Choose
 and l with  < /3 and l > 1/ . By the Regularity Lemma, there exist integers
M and N such that for n > N we are guaranteed an equipartition {V1 , . . . , V¾ }
of V(G) with l ≤  ≤ M such that all but   2 pairs of parts are  -regular.
With respect to this partition, we first remove a set of edges that hits all tri-
angles except a certain type. We remove all edges that (1) lie inside parts, (2) join
parts that do not form an  -regular pair, or (3) join parts forming an  -regular
pair with density less than 2  . The number of edges of type (1) is at most  (⌈n/2¾⌉),
which is less than  n2 since  > 1/ . There are fewer than   2(n/)2 edges of
type (2) and fewer than (¾2)2  ⌈ n/ ⌉ 2 of type (3), both less than  n2 . In total, we
have removed fewer than  n2 edges.
By the hypothesis on G, at least one triangle T remains. By the types of
edges removed, T has its vertices in three distinct parts such that each pair
among these three is  -regular with density at least 2  . Also, each part has size
at least ⌊ n/ ⌋ , which is at least n/M. By the Triangle Counting Lemma, among
these three parts the number of triangles remaining is at least (1 − 2  )( )3(n/M)3 .
Thus the desired conclusion holds with  = (1 − 2  )( /M)3 .

11.1.23. Theorem. (Roth’s Theorem; Roth [1953]) A subset of [n] containing

no 3-term arithmetic progression has size at most o(n).
486 Chapter 11: Extremal Problems

Proof: (Ruzsa–Szemerédi [1978]) The meaning of the statement is that for each
positive , for large enough n every set S ⊆ [n] with | S| = n contains a 3-term
arithmetic progression; this is what we must prove.
We construct a graph G with vertex set A ∪ B ∪ C, where A = {a1 , . . . , an},
B = {b1 , . . . , b2n}, and C = {c1 , . . . , c3n}. For s ∈ S and j ∈ n, add to E(G) the
triangle with vertices {aj , bj + s , cj +2s}. Let Ì be this set of triangles.
The n2 triangles in Ì are edge-disjoint; hence it takes at least n2 edges to
hit all the triangles in G. By the Triangle-Removal Lemma, G has at least n3
triangles (for some constant ). For large enough n, we have n3 >  n2 , and hence
G has a triangle T not in Ì ; it uses edges from distinct triangles in Ì.
We gave G no edges within A, B, or C. Let V(T) = {aj , bj + s , cj +t }. By the
construction of G, these three edges require s ∈ S and t − s ∈ S, and also t = 2s
for some s ∈ S. Also, s and s are distinct, since T ∈ / Ì. Since 2s − s = t − s ∈ S,

the numbers s , s , 2s − s form an arithmetic progression in S.

PROOF OF THE REGUL ARITY LEMMA (optional)

It remains to prove the Regularity Lemma. For this we will need several lem-
mas about the behavior of densities of pairs of vertex sets.

11.1.24. Definition. For an A , B-bigraph G with density d, let  (G) = | A| | B| d2 .

For a partition Π of V(G) into V1 , . . . , V¾ , let G i , j be the subgraph of G gen-
erated by (Vi , Vj ), and let  (G , Π) = ∑i , j  (G i , j ).

We show first that if Π refines Π, then  (G , Π) ≤  (G , Π ) ≤ | E(G)|. Note

that  (G , Π ) = | E(G)| when the partition consists of singleton sets. The idea is
to iteratively refine pairs that are not -regular. Each refinement increases  ,
but the process cannot continue forever, since  is bounded by | E(G)|.

11.1.25. Lemma. Let G be an A , B-bigraph with density d, and let { A1 , A2 } par-

tition A, with ai = | Ai |. If the subgraph G i generated by (Ai , B) has density
d i , then  (G 1) +  (G 2) =  (G) + (a1/a2) | A| | B| (d − d1)2 .
Proof: Let a = | A| and b = | B| , and let m , m1 , m2 be the numbers of edges in
G , G 1 , G 2 , respectively. Note that  (G) = abd2 = m2/(ab) = md. Letting x =
m1/m and  = a1/a, we compute
ba1 a2 ba1 a2 m12 m22 m2
( (G 1) +  (G 2) −  (G)) = ( + − )
m2 a m2 a a1 b a2 b ab
m1 2 a2 m 2a a a
=( ) + ( 2) 1 − 122
m a m a a
= x2(1 − ) + (1 − x)2  − (1 − ) = x2 − 2x  + 2 = (x − )2 .
Therefore,
m2 a a m 2 a1 m 2
 (G 1) +  (G 2) −  (G) = (x − )2 = 1 ab ( ) ( − 1)
ba1 a2 a2 ba1 a m
a m m 2 a
= 1 ab ( − 1 ) = 1 ab(d − d1)2 .
a2 ab a1 b a2
Section 11.1: Forced Subgraphs 487

11.1.26. Lemma. Let G be an A , B-bigraph with density d. Let { A1 , A2} and

{B1 , B2 } be partitions of A and B, with | A1 | ≥ | A| and | B1 | ≥ | B| . Let i , j =
(G[Ai ∪ Bj ]). If | (A1 , B1) − d| ≥  , then ∑ i , j ≥ (G) + 4 | A| | B| .
Proof: Let a , b , ai , bj be the sizes of A , B , Ai , Bj , respectively. Also let d∗ =
(A1 , B) and d i , j = (Ai , Bj). By Lemma 11.1.25,
a
(G) = (G[A1 ∪ B]) + (G[A2 ∪ B]) − 1 ab(d − d∗)2
a2
b1 a
≤ 1 ,1 + 1 ,2 − a1 b(d∗ − d1 ,1)2 + 2 ,1 + 2 ,2 − 1 ab(d − d∗)2 .
b2 a2
Here we have ignored the gain due to the second subsplit.
It now suffices to show bb12 aa1 (d∗ − d1 ,1)2 + aa12 (d − d∗)2 ≥ 4 . Since b1 ≥ b and
a1 ≥  a, we only need 1− (d∗ − d1 ,1)2 + 1− (d − d∗)2 ≥ 4 . For fixed d and d1 ,1 , we
d+ d
minimize the left side by setting d∗ = 1+ 1 ,1 , since then

d∗
[(d∗ − d1 ,1)2 + (d − d∗)2 ] = 2 (d∗ − d1 ,1) − 2(d − d∗) = 0.
d− d1 ,1
With d∗ so chosen, d∗ − d1 ,1 = 1+ and d − d∗ = 1+
(d − d1 ,1). Using
|d1 ,1 − d| ≥  , we have
⎡ 2⎤
   ⎢ (  ) + (  ) ⎥
⎢ 2 2 ⎥
(d∗ − d1 ,1)2 + (d − d∗)2 ≥ ⎢
1− 1− 1−⎢ 1+ 1+ ⎥⎥
⎣ ⎦
4 (1 + ) =
4
= ≥ 4 .
(1 + )2(1 − ) 1 − 2

11.1.27. Lemma. Let G be an n-vertex graph, and let Π be an  , -partition of

G, where 0 <  < 15 . If Π is not -regular and | V0 | < ( − 6)n, then there is
# #
an  ,  -partition Π of G with  ≤  ≤  2+2/6 such that #### V0 #### ≤ | V0 | + n/2
and (G , Π ) ≥ (G , Π) + 5 n2/2.
Proof: If Π is not -regular, then too many pairs of parts are not -regular. Let m
be their common size. Each bad pair (Vi , Vj ) yields sets X ⊆ Vi and Y ⊆ Vj with
| X | , | Y | ≥  m such that |(X , Y) − (Vi , Vj)| ≥ . By Lemma 11.1.26, replacing Vi
with {X , Vi − X} and Vj with {Y, Vj − Y} increases (G , Π) by at least 4 m2 .
The idea is to capture this gain simultaneously for all the pairs that are not
-regular. For each pair involving Vi that is not -regular, we obtain a partition
{X , Vi − X} of Vi . With fewer than  such pairs, the least common refinement
of these partitions has fewer than 2 parts. Apply this to each Vi , obtaining a
partition P consisting of V0 and at most  2 other parts, with unequal sizes.
To compare (G , P) with (G , Π), we group the pairs of parts according to
the pairs in Π generating them. When (Vi , Vj ) is not -regular, witnessed by
(X , Y), the contributions from subsets of Vi and Vj can be obtained by first split-
ting Vi and Vj according to X and Y and then continuing the refinement. The
gain of at least 4 m2 is followed by further gains we ignore. When computing the
contribution for subsets of Vi with subsets of Vj , it does not matter that the ini-
tial split is different, since we are obtaining a lower bound on the contributions
from a set of subpairs disjoint from those arising from {Vi , Vj }. Since more than
 2 pairs of parts in Π failed to be -regular, we have (G , P) ≥ (G , Π) + 5  2 m2 .
488 Chapter 11: Extremal Problems

We still need a partition whose nonexceptional parts have equal size. Let m
be the desired part-size, to be specified later. Break each part in P other than V0
into blocks of size m , with fewer than m left over. Combine the leftover vertices
with V0 to form V0 ; this produces Π with ¾ parts other than V0 , where ¾ ≤
# #
n/m . Note that #### V0 #### ≤ | V0 | + m ¾ 2¾ .
Splitting into blocks of size m further increases (we ignore this gain), but
combining the leftovers into V0 reduces . Combining singletons with V0 would
lose the most. The contributions of a singleton part with all other parts sum to
at most the degree of the singleton vertex. Since m  2¾ is an upper bound on
### V − V ### , the amount lost by forming V is at most m  2¾ n.
## 0 0 ## 0
Since  m ≥ n(1 − ), we now have (G , Π ) ≥ (G , Π) + 5 n2(1 − )2 − m  2¾ n.
The gain is at least 5 n2 − [5 n2(2  − 2) + m  2¾ n]. To ensure net gain at least
5 n2/2, we want 26 n + m  2¾ < 5 n/2. With  < 1/5, it suffices to choose m so
that m  2¾ ≤ 6 n/2. Set m = ⌊ 6 n/( 2¾+1 )⌋ . Now Π is an  ,  -partition with 
bounded by a bit more than  2¾+1/6 . Doubling this value suffices. Finally, note
that the exceptional part has gained at most 6 n/2 vertices.

The simple formula for the gain per split is why we keep equal size for the
nonexceptional parts. Later the exceptional part can be distributed to the others.

Proof of the Szemer édi Regularity Lemma (as Theorem 11.1.15) : Given pos-
itive constants  and l, we seek positive integers M and N such that every graph
with at least N vertices has an -regular  , -partition for some  ∈ [l , M].
To apply Lemma 11.1.27, we need  < 15 . Since an -regular  , -partition is
also an  -regular  , -partition when  >  , we may assume  < 15 . Similarly, if
the claim is true for a given value of l, it remains true for smaller values, so we
may assume l > 2/.
Given such  and l, define M and N as follows. Let L0 = l and Li+1 =
⌈ Li 2 Li +2/6 ⌉ . Let M = L⌈1/ 5 ⌉ , and choose N with N > max{M, l/6 }. Let G be
an n-vertex graph, where n ≥ n. Since l > 2/ , we can form an  , l-partition of
G by breaking V(G) arbitrarily into l parts of size ⌊ n/l⌋ , with fewer than  n/2
vertices left over for the exceptional part.
Since | E(G)| < n2/2, fewer than −5 iterations of the refinement procedure in
Lemma 11.1.27 produce an -regular  , -partition of G. The initial exceptional
set has fewer than  n/2 vertices, and the exceptional set gains fewer than 6 n/2
with each iteration. With fewer than −5 iterations, the final exceptional set has
at most  n vertices. Also, the final partition has at most M parts.

11.1.28.* Remark. Since the upper bound on the number of parts in Lemma
11.1.27 is  2+2/6 , iterating −5 times makes N an exponential tower with height
−5 . Thus the Regularity Lemma applies only to enormous graphs.
We have given only a few applications of Regularity; there are many others.
Komlós–Simonovits [1996] is an influential early survey; others include Komlós–
Shokoufandeh–Simonovits–Szemerédi [2002], Kohayakawa–Rödl [2003], Rödl–
Schacht [2010]. K ühn–Osthus [2009] explains the use of the regularity/blow-
up method in general. The original lemma gives useful results only for dense
graphs; Kohayakawa [1997] developed an analogue for sparse graphs (see also
Kohayakawa–Rödl [2003], Gerke–Steger [2005], and Scott [2011]). Frankl–Rödl
Exercises for Section 11.1 489

[2002] extended the lemma to 3-uniform hypergraphs, and Rödl–Skokan [2004,

2006] extended it to ¾-uniform hypergraphs. Textbooks discussing Regularity
include Bollobás [1998] and Diestel [1997, etc.] (from the 2000 edition onward).
There are lecture notes by Lee [2015], Palmer [2015], and Shapira [2016]. See
also Martin [2012, Chapters 6–9] and Tao–Vu [2006, Section 10.6].

EXERCISES 11.1

11.1.1. (−) Complete the computation in Corollary 11.1.7.

11.1.2. (−) Find the minimum size of a maximal triangle-free n-vertex graph.
11.1.3. (−) Determine ex (n; P3) and ex (n; P4).
11.1.4. Determine the maximum size of n-vertex graphs of the following types.
(a) Graphs having an independent set of size ¾ .
(b) Graphs having ¾ components.

11.1.5. Let G be a connected graph having neither P4 nor the 4-vertex “paw” obtained by
adding one edge to K1 ,3 as an induced subgraph. Prove that G is complete multipartite.

11.1.6. Prove that every n-vertex graph not containing K r+1 has at most (1 − 1r ) n2 edges.
2

11.1.7. (♦) Turán’s proof of Turán’s Theorem. Recall that tr(n) = | E(T n ,r)| .
(a) Prove that a maximal graph with no (r + 1)-clique has an r-clique.
(b) For n ≥ r, prove that tr(n) = (2r ) + (n − r)(r − 1) + tr(n − r).
(c) Use parts (a) and (b) to prove Turán’s Theorem by induction on n, including the
uniqueness of graphs achieving the bound. (Turán [1941])
11.1.8. Let S be the set of nonnegative vectors in n with sum 1. Given a graph G with
vertex set {v1 , . . . , vn}, let (x) = ∑vi v ∈ E(G) xi xj for x ∈ S, and let  = max x∈S (x).
j
(a) Prove that is maximized by some vector whose nonzero coordinates correspond
to the vertices of a clique. Conclude  = 12 (1 − 1/(G)).
(b) Prove that  is attained by a vector x with all coordinates nonzero if and only if G
is a complete multipartite graph.
(c) Prove Turán’s Theorem for K r+1 -free n-vertex graphs by induction on n − ⌊ n/r⌋ ,
using parts (a) and (b) for the base case. (Motzkin–Straus [1965])
11.1.9. (♦) The Turán graph T n ,r with size tr(n) is the complete r-partite graph with b
parts of size a + 1 and r − b parts of size a, where a = ⌊ n/r⌋ and b = n − ra.
(a) Prove that tr(n) = (1 − 1/r)n2/2 − b(r − b)/(2r).
(b) Since tr(n) is an integer, part (a) yields tr(n) ≤ ⌊ (1 − 1/r)n2/2⌋ . Determine the
smallest value of r such that strict inequality occurs for some n. For this value of r, deter-
mine the values of n such that tr(n) < ⌊ (1 − 1/r)n2/2⌋ .

11.1.10. Let a = ⌊ n/r⌋ . Compare the Turán graph T n ,r with the graph K a + K n− a to prove
directly that tr(n) = ( n−2 a) + (r − 1)(a+2 1 ) .

11.1.11. Given positive integers n and  , let q = ⌊ n/ ⌋ , r = n − q , s = ⌊ n/( + 1)⌋ , and
t = n − s( + 1). Use Turán’s Theorem to prove (2q) + rq ≥ (2s)( + 1) + ts. (Richter [1993])

11.1.12. Let G be a graph with n vertices and m edges. Determine lower bounds on (G)
and (G) in terms of n and m that are sometimes sharp.
490 Chapter 11: Extremal Problems

11.1.13. Let S be a set of n points in a circular region with radius 1√, and consider the
distance between any two points in S. Prove that the distance exceeds 2 for at most n2/3
of the pairs. (Bondy–Murty [1976, p. 114])
11.1.14. (♦) Let G be a graph with n vertices that has tr(n)− ¾ edges and at least one (r + 1)-
clique, where ¾ ≥ 0. Prove that G has at least r(n) + 1 −  cliques of size r + 1, where
r(n) = n − ⌈ n/r⌉ − r. (Hint: Prove that a graph with exactly one (r + 1)-clique has at most
tr(n) − r(n) edges.) (Erdős [1964], Moon [1965c])
11.1.15. (♦) Let ¾(n) be the minimum size of a family F of subsets of [n] such that every
subset of [n] with size at most  is the union of two (possibly equal) members of F . Note
that 1 (n) = 2√(n) = n + 1. (Comment: Füredi–Katona [2006] determined also 4(n).)
(a) Prove 2 · 2 n/2 − 1/2 ≤ n(n) ≤ 2⌈n/2⌉ + 2⌊n/2⌋ − 1.
(b) Prove 3(n) = 1 + n + ( n2) − ⌊ n2/4⌋ .

11.1.16. Extensions of Turán’s Theorem. As usual, let tr(n) = | E(T n ,r )|.

(a) If n ≥ r + 1 ≥ 4, prove that every n-vertex graph with tr−1 (n) + 1 edges contains
K r+1 − e as a subgraph. (Hint: Remove a vertex of minimum degree.)
(b) If n ≥ n > r ≥ 2, prove that every n-vertex graph with tr(n) + p edges has an n -
vertex subgraph with tr(n ) + p edges, where p ≤ 1.
(c) If 1 ≤ p ≤ r − 1, prove that every n-vertex graph with tr(n) edges other than Tr,n
has an (r + p)-vertex subgraph with more than (r+2 p) − p edges. (Dirac [1963])
11.1.17. (+) The Turán graph T n ,r is regular when r divides n. Prove that there is a -
regular triangle-free graph on n vertices if and only if n is even and at least 2  or n is odd
and at least 5 /2, with  even when n is odd. (Comment: Bauer [1983] determined all
triples (r, n , ) such that there is a -regular K r+1 -free graph on n vertices.)
11.1.18. Let G be the Petersen graph. For n ≥ 2, prove ex (n , G) ≥ 2n − 3 + ⌊ (n − 2)2/4⌋ .
(Comment: Simonovits [1999] proved that equality holds.)
11.1.19. Prove that the minimum number of edges in a connected n-vertex graph where
every edge lies in a triangle is 3(n − 1)/2 if n is odd, 3n/2 − 1 if n is even. (Erdős [1988a])
11.1.20. (+) Determine the minimum number of edges in a triangle-free graph on 2n ver-
tices whose complement contains no n-clique. (Erdős [1988b])
11.1.21. Determine the maximum number of edges in a 10-vertex graph with no 4-cycle.
(Bialostocki–Schönheim [1984])
11.1.22. For n ≥ 6, prove that the maximum number of edges in an n-vertex graph not
having two edge-disjoint cycles is n + 3. (Erdős–P ósa [1962])
11.1.23. (♦) Prove that if | V(G)| ≥ 6 and (G) ≥ 4, then G has two disjoint cycles. Con-
clude that for n ≥ 6 the maximum number of edges in an n-vertex graph not having two
disjoint cycles is 3n − 6. (Erdős–P ósa [1962]) (Comment: Corradi–Hajnal [1963] proved
more generally that if | V(G)| ≥ 3 and (G) ≥ 2  , then G has  pairwise disjoint cycles.)
11.1.24. Let G be a connected graph with m edges and more than one vertex.
(a) Prove that if G is P4 -free, then G is disconnected. (Seinsche [1974])
(b) Prove that if (G) = D and G is disconnected, then m ≤ D2 , with equality only for
K D ,D . (Chung–West [1993])
11.1.25. Let G be a self-complementary n-vertex graph, and let  = ⌊ n/4⌋ . Conclude from
Exercise 27 that G has at least ( − 1)(4  − 2)/3 triangles if n = 4  and that G has at least
( − 1)(4 + 1)/3 triangles if n = 4 + 1.
11.1.26. (♦) Let G be a graph on n vertices such that G has no triangles. For n ≥ 6, prove
that the minimum possible number of triangles in G is (⌊n/2⌋
3
) + (⌈n/2⌉
3
).
Exercises for Section 11.1 491

11.1.27. For an n-vertex graph G, prove the bounds below and show that they are sharp.
⎧
⎪ 2( l ) if n = 2l
⎪
⎪2 3
¾3(G) + ¾3(G) ≥ ⎨ 3 ¾(¾ − 1)(4 ¾ + 1) if n = 4 ¾ + 1 .
⎪
⎪
⎪
⎩ 3 ¾(¾ + 1)(4 ¾ − 1) if n = 4¾ + 3
2

11.1.28. Let H¾ = K1 ¾ K 2 . Prove that ex (n; H2) = ⌊ n2/4⌋ + 1 for n ≥ 5. (Comment:

Erdős–Füredi–Gould–Gunderson [1995] proved ex (n; H¾) = ⌊ n2/4⌋ + ¾ 2 − ¾ for odd ¾ and
ex (n; H¾) = ⌊ n2/4⌋ + ¾ 2 − 3¾/2 for even ¾ .)
11.1.29. A graph G with vertex degrees d1 , . . . , d n is r-majorizable if there is an r-partite
graph H with vertex degrees d 1 , . . . , d n such that d i ≥ d i for 1 ≤ i ≤ n. By Erdős’ proof of
Turán’s Theorem (Theorem 5.2.11), every K r+1 -free graph is r-majorizable.
(a) Construct non-r-majorizable n-vertex graphs containing only one copy of K r+1 .
(b) Prove that G is r-majorizable if (G) ≤ n(1 − 1/r).

(c) Consider n , D , r ∈ such that n − 1 ≥ r(n − D) + (r−2 t) , where 0 ≤ t < r. Given
n1 , . . . , nr with sum n − 1 such that n − D = n1 = · · · = nt = nt+1 < · · · < nr , construct
G from K n1 ,... ,nr by adding one vertex adjacent to all vertices in t smallest parts and one
vertex in each other part. Prove that G is not r-majorizable.
(d) The graph of part (c) has (n − D)t copies of K r+1 . For r = 2, prove that this is the
fewest among non-r-majorizable n-vertex graphs with (G) = D. (Comment: West [1982b]
proved this for r = 2 and also for (n , D , r) = (7 , 5 , 3); does it always hold?)
•

• • • • •
• • • • •
• • • • •
• •
•

11.1.30. Construct a 3-uniform hypergraph with asympotically 95 (n3) edges that does not
(3)
contain K4 , the complete 3-uniform hypergraph with four vertices. (Comment: Turán
[1941] conjectured that this is asymptotically the most edges in such an n-vertex hyper-
(3)
graph: that is, ex (n; K4 )/(3n) → 5/9. Erdős offered $1000 for a proof. Chung–Lu [1999]
proved an upper bound of .593, and Razborov [2011] lowered it to .561.)
11.1.31. (♦) Erdős–Stone Theorem (Theorem 11.1.10). Fix ∈ (0 , 1), and consider r, t ∈ . 
(a) Prove that when n is sufficiently large, every n-vertex graph with minimum degree
at least (1 − 1/r + )n contains K r+1 [t]. (Hint: Use induction on r.) (See Lovász [1993])
(b) Prove ex (n; K r+1 [t]) ≤ (1 − 1/r + )n2/2 for sufficiently large n. (Erdős–Stone [1946])
11.1.32. An ordered graph is a graph with linear order on the vertices. The interval
chromatic number of an ordered graph is the least number of independent sets of consec-
utive vertices needed to cover the vertices (ranging from 2 to n for n-vertex ordered paths).
An ordered graph G avoids an ordered graph H if H does not appear in G via an order-
preserving injection on the vertices. Let ex (n; H) be the maximum number of edges in an
n-vertex ordered graph that avoids H . Use the Erdős–Stone Theorem to prove that if H has
interval chromatic number  , then ex (n; H) = (1 − ¾−1 1 )(n2) + o(n2). (Pach–Tardos [2006])
11.1.33. A graph G is F-saturated if it does not contain F , but adding any edge creates
a copy of F . For n ≥ t − 1, prove that K t−2 K n− t+2 has the fewest edges among all K t -
saturated graphs with n vertices. (Erdős–Hajnal–Moon [1964])
492 Chapter 11: Extremal Problems

11.1.34. (♦) Prove that there is an n-vertex F-saturated graph (see Exercise 11.1.33 for def-
inition) having at most ( (F) − 1)n + (F)n/2 edges, where (F) is the vertex cover number
of F . (K ászonyi–Tuza [1986]; clarified in Füredi [2007])
11.1.35. Let sat(n; F) be the minimum number of edges in an F-saturated graph with n
vertices (see Exercise 11.1.33 for definition).
(a) Determine sat(n; K3).
(b) With P being the Petersen graph, show that sat(n; P) ≤ 4n − 4.
11.1.36. Given the Regularity Lemma (Theorem 11.1.15), prove that one can instead re-
quire an -regular partition in which each part belongs to at most  irregular pairs (in-
stead of a total of  2 irregular pairs).
11.1.37. (♦) The Ramsey–Tur án problem. (Szemerédi [1972])
(a) Let G be a graph with 2n vertices, partitioned into sets A and B of size n. Show
that if | [ A , B]| ≥ (1/2 + )n2 , then G contains K4 or an independent set of size  n/2.
(b) Let G be a graph with 3n vertices partitioned into sets A, B, and C, each of size n.
Show that if each pair among { A , B , C} is -regular and has density more than 2  , then
G contains K4 or an independent set of size at least 2 n.
(c) (+) Let G be an n-vertex graph. Prove that if G does not contain K4 , and (G) ∈
o(n), then | E(G)| < (1/8)n2 + o(n2).
11.1.38. The Slicing Lemma. Let (A , B) be an -regular pair with density d, where
0 <  ≤  ≤ 1/2. Prove that for any A ⊆ A and B ⊆ B with | A | ≥  | A| and | B | ≥  | B| ,
the pair (A , B ) is /-regular with density greater than d − .
11.1.39. (♦) Let G be an A , B-bigraph with | A| = | B| = n.
(a) Prove that if G has no independent set consisting of (G) vertices in A and (G)
vertices in B, then G has a perfect matching.
(b) Obtain a perfect matching when (A , B) is an -regular pair and (G) >  n.
11.1.40. (♦) Given an -regular pair (A , B) with density d and | A| = | B| = m, let a vertex
v in A ∪ B be good for the pair if it has at least (d − )m neighbors in the other part.
(a) Let x ∈ A and y ∈ B be fixed good vertices in the pair (A , B) as specified above.
Prove that if d > 5  , then when m is sufficiently large there is an x , y-path of length at
least (1 −  − d− )2m that avoids any K specified vertices.
(b) Given d > 5  , let R be the reduced graph of -regular pairs with density at least d
resulting from an -regular partition of an n-vertex graph G. Prove that if R has  vertices
and some component of R has a matching of size t, then G has a cycle of length at least
(1 −  − d− ) 2t

n. (See Figaj–Łuczak [2007] for similar results.)
11.1.41. (+) Let G be an A , B-bigraph √ with | A| = | B| = n and (G) ≥ dn. Prove that if
(A , B) is an -regular pair and d > 3 +  , then G has a spanning cycle. (Haxell [1997])
11.1.42. Let (A , B) be an -regular pair with density d and | A| = | B| = n.

(a) For  ∈ , consider Y ⊆ B such that (d − )−1 | Y | ≥  | B|. Prove that the number
of -tuples (x1 , . . . , x) of elements of A having at most (d − ) | Y | common neighbors in Y
is at most  | A| . (Komlós–Simonovits [1996]) (Comment: Lemma 11.1.20 is  = 1.)
(b) Prove that at least 14 (1 − 4 )(d − )4 n4 4-cycles alternate between A and B.
11.1.43. (♦) (6,3)-Theorem (Ruzsa–Szemerédi [1978]).
(a) Let G be an n-vertex graph with every edge in exactly one triangle. Use the
Triangle-Removal Lemma (Lemma 11.1.22) to prove that G has o(n2) edges.
(b) Let G be an n-vertex graph that decomposes into at most n induced matchings (ver-
tices of distinct edges in such a matching are not adjacent). Prove that G has o(n2) edges.
(c) Let H be a 3-uniform n-vertex hypergraph having no six vertices that induce at
least three edges. Prove that H has o(n2) edges.
Section 11.2: Families of Sets 493


11.1.44. (♦) Fix c ∈ (0 , 1). Prove that for some n ∈ , any set of at least cn2 points in [n]2
contains three distinct points of the form {(x , y) , (x + a , y) , (x , y + a)}. (Hint: Use Lemma
11.1.22 on a suitable graph.) (Ajtai–Szemerédi [1974], Solymosi [2003])

11.2. Families of Sets

In this section, we consider various extremal problems for families of subsets
of [n]. West [1982a], Frankl [1987], and K leitman [1994] survey these and re-
lated topics. See also the books Engel [1997], Jukna [2011], and Gerbner–Patkós
[2019]. Frankl–Tokushige [2016] is an authoritative survey. At the end, we dis-
cuss entropy as a tool for extremal problems in enumeration.

THE KRUSK AL–K ATON A THEOREM

In extensions of the Turán problem, we wanted to choose m edges to have the

fewest r-cliques. Similarly, we may ask for m vertex sets of size ¾ that contain
the fewest sets of size ¾ − 1.

11.2.1. Definition. A family of sets or set system is a set whose members are
sets. A ¾-uniform family is a family of ¾-sets.

We use “member of family” to avoid confusion with “element of set ”. A fam-

ily of subsets of [n] is the edge set of a hypergraph with vertex set [n], and “ ¾-
uniform” has the same meaning in both contexts. In Section 10.2 we used hyper-
graph language to generalize notions from graph theory. Some authors use the
term “hypergraph” just for a family of sets without specifying the vertex set.

11.2.2. Definition. The t-shadow of a family F is the family of all t-sets con-
tained in members of F . The shadow F of a -uniform family F is its ( − 1)-
shadow. The shade is the family of all ( + 1)-sets containing members of F .

Among all -uniform families of size m, we seek the family F with smallest
shadow. Since each -set contributes exactly  sets to the shadow and | F | is fixed,
the size of the shadow is minimized when the sets in the shadow arise from many
members of F . This can be achieved by confining F to subsets of a small subset of
[n]. The amazing result is that there exists an ordering of the -sets such that,
for each m, the first m sets in this ordering form an optimal family.
To confine the initial sets to a small subset of [n], we define an ordering of
-sets in which x precedes y if the largest element of x is less than the largest
element of y. We use notation like x and y for -element subsets to emphasize
their interpretation as binary incidence vectors.

11.2.3. Definition. The co-lexicographic or colex ordering on a family of -

sets is obtained by putting x < y if x i < yi in the highest coordinate where
their binary incidence vectors differ.
494 Chapter 11: Extremal Problems

11.2.4. Example. Colex ordering on ([5] 3

). It is convenient to start the indexing at
0 for both the list of vectors and the coordinates of the vectors.
index set positions of 1s index set positions of 1s
0 11100 012 5 10101 024
1 11010 013 6 01101 124
2 10110 023 7 10011 034
3 01110 123 8 01011 134
4 11001 014 9 00111 234

To facilitate the proof that the first m sets in this order form an optimal fam-
ily, we study its shadow.

11.2.5. Lemma. If the vector with index m in the colex ordering on ([n] ¾ ) has 1 in
positions m1 , . . . , m¾ , with m1 < · · · < m¾ , then m = (m¾¾ ) + (m¾−¾−11 ) + · · · + (m11 ).
Proof: Let be the vector with index m. Indexing starts with 0, so m counts the
vectors preceding . We count them another way. To reach , we must skip all
vectors whose th 1 appears before position m¾ . There are (m¾¾ ) of these, since
the first position is position 0. Some vectors with the last 1 in position m¾ also
precede  . These begin with (m¾−¾−11 ) vectors having − 1 1s in positions before m¾−1 .
Continuing, the jth term in the summation counts the vectors that appear before
 in the ordering and have their last j − 1 1s in the same positions as  .

11.2.6. Corollary. For m , ∈ , there is a unique expression of m in the form
(m¾¾ ) + (m¾−¾−11 ) + · · · + (mi i ) with m¾ > · · · > mi ≥ i.
Proof: In the colex ordering on ([n]¾ ), appending 0s increases the number of co-
ordinates without changing the position of the first (¾n) vectors. Lemma 11.2.5
thus applies for any n with (¾n) > m to obtain an expression for m using the vector
with index m in the ordering. Left-justified 1s in the vector yield mj = j − 1 and
contribute 0; we drop those terms from the expression in Lemma 11.2.5 to obtain
an expression of the desired form. Two such expansions correspond to distinct
vectors in the colex ordering and hence distinct values of m.

11.2.7. Definition. The unique representation of m described in Corollary 11.2.6

is the -binomial expansion of the integer m. For an integer m with -
binomial expansion m = (m¾¾ ) + (m ¾ −1 ) + · · · + ( i ) such that m¾ > m¾ −1 · · · >
¾ −1 mi

mi ≥ i, let ¾ (m) = (¾m−¾1 ) + (m¾−¾−21 ) + · · · + (im

−1 ).
i

11.2.8. Lemma. The shadow of the first m vectors in the colex order on ([n]
¾ ) con-
sists of the first ¾ (m) vectors in the colex order on (¾[n]
−1 ).
Proof: We count the shadow using Lemma 11.2.5. The family consisting of all
elements of ([n]
¾ ) whose last 1 appears before position m¾ has shadow consisting of
all vectors of weight − 1 whose last 1 precedes position m¾ . In general, the j-th
term in the summation for ¾ (m) considers all vectors before the indexed vector vm
in ([n]
¾ ) that have their rightmost j − 1 ones in the same positions as vm . It counts
the portion of their shadow consisting of sets whose last j − 1 ones are in those
same positions. This is the portion of the shadow not counted by earlier terms.
Thus the full sum counts the entire shadow exactly.
Section 11.2: Families of Sets 495

11.2.9. Definition. On a family F in 2[n] , the shift operator Ìi , j is defined by

x − { j} + {i} if j ∈ x and i ∈
Ìi , j (x) = {
/ x and x − { j} + {i} ∈
/ F,
x otherwise.
Let Ìi , j (F) = {Ìi , j (x): x ∈ F}.

11.2.10. Lemma. If F ⊆ 2[n] and i < j , then (i , j (F)) ⊆ i , j ( F).

Proof: Let G = F . Let  F and  G denote i , j on F and G, respectively. It suffices
to show y ∈  G(G) when y ∈ ( F (F)). For y ∈ ( F (F)), we have y = x − s for some
s ∈ x ∈  F (F). Note also that x =  F (x) for some x ∈ F .
Case 1: x = x. In this case y ∈ G, and it suffices to show  G(y) = y. If not,
then i ∈ / y and j ∈ y. Since j ∈ y and y = x − s, we have s
= j and j ∈ x. If s = i,
then y − j + i = x − j ∈ G, so  G(y) = y. If s
= i, then i ∈ / y yields i ∈
/ x; now
 F (x) = x implies x − j + i ∈ F . Hence y − j + i ∈ F = G, so again  G(y) = y.
Case 2: x
= x. Now x contains j but not i, and x = x − j + i. Since y = x − s,
we have j ∈ / y, and hence  G(y) = y if y ∈ G. If i ∈
/ y, then i = s and y = x − j ∈ G.
Hence we may assume i ∈ y and y ∈ / G. Let ŷ = y + j − i, so now ŷ ∈ x ⊆ G. With
ŷ ∈ G and ŷ − j + i = y ∈/ G, we have  G(ŷ) = y, so y ∈  G(G) as desired.

11.2.11. Theorem.(Kruskal–Katona Theorem; Kruskal [1963], K atona [1968])

For F ⊆ ([n]
¾ ) with | F | = m, the minimum of | F | occurs when F consists of
the first m members of the colex ordering on ([n]
¾ ), and then | F | = ¾
(m).
Proof: (Frankl [1984]) By Lemma 11.2.8, the shadow of the specified family has
size ¾ (m), so it suffices to show | F | ≥ ¾ (m) for all F ⊆ ([n] ¾ ) with | F | = m.
We use induction on  ; trivially the claim holds for a 1-uniform family. For
the induction step, we use induction on m. Note that m = (¾¾ ) when m = 1, and
indeed then | F | =  = (¾−¾ 1 ). Now consider m > 1.
Fix i ∈ [n]. Applying any i , j does not change the number of sets, and it
increases the number of sets containing i if it produces a change. From F , appli-
cation of finitely many such operators therefore leads to a family F ∗ , invariant
under all i , j , in which i is a dominant element, meaning x − { j} + {i} ∈ F ∗ for all
j ∈ x ∈ F ∗ . By Lemma 11.2.10, | F ∗ | ≤ | F | . Hence it suffices to prove | F | ≥ ¾ (m)
in the case where F has a dominant element i.
From F define F = {x: i ∈ / x ∈ F} and F − = {x − i: i ∈ x ∈ F}. Note F ⊆ ([n]
¾ ),
− [n] −
F ⊆ (¾−1 ), and | F | + | F | = m. If j ∈ x ∈ F and i ∈ / x, then x − j + i ∈ F , since i
is dominant. Hence x − j ∈ F − and F ⊆ F − . We conclude that F consists of F −
plus the sets formed by adding i to members of the shadow of F − .
¾
Let ∑ j =i (mj j ) be the -binomial expansion of m, as in Definition 11.2.7. Let
m = ∑ j =i (mj j−1). If | F − | < ¾ (m ), then | F | = m − | F − | > m − ¾ (m ) = m . Now,
¾

since | F | < | F | = m, the induction hypothesis on m yields | F | ≥ ¾ (m ) > | F − | ,

which contradicts F ⊆ F − .
Hence we may assume | F − | ≥ ¾ (m ). In this case the induction hypothesis on
 yields | F − | ≥ ¾−1 ¾ (m ) ≥ ∑¾j =i (mj j−−21). Finally, we compute
¾ m −1 ¾
| F| ≥ |F−| + | F−| ≥
¾ (m ) + ∑ j =i ( j j−2 ) ≥ ∑ j =i ( j −j1) = ¾(m).
m
496 Chapter 11: Extremal Problems

Note that the optimal family depends only on m and ¾ , not n, as long as (¾n) ≥
m. The Kruskal–K atona Theorem is a very precise statement. For many appli-
cations, a smoothed version due to Lovász [1979] suffices. It treats m as a value
of the “choose ¾ ” polynomial and can also be proved using shift operators (Ex-
ercise 9). Instead, we give a short more recent proof. Exercise 10 requests the
characterization of when equality holds.

11.2.12. Theorem. (Lovász [1979, Ex. 13.31(b)]) For u ≥ ¾ , let (u¾ ) =

u(¾)
¾! . If F ⊆
([n]
¾ ) with | F | = (¾ ), then | F | ≥ (¾ −1 ).
u u

Proof: (Keevash [2008]) Note that F is contained among the -sets whose ( − 1)-
subsets are all present in F . Letting r =  − 1, it therefore suffices to show that
when G is a family of size (ur) in ([n] r
), the number of (r + 1)-sets whose r-sets are
u
all present in G is at most (r+1 ).
We use induction on r; the claim is immediate when r = 1. For r > 1, we
treat G as the edge set of a hypergraph. We may assume that each v ∈ [n] lies in
some member of G. Let d(v) be the number of edges of G containing v. Let G (v)
be the family of (r − 1)-sets obtained by deleting v from edges of G. Let q(v) be the
number of (r + 1)-sets containing v whose r-sets all lie in G.
Note that S ∪ {v} is counted by q(v) if and only if S ⊆ G and all (r − 1)-subsets
of S lie in G (v). The first condition implies q(v) ≤ | G | − d(v), while the second
bounds q(v) by the number of r-sets whose (r − 1)-subsets all lie in G (v).
We claim that for all v, at least one of these bounds is at most u−r r d(v). If
d(v) ≥ (ur−−11), then (ur) − d(v) ≤ ur (ur−−11) − d(v) ≤ u−r r d(v). If d(v) ≤ (ur−−11), then define

u by d(v) = (ur−−11 ); note that u ≤ u. Note also that | G (v)| = d(v). By the induction

hypothesis, at most (u r−1 ) elements of ([n] r
) have all their (r − 1)-sets in G (v). We
u −1 u −r u −1 u−r
compute ( r ) = r ( r−1 ) ≤ r d(v).
Finally, every (r + 1)-set whose r-sets are all present in G is counted r + 1
times in ∑v∈[n] q(v). Hence the number of these sets is bounded by r(ru−+r1) ∑v∈[n] d(v).
Since the degree sum is r | G | , which equals r(ur), the bound simplifies to (r+u1 ).

According to Engel [1997, p. 46], Theorem 11.2.11 was “formulated” by

Schützenberger [1959], independently “found” by Kruskal [1963], Harper [1966],
K atona [1968], and Clements–Lindström [1969], and reproved later by Hansel
[1972], Daykin [1974b], Eckhoff–Wegner [1975], Greene–K leitman [1978], Hilton
[1979], and Frankl [1984]. Clements–Lindström [1969] extended it to multisets
with bounded multiplicity. Related exercises include Exercises 7–11.

ANTICHAINS AND INTERSECTING FAMILIES

We began with the Kruskal–K atona Theorem due to its relation to Turán-
type problems, but the most fundamental problem in extremal set theory is to
maximize the size of a family in which no member contains another. Sperner
[1928] gave the answer. The theorem has many proofs, most of which extend to
interesting additional results. Chapter 12 presents some of these extensions.
Section 11.2: Families of Sets 497

11.2.13. Definition. An antichain of sets is a family of sets in which no member

contains another.

11.2.14. Theorem. (Sperner ’s Theorem; Sperner [1928]) The maximum size

of an antichain of subsets of [n] is (⌊n/2⌋
n
), achieved only by the family of ⌊ n/2⌋-
sets or the family of ⌈ n/2⌉-sets.
Proof: Let ¾ and l be the sizes of the smallest and largest members of a maximum
antichain F . Let A = F ∩ ([n] l
). Each l-set has l sets immediately below it, and
each (l − 1)-set lies under n − l + 1 sets of size l. Thus | A| ≥ | A| n− ll+1 . If l >
(n + 1)/2, then replacing the l-sets in F by their shadow yields a larger antichain.
n−¾
Similarly, the shade of F ∩ ([n]
¾ ) has size at least | A| ¾ +1 . If  < (n − 1)/2, then
replacing the -sets in F by their shade yields a larger antichain. Thus (n − 1)/2 ≤
 ≤ l ≤ (n + 1)/2.
When n is even, we obtain  = l. When n is odd, let A = F ∩ ((n+[n] 1)/2
). The
argument to enlarge F still works unless | A| = | A| , which requires all sets above
| A| to lie in A. Since one can move from any i-set to any other by alternately
adding and deleting elements, | A| = | A| occurs only when A is ∅ or ((n+[n] 1)/2
). Thus
the sets in F all have the same size.

We can also restrict intersections of members of our family.

11.2.15. Definition. An intersecting family is a family F such that A , B ∈ F

implies A ∩ B
= ∅; it is t-intersecting if always | A ∩ B| ≥ t. A star is a
family of sets having a common element; a t-star is a family having t common
elements.

This definition extends the graph-theoretic notion of star to the hypergraph

context. Under various conditions, the large intersecting families are stars.

11.2.16. Example. Intersecting families. An intersecting family of subsets of [n]

has at most 2 n−1 members, since a set and its complement cannot both appear.
The bound is achieved by the star consisting of all sets containing a fixed element.
Other largest intersecting families consist of all sets with more than half
the elements, plus (for even n) one from each complementary pair of sets of size
n/2. K atona [1964] (Exercise 15) generalized this to determine the largest t-
intersecting families (see Ahlswede–K hachatrian [2005] for other proofs).
The family of subsets of [n] containing one element and omitting another is
an intersecting family whose complements also form an intersecting family. It is
a largest such family, with size 2 n−2 (Seymour [1973], Schönheim [1974], Ander-
son [1976], Daykin–Lovasz [1976], Hilton [1976]; see Exercises 28–29.)

We next consider intersecting families of special types. Frankl [1995] sur-

veys many such classical problems. We emphasize the Erdős–Ko–Rado Theorem
on t-intersecting families consisting of an antichain of subsets of size at most  .

11.2.17. Definition. An EK R( , t)-family is an antichain F that is also a t-

intersecting family whose members have size at most  .
498 Chapter 11: Extremal Problems

We will see that a largest such family uses only ¾-sets. The ¾-sets contain-
ing t particular elements form a t-star that is an EK R(¾ , t)-family of size (¾n−−tt).
When n is sufficiently large, this is the extremal family. We first consider only the
case t = 1, using a counting argument. In this case we may express the problem
as a Turán-type problem; we seek a largest ¾-uniform hypergraph not contain-
ing the hypergraph consisting of two disjoint edges. Published in 1961, the EKR
Theorem was actually proved in 1938, as noted by Erdős [1990].

11.2.18. Theorem. (Erdős–Ko–Rado Theorem t = 1; Erdős–Ko–Rado [1961])

For n ≥ 2 ¾ , the maximum size of an EK R(¾ , 1)-family is (¾n−−11). For n > 2 ¾ ,
equality holds only for stars.
Proof: (K atona [1972b]) If some member of such a family F has size less than ¾ ,
then replacing all smallest members by their shade strictly enlarges the family
(since ¾ ≤ n/2) while preserving the conditions. Thus we may assume F ⊂ ([n] ¾ ).
Given a circular ordering of [n], we ask how many members of F occur in
as a consecutive segment of elements. Write such a member x as (a1 , . . . , a¾),
indexed in order in . Every -element segment that intersects x has a boundary
immediately following some a1 with 1 ≤ i <  . Since  ≤ n/2, two -sets having
the same boundary within x cannot intersect, as shown below. Hence at most
 − 1 members of F other than x appear in .
Summing over all (n − 1)! circular orderings yields at most (n − 1)!  appear-
ances of members of F . Since each member of F appears in  !(n − )! such order-
−1)!¾ n−1
ings, | F | ≤ ¾(n!(n−¾)! = (¾−1 ).

←x→

Equality requires having exactly  members of F in every circular ordering

of [n]. With x occuring in , for 1 ≤ i <  the location between ai and ai+1 must
be the boundary of one member x i of F . If always x i extends clockwise (or always
counterclockwise) from there, then the  appearances successively shift by one
element. Otherwise, some x i−1 and x i extend in opposite directions. Since n >
2  , they intersect only if they both contain ai . Now x i , . . . , x ¾−1 , x , x1 , . . . , x i−1
successively shift by one element. The argument holds for all orderings.
Let b be the element before a1 in . Consider all circular orderings such
that b is followed immediately by a1 , . . . , a¾−1 in some order and then a¾ . Since
/ F but x ∈ F , all -element segments in that contain a¾
{b} ∪ {a1 , . . . , a¾−1 } ∈
must lie in F . Using all such , every -set containing a¾ belongs to F .

The original proof pushed an EK R( , t) family toward the claimed extremal
family using shift operators.

11.2.19. Lemma. If F is a -uniform t-intersecting family, then also i , j (F) is

t-intersecting, where i , j is the shift operator of Definition 11.2.9.
Section 11.2: Families of Sets 499

Proof: View x , y ∈ F as binary vectors, and let x = Ìi , j (x) and y = Ìi , j (y). By

symmetry, suppose i < j . If x or y has 11 or 00 in these positions, or if they both
have 10, then | x ∩ y | = | x ∩ y| . If they have 10 and 01, then x ∩ y ⊇ x ∩ y.
This leaves the case where x and y both have 01 in these positions. If exactly
one of x and y changes to 10, then possibly | x ∩ y | < t = | x ∩ y| . If y = y (by
symmetry), then ∈ F , where = y − j + i. Since F is t-intersecting, x and
have at least t common elements outside {i , j}. Hence this holds also for x and
and for x and y , since and y agree outside {i , j}.

11.2.20. Theorem. (Erdős–Ko–Rado Theorem t = 1, revisited) For n ≥ 2  ,

the maximum size of an EK R( , 1)-family is (¾n−−11).
Proof: As in K atona’s proof, we may let F be a -uniform intersecting family.
We use induction on n +  . When n = 2  , in F there is at most one member of each
complementary pair, and choosing one from each such pair yields an intersecting
family. Since (2¾¾−−11 ) = 12 (¾n), the bound holds.
Now suppose n > 2  . Since the shift operator i , j does not change the size of
a family, by Lemma 11.2.19 we may assume that F is unchanged under all i , j
with i < j . Partition F into subfamilies F1 and F0 , with F1 consisting of those
members containing element n. Inductively, | F0 | ≤ (¾n−−12). Trivially | F1 | ≤ (¾n−−11);
we obtain the desired bound if we can improve this bound to (¾n−−22).
Let F = { A − n: A ∈ F1 }. By the induction hypothesis, it suffices to show
that F is an intersecting family. Otherwise, we have A , B ∈ F with A ∩ B = {n}.
Since n > 2  , their union omits some other element i. Since F is invariant under
i ,n , we have A − n + i ∈ F . Now B does not intersect A − n + i, a contradiction.

Frankl [1987] proved via shift operators that EK R( , t)-families have size
at most (¾n−−tt) when n is sufficiently large. Forbidding t-stars costs a lot; an
−1
EK R( , 1)-family with no common element has size at most (n¾− ) (n−¾−¾−1 1 ) + 1
1 −
(Hilton–Milner [1967]; see Exercise 14). Bollobás [1986] considered larger t.

11.2.21.* Theorem. (Bollobás [1986]) For 2 ≤ t <  , let F be a t-intersecting

-uniform family of subsets of [n] that is not a t-star. If n > 2  − t, then
t 2
n− t−1 t −t n− t− j
| F | ≤ ( ) + ∑ ( )( ) ( ).
−t−1 j j −t− j
j =1

Proof: Let F be a maximal such family. Since n > 2  − t, we have F

= ([n] ¾ ). If
| A ∩ B| > t for all A , B ∈ F , then adding a -set that differs from a member of F
by one element enlarges the family. Hence there exist A , B ∈ F with | A ∩ B| = t;
let T = A ∩ B. Since F is not a t-star, we find C ∈ F with T
⊆ C.
For 0 ≤ j ≤ t, let F j be the subfamily of members of F omitting j elements of
T . For j ≥ 1, a member of F j must have at least j elements of A and j elements
of B outside T . After picking the j elements of T to omit, we thus have
−t
2
t n− t− j
| F j | ≤ ( )( ) ( ).
j j −t− j
For j = 0, a member of F j contains all of T and needs at least one element of C
outside T , so | F0 | ≤  (¾n−−tt−−11 ). Since | F | = ∑ j =0 | F j | , the bound is complete.
t
500 Chapter 11: Extremal Problems

11.2.22.* Theorem. For 2 ≤ t < ¾ , let F be a t-intersecting ¾-uniform family of

subsets of [n]. If n ≥ 2t ¾ 3 , then | F | ≤ (¾n−−tt), with equality only for a t-star.
Proof: The claim is tivial for t-stars. By Theorem 11.2.21, it suffices to show
t 2
t ¾−t n− t− j n− t n− t−1
∑ ( j )( j ) (¾ − t − j ) < (¾ − t) − [¾ + t(¾ − t)2 ](¾ − t − 1 ).
j =2

2 t− j ¾ − t− j 2 ¾ − t− j
With Sj = ( tj )(¾−j t) (n¾−
− t− j
t− j ), we have Sj = j +1 ( j +1 ) < (j + 1)−3 . Since
S
j +1
n− t− j
∞
∑ j =2(j + 1)−3 < 1, we have ∑ j =2 Sj < 2S2 , so
t

t
n−t−2 t2(¾ − t)5 n − t − 1
∑ Sj < t2(¾ − t)4 (¾ − t − 2) < n− t
(
¾−t−1
).
j =2

t2 (¾ − t)5 1 n− t n− t−1
Since n− t < 14 ¾n−−tt when n > 2t¾ 3 , we obtain ∑tj =2 Sj < (
4 ¾ − t ¾ − t−1
), and then
1 n− t n− t
4 ¾−t < ¾−t − ¾ − t(¾ − t)2 (for n > 2t¾ 3) finishes the proof.

3
At first, “sufficiently large” meant n ≥ t + (¾ − t)(¾t ) ; the threshold in Theo-
rem 11.2.22 is better. Frankl [1976] and Wilson [1984] together showed that the
t-star of ¾-sets is optimal when n ≥ (t + 1)(¾ − t + 1), extending Theorem 11.2.18.
For smaller n, other EK R(¾ , t)-families are larger, such as (⌈(n+[n]
1)/2⌉
) when t = 1
and n < 2 ¾ . Ahlswede–K hachatrian [1997] determined the extreme in all cases
(see Exercise 22), as conjectured by Frankl [1978].
The Erdős–Ko–Rado Theorem for t = 1 follows easily from the Kruskal–
K atona Theorem, as shown by Daykin [1974a]). The earlier result below uses
the same ideas and is stronger (simply set F = G).

11.2.23.* Theorem. (K leitman [1968]) Let ¾ , l , n be positive integers such that

¾ + l ≤ n. If F ⊂ ([n] [n]
¾ ) and G ⊂ ( l ) satisfy x ∩ y
= ∅ whenever x ∈ F and
−1 n−1
y ∈ G, then | F | < (n¾− 1 ) or | G | ≤ ( l−1 ).

Proof: Let F ⊂ (n[n] −¾) consist of the complements of members of F . Note that
n − ¾ ≥ l, since ¾ + l ≤ n. The family G ∪ F is an antichain. Thus the l-shadow of
F is disjoint from G. By the Kruskal-K atona Theorem, the size of the l-shadow
of F is at least l+1 l+2 · · · n−¾ (| F |). Thus | G | + l+1 l+2 · · · n−¾ (| F |) ≤ (nl).
If | F | ≥ (¾n−−11), then also | F | ≥ (nn−
−¾ ), and the first term in the (n − )-binomial
1

n−1
expansion of | F | is (n−¾). Regardless of what else is present, iteration of yields
(n−l 1 ) as a lower bound on the second term in the inequality above. Thus | G | ≤
(nl) − (n−l 1) = (nl−−11 ).

We refer the reader also to Frankl [1978], Wilson [1984], Frankl–Graham

[1989], Ahlswede–K hachatrian [1997], Frankl–Füredi [2012], Godsil–Meagher
[2016, Chapter 1], and Borg–Meagher [2016] for proofs and extensions of the
Erdős–Ko–Rado Theorem. Mubayi–Verstraëte [2016] surveys a related topic.
Section 11.2: Families of Sets 501

CHV ÁTAL’S CONJECTURE

Every star is an intersecting family. The Erdős–Ko–Rado Theorem implies

that in 2[n] with n ≥ 2 ¾ , the star F consisting of all sets with size at most ¾
containing 1 is a largest intersecting family of sets with size at most ¾ ; it si-
multaneously achieves equality in the EKR bound for sets of each size. Chvátal
conjectured a tantalizing generalization.

11.2.24. Definition. A family I is an ideal if x ∈ I and y ⊆ x imply y ∈ I. The

maximal members of an ideal are its bases. A family has the star property
if some largest intersecting subfamily is a star.

11.2.25. Conjecture. (Chvátal [1974]) Every ideal of subsets of [n] has the star
property.

The term “ideal” is traditional for this topic and is used in the rest of this
chapter. Since “ideal” is used in other ways in other areas of mathematics, “down-
set ” has grown in popularity as a alternative term without alternative meanings.
As we observed before Definition 11.2.24, the Erdős–Ko–Rado Theorem im-
plies the conjecture for the ideal consisting of all subsets of [n] with size at most
¾ , where ¾ ≤ n/2. Example 11.2.16 proves it for the family of all subsets of
[n]. Other partial results appear in Chvátal [1974], K leitman–Magnanti [1974],
Berge [1975], and Wang–Wang [1978]; they are surveyed in K leitman [1979].
We show first that an intersecting family cannot contain more than half of
an ideal, because ideals split into pairs of disjoint sets. The proof removes some
disjoint pairs from the “top” of the ideal and applies induction to the rest.

11.2.26. Theorem. (Berge [1976]) If I is an ideal of subsets of [n], then I or

I − {∅} can be partitioned into pairs of disjoint sets.
Proof: (Daykin–Hilton–Miklós [1983], simplified by M. Pelsmajer) We use induc-
tion on | I | , with a trivial basis for size 0 or 1. For | I | > 1, choose a maximal set
T ⊆ [n] such that T is a union of disjoint sets A , B ∈ I. Such a set exists, since
the empty set qualifies.
Let H = { A ⊆ T : A , T − A ∈ I}. The choice of T yields H
= ∅. Observe that
if A ∈ H and A ⊆ C ∈ I, then T − C ⊆ T − A ∈ I. Also, if C
⊆ T , then C ∪ (T − C)
is larger than T , contradicting the choice of T . Hence C ∈ H , by the definition
of H. We conclude that I − H is an ideal.
By construction, H is a subfamily of I partitioned into pairs of disjoint sets
(complements within T). Since I − H is an ideal, we complete the desired partition
by applying the induction hypothesis to I − H.

H
I−H
502 Chapter 11: Extremal Problems

By Theorem 11.2.26, Chvátal’s Conjecture holds under any conditions that

guarantee a star with half of I, such as when some element belongs to all bases
(Exercise 30). Chvátal [1974] proved the conjecture in the case where there is a
fixed linear order L on [n] such that {a1 , . . . , a¾ } ∈ I whenever {b1 , . . . , b¾ } ∈ I
and ai ≤ bi for 1 ≤ i ≤ ¾ . In this case, the star generated by the initial element
in L is a largest intersecting family. It need not contain half of I, as shown when
I is generated by singleton sets.
Several survey papers (including one by this author) mistakenly stated that
what Chvátal [1974] proved was the stronger result by Snevily that we prove
next. Snevily ’s result is a common generalization of the results of Chvátal and
Schönheim. Note the use of + and − for addition and deletion of single elements.

11.2.27. Theorem. (Snevily [1991]) If I is an ideal of subsets of [n], and x is an

element of [n] such that A − a + x ∈ I whenever a ∈ A ∈ I, then the star
generated by x is a largest intersecting family in I.
Proof: Consider a minimal ideal I violating the claim; note that | I | > 1. The fig-
ure below shows various parts of I that we define. Let B be the set of bases of I
not containing x. Let F be a largest intersecting family in I.

B
F− F F
↑
I
H ↓
x H

Let a free base be a base of I that is not expressible as B − a + x with a ∈ B ∈

B. If C is a free base, then let I ∗ = I − {C}; we claim that I ∗ is a smaller ideal
satisfying the hypothesis. This holds because if C is needed as A − a + x for some
A, then A is below some B ∈ B, and then C is below B − a + x, which is in I. This
contradicts that C is a base.
By the induction hypothesis, the star F ∗ generated by x is a maximum inter-
secting family in I ∗ . Since F ∗ ⊆ I, the minimality of I yields | F | > | F ∗ |. Since
F − {C} ⊆ I ∗ , we obtain C ∈ F and | F | = | F ∗ | + 1.
If x ∈ C, then I satisfies the claim, since F ∗ ∪ {C} is a star with center x in I
and has size | F |. Hence we may assume that every free base omits x and belongs
to F . Thus B is the set of free bases, and B ⊆ F .
Now let I be the ideal generated by B. Every member of I − I lies below a
base of the form B − a + x for some B ∈ B and must contain x; otherwise it would
lie below B and belong to I .
Let F = F ∩ I . Every member of F omits x, since it lies in I . To show that
the star generated by x has size at least | F | , we will replace F in F with | F | sets
outside F that contain x. We do this in two steps: first we map the members of
F to their mates in the pairing of I guaranteed by Theorem 11.2.26, and then
we add x to each such mate.
Let H be the set of mates of members of F , and let H = {Y ∪ {x}: Y ∈ H }
Section 11.2: Families of Sets 503

(it does not matter whether H is an ideal). Since F is pairwise intersecting and
Y ∈ H is disjoint from its mate, H ∩ F = ∅. Furthermore, Y ∪ {x} also is
disjoint from the mate of Y in F , so Y ∪ {x} ∈ / F − F .
It remains only to show Y ∪ {x} ∈ I to obtain (F − F ) ∪ H as a star generated
by x with size | F | . Since B ⊆ F and F ∩ H = ∅, we have Y ∈
/ B. Thus Y is non-
maximal in I and has the form A − a for some A ∈ I ⊂ I. The hypothesis then
yields Y ∪ {x} ∈ I.

Subsequent decades have not seen much work on Chvátal’s Conjecture. We

mention Wang [2002], Borg [2011] (a weighted version), K amat [2011], and
Friedgut–K ahn–K alai–Keller [2018].
In addition to considering families closed under taking subsets, we can also
consider families closed under taking unions. This leads to a conjecture at least
as notorious as that of Chvátal. A family of sets is union-closed if the union of
every two members of the family also belongs to the family.

11.2.28. Conjecture. (Union-Closed Sets Conjecture) If F is a finite union-

closed family of finite sets with | F | ≥ 2, then some element appears in at least
half the members of F .

The conjecture is generally attributed to Frankl in 1979 and first appeared

in print in 1985. Early special cases were proved by Sarvate–Renaud [1989] (Ex-
ercise 32), Poonen [1992], and Johnson–Vaughan [1998]. Bruhn–Schaudt [2015]
presents a thorough survey. The fundamental nature of the conjecture is shown
by reformulations such as a version in terms of bipartite graphs in Exercise 33.

SUNFLOWERS (optional)

In the problems we have considered thus far, our notion of “star ” was a family
of sets having a common element. Now we consider a more restricted notion.

11.2.29. Definition. A -system or sunflower is a family of sets in which any

two members have the same intersection.

In a 2-uniform family (a graph), the sunflower condition reduces to a star or

a family of disjoint edges, where the common intersection and pairwise intersec-
tion is always the empty set. The original term is -system, but “ ” does not
evoke the meaning in any way, so we will use the term “sunflower ”. An early use
of this term is in Alon–Frankl–Lovász [1986].

11.2.30. Theorem. (Sunflower Theorem; Erdős–Rado [1960]) Every ¾-uniform

family with more than ¾ !(s − 1)¾ members contains a sunflower of size s, and
there is a family of size (s − 1)¾ having no such sunflower.
Proof: We use induction on ¾ . For ¾ = 1, the sets are disjoint, and we have more
than s − 1 of them.
For ¾ ≥ 2, let F be the ¾-uniform family, and let H be a maximal subfamily
consisting of disjoint sets. If | H | ≥ s, then H is the desired sunflower. Otherwise,
504 Chapter 11: Extremal Problems

let B be the union of the sets in H ; note that | B| ≤ ¾(s − 1). Every member of
F intersects B. By the Pigeonhole Principle, some element of B lies in at least
| F | /| B| members of F . We compute
| F | ¾ !(s − 1)¾
> = (¾ − 1)!(s − 1)¾−1 .
| B| ¾(s − 1)
Deleting x from the sets containing it yields a (¾ − 1)-uniform family F with
more than (¾ − 1)!(s − 1)¾−1 members. By the induction hypothesis, F contains a
sunflower of size s. Returning x to its sets yields a sunflower in F of size s.
For the construction, let X 1 , . . . , X ¾ be disjoint sets of size s − 1. Let F
be the family of all transversals (¾-sets consisting of one element from each of
X 1 , . . . , X ¾). If F contains a sunflower A1 , . . . , A s , then some element x belongs
to exactly one of A1 , . . . , A s . By symmetry, we may assume x ∈ A1 ∩ X 1 . Since
the pairwise intersections are the same, the sets A1 , . . . , A s must intersect X 1
in distinct elements. Since | X 1 | = s − 1, this is impossible.

It is not known whether the bound ¾ !(s − 1)¾ in Theorem 11.2.30 is sharp.


11.2.31. Conjecture. (Erdős–Rado [1960]) For ¾ ∈ there is a constant c¾ such
that every ¾-uniform family F with | F | > cs¾ contains a sunflower of size s.

Kostochka [2000] surveyed results on this and related problems. Abbott–

Hanson–Sauer [1972] solved the case ¾ = 2. Erdős [1975] offered $1000 for set-
tling the case ¾ = 3. For the case ¾ = 3, a more specific conjecture was posed.

11.2.32. Conjecture. (Erdős–Szemerédi [1978]) There is a constant c, less than

2, such that every family of subsets of [n] containing no sunflower of size 3
has at most cn sets.

By results of Erdős–Szemerédi [1978], Conjecture 11.2.31 implies Conjec-

ture 11.2.32. Using algebraic methods like those in Chapter 15, Naslund–Sawin
[2016] proved Conjecture 11.2.32 using a constant c that is just above 3/22/3 .
The most famous applications of the Sunflower Theorem (and various modifi-
cations) are to complexity theory; see Razborov [1985] and Alon–Boppana [1987].
More recently, there is an application to matrix multiplication (Alon–Shpilka–
Umans [2013]). Alon–Pudlak [2001] used the method to prove an explicit lower
bound for off-diagonal Ramsey numbers: There is a fixed constant c such that for
every fixed s and sufficiently large m, the construction produces a graph having
neither an s-clique nor an independent
√ m-set. Thus R(s , m) exceeds the order of
c log s/log log s
the constructed graph, which is m .

ENTROPY (optional)

The notion of entropy is quite different from the rest of this section. We
include it here because it can yield extremal results on enumerative problems
related to topics studied in this chapter. Our treatment draws on Alon–Spencer
[2008] and an excellent tutorial by Galvin [2014]; see also Radhakrishnan [2003].
Section 11.2: Families of Sets 505

Imagine an experiment having two possible outcomes, one with probability p

and one with probability 1 − p. If p = 1/2, then we can treat the “information”
gained by knowing the outcome as one unit, since it takes one bit to specify which
outcome occurred. If p ∈ {0 , 1}, then there is no information in the outcome,
since we know before the experiment which will occur.
Now consider three possible outcomes, one with probability 1/2 and two with
probability 1/4. Using natural binary lists to represent the outcomes, we specify
the outcome in one bit with probability 1/2 and in two bits with probability 1/2,
so the expected information in the outcome is 3/2. (We will define expectation
formally in Chapter 14; here we just use the idea as informal motivation.)

11.2.33. Definition. A finite probability distribution p is a list (p1 , . . . , pn) of

nonnegative values with sum 1. A random variable X with distribution
p takes on n possible values, the ith occurring with probability pi . With lg
denoting the base-2 logarithm, the entropy of X is a value H(X) defined by
H(X) = − ∑ i=1 pi lg pi .
n

We think of entropy intuitively as the expected information in the outcome of

an experiment with distribution p, motivated by the fact that when the probabil-
ities are powers of 1/2, the entropy is the expected length of the natural binary
word specifying the outcome (Exercise 36). The function can also be obtained as
the unique function satisfying certain natural axioms.
The relevance of entropy for extremal problems in enumeration is based on
the following lemma.

11.2.34. Lemma. The entropy H(X) of a random variable X with n outcomes is

largest only when the outcomes are equally likely, and then H(X) = lg n.
Proof: When the outcomes have the distribution p1 , . . . , pn , note that H(X) is
a weighted average of the values of lg(1/pi). For a strictly convex function º ,
any weighted average of values º (x i) , . . . , º (x n) is less than the value at the same
weighted average of the arguments x1 , . . . , x n , unless x1 = · · · = x n . (In par-
ticular, ë º (x)ë++ÞÞº (y) < º ( ë ëx+
+Þy
Þ ) when x
= y. The general statement is known as
Jensen’s Inequality.)
Since lg is strictly convex, when X is not uniform we have
n n
H(X) = ∑ pi lg(1/pi) < lg(∑ pi(1/pi)) = lg n.
i=1 i=1

When X is uniform, H(X) = ∑ pi lg n = lg n.

n
i=1

The point is this: if we have a random variable X defined over some set S,
and we can compute H(X) without knowing | S| , then by Lemma 11.2.34 we have
|S| ≥ 2 H(X) , with equality if X actually is uniform over S.
We need some additional notions about random variables. When X and Y
are random variables, with outcomes respectively in S and T , we denote by
(X , Y) the joint random variable; it has outcomes in S × T , with (X = i) =
∑ j ∈ T ((X , Y) = (i , j)) and (Y = j) = ∑i∈S ((X , Y) = (i , j)).
506 Chapter 11: Extremal Problems

11.2.35. Lemma. If X and Y are random variables, then H(X) ≤ H(X , Y).
Proof: The joint distribution breaks events for X into subevents; intuitively, this
cannot reduce the expected information in the outcome. In the entropy computa-
tion, breaking an outcome with probability p+ p into outcomes with probabilities
p and p replaces −(p + p ) lg(p + p ) with − p lg p − p lg p , which is larger since
− lg is a decreasing function. Repeating this leads us from H(X) to H(X , Y).

More importantly, H(X , Y) ≤ H(X) + H(Y). Exercise 37 requests a direct

proof; we prove this and a generalization via other useful tools (see Galvin [2014]).

11.2.36. Definition. The conditional probability of an event F given event

E, written (F | E), is defined by (F | E) = (F & E)/(E). Write X | E
to mean the random variable X normalized by the condition that E occurs.
Let p(x | E) denote the probability of X = x given E; that is, p(x | E) =
(X = x & E)/(E). This yields a distribution, since ∑ x p(x | E) = 1. Its en-
tropy is given by H(X | E) = − ∑ x p(x | E) lg p(x | E). When Y is a random
variable and E is the event Y = y, taking the weighted average of these dis-
tributions over the distribution of Y yields the conditional entropy
H(X | Y) = (H(X | (Y = y))) = − ∑ p(Y = y) ∑ p(x | Y = y) lg p(x | Y = y).
y x

Intuitively, the information gained in knowing the outcome for the joint vari-
able (X , Y) is the information gained in knowing X plus the information gained
in knowing Y given that we already know X . We formalize this.

11.2.37. Lemma. (Chain Rule for entropy) H(X , Y) = H(X) + H(Y | X).
Proof: For simplicity, we just write the value for the event that the variable
takes that value. Since (Y = y | X = x) = (X = x & Y = y)/(X = x), we thus
write p(y | x) = p(x , y)/p(x). Note also ∑ y p(y | x) = 1. We compute

H(X , Y) − H(X) = − ∑ ∑ p(x , y) lg p(x , y) + ∑ p(x) lg p(x)

x y x

= − ∑ p(x) ∑ p(y | x) lg[p(x)p(y | x)] + ∑ p(x) ∑ p(y | x) lg p(x)

x y x y

= − ∑ p(x)(∑ p(y | x) lg p(x)p(y | x) − p(y | x) lg p(x))

x y

= − ∑ p(x)(∑ p(y | x) lg p(y | x)) = H(X | Y).

x y

Another intuitive notion is that the information gained by knowing X is at

least that gained by knowing X subject to the restriction of first knowing Y .

11.2.38. Lemma. (Deconditioning) H(X | Y) ≤ H(X).

Proof: Note that the quantities p(y)p(x | y) and p(x)p(y | x) both equal p(x , y).
Also, since ∑ y p(y | x) = 1 for fixed x, we can apply Jensen’s Inequality (concavity)
to the logarithm function to obtain the inequality below. We compute
Section 11.2: Families of Sets 507

1
H(X | Y) = − ∑ p(y) ∑ p(x | y) lg p(x | y) = ∑ p(x) ∑ p(y | x) lg
y x x y
p(x | y)
p(y | x) p(y)
≤ ∑ p(x) lg ∑
p(x | y) ∑
= p(x) lg ∑
x y x y
p(x)
1
= ∑ p(x) lg = H(X).
x
p(x)

11.2.39. Lemma. (Subadditivity) H(X , Y) ≤ H(X) + H(Y).

¾
More generally, H(X1 , . . . , X ¾) ≤ ∑ i=1 H(X i).
Proof: The proof is immediate from the Chain Rule (Lemma 11.2.37) and Decon-
ditioning (Lemma 11.2.38), with induction on ¾ for the generalization.

Already we can use entropy to give a good bound on the number of subsets of
[n] with size at most n. For 0 ≤ ≤ 1, let H( ) = − lg − (1 − ) lg(1 − ); note
that H( ) is the entropy of a 0 , 1-variable with success probability .

11.2.40. Theorem. Fix ∈ (0 , .5). For n ∈  ,

)n n
2 H( n
√ ≤ ∑ ( ) ≤ 2 H( )n
.
2 n (1 − ) i=0
i

Proof: (Massey [1974]) The lower bound is the lower bound for the top term from
Exercise 2.3.14, using Stirling ’s Formula. For the upper bound, let F be the
family of subsets of [n] with size at most n. Choose a member of F uniformly at
random, generating the random binary n-tuple X where X i = 1 if and only if i is
in the chosen set. By Lemma 11.2.34, it suffices to show H(X) ≤ H( )n.
By subadditivity and symmetry, H(X) ≤ ∑ i=1 H(X i) = nH(X 1). Since X 1 is
n

a 0 , 1-variable, we have H(X 1 ) = H(p), where p = (X i = 1). An element i ∈ [n]

appears in the fraction /n of the -element subsets of [n]. Since we choose only
from F , we have  ≤ n for the restriction to each  , so p ≤ . Since H(p) is
monotone over [0 , .5] and ≤ .5, we have H(p) ≤ H( ), completing the proof.

Our next application is in extremal set theory, from K leitman–Shearer–

Sturtevant [1981]. We begin with a corollary of subadditivity.

11.2.41. Corollary. Let F be a family of subsets of [n]. If pi is the fraction of

sets in F that contain i, for all i ∈ [n], then | F | ≤ 2 i H(pi) .
Proof: Choose a member of F uniformly at random, generating the random bi-
nary n-tuple X where X i = 1 if and only if i is in the chosen set. Note that
(X i = 1) = pi for all i. By Lemma 11.2.39, H(X) ≤ ∑i=1 H(pi). The bound on | F |
now follows from Lemma 11.2.34.

11.2.42. Theorem. If F is a family of -subsets of [n] in which no two pairs of

sets have the same intersection, then | F | ≤ 2 .8114+o() .
Proof: The condition implies that the intersections of any one set with the others
are distinct, and hence | F | ≤ 2 . To improve the bound, let pi be the fraction of
508 Chapter 11: Extremal Problems

the sets in F that contain i, and let m = | F | . Let G be the family of all pairwise
intersections of sets in F . By the condition on F , we have | G | = (m2 ). Element i
pi m
appears in ( 2 ) of the members of G; that is, in a proportion less than p2i . By
Corollary 11.2.41, | G | ≤ 2 i H(pi ) .
2

Since all sets in F have size ¾ , we have ∑ pi = ¾ . Hence dividing by ¾ gives a

distribution, and we can apply Jensen’s Inequality to the function H(p2)/p that
is concave for p ∈ [0 , 1]. We compute
pi H(p2i ) H(p2)
∑ H(p2i ) = ¾ ∑ ¾ pi
≤¾
p
,

where p = ∑ p¾i pi . The maximum of H(p2)/p occurs at p ≈ .4914, where the value
is about 1 .6228. Thus (m 2 ≤ 2
) 1 .6228¾
, so roughly | F | ≤ 2 .8114¾+o(¾) .

The next lemma is a substantial generalization of subadditivity, which is just

the special case where F consists of the n singleton sets and r = 1. A proof first
appeared in Chung–Frankl–Graham–Shearer [1986].

11.2.43. Lemma. (Shearer ’s Lemma) Let F be a list of subsets of [n]. If each

i ∈ [n] appears in at least r members of F , then the joint distribution of ran-
dom variables X 1 , . . . , X n satisfies
1
H(X 1 , . . . , X n) ≤
r ∑ H(X i1 , . . . , X i¾ ).
{i1 ,... ,i¾ }= A∈ F

Proof: (Radhakrishnan [2003], with Llewellyn) Index the elements of A ∈ F as

i1 , . . . , i¾ in increasing order. Using the Chain Rule and Deconditioning,
¾ ¾
H(X i1 , . . . , X i¾ ) = ∑ H(X ij | (X is : s < j)) ≥ ∑ H(X ij | X 1 , . . . , X ij −1).
j =1 j =1

Summing over all A ∈ F yields a lower bound on ∑{i1 ,... ,i¾ }= A∈ F H(X i1 , . . . , X i¾ ).
In the lower bound, each term of the form H(X i | X 1 , . . . , X i−1) appears at least r
times. Hence
n

∑ H(X i1 , . . . , X i¾ ) ≥ r ∑ H(X i | X 1 , . . . , X i−1) = rH(X 1 , . . . , X n) ,

{i1 ,... ,i¾ }= A∈ F i=1

where the final equality is by the Chain Rule.

The next application requires a combinatorial version of Shearer ’s Lemma.

11.2.44. Lemma. (Combinatorial Shearer ’s Lemma) Let F be a list of subsets

of [n] such that each i ∈ [n] appears in at least r members of F . Given a
family G of subsets of [n], let t A(G) = {B ∩ A: B ∈ G}. The following holds:

| G | ≤ ∏ | t A(G) |1/r .
A∈ F
Section 11.2: Families of Sets 509

Proof: Choose X uniformly at random from G, so | G | ≤ 2 H(X). Treat X as its

incidence n-tuple X 1 , . . . , X n , with X i indicating whether i ∈ X . By Shearer ’s
Lemma, H(X) ≤ 1r ∑ A∈ F H(X A), where X A = {X i : i ∈ A}.
To bound H(X), we use Lemma 11.2.34 to bound H(X A) by lg | S| , where S is
the set of possible values of X A . This set is the set of subsets of A that can arise
when X is chosen from G, which is precisely t A(G). Thus

| G | ≤ 2 H(X) ≤ 2( ∑ A∈ F H(X A)) /r = 2( ∑ A∈ F lg|t A(G)|) /r = ∏ |t A(G)|1/r .

A∈ F

We now consider a restricted family of sets whose pairwise intersections have

size at least 3. The sets are restricted to be edge sets of graphs with vertex set [n],
and the intersections must contain three edges forming a triangle. By taking all
the graphs containing a fixed triangle, a family of size 2(2)−3 can be formed. Ellis–
n

Filmus–Friedgut [2012] later proved that this family is extremal, as conjectured

originally by Simonovits and Sós in 1976. We present the use of Shearer ’s Lemma
in Chung–Frankl–Graham–Shearer [1986] to obtain an easy upper bound.
11.2.45. Theorem. Let be a family of graphs with vertex set [n]. If any two
graphs in share a triangle, then | | ≤ 2(2)−2 .
n

Proof: View as a family of subsets of ([n] 2

). In order to use Lemma 11.2.44 to
bound | | , we introduce a family of graphs. We let consist of all copies of
K ⌊n/2⌋ + K ⌈n/2⌉ with vertex set [n]. The size s of each member of is (⌊n/2⌋
2
) + (⌈n/2⌉
2
);
note that s ≤ 2 (2).
1 n

The total size of all members of is s | |. By symmetry, each vertex pair lies
in the same number of members of , so it occurs in exactly s | | /(n2) such graphs.
This is the value of r in Lemma 11.2.44.
For each A ∈ , the family t A() is an intersecting family. In particular, any
B , B ∈  share a triangle, but the complement of A has no triangle, so B ∩ A and
B ∩ A must share an edge of that triangle.
Since t A() is an intersecting family of subsets of A, and | A| = s, we conclude
|t A()| ≤ 2 s−1 (recall Example 11.2.16). Since s | | /r = (n2) and s ≤ 12 (n2), using
Lemma 11.2.44 yields

|| ≤ ∏ | t A(G)|1/r ≤ (2 s−1)| |/r = 2(2)−(2)/s ≤ 2(2)−2 .

n n n

A∈

We have presented here only the basic ideas of using entropy to prove bounds
on extremal combinatorial problems: the relationship between entropy and the
number of outcomes, subadditivity, and Shearer ’s Lemma. Similar applications
appear in Exercises 38–41. Meanwhile, Galvin [2014] presents an impressive va-
riety of applications and pointers to applications of entropy. We describe a few
below to encourage readers to explore more deeply.
11.2.46. Example. Antichains of subsets of [n]. By taking families of n/2-sets
from [n], one can form 2(⌈n/2⌉) antichains. K leitman [1969] was the first to show
n

n
that the logarithm of the total number of antichains is asymptotic to (⌈n/2⌉) (see
Section 12.2). Pippenger [1999] showed how to prove this using the properties of
entropy discussed here, though his lower-order terms were not as sharp.
510 Chapter 11: Extremal Problems

11.2.47. Example. Maximizing the permanent. The permanent per(A) of an n-

n
by-n matrix A is the sum of ∏i=1 ai , (i) over all permutations of [n] (Chapter 15).
For the matrix of a bipartite multigraph, this counts the perfect matchings.
An X , Y -subgraph of K n ,n trivially has at most ∏ x∈ X d(x) perfect matchings.
In 1963, Minc conjectured that a binary matrix with row-sums d1 , . . . , d n has
n
permanent at most ∏i=1 (d i !)1/d i , reducing the bound by a factor almost e n . When
d divides n, equality holds for the graph (n/d)K d ,d . Brègman [1973] proved the
conjecture. Radhakrishnan [1997] gave a proof using subadditivity of entropy.
See Alon–Friedland [2008] for the K ahn–Lovász extension to general graphs.

11.2.48. Example. Embeddings of H into G. An embedding of a graph H into a

graph G maps V(H) injectively into V(G) so that all edges of H map into edges
of G. Let H (m) denote the maximum number √ of embeddings of H into a graph
with m edges. It is easy to show K3 (m) ≤ 2 2m3/2 (Exercise 42), sharp by K √2m .
∗
Alon [1981] proved the existence of a constant c such that H (m) ≤ cm (H) for
all H and m. Here ∗(H) is the maximum sum of nonnegative vertex weights such
that each edge has total weight at most 1 (note that ∗(K 3) = 3/2). Friedgut–
K ahn [1998] proved Alon’s result using Shearer ’s Lemma.

EXERCISES 11.2


11.2.1. (−) Fix m , ¾ ∈ . When n is sufficiently large, how can a family of m sets of size ¾
be chosen in [n] to maximize the size of the shadow of the family?
11.2.2. (−) Permutations (in word form) intersect if they agree in some position. Find the
maximum size of an intersecting family of permutations of [n]. (Deza–Frankl [1977])
11.2.3. (−) Does every maximal intersecting family in 2[n] have size 2 n−1 ?
11.2.4. (−) Let I be an ideal of subsets of [n]. Prove that every element of [n] appears in
at most half the members of I.
11.2.5. (−) Use the Sunflower Theorem to prove that every graph with more than 2(¾ − 1)2
edges contains a matching of size ¾ or a star of size ¾ .
11.2.6. Determine the maximum size of a family of subsets of [n] such that no element
belongs to at least two sets in the family and is omitted by at least two sets in the family.
11.2.7. Use the Kruskal–Katona Theorem to prove that the size of the shade of a family
of m elements of ([n]
¾ ) is minimized by the family consisting of the first m elements of the

¾ ) in which x < y if x i > yi in the highest coordinate where they differ.

order on ([n]


11.2.8. (♦) Let G be an n-vertex graph with m edges. For each ¾ ∈ , determine the max-
imum number of ¾-cliques in G, and construct an example achieving equality. (Bollobás)
11.2.9. Explain how to modify the shift-operator proof of the Kruskal–Katona Theorem
(Theorem 11.2.11) to give a direct proof of Theorem 11.2.12 that the shadow of a ¾-uniform
family with size (u¾) has size at least (¾−u 1 ). (Frankl [1984])

11.2.10. Let F be a ¾-uniform family of size (u¾) . Refine the proof of Theorem 11.2.12 to
show that the shadow of F has size exactly (¾−u 1 ) if and only if u is an integer and F consists
of all ¾-sets in a set of size u. (Lovász [1979])
Exercises for Section 11.2 511

11.2.11. Prove that the list 0 , . . . , 0 , º , º ¾ +1 , . . . , º , 0 , . . . , 0 is realizable as ºi =

## F ∩ ([n])## for some antichain F of subsets ¾of [n] l
### i # #
#
if and only if

º¾ + ¾ +1 (¾ +1 + ··· + 
l −1 ( l −1 + l(l))) ≤ (¾n).
(Clements [1973], Daykin–Godfrey–Hilton [1974])

11.2.12. Fix r ≥ 2. For what values of n is it possible to color every square in an n-by-n
grid with one of r colors so that , for all i , j ,  between 1 and n with i
= j and j
=  , the
square in row i and column j is assigned a different color from the square in row j and
column  . (Hint: Use Sperner ’s Theorem.) (Propp [1998])
11.2.13. Define a rising antichain to be an antichain of subsets of [n] whose sizes are
distinct. Let two rising antichains be equivalent if one is obtained from the other by com-
plementing all sets and/or relabeling the elements of [n].
(a) Determine the maximum size of a rising antichain of subsets of [n].
(b) Prove that when n ≥ 5, the largest rising antichains are all equivalent to each
other. (Bey–Griggs [2002])
11.2.14. For n > 2  , construct an intersecting family of -subsets of [n] with size (¾n−−11) −
(n−¾−¾−1 1) + 1 that is not a star. (Comment: Hilton–Milner [1967] proved that none is larger.)
11.2.15. For n ≥ t ≥ 1, let q = ⌈ (n + t)/2⌉ . Prove that a largest t-intersecting family in 2 n
has size ∑i= q ( ni) when n + t is even and ∑i= q (ni) + (nq−−11 ) when n + t is odd. (Katona [1964])
n n

11.2.16. Let F be a family of 4-sets of [n] that pairwise intersect at most twice. Prove the
existence of S ⊂ [n] with | S| ≥ ⌈ (6n − 6)1/3 ⌉ that contains no member of F . (Adrian [1991])
11.2.17. Let F be an intersecting family of subsets of [n] whose members have size at most
¾
 . Prove that | F | ≤ ∑i=1 (ni−−11 ) , with equality only for stars if  < n− 1.
11.2.18. (♦) Prove that the largest antichain of subsets of [n] consisting of pairs of com-
n−1
plementary sets has size 2(⌈n/2⌉). (Bollobás [1973])
11.2.19. Let  and  be simple -words from [n] (lists of  distinct elements). Say that 
and  intersect if  i =  i for some i. Prove that the maximum size of a pairwise intersecting
family of simple -words from [n] is (n − 1)!/(n − )!. (see Deza–Frankl [1978], Lovász–
Nešetril–Pultr [1980])
11.2.20. Let F be a family of subsets of [n] such that each member has size at least l and
each pair of members have at most  common elements, where l ≤  . Prove that | F | is
maximized uniquely by F = {A ⊆ [n]: l ≤ | A| ≤  + 1}.
11.2.21. Let A be an intersecting antichain of subsets of [n] with | x| ≤ n/2 for all x ∈ A.
−1
−1
Prove that ∑ x∈ A (|nx|− 1)
≤ 1. (Hint: Assign weight | x|−1 to each x ∈ A and prove that the
total weight on members of A appearing as consecutive strings in a single cyclic permuta-
tion is at most 1.) (Bollobás [1973], Greene–Katona–Kleitman [1976])
11.2.22. For 0 ≤ r ≤  − t, let Sr be the family of -subsets of [n] that contain at least t + r
of the smallest t + 2r elements.
(a) Show that Sr is a t-intersecting family.
(b) Prove that | Sr | = ∑i=0 (tt++r2r n− t−2r
r
+i)(¾ − t−r−i).
(c) Prove that | S1 | > | S0 | if and only if n < ( − t + 1)(t + 1). Thus the t-star S0 is not a
largest EKR( , t)-family when n < ( − t + 1)(t + 1). (Frankl [1978]) (Comment: Ahlswede–
Khachatrian [1997] proved that Sr is a largest EKR( , t)-family when ( − t + 1)(2 + rt−+11 ) ≤
n < ( − t + 1)(2 + t−r 1 ). A proof appears in Engel [1997, pp. 50–60].)
512 Chapter 11: Extremal Problems

11.2.23. (♦) Let A1 , . . . , A n be r-sets such that | Ai ∩ Aj | ≤ ¾ for distinct i , j ∈ [n]. Prove
### n A ### ≥ nr 2/[r + (n − 1)¾]. Interpret for proper edge-colorings of K . (Corrádi [1969])
# #⋃ i = 1 i #
# n

11.2.24. Let X 1 , . . . , X n be independent 0 , 1-random variables with (X i = 1) = p ≥ .5.

Let Z be a convex combination of {X i }; that is, Z = ∑ i X i with i ≥ 0 and ∑ i = 1.
Prove that P(Z ≥ .5) ≥ p when no subset of { i } sums to .5. (Hint: Use the EKR Theorem
to bound the number of -subsets of { i } whose sum exceeds .5. Comment: The claim holds
without the restriction that no subset of { i } sums to .5, but eliminating that restriction
requires a limit argument.) (Liggett [1977])
11.2.25. Let F and G be ideals in 2 × 3. Show that it is not always possible to mini-
mize | F ∩ G| by letting F be the first | F | vectors in lexicographic order and letting G be
the lexicographically first | G| vectors when the components are read in the opposite order.
(Daykin–Kleitman–West)
11.2.26. Find the largest  such that every -coloring of the subsets of [n] makes all of
{A , B , A ∪ B , A ∩ B} the same color for some distinct sets A and B. Solve the same problem
for n = 6 when A − B and B − A must be nonempty. (Tomescu [1987])
11.2.27. Let I be an ideal of subsets of [n], and let I = { A: A ∈ I}. Prove that there exists
a bijection  : I → I such that A ⊆  (A) for all A ∈ I. (Erdős–Herzog–Schönheim [1970],
Marica [1971], Daykin–Hilton–Miklós [1983])
11.2.28. (♦) Use Berge’s Theorem (Theorem 11.2.26) to prove the following:
(a) The average size of the sets in any ideal of subsets of [n] is at most n/2.
(b) The maximum size of an intersecting family of subsets of [n] whose complements
also form an intersecting family is 2 n−2 .
11.2.29. Let F be an intersecting family of subsets of [n] whose complements also form an
intersecting family. Use the shift operators 1 , j and Theorem 11.2.27 to prove | F | ≤ 2 n−2 .
11.2.30. Special cases of Chvatál’s Conjecture. (Schönheim [1976])
(a) Prove the conjecture for ideals whose bases have a common element.
(b) Prove the conjecture for ideals with two bases.
(c) Use Theorem 11.2.27 to give a proof without the Erdős–Ko–Rado Theorem for the
ideal consisting of all subsets of [n] with size at most  .
11.2.31. For a set B ⊆ [n] and a family F of subsets of [n], the translate of F by B, writ-
ten F(B), is {A B: A ∈ F}, where  denotes symmetric difference.
(a) Prove that the translate of an ideal by {x} has the star property. (B. Reiniger)
(b) Prove that the translate of an ideal I by a set B is an ideal if and only if B is con-
tained in every element of I that is maximal (contained in no other). (P. Wenger)
(c) Let I be an ideal. Prove that every translate of I by a nonempty set B contains a
star at least as large as the largest intersecting family in I. (Snevily)
11.2.32. Prove that the Union-Closed Sets Conjecture (Conjecture 11.2.28) is true for
families having a member with size at most 2. (Sarvate–Renaud [1989])
11.2.33. (♦) Prove each statement below equivalent to the Union-Closed Sets Conjecture.
(a) If a finite family F with | F | ≥ 2 is closed under taking intersections, then ⋃ A∈ F A
has an element belonging to at most half of the members of F .
(b) Every nontrivial X , Y -bigraph G contains in both X and Y a vertex that lies in at
most half of the maximal independent sets of G. (Bruhn–Charbit–Telle [2013])
11.2.34. Given a graph G, let H be the graph with vertex set (V(G)
¾ ) such that vertices A
and B are adjacent if and only if G has an edge uv with u ∈ A − B and v ∈ B − A. Suppose
that every induced subgraph of G having at most s vertices has fewer than (2s ) edges. Use
the Sunflower Theorem to prove (G) ≤  !(s − 1)¾ .
Section 11.3: Matroids 513

11.2.35. For n − s + 1 < ¾ ≤ n, prove that the family ([n]

¾ ) has no sunflower of size s.

11.2.36. Let X be a random variable such that for the ith outcome, the probability pi that
it occurs is a power of 1/2. Associate distinct binary words with each outcome so that the
word associated with the ith outcome has length − lg pi . Show that the expected length of
the word specifying the outcome is the entropy H(x).
11.2.37. For the entropy of the joint variable (X , Y), prove H(X , Y) ≤ H(X) + H(Y) di-
rectly from the definition. (Hint: Express H(X) + H(Y) − H(X , Y) as the double sum
∑ ∑ (X = i)(Y = j) i , j lg i , j , where i , j = (X
[(X ,Y)=(i , j)]
=i)(Y = j) .)

11.2.38. Example 10.1.5 used the Pigeonhole Principle to show that lg n bipartite sub-
graphs are needed to cover K n . Use entropy to generalize: The number of -partite sub-
graphs needed to cover K n is at least (lg n)/(lg ).
11.2.39. (♦) A distinguishing family for [n] is a family of subsets of [n] such that for
any two subsets A , B ⊆ [n], there is some D ∈ such that | A ∩ D|
= | B ∩ D|. Let  (n) be
the minimum size of such a family.
(a) Use entropy to prove  (n) ≥ n/lg(n + 1). (Comment: There is an improvement by
an asymptotic factor of 2; see Erdős–Rényi [1963] and Moser [1970]. That matches the up-
per bound  (n) ≤ (1 + o(1))2n/lg n by Lindström [1965] and Cantor–Mills [1966].)
(b) Given n coins of the same weight , except that some counterfeit coins weigh a fixed
smaller amount , prove that  (n) is the minimum number of subsets of the coins whose ac-
curate weighing determines the counterfeit coins. The outcome of a weighing determines
how many in the weighed set are light. (Söderberg–Shapiro [1963])
11.2.40. (♦) Let S be a family of n points in d . Let nj be the number of distinct points
obtained by setting the jth coordinates of these points to 0 (projecting S on the hyper-
plane xj = 0). Prove nd−1 ≤ ∏ j =1 nj , and show that this is sharp for all d. (Hint: Use
d

Lemma 11.2.43, Shearer ’s Lemma. Comment: Without entropy, Loomis–Whitney [1949]

generalized to a measurable body B in d . With Bj being the projection of B on the hyper-
plane xj = 0, they proved Vol(B)d−1 ≤ ∏ j =1 (Vol(Bj )). The general case can be obtained by
n

approximating B as a union of axis-parallel cubes centered at points of a fine grid.)

11.2.41. (♦) Let G be a family of subsets of [n] such that any two members of G have  con-
secutive common elements. Determine the maximum size of G. (Hint: Design a suitable
family F in order to apply the combinatorial version of Shearer ’s Lemma, Lemma 11.2.44.)
(Comment: Note that the families in Exercise 11.2.15, where more general -intersecting
families are allowed, are much larger.)
√
11.2.42. Let G be a graph with m edges. Prove that there are at most 2 2m3/2 embed-
dings of K3 into G (see Example 11.2.48).

11.3. Matroids
The defining properties of matroids are general enough to occur in many con-
texts and special enough to yield rich combinatorial structure. Many elegant re-
sults from graph theory, linear algebra, and elsewhere generalize in the theory
of matroids. These include the greedy algorithm for minimum spanning trees,
min-max relations for systems of distinct representatives, dimension properties
of vector spaces, and duality properties of planar graphs. When a theorem on a
special class holds for all matroids, it immediately yields results for other classes.
514 Chapter 11: Extremal Problems

Matroids were independently introduced by Whitney [1935] (to study planar

graphs) and by Nakasawa [1935, 1936] (to study linear dependence; reprinted and
translated in Nishimura–Kuroda [2009]). They were reinvented by van der Waer-
den [1937] generalizing independence in vector spaces, and they also arose (from
a structural viewpoint) in the theory of geometric lattices (MacLane [1936]). The
first modern textbook on the subject was Welsh [1976], later Oxley [1992, 2011].
In this brief treatment, we focus on the application of matroids to extremal
problems, leaving deeper structural aspects and more subtle matroidal aspects
of graphs to a more advanced book. We start from the notions of ideals and an-
tichains defined in the preceding section.
In many mathematical contexts, sets that avoid conflict are called “indepen-
dent ”. Inherently, subsets of independent sets (and the empty set) are indepen-
dent. Thus the family of independent sets is closed under taking subsets. The
same family can be specified by its antichain of maximal independent sets, its an-
tichain of minimal nonindependent sets, or other aspects we will describe later.
To obtain the elegant behavior of matroids, we need an additional property. A
restriction on the family of independent sets can be translated into a correspond-
ing restriction on some other specification of the system. Because we can specify
the system in many ways, we have many equivalent definitions of matroids.

HEREDITARY SYSTEMS AND EX AMPLES

Before discussing the added properties that yield matroids, we begin with
the more general notion of hereditary systems, familiar from our study of ideals.
Given a finite set E of elements, we use 2 E to denote the family of subsets of E,
ordered by inclusion; it has size 2| E| . We view a nonempty ideal in 2 E and the
other ways of specifying it as a single structure. A matroid will then be such a
structure that satisfies one of various equivalent additional constraints.

11.3.1. Definition. A hereditary system M on E consists of a nonempty ideal

I M in 2 E and all ways of specifying I M . The ideal I M is the family of indepen-
dent sets of M ; the other subsets of E are dependent. The bases B M are
the maximal independent sets. The circuits C M are the minimal dependent
sets. The rank function r M (X) of M assigns to each X ⊆ E a value called its
rank, which is the maximum size of a member of I M contained in X .

11.3.2. Example. On the left below we sketch the inclusion order on subsets of
E, with the full set E at the top and the empty set at the bottom. We sketch the
relationships among the independent sets, bases, circuits, and dependent sets of
a hereditary system. The bases are the maximal elements of I, and the circuits
are the minimal elements not in I. Always ∅ ∈ I. If every set is independent,
then there is no circuit, but every hereditary system has at least one base.
On the right we obtain a hereditary system from a multigraph with three
edges. The independent sets are the acyclic edge sets. The only dependent sets
are {1 , 2} and {1 , 2 , 3}, the only circuit is {1 , 2}, and the bases are {1 , 3} and
{2 , 3}. The rank of an independent set is its size. For the dependent sets, we
have r({1 , 2}) = 1 and r({1 , 2 , 3}) = 2.
Section 11.3: Matroids 515

E •
123
•
D 3
12
• 13
• 23
•
B •
C
1
• •2 •3
I 1 2
•
• ∅
∅

11.3.3. Remark. Aspects of hereditary systems. An aspect of a hereditary sys-

tem M is a way of specifying it. For example, M can be specified by any of I M ,
B M , C M , or r M , because each of these determines the others. We have expressed
B M , C M , r M in terms of I M . Conversely, if B M is given, then I M consists of the
sets contained in members of B M . If C M is given, then I M consists of the sets con-
taining no member of C M . If r M is given, then I M and the other aspects are found
by setting I M = {X ⊆ E: r M (X) = | X |}. This justifies our view of a hereditary sys-
tem as a unified structure (I , B , C , r, · · ·). We drop the subscripts on I , B , C , r, · · ·
when discussing only one hereditary system. Later we introduce additional as-
pects that can specify a hereditary system.

Most terminology in matroid theory comes from the motivating contexts that
led to the discovery of matroids, particularly graphs and linear algebra. We begin
with the fundamental example from graphs.

11.3.4. Definition. The cycle matroid M(G) of a multigraph G is the heredi-

tary system on E(G) whose circuits are the edge sets of cycles of G. A hered-
itary system that can be specified in this way is a graphic matroid.

11.3.5. Example. Bases in a cycle matroid M(G). The graph K 4− , with four ver-
tices and five edges, arises by deleting one edge from K 4 . Spanning trees in K 4−
have three edges, so every set with more than three edges is dependent. Also the
two triangles are dependent. This yields eight dependent sets and 24 indepen-
dent sets among the subsets of E(K 4−). There are three minimal dependent sets
(the cycles) and eight maximal independent sets (the spanning trees).
For a general multigraph G, the bases of the cycle matroid M(G) are the edge
sets of the maximal forests in G. Each contains a spanning tree of each component
of G, so they have equal size. Consider B1 , B2 ∈ B and e ∈ B1 − B2 . Deleting e
from B1 disconnects a component; since B2 contains a tree spanning the same
component of G, some edge º ∈ B2 − B1 can be added to B1 − e to reconnect it.
The base exchange property is
If B1 , B2 ∈ B M , then for all e ∈ B1 − B2
there exists º ∈ B2 − B1 such that B1 − e + º ∈ B M .
Matroids are the hereditary systems satisfying the base exchange property.
516 Chapter 11: Extremal Problems

11.3.6. Remark. Notational conventions. In this subject, we often discuss modi-

fying a set by adding or deleting a single element. For symmetry and simplicity,
we use the symbols + and − instead of ∪ and − for this, and we drop the set brack-
ets on singleton sets.
Using bold I , B , C for families of subsets of E allows using I ∈ I, B ∈ B,
C ∈ C, respectively, to denote membership. We use roman I, B, C, R to denote
properties, e , º , x , y for elements of E, and X , Y, F for subsets of E.

11.3.7. Example. Rank in cycle matroids. Let G be an n-vertex graph. For X ⊆

E(G), let G X be the spanning subgraph of G with edge set X . In M(G), an inde-
pendent subset of X is the edge set of a forest in G X . When G X has ¾ components,
the maximum size of such a forest is n − ¾ . Hence r(X) = n − ¾ . Below we show
such a forest Y in X .
If r(X + e) = r(X) for some e ∈ E − X , then the endpoints of e lie in a single com-
ponent of G X ; adding e does not combine components. If we add two such edges,
then again we do not combine components. Therefore, r(X) = r(X + e) = r(X + º )
implies r(X) = r(X + e + º ).
The (weak) absorption property (name suggested by A. K ézdy) is
If X ⊆ E and e , º ∈ E,
then r(X) = r(X + e) = r(X + º ) implies r(X + e + º ) = r(X).
Matroids are the hereditary systems satisfying the absorption property.

• • • • Y
e
• • • • X−Y

• • • • º • E− X

Multigraphs may have loops and multiedges. Analogous terminology in

hereditary systems captures the behavior in cycle matroids of elements arising
from loops and multiedges in multigraphs.

11.3.8. Definition. In a hereditary system, a loop is an element forming a cir-

cuit of size 1. Two non-loops are parallel elements if they form a circuit. A
hereditary system is simple if it has no loops or parallel elements.

Another motivating context for matroids is linear independence in vector

spaces. Matroids can be obtained from matrices.

11.3.9. Definition. The vector matroid on a finite set E of vectors in a vector

space (over a field K) is the hereditary system whose independent sets are
the linearly independent sets of vectors in E. A matroid expressible in this
way is a linear matroid (or representable matroid) over K . The column
matroid M(A) of a matrix A is the vector matroid defined on its columns.

11.3.10. Example. Circuits in vector matroids. Technically, one vector (or multi-
ples of it) may be used to represent distinct elements of E, just as a matrix may
Section 11.3: Matroids 517

have repeated columns; these yield parallel elements. The circuits are the min-
imal sets {x1 , . . . , x ¾ } ⊆ E such that ∑ ci x i = 0 using coefficients not all zero.
Minimality forces all ci
= 0.
Let C1 and C2 be distinct circuits containing x. The equations of dependence
for C1 and C2 let us write x as a linear combination of C1 − x and of C2 − x. Equat-
ing these expressions yields an equation of dependence for (C1 ∪ C2) − x; thus
(C1 ∪ C2) − x contains a circuit.
The (weak) elimination property is
If C1 and C2 are distinct circuits and x ∈ C1 ∩ C2 ,
then another member of C M is contained in (C1 ∪ C2) − x.
Matroids are the hereditary systems satisfying the weak elimination property.
The column matroid of the matrix below is also the cycle matroid of K 4− .

⎛0 0 0 1 1⎞
⎜0 1 1 0 1⎟
⎝1 1 0 0 0⎠
The weak elimination property implies that in a matroid the relation of being
parallel is transitive, but this does not hold for all hereditary systems.

Another class of matroids important for applications arose much later from
families of sets. Edmonds–Fulkerson [1965] and Mirsky–Perfect [1967] indepen-
dently discovered that these are matroids.

11.3.11. Definition. The transversal matroid induced by sets A1 , . . . , Am with

union E is the hereditary system on E in which the independent sets are the
systems of distinct representatives (SDR) for subsets of { A1 , . . . , Am}.

11.3.12. Remark. Transversal matroids and bipartite graphs. The name “trans-
versal” arises from using this word to mean SDR. However, in the study of op-
timization problems in hypergraphs, the word “transversal” is used without re-
quiring the representatives to be distinct. We therefore use “SDR” here.
Given A1 , . . . , Am with union E, consider the bipartite incidence graph G.
The parts are E and [m], with e ∈ E adjacent to i ∈ [m] if and only if e ∈ Ai . A
set X ⊆ E is independent in the transversal matroid induced by A1 , . . . , Am on E
if and only if X is covered by some matching in G.
Also, given an E , F-bigraph G, we can associate with vi ∈ F the set NG(vi).
Letting Ai = NG(vi) expresses G as the incidence graph of the family of sets.
Hence every bipartite graph induces a transversal matroid on each of its parts.
The bipartite graph below yields the transversal matroid of the family A =
{{1 , 2} , {2 , 3 , 4} , {4 , 5}}. This matroid is again M(K 4−).
1 2 3 4 5
E • • • • •

[m] • • •
{1 ,2} {2 ,3 ,4} {4 ,5}
518 Chapter 11: Extremal Problems

11.3.13. Example. Independent sets in transversal matroids. The symmetric dif-

ference of matchings M and M in a bipartite graph consists of alternating paths
and even cycles. The only components with more edges from M than from M are
the M-augmenting paths. When | M | > | M | , there must be some such path. For
an M-augmenting path P , replacing M ∩ P with M ∩ P yields a matching of size
| M | + 1 that covers all vertices of M plus the endpoints of P.
For independent sets I1 and I2 in the transversal matroid generated by
A1 , . . . , Am , let M1 and M2 be matchings that saturate I1 and I2 , respectively,
in the associated bipartite graph (below, M1 is bold and M2 is solid). If | I2 | > | I1 | ,
then the matching obtained from M1 via an M1 -augmenting path in M2 M1 cov-
ers I1 plus an element e ∈ I2 − I1 ; this “augments” I1 .
For a hereditary system on E, the augmentation property is
For distinct I1 , I2 ∈ I with | I2 | > | I1 | ,
there exists e ∈ I2 − I1 such that I1 + e ∈ I.
Matroids are the hereditary systems with the augmentation property.
I1 I1 I1 I1 e
E • • • • • • •

P
[m] • • • • • •

We have defined matroids as hereditary systems with an additional property,

but we gave four candidates for this “additional property”, claiming that each
defines matroids. In fact, the properties are equivalent for hereditary systems;
every matroid satisfies them all. After we prove the properties equivalent for
hereditary systems, we only need to verify one such property to use them all. To
illustrate that the various properties all hold, consider cycle matroids.

11.3.14. Example. Augmentation in cycle matroids. Consider I1 , I2 ∈ I M(G) with

| I2 | > | I1 |. The spanning subgraph G Ii has n − | Ii | components. Therefore, the
forest I2 has some edge with endpoints in two components of G I1 . This edge can
be added to I1 to form a larger forest. Hence the augmentation property holds.

11.3.15. Example. Weak elimination in cycle matroids. The circuits of M(G) are
the edge sets of cycles in G. Cycles have even degree at each vertex. For C1 , C2 ∈
C, the symmetric difference C1  C2 also has even degree at each vertex. If C1
=
C2 , then C1  C2 contains a cycle. This is stronger than weak elimination, since
C1 C2 ⊆ C1 ∪ C2 − x. In the picture below, C1 and C2 are face boundaries of length
9 sharing the dashed edges, and C1 C2 is the union of two disjoint cycles.

• •
C2 •
• • • • x • •
C1 •
• •
Section 11.3: Matroids 519

For linear matroids, direct proof of the augmentation or base exchange prop-
erty uses the fact that ¾ linearly independent vectors cannot all be expressed as
linear combinations of a smaller set. On the other hand, since we have verified
the weak elimination property for linear matroids, this theorem of linear algebra
follows from matroid axiomatics!

11.3.16. Lemma. For the rank function r of a hereditary system on E,

(r1) r(∅) = 0.
(r2) r(X) ≤ r(X + e) ≤ r(X) + 1 whenever X ⊆ E and e ∈ E.
Proof: From the definition r(X) = max{| Y | : Y ⊆ X , Y ∈ I}, we have r(∅) = 0.
Because X + e contains every independent subset of X , we have r(X + e) ≥ r(X).
Because the independent subsets of X + e not contained in X consist of e together
with an independent subset of X , we have r(X + e) ≤ r(X) + 1.

11.3.17.* Remark. Every nonempty ideal is the family of independent sets of a

hereditary system. A family B is the set of bases of some hereditary system if
and only if it is a nonempty antichain. A family C is the set of circuits of some
hereditary system if and only if it is an antichain of nonempty sets.
Characterizing rank functions of hereditary systems is more subtle. Proper-
ties (r1) and (r2) above are used to study matroids but do not suffice to make r the
rank function of a hereditary system. (Consider E = {1 , 2}, r(∅) = r(1) = r(2) =
0, r({1 , 2}) = 1; the set {1 , 2} contains no independent set of size 1, even though
it has rank 1.) Another technical condition is needed to characterize rank func-
tions of hereditary systems. Fortunately, we do not need this characterization,
because when studying matroids we always start with a hereditary system.

A XIOMATICS OF MATROIDS

We view a matroid as a hereditary system satisfying an additional structural

property. A constraint on one aspect of a hereditary system yields corresponding
constraints on other aspects; thus we have many equivalent definitions. We can
show that a hereditary system is a matroid by verifying any of them, and then
we can use them all without additional proof.
Adding an edge to a forest creates at most one cycle. This is the property
needed to show that the greedy algorithm works to find minimum spanning trees
in weighted graphs. The creation of at most one circuit by adding one element to
an independent set is in fact a characterization of matroids, as is the effectiveness
of the greedy algorithm itself! Both appear in our list.
Given nonnegative weights on the elements, the greedy algorithm itera-
tively picks a heaviest element whose addition to those already selected yields an
independent set. Rado [1957] (see also Bor u̇vka [1926]) proved that matroids are
precisely the hereditary systems such that for each nonnegative weight function,
the greedy algorithm selects a heaviest independent set.

11.3.18. Definition. A hereditary system M on E is a matroid if it satisfies any

of the following additional properties, where I , B , C , r are the independent
sets, bases, circuits, and rank function of M.
520 Chapter 11: Extremal Problems

I: augmentation - if I1 , I2 ∈ I with | I2 | > | I1 | , then I1 + e ∈ I for some e ∈ I2 − I1 .

B: base exchange - if B1 , B2 ∈ B, then for all e ∈ B1 − B2 there exists º ∈ B2 − B1
such that B1 − e + º ∈ B.
B : dual base exchange - if B1 , B2 ∈ B, then for all e ∈ B1 − B2 there exists
º ∈ B2 − B1 such that B2 + e − º ∈ B.
A: weak absorption - r(X) = r(X + e) = r(X + º ) implies r(X + e + º ) = r(X)
whenever X ⊆ E and e , º ∈ E,
A : strong absorption - if X , Y ⊆ E, and r(X + e) = r(X) for all e ∈ Y , then
r(X ∪ Y) = r(X).
U: uniformity - for every X ⊆ E, the maximal subsets of X belonging to I have
the same size.
R: submodularity - r(X ∩ Y) + r(X ∪ Y) ≤ r(X) + r(Y) when X , Y ⊆ E.
C: weak elimination - if C1 , C2 ∈ C are distinct with e ∈ C1 ∩ C2 , then (C1 ∪
C2) − e contains a circuit.
J: induced circuits - if I ∈ I, then I + e contains at most one circuit.
G: greedy algorithm - for each nonnegative weight function on E, the greedy
algorithm selects a maximum-weight independent set.

11.3.19. Remark. Submodularity. Submodularity arises naturally in linear al-

gebra. The rank of a set X ⊆ E in a vector matroid is the dimension of the
space spanned by X . For subspaces U and V , the result in linear algebra is
dim U ∩ V + dim W = dim U + dim V , where W is the space spanned by U ∪ V .
When X and Y are spanning sets of vectors in U and V , the dimension of the space
spanned by X ∩ Y may be less than dim U ∩ V . Nevertheless, dim U ∩ V + dim W ≤
dim U + dim V is proved naturally for vector spaces using the proof of U ⇒ R be-
low. Exercise 31 obtains submodularity directly for cycle matroids.

Various of these properties (plus being a hereditary system) have been used
to define matroids. Examples include I (Welsh [1976], Schrijver [2003]), U (Ed-
monds [1965a,b], Bixby [1981], Nemhauser–Wolsey [1988]), A (Whitney [1935]),
C (Tutte [1970]), G (Papadimitriou–Steiglitz [1982]). van der Waerden [1937],
Rota [1964], Crapo–Rota [1970], and Aigner [1979] use other conditions.
Many authors list properties of hereditary systems when characterizing some
aspect of a matroid. This obscures the essence of the characterization and leads
to extra work. Working in the context of hereditary systems is simpler. All prop-
erties of hereditary systems are always available; we need not verify them when
introducing another aspect. Our chain of implications may seem long, so it is
worth noting that augmentation (I) and uniformity (U), for example, are easy to
show equivalent, and hence the proof can be given in shorter implication chains.

11.3.20. Theorem. For a hereditary system M , the conditions defining matroids

in Definition 11.3.18 are equivalent.
Proof: Property I is often used to show that some hereditary system is a matroid.
I ⇒ B. Since a smaller base could be augmented from a larger base, bases
have equal size. Now consider B1 , B2 ∈ B with e ∈ B1 − B2 . Since B1 − e ∈ I and
| B2 | > | B1 | − 1, the desired º exists.
B ⇒ A. Bases cannot differ in size; otherwise, repeated base exchange yields
one base inside another. Now let X = X + e + º with r(X) = r(X + e) = r(X + º )
Section 11.3: Matroids 521

and r(X ) > r(X). Among bases containing largest independent sets in X and X ,
choose B and B with largest intersection. Since r(X + e) = r(X + º ) = r(X), we
have e , º ∈/ B. Since | B ∩ X | > | B ∩ X | and | B| = | B | , there exists x ∈ B − B
with x ∈ / X . Base exchange guarantees a base B − x + x for some x ∈ B − B.
Since x ∈ / X , we have | B − x + x ∩ X | = r(X), but |(B − x + x ) ∩ B | > | B ∩ B |.
A ⇒ A . We use induction on | Y − X | . There is nothing to prove when
| Y − X | = 1. When | Y − X | > 1, choose e , º ∈ Y − X and let Y = Y − e − º . Apply-
ing the induction hypothesis using Y or Y + e or Y + º yields r(X) = r(X ∪ Y ) =
r(X ∪ Y + e) = r(X ∪ Y + º ). Now weak absorption yields r(X) = r(X ∪ Y).
A ⇒ U. If Y is a maximal independent subset of X , then r(Y + e) = r(Y) for
all e ∈ X − Y . By strong absorption, r(X) = r(Y) = | Y |. Hence all such Y have
the same size, r(X).
U ⇒ R. Given X , Y ⊆ E, choose a maximum independent set I1 from X ∩ Y .
By uniformity, I1 can be enlarged to a maximum independent subset of X ∪ Y ;
call this I2 . Consider I2 ∩ X and I2 ∩ Y ; these are independent subsets of X and
Y , and each includes I1 . Hence
r(X ∩ Y) + r(X ∪ Y) = | I1 | + | I2 | = | I2 ∩ X | + | I2 ∩ Y | ≤ r(X) + r(Y).

X I2 I1 I2 Y

U⇒R

R ⇒ C. Consider distinct circuits C1 , C2 ∈ C with e ∈ C1 ∩ C2 . We have

r(C1) = | C1 | − 1 and r(C2) = | C2 | − 1. Also r(C1 ∩ C2) = | C1 ∩ C2 | , since every proper
subset of a circuit is independent. If (C1 ∪ C2)− e contains no circuit, then its rank
is | C1 ∪ C2 | − 1, and hence r(C1 ∪ C2) ≥ | C1 ∪ C2 | − 1. Applying submodularity to
C1 and C2 now yields the contradiction
| C1 ∩ C2 | + | C1 ∪ C2 | − 1 ≤ | C1 | + | C2 | − 2.
C ⇒ J. If I + e contains circuits C1 and C2 for some I ∈ I, then C1 and C2 both
contain e. Now weak elimination guarantees a circuit in (C1 ∪ C2) − e. However,
(C1 ∪ C2) − e is independent, being contained in I.
J ⇒ B . For distinct B1 , B2 ∈ B and e ∈ B1 − B2 , by J there is a unique circuit
C in B2 + e. Since C
⊆ B1 , there exists º ∈ C − B1 , and now B2 + e − º ∈ I. To prove
B2 + e − º is a base, we show that no two bases have distinct sizes. Otherwise,
choose bases B1 and B2 with largest intersection such that ##### B1 ##### < ##### B2 #####. Since
B1
⊆ B2 , there is e ∈ B1 − B2 . As before, we find º ∈ B2 − B1 with B2 + e − º ∈ I.
Any base containing B2 + e − º is as large as B2 but shares more with B1 than B2 .
B ⇒ G. With nonnegative weights, some optimum is a base, and the algo-
rithm chooses a base B. Let B∗ be an optimal base having largest intersection
with B. If B∗
= B, then let e be the first element chosen for B that is not in
B∗ . Dual base exchange yields º ∈ B∗ − B such that B∗ + e − º ∈ B. The choice
of B∗ yields w(º ) > w(e). Since the algorithm chose e when º was available,
w(e) ≥ w(º ). Hence B∗ + e − º contradicts the choice of B∗ , so B = B∗ .
522 Chapter 11: Extremal Problems

G ⇒ I. Given I1 , I2 ∈ I with ¾ = | I1 | < | I2 | , we design a weight function so

that the greedy algorithm finds the desired augmentation. Let w(e) = ¾ + 2 for
e ∈ I1 , and let w(e) = ¾ + 1 for e ∈ I2 − I1 . Let w(e) = 0 for e ∈ / I1 ∪ I2 . Now
w(I2) ≥ (¾ + 1)2 > ¾(¾ + 2) = w(I1), so I1 does not have maximum weight. However,
the greedy algorithm chooses all of I1 before any element of I2 − I1 . Because it
finds a maximum-weight independent set, after absorbing I1 it adds an element
of I2 − I1 such that I1 + e ∈ I.

11.3.21. Example. Uniform matroids. The uniform matroid of rank ¾ , denoted

U ¾ ,n when | E| = n, is defined by I = {X ⊆ E: | X | ≤ ¾}. This immediately sat-
isfies the base exchange and augmentation properties. The free matroid is the
uniform matroid of rank | E|. Few uniform matroids are graphic, and few graphic
matroids are uniform (Exercise 27). Neither M(K 4−) nor M(K 4) is uniform.

11.3.22. Example. Partition matroids. The partition matroid on E induced by

a partition of E into blocks E1 , . . . , E¾ is defined by I = {X ⊆ E: | X ∩ Ei | ≤ 1
for all i}. Since ∅ ∈ I, and since X ∈ I if and only if its elements lie in distinct
blocks, I is a nonempty ideal. Given I1 , I2 ∈ I with | I2 | > | I1 | , the set I2 must
intersect more blocks than I1 ; an element of I2 in a block that I1 misses yields the
desired augmentation of I1 . Alternatively, r(X) is the number of blocks having
elements in X ; this satisfies the absorption property. Every partition matroid
is a transversal matroid; consider the E , [¾]-bigraph consisting of ¾ stars whose
leaf sets are the blocks of the partition.
Given a U, V-bigraph G, each part induces a partition matroid with ground
set E(G), partitioning the edges by their endpoints in one part. A set X ⊆ E(G)
is a matching in G if and only if X is independent both in the partition matroid
induced by U and in the partition matroid induced by V . This is the motivation
for the idea of matroid intersection discussed later in this section.
E1 E2 E3 E4 E5
• • • • • U

• • • • • V
F1 F2 F3 F4 F5

When G has an odd cycle, G has no set of vertices whose incident sets par-
tition E(G); thus M(K 4−) is not a partition matroid, for example. In a digraph,
however, each edge has a head and a tail, and we can define the head partition
matroid and the tail partition matroid using the edge partitions induced by
incidences with heads and by incidences with tails. For example, the matroid of
Example 11.3.2 arises as the partition matroid on E induced by U in the bipar-
tite graph on the left below, as the head partition matroid in the first digraph,
and as the tail partition matroid in the second digraph.
U
• • • • • •
1 2 3 1 2 3 1 2 3
• • • • • • • • •
Section 11.3: Matroids 523

DUALITY AND MINORS

Duality in matroids generalizes duality in planar graphs (Section 9.1). Every

connected plane graph G has a natural dual graph G ∗ such that (G ∗)∗ = G. The
dual associates a vertex of G ∗ to each face of G and an edge of G ∗ to each edge of
G, joining the faces incident to it. Loops and multiedges can arise in G ∗ (from
cut-edges and pairs of faces sharing more than one boundary edge, respectively),
so when discussing graphic matroids we allow the generality of multigraphs.
A set of edges forms a spanning tree in a plane graph G if and only if the
duals to the remaining edges form a spanning tree in G ∗ (Exercise 9.1.25). Hence
the bases in the cycle matroid M(G ∗) are the complements of the bases in M(G). We
define duality for matroids and for hereditary systems to extend these duality
properties of planar graphs.

11.3.23. Definition. The dual of a hereditary system M on E is the hereditary

system M ∗ defined by B M ∗ = {B: B ∈ B M }. We may write B∗ for B M ∗ when
M is understood; these are the cobases of M. Similarly, the circuits of M ∗
are the cocircuits of M , denoted C∗ , etc.
A set in E is spanning if it contains a base; S M denotes the family of
spanning sets. A set is a hyperplane if it is a maximal set containing no
base; H M denotes the family of hyperplanes.

11.3.24. Remark. If B is a nonempty antichain, then also {B: B ∈ B} is a

nonempty antichain (since containment reverses under complementation), so the
dual of a hereditary system is a hereditary system. Also it is immediate that
(M ∗)∗ = M , which explains the name dual.
A set is independent in M ∗ if and only if it is contained in a cobase of M.
Hence I M ∗ = {S: S ∈ S M }, and similarly S M ∗ = {I: I ∈ I M }.
A set is a circuit in M ∗ if and only if it is a minimal set contained in no cobase
of M. Hence C M ∗ = {H: H ∈ H M } and H M ∗ = {C: C ∈ C M }.

Duality is useful because the dual of a matroid is a matroid. This follows

from the dual base exchange property (B ). In essence, the statement of the dual
base exchange property for M is that of the base exchange property for M ∗ .

11.3.25. Theorem. (Whitney [1935]) The dual of a matroid M is a matroid.

Proof: We have seen that M ∗ is a hereditary system; now we prove the base ex-
change property for M ∗ . If B1 , B2 ∈ B∗ and e ∈ B1 − B2 , then B1 , B2 ∈ B, with
e ∈ B2 − B1 . By dual base exchange for M , there exists º ∈ B1 − B2 such that
B1 + e − º ∈ B. Now B1 − e + º ∈ B∗ , proving base exchange for M ∗ .

Computing the rank function of the dual is easy using an alternative notion of
the rank of a set. Instead of viewing r(X) as the maximum size of an independent
subset of X , it can be helpful to view it as the maximum size of the intersection
of X with a base.

11.3.26. Proposition. The rank function r∗ of the dual of a matroid M on E is

given by r∗(X) = | X | − (r(E) − r(X)).
524 Chapter 11: Extremal Problems

Proof: The rank of a set X is the maximum size of its intersections with bases.
Choose B∗ ∈ B∗ so that r∗(X) = | X ∩ B∗ | , and let B = B∗ . Now B is a base hav-
ing smallest intersection with X and hence largest intersection with X . Hence
r(X) = | B − X | . Also | B| = r(E), so r∗(X) = | X | − | X ∩ B| = | X | − (r(E) − r(X)).
E
X X
r∗(X) r(X)
X X
B∗ B

The argument of Proposition 11.3.26 is not valid for arbitrary hereditary

systems, because | B| = r(E) uses the uniformity property.
To understand how duality of matroids extends notions of graph theory, we
need to know what the cocircuits are in graphic matroids. We begin by charac-
terizing the cocircuits for all matroids.

11.3.27. Proposition. The cocircuits of a matroid are the minimal sets intersect-
ing every base. The bases are the minimal sets intersecting every cocircuit.
Proof: Cocircuits are the minimal sets in no cobase. Since the cobases are the
complements of the bases, a set lies in no cobase if and only if it intersects ev-
ery base. Similarly, cobases are the maximal sets containing no cocircuit, so the
complements of the cobases are the minimal sets intersecting every cocircuit.

11.3.28. Corollary. The cocircuits of a cycle matroid M(G) are the bonds of G.
Proof: By Proposition 11.3.27, the cocircuits are the minimal sets intersecting
every maximal forest. Hence they are the minimal sets whose deletion increases
the number of components; these are the bonds.

11.3.29. Definition. The bond matroid or cocycle matroid of a graph G is the

hereditary system whose circuits are the bonds of G.

By Corollary 11.3.28, the bond matroid of G is the dual of the cycle matroid
M(G). By Theorem 9.1.11, a set of edges forms a bond in a plane graph G if and
only if the corresponding dual edges form a cycle in G ∗ . Hence we have shown
that the dual of the cycle matroid of G is in fact the cycle matroid of G ∗ . We
will use this to characterize planar graphs. For the proof, we must generalize
edge-deletion and edge-contraction to matroids.
Given a graph G, the acyclic subsets of E(G − e) are just the acyclic subsets of
E(G) that omit e. To make contraction a dual notion, it makes sense to describe
its effect in terms of sets containing bases instead of sets contained in bases. A
set X ⊆ E(G · e) contains a spanning tree of each component of G · e if and only if
X + e contains a spanning tree of each component of G.
This motivates our notation for hereditary systems. Just as we express the
subgraph of G induced by S as both G[S] and G − S, for matroids we use different
notation to emphasize the set of elements remaining or one element eliminated.
Section 11.3: Matroids 525

11.3.30. Definition. Given a hereditary system M on E and a set F ⊆ E, the re-

striction of M to F , denoted M | F and obtained by deleting F , is the hered-
itary system defined by I M | F = {X ⊆ F: X ∈ I M }. The contraction of M to
F , denoted M.F and obtained by contracting F , is the hereditary system
defined by S M.F = {X ⊆ F: X ∪ F ∈ S M }. When F = E − e, we write M − e
for M | F and M · e for M.F . The minors of a hereditary system M are the
hereditary systems arising from M by applying deletions and contractions.

From the definitions, M | F and M.F are hereditary systems. The notations
M | F and M.F appear (briefly) in Oxley [1992] (pages 22 and 104, respectively).
Note the distinction between “ .” and “ · ”; the former emphasizes “contracting to”
a specified set by “contracting away” the other elements. We use “ − ” or “ · ” when
eliminating one specified element and “ | ” or “ .” when specifying the elements that
remain. Our notation for M − e and M · e is consistent with our usage for graphs
but is nonstandard in the matroid community, where M \ e for our M − e and M/e
for our M · e are most common.
Defining contraction via spanning sets yields a natural duality.

11.3.31. Proposition. For hereditary systems, restriction and contraction are

dual operations: (M.F)∗ = (M ∗ | F) and (M | F)∗ = (M ∗ .F).
Proof: I(M.F)∗ = {X ⊆ F: F − X ∈ S M.F } = {X ⊆ F: (F − X) ∪ F ∈ S M }
= {X ⊆ F: X ∈ S M } = {X ⊆ F: X ∈ I M ∗ } = I M ∗ | F .
For the second statement, apply the first to M ∗ and take duals.

As expected, restrictions and contractions of matroids are matroids.

11.3.32. Theorem. Given F ⊆ E and a matroid M on E, both M | F and M.F are

matroids on F . Their bases and rank functions are given by
B M | F = {B ∩ F: B ∈ B M and | B ∩ F | = r M (F)} , r M | F (X) = r M (X) ,
B M.F = {B ∩ F: B ∈ B M and ##### B ∩ F ##### = r M (F)} , r M.F (X) = r M (X ∪ F) − r M (F).

Proof: The augmentation property from M applies to sets in I M | F ; thus M | F sat-

isfies the augmentation property and is a matroid. Since M.F = (M ∗ | F)∗ , duality
implies that M.F is also a matroid.
The expressions for the bases and rank function of M | F follow from the defi-
nition of I M | F ; we do the same for M.F from the definition of S M.F . A base of M.F
is a minimal set B̂ ⊆ F such that B̂ ∪ F ∈ S M . Thus B̂ ∪ F contains a base B of
M. The minimality of B̂ implies that B̂ ⊆ B and that B ∩ F is a maximal inde-
pendent subset of F (by uniformity on B̂ ∪ F). Thus B M.F consists of the sets in
F whose addition to a maximal independent subset of F yields a base of M.
To compute r M.F (X), recall that r M.F (X) = | B ∩ X | for some B ∈ B M.F ; let
# #
Y = B ∩ X . Also B = B ∩ F for some B ∈ B M such that #### B ∩ F #### = r M (F); let
Z = B ∩ F . Since B ∈ B M , we have Y ∪ Z ∈ I M . If Y ∪ Z is not a maximal
independent subset of X ∪ F , then it augments to a base in M that contains Z
and has larger intersection with X than B , contradicting the choice of B . Hence
uniformity in M yields | Y ∪ Z | = r M (X ∪ F). Now | Y | = r M (X ∪ F) − r M (F).
526 Chapter 11: Extremal Problems

11.3.33. Remark. The formula for r M.F yields a description of the independent
sets: X ∈ I M.F if and only if adding X to F increases the rank by | X | . Note that
when F is an independent set {e} of size 1, we have r M.F (X) = r M (X + e) − 1. In
particular, when we contract a nonloop edge in a graph, the maximum size of a
forest among the edges of any set containing that edge decreases by 1.
The duality between deletion and contraction is familiar in plane graphs.
Deleting an edge e in a plane graph G contracts the corresponding edge in G ∗ ;
contracting e deletes the edge in the dual. The fate of the 4-cycle below illustrates
that when a circuit of M intersects F , its intersection with F need not be a cir-
cuit of M.F , even when only one element has been contracted. In fact, the circuits
of M.F are the minimal nonempty sets in {C ∩ F: C ∈ C M } (Exercise 47).
Also, restriction and contraction commute (Exercise 49).

• •
← →
• • • • • • • • • • • •
delete contract
• •

11.3.34. Corollary. The behavior of cycle matroids and bond matroids under
deletion or contraction of an edge e ∈ E(G) is
M(G − e) = M(G) − e , M ∗(G − e) = M ∗(G) · e ,
M(G · e) = M(G) · e , M ∗(G · e) = M ∗(G) − e.

Proof: Matroid deletion and contraction are defined so that the statements in the
first column describe the behavior of cycle matroids. Using these and Proposition
11.3.31, for the second column we compute
M ∗(G − e) = [M(G − e)]∗ = [M(G) − e]∗ = M ∗(G) · e ,
M ∗(G · e) = [M(G · e)]∗ = [M(G) · e]∗ = M ∗(G) − e.

Now we characterize planar graphs using matroids. In Theorem 9.1.11, we

proved that a set of edges in a plane graph G forms a cycle if and only if the
corresponding dual edges form a bond in G ∗ . Thus under the natural bijection
from edges to dual edges, the cycle matroid of a plane graph G is (isomorphic to)
the bond matroid of G ∗ . By Corollary 11.3.28, the bond matroid of a graph H is
[M(H)]∗ . Applying this to G and G ∗ tells us that the bond matroid of G is (isomor-
phic to) the cycle matroid of G ∗ . In particular, the bond matroid of G is graphic.
Using Kuratowski’s Theorem, we will use this to characterize planarity.
Whitney [1932c] approached this question by defining a non-geometric notion
of the dual of a graph. Changing his definition slightly, we say that H is an ab-
stract dual of G if there is a bijection
: E(G) → E(H) such that X ⊆ E(G) is
a bond in G if and only if
(X) is the edge set of a cycle in H. With this defini-
tion, the statement that G has an abstract dual H is the same as the statement
that the bond matroid of G is graphic; the bijection
establishes an isomorphism
between M ∗(G) and M(H).

11.3.35. Theorem. (Whitney [1932c]) A graph G is planar if and only if its bond
matroid M ∗(G) is graphic.
Section 11.3: Matroids 527

Proof: We have observed that planar graphs have abstract duals; this proves
necessity. For sufficiency, we first prove that existence of an abstract dual is pre-
served under deletion and contraction of edges. Suppose that G has an abstract
dual H , so that M(H) =∼ M ∗(G). Let e be the edge of H corresponding to e under
the bijection. To prove that H · e is an abstract dual of G − e and that H − e is
an abstract dual of G · e, we use Corollary 11.3.34 to compute
M ∗(G − e) = M ∗(G) · e ∼
= M(H) · e = M(H · e ) ,
M ∗(G · e) = M ∗(G) − e ∼
= M(H) − e = M(H − e ).
By Kuratowski’s Theorem, a nonplanar graph contains a subdivision of K 5
or K 3 ,3 . Hence K 5 or K 3 ,3 is a minor of it. Since existence of abstract duals is pre-
served under deletion and contraction, showing that K 5 and K 3 ,3 have no abstract
dual implies that every nonplanar graph has no abstract dual.
If H is an abstract dual of G, then also G is an abstract dual of H , since
M ∗(G) ∼ = M(H) if and only if M(G) ∼ = M ∗(H). If G has girth ½ , then bonds of H
have size at least ½ , so (H) ≥  . Letting n = | V(H)| and m = | E(H)| , we have also
| E(G)| = m. Thus the Degree-Sum Formula yields n ≤ ⌊ 2m/ (H)⌋ ≤ ⌊ 2m/ ⌋ .
Let H be an abstract dual of K 5 . Since K 5 has girth 3, n ≤ ⌊ 20/3⌋ = 6. Since
all bonds of K 5 have four or six edges, all cycles of H have four or six edges, and
thus H is a simple bipartite graph. However, no simple bipartite graph with at
most six vertices has ten edges.
Let H be an abstract dual of K 3 ,3 . Since K 3 ,3 has girth 4, n ≤ ⌊ 18/4⌋ = 4.
Since all bonds of K 3 ,3 have at least three edges, all cycles of H have at least three
edges, and thus H is a simple graph. However, no simple graph with at most four
vertices has nine edges.

Planar graphs have no K 5- or K 3 ,3-minor, since planarity is preserved under

deletion and contraction of edges. Kuratowski’s Theorem implies that nonplanar
graphs do have such minors. This yields the characterization by Wagner [1937]:
K 5 and K 3 ,3 are the minimal forbidden minors for planar graphs.

THE SPAN FUNCTION

In addition to the notion of “spanning set ” as “set containing a base”, another

notion of span is suggested by vector spaces. A set S of vectors spans a vector v
if v is a linear combination of elements of S, which means that S + v contains a
dependent set. This notion defines another aspect of hereditary systems.

11.3.36. Definition. The span function of a hereditary system M on E is the

function  M on 2 E defined by  M (X) = X ∪ {e ∈ E: Y + e ∈ C M for some
Y ⊆ X}. If e ∈  (X), then X spans e.

A set in a hereditary system is dependent if and only if it contains a circuit,

which by Definition 11.3.36 means that e ∈  (X − e) for some e ∈ X . Hence we
can find the independent sets (and other aspects of M) from the span function via
I M = {X ⊆ E: (e ∈ X) ⇒ (e ∈ /  M (X − e))}. The properties of  that we use in
studying matroids are (s1, s2, s3) below.
528 Chapter 11: Extremal Problems

11.3.37. Proposition. If is the span function of a hereditary system on E, and

X , Y ⊆ E, then
s1) X ⊆ (X) ( is expansive).
s2) Y ⊆ X implies (Y) ⊆ (X) ( is order-preserving).
/ (X) and e ∈ (X + ) imply ∈  (X + e) (Steinitz exchange).
s3) e ∈
Proof: Definition 11.3.36 implies immediately that  is expansive and order-
preserving. If e ∈ /  (X), then e ∈/ X . With e ∈  (X + ), also e belongs to a
circuit C in X + + e. With e ∈ /  (X), we must have ∈ C. This circuit yields
∈  (X + e), and hence  satisfies the Steinitz exchange property.
11.3.38. Example. Steinitz exchange in cycle matroids. In the cycle matroid
M(G), the meaning of e ∈/  (X) is that X (the solid edges below) contains no path
/  (X) but e ∈  (X + ), then adding completes
between the endpoints of e. If e ∈
such a path. Adding e to the path completes a cycle, so also ∈  (X + e).
• • •
e
• • •

Definition 11.3.36 immediately implies that the circuits of a hereditary sys-

tem satisfy the strong dependence property: e ∈ C implies e ∈  (C − e). The
natural notion that an element is spanned by a set X if adding it to X does not
increase the rank is valid for all hereditary systems.

11.3.39. Lemma. In a hereditary system, [r(X + e) = r(X)] ⇒ e ∈  (X).

Proof: Let Y be a maximum independent subset of X . Since | Y | = r(X) = r(X +
e), also Y is a maximum independent subset of X + e. Hence e completes a circuit
with some subset of X contained in Y , and e ∈  (X).

In fact, the converse characterizes matroids! We call the converse of Lemma

11.3.39 the “incorporation property”.

11.3.40.* Theorem. If M is a hereditary system, then each condition below is

necessary and sufficient for M to be a matroid.
P: incorporation - r( (X)) = r(X) for all X ⊆ E.
S: idempotence -  2(X) =  (X) for all X ⊆ E.
T: transitivity of dependence - if e ∈  (X) and X ⊆  (Y), then e ∈  (Y).
C : strong elimination - whenever C1 , C2 ∈ C, e ∈ C1 ∩ C2 , and ∈ C1 C2 ,
there exists C ∈ C such that ∈ C ⊆ (C1 ∪ C2) − e.
Proof: See Exercise 50.

Incorporation, idempotence, and transitivity of dependence are well-known

properties of matroids. The equivalence of C and C for hereditary systems was
first proved by Lehman [1964]. Brylawski [1986] gave another way to obtain C .
Idempotence is natural for linear matroids; the span of a set of vectors has noth-
ing more in its own span. This suggests another aspect of hereditary systems.

11.3.41. Definition. The closed sets of a hereditary system on E are the sets
X ⊆ E such that  (X) = X (also called flats or subspaces).
Section 11.3: Matroids 529

In a matroid on E, the sets whose span is E are the sets containing bases;
hence the term “spanning sets”. Similarly, in a matroid the hyperplanes are the
maximal proper closed subsets of E. Both fail for general hereditary systems.
The span function of a matroid is also called its closure function. A closure
operator on 2 E is an expansive, order-preserving, idempotent function from 2 E
to 2 E . Such an operator is the span function of a matroid if and only if it satisfies
Steinitz exchange. Since every hereditary system satisfies Steinitz exchange, our
approach to matroids as hereditary systems with an added property is not well
suited for studying closure operators. The span function of a hereditary system
M is a closure operator if and only if M is a matroid.
Matroids are developed from lattice theory (Section 12.4) in MacLane [1936],
Rota [1964], and Aigner [1979]. Brylawski [1986] described the transformations
among about a dozen aspects of matroids, calling these maps cryptomorphisms.

MATROID INTERSECTION

Matroid theory took a great leap forward with the Matroid Intersection and
Union Theorems. They provided a unified context for many well-known min-max
relations, which became corollaries. We proved some of these in earlier chapters.
Yielding a unified proof for many important theorems, the Matroid Intersection
Theorem can be considered among the most beautiful theorems of combinatorics.
The Matroid Intersection Theorem is a min-max relation for the maximum
size of common independent sets in two matroids on the same set E. The intersec-
tion of two matroids is a hereditary system but generally not a matroid. For mul-
tiple matroids on a set E, we typically use subscripts to distinguish corresponding
aspects, as in B i for the bases of Mi , etc. We still use X for the complement of X
within the full set E.

11.3.42. Definition. The intersection of hereditary systems M1 and M2 on E

is the hereditary system whose independent sets are {X ⊆ E: X ∈ I1 ∩ I2 }.

11.3.43. Example. Since I1 and I2 are closed under taking subsets, the common
independent sets in two hereditary systems also form a hereditary family.
In a bipartite graph G with edge set E, each part induces a partition matroid
on E. A set of edges forms an independent set in one of these partition matroids
if and only if its endpoints in the corresponding part of G are distinct. A set of
edges is independent in both matroids if and only if it is a matching in G.
The hereditary system whose independent sets are the matchings in a bi-
partite graph is not a matroid. The central edge in a path of length 3 forms an
independent set, and the two end edges form a larger one, but the smaller set can-
not be augmented from the larger, so augmentation fails. This is why the greedy
algorithm does not solve maximum matching in bipartite graphs.

The Matroid Intersection Theorem expresses the maximum size of a common

independent set in two matroids M1 and M2 in terms of their rank functions. We
prove it inductively, using restrictions and contractions. Recall that a loop is an
element forming a circuit of size 1.
530 Chapter 11: Extremal Problems

11.3.44. Theorem. (Matroid Intersection Theorem; Edmonds [1970]) For

matroids M1 and M2 on E,
max I ∈I1 ∩I2 | I | = min X ⊆ E(r1 (X) + r2(X)).

Proof: (Woodall; see Seymour [1976]) For the upper bound on | I | , consider I ∈
I1 ∩ I2 and X ⊆ E. The sets I ∩ X and I ∩ X are also common independent sets,
and | I | = | I ∩ X | + ##### I ∩ X ##### ≤ r1 (X) + r2(X). Hence max{| I |} ≤ min{r1 (X) + r2(X)}.

X X
I

To achieve equality, we use induction on | E| ; when | E| = 0 both sides are 0. If

every element of E is a loop in M1 or in M2 , then max | I | = 0 = r1 (X)+ r2 (X), where
X consists of all loops in M1 . Hence we may assume that | E| > 0 and that some
e ∈ E is a non-loop in both matroids. Let F = E − e, and consider the matroids
M1 | F , M2 | F , M1 .F , M2 .F . By Theorem 11.3.32, a set X is independent in M1 .F
if and only if X + e is independent in M1 , and similarly for M2 .F .
Let ¾ = min X ⊆ E {r1(X) + r2(X)}; we seek a common independent ¾-set in M1
and M2 . If none exists, then M1 | F and M2 | F have no common independent ¾-set,
and M1 .F and M2 .F have no common independent (¾ − 1)-set. By the induction
hypothesis and rank formulas (Theorem 11.3.32),
r1 (X) + r2(F − X) ≤ ¾ − 1 for some X ⊆ F ,
r1 (Y + e) − 1 + r2(F − Y + e) − 1 ≤ ¾ − 2 for some Y ⊆ F.

We use (F − Y) + e = Y and F − X = X + e and sum the two inequalities:

r1 (X) + r2(X + e) + r1 (Y + e) + r2(Y) ≤ 2 ¾ − 1 .
Write X = X + e and Y = Y + e. We apply submodularity of r1 to X and Y
and submodularity of r2 to Y and X . With the preceding inequality, this yields
r1 (X ∪ Y ) + r1 (X ∩ Y ) + r2(Y ∪ X ) + r2(Y ∩ X ) ≤ 2 ¾ − 1 .
Since Y ∩ X = X ∪ Y and Y ∪ X = X ∩ Y (see Venn diagram below), the left
side sums two instances of r1 (Z) + r2(Z). Hence the hypothesis ¾ ≤ r1 (Z) + r2(Z)
for all Z ⊆ E yields 2 ¾ ≤ 2 ¾ − 1. This contradiction implies that M1 and M2 do
have a common independent ¾-set.

X Y
•
e

We have already proved various special cases of the Matroid Intersection The-
orem, including the K önig–Egerváry Theorem and the Ford–Fulkerson Theorem
Section 11.3: Matroids 531

(Theorem 7.2.15). Given any two matroids on the same set, the Matroid Intersec-
tion Theorem guarantees a min-max relation for the maximum size of a common
independent set, says what the result should be, and provides a proof.

11.3.45. Corollary. (K önig [1931], Egerváry [1931]) In a bipartite graph, a

largest matching and smallest vertex cover have equal size.
Proof: In the partition matroids M1 and M2 on E(G) induced by the partite sets
U1 and U2 , the common independent sets are the matchings in G.
For X ⊆ E, the number of vertices of Ui incident to edges of X is ri(X). Hence
r1 (X) + r2(X) is the size of a vertex cover, using U1 for X and U2 for X . On the
other hand, if T1 ∪ T2 is a vertex cover with T i ⊆ Ui , and X is the set of edges
incident to T1 , then r1 (X) + r2(X) ≤ | T1 ∪ T2 |. We conclude that min{r1(X) +
r2(X)} is the minimum size of a vertex cover.
With (G) and (G) denoting the maximum size of a matching and the min-
imum size of a vertex cover, we obtain
 (G) = max{| I | : I ∈ I1 ∩ I2 } = min{r1(X) + r2(X)} = (G).
T1
• • • • •

• • • • •
T2
The next application uses the rank function for transversal matroids.

11.3.46. Example. Transversal matroids (Definition 11.3.11). The transversal

matroid induced by subsets A1 , . . . , Am of E has as its independent sets the par-
tial systems of distinct representatives. Equivalently, these are subsets of E that
can be matched into [m] in the bipartite incidence graph G with parts E and [m].
If N(S) < | S| for some S ⊆ X ⊆ E, then every matching in G leaves at least
−
| | | N(S)| elements of X uncovered. Ore ’s Defect Formula (Corollary 6.1.12)
S
gives r(X) = min S⊆ X {| X | − (| S| − | N(S)|)} = minS⊆ X {| N(S)| + | X − S|}.
Ore [1955] gave another useful expression for r(X). For J ⊆ [m], let A(J) =
⋃i∈ J Ai ; in terms of the graph, A(J) = N(J) (see figure below). Since r(X) =
 (G[X ∪ [m]]), we can write the defect formula in terms of neighborhoods of sub-
sets of [m] instead of subsets of X to obtain
r(X) = min J ⊆[m]{| A(J) ∩ X | + m − | J |}. (∗)
The upper bound here holds because for any J ⊆ [m], at most | A(J) ∩ X | el-
ements of X can be matched into J, and at most m − | J | can be matched into
[m] − J. The statement that equality holds is equivalent to the K önig–Egerváry
Theorem, since (A(J) ∩ X) ∪ ([m] − J) is a vertex cover of G[X ∪ [m]], and any
minimal vertex cover Q can be achieved in this way by setting J = [m] − Q.
N(Y) J
• • • • • [m]

• • • • • • • • • • • E
Y X A(J)
532 Chapter 11: Extremal Problems

11.3.47. Corollary. (Ford–Fulkerson [1958]) Families A = { A1 , . . . , Am} and

B = {B1 , . . . , Bm} have a common system of distinct representatives (CSDR)
if and only if, for each I , J ⊆ [m],
## #
###(⋃ Ai ) ∩ (⋃ Bj )#### ≥ | I | + | J | − m.
## ##
i∈ I j∈J

Proof: Let E be the union of all the sets in A ∪ B. A common partial transversal
is a common independent set in the two transversal matroids M1 , M2 induced on
E by A and B. A common transversal is a common independent set of size m. We
need only restate the condition r1 (X)+ r2 (X) ≥ m to find the appropriate condition
on the set systems.
The rank formula (∗) from Example 11.3.46 yields
# #
r1 (X) + r2(X) = min {| A(I) ∩ X | − | I | + m} + min {#### B(J) ∩ X #### − | J | + m}.
I⊆ [m] J⊆ [m]

Hence r1 (X) + r2(X) ≥ m for all X if and only if

| A(I) ∩ X | + ##### B(J) ∩ X ##### ≥ | I | + | J | − m for all X ⊆ E and I , J ⊆ [m].
Since | A(I) ∩ X | + ##### B(J) ∩ X ##### ≥ | A(I) ∩ B(J)| , the Ford–Fulkerson condition
is sufficient. To see that it is necessary, let X be a subset of E such that A(I) −
B(J) ⊆ X and B(J) − A(I) ⊆ X . Given I , J, consider the contribution by elements
# #
of E to the left side. In this case | A(I) ∩ X | + #### B(J) ∩ X #### = | A(I) ∩ B(J)| , so the
condition on r1(X) + r2(X) shows that the condition on A(I) ∩ B(J) is necessary.

11.3.48.* Remark. The augmenting path approach to maximum bipartite

matching generalizes to matroid intersection. The algorithm yields a common
independent set I of maximum size and a set X such that r1 (X) + r2(X) = | I | (see
Lawler [1976], Edmonds [1979], Faigle [1987]). Finding a maximum common in-
dependent set in three matroids is NP-complete (Exercises 25–26).
Maximum matching in general graphs also extends to a matroid context. A
common generalization of matroid intersection and general graph matching is
known in different phrasings as the Matroid Matching Problem or the Ma-
troid Parity Problem. It is solvable in polynomial time for linear matroids, in
which case it yields a min-max relation (Lovász [1978]). Lovász also proved that it
is not solvable in polynomial time for general matroids in the same computational
model. The solution technique over linear matroids involves “polymatroids”, a
polyhedral generalization of matroids.

Next we consider a related min-max relation for “union” of matroids. In this

and other applications of matroid intersection, it can be helpful to restrict the
range of the minimization.

11.3.49. Corollary. The maximum size of a common independent set in matroids

M1 , M2 on E is the minimum of r1 (X 1) + r2(X 2) over sets X 1 , X 2 such that
X 1 ∪ X 2 = E and each X i is closed in Mi .
Proof: The incorporation property implies ri( i(X)) = ri(X).
Section 11.3: Matroids 533

MATROID UNION

The intersection of two matroids is seldom a matroid, but a natural con-

cept of union does yield a matroid. The Matroid Union Theorem states this and
gives a useful min-max relation for the rank function. The Matroid Intersection
and Union Theorems are equivalent; they can be derived from each other. Welsh
[1976] proves the Matroid Union Theorem first; here we obtain it from the Ma-
troid Intersection Theorem.

11.3.50. Definition. The union M1 ∪ · · · ∪ M¾ of hereditary systems M1 , . . . , M¾

on E is the hereditary system M on E defined by I M = {I1 ∪ · · · ∪ I¾ : Ii ∈ Ii }.
The direct sum M1 ⊕ · · · ⊕ M¾ of hereditary systems M1 , . . . , M¾ on disjoint
sets E1 , . . . , E¾ is the hereditary system M on E1 ∪ · · · ∪ E¾ defined by I M =
{I1 ∪ · · · ∪ I¾ : Ii ∈ Ii }.

Since I M is a nonempty ideal, M1 ∪ · · · ∪ M¾ is indeed a hereditary system.

The direct sum M1 ⊕ · · · ⊕ M¾ on E1 , . . . , E¾ can be expressed as a matroid union.
Let E = E1 ∪ · · · ∪ E¾ . For 1 ≤ i ≤ ¾ , define Mi on E by starting with Mi and
addiing each element outside Ei as a loop. Now M1 ⊕ · · · ⊕ M¾ = M1 ∪ · · · ∪ M¾ .
When each Mi is a uniform matroid, the direct sum is a generalized parti-
tion matroid. Such a matroid on E consists of a partition E1 , . . . , E¾ of E and
positive integers r1 , . . . , r¾ such that X ∈ I if | X ∩ Ei | ≤ ri for all i. The partition
matroids defined earlier arise when ri = 1 for all i.

11.3.51. Proposition. Given matroids M1 , . . . , M¾ on disjoint sets E1 , . . . , E¾ ,

the direct sum M = M1 ⊕ · · · ⊕ M¾ is a matroid.
Proof: Since the E1 , . . . , E¾ are pairwise disjoint, the intersection of any I ∈ I
with each Ei is independent in Mi . If I1 , I2 ∈ I with | I2 | > | I1 | , then | I2 ∩ Ei | >
| I1 ∩ Ei | for some i. Since both sets are independent in Mi , we can augment I1 ∩ Ei
from I2 ∩ Ei and therefore I1 from I2 . Hence M1 ⊕ · · · ⊕ M¾ satisfies the augmen-
tation property.

Using a direct sum, we prove that the union of matroids is always a matroid,
and we compute the rank function.

11.3.52. Theorem. (Matroid Union Theorem; Edmonds–Fulkerson [1965],

Nash-Williams [1966]) If M1 , . . . , M¾ are matroids on E with rank functions
r1 , . . . , r¾ , then the union M = M1 ∪ · · · ∪ M¾ is a matroid with rank function
r(X) = min Y ⊆ X (| X − Y | + ∑ ri(Y)).
Proof: (following Schrijver [2003]). After proving the formula for the rank func-
tion, we will verify the submodularity property to prove that M is a matroid.
First we reduce the computation of the rank function to the computation of r(E).
In the restriction of the hereditary system M to the set X , we have I M | X = {Y ⊆
X : Y ∈ I M } and r M | X (Y) = r M (Y) for Y ⊆ X . Thus M | X = ∪i(Mi | X), and applying
the formula for the rank of the full union to M | X yields r M (X).
Consider a ¾-by-| E| grid of elements E in which the jth column Ej consists
of ¾ copies of the element e j ∈ E. We define two matroids N1 , N2 on E such that
534 Chapter 11: Extremal Problems

the maximum size of a set independent in both N1 and N2 equals the maximum
size of a set independent in M. We then compute r M (E) by applying the Matroid
Intersection Theorem to N1 and N2 . Let Mi be a copy of Mi defined on the ele-
ments E i of row i in E . Let N1 be the direct sum matroid M1 ⊕ · · · ⊕ M¾ , and let
N2 be the partition matroid induced on E by the column partition {Ej }.
Each set X ∈ I M has a decomposition into disjoint subsets X i ∈ Ii , since Ii is
an ideal. Given a decomposition {X i } of X ∈ I M , let X i be the copy of X i in E i .
Since {X i } are disjoint, ∪ X i is independent in N2 , and X i ∈ Ii implies that ∪ X i
is also independent in N1 . From X ∈ I M , we have constructed ∪ X i of size | X | in
I N1 ∩ I N2 . Conversely, any X ∈ I N1 ∩ I N2 corresponds to a decomposition of a set in
I M of size | X | when the sets X ∩ E i are transferred back to E, because N2 forbids
multiple copies of elements.

E
ej

1 • X1 • • • • • • • ← M1
Ei → • • • • X 2 • • • • ..
Ej .
• • • • • • • •
¾ • • • • • • • X ¾ • ← M¾

Hence r(E) = max{| I | : I ∈ I N1 ∩ I N2 }. Let the rank functions of N1 , N2 be

q1 , q2 , and let r i be the rank function of the copy Mi of Mi on E i . We have q1 (X ) =
∑ r i (X ∩ E i), and q2(X ) is the number of elements of E that have copies in X .
The Matroid Intersection Theorem yields r(E) = min X ⊆ E {q1 (X ) + q2(E − X )}.
By Corollary 11.3.49, the minimum is achieved by a set X such that E − X
is closed in N2 . The closed sets in the partition matroid N2 are the sets contain-
ing all or no copies of each element e j ; these are the unions of full columns of
E . Given X with E − X closed in N2 , let Y ⊆ E be the set of elements whose
copies comprise X . Then q2(E − X ) = | E − Y | , and X contains all copies of
the elements of Y , so q1 (X ) = ∑ r i (X ∩ E i) = ∑ ri(Y). We conclude r(E) =
min Y ⊆ E {| E − Y | + ∑ ri(Y)}.
To show that M is a matroid, we verify submodularity for r. Given X , Y ⊆ E,
the formula for r yields U ⊆ X and V ⊆ Y such that

r(X) = | X − U | + ∑ ri(U); r(Y) = | Y − V | + ∑ ri(V).

Since U ∩ V ⊆ X ∩ Y and U ∪ V ⊆ X ∪ Y , we also have

r(X ∩ Y) ≤ |(X ∩ Y) − (U ∩ V)| + ∑ ri(U ∩ V);

r(X ∪ Y) ≤ |(X ∪ Y) − (U ∪ V)| + ∑ ri(U ∪ V).

After applying the submodularity of each ri and the diagram below, these in-
equalities yield r(X ∩ Y) + r(X ∪ Y) ≤ r(X) + r(Y).
Section 11.3: Matroids 535

X 2 Y
1 0 1 0 1 0 1
U V
2

|(X ∩ Y) − (U ∩ V)| + |(X ∪ Y) − (U ∪ V)| = | X − U | + | Y − V |

The Matroid Union Theorem yields short proofs of formulas for packing and
covering problems. In each formula below, the optimal subset is closed, because
switching from X to (X) improves the numerator without changing the denomi-
nator. The graph corollaries were originally proved by difficult ad hoc arguments.

11.3.53. Corollary. (Matroid Covering Theorem; Edmonds [1965a])

In a loopless matroid M on E, the minimum number of independent sets
|X|
whose union is E is max∅
= X ⊆ E ⌈ r(X) ⌉.
Proof: Let M1 , . . . , M¾ be copies of M on E. The set E is the union of ¾ inde-
pendent sets in M if and only if E is independent in M = M1 ∪ · · · ∪ M¾ . By
the Matroid Union Theorem, r (E) ≥ | E| is equivalent to | E| − | Y | + ∑ ri(Y) ≥ | E|
for all Y ⊆ E. Since ri(Y) = r(Y) for all i, we conclude that E is the union of ¾
independent sets if and only if ¾r(Y) ≥ | Y | for all Y ⊆ E.

11.3.54. Corollary. (Nash-Williams [1964]) The number of forests needed to

| E(H)|
cover the edges of a graph G (its arboricity) is max H ⊆ G: |V(H)|≥2 ⌈ | V(H) ⌉
|−1 .
Proof: (Edmonds [1965a]) This follows immediately by applying Corollary 11.3.53
to M(G). The best lower bound arises from a connected induced subgraph H (cor-
responding to a closed set in M(G)).

11.3.55. Corollary. (Matroid Packing Theorem; Edmonds [1965b])

For a matroid M on E, the maximum number of pairwise disjoint bases equals
| E|−| X |
min X : r(X)<r(E)⌊ r(E) −r(X) ⌋ .
Proof: There are ¾ disjoint bases if and only if r (E) ≥ ¾r(E) in the union M of ¾
matroids M1 , . . . , M¾ that are copies of M on E. By the Matroid Union Theorem,
this requires | E| − | Y | + ∑ ri(Y) ≥ ¾r(E) for all Y ⊆ E. Since ri(Y) = r(Y) for all i,
there are ¾ disjoint bases if and only if | E| − | Y | ≥ ¾(r(e) − r(Y)) for all Y ⊆ E.

11.3.56. Corollary. (Nash-Williams [1961], Tutte [1961]) A graph G has ¾ edge-

disjoint spanning trees if and only if, for every partition P of V(G), at least
¾(| P | − 1) edges have endpoints in different parts.
Proof: (Edmonds [1965b]) We may assume G is connected. By Corollary 11.3.55,
we need | E| − | X | ≥ ¾(r(E) − r(X)) for each closed set X in M(G). The closed sets
correspond to partitions of V(G) into sets inducing connected subgraphs. For
each such partition V1 , . . . , V p , the corresponding closed set X is ⋃ E(G[Vi]) with
rank n − p. Since | E| − | X | counts the edges joining parts of the partition and
r(E) − r(X) ≥ p − 1, the graph has ¾ disjoint spanning trees if and only if the
condition holds.
536 Chapter 11: Extremal Problems

Corollary 11.3.56 implies that every 2 ¾-edge-connected graph has ¾ edge-

disjoint spanning trees, and this is sharp (Exercise 63). Nash-Williams and Tutte
proved Corollary 11.3.56 independently, by different methods, a few months
apart, and the two papers were published in the same volume of the same journal.

EXERCISES 11.3
11.3.1. (−) Let M be the hereditary system on [4] whose bases are {1 , 2} and {3 , 4}. Show
that all the properties listed in Definition 11.3.18 fail for M.
11.3.2. (−) For each family C below, determine whether it is the family of circuits of a
hereditary system on [6]. If it is, determine whether the system is a graphic matroid.
(a) C = {{1 , 2 , 3} , {3 , 4 , 5} , {5 , 6 , 1}}.
(b) C = {{1 , 2 , 3} , {3 , 4 , 5} , {1 , 2 , 4 , 5}}.
11.3.3. (−) Characterize the G graphs that yield matroids as follows:
(a) The stable sets of G are the independent sets of a matroid on V(G).
(b) The matchings in G are the independent sets of a matroid on E(G).
11.3.4. (−) Show that the stable sets of a graph need not be the independent sets of a
matroid by finding vertex-weighted graphs where the ratio between the set found by the
greedy algorithm and the maximum weight of a stable set is arbitrarily large.
11.3.5. (−) Explain how to obtain the rank function of a hereditary system from the bases.
11.3.6. (−) Let B be a base of a matroid. For e ∈ B, prove that B − º + e is a base if and
only if º belongs to the circuit formed by adding e to B.
11.3.7. (−) Let e be an element of a circuit C in a matroid. Prove that C is the unique
circuit created by adding e to some base.
11.3.8. (−) Prove the following implications directly for hereditary systems.
(a) The augmentation (I) and uniformity (U) properties are equivalent.
(b) The uniformity (U) and base exchange (B) properties are equivalent.
(c) Submodularity (R) implies weak absorption (A).
(d) Strong absorption (A ) implies base exchange (B).
(e) Augmentation (I) implies weak elimination (C).
11.3.9. (−) Prove that if r(X) = r(X ∩ Y) for some X , Y ⊆ E in a matroid on E, then
r(X ∪ Y) = r(Y).
11.3.10. (−) A set of | E| − r(E) circuits of a matroid on E form a fundamental set of
circuits if the elements e1 , . . . , e n can be ordered so that the last element of Ci is er(E)+i .
Prove that every matroid has a fundamental set of circuits. (Whitney [1935])
11.3.11. (−) Describe the circuits of a partition matroid M. Use this description to prove
directly that partition matroids satisfy the weak elimination property.
11.3.12. (−) Prove that every partition matroid is a transversal matroid.
11.3.13. (−) Determine whether the cycle matroid of G below is a transversal matroid.
c
• •
d
a b e º
G
• •
½
Exercises for Section 11.3 537

11.3.14. (−) Let B1 and B2 be bases of a matroid such that | B1  B2 | = 2. Prove that there
is a unique circuit C such that B1  B2 ⊆ C ⊆ B1 ∪ B2 .
11.3.15. (−) Prove that the cycle matroid of a graph G is the column matroid over 2 of
the vertex-edge incidence matrix of G.
11.3.16. (−) Use matroid duality to prove Euler ’s Formula for plane graphs.
11.3.17. (−) Let M be the hereditary system on E(K n) whose independent sets are the edge
sets of planar graphs. Determine whether M is a matroid.
11.3.18. (−) Determine whether a set can be a circuit and a cocircuit in the same matroid.
11.3.19. (−) Let C and C∗ be a circuit and a cocircuit in a matroid on n elements. Deter-
mine the minimum and maximum possible values of | C| + | C∗ | .
11.3.20. (−) Let M be a matroid on E, and fix A ⊆ E. Let I be the family of sets X ⊆ E
such that X ∈ I and X ∩ A = ∅. Prove that I is the family of independent sets of a matroid.
11.3.21. (−) Let r and be the rank function and span function of a matroid. Prove that
r(X) = min{| Y | : Y ⊆ X and (Y) = (X)}.
11.3.22. (−) Prove that a matroid of rank r has at least 2 r closed sets. (Lazarson [1957])
11.3.23. (−) Let G be an n-vertex graph, and let E1 , . . . , En−1 be a partition of E(G) into
n − 1 sets. Show that matroids can be used to test whether G has a spanning tree with
exactly one edge in each subset Ei .
11.3.24. (−) Given matroids M1 , . . . , M¾ on E, the Matroid Partition Problem asks
whether an input set X ⊆ E partitions into sets I1 , . . . , I¾ with Ii ∈ Ii . Prove that X is
partitionable if and only if | X − Y | + ∑ ri(Y) ≥ | X | for all Y ⊆ X .
11.3.25. (−) Use HAMILTONIAN PATH in directed graphs to prove that 3-MATROID
INTERSECTION is NP-complete.
11.3.26. (−) Use 3-D MATCHING to prove that 3-MATROID INTERSECTION is NP-
complete. Given disjoint sets V1 , V2 , V3 and a family of triples that each consists of one
element from each Vi , 3-D MATCHING is the problem of finding the maximum number of
triples that together use each element at most once. (In this terminology, ordinary bipar-
tite matching is 2-D MATCHING.)
11.3.27. Graphic matroids.
(a) Determine which graphic matroids are also uniform matroids.
(b) Determine which graphic matroids are also partition matroids.

11.3.28. (♦) The Fano matroid is the vector matroid of the matrix below over the 2-
element field 2 . Prove that the Fano matroid is neither a graphic matroid nor a tranver-
sal matroid. (Hint: Use the circuits.)
1001011
0101110
0010111

11.3.29. (♦) Let E be the edge set of a graph G. Say that a set X ⊆ E is weakly acyclic if
the spanning subgraph of G with edge set X has at most one cycle. Prove that the weakly
acyclic sets are the independent sets of a matroid on E.
11.3.30. Let s and t be vertices in a digraph D. Let E = V(D) − {s , t}. For X ⊆ E, let r(X)
be the number of edges from s ∪ X to X ∪ t. Prove that r is submodular.
538 Chapter 11: Extremal Problems

11.3.31. (♦) Submodularity for cycle matroids. Given a graph G, let ¾(X) be the number
of components of the spanning subgraph G X with edge set X . Let U and V be the sets
of components of G X and G Y , respectively. Let H be the U, V-bigraph with Ui adjacent
to Vj if Ui in G X and Vj in G Y have a common vertex. Relate the numbers of vertices,
components, and edges in H to the numbers ¾(X) , ¾(Y) , ¾(X ∩ Y) , ¾(X ∪ Y), and conclude
directly that the rank function of the cycle matroid of G is submodular. (Aigner [1979])

11.3.32. Without using other characterizations of matroids, prove directly that the base
exchange and dual base exchange properties are equivalent.

11.3.33. (♦) Prove the following implications directly for hereditary systems.
(a) The base exchange property (B) implies the augmentation property (I).
(b) The augmentation property (I) implies the absorption property (A).
(c) The strong absorption property (A ) implies the submodularity property (R) (with-
out using uniformity).

11.3.34. Using only linear dependence, prove that vector matroids satisfy the induced
circuit property: adding an element to a linearly independent set of vectors creates at most
one minimal dependent set.

11.3.35. (♦) Prove the following implications directly for hereditary systems.
(a) The base exchange property (B) implies the induced circuit property (J).
(b) The induced circuit property (J) implies the augmentation property (I).
(c) The induced circuit property implies the weak elimination property (C).

11.3.36. Prove that a hereditary system is a matroid if and only if it satisfies the following
“ultra-weak” augmentation property: If I1 , I2 ∈ I with | I2 | > | I1 | and | I1 − I2 | = 1, then
I1 + e ∈ I for some e ∈ I2 − I1 . (Chappell)

11.3.37. Given a matroid on a set E, let C1 , . . . , C¾ be circuits such that none is contained
¾
in the union of the others. Let X be a subset of E with | X | < ¾ . Prove that ⋃i=1 Ci − X
contains a circuit. (Welsh [1979])

11.3.38. (♦) Let e be an element in a matroid M. Prove that if C is a circuit in M · e, then

C or C + e is a circuit in M.

11.3.39. A refinement of a matroid M is a matroid N on the same elements such that

every circuit of M is a circuit of N . Prove that M has a refinement N different from M if
and only if no circuit of M has size r(M) + 1.

11.3.40. (♦) Let B1 and B2 be bases of a matroid M.

(a) Prove that for each e ∈ B1 , there exists ∈ B2 such that B1 − e + and B2 − + e
are bases of M. (Brualdi [1969])
(b) Use the cycle matroid M(K4) to show that there may be no bijection  : B1 → B2
such that setting =  (e) satisfies part (a) for all e ∈ B1 .
(c) Given X 1 ⊆ B1 , prove that for some X 2 ⊆ B2 both (B1 − X 1) ∪ X 2 and (B2 − X 2) ∪ X 1
are bases. (Brylawski [1973], Greene [1973], Woodall [1974], Greene–Magnanti [1975])

11.3.41. (+) Let B1 and B2 be bases of a matroid M. Prove that there is a bijection  : B1 →
B2 such that for each e ∈ B1 , the set B2 −  (e) + e is a base of M . (Brualdi [1969]) (Hint:
Define a B1 , B2 -bigraph by making e ∈ B1 and ∈ B2 adjacent when B2 + e − ∈ B.)


11.3.42. (♦) Let M be a matroid on a set E, and let w: E → 0 . Use the greedy algorithm
to prove a min-max formula for maximum weighted independent set: max I∈I ∑ e∈ I w(e) =
min ∑i r(X i), where the minimum is over all chains (by inclusion) of sets in E such that
each element e ∈ E appears in at least w(e) sets in the chain (sets may repeat in chains).
Exercises for Section 11.3 539

11.3.43. (♦) Circuits and cocircuits. Consider a matroid M and its dual M ∗ .
(a) Dual augmentation property. Given disjoint sets X and X ∗ with X ∈ I and
X ∈ I∗ , prove that there are disjoint sets B and B∗ with X ⊆ B ∈ B and X ∗ ⊆ B∗ ∈ B∗ .
∗
(b) Let e be an element of a base B. Prove that M has exactly one cocircuit disjoint
from B − e and that it contains e.
(c) Prove that the cocircuits of M are the minimal nonempty sets C∗ such that
| C∗ ∩ C|
= 1 for every C ∈ C.
(d) For distinct elements x and y of a circuit C, prove that there is a cocircuit C∗ such
that C∗ ∩ C = {x , y}. (Minty [1966])
11.3.44. The ¾-truncation M¾ of a matroid M is defined by I M¾ = {X ∈ I M : | X | ≤ ¾}.
(a) Prove that M¾ is a matroid.
(b) Prove that a matroid can be covered by t independent sets of size at most ¾ if and
only if max| X |≥1 min{|¾X,r(X)}
|
≤ t.
(c) Prove that a matroid of rank at least ¾ with ground set E has t pairwise disjoint
independent sets of size ¾ if and only if min X : r(X)<¾ |¾E−|−| X|
r(X)
≥ t. (Chen–Lai [1996])
11.3.45. The ¾-elongation of a matroid M is the hereditary system M ¾ whose bases are
the spanning sets of M with size ¾ .
(a) Prove that M ¾ is a matroid.
(b) Prove that (M¾)∗ = (M ∗)| E|−¾ if ¾ ≤ r(M), where M¾ is the ¾-truncation of M defined
in Exercise 11.3.44. (Welsh [1979])
11.3.46. Prove that a matroid has no circuits of size at most 2 if and only if (1) no element
is in every hyperplane, and (2) for any two elements some hyperplane contains exactly one.
11.3.47. Let M be a matroid on E. For F ⊆ E, prove that C M | F = {C ⊆ F : C ∈ C M } and
that C M.F is the set of minimal nonempty members of {C ∩ F : C ∈ C M }.
11.3.48. By duality and matroid restriction, prove r M.F (X) = r M (X ∪ F) − r M (F).
11.3.49. (♦) Prove that any minor of a matroid obtained by restricting and then con-
tracting can also be obtained by contracting and then restricting. In particular, if M
is a matroid on E and Y ⊆ X ⊆ E, prove that (M | X).Y = (M.X − Y)| Y and (M.X)| Y =
(M | X − Y).Y . (Hint: Use the rank function.)
11.3.50. (♦) Prove that the properties in Theorem 11.3.40 (involving the span function)
are equivalent and characterize matroids: incorporation (P), idempotence (S), transitivity
of dependence (T), and strong elimination (C ). (Hint: Prove U ⇒ P ⇒ S ⇒ T ⇒ C ⇒ C.)
11.3.51. Prove the following properties of closed sets of a matroid.
(a) The closed sets are the complements of the unions of cocircuits.
(b) The intersection of two closed sets is closed.
(c) The span of a set is the intersection of all closed sets containing it.
(d) The union of two closed sets need not be a closed set.
11.3.52. (+) Prove directly that in a hereditary system, the weak elimination property
implies the strong elimination property, using induction on | C1 ∪ C2 |. (Lehman [1964])
11.3.53. Given a matroid M on E and e ∈ E, the Shannon switching game (M, e) played
by Spanner and Cutter is as follows. On each round Cutter deletes an element of E − e, and
then Spanner seizes an element of E − e. Spanner wants a set that spans e; Cutter aims to
prevent this. Prove that Spanner has a winning strategy when there are disjoint subsets
X 1 , X 2 of E − e such that e ∈ (X 1 ) = (X 2). (Lehman [1964]) (Comment: A lengthy proof
using the Matroid Union Theorem (Theorem 11.3.52) shows that this sufficient condition
is also necessary. A special case using the cycle matroid for a union of two edge-disjoint
trees was marketed by Hasbro under the name “Bridg-It ” ; see West [2001, p.74].)
540 Chapter 11: Extremal Problems

11.3.54. (+) Given a matroid M , the base exchange graph (M) has a vertex for each
base of M , with two bases adjacent when their symmetric difference has size 2. Prove
that (M) is Hamiltonian when M has at least three bases. (Hint: Prove the stronger
statement that (M) has a spanning cycle through any edge.) (Holzmann–Harary [1972])
11.3.55. Use the formula r(X) = min J ⊆[m] {| A(J) ∩ X | + m − | J |} for the rank function of
the transversal matroid on E induced by subsets A1 , . . . , A m (Example 11.3.46) to prove
directly that the rank function satisfies r(∅) = 0 and r(X) ≤ r(X + e) ≤ r(X) + 1.
11.3.56. (♦) Given a bipartite graph G with E being one part , let M be the transversal
matroid on E whose independent sets are the subsets of E that can be covered by matchings
in G. By Ore’s Defect Formula (see Example 11.3.46), r(X) = min S⊆ X {| N(S)| + | X − S|}.
Prove directly that r is submodular.
11.3.57. Prove that restrictions and unions of transversal matroids are transversal ma-
troids, but contractions and duals of transversal matroids need not be.
11.3.58. (♦) Let M be the transversal matroid on E induced by subsets A1 , . . . , A m . Use
the Matroid Union Theorem to prove r M (X) = min Y ⊆ X {| X − Y | + | N(Y)|}.
11.3.59. (♦) Common independent and spanning sets.
(a) For matroids M1 and M2 on a set E, prove | I | + | S| = r1 (E) + r2(E), where I is a
largest common independent set and S is a smallest common spanning set.
(b) Let G be an n-vertex bipartite graph with no isolated vertices. Prove  (G) +
(G) = n (Gallai’s Theorem, Exercise 6.1.50).
(c) Let G be a bipartite graph with no isolated vertices. Without using other results,
use part (a) directly to prove (G) = (G) (K önig ’s Other Theorem, Theorem 6.1.16).
11.3.60. (♦) Use the Matroid Intersection Theorem to prove that in every acyclic digraph,
the vertices can be covered by at most  pairwise disjoint paths, where  is the indepen-
dence number of the underlying graph. (Chappell)
11.3.61. Given matroids M1 and M2 whose families of spanning sets are S 1 and S 2 ,
prove that the matroid (M1∗ ∪ M2∗)∗ is the hereditary system whose spanning sets are
{X 1 ∩ X 2 : X 1 ∈ S 1 , X 2 ∈ S 2 }.
11.3.62. Matroid Intersection from Matroid Union.
(a) Without the Matroid Intersection Theorem, prove that the maximum size of a com-
mon independent set in matroids M1 and M2 on E is r M1 ∪ M ∗ (E) − r M ∗ (E).
2 2
(b) Prove the Matroid Intersection Theorem by applying Matroid Union to M1 ∪ M2∗ .
11.3.63. (♦) Connectivity and spanning trees.
(a) Let F be a set of at most  edges in a 2 -edge-connected graph G. Use Corollary
11.3.56 to prove that G − F has  edge-disjoint spanning trees. (Nash-Williams [1961])
(b) For each  , construct a (2  − 1)-edge-connected graph that does not have  edge-
disjoint spanning trees.
11.3.64. Colored trees and b-detachments.
(a) Let G be a connected edge-colored graph with color classes E1 , . . . , E¾ . Prove that
G has a spanning tree with distinct colors if and only if G − F has at most t + 1 components
whenever F consists of t color classes. (Hint: Use the Matroid Intersection Theorem.)
(b) A split replaces a vertex x with two new vertices x1 and x2 whose neighborhoods

in the new graph partition N(x). Given b: V(G) → , a b-detachment of a graph G arises
by performing splits until there are b(v) copies of each vertex v. Use part (a) to prove that
G has a connected b-detachment if and only if for all U ⊆ V(G), graph G − U has at most
 (U) + 1 − b(U) components, where b(U) = ∑v∈U b(v) and  (U) is the number of edges inci-
dent to U. (Schrijver [2003, p. 704]; see Nash-Williams [1985] for a more general result.)
Chapter 12

Partially Ordered Sets

In a totally ordered set, such as the natural numbers under “ ≤ ”, any two
elements are related. Other orderings are “partial”; examples include the di-
visibility relation on positive integers and the inclusion relation on subsets of a
set. Partially ordered sets can model precedence constraints in scheduling, pref-
erences among objects, partial information about numbers being sorted, etc.

12.1. Structure of Posets

DEFINITIONS AND EX AMPLES

12.1.1. Example. The subsets of a finite set, ordered by inclusion, form a natu-
ral poset. We spent most of Chapter 11 studying aspects of it. Elementary under-
standing of containment yields three natural properties: (1) A ⊆ A, (2) A ⊆ B
and B ⊆ A together imply A = B, and (3) If A ⊆ B and B ⊆ C, then A ⊆ C.
Similarly, the divisibility relation on the divisors of an integer N satisfies:
(1) x | x, (2) x | y and y | x imply x = y, and (3) If x | y and y | , then x | .

12.1.2. Definition. A relation R on a set X is a subset of the cartesian product

X × X . We write (x , y) ∈ R or xRy or say (x , y) satisfies R. An order relation
(or partial order) on X is a relation that is
reflexive (xRx for all x),
antisymmetric (xRy and yRx imply x = y), and
transitive (xRy and yR imply xR ).
A partially ordered set (or poset) is a set P plus an order relation on P.
When (x , y) satisfies the relation, we write x ≤ P y or simply x ≤ y. If x ≤ y
or y ≤ x, then x and y are comparable (or related) in P ; otherwise they are
incomparable (or unrelated). All of ≤ , <, ≥ , > describe ways that elements
can be comparable. We write x y to mean that x and y are incomparable.

The word “poset ” illustrates the evolution of terminology. Originally it was

written as “PO-set ”, emphasizing grammatically that a poset is a set equipped
with a partial order.

541
542 Chapter 12: Partially Ordered Sets

Like any binary relation, an order relation is specified by a 0 , 1-matrix, with

1 in position (x , y) if x ≤ y. To facilitate study, we seek visual representations.
Since an order relation on P is a set of ordered pairs from P , we can view it as
a directed graph with vertex set P. Ignoring which way pairs are ordered treats
the edges as unordered pairs.

12.1.3. Definition. The comparability digraph of a poset P is the digraph

whose vertices are the elements of P and whose edges are the ordered pairs
xy such that x ≤ P y. The comparability graph is the graph whose vertices
are the elements of P and whose edges are the unordered pairs of distinct
vertices that are comparable in P.

12.1.4. Remark. Comparability graphs and digraphs. The comparability digraph

specifies a poset completely. Since order relations are reflexive, a comparability
digraph has a loop at each vertex. Since this is understood, we omit the loops
when drawing a comparability digraph and just draw the digraph of the strict
order relation: the pairs xy with x < y.
The comparability graph discards information and does not specify the poset.
Below we show distinct posets with the same comparability graph. Reversing all
comparable pairs also yields another poset with the same comparability graph.
A comparability graph becomes a comparability digraph by transitively ori-
enting the edges. A transitive orientation of a graph G is an orientation such
that if xy and y are (directed) edges, then x is an edge; that is, x and are
adjacent in G and the orientation directs the edge from x to . The transitive ori-
entations of G correspond to the posets for which G is the comparability graph.
Exercise 6 gives a necessary condition for comparability graphs that turns
out also to be sufficient. Another characterization in Exercise 7 leads to a fast
algorithm for testing whether a graph is a comparability graph. Nevertheless,
our focus in this chapter is on the order-theoretic properties of posets, so we will
not study comparability graphs further.

d d
• •
• •
• a• •b
c
• •
a• •b •
c
For a useful visual presentation, transitivity of the order relation allows us
to describe a poset completely while drawing only some of the comparable pairs.

12.1.5. Definition. If x < y in P and there is no with x < < y, then y covers
x (in P), written as x ≺ y or y ! x. The cover relation is the set of pairs
(x , y) such that x ≺ y. The cover digraph is the digraph on the elements of
P whose edge set is {xy: x ≺ y}.
A cover diagram (or Hasse diagram or diagram) of P is obtained by
erasing the directions on edges after drawing the cover digraph in the plane
such that each (straight-line) edge points upward. The cover graph is the
graph on the elements of P whose edge set is {xy: x ≺ y or y ≺ x}.
Section 12.1: Structure of Posets 543

12.1.6. Example. The diagram. The comparability digraphs in Remark 12.1.4

direct all edges upward. Erasing the edges implied by transitivity yields the cover
digraphs. Since the edges all point upward, erasing the arrowheads then yields
the diagrams, which appear on the left below. One of these is the poset of subsets
of {1 , 2}, ordered by inclusion.
[3]
• • •
{1 , 2}
• {1
• , 3} •{2 , 3}
• • •
{1}
• •{2} •{3}
• • • •
∅

On the right is the poset of subsets of [3], ordered by inclusion. The diagram
specifies the poset completely; x < y in P if and only if we can reach y from x in
the diagram by moving only upward along edges. Omitting the edges implied by
transitivity makes it easier to see the structure. By the convention that the edges
for the order relation point upward, the diagram provides the cover digraph and,
by transitive closure, the comparability digraph.

The cover graph (but not the diagram) discards the order information and
does not specify the poset. For example, C4 is the cover graph of two posets (one
shown above), and C6 is the cover graph of seven posets (Exercise 3). In contrast
to comparability graphs, testing whether a graph is a cover graph is NP-complete
(Nešet řil–Rödl [1987, 1993]; Brightwell [1993] gave a simpler proof).
The simplest posets are totally ordered or totally unordered. Indeed, we say
“partial order ” as a generalization of total order.

12.1.7. Definition. A chain is a poset whose elements are linearly ordered so

that x < y if and only if x comes before y in that order. The chain with ¾
elements is denoted k and called a ¾-chain. An antichain is a poset in which
no two elements are comparable.
An element of a poset is maximal if no other element is greater than it,
minimal if no other element is less than it.

The first poset in Remark 12.1.4 has one maximal element (d) and two min-
imal elements (a and b). A ¾-chain has one maximal and one minimal element.
In an antichain, every element is maximal and minimal. The cover graph of k is
the path P¾ ; its comparability graph is K ¾ .

12.1.8. Definition. Posets P and Q are isomorphic if some bijection from the
set P to the set Q preserves the order relation. A subposet of a poset P is a
poset R on a subset of P defined by restricting the comparability relation to R.
If Q is isomorphic to a subposet of P , then P contains Q or Q embeds in P.
The dual of P , written P ∗ , is the poset on the elements of P defined by
y ≤ P ∗ x if and only if x ≤ P y. A poset isomorphic to its dual is self-dual. A
poset is finite if it has finitely many elements.
544 Chapter 12: Partially Ordered Sets

The notion of isomorphism allows us to view a diagram of a poset as the poset

itself, just as a drawing is a graph. We use k for a ¾-element chain, as an isomor-
phism class, just as C¾ denotes a ¾-vertex cycle.

12.1.9. Example. Subposets. If n ≤ m, then n embeds in m, and all n-element

subposets of m are isomorphic to n.
We write 2 n for the inclusion poset on the subsets of [n], rather than 2[n] as in
Section 11.3; we will soon explain why. In 23 (Example 12.1.6) we find six chains
of size 4, two antichains of size 3, and 15 subposets isomorphic to 2 2 (Exercise 2).
Furthermore, 23 is self-dual, via complementation of subsets of [3].
A chain in a poset is a set of pairwise comparable elements; an antichain is a
set of pairwise incomparable elements. They become cliques and independent sets
in the comparability graph. When x1 , . . . , x ¾ is a chain in P , with x1 < · · · < x¾ ,
it need not hold that x i+1 covers x i . For example, 23 contains 19 chains of size 2.

12.1.10. Remark. Subposet in posets corresponds to induced subgraph in graph

theory. When G is the comparability graph of P , and Q is a subposet of P , the
comparability graph of Q is the subgraph of G induced by the set Q. However,
the cover graph of Q need not be a subgraph of the cover graph of P.
A graph consists of a vertex set and an adjacency relation on it. Similarly, a
poset consists of a set of elements with an order relation on it. Nevertheless, we
generally use the same notation (P) for a partially ordered set and for its set of
elements. This abuse of notation works because we treat P as a partially ordered
set and study as “subposets” only the structures that inherit all comparabilities.
For graphs we manipulate both vertices and edges, so we write v ∈ V(G) and
e ∈ E(G) instead of v ∈ G and e ∈ G to avoid ambiguity. For posets, we just write
x ∈ P when x is in the set of elements, and we use | P | for the size of that set. We
rarely consider analogues of edge deletion or contraction for posets.

12.1.11. Definition. Let P be a poset. Its width w(P) is the size of a largest
antichain in P. Its height h(P) is the size of a largest chain in P. Its length
is one less than its height.

12.1.12. Example. The poset 23 has width 3, height 4, and length 3. An an-
tichain of size w(P) in P is a maximum antichain, but a maximal antichain
(one not contained in a larger antichain) may be smaller. For example, the an-
tichain {{1} , {2 , 3}} in 23 is a maximal antichain but not a maximum antichain.
A chain in a poset with finite height is a maximal chain if and only if it
extends from a minimal element to a maximal element and its successive pairs
satisfy the cover relation. For elements, “maximal” and “minimal” refer to the
order relation in P , not to containment on sets of elements.

12.1.13. Definition. In a poset P , a down-set is a subset D such that x ∈ D and

y < x imply y ∈ D. The complement of a down-set is an up-set. The down-set
D[x] generated by x is {y ∈ P: y ≤ x}. The down-set D[A] generated by a
family A is {y: y ≤ x for some x ∈ A}. We also write D(x) = {y ∈ P: y < x}.
Similarly define U[x], U[A], and U(x).
Section 12.1: Structure of Posets 545

Down-sets have also been called ideals or order ideals, and up-sets have
been called dual ideals or filters. The terms “down-set ” and “up-set ” are less
sophisticated, but they are explicit and do not conflict with other uses of “ideal”
and “filter ” in mathematics. In Chapter 11, we used “ideal of sets” for a down-set
in 2 n because this is common and is consistent with algebraic notions.
In 23 , the down-set generated by {1 , 23} is {1 , 23 , 2 , 3 , ∅}. There are exactly
20 antichains and 20 down-sets in 23 ; one of each is empty.

12.1.14. Remark. There is a one-to-one correspondence between the antichains and

the down-sets in any poset. The natural bijection maps an antichain A to the down-
set D[A]. The inverse assigns to each down-set the antichain consisting of its
maximal elements. The empty antichain corresponds to the empty down-set.

The product operation is a fundamental way to combine posets.

12.1.15. Definition. The product P × Q of two posets P, Q is the poset on

{(x , y): x ∈ P, y ∈ Q} defined by (x , y) ≤ (x , y ) if and only if x ≤ x and
y ≤ y . The sum P + Q consists of disjoint copies of P and Q with no compa-
rabilities between them.
• •
• • •
• • • • • •
• • • •
P Q • • P×Q • •

12.1.16. Remark. Products of posets. The product P × Q has disjoint copies of Q

for each element of P and disjoint copies of P for each element of Q. If the cover
graphs of P and Q are G and H , then the cover graph of P × Q is the cartesian
product G H. However, the comparability graph of P × Q is not the cartesian
product of the comparability graphs of P and Q.
In the sense of isomorphism, the operation is commutative and associative,
and we write the product of n copies of P as P n .

12.1.17. Example. The subset poset. The poset 2 is a 2-element chain; call its el-
ements 0 and 1, with 0 < 1. With each element of [n], associate a copy of 2. Let
P be the poset of subsets of [n], ordered by inclusion. Mapping each member of
P to its incidence vector expresses P as 2 n . Thus we denote P as 2 n . (We suggest
pronouncing 2 n as “ 2 sup n”.) For the inclusion poset on the subsets of a set X ,
we write 2 X and call X the ground set. The poset 2 n is also called the Boolean
algebra on n elements, sometimes denoted Bn .
Since 2 n is the product of n copies of 2, its cover graph is the n-dimensional
cube. Note that if x is a ¾-set and y is an l-set in [n] with x < y, then the sub-
poset of 2 n consisting of { : x ≤ ≤ y} is isomorphic to 2 l−¾ . Also, 2 n is self-dual
(via complementation of subsets of [n]). Finally, when writing elements of 2 n as
sets or as their incidence vectors, we typically drop set brackets, commas, and
parentheses and just write strings of elements.
546 Chapter 12: Partially Ordered Sets

DILWORTH’S THEOREM AND BEYOND

The archetypal extremal problem for posets is finding a largest antichain.

Explicit answers exist only for special classes; we discuss these later. First, we
obtain a min-max relation. Since an antichain has no two elements on any chain,
the width is bounded by the number of chains needed to cover the elements. Dil-
worth proved that equality holds.

12.1.18. Theorem. (Dilworth’s Theorem; Dilworth [1950]) If P is a finite

poset, then the maximum size of an antichain in P equals the minimum num-
ber of chains needed to cover the elements of P.
Proof: (Perles [1963]) We use induction on | P | , with trivial basis. For | P | > 1,
suppose first that some largest antichain A omits both a maximal element and
a minimal element of P. Thus neither the down-set D[A] nor the up-set U[A]
generated by A contains all of P , and we can apply the induction hypothesis to
each subposet. Also, they share only A.
Let ¾ = w(P). Since both D[A] and U[A] are subposets of P , they have width
at most ¾ . Since they both contain A, equality holds. From the induction hypoth-
esis, we obtain ¾ chains covering D[A] and ¾ chains covering U[A], the former
with elements of A at the top and the latter with elements of A at the bottom.
These combine to form ¾ chains. These chains cover all of P , because the maxi-
mality of A implies that D[A] ∪ U[A] = P. This case is shown on the left below.

U[A]
• • •x
A • • • • • • • • • •
• • •y
D[A]

In the remaining case, every maximum antichain of P consists of all maximal

elements or all minimal elements. Thus w(P − {x , y}) ≤ ¾ − 1 if x is a minimal
element and y is a maximal element. Choose a maximal element x and a minimal
element y below it (they may be equal). Since P − {x , y} is smaller (and has width
¾ − 1), we can apply the induction hypothesis to cover P − {x , y} with ¾ − 1 chains
and add the chain consisting of x and y to complete the desired covering.

Since subposets of chains are chains, we conclude that P is covered by w(P)

pairwise disjoint chains. A partition of P into chains is a chain decomposition
of P. In honor of Theorem 12.1.18, a smallest chain decomposition (size w(P)) is
called a Dilworth decomposition of P.
Perles’ proof is like Pym’s proof of Menger ’s Theorem. Each uses splicing of
paths, with a special argument for a degenerate case. Here we seek a maximum
cut (the antichain) instead of a minimum cut. In the language of network flows,
the arguments are essentially equivalent.
Section 12.1: Structure of Posets 547

Meanwhile, we relate Dilworth’s Theorem to the min-max relations of graph

theory. A poset is bipartite if every element is minimal or maximal (or both);
equivalently, the comparability graph is bipartite.

12.1.19. Theorem. (Fulkerson [1956]) Dilworth’s Theorem is equivalent to the

K önig–Egerváry Theorem on matchings in bipartite graphs.
Proof: To apply Dilworth’s Theorem to bipartite matching, view an n-vertex bi-
partite graph G as a bipartite poset. The vertices of one part become maximal
elements; the others are minimal. Chains have one or two elements. Every cover-
ing of the poset by n − ¾ chains uses ¾ disjoint 2-chains; these yield a matching of
size ¾ in G. Each antichain corresponds to an independent set in G; if it is maxi-
mal, then the remaining vertices form a vertex cover. Hence Dilworth’s equality
between the sizes of a maximum antichain and minimum chain-covering yields a
matching and a vertex cover of the same size in G. Since every vertex cover is at
least as large as every matching, this proves the K önig–Egerváry Theorem.
For the converse implication, consider any poset P of size n. We define a bi-
partite graph S(P), the split of P , in which to study matchings. For each element
x ∈ P , create two vertices x− and x+ (see figure below). The parts of the biparti-
tion of S(P) are {x− : x ∈ P} and {x+ : x ∈ P}. The edge set is {x− y+ : x <P y}.
Every matching in S(P) yields a chain-covering in P as follows: if x− y+ is in
the matching, then y is immediately above x on a chain in the cover. If x− or x+
is unmatched, then x is the top or bottom of its chain, respectively. Since each
vertex of S(P) appears in at most one edge of the matching, this defines disjoint
chains covering P. If the matching has ¾ edges, then the cover has n − ¾ chains,
since each additional edge links the top of one chain with the bottom of another
to form a single chain.
Given a vertex cover R of S(P), let A = {x ∈ P: x− , x+ ∈ / R}. A relation
between elements of A would yield an edge of S(P) uncovered by R, so A is an
antichain. No minimal vertex cover of S(P) uses both of {x+ , x−}, because by
transitivity the sets { − : ∈ D(x)} and {y+ : y ∈ U(x)} induce a complete bipar-
tite subgraph in S(P). A vertex cover of a complete bipartite graph must use all
of one part. Since x+ and x− have no other neighbors, we can drop from R the one
of {x+ , x−} that is not in that part.
Thus a minimum vertex cover of size  yields an antichain of size n −  . Now
the existence of a matching and a vertex cover in S(P) of equal size yields an an-
tichain and a chain-covering of equal size in P.

d e a+ b+ c+ d+ e+  +
• • • • • • • •
•c •
• • • • • • • •
a b a− b− c− d− e−  −

12.1.20. Remark. A simpler min-max relation holds for maximum chains and
minimum antichain coverings (Mirsky [1971]). The antichain of maximal ele-
ments intersects each maximal chain, so by induction on h(P) we can cover P
with h(P) antichains. Via the Perfect Graph Theorem (Corollary 8.3.35), this
easy observation implies Dilworth’s Theorem.
548 Chapter 12: Partially Ordered Sets

Dilworth’s original proof was a complicated induction, cutting and pasting

chains. An appropriate generalization is simpler. Here (D) denotes the indepen-
dence number of the graph obtained by viewing the edges as unordered pairs.

12.1.21. Theorem. (Gallai–Milgram [1960]) The vertices of an n-vertex digraph

D can be covered using at most (D) disjoint paths.
Proof: Since V(D) is covered by n disjoint paths of length 0, it suffices to prove
a stronger claim: If C is a set of disjoint paths covering V(D), with > (G),
and S is the set of sources (initial vertices) of these paths, then V(D) can be cov-
ered using fewer than disjoint paths with sources in S. The statement holds
vacuously when n = 1; we proceed by induction. The added requirement about
sources simplifies the induction step.
Suppose n > 1. Since > (D), there is an edge xy with x , y ∈ S. Let A and
B be the paths in C starting with x and y. If A has no edge, then moving x to the
beginning of B saves one path. Hence A has an edge x . Deleting x from A yields
a cover C of V(D − x) by paths with sources in S , where S = S − x + . Since
(D − x) ≤ (D), the induction hypothesis yields a cover C of V(D − x) by fewer
than paths with sources in S .
All of S is in S except . If  is a source in C , then prepend x to the path
starting at  to cover V(D) using fewer than paths. If  is not a source but y
is, then prepend x to the path starting at y. If neither y nor  is a source, then
|S | = implies that C has at most − 2 paths; let x be a path by itself. In all
cases, the resulting paths are disjoint and have sources in S.
x
•
S
•y •
S
In the special case where D is the comparability digraph of a poset P , the
initial covering is the set of | P | trivial paths. Since (D) = w(P), the paths given
by Theorem 12.1.21 form the desired chain-covering, a Dilworth decomposition.
The special case where D is acyclic is in Exercise 11.3.60.
Dilworth’s Theorem also generalizes beyond antichains.

12.1.22. Definition. A -family in a poset is a subposet containing no ( + 1)-

chain. The -norm of a set partition C1 , . . . , Cm is ∑i=1 min{ , | Ci |}. A chain
m

decomposition of a poset P is -saturated if its -norm equals the maximum

size of a -family in P.

Dilworth decompositions are 1-saturated partitions. By Mirsky ’s observa-

tion, -families are just the unions of (at most) antichains. Each chain C in a
chain partition contributes at most min{ , | C|} to a -family. Hence the size of a
-family is bounded by the -norm of a chain partition.

12.1.23. Theorem. (Greene–Kleitman Theorem; Greene–K leitman [1976a])

For every poset P and natural number , there is a chain decomposition of P
that is both -saturated and ( + 1)-saturated.
Exercises for Section 12.1 549

The original proof of the Greene–K leitman Theorem was quite long. There
are now shorter combinatorial proofs(Saks [1979]) and proofs applying other min-
max relations (such as Frank [1980b] via minimum-cost network flow). We leave
these to a more advanced book.
For acyclic digraphs, there are extensions of the Greene–K leitman Theorem
due to Linial [1981], Saks [1980], Cameron [1982], and Hoffman [1983]. Berge
conjectured an extension for arbitrary digraphs. Let C be a partition of the ver-
tices of a digraph D using paths. A partition with smallest ¾-norm is ¾-optimal.
On the other hand, a partial ¾-coloring is a union of ¾ independent sets of ver-
tices; view each set as a color class. The number of colors on a path C is bounded
by min{¾ , | V(C)|}. We seek a partial ¾-coloring where equality holds.

12.1.24. Conjecture. (Berge [1982]) For every digraph D, integer ¾ , and ¾-

optimal path partition C of D, there is a partial ¾-coloring such that each
path C in C intersects min{¾ , | V(C)|} color classes.

Berge [1985] noted that the Greene–K leitman Theorem is the special case for
transitive digraphs; he proved it also for ¾ = 1 in general. Cameron [1986] proved
it for acyclic digraphs, Berger–Hartman [2008] proved it for ¾ = 2, and Berger–
Hartman [2012] gave a unified proof for various cases using network flows. The
general conjecture remains open; Hartman [2006] surveyed partial results.

EXERCISES 12.1

12.1.1. (−) For n = 3 and n = 4, list the isomorphism classes of posets with n elements.
Determine how many are self-dual.
12.1.2. (−) Show that 23 contains 15 copies of 2 2 .
12.1.3. (−) Draw the diagrams of the seven posets (isomorphism classes) whose cover graph
is a 6-cycle. Determine their widths.
12.1.4. (−) By Remark 12.1.20, every ¾-family decomposes into ¾ antichains. Use the
poset below to show that it may not be possible to obtain a maximum 2-family by adding
an antichain to a maximum antichain.
• •
• • • • •
• •

12.1.5. Prove that the graphs below are comparability graphs by exhibiting cover dia-
grams of posets for which they are the comparability graphs. Use the necessary condition
in the next exercise (12.1.6) to prove that their complements are not comparability graphs.

• • •
• •
• •
• •
• • • • • •
• • • • •
550 Chapter 12: Partially Ordered Sets

12.1.6. (♦) For a closed walk [x1 , . . . , x¾ ] in a graph, a triangular chord is an edge of
the form xi xi+2 (with indices modulo ¾). Prove that in a comparability graph, every closed
walk of odd length has a triangular chord. Conclude that the complement of the cycle Cn is
not a comparability graph when n ≥ 5. (Comment: Gilmore–Hoffman [1964] showed that
this necessary condition is also sufficient.)
12.1.7. (♦) For a graph G, the Ghouilà-Houri graph G is the graph defined on the or-
dered pairs of adjacent vertices of G by putting (x , y) ↔ (y , ) in G if and only if x ∈
/ E(G).
Prove that G is bipartite if and only if every closed odd walk of G has a triangular chord.
(Comment: With Exercise 12.1.6, this yields a polynomial-time recognition algorithm for
comparability graphs.) (Ghouilà-Houri [1962])
x
•u
xu
• • •w
yx
• • y
•x
G vy • • yv G
• • • • xy
• y • y
w v
w
• • • ux
x

12.1.8. (♦) Prove that a graph is the cover graph of some poset if and only if it has an
acyclic orientation without dependent edges, where a dependent edge in a digraph is an
edge whose reversal completes a cycle. Conclude that if the chromatic number of a graph
is less than its girth, then it is a cover graph.
12.1.9. Use Exercise 12.1.8 to prove that the Gr ötzsch graph below is not a cover graph.
(Fisher–Fraughnaugh–Langley–West [1997])
•
•
• • • •
•

• •
• •

12.1.10. A digraph D is a cover digraph if E(D) is the cover relation of some poset. De-
rive a polynomial-time checkable characterization of cover digraphs. Assume that one can
find in polynomial time all vertices reachable from a specified vertex. (Comment: No effi-
cient algorithm is known to check whether an undirected graph is a cover graph.)
12.1.11. Obtain simple formulas (constant number of terms) for
(a) the number of chains of size 2 and the number of chains of size 3 in 2 n .
(b) the numbers of antichains of size 2 and size 3 in 2 n . (Popadić [1970])
12.1.12. A cutset in a poset is a subset intersecting every maximal chain. The family
[n]
(⌈n/2⌉) is a minimal cutset in 2 n ; is it a largest one? (Füredi–Griggs–Kleitman [1989])
12.1.13. There are 16 sentences of the form “In every poset , A B chain intersects C D
antichain” , where A , C ∈ {some, every} and B , D ∈ {maximal, largest}. For each such
sentence, determine whether it is true or false.
12.1.14. Suppose that the red subgraph in a red/blue-coloring of E(K n) has a transitive
orientation.
√ Prove that the coloring has a monochromatic complete subgraph with at least
n vertices. Show that the bound is sharp.
Exercises for Section 12.1 551

12.1.15. (♦) Prove that a poset of size greater than mn has a chain of size greater than
m or an antichain of size greater than n. Use this to prove the Erdős–Szekeres Theorem:
every list of mn + 1 distinct integers has an increasing sublist with more than m elements
or a decreasing sublist with more than n elements.
12.1.16. (♦) A family of sets is union-free if it has no two distinct members whose union
is a third member. Moser asked for º (n), the maximum size of a union-free subfamily that
can be guaranteed to exist in any family of n sets.√
(a) Use Dilworth’s Theorem to prove º (n) ≥ n. (Riddell, Erdős–Komlós [1970])
√
(b) Prove º (n) > 2n − 1. (Erdős–Shelah [1972], Kleitman [1973])
(c) Let Ai , j consist of the integers from −i to + j , and let F = {Ai , j : (i , j) ∈ [t]2 }. Prove
√
that the maximum size of a union-free family in F is 2t − 1. Thus º (n) ≤ 2 n − 1 when n
is a perfect square. (Erdős–Shelah [1972]) √
(Comment: Fox–Lee–Sudakov [2012] proved º (n) = ⌊ 4n + 1⌋ − 1 for all n.)

12.1.17. Given n1 , . . . , n¾ ∈ , let P denote the poset whose elements are the -tuples x
with xi ∈ [ni ] for all i, ordered by x < y if and only if xi < yi for all i. Prove that every
maximal antichain in P is a maximum antichain, and determine its size. (Tsai [2017])
12.1.18. Exercise 6.1.23 used Hall’s Theorem to prove that if G is an X , Y -bigraph with no
isolated vertices, and d(x) ≥ d(y) whenever xy ∈ E(G) with x ∈ X and y ∈ Y , then G has a
matching that covers X . Use this to show that the family of subsets of [n] decomposes into
(⌈n/2⌉
n
) inclusion chains, which therefore form a Dilworth decomposition of 2 n . (F. Galvin)
12.1.19. Another proof of Dilworth’s Theorem. The poset on the right below arises from
that on the left by deleting the central covering pair. All other comparabilities remain.
(a) Suppose that x covers y and  in P. Let Q and R be the posets obtained by deleting
(y , x) and ( , x), respectively, from the set of relations in P. Prove that min{w(Q) , w(R)} =
w(P). (Hint: Take maximum antichains in Q and R, and consider the maximal elements
and the minimal elements of the union of these antichains as a subposet of P.)
(b) Use part (a) to prove Dilworth’s Theorem. (Harzheim [1983])
• •
• • • •
• • • •
• •

12.1.20. For each poset below and each , find a chain partition that is both -saturated
and ( + 1)-saturated. Is some chain partition -saturated for all ?
•
• • •
• •
• • • • • •
• • •
• • • • • •
• •
• • •
•

12.1.21. (♦) Let d ¾ be the maximum size of a -family in a poset P , and let ¾ = d ¾ − d ¾−1 .
(a) In a -saturated chain partition C of P , let  be the number of chains of size at least
. Prove ¾ ≥  ≥ ¾+1 , (Comment: The Greene–Kleitman Theorem thus implies 1 ≥
· · · ≥ h , where P has height h. No direct proof is known.) (Greene–Kleitman [1976a])
(b) Prove  = ¾ when C is both -saturated and ( − 1)-saturated.
552 Chapter 12: Partially Ordered Sets

12.1.22. A poset P is polyunsaturated (West [1986]) if for 1 ≤ ¾ < l − 1 < h(P) no chain
partition of P is both ¾-saturated and l-saturated.
(a) Use the Greene–Kleitman Theorem and Exercise 12.1.21 to prove that h(P) ≤
w(P) + 2 when P is polyunsaturated.
(b) Construct a poset P¾ iteratively as follows. For ¾ = 1, let P¾ = 3. For ¾ > 1, ob-
tain P¾ from P¾−1 by adding a chain C of ¾ + 1 new elements and making the element just
below the maximal element on C cover the element just below the maximal element of the
(unique) longest chain in P¾−1 . The diagrams of P2 and P3 appear below. Prove that P¾
is polyunsaturated. (Comment: Thus the maximum height of a polyunsaturated poset of
width ¾ is ¾ + 2.) (Chappell [2002])
•
• • •
• • • • •
• • • • •
• •

12.2. Symmetric Chains and LYM Orders

Although Dilworth’s Theorem applies to all (finite) posets, much of the study
of partially ordered sets concerns special families of posets. When suitable con-
straints are placed on posets, much more can be said about their maximum an-
tichains and other structural aspects.

GRADED POSETS

In this section we consider only finite posets. This property makes the notion
of the “rank” of an element well defined.

12.2.1. Definition. In a poset, the rank of an element x, written r(x), is the max-
imum length of a chain having x as its top element. A poset P is graded if all
its maximal chains have the same length, and its rank r(P) is that length.
• •2 •2
•
• •1 •1
•
• •0 •0
not graded graded

In a graded poset, r(x) = r(y) + 1 when x covers y. The rank of the poset is
the rank of its maximal elements. The notion of rank function can be extended
to more general posets, but we will study only graded posets in this section.
Section 12.2: Symmetric Chains and LYM Orders 553

12.2.2. Definition. If P is graded, then the elements with rank ¾ are the ¾ th
rank or ¾ th level P¾ , and we write N¾ (P) for the rank size | P¾ | (also called
the ¾ th Whitney number of P). A graded poset is rank-symmetric if N¾ =
Nr(P)−¾ for all ¾ . It is rank-unimodal if there is a rank ¾ such that Ni ≤ Nj
whenever i ≤ j ≤ ¾ or i ≥ j ≥ ¾ . The rank generating function is the
formal power series ∑¾≥0 N¾ x¾ .

12.2.3. Example. Subsets. The poset 2 n is graded, with r(x) = | x| and r(2 n) =
n. The ¾ th rank is ([n]
¾ ). The poset is rank-symmetric and rank-unimodal. Since
N¾ (2 n)
= ¾ the rank generating function is (1 + x)n .
( n
),
Every maximal chain in 2 n has length n (size n + 1). There are maximal an-
tichains of size 1, but for maximum antichains we will again prove Sperner ’s
Theorem (Theorem 11.2.14) that w(2 n) = (⌊n/2⌋n
).

12.2.4. Example. Divisors of an integer, or multisets. The divisors of a positive

integer N form a poset D(N) under divisibility. It is graded; the rank of an ele-
ment is the number of primes (with multiplicity) in its factorization. It is rank-
symmetric; mapping x to N/x also shows that D(N) is self-dual.
When N is a product of n distinct primes, D(N) = ∼ 2 n ; the subsets select the
e i −1
prime factors. When N = ∏i=1 pi , the divisors of N correspond to integer lists
n

(a1 , . . . , an) with 0 ≤ ai < e i , ordered by a ≤ b if and only if ai ≤ bi for all i.

Thus D(N) can be viewed as the containment order on multisets from [n] with
at most e i − 1 copies of the ith element, where A ⊆ B if each element appears at
least as many times in B as in A. Since e1 , . . . , e n determine the multiset poset,
we denote it by M e , where e = (e1 , . . . , e n).
The multiset description expresses the poset as a product of chains: M e = ∼ e1 ×
· · ·× en . Study of M arose from divisibility questions, and some early proofs were
e

phrased using divisibility, but the structure is completely captured by the chain-
product description, and we usually view the elements as n-tuples.
For a ∈ M e , we have r(a) = ∑ ai . The rank generating function enumerates
multisets, indexed by size: ∏i=1 (1 + x + · · · x ei −1). When the ith multiplicity is
n

unbounded, we let e i = ∞ and use 1/(1 − x) as the ith factor.

SYMMETRIC CHAIN DECOMPOSITIONS

In light of Dilworth’s Theorem, we can determine the width of a poset by

exhibiting an antichain and a chain decomposition of the same size. Given a rank-
symmetric rank-unimodal poset, we can hope for a special decomposition.

12.2.5. Definition. A chain in a graded poset P is symmetric if it has an ele-

ment of rank r(P) − ¾ whenever it has an element of rank ¾ . A chain is con-
secutive or skipless if its elements lie in consecutive ranks. A symmetric
chain decomposition is a partition into symmetric skipless chains. A poset
with a symmetric chain decomposition is a symmetric chain order.
A graded poset has the Sperner property if a largest-sized rank is a
largest antichain. It has the strong Sperner property if for all ¾ the union
of any ¾ largest ranks form a largest ¾-family.
554 Chapter 12: Partially Ordered Sets

12.2.6. Example. Since every chain in a symmetric chain decomposition inter-

sects the middle rank (or both middle ranks), it is a Dilworth decomposition. A
symmetric chain order must be rank-symmetric and rank-unimodal. The poset on
the left fails, although it is graded and has the Sperner property. The poset on
the right is graded but does not have the Sperner property. It has an antichain
of size 4, but the maximum rank-size is 3.

•
• •
• • •
•
• • •
• •
•

12.2.7. Proposition. Every symmetric chain order satisfies the strong Sperner
property.
Proof: In a graded poset P , any ¾ largest ranks form a ¾-family. Every chain C in
a symmetric chain decomposition contributes exactly min{¾ , | C|} elements to the
union of the ¾ largest ranks. No chain C can contribute more than min{¾ , | C|}
elements to a ¾-family, so these ranks form a maximum ¾-family.

12.2.8. Theorem. (de Bruijn–Tengbergen–Kruyswijk [1951]) 2 n is a symmetric

chain order.
Proof: We use induction on n. When n = 0, there is only one element. For n > 0,
take two copies of a chain decomposition of 2 n−1 . Add {n} to each member of each
chain in the second copy. This decomposes 2 n into skipless chains, but they are
not symmetric and are too many when n is odd. Alter the two copies of a chain
C by transferring the top element of the second copy to the top of the first copy.
Since the element moved is the union of {n} with the previous top element, the
result is still skipless. All chains are now symmetric within the full poset.

• •
• • • • • • •
• • → • • • → • • •
• • •

The result was proved more generally for the divisor order, solving a problem
posed by the Dutch Wiskundig Genootschap in 1949. K atona observed that the
method applies to any product of symmetric chain orders.

12.2.9. Theorem. (K atona [1972a]) Products of symmetric chain orders are sym-
metric chain orders.
Proof: We need only consider a product of two such orders. If P and Q have sym-
metric chain decompositions B1 , . . . , B¾ and D1 , . . . , Dl , then the products Bi × Dj
Section 12.2: Symmetric Chains and LYM Orders 555

partition P × Q into “symmetric rectangles”. The product Bi × Dj has elements

from ranks ¾ through r(P × Q) − ¾ for some ¾ , and it has as many elements from
rank j as from rank r(P × Q) − j , for each j . Partitioning each such rectangle
into symmetric chains provides a symmetric chain decomposition of P × Q.
This reduces the problem to P and Q being chains of sizes s + 1 and t + 1,
with s ≥ t. The product consists of {(i , j): 0 ≤ i ≤ s and 0 ≤ j ≤ t}. Use chains
C0 , . . . , Ct , where C¾ consists of (0 , ¾) , (1 , ¾) , . . . , (s − ¾ , ¾) , (s − ¾ , ¾ + 1) , . . . , (s −
¾ , t). This chain is skipless; it is symmetric because the top and bottom ranks
sum to s + t. If i + j ≤ s, then (i , j) belongs only to chain Cj . If i + j > s, then
(i , j) belongs only to chain Cs−i . Thus the chains are disjoint, as shown below.
Theorem 12.2.8 is the case t = 1.

•
• •
• • •
• • • •
• • • •
• • •
• •
•

Theorem 12.2.8 is easy but does not explicitly describe the chains. How can
we tell whether two given elements lie in the same chain?

12.2.10. Example. Bracketing decomposition of 2 n. Greene–K leitman [1976b]

and Leeb [unpublished] gave an explicit locally described symmetric chain de-
composition of 2 n (see Exercise 12 for multisets); we will see that it is the same
as the decomposition in Theorem 12.2.8.
View S ∈ 2 n as a binary vector x, with x i = 1 if and only if i ∈ S. Encode each
0 as a left bracket and each 1 as a right bracket. Iteratively pair some positions
as follows. As long as some unmatched left bracket precedes an unmatched right
bracket, there is some closest such pair, say positions l and r with l < r. Add
(l , r) to the list of paired positions. For any closest unmatched pair, all positions
between them are already matched to other positions between them.
The resulting set of position pairs is the bracketing structure of x. Having
the same bracketing structure is more restrictive than having the same set of
matched positions. Below we group the sets in 2 4 by their bracketing structure;
those in the first column have no paired positions, then three columns have one
pair, and finally the two rightmost sets show two ways to pair all four positions.
1234 = ))))
123 = )))( 234 = ())) 134 = )()) 124 = ))()
12 = ))(( 23 = ())( 13 = )()( 14 = )(() 24 = ()() 34 = (())
1 = )((( 2 = ()(( 3 = (()( 4 = ((()
∅ = ((((
When the process ends, all unmatched rights occur before all unmatched
lefts. Let C be the family of subsets of [n] with a fixed bracketing structure. The
positions with matched rights correspond to elements of [n] belonging to each set
in C. Order C by the number of unmatched rights, starting with the set whose
556 Chapter 12: Partially Ordered Sets

unmatched parentheses are all lefts. This expresses C as a chain in 2 n . To move

up the chain, change the leftmost unmatched left parenthesis to a right parenthe-
sis, changing a 0 to a 1 in the incidence vector. At the bottom of C, all unmatched
positions are left parentheses; at the top they are rights.
When the bracketing structure has j matched pairs, the chain extends from
rank j to rank n − j (with n − 2 j + 1 members), since every member contains the
elements for the j matched right parentheses and omits those for the j matched
left parentheses. Hence the chains are skipless and symmetric. Every member
of 2 n occurs on exactly one chain, so this is a symmetric chain decomposition.

12.2.11. Theorem. The inductive (Theorem 12.2.8) and bracketing (Example

12.2.10) decompositions of 2 n are the same.
Proof: We use induction on n. For n ≤ 1, the poset is a single chain. For n > 1,
the induction hypothesis states that each chain in the inductive decomposition of
2 n−1 has a fixed bracketing. It suffices to show that applying the inductive con-
struction to any chain C in the bracketing decomposition of 2 n−1 yields chains in
the bracketing decomposition of 2 n .
From C define a chain C by adding {n} to each element of C. In position n,
each vector for C gains a left parenthesis, while the corresponding vector for C
gains a right parenthesis. The bracketing structure for C does not change by the
added left parenthesis. The top member of C also has this bracketing structure,
since its unmatched positions all have right parentheses, so adding a right at the
end does not create another match. Thus this element moves from the top of C
to the top of C in the bracketing decomposition of 2 n.
The remaining members of C all had a rightmost unmatched left parenthe-
sis in the same position. Thus the new right parenthesis matches to the same
position to complete the bracketing structure for each member of C below the
top. Hence they lie on the same chain in the bracketing decomposition of 2 n .
Furthermore, when the bracketing structure of C consists of j pairs, C starts
with n − 2 j members. It gains one to reach n + 1 − 2 j members (still with j pairs in
its bracketing structure), while C ends with j + 1 pairs and n − 1 − 2 j members.
Hence these are full chains in the bracketing decomposition of 2 n .
We have proved that every chain of size at least 2 in the bracketing decompo-
sition of 2 n−1 becomes two chains in the decomposition of 2 n when the induction
step is applied. By induction on n, the two decompositions are thus the same.

The Dilworth decomposition we have described for 2 n arises in many ways.

We present a third description, and Exercise 19 has yet a fourth.

12.2.12.* Example. (Aigner [1973]) We use a greedy lexicographic rule. List the
¾-sets in lexicographic order, from [¾] to [n] − [n − ¾]. Match each ¾-set A in turn
to the (¾ + 1)-set that is earliest in lexicographic order on level ¾ + 1 among the
(¾ + 1)-sets that remain unmatched and contain A. For example, in 24 we match
∅ to 1, then 1 → 12, 2 → 23, 3 → 13, and 4 → 14, then 12 → 123, 13 → 134,
14 → 124, 23 → 234, with none available for 24 and 34, and finally 123 → 1234.
The resulting chains (shown below for n = 4) are the same as those in Exam-
ple 12.2.10 (see Exercise 20). It is not even obvious that this produces symmetric
Section 12.2: Symmetric Chains and LYM Orders 557

chains, and yet we obtain the familiar bracketing decomposition. The construc-
tion also extends to multisets.

• 1234
123 • •124 •134 •234
12• 13• 14• • 23 24 34
1 • 2• •3 •4
•∅

For many classical rank-symmetric rank-unimodal posets, the existence of

symmetric chain decompositions remains unknown.

12.2.13. Example. L(m , n): Bounded integer partitions. The elements of L(m , n)
are the integer lists (a1 , . . . , am) such that 0 ≤ a1 ≤ · · · < am ≤ n. These corre-
spond to Ferrers diagrams contained in an m by n rectangle. The order relation
puts a ≤ b if and only if ai ≤ bi for all i. Hence L(m , n) is a subposet of (n + 1)m .
The complement of a Ferrers diagram in a rectangle fits in that rectangle,
so L(m , n) is rank-symmetric (actually, self-dual). The rank generating function
1− qn+ j
(with formal variable q for algebraic reasons), is ∏m j =1 1 − q j
(Exercise 3.4.40).
Rank-unimodality is difficult; algebraic proofs began with Sylvester (see Proctor
[1982]). O ’Hara [1990] found an intricate combinatorial proof, presented also in
Zeilberger [1989]. Stanley [1982] observed that L(m , n) has the Sperner prop-
erty, using results of Griggs [1977].
Stanley conjectured that L(m , n) is a symmetric chain order. This is easy for
m ≤ 2 (see Exercises 5–8) and is known for m ≤ 4 (Riess [1978], Lindström [1980],
West [1980]). Solutions for L(m , n) with m = 5 and n odd have been rumored.

Finally, we present an application of the bracketing decomposition. We want

to count antichains in 2 n ; this is known as Dedekind’s Problem. By Remark
12.1.14, antichains correspond to down-sets. There are 2(⌈n/2⌉) down-sets whose
n

maximal elements all have size ⌈ n/2⌉ ; this yields a lower bound. For the upper
bound, we view down-sets as monotone Boolean functions, defined to be order-
preserving functions from 2 n to {0 , 1}.

12.2.14. Theorem. (Hansel [1966]) The number of monotone Boolean functions

is at most 3(⌈n/2⌉).
n

Proof: We construct a monotone Boolean function º by specifying its values in or-

der. Consider chains in the bracketing decomposition in increasing order of size.
For each new chain, the values of º on some of it are already forced. If x contains
a set already assigned 1, then also º (x) = 1. Similarly, a subset of a set assigned
0 must be assigned 0. Since there are (⌈n/2⌉n
) chains, it suffices to show that always
there are at most three ways to extend º to the next chain C. To do this, we show
that C has at most 2 elements with undetermined labels. When these are x and
y with x > y, the options (º (y) , º (x)) will be only (0 , 0), (0 , 1), and (1 , 1).
558 Chapter 12: Partially Ordered Sets

Consider sets x and y in C, with x > y and | x − y| > 1. To show that º is

known for x or y, it suffices to construct between them in 2 n such that  ( ) is
known. Then  ( ) = 1 implies  (x) = 1, and  ( ) = 0 implies  (y) = 0.
If C has size  , then its bracketing has n −  + 1 matched positions. Shorter
chains have more matched positions. The unmatched positions for x and y have
right parentheses followed by left parentheses, with at least two more rights in
x than in y. Construct from x by changing the next-to-last unmatched right
parenthesis in x to a left parenthesis.
Now y < < x. In the bracketing of , the positions corresponding to the
last two unmatched right parentheses in x match, since the positions between
them are all matched. The other matches are unchanged, so has more matched
positions and appears on a shorter chain, as desired.
x = − − −) − − − −) − − − −−
= − − −(− − −−) − − − −−
y = − − −(− − − − (− − − − −

By a more detailed argument along the same lines, showing that almost
always the new chain contains at most one undetermined element, K leitman–
Markowsky [1975] improved this upper bound to 2(⌈n/2⌉)(1+O(log n/n)) (strengthening
n

K leitman [1969]). Korshunov [1981] obtained a more detailed asymptotic for-

mula, which for even n is
n
(n/2) (n/2n−1)(2−n/2 + n2 2−n−5 − n2−n−4)
2 e .

Exact values for n ≤ 7 are 3, 6, 20, 168, 7581, 7828354, and 2,414,682,040,998,
with Korshunov ’s formula giving 7996118 for n = 6.

LYM AND SPERNER PROPERTIES

A short direct proof of Sperner ’s Theorem counts the maximal chains in 2 n

by where they hit an antichain. The proof identifies a property for graded posets
that implies the Sperner property and much more.

12.2.15. Theorem. (Sperner [1928]) In 2 n , the elements of rank ⌊ n/2⌋ form a

maximum antichain.
Proof: (Lubell [1966]) A maximal chain has at most one element in an antichain
F . We count the chains according to which element of F (if any) they contain.
There are n! maximal chains, since they are specified by acquiring the elements
of [n] in some order. To pass through x, they must acquire the elements of x first,
so exactly | x|!(n −| x|)! maximal chains pass through x. Thus ∑ x∈ F | x|!(n −| x|)! ≤ n!.
Dividing both sides by n! yields ∑ x∈ F 1/Nr(x) ≤ 1. Replacing Nr(x) with the
largest binomial coefficient preserves the inequality, and then clearing the frac-
tion yields | F | ≤ (⌈n/2⌉
n
).

−1 ≤ 1 was discovered independently for antichains in

The inequality ∑ x∈ F Nr(x)
2 n by Lubell [1966], Yamamoto [1954], and Meshalkin [1963].
Section 12.2: Symmetric Chains and LYM Orders 559

−1 ≤
12.2.16. Definition. For a family F in a graded poset, the inequality ∑ x∈ F Nr(x)
1 is the LYM inequality. A graded poset satisfies the LYM property if its
antichains all satisfy the LYM inequality. Such a poset is an LYM order.

By the argument of Theorem 12.2.15, the LYM property implies the Sperner
property. An analogue of this argument using circular arrangements yields the
strong Sperner property for 2 n (Exercise 28). We will see later that the LYM
property implies the strong Sperner property.
We will also see that the argument in Theorem 12.2.15 generalizes; it needs
only a list of maximal chains such that, in each rank, each element appears
equally often. We introduce a name for such a list of maximal chains in order
to use the argument in more generality.

12.2.17. Definition. A nonempty list of maximal chains in a graded poset P is a

regular covering of P if, for each rank P¾ , each element of P¾ lies in exactly
the fraction 1/ | P¾ | of these chains.

12.2.18. Example. A regular covering of P is a list; chains may be used repeat-

edly. The list may omit some maximal chains and use some more than others. In
3 × 4, shown below, the number of maximal chains containing a specified element
of a middle rank is 1, 6, or 3. The full set of maximal chains is not a regular cov-
ering. On the other hand, there is a regular covering with 6 chains consisting of
two copies of the “outer ” indicated chains plus one copy of each “inner ” chain. We
will prove later that every (finite) product of chains has a regular covering.
•
• •
• • •
• • •
• •
•
LYM orders are characterized by the existence of a regular covering. To prove
this, we use a third and equally important property.

12.2.19. Definition. (Graham–Harper [1969]) A graded poset P has the nor-

malized matching property if | A∗ | /N¾+1 ≥ | A| /N¾ for all ¾ and all A ⊆ P¾ ,
where A∗ denotes U[A] ∩ P¾+1 . The set A∗ is called the shade of A or its
“shadow at the rank above”.

12.2.20. Example. 2 n has the normalized matching property. For A ⊂ ([n] ¾ ), each
element of A extends to n − ¾ elements of A∗ . Since each element in A∗ can lose
an element in ¾ + 1 ways, (¾ + 1) | A∗ | ≥ (n − ¾) | A| . Dividing both sides by (n − ¾)(¾n)
yields | A| /N¾ ≤ | A∗ | /N¾+1 . This was the essence of the argument we used to prove
Sperner ’s Theorem (Theorem 11.2.14).

The subgraph of the diagram induced by P¾ ∪ P¾+1 is a P¾ , P¾+1 -bigraph with

N(A) = A∗ . When N¾+1 ≥ N¾ , normalized matching is stronger than Hall’s condi-
tion for a matching that covers P¾ (Section 6.1). We use Hall’s Theorem to prove
equivalence for these properties.
560 Chapter 12: Partially Ordered Sets

12.2.21. Theorem. (K leitman [1974]) For a graded poset P , the following state-
ments are equivalent:
(A) P has a regular covering.
(B) P has the LYM property.
(C) P has the normalized matching property.
Proof: A ⇒ B. As in the proof of Theorem 12.2.15, counting the chains in
a regular covering C that are hit by the elements of an antichain F yields
∑ x∈ F |C| /Nr(x) ≤ |C| and thus the LYM inequality.
B ⇒ C. If A ⊆ P¾ , then A ∪ (P¾+1 − A∗) is an antichain in P. The LYM
inequality yields | A| /N¾ +| P¾+1 − A∗ | /N¾+1 ≤ 1. Since | P¾+1 | = N¾+1 , this becomes
| A| /N¾ ≤ | A∗ | /N¾+1 .
A∗ P¾+1 − A∗ P¾+1

A P¾

C ⇒ A. Letting M = ∏ N¾ (P), define the blowup poset P with M/Nr(x)

copies of x for each x ∈ P. A copy of x is less than a copy of y in P if and only
if x < y in P. Use the normalized matching property in P to find perfect match-
ings joining adjacent ranks in P ; such matchings combine to partition P into
disjoint maximal chains, and these chains collapse to form a regular covering of
P with x appearing M/Nr(x) times.
• • • • •
S∗ = B∗
• • • • • •
P → P
• • • • • •
S B
• • • • •
By Hall’s Theorem, it suffices to show for S ⊂ P¾ that | S∗ | ≥ | S| , where S∗
is the shade of S (in P¾ +1 ). Let B be the subset of P¾ consisting of all copies of
each element of P¾ that has at least one copy in S. Now S∗ = B∗ and S ⊆ B, so it
suffices to show that | B∗ | ≥ | B| .
Let A be the sets of elements in P that have copies in B. Note that in P the
set B∗ consists of all copies of all elements of P¾ covering elements of A; that is,
all copies of elements of A∗ . Note that | B| = | A| M/N¾ and | B∗ | = | A∗ | M/N¾+1 , since
all copies of an element appear when any copies appear. The normalized matching
property in P now yields | B∗ | ≥ | B| .

12.2.22.* Remark. (1) The LYM inequality and regular coverings are self-dual,
so normalized matching is equivalent to requiring | D[A] ∩ P¾−1 | ≥ | A| N¾−1/N¾
when A ⊆ P¾ (Exercise 29 requests a direct proof without duality).
(2) Theorem 12.2.21 guarantees short proofs. Having the LYM property can
be proved by giving a regular covering. Failing the LYM property can be proved
by giving a violation of the normalized matching property (Exercises 26–27).
(3) Since the number of chains is divisible by each rank-size, every regular
covering has at least lcm {N¾ } chains, achieved by blowing up x to lcm {N¾ }/Nr(x)
copies (Exercise 4). However, every LYM poset has a regular covering using at
most | P | − r(P) distinct chains, and this is best possible (Exercise 30).
(4) Graham–Harper [1969] found an efficient method to find a regular cover-
ing. We seek nonnegative integer weights on the edges from P¾ to P¾+1 such that
Section 12.2: Symmetric Chains and LYM Orders 561

the total weight on edges leaving each element of P¾ is N¾+1 and the total weight
on edges entering each element of P¾+1 is N¾ . A solution forms a regular covering
for the subposet P¾ ∪ P¾+1 . The weights give the relative usage of the covering
pairs in the full regular covering. Feasibility can be tested by adding a source
and sink, modeling the weights with multiedges, and using Menger ’s Theorem;
thus we can test the LYM property in polynomial time.
(5) When the number of elements an element x covers depends only on the
rank of x, the set of all maximal chains is a regular covering (Baker [1969]).

We have seen that the LYM property implies the Sperner property. The LYM
property (with its equivalence to regular covering) also implies a more general
statement that implies the strong Sperner property.

12.2.23. Theorem. (Greene–K leitman [1978]) Let ë: P → be a weight func-

tion defined on the elements of P. If P has the LYM property, then for every
subset G ⊂ P and every regular covering C of P ,
ëx
∑ Nr(x) ≤ max
C∈C
∑ ëy .
x∈ G y ∈ C∩ G

Proof: We interpret the inequality probabilistically. Choose a chain from C uni-

formly at random, and define a random variable X = ∑ y∈ C∩ G ë y . The expectation
of X is ∑ x∈ G ë x (x ∈ C). Since C is a regular covering, (x ∈ C) = 1/Nr(x) . Thus
the left side of the desired inequality is the expected value of X , while the right
side is its maximum value.

Suitable choices of ë yield various applications. First let ë x = Nr(x).

12.2.24. Corollary. If P is an LYM order with regular covering C and G is any

subset of P , then | G | ≤ max C∈C ∑ y∈ C∩ G Nr(y) .

Bounds on | G | follow from Corollary 12.2.24 for various chain conditions on

| G |. Erdős [1945] proved the next corollary for 2 n by generalizing the argument
of Sperner [1928] for the Sperner property.

12.2.25. Corollary. (Erdős [1945]) The LYM property implies the strong Sperner
property.
Proof: If G is a ¾-family, then Corollary 12.2.24 limits its size to the sum of the
¾ largest rank sizes.

Now it is natural to ask whether the LYM property is so strong that it also im-
plies the structural “holy grail” of symmetric chain decomposition. The answer
is “yes” when the obvious necessary conditions hold.

12.2.26. Theorem. (Anderson [1976], Griggs [1977]) Every rank-unimodal

rank-symmetric LYM poset is a symmetric chain order.
Proof: We use induction on the height of the poset P. For height 1, there is
nothing to prove. For even height, the two middle levels have equal size, and nor-
malized matching between them reduces to Hall’s Condition. Collapse the two
562 Chapter 12: Partially Ordered Sets

levels to one along the resulting matching. Since normalized matching considers
only consecutive levels, the smaller poset P is still an LYM order, and it also has
symmetric and unimodal rank sizes. The guaranteed chain decomposition for P
expands into a symmetric chain decomposition of P using the central matching.
For odd height, let the middle rank be P¾+1 . If we can match P¾ to P¾+2
through P¾+1 , then the unused elements of P¾+1 will become singleton chains.
After discarding them, the three levels collapse to one level along the matched
edges to obtain a smaller LYM order P . After obtaining a symmetric chain de-
composition of P , expand each element of the middle level into a 3-element chain
as the middle of a skipless symmetric chain in P. Together with the discarded
singletons, these chains form a symmetric chain decomposition of P.
The covering pairs down from P¾+2 and up from P¾ yield two families A and
B of subsets of P¾+1 , each with N¾ sets. The two desired matchings exist when
there is a common system of distinct representatives for A and B. For subsets
I , J ⊆ [N¾ ], let A = ∪i∈ I Ai and B = ∪ j ∈ J Bj . By the Ford–Fulkerson Condition
(Theorem 7.2.15), | A ∩ B | ≥ | I |+| J |− N¾ suffices. Since the dual of an LYM order
is an LYM order, both | A | ≥ | I | (N¾+1/N¾) and | B | ≥ | J | (N¾+1/N¾). Thus
| A ∩ B | = | A | + | B | − | A ∪ B | ≥ (| I | + | J |)(N¾+1/N¾) − N¾+1
= (| I | + | J | − N¾)(N¾+1/N¾) ≥ | I | + | J | − N¾ .

The LYM property has many consequences. Unfortunately, many graded

posets of interest are not LYM orders (see Exercises 25–27). For these, one can
still ask whether consequences such as symmetric chain decomposition hold. For
the poset L(m , n) (Example 12.2.13), this remains a fascinating open problem.

PRODUCTS OF LYM ORDERS (optional)

So far, our only criteria for the LYM property are the three equivalent defin-
ing properties. We have not yet proved that chain-products are symmetric chain
orders. We consider the behavior of the LYM property under products. First note
the behavior of the rank sizes.

12.2.27. Remark. Rank of poset products. If P and Q are graded, then their
product is also graded and has rank function given by r P × Q(x , y) = r P (x) + r Q(y).
Thus N¾ (P × Q) = ∑i Ni(P)N¾−i(Q).

12.2.28. Example. The product of LYM posets need not be an LYM poset. In the
example below, the set {a1 , b1} occupies 32 of its rank, but its shadow at the rank
above is only 12 of that rank.

• d1
d•
1• d0 • • c1
c• × =
0• c0 • • •
• • a1 b1
a b • •
a0 b0
Section 12.2: Symmetric Chains and LYM Orders 563

12.2.29. Remark. Necessary condition for LYM property in a special product. Let
Q = P × 2, where P is an LYM poset. Let A = {(x , 1): x ∈ P¾−1 }. The set A lies in
rank ¾ in Q. Note that A∗ ⊆ {(y , 1): y ∈ P¾ }.
Write Nj for Nj (P). Since 2 has one element at each rank, Q has Nj −1 + Nj
elements at rank j . Normalized matching thus requires
| A∗ | | A|
≥ .
N¾ + N¾+1 N¾−1 + N¾
Since | A| = N¾−1 and | A∗ | = N¾ , we need N¾2 ≥ N¾−1 N¾+1 . This condition on the
sequence of rank sizes is necessary for the product with 2 to be an LYM order. We
will show that it is sufficient for products in general.

12.2.30. Definition. A sequence ⟨N⟩ is log-concave if N¾2 ≥ N¾−1 N¾+1 for all ¾ .

The condition for log-concavity is that the logarithms of the terms form a
concave sequence. Every concave sequence is also log-concave.

12.2.31. Theorem. (Harper [1974], Hsieh–K leitman [1973]) The family of LYM
orders with log-concave rank sizes is closed under taking products.
¾
Proof: We have r(P1 × P2) = r(P1) + r(P2) and N¾ (P1 × P2) = ∑ i=0 Ni(P1)N¾−i(P2).
Log-concavity of the convolution of two log-concave sequences is an exercise in al-
gebraic manipulation (Exercise 31).
To prove the LYM inequality for P1 × P2 , we generalize Theorem 12.2.23.
There we picked a random chain from a regular covering. Here, we pick a ran-
dom pair C1 , C2 from regular coverings C1 and C2 of P1 and P2 . This produces a
random rectangle in the product.
Given any subset G of P1 × P2 and weights ë x for x ∈ P , let X =
∑ x∈ G∩(C1 × C2) ë x . The expectation is ∑ x∈ G ë x (x ∈ C1 × C2). Since C1 and C2 are
regular coverings, (x ∈ C1 × C2) = Nr(x1)(P1)−1 Nr(x2)(P2)−1 , where x = (x1 , x2).
Thus the left side of the inequality below is the expectation of X , while the right
side is its maximum.
ëx
∑ Nr(x )(P1)Nr(x )(P2) ≤ max
C ∈C i i
∑ ëy
x∈ G 1 2 y ∈ G ∩ C1 × C2
Setting ë x = Nr(x1)(P1)Nr(x2)(P2)/Nx yields
1 Nr(y1)(P1)Nr(y2)(P2)
∑ Nr(x) ≤ max
C ∈C i i
∑ Nr(y)
.
x∈ G y ∈ G ∩ C1 × C2
To obtain the LYM inequality, it suffices to show that the right side of this in-
equality is at most 1 when G is an antichain.
If G lies entirely in the ¾ th rank of P1 × P2 , then we compute a bound over
rank ¾ for arbitrary C1 × C2 . The elements of rank ¾ in P1 × P2 that lie in C1 × C2
have the form (y1 , y2), where y1 is the element of C1 with rank i in P1 and y2 is
the element of C2 with rank ¾ − i in P2 . There is at most one such element for
each i ∈ {0 , . . . , ¾}. Thus
¾
Nr(y1)(P1)Nr(y2)(P2) N (P )N (P )
∑ ≤ ∑ i 1 ¾−i 2 = 1 .
Nr(y) N¾ (P1 × P2)
y ∈ G ∩ C1 × C2 i=0
564 Chapter 12: Partially Ordered Sets

Suppose that G has elements from more than one rank in P1 × P2 . We show
that pushing those in the lowest occupied rank upward toward the highest occu-
pied rank cannot decrease the specified sum. This suffices, since we have shown
that the sum is bounded by 1 when G lies in a single rank.
Let T be a maximal set of “consecutive” elements of G at rank ¾ in C1 × C2 ;
“consecutive” means T = {(x i , y¾−i): r ≤ i ≤ s} for some 0 ≤ r ≤ s, where x i is
the element of rank i in C1 and yj is the element of rank j in C2 . Let T ∗ be the
set of elements covering members of T in C1 × C2 ; this is another consecutive
set. Because T is a maximal consecutive set, replacing all of T with T ∗ does not
violate the antichain condition.
T∗
¾+1◦ • • • ◦ ◦
¾→ ◦ • T • ◦ •
s+1

Let ½ ¾ (r, s) = i=r Ni(P1)N¾−i(P2). The contribution of T to the original sum

s
∑
+1)
is ½ ¾ (0 ,∞) . The contribution of T ∗ to the new sum is ½½¾¾++11(r,s
½ ¾ (r,s)
(0 ,∞) . We need only show
+1) ½ ¾+1 (r,s+1)
that ½½¾¾++11(r,s ½ ¾ (r,s)
(0 ,∞) ≥ ½ ¾ (0 ,∞) . Since this is equivalent to ½ ¾ (r,s)
≥ ½½¾+¾ (0
1 (0 ,∞)
,∞)
, it suffices
½ ¾+1 (r,s+1)
to show that ½ ¾ (r,s) is increasing in r and decreasing in s. See Exercise 32.

12.2.32. Corollary. M e is an LYM order.

Proof: M e is a product of chains, each of which is an LYM order with log-concave
rank sizes.

Algebraic techniques can yield log-concavity of sequences. When the rank

generating function is known, the following result may apply (see Stanley [1989]).

12.2.33. Theorem. If all roots of a polynomial with real coefficients are real,
then the sequence of coefficients is log-concave.

12.2.34. Example. The rank generating function for 2 n is (1 + x)n , so Theo-

rem 12.2.33 applies. However, for M e it is ∏i=1 (1 + x + · · · + x ei −1), which has
n

complex roots when max e i > 2. Log-concavity follows because the convolution of
log-concave sequences is log-concave.

EXERCISES 12.2

12.2.1. (−) Use Sperner ’s Theorem and Dilworth’s Theorem to prove (weaker than Theo-
rem 12.2.14) that the number of antichains in 2 n is at most (n + 1)(⌈n/2⌉). (Gilbert [1954])
n

12.2.2. (−) Let P be a rank-symmetric LYM order of rank n. Let S be a down-set in P.

Prove that the elements of S have average rank at most n/2.
12.2.3. (−) Show that the sequence of rank sizes of a product poset may be log-concave
even though the sequences for the factors are not both log-concave.
Exercises for Section 12.2 565

12.2.4. (−) Prove that the minimum number of chains in a regular covering of an LYM or-
der is lcm N¾ . (Hint: Blow up each element x to (lcm N¾)/Nr(x) copies and use the argument
of Theorem 12.2.21.) (West–Harper–Daykin [1983])
12.2.5. (−) Prove that L(2 , n) (Example 12.2.13) is a symmetric chain order.
12.2.6. (−) An automorphism of a poset is an order-preserving permutation of the ele-
ments. Let P be a graded poset such that whenever x and y have the same rank in P , some
automorphism maps x to y. Prove that P is an LYM order.
12.2.7. Prove w(n × n × n) = ⌊ (3n2 + 1)/4⌋ .
12.2.8. (+) Construct a symmetric chain decomposition of L(3 , n) for odd n. (West [1980])
12.2.9. Let P be a finite poset whose diagram is connected.
(a) Show that P may have a maximal chain that is not a longest chain even when every
element in P lies in a longest chain of P.
(b) Suppose that , whenever y covers x in P , some longest chain contains both x and y.
Prove that P is a graded poset. (Stanley [1971, p. 19–20])
12.2.10. ¾ -families in chain products.

(a) For i , j , ¾ ∈ , find a formula without summations for the maximum size of a ¾-
family in the chain-product i × j. (Hint: Consider three cases for {i , j , ¾}.)
(b) Consider three sets of parallel lines in the plane, forming equilateral triangles.
Within the sets, the lines need not be equally spaced. Let the sets have sizes r, s , t. De-
termine the maximum number of points that occur as the intersection of three lines, one
from each set. (Matsko–West–Wetzel [2001])
12.2.11. (♦) A semiantichain in P × Q is a subset S having (u , v) < (u , v ) for two ele-
ments of S only if u < u and v < v . A product poset has the 2-part Sperner property
when some single rank is a largest semiantichain.
(a) Prove that the product of two symmetric chain orders is 2-part Sperner.
(b) Use part (a) to show that for any 2-coloring of the elements of [n], the largest sub-
poset of 2 n having no related pair with monochromatic difference consists of the middle
rank. (Kleitman [1965], Katona [1966])
12.2.12. Bracketing decomposition of M e . Dedicate ei − 1 positions for the ith coordinate.
View a ∈ M e as a list with 0 ≤ ai < ei . In the positions for the ith coordinate, put ai
right parentheses followed by ei − 1 − ai left parentheses. Match parentheses as in Ex-
ample 12.2.10. Associate a single chain to the elements having the same matched pairs.
Prove that these chains form a symmetric chain decomposition of M e . (Below is a chain
from M (5 ,4) with one matched pair; it is the second chain in the figure for Theorem 12.2.9.)
(Greene–Kleitman [1976b], Leeb [unpublished])
(((( )(( )((( )(( ))(( )(( )))( )(( )))( ))( )))( )))
< < < < <
(0 , 1) (1 , 1) (2 , 1) (3 , 1) (3 , 2) (3 , 3)

12.2.13. A finite set of integers is balanced if the numbers of even and odd elements differ
by at most 1. Let F be a family of subsets of [n] such that whenever A , B ∈ F with A ⊆ B,
the set B − A is not balanced. Prove that | F | ≤ (⌈n/2⌉
n
) and that this bound is best possible.
(Greene–Kleitman [1976b])
12.2.14. (♦) A skew chain order is a poset having a rank function in which every mini-
mal element has rank 0 and a decomposition into skipless chains starting at rank 0. De-
scribe (with proof) the largest semiantichain (see Exercise 12.2.11) in the product of two
skew chain orders. Let P¾ be the inclusion order on the set of intervals in with end-
points in [¾]. Give a geometric description of the maximum semiantichain in Pm × Pn .
(West–Kleitman [1979])
566 Chapter 12: Partially Ordered Sets

12.2.15. Use a chain decomposition of 2 n to construct a spanning subgraph of the hyper-

cube graph Qn that has diameter n and has only 2 n + (⌈n/2⌉
n
) − 2 edges. (Comment: This
subgraph has the same diameter as Qn while keeping only a vanishing fraction of the
edges.) (Graham–Harary [1992])
12.2.16. (♦) A universal subset list on an alphabet S is a word having every subset of S
as a consecutive substring. For example, 1231 is such a list on [3], and 123421341 is such
a list on [4]. Listing all 2 n subsets successively yields a universal subset list on [n] with
length n2 n−1 , since the average size is n/2.
(a) For even n, use symmetric chain decompositions of two copies of 2 n/2 to construct
a universal subset list on [n] with length asymptotically at most (4/ )2 n . (Hint: Use Stir-
¾
ling ’s formula to approximate (⌊¾/2⌋ ).)
(b) Show that the top element on chains in a symmetric chain decomposition of 2[¾]
has average size /2 + O( 1/2). Use this to reduce the bound in√(a). (Lipski [1978])
(c) Prove that these upper bounds are within a factor of c n of being optimal.
12.2.17. (♦) The Littlewood–Offord Problem. Let a1 , . . . , an be vectors in d , each having
length at least 1. Let R1 , . . . , R¾ be regions in d , each having diameter less than 1 (i.e.,
contained in a sphere of diameter 1), and let R be their union. Let d ¾(n) = ∑i=r ( ni) , where
s

r = ⌊ (n −  + 1)/2⌋ and s = ⌊ (n +  − 1)/2⌋ (this counts the  middle ranks in 2 n).

(a) Prove that d ¾(n) = d ¾+1 (n − 1) + d ¾−1(n − 1).
(b) Prove that the number of 0 , 1-vectors x such that ∑ xi ai ∈ R is at most d ¾(n).
(Hint: To apply part (a) in an inductive proof, one must group these vectors x into sets
corresponding to two problems of the same type with n − 1 vectors, one having  + 1 regions
and one having  − 1 regions.) (Kleitman [1970])
12.2.18. Let G be an X , Y -bigraph with X = {x1 , . . . , x m} and Y = {y1 , . . . , yn}. Define
the greedy matching of X into Y as follows: having processed each xi with i < r, match
xr to its least-indexed available neighbor in Y , if any is available. Prove that the greedy
matchings of X into Y and Y into X are the same.
12.2.19. (♦) For X ⊆ [n], let X = {x1 , . . . , x¾ } with x1 < · · · < x¾ , and set x0 = 0. Let
t be the largest nonnegative index i minimizing xi − 2i. If x t < n, then form  (X) from
X by adding the element 1 + x t , but leave  (X) undefined if x t = n. Prove that the pairs
(X ,  (X)) form the chains in the bracketing decomposition of 2 n . (For example, applica-
tions of  in 24 starting with {3} yield {1 , 3}, then {1 , 3 , 4}, then nothing. Other exam-
ples are  ({1 , 3 , 4 , 7}) = {1 , 3 , 4 , 5 , 7} with t = 3 and  ({3 , 5 , 7 , 9}) = {1 , 3 , 5 , 7 , 9} with
t = 0.) (White–Williamson [1977])
12.2.20. Prove that Aigner ’s lexicographically-generated chains (Example 12.2.12) are
the same as the chains in Theorem 12.2.8 or the chains in Exercise 12.2.19.
12.2.21. (♦) Two chain partitions are orthogonal if no two elements appear in the same
chain in both partitions. (Kleitman–Shearer [1979])
(a) Define D from the bracketing decomposition C of 2 n by changing each set to its
complement and reversing the chains. For n ≥ 4, prove that a slight change in D yields
chain partition orthogonal to C.
(b) Construct two orthogonal Dilworth decompositions for each of 2 2 and 23 , and con-
struct three pairwise orthogonal Dilworth decompositions for 24 .
(c) Prove that 2 n has at most ⌈ n+2 1 ⌉ pairwise orthogonal Dilworth decompositions.
12.2.22. Prove that 2 n has n! Dilworth decompositions such that no two elements lie on
the same chain in more than n!⌈ (n + 1)/2⌉ −1 of them. (Kleitman–Shearer [1979])
12.2.23. For x and y chosen from n × m according to some probability distribution p, prove
(x ≤ y) ≥ n2nm
+m
, with equality possible when n divides m. (Hint: For a and b chosen from
a distribution over  choices, prove (a = b) ≥ 1/ .) (Kleitman–Shearer [1979])
Exercises for Section 12.2 567

12.2.24. (♦) Let G be an X , Y -bigraph with | Y | ≥ | X | and no isolated vertices. The defi-
ciency (A) of a set A ⊆ V(G) is | A| − | N(A)|. Say that G has the strong Hall property if
(A) + (B) ≤ | Y | − | X | when A , B ⊆ Y with | A| + | B| ≤ | Y |. Let H be a graph formed from
G + G by adding a matching joining the copies of Y .
(a) Prove that if G satisfies the strong Hall property, then H has a perfect matching.
(b) Let P be a self-dual rank-symmetric rank-unimodal poset. Prove that if the bi-
partite graphs joining consecutive levels all have the strong Hall property, then P has a
symmetric chain decomposition. (Lu–Wang–Wong [1998])
12.2.25. The Weak Order Wn on the permutations of [n] is defined by letting  cover  if
 is obtained from  by transposing two consecutive elements to make them an inversion.
The poset is graded: r( ) is the number of inversions in  , so the rank generating function
is 1(1 + x)(1 + x + x2) · · · (1 + x + · · · + x n−1 ), with unimodal coefficients (Example 12.2.34).
(a) Show that Wn is a symmetric chain order for n ≤ 4. (Comment: Also W5 is a sym-
metric chain order by using Exercise 12.2.24, but for n > 5 the answer is unknown.)
(b) Show that W4 is not an LYM order by considering the permutation 2143. Gener-
alize the argument to prove that W2m is not an LYM order for m ≥ 2.
12.2.26. Prove that L(m , n) is an LYM order if and only if m = n = 3 or min{m , n} ≤ 2.
(Hint: Find the lowest violation of normalized matching in L(3 , 4) and generalize this for
m ≥ 3 and n ≥ 4. A different violation is needed for L(4 , 4); consider 1114.)
12.2.27. (♦) Let Π n denote the poset of partitions of [n], with  ≤  if  is a union of par-
titions of the blocks of  . Partitions with  blocks have rank n −  , so Nn−¾(Π n) = S(n , )
(the Stirling number). For even n, let A be the set of partitions of [n] into two blocks of
size n/2. Use A and its shadow to prove that Π n with n even is not an LYM order when
n ≥ 20. (Spencer [1974]) (Comment: Rota asked whether Π n always has the Sperner prop-
erty; Canfield [1978] showed that it doesn’t. Shearer [1979] and Jichang–Kleitman [1984]
reduced the least n such that Π n is not Sperner to 4 × 109 and then to 3.4 × 106 .)
12.2.28. (♦) Let F be a -family in 2 n .
(a) Prove that every circular permutation of [n] has at most n substrings in F .
−1 ≤  .
(b) Use part (a) to prove ∑ x∈ F Nr(x)
(c) Use part (b) to prove the strong Sperner property for 2 n . (Füredi–Katona)
12.2.29. Prove directly that the dual of a graded poset satisfying the normalized match-
ing property also satisfies that property. That is, given an X , Y -bigraph satisfying
| N(S)| / | Y | ≥ | S| / | X | for all S ⊆ X , prove | N(T)| / | X | ≥ | T | / | Y | for all T ⊆ Y .
12.2.30. A minimal LYM order is an LYM order such that deleting any covering pair
from the order relation destroys the LYM property.
(a) Prove that the relations between adjacent ranks of a minimal LYM order form a
forest , connected when the rank sizes are relatively prime. Construct an example to show
that the forest may or may not be connected when the rank sizes are not relatively prime.
(Hint: Consider the Graham–Harper version of the normalized matching condition.)
(b) Prove that every LYM order P has a regular covering using at most | P | − r(P) dis-
tinct chains. Prove that this is sharp. (West–Harper–Daykin [1983])
12.2.31. Let ⟨a⟩ and ⟨b⟩ be log-concave sequences.
(a) Prove that ai aj ≥ ai+1 aj −1 whenever i ≥ j .
(b) Prove that ∑i ∑j (ai aj − ai+1 aj −1)(b¾−i b¾− j − b¾−1 −i b¾+1 − j ) ≥ 0.
(c) (+) Use part (b) to prove that the product of two graded posets with log-concave
rank sizes has log-concave rank sizes.
12.2.32. Given log-concave sequences ⟨a⟩ and ⟨b⟩, let  ¾(r, s) = ∑i=r ai b¾−i . Prove that
s

½ ¾+1 (r,s+1)
½ (r,s)
is increasing in r and decreasing in s. (This completes Theorem 12.2.31.)
¾
568 Chapter 12: Partially Ordered Sets

12.2.33. (♦) Given s , t ∈ 0 , let an = ∑it=s (ni). Prove that ⟨a⟩ is log-concave. (Hint: Let
ºn(x) = ∑¾= s (¾) x
t n ¾
, and study the coefficients of [ºn(x)]2 − ºn−1 (x)ºn+1 (x). (Knuth [2017])
12.2.34. Let Ln(q) be the inclusion order on the subspaces of an n-dimensional vector space
over a field with q elements.
(a) Count the bases in a subspace of dimension ¾ .
(qn − 1)(qn−1 − 1) · · · (qn−¾+1 − 1)
(b) Prove N¾(Ln(q)) = .
(q¾ − 1)(q¾−1 − 1) · · · (q1 − 1)
q¾ −1
(c) Prove that elements of rank ¾ cover q−1 elements. Conclude that Ln(q) is an LYM
order. (Comment: N¾(Ln(q)) is the Gaussian polynomial [¾n]q of Exercise 3.4.40.)

12.2.35. A graded poset is regular if all elements of rank ¾ are covered by the same num-
ber of elements at rank ¾ + 1 and cover the same number at rank ¾ − 1. A graded poset is
strictly Sperner if every maximum antichain consists of one rank. A regular covering is
exhaustive if any two comparable elements both lie on some chain in the covering.
(a) Show that the product of two chains of different lengths is not strictly Sperner (al-
though it is an LYM order).
(b) Prove that a poset with an exhaustive regular covering has the strict Sperner prop-
erty if and only if for every pair of maximum-sized ranks, the bipartite graph of relations
between them is connected. (Broline)
(c) Prove that a regular poset has an exhaustive regular covering, and conclude that
2 n has the strict Sperner property.
12.2.36. Derived LYM posets.
(a) Let P be an LYM order of rank r. For S ⊆ [r], let Q be the subposet consisting of
all elements whose rank in P belongs to S. Prove that Q is an LYM order.
(b) Let ⟨a¾ ⟩ be a log-concave sequence. Prove that a2¾ ≥ a¾+ j a¾− j for all ¾ , j .
(c) Let A1 , . . . , Ar be a partition of [n] into r blocks, and let S1 , . . . , Sr be r arithmetic
progressions. Let P = {X ⊂ [n]: | X ∩ Ai | ∈ Si for 1 ≤ i ≤ r}. Prove that the inclusion order
on P is an LYM order. (Griggs [1982])
12.2.37. Bollobás’ Inequality.
(a) Let A1 , . . . , A m and B1 , . . . , Bm be subsets of [n] such that Ai ∩ Bj = ∅ if and only
−1
if i = j . Prove that ∑ (| Ai||+|
Ai |
Bi |
) ≤ 1. (Hint: Consider instances of B¾ completely after A¾
in permutations of [n].) (Bollobás [1965])
(b) Use part (a) to prove the LYM property for 2 n .
(c) Use part (a) to prove that (⌊(nn−−¾)/2⌋
¾
) is the maximum t such that 2 n contains (¾ + 1)-
chains C1 , . . . , Ct such that every member of each chain is incomparable to all members of
all the other chains in the list. (Griggs–Stahl–Trotter [1984])
12.2.38. Let be a family of finite sets A1 , . . . , A m . Let Ì( ) denote the minimum size of
a set intersecting each Ai . A family is Ì-critical if Ì( − {Ai }) < Ì( ) for all i.
−1
(a) Prove ∑i=1 (| Ais|+ s) ≤ 1 for Ì-critical with Ì( ) = s+1. (Hint: Exercise 12.2.37a.)
m

(b) Prove that if is r-uniform and every Ì(A) ≤ s for every subfamily A with | A| ≤
(r+s s) , then Ì( ) ≤ s. State the special case for graphs, using the vertex cover number Î(G).

12.3. Linear Extensions and Dimension

Few partial orders are chains, but chains are useful in understanding more
complicated posets. We can describe a poset using chains that are consistent with
the order relation, called “linear extensions”.
Section 12.3: Linear Extensions and Dimension 569

ORDER DIMENSION
When purchasing a new car, a buyer considers many models. Between any
two, the buyer may prefer one or be undecided. Assuming that the preference
relation is a partial order, let P be the resulting poset. The buyer may try to
encode P by rating the cars on criteria such as price, reliability, mileage, roomi-
ness, color, styling, etc. Assume that cars are ranked linearly on each scale. The
scales “realize” P when car x is preferred to car y in P if and only if x is preferred
to y on each scale. We try to realize a poset using a small number of linear orders.

12.3.1. Definition. An extension of a poset P is a partial order on the elements

of P that contains all the relations of P. A linear extension is an extension
that is a chain. The intersection of partial orders on a given set is the set
of relations that appears in each of them.

12.3.2. Proposition. A finite poset is the intersection of its linear extensions.

Proof: A linear extension iteratively lists (and deletes) a minimal unlisted ele-
ment. All linear extensions arise in this way. For every down-set F in P , this
procedure can produce a linear extension that lists all of F before all of P − F . If
x and y are incomparable elements of P , then x ∈ / D[y] and y ∈ / D[x]. We thus
have a linear extension listing x after y (and all of D(y)) and another listing y af-
ter x (and all of D(x)). Hence x < y in P if and only if x precedes y in every linear
extension. Thus P is the intersection of all its linear extensions.

In the algorithmic literature, linear extensions are called topological or-

derings. Proposition 12.3.2 is more subtle for infinite posets; one applies Zorn’s
Lemma to the family of all extensions, ordered by inclusion (Szpilrajn [1930]).

12.3.3. Definition. A realizer of a poset P is a set of extensions whose intersec-

tion is P. The (order) dimension dim P is the minimum number of linear
extensions forming a realizer of P. Let I(P) denote the set of ordered pairs of
incomparable elements in P. For (x , y) ∈ I(P), an extension L establishes
the pair (x , y) if x < y in L.

A set of extensions realizes P if and only if for every (x , y) ∈ I(P), some exten-
sion in the set has x < y. That is, every incomparable pair must be established.

12.3.4. Example. A poset has dimension 1 if and only if it is a chain. Antichains

have dimension 2 (list the elements in some order and the reverse order).
The poset below also has dimension 2. Using 123456 as one extension, the
other must have 4 < 1 and 6 < 3; the extension 412563 completes a realizer. Not
every linear extension is part of a realizer of size 2 (see Exercise 3). For example,
starting with 124356, we cannot establish all of 4 < 1, 3 < 4, and 6 < 3 with one
additional extension, since 6 > 1.
•6
•5 •3
4• 2•
1•
570 Chapter 12: Partially Ordered Sets

12.3.5. Theorem. (Dushnik–Miller [1941]) A poset P with comparability graph

G has dimension at most 2 if and only if G is also a comparability graph.
Proof: Necessity. Let L1 and L2 be two linear extensions of P with intersection P.
We have x y if and only if x and y appear in opposite order on L1 and L2 . Hence
x y if and only if x and y are in the same order on L1 and L 2 , where L 2 is the
reverse of L2 . Now G is the comparability graph of the intersection of L1 and L 2 .
Sufficiency. Let F be the comparability digraph of P. Let F̂ be a transitive
orientation of G (edges point up), and let F̂ be its reverse (also transitive). Both
F ∪ F̂ and F ∪ F̂ are orientations of K n , where n = | P | . A non-transitive orien-
tation of K n has a directed 3-cycle, but this would violate the transitivity of F or
F̂ or F̂ . Hence F ∪ F̂ and F ∪ F̂ are transitive orientations of K n .
A transitive orientation of K n linearly orders its vertices by outdegree. Hence
there are chains L and L on V(G) whose comparability digraphs are F ∪ F̂ and
F ∪ F̂ . The intersection of these two chains is P , since two elements are ordered
in the same way on L and L if and only if they are adjacent in F , which is the
comparability digraph of P.

1 5 6•
• •
4 5 6 4 5 6 5• •3
3• •4 • • • • • •
4• •2
• • • • • • • •
2 6 1 2 3 1 2 3 •1
F F̂ F̂ P

F ∪ F̂ yields 1 < 2 < 3 < 4 < 5 < 6.

F ∪ F̂ yields 4 < 1 < 2 < 5 < 6 < 3.

On the other hand, dimension can be arbitrarily large.

12.3.6. Example. (Dushnik–Miller [1941]) The standard example Sn is the

subposet of 2[n] induced by the singletons and their complements, denoted by i
and i for i ∈ [n] (S4 appears below). In a realizer of Sn , for each i there must be
a linear extension in which i appears above i, establishing the incomparable pair
(i , i). This forces the other singletons to appear below this pair and the other sets
of size n − 1 to appear above them. Thus n distinct extensions are needed. Any n
extensions establishing these pairs also establish all others.

1 2 3 4
• • • •

• • • •
1 2 3 4

One motivation for studying dimension is compact encoding of n-element

posets. A 0 , 1-matrix for the order relation takes n2 bits, testing x < y in unit
Section 12.3: Linear Extensions and Dimension 571

time by checking an entry. A realizer of size ¾ uses ¾ n log2 n bits. This takes less
storage when dim P ≤ O(n/log2 n), although testing x < y takes ¾ comparisons.
Encoding each element by its heights on the ¾ extensions of a realizer embeds
a poset in ¾ under the product ordering. Indeed, the least such ¾ is another
definition of dimension, often attributed to Ore but given earlier by Hiraguchi.

12.3.7. Theorem. (Hiraguchi [1955], Ore [1962]) A partial order P has a real-
izer of size ¾ if and only if it embeds in the product of ¾ chains.
Proof: Given a realizer, we obtain such an embedding using the heights on the
extensions to map each element of P to an element of the product. Conversely,
given an embedding that maps x ∈ P to (x1 , . . . , x ¾), we seek a realizer by placing
the elements on the ith extension in the order of their ith coordinates.
These values need not be distinct; there may be “ties” in coordinate i. Let
S = {x ∈ P: x i = t}. To break the ties, expand this position on the ith extension
into any linear extension of the subposet S. This preserves all the relations of P ,
so we have obtained linear extensions.
To show that this is a realizer, consider x , y ∈ S with x y. Because the encod-
ing embeds P , there are coordinates j and ¾ such that x j < yj and x¾ > y¾ . The
extensions corresponding to these coordinates have x < y and y < x, respectively,
which is not affected by breaking ties in another coordinate.

Theorem 12.3.7 yields dim 2[n] ≤ n, since 2[n] is a product of n chains. Komm
[1948] showed that dim 2[n] = n. To see that 2[n] cannot embed in a product of
fewer chains, we use Sn and the next observation.

12.3.8. Corollary. If Q is a subposet of P , then dim Q ≤ dim P.

Proof: A subposet of P embeds wherever P embeds. Also, dropping P − Q from
the extensions in a realizer of P yields a realizer of Q.

By Example 12.3.6, Theorem 12.3.7, and Corollary 12.3.8, products of n non-

trivial chains have dimension n (Ore [1962]). We next prove a more general result.
The upper bound was observed by Hiraguchi [1951]. The sufficient condition for
equality was originally proved using completion of lattices; we present a later
more explicit proof. A poset is bounded if it has a unique minimal element 0̂ and
a unique maximal element 1̂.

12.3.9. Theorem. (Baker [1961]) If P and Q are posets, then dim (P × Q) ≤

dim P + dim Q, with equality if P and Q are bounded posets of size at least 2.
Proof: (Kelly [1981]) Let º : P → m and ½ : Q → n be optimal embeddings.
Define h: P × Q → m+n by letting h((p , q)) be the concatenation of º (p) and ½(q).
This is an embedding of P × Q, so dim P × Q ≤ m + n.
Given bounded posets P and Q with size at least 2, let L1 , . . . , Lt be a realizer
of P × Q, indexed so L1 , . . . , Lm are the extensions with (0̂ , 1̂) < (1̂ , 0̂). Using
(0̂ , 1̂)(1̂ , 0̂), we force a realizer of P × {0̂} among the m extensions with (0̂ , 1̂) <
(1̂ , 0̂) and a realizer of {0̂} × Q among the t − m extensions with (0̂ , 1̂) > (1̂ , 0̂).
Thus m ≥ dim P and t − m ≥ dim Q, so dim (P × Q) = t ≥ dim P + dim Q.
If P is a chain, then already m ≥ dim P , so we may choose (a , b) ∈ I(P). Since
(a , 1̂)(b , 0̂), we must have (a , 1̂) < (b , 0̂) on some extension. Since (0̂ , 1̂) < (a , 1̂)
572 Chapter 12: Partially Ordered Sets

and (b , 0̂) < (1̂ , 0̂), such an extension will have (0̂ , 1̂) < (1̂ , 0̂) and must be one of
L1 , . . . , Lm . Furthermore, since (a , 0̂) < (a , 1̂), this extension has (a , 0̂) < (b , 0̂).
Since (a , b) ∈ I(P) was arbitrary, L1 , . . . , Lm contains a realizer of the copy P ×
{0̂} of P , so m ≥ dim P. By analogous reasoning, the extensions with (0̂ , 1̂) >
(1̂ , 0̂) realize {0̂} × Q, so there are at least dim Q of those.

To see that dim P × Q can be less than dim P + dim Q, consider the simplest
product involving a non-bounded poset: dim (• •) × (•) = 2 = dim (• •) + dim (•) − 1.
However, it seems that it cannot be much less.

12.3.10. Conjecture. (Kelly–Trotter [1982]) If P and Q are posets, then dim P ×

Q ≥ dim P + dim Q − 2.

Little is known about this conjecture. Trotter [1985] proved dim Sn × Sn =

2n − 2, and Reuter [1989a] extended this to dim Sm × Sn = m + n − 2 (he also proved
dim P × P ≥ 4 when dim P = 3).

COMPUTATION AND BOUNDS

Since a set of extensions realizes P if and only if each ordered pair (x , y) ∈ I(P)
appears in some extension, computing dimension is equivalent to covering of I(P)
by the fewest sets of pairs that can appear in one extension. This expresses dimen-
sion as hypergraph coloring (recall that the chromatic number of a hypergraph is
the minimum size of a vertex partition into sets containing no edge).

12.3.11. Definition. An alternating cycle of incomparable pairs in P is a set

{(x i , yi)}¾i=1 in I(P) such that yi ≤ x i+1 in P for all i (indices modulo ¾).

If x y, then (x , y) and (y , x) together form an alternating cycle.

12.3.12. Lemma. The dimension of P is the minimum number of classes covering

I(P) such that no class contains an alternating cycle.
Proof: A single extension cannot establish all pairs in an alternating cycle, be-
cause the relations x i < yi together with yi ≤ x i+1 violate transitivity. Thus the
number of classes needed is a lower bound on dim P.
For the opposite inequality, let S be a subset of I(P) containing no alternat-
ing cycle. The digraph having y → x whenever x < y in P or (x , y) ∈ S is acyclic.
A linear extension of the transitive closure of this digraph is a linear extension
of P that contains all relations in S. Thus dim P is at most the number of classes
in such a covering.

Thus dim P is the chromatic number of a hypergraph where the vertices are
I(P), the edges are the alternating cycles, and the colors are the linear exten-
sions. We do not need all of this hypergraph to compute dim P. In computing
dim Sn it was enough to establish the incomparable pairs (ī , i); the others were
then necessarily also established. In general, it suffices to establish the “crucial”
incomparable pairs.
Section 12.3: Linear Extensions and Dimension 573

12.3.13. Definition. Among ordered incomparable pairs, (a , b) forces (c , d) if

c ≤ a and b ≤ d (every extension with a < b has c < d). A pair (x , y) ∈ I(P)
is unforced if x < y is not implied by adding any other pair from I(P) to P.
Let C(P) denote the set of unforced pairs.

In the literature, when (x , y) is an unforced pair, the ordered pair (y , x) is

called a critical pair. Realizing a poset requires reversing the critical pairs,
which is equivalent to establishing the unforced pairs.

12.3.14. Proposition. An ordered incomparable pair (x , y) is an unforced pair if

and only if D(y) ⊆ D(x) and U(x) ⊆ U(y).
Proof: The pair (x , y) fails to be unforced if and only if there is an incomparable
pair (a , b) other than (x , y) such that adding the relation a < b forces x < y. Such
forcing occurs if and only if x ≤ a and b ≤ y. Hence it will occur if any element of
U[x] is incomparable to any element of D[y]. Since x and y must remain incom-
parable, the forcing fails if and only if all of U[x] is above all of D[y], except for
the pair (x , y) itself. This is equivalent to U(x) ⊆ U(y) and D(y) ⊆ D(x).

12.3.15. Theorem. The dimension of P is the minimum number of linear exten-

sions establishing the unforced pairs of P. This equals the chromatic number
of the hypergraph H(P) with vertex set C(P) whose edges are the minimal al-
ternating cycles consisting of unforced pairs.
Proof: By Definition 12.3.13, extensions that establish x < y for all (x , y) ∈ C(P)
also establish all incomparable pairs.

The hypergraph H(P) is useful for computing dimension in special classes be-
cause the lower bound can be established by exhibiting any subgraph of H(P) with
the desired chromatic number. The upper bound is then verified by exhibiting a
realizer that has all the unforced pairs, rather than by verifying that every edge
of H(P) is properly colored. For the standard example, C(Sn) = {(ī , i): 1 ≤ i ≤ n}.
The minimal alternating cycles are precisely the sets of two unforced pairs so
H(Sn) = K n and Ò(H(Sn)) = n.

12.3.16.* Remark. Since computing dim P is a hypergraph coloring problem,

it is not surprising that testing dim P ≤ 3 is NP-complete. Yannakakis [1982]
showed this by constructing from any graph a poset that has dimension 3 if and
only if the graph is 3-colorable. Recognition of 2-dimensional posets runs in time
linear in the number of comparabilities (McConnell–Spinrad [1999]). Structural
descriptions of 2-dimensional posets allow problems that are hard in general to
run quickly on this class (see Möhring [1985], Spinrad [1982, 2003]).
We can also seek a forbidden subposet characterization of d-dimensional
posets. A poset P is irreducible if deleting any element reduces its dimension; it
is ¾-irreducible if also dim P = ¾ . The only 2-irreducible poset is the 2-element
antichain. Kelly [1977] and Trotter–Moore [1976b] independently found all 3-
irreducible posets (see the dimension survey in Kelly–Trotter [1982]). The list
includes seven infinite families and ten small examples.

Although exact computation of dim P is difficult, there are bounds in terms

of other parameters. For Sn , the dimension equals the width and is half the size.
574 Chapter 12: Partially Ordered Sets

Hiraguchi proved that this is extremal for both parameters. The proof of dim P ≤
| P | /2 developed here is due independently to K imble [1973] and Trotter [1975].
12.3.17. Definition. A linear extension L puts Y over X if X and Y are disjoint
subposets and y is above x in L whenever x y, x ∈ X , and y ∈ Y . For Q ⊆ P ,
an upper extension of Q is a linear extension that puts P − Q over Q; a
lower extension puts Q over P − Q.

12.3.18. Lemma. Every chain in a poset has upper and lower extensions.
Proof: By symmetry, it suffices to find upper extensions. Let C be a chain in a
poset P consisting of x1 , . . . , x ¾ from bottom to top. Let L be a linear extension of
P − C. For 1 ≤ i ≤ ¾ − 1, let Yi be the set of elements of P − C that are less than
x i+1 but not less than x i ; also let Y0 = D(x1) and Y¾ = P − D[C]. Form a linear
ordering L of the elements of P by inserting between x i and x i+1 all elements of
Yi in the order that they have on L. Similarly insert Y0 before x1 and Y¾ after x¾ .
If y < y for y ∈ Yi and y ∈ / C, then transitivity yields y ∈ Yj for some j ≤ i,
so L puts y and y in the right order. Also x i and y appear in the right order
when they are comparable. If x i  y, then y ∈ Yj for some j ≥ i. Thus L is an
upper extension of C.

Not all subposets have upper or lower extensions. Rabinovitch [1978] deter-
mined when there is an extension putting Y over X (Exercise 16).

12.3.19. Theorem. (Dilworth [1950], Hiraguchi [1955]) dim P ≤ w(P).

Proof: Start with a Dilworth decomposition of P (a partition into w(P) chains),
and take an upper extension of each chain. The resulting extensions form a real-
izer, since incomparable elements appear on different chains in the original par-
tition C. If x y, then x appears above y on the extension arising from the chain
of C containing y, and y appears above x on the extension arising from the chain
of C containing x.

In addition to the notations U(x) and D(x) for the sets of elements above and
below x, we also use I(x) = {y ∈ P: y x}. Thus P − x = U(x) ∪ D(x) ∪ I(x).

12.3.20. Theorem. (One-Point Removal Theorem; Hiraguchi [1951]) If x is

an element of a poset P , then dim P ≤ 1 + dim (P − x).
Proof: We construct a realizer of P from a realizer L of P − x. Extract from P the
subsets U(x), D(x), U(x) ∪ I(x), and D(x) ∪ I(x). Using in each of these subposets
the order given by an extension L in L, form the two extensions U(x) ∪ I(x) > x >
D(x) and U(x) > x > D(x) ∪ I(x). Replace L in L by these two extensions, and
insert x anywhere between U(x) and D(x) on the other extensions.

U(x)
U(x) ∪ I(x) U(x)
x• I(x) x ←→ x
D(x) D(x) ∪ I(x)
D(x)
L
Section 12.3: Linear Extensions and Dimension 575

We claim that these 1 + dim (P − x) extensions realize P. The two new exten-
sions establish the incomparable pairs involving x. We must also consider incom-
parable pairs (y , ) with y ∈ I(x) and ∈ U(x) ∪ D(x) that were established by L.
Since the two new extensions have U(x) ∪ I(x) and D(x) ∪ I(x) in order as on L,
the pair (y , ) appears as it did in L in one of the new extensions.

12.3.21. Theorem. If A is an antichain of P , then dim P ≤ max{2 , | P − A|}.

Proof: As a basis for induction, we want dim P ≤ 2 when | P − A| ≤ 2. This re-
duces by easy remarks to several cases like the posets shown below, where A is
the marked antichain (see Exercise 13).
For | P − A| > 2, deleting all but two elements of P − A from P leaves a poset
with only two elements outside A. Hence the One-Point Removal Theorem induc-
tively yields dim P ≤ | P − A| when | P − A| > 2.

• • • •
•
• • • • • • • • • • •
•

12.3.22. Corollary. (Hiraguchi’s Inequality; Hiraguchi [1955]) If | P | ≥ 4,

then dim P ≤ | P | /2.
Proof: To obtain dim P ≤ | P | /2 when w(P) > | P | /2, apply Theorem 12.3.21; oth-
erwise apply Theorem 12.3.19.

The technique of Theorem 12.3.20 yields other removal theorems, bounding

dim P in terms of the dimension of a subposet Q. Start with a realizer of Q, mod-
ify and/or add extensions appropriately, and show that all the incomparable pairs
of P are established (Exercises 17–22).
Theorem 12.3.20 and Corollary 12.3.22 suggest a famous conjecture.

12.3.23. Conjecture. (Two-Point Removal Conjecture) Every poset with at

least three elements has a removable pair of elements; a pair {x , y} such
that dim (P − {x , y}) ≥ dim P − 1.

Removable pairs are discussed in Exercises 19–26. Tator [1983] proved a

weaker statement than Conjecture 12.3.23: always there exist four points in P
whose removal decreases the dimension by at most 2 (Exercise 22).

BIPARTITE POSETS

The standard examples Sn are posets whose elements are all maximal or min-
imal. We consider more general such posets.

12.3.24. Definition. A bipartite poset is a poset having no 3-element chains

(bipartite comparability graph). A bipartite poset is normal if (1) its compa-
rability graph is connected, and (2) when x and y are on the same level, some
element other than x is comparable to x but not to y.
576 Chapter 12: Partially Ordered Sets

12.3.25. Remark. In a normal bipartite poset, the unforced pairs are the pairs
(x , y) ∈ I(P) such that x is maximal and y is minimal.

Various bipartite subposets of 2 n generalize Sn and lead to natural dimension

problems. We study those consisting of two ranks.

12.3.26. Definition. Write 2 nl,¾ for the subposet of 2 n induced by the l-sets and
¾-sets. Let d n(l , ¾) = dim 2 nl,¾ .

The case of most interest is d n(1 , ¾). Spencer [1971] showed that d n(1 , ¾) ∼
c¾ lg lg n when ¾ is constant (we henceforth use lg√for log2 and ln for log e). Dush-
nik [1950] computed d n(1 , ¾) exactly when ¾ ≥ 2 n; the exact result appears in
Theorem 12.3.29. After a slow√ decline in d n(1 , ¾) as ¾ decreases from n, the drop
becomes rapid for ¾ below 2 n. The upper and lower bounds when ¾ ∈ o(n) differ
by a factor of ln n.
¾ d n(1 , ¾) reference
n−1 n standard example Sn
n−1
2 ≤ ¾ ≤ n− 2 n−1 Dushnik [1950]
n
3< ¾ ≤ n2 −
√1 n− 2 Dushnik [1950]
¾=n ≥2 n n − n1− + O(n2−3 ) Dushnik [1950]
general
√ < ¾(¾ + 1) ln(ne/¾) Füredi–K ahn [1986]
2 ≤ ¾ ≤ 2 n−4 ≥ 14 (¾ + 2)2 Exercise 29
constant ∼ c lg lg n Spencer [1971]
2 ∼ lg lg n Spencer [1971]
These results use different methods. Dushnik’s lower bound uses the Pigeon-
hole Principle, Spencer ’s bounds are by relating the problem to other questions,
and the Füredi–K ahn upper bound is probabilistic.
For bipartite posets in general, the computation of dimension can be reduced
to constructing an appropriate set of permutations of the minimal elements. By
“permutation”, here we mean a linear ordering written out as a list of elements
in order; we use “permutation” to distinguish a linear ordering of some of the
elements from a linear extension of the full poset. An element of a permutation
“comes later than” or “follows” all the elements to its left.

12.3.27. Definition. Given a bipartite poset P , let X and Y be the sets of mini-
mal and maximal elements, respectively. For y ∈ Y , let Sy = {x ∈ X : x < y}.
A set {L1 , . . . , Lt } of permutations of X is a suitable set for P if whenever
y  x with y ∈ Y and x ∈ X , some Li puts x later than all of Sy .

12.3.28. Lemma. If P is a normal bipartite poset, then dim P equals the mini-
mum size of a suitable set for P.
Proof: Let t be the minimum size of a suitable set for P. Given a suitable set of
size t, in each permutation we insert each maximal element y immediately after
the last element of Sy . The resulting linear orderings can be viewed as linear
extensions of P. In fact, they form a realizer of size t, since the unforced pairs
are the max-min pairs (y , x) with x ∈
/ Sy . Hence dim P ≤ t.
Section 12.3: Linear Extensions and Dimension 577

For the opposite inequality, consider a smallest realizer of P ; this is a set of

dim P linear extensions of P. In each such extension, any maximal element y of P
must come later than each element of Sy , since extensions of P preserve the order
relation. Hence in an extension where x follows y, also x follows all of Sy . If x y,
then x must follow y in some extension in the realizer. Therefore, deleting the
maximal elements from each extension in the realizer yields a set of permutations
of the minimal elements that by definition is a suitable set. Hence t ≤ dim P.
y
Y • extension
Sy y x
X •
Sy x

In studying d n(1 , ¾), we thus seek realizers as suitable sets of permutations

of [n]. Lemma 12.3.28 immediately implies that d n(1 , ¾) is nondecreasing √ in ¾ .
Dushnik’s result below thus yields d n(1 , ¾) exactly for all ¾ with ¾ ≥ 2 n.
√
12.3.29. Theorem. (Dushnik [1950]) If 1 ≤ r ≤ n, then d n(1 , ¾) ≤ n − r if and
only if ¾ ≤ n/r + r − 3.
Proof: We show that a suitable set of n − r permutations of [n] exists for 21n ,¾ if
and only if ¾ ≤ n/r + r − 3. Let t = n − r.
Necessity. Let L1 , . . . , Lt be a suitable set for 21n ,¾ . By symmetry, we may
assume that 1 is last in L1 . Thus in L1 , element 1 follows all ¾-sets omitting 1.
Hence in L2 , . . . , Lt we may move 1 to the beginning. By arguing similarly for
2 , . . . , t on L2 , . . . , Lt , we may assume that each Li ends with i.
Let R = {t + 1 , . . . , n}, so | R| = r. For a fixed element x ∈ R, let S be the set
of indices i ∈ [t] such that x appears last among R in Li . Let A = (R − {x}) ∪ S.
Note that x does not follow A x in any Li (if i ∈ S, then i follows x in Li ; otherwise,
x is not last among R − {x} in Li). Every ¾-set not containing x must precede x on
some Li , so | A| > ¾ . Thus r − 1 + | S| > ¾ , so | S| > ¾ − r + 2.
On the other hand, the Pigeonhole Principle guarantees that some x ∈ R
appears last among R at most t/r times. Thus | S| ≤ t/r for some x. Now ¾ − r + 2 ≤
min | S| ≤ t/r. Using t = n − r, we obtain ¾ ≤ n/r + r − 3. Hence this inequality is
necessary for d n(1 , ¾) ≤ t.
Sufficiency. We define permutations L1 , . . . , Lt of [n] that meet the √ necessity
conditions above and form a suitable set if ¾ ≤ n/r + r − 3 and r ≤ n. The last
element of Li is i, preceded immediately by R in some order. Each element of R
is next-to-last in ⌊ t/r⌋ or ⌈ t/r⌉ of the permutations.
√
Since r ≤ ⌊ n⌋ ≤ ⌊ n/r⌋ , we have r − 1 ≤ ⌊(n − r)/r⌋ = ⌊ t/r⌋ . Thus for x ∈ R,
we can make each element of R − {x} appear immediately before x in one of the
permutations where x is next-to-last. Thus for distinct x and y in R, there is a
permutation L j that ends y , x , j .
Let A be a set that does not appear before x in any L j . If x is next-to-last
in Li , then i ∈ A. If x is third-to-last in L j , followed by and j , then A must
contain or j (these pairs are disjoint and omit those i where x is next-to-last on
Li). Thus | A| ≥ ⌊ t/r⌋ + r − 1 > n/r + r − 3 ≥  . Hence L1 , . . . , Lt is a suitable set,
and  ≤ n/r + r − 3 yields d n(1 , ) ≤ t.
578 Chapter 12: Partially Ordered Sets
√
When ¾ ≥ 2 n, Theorem 12.3.29 gives d n(1 , ¾) exactly. For smaller ¾ , it gives
√
no better upper bound than n − n. A technique like that in Theorem 12.3.29
√
yields a lower bound d n(1 , ¾) ≥ ¾ 2/4 when ¾ ≤ n (Exercise 29). We present a
general upper bound.

12.3.30. Theorem. (Füredi–K ahn [1986]) d n(1 , ¾) ≤ ⌈ ¾(¾ + 1) ln(ne/¾)⌉ .

Proof: Generate permutations L1 , . . . , Lt of [n] by selecting each at random from
all n! orders. We show that if t ≥ ¾(¾ + 1) ln(ne/¾), then with positive probabil-
ity these form a suitable set of permutations for 21n ,¾ . Hence some outcome of the
experiment is a realizer of the desired size.
For each ¾-set S and each x ∈ [n] with x ∈ / S, the probability is 1/(¾ + 1) that
x follows all of S on L j . Hence the probability that x follows all of S in none of
the random permutations is ( ¾+¾ 1 )t . There are n(n−¾ 1) such pairs (S, x). Hence we
bound the probability that some pair is not established:

(failure) ≤ n(n −¾ 1 ) ( ¾ +¾ 1 )
t t
n 1
< ( ) (1 − )
¾ ¾+1
ne ¾ ne ¾
<( ) e−t/(¾+1) ≤ ( ) e−¾ ln(ne/¾) = 1 .
¾ ¾
¾ − x (Chapter 14), slightly
We used standard inequalities (¾n) ≤ ( ne
¾ ) and 1 − x ≤ e
weakening the Füredi–K ahn bound to simplify computation.

In the realm of constant ¾ , Spencer ’s construction of small realizers used

special families of sets (he attributed this argument to A. Hajnal).

12.3.31. Definition. A family F of sets is ¾-scrambling if for all choices

S1 , . . . , S¾ of ¾ distinct sets in F and all subsets A of the index set [¾], the
set (⋂r∈ A Sr) ∩ (⋂r/∈ A S r) is nonempty.

The ¾-scrambling condition can be stated in several ways. If F is 2-scrambling

and X , Y ∈ F , then X ∩ Y , X − Y , Y − X , and X ∩ Y are nonempty; for ¾ in gen-
eral, the condition states that in the Venn diagram on any ¾ sets in F , every cell
is nonempty. In terms of elements, the condition is that for S1 , . . . , S¾ ∈ F and
A ⊆ [¾], there exists an element x such that x ∈ Sr for r ∈ A and x ∈/ Sr for r ∈
/ A.
We will use ¾-scrambling families to prove an upper bound on d n(1 , ¾).

12.3.32. Example. The four permutations of the elements 0 through 7 shown

below form a suitable set for 281 ,2 , proving d8(1 , 2) ≤ 4.
L1 : 0 , 1 , 2 , 3 , 4 , 5 , 6 ,7
L2 : 3 , 2 , 1 ,0 ,7 ,6 ,5 ,4
L3 : 5,4,7,6,1 ,0,3,2
L4 : 6,7,4,5,2,3,0,1
It is easy to show directly that this is a suitable set. An element at the end
of a permutation follows all pairs among the other elements in that permutation.
This takes care of the singletons 1 , 2 , 4 , 7. The elements 0 , 3 , 5 , 6 appear next-
to-last. When r is next-to-last in a permutation, on that permutation it follows
Section 12.3: Linear Extensions and Dimension 579

all pairs of other elements except the pairs involving the last element s. In each
case, there are two permutations where s immediately precedes r, and each of the
other six elements precedes s in one of those permutations.

These carefully structured permutations arose as a special case of a general

construction using ¾-scrambling sets.

12.3.33. Lemma. Given a ¾-scrambling family S1 , . . . , Sm of subsets of [t], there

exist permutations L1 , . . . , Lt of the numbers 0 through 2 m − 1 such that if
a < b and j is the leftmost position where the m-bit binary expansions of a
and b differ, then b follows a on Li if and only if i ∈ Sj .
Proof: Begin with all the m-bit binary integers in one “group”. To produce Li ,
perform Steps 1 through m in order as follows. On Step j , each current group
splits into two smaller groups. In a current group X , let X r be the numbers whose
expansion has r in coordinate j , for r ∈ {0 , 1}. Replace X with X 1 after X 0 if i ∈
Sj ; otherwise put X 0 after X 1 .
After Step j , the groups all have size 2 m− j ; thus an explicit ordering Li is
produced after Step m. Furthermore, the relative ordering between a and b on
Li is determined in Step j , where j is the leftmost position where the expansions
of a and b differ (a has 0 there; b has 1). The procedure explicitly puts b in a group
after a if and only if i ∈ Sj .

Example 12.3.32 arises from Lemma 12.3.33 using the family given by S1 =
{1 , 2}, S2 = {1 , 3}, and S3 = {1 , 4}. This is a 2-scrambling family of subsets
of [4]; here m = 3. For each Li , Step 1 splits the elements into the two groups
{0 , 1 , 2 , 3} and {4 , 5 , 6 , 7}. The lower group goes first when i ∈ S1 , which holds
for i ∈ {1 , 2}; in L3 and L4 the lower group goes last. Step 2 splits the lower
group into {0 , 1} and {2 , 3} and the upper group into {4 , 5} and {6 , 7}. Within
each group, the lower subgroup goes first when i ∈ S2 , which holds for i ∈ {1 , 3}.
Step 3 finishes the job, deciding which goes first in each pair.
We will use this construction for the upper bound in the next theorem. Given
m sets forming a ¾-scrambling family of subsets of [t], we will obtain d n(1 , 2) ≤ t
when n = 2 m . Hence we seek a large ¾-scrambling family.
Let M(t , ¾) be the maximum size of a ¾-scrambling family in 2 t . We are par-
ticularly interested in ¾ = 2. The family of all ⌊ t/2⌋-sets in [t] that contain the
√
t−1
element 1 is 2-scrambling, so M(t , 2) ≥ (⌊t/2⌋ −1 ) > 2 / 2 t. For  ∈
t
, there is a
constant c¾ such that M(t , ) ≥ ct¾ (Exercise 31).

12.3.34. Theorem. (Spencer [1971]) If c is a constant such that [t] has a -

scrambling family of size greater than ct , then
lg lg(n − 1) < d n(1 , ) < 1
lg c lg lg n.

Proof: Upper bound. We prove d n(1 , ) ≤ t for n = 2 M(t ,¾) . Since M(t , ) > ct , this
yields d n(1 , ) < lg1 c lg lg n. For convenience, let m = M(t , ). Let {S1 , . . . , Sm} be
a largest -scrambling family of subsets of [t]. Lemma 12.3.33 provides orderings
L1 , . . . , Lt of 0 through n − 1 such that if a < b and j is the leftmost position
where a and b differ as vectors, then b follows a on Li if and only if i ∈ Sj .
580 Chapter 12: Partially Ordered Sets

We claim that {L1 , . . . , Lt } is a suitable set for 21n ,¾ . Viewing the elements
as binary m-vectors, consider a vector b and vectors a1 , . . . , a¾ other than b. For
1 ≤ r ≤ ¾ , let jr be the first coordinate where ar and b differ. Let A = {r: bjr = 1},
so A ⊆ [¾]. Since {S1 , . . . , Sm} is a ¾-scrambling family and { j1 , . . . , j¾ } is a set
of at most ¾ indices, there is a value i ∈ [t] such that i ∈ Sjr for r ∈ A and i ∈ / Sjr
for r ∈/ A. We claim that b occurs after all of a1 , . . . , a¾ on Li .
In constructing Li , element b is compared with ar when processing coordinate
jr . If bjr = 1, then r ∈ A and i ∈ Sjr . If bjr = 0, then r ∈ / A and i ∈ / Sjr . In either
case, b is placed after ar at stage j . Hence b follows each of a1 , . . . , a¾ in Li . We
conclude that L1 , . . . , Lt is a suitable set.
Lower bound. A suitable set of permutations for 21n ,¾ is also a suitable set for
21n ,¾−1 . Hence by Lemma 12.3.28 it suffices to prove the lower bound for ¾ = 2.
We prove that if n ≥ 22 + 1, then any t orderings of [n] yield a triple that ap-
t

pears monotonically (increasing or decreasing) in each ordering. For such a triple

{x , y , } with x < y < , we have y between x and in each permutation. Hence y
never appears after {x , }. We conclude that a suitable set must have more than
t permutations, so d n(1 , 2) > t if t ≤ lg lg(n − 1).
We prove the claim by induction on t using the Erdős–Szekeres Theorem (The-
orem 10.1.13): in every list of m2 + 1 distinct numbers some m + 1 numbers appear
monotonically (Exercise 12.1.15 requests a proof using Dilworth’s Theorem). For
t = 1, five elements suffice to guarantee a monotone triple. For t > 1, consider
orderings L1 , . . . , Lt on [22 + 1]. The Erdős–Szekeres Theorem yields a set S of
t

2 t−1
size 2 + 1 that appears monotonically in Lt . Within S, the induction hypothe-
sis yields a triple {x , y , } that appears monotonically in each of L1 , . . . , Lt−1 . By
the choice of S, this triple is also monotone in Lt .

12.3.35. Corollary. (Spencer [1971])

lg lg n ≤ d n(1 , 2) < lg lg n + 1
2 lg lg lg n + O(1).

t−1
√
Proof: We have M(t , 2) ≥ (⌊t/2⌋ −1) > 2 / 2 t (using Stirling ’s Formula). By the
t

argument√in Theorem 12.3.34, d n(1 , ) ≤ t when lg n = M(t , ). Solving for t in

lg n = 2 t/ 2 t yields the more precise upper bound.

Note the effect of the double exponential. Since M(t , 2) is the number of ⌊ t/2⌋-
subsets of [t] containing element 1, for t = 6 we have M(t , 2) = 10. The lower and
upper bounds thus yield 4 ≤ d1024(1 , 2) ≤ 6.
For d n(1 , 2), Corollary 12.3.35 establishes the asymptotic behavior. We
provided the next term of the upper bound because this is in fact sharp. In-
deed, there are four natural related problems whose answers in terms of n are
lg lg n + ( 21 + o(1)) lg lg lg n. We discuss three in the rest of this section and the
fourth in Exercise 12.4.7.
Meanwhile, we note that Hoşten–Morris [1999] determined exact values of
n where d n(1 , 2) increases. Biró–Hamburger–P ór–Trotter [2016] observed that
with the result of K leitman–Markowsky [1975], this determines d n(1 , 2) ex-
actly for almost all n, and within 1 otherwise. In particular, given any posi-
tive  , for sufficiently large n we have s −  < d n(1 , 2) < s + 1 +  , where s =
lg lg m + 12 lg lg lg n + 12 lg  + 12 .
Section 12.3: Linear Extensions and Dimension 581

12.3.36. Definition. The shift graph G n is the graph with vertex set ([n] )
2 and
edges defined by i j ↔ j ¾ if and only if i < j < ¾ (disjoint pairs are non-
adjacent). The double shift graph G n is the graph with vertex set ([n])
3 and
edges defined by i j ¾ ↔ j ¾ l if i < j < ¾ < l.

12.3.37. Lemma. (A. Hajnal) The chromatic number of the shift graph G n is
⌈ lg n⌉ ; this is the least t such that 2 t has at least n elements.
Proof: Given that 2 t has at least n elements, we properly color G n using [t] as
colors. Let A1 < A2 < · · · < An < · · · be a linear extension of 2 t . For each pair
i j ∈ V(G n) with i < j , color i j with some element of A j − Ai . Since the ordering is
an extension of 2 t , such an element exists. Since no element of A j − Ai can belong
to A¾ − A j , the coloring is proper.
Conversely, when Ò(G n) = t, we show that [t] has at least n subsets. Consider
a proper coloring of G n using [t] as colors. For each i ∈ [n], let Si be the set of
colors used on vertices of the form i j with j > i. If Si = Sj with j > i, then the
color c that appears on i j also appears on j ¾ for some ¾ > j . This is impossible in
a proper coloring, since i j ↔ j ¾ in G n . Hence S1 , . . . , Sn are distinct.

12.3.38. Theorem. (Füredi–P.Hajnal–Rödl–Trotter [1992]) The chromatic num-

ber of the double shift graph G n is the least t such that there are at least n
antichains in 2 t .
Proof: For the upper bound, we give a proper coloring. Instead of antichains,
consider the down-sets they generate. Let D1 , D2 , . . . be a linear extension of the
poset of down-sets in 2 t , ordered by inclusion. Associate with each pair {p , q} ∈
([n]
2
) having p < q a set Apq ∈ Dq − Dp . Given i < j < ¾ , color i j ¾ ∈ V(G n) with
an element of Aj¾ − Ai j . Such an element exists, since otherwise A j¾ ⊆ Ai j , but
the set Aj¾ not contained in Dj cannot be a subset of the set Ai j contained in the
down-set Dj . Now suppose that i j ¾ and j ¾ l both have color c. The first requires
c ∈ A j¾ , and the second requires c ∈
/ A j¾ . Thus the coloring is proper.
Conversely, let t = Ò(G n) and consider an optimal coloring; we show that 2 t
has at least n down-sets. For each pair i , j ∈ [n], let Si j be the set of colors ap-
pearing on vertices of the form i j ¾ with ¾ > j , and let Bi = {Si j : j > i}. Let Di be
the down-set consisting of all subsets of [t] contained in elements of Bi . We must
show that these down-sets are distinct. Note that Dn = ∅ and Dn−1 = {∅}; the
other down-sets contain non-empty sets.
Consider Di = Dj with i < j ≤ n − 2. Since Si j is nonempty and contained in
Di , and Dj is generated by Bj , the condition Di = Dj requires a set of the form
Sj¾ that contains Si j . This contradicts the coloring of G n , because the color in Si j
that is used on i j ¾ cannot appear in Sj¾ .

12.3.39. Corollary. (Erdős–A.Hajnal) The chromatic number of the double shift

graph G n is lg lg n + ( 12 + o(1)) lg lg lg n.
Proof: This follows from the K leitman–Markowsky expression 2(⌊n/2⌋)(1+o(1)) for the
n

number of down-sets in 2 n (see Section 11.3).

12.3.40. Corollary. The dimension of the subposet of 2 n induced by the sets of

sizes 1 and 2 is lg lg n + ( 12 + o(1)) lg lg lg n.
582 Chapter 12: Partially Ordered Sets

Proof: Corollary 12.3.35 establishes the upper bound. We prove that Ò(G n) is a
lower bound. Let L1 , . . . , Lt be a realizer. For i j ¾ with i < j < ¾ , choose c ∈ [t]
such that the singleton j is above the doubleton i¾ in Lc . Let c be the color of i j ¾ .
We cannot also give color c to j ¾ l with ¾ < l, because this would place i¾ < j <
jl < ¾ < i¾ on Lc . Hence this defines a proper t-coloring of G n , and Ò(G n) is a
lower bound on the dimension.

Although we have emphasized d n(1 , ¾), the dimension of the subposet of ¾-

sets and (n − ¾)-sets in 2 n is also of interest. Füredi [1994] proved d n(¾ , n − ¾) ≥
n − 2 ¾ + 2 for n > 2 ¾ (Exercise 16.2.18).

EXERCISES 12.3

12.3.1. (−) Prove that the intersection of two order relations on the same set is an order
relation.
12.3.2. (−) Prove that a set of elements forms a down-set in a poset P if and only if it is
an initial segment of some linear extension of P.
12.3.3. (−) List all linear extensions of the poset in Example 12.3.4. Determine which
belong to realizers of size 2.
12.3.4. (−) Describe all pairs of linear extensions realizing m + n.
12.3.5. (−) Prove that an n-vertex graph G is the comparability graph of a 2-dimensional
poset if and only if its vertices can be named v1 , . . . , vn so that there is a permutation of
[n] such that vi vj ∈ E(G) for i < j if and only if i > j .
12.3.6. Prove that the posets below have dimension 3. List four techniques that can be
used for the lower bound.
• • • • • • • • •
•
• • • • •
• •
• • • • • • • •
12.3.7. The “fence” poset F n has minimal elements x0 , . . . , x n and maximal elements
y1 , . . . , yn with comparable pairs given by xi−1 , xi < yi for 1 ≤ i ≤ n. Use an embedding of
F n in 2 to find all pairs of linear extensions realizing F n .
12.3.8. Prove that the poset below has dimension 3.
9• •10
7• •8
5• • 11 •6
3• •4
1• •2

12.3.9. The poset in Exercise 12.3.8 is 3-irreducible (deleting any element leaves a
2-dimensional poset), and it generalizes easily to 3-irreducible posets of all nontrivial
heights. Use this and Theorem 12.3.9 to prove the existence of n-irreducible posets of
large height.
Exercises for Section 12.3 583

12.3.10. Prove that a graph G is the complement of a comparability graph if and only if G
is the intersection graph of the curves graphing a set of continuous real-valued functions
on [0 , 1]. (Golumbic–Rotem–Urrutia [1983])
12.3.11. Let P be the graded poset with rank sizes a0 , . . . , ar such that elements are in-
comparable if and only if they have the same rank. Prove dim P = 2. Describe the minimal
alternating cycles of incomparable pairs and a proper 2-coloring of the hypergraph H(P).
12.3.12. Let {(xi , yi): 1 ≤ i ≤ ¾} be a minimal alternating cycle of incomparable pairs in
a poset P. Prove that {x1 , . . . , x¾ } and {y1 , . . . , y¾ } are antichains in P.
12.3.13. Complete the proof that dim P ≤ max{2 , | P − A|} by proving that dim P ≤ 2 if P
has an antichain A such that | P − A| = 2.
12.3.14. (♦) The composition (also called lexicographic product) Q[P1 , . . . , P¾ ] is
formed from a poset Q of size ¾ by expanding each xi ∈ Q to a copy of Pi ; elements expanded
from xi and xj are related as xi and xj are related in Q. Prove that if P = P0 [P1 , . . . , P¾ ],
then dim P = max i≥0 dim Pi . (Hiraguchi [1951])
12.3.15. (♦) Given a connected graph G, let P be the poset of subsets of V(G) that in-
duce connected subgraphs of G, ordered by inclusion. Prove that dim P is the number of
non-cut vertices in G. (Hint: Use distance from non-cut-vertices to partition the ordered
incomparable pairs into classes that avoid alternating cycles.) (Trotter–Moore [1976a])
12.3.16. (♦) Let X and Y be disjoint subposets of a poset P. Prove that P has a linear
extension putting Y over X if and only if P contains no copy of 2 + 2 with minimal elements
in Y and maximal elements in X . (Hint: Use induction on the number of incomparable
pairs (x , y) with x ∈ X and y ∈ Y .) (Rabinovitch [1978])
12.3.17. In a poset P that is not an antichain, let C be a chain, M be the antichain of max-
imal elements, and A be an antichain. Prove the following inequalities.
(a) dim P ≤ 2 + dim (P − C). (Hiraguchi [1951])
(b) dim P ≤ 1 + w(P − M). (Trotter [1975])
(c) dim P ≤ 1 + 2w(P − A). (Trotter 1975]; sharpness in Trotter [1974b])
12.3.18. (+) Let C be a chain in a poset P such that each element of P − C is incomparable
to at most one element of C. Prove that dim P ≤ 1 + dim (P − C). (Bogart–Trotter [1973])
12.3.19. (♦) Suppose that a is a maximal element in P , b is a minimal element in P , and
ab. Prove that dim P ≤ 1 + dim (P − {a , b}).
12.3.20. (♦) Given a < b in a poset P , let r(a , b) count the ordered pairs (x , y) ∈ I(P) such
that a < x and y < b. Prove that (a , b) is a removable pair if r(a , b) ≤ dim P − 3. (Hint: In a
realizer of P − {a , b}, replace a well-chosen extension with two others.) (Hiraguchi [1951])
12.3.21. (♦) Given incomparable elements a and b in a poset P , let r(a , b) count the ordered
pairs (x , y) ∈ I(P) such that x is comparable to both a and b and y is incomparable to both
a and b. Prove that (a , b) is a removable pair if r(a , b) ≤ dim P − 3. (Hint: In a realizer of
P − {a , b}, replace a well-chosen extension with two others.) (Kelly–Trotter [1982])
12.3.22. (♦) Four-Point Removal Theorem. Let P be a poset.
(a) Let C and D be chains in P such that x y for all x ∈ C and y ∈ D. Prove that P has
a linear extension that puts P − C − D over C and puts D over P − C − D. (Hint: Partition
P − C − D into the set X 1 below the top of C, the set X 2 above the bottom of D, and the
remainder X 3 . Combine linear extensions of X 1 ∪ C, X 2 ∪ D, and X 3 .) (Hiraguchi [1955])
(b) For C and D as in part (a), prove dim P ≤ 2 + dim (P − C − D). (Hiraguchi [1955])
(c) For x and y maximal in P , with D(x) ⊆ D(y), prove dim P ≤ 1 + dim (P − x − y).
(d) Given | P | ≥ 4, use parts (b) and (c) to prove that x , y , , w ∈ P exist such that
dim P ≤ 2 + dim (P − {x , y , , w}). (Tator [1983])
584 Chapter 12: Partially Ordered Sets

12.3.23. (♦) The ¾-dimension dim ¾(P) of a poset P is the minimum t such that P em-
beds in k t . Prove that if P decomposes into t chains of size less than ¾ , then dim ¾ P ≤ t.
(Comment: Thus ¾-dimension is well defined.) (Trotter [1976])
12.3.24. (♦) Prove that the maximum size of a minimal realizer of an n-element antichain
is ⌊ n2/4⌋ (for n ≥ 4). (Maurer–Rabinovitch [1977])
12.3.25. In the poset below, prove that (x , y) is an unforced pair whose removal decreases
the dimension by 2. (Reuter [1989b])
x
• • •
• •
•
• •y •
• • •
12.3.26. For n ≥ 5, we construct an n-dimensional poset Pn with 4n − 4 elements and an
unforced pair (y , x) such that dim (Pn − {x , y}) = n − 2 (see P5 below). Begin with disjoint
copies A and B of Sn−2 , with sets A1 and B1 of minimal elements, A2 and B2 of maximal
elements. Add four elements x , y , , w plus covering pairs B1 ≺ x ≺ A2 and A1 ∪ B1 ≺ y
and { , w} ≺ A2 ∪ B2 and w ≺ y. Prove that dim Pn = n, that (y , x) is an unforced pair, and
that dim (Pn − {x , y}) = n − 2. Which property fails for n = 4? (Kierstead–Trotter [1991])
y
A2 • • • • • • • B2
x
•

A1 • • • • • • • • B1
w

12.3.27. Prove that the n-dimensional standard example Sn is n-irreducible.

12.3.28. The crown Sn¾ is a bipartite poset with minimal elements i and maximal ele-
ments i for i ∈ [n + ], with i j when j ∈ {i , . . . , i + } (modulo n + ) and otherwise i < j .
(a) Show that  + 1 <  + 1 < · · · < 1 < 1 and  + 2 < 2 < · · · < 2  + 2 <  + 2 together
+¾)
establish the unforced pairs involving {1 , . . . ,  + 2}. Conclude dim Sn¾ ≤ ⌈ 2(n ¾ +2 ⌉ .
+¾)
(b) Use alternating cycles of length 2 to prove dim Sn¾ ≥ ⌈ 2(n ¾ +2 ⌉ .
(Trotter [1974a])
√
12.3.29. (♦) For 2 ≤  ≤ 2 n − 4, prove d n(1 , ) ≥ 14 ( + 2)2 . (Hint: Modify the argument
√
for necessity in the proof of Theorem 12.3.29 by letting R = {t + 1 , . . . , ⌈ t + t ⌉ }.)
12.3.30. (♦) For  < l, prove that d n( , l) ≥ n −  if l > (n +  − 1)/2, and d n( , l) ≥ l if
l ≥ 2  . (Hint: Start with Dushnik’s Theorem when  = 1.) (Füredi [1994])
12.3.31. (♦)  -scrambling sets.
(a) Prove that the ⌊ t/2⌋ -sets containing a fixed element form a maximum 2-scrambling
family in 2 t . (Hint: For odd t, use the Erdős–Ko–Rado Theorem (Theorem 11.2.18).)
(b) Prove existence of a -scrambling family in 2 t with size at least 12 [(1 − 2−¾)−1/¾ ]t .
(Hint: Generate m subsets of [t] independently at random, and prove that the probability
they are not -scrambling is less than 1 when m is smaller than desired.) (Spencer [1971])
12.3.32. Let  (r, ) be the maximum size of a -scrambling family of subsets of [r]. Let
(n , m) be the minimum size of a set of vertices in the hypercube Qn intersecting every m-
dimensional subcube. (Graham–Harary–Livingston–Stout [1993])
(a) Determine (4 , 2).
(b) Prove (n , m) = min{r:  (r, n − m) ≥ n}.
Section 12.4: Special Families of Posets 585

12.4. Special Families of Posets

We begin this section with “chain-like” posets, which maintain some aspects
of linear orders as tools of measurement and comparison. Subsequently, we con-
sider posets with special algebraic properties.

SEMIORDERS AND INTERVAL ORDERS

A ranking or weak order is a partial order whose elements occur in ranks

P1 , . . . , P¾ such that elements are incomparable if and only if they belong to the
same rank. Rankings are used in voting theory; one seeks a consensus ranking
among all voters. Paradoxes abound, and Arrow [1951] proved that no function
producing a consensus ranking can satisfy a particular set of four natural axioms.
Rankings are not general enough to model preferences realistically. The big
problem is that they require transitivity of indifference. A person given cups of
coffee with different amounts of sugar is likely to be indifferent when the amount
differs by one or two grains, but a large enough difference yields a preference.
A difference of a few dollars in the price of a house won’t affect one ’s attitude
toward it, but thousands of dollars will. Luce [1956] introduced a model for “ just-
noticeable” difference.

12.4.1. Definition. A semiorder is a poset representable by a function º and

fixed threshold ≥ 0 so that x < y if and only if  (y) −  (x) > (this is a
semiorder representation).

The rankings are the posets having semiorder representations with = 0. By

scaling, we may assume that the threshold is 1 when it is nonzero. The terms
weak order and semiorder suggest weakening the conditions for an order relation,
but these are quite restricted posets; what they weaken is the condition of total
order. We will characterize semiorders as the posets not containing 1 + 3 or 2 + 2.
There may be uncertainty not only in comparison of elements, but also in
assignment of values. For example, the skill of a tennis player may vary from day
to day, leading us to represent a player  by an interval [a x , bx]. We might then
conclude that y beats x if the interval for y is wholly above the interval for x.
That is, when a y > bx we expect y to win.

12.4.2. Definition. An interval order is a poset representable by assigning an

interval [a , b ] to each element  so that x < y if and only if bx < a y (the
assignment is an interval representation).

12.4.3. Example. Among topics to be discussed in a committee, we set x < y if

topic x must be settled before topic y is discussed. On the other hand, we set x y
if x and y will be available for discussion at the same time. A schedule assigns an
interval of time to each topic during which it is available for discussion.
586 Chapter 12: Partially Ordered Sets

12.4.4. Remark. Every semiorder is an interval order. From a semiorder represen-

tation º , letting [a x , bx] = [º (x) − 2 , º (x) + 2 ] yields an interval representation.
The incomparability graph of an interval order is an interval graph, because el-
ements are incomparable when their intervals in an interval representation in-
tersect. The incomparability graph of a semiorder has an interval representation
using intervals of length 1.

We will characterize interval orders as the posets not having 2 + 2 as a sub-

poset. The table below compares the four classes we have discussed.

class representing function(s) forbidden subposet

chain distinct values 1+1
ranking x < y if º (x) < º (y) 1+ 2
semiorder x < y if º (x) < º (y) − 1 + 3 and 2 + 2
interval order x < y if  (x) ≤ (x) <  (y) ≤ (y) 2+2

In specifying an interval order, the functions  and  above give the left and
right endpoints of the corresponding interval, respectively.
The characterizations of interval orders and semiorders by forbidden sub-
posets can be used to construct representations. Semiorders were characterized
much earlier, but it is convenient to characterize interval orders first and then
characterize semiorders among them. Fishburn–Monjardet [1992] noted that
posets without 2 + 2 were studied as early as Wiener [1914], who gave a char-
acterization in much different terminology.

12.4.5. Theorem. (Fishburn [1970], Mirkin [1972]) A poset is an interval order

if and only if it does not contain 2 + 2 as a subposet.
Proof: (Balof–Bogart [2003]) Suppose that P has an interval representation and
that x , y ,  , w are four elements inducing 2 + 2 with x < y and  < w. Let [ai , bi]
be the interval representing i ∈ {x , y ,  , w}. Because x < y and  < w, we have
bx < a y and b < aw . From xw and  y, we have bx ≥ aw and b ≥ a y . This yields
the contradiction bx < a y ≤ b < aw ≤ bx . Thus the condition is necessary.
For the converse, we use induction on | P | to produce an interval representa-
tion for a poset P without 2 + 2. Choose x ∈ P to maximize | D(x)| ; note that x is
a maximal element. Let P = P − x. Since P has no copy of 2 + 2, the induction
hypothesis yields an interval representation of P .
The key claim is that elements of I(x) are maximal. If  is incomparable to x
but not maximal, then choose w ∈ U(). Since | D(x)| ≥ | D(w)| and  ∈ D(w) − D(x),
there exists y ∈ D(x) − D(w). Since  x, also y is incomparable to  and w, and
the subposet formed by {x , y ,  , w} is 2 + 2.
Conversely, maximal elements are incomparable to x, so I(x) is the set of max-
imal elements in P other than x. In an interval representation of P , extend the
intervals for all maximal elements rightward to a common endpoint a. Since the
elements are maximal, this changes no intersections, and a is now the rightmost
point in the representation. Among these maximal elements, extend the inter-
vals for those that are also maximal in P further rightward to b. Adding an
interval for x that starts at (a + b)/2 completes an interval representation of P.
Section 12.4: Special Families of Posets 587

The earlier proof of Theorem 12.4.5 in Bogart [1993] produced an explicit

interval representation with the minimum number of endpoints (Exercise 5). We
build on Theorem 12.4.5 to characterize semiorders, as did Balof–Bogart [2003].

12.4.6. Theorem. (Scott–Suppes Theorem; Scott–Suppes [1958]) A poset is a

semiorder if and only if neither 2 + 2 nor 3 + 1 is a subposet.
Proof: (Bogart–West [1999]) A semiorder is an interval order, so 2 + 2 is forbid-
den. Given a semiorder representation º with threshold 1, if 3 + 1 is a subposet
with x < y < and incomparable element w, then  (w) ≥  (y) contradicts xw
and  (w) ≤  (y) contradicts w. Hence the condition is necessary.
For sufficiency, suppose that P has no 2 + 2 or 3 + 1. By Theorem 12.4.5, P is
an interval order, representable as in that proof. We convert this representation
to one whose intervals have the same length. Since P has no 3 + 1, there is no
pair x , y ∈ P such that (1) Iy ⊂ Ix and (2) Ix intersects intervals to the left and
right of Iy that do not intersect Iy . This enables us to alter the representation so
that no interval properly contains another. If Ix = [a , b] and Iy = [c , d] with a <
c ≤ d < b, then we know that [a , c] or [d , b] contains no endpoint of an interval
not intersecting Iy . Hence we can extend Iy past the end of Ix on one end.
Doing this until no more pairs of intervals are related by inclusion yields a
proper interval representation. From this we obtain a representation using inter-
vals of length 1: a semiorder representation. When no interval properly contains
another, the left ends appear in the same order as the right ends. We process the
representation from left to right, adjusting all intervals to have length 1.
Of the remaining unadjusted intervals, let Ix be one with leftmost left end-
point, with Ix = [a , b]. If some interval has right end in [a , b), then its left end is
before a, and the interval already has length 1. If this occurs, then let  be the
largest such right end; otherwise, let  = a. In either case,  ∈ [a , a + 1).
Now, adjust the portion of the representation in [a , ∞) by shrinking or ex-
panding [ , b] to [ , a + 1] and translating [b , ∞) to [a + 1 , ∞). The order of
endpoints does not change, intervals that begin before a still have length 1, and
Ix also now has length 1. Iterating produces the desired representation.

Iw Ix Iy I
↑ before
a  a+1 b
↓ after
Iw Ix Iy I

More generally, Fishburn [1984b] proved that the interval orders having in-
terval representations using intervals with lengths in [n − 2] are those not con-
taining the poset n + 1 (see also Fishburn [1985, Theorem 8.3]).
Counting interval orders is hard (Hanlon [1982]); but there are precisely
1
( 2n
)
n+1 n semiorders on n elements, bijectively, using the Catalan numbers
(Dean–
Keller [1968]; Exercise 9). For semiorders representable using open intervals of
length  with integer endpoints, Mitas [1994] obtained a forbidden subposet char-
acterization; the number of forbidden subposets is the ( + 1)th Catalan number!
588 Chapter 12: Partially Ordered Sets

L ATTICES

Some special posets admit algebraic operations that generalize the notions of
intersection/union for subsets and gcd/lcm for divisibility.

12.4.7. Definition. If x ≤ y, then x is a lower bound for y and y is an upper

bound for x. A poset L is a lattice if for all x , y ∈ L there is a unique max-
imal common lower bound (the meet x ∧ y) and a unique minimal common
upper bound (the join x ∨ y). Recall that a poset is bounded if it has one
minimal element 0̂ and one maximal element 1̂.

“Unique maximal common lower bound” means that x ∧ y is an upper bound

for every common lower bound of x and y; that is, the subposet of common lower
bounds has a unique maximal element. The definition immediately implies that
the meet and join operations are commutative and associative.

12.4.8. Example. Subsets and divisors. Consider sets x , y ∈ 2 n . A set is con-

tained in both x and y if and only if it is contained in x ∩ y. Hence x ∩ y is the
unique maximal common lower bound. Similarly, x ∪ y is the unique minimal
common upper bound. Hence meets and joins exist, and 2 n is a lattice.
Let D(n) be the divisibility poset on divisors of n. In D(n), meet and join are
least common multiple and greatest common divisor, respectively.

12.4.9. Proposition. A product of lattices is a lattice.

Proof: Let P = L1 ×· · ·× Ln . A lower bound for an element must be a lower bound
coordinate by coordinate, and similarly for upper bounds. Thus meets and joins
arise componentwise, with (x ∧ y)i = x i ∧ Li yi and (x ∨ y)i = x i ∨ Li yi .
The divisibility poset on divisors of an integer and the containment poset on
multisets with bounded multiplicities can both be expressed as products of chains.
We can write an element a in a product of n chains as (a1 , . . . , a¾), where ai is the
height of the ith coordinate of a on its chain. The meet and join operations then
have simple formulas: (a ∧ b)i = min{ai , bi} and (a ∨ b)i = max{ai , bi}.

12.4.10. Example. The partition lattice Π n on partitions of [n]. A refinement

of a partition replaces each block with a partition of that block. Put < in Π n
when  is a refinement of (Π3 and Π4 appear below).
Rank is given by r( ) = n − b( ), where b( ) is the number of blocks in  . The
meet  ∧ is the common refinement of  and with the fewest blocks, and  ∨
is the partition with the most blocks that does not “split ” any block of  or .
1234
•
123
•
123:4 12:34 134:2 24:13 124:3 14:23 234:1
• • • • • • •
1:23• 2:13• 3:12•
• • • • • •
13:2:4 12:3:4 34:1:2 14:2:3 23:1:4 24:1:3
•
1:2:3
•
1:2:3:4
Section 12.4: Special Families of Posets 589

12.4.11. Definition. A subposet P of a lattice L is a sublattice of L if x ∧ L y

and x ∨ L y are in P for all x , y ∈ P. Equivalently, a lattice L is a sublattice
of a lattice M if there is an embedding º : L → M such that º (x ∧ L y) =
º (x) ∧ M º (y) and similarly for join.
12.4.12. Example. J(P): The poset of down-sets. Let J(P) denote the contain-
ment poset on the family of down-sets in P. Each down-set is a subposet of P , so
J(P) ⊆ 2| P | . Since J(P) is a containment poset, common lower bounds of elements
x and y in J(P) are contained in x ∩ y, and common upper bounds contain x ∪ y.
Since the intersection and union of down-sets in P are down-sets in P , we have
x ∧ y = x ∩ y and x ∨ y = x ∪ y. Hence meet and join in J(P) agree with meet
and join when viewed in all of 2 P . Thus J(P) is a sublattice of 2|P | . The poset is
bounded, with 1̂ = P and 0̂ = ∅. Indeed, J(P) is graded, with r(I) = | I | .
Antichains in P correspond to down-sets in P. Consider antichains A and B
generating down-sets D[A] and D[B]. We have D[A] ⊆ D[B] if and only if, for
every x ∈ A, there exists y ∈ B such that x ≤ y. This condition just rephrases
the containment condition on down-sets, so J(P) and the resulting lattice on an-
tichains are isomorphic. The maximum antichains induce a sublattice (Exercise
26), which was used in the original proof of the Greene–K leitman Theorem.

12.4.13. Example. L(m , n), again. The poset L(m , n) of Example 12.4.13 arises
as a lattice of down-sets: L(m , n) =∼ J(m × n) (Exercise 14). The isomorphism
maps a ∈ L(m , n) to the down-set of m × n generated by {(m + 1 − i , ai): ai > 0}.
The down-set in 4 × 5 corresponding to (0 , 1 , 5 , 5) ∈ L(4 , 5) is shown below.
Another proof that L(m , n) is a lattice is by applying Definition 12.4.11 to
the lattice (n + 1)m that contains it (Exercise 15).
•
• •
• • •
• • • •
• • • •
• • •
• •
•

We develop several properties that hold for all lattices.

12.4.14. Lemma. For elements x , y , of a lattice L,

(a) x ∧ y ≤ x ≤ x ∨ .
(b) If x ≤ , then x ∧ y ≤ ∧ y and x ∨ y ≤ ∨ y.
(c) (4-point Lemma) If , w ≤ x , y, then ∨ w ≤ x ∧ y.
(d) (x ∧ y) ∨ (x ∧ ) ≤ x ∧ (y ∨ ).
(e) If ≤ x, then x ∧ (y ∨ ) ≥ (x ∧ y) ∨ .
Proof: (a): This holds by definition.
(b): Since x ∧ y is a common lower bound for y and , it lies below the unique
greatest lower bound (the second conclusion is symmetric).
(c): Since both and w are lower bounds for each of x and y, they are lower
bounds for x ∧ y. Hence x ∧ y is a common upper bound for and w, which yields
x ∧ y ≥ ∨ w (see figure below).
590 Chapter 12: Partially Ordered Sets

(d): By statement (a), x is an upper bound for x ∧ y and x ∧ , and y ∨ is an

upper bound for both y and and hence for x ∧ y and x ∧ . Hence (d) follows from
(c) by using x and y ∨ as {x , w} in (c) and x ∧ y and x ∧ as {x , y} in (c).
(e): When ≤ x, we have x ∧ = , so (e) follows immediately from (d).
x y
• • • •
4-point • x∧y •x
•x •y • •y
Lemma • ∨w •
• • • •
w Π3 = M5 N5

12.4.15. Example. Examples of strict inequality in Lemma 12.4.14(d) and (e) oc-
cur in the lattices M5 and N5 shown on the right above; we will explain their
names later. Note that M5 is the partition lattice Π3 .

12.4.16. Definition. For elements x , y ∈ P with x ≤ y, the interval [x , y] is

{ ∈ P: x ≤ ≤ y}. A poset is locally finite if every interval is finite.

12.4.17. Example. Every interval [x , y] in 2 n with r(x) = ¾ and r(y) = l is iso-

morphic to 2 l−¾ . Also, every interval in a chain-product is a chain-product. Meets
and joins of elements in an interval also lie in the interval.
All intervals in lattices are sublattices, but some sublattices may not be in-
tervals. For example, three of the nine sublattices of 23 isomorphic to 2 2 are not
intervals in 23 . Also, a subposet of P that is a lattice need not be a sublattice of
P. For example, N5 is a subposet of 2[3] but is not a sublattice of 2[3] .

The next lemma saves some work in proving that posets are lattices. Given
an upper bound (or a lower bound), one need not construct both meets and joins.

12.4.18. Lemma. If P is locally finite, has an upper bound (the element 1̂), and
has a well-defined meet operation, then P is a lattice.
Proof: It suffices to prove that joins exist. Consider x , y ∈ P. The upper bound 1
is a common upper bound for x and y. Since P is locally finite, the interval from
x ∧ y to 1̂ is finite. Thus we can consider the minimal elements among the set of
common upper bounds of x and y; they lie in the interval [x ∧ y , 1̂].
Let u and v be minimal common upper bounds for x and y. Since x and y are
common lower bounds for u and v, we have x < u ∧ v and y < u ∧ v. Thus u ∧ v
is a common upper bound for x and y. Since u and v are minimal such elements,
u ∧ v ∈ {u , v}. Thus u and v are comparable. Since they are minimal elements in
a subposet of P , they must therefore be equal. Hence there is a unique minimal
common upper bound for x and y.
•1̂
u• •v
•
u∧v
x• •y
•
x∧y
Section 12.4: Special Families of Posets 591

DISTRIBUTIVE L ATTICES

Lemma 12.4.14(d) states that (x ∧ y) ∨ (x ∧ ) ≤ x ∧ (y ∨ ) for all x , y , in any

lattice. In the subset lattice, equality holds. We study the class of lattices where
equality always holds, seeking characterizations of the class and properties of the
subset lattice that extend to such lattices.

12.4.19. Definition. A lattice L is distributive if meet distributes over join in

L; that is, x ∧ (y ∨ ) = (x ∧ y) ∨ (x ∧ ) for all x , y , ∈ L.

Exercise 30 requests a direct proof that a lattice L is distributive if and only if

its dual L∗ is distributive, by showing that the defining condition for distributiv-
ity is equivalent to the property that x ∨ (y ∧ ) = (x ∨ y) ∧ (x ∨ ) for all x , y , ∈ L.
We will also see other ways to prove the equivalence.

12.4.20. Example. Subset and divisor lattices. Distributivity for the subset lat-
tice can be seen by marking x ∧ (y ∨ ) and (x ∧ y) ∨ (x ∧ ) in a Venn diagram.

For the divisor lattice, we can argue directly about divisors using gcd and lcm to
show distributivity. Alternatively, since every chain is a distributive lattice (min
distributes over max for integers), the conclusion that M e is distributive follows
immediately from the next lemma.

12.4.21. Lemma. A product of lattices is distributive if and only if each factor is

distributive.
Proof: The order relation in a product is defined componentwise, so (x ∨ y)i =
x i ∨ yi and (x ∧ y)i = x i ∧ yi . Thus distributivity holds for the full lattice if and
only if it holds in each factor.

Recall that P is a sublattice of a lattice L if and only if P is closed under the

taking of meets and joins in L.

12.4.22. Lemma. Every sublattice of a distributive lattice is distributive.

Proof: The sublattice inherits the distributivity condition from the full lattice,
since meets and joins are computed in the full lattice.

Lemma 12.4.22 implies that there is a forbidden sublattice characterization

of distributive lattices. The two 5-element lattices M5 and N5 in Example 12.4.15
are not distributive; they violate the condition with x , y , as illustrated. For-
bidding M5 and N5 as sublattices is thus necessary for distributivity. It is also
sufficient, but we will not prove this.
592 Chapter 12: Partially Ordered Sets

12.4.23. Theorem. A lattice is distributive if and only if it does not have M5 or

N5 of Example 12.4.15 as a sublattice.

Theorem 12.4.23 is quite strong in proving properties of distributive lattices.

For example, since M5 and N5 are self-dual, Theorem 12.4.23 implies immediately
that a lattice L is distributive if and only if L∗ is distributive. Thus interchang-
ing meet and join in Definition 12.4.19 yields an equivalent condition.
Lemma 12.4.22 and Example 12.4.20 imply that every sublattice of a subset
lattice (or of a chain product) is distributive. For example, since J(P) is a sub-
lattice of 2|P | , always J(P) is distributive, for any poset P. Our main objective in
this discussion of distributive lattices will be a proof that every finite distributive
lattice L can be expressed as J(P) for an appropriate poset P.

12.4.24. Definition. An element p of a lattice L is join-irreducible if it is non-

minimal and is not the join of two other elements; equivalently, p = x ∨ y
implies p ∈ {x , y}. Similarly, p is meet-irreducible if it is not maximal and
p = x ∧ y implies p ∈ {x , y}. In a lattice L, we write P(L) and Q(L) for the
subposets formed by the join-irreducible elements and the meet-irreducible
elements, respectively.

12.4.25. Example. In a finite lattice, the join-irreducible elements are those cov-
ering exactly one element. In the subset lattice these are the 1-sets. In the divisor
lattice these are the powers of primes.
In the lattice L on the left below, the elements labeled by single letters are
the join-irreducible elements. The resulting subposet P(L) is on the right. The
label for each x ∈ L is the set of join-irreducible elements whose join is x. Unique-
ness of such expressions is our next objective. The minimal element of L is not
considered join-irreducible, and the maximal element is not meet-irreducible.

be•

bcd • •e
bc• bd• •cd e•

b• c• •d b• c• •d
•a •a
L P(L)
•∅

We will see in Theorem 12.4.38 that distributive lattices are the appropriate
general setting for many results on 2 n and M e . It helps to keep the divisor lattice
in mind when discussing them. The proof that L = ∼ J(P(L)) when L is distribu-
tive generalizes the proof that integers have unique prime factorizations. In this
discussion, we abbreviate “ join-irreducible” to irreducible.
By induction on the size of the set, in a lattice every finite set has a unique
least common upper bound, so we write joins of finite sets without parentheses.
Induction also shows that meet distributes over a join of a finite set of elements.
Section 12.4: Special Families of Posets 593

12.4.26. Lemma. If p ≤ a1 ∨ · · · ∨ a¾ for some irreducible element p in a distribu-

tive lattice L, then p ≤ ai for some i.
Proof: The definition of meet yields x ≤ y if and only if x ∧ y = x. With x = p and
y = a1 ∨ · · · ∨ a¾ , we have
p = p ∧ (a1 ∨ · · · ∨ a¾) = (p ∧ a1) ∨ · · · ∨ (p ∧ a¾).
This expresses p as the join of several elements. Since p is irreducible, it must
equal one of them. Thus p = p ∧ ai for some i, which in turn yields p ≤ ai .

12.4.27. Definition. An irredundant representation of a is a minimal expres-

sion of a as a join of a set of irreducible elements; that is, a = p1 ∨ · · · ∨ p¾
and no proper subset of {p1 , . . . , p¾ } has join a.

The irreducible elements in an irredundant representation are incompara-

ble, since if p > q then q is redundant in any join of a set containing p. To prove
existence and uniqueness of irredundant representations, we need a unique min-
imal element and a finiteness condition for inductive arguments. For clarity and
simplicity, we restrict to finite lattices our discussion of this lemma and its ap-
plication to characterize distributivity. The results can be extended to infinite
lattices satisfying appropriate local finiteness conditions (such as the divisibility

order on ), but we will not discuss this.

12.4.28. Lemma. In a finite distributive lattice L having a lower bound 0̂, every
element has a unique irredundant representation.
Proof: We first prove existence. Note that the identity element for join is 0̂; fol-
lowing our usual convention, 0̂ is thus the join of the empty set of irreducible
elements. Now, if some element has no irredundant representation, then there
is a minimal such element x, and x is the join of two lower elements. By mini-
mality, those elements have irredundant representations, and x is the join of the
union of those two sets. Deleting the non-maximal elements in the union yields
an irredundant representation.
Now let p1 ∨ · · · ∨ p¾ and q1 ∨ · · · ∨ ql be irredundant representations of a.
Since pi ≤ a, Lemma 12.4.26 implies that each pr satisfies pr ≤ qs for some s.
Similarly each qs satisfies qs ≤ pt for some t. This yields a relation pr ≤ pt , which
forces pr = qs = pt since p1 , . . . , p¾ form an antichain. Hence each pi belongs to
{qj }, and similarly each qj belongs to {pi}, and the sets are the same.

12.4.29. Definition. The unique irredundant representation of an element x in a

finite distributive lattice is the factorization of x. For a lattice L, the ideal
map
: L → J(P(L)) assigns to each x ∈ L the down-set of join-irreducibles
defined by
(x) = {p ∈ P(L): p ≤ x}.

We use the word “factorization” because for a divisor lattice, the factoriza-
tion of an element is the set of prime powers in its numerical prime factorization.

12.4.30. Lemma. For x in a finite distributive lattice L, the factorization of x is

the antichain of maximal elements in
(x), where
is the ideal map on L.
594 Chapter 12: Partially Ordered Sets

Proof: Let A be the antichain of elements in the factorization of x. Since x = ⋁ A

and A ⊆ P(L), we have A ⊆
(x). Hence ⋁ A ≤ ⋁ B, where B is the set of maximal
elements in
(x), since every element of A is bounded above by some element of
B. Now ⋁ B ≤ x, since B ⊆
(x) and x is an upper bound for
(x). We have proved
x = ⋁ A ≤ ⋁ B ≤ x, and hence ⋁ A = ⋁ B = x. By Lemma 12.4.28, A = B.

12.4.31. Theorem. (Birkhoff [1935]) A finite lattice L is distributive if and only

if L ∼
= J(P(L)) (and hence L is a sublattice of 2 P(L)).
Proof: Because every sublattice of a distributive lattice is distributive, the con-
dition is sufficient. For necessity, suppose that L is distributive. We prove that
the ideal map
: L → J(P(L)) is a lattice isomorphism.
Since the elements in the factorization of x form the antichain of maximal
elements in
(x), we have ⋁
(x) = x. Hence
is injective.
For surjectivity, let D be a down-set in P(L), and let x = ⋁ D. Each element
of D is bounded above by x, so D ⊆
(x). Since x = ⋁ D, we also have x = ⋁ A,
where A is the antichain of maximal elements in D. An expression of x as a join
of an antichain of irreducible elements is irredundant, so Lemma 12.4.28 and
Lemma 12.4.30 imply that D =
(x).
If x ≤ y, then
(x) ⊆
(y), and
(x) ⊆
(y) implies x = ⋁
(x) ≤ ⋁
(y) = y,
so
is a poset isomorphism. For a lattice isomorphism, it remains only to check
that
preserves meets and joins. For meets,

(x) ∧ J(P(L))
(y) =
(x) ∩
(y) = { ∈ P(L): ≤ x and ≤ y}
= { ∈ P(L): ≤ x ∧ L y} = (x ∧ L y)
The computation for joins is analogous.

12.4.32. Corollary. (1) Every distributive lattice L is graded, with r(L) = | P(L)| .
(2) Under  , the irreducible elements of L map to members of J(P(L)) gener-
ated by single elements.
Proof: (1) J(P(L)) is ranked by cardinality of the down-sets in P(L) (for example,
compare L and P(L) in the drawing of Example 12.4.25).
(2) The factorization of an irreducible element p ∈ P(L) is {p}.

∼ Q(L);
Writing this discussion using meets instead of joins would prove P(L) =
Exercise 33 obtains this statement from Theorem 12.4.31.

12.4.33.* Remark. Like our characterization of semiorders by first characteriz-

ing interval orders, Theorem 12.4.23 characterizing distributive lattices by for-
bidden sublattices is proved by first characterizing a larger class.
A lattice is modular if equality always holds in the inequality of Lemma
12.4.14(e). This fails for N5 with x , y , as in Example 12.4.15. On the other
hand, it holds for M5 . This is the source of the notation: M5 is modular, and N5
is non-modular. Using various properties of modular lattices, one can show that
a lattice is modular if and only if it does not have N5 as a sublattice.
Every distributive lattice is modular: when ≤ x in a distributive lattice,
x ∧(y ∨ ) = (x ∧ y)∨(x ∧ ) = (x ∧ y)∨ . Theorem 12.4.23 is proved by showing that
a modular lattice is distributive if and only if it does not have M5 as a sublattice.
Section 12.4: Special Families of Posets 595

Modular lattices lie in a still larger class. A lattice is semimodular if x ∨ y

covers y whenever x covers x ∧ y. Semimodular lattices are characterized by hav-
ing submodular rank functions, meaning r(x ∧ y) + r(x ∨ y) ≤ r(x) + r(y). This
links them closely to matroids (Section 11.3); in any matroid, the inclusion order
on the family of closed sets is a semimodular lattice.

CORREL ATION AL INEQUALITIES

A natural probability space arises from linear extensions of posets. We view a

poset Q as partial information about an underlying linear order on the elements.
This indeed is the setting when we have partially sorted numbers via pairwise
comparisons. The set S of possible outcomes is the set of linear extensions of Q.
We assume that each linear extension of Q is equally likely to be the true
ordering. An event is a subset of the linear extensions. When we specify an event
A by a condition (like “ x < y ”), we mean the set of linear extensions in which the
condition occurs. The probability (A) of an event A is then | A| / | S|.
Events A and B are independent if (A ∩ B) = (A)(B). Events A and B
are positively correlated if (AB) ≥ (A)(B). Our goal is the “XYZ Inequal-
ity” (Theorem 12.4.41): for elements x , y , in any poset Q, the events “ x < y ”
and “ x < ” are positively correlated.
Sampling randomly from a poset raises similar questions. For F , G ⊆ P , we
ask whether “membership in F ” and “membership in G ” are positively correlated.

12.4.34. Theorem. (Kleitman’s Inequality; K leitman [1966]) If F and G are

down-sets in 2 n , then | F|∩P |G| > || FP || ||G |
P | . Equivalently, membership in a down-set
and an up-set are negatively correlated.

K leitman was motivated by an extremal problem: how large can | F ∩ G | be

for a down-set F and an up-set G of specified sizes? Later, Anderson [1976] proved
that K leitman’s Inequality also holds in the multiset lattice M e . The proof we
present involves Chebyshev ’s Inequality (a correlational inequality for real num-
bers) and illustrates an inductive technique for chain-products.

12.4.35. Lemma. (Chebyshev’s Inequality) If x1 , . . . , x m and y1 , . . . , ym are

both nonincreasing or both nondecreasing sequences, then
∑ x i yi ≥ ∑ x i ∑ yi .
m m m
Furthermore, if Þ is a nonnegative weight function on [m], then
∑ x i yi Þ(i) ≥ ∑ x i Þ(i) ∑ yi Þ(i) .
∑ Þ(i) ∑ Þ(i) ∑ Þ(i)
Proof: Form the double sum ∑ i , j (x i − x j )(yi − yj ). Since both sequences are mono-
tone, both factors in any term have the same sign (or 0), so each term is nonneg-
ative. Multiplying out and moving the terms expressed with minus signs to the
other side yields 2m ∑ x i yi ≥ 2 ∑ x i ∑ yj . For the weighted version, start with
∑i , j (x i − x j)(yi − yj)Þ(i)Þ(j) and proceed in the same way.
596 Chapter 12: Partially Ordered Sets

We use this to prove K leitman’s Inequality for chain-products.

12.4.36. Theorem. (Anderson [1976]) If F and G are down-sets in L, a product

of n chains, then

| L| | F ∩ G | ≥ | F | | G | .

Proof: (Daykin–K leitman–West [1979]) Let e1 , . . . , e n be the chain sizes. We use

induction on n. If n = 1, then | F ∩ G | = min{| F | , | G |} and | L| ≥ max{| F | , | G |}.
For n > 1, partition these sets using the last coordinate. Each x ∈ L is a
vector (x1 , . . . , x n) with 0 ≤ x j < e j for j ∈ [n]. For X ⊆ L and 0 ≤ i ≤ e n − 1, let
X i = {x ∈ X : x n = i}. Note that Li is isomorphic to the product L of the first n − 1
chains. If X is a down-set in L, then X i is a down-set in Li , and | X 0 | ≥ · · · ≥ | X e n−1 |.
Since F , G, and F ∩ G are down-sets, applying this chain of inequalities
yields nonincreasing lists for the sizes of their “slices” | F i | , | G i | , and |(F ∩ G)i |.
Also, (F ∩ G)i = F i ∩ G i . Now we apply Chebyshev ’s Inequality and then the in-
duction hypothesis (for the subsets of L obtained by deleting the last coordinate
from the elements of F i , G i , and F i ∩ G i). The computation is
| F | | G | = ∑ | F i | ∑ | G i | ≤ e n ∑ | F i | | G i | ≤ e n ∑ | Li | | F i ∩ G i |
= e n ∏ e i ∑ |(F ∩ G)i| = | L| | F ∩ G | .
i<n

X e n −1
X1 L e n −1
X0
L1
L0

Daykin [1977] proved a more general inequality for more general sets in dis-
tributive lattices and thereby characterized distributive lattices (Exercise 43).
As a common extension of Theorem 12.4.36 and Chebyshev ’s Inequality, we
will prove the FKG Inequality for monotone functions on lattices. Discovered
jointly by Fortuin, K asteleyn, and Ginibre, this is the central result about cor-
relational inequalities.
Chebyshev ’s Inequality can be viewed as a statement about random vari-
ables. The weight function Þ(i) gives the probability that the outcome is i. The
sequences x and y are monotone functions º (i) and ½(i), and we compare the ex-
pectation of their product and the product of their expectations. To extend this
to distributive lattices, we need a technical condition on the weight function Þ.

12.4.37. Definition. A function º : P → is order-preserving if x ≤ y implies

º (x) ≤ º (y), order-reversing if x ≤ y implies º (x) ≥ º (y). A real-valued
function Þ on a lattice is log-supermodular if Þ(x ∧ y) Þ(x ∨ y) ≥ Þ(x) Þ(y).
Section 12.4: Special Families of Posets 597

All weight functions on chains are log-supermodular, since in a chain always

{x ∧ y , x ∨ y} = {x , y}. Chebyshev ’s Inequality is the special case of the FKG
Inequality where the lattice is a chain.

12.4.38. Theorem. (FKG Inequality; Fortuin–K asteleyn–Ginibre [1971]) In a

distributive lattice L, functions º and that are both order-preserving or
both order-reversing are positively correlated with respect to each nonnega-
tive log-supermodular weight function  , meaning that

∑  (x) (x) (x) ∑ (x) ≥ ∑  (x) (x) ∑ (x) (x).

The FKG Inequality yields K leitman’s Inequality for all distributive lattices
by setting  = 1 and letting  and be the (order-reversing) characteristic func-
tions on the down-sets F and G, where the characteristic function of a set A
is the function  A having value 1 on A and value 0 outside A. Nevertheless,
generalizations of K leitman’s Inequality continued to appear long after the FKG
Inequality, because the FKG Inequality appeared in the literature of statistical
mechanics and discrete mathematicians were unaware of it for years.
Later, Ahlswede and Daykin found a generalization of the FKG Inequality
having an easier inductive proof. We give this proof, following the presentation of
Graham [1982]. The theorem is known both as the Ahlswede–Daykin Inequal-
ity and as the Four Function Inequality. A result intermediate between the
FKG Inequality and the Four Function Inequality appeared in Holley [1974].

12.4.39. Definition. For a set X in a lattice L and a function  : L → , define

 (X) = ∑ x∈ X  (x). For X , Y ⊆ L, define X ∧ Y = {x ∧ y: x ∈ X , y ∈ Y } and
X ∨ Y = {x ∨ y: x ∈ X , y ∈ Y}.

12.4.40. Theorem. (Ahlswede–Daykin [1978]) If  ,  ,  ,  are four nonnegative

functions on a distributive lattice L such that
(x)(y) ≤ (x ∧ y)(x ∨ y) for all x , y ∈ L,
then
(X)(Y) ≤ (X ∧ Y)(X ∨ Y) for all X , Y ⊆ L.

Proof: We first reduce to the case of 2 n , using that distributive lattices are sub-
lattices of such posets (Theorem 12.4.31). If the claim holds for 2 n , then it holds
for a sublattice L of 2 n as follows. Given  ,  ,  ,  defined on L, extend them to
2 n by giving them value 0 on 2 n − L. Whenever (x)(y)
= 0, we have x , y ∈ L.
Since L is a sublattice, L also contains x ∧ y and x ∨ y, and the hypothesis for
2 n follows from its truth for L. Hence the conclusion holds for any X , Y ⊆ 2 n ,
including when X , Y ⊆ L.
The proof for 2 n uses induction on n. For n = 1, we check several cases. The el-
ements of L are 0 and 1 (representing ∅ and [1]). The hypothesis gives (x)(y) ≤
(min{x , y})(max{x , y}) for the four possibilities of x , y ∈ {0 , 1}. The conclusion
is easy when | X | = 1 or | Y | = 1. The case X = Y = {0 , 1} requires a numerical
optimization (Exercise 45).
598 Chapter 12: Partially Ordered Sets

For n > 1, consider fixed X , Y ⊆ 2 n . Let L = 2 n−1 . In terms of X and Y , we

define , ,  ,  on L so that the desired inequality (X) (Y) ≤ (X ∧ Y)(X ∨
Y) will become  (L ) (L ) ≤  (L ∧ L )(L ∨ L ). Note that L ∧ L = L = L ∨ L
in the lattice L . For x ∈ L , we define
 (x) = (x) X (x) + (x ∪ {n}) X (x ∪ {n})
(x) = (x) (x) + (x ∪ {n}) (x ∪ {n})
Y Y
 (x) = (x) X ∧ Y (x) + (x ∪ {n}) X ∧ Y (x ∪ {n})
 (x) = (x) X ∨ Y (x) + (x ∪ {n}) X ∨ Y (x ∪ {n}).
For t ∈ { , ,  , }, this definition accumulates in t (x) the contributions of x and
x ∪ {n} to the “relevant set ” Z, which is X , Y , X ∧ Y , or X ∨ Y when t is  , ,  ,
or  , respectively. Including all contributions, t (L ) = t(Z), and thus the desired
inequality becomes  (L ) (L ) ≤  (L )(L ).
To obtain this conclusion from the induction hypothesis, we show that the
functions  , ,  ,  satisfy  (x) (y) ≤  (x ∧ y) (x ∨ y) for all x , y ∈ 2 n−1 . For
fixed x and y, each quantity in this inequality is computed from two elements of
L having the form  and  ∪ {n}. Since each of {  ,  ∪ {n}} might or might not
belong to the relevant set Z, we could complete the proof by checking 16 cases.
We can reduce the verification to four cases by using the induction hypothesis
for n = 1. We split the contributions to t by defining new functions on 21 .
 = x:  (∅) = () X ()  ([1]) = ( ∪ {n}) X ( ∪ {n})
 = y: (∅) = () Y () ([1]) = ( ∪ {n}) ( ∪ {n})
Y
 = x ∧ y:  (∅) = () X ∧ Y ()  ([1]) = ( ∪ {n}) X ∧ Y ( ∪ {n})

 = x ∨ y:  (∅) = () X ∨ Y ()  ([1]) = ( ∪ {n}) X ∨ Y ( ∪ {n}).
For t ∈ { , ,  , }, we have defined t on {∅ , [1]} so that t (21) = t (), where 
is the “relevant element ” of 2 n−1 as listed above.
If  , ,  ,  satisfy  (u) (v) ≤  (u ∧ v) (u ∨ v) for all u , v ∈ 21 , then
the induction hypothesis for the case n = 1 yields  (21) (21) ≤  (21) (21),
which is the needed inequality  (x) (y) ≤  (x ∧ y) (x ∨ y) for elements of 2 n−1 .
Using the definition of each t , the inequalities for u , v ∈ 21 are now statements
about elements of L, and we can apply the original hypothesis about { , ,  , }.
Since  , ,  ,  are nonnegative, in checking the hypothesis on 21 we may
assume that the left side of the needed inequality is positive, which requires that
the arguments on the left belong to X and Y , respectively. The corresponding
arguments to  and  then belong to X ∧ Y and X ∨ Y , respectively, so all the
values of the characteristic functions in the definitions of t may be assumed to
be 1. For x , y ∈ L , evaluating t at the relevant elements yields the four needed
inequalities below.
u v
∅ ∅ (x) (y) ≤ (x ∧ y)(x ∨ y)
∅ [1] (x) (y ∪ {n}) ≤ (x ∧ y)((x ∨ y) ∪ {n})
[1] ∅ (x ∪ {n}) (y) ≤ (x ∧ y)((x ∨ y) ∪ {n})
[1] [1]  ∪
(x {n}) (y ∪ {n}) ≤ ((x ∧ y) ∪ {n})((x ∨ y) ∪ {n})
In each nontrivial case,  (u) (v) ≤  (u ∧ v) (u ∨ v) thus reduces to an
instance of the hypothesis given on 2 n .
Section 12.4: Special Families of Posets 599

The Ahlswede–Daykin Inequality now yields the FKG Inequality.

Proof of FKG Inequality (Theorem 12.4.38): Set = =  =  = . By negat-

ing the functions if necessary, we may assume that  and  are order-reversing.
Also, adding a constant to one of the functions does not affect the inequality, so
we may assume that  and  are nonnegative.
We first prove the FKG Inequality when  and  are the characteristic func-
tions of down-sets X and Y . In this case, ∑ x∈ L  (x) (x) = ∑ x∈ X (x) = (X), and
similarly ∑ x∈ L (x) (x) = (Y). The hypothesis that  is log-supermodular im-
plies the hypothesis for the Ahlswede–Daykin Inequality, and the conclusion of
the Ahlswede–Daykin Inequality for X , Y is the desired statement for  and  :
(X)(Y) ≤ (X ∧ Y)(X ∨ Y) ≤ (X ∩ Y)(L) = ∑  (x)(x)(x) ∑ (x).

To complete the proof, we need only show (1) every nonnegative monotone
nonincreasing function on L is a linear combination of at most | L| characteris-
tic functions on down-sets, and (2) the FKG Inequality is preserved by taking a
positive linear combination of order-reversing functions 1 and 2 .
(1) follows by induction on the number of nonzero values of  , with a trivial
basis. Let  be the least nonzero value of  , let S = {x ∈ L:  (x)
= 0}, and let  S
be the characteristic function of S. Now  −  S is a nonnegative nonincreasing
function on L with fewer nonzero values than  .
For (2), let  be a positive linear combination of 1 and 2 , which each satisfy
the FKG Inequality with  . Term-by-term linearity of real number arithmetic
yields the FKG Inequality for  and  .

The FKG Inequality was notably applied to prove a conjecture by Rival and
Sands about positive correlation in random linear extensions of a poset Q. Recall
that (A) is the fraction of the linear extensions in which event A occurs. For
example, when Q is an antichain, (x < y) = 12 for any x and y. If Q = 2 + 1, with
x < and y unrelated to both, then (x < y) = (y < ) = 23 and (x < ) = 1.

12.4.41. Theorem. (XYZ Inequality; Shepp [1982]) For any elements x , y , in

any poset Q, the events “ x < y ” and “ x < ” are positively correlated.
Proof: Roughly speaking, we seek a distributive lattice where the events of inter-
est correspond to down-sets and will be positively correlated. Let the elements of
Q be {q1 , . . . , qn}, with (q1 , q2 , q3) = (x , y , ). Henceforth, we use x and y instead
as elements of the eventual distributive lattice, as in Theorem 12.4.38.
Let N be a large positive integer, and write x ∈ [N]n as (x1 , . . . , x n). View x
as a map
: Q → [N] in which x i =
(qi). If the coordinates of x are distinct (
is
injective), then x yields an ordering of Q. There are ( Nn ) such vectors x for each
ordering. When
is order-preserving, the ordering is a linear extension.
Using [N]n rather than the extensions themselves makes it easier to de-
fine a distributive lattice. We choose N large to make the effect of non-injective
mappings negligible. We return to this detail later; for now we view the order-
preserving mappings in [N]n as corresponding to extensions of Q.
Application of the FKG Inequality needs a weight function; let  be the
characteristic function of the set of order-preserving mappings. So that  and
600 Chapter 12: Partially Ordered Sets

½ capture the desired events, let them be the characteristic functions of F and G,
where F = {x ∈ [N]n : x1 < x2 } and G = {x ∈ [N]n : x1 < x3}. Roughly,

∑ (x)(x)/ ∑ (x) ≈ (q1 < q2) ,

∑ (x)(x)/ ∑ (x) ≈ (q1 < q3) ,
∑ (x)(x)(x)/ ∑ (x) ≈ (q1 < q2 and q1 < q3).
To apply the FKG Inequality, we need a partial order on [N]n so that and
 are order-preserving,  is log-supermodular, and the poset is a distributive lat-
tice. For x , y ∈ [N]n , set x ≤ y when x1 ≥ y1 and x i − x1 ≤ yi − y1 for i > 1.
We have (x) = 1 when x1 ≤ x2 . If x ≤ y, then also y1 ≤ y2 − x2 + x1 ≤ y2 , so
(y) = 1. Thus is monotone increasing, and similarly for  .
To see that [N]n is a distributive lattice under this ordering, consider the
map  (x) = (− x1 , x2 − x1 , . . . , x n − x1). We have x ≤ y in our ordering if and only if
 (x) ≤  (y) in the usual ordering on a product of n chains of size 2N + 1, indexed
from − N to N . Meets and joins are preserved under  , which follows from
(x ∧ y)i = min(x i − x1 , yi − y1) + max(x1 , y1)
(x ∨ y)i = max(x i − x1 , yi − y1) + min(x1 , y1).
For example, if i > 1, then
[ (x) ∧  (y)]i = min(x i − x1 , yi − y1) = (x ∧ y)i − (x ∧ y)1 = [ (x ∧ y)]i .
Thus our poset is a sublattice of a chain-product and hence a distributive lattice.
To apply the FKG Inequality, we also need  to be log-supermodular. We de-
fined (x) = 1 when x is order-preserving (x i ≤ x j for qi ≤ qj ). We need only check
that x ∧ y and x ∨ y are order-preserving when x and y are. This is easy from the
formulas given above for x ∧ y and x ∨ y, and it is the reason for subtracting the
first coordinate from the others in defining the order relation on [N]n .
From the FKG Inequality, ∑ (x)(x) (x) ∑ (x) ≥ ∑ (x) (x) ∑ (x) (x) ,
where the sums run over x ∈ [N]n . From our definitions of ,  ,  , this is what
we want, except that some vectors are not injective. We show that the relative
contribution to each term from non-injective x tends to 0 as N → ∞.
Let A1 ⊂ [N]n be the set of injective n-tuples (distinct coordinate values), and
let A2 = [N]n − A1 . For i ∈ {1 , 2}, define
F i = ∑ x∈ Ai (x) (x), G i = ∑ x∈ Ai (x) (x),
Hi = ∑ x∈ Ai (x)(x) (x), Mi = ∑ x∈ Ai (x).

Note that M1 is ( Nn ) times the number of extensions of Q. There is at least one

extension, so M1 ≥ ( Nn ). On the other hand, M2 ≤ | A2 | = N n − N(n). With n fixed,
we have M1 ∼ N n/n! and M2 = O(N n−1), so M2/M1 → 0. Since F i and G i and Hi
are all bounded by Mi , also F 2/M1 and G 2/M1 and H2/M1 all tend to 0.
Since each linear extension is represented equally often in A1 , what we want
is H1 M1 ≥ F1 G 1 . The FKG Inequality yields
(H1 + H2)(M1 + M2) ≥ (F1 + F 2)(G 1 + G 2).
Dividing by M12 and taking the limit as N → ∞ yields H1
M1 ≥ F1 G 1
M1 M1 , which is
precisely the desired inequality.
Section 12.4: Special Families of Posets 601

The XYZ Inequality is quite special. Many similar-sounding statements are

not true. For example, given x , y , , w ∈ Q, one might expect that the events
x < y < w and x < < w are positively correlated, but this does not hold for all Q
(Exercise 47). Winkler [1986] characterized the very few pairs of subposets that
are always positively correlated. Brightwell–Trotter [2002] gives a combinatorial
proof of a stronger version of the XYZ Inequality due to Fishburn [1984a].

A PROBLEM IN RAMSEY THEORY (optional)

Here we apply the lattice of down-sets to a problem in ordered Ramsey theory.

No aspects of lattice theory are needed, just the concept of the poset of down-sets.
12.4.42. Definition. An ordered hypergraph is a hypergraph on a linearly
ordered vertex set. An ordered hypergraph G occurs as a subhypergraph
of an ordered hypergraph H if some order-preserving injection from V(G) to
V(H) also preserves edges. We then say that H contains (a copy of) G.
Let K nr denote the complete r-uniform hypergraph with n vertices. By Ram-
sey ’s Theorem, when n is large enough every -coloring of E(K nr) contains a
monochromatic copy of K pr . Because complete ordered hypergraphs with the same
number of vertices are isomorphic, the same statement holds in the ordered sense,
and hence Ramsey numbers for ordered hypergraphs are well-defined.
12.4.43. Definition. Let G 1 , . . . , G ¾ be r-uniform ordered hypergraphs. The
-color Ramsey number R(H1 , . . . , H¾) is the least n such that every -
coloring of an n-vertex complete r-uniform ordered hypergraph contains a
copy of Hi in color i for some i. The r-uniform monotone path Ptr is the t-
vertex ordered hypergraph whose edges are the sets of r consecutive vertices.
When discussing only ordered hypergraphs, we can use the same notation
and terminology as for classical Ramsey numbers. Indeed, when G 1 , . . . , G ¾ are
complete, the Ramsey numbers in the classical and ordered senses are the same.
However, even when r = 2 the Ramsey numbers for ordinary paths and mono-
tone paths differ greatly. For a 3-vertex path, the classical -color Ramsey number
is + 2 (with + 1 edges at a vertex, two have the same color), but for the 3-vertex
monotone (ordered)path the Ramsey number is 2¾ + 1. This generalizes as follows,
using the same idea as the Erdős–Szekeres Theorem (Theorem 10.1.13).
¾
12.4.44. Proposition. For ordered paths, R(Pt21 , . . . , Pt2¾ ) = 1 + ∏i=1 (t i − 1).
Proof: Fix a -coloring of E(K n). For x ∈ V(K n) and i ∈ [ ], let x i be the maximum
number of vertices in a monotone increasing path with color i ending at x. Note
that x i ≥ 1 for all x and i, and avoiding long paths in color i requires x i ≤ t i − 1.
If y comes before  and edge y  has color i, then i > yi , so no two -tuples can
¾
be the same. If there is no sufficiently long path, then only ∏i=1 (t i − 1) vectors
are available. Hence with larger n we have a monotone path in color i for some i.
For the lower bound, assign the th vertex to the th element on a linear ex-
tension of the product of chains with sizes t1 − 1 , . . . , t¾ − 1. When y is before  ,
we have yi < i for some i; use color i on edge y . Since coordinate i must increase
along any monotone path in color i, no monotone path in color i has t i vertices.
602 Chapter 12: Partially Ordered Sets

This proof is the first instance of the induction step in a characterization of

R(Ptr1 , . . . , Ptr¾ ) as the number of down-sets in an appropriate poset, leading to up-
per and lower bounds that are exponential towers of height r − 2 for the Ramsey
number. The problem of computing R(Ptr1 , . . . , Ptr¾ ) was introduced in Fox–Pach–
Sudakov–Suk [2012] with a geometric application, though it had actually been
considered by Duffus–Lefmann–Rödl [1995] in the language of shift graphs (Defi-
nition 12.3.36) in order to give a lower bound on classical Ramsey numbers. Recall
that J(Q) denotes the lattice of down-sets in a poset Q, ordered by inclusion.

•
• •
• • • •
• • • • •
• • • •
•
Q J(Q)

12.4.45. Theorem. (Moshkovitz–Shapira [2014]) Fix t1 , . . . , t¾ larger than r.

Let Q consist of disjoint chains with sizes t1 − r, . . . , t¾ − r. With Q1 = Q and
Qi = J(Qi−1) for i > 1, the monotone paths satisfy R(Ptr1 , . . . , Ptr¾ ) = | Qr | + 1.
Proof: (Milans–Stolee–West [2015]) For r = 1, we use the Pigeonhole Principle.
¾ ¾
In this case | Qr | = ∑i=1 (t i − 1). Any ¾-coloring of 1-sets chosen from 1 + ∑ i=1 (t i − 1)
vertices has t i vertices with color i for some i, forming a monotone copy of Pt1i in
color i. We proceed by induction on r.
Lower Bound. Let n = | Qr | ; we construct a ¾-edge-coloring of K nr that avoids
Ptri in color i for each i. First consider x , y ∈ Qj , where 2 ≤ j ≤ r. The meaning of
x
≥ y is that x does not contain y when they are viewed as down-sets in Qj −1 . In
this case, let (x , y) be a fixed element of the family y − x in Qj −1 . For x , y ,  ∈ Qj ,
if x
≥ y and y
≥  , then y (in Qj −1) contains (x , y) but not (y , ). Since y is a
down-set in Qj −1 , we obtain (x , y)
≥ (y , ) in Qj −1 .
A list x1 , . . . , x s of elements in a poset is descent-free if x i
≥ x i+1 for 1 ≤
i ≤ s − 1. We extend to descent-free lists in Qj by setting (x1 , . . . , x s) =
( (x1 , x2) , . . . , (x s−1 , x s)). Thus the image under of a descent-free s-list in
Qj is a descent-free (s − 1)-list in Qj −1 . Let 0 be the identity map. For a
descent-free s-list x1 , . . . , x s in Qj , where s , j > d > 1, define d(x1 , . . . , x s) =
( d−1 (x1 , . . . , x s)); now d(x1 , . . . , x s) is a descent-free (s − d)-list in Qj −d .
Let y1 , . . . , yn be a linear extension of Qr , so yi ≤ yj implies i ≤ j . Each
sublist of a linear extension is descent-free. For a sublist x1 , . . . , x r , note that
r −1 (x
1 , . . . , x r) is a single element in Q. Color {x1 , . . . , x r } (as an edge in K n)
r
r −1 (x
with the index of the chain in Q that contains 1 , . . . , x r).
Suppose that this coloring has a monotone copy of P sr in color i. Let x1 , . . . , x s
be its vertices, in increasing order. Since x1 , . . . , x s is a sublist of a linear ex-
tension, it is descent-free. Since each set of r consecutive vertices from the list
x1 , . . . , x s was given color i, r−1 (x1 , . . . , x s) is a descent-free (s − r + 1)-list in the
ith chain of Q. A descent-free list in a chain is strictly increasing (because equal-
ity is also considered a descent), so s − r + 1 ≤ t i − r. Thus s < t i . We conclude
that the coloring avoids Ptri in color i for each i, so R(Ptr1 , . . . , Ptr¾ ) > n.
Section 12.4: Special Families of Posets 603

Upper bound. Given a ¾-edge-coloring á of E(K nr) that avoids Ptri in color i for
each i, it suffices to define an injection from [n] to Qr . View each vertex subset
Y ⊆ [n] as an increasing list, with Y − = Y − max Y and Y + = Y − min Y .
For 1 ≤ j ≤ r < n, we construct ½ j : ([n]
j
) → Qr− j +1 such that
½ j (Y −)
≥ ½ j (Y +) in Qr− j +1 when Y ∈ ( j[n]
+1 ). (∗)
This suffices, since ½ 1 will then be the desired injection.
We first define ½ r . For X ∈ ([n] r
), let i = á(X), and let w1 , . . . , wt −r be the ith
i
chain in Q. Set ½ r (X) = wh , where h is the largest integer such that some copy of
P hr +r−1 in color i has last edge X . Note that h ≤ t i − r, since á has no copy of Ptr in i

color i. If á(Y −) = á(Y +) for some Y ∈ (r[n] +1 ), then ½ r (Y ) > ½ r (Y ), and otherwise
+ −
½ r(Y ) and ½ r(Y ) are incomparable in Q. In either case, ½ r(Y )
≥ ½ r(Y +).
+ − −

Now consider smaller j , with ½ j +1 : ( j[n] +1 ) → Qr − j already defined and satisfy-

ing (∗). For X ∈ ([n]
j ), let the precursors of X be the
(j + 1)-sets obtained from X by
adding an element smaller than all of X ; that is, {Z ∈ ( j[n] +
+1 ): Z = X}. (Note that
sets containing the vertex 1 have no precursors.) Define ½ j (X) to be the down-set
in Qr− j generated by the set of elements ½ j +1 (Z) such that Z is a precursor of X .
Since Qr− j +1 = J(Qr− j ), by definition ½ j (X) ∈ Qr− j +1 .
To check (∗) for ½ j , consider Y ∈ ( j[n] +1 ). Since Y is a precursor of Y , the
+

definition of ½ j yields ½ j +1(Y) ∈ ½ j (Y +). If ½ j +1 (Y) also lies in ½ j (Y −), then by the
definition of ½ j (Y −) (using X = Y −), the element ½ j +1(Y) in Qr− j lies below some
element ½ j +1 (Z) such that Z + = Y − . Letting W = Z ∪ Y , we have Z = W − and
Y = W + . Thus ½ j +1 (W −) ≥ ½ j +1 (W +), which contradicts (∗) for ½ j +1 .
We conclude ½ j +1 (Y) ∈ ½ j (Y +) − ½ j (Y −). Hence the down-set ½ j (Y +) in Qr− j
does not contain the down-set ½ j (Y −), which means ½ j (Y −)
≥ ½ j (Y +) in Qr− j +1 .

Z• • • • •
• • • • •Y
Theorem 12.4.45 yields inductive upper and lower bounds for R¾ (Ptr). For
R¾ (P43), note that Q2 = 2 ¾ , so computing | Q3 | for R¾ (P43) is just Dedekind ’s Prob-
lem (see Theorem 12.2.14). For an application, see Exercise 49.

12.4.46. Corollary. (Moshkovitz–Shapira [2014]) Let tow h(x) be x when h = 0

and 2tow h−1(x) when h ≥ 1. With m = t − r + 1, the monotone paths satisfy
√
towr−2(m¾−1/2 ¾) ≤ R¾(Ptr) ≤ towr−2(2m¾−1) .
Proof: We prove weaker bounds, still exponential towers of the same height.
We bound | Qr | of Theorem 12.4.45. Since elements of Qj are subsets of Qj −1 ,
we have | Qj | ≤ 2| Qj −1 | . With | Q2 | = m¾ , we obtain | Qr | ≤ towr−2(m¾).
For the lower bound, recall that | J(Q)| also counts the antichains in Q. The
subsets of an antichain A form 2| A| antichains. Since | Q1 | = ¾(t − r), the down-sets
in Q1 range in size from 0 to ¾(t − r). Thus Q2 has fewer than ¾ m ranks and has
an antichain of size at least m¾−1/¾ . Also, the elements of this antichain have the
same size in Q1 . (The chain-product Q2 is a symmetric chain order, so its middle√
rank is a largest antichain. Sharper analysis reduces the denominator to 2 ¾ .)
604 Chapter 12: Partially Ordered Sets

Given an antichain of size M in Qj , the subsets of size M/2 generate pair-

wise incomparable down-sets and hence an antichain in Qj +1 . The resulting lower
bound uses an analogue of exponential towers involving middle binomial√coeffi-
1 (x)
cients. Let b0(x) = x, and for h ≥ 1 let bh(x) = (⌊bbh−h−1 (x)/2⌋ n
). Since (⌊n/2⌋ ) ∼ 2 n/ n/2
(Example 2.3.10), inductively we have bh(x) ≥ tow h(x − O(lg x)).
Let Aj be a largest antichain in Qj given by down-sets in Qj −1 having the
same size; let aj = | A j |. Thus | Qj +1 | ≥ 2 aj and aj +1 ≥ ( aa/2j
j
). Thus aj ≥ bj −2(a2).
With a2 ≥ m¾−1/¾ , we obtain | Qr | ≥ towr−2(m¾−1/¾ − O(¾ lg m)).

EXERCISES 12.4

12.4.1. (−) Prove that a poset is a ranking if and only if no subposet is isomorphic to 2 + 1.
12.4.2. (−) Prove that a poset is a chain if and only if every subposet is a lattice.
12.4.3. (−) Draw the distributive lattice whose poset of join-irreducible elements is below.

e•

b• c• •d
º• •a

12.4.4. Let P be a poset such that P − x is an interval order. Can relations involving x be
added to P to obtain an interval order P such that P − x = P − x?
12.4.5. Alternative proof of Theorem 12.4.5. Let P be a poset without 2 + 2.
(a) Prove that the “upper holding” sets U(x) for x ∈ P are linearly ordered by inclu-
sion, and similarly for the sets D(y).
(b) For the 0 , 1-matrix of the order relation, prove that ordering the rows as x1 , . . . , x n
in decreasing order of |U(xi)| and the columns as y1 , . . . , yn in decreasing order of | D(yj )|
puts the 1s of the matrix in the positions of a Ferrers diagram.
(c) For x ∈ P , let u(x) count distinct sets of the form U(y) that are proper subsets of
U(x), and let d(x) count distinct sets of the form D(y) that are proper subsets of D(x). Let
h be the total number of distinct nonempty sets of the form U(x). Prove that assigning
each x ∈ P the interval [d(x) , h − u(x)] produces an interval representation of P. (Bogart ,
motivated by Rabinovitch [1977])
12.4.6. (♦) Let In be the poset of nontrivial intervals with integer endpoints in [n], ordered
by [a , b] < [c , d] if b < c. Prove that dim In > ¾ when n is sufficiently large. (Hint: Given
Ramsey’s Theorem, use n ≥ R¾(4; 3).) (Rabinovitch [1973])
12.4.7. (♦) Given the interval order In as in Exercise 12.4.6 and the “double-shift graph”
G n as in Definition 12.3.36, prove dim In ≥ Ò(G n). (Comment: Thus dim In ≥ lg lg n +
( 12 + o(1)) lg lg lg n, by Corollary 12.3.39. With some care, Füredi–P.Hajnal–Rödl–Trotter
[1992] constructed linear extensions to prove that the lower bound is tight.)
12.4.8. Representations of interval orders. Fix ¾ ∈ . 
(a) Construct an interval order such that every interval representation uses at least
¾ different lengths of intervals. (Fishburn [1983])
(b) Construct an interval order such that every representation has an interval of
length more than ¾ times the length of its shortest interval. (Fishburn–Graham [1985])
Exercises for Section 12.4 605

12.4.9. Establish a one-to-one correspondence between the isomorphism classes of n-

element semiorders and the ballot lists of length 2n (Definition 1.3.16). (Comment: Thus
there are n+1 1 (2n
n
) isomorphism classes of semiorders on n elements.) (Dean–Keller [1968])
12.4.10. (♦) Since a semiorder P has no 2 + 2, by Exercise 12.3.16 every subposet of P has
upper and lower extensions. Let Q be the subposet of P consisting of the elements with
even rank. Using upper and lower extensions of Q, prove dim P ≤ 3. (Rabinovitch [1978])

12.4.11. (♦) Although interval orders are more general than chains (or semiorders), prove
that the standard example Sn cannot be an intersection of fewer than n interval orders.
12.4.12. (♦) A general binary relation is modeled by a digraph D, with adjacency matrix
A(D). A biorder representation of D consists of real-valued functions º and ½ on V(D)
such that uv ∈ E(D) if and only if º (u) > ½(v). For a digraph D, prove that the following
five conditions are equivalent.
(A) A(D) has no 2-by-2 submatrix that is a permutation matrix.
(B) The successor sets of D are ordered by inclusion.
(C) The predecessor sets of D are ordered by inclusion.
(D) The rows and columns of A(D) can be permuted independently so that every entry
below or to the left of a 1 is a 1.
(E) D has a biorder representation.
(Comment: The equivalences are due to various authors, with a short proof in West [1998].
Such relations are called biorders, Ferrers relations, or Ferrers digraphs (Riguet
[1951], Wiener [1914]). If also º (x) ≤ ½(x) for all x, then D is an interval order.)

12.4.13. (♦) The completion of a graph G with V(G) = [n] is the partition of [n] whose
blocks are the vertex sets of components of G. Let Q(G) denote the subposet of Π n whose
elements are the completions of spanning subgraphs of G. The partition of rank 1 with
block {i , j} lies in Q(G) if and only if i j ∈ E(G), so these posets are distinct for distinct G.
(a) Show that Q(G) need not be a sublattice of Π n .
(b) Prove that Q(G) is a sublattice of Π n isomorphic to 2 n−1 when G is a tree.
(c) Prove that if P is a copy of 2 n−1 in Π n , then P = Q(G) for an n-vertex tree G.
(d) Conclude that Π n contains nn−2 copies of 2 n−1 .

12.4.14. Let J(P) denote the inclusion poset on the down-sets of P. Using the mapping
∼ J(m × n) =
suggested in Example 12.4.12, prove L(m , n) = ∼ L(n , m).

12.4.15. As a component-wise order on m-tuples from {0 , . . . , n}, the poset L(m , n) is a

subposet of (n + 1)m . Prove that the meet and join in (n + 1)m of elements of L(m , n) is also
in L(m , n), thereby proving by Definition 12.4.11 that L(m , n) is a sublattice of (n + 1)m .
12.4.16. (♦) Write subsets of [n] as strictly decreasing lists, n ≥ a1 > · · · > a¾ > 0. Let Mn
be the poset of subsets of [n], ordered by a ≤ b if and only if ai ≤ bi for all i, with trailing
0s added to permit comparison, so a subset is never less than a smaller subset.
(a) Describe the rank function and the cover relation in Mn .
(b) Describe Mn+1 in terms of Mn .
(c) Prove Mn+1 ∼
= J(L(2 , n − 1)).
321
32
•
31
•
•
21
• •3
•2
•1
•∅
606 Chapter 12: Partially Ordered Sets

12.4.17. (♦) Let Mn be the poset of compositions of the integer n + 1, ordered by a ≤

¾ ¾
b if ∑i=1 ai ≥ ∑i=1 bi for all ¾ (add trailing 0s as needed to test the order relation).
∼
Prove Mn = Mn , where Mn is the poset of Exercise 12.4.16. (The compositions of 4 are
{4 , 31 , 22 , 211 , 13 , 121 , 112 , 1111}.)

12.4.18. Let Mn be the poset of Exercise 12.4.16. Find an explicit symmetric chain de-
composition for n ≤ 5. (Comment: Lindström conjectured that Mn is always a symmetric
chain order. Stanley [1982] observed that Mn satisfies the strong Sperner property.)

12.4.19. (♦) Prove that if (x , y) is an unforced pair in a lattice L, then x is meet-irreducible

and y is join-irreducible. Conclude that dim L = dim R, where R is the subposet of meet-
irreducible and join-irreducible elements. (Kelly [1981])

12.4.20. (♦) The Young lattice. Let Y be the poset of all partitions of all integers, ordered
by a ≤ b if and only if ai ≤ bi for all i (trailing zeros appended as needed).
(a) Prove that Y is a lattice.
(b) Prove that every a ∈ Y is covered by one more element than it covers.
(c) Prove that Y is a distributive lattice.
(d) Describe the join-irreducible elements of Y.

12.4.21. (♦) Prove that the dominance order on the partitions of n (written in decreasing
order) is a lattice. Here Þ ≤ ë if ∑i=1 Þi ≤ ∑i=1 ë i for all j . (Hint: Apply Lemma 12.4.18.)
j j

12.4.22. (♦) Obtain a necessary and sufficient condition for ë to cover Þ in the dominance
order on all integer partitions.

12.4.23. For integers m and ¾ with 0 ≤ ¾ ≤ m − 1, let C(m , ¾) = {im + ¾ : i ∈ }; these

sets are the congruence classes. Let P be the poset of all congruence classes, ordered by
inclusion, with the empty set added as a minimal element. Prove that P is a lattice, and
describe the meet and join of C(m , ¾) and C(n , l).

12.4.24. The poset Ln(q) is the containment poset on the set of subspaces of an n-
dimensional vector space over a q-element field. Prove that Ln(q) is a graded lattice but in
general is not a distributive lattice.

12.4.25. Let Λ n denote the poset of partitions of the integer n, ordered by refinement.
That is, for partitions ë and Þ of n, we put ë ≤ Þ if the multiset of parts in ë is obtained
by replacing each integer part in Þ with a partition of that part.
(a) Prove that Λ n is a graded poset.
(b) Prove that Λ 7 is not a lattice. (Comment: This holds also for n > 7.)

12.4.26. (♦) Let A and B be maximum antichains in a poset P , and let A and B be the
down-sets generated by A and B. Prove that the maximal elements of A ∪ B form a max-
imum antichain, and similarly for the maximal elements of A ∩ B . Conclude that the
maximum-sized antichains in a poset form a sublattice of the lattice of antichains, ordered
by A ≤ B if and only if for each x ∈ A there exists y ∈ B such that x ≤ y. (Dilworth [1960])

12.4.27. Given the result of Exercise 12.4.26, design a polynomial-time algorithm to con-
struct the maximal element of the lattice of maximum-sized antichains. (Hint: Use the
relation of Dilworth’s Theorem to maximum matching.)

12.4.28. An automorphism of a poset permutes elements but preserves the order rela-
tion. Prove that the automorphism group of Π n is isomorphic to the symmetric group Ën .

12.4.29. Given a connected graph G, let C(G) be the poset of the connected induced sub-
graphs of G, ordered by inclusion. Prove that C(G) is a lattice if and only if every block of
G is a complete graph. (Klavžar–Petkovšek [1988])
Exercises for Section 12.4 607

12.4.30. For a lattice L, prove directly that x ∨ (y ∧ ) = (x ∨ y) ∧ (x ∨ ) for all x , y , ∈ L

if and only if x ∧ (y ∨ ) = (x ∧ y) ∨ (x ∧ ) for all x , y , ∈ L.
12.4.31. (♦) Prove that a lattice L is distributive if and only if (x ∧ y) ∨ (y ∧ ) ∨ ( ∧ x)
= (x ∨ y) ∧ (y ∨ ) ∧ ( ∨ x) for all x , y , ∈ L. (Comment: There are many proofs.)
12.4.32. Two elements x and different from y are y-equivalent in a lattice L if x ∧ y =
y ∧ and x ∨ y = y ∨ . Prove that L is distributive if and only if for each y ∈ L there is
no pair of y-equivalent elements. (Hint: Use Exercise 12.4.31 and Theorem 12.4.23.)
12.4.33. (♦) The dual R∗ of a poset R is obtained by reversing all relations.
(a) Prove that J(R∗) = J(R)∗ for every poset R.
(b) Prove that P and Q are isomorphic if J(P) and J(Q) are isomorphic.
(c) Conclude that in a finite distributive lattice L the posets P(L) and Q(L) (join-
irreducibles and meet-irreducibles) are isomorphic.
12.4.34. Applications of the ideal map (Theorem 12.4.31) on a distributive lattice L.
(a) For x , y ∈ L, prove that { ∈ L: x ∧ = y ∧ } is a down-set in L. (Birkhoff [1967])
(b) Prove that the number of elements of L that cover exactly  elements equals the
number of elements of L covered by exactly  elements. (Hint: Use Exercise 12.4.33.)
12.4.35. (♦) Distributive lattices. For an element x in a finite lattice L, let (x) =
{p ∈ P: p ≤ x}, where P is the poset of join-irreducible elements of L, and let (x) =
{q ∈ Q: q ≥ x}, where Q is the poset of meet-irreducible elements of L.
(a) Prove that : L → J(P) embeds L as a subposet of J(P).
(b) Prove that the following are equivalent (via A ⇒B⇒C⇒D⇒ A.)
(A) L is distributive.
(B) L is graded and | P | = r(L) = | Q| .
(C) | P | = | (x)| + |(x)| = | Q| for all x ∈ L.
(D) |(x ∨ y)| = |(x) ∪ (y)| for all x , y ∈ L.

12.4.36. In a lattice, y is a complement of x if x ∧ y = 0̂ and x ∨ y = 1̂.

(a) Using the isomorphism : L → J(P), characterize the elements of a distributive
lattice that have complements. Conclude that in any distributive lattice the number of
elements having complements is a power of 2.
(b) In a product of chains, which elements have complements?
(c) A lattice is complemented if every element has a unique complement. Show that
the only complemented distributive lattices are the subset lattices.
12.4.37. Prove that the order dimension (Section 12.3) of a distributive lattice equals the
width of its poset of join-irreducible elements. (Comment: The famous “Dilworth’s Theo-
rem” was proved as a lemma for this result.) (Dilworth [1950])
12.4.38. Let L be a distributive lattice. Prove that the smallest dimension of a subset
lattice in which L appears as a subposet is the number of join-irreducible elements in L.
12.4.39. A graded lattice L is semimodular if its rank function is submodular, meaning
r(x ∧ y) + r(x ∨ y) ≤ r(x) + r(y) for all x , y ∈ L. Prove that the product of two semimodu-
lar lattices is a semimodular lattice. Conclude that the divisor lattice D(N) is modular,
meaning that equality always holds in the submodularity inequality.
12.4.40. (♦) Semimodularity of the partition lattice Π n.
(a) Prove by induction that r( ∧  ) + r( ∨  ) ≤ r() + r( ) for  ,  ∈ Π n .
(b) Given two partitions  ,  of [n], let G( ,  ) be the bipartite graph whose parts are
the blocks of  and  , respectively, with vertices adjacent if they intersect as sets. In terms
of the number of blocks in  ,  ,  ∧  ,  ∨  , compute the number of vertices, edges, and
components of G( ,  ). Use this to give another proof of part (a). (Aigner [1979])
608 Chapter 12: Partially Ordered Sets

12.4.41. Applications of Kleitman’s Inequality (Theorem 12.4.34).

(a) Prove that 2 n−2 is the maximum size of an intersecting family whose complements
also form an intersecting family. (Seymour [1973], Schönheim [1974], Daykin–Lovász
[1976], Hilton [1976], Anderson [1976], Greene–Kleitman [1978])
(b) Prove that 2 n − 2 n−¾ is the maximum size of the union of ¾ intersecting families in
n (Kleitman [1966])
2 .
12.4.42. A use of the up-set/down-set form of Kleitman’s Inequality (Theorem 12.4.34).
(a) Let X be an up-set and Y a down-set in 2 n . Let X and Y be their complements,
and let X − Y and Y − X be their differences as sets. Use Kleitman’s Inequality to prove
# # # #1/2
| X − Y | | Y − X | ≥ | X ∩ Y | #### X ∩ Y #### , and conclude | X ∩ Y |1/2 + #### X ∩ Y #### ≤ 2 n/2 .
(b) Let F , G be subsets of 2 such that no member of either contains a member of the
n

other. Apply (a) to prove | F | + | G| ≤ 2 n/2 . (Seymour [1973])

1/2 1/2

12.4.43. (♦) Using Theorem 12.4.23 and Theorem 12.4.40, prove that a lattice L is
distributive if and only if | F | | G| ≤ | F ∧ G| | F ∨ G| for all F , G ⊆ L (Definition 12.4.39).
Conclude that Kleitman’s Inequality holds for distributive lattices. (Daykin [1977])
12.4.44. For X , Y ⊆ 2 n , let X − Y = {A − B: A ∈ X , B ∈ Y }. Apply Exercise 12.4.43
to prove | X − X | ≥ | X | for any family X of finite sets. (Comment: Daykin–Lovász [1976]
showed in general that every nontrivial boolean function takes at least m distinct values
when evaluated over m distinct sets.)
12.4.45. Complete the proofs of the FKG and Ahlswede–Daykin Inequalities by proving
the latter for 21 . (Hint: For the case X = Y = {∅ , [1]}, let w = (∅)([1]), x = ([1])(∅),
y = (∅)([1]), and  = ([1])(∅). Reduce the problem to inequalities involving w , x , y , .)
12.4.46. Suppose that the outcome of matches between tennis players is determined by an
unknown linear ordering of ability. Suppose that A and B are two teams with two players
each, and that initially we know nothing about the relative abilities of the two players on
a team. However, we do know a2 < b1 . Determine whether the events a1 < b1 and a2 < b2
are positively correlated, assuming that unknown information is random. (Shepp [1980])
12.4.47. Let Q = 2 + 4. Let  and y be the top and bottom of the 4-chain, and let w and x
be the top and bottom of the 2-chain. Prove that the events x < y < w and x <  < w are
not postively correlated on Q. (C. Mallows)
12.4.48. (♦) For x ∈ Q, let the random variable Hx be the height of x on a random linear
extension of Q, and let h(x) = (Hx). Let y be incomparable to x. Use the XYZ Inequality
to prove h(x|(x > y)) ≥ 1 + h(x|(x < y)). (Winkler [1982])
12.4.49. (+) The track number of a graph G, written (G), is the least t such that G
is the union of t interval graphs. The interval number of G, written i(G), is the least t
such that G is the intersection graph of sets that are unions of t intervals in . Note that
always i(G) ≤ t(G). We show that t(G) is unbounded even when i(G) = 2, as conjectured by
Heldt , Knauer, and Ueckerdt (see Milans–Stolee–West [2015]).
(a) Prove that i(G) ≤ 2 when G is a line graph.
(b) Prove that if n ≥ R¾(P ), then (L(K n)) >  , where P is the ordered hypergraph
with vertices 1 , 2 , 3 , 4 , 5 , 6 in order obtained from the monotone path P63 by adding the
two edges {1 , 2 , 5} and {2 , 5 , 6}.
(c) Prove that if n < R¾(P43), then (L(K n)) ≤  + 2.
(d) For the complete 3-uniform hypergraph K63 , Erdős–Rado [1952] proved the bound
O(¾ lg ¾) (4+ o(1))¾ lg ¾
R¾(K63) ≤ 2 2 (improved by Conlon–Fox–Sudakov [2010] to 2 2 ). Use this and the
√
¾ −
bound R¾ (Pt ) ≥ 2 ¾
from Corollary 12.4.46 (where m = t − r + 1) to prove
r m 1 /2

( lglglglglgnn ) ≤ (L(K n)) ≤ lg lg n + 12 lg lg lg n + O(lg lg lg lg n).

Chapter 13

Combinatorial Designs
In this chapter we study highly structured combinatorial arrangements.
Block designs are families of sets (actually, uniform hypergraphs) with special
conditions on pairs of elements. They are useful in constructions for extremal
problems and in designing statistical experiments.
We study general designs and special configurations such as Latin squares,
Hadamard matrices, and projective planes. Tools involve finite fields and differ-
ence sets. The final more technical section leads to a constructive disproof of a
famous conjecture of Euler.
Many texts and monographs treat classical design theory. For example, An-
derson [1997], Cameron–van Lint [1991], Lindner–Rodger [1997, 2009], Stinson
[2004], van Lint–Wilson [1992, 2001], and Wallis [1988, 2007] are patient and ac-
cessible. Beth–Jungnickel–Lenz [1986, 1999] and Colbourn–Dinitz [1996, 2007]
are encyclopedic. Dinitz–Stinson [1992] and Wallis [1996, 2003] provide surveys
on special topics.

13.1. Arrangements
For hundreds of years, mathematicians have been fascinated by questions
about the arrangements of labels that use all labels or all combinations of labels
in equal and symmetric ways. The requirements are quite rigid. For example, so-
lutions to “Sudoku” puzzles are Latin squares satisfying additional constraints.

L ATIN SQUARES

We begin with Latin squares and the design of experiments.

13.1.1. Definition. A Latin square of order n is an n-by-n matrix with n dis-

tinct entries such that each label occurs exactly once in each row and each
column. Two such squares are orthogonal if the n2 ordered pairs of labels in
corresponding positions are distinct. A family of pairwise orthogonal Latin
squares is an orthogonal family.

609
610 Chapter 13: Combinatorial Designs

13.1.2. Example. Growing corn. Suppose we wish to test four different types of
corn seeds (A , B , C, D). We want to subdivide our field into plots in order to min-
imize the effect of the geography of the field on the test results. To guard against
the possibility of an east-west or north-south gradient in soil quality, we can di-
vide the field into a four by four grid of subplots and assign seed types to subplots
using a Latin square design.
Given also four types of fertilizer to test ( , ,  , ), we can use another Latin
square for those. Ideally, we also want to test each fertilizer with each seed type.
We can do this using a Latin square orthogonal to the first square. We can also
test levels of irrigation (1,2,3,4) at the same time; the three Latin squares below
are pairwise orthogonal. There is no larger family of pairwise orthogonal Latin
squares of order 4.
ABCD    1234
DCBA    341 2
BADC    4321
CDAB    2143

Versatile experimental designs require large orthogonal families. There is a

trivial upper bound on the size of such a family.

13.1.3. Lemma. Every orthogonal family of order n has size at most n− 1.

Proof: By relabeling elements, we can normalize each square so that the entries
in each first row are the integers 1 , . . . , n in order (see Example 13.1.2). With
this normalization, the pairs (i , i) have already all occurred in the first row for
all i and every pair of squares in the family.
Since these pairs already occur, in a particular position below the first row
the values in all squares in the family must be distinct. They also must differ
from the column index, since that value already appears at the top of the column
and these are Latin squares. There are only n − 1 such values and hence at most
n − 1 squares in the family.

13.1.4. Definition. An orthogonal family of order n and size n − 1 is a complete

family. A family of  pairwise-orthogonal Latin squares of order n is indi-
cated by MOLS(n , ), read as “mutually orthogonal family of Latin squares”.

This use of “mutually” to mean “pairwise” is historical; “pairwise” is more

accurate. This differs from the notion of mutual independence in Section 14.2,
which is decidedly not pairwise. Dénes–Keedwell [1974, 1991, 2015] provides ex-
tensive material on Latin squares.
Complete families arise from finite fields. There is a finite field of size (order)
n if and only if n is a power of a prime. The field is unique up to isomorphism and
is denoted n . The additive identity is 0, the multiplicative identity is 1, and
every nonzero element has both an additive and a multiplicative inverse.

13.1.5. Theorem. (Moore [1896], Bose [1939], Stevens [1939]) If n is a power of

a prime, then there is a complete family MOLS(n , n − 1).
Section 13.1: Arrangements 611

Proof: Let the elements of n be x1 , . . . , x n , with x n the additive identity. The

labels in each square are {x i}; the field arithmetic determines where to put them.
Define the ¾ th square by putting x¾ x i + x j in position (i , j), for 1 ≤ ¾ ≤ n − 1.
Because addition of a single element or multiplication by a single nonzero
element permutes the elements of the field, we have defined n − 1 Latin squares.
To prove orthogonality, suppose that the pair in position (i , j) of squares ¾ and l
is the same as the pair in position (r, s) of squares ¾ and l. That is,
x¾ x i + x j = x¾ xr + x s ,
x l x i + x j = x l xr + x s .
Subtracting the two equations and canceling x¾ − x l yields xr = x i , which in turn
yields x s = x j . Thus (r, s) = (i , j).

When n is not a prime power, this construction still yields some orthogonal
squares (Exercise 6). Next we give a combining theorem for orthogonal families.

13.1.6. Theorem. (Moore–MacNeish Theorem; Moore [1896], MacNeish

[1922]) Given h pairwise orthogonal Latin squares of order m and h pair-
wise orthogonal Latin squares of order n, there are h pairwise orthogonal
Latin squares of order mn.
Proof: Let A1 , . . . , A h and B1 , . . . , Bh be the two families. Let C1 , . . . , Ch be h
matrices defined as follows. The elements of C¾ are pairs of elements from A¾
and B¾ . The matrix C¾ splits into m2 square blocks of order n (shown below) such
that the first coordinates of the entries in the (i , j)th block all equal ai¾, j and the
( )

second coordinates form a copy of B¾ .

Because A¾ and B¾ are Latin squares, C¾ is also. If two pairs of corresponding
positions in C¾ and Cl have the same entries, then
([a(i¾,)j , b(r,s
¾)
] , [a(l) (l)
]) ([a(p¾,q
i , j , br,s =
)
, bt¾,u] , [ap ,q , bt ,u]).
( ) (l) (l)

Now the orthogonality of A¾ and A l yields (i , j) = (p , q), and the orthogonality of

B¾ and Bl yields (r, s) = (t , u).

(a(1¾,1
) (¾)
, br,s) ··· (a(1¾,m
) (¾)
, br,s)

.. .. ..
C¾ = . . .

(a(m¾),1 , b(r,s
¾)
) ··· (a(m¾),m , b(r,s
¾)
)

13.1.7. Corollary. If n = ∏ pei i with each pi prime, then there exist mini(pei i − 1)
pairwise orthogonal Latin squares of order n.
612 Chapter 13: Combinatorial Designs

In general, there are larger families of MOLS than guaranteed by Corollary

13.1.7. When n is twice an odd number, Corollary 13.1.7 does not guarantee any
pair of orthogonal Latin squares at all.

13.1.8. Example. Euler’s problem. The problem of finding MOLS(6 , 2) was pub-
lished by Euler in 1779 as the Problem of the 36 Officers. On parade day, there
are 36 officers: six ranks, six regiments, one officer of each rank from each regi-
ment. They desire to march with one officer of each rank in each row and column,
and one officer of each regiment in each row and column. Thus the appearances of
each rank must form a Latin square, as must the appearances of each regiment.
Since there is only one officer of a given rank from a given regiment, these Latin
squares must be orthogonal.
Euler conjectured impossibility. By normalizing the top rows to the labels
123456 in order, Tarry [1900, 1901] reduced the problem to an exhaustive exam-
ination of 9408 pairs, proving Euler right. Stinson [1984] found a much shorter
proof (see Anderson [1997, pp. 130–133]).

13.1.9. Remark. Due to Euler ’s use of Greek and Latin letters for orthogonal
squares, orthogonal pairs are also called Graeco–Latin squares.
Euler conjectured also that there are no orthogonal n-by-n Latin squares
whenever n is an odd multiple of 2, but here he was very wrong. Bose–
Shrikhande–Parker [1960] proved that for n ∈ / {1 , 2 , 6}, orthogonal Latin
squares of order n exist. We will discuss this in Section 13.3.
It is a long way from 2 to n − 1. The next value after 6 that is not a prime
power is 10. A long-running supercomputer search concluded that no complete
family of order 10 exists (see Lam–Thiel–Swiercz [1989]). Even after this long
search, the maximum size of an orthogonal family is not known; in fact, it is not
known whether there are 3 pairwise orthogonal 10-by-10 Latin squares.

We will see that complete families of order n are equivalent to various other
combinatorial configurations, including projective planes of order n, which can be
viewed as a special class of block designs. To indicate more of the scope of design
theory, we next introduce the general object.

BLOCK DESIGNS

In the setting of Example 13.1.2, the treatments form a set V . For complete
testing, we would like to test each treatment on each experimental unit, but this
may be expensive or infeasible. In a more general setting, each unit receives only
a subset of V , called a block. Several conditions are imposed to balance the usage
of treatments and units.

13.1.10. Definition. A (v , ¾ , ë)-design is a family of blocks of size ¾ from a set V

of size v, such that each element of V appears in the same number of blocks
and every two elements of V appear in ë common blocks. Let b denote the
number of blocks and r denote the number of blocks containing a given ele-
ment. When ¾ < v, this is a balanced incomplete block design (BIBD) .
Section 13.1: Arrangements 613

Having each treatment appear in r blocks ensures that treatments are tested
equally often; keeping the blocks the same size ensures that no experimental unit
is overused. Fixing ë helps to control for interaction between treatments. These
conditions were introduced formally by Yates [1936], but special cases of designs
were studied at least as early as 1844, with a seminal paper by K irkman [1847].
Originally, all five parameters were listed in naming a design (how fast can
you say “(v , b , r, ¾ , ë)-design”?). However, two simple counting arguments reduce
the number of independent parameters to three (Proposition 13.1.14). Thus we
simply call the object a (v , ¾ , ë)-design. (Even the hypothesis that all elements
appear in equally many blocks can be dropped; see Exercise 5.)
Historically, the treatments were called varieties, and thus the notation re-
flects the terminology: v counts the varieties, b the blocks, r the “replications”
(usage of each element), ¾ the “kardinality” of the blocks, and ë the “linkage”
(common appearances) of any two elements. Because we can view the blocks as
the edge set of a hypergraph with vertex set V , we usually call V the set of ver-
tices rather than the set of varieties.

13.1.11. Example. Complete ¾ -uniform hypergraph. When | V | = v, the family (V¾ )

is the set of blocks of a (v , ¾ , (¾v−−22))-design. There are (¾v) blocks, each element
appears in (¾v−−11 ) blocks, and each pair of elements appears in (¾v−−22) blocks.

13.1.12. Example. Fano plane. The Fano plane is the (7 , 3 , 1)-design with ver-
tex set [7] and blocks {124 , 235 , 346 , 457 ,561 ,672 ,713} (for clarity, we delete
commas within blocks and drop set brackets on blocks). Viewing the elements as
congruence classes modulo 7, the blocks have the form {i , i + 1 , i + 3}. Every pair
{r, s} of congruence classes is one apart, two apart, or three apart. It thus occurs
uniquely in the block generated by i such that {r, s} is {i , i + 1} or {i + 1 , i + 3} or
{i , i + 3}, respectively.
We can represent the Fano plane geometrically by letting the blocks corre-
spond to lines through the points; hence the use of “plane” in the name. As shown
in the famous diagram below, one line must “bend”.
If we delete one of these blocks and delete its points from all other blocks,
then we are left with six blocks consisting of all pairs from a set of four points,
which is an instance of Example 13.1.11.

1 B
• ⎛1 0 0 0 1 0 1⎞
⎜1
⎜ 1 0 0 0 1 0⎟
⎟
⎜0
⎜ 1 1 0 0 0 1⎟
⎟
6 2
• 3 • V⎜
⎜1 0 1 1 0 0 0⎟
⎟
• ⎜0
⎜ 1 0 1 1 0 0⎟
⎟
⎜ ⎟
⎜0 0 1 0 1 1 0⎟
• • • ⎝0 0 0 1 0 1 1⎠
5 7 4

13.1.13. Example. 9-point triple system. From nine points in a square array, we
form four sets of three triples. We use the rows, the columns, the rightward diag-
onals, and the leftward diagonals. Two points in the same row or column clearly
614 Chapter 13: Combinatorial Designs

appear together in exactly one block. If they are not in the same row or column,
then there is exactly one way to reach one from the other along a diagonal. Hence
the 12 triples form a (9 , 3 , 1)-design.
123 147 159 168
456 258 267 249
789 369 348 357
• • •
• • •
• • •

The incidence matrix of a design is the matrix A of the incidence relation

between the set V of elements and the set B of blocks. Position (i , j) is 1 if the
ith element belongs to the jth block, 0 otherwise. The incidence matrix is a fun-
damental tool in the study of designs, facilitating counting arguments. The first
such arguments yield necessary conditions on the parameters and an algebraic
interpretation of the definition condition for usage of pairs of vertices.

13.1.14. Proposition. Every (v , ¾ , ë)-design having b blocks and r appearances

of each element satisfies the equations below, where A is the incidence matrix
of the hypergraph of blocks and J is the all-1 square matrix.
(a) b¾ = vr.
(b) r(¾ − 1) = ë(v − 1).
(c) A A T = (r − ë)I + ë J.
Proof: The matrix A has ¾ 1s in each column and r 1s in each row, so each side
of (a) counts the 1s in A.
For an element x, each side of (b) counts all appearances of elements other
than x in blocks containing x, on the left by blocks and on the right by elements.
For (c), the entry in position (i , j) of A A T is the dot product of rows i and j
of A. By definition, this is ë when i
= j and is r when i = j .

Thus r = ë(v − 1)/(¾ − 1) and b = vr/¾ for a (v , ¾ , ë)-design. Next we use

algebraic properties of the incidence matrix to prove a fundamental inequality.

13.1.15. Theorem. (Fisher ’s Inequality; Fisher [1940]) If ¾ < v in a (v , ¾ , ë)-

design, then b ≥ v.
Proof: Let A denote the incidence matrix. We use Proposition 13.1.14c to show
linear independence of the rows of A. If w T A = 0, then associativity of matrix
multiplication yields
0 = w T A A T w = w T [(r − ë)I + ë J]w = (r − ë)(∑ w2i ) + ë(∑ wi)2 .
Since r(¾ − 1) = ë(v − 1) always, ¾ < v implies r > ë. Hence the terms on the
right are both nonnegative and hence both 0. From ∑ w2i = 0 we obtain w = 0.
Hence the rows of A are linearly independent, and therefore A has at least as
many columns as rows.
Section 13.1: Arrangements 615

SYMMETRIC DESIGNS

When b = v, the incidence matrix is square, and by the proof of Fisher ’s

Inequality it is nonsingular. Such designs have many additional nice properties.

13.1.16. Definition. A (v , ¾ , ë)-design is a symmetric design if b = v.

By Proposition 13.1.14a, b = v in a symmetric design implies r = ¾ ; the num-

ber of elements in each block is the same as the number of blocks containing each
element. We will also show that not only does each pair of elements appear in ë
common blocks, but also each pair of blocks has ë common elements.

13.1.17. Example. With 7 blocks and 7 elements, the Fano plane (Example
13.1.12) is a symmetric design. In the Fano plane, we compare two blocks A =
{i , i + 1 , i + 3} and B = { j , j + 1 , j + 3} by subtracting each element of A from
each element of B. The difference j − i appears three times, and the other differ-
ences are the six other values from j − i − 3 through j − i + 3. This means that
when j
= i the difference 0 appears exactly once, and the two blocks have exactly
one common element.

The statement that every two blocks have ë common elements is just A T A =
(¾ − ë)I + ë J. This statement is nontrivial, whereas the statement A A T =
(¾ − ë)I + ë J from Proposition 13.1.14c is essentially by definition. We conclude
that for a symmetric design, A and A T commute.

13.1.18. Theorem. (Bose [1939]) In a symmetric (v , ¾ , ë)-design, every two blocks

have ë common elements.
Proof: Let B be a block in the design, and let x1 , . . . , x v−1 be the sizes of the
intersections of B with the other blocks. To prove that each x i equals ë, we will
show that ∑(x i − ë)2 = 0.
Since each element of B appears in ¾ blocks, ∑ x i = ¾(¾ − 1) = ë(v − 1) (ap-
plying Proposition 13.1.14). Counting the appearances of pairs of elements from
B yields (¾2) + ∑ ( x2i ) = (¾2) ë, which we rewrite as ∑(x2i − x i) = ¾(¾ − 1)(ë − 1) =
ë(ë − 1)(v − 1) (applying Proposition 13.1.14). Summing the two identities yields
∑ x2i = ë2(v − 1). Now
∑(xi − ë)2 = ∑ x2i − ∑ 2 ë x i + ∑ ë2
= ë2(v − 1) − (2 ë)ë(v − 1) + (v − 1)ë2 = 0 ,
as desired.

Next we obtain necessary conditions for symmetric (v , ¾ , ë)-designs.

13.1.19. Proposition. If a symmetric (v , ¾ , ë)-design exists, then

¾(¾ − 1) = ë(v − 1).
Proof: Symmetric designs require r = ¾ (Proposition 13.1.14a), so Proposition
13.1.14b becomes ¾(¾ − 1) = ë(v − 1).
616 Chapter 13: Combinatorial Designs

13.1.20. Example. Symmetric designs. Proposition 13.1.19 requires ¾ > when

 < v (we exclude ( ,  , )-designs, which are “complete” designs but can be called
trivial designs). Some such triples are easy to achieve.
( + 1 ,  ,  − 1): The complete -uniform hypergraph with  + 1 vertices. Note
that any two -sets in [ + 1] have  − 1 common elements.
(( − 1) + 1 ,  , 1): Symmetric designs with = 1 are called “projective
planes”, studied in Section 13.2. They exist when  − 1 is a power of a prime.
The case (7 , 3 , 1) is the Fano plane of Example 13.1.12. The case (43 , 7 , 1) will be
forbidden by Theorem 13.1.23 in Example 13.1.24.
For  ≤ 7, the triples with 2 ≤ ≤  − 2 that satisfy Proposition 13.1.19 are
(7 , 4 , 2), (11 , 5 , 2), (11 , 6 , 3), (16 , 6 , 2), (15 , 7 , 3), and (22 , 7 , 2). The last one will
be forbidden by the next proposition. The other cases listed here with = 2 exist
and are called biplanes.

Our remaining necessary conditions use the algebraic statements we have

proved about symmetric designs.

13.1.21. Proposition. (Schützenberger [1949], Shrikhande [1950]) If a symmet-

ric (v ,  , )-design exists with v even, then  − is a square.
Proof: Let A be the incidence matrix. For a symmetric design, we have b = v and
r =  . By Proposition 13.1.14c, A A T = ( − )I + J. The eigenvalues of J are
v with multiplicity 1 and 0 with multiplicity v − 1. Adding a multiple of I shifts
the eigenvalues, so the eigenvalues of A A T are  + (v − 1) with multiplicity 1
and ( − ) with multiplicity v − 1.
Since ( − 1) = (v − 1) (Proposition 13.1.14b), the largest eigenvalue is  2 .
The determinant is the product of the eigenvalues, so det(A A T ) =  2( − )v−1 .
Since det(A A T ) = det A det A T = (det A)2 , we have det A = ( − )(v−1)/2 . Since A
is an integer matrix, det A is an integer, and thus ( − )(v−1)/2 is rational. With
v even, this requires  − to be a square.

Proposition 13.1.21 prohibits (22 , 7 , 2)-designs, since 5 is not a square.

The next theorem is the work of Bruck–Ryser [1949] and Chowla–Ryser
[1950]. The proof uses Lagrange ’s Theorem (every positive integer is a sum of
four squares) and a special matrix construction.

13.1.22. Lemma. The matrix below satisfies HH T = (b12 + b22 + b33 + b42)I4 .

⎛ b1 b2 b3 b4 ⎞
⎜ −b b1 −b4 b3 ⎟
H=⎜ 2 ⎟
⎜ −b3 b4 b1 −b2 ⎟
⎝ −b4 −b3 b2 b1 ⎠

13.1.23. Theorem. (Bruck–Chowla–Ryser Theorem) If a symmetric (v ,  , )-

design exists, then
(a) If v is even, then  − is the square of an integer.
(b) If v is odd, then 2 = ( − )x2 + (−1)(v−1)/2 y 2 has a nonzero solution in
integers x , y , .
Section 13.1: Arrangements 617

Proof: Statement (a) is Proposition 13.1.21. For odd v, let n = ¾ − . Let A be

the incidence matrix, and let x = (x1 , . . . , x v)T ; this is a column vector of v free
variables. Let (1 , . . . , v)T = Ax. Theorem 13.1.18 yields A T A = nI + J. Hence
v v v 2
⎛ ⎞
∑  = (x
2
i
T
A )(Ax) = n ∑
T
x 2i + ∑ xi .
⎝ i=1 ⎠
(∗)
i=1 i=1

We will modify this equation to obtain the desired equation satisfied by inte-
gers x , y , . First, Lagrange ’s Theorem expresses n as the sum of four squares:
n = b12 + b22 + b32 + b42 . Let (y1 , y2 , y3 , y4)T = H T (x1 , x 2 , x3 , x4)T , where H is the
matrix of Lemma 13.1.22 defined on b1 , b2 , b3 , b4 . Since HH T = nI4 , we have
∑i=1 yi2 = n ∑i=1 x2i . We call such a replacement H-substitution.
4 4

Case 1: v ≡ 1 (mod 4). With w =

v
∑i=1 x i , applying H-substitution to sets of
four variables yields
v v−1

∑ i2 = ∑ yi2 + nx2v + w2 . (∗∗)

i=1 i=1

Since H is invertible and has integer entries, we can write x1 , . . . , x v−1 as

rational linear combinations of {y1 , . . . , yv−1 }. Thus also 1 , . . . , v−1 are ration-
al linear combinations of {y1 , . . . , yv−1 , xv}, treating these as the freely chosen
v−1
variables. Consider 1 = ∑ i=1  i yi + v x v . If 1
= 1, then restricting the choice of
v−1
y1 by requiring y1 = 1−1 1 [∑i=2 ai yi + av xv ] yields 1 = y1 . If 1 = 1, then setting
v−1
y1 = −11− 1 [∑i=2 ai yi + av xv ] yields 1 = − y1 . In either case, we have 12 = y12 and
can cancel 12 and y12 from (**), leaving an expression in the remaining variables.
With this substitution for y1 , we have expressed 2 as a rational linear combi-
nation of y2 , . . . , yv−1 and xv . Suitably restricting y2 in terms of the free variables
y3 , . . . , yv−1 and xv yields 22 = y22 and allows us to cancel 22 and y22 from (**).
v−1 v−1
Repeating this substitution procedure leads to canceling ∑i=1 i2 and ∑ i=1 yi2
from (**) by restricting y1 , . . . , yv−1 as appropriate rational multiples of the re-
maining free variable x v , leaving the equation v2 = nx2v + w2 . Since w = ∑i=1 x i
v

and (1 , . . . , v)T = Ax, both w and v are linear combinations of x1 , . . . , x v . We

expressed x1 , . . . , x v−1 as rational combinations of y1 , . . . , yv−1 , which in turn
were reduced to rational multiples of xv . Now multiplying by the least common
multiple of all the denominators yields an integer solution to 2 = ( − )x2 + y 2 .
Case 2: v ≡ 3 (mod 4). First add nx 2v+1 to both sides of (*), where xv+1 is a
new free variable. Now the variables x1 , . . . , x v+1 are transformed to y1 , . . . , yv+1
in sets of four by H-substitution. The equation (**) becomes nx 2v+1 + ∑i=1 i2 =
v

v+1
∑i=1 yi2 + w2 in terms of the free variables y1 , . . . , yv+1 . Iterative restriction
and cancellation reduces this to nx 2v+1 = yv2+1 + w2 in v steps. Now the clearing
of fractions yields an integer solution of ( − )x2 = 2 + y 2 , as desired.

13.1.24. Example. Except for (16 , 6 , 2) and (43 , 7 , 1), the unresolved triples from
Example 13.1.20 have the form (4m − 1 , 2m , m) or (4m − 1 , 2m − 1 , m − 1), where
m is an integer greater than 1. Both types are allowed by Theorem 13.1.23.
618 Chapter 13: Combinatorial Designs

For the first, we need 2 = mx 2 − my 2 = m(x − y)(x + y); set x = m + 1,

y = m − 1, and = 2m. For the second, we need 2 = mx 2 − (m − 1)y 2 ; set x = y =
= 1. The numerical conditions are only necessary conditions, but these designs
do exist for many m (see Proposition 13.1.29).
For (16 , 6 , 2), since v is even, our only necessary condition is that 6 − 2 is a
square. Such designs exist, but we will not construct them.
However, consider (43 , 7 , 1). We need a solution to y 2 + 2 = 6x 2 . We may
assume that x , y , have no common factor. The sum of two squares is a multiple
of 3 only if they are themselves multiples of three. Hence 3 divides both y and ,
so 9 divides the sum of their squares. This makes x divisible by 3, contradicting
the absence of common factors. (See Example 13.2.9 for a generalization.)

HADAMARD MATRICES

Designs with = 1 are studied the most. We pause to discuss special ma-
trices that yield symmetric designs with larger . Hadamard [1893] asked for
the largest magnitude of the determinant of an n-by-n matrix with entries ±1.
The bound of nn/2 is attained only by matrices of the following type (Exercise 13).
(Actually, these matrices were first studied by Sylvester [1867].)

13.1.25. Definition. A Hadamard matrix of order n is an n-by-n matrix H with

entries in {1 , −1} such that HH T = nI.

With entries ±1, the dot product of any row with itself is n. The rows
(or the columns, equivalently) must be pairwise orthogonal. We will obtain a
(4m − 1 , 2m − 1 , m − 1)-design from a Hadamard matrix of order 4m. First we
construct arbitrarily large Hadamard matrices.

13.1.26. Proposition. If A and B are Hadamard matrices of orders m and n,

then the matrix H with row indices (i , ), column indices (j , l) and entries
h(i ,¾) ,(j ,l) = ai , j · b¾ ,l is a Hadamard matrix of order mn.
Proof: This is the same construction as in Theorem 13.1.6, except that we record
the product of entries instead of the ordered pair. Grouping entries of H by the
first coordinate of each index expresses H as a matrix of n-by-n blocks, where the
(i , j)th block is ai , j B. With br denoting row r of B, the computation of the dot
product of row (i , ) and row (i ,  ) of H becomes

∑ ai , j ai , j (b¾ · b¾ ) = ¾ ,¾ n ∑ ai , j ai , j = ¾ ,¾ i ,i mn.
j j

13.1.27. Example. Note that (1) is a 1-by-1 Hadamard matrix. Because (11 −11 ) is
a Hadamard matrix, products yield Hadamard matrices for all powers of 2 (ob-
served by Sylvester [1867]). Below is the resulting matrix of order 4.
⎛1 1 1 1⎞
⎜1
⎜
−1 1 −1 ⎟
⎟
⎜1 1 −1 −1 ⎟
⎝1 −1 −1 1⎠
Section 13.1: Arrangements 619

13.1.28. Definition. A normalized Hadamard matrix is a Hadamard matrix

in which the first entry of every row and column is 1.

Multiplying a row or column of a matrix by −1 does not change the magni-

tude of the dot products of any two rows, so a Hadamard matrix of order n yields
a normalized Hadamard matrix of order n.

13.1.29. Proposition. (Hadamard [1893]) If n > 2 and an n-by-n Hadamard ma-


trix exists, then m = n/4 ∈ , and a (4m − 1 , 2m − 1 , m − 1)-design exists.
Proof: First normalize the Hadamard matrix. Since the dot product of the first
row with any other is 0, each row other than the first has n/2 elements of each
sign. Since n > 2, there are two other rows. Let m be the number of columns
where they both have +1. Since they each have n/2 elements of each sign, the
product is now 2m − 2(n/2 − m), as shown below. This must equal 0, so m = n/4,
and n is divisible by 4.
n/2 n/2
1 −1
1 −1 1 −1
m n/2 − m n/2 − m m
From the normalized Hadamard matrix, discard the first row and column
and change each −1 to 0. By the observations above, each row and column has
2m − 1 1s, and each pair of rows has 1s in exactly m − 1 common columns.

Conversely, a (4m − 1 , 2m − 1 , m − 1)-design yields a Hadamard matrix of

order 4m (Exercise 3). Hadamard matrices of order 4m are also equivalent to the
existence of (4m − 1 , 2m , m)-designs (Exercise 16). Designs with parameters of
these two forms are called Hadamard designs. The rows of the incidence matrix
of a Hadamard design are binary vectors that differ in many positions.
The fundamental problem of coding theory can be described as finding large
sets of vectors that pairwise are far apart. For binary vectors, the length is the
number of positions, the weight is the number of positions with 1, and the distance
between vectors is the number of places where they differ (Hamming distance).
When the distance between words (vectors) in the code is large, more transmission
errors can be tolerated, since they won’t change one codeword into another.
Let A(n , d) be the maximum size of a set of binary n-tuples with pairwise dis-
tance at least d. Hadamard matrices provide excellent codes, essentially achiev-
ing an upper bound on the code size.

13.1.30. Theorem. (Plotkin) If d > n/2, then A(n , d) ≤ 2d/(2d − n).

Proof: Consider a code of size N , and let s be the sum of the distances between
all pairs of code words. Since the distance is at least d for each pair, s ≥ d( N2 ).
We obtain an upper bound on s by considering the contributions to distances
from the n coordinates. A coordinate contributes to the distance between two
words if and only if the words differ in that position. If ¾ code words have 1 in po-
sition j and N − ¾ have 0 there, then position j contributes ¾(N − ¾) to the total
of the distances. Counting by positions, we thus have s ≤ nN 2/4.
Together, the bounds on s yield N ≤ 2d/(2d − n).
620 Chapter 13: Combinatorial Designs

13.1.31. Proposition. If n = 4m − 1 and there is a Hadamard matrix of order

n + 1, then A(n , (n + 1)/2) = n + 1.
Proof: In a Hadamard matrix of order n + 1, every two rows differ in (n + 1)/2
places. In a normalized Hadamard matrix, they agree in the first column. We
discard the first column and change each −1 to 0. We now have n + 1 words of
length n with pairwise distance (n + 1)/2. This achieves the Plotkin Bound for
d = (n + 1)/2 and hence is optimal.

When d = n/2, the Plotkin Bound is meaningless. Nevertheless, here again

Hadamard matrices give the largest possible codes.

13.1.32. Proposition. If d = n/2, then A(n , d) ≤ 2n, and this is achieveable if

there is a Hadamard matrix of order n.
Proof: Consider a code of size N , and let S and T be the sets of code words with
first coordinate 1 or 0, respectively. Deleting the first coordinate from every code
word in S creates a code of length n − 1 with minimum distance d, and similarly
for T . Hence N ≤ 2 A(n − 1 , d) ≤ 2n, by Proposition 13.1.31.
We obtain a code of size N by using the rows of the Hadamard matrix
(changing each −1 to 0) and the complements of these words. Since the pairs
(1 , 1) , (1 , −1) , (−1 , 1) , (−1 , −1) occur equally often between two rows, the dis-
tance between any two words in this code is n/2 or n.

To illustrate the application of Hadamard matrices and presage similar ap-

plications of projective planes, we consider one extremal problem. In applications
of designs to extremal problems, there is a bound that arises from a counting ar-
gument, and the bound can be attained only when the design-theoretic structure
is used in the construction.

13.1.33. Application. A bipartite Ramsey problem. (Beineke–Schwenk [1976])


Fix s , t ∈ . When n is large, every 0 , 1-matrix of order n has a constant s-by-
t submatrix; let b(s , t) be the least such n. In graph-theoretic language, this is
a bipartite analogue of Ramsey numbers (Section 10.2). Consider K n ,n with bi-
partition X , Y . A 0 , 1-matrix is a 2-coloring of E(K n ,n), and we want to force a
monochromatic copy of K s ,t with the part of size s in X .
Fix s = 2. For n ≥ 4t − 3, we prove that a color used on a subgraph G with at
least n2/2 edges of K n ,n yields such a copy of K s ,t . Let m = | E(G)|.
Consider the number p of copies of P3 in G whose center lies in Y . Since
any two vertices in X have at most t − 1 common neighbors in G, we have p ≤
(t − 1)(n2) if K 2 ,t does not occur in G. On the other hand, p = ∑ y∈ Y (d(y)
2
), since each
y ∈ Y extends in (d(y) 2 ) ways to form P3 . Since ∑ y ∈ Y d(y) = m, and the average
degree among vertices of Y is m/n, the convexity of quadratic functions yields
∑ y∈ Y (d(y)
2
) ≥ n(m/n
2
). Since m ≥ n2/2, we have m/n ≥ n/2. Thus
n/2 d(y) n
n( ) ≤ ∑( ) = p ≤ (t − 1)( ).
2 2 2
y∈ Y

This inequality simplifies to n nn−−2 ≤ 4t − 4, which requires n ≤ 4t − 3 for

1
integer n. Thus if n > 4t − 3, then a monochromatic K 2 ,t occurs. However, when n
Section 13.1: Arrangements 621

is odd, the numbers n2/2 and n/2 are not integers, which allows us to strengthen
the lower bound on p. We leave to Exercise 19 the completion of the proof that
also when n = 4t − 3, a monochromatic K 2 ,t occurs. Thus b(2 , t) ≤ 4t − 3.
Beineke–Schwenk [1976] proved that b(2 , t) = 4t − 3 when t ∈ {2 , 4} and
when there is a Hadamard matrix H of order 2t − 2 (this requires t odd). To prove
b(2 , t) > 4t − 4, we construct from H a matrix of order 4t − 4 with no constant
2-by-t submatrix. Obtain H from H by converting each −1 to 0, and let M =
H H
( H H ). Since any two rows of H have (t − 1)/2 common columns with 1s and
(t − 1)/2 common columns with 0s, any two rows of M have constant 2-by-(t − 1)
submatrices in both 0 and 1, but there is no constant 2-by-t submatrix.

Thus far we have constructed Hadamard matrices only of orders that are
powers of 2; others are more difficult. As observed in Example 13.1.24, Theorem
13.1.23 does not prohibit (4m − 1 , 2m − 1 , m − 1)-designs for any m. It is conjec-
tured that there is a Hadamard matrix of each order that is a multiple of 4. We
will show how to construct Hadamard matrices of order n = 2¾ (q + 1), where q is
an odd prime power and n is a multiple of 4, and then Proposition 13.1.26 yields
matrices of many other orders.

13.1.34.* Example. The Paley construction. Below are Hadamard matrices of or-
der 12 from the construction in Theorem 13.1.37 with q = 11 and q = 5.
+ + + + + + + + + + + + + + + + + + − + + + + +
− + + − + + + − − − + − + + + − − + + − + − − +
− − + + − + + + − − − + + + + + − − + + − + − −
− + − + + − + + + − − − + − + + + − + − + − + −
− − + − + + − + + + − − + − − + + + + − − + − +
− − − + − + + − + + + − + + − − + + + + − − + −
− − − − + − + + − + + + − + + + + + − − − − − −
− + − − − + − + + − + + + − + − − + − − − + + −
− + + − − − + − + + − + + + − + − − − − − − + +
− + + + − − − + − + + − + − + − + − − + − − − +
− − + + + − − − + − + + + − − + − + − + + − − −
− + − + + + − − − + − + + + − − + − − − + + − −

13.1.35.* Definition. A conference matrix C is a matrix of order n with diag-

onal entries 0 and off-diagonal entries ±1 such that CC T = (n − 1)I; that is,
the rows are pairwise orthogonal (Belevitch [1950]).

13.1.36.* Lemma. If C is an antisymmetric conference matrix, then C + I is

a Hadamard matrix. If C is a symmetric conference matrix, then H =
(CC−+II −CC−−II) is a Hadamard matrix.
Proof: When C is antisymmetric, meaning that C T = − C, we have
(C + I)(C + I)T = CC T + C + C T + I = (n − 1)I + I = nI.
When C is symmetric, computing HH T again suffices (Exercise 20).

13.1.37.* Theorem. (Paley [1933]) If q is an odd prime power and n = 2¾ (q + 1) ≡

0 (mod 4), then there is a Hadamard matrix of order n.
622 Chapter 13: Combinatorial Designs

Proof: By the construction of Proposition 13.1.26, it suffices to consider n = q + 1

when q ≡ 3 (mod 4) and n = 2(q + 1) when q ≡ 1 (mod 4). Using Lemma 13.1.36,
it suffices to construct a conference matrix of order q + 1, antisymmetric when
q ≡ 3 (mod 4) and symmetric when q ≡ 1 (mod 4).
The construction uses the notion of quadratic residues in finite fields. For an
element x of q , the finite field of order q, we define the character (x) to be 1 if
x is a nonzero square, −1 if x is not a square, and 0 if x = 0.
Observe that (x)(y) = (xy) and that ∑ x∈ q (x) = 0, since half of the
nonzero elements are squares. For a nonzero element c of q , we have

∑ (x)(x + c) = ∑ (x)(x)(1 + cx−1) = ∑ (1 + cx−1) = −1 ,

x∈ q x
=0 x
=0

since the last sum includes characters of all elements except 1.

List q as x0 , . . . , x q−1 with x0 = 0, and define a matrix A by ai j = (x i − x j ).
Since −1 is a square if and only if q ≡ 1 (mod 4), the matrix A is symmetric if q ≡
1 (mod 4) and antisymmetric if q ≡ 3 (mod 4). We have computed A A T = qI − J.
Now enlarge A to a matrix C of order q + 1 by adding a first row and column.
The first row is (0 1 1 · · · 1). The first column is (0 1 1 · · · 1)T when q ≡ 1 (mod 4)
and (0 −1 −1 · · · −1)T when q ≡ 3 (mod 4). The result is a conference matrix, since
each row or column of A sums to 0. The matrix C is symmetric or antisymmetric
as desired.

The conference matrices of Theorem 13.1.37 are Paley matrices. Scarpis

[1898] did this for prime q. Two resulting Hadamard matrices appear in Exam-
ple 13.1.34. With Proposition 13.1.26, Paley ’s Theorem guarantees Hadamard
matrices of all multiples of 4 through n = 128 except 92 and 116 (Exercise 4).
A Hadamard matrix of order 92 was found in Baumert–Golomb–Hall [1962]
by computer, based on a method of Williamson [1944] that uses the matrix of
Lemma 13.1.22 (see van Lint–Wilson [1992, pp. 177–179]). In 1978, the smallest
multiple of 4 for which no Hadamard matrix was known was 268 (Seberry [1978]).
In 1992, it was 428 (van Lint–Wilson [1992]). Later, K haraghani–Tayfeh-Rezaie
[2005] constructed one of order 428, leaving 668 as the next open case.

EXERCISES 13.1

13.1.1. (−) Explain how to construct a pair of orthogonal Latin squares of order 15. Include
all needed steps, but do not write out the final pair of squares.

13.1.2. (−) Show that there is a (v , ¾ , ë)-design with b blocks and each element appearing
in r blocks if and only if there is a (v , v − ¾ , b + ë − 2r)-design.

13.1.3. (−) Prove that if a (4m − 1 , 2m − 1 , m − 1)-design exists, then there is a Hadamard
matrix of order 4m.

13.1.4. Use Proposition 13.1.26 and Paley’s Theorem (Theorem 13.1.37) to show that
Hadamard matrices exist of all orders through n = 128 that are multiples of 4, except
possibly for 92 and 116.
Exercises for Section 13.1 623

13.1.5. Generalized block designs.

(a) Suppose that elements need not appear equally often. Each block has size ¾ and
each pair among the v elements appears together ë times, but the ith element appears ri
times. Prove that in fact r1 = · · · = rv ; no generalization.
(b) Suppose that blocks need not have equal size. Any two elements of [v] appear to-
gether ë times, and each element appears in r blocks. Construct an infinite family of
examples with ë = 1 in which the block sizes are not all equal to show that this is really a
generalization. Blocks of size 1 are not allowed, since they could be added to any design.

13.1.6. For n ∈ and 1 ≤ ¾ < n, define A(¾) by a¾i , j = ¾i + j (mod n) (see Theorem 13.1.5).
(a) Prove that A(¾) is a Latin square if and only if n and ¾ are relatively prime.
(b) When A(¾) and A(l) are Latin squares, prove that they are orthogonal if and only if
¾ − l is relatively prime to n.
(c) Prove that this construction cannot generate any larger families of pairwise or-
thogonal Latin squares of order n than Corollary 13.1.7.
13.1.7. (♦) A transversal in a Latin square is a set of positions, one in each row and col-
umn, containing distinct elements.
(a) Prove that a Latin square of order n belongs to a pair of orthogonal Latin squares
if and only if it consists of n disjoint transversals. (Laywine–Mullen [1998, p. 33])
(b) Prove that no Latin square is orthogonal to the Latin square below.

⎛ a b c d⎞
⎜b d a c⎟
⎜ c a d b⎟
⎝d c b a⎠

13.1.8. (♦) Let M be a Latin square that can be written as ( YX YX ) with X and Y being Latin
squares of odd order. Prove that M has no transversal. Use this to prove that there is no
Latin square orthogonal to M. (Comment: Maillet [1894] proved that the Cayley table of
a group of even order cannot have a transversal.) (Mann [1950])
13.1.9. Determine all n such that there is a Latin square of order n where each row is a
rotation of the first and the main diagonal consists of 1 , . . . , n in order. (Dályay [2012])
13.1.10. (♦) For n ≥ 3, prove that if MOLS(n , n − 2) exists, then a complete family
MOLS(n , n − 1) exists. (Comment: Shrikhande [1961] proved that MOLS(n , n − 3) suf-
fices. Bruck proved that if n > p(n − ¾), where p(x) = 12 x4 + x3 + x2 + 32 x, then a family
MOLS(n , ¾ − 2) can be completed to MOLS(n , n − 1); see Denes–Keedwell [1974, sec. 9.3].)
13.1.11. Prove that if MOLS(n , ¾) exists, then there exist ¾ − 1 pairwise orthogonal Latin
squares whose diagonals all consist of the elements 1 through n in order.
13.1.12. Let S be a ¾-set of elements in a symmetric (v , ¾ , ë)-design. Prove that if S in-
tersects each block at least ë times, then S is a block in the design. (Hint: Consider the
counting argument in Theorem 13.1.18.) (Lander [1981])
13.1.13. Let A be a matrix of order n whose entries are real and have absolute value at
most 1. Prove that |det A| ≤ nn/2 , with equality only when A is a Hadamard matrix.
13.1.14. (+) For distinct integers i , j ∈ [4], let P(i , j , ¾ , l) = {s ∈ 4n : si = ¾ and sj = l}.
That is, P(i , j , ¾ , l) is a plane specified by giving coordinates i and j fixed values ¾ and l.
Prove that there is a function á: 4n → 2n that maps each such plane bijectively onto 2n if
and only if there are two orthogonal Latin squares of order n. (Stong [2007])
13.1.15. Use the Fano plane to prove that K7 ,7 is not 3-choosable.
13.1.16. (♦) Prove that there is a Hadamard matrix of order 4m if and only if there is a
(4m − 1 , 2m , m)-design.
624 Chapter 13: Combinatorial Designs

13.1.17. A [¾]-pair-covering of [n] is a list º1 , . . . , ºm of functions from [n] to [¾] such

that for every {u , v} ∈ ([n]
2)
and every ordered pair (r, s) ∈ [¾]2 , there is some t such that
ºt(u) = r and ºt(v) = s. Prove that there is a [¾]-pair-covering of [n] with ¾ 2 functions if
and only if there is a family of n − 2 pairwise orthogonal Latin squares of order ¾ .
13.1.18. (+) The product dimension of a graph is the minimum ¾ such that each vertex
can be encoded as an integer ¾-tuple so that vertices are adjacent if and only if their codes
differ in every coordinate (see Definition 15.1.8). Let G = nK m (the disjoint union of n
copies of K m), with n ≥ 3. Prove that the product dimension of G is m if and only if there is
a family of (n − 1) pairwise orthogonal Latin squares of order m. (M. Sohoni [unpublished],
Evans–Isaak–Narayan [2000])
13.1.19. (♦) Bipartite Ramsey numbers. Define b(2 , t) as in Application 13.1.33.
(a) Strengthen the counting argument in Application 13.1.33 when n = 4t − 3 to com-
plete the proof that b(2 , t) ≤ 4t − 3. (Beineke–Schwenk [1976])
(b) Prove that b(2 , t) ≥ 4t − 4 when a Hadamard matrix of order 4t − 4 exists.
13.1.20. Complete the proof of the construction of a Hadamard matrix of order 2n from a
symmetric conference matrix of order n (Lemma 13.1.36).

13.2. Projective Planes

Symmetric designs with ë = 1 have a geometric interpretation. View the el-
ements as points and the blocks as lines. The condition ë = 1 says that any two
points determine exactly one common line; this suggests the terminology. Sym-
metrically, any two lines having one common point.

13.2.1. Definition. A projective plane is an incidence relation between points

and lines such that
(P1) every two points lie in one common line,
(P2) every two lines have one common point, and
(P3) there exist four points of which no three lie in a common line.

Property (P3) avoids degeneracy. (P2) is the non-Euclidean aspect; projective

planes have no parallel lines. Definition 13.2.1 has infinite realizations, but we
study finite projective planes, with finite sets of points and lines. The basic exis-
tence question is: for which values of n do projective planes with n points exist?

13.2.2. Example. Small projective planes. We will see that the Fano plane of
Example 13.1.12 is the smallest projective plane. With 13 points (0 to 9 plus
{A ,B,C}), the following 13 lines form a projective plane: 0139, 124A , 235B, 346C,
4570, 5681, 6792, 78A3, 89B4, 9AC5, AB06, BC17, C028.

REL ATION TO DESIGNS

The postulates in Definition 13.2.1 are quite strong. We will see that the
incidence matrix of a projective plane is that of a symmetric design with special
parameters and yields a complete family of orthogonal Latin squares.
Section 13.2: Projective Planes 625

We begin with point /line duality. Since (P1) and (P2) are dual as stated, we
need only show that the dual of (P3) is implied. The duality corresponds to trans-
posing the incidence matrix.

13.2.3. Proposition. In a finite projective plane, there are four distinct lines of
which no three contain a common point. Hence any statement about projec-
tive planes remains true when the roles of points and lines are interchanged
(simply transpose the incidence matrix).
Proof: Let a , b , c , d be four points with no three on a line (guaranteed by (P3)).
Let A , B , C, D be the lines guaranteed by (P1) to contain {a , b}, {b , c}, {c , d},
{d , a}, respectively. By the condition on a , b , c , d, these lines are distinct. If they
do not satisfy the claim, then by cyclic symmetry we may assume that A , B , C
have a common point x. The condition on a , b , c , d implies that x ∈ / {a , b , c , d},
but now A , B have two common points b , x and B , C have two common points c , x,
contradicting (P2).

a A b

D B
d C c

We next prove that the incidence relation of a projective plane has the defin-
ing properties of a design, with the points and lines as the elements and blocks
of a design. The number of points becomes the number of varieties v, the size of
each line becomes the block size ¾ , and ë = 1.

13.2.4. Theorem. In a finite projective plane, every point lies in the same num-
ber of lines, and every line has the same number of points.
Proof: By duality, it suffices to prove the latter, which we do by establishing an
injection from one arbitrary line A to another line B. This yields | B| ≥ | A| , and
interchanging them in the argument yields | A| ≥ | B| .
To define the map, we use a point not on A or B; we first obtain such a point.
(P3) provides four points a , b , c , d with no three on a line. If one of these is not
in A ∪ B, then it can be x. Otherwise, A and B each contain exactly two of them,
say a , b ∈ A and c , d ∈ B. Now (P1) provides a line containing {a , c} and another
line containing {b , d}, and (P2) guarantees that they have a common point x. If
x is in A or B, then we have two points on two distinct lines, violating (P1).
We map each p ∈ A to a point º (p) ∈ B using the unique line L(p) containing
p and x. Let º (p) be the point where L(p) meets B. If p
= q, then L(p)
= L(q),
else A and L(p) meet twice. This further implies º (p)
= º (q), else x and º (p) lie
on both L(p) and L(q). Hence | B| ≥ | A| , and by symmetry also | A| ≥ | B| .

b p
a B q B
x
d A º (p) A
c º (q)
x
626 Chapter 13: Combinatorial Designs

13.2.5. Theorem. In a finite projective plane, the number of lines containing

each point is the same as the number of points in each line.
Proof: By (P3), there are a point x and a line L such that x is not in L. By The-
orem 13.2.4, it suffices to define a bijection mapping the points in L to the lines
through x. For each y ∈ L, there is a unique line containing x , y; let this be º (y).
Let L be an arbitrary line through x. Since L and L meet exactly once, there is
a unique y for which L = º (y).
L

13.2.6. Definition. The order of a projective plane is one less than the number
of points in each line (or lines through each point).

13.2.7. Theorem. A family of (q + 1)-sets is the set of lines of a projective plane of

order q (≥ 2) if and only if it is the set of blocks of a (q2 + q + 1 , q + 1 ,1)-design.
Proof: Necessity. For a plane with v points and b blocks, Theorems 13.2.4–13.2.5
imply that the incidence relation is that of a (v , q+1 , 1)-design. Using Proposition
13.1.14, r = implies b = v, and then r( − 1) = (v − 1) implies v = 1 + (q + 1)q.
Sufficiency. Given a (q2 + q + 1 , q + 1 , 1)-design, call the blocks lines. Since
(q + q + 1) − 1 = (q + 1)q, by Proposition 13.1.14 the design is symmetric. Hence
2

(P1) and (P2) both hold. When q = 1, there are no four points to satisfy property
(P3) and hence no projective plane. It suffices to show that if q ≥ 2, then every
(q2 + q + 1 , q + 1 , 1)-design has four points with no three on a line.
Begin with one line L1 and distinct points a , b ∈ L1 . With q2 + q + 1 > q +
1, we can choose c outside L1 . Now there are unique lines L2 and L3 containing
{c , a} and {c , b}, respectively. Since any two lines have exactly one common point,
| L1 ∪ L2 ∪ L3 | = 3(q + 1) − 3 = 3q. Since q ≥ 2, we have 3q < q2 + q + 1. Hence a
point d remains outside L1 ∪ L2 ∪ L3 , and now {a , b , c , d} is the desired 4-set.

13.2.8. Example. Theorem 13.2.7 implies that (P1) and (P2) suffice when lines
have more than two elements, and (P3) becomes unnecessary.
No projective plane has order 1. Every projective plane of order 2 is isomor-
phic (by renaming points) to the Fano plane (Exercise 1). A projective plane of
order 3 (Example 13.2.2) has 13 points.

Theorem 13.2.7 states that a projective plane is just a symmetric design with
special parameters. We can thus use the Bruck–Chowla–Ryser Theorem (Theo-
rem 13.1.23) to prohibit some values of q as orders of projective planes.

13.2.9. Theorem. If q is congruent to 1 or 2 modulo 4, and the prime factoriza-

tion of q has odd exponent on some prime congruent to 3 modulo 4, then there
is no projective plane of order q.
Proof: A projective plane of order q is a symmetric (q2 + q + 1 , q + 1 , 1)-design.
If q ≡ 1 (mod 4) or q ≡ 2 (mod 4), then q2 + q + 1 ≡ 3 (mod 4) and (v − 1)/2 is odd,
where v = q2 + q + 1. The necessary condition for a (v , q + 1 , 1)-design in Theorem
13.1.23b now reduces to having an integer solution to y 2 + 2 = qx2 .
Section 13.2: Projective Planes 627

In Theorem 10.1.10 we proved Fermat ’s Two Squares Theorem: A natural

number is the sum of two squares if and only if all primes congruent to 3 modulo
4 occur with even exponent in its factorization. Hence when q is congruent to 1
or 2 modulo 4 the stated condition forbids projective planes of order q.

As observed by Bruck–Ryser [1949], Theorem 13.2.9 prohibits a projective

plane of order q when q ∈ {6 , 14 , 21 , 22 , 30 , 33 , 38 ,42 ,46 , 54 , . . .}. For order
10, we seek a (111 , 11 , 1)-design. The equation y 2 + 2 = 10x 2 has the solution
(x , y , ) = (1 , 3 , 1), so Theorem 13.2.9 does not forbid such a design.
Consider prime powers. When q = p¾ with p ≡ 3 (mod 4), we have q ≡
1 (mod 4) if and only if is even. Hence Theorem 13.2.9 cannot exclude any prime
power from being the order of a projective plane. In fact, we will see next that
a projective plane of order q exists whenever q is a prime power (this is why we
define the order of a plane to be one less than the size of the lines). This uses a con-
nection with Latin squares. We proved in Theorem 13.1.5 that there is a complete
family MOLS(q , q − 1) of Latin squares when q is a prime power.

13.2.10. Theorem. (Bose [1939]) If q ≥ 2, then a projective plane of order q exists

if and only if MOLS(q , q − 1) exists (such as when q is a prime power).
Proof: We first construct an orthogonal family A(1) , . . . , A(q−1) from a projec-
tive plane. In the plane, consider a line L with points x , y , w1 , . . . , wq−1 . Let
X 1 , . . . , X q be the other lines through x, let Y1 , . . . , Yq be the other lines through
y, and let W¾ ,1 , . . . , W¾ ,q be the other lines through w¾ .
Each X i and Yj have one common point; call it i , j . This uses q2 points, which
are the remaining points of the plane. We use the lines determined by i , j and
points on L to define Latin squares. Let Ai¾, j = t, where W¾ ,t is the line deter-
( )

mined by i , j and w¾ . This is well-defined, since two points determine one line.

•y
wq−1
•

•
L
Xq •w1
x
i , j •
X1

Y1 Yq

Since i , j and i , j both lie on X i , which avoids w¾ , they cannot both lie on
a line with w¾ . Thus the entries in row i of A(¾) are distinct. Similarly, column
entries are distinct, so each A(¾) is a Latin square.
For orthogonality, let ( , ) = (Ai¾, j , Ai , j ). We have i , j on W¾ , and Wl ,  .
( ) (l)

Since the two lines W , and Wl ,  have only one common point, the only position
that can yield the pair ( , ) in A and A l is (i , j).
628 Chapter 13: Combinatorial Designs

The converse reverses the construction. Begin with squares A(1) , . . . , A(q−1) .
Introduce one point w¾ for each square A¾ . Introduce q2 points { i , j }iq, j =1 . Finally,
(¾)
add two points x and y. The line W¾ ,l consists of w¾ and all i , j such that Ai , j = l.
The line X i consists of x and all i , j (first index fixed). The line Yj consists of y and
all i , j (second index fixed). Finally, the line L consists of x, y, and w1 , . . . , wq−1 .
We have defined q2 + q + 1 points and q2 + q + 1 lines, each consisting of q + 1
points. For q ≥ 2, points 1 ,1 , 1 ,2 , 2 ,1 , 2 ,2 have no three on a line. It suffices to
verify that every two points lie in one common line and every two lines have one
common point. We leave this to Exercise 4.

Finite fields of order q exist only when q is a power of a prime, yielding pro-
jective planes of such orders. Theorem 13.2.9 forbids projective planes of many
non-prime-power orders, but not orders such as 10. In light of Theorem 13.2.10,
the computer search mentioned in Remark 13.1.9 also prohibits projective planes
of order 10. It remains possible that a complete family of Latin squares (or a
projective plane) does not require a finite field of order q. No projective plane of
non-prime-power order is known, and none is believed to exist. For q ∈ , below 
we summarize the relationship of ∃ MOLS(q , q − 1) to the other two properties.
q is a prime power ⇒ ∃ MOLS(q , q − 1) ⇔ ∃ proj . plane of order q

13.2.11.* Example. Affine planes. In the picture for Theorem 13.2.10, deleting L
and its points leaves q2 points and leaves q2 + q lines, each of size q. Any two points
that remain lie on exactly one line that remains, but some lines that remain are
now “parallel”; they met in a point now deleted. The result is a (q2 , q , 1)-design
(not symmetric) whose lines group into q “parallel classes”, each of which parti-
tions the set of points. Such a design is an affine plane (see Example 13.3.23).
The (9 , 3 , 1)-design in Example 13.1.13 is the affine plane that results from
deleting a line from the projective plane of order 3. What is the affine plane ob-
tained by deleting a line from the Fano plane?

APPLICATIONS TO EXTREMAL PROBLEMS

The delicate structural properties of projective planes lead to constructions

that meet the counting bounds in some extremal problems.
The point-line incidence matrix of a projective plane yields a bipartite graph
called the incidence graph or incidence bigraph of the plane. It is an X , Y -
bigraph, where X is the set of points, Y is the set of lines, and x and y are adja-
cent when x belongs to y. Since any two lines have exactly one common point, this
graph has no 4-cycle, even though it has many edges. Since any two lines have a
common point, the diameter is only 3, although no vertex has high degree.
To obtain such a graph with diameter 2, we merge points with lines. First we
construct projective planes directly from finite fields. Since finite fields yield com-
plete orthogonal families (Theorem 13.1.5), which yield projective planes (Theo-
rem 13.2.10), such a construction must exist. The direct construction is simple
and provides a natural way to merge points and lines in the incidence graph.
Section 13.2: Projective Planes 629

13.2.12. Theorem. If q is a prime power, there is a projective plane of order q.

Proof: Put triples of elements from the field q into multiplicative classes; The
class containing a , b , c is { a , b , c: ∈ q − {0}}. The number of classes of
nonzero triples is (q3 − 1)/(q − 1), which equals q2 + q + 1. We use one copy of these
classes as the points and another as the lines, letting (a , b , c) denote the point con-
taining the triple a , b , c and [a , b , c] denote the line containing the triple a , b , c.
To define the incidence relation, put (x , y , ) ∈ [a , b , c] if and only if ax +
by + c = 0. (Geometrically, view [a , b , c] as multiples of a vector normal to a
plane through the origin, with the classes (x , y , ) being the points in the plane.)
Consider a line [a , b , c]; by symmetry, we may assume c
= 0. For each of the q2 − 1
ways to choose x and y not both 0, there is a unique solution = −(ax + by)/c.
Solutions equivalent to (x , y , ) arise q − 1 times, since multiplying x , y , by a
nonzero field element still yields a solution. Thus each line consists of q + 1 points.
The two lines [a , b , c] and [d , e , ] share any point (x , y , ) satisfying both
ax + by + c  = 0 and dx + ey +  = 0. With  = −(ax + by)/c again, the other
equation leaves one multiplicative class as a common solution. Hence two lines
have one common point. Since the incidence relation is symmetric in points and
lines, also each point is in q + 1 lines and every two points lie on one common line.
We thus have a symmetric (q2 + q + 1 , q + 1 , 1)-design. When q ≥ 2, it is a
projective plane (Theorem 13.2.7).

The construction extends to higher-dimensional projective geometries by using

multiplicative classes of nonzero (d + 1)-tuples.

13.2.13. Definition. A polarity of a projective plane is an involution  exchang-

ing points and lines having the property that point p and line l are incident
if and only if  (p) and  (l) are incident. Let V be the set of q2 + q + 1 points
of the projective plane of order q constructed above. Given a polarity  , the
polarity graph has vertex set V , with (a , b , c) and (x , y , ) adjacent if and
only if (x , y , ) ∈ [a , b , c] and  (x , y , )
= [a , b , c]. For the natural polarity
that pairs (a , b , c) with [a , b , c], the polarity graph puts (a , b , c) adjacent to
(x , y , ) if and only if these vertices lie in distinct multiplicative classes and
ax + by + c  = 0. We will consider only this polarity and this polarity graph,
shown below for orders 2 and 3.

• 100:200
012:021
001 010 •
• • 211:122
110 101 • • 011:022
• •
•
111 110:220 • • 121:212
•
• 011 001:002 • 111:222
• • 201:102
120:210
•
100 • •
112:221 101:202 •
010:020
630 Chapter 13: Combinatorial Designs

13.2.14. Example. Diameter 2, small degree, no 4-cycles. The polarity graph G

collapses the incidence graph by identifying (a , b , c) with [a , b , c] and deleting
loops. Given vertices (a , b , c) and (d , e , º ) in G, the system of two homogeneous
equations ax + by + c = 0 and dx + ey +  = 0 always has a nonzero solution
(the line [x , y , ] containing the points (a , b , c) and (d , e ,  )). Hence (x , y , ) is a
common neighbor, and diam G = 2.
The incidence graph is (q + 1)-regular. Since (x , y , ) ∈ [a , b , c] if and only if
(a , b , c) ∈ [x , y , ], vertices of G also have degree q + 1, except that one edge is lost
when a2 + b2 + c2 = 0, since we exclude loops. It is a nontrivial algebraic fact that
in a finite field of order q there are q + 1 multiplicative classes of nonzero triples
solving this equation; see Exercises 7–8. Hence G has√q + 1 vertices of degree
q. The rest have degree q + 1, which slightly exceeds n. By the Degree-Sum
Formula, G has q(q + 1)2/2 edges.
The properties of the projective plane forbid distinct multiplicative classes
as common solutions of u · x = 0 and v · x = 0. Thus no two triples u , v have two
common neighbors, and G has no 4-cycle.

13.2.15. Proposition.√ Every n-vertex graph G with diameter 2 has maximum

degree at least ⌈ n − 1⌉ , with equality for the polarity graph when n equals
q2 + q + 1 for some prime power q.
Proof: Let  = (G). From one vertex, at most  + ( − 1) others are reachable
√ Hence diameter 2 requires n ≤ 1 +  .
within two steps (see Proposition 5.4.12). 2

When n = q + q + 1, we have q < n − 1 < q + 1. Hence the polarity graph,

2

which has diameter 2, achieves equality in the bound.

The polarity graph also solves the Turán problem for 4-cycles. In Section 10.1
we defined ex (n; H) to be the maximum number of edges in an n-vertex graph not
containing H and computed ex (n; K r). The problem of computing ex (n; C4) was
raised by Erdős in 1938. Perhaps surprisingly, 4-cycles are forced much earlier
than triangles; a simple counting argument shows that ex (n; C4) ∈ O(n3/2).
√
13.2.16. Proposition. If m > 14 n(1 + 4n − 3), then every simple n-vertex graph
G with m edges contains C4 . When n = q2 + q + 1, the polarity graph has
nearly this many edges.
Proof: Let x count the pairs of incident edges in G. If G has no 4-cycles, then no
two vertices have more than one common neighbor, so ∑ (d2i ) = x ≤ (n2). Using the
average degree and the convexity of (u2) in terms of u, we have ∑ (d2i ) ≥ n(2m/n
2 ).
Combining these inequalities √ yields m(2m/n − 1) ≤ n(n − 1)/2. The quadratic
formula yields m ≤ 14 n(1 + 4n − 3).
When n = q2 + q + 1, the upper bound reduces to 12 n(q + 1). Since the polarity
graph has q2 vertices of degree q + 1 and q + 1 vertices of degree q, its√number
of edges is just (q + 1)/2 less than the bound. The difference is about 12 n out of
approximately 12 n3/2 edges.

Füredi [1996a] improved the upper bound to m ≤ 12 q(q + 1)2 when n = q2 + q +

1 and q ≥ 15. Thus when q is a prime power greater than 13, the polarity graph is
Section 13.2: Projective Planes 631

a largest graph not containing C4 . When q is an even prime power, Füredi [1983]
proved that only polarity graphs achieve the extreme; the condition for equality
is the existence of a polarity with q + 1 fixed points on the projective plane of order
q. This holds also for odd prime powers, but that case is more difficult; Füredi has
not published the proof. The case n = q2 + q for even prime powers is discussed in
Firke–Kosek–Nash–Williford [2013].
The incidence graph of a projective plane provides extremal constructions for
an analogous bipartite problem. In the context of reduced adjacency matrices,
extremal problems for complete bipartite subgraphs of K n ,n are natural.

13.2.17. Definition. Zarankiewicz’s Problem asks for the maximum number

of 1s in an m × n matrix with no constant s × t submatrix of 1s. Equivalently,
this is the maximum size of an X , Y -bigraph having no copy of K s ,t with s
vertices in X and t vertices in Y , where | X | = m and | Y | = n. Let (m , n; s , t)
denote this value, and put (n; t) = (n , n; t , t).

Zarankiewicz asked for the first few values of (n; 3). Bounds on apply to
other incidence relations. For example, in a set of n points and m circles in the
plane (with no three points collinear and no four cocircular), the maximum num-
ber of incidences between points and circles is bounded by (m , n; 3 , 2), since any
three points determine a unique circle.
The incidence-counting idea in Proposition 13.2.16 also provides an upper
bound for (m , n; s , t). First we observe the connection between (n , n; s , t) and
the corresponding extremal problem for subgraphs of K n .

13.2.18. Proposition. For s , t ≥ 1, ex (n; K s ,t) ≤ 1

2
(n , n; s , t).
Proof: Let G be a simple n-vertex graph with no K s ,t . We form a subgraph G of
K n ,n with no K s ,t (or K t ,s). Let V(G ) be the union of copies V1 and V2 of V(G).
Create edges x1 y2 and y1 x2 in G for each edge xy in G. If G contains K s ,t , then
since G has no loops the vertices in the two parts of the subgraph arise from dis-
tinct vertices of G, and G also contains K s ,t .

13.2.19. Theorem. (K ővári–Sós–Turán [1954]) For s , t > 1,

(m , n; s , t) < (s − 1)1/t(n − t + 1)m1−1/t + (t − 1)m.
ex (n; K s ,t) < 12 (s − 1)1/t n2−1/t + t−21 n.

Proof: In a K s ,t -free X , Y -bigraph G with | X | = m and | Y | = n, each set of t

vertices in Y has at most s − 1 common neighbors in X . Hence there are at most
(s − 1)(nt) choices of a vertex x ∈ X and a t-set T ⊆ Y such that T ⊆ N(x). Since
we obtain such a pair for every choice of t vertices from the neighborhood of any
x ∈ X , we have ∑ x∈ X (d(x)
t
) ≤ (s − 1)(nt).
Let = | E(G)|. As in Proposition 13.2.16, we obtain ∑ (d(x) t
) ≥ m( /m t
) by
y
replacing each degree with the average degree /m, since ( t ) is a convex function
of y. Multiply both sides by t!, let = /m < n, and use n > n− − t+1
−1 > · · · > n− t+1 ,
1

which holds when < n. This yields m( m − t + 1) < (s − 1)(n − t + 1) , which
t t

simplifies to the bound claimed.

For the bound on ex (n; K s ,t), apply Proposition 13.2.18.
632 Chapter 13: Combinatorial Designs

The better bound arises by naming the parameters so that t ≤ s. Znám [1965]
improved the leading coefficient when s = t, obtaining
(n; t) < (t − 1)1/t n2−1/t + (t − 1)n/2
(see Bollobás [1978, p. 311]). Guy [1968] collected and enhanced early results
on the problem. Roman [1975] extended the counting arguments for the upper
bound on (m , n; s , t) to a family of bounds, using a parameter , each bound op-
timal for infinitely many pairs (m , n). In particular, for any at least t − 1 he
proved (m , n; s , t) ≤ (s − 1)(nt)/ (t−¾1 ) + ( + 1)(t − 1)m/t.
Füredi further improved the leading coefficient for ex (n; K s ,t). This yields
asymptotic optimality of the construction for s = t = 3 we will describe.

13.2.20. Theorem. (Füredi[1996b]) ex (n; K s ,t) ≤ 12 (s − t + 1)1/t n2−1/t + o(n2−1/t).

t ) ≤
Equality in the counting bound ∑ x∈ X (d(x) (s − 1)(nt) requires each t-set
in Y to have exactly s − 1 common neighbors in X and all vertices in X to have
the same degree d. When t = 2, these are the conditions for the neighborhoods
of vertices in X to form an (n , d , s − 1)-design on Y . Brown [1966] used finite
geometries to show that the counting bound is asymptotically optimal for (n; 3),
but the method does not handle t > 3. The upper bound on ex (n; C4) below was
proved earlier by Reiman [1958].

13.2.21. Theorem. (Erdős–Rényi–Sós [1966], Brown [1966]) If q is a power of an

odd prime, then (q2 + q + 1; 2) ≥ (q2 + q + 1)(q + 1) and ex (q2 + q + 1; C4) ≥
2 q(q + 1) . Thus when n is sufficiently large,
1 2

√
n3/2 − n4/3 ≤ (n; 2) ≤ 12 n(1 + 4n − 3) ,
√
1 (n3/2
2 − n4/3) ≤ ex (n; C4) ≤ 14 n(1 + 4n − 3) .
Proof: The incidence graph and polarity graph of the projective plane yield the
lower bounds in terms of q. The second statement uses the number-theoretic re-
sult that if m is sufficiently large, then there is a prime between m − 101
m2/3 and
√
m. Applying this to m = 12 ( 4n − 3 − 1) and restricting the projective plane con-
struction for this prime to a subset of its vertices yields the bound.

Füredi [1996c] also generalized the polarity construction to ex (n; K 2 ,s).

13.2.22.* Example. Large graphs with no K 2 ,s+1 (Füredi [1996c]). When q is a

prime power and n = (q2 − 1)/s, we use q to construct an n-vertex graph G with
√
about 12 sn3/2 edges and no copy of K 2 ,s+1 .
The vertices of G are the equivalence classes of a relation on the nonzero or-
dered pairs of elements of q . Since s divides q2 − 1, there is an element  in q
with multiplicative order s. The s elements of S = {1 , w , . . . , ws−1 } are distinct,
and ws = 1. We make (a , b) equivalent to (c , d) if c = wi a and d = wi b for some i.
This defines an equivalence relation (Exercise 14), and the classes all have size s.
Let ⟨a , b⟩ denote the class containing (a , b). Define G by ⟨a , b⟩ ↔ ⟨u , v⟩ if
au + bv ∈ S; this is symmetric. If au + bv = wi , then (w j a)u + (w j b)v = wi+ j , so the
relation is consistent on equivalence classes. The graph K 2 ,s+1 is not a subgraph
of G, and the vertices of G have degree q − 1 or q (Exercise 14).
Section 13.2: Projective Planes 633

13.2.23.* Theorem. ex (n , K 3 ,3) = ( 12 + o(1))n5/3 .

Proof: (sketch) The asymptotic statements again use density results about prime
powers. The upper bound is from Theorem 13.2.20 (Füredi [1996b]).
The lower bound is due to Brown [1966]. Given a field of order q, where q ≡
−1 (mod 4), we create a graph whose vertices are the triples of field elements. We
put xy ∈ E(G) if and only if ∑i=1 (x i − yi)2 = 1.
3

The idea is that the points at distance 1 from a given point form a sphere. A
copy of K 3 ,3 would correspond to three spheres having three common points. In
such a three-dimensional space, it is not possible for three spheres to have three
common points (visualize this in 3).
For each solution to ∑i=1 a2i = 1, one can solve x i − yi = ai in q ways. Since
3

there are n = q3 vertices, it thus suffices to show that there are about 12 q2 solu-
tions in the field to ∑i=1 a12 = 1.
3

The gap between the bounds on (n; t) is large when t > 3. The current best
general lower bound arises from a simple non-constructive counting argument. A
similar argument gives a lower bound for ex (n; K t ,t).
s −1
13.2.24. Theorem. If = st−1 and  = stt−−11 , then
(m , n; s , t) ≥ ⌊(1 − 1
s!t!
)m1− n1− ⌋ .

Proof: Let  = ⌊ m1− n1− ⌋ . We count the copies of K s ,t over all subgraphs of K m ,n
−1
with  edges; let the total be T . Some such graph then has at most T (mn  ) copies
of K s ,t . Deleting one edge from each copy of K s ,t in this graph yields a K s ,t -free
−1
graph with at least  − T (mn  ) edges.
By summing the number of graphs containing each possible copy of K s ,t , we
− st mn(mn−1)···(mn− st+1) mn− st
obtain T = (ms)(nt)(mn − st ). Since (  ) =
mn
(−1)···(−t+1) ( − st ), the average
m n 
is bounded by ( s )( t )( mn) , which in turn is bounded by s!t!
st 1
ms− st nt− st . Hence
(m , n; s , t) ≥ e − s!t! m1− n1− .
1

For s much larger than t, the exponents in Theorems 13.2.19–13.2.20 are

sharp, by explicit constructions. Further improvements are in Alon–Ronyai–
Szabó [1999] and Ball–Pepe [2012, 2016].

13.2.25.* Theorem. (Kollár–Rónyai–Szabó [1996]) For t ≥ 4 and s > t!,

ex (n; K s ,t) = (n2−1/t).
Proof: (sketch) When n has the form qt with q a prime power, we construct a
graph whose vertices are the elements of the field of size qt . We put x , y adjacent
t−1
if N(x + y) = 1, where N() =  ·  q · · ·  q . Here N() is called the norm of  , and
the resulting graph G is the norm graph.
Except for N(0) = 0, N() is distributed equally over a subfield with q ele-
ments. Thus each vertex in the graph has degree (qt − 1)/(q − 1), except that x has
degree one less when N(2x) = 1 (since we don’t include loops). We thus obtain
1 t qt − 1 1 1
| E(G)| ≥ q ( − 1) ≥ q2t−1 = n2−1/t .
2 q−1 2 2
634 Chapter 13: Combinatorial Designs

The norm graph contains K s ,t if and only if it has t vertices 1 , . . . , t with s

common neighbors. Common neighbors are simultaneous solutions to N(x + i) =
1 for 1 ≤ i ≤ t. Kollár–Rónyai–Szabó [1996] proved that there are at most t!
solutions by proving this bound in the more general setting where there are t
equations of the form ∏ j =1 (x j − ai , j ) = bi over a field K , and the coefficients are
t

distinct in each column (that is, ai , j = ai , j if and only if i = i ).

DIFFERENCE SETS

Some projective planes and some more general symmetric designs have the
simple structure that the blocks are cyclic translates of a single block. A familiar
example is the usual description of the Fano plane, in which the blocks are cyclic
translates of {0 , 1 , 3} modulo 7.
Here we describe a way of obtaining such a description. It is known that not
all symmetric (v , ¾ , ë)-designs arise in this way. In fact, not all projective planes
of order 16 can be expressed cyclically.

13.2.26. Definition. A set {a1 , . . . , a¾ } is a (v , ¾ , ë)-difference set if for every

d ∈ v − {0}, there are ë ordered pairs (ai , aj) such that d ≡ aj − ai (mod v).

Since there are ¾(¾ − 1) ordered pairs of elements in a ¾-set, the definition
requires ë(v − 1) = ¾(¾ − 1), which is a familiar necessary condition for symmetric
designs (Proposition 13.1.14).

13.2.27. Example. The set {0 , 1 , 3} is a (7 , 3 , 1)-difference set. We have 1 = 1 − 0,

2 = 3 − 1, 3 = 3 − 0, 4 = 0 − 3, 5 = 1 − 3, 6 = 0 − 1. The next result now implies
that the Fano plane is a symmetric (7 , 3 , 1)-design.

13.2.28. Definition. Given a set D ⊆ v , the translate D + x (modulo v) is the

set {a + x: a ∈ D}, with entries reduced modulo v.

13.2.29. Proposition. A set D of size ¾ is a (v , ¾ , ë)-difference set if and only if

the translates of D modulo v form (the blocks of) a symmetric (v , ¾ , ë)-design.
Proof: Let D be a (v , ¾ , ë)-difference set. If a − a ≡ t (mod v) for a , a ∈ D, then
the translates of D provide all pairs r, s ∈ v such that s − r ≡ t (mod v). Further-
more, there one such set of v pairs (r, s) for each pair in D with difference t, so
each pair in v arises exactly ë times.
Conversely, if the translates D + x form a (v , ¾ , ë)-design, then the ë trans-
lates in which {r, s} occurs yield ë ordered pairs (i , j) such that aj − ai ≡ (s −
r) (mod v). Because ë(v − 1) = ¾(¾ − 1), this accounts for all ordered pairs from D,
and D is a difference set.

Thus, difference sets correspond to cyclically invariant designs. We next dis-

cuss a technique to construct a (v , ¾ , ë)-difference set or show that none exists.

13.2.30. Definition. A multiplier of a (v , ¾ , ë)-difference set D is an element t

of v such that tD is a translate of D. If tD = D, then t fixes D.
Section 13.2: Projective Planes 635

13.2.31. Proposition. All translates of D have the same multipliers.

Proof: If t is a multiplier of D, then tD = D + x for some translate D + x. For
any translate D + y, we have t(D + y) = (tD) + ty = (D + x) + ty. Since x + ty =
y + (x + (t − 1)y), we conclude that t(D + y) is a translate of D + y, making t also
a multiplier of D + y.

13.2.32. Proposition. If gcd(v , ¾) = 1 and D is a (v , ¾ , )-difference set, then

some translate of D is fixed by every multiplier of D.
Proof: Let D = {a1 , . . . , a¾ }. Let  (S) denote the sum (modulo v) of a set S. For
any translate D + x, we have  (D + x) =  (D) +  x. Since gcd(v , ) = 1, multipli-
cation by  permutes the nonzero congruence classes, and hence there is exactly
one choice of x such that  (D + x) ≡ 0 (mod v).
If t is a multiplier of D, then also t is a multiplier of this translate D + x, and
t(D + x) = D + y. With x chosen in this way,  (D + y) =  (t(D + x)) = t  (D + x) ≡
0 (mod v). Since only one translate of D sums to a multiple of v, we have y = x,
and t fixes D + x.

We will show that every prime p that exceeds and divides  − but not v
is a multiplier of every (v ,  , )-difference set. This statement is the Multiplier
Theorem. Knowing that a number p is a multiplier helps to find a difference set
or forbid its existence, because some difference set must consist of complete orbits
modulo v under multiplication by p.

13.2.33. Example. Multipliers and difference sets. Given a triple (v ,  , ), we

seek a (v ,  , )-difference set in v .
(7 , 3 , 1). By the Multiplier Theorem, 2 is a multiplier. The orbits under mul-
tiplication by 2 are {0}, {1 , 2 , 4}, {3 , 6 , 5}. Each of these triples is a difference
set. In general, the lines of a projective plane of order q form a (q2 + q + 1 , q + 1 ,1)-
design. If such a design is generated by a difference set, then by the Multiplier
Theorem every prime dividing q is a multiplier of that difference set. For exam-
ple, when q = 3 the orbits modulo 13 are {0}, {1 , 3 , 9}, {2 , 6 , 5}, {4 , 12 , 10}. To
form a set of size 4 we combine two orbits: {0 , 1 , 3 , 9} is a difference set.
(11 , 5 , 2). By the Multiplier Theorem, 3 is always a multiplier. The orbits
are {0}, {1 , 3 , 9 , 5 , 4}, {2 , 6 , 7 , 10 , 8}. Sure enough, {1 , 3 , 4 , 5 , 9} is a difference
set, with every difference appearing twice.
(31 , 10 , 3). By the Multiplier Theorem, 7 is a multiplier of every difference
set. Under multiplication by 7, orbits in 31 have sizes 1, 15, and 15. No combi-
nation of these has size 10, so there is no (31 , 10 , 3)-difference set.
(9 , 3 , 1). By the Multiplier Theorem, 2 is a multiplier of every difference set.
The orbits are {0}, {1 , 2 , 4 , 8 , 7 , 5} and {3 , 6}. The union {0 , 3 , 6} is the only
union of size 3, but it is not a difference set. (Of course, here ( − 1) = 6
= 8 =
(v − 1), so this triple already fails the elementary counting condition.)

As shown by the last part of Example 13.2.33, not every union of orbits with
the desired size is a difference set (see also Exercise 16).
The remainder of this section is the proof of the Multiplier Theorem and can
be skipped without loss of continuity. We note only that although the proof ap-
636 Chapter 13: Combinatorial Designs

plies only when the specified prime p satisfies p > ë, it is conjectured that this
restriction is unnecessary for the conclusion.
We begin by defining a polynomial from a difference set, using the elements
¾
as exponents of monomials. Throughout this discussion, we let R(x) = ∑i=1 x d i
v−1
and Q(x) = ∑i=0 x i ; note that Q records all the congruence classes as exponents.

¾
13.2.34. Lemma. Given D = {d1 , . . . , d¾ }, let R(x) = ∑i=1 x di , and let Q(x) =
v−1
∑i=0 x i . If D is a (v , ¾ , ë)-difference set, then
R(x)R(x−1) ≡ ¾ − ë + ë Q(x) (mod (xv − 1)).

Proof: By definition, R(x)R(x−1) = ¾ + ∑ i

= j x d i −d j . Since D is a difference set, each
nonzero congruence class modulo v occurs exactly ë times as d i − d j . By treating
v−1
xv as equivalent to x0 , we thus obtain R(x)R(x−1) ≡ ¾ + ë(∑ i=1 x i) (mod (xv − 1)).
To introduce Q(x), we add and subtract ë.

v−1
13.2.35. Lemma. Let Q(x) = ∑i=0 x i . If º is a polynomial, then
º (x)Q(x) ≡ º (1)Q(x) (mod (xv − 1)).

Proof: Working modulo x v − 1, we have xv ≡ 1, so xQ(x) ≡ Q(x). Using induction

on j , we have x j Q(x) ≡ Q(x) (mod (xv − 1)). For º (x) = ∑i=0 ci x i , we compute
r

r r
º (x)Q(x) = ∑ ci x i Q(x) ≡ ∑ ci Q(x) ≡ º (1)Q(x) (mod (xv − 1)).
i=0 i=0

These lemmas yield a sufficient condition for multipliers.

13.2.36. Lemma. Let D be a (v , ¾ , ë)-difference set, and let n = ¾ − ë. If

R(x p)R(x−1) ≡ nx s + ë Q(x) (mod (xv − 1)) for some p , s ∈ v , then pD = D + s.
Proof: All congruence expressions are modulo xv − 1. Multiplying the assumed
congruence by R(x) and applying Lemmas 13.2.34–13.2.35 yields
R(x p)(n + ë Q(x)) ≡ nx s R(x) + R(1)ë Q(x). (∗)
Viewing R(x as a polynomial in x and using 1 = 1, Lemma 13.2.35 yields
p) p

R(x p)Q(x) ≡ R(1)Q(x). Subtracting R(1)ë Q(x) from both sides of (∗) now yields
nR(x p) ≡ nx s R(x). This is just the statement that the congruence classes in pD
are the same as the congruence classes in D + s.

13.2.37. Theorem. (Multiplier Theorem) Let D be a (v , ¾ , ë)-difference set.

Any prime p that exceeds ë and divides ¾ − ë but not v is a multiplier of D.
Proof*: All congruences are modulo x v − 1; exponents are elements of v . To apply
Lemma 13.2.36, we seek s such that R(x p)R(x−1) ≡ nx s + ë Q(x), where n = ¾ − ë.
In pD, the differences are p times the differences in D. Since p is relatively
prime to v, this multiplication just permutes the congruence classes, and pD is a
difference set. By Lemma 13.2.34, n + ë Q(x) is thus congruent to both R(x)R(x−1)
and R(x p)R(x− p). In particular,
R(x)R(x−1)R(x p)R(x− p) ≡ [n + ë Q(x)]2 . (1)
Section 13.2: Projective Planes 637

We will obtain another expression for this product to study R(x p)R(x−1).
Let m = n/p. Since Q(x) divides xv − 1, we can write R(x)R(x−1) ≡ n + ë Q(x)
as R(x)R(x−1) = pm + Q(x)A(x) for some polynomial A. Multiplying by [R(x)] p−1
yields [R(x)] p R(x−1) = pB(x) + Q(x)C(x) for some polynomials B and C.
Since p is prime, p is a factor of all multinomial coefficients in the expan-
sion of [R(x)] p except when we take the same monomial from each factor. Thus
[R(x)] p = R(x p) + pE(x) for some polynomial E. We now have
R(x p)R(x−1) = p[B(x) − E(x)R(x−1)] + Q(x)C(x).
We reduce this modulo xv − 1. Since x− j ≡ xv− j , we have B(x) − E(x)R(x−1) ≡ F(x)
for some polynomial F . Applying Lemma 13.2.35 to C(x) yields
R(x p)R(x−1) ≡ pF(x) + Q(x)C(1). (2)
When setting x = 1, adding a multiple of x − 1 changes nothing, and congru-
v

ence becomes equality. Since R(x) has | D| terms with coefficient 1, setting x = 1
in (2) thus yields ¾ 2 = pF(1) + vC(1). Since ë(v − 1) = ¾(¾ − 1), we have ¾ 2 = n + ë v.
Since p divides n, we conclude that vC(1) − ë v is divisible by p. Since v and p are
relatively prime, p divides C(1) − ë.
Thus we can set C(1) = ë + pt in (2) for some integer t. Now
R(x p)R(x−1) ≡ pF(x) + ptQ(x) + ë Q(x) ≡ pG(x) + ë Q(x)
for some polynomial G. Setting x = 1 yields ¾ 2 = pG(1) + ë v, and thus pG(1) = n.
Since Q(x−1) ≡ Q(x), evaluating at x−1 yields R(x− p)R(x) ≡ pG(x−1) + ë Q(x). We
have now obtained our second expression for the four-way product:
R(x p)R(x−1)R(x− p)R(x) ≡ (pG(x) + ë Q(x))(pG(x−1) + ë Q(x)). (3)
Since pG(x) and pG(x−1) are both congruent to polynomials in x, Lemma
13.2.35 yields pG(x)Q(x) ≡ pG(1)Q(x) = nQ(x), and similarly pG(x−1)Q(x) ≡
nQ(x). Combining (1) with (3) now yields
p2 G(x)G(x−1) + 2 ë nQ(x) + ë2 Q2(x) ≡ n2 + 2 ë nQ(x) + ë2 Q2(x),
which simplifies to
p2 G(x)G(x−1) ≡ n2 . (4)
Recall that R(x p)R(x−1) ≡ pG(x) + ë Q(x). Our aim, allowing us to apply
Lemma 13.2.36, is to show that G(x) = ax s for some constants a and s with s ≥ 0.
Since pG(1) = n, this will yield a = m and complete the proof.
In the product R(x p)R(x−1), all coefficients are nonnegative. The coefficients
do not become negative when we reduce exponents to at most v − 1 by applying
xv ≡ 1. Do this with both R(x p)R(x−1) and G(x), obtaining R (x) = pG (x) + ë Q(x)
(note that Q(x) does not change). Since each coefficient in ë Q(x) is ë and R (x) ≡
pG (x) + ë Q(x), each coefficient in R (x) is congruent to ë modulo p. Since p > ë
and each coefficient is nonnegative, each coefficient is at least ë. Therefore, the
polynomial pG(x) congruent to R(x p)R(x−1) − ë Q(x) has nonnegative coefficients.
Since G(x) has nonnegative coefficients, the expression p2 G(x)G(x−1) will
have a nonconstant term if G(x) has more than one nonzero term. This would
contradict (4). Also (4) implies that G is nonzero. Thus we have shown that G has
one monomial term, as desired.
638 Chapter 13: Combinatorial Designs

The Multiplier Theorem was proved by Hall [1947] for cyclic difference sets
yielding projective planes. Hall–Ryser [1951] extended it for ë > 1. Extensions
appear in McFarland–Mann [1965], Baumert [1971], and Lander [1983].

EXERCISES 13.2

13.2.1. (−) Prove that every projective plane with seven points is isomorphic to the Fano
plane, in the sense that renaming elements and permuting lines can turn it into the Fano
plane as described in Example 13.1.12.
13.2.2. (−) Prove that the Heawood graph (shown below) is the incidence bigraph of the
Fano plane.
• •
• •
• •
• •
• •
• •
• •

13.2.3. (−) Check that {1 , 2 , 5 , 15 , 17} is a (21 , 5 , 1)-difference set and find a translate
that is fixed by every multiplier.
13.2.4. Complete the proof that the construction in Theorem 13.2.10 of points and lines
from a complete orthogonal family produces a projective plane.
13.2.5. A hypergraph is ¾-colorable if its vertices can be partitioned into ¾ sets contain-
ing no edge. Prove that a hypergraph whose edges form the set of lines of a projective plane
of order q is 2-colorable if and only if q > 2.
13.2.6. The transversal number of a hypergraph is the minimum size of a set of vertices
that intersects every edge. Let q be a prime power. Let H be the hypergraph with q2 + q + 1
vertices whose edges are the lines of a projective plane of order q on the vertex set. Let H
be the hypergraph with the same vertex set whose edges are the complements of the edges
in H . Determine the transversal numbers of H and H .
13.2.7. Let q be a prime power congruent to 0 or 1 modulo 4. Prove that a2 + b2 + c2 = 0
has exactly q2 solutions with a , b , c ∈ q , and conclude that there are exactly q + 1 multi-
plicative classes of nonzero solutions. (Hint: When q ≡ 1 (mod 4), show that there exists
j ∈ q with j 2 = −1. Use j to reduce the problem to finding solutions of uv = −c2 .)
13.2.8. Let q be a prime power with q ≡ 3 (mod 4). Let S0 and S1 be the set of nonzero
squares and the set of non-squares in q . Note that −1 ∈ S1 , and hence x ∈ S0 if and only
if − x ∈ S1 for x
= 0. For i , j ∈ {0 , 1}, let Ti , j = {x ∈ Si : x + 1 ∈ Sj } and ti , j = | Ti , j |.
(a) Prove that t0 ,0 = t1 ,0 = t1 ,1 = (q − 3)/4 and x0 ,1 = (q + 1)/4. (Hint: Prove t1 ,0 = t1 ,1
by showing that the map taking x to 1/x is a bijection from T1 ,0 to T1 ,1 .)
(b) For each multiplicative class of nonzero triples solving a2 + b2 + c2 = 0, take the
representative whose
√ first nonzero coordinate is 1. Prove that each such triple has the
√
form (1 , ± x , ± − x − 1) for some x ∈ T0 ,1 . Use this to conclude that there are exactly
q + 1 multiplicative classes of nonzero solutions. (Comment: This approach also yields
q + 1 classes when q ≡ 1 (mod 4), where the values of ti , j are different.)
Exercises for Section 13.2 639

13.2.9. (♦) A dominating set in a graph is a vertex subset S such that every vertex
outside S has a neighbor in S. Determine the minimum size of a dominating set in the in-
cidence bigraph of a projective plane of order q. (Comment: A total dominating set in
G is a set that contains a neighbor of every vertex; it is a transversal in the hypergraph
whose edges are the neighborhoods in G; see Henning–Yeo [2013b].)
13.2.10. Zarankiewicz problem for forbidden K 2 ,t . Let G be an n-vertex graph.
(a) Prove that if G is simple and ∑v∈V(G) (d(v)
2 )
> (t − 1)(2n) , then G contains K 2 ,t .
(b) Prove that ∑v∈V(G) ( d(v)
2
) ≥ m(2m/n − 1), where G has m edges.
(c) Use (a) and (b) to prove K 2 ,t ⊆ G when m > 12 (t − 1)1/2 n3/2 + n/4.
(d) Application: Given n distinct points in the plane, prove that the distance is exactly
1 for at most √1 n3/2 + n/4 pairs. (Bondy–Murty [1976, pp. 111–112])
2

13.2.11. (♦) Prove that every ¾-regular graph with girth 6 has at least 2 ¾ 2 − 2 ¾ + 2 ver-
tices. For ¾ ≥ 3, prove that some ¾-regular graph of girth 6 has 2 ¾ 2 − 2 ¾ + 2 vertices if and
only if there is a projective plane of order ¾ − 1. (Karteszi [1960], Singleton [1966])
13.2.12. Use Theorem 13.1.23 to prove that there is no symmetric (29 , 8 , 2)-design. Other
triples (v , ¾ , ) excluded by such arguments are (43 ,7 ,1), (22 ,7 ,2), (46 ,10 ,2), (67 ,12 ,2),
(92 ,14 ,2), (106 ,15 ,2), (137 ,17 ,2), (53 ,13 ,3), (103 ,18 ,3), (34 ,12 ,4), (43 ,15 ,5), (72 ,20 ,5).
13.2.13. (♦) The degeneracy  (G) of a graph G is defined by  (G) = max √ H⊆ G (H).
¾
(a) Let {G1 , . . . , G ¾ } be a decomposition of K n . Prove ∑i=1  (G i) ≤  n. (Hint: Each
G i has a subgraph with minimum degree  (G i).)
(b) Prove that part (a) is almost sharp, as follows: When q is a power of a prime,  =
q2 + q + 1, and n =√m for some integer m, construct a decomposition {G1 , . . . , G ¾ } such
¾
that ∑i=1  (G i) ≥ (  − 1)n. (Füredi–Kostochka–Škrekovski–Stiebitz–West [2005])
13.2.14. Prove that the construction in Example 13.2.22 produces a (q2 − 1)/s-vertex graph
not containing K 2 ,s+1 , with all vertices having degree q or q − 1. (Füredi [1996c])
13.2.15. A (v ,  , )-difference family in v is a family of -subsets of v whose differ-
ences cover each nonzero element of v exactly times. Proposition 13.2.29 generalizes
immediately to show that the translates modulo v of the sets in a (v ,  , )-difference fam-
ily form a (v ,  , )-design. Show that the translates modulo 41 of {0 , 1 , 4 , 11 , 29} and
{0 , 5 , 14 , 20 , 22} form a (41 , 5 , 1)-design. (Hanani [1975])
13.2.16. (♦) Applications of the Multiplier Theorem.
(a) Use the Multiplier Theorem to obtain a difference set that generates a projective
plane of order 4.
(b) Use the Multiplier Theorem to obtain a difference set that generates a projective
plane of order 5. (Comment: Not all unions of orbits of the desired size are difference sets.)
13.2.17. Use the Multiplier Theorem to obtain a (37 , 9 , 2)-difference set and a (73 , 9 , 1)-
difference set.
13.2.18. Use multipliers to show that there is no (56 , 11 , 2)-difference set.
13.2.19. Show that every (n2 + n + 1 , n + 1 , 1)-difference set has n as a multiplier.
13.2.20. (♦) Prove that there is no (111 , 11 , 1)-difference set. (Hint: Without generating
orbits of multipliers, use two multipliers to restrict the set of values that can appear in
a difference set fixed under all multipliers. Comment: The impossibility of a (111 , 11 , 1)-
difference set does not (yet) prohibit a projective plane of order 10.)
13.2.21. Prove that there is no (n2 + n + 1 , n + 1 , 1)-difference set when n is divisible by 14
or 15 or 21. (Comment: For n ≤ 3600, it is known that an (n2 + n + 1 , n + 1 , 1)-difference
set exists only when n is a prime power (van Lint–Wilson [1992, p. 348]).)
640 Chapter 13: Combinatorial Designs

13.3. Further Constructions

Beyond Latin squares, projective planes, and Hadamard matrices, design
theorists study many other structures. Decomposing an object into isomorphic
copies of a smaller object has the flavor of and generalizes design theory.

13.3.1.* Remark. Decomposition into isomorphic copies of a graph or hypergraph

F. A (v , ¾ , 1)-design decomposes K v into copies of K ¾ . More generally, when does
a graph G decompose into copies of a given graph F ? Obviously | E(F)| must divide
| E(G)| , and the greatest common divisor of the vertex degrees in F must divide
each vertex degree in G. A graph G is F-divisible if these conditions hold.
A deep and difficult theorem of Wilson [1975] states that when v is sufficiently
large, F-divisibility of K v suffices for K v to decompose into copies of F . When F =
K ¾ these are the familiar necessary conditions (¾ − 1) | (v − 1) and ¾(¾ − 1) | v(v − 1)
for a (v , ¾ , 1)-design. For F = K 3 (triple systems) the necessary conditions are
always sufficient, but in general the threshold on v for sufficiency is unknown.
The problem generalizes to hypergraphs. A t-design is a ¾-uniform hyper-
graph on n points such that each t-set of points appears together in exactly ë
edges. Setting t = 2 yields the classical (n , ¾ , ë)-designs that we have discussed.
For ë = 1, a t-design is called a Steiner system and must have exactly (nt)/(¾t )
blocks, since each ¾-set provides exactly (¾t ) of the t-sets. It was conjectured in the
mid-19th century that t-designs exist for all ¾ and t when n is sufficiently large
and satisfies the divisibility conditions that (¾t−−ii) divides ë(nt−−ii) for 0 ≤ i < t.
Wilson’s Theorem mentioned above is the case t = 2. Its proof was alge-
braic. In a stunning breakthrough, Keevash [2014+] used “randomized algebraic
constructions” to prove the conjecture for general t (an exposition of a special case
appears in Keevash [2018]). Stronger and more general results were obtained in
Glock–K ühn–Lo–Osthus [2016] by the method of “iterative absorption”.
Generalizing beyond decomposition of complete graphs, Nash-Williams [1970]
conjectured that when n is sufficiently large, every K 3-divisible graph with min-
imum degree at least 3n/4 decomposes into triangles; Exercise 5.3.51 shows that
the degree threshold would be sharp. Gustavsson [1991] proved that for every
graph F there is a fraction c (perhaps only slightly less than 1) and a threshold
n0 such that when n ≥ n0 every n-vertex F-divisible graph with minimum de-
gree at least cn decomposes into copies of F . For F = K 3 , the value c = .9 suffices
(Dross [2016]). Results of Barber–K ühn–Lo–Osthus [2016] and those authors
with Montgomery in [2017] improved the values of c and n0 for general graphs
and provided a purely combinatorial proof of Wilson’s Theorem.

Wilson’s Theorem, t-designs, and the general theory of isomorphic decom-

position are well beyond the scope of this book. Hence we generally stick with
decomposition of complete graphs and t = 2 (but see t = 3 in Exercise 28). We con-
struct triple systems, we construct decompositions into isomorphic 2-factors, and
we construct orthogonal Latin squares to disprove the Euler Conjecture.
Section 13.3: Further Constructions 641

STEINER TRIPLE SYSTEMS

For a (v , ¾ , )-design with = 1, Fisher ’s Inequality b ≥ v (Theorem 13.1.15)

yields  ≤ r and thus ( − 1) ≤ v − 1, by Proposition 13.1.14. Thus projective
planes are the designs with the largest possible blocks, given v. At the other end
are designs with  = 3 (those with  = 2 are trivial). A 3-uniform hypergraph is
also called a triple system; these were the first objects studied in design theory.

13.3.2. Definition. A Steiner triple system on v elements, denoted STS(v), is

a (v , 3 , 1)-design.

By Wilson’s Theorem, the obvious necessary conditions for existence of

(v , 3 , 1)-designs are sufficient when v is sufficiently large. Fortunately, triple sys-
tems are simple enough to be analyzed without this deep result. Steiner triple
systems exist for all values of v that satisfy the necessary conditions, which re-
quire v to be congruent to 1 or 3 modulo 6.
Steiner triple systems were introduced by Woolhouse [1844], who asked for
which v they exist. K irkman [1847] solved the problem, but it seems no one no-
ticed. Steiner [1853] later publicized the notion and conjectured that the obvious
necessary conditions are sufficient, but he gave no construction. The construc-
tions by K irkman [1847] and Moore [1893] were inductive. We present more re-
cent constructions by Bose [1939] and Skolem [1958] of explicit Steiner triple
systems for the two congruence classes. Wilson [1974] gave another construction
(see Lindner–Rodger [1997, 27–31]). For the history and additional material on
triple systems, see Colbourn–Rosa [1999].
The existence problem for (v ,  , 1)-designs was completely solved in the cases
 = 4 and  = 5 by Hanani [1972]. Wilson’s Theorem does not apply for small v;
there remain about 30 values of v (between 51 and 801) for which the existence of
a (v , 6 , 1)-design is unknown.

13.3.3. Proposition. If an STS(v) exists, then v ∈ {1 , 3} (mod 6).

Proof: Each element in a (v , 3 , 1)-design appears in v−2 1 blocks, and then there
are 13 (2v) blocks of size 3 (these are just special cases of the usual relationships in
Proposition 13.1.14). Since the numbers of occurrences and blocks are integers,
v is odd, and v or v − 1 is a multiple of 3.

13.3.4. Example. STS(2 l − 1). The Fano plane is an STS(7). When v = 2 l − 1, we

can form a Steiner triple system on the set of nonzero binary vectors of length l by
letting distinct x and y form a triple with the vector  such that  = x + y (mod 2).
Since also x = y +  and y = x +  , these triples are well-defined, and each pair
belongs to exactly one triple.
This construction can also be described iteratively. The construction for l
contains the construction for l − 1 on the set of vectors having 0 in the first coor-
dinate. The sum of any two vectors with 1 in the first coordinate is in this set, so
the remaining triples consist of one vector starting with 0 and two vectors start-
ing with 1.
642 Chapter 13: Combinatorial Designs

Iterative constructions build a design by adding to a smaller design. We will

use the iterative approach to disprove the Euler Conjecture. To construct triple
systems explicitly, however, we use the algebraic approach. Our presentation is
based on Lindner–Rodger [1997, p. 1–14].

13.3.5. Definition. A quasigroup is a set Q equipped with a binary operation ◦

such that for all a , b ∈ Q, both a ◦ x = b and y ◦ a = b have unique solutions.
A quasigroup is idempotent if a ◦ a = a for all a.

We consider only finite quasigroups. Applying a via ◦ (on the right or the
left) permutes Q. Equivalently, the matrix recording ◦ by putting x ◦ y in the
row indexed by x and the column indexed by y is a Latin square with elements
Q. Hence quasigroups are equivalent to Latin squares. Quasigroup terminol-
ogy facilitates algebraic construction and the notation i ◦ j for the (i , j)-entry in
a Latin square. A quasigroup is commutative if and only if the corresponding
Latin square is symmetric. The lemma has a nice application to total coloring of
complete graphs (Exercise 5).

13.3.6. Lemma. There is an idempotent commutative quasigroup of order v if

and only if v is odd.
Proof: When v is odd, use v for the elements, and index the rows and colums
from 0 to v − 1. Since v is odd, the element 2 has a multiplicative inverse. In
position (i , j), put the congruence class of (i + j)/2. Since adding a constant or di-
viding by 2 permutes j , the elements are distinct in each row or column. Also
the formula is symmetric in i and j , and (i + i)/2 = i. Hence we have an idempotent
commutative quasigroup; for v = 5 this yields the Latin square below.
In the Latin square table of an idempotent commutative quasigroup, idem-
potence requires each symbol to be used once on the diagonal, and commutativity
requires an even number of appearances off the diagonal. Hence each symbol is
used an odd number of times, including once in each row, so v is odd.

0 3 1 4 2
3 1 4 2 0
1 4 2 0 3
4 2 0 3 1
2 0 3 1 4

13.3.7. Example. The Bose Construction. When v = 6n + 3, let the set of

elements consist of three copies of the elements of an idempotent commutative
quasigroup Q of order 2n + 1, indicating the three copies by subscripts modulo 3.
We use 2n + 1 Type 1 triples of the form (i0 , i1 , i2) for i ∈ Q. The Type 2 triples
have the form (i¾ , j¾ , (i ◦ j)¾+1 ), where i , j ∈ Q with i
= j and ¾ ∈ 3 . Below we
illustrate a Type 1 triple and two Type 2 triples.

• • • • • • •0
• • • • • • •1 3

• • • • • • •2
7
Section 13.3: Further Constructions 643


13.3.8. Theorem. (Bose [1939]) For n ∈ 0 , the triples in the Bose Construction
(Example 13.3.7), defined using any idempotent commutative quasigroup of
order 2n + 1, form an STS(6n + 3).
Proof: We have specified 2n + 1 Type 1 triples and 3(2n2+1 ) Type 2 triples. With
v = 6n + 3, we compute
2n + 1 (6n + 2)(6n + 3) 1 v
(2n + 1) + 3( ) = (3n + 1)(2n + 1) = = ( ).
2 6 3 2
The number of triples in an STS(v) is 13 (2v), since each pair must appear in
exactly one triple. Since we have 13 (2v) triples, it suffices to show that every pair
of elements appears in some triple.
If x and y are copies of the same element of Q, then they appear together in
a Type 1 triple (note that idempotence ensures that they appear together in no
other triple). If x = i¾ and y = j¾ with i
= j , then they appear together in the
Type 2 triple (i¾ , j¾ , (i ◦ j)¾+1 ).
The remaining case is x = i¾ and y = j l , with i
= j and ¾
= l. Since ¾ and l
are distinct in 3 , we may assume l = ¾ + 1. Since Q is a quasigroup, there is a
unique h ∈ Q such that h ◦ i = j . Now x and y appear in the triple (h¾ , i¾ , j l).

When v is odd, there are many idempotent commutative quasigroups (Lemma

13.3.6 presents just one to prove existence). Thus the Bose construction generates
many Steiner triple systems of a given order. The original construction of Bose
[1939] was more specialized.
When 3  v, the Bose construction is not valid, but there is another construc-
tion. Again we use three copies of a quasigroup. With v ≡ 1 (mod 6), we designate
a special element ∞ and partition the other elements into three copies of a quasi-
group of order 2n. Since commutativity requires each element to appear an even
number of times off the diagonal, a commutative quasigroup of even order must
also have each element appear an even number of times on the diagonal. We make
each element in the “lower half ” appear twice.

13.3.9. Definition. A quasigroup on 2n is half-idempotent if i ◦ i = (n + i) ◦

(n + i) = i for 0 ≤ i ≤ n − 1. Let {0 , . . . , n − 1} be the small elements of 2n ;
the remainder are the large elements.

Construction of a half-idempotent quasigroup is similar to Lemma 13.3.6.

Note that in 2n , when a number is a multiple of 2 it is twice a small element and
twice a large element.

13.3.10. Lemma. For n ∈  , there is a half-idempotent commutative quasigroup

on 2n .
Proof: When i + j is even, define i ◦ j to be the small element x such that 2x ≡
i + j (mod 2n). When i + j is odd, define i ◦ j to be the large element x such that
2x + 1 ≡ i + j (mod 2n). The result for n = 3 appears below.
By construction, the diagonals with i + j even are constant in small entries,
the diagonals with i + j odd are constant in large entries, and the diagonals cycle.
That is, each successive row is a cyclic shift to the left of the previous row. Hence
the quasigroup table is a symmetric Latin square, as needed.
644 Chapter 13: Combinatorial Designs

Since the positions on the diagonal have even coordinate sum, all diagonal
entries are small. In particular, i + i = 2i ≡ (n + i) + (n + i) (mod 2n), so the
quasigroup is half-idempotent.

0 3 1 4 2 5
3 1 4 2 5 0
1 4 2 5 0 3
4 2 5 0 3 1
2 5 0 3 1 4
5 0 3 1 4 2

13.3.11. Example. The Skolem Construction. For v = 6n + 1, let the elements

consist of one special element ∞ plus three copies of a half-idempotent commuta-
tive quasigroup Q on 2n , indicating the three copies by subscripts modulo 3. We
use n Type 1 triples of the form (i0 , i1 , i2) for 0 ≤ i ≤ n − 1. The Type 2 triples
have the form (∞ , (n + i)¾ , i¾+1), where 0 ≤ i ≤ n − 1 and ¾ ∈ 3 . Finally, Type 3
triples have the form (i¾ , j¾ , (i ◦ j)¾+1 , where i , j ∈ Q with i
= j and ¾ ∈ 3 .

13.3.12. Theorem. (Skolem [1958]) For n ∈ , the triples in the Skolem Con-
struction (Example 13.3.11), defined using any half-idempotent commutative

quasigroup on 2n , form an STS(6n + 1).
Proof: We have specified n Type 1 triples, 3n Type 2 triples, and 3(2n
2
) Type 3
triples. With v = 6n + 1, we compute
n + 3n + 3(2n
2 = 4n + 3n(2n − 1) = n(6n + 1) = 3 2
) ( ).
1 v

As in Theorem 13.3.8, since we have 13 (2v) triples, it suffices to show that every
pair of elements appears in some triple.
First suppose that ∞ ∈ / {x , y}. If x and y are copies of the same small element
of Q, then they appear together in a Type 1 triple. If x = i¾ and y = j¾ with i
= j ,
then they appear together in the Type 3 triple (i¾ , j¾ , (i ◦ j)¾+1 ).
Hence we may assume that x = i¾ and y = j l with ¾
= l and that {i , j}

consists of two distinct elements or the same large element. Since ¾ , l ∈ 3 , by
symmetry we may assume l = ¾ + 1. If i = n + j with j small, then x and y appear
in the Type 2 triple (∞ , i¾ , j l). Otherwise, let h ∈ Q be the unique element such
that h ◦ i = j , which exists since Q is a quasigroup. Since Q is half-idempotent,
i ◦ i is i when i is small, i − n when i is large; we have already considered those
cases. Hence h
= i, so x and y appear in the Type 3 triple (h¾ , i¾ , j l).
Finally, consider the pairs involving ∞. This element appears with one small
element and one large element in each of the 3n Type 2 triples, yielding the de-
sired 6n pairs.


There are many half-idempotent commutative quasigroups on 2n and hence
many Steiner triple systems of order 6n + 1 from the Skolem construction (the
original construction of Skolem [1958] was more specialized). Nevertheless, these
constructions lack cyclic invarance.

13.3.13. Definition. A Steiner triple system of order v is cyclic if its set of ele-

ments is v and the set of triples is cyclically invariant (that is, adding 1 to
each element of any triple yields another triple).
Section 13.3: Further Constructions 645

13.3.14. Example. Cyclic Steiner triple systems. We generated cyclic projective

planes as translations of a single difference set. In a cyclic STS(v), we have more
than one class of triples, but each class is invariant under translations modulo v.
Define a difference triple in v to consist of three nonzero elements that
sum to 0 modulo v or such that one element is the sum of the other two. Heffter
[1896] gave a sufficient condition for a family of difference triples to generate a
cyclic STS(v) via its translations, each difference triple generating one cyclically
invariant class. For v = 6n + 1, the condition is a partition of {1 , 2 , . . . , 3n} into
difference triples. For v = 6n + 3, it is a partition of {1 , 2 , . . . , 3n + 1} − {2n + 1}
into difference triples.
When v = 6n + 1, there are n triples in the partition, which yields nv triples
under cyclic translations; this is the desired number of triples, 13 (2v). When v =
6n + 3, again there are n triples, but we add a “short orbit ”: the 2n + 1 trans-
lations of (0 , 2n + 1 , 4n + 2). The resulting number of triples is (3n + 1)(2n + 1),
which again is 13 (2v). Hence the specified partition yields the needed number of
triples. We leave to Exercise 9 the job of checking that every pair appears, thereby
proving Heffter ’s result.
Heffter did not find the needed partitions into difference triples, but Pelte-
sohn [1939] did. The specified partitions exist for all v congruent to 1 or 3 modulo
6 except for v = 9. There is no cyclic STS(9) (Exercise 11).

GRAPHICAL DESIGNS

Thinking about (v , ¾ , ë)-designs only as families of ¾-sets in a [v]-set does not

provide much assistance in actually finding designs. It is also difficult to under-
stand the structure of a design just from its incidence matrix, even though a few
computational operations on the incidence matrix will confirm that it is a design.
We used algebraic tools in constructing Steiner triple systems above, but
there are other ways to find structure that can be exploited. For example, if v
has the form (n2), then we can view the elements of the design as pairs in an [n]-
set, or equivalently as edges in K n . Now ¾-element blocks become subgraphs with
¾ edges. Confining our attention to such subgraphs with particular structure
can help lead us to the desired design. We begin with the example of finding an-
other Steiner triple system on 15 elements (recall that a Steiner triple system on
v elements is a (v , 3 , 1)-design).

13.3.15. Example. A (15 , 3 , 1)-design. The graph K 6 has 15 edges; these edges
will be the elements of the design. A natural way to form triples of edges is to
use the edge-sets of triangles as blocks. There are 20 triangles in K 6 ; each edge
appears in four of them. Although incident edges appear together in exactly one
triangle, we do not yet have blocks containing non-incident pairs of edges.
Natural triples containing non-incident pairs are the perfect matchings.
There are 15 perfect matchings, every non-incident pair appears in exactly one
of them, and no two incident edges appear together in a perfect matching. Thus
our set of 35 triples now satisfies the definition of a (15 , 3 , 1)-design. Each edge
appears in three perfect matchings and in a total of seven blocks, which confirms
the equation r(¾ − 1) = ë(v − 1).
646 Chapter 13: Combinatorial Designs

13.3.16. Example. An (11 , 6 , 3)-design. It is hard to think of a natural set with

11 elements. We use the edges of K 5 plus a special element ∗. To keep the goal in
mind, use r(¾ − 1) = ë(v − 1) and b¾ = vr to obtain r = 6 and b = 11; we seek a
symmetric design. We will design blocks that pairwise share three elements; the
transpose of the incidence matrix is the incidence matrix of a design where any
two elements appear together in three blocks.
The edge sets of the five copies of K 4 have six elements, and any two of these
blocks share three edges (a triangle). Add six blocks that consist of ∗ plus a 5-cycle
in K 5 . Fixing one 5-cycle, the other five each share two nonconsecutive edges with
it. There is one way to complete each such cycle.
We have ensured that the fixed cycle-block shares three elements with each
other cycle-bloc (two edges plus ∗). Also, since a 5-cycle visits each vertex in K 5 ,
each cycle-block shares three edges with each clique-block. Note that any two
rotations of the figure on the right below have two common edges (and share ∗).
Thus each of the 11 blocks has size 6, and every two blocks have exactly three
common elements.
• •
• • • •
∗ ∗
• • • •

These constructions are examples of “graphical designs”. Others can be found

in Exercise 10 and the chapter on graphical designs in Colbourn–Dinitz [1996,
2007]. Building designs from the edges of K n opens an entire new area of inves-
tigation. It is natural to require isomorphic subgraphs. When each subgraph
intersects each other exactly once, the subgraphs are sometimes said to be “or-
thogonal”. The transpose of the incidence matrix then yields a design with ë =
1. This is then a graphical interpretation of an ordinary design.

13.3.17. Example. Graphical designs. Classical (v , ¾ , 1)-designs can be inter-

preted as decompositions of K v into edge-disjoint copies of K ¾ ; decomposition
means that the subgraphs are pairwise edge-disjoint. From the graphical view-
point, it is natural to allow more general subgraphs in seeking decompositions of
K n into isomorphic subgraphs. For example, Exercise 5.3.48 decomposes K n into
spanning cycles (for odd n) or spanning paths (for even n).
Ringel [1964] famously conjectured that K 2m+1 decomposes into 2m + 1 iso-
morphic copies of any fixed tree T with m edges. Kotzig extended this to say that
some such decomposition is rotationally invariant when V(K 2m+1) is viewed as the
family of congruence classes modulo 2m + 1 (appearing first in Rosa [1967]). This
became known as the Ringel–Kotzig Conjecture, eventually proved for large m
by Montgomery–Pokrovskiy–Sudakov [2020] using probabilistic methods.
A further restriction requires vertex labels 0 through m so that the differ-
ences between endpoint labels on the edges are 1 through m. Such a labeling is
called a graceful labeling, and the famous Graceful Tree Conjecture asserts
that every tree has a graceful labeling. The ever-growing dynamic survey by Gal-
lian [1998] lists hundreds of papers on graceful and related labelings of graphs.
We include just a few exercises about graceful labeling: Exercises 21–24.
Section 13.3: Further Constructions 647

Many papers have studied decomposition of K n into ¾-cycles when n is odd;

here the problem is completely solved. The problem extends to even n by omitting
one perfect matching. If the necessary divisibility conditions hold (¾ divides (n2)
when n is odd, ¾ divides 12 n(n − 2) when n is even), then the desired decomposition
exists; this was proved by Alspach–Gavlas [2001] for odd ¾ and by Šajna [2002]
for even ¾ (see also an algebraic proof by Buratti [2003] for odd ¾). Much more
generally, Alspach [1981] conjectured that whenever n is odd and m1 , . . . , mt are
integers at least 3 that sum to (n2), the complete graph K n decomposes into cycles
of lengths m1 , . . . , mt . When n is even, the corresponding conjecture is a decom-
position of K n into a perfect matching plus such cycles, when ∑ mi = 12 n(n − 2).
Bryant–Horsley–Pettersson [2014] proved these conjectures.

13.3.18. Example. The Oberwolfach Problem. A conference center in Ober-

wolfach, Germany, holds small weekly mathematical conferences and seats the
participants at round tables of specified sizes for meals. Ideally, each participant
should sit next to each other participant exactly once during the week. In 1967,
at such a meeting, Ringel asked when it is possible to satisfy this requirement.
This is the Oberwolfach Problem: given a partition ë of n, we seek OP(ë),
meaning a decomposition of K n into 2-factors with component lengths ë 1 , . . . , ë t
(plus a leftover 1-factor if n is even). It is easy to see that OP(3 , 3) does not ex-
ist, and it is easy to construct OP(3 , 4) (Exercise 26). Kotzig–Rosa [1974] showed
that OP(3 , 3 , 3 , 3) does not exist; the same is true for OP(4 , 5) (K öhler [1977])
and OP(3 , 3 , 5) (Piotrowski [1979], by computer).
Although the Walecki decomposition is OP(n) with spanning cycles, in gen-
eral the existence of OP(¾ , . . . , ¾) (generally written as OP(¾ t)) is stronger than
the ¾-cycle decompositions mentioned earlier, because here the ¾-cycles must be
grouped into 2-factors, forming what are called resolvable cycle systems. In
general, existence of OP(¾ t) has been solved: for ¾ = 3 by Ray-Chaudhuri–Wilson
[1971] (odd n) and Kotzig–Rosa [1974] (even n), for even ¾ by Häggkvist [1985]
(when ¾ t ≡ 2 (mod 4)) and Alspach–Häggkvist [1985] (when ¾ t ≡ 0 (mod 4)),
and for odd ¾ by Alspach–Schellenberg–Stinson–Wagner [1989] (when t
= 4) and
Hoffman–Schellenberg [1991] (when t = 4).
There is also a bipartite version for decomposing K n ,n ; see Piotrowski [1991].
Bryant–Rodger [2007] presents a survey on 2-factorizations.
For ë ∈ {(3 , 3) , (3 , 3 , 3 , 3) , (4 , 5) , (3 , 3 , 5)}, no OP(ë) exists. The general con-
jecture is that OP(ë) exists in all other cases. The bipartite case has been solved:
when all parts in ë are even, n must be even, and we seek a decomposition of
K n minus a perfect matching into 2 factors with cycle lengths given by ë. Here
Häggkvist [1985] solved the case n ≡ 2 (mod 4) (see below), and Bryant–Danziger
[2011] completed the case n ≡ 0 (mod 4).
For the general case, the conjecture has been confirmed when n is sufficiently
large; see Glock–Joos–K im–K ühn–Osthus [2018].

13.3.19. Lemma. (Häggkvist [1985]) If G is a path or cycle with m edges, and H

is a 2-regular bipartite graph with 2m edges, then G(2) decomposes into two
copies of H , where G(2) is obtained from G by expanding each vertex into an
independent set of size 2 and replacing each edge xy in G with a copy of K 2 ,2
whose parts are the vertex pairs expanding x and y.
648 Chapter 13: Combinatorial Designs

Proof: Decompose G into paths whose lengths are half the length of the cycles
in H. Each such segment P in G with length l expands to 4l edges in G(2) and
decomposes into two cycles C and C of length 2l, as illustrated below. Each end
edge of P expands in each of C and C into two edges at one vertex in the expan-
sion, and each internal edge of P expands in each of C and C into a 1-factor in
the corresponding copy of K 2 ,2 . Each expansion can be distributed to C and C in
either way. The resulting cycles decompose G(2) into two copies of H.

··· • • • • • • • • • ···
• • • • • • • • •
• • • • • • • • •

13.3.20. Theorem. (Häggkvist [1985]) If ë is a partition of n and n ≡ 2 (mod 4),

then OP(ë) exists.
Proof: Begin with a decomposition of K n/2 into (n − 2)/4 spanning cycles. Dou-
ble each cycle by replacing each edge with four edges as in Lemma 13.3.19; each
doubled cycle then has 2n edges. Since each cycle visits each vertex of K n/2 , we
can arrange that in each doubling we omit the same perfect matching in K n . The
union of the doubled cycles then decomposes K n minus the perfect matching. Us-
ing Lemma 13.3.19 to decompose each doubled cycle into two copies of the 2-factor
with cycle lengths specified by ë then completes the desired decomposition.

The solution of the case n ≡ 0 (mod 4) in Bryant–Danziger [2011] also uses

Lemma 13.3.19.

RESOLVABLE DESIGNS AND OTHER TOOLS

For existence theorems in design theory, one tries to prove that construc-
tions of a desired type exist whenever appropriate necessary conditions are sat-
isfied. In an iterative approach, methods are developed for building larger con-
structions from smaller ones. We seek enough such combining theorems to obtain
a proof that works for all sufficiently large values of the parameters. The Moore–
MacNeish Theorem (Theorem 13.1.6) is a combining theorem of this type for or-
thogonal families of Latin squares. Many constructions for small values may be
needed to start the induction.
We use this method to disprove Euler ’s Conjecture. Euler conjectured that
there is no pair of orthogonal Latin squares of order n whenever n ≡ 2 (mod 4).
In fact, Euler was completely wrong; such pairs of squares exist if and only if
n∈/ {1 , 2 , 6}. We will develop resolvable designs, pairwise-balanced designs, and
orthogonal arrays as tools, transforming the problem from the construction of
Latin squares into the construction of certain arrays.
We start with a famous request for a resolvable design.
Section 13.3: Further Constructions 649

13.3.21. Example. The Kirkman Schoolgirls Problem (K irkman [1850]) is

Fifteen young ladies in a school walk out three abreast for seven days in
succession; it is required to arrange them daily, so that no two shall walk
abreast twice.
Each day uses five triples; we want the 35 triples to contain the 105 pairs once
each. Thus the triples form a (15 , 3 , 1)-design. The problem asks for more: the
35 triples must group into seven sets of five disjoint triples.

13.3.22. Definition. A resolvable design is a (v , ¾ , ë)-design whose blocks

group into sets of size v/¾ such that the blocks in each set cover each element
exactly once. The groups are called parallel classes or resolution classes.

In a resolvable design, v/¾ must be an integer. Since v/¾ = b/r it may then
be possible to partition the blocks into r parallel classes. The number of blocks is
just right for each element to appear once in each parallel class. However, not all
designs where ¾ | v are resolvable. The triple systems in the Bose construction
(Theorem 13.3.8) are not, for example.

13.3.23. Example. Resolvable designs. The set of all pairs in [2n] is a (2n , 2 , 1)-
design; the n(2n − 1) blocks are the edges of a complete graph. In Example 8.3.5,
we partitioned these edges into perfect matchings. Thus the design is resolvable.
The designs with the largest blocks relative to the number of points are the
projective planes. Projective planes are not resolvable, since q + 1 does not divide
q2 + q + 1, but they lead to resolvable designs. Given a projective plane of order q,
delete one line L and all appearances of its q + 1 elements. Since every two lines
have one common element, the remaining blocks have size q. All q + 1 appear-
ances of the remaining elements remain. They appeared together only outside L,
so they still appear together once. Thus the family is a (q2 , q , 1)-design; it is an
affine plane. Exercise 13 constructs these explicitly from finite fields.
For each element x of the line L deleted to obtain an affine plane, the remain-
ing n lines through x partition the remaining n2 points into n sets, forming a
parallel class. Each line other than L lies in exactly one of these classes, because
it intersects L in one point. Thus an affine plane is a resolvable design.
The blocks of an affine plane obtained from a projective plane of order 3 ap-
pear below. The columns in the list of blocks form parallel classes.
⎧
⎪ 123 , 147 , 159 , 168 ⎫
⎪
⎪ ⎪
⎨ 456 , 258 , 267 , 249 ⎬
⎪
⎪ ⎪
⎩ 789 , 369 , 348 , 357 ⎪
⎭
Like projective planes, affine planes have been used to construct other objects.
For example, Fon-Der-Flaass [2002] used them to produce large new families of
strongly regular graphs (Definition 15.3.19).

K irkman did find a construction for the Schoolgirl problem. When supplied
with a grouping into parallel classes, resolvable Steiner triple systems become
Kirkman triple systems. The number of points must be divisible by 3. With
that restriction, the full problem was solved by Ray-Chaudhuri–Wilson [1971]
650 Chapter 13: Combinatorial Designs

using pairwise balanced designs: there is a K irkman triple system of order n if

and only if n ≡ 3 (mod 6) (see Anderson [1997, pp. 198–200]).
The decomposition of K 2n into 1-factors in Example 8.3.5 uses one special el-
ement and cyclic symmetry among the rest (see also Theorem 13.3.12). A similar
idea leads to resolvability for the schoolgirl problem.

13.3.24. Example. The Schoolgirl design; a resolvable STS(15). Let the school-
girls be { A , B1 , . . . , B7 , C1 , . . . , C7}. On the first day, use blocks
1 2 3 4 5
AB1 C1 B2 B4 C3 B3 B7 C5 B5 B6 C2 C4 C6 C7
The blocks used on the other days are obtained by adding a constant to all the
subscripts and treating them modulo 7.
Each such set of five blocks covers each element once. Since there are 35
blocks, it suffices to show that each of the 105 pairs appears. The classes of pairs
are covered as follows, with appropriate shifts for i.
1 2 3 4 5
Bi Ci Bi Ci+1 Bi Ci+2 Bi Ci+3 Ci Ci+1
ABi Bi Ci+6 Bi Ci+5 Bi Ci+4 Ci Ci+2
ACi Bi Bi+2 Bi Bi+3 Bi Bi+1 Ci Ci+3

The interpretation of affine planes as resolvable designs is reversible.

13.3.25. Example. Every resolvable (q2 , q , 1)-design is an affine plane. For each
parallel class, we add one element in those blocks, and finally we add one block
consisting of the q+ 1 new elements. By construction, the new family of blocks has
the properties required to be the set of lines in a projective plane, so the original
design was an affine plane.

We define more general types of designs.


13.3.26. Definition. For K ⊆ , a pairwise balanced design with parame-
ters (v , K , ë) is a family of blocks on a set of v elements such that the size of
each block is in K and any two elements appear together exactly ë times. Let
(v , K , ë)-PBD denote such a family.

13.3.27. Example. Constructions of PBDs.

(1) By deleting one element from a design wherever it appears, we obtain two
possible block sizes, but still each two remaining elements appear together the
same number of times. Thus deleting one element from a (v , ¾ , ë)-design yields a
(v − 1 , {¾ , ¾ − 1} , ë)-design.
(2) Given any resolvable (v , ¾ , 1)-design, pick s of the parallel classes. Add
elements x1 , . . . , x s , with each new element entering all the blocks of one paral-
lel class. Add {x1 , . . . , x s} as one additional block. The result is a (v + s , {s , ¾ +
1 , ¾} , 1)-PBD (no blocks of size ¾ remain if s equals the number of parallel classes).
For example, adding one element to each of the seven parallel classes in the
schoolgirl design (Example 13.3.24) yields a (22 , {7 , 4} , 1)-PBD.
Section 13.3: Further Constructions 651

We rephrase the orthogonality conditions for Latin squares in order to de-

velop an easier way to generate orthogonal families.

13.3.28. Definition. Two n-by-n matrices (or two vectors of length n2) are or-
thogonal if the corresponding entries form n2 distinct ordered pairs. Define
special n-by-n matrices Rn and Cn by letting the entry in each position be the
row index or the column index, respectively.

Storing a matrix in row-major order lists its entries row-by-row, across each
row. We can thus interpret each matrix as a vector. In row-major order,
Rn = (1 , . . . , 1 , 2 , . . . , 2 , · · · , n . . . , n) ,
Cn = (1 , . . . , n , 1 , . . . , n , · · · , 1 . . . , n) .

13.3.29. Definition. An orthogonal array OA(s , n) is an s-by-n2 matrix with

entries in [n] in which the rows are pairwise orthogonal.

13.3.30. Proposition. There is an orthogonal family of ¾ Latin squares of order

n if and only if there is an orthogonal array OA(¾ + 2 , n).
Proof: An n-by-n matrix M with entries in [n] is a Latin square if and only if it
is orthogonal to both Rn and Cn . Therefore, from an orthogonal family, we obtain
an orthogonal array by listing Rn , Cn as the first two rows and listing the squares
in the family in row-major order as the other rows.
In an orthogonal array, orthogonality and the Pigeonhole Principle require
n copies of each entry in each row. By reordering columns, we can convert the
first two rows to Rn and Cn without disturbing orthogonality. Now our initial
observation implies that the remaining rows are row-major listings of the Latin
squares in an orthogonal family.

The book Hedayat–Sloane–Stufken [1999] provides extensive material on or-

thogonal arrays and their applications.

THE EULER CONJECTURE (optional)

In the remainder of this section, we apply orthogonal arrays and other tools
to disprove Euler ’s Conjecture. The details are somewhat involved.

13.3.31. Conjecture. (Euler Conjecture) When n ≡ 2 (mod 4), there is no pair

of orthogonal Latin squares of order n.

Let N(n) be the maximum size of an orthogonal family of n-by-n Latin

squares. Since 7 and 4 are prime powers, N(7) = 6 and N(4) = 3. Applied to
the (22 , {7 , 4} , 1)-PBD found in Example 13.3.27, the next theorem yields an or-
thogonal pair of order 22. The value 22 was the first known counterexample to
Euler ’s Conjecture, proved in Bose–Shrikhande [1959].
652 Chapter 13: Combinatorial Designs

13.3.32. Theorem. An (n , {¾1 , . . . , ¾ s} , 1)-PBD yields N(n) ≥ mini {N(¾i) − 1}.

Proof: Let m = mini {N(¾i)}. For each i, take an orthogonal family of m Latin
squares of order ¾i , written as an orthogonal array OA(m + 2 , ¾i).
In OA(m + 2 , ¾i), we may assume that R¾i and C¾i are the first two rows. In
the other rows, we permute labels so that each row agrees with C¾i in the first
¾i columns, having labels 1 , . . . , ¾i in order. In the remaining ¾i2 − ¾i columns,
each pair of rows (not using the first row) has all ordered pairs of distinct labels
from [¾i ]. Let A¾i be the matrix consisting of the last m + 1 rows and last ¾i2 − ¾i
columns of the orthogonal array written in this form.
For each block B of the given pairwise-balanced design, consider the matrix
A| B| , and change the entries 1 , . . . , | B| in A| B| into the elements of x1 , . . . , x| B| ,
respectively. Let the resulting matrix be MB .
Given the blocks B1 , . . . , Bb of the (n , {¾1 , . . . , ¾ s} , 1)-PBD on elements [n],
we form a matrix with m + 1 rows by concatenating b + 1 matrices with m + 1
rows each. The first matrix has n columns and has entries 1 , . . . , n in each row.
The subsequent matrices are MB1 , . . . , MBb .
Consider rows r and s of this matrix, and consider x , y ∈ [n]. If x = y, then
x and y appear in a common column in one of the first n columns. If x
= y, then
x and y appear together in exactly one block Bt of the design. Thus x and y can
appear in a common column only in the block MBt . By the properties of the matrix
A| Bt | , rows r and s have x and y (in that order) in a common column exactly once.
Thus every ordered pair of entries occurs once in rows r and s. Hence the
matrix has n2 columns and is an orthogonal array OA(m + 1 , m). By Proposition
13.3.30, there exists an orthogonal family of m − 1 Latin squares of order n.

Parker [1959] built orthogonal 10-by-10 squares using orthogonal arrays and
a computer. Now Exercise 18 can do this using the general construction below.

13.3.33. Theorem. (Bose–Shrikhande–Parker [1960]) If N(m) ≥ 2, then also

N(3m + 1) ≥ 2.
Proof: It suffices to construct an OA(4 , 3m + 1). Our 3m + 1 symbols are the
integers 0 , . . . , 2m and the labels x1 , . . . , x m . We concatenate 3 + 2m blocks with
four rows each. The first block consists of 0 , . . . , 2m in order in each row. The
second block is an orthogonal array OA(4 , m) using the labels x1 , . . . , x m . In the
remainder we must get the distinct ordered pairs of integers and the pairs (x i , j)
and (j , x i) as columns in each pair of rows.
First define four row vectors of length m: C0 = (0 , . . . , 0), U0 = (1 , . . . , m),
D0 = (2m , . . . , m + 1), and X = (x1 , x2 , . . . , x m) (note that C, U, D stand for “con-
stant ”, “up”, “down”, respectively). Obtain Ci , Ui , Di from C0 , U0 , D0 by adding
i to each coordinate value and reducing modulo 2m + 1. Define

⎛ Ci Ui Di X ⎞
⎜U Ci X Di ⎟
Ai = ⎜ i ⎟ .
⎜ Di X Ci Ui ⎟
⎝X Di Ui Ci ⎠

Complete the array by adding blocks A0 , . . . , A2m after the first two blocks. Ver-
ification that this constructs an OA(4 , 3m + 1) is left to Exercise 17.
Section 13.3: Further Constructions 653

13.3.34. Corollary. For each nonnegative integer t, N(12t + 10) ≥ 2.

Proof: Let m = 4t + 3. Since m is odd, the Moore–MacNeish Theorem (Theorem
13.1.6) yields N(m) ≥ 2. Now apply Theorem 13.3.33.

To kill the Euler Conjecture for all n, we also strengthen Theorem 13.3.32.

13.3.35. Theorem. (Bose–Shrikhande–Parker [1960]) If an (n , {¾1 , . . . , ¾m} , 1)-

PBD exists whose blocks with sizes in {¾1 , . . . , ¾r } are pairwise disjoint, then
N(n) ≥ min{N(¾1) , . . . , N(K r) , N(¾r+1) − 1 , . . . , N(¾m) − 1}.

Proof: Let m be the claimed value plus 1. For i > r, construct A¾i with m + 1 rows
as in Theorem 13.3.32. For i ≤ r, let A¾i be an orthogonal array OA(m + 1 , ¾i).
For each block B, form MB by substitution in A| B| as in Theorem 13.3.32. Con-
catenate the matrices MBi as before, plus a matrix C consisting of one constant
column for each element outside all blocks with sizes in {¾1 , . . . , ¾r }.
If x
= y, then x and y appear together in one block B, and ( yx) occurs as a
column in rows r and s in MB exactly once (and nowhere else). If x = y and x
appears in a block B with size in {¾1 , . . . , ¾r }, then ( yx) appears in rows r and s in
MB ; otherwise, it appears in C.

13.3.36. Example. Orthogonal pair of order 18. A projective plane of order 4 is

a (21 , 5 , 1)-design. Deleting three points not on a line yields an (18 , {3 , 4 , 5} , 1)-
PBD. Since the three deleted points determined three distinct lines, this has
three blocks of size 3. Furthermore, these three blocks are pairwise disjoint,
since otherwise two of the original lines would have two common points. Thus we
can apply Theorem 13.3.35 to obtain N(18) ≥ min{N(3) , N(4) − 1 , N(5) − 1} = 2.
To obtain N(38) ≥ 2, apply this idea to a (41 , 5 , 1)-design, such as constructed
in Exercise 13.2.15.

We need one more “combining ” theorem.

13.3.37. Theorem. If N(m) ≥ ¾ − 1 and x < m, then
N(¾ m + x) ≥ min{N(m) , N(x) , N(¾) − 1 , N(¾ + 1) − 1}.
Proof: Since N(m) ≥ ¾ − 1, there exists an orthogonal array OA(¾ + 1 , m) with
last row Rm . From this we produce a special pairwise balanced design. Since the
last row is Rm , in each other row the m distinct entries appear in each successive
group of m columns. Delete the last row and add (j − 1)m to the entries in row j ,
for 1 ≤ j ≤ ¾ . Row j now consists of entries in the interval Ij = [1 + (j − 1)m , jm].
Treat the columns as m2 blocks of size ¾ on the set 1 , . . . , ¾ m. Each group of
m columns contains each element once and is a parallel class. Each block contains
one element from each Ij . An ordered pair from Ij and Ij can only arise from rows
j and j ; the orthogonality of the original array yields a unique column in which
the desired ordered pair of congruence classes modulo m appeared in these rows.
To complete a pairwise balanced design, we add ¾ further blocks that are pre-
cisely the intervals Ij ; these form a parallel class of blocks of size m. Thus we have
a resolvable (¾ m , {¾ , m} , 1)-PBD.
Next we transform this into a (¾ m + x , {x , m , ¾ , ¾ + 1} , 1)-PBD. We add x new
654 Chapter 13: Combinatorial Designs

elements. Each new element is added to the blocks of one of the parallel classes
of size ¾ , and we add one more block consisting of the new elements. This block of
size x and the special class of ¾ blocks of size m together partition the elements.
If m , x ∈
/ {¾ , ¾ + 1}, then the claim now follows from Theorem 13.3.35; otherwise,
we drop N(m) or N(x) from the minimization and obtain the same result.
When x = 1 there are no pairs involving the new elements, so we don’t need
the extra block. Now x is not a block size, and again we drop N(x) from the mini-
mization, which we can do by setting N(1) = ∞.
13.3.38. Corollary. If N(m) ≥ 3, N(x) ≥ 2, and 1 ≤ x < m, then N(4m + x) ≥ 2.
Proof: Take ¾ = 4 in Theorem 13.3.37.
In applying Corollary 13.3.38, it is helpful to take m as a multiple of 4.
13.3.39. Lemma. (see Anderson [1990], p. 129) If 4 | m, then N(m) ≥ 3.
Proof: The Moore–MacNeish Theorem (Theorem 13.1.6) yields the bound unless
m is divisible by 3 but not by 9. Thus m is divisible by 12 but not by 36. We
can express m as 12 or 24 times an even power of 2 times an odd number other
than 3. By the Moore–MacNeish Theorem, it suffices to prove that N(12) ≥ 3 and
N(24) ≥ 3. For n = 12, there is an explicit family of size 5 (Dulmage–Johnson–
Mendelsohn [1961]). For n = 24, see Exercise 15.
These combining theorems can yield better lower bounds on N(m) than the
Moore–MacNeish Theorem, but we only need N(m) ≥ 2.
13.3.40. Theorem. (Bose–Shrikhande–Parker [1960]) N(n) ≥ 2 for all positive
integers outside {1 , 2 , 6}.
Proof: We need consider only n ≡ 2 (mod 4) with n > 6. Corollary 13.3.34 handles
the case n ≡ 10 (mod 12). Since 12t + 6 = 3(4t + 2), the Moore–MacNeish Theorem
takes care of n = 12t + 6 given an orthogonal pair with order 4t + 2. There is no
pair when n = 6, so we need an explicit construction for n = 18, done in Example
13.3.36. When 18 < 4t + 2 ≡ 6 (mod 12), we reduce by another factor of 3.
This leaves only n ≡ 2 (mod 12). Explicit pairs are known for n = 14 and
n = 26 (we omit these), and we covered n = 38 in Example 13.3.36. We use the
Moore–MacNeish Theorem again for 50 = 5 · 10, 98 = 7 · 14, 110 = 10 · 11. We
use Corollary 13.3.38 for 62 = 4 · 13 + 10, 74 = 4 · 16 + 10, 86 = 4 · 19 + 10,
122 = 4 · 27 + 14, 134 = 4 · 27 + 26. We have now discussed the cases up to n = 144.
We express the other values as a multiple of 16 plus one of {18 , 22 , 26 , 30}.
Since n > 144, we have n = 16t + x = 4m + x, where 4 ÷ m and m ≥ 32. Thus
x < m. Also N(m) ≥ 3, by Lemma 13.3.39. Now N(n) ≥ 2, by Corollary 13.3.38.

EXERCISES 13.3
13.3.1. (−) For even v, construct a resolvable (v , 2 , 1)-design.
13.3.2. (−) Prove that K6 decomposes into three edges and four triangles.
13.3.3. (−) Prove that K9 decomposes into 4-cycles.
13.3.4. (−) A caterpillar is a tree having a single path incident to or containing every
edge. Prove that every caterpillar is graceful.
Exercises for Section 13.3 655

13.3.5. (♦) A total coloring of G color each vertex and edge so that objects adjacent or
incident have different colors; Ò (G) is the number of colors needed. Prove that an idempo-
tent commutative quasigroup of order n (Definition 13.3.9) exists if and only if Ò (K n) = n.
Conclude from Lemma 13.3.10 that Ò (K n) is n when n is odd and n + 1 when n is even.
(Behzad [1965]; see also Yap [1989]) (Comment: Since (K n) = n − 1, the Total Coloring
Conjecture (Exercise 8.3.33) holds for complete graphs.)
13.3.6. Prisoners A, B, and C stand with C in front of B and B in front of A. From ¾
hats of different colors, each receives one hat , selected at random. The prisoners know all
the colors but see only the hats in front of them. They guess the colors of their own hats,
first A, then B, then C, hearing what is said. If they all guess correctly, they will all be
freed. They know the rules and plan a strategy in advance. What strategy maximizes the
probability of success? (Hint: Use an idempotent quasigroup.) (Wagon–Zielinski [2018])
13.3.7. View v × m as {ai : a ∈ v , i ∈ [m]} An (i , j)-difference is a difference in v be-
tween elements with subscripts i and j . A (v , ¾ , ë , m)-mixed difference system is a fam-
ily D1 , . . . , Dt of ¾-subsets of v × m such that each element of v occurs exactly ë times as
an (i , j)-difference for each (i , j) ∈ [m]2 , except that difference 0 does not occur with i = j .
(a) Prove that the translates (in v) of the t sets in a (v , ¾ , ë , m)-mixed difference sys-
tem D1 , . . . , Dt form a (vm , ¾ , ë)-design.
(b) Let u = 2n + 1. Consider sets {01 , 02 , 03 } and { {ri , −ri , 0i+1 }: i ∈ 3 and 1 ≤ r ≤ n} .
Prove that these 3n + 1 sets form a (u , 3 , 1 , 3)-mixed difference system.
(c) Obtain an STS(v) when v ≡ 3 (mod 6) (see also Example 13.3.7). (Bose [1939])
13.3.8. (♦) Prove that the complete ¾-partite graph with parts of size ¾ decomposes into
¾ 2 copies of K¾ if and only if there is an affine plane of order ¾ .
13.3.9. Complete the proof that if the small numbers in v have a partition into difference
triples as specified in Example 13.3.14, then there exists a cyclic STS(v).
13.3.10. (♦) Construct a (6 , 3 , 2)-design. (Hint: Define 10 triangles on six points.)
13.3.11. Prove that there is no cyclic STS(9).
13.3.12. Fisher ’s Inequality states b ≥ v for the number b of blocks in a (v , ¾ , ë)-design.
Improve this to b ≥ v + p − 1 for a resolvable (v , ¾ , ë)-design with p parallel classes.
13.3.13. Let q be the finite field of size q. Let V be the set of ordered pairs of elements
from q . For a , b ∈ q , let La ,b = {(x , y) ∈ V : y = ax + b}, and let Ma = {(x , y) ∈ V : x = a}.
Prove that the q2 sets La ,b and the q sets Ma together form a (q2 , q , 1)-design.
13.3.14. (♦) Affine planes.
(a) Prove that if B is a block in a (q2 , q , 1)-design, and x is an element not in B, then
exactly one block in the design contains x and is disjoint from B.
(b) Use part (a) to prove that every (q2 , q , 1)-design is resolvable.
(c) Prove that if there is a (q2 , q , 1)-design, then there is a (q2 + 1 , q , q − 1)-design.
(Hint: Make q − 1 copies of the blocks in all but one class.) (Rasch–Herrendörfer [1977])
13.3.15. Let q be a prime power. Obtain a (q2 , q , 1)-design. Delete one element to obtain a
(q2 − 1 , {q , q − 1} , 1)-PBD in which the blocks of size q − 1 are pairwise disjoint. Conclude
that N(q2 − 1) ≥ N(q − 1); in particular, N(24) ≥ 3.
13.3.16. Pairwise-balanced designs.
(a) Show that a (v , {3 , 5} , 1)-PBD exists only when v is odd, and construct one for
v = 11. (Hint: For the construction, start with a resolvable (6 , 2 , 1)-design.)
(b) Show that a (v , {4 , 5} , 1)-PBD exists only when v or v − 1 is divisible by 4, and
construct one for v = 17.
13.3.17. (♦) Complete the proof of Theorem 13.3.33 that N(m) ≥ 2 implies N(3m + 1) ≥ 2.
656 Chapter 13: Combinatorial Designs

13.3.18. (♦) Use Theorem 13.3.33 to construct an explicit pair of 10-by-10 orthogonal Latin
squares. Express it as a “Graeco-Latin” square with the 100 numbers from 00 to 99 in the
100 positions. Explain the steps used (no proof needed).
13.3.19. Without restricting block sizes, a pairwise balanced design on a set E becomes
just a family F of blocks (subsets of E) such that every two elements appear in common
blocks. Prove that a family F having no block of size 1 is a pairwise balanced design with
= 1 if and only if F is the set of hyperplanes of a matroid of rank 3 with no circuits of
size 1 or 2. (G. Chappell)
13.3.20. (♦) Prove that if there exist n − 1 pairwise orthogonal Latin squares of order n,
then there exists a set of n(n − 1) permutations of [n] such that no two of them have the
same ordered pair in any two corresponding positions. The columns of the array below
form such a set of permutations when n = 4. (Comment: The converse also holds.)
123412341234
432121433412
214334124321
341243212143

13.3.21. (♦) A graceful labeling of a graph G with m edges is an injection  : V(G) →

{0 , . . . , m} such that {| (u) −  (v)| : uv ∈ E(G)} = {1 , . . . , m}. Kotzig conjectured that ev-
ery tree has a graceful labeling (see Ringel [1964]).
(a) Prove that if a graph G with m edges has a graceful labeling, then K 2m+1 decom-
poses into 2m + 1 copies of G. (Rosa [1967])
(b) Prove that if G is a tree and all trees have graceful labelings, then also K 2m de-
composes into 2m − 1 copies of G.
13.3.22. An -labeling of a bipartite graph is a graceful labeling where all labels on one
part are smaller than all labels on the other. Prove that if a bipartite graph G with m

edges has an -labeling, and  ∈ , then K 2m¾+1 decomposes into copies of G. (Rosa [1967])
13.3.23. (+) Graceful Eulerian graphs.
(a) Prove that if an Eulerian graph G with m edges is graceful, then 4| m or 4|(m + 1).
(b) Prove that the condition in part (a) is sufficient for cycles. (Hebbare [1976])
13.3.24. (+) Let G be the graph consisting of  4-cycles with one common vertex. Prove
that G is graceful. (Hint: Put 0 at the vertex of degree 2  .)
13.3.25. Prove that K9 decomposes into 6-cycles.
13.3.26. For the Oberwolfach Problem, prove that OP(3 , 3) does not exist. Using vertex
set 6 ∪ {∗}, construct OP(3 , 4).
13.3.27. (♦) A (6 , 3 , 2)-design. Let the points be the vertices of K6 . Distinguish one vertex
u of K6 . Using five triangles containing u and five triangles not containing v, obtain a
(15 , 3 , 1)-design. Is this design resolvable? (Füredi)
13.3.28. (♦) A (21 , 5 , 3)-design for t = 3. We construct 5-element blocks on 21 elements.
The elements are E(K7). We use the copies of K1 ,5 , the 5-cycles, and the spanning sub-
graphs consisting of a triangle and two isolated edges. Prove that these blocks form a
3-design by showing that every triple of elements appears in exactly three blocks. (Hint:
There are five distinguishable ways to arrange three edges in K7 .)
13.3.29. A ( , ; n)-transversal design is a set of  disjoint n-sets (“groups”) and a
family  of -sets (“blocks”) such that each block intersects each group and any two ele-
ments from different groups appear together in blocks. Prove that there exists a ( , 1; n)-
transversal design if and only if there exists an orthogonal array OA( , n) (and hence a
family of  − 2 pairwise orthogonal Latin squares of order n). (Comment: Transversal
designs were introduced by Hanani [1961]; see Laywine–Mullen [1998, pp. 28–29].)
Chapter 14

The Probabilistic Method

In its simplest form, the probabilistic method in combinatorics is a non-
constructive method for proving the existence of desired combinatorial objects.
A discrete probability space is defined where occurrence of the desired structure
is an event. If the event has positive probability, then the desired structure oc-
curs for some point in the space, and hence it exists.
Random variables add considerable power. For example, we proved inequali-
ties for LYM orders in Section 12.2 by comparing average and maximum values
of a random variable. Here we will also use random variables to study “threshold
functions” in order to make precise statements about “typical behavior ” of large
structures. Finally, we will also discuss the benefits of showing that random vari-
ables are highly concentrated around their expectations.
Several texts and monographs present the probabilistic method in more detail
than we can here. Prominent among these is Alon–Spencer [1992, 2000, 2008,
2016]. Others emphasize probabilistic methods in graph theory: Bollobás [1985,
2001], Palmer [1985], Janson–Łuczak–Rucinski [2000], Molloy–Reed [2002], and
Frieze–K aroński [2016]. Motwani–Raghavan [1995] discusses probabilistic al-
gorithms. Habib–McDiarmid–Ramirez-Alfonsin–Reed [1998] surveys additional
uses of probabilistic methods in algorithmic discrete mathematics.

14.1. Existence and Expectation

In this section we begin with the simplest form of the probabilistic method
and then show how to strengthen it by using random variables and expectation.
We assume familiarity with the notions of events, probability spaces, and condi-
tional probability as defined in Chapter 0.
Proving that an event occurs with positive probability involves proving an
inequality. We use several fundamental tools in proving inequalities, some from
elementary calculus. We begin with useful facts about exponentials, where e is
the base 2.71828... of the natural logarithm. These inequalities are important
in asymptotic arguments, and we use them often, so we distinguish e from the
notation for an edge e in a graph by setting the numerical value e in roman font.

657
658 Chapter 14: The Probabilistic Method

14.1.1. Proposition. If x ∈ and n ∈ , then 1 + x ≤ e x and (1 + nx )n ≤ e x , with

equality in each only when x = 0.
Proof: The first inequality is trivial for x ≤ −1 and follows from the series ex-
pansion of e x when x > −1. For x
= 0, it yields (1 + nx )n < e(x/n)n = e x .

We will need bounds on factorials and binomial coefficients. For (¾n), the
bound n¾/¾ ! may suffice for constant ¾ , but often we need a better upper bound.

14.1.2. Proposition. If ¾ ∈ , then (¾n) < ( ne¾ )¾ .

Proof: Using the factorial expression for the binomial coefficient,
¾ −1 ¾ −1
n n− i n ¾ 1 − i/n
( )=∏ =( ) ∏ .
¾ ¾−i ¾ 1 − i/¾
i=0 i=0
¾ −1 1− i/n
It suffices to prove ∏ i=0 1− i/¾ < e¾ . This is implied by () < e¾ , where () =
¾ −1
∏ 1
i=0 1− i/¾ . Since (0) = 1 (empty product), the claim follows from
¾ ¾ −1¾
( + 1) +1 −i +1 1 ¾
=∏ ∏ =∏ = (1 + ) < e ,
()
i=0
 + 1 − i i=0  i=1
 
where the last inequality uses Proposition 14.1.1.
√
Stirling ’s Formula states n! ∼ nne−n 2 n (Application 2.3.8). It approx-
¾ n−¾ √
imates (¾n) by ( ¾n ) ( n−n ¾ ) n/[2(n − )]. The last factor is less than 1 for
n−¾
1 ≤  ≤ n − 1, and for the middle factor Proposition 14.1.1 yields ( n−n ¾ ) =
¾ n−¾
(1 + n−¾ ) < e . This is not a rigorous proof of Proposition 14.1.2, because it
¾
√
does not analyze the error in Stirling ’s Formula and we never derived the 2 .

THE UNION BOUND

The probabilistic method was popularized by Erdős, who used it in 1947 to

obtain lower bounds on Ramsey numbers. Recall that R( , ) is the least n such
that every 2-coloring of E(K n) has a homogeneous set of  vertices, meaning a set
whose (¾2) edges all have the same color. The inductive proof of Ramsey ’s Theorem
√ √ ¾
yields R( , ) ≤ (2¾¾−−12) ∼ 4¾/  . Erdős showed that 2 vertices do not suffice.

14.1.3. Theorem. (Erdős [1947]) R( , ) > √1  2¾/2 .

e 2
Proof: Consider a random 2-coloring of E(K n), with each edge having each color
with probability 12 , independently. A set of  vertices is homogeneous with prob-
ability 2 · 2−(2). With (n) such events, the probability that at least one occurs is
¾

¾
at most (n¾ )21−(2). If this is less than 1, then some outcome of the process has no
¾

homogeneous -set, and hence R( , ) > n. Since (¾n) < ( ne ¾

)¾ , it suffices (roughly)
to have ¾ ≤ 2
ne (¾ −1)/2
, which is equivalent to n ≤ √  2 .
1 ¾/2
e 2
Section 14.1: Existence and Expectation 659
√ ¾
14.1.4. Remark. There is a large gap between 2 and 4¾ . More sophisticated
probabilistic methods have made only small improvements in the lower bound.
Nevertheless, the bounds from known constructions are sub-exponential, so this
is a triumph for the probabilistic method.
The proof is the same as the counting argument in Theorem 10.2.14, just
phrased in probabilistic language. Many probabilistic arguments with finite sam-
ple spaces are just weighted counting arguments, but the tools of probability do
the job more clearly and efficiently and go well beyond what counting can do.
Existence arguments yield probabilistic constructions. For example, a ran-
dom 2-coloring of E(K 64) has a homogeneous 10-set with probability less than
((26)10/10!)2−44 , which is about .018. If a randomly generated coloring has a ho-
mogeneous 10-set, just try again. The probability of repeated bad outcomes is a
product of small numbers and soon becomes incomprehensibly small.

The fundamental probabilistic argument in Theorem 14.1.3 has a name and

underlies many easy applications of the probabilistic method. Note in particular
that this bound does not require independence; it holds for any events A1 , . . . , An .

14.1.5. Proposition. (The Union Bound) For any events A1 , . . . , An in a prob-

ability space, (⋃i=1 Ai) ≤ ∑ i=1 (Ai).
n n

We next study proper 2-colorings of ¾-uniform hypergraphs. Recall that the

edge set of a ¾-uniform hypergraph H is a family of ¾-sets, and a t-coloring of
the vertex set V(H) is proper if no edge is monochromatic. We can prove that a
hypergraph is 2-colorable by showing that a random coloring of the vertex set has
positive probability of being proper.

14.1.6. Proposition. (Erdős [1963]) If () denotes the minimum number of

edges in a -uniform hypergraph that is not 2-colorable, then
2 − 1 4¾
2 ¾ −1 ≤ () ≤ ( )∼ √ .
 2 
Proof: A random coloring makes a specified edge monochromatic with probabil-
ity 21−¾ , so with fewer than 2¾−1 edges the total probability of the bad events is
less than 1 and a proper 2-coloring exists.
By the Pigeonhole Principle, the hypergraph with edge set ([2¾¾−1]) is not 2-
colorable, since some  vertices must have the same color. Stirling ’s Formula
yields the asymptotics of (2¾¾−1 ) (see Example 2.3.10).

In this context, 2-colorability was named “Property B” by Erdős in honor of

Bernstein. The problem of finding () was posed by Erdős and Hajnal in 1961.
Beck [1978] improved the lower bound to () ≥  1/3−o(1) 2¾ (see Corollary 14.2.7).
A probabilistic argument yields an improved upper bound, logarithmically
asymptotic to the lower bound. Instead of seeking existence of a proper color-
ing, we seek existence of a hypergraph with no proper coloring. Hence we gener-
ate a random hypergraph instead of a random coloring. In Proposition 14.1.6 we
showed that some outcome of forming a random coloring has no bad edge; now we
show that some outcome of forming a random hypergraph has no good coloring.
660 Chapter 14: The Probabilistic Method

14.1.7. Theorem. (Erdős [1964]) There exists a ¾-uniform hypergraph with

(1 + o(1)) e ln4 2 ¾ 2 2¾ edges that is not 2-colorable.
Proof: We form a ¾-uniform hypergraph by choosing m edges at random from ver-
tex set [n], where n and m will be chosen later to optimize the resulting bound.
Repeated edges cause no problem; indeed, they would improve the bound. We ap-
ply the Union Bound to the 2 n events that particular red/blue colorings of the
vertex set are proper colorings of the resulting hypergraph.
For a given coloring with r points in one color and s in the other, where r + s =
n, the probability that a random ¾-set is monochromatic is [(¾r ) + (¾s )] / (n¾ ), which
is minimized when r = s. Let p = 2(n/2 n
¾ )/( ¾ ). For a given coloring, the proba-
bility that a random ¾-set is monochromatic is at least p. For m edges chosen
independently, the probability that none is monochromatic is at most (1 − p)m .
By the Union Bound, our random hypergraph is 2-colorable with probability
at most 2 n(1 − p)m . If 2 n(1 − p)m < 1, then some ¾-uniform hypergraph with n
vertices and m edges is not 2-colorable. Since (1 − p) ≤ e− p , it suffices to have
2 n e−mp < 1. To minimize m such that 2 n < emp , we set m = ⌈ np ln 2⌉ .
¾ −1
Now we choose n in terms of ¾ to minimize n/p. Rewrite p as 21−¾ ∏i=0 nn−−2ii .
Using 1−1i/n = 1 + ni + O(( ni )2), we have
n − 2i 1 − 2i/n i i2 i2
= = 1 − + O ( 2 ) = e−i/n + O ( 2 ) .
n− i 1 − i/n n n n
¾ −1
Using ∑i=0 i = ¾(¾ − 1)/2 makes p asymptotic to 21−¾ e−¾ /(2n). To minimize
2

n2¾−1 e¾ /(2n) , calculus suggests n = ¾ 2/2, yielding m = (1 + o(1)) e ln 2 2 ¾

4 ¾ 2 .
2

The 2-colorability of ¾-uniform hypergraphs is closely related to ¾-choosability

of bipartite graphs. Recall that a graph G is ¾ -choosable if a proper coloring can
be chosen from any color lists of size ¾ assigned to the vertices. The choice num-
ber Ò l(G) is the least ¾ such that G is ¾-choosable. Combinatorial proof of the
relationship below is requested in Exercise 15.

14.1.8. Theorem. (Erdős–Rubin–Taylor [1979]) If n¾ is the smallest order of

a non- ¾-choosable bipartite graph, and m¾ is the smallest size of a non-2-
colorable ¾-uniform hypergraph, then m¾ ≤ n¾ ≤ 2m¾ .

Example 8.2.7 proved Ò l(K r,r) > ¾ for r = (2¾¾−1 ), so Ò l(K r,r) ≥ ( 12 + o(1)) lg r.
Exercise 16 requests an upper bound about twice that. Both have been improved.

14.1.9. Corollary. lg r − (2 + o(1)) lg lg r ≤ Ò l(K r,r) ≤ lg r − ( 12 − o(1)) lg lg r.

Proof*: The lower bound uses Theorems 14.1.7–14.1.8. Since some hypergraph
with at most O(¾ 2 2¾) edges is not 2-colorable, ¾-choosability of K r,r requires
r ≤ O(¾ 2 2¾). Taking logs and inverting the relationship yields the lower bound.
The improvement to the upper bound combines Theorem 14.1.8 with a lower
bound on the number m¾ of edges in a non-2-colorable ¾-uniform hypergraph.
Radhakrishnan–Srinivasan [2000] proved m¾ ≥ ((¾/log ¾)1/2 2¾) (another proof
appears in Kozik–Cherkashin [2015]). For Ò l(K r,r) > ¾ , we have 2r ≥ n¾ ≥ m¾ .
Thus there is a constant c such that r ≤ c(¾/lg ¾)1/2 2¾ yields Ò l(K r,r) ≤ ¾ . Invert-
ing the relationship between r and ¾ yields the upper bound.
Section 14.1: Existence and Expectation 661

Having  l(G) > requires existence of a -uniform list assignment from

which no proper coloring can be chosen. Hence to apply the existence argument
we generate a random list assignment. The subtlety is how to do this so that with
positive probability no proper coloring can be chosen. It will suffice to consider a
bipartite subgraph with large minimum degree.

14.1.10.* Theorem. (Alon [1993]) For some constant c, every graph G with av-
erage degree d has choice number at least c logloglogd d .

> (ss ) log (2(ss )) implies  l(G) > s, since taking

4 4
d
Proof: It suffices to show that 4
logarithms yields log d ∼ c s log s, which is equivalent to s ∼ c logloglogd d for some c.
Since G has average degree d, it has a subgraph G with minimum degree at
least d/2 (iteratively delete vertices with smaller degree; Exercise 5.2.10). The
subgraph G in turn has a spanning bipartite subgraph H such that d H (v) ≥
1
(v) for all v ∈ V(G ) (Exercise 27, Theorem 5.2.9), so (H) ≥ d/4.
2 dG
We generate a random list assignment L for H with lists of size s (from a set
of s4 colors). We show that with positive probability, H has no L-coloring (Defini-
tion 8.2.6). This yields  l(H) > s and hence  l(G) > s.
Let A and B be the parts of H , with | A| ≥ | B| . Let S = [s4 ], and let t = (ss ).
4

Each vertex receives a random s-subset of S as a list, with all t such sets equally
likely. Say that a vertex of A is full if all t possible lists appear on its neighbors.
The probability that a particular s-set T fails to appear on the neighbors of x is
(1 − 1/t)d H (x) . Since there are t such sets and d/4 > t log(2t), we obtain

(x is not full) ≤ t (1 − t−1 )d/4 < te−t −1 d/4

< te− log(2t) = 12 .
Hence the expected number of full vertices is at least | A| /2, and there is some
outcome of the random list assignment such that at least | A| /2 vertices of A are
full. Fix such an assignment for B.
We now claim that extending this assignment by randomly chosen lists for A
produces with positive probability a list assignment from which no proper color-
ing can be chosen. Let  be a particular choice of colors from the lists on B (recall
that B is independent in H). For a full vertex x in A, since all s-sets appear on
its neighbors, at most s − 1 colors fail to be chosen on its neighbors. Hence  can
be properly extended to x only if L(x) contains one of these missing colors. There
are at most s − 1 ways to name a usable color, and then L(x) must be filled from
the remaining colors, so
(s − 1)( ss−−11 )
4
s−1
(x can be colored) ≤ t
= 3 < 2.
s
1
s
To extend  to an L-coloring of H , all full vertices must be colored; this has
probability at most (1/s2)| A|/2 , which equals s−| A| . With s| B| choices for  on B from
the assignment on B, the probability that some choice  on B extends to an L-
coloring is at most s| B| s−| A| . Since | A| ≥ | B| , the bound is less than 1. Hence for
some outcome of the random lists on A, no proper coloring can be chosen.

Alon [2000b] later improved Theorem 14.1.10 to  l(G) ≥ ( 12 − o(1)) ln d when

G has average degree d. It is conjectured that there is a constant c such that
 l(G) ≤ c log d when G is d-regular and bipartite; only  l(G) ≤ O( logd d ) is known.
662 Chapter 14: The Probabilistic Method

RANDOM VARIABLES

Often we study numerical aspects of points in a probability space. Typically

we generate an object randomly, and we consider the value of some parameter of
the resulting object. We repeat the relevant definitions from Chapter 0.

14.1.11. Definition. A random variable on a discrete probability space S is

a function X : S → . The expectation or expected value (X) of X
is ∑ a∈S X(a)(a) (when this converges). A discrete random variable X
has countable range, usually 0 . For such X , let X = ¾ denote the event
{a ∈ S: X(a) = ¾} and write (X) = ∑¾=0 ¾ · (X = ¾). The pigeonhole
∞

property of the expectation is the statement that there is an element of the

probability space for which the value of X is as large as (or as small as) (X).

Using the pigeonhole property requires a value or bound for (X). Often we
obtain this by expressing X as a sum of simpler random variables. We repeat
Lemma 0.13, restricting our attention to finite sums of random variables on dis-
crete probability spaces.

14.1.12. Lemma. (Linearity of expectation) If X and X1 , . . . , X ¾ are random

variables on the same space such that X = ∑ X i , then (X) = ∑ (X i). Also
(cX) = c (X) for any constant c.
Proof: In a discrete probability space, each sample point contributes the same
amount to each side of each of these equations.

14.1.13. Remark. We will apply Lemma 14.1.12 to sums of variables. A count-

ing variable X is a sum of variables X i indicating whether the ith event in some
set occurs. An indicator variable or 0 , 1-variable takes values in {0 , 1}. When
X i is an indicator variable, (X i) = (X i = 1). Thus the expected number of
events that occur is the sum of their probabilities: (X) = ∑ (X i = 1).

Linearity and pigeonholing capture the existence arguments we have given.

For example, to show that ¾-uniform hypergraphs with fewer than 2¾−1 edges are
2-colorable, we chose a random vertex partition. Each edge is monochromatic with
probability 21−¾ . With fewer than 2¾−1 edges, the expected number of monochro-
matic edges is less than 1, and hence some coloring has no monochromatic edge.
Use of linearity often simplifies computations of expectation; the exercises
present many examples. Also, comparing the expected and maximum values of a
random variable yields results such as the next theorem.

14.1.14. Theorem. (Caro [1979], Wei [1981]) For a graph G,

1
(G) ≥ ∑ d(v) + 1
.
v∈ V(G)

Proof: (Alon–Spencer [1992]) In an ordering of V(G), the vertices that appear

before all neighbors form an independent set S. In a random ordering, the prob-
ability that v precedes all its neighbors is (d(v) + 1)−1 . Thus the right side of the
inequality is the expected size of S. The left side is the maximum size of S.
Section 14.1: Existence and Expectation 663

Theorem 14.1.14 yields a proof of Turán’s Theorem (Theorem 11.1.3).

14.1.15. Theorem. Let G be an n-vertex graph with no (r + 1)-clique. The num-

ber of edges of G is maximized (uniquely) when G is the complete r-partite
graph T n ,r whose part-sizes differ by at most 1.
Proof: Maximizing | E(G)| such that K r+1
⊆ G is equivalent to minimizing | E(H)|
such that (H) ≤ r, where H = G. Note that T n ,r is a disjoint union of complete
# #
graphs. Let m = #### E(T n ,r)####. It suffices to show that T n ,r is the only graph with
m edges having independence number as small as r, since deleting edges never
reduces the independence number and deleting an edge from T n ,r increases it.
Fix | E(H)| = m. The lower bound on (H) in Theorem 14.1.14 is now min-
imized only when the degrees differ by at most 1, by convexity of the reciprocal
function. Since the degree-sum is 2m, the only such list is the degree list for T n ,r .
Groups of vertices with degree = 1 contribute 1 to the sum, so the sum is r.
We have proved that (H) ≥ r when H has m edges (and n vertices).
To show that equality holds only when H = T n ,r , we observe that equal-
ity requires not only the specified degrees, but also equality in the argument of
Theorem 14.1.14. To avoid having an independent set larger than (| S|), every or-
dering must generate an independent set of the same size. If there exist x , y ,  ∈
V(H) with x ,  ∈ N(y) and x  ∈ / E(H), then an ordering  that starts x , y ,  puts
x ∈ S but y ,  ∈ / S. For an ordering  that agrees with  except for starting
x ,  , y, the independent set will be the same except that  will be added. Hence
(H) > E(| S|) = r. A graph with no induced P3 is a disjoint union of complete
graphs. With vertex degrees as in T n ,r , the only such graph is T n ,r .

Applications of expectation and linearity often yield strong results with sur-
prisingly simple proofs. Many occur in the exercises, and another example is in
Theorem 16.1.28. By choosing random subgraphs, we show there that when m ≥
4n, every drawing in the plane of a graph with n vertices and m edges has at least
m3/(64n2) crossing pairs of edges.
The pigeonhole property of expectation plays a crucial role in an upper bound
on the Ramsey number R(3 , ). We seek that guarantees (G) ≥ in every
triangle-free n-vertex graph G. Weakening the Caro–Wei Theorem (Theorem
14.1.14) yields (G) ≥ n/(d + 1) when (G) = d, which is easily proved directly.
The breakthrough by Ajtai–Komlós–Szemerédi [1980] showed (G) ≥ (cn lg d)/d
when G is triangle-free. Simplifications and improvements were due to Shearer
[1983] (achieving c = 1 + o(1)), Shearer [1995], and Alon [1996]. We present the
short proof in Alon–Spencer [2008] based on the earlier papers.

14.1.16. Theorem. If G is a triangle-free n-vertex graph with (G) ≤ d, where

d ≥ 1, then (G) ≥ (n lg d)/(8d).
Proof: If d ≤ 16, then d+d 1 ≥ 12 ≥ lg8d , which implies d+n 1 ≥ n8d lg d
, so the desired
bound follows from (G) ≥ d+n 1 . Hence we may assume d ≥ 16.
Let W be a random independent set in G, with all independent sets equally
likely. To prove (G) ≥ n8d
lg d
, it suffices to show (| W |) ≥ n8dlg d
. To study (| W |),
we introduce additional random variables. For each vertex v in G, define X v by
664 Chapter 14: The Probabilistic Method

d if v ∈ W
Xv = {
| N(v) ∩ W | if v ∈
/ W.
Let X = ∑v∈V(G) X v . In computing X , each vertex u ∈ W contributes d to X u
and contributes 1 to X v for each v ∈ N(u), since N(u) ∩ W = ∅. Hence the total
contribution to X from u is at most 2d. This yields X ≤ 2d | W | no matter what
W is. Taking expectations, we have (| W |) ≥ (X)/(2d). It therefore suffices to
prove (X v) ≥ 14 lg d for each vertex v, taken over all (equally likely) choices of W .
Fix v ∈ V(G). We will group the choices of W and show that over each group,
the expectation of X v is at least 14 lg d. The desired lower bound on (X v) then
follows, regardless of the sizes of the groups, since all W are equally likely.
We group the choices of W according to the subset of W outside the closed
neighborhood of v. This set W − (NG(v) ∪ {v}) is another random variable de-
termined by W ; call it S. For a given choice of S, let Y be the set of common
nonneighbors of S that lie in N(v), and let t = | Y |. The independent set W that
generates S may arise by adding v to S (since S ∩ N(v) = ∅) or by adding to S any
subset of Y . These 1 + 2 t possibilities for W are equally likely, since all W are
equally likely. In the first case, the value of X v is d. Over the other 2 t possibil-
ities the average value of X v is t/2, since the value of X v is then | Y ∩ W | . Thus
over this group the expected value of X v is the left side of the desired inequality
d (t/2)2 t
+ ≥ 14 lg d.
2t + 1 2t + 1
If t = 0, then the left side simplifies to d/2, which is big enough. Hence we
may assume t ≥ 1. If the inequality fails, then clearing fractions and rearranging
terms yields 4d − lg d < 2 t(lg d − 2t). Since
√ the left side is positive,
√ we have lg d >
2t ≥ 2. With t < 12 lg d, we have 2 t < d. Thus 4d − lg d < d(lg d − 2), which
fails when d ≥ 16.

14.1.17. Corollary. (Ajtai–Komlós–Szemerédi [1980]) There is a constant c such

that R(3 , ¾) ≤ c¾ 2/lg ¾ whenever ¾ > 1.
Proof: We prove that c = 8 suffices. Consider a graph G with 8¾ 2/lg ¾ vertices.
Suppose that G contains no triangle. If (G) ≥ ¾ , then the neighborhood of a
maximum-degree vertex contains an independent set of size ¾ . Otherwise, with
(G) < ¾ , Theorem 14.1.16 applies to guarantee an independent set of size at
least (| V(G)| lg ¾)/8¾ . With | V(G)| = 8¾ 2/lg ¾ , we obtain (G) ≥  .

It turns out that this relatively easy upper bound has the right order. We
will obtain the weaker lower bound R(3 , ) ≥ ( 2/lg 2 ) from Erdős [1961] in
Theorem 14.2.22, but K im [1995] proved R(3 , ) = ( 2/lg ).
We close this section with two unusual applications, showing the flexibility
and creativity that may be involved in applying probabilistic arguments to com-
binatorial problems. First we obtain Binet ’s Formula for the adjusted Fibonacci
numbers without directly solving the Fibonacci recurrence.

14.1.18. Theorem. If an is the number of 1,2-lists with sum n, then

√ n+1 √ n+1
1 ⎛1 + 5⎞ 1 ⎛1 − 5⎞
an = √ −√ .
5⎝ 2 ⎠ 5⎝ 2 ⎠
Section 14.1: Existence and Expectation 665

Proof: (Benjamin–Levin–Mahlburg–Quinn [2000]) Model each 1 , 2-list as a row

of squares and dominoes (see Example 2.1.2), forming a tiling. Generate a tiling
randomly by iteratively letting the next tile be a√square with probability 1/
and
a domino with probability 1/
2 , where
= (1 + 5)/2. Note that 1/
+ 1/
2 = 1.
The experiment generates an infinite tiling. The probability that it gener-
ates a fixed 1 , 2-list with sum n in the initial positions is 1/
n (this is the product
of the probabilities for the successive tiles, given what has already happened).
Since there are an such lists, which are disjoint events, the probability qn that
the experiment produces a break in the tiling after position n is an/
n .
A tiling is unbreakable after length n if and only if it adds a domino following
a tiling of length n − 1; the probability of this is qn−1/
2 . Thus qn = 1 − qn−1/
2 .
√ yields qn = 1 −
√ +
− · · · + (−
) .
Since q0 = 1, iteration −2 −4 −2 n

With = (1 − 5)/2 and  = (1 − 5)/2, we have  = −1, so −−2 = /.

Evaluating the geometric sum yields
1 − (−−2)n+1 1 − (/)n+1
qn = = .
1 − (−−2) 1 − /)
Now multiplying by n yields the same formula as Example 2.2.4:
n+1 −  n+1 1 n+1
an = n qn = =√ [ −  n+1 ] .
−  5

The next problem, posed by Kearnes–K iss [1999], succumbs surprisingly eas-
ily to a clever probabilistic argument. This time, instead of using max X ≥ (X),
we use (X) ≥ min X , but the definition of the random variable is not obvious.

14.1.19. Definition. Let A be the cartesian product of n finite sets A1 , . . . , An .

A proper subproduct of A is a set B of the form B1 × · · · × Bn , where Bi is
a proper subset of Ai for 1 ≤ i ≤ n. A proper dissection of A is a partition
of A into proper subproducts; we call these “pieces” of the dissection.

It is not hard to show that proper dissections of 1- and 2-dimensional grids

need at least two and four pieces, respectively. Note that the definition of sub-
product places no order on the elements of the factors, so the pieces need not be
geometrically connected.

• • • • •
• • • • •
• • • • •
• • • • •

14.1.20. Theorem. (Alon–Bohman–Holzman–K leitman [2002]) A proper dissec-

tion of a cartesian product of n finite sets has at least 2 n pieces.
Proof: Let B1 , . . . , B¾ be the pieces in a proper dissection of A1 × · · · × An ; recall
that B j = B1j ×· · ·× B¾j . We show that  ≥ 2 n . For each i ∈ [n], let Ri be a subset of
Ai chosen uniformly at random from among the subsets of Ai with odd size, and
let R = R1 × · · · × Rn .
666 Chapter 14: The Probabilistic Method

For j ∈ [¾], let X j be the 0,1-variable that is 1 if ##### B j ∩ R##### is odd. Let X =
# # # j #
∑j X j . Since #### B j ∩ R#### is odd if and only if #### Bi ∩ Ri #### is odd for all i ∈ [n], and
# #
(#### Bi ∩ Ri#### is odd) = 12 because Bi is a nonempty proper subset of Ai , we have
j j

(##### B j ∩ R##### is odd) = 2−n. Hence (X j) = 2−n and (X) = ¾ 2−n.

On the other hand, since B1 . . . , B¾ is a dissection of A1 × · · · × An ,

X = ∑ X j ≡ ∑ #### B j ∩ R#### ≡ | R| ≡ 1 (mod 2).

# #
j j

Thus always X ≥ 1. We conclude ¾ 2−n = (X) ≥ 1, so ¾ ≥ 2 n .

EXERCISES 14.1

14.1.1. (−) Choose a random ¾-subset of [n], with each ¾-set equally likely. Determine the
probability that the element 1 is chosen.
14.1.2. (−) Consider clubs A and B of size n. Each A sends an invitation to a B chosen at
random. What is the asymptotic probability that each B receives one?
14.1.3. (−) Construct a random variable and distribution on  so that the expectation of
the random variable is undefined (infinite).
14.1.4. (−) A university study finds that on any given day, permit holders fail to drive to
campus with probability .1, independently. The university therefore sells ten permits for
a lot with nine spaces and 20 permits for a lot with 18 spaces. Which lot is more likely to
turn away the last permit holder who arrives? (adapted from Grimmett–Stirzaker [1992])
14.1.5. (−) Compute the expectations of the following quantities.
(a) The number of fixed points in a random permutation of [n].
(b) The number of male/female pairs when 2n people consisting of n men and n women
are randomly partitioned into n pairs.
(c) The number of vertices of degree ¾ in a graph with vertex set [n] where each edge
occurs with probability p, independently.
14.1.6. (−) Determine the expected number of monochromatic copies of K r,r in a random
2-coloring of the edges of K n ,n .
14.1.7. (−) Obtain an exponential lower bound on the minimum number of edges in a ¾-
uniform hypergraph that is not t-colorable.
14.1.8. (−) Prove that some n-vertex tournament has at least n!/2 n−1 spanning paths.
(Comment: Alon [1990] proved that the maximum number of Hamiltonian paths in an
n-vertex tournament is at most n!/(2 − o(1))n .) (Szele [1943])
14.1.9. Given a graph G with p vertices, q edges, and s automorphisms, let n = (s¾ q−1)1/p .
Prove that some ¾-coloring of E(K n) has no monochromatic G. (Chvátal–Harary [1973])
14.1.10. Sperner’s Theorem by random variables. Let F be an antichain of subsets of [n].
Let X be the number of members of F that occur as initial segments of a random permu-
tation of [n]. Use X to prove | F | ≤ (⌈n/2⌉
n
).
14.1.11. Fix  , r ∈ with r ≥  . Determine the value m such that a random -coloring
of the vertices guarantees that every r-uniform hypergraph with at most m edges has a
vertex -coloring such that every color appears in every edge. (Alon–Spencer [2008])
Exercises for Section 14.1 667

14.1.12. (♦) Determine the expected number of descents in a random permutation (Defi-
nition 3.1.22). (Comment: Since the number of runs in a permutation is one more than
the number of descents, this gives an easier proof for Exercise 3.1.36.)
14.1.13. Let S be a finite set of finite binary words such that none is a prefix of another.
Prove that ∑a∈S 2−|a| ≤ 1.
14.1.14. An instance of SATISFIABILITY is a list of clauses, where each clause is a
set of literals (a literal is a variable or its negation). The instance is satisfied when each
variable is set to be TRUE or FALSE so that each clause has at least one true literal.
(a) Prove that an instance with clauses e1 , . . . , e m is satisfiable if ∑i=1 2−|ei | < 1.
m

(b) Construct an unsatisfiable instance with 2 ¾ clauses when all clauses have size ¾ .
14.1.15. (♦) Let n¾ be the least number of vertices in a non- ¾-choosable bipartite graph,
and let m¾ be the least number of edges in a non-2-colorable ¾-uniform hypergraph.
(a) Let H be a ¾-uniform hypergraph with n edges. Prove that H is 2-colorable if and
only if K n ,n is L-colorable when each part has E(H) as its color lists. Conclude n¾ ≤ 2m¾ .
(b) Let G be a non- ¾-choosable X , Y -bigraph. Let H be the hypergraph whose edges
are the lists in a ¾-uniform assignment L such that G has no L-coloring. Prove m¾ ≤ n¾ .
(Erdős–Rubin–Taylor [1979])
14.1.16. (♦) Apply 2-colorability of hypergraphs to prove that if ¾ > 1 + lg n, then every
n-vertex bipartite graph is ¾-choosable. Obtain the same result when ¾ > lg n.
14.1.17. Let H be a hypergraph with no edges e1 , . . . , e¾ and vertex x such that ei ∩ ej =
{x} for i , j ∈ [¾]. Prove that H is ¾-choosable. Strengthen the statement to an analogue of
degeneracy in graphs. (Hint: No probability needed.) (Berge [1973], from Tomescu [1968])
14.1.18. A tripartite 3-uniform hypergraph is a hypergraph whose vertex set parti-
tions into sets X , Y , and Z such that each edge has one vertex in each of those sets. Con-
struct a tripartite 3-uniform hypergraph that is not ¾-choosable.
14.1.19. The first passenger to board a plane loses her boarding pass and sits in a random
seat. Each subsequent passenger sits in his or her own seat if it is available and other-
wise chooses a random seat from those that remain. What is the probability that the last
passenger sits in his or her own seat? (Winkler [2004])
14.1.20. (♦) Prove that a tournament cannot contain three 3-cycles on four vertices. Con-
clude that the maximum number of edges in an n-vertex 3-uniform hypergraph having no
set of four vertices containing three edges is at least 14 (3n).
14.1.21. A bag has 1000 tiles; on each is printed one of the 26 letters of the alphabet.
Player A picks a tile at random and then replaces it. Player B then picks a tile at random.
Prove that the probability of the two players picking the same letter is at least 1/26.
14.1.22. A communication network is configured as a complete ¾-ary tree with leaves at
distance l from the root. At any time, each node is working with probability p, indepen-
dently of all others. When a node is not working, the subtree rooted at it is inaccessible.
What is the expected number of accessible nodes?
14.1.23. (♦) Let all nn−2 trees with vertex set [n] be equally likely. Define the following
random variables: X is the number of leaves, Y is the number of vertices of degree 2, and
Z is the number of leaves having neighbors of degree 2. Determine the limits as n → ∞ of
(X)/n, (Y)/n, and (Z)/n. (Hint: Use an encoding of the trees as n-ary lists.)
14.1.24. (♦) A pill bottle has m large pills and n small pills. Each day a patient chooses a
random pill; a small pill is consumed, but a large pill is split , with half consumed and the
other half becoming a small pill. Determine the expected number of small pills after the
last large pill is split and the expected day this occurs. (Knuth–McCarthy [1991])
668 Chapter 14: The Probabilistic Method

14.1.25. Let n distinct pairs of socks be put into the laundry. After washing, socks are
drawn successively at random (without replacement). Compute the expected number of
pairs among the first ¾ socks drawn.

14.1.26. For n , ¾ ∈ 0 with 0 ≤ ¾ ≤ n, form a permutation of [n] by choosing the first ¾
positions at random and filling the remaining n − ¾ positions in ascending order. Let En ,¾
be the expected number of left-to-right maxima (for example, E3 ,1 = 2 and E3 ,2 = 11/6).
Compute En ,¾ . Conclude that En+1 ,¾ − En ,¾ = 1/(¾ + 1). (Deshpande–Deshpande [2008])
14.1.27. (♦) Bipartite subgraphs.
(a) Use a random partition of the vertices to prove that every graph has a bipartite
subgraph with at least half its edges.
(b) Improve part (a) by showing that if G has m edges and n vertices, then G has a bi-
⌈n/2⌉
partite subgraph with at least m2 ⌈n/2⌉−1
edges whose part-sizes differ by at most 1.
(c) Prove that every graph with m edges that has a matching with ¾ edges has a bi-
partite subgraph with at least (m + ¾)/2 edges. (Molloy–Reed [2002, p. 37])
14.1.28. Let G be an n-vertex graph with average degree d and minimum degree ¾ . Prove
that G has an induced subgraph with at least d¾+n1 vertices that does not contain K ¾+1 .
14.1.29. (♦) Bollobás’s Inequality. Let {Ai }im=1 and {Bi }im=1 be subsets of [n], with | Ai | =
−1
ai , | Bi | = bi , and Ai ∩ Bj = ∅ if and only if i = j . Prove ∑i=1 ( aia+ibi ) ≤ 1. Conclude
m

that the maximum size of an antichain in 2 n is (⌈n/2⌉n

) (Sperner ’s Theorem). (Hint: Define
appropriate events over random permutations of [n].) (Bollobás [1965])
14.1.30. (♦) The Coupon Collector Problem (Geometric random variable).
(a) An experiment has success probability p on each trial, independently. Obtain a
simple formula (no summation) for the expected number of trials up to the first success.
(b) Every box of a type of candy contains one of n prizes, each with probability 1/n.
The grand prize requires obtaining each of these at least once. Prove that the expected
number of the box on which the last prize is obtained is n ∑i=1 1/i. (Feller [1968, p. 255])
n

14.1.31. (♦) Given a distribution of pebbles to the vertices of a graph, a pebbling move
removes two pebbles from some vertex and adds one pebble to a neighboring vertex. A ver-
tex is reachable from a distribution if some (possibly empty) list of pebbling moves results
in it having a pebble. Let D be a distribution on the ¾-cube Q¾ from which every vertex
is reachable. Moews [1998] proved that D must have at least (4/3)¾ pebbles. Here we de-
velop the proof in Bunde–Chambers–Cranston–Milans–West [2008].
(a) Prove ∑t≥0 ar,t 2− t ≥ 1 for any r ∈ V(Q¾), where ar,t is the number of pebbles in D
at distance t from r. Thus ∑t≥0 (ar,t)2− t ≥ 1 when r is chosen uniformly at random.
(b) Prove (ar,t) = | D| (¾t )2−¾ , and conclude | D| ≥ (4/3)¾ .
14.1.32. A coloring of a graph G is r-dynamic if it is proper and each vertex neighborhood
N(v) has at least min{r, d G(v)} colors. Prove that if (G) > sr++1s r ln n, where n = | V(G)| ,
then G has an r-dynamic coloring with at most (G)+ r + s colors. (Hint: Color the vertices
properly in the order v1 , . . . , vn , at random.) (Jahanbekam–Kim–O–West [2016])
14.1.33. Two players alternately flip a coin that has probability p of landing heads up.
The winner is the first to obtain heads. What is the probability that Player 1 wins?
14.1.34. (♦) A standard random walk moves distance 1 to the left or right on the real line
with each unit of time, with probability 12 in each direction.
(a) From position 0, prove that the expected number of steps taken to reach position
−1 or position n for the first time is n.
(b) Let X n be the number of steps until the difference between the maximum and min-
imum positions visited first equals n. Determine (X n). (Palacios–Sandell [1991])
Exercises for Section 14.1 669

14.1.35. (♦) A random walk on a graph G starts at a fixed vertex v and at each step moves to
a random neighbor of the current vertex, chosen uniformly. When G is a cycle, prove that
all vertices other than v are equally likely to be the last vertex visited for the first time.
(Comment: This folklore result appeared in Lovász–Winkler [1993], which also showed
that cycles and complete graphs are the only graphs with this property.)
14.1.36. An unbiased coin is flipped until ¾ heads occur. Let X be the number of isolated
heads in the list , and let Y be the number of runs (when ¾ = 4, the list HHTTHTTH
yields X = 2 and Y = 5). Compute (X) and (Y). (Bhanu–Deshpande [2010])
14.1.37. Consider a committee having both Senators from each of n states. Partition the
2n people into pairs at random, with each matching equally likely. All pairs are allowed.
(a) What is the probability that no Senators from the same state are paired?
(b) What is the limit of this probability as n → ∞ ?
14.1.38. (♦) An urn has n balls, of which ¾ are white. A ball is chosen at random until
a white ball is obtained. Determine the expected number of the step on which the first
white ball is obtained, under each of the following two scenarios.
(a) Non-white balls are put back in the urn before the next drawing.
(b) Non-white balls are discarded before the next drawing.
14.1.39. Let Q be a set of n points chosen independently and uniformly in the unit square.
Find the expected number of sets S ⊆ Q such that there is an increasing curve with S on
or below the curve and Q − S above it. (Stanley–Steele [1989])
14.1.40. (♦) Let v1 , . . . , vn √ be unit vectors in n . Prove that √ 1 , . . . , n can be chosen in
{−1 , +1} so that ####∑ i vi #### ≥ n and also so that ####∑ i vi #### ≤ n. Prove that both results are
sharp. (Alon–Spencer [1992])
14.1.41. (♦) Fix p1 , . . . , pn ∈ [0 , 1], and let v1 , . . . , vn be unit vectors in n . Let w = ∑ pi vi .
√
Prove that some subset of v1 , . . . , vn sums to a vector u such that |w − u| ≤ n/2. (Alon–
Spencer [1992])

14.1.42. (♦) Let A be an n-set in n2 . Prove that there is an n-set B in n2 such that at

least half of n2 can be expressed in the form a + b with a ∈ A and b ∈ B.
14.1.43. (+) A set S is sum-free if no distinct x , y ,  ∈ S satisfy x + y = . Prove that
every set A of nonzero integers contains a sum-free set S with | S| > | A| /3. (Erdős [1965])
14.1.44. Let A1 , . . . , A t be events in a space of n equally likely outcomes, with pi = (Ai) =
| Ai | /n and s = ∑i=1 pi . Given r ∈  with 1 ≤ r ≤ s, prove that for some set of r events in
t

s
{Ai }, the probability that they all occur is at least r!(r) / (rt ) . (Ford [1994])
14.1.45. Determine the probability that n points independently chosen uniformly at ran-
dom on a circle all lie in a semicircle. (Comment: This result appeared in at least three
papers and six books by the 1950s; Jordan [1872/1873] and Wendel [1962] generalized it.)
14.1.46. Fix n ≥ 1. For each of n independent trials, a ball is placed in some indexed box; it

is box j with probability 2− j , for j ∈ . Let X denote the number of empty boxes below the

highest-indexed occupied box. Compute (X). (Comment: Ferguson–Melolidakis [1984]
determined the complete distribution of X for each n.)
14.1.47. (♦) A binary maze is a digraph in which every vertex has two exiting edges, one
labeled 0 and one labeled 1. A search list specifies the label of each successive edge to be
followed. A search list b is n-universal if for every vertex u in every n-vertex strongly
connected binary maze G, following b from u reaches every vertex of G. Prove that the
minimum length of an n-universal search list is between 2 n−1 and n2 2 n ln n. (Hint: Bound
the number of triples (G , u , v) such that G is a maze and a random binary list of length m
never reaches v when starting from u.) (Knuth [1999])
670 Chapter 14: The Probabilistic Method

14.2. Refinements of Basic Methods

In this section we continue to study existence-type arguments, introducing
some refinements that lead to stronger results.

DELETIONS AND ALTERATIONS

When a randomly generated object is close to having a desired property, a

slight alteration may produce that property. This technique has been called the
deletion method, the alteration principle, or the two-step method. Ramsey
numbers provide a classic example.

14.2.1. Theorem. R(¾ , ¾) > n − (¾n)21−(2) , where n ∈

¾
. In particular, R(¾ , ¾) >
(1 − o(1)) ¾e 2¾/2 .
Proof: Let X be the number of homogeneous ¾-sets in a random 2-coloring of
E(K n). Expressing X as a sum of indicator variables for each ¾-set and applying
linearity yields (X) = (¾n)21−(2). The pigeonhole property guarantees a coloring
¾

with at most this many homogeneous ¾-sets. By deleting a vertex of each homoge-
neous ¾-set, we retain a graph with at least the specified number of vertices but
no homogeneous ¾-set.
Thus R(¾ , ¾) > n− (X), for each n. Since (¾n) < ( ne
¾
)¾ , we also have the simpler
n − ( ne
¾
)¾ 21−¾(¾−1)/2 as a lower bound. We seek n to maximize this bound. Differ-
¾ −1
entiating suggests choosing n to satisfy 1 = ¾ ¾e ( ne ¾ ) 21−¾(¾−1)/2 , which requires
−
n = e ¾2
1 ¾/2(2e)−1/(¾ −1)
. The factor (2e)− 1/(¾ − 1)
is near 1 when ¾ is large.
For simplicity, we set n = e−1 ¾ 2¾/2 (actually a nearby integer) to obtain the
claimed bound. We have
ne ¾ ¾ 1 2e
n−( ) 21−¾(¾−1)/2 = 2¾/2 − 2¾ /2 1−¾(¾ −1)/2
= ¾ 2¾/2 (1 −
2
2 ).
¾ e e ¾
Since 2e/¾ tends to 0 for large ¾ , the bound is as claimed.

Deletion arguments can improve basic existence results,

√ although here the
improvement from Theorem 14.1.3 is only by a factor of 2. Both arguments can
also be applied to R(¾ , l) with ¾
= l (Exercise 16).
Next we add to a random almost-good-enough structure. In a graph G, a set
S ⊆ V(G) is dominating if every vertex outside S has a neighbor in S.

14.2.2. Theorem. (Arnautov [1974], Payan [1975]) For ¾ > 1, every n-vertex
graph with minimum degree ¾ has a dominating set of size at most n 1+ln(¾ +1)
¾ +1 .
Proof: (Alon [1990]) Form a random vertex subset S in such a graph by including
each vertex independently with probability p = ln(¾¾++11) . Given S, let T be the set
Section 14.2: Refinements of Basic Methods 671

of vertices outside S having no neighbor in S; adding T to S yields a dominating

set. The experiment provides both S and T ; we seek the expected size of the union
as an upper bound on the minimum size of a dominating set.
Since vertices appear in S with probability p, linearity yields (| S|) = np.
The random variable | T | is the sum of n indicator variables for whether individual
vertices belong to T . We have v ∈ T if and only if v and its neighbors all fail to
be in S. This has probability bounded by (1 − p)¾+1 , since v has degree at least ¾ .
Since (1 − p)¾+1 < e− p(¾+1) , we have (| S| + | T |) ≤ np + ne− p(¾+1) = n 1+ln(¾ +1)
¾ +1 . The
pigeonhole property of the expectation completes the proof.

A greedy algorithm that iteratively adds the vertex dominating the most
currently undominated vertices proves the same upper bound constructively (Ex-
ercise 8). The coefficient on n in Theorem 14.2.2 is asymptotically sharp; Alon–
Wormald [2010] gave a probabilistic construction of ¾-regular graphs with no
dominating set of size less than (1 − o(1)) ln¾¾ n.
The most famous use of the deletion method may be Erdős’s proof that graphs
with large girth and chromatic number exist. Theorem 10.1.31 presents an ex-
plicit construction. Here we present Spencer ’s simplification of Erdős’ proof, with
apologies for advance use of Lemma 14.3.8.

14.2.3. Theorem. (Erdős [1959]) Given ¾ ≥ 3 and ≥ 3, there exists a graph

with girth at least and chromatic number at least  .
Proof: We generate graphs with vertex set [n] and edge probability p. Since
(G) ≥ (G) n
, we choose p large to make large independent sets unlikely. We also
choose p small to make the expected number of short cycles (length less than )
small. From a graph satisfying both conditions, we will delete a vertex from each
short cycle to obtain the desired graph.
To have few short cycles, let p = nt−1 , where t < 1/ . Each possible cycle of
length j occurs with probability p j . By choosing and cyclically ordering j ver-
n
tices, 2(j)j such cycles are possible, so the total number X of cycles of length less
than satisfies
 −1  −1
n(j) j nt j
(X) = ∑ p ≤∑ < n( −1)/ .
2j 2j
j =3 j =3

1−1/
With (X) < n , we have (X)/n → 0 as n → ∞. From Markov ’s Inequality
(Lemma 14.3.8), we conclude (X ≥ n2 ) → 0 as n → ∞. For n large, (X ≥ n2 ) < 12 .
Since (G) cannot grow when we delete vertices, at least n−(G) X
independent
sets are needed to color the remaining graph after we delete a vertex of each short
cycle. If X < n2 and (G) ≤ 2n , then at least  colors are needed for the graph
remaining. By the Union Bound,

((G) ≥ r) ≤ (nr)(1 − p)( ) < [ne− p(r−1)/2]r .

r
2

The bound tends to 0 as n → ∞ when, for example, r = ⌈ 3p ln n⌉ .

With r = ⌈ 3n1−t ln n⌉ and  fixed, we can choose n large enough to obtain
r < 2n . If also n is large enough so that (X ≥ n2 ) < 12 and ((G) ≥ r) < 12 , then
672 Chapter 14: The Probabilistic Method

there exists an n-vertex graph G such that (1) (G) ≤ 2n¾ and (2) G has fewer than
2 cycles of length less than ½ . We delete a vertex from each short cycle and retain
n

a graph with girth at least ½ and chromatic number at least ¾ .

The deletion method is also used in the proof of a technique called “depen-
dent random choice”. To motivate this, recall that the Tur án number ex (n; H)
is the maximum number of edges in an n-vertex graph not containing H. In
Theorem 13.2.19, a counting argument showed ex (n; K r,s) ≤ cn2−1/r when r ≤ s
(K óvari–Sós–Turán [1954]), which is the right order of growth when s > r! (The-
orem 13.2.25, Kollár–Rónyai–Szabó [1996]). Using a counting argument, Füredi
[1991] extended this bound on ex (n; H) to every A , B-bigraph H such that ver-
tices in B have degree at most r. Here we develop a later probabilistic proof.
The elementary counting argument shows that if G has too many edges,
then some set of r vertices has s common neighbors, yielding K r,s . That is, with
nonzero probability a random r-set from V(G) has at least s common neighbors.
In the Dependent Random Choice Lemma, we introduce some dependence among
the chosen vertices so that r-sets with many common neighbors are more likely to
be kept. In this way a smaller number of edges yields nonzero probability of hav-
ing many r-sets with at least s common neighbors. The comparison with Theorem
14.2.3 is that there we deleted a vertex from each bad cycle (too short), while here
we will delete a vertex from each bad r-set (too few common neighbors).
Versions of the lemma were proved in Kostochka–Rödl [2004] and Gowers
[1998]. We give roughly the proof by Alon–Krivelevich–Sudakov [2003] (see
Alon–Spencer [2008], Lecture 2 of Lee [2015], and Fox–Sudakov [2011]; the lat-
ter is an excellent survey presenting a variety of applications of the technique).

14.2.4. Lemma. (Dependent Random Choice) Let G be an n-vertex graph


with m edges. For a , b ∈ , if there is a positive integer t such that
(2m/n)t n b t
− ( )( ) ≥ a
nt−1 r n
then G contains a set U of at least a vertices such that every r vertices in U
have at least b common neighbors in G.
Proof: We use linearity of expectation repeatedly in the proof.
Let T consist of t vertices chosen independently at random (repetition al-
lowed). Let S be the set of vertices neighboring all of T , and let X = | S|. Since
v ∈ S if and only if T ⊆ NG(v), we have (v ∈ S) = (d(v)/n)t . Therefore (X) =
n−t ∑v∈V(G) d(v)t . By convexity, 1n ∑ d(v)t ≥ (∑ d(v)/n)t , so (X) ≥ (2m/n)
t

n t−1
.
Now let Y count the r-subsets of S having fewer than b common neighbors; we
want to eliminate such sets. A given r-set R lies in S if and only if each vertex of T
t
lies in N(R). If | N(R)| < b, then (R ⊆ S) < ( nb ) . This is where dependence helps;
r-sets having more common neighbors are more likely to lie in S. Considering all
r-sets, we have (Y) < (nr) ( nb ) .
r

Under the given hypothesis, (X − Y) ≥ a. Hence X − Y ≥ a for some outcome

T of the experiment. If we delete, from the resulting set S, one vertex of each r-
set having fewer than b common neighbors, we will still be left with a set U of at
least a vertices in which every r-set has at least b common neighbors.
Section 14.2: Refinements of Basic Methods 673

14.2.5. Theorem. (Füredi [1991]) If H is an A , B-bigraph such that d H (v) ≤ r

for all v ∈ B then ex (n; H) < c H n2−1/r for some constant c H .
Proof: (Alon–Krivelevich–Sudakov [2003]) Let a = | A| and b = | V(H)| , and set
t = r. Let G be an n-vertex graph with at least 12 cn2−1/t edges, where ct ≥ br! + a.
t

This hypothesis on the constant c yields the hypothesis on G for Lemma 14.2.4.
Lemma 14.2.4 therefore provides a set U of at least a vertices in G such that
every r-subset of U has at least b common neighbors. We now find H in G. Use
a vertices of U as the vertices of A, assigned arbitrarily. Now, each vertex of B
needs to be made adjacent to some set of at most r vertices of A. For each vertex of
B in turn, find an r-set of vertices mapped to A containing its desired neighbors,
and find a common neighbor of this r-set that has not yet been used. Since the
r-set has at least b common neighbors, and b = | V(H)| , there is always such a
neighbor available until the process is finished.
Hence it suffices to let c H = c/2.

Other applications of dependent random choice appear in Exercises 12–13.

Fox–Sudakov [2011] described further enhancements of the method with many
impressive applications in additive number theory and extremal graph theory.

We mention only a result on Ramsey numbers. For d ∈ , Theorem 11.1.18 guar-
antees a constant cd such that R(G , G) ≤ cd | V(G)| whenever (G) = d. Proved
using the Regularity Lemma, the constant there is a tower of height d. Re-
mark 11.1.19 discusses improvements avoiding the Regularity Lemma. In par-
ticular, Fox–Sudakov [2009] used dependent random choice to prove R(G , G) ≤
8d2 d | V(G)| when G is bipartite with (G) = d ≥ 1; this is essentially sharp by a
construction of Graham–Rödl–Rucinski [2001].
We consider one more alteration method: a second random step. When a
random coloring is bad in a few places, we can recolor those elements. For the
smallest non-2-colorable ¾-uniform hypergraph, Beck [1978] used this technique
to improve the simple lower bound 2¾−1 (Proposition 14.1.6) to (2¾ ¾ 1/3). We
present a simplification from Alon–Spencer [1992] that yields a slightly weaker
lower bound. A further improvement to (2¾ ¾ 1/2(ln ¾)−1/2) by Radhakrishnan–
Srinivasan [2000] appears in Alon–Spencer [2000, 2008], with another proof by
Kozik–Cherkashin [2015] in Alon–Spencer [2016].

14.2.6.* Theorem. (Beck [1978]) If 2te− p¾ + t2 pe p¾ < 1, where 0 ≤ p ≤ 1

2, then
every ¾-uniform hypergraph with 2¾−1 t edges is 2-colorable.
Proof: (Alon–Spencer [1992]) Given such p and t, consider a ¾-uniform hyper-
graph H with 2¾−1 t edges. Let º : V(H) → {0 , 1} be a random 2-coloring; each
vertex gets each color with probability 12 , independently. Call a monochromatic
edge a bad edge. For each vertex in at least one bad edge under º , change its color
with probability p (independently), obtaining a new coloring º . We show that
with positive probability º is a proper 2-coloring.
First consider the probability that an edge S is bad under both º and º .
Colors under º are the same, and then the vertices in S all keep or all change
their color. We compute
2 · 2−¾ (p¾ + (1 − p)¾) ≤ 21−¾ · 2(1 − p)¾ ≤ 21−¾ · 2e− p¾ .
674 Chapter 14: The Probabilistic Method

With 2¾−1 t edges, this occurs for some edge with probability at most 2te− p¾ , by
the Union Bound.
Now consider an edge S that is bad (in one color) under º but not º . Since S
requires a color change to go bad, some edge T intersecting S has the other color
on all its vertices (under º ). Given edges S and T that intersect, let AS,T be the
event that S becomes monochromatic and the vertices of S ∩ T change color.
Let U be the subset of S that changes color in the recoloring step; occurrence
of AS,T requires S ∩ T ⊆ U. The original coloring º gives one color to all of U ∪ T
and the other to all of S − U ; this has probability 2( 12 )2¾−r , where r = | S ∩ T |.
With j = |U | , the probability that the colors on U change in the recoloring step
is at most p j , and the probability of no change on S − U is at most 1.
For specified S, T, U , the probability is thus at most 21−2¾+r p j . We sum over
U with S ∩ T ⊆ U ⊆ S to bound (AS,T ). Letting i = j − r and using the Binomial
Theorem and 1 + x < e x , the bound becomes
¾ −r
−r 2p r
∑( ) (21−2¾+r pr ) pi = 21−2¾+r pr (1 + p)¾−r ≤ 21−2¾ e p¾ ( ) .
i 1+p
i=0

Since 0 ≤ p ≤ 12 , the last expression is maximized when r = 1. Thus (AS,T ) ≤

22−2¾ e p¾ p for every intersecting pair (S, T) of edges. Since there are fewer than
(2¾−1 t)2 intersecting ordered pairs of edges, by the Union Bound the probability
that some edge is bad under  but not under  is at most t2 e p¾ p, which is the
second term in the hypothesis.
Therefore, when the specified hypotheses hold, the probability that  is a
proper 2-coloring is positive.

14.2.7.* Corollary. For  > 0, when is sufficiently large the minimum num-
ber
√ of edges in a -uniform hypergraph that is not 2-colorable is at least
( 3 − )(2 ¾ − 1 1/3(ln )−1/2).
√
Proof: Choosing t < ( 3 − ) 1/3(ln )−1/2 and p = ln ¾
3¾ yields the inequality
2te− p¾ + t2 pe p¾ < 1 needed for Theorem 14.2.6.

THE SYMMETRIC LOCAL LEMMA

In simple existence arguments showing that the probability of having some

bad event is less than one, parameters are chosen so weakly that even when the
events are disjoint there remains probability outside their union. In constrast,
independent events always have nonzero probability in the complement. We seek
a method that can accommodate some lack of independence and still guarantee
nonzero probability outside all the events when the events have sufficiently small
probability (we will make this precise).
The resulting technique is the Lovász Local Lemma or simply the Local
Lemma. Discussion of the lemma and its applications appears in Alon–Spencer
[1992, 2000, 2008, 2016], Molloy–Reed [2002], and Szegedy [2013]. We follow
Molloy–Reed by starting with the simpler symmetric version that is easier to un-
derstand and suffices for many applications. Subsequently we prove the general
version (Theorem 14.2.18) and derive the symmetric version from it.
Section 14.2: Refinements of Basic Methods 675

14.2.8. Definition. Let A1 , . . . , An be events. For disjoint subsets S and T of

[n], the compound event A(S, T) specifies the occurrence of Ai for i ∈ S and
the non-occurrence of A j for j ∈ T . An event B is mutually independent of
A1 , . . . , An if B is independent of each compound event specified by disjoint
subsets of [n] (meaning (B | A(S, T)) = (B)).

14.2.9. Theorem. (Symmetric Local Lemma; Erdős–Lóvasz [1975]) Consider

events A1 , . . . , An such that (1) (Ai) ≤ p for all i, and (2) each Ai is mutually
independent of a set of all but d events. If epd ≤ 1, then  (⋂ Ai ) > 0.

Shearer [1985] proved that replacing the constant e with any smaller con-
stant in the condition epd ≤ 1 no longer guarantees  (⋂ Ai ) > 0.
Most treatments of the Symmetric Local Lemma replace “all but d events”
with “all but d other events”) but then must require ep(d + 1) ≤ 1. We phrase the
condition using “all but d events” because the formula to count events omitted
from the set that is mutually independent of Ai often also counts Ai .
The Local Lemma has many applications. We use it first to improve the lower
bound on diagonal Ramsey numbers produced by the elementary existence argu-
ment, but only by a factor of 2. Gaining only a factor of 2 is disappointing; the
Local Lemma is more effective when dependence is rarer. The improvement for
van der Waerden numbers is more dramatic (Exercise 22). The general version of
the Local Lemma does provide a significant improvement for off-diagonal Ramsey
numbers with one parameter fixed (see Theorem 14.2.22).
√
14.2.10. Theorem. (Spencer [1975]) R(¾ , ¾) > (1 + o(1)) e ¾2
2 ¾/2 .

Proof: Form a random 2-coloring of E(K n), as usual. For each ¾-set S of vertices,
let AS be the event that S is homogeneous. Knowing the color of all edges outside
(S2 ) has no effect on the probability of AS . Hence we can let d in Theorem 14.2.9
be the number of ¾-sets in [n] (including S) that share at least two elements with
S; the event AS is mutually independent of the set of all other events.
To obtain an upper bound on d, pick two elements of S and pick ¾ − 2 elements
outside these two. Sets that share more than two elements with S are counted
many times. Nevertheless,
¾ n−2 ¾ n ¾2 ne ¾−2
d < ( )( ) < ( )( )< ( ) .
2 ¾−2 2 ¾−2 2 ¾−2

As in Theorem 14.1.3, (AS) = 2 · 2−(2) for all S; let this probability be p.

¾

Fix ¾ , so p is fixed. By making n small, we can make d small. To apply Theo-

rem 14.2.9, it suffices to make n small enough so that
¾2 ne ¾ −2 1 1
( ) < = 2¾/2(2¾/2)¾−2 .
2 ¾−2 ep 2e
When this holds, the probability of having no monochromatic copy of K ¾ is
nonzero, and hence R(¾ , ¾) > n.
√ ¾ −2 √
1/(¾ −2) ¾ −2
Since 2¾/2 = 2 2 , it suffices to have n ≤ c e2 ¾ 2¾/2 , where c = ( e2¾ 2 ) ¾ .
Since c → 1 as ¾ → ∞ , the claimed bound holds.
676 Chapter 14: The Probabilistic Method

In the proof of Theorem 14.2.10, we could tighten the bound on d to

(n¾ ) − (n−¾ ¾ ) − ¾ (¾n−−1¾), but this would not asymptotically improve the result.

14.2.11. Remark. Independence vs. mutual independence. It is wrong to say “ B

is independent of A1 , . . . , An ” when applying the Local Lemma, because it is not
generally true that B is mutually independent of A1 , . . . , An when B is indepen-
dent of each of A1 , . . . , An .
For example, in Theorem 14.2.10, each event AS is (pairwise) independent of
all events for sets sharing at most two vertices with it, because when the color of
one edge is known, the probability that S is homogeneous remains ( 21 )(2)−1 . How-
¾

ever, AS is not mutually independent of that set of events. Let T and U be ¾-sets
such that S ∩ T = {v , x}, S ∩ U = {v , y}, and T ∩ U = {v , } (see figure below).
The events AS , A T , and AU are pairwise independent, as noted above. However,
AS is not mutually independent of {A T , AU }. If A T and AU both occur, then vx
and vy have the same color, and (AS | { A T , AU }) = ( 12 )(2)−2
= (AS).
¾

An intuitive illustration of pairwise independence not guaranteeing mutual

independence is given by the “Borromean rings”, shown below on the right. Any
two of the three rings are “independent ” in the sense that they are not linked.
However, the three rings cannot be separated.

S
x
• v
•y
•
T U
•

The Local Lemma is effective when the sets on which an event Ai depends
are somehow “near ” Ai ; hence the name “Local Lemma”. Often the needed state-
ment about mutual independence uses the following principle, named by Molloy
and Reed. It describes a setting under which mutual independence is automatic.

14.2.12. Proposition. (Mutual Independence Principle) Let Z1 , . . . , Zm be

independent experiments and A1 , . . . , An be events such that occurrence of
each Ai is determined by a subset Si of Z1 , . . . , Zm . If Si is disjoint from
Sj 1 ∪ · · · ∪ Sj ¾ , then Ai is mutually independent of { A j 1 , . . . , A j ¾ }.

Intuitively, Proposition 14.2.12 is making the observation that no specifica-

tion of whether each of A j 1 , . . . , A j ¾ does or does not occur can affect the prob-
ability of Ai . In Theorem 14.2.10, Z1 , . . . , Zm are the independent “coin-flips”
used to determine the colors on the m edges of K n . Hence Proposition 14.2.12 ap-
plies to show that AS is mutually independent of all sets of events whose members
share at most one edge with S.
Section 14.2: Refinements of Basic Methods 677

We next apply the Local Lemma to list coloring. As in Proposition 8.1.12,

greedy coloring with vertices in a least-last degree order shows that every ¾-
degenerate graph is (¾ + 1)-choosable. This bound makes no use of the structure
of the vertex lists. When the color lists do not overlap much, smaller lists suffice,
regardless of the vertex degrees.

14.2.13. Theorem. (Reed [1999]) Let L be a list assignment for a graph G such
¾
that each list has size at least ¾ . If every color appears in at most 2e of the
lists in any vertex neighborhood, then G has a proper coloring from the lists.
Proof: We may assume that each list has size exactly ¾ . Form a random coloring
by choosing uniformly and independently from each list. For each edge xy and
color c ∈ L(x) ∩ L(y), let A xy ,c be the event that x and y both receive color c. The
probability p of each such event is ¾ −2 .
Occurrence of an event A xy ,c is determined by the colors chosen for x and y.
By the Mutual Independence Principle, A xy ,c is mutually independent of the set
of all these events indexed by edges not incident to xy. We obtain an upper bound
on d by counting the events corresponding to edges incident to xy. Note that
there is no bound on the vertex degrees.
For an edge x y incident to xy, the event A x y ,c is defined if and only if c is
in L(x ) ∩ L(y ). By symmetry, suppose that x = x and y
= y. The hypothesis
allows at most 2e ¾
such events for each color c ∈ L(x), one of which is eliminated as
A xy ,c . The same computation holds for edges incident to y. We conclude that each
event A xy ,c is mutually independent of a set of all but d − 1 of the other events,
where d < ¾ 2/e. Since epd < 1, the Local Lemma implies that in some coloring
no such event occurs, and hence G has a proper coloring from the lists.

Choosing appropriate events for application of the Local Lemma can be tricky.
In Theorem 14.2.13, one might try defining an event A xy for having the same color
on x and y, but this fails (Exercise 24). Molloy–Reed [2002] gave various appli-
cations of Theorem 14.2.13. By other techniques, Haxell [2001] showed that the
conclusion still holds when each color appears in at most ¾/2 lists in each neigh-
borhood. By iterative use of the Local Lemma, Reed–Sudakov [2002] relaxed this
to ¾(1 − o(1)). Reed [1999] conjectured that ¾ − 1 suffices.
Before proving the general Local Lemma, we apply the symmetric version
again. Consider decomposition of graphs into linear forests, which are forests
whose components are paths. The minimum number of such forests needed is the
linear arboricity of the graph.

14.2.14. Conjecture. (Akiyama–Exoo–Harary [1981]) Every ¾-regular graph

has linear arboricity at most ⌈(¾ + 1)/2⌉ .

The bound is sharp, because a ¾-regular n-vertex graph has 12 ¾ n edges, and
each linear forest contains at most n − 1. Alon [1988] used the Local Lemma to
prove the conjecture asymptotically, showing that for some constant c the linear
arboricity is always at most 12 ¾ + c¾ 3/4(ln ¾)1/2 .
We present only the proof of the exact bound for graphs with large girth. As
a lemma, we need a variation on the greedy selection that finds independent sets
678 Chapter 14: The Probabilistic Method

of size ¾+n1 . We will be content with smaller independent sets but have some con-
trol on the locations of the vertices. (Haxell [2001] strengthened this result: the
sets only need to have size at least 2 ¾ .)

14.2.15. Lemma. (Alon [1994]) Given a graph G with maximum degree ¾ and a
partition of V(G) into sets of size at least 2e ¾ , there is an independent set in
G consisting of one vertex from each set in the partition.
Proof: Let a = ⌈ 2e¾ ⌉ , and let V1 , . . . , Vr be the partition. By discarding ver-
tices from G, we may assume that each Vi has size a. Pick one vertex from each
Vi , uniformly and independently. We want the resulting set S to have nonzero
probability of being independent.
For each edge xy with endpoints in distinct parts, let A xy be the event that
x , y ∈ S. We have (A xy) = a12 . By the Mutual Independence Principle, if x ∈ Vi
and y ∈ Vj , then A xy is mutually independent of the set of events for all edges
with both endpoints outside Vi ∪ Vj . The bound on d for the application of the
¾
Local Lemma is thus 2a¾ − 1. It suffices to have epd < 1. We have epd < e 2a a2
< 1,
by the choice of a.

Lemma 14.2.15 implies, for example, that when 11t points on a circle are
colored, with 11 points in each of t colors, there is a set of t points with distinct
colors such that no two are consecutive among the colored points.
Every connected 2 ¾-regular graph has an Eulerian circuit and hence an ori-
entation in which indegree and outdegree both equal ¾ at every vertex. Such a di-
graph is called ¾-regular. Decomposing the resulting digraph into ¾ + 1 forests of
directed paths yields the desired decomposition of the original 2 ¾-regular graph.
Thus it suffices to prove the following.

14.2.16. Conjecture. (Nakayama–Peroche [1987]) Every ¾-regular digraph de-

composes into ¾ + 1 linear forests.

14.2.17. Theorem. (Alon [1988]) Every ¾-regular digraph having no cycle of

length less than 8e¾ decomposes into ¾ + 1 linear forests.
Proof: Let G be such a digraph. Let H be the ¾-regular bipartite graph obtained
by splitting each vertex v of G into two vertices v− and v+ , inheriting the in-
coming and outgoing edges at v, respectively. The Marriage Theorem (Corollary
6.1.6) decomposes H into 1-factors, which correspond to a decomposition of G into
1-regular subdigraphs (the argument is the same as that in Theorem 6.2.13).
Each subgraph in the decomposition of G is a spanning union of disjoint cy-
cles. Deleting one edge from each cycle in each factor yields ¾ linear forests. If
the deleted edges form a matching, then they form a linear forest to complete the
desired decomposition.
Let G be the underlying undirected graph of G. A matching in G corre-
sponds to an independent set in the line graph H of G . Since G is 2 ¾-regular, H
is (4¾ − 2)-regular. Let V1 , . . . , Vr be the partition of V(H) in which each block is
the edge set of one of the cycles in the decomposition of G into 1-regular factors.
By the hypothesis, each Vi has size at least 8e ¾ . By Lemma 14.2.15, H has an in-
dependent set with one vertex in each Vi . This independent set forms the desired
matching in G.
Section 14.2: Refinements of Basic Methods 679

THE GENERAL LOCAL LEMMA (Optional)

The proof of the general form of the Local Lemma is a subtle but short in-
duction using conditional probability (see Definition 0.10). The set Di includes
indices of all events other than Ai on which Ai is dependent, and possibly more.

14.2.18. Theorem. (General Lovász Local Lemma; Erdős–Lovász [1975])

Consider events A1 , . . . , An in a probability space. For each i, let Di ⊆ [n]−{i}
be a set such that Ai is mutually independent of { A j : j ∈ / Di ∪ {i}}. If there
are weights x1 , . . . , x n such that 0 ≤ x i < 1 and (Ai) ≤ x i ∏j ∈ Di (1 − x j ) for
all i, then  (⋂t=1 A t ) > 0.
n

Proof: To simplify the notation, let IS denote ⋂ j ∈S A j , the event that no event
indexed by S occurs. To prove  (⋂t=1 At ) > 0, it suffices to prove both (IS) > 0
n

and  ( At ##### IS ) ≤ x t whenever t ∈

/ S ⊆ [n], since then
n n n
(⋂ At ) = ∏  ( At ##### I[t−1] ) ≥ ∏(1 − x t) > 0.
t=1 t=1 t=1

We use induction on | S| . For | S| = 0, both claims are immediate: (I∅) = 1

and (At ##### I∅) = (At) ≤ x t . Now assume | S| > 0.
For the first claim, choose t ∈ S and let S = S − {t}. Using the induction
#
hypothesis, (IS) = (At #### IS )(IS ) > 0.
For the second claim, choose t ∈ / S and let T = S ∩ Dt . If T = ∅, then At
is mutually independent of {A j : j ∈ S}, and hence (At ##### IS) = (At) ≤ x t . For
larger T , since IS = IT ∩ IS− T , the definition of conditional probability yields

 (At | IS) = (A t ∩ IS)

(I ) =
(At ∩ IT | IS− T)(IS− T) = (At ∩ IT | IS− T ) .
(I | I )(I ) (I | I ) (∗)
S T S− T S− T T S− T

Intuitively, At | IS is the conditional event that results from conditioning on IT

within the probability space obtained by restricting to IS− T .
We seek an upper bound on (At | IS), expressed as a ratio in (∗). The numer-
ator is bounded above by (At | IS− T ), but this equals (At) since S − T ⊆ Dt − {t}.
With T(j) = {i ∈ T : i < j}, the induction hypothesis yields a lower bound on the
denominator:
 (IT ##### IS− T ) = ∏ j ∈ T  ( Aj ##### IT(j) ∩ IS− T ) ≥ ∏ j ∈ T (1 − x j).
With these bounds and the hypothesis (At) ≤ x t ∏ j ∈ Dt (1 − x j ),

(At | IS) = (A(I

t ∩ IT | IS − T )
|I )
≤
(At) ≤ x t ∏ (1 − x j) ≤ x t .
T S− T ∏ j ∈ T (1 − x j) j ∈ Dt − T

When the events are completely independent, automatically  (⋂ Ai ) > 0,

which also follows from the Local Lemma (with each Di being empty) using any
weights such that (Ai) < x i < 1. Larger sets Di impose tighter restrictions on
(Ai) to still guarantee  (⋂ Ai) > 0. If the events are pairwise disjoint, then
Di = [n] − {i} for all i, and finding suitable weights is hard. It becomes impossible
680 Chapter 14: The Probabilistic Method

when ⋃i Ai absorbs enough probability, such as when n = 2 and (A1) = (A2) =

.25 (Exercise 19), even though in this case  (⋂ Ai ) = .5.
The symmetric version is an easy corollary. Note that (1 − 1/d)d < e−1 (Propo-
sition 14.1.1), but (1 − 1/d)d−1 > e−1 when d ≥ 2 (Exercise 4). Fortunately, the
factor 1 − x i does not appear in the product over Di in Theorem 14.2.18.

Proof of Theorem 14.2.9 (Symmetric Local Lemma) . We are given A1 , . . . , An

such that (Ai) ≤ p and Ai is mutually independent of {A j : j ∈ / Di ∪ {i}}, where
D1 , . . . , Dn are subsets of [n] with | Di | < d. We are also given epd ≤ 1.
To show that  (⋂ Ai ) > 0, we set x1 = · · · = x n = 1/d. Let x be this common
value 1/d. We compute
(Ai) ≤ p ≤ d−1 e−1 < d−1(1 − 1/d)d−1 ≤ x ∏ j ∈ D (1 − x j).
i

Thus Theorem 14.2.18 applies. The inequality e −1

< (1 − 1/d)d−1 holds because
−
the right side decreases to e as d → ∞ (Exercise 4).
1

14.2.19.* Remark. The proof given here for the Local Lemma guarantees a point
outside all the bad events, called “flawless” in Achlioptas–Iliopoulos [2016]. How-
ever, the probability of flawless points may be exponentially small, and we are not
given an efficient way to find one. Much effort was devoted to “constructivizing ”
the proof to obtain such a point in polynomial time.
Work in this direction began with Beck [1991]. Subsequent work obtained
randomized algorithms to produce a flawless point in various special cases or un-
der additional restrictions, with a particularly simple example in Moser [2009].
Moser–Tardos [2010] then produced an algorithm that works whenever the sam-
ple space can be described as a product of independent choices over a set of vari-
ables, as in Proposition 14.2.12.
In particular, if the hypothesis of Theorem 14.2.18 holds for events struc-
tured as in Proposition 14.2.12, then one can start with a random point in the
probability space and proceed as follows, iteratively. If some bad event Ai occurs,
then randomly re-sample the variables in the underlying subset Si that determine
Ai . Moser–Tardos [2010] proved that this leads to a flawless point with expected
number of steps at most ∑ i 1−xixi (see Sinclair [2018, Lecture 25] for a proof).
The focus in this area has shifted to efficient local search algorithms de-
scribed in general settings that go beyond the probabilistic framework of the Local
Lemma. For example, Achlioptas–Iliopoulos [2016] replaced the probabilistic for-
mulation by a directed graph on the underlying space and sought a point outside
the bad events via a random walk on the digraph.
The original proof in Moser–Tardos [2010] used a technique called “witness
trees” to track dependencies and progress toward a flawless point. An alternative
approach presented by Moser was named the Entropy Compression Method
by Tao [2009]. This method has been applied in many situations and some-
times gives stronger results than the Local Lemma (see for example Grytczuk–
Kozik–Micek [2013], Esperet–Parreau [2013], and Dujmović–Joret–Kozik–Wood
[2016]). Bernshteyn [2017] further strengthened this method with his Local
Cut Lemma, which captures the sometimes lengthy computations of the Entropy
Compression Method and makes it easier to prove the same results.
Section 14.2: Refinements of Basic Methods 681

A different restriction of the general Local Lemma can be useful when there
are several types of events.

14.2.20. Theorem. (Neighborhood Local Lemma; Molloy–Reed [2002]) Let

A1 , . . . , An be events in a probability space. For each i, let Di ⊆ [n] − {i} be
such that Ai is mutually independent of {A j : j ∈ / Di ∪ {i}}. If (Ai) ≤ 14 and
∑ j ∈ Di (Aj) ≤ 4 for all i, then  (⋂t=1 At ) > 0.
1 n

Proof: With (Ai) ≤ 14 , we have 0 ≤ x i ≤ 12 , where x i = 2(Ai). By Theorem

14.2.18, it suffices to prove (Ai) ≤ x i ∏ j ∈ Di (1 − x j).
Let = 2 ln 2. When 0 ≤ x ≤ 12 , we have e− x ≤ 1 − x (equality holds at both
ends, and e− x is convex on the interval). Using 2(Ai) = x i and 12 ≥ ∑ j ∈ Di 2(A j ),

(Ai) = 2(Ai)e− /2
≤ x i e− j ∈ Di 2 (Aj ) = x
i ∏e
− xj
≤ x i ∏(1 − x j).
j ∈ Di j ∈ Di

Molloy and Reed call Theorem 14.2.20 the Asymmetric Local Lemma. It is
useful when events have different probabilities and each “neighborhood” Di has
few high-probability events. Because the Symmetric Local Lemma requires the
same upper bound on the probabilities of all the events, that bound must be very
small: ed1
. Since here we only require ∑ j ∈ Di (A j ) ≤ 14 , the probabilities of indi-
vidual events can also be near 14 . Exercise 26 further underscores the distinction.
Exercise 27 develops a generalization.
We illustrate the technique. Hind–Molloy–Reed [1999] used the next result
to prove that every graph with maximum degree D has a total coloring (see Exer-
cise 8.3.33) using at most D + 8 log8 D colors. A coloring is ¾-frugal if it is a proper
coloring and every color appears at most ¾ times in each vertex neighborhood.

14.2.21. Theorem. (Hind–Molloy–Reed [1997]) If (G) ≥ ¾ ¾ , then G has a ¾-

frugal coloring using at most 16( (G))1+1/¾ colors.
Proof: (Molloy–Reed [2002]) For ¾ = 1, a ¾-frugal coloring is just a proper color-
ing of G 2 , and the claim follows because (G 2) ≤ ( (G))2 .
For ¾ ≥ 2, let q = ⌊ 16( (G))1+1/¾ ⌋ . Color V(G) uniformly at random from [q]
(this coloring need not be proper). Define events AS for all S ⊆ V(G) such that S
is a pair of adjacent vertices or a (¾ + 1)-set contained in a vertex neighborhood;
the event AS occurs when S is monochromatic. Note that (AS) = 1/q| S|−1 < 14 .
By the Mutual Independence Principle, each AS is mutually independent of
the set consisting of all events A T such that T ∩ S = ∅. That is, let DS be the set of
all T other than S such that T ∩ S
= ∅ and A T is an event. By the Neighborhood
Local Lemma, it suffices to show ∑ T ∈ DS (A T ) ≤ 14 for each event AS .
Every set in DS is an edge incident to S or a (¾ + 1)-set intersecting S that
is contained in a vertex neighborhood. There are at most | S| (G) sets of the first
type and at most | S| (G)( (G)¾ ) of the second type. Since | S| ≤ ¾ + 1, we compute

(G) 1
∑ (A T ) ≤ (¾ + 1)
1
(G) + (¾ + 1) (G)( )
T ∈ DS
q ¾ q¾
(¾ + 1) (G) (¾ + 1)( (G))¾+1 ¾+1 ¾+1 1
< + ≤ + < .
q ¾ !q¾ 16( (G))1/¾ ¾ !16¾ 4
682 Chapter 14: The Probabilistic Method

The original proof of Theorem 14.2.21 used the general Local Lemma. The
neighborhood version gives a simpler proof, but the symmetric version is not
strong enough. It would require a single value p bounding the probabilities of
all the events; thus p ≥ 1/q2 is needed, where q = ⌊ 16( (G))1+1/¾ ⌋ . However,
in DS we must include more than (¾ + 1) (G)( (G) ¾ ) events, so we cannot satisfy
epd ≤ 1. The Neighborhood Local Lemma takes advantage of the fact that not
many of the events with “high” probability 1/q2 occur in each neighborhood.
The Neighborhood Local Lemma is not strong enough for the next applica-
tion to Ramsey numbers. In Corollary 14.1.17, we proved R(3 , ¾) ≤ 8¾ 2/lg ¾ . The
lower bound needs a graph with many vertices having no triangle or independent
¾-set. The deletion method yields R(3 , ¾) > ¾ 3/2+o(1) (Exercise 16).
We will improve the lower bound to c¾ 2/(ln ¾)2 . K im [1995] raised it to
R(3 , ¾) ≥ c¾ 2/ln ¾ , matching the growth rate of the upper bound. More
recent improvements include Bohman [2009], Bohman–Keevash [2013], and
Fiz Pontiveros–Griffiths–Morris [2013], the latter bringing the lower bound
asymptotically within a factor of 4 of the upper bound in Shearer [1983]. (For
analogous discussion of R(4 , ¾), see Exercise 17.)
The lower bound of c¾ 2/(ln ¾)2 was originally proved by Erdős [1961] using an
intricate deletion argument. As noted in Remark 8.1.19, it yields a triangle-free
r-chromatic graph with at most c (r ln r)2 vertices (Exercise 5). Spencer [1977]
gave a simpler proof using the general Local Lemma (see also Spencer [1987]).

14.2.22. Theorem. (Erdős [1961]) R(3 , ¾) ≥ c¾ 2/ln2 ¾ for all ¾ when c is a suffi-
ciently small constant.
Proof: (Spencer [1977]) Form a random graph with vertex set [n] by generating
each edge with probability p, independently; p will be chosen later. We seek n as
large as possible such that the probability of having no triangle or independent
¾-set is nonzero. Since we can adjust c if needed, we may assume that ¾ is large.
Define events AS for all S ⊆ [n] with | S| ∈ {3 , ¾}; event AS occurs if S is an
independent ¾-set or induces a triangle. The (n3) events AS with | S| = 3 each have
probability p3 , and the (n) events AS with | S| = ¾ each have probability (1 − p)(2).
¾

¾
By the Mutual Independence Principle, each AS is mutually independent of
the set consisting of all A T such that | T ∩ S| ≤ 1. That is, let DS be the set of all
T other than S such that | T ∩ S| ≥ 2.
To simplify the computation, we bound a slightly larger quantity. For each
S we include all ¾-sets in DS . Taking a closer look at the number of triples, we
must include (| S2 |)(n − |S|) triples in DS . Since we only have two types of events, to
prove R(3 , ¾) > n it suffices to choose y and in (0 , 1) such that for some p ∈ (0 , 1),

p3 ≤ y(1 − y)3n(1 − )(¾) (1 − p)(2) ≤ (1 − y)(2)n(1 − )(¾).

n ¾ ¾ n
and
We seek n (in terms of ) such that y, , and p can be chosen to satisfy these
inequalities. We will show that n = c0  2/(ln )2 is suitable for some constant c0 .
We may assume that  is sufficiently large to make the argument work, since later
replacing c0 by a suitably small constant c will make the conclusion hold also for
all smaller  .
For large  , we guarantee the first inequality by setting y = p3  (where  is
a constant larger than 1) and making p and small enough that (1 − y)3n → 1
Section 14.2: Refinements of Basic Methods 683

and (1 − )(¾) → 1. By expanding n ln(1 − x), it follows that (1 − x)n → 1 when

n

nx → 0 as n → ∞. Thus the first limit follows from ny → 0, which requires

p = o(n−1/3); we set p = c1 n−1/2 . To obtain (1 − )(¾) → 1, we ensure (¾n) → 0 by
n

−
setting = (¾n) .
Because (1 − p)(2) < e− p(2) and (1 − )(¾) → 1, for the second inequality we study
¾ ¾ n

e− p(2) < (1 − y)(2)n . To simplify the right side, note that (2)ny 2 ∼ c (ln2 ) → 0,
¾ ¾ 4

which yields (1 − y)(2)n ∼ e− y(2)n (see Lemma 14.3.11). Note that ln = − ln (n).
¾ ¾

By comparing exponents in the desired inequality and canceling −(2), it thus suf-
fices to show p > [ln (n)] / (2) + ny.
Using n > (n) and 2.5 2 < ( ), it suffices to show p > 2.5 ln n + ny. Balancing
2 
ny and p motivated our choice of p, which yields ny = pc21 . Also, p = c1 n−1/2 =
c1 c−0 1/2 · ln . Finally, the leading term in ln n as a function of  is 2 ln  . If we
choose the constants to yield the desired strict inequality in the leading terms,
then the lower-order terms can be ignored for sufficiently large  (this also ex-
plains why we can ignore the last factor in the second desired inequality). Thus
it suffices to have c1 > (5c1/2
0 + c1
3 ).
When c0 and c1 are made sufficiently small,
the inequality holds.

The computations in Theorem 14.2.22 are delicate. It would be nice to obtain

the result more simply from the Neighborhood Local Lemma, but this fails.

14.2.23.* Remark. In attempting to use the Neighborhood Local Lemma to

prove R(3 , ) ≥ c 2/(ln )2 , we define the same events to be avoided as in Theo-
rem 14.2.22, with AS for all S ⊆ [n] such that | S| ∈ {3 , }. Again we generate
a random graph using edge-probability p. The formulas for (AS) are as before,
and again AS is mutually independent of { A T : | T ∩ S| ≤ 1}.
By the Neighborhood Local Lemma, choosing p so that ∑ T ∈ DS (A T ) ≤ 14 for
each AS suffices. Again we do not lose much by including all -sets in DS . Taking
a closer look at the number of triples, we must include (|S2 |)(n − | S|) triples in DS .
Since  > 3 and 1 − p < e− p , it suffices to have

 n 1
( )np3 + ( )e− p(2) ≤ .
¾

2  4

We want to maximize n in terms of  such that p ∈ (0 , 1) can be chosen to

make the inequality true. Each term must be less than 14 . The first term bounds
p from above, and the second bounds it from below. Up to constants c1 and c2 (if
n is superlinear in ), the bounds are
c1 ln n c2
< p< .
 ( 2 n)1/3

In order for p to satisfy both inequalities, n(ln n)3 < c is required. Hence
this method cannot even yield a linear lower bound! The Symmetric Local Lemma
is even worse, providing only a constant lower bound. In this problem the deletion
method does better; see Exercise 16.
684 Chapter 14: The Probabilistic Method

EXERCISES 14.2
2
14.2.1. (−) Given natural numbers n and t, let m = n − (nt) 21 − t . Prove that there is a
2

2-coloring of the edges of K m ,m having no monochromatic copy of K t ,t .

14.2.2. (−) Let H be a hypergraph in which every edge has size at least ¾ and intersects
fewer than ⌊ 2 ¾−1/e⌋ other edges. Use the Local Lemma to prove that H is 2-colorable.
14.2.3. (−) Let H be a ¾-uniform hypergraph in which every vertex appears in exactly ¾
edges. Use the Local Lemma to prove that H is 2-colorable if ¾ ≥ 9.
14.2.4. (−) Prove (1 − 1/d)d−1 > e−1 for d ≥ 2.
14.2.5. (−) Given R(3 , ¾) ≥ c¾ 2/(ln ¾)2 for some constant c , prove that there is a triangle-
free r-chromatic graph with at most O((r ln r)2) vertices.
14.2.6. (♦) Let G be an n-vertex graph with average degree d, where d ≥ 1. Prove
(G) ≥ n/(2d) by first selecting vertices with some probability p and then discarding some.
(Comment: This is weaker than Theorem 14.1.14, the Caro–Wei Theorem.)
14.2.7. (♦) Let G be an n-vertex graph having a proper coloring such that every vertex
has neighbors in at most  color classes. Prove that (G) ≥ n/( + 1). (Hint: Modify the
argument of Theorem 14.1.14.)
14.2.8. Dominating set algorithm. Let (G) =  , and let N[v] = N(v) ∪ {v}.
(a) Given S ⊆ V(G), let U = V(G) − ⋃v∈S N[v]. Prove that some vertex of G − S
dominates at least |U | ( + 1)/n vertices in U.
(b) Construct S ⊆ V(G) by iteratively including a vertex that dominates the most
vertices yet undominated. Prove that at most n/( + 1) vertices remain undominated after
n ln( + 1)/( + 1) steps, so adding them yields a dominating set of size at most ⌈ n 1 +ln(¾ +1)
¾ +1 ⌉.
+1
14.2.9. (♦) Prove that an n-vertex graph with minimum degree  has 2¾ln n
pairwise dis-
joint dominating sets. (Feige–Halldórsson–Kortsarz–Srinivasan [2002])
14.2.10. (♦) A graph is locally linear (Fronček [1989]) if every edge lies in exactly one
triangle; equivalently, the subgraph induced by each vertex neighborhood is a matching.
(a) Prove that a union of disjoint triangles is locally linear if and only if it has no two
triangles on four vertices and no (6 , 3)-configuration (three triangles on six vertices).
(b) Use the deletion method to prove that there is a locally linear n-vertex graph with
at least n3/2/c − O(n1/2) edges, where c is any constant greater than 18. (Comment: Sós–
Erdős–Brown [1973] proved existence of cn3/2 triangles with no (6 , 3)-configuration;
√
the
“(6 , 3)-Theorem” of Ruzsa–Szemerédi [1978] showed that a multiple of n2−c/ log n triangles
can occur and that o(n2) is an upper bound; see Exercise 11.1.43.)
14.2.11. (♦) Use the deletion method to prove ex (n; C¾) ∈ (n1 +1/(¾−1)) (see Definition
11.1.9). In other words, show that for some constant c¾ there is an n-vertex graph with
at least c¾ n1 +1/(¾−1) edges that has no -cycle. (Comment: For even  , counting arguments
yield ex (n; C¾) ∈ O(n1 +2/¾).) (Bondy–Simonovits [1974])
14.2.12. (♦) A -subdivision of a graph H is formed by turning each edge of H into a path
of length  + 1. Prove that every n-vertex
√ graph with  n edges contains a 1-subdivision of
2

some complete graph K a with a ≥ 3/2 n. (Comment: Alon–Krivelevich–Sudakov [2003]

√
showed that a ≥  n can be guaranteed, and Fox–Sudakov [2009] considered 1-subdivision
of general graphs. See Fox–Sudakov [2011].)
14.2.13. (♦) For the Ramsey number of the -cube Q¾ , prove R(Q¾ , Q¾) ≤ 23¾ . (Fox–
Sudakov [2011]) (Comment: Fox–Sudakov [2009] proved R(G , G) ≤ 8d2 d | V(G)| when G
is bipartite with (G) = d, which improves this bound to 2 2¾+ o(¾) .)
Exercises for Section 14.2 685

14.2.14. To show the existence of triangle-free graphs with large chromatic number, con-
sider random n-vertex graphs in which each vertex pair is an edge with probability n−2/3 ,
independently. Use these to argue that there is a triangle-free graph with 2n/3 vertices
and chromatic number at least n1/3/(2 ln n). (Comment: This easy result is weaker than
what is guaranteed from R(3 , ¾) in Exercise 14.2.5.)

14.2.15. (♦) Reed [1998] conjectured (G) ≤⌈ (G)+21 + (G) ⌉ for every graph G. He proved
the existence of a positive constant  such that (G) ≤ ⌈ (G) + (1 − )( (G) + 1)⌉ . Reed
[1998] showed that  ≤ 1/2 is necessary. Sufficiency remains open.
(a) Generate a random n-vertex graph H with independent edge-probability p. Prove
((H) < 2+p c ln n) → 1 as n → ∞ , where c is a positive constant.
(b) Prove that there is a graph H with n − (1/3)n3/4 vertices having independence
number 2 and clique number less than (2 + c)n3/4 ln n.
(c) Prove that any  such that Reed’s Conjecture holds for all graphs satisfies  ≤ 1/2.
(Comment: As observed by Kostochka, part (c) also follows from the explicit graph
G t obtained by expanding each vertex of C5 into a t-clique; (G t) ≥ 5t/2 while (G t) =
2t and (G) = 3t − 1. Toward Reed’s Conjecture, when (G) is large it suffices to have
 < 1/130,000 (King–Reed [2016]),  < 1/26 (Bonamy–Perrett–Postle [2018]), or  < 1/13
(Delcourt–Postle [2017]). The last result also works for the list chromatic number.)

14.2.16. Off-diagonal Ramsey numbers. (Note R(3 , ) ≥ c 2/ln2  from Theorem 14.2.22.)
¾
(a) Prove that if (n) p(2) + ( nl)(1 − p)(2) < 1 for some p ∈ (0 , 1), then R( , l) > n.
l

¾

Prove also that R( , l) > n − (n) p(2) − ( nl)(1 − p)(2) for all n ∈ and p ∈ (0 , 1).
l

(b) Choose n and p in the second part of (a) to prove R(3 , ) >  3/2− o(1) . What lower
bound on R(3 , ) can be obtained from the first part of (a)? (Spencer [1977])
14.2.17. (+) Lower bounds on the Ramsey number R(4 , ).
(a) Use part (a) of Exercise 14.2.16 to show that R(4 , ) ≥ ((/ ln )3/2).
(b) Use the Deletion Method to show that R(4 , ) ≥ ((/ ln )2).
(c) Use the Local Lemma to show that R(4 , ) ≥ ((/ ln )5/2). (Spencer [1977])
(Comment: Bohman [2009] proved R(4 , ) ≥ ( 5/2/(ln )2), and R(4 , ) ≤  3+ o(1) is known.)

14.2.18. If n ≥ m, then trivially there is an injective function from [m] to [n]. Consider
functions where elements of [m] are mapped to uniformly random elements of [n], inde-
pendently (all nm functions are equally likely). Show the power of the Local Lemma by
comparing the lower bounds on n given by the Union Bound and by the Local Lemma for
positive probability of obtaining an injective function. (P. Prałat)
14.2.19. Let A1 and A2 be two dependent events, each having probability p. Prove that
the hypotheses of the General Local Lemma can hold for A1 and A2 if and only if p ≤ .25.
√
14.2.20. Let H be a graph with maximum degree  . Prove that if  ≤ s3/2/ 4e, then H
has an s-edge-coloring in which no 4-cycle is monochromatic. (Comment: Jiang–Milans–
West [2013] more generally considered avoiding monochromatic Kp ,q .)
14.2.21. (♦) Let G be a graph with (G) =  − 1. Let S be a set of disjoint cliques in G, each
with size at least c . Use the Local Lemma to prove that if c is large enough, then G has
an independent set consisting of one vertex from each clique in S. Determine how large c
must be to allow the argument to work. (Hint: 11/13 is slightly above the threshold.)
14.2.22. (♦) The van der Waerden number w(l , ) is the least n so that every -coloring
of [n] has an l-term arithmetic progression in one color (Theorem 10.3.8).
(a) Use the existence method to prove that w(l , ) ≥ (l l−1)1/2 .
(b) Use the Local Lemma to prove that w(l , ) ≥ (el)−1  l−1 . (Comment: When l is
prime, there is a construction for w(l , 2) > l2 l using finite fields.)
686 Chapter 14: The Probabilistic Method

14.2.23. (♦) Let G be a digraph in which every vertex has outdegree ¾ and indegree ¾ . Let
r = ⌊ ¾/(2 .25 + 2 ln ¾)⌋ . Partition V(G) into r nonempty sets V1 , . . . , Vr by an appropriate
experiment. Use the Local Lemma to prove that with positive probability every vertex
has a successor in the set containing it. Conclude that every ¾-regular directed graph
has a family of r pairwise disjoint cycles. (Hint: Be careful to ensure that V1 , . . . , Vr
are nonempty. Comment: Alon [1996] proved that +(G) ≥  guarantees /64 pairwise
disjoint cycles (see also Alon–McDiarmid–Molloy [1996]), later improved to /18 (Bucić
[2018]). Bermond–Thomassen [1981] conjectured that /2 can be guaranteed, proved for
tournaments in Bang-Jensen–Bessy–Thomassé [2014].)
14.2.24. Let L be a -uniform list assignment for a graph G. Color vertices uniformly at
random from their lists, independently. For each edge xy, let A x y be the event that x and
y receive the same color. Determine what  is needed for the Symmetric Local Lemma to
guarantee a proper coloring from the lists. How useful is this result? (Molloy–Reed [2002])
14.2.25. Let  (n , r, s) be the minimum number of colors on E(K n) so that every copy of
K r has at least s colors. Let (n , r, s) be the minimum number of colors on E(K n ,n) so that
every copy of K r,r has at least s colors. Use the Local Lemma to prove the following.
(a)  (n , r, s) ≤ cr,s n(r−2)/[(2)− s+1] . (Erdős–Gyárfás [1997])
r

(b) (n , r, s) ≤ c r,s n2(r−1)/[r − s+1] . (Axenovich–Füredi–Mubayi [2000])

2

14.2.26. Let H be a hypergraph in which every edge has size at least 3 and each edge of H
intersects at most ar (other) edges of size r. Use the Neighborhood Local Lemma to prove
that if ∑ ar 2−r ≤ 18 , then H is 2-colorable. How would this condition need to change to
obtain the conclusion from the Symmetric Local Lemma? (Molloy–Reed [2002])
14.2.27. Weighted Neighborhood Local Lemma (Molloy–Reed [2002]).
(a) Let A1 , . . . , A n be events in a probability space. For each i, let Di ⊆ [n] − {i} be such
that Ai is mutually independent of {Aj : j ∈ / Di ∪ {i}}. Let  and t1 , . . . , t n be real numbers
such that 0 ≤  ≤ 14 and ti ≥ 1 for all i. Prove that if (Ai) ≤ ti and ∑ j ∈ D (2 )tj ≤ t2i for all
i
i, then  (⋂i=1 Ai ) > 0. (Hint: Mimic the proof of Theorem 14.2.20, with xi = (2 )ti .)
n

(b) Let H be a hypergraph in which every edge has size at√least 3 and each vertex of
H lies in at most bi edges of size i. Prove that if ∑ bi 2−i/2 ≤ 1/(6 2), then H is 2-colorable.

14.3. Moments and Thresholds

In this section we develop further techniques for random processes. We focus
on random graphs, which have many important applications.
14.3.1. Application. Melting points. View a physical solid as a three-dimensional
grid of molecules. Neighboring molecules are joined by bonds, as in the cartesian
product of three paths. Adding energy excites molecules and breaks bonds, ran-
domly. The fraction of bonds broken increases with temperature. When the graph
remains well connected, the material acts solid. Breaking off small pieces doesn’t
change this, but when all components are small the global nature of the material
changes. Small components float freely, like a liquid or gas.
Mathematically, there is a threshold for the number of bonds to be broken
(in terms of the size of the grid) such that breaking somewhat fewer bonds almost
always leaves a giant component, while breaking somewhat more almost always
leaves only tiny components. Below the threshold temperature the material will
almost certainly be solid, while above it it will almost certainly not be.
Section 14.3: Moments and Thresholds 687

We will need precise notions of “almost always” and “threshold”.

14.3.2. Application. Analysis of algorithms. Worst-case complexity does not

fairly judge algorithms that run quickly on most inputs. Probabilistic analysis
assumes a probability distribution on the inputs to study the expected running
time. Finding a realistic distribution that is easy to analyze can be very hard.
Given a distribution over graphs of each order, we study the expected running
time as a function of the number of vertices.
Naive algorithms may have good expected behavior. For example, we have no
polynomial-time algorithm to find maximum cliques in graphs. If “almost every”
graph has clique number about 2 lg n, then we can check all vertex subsets up
to size 3 lg n. If (G) < 3 lg n, then this will compute the clique number (every
set of size (G) + 1 is not a clique); otherwise it fails. The algorithm rarely fails.
There are somewhat too many subsets of size 3 lg n for this to be a polynomial-time
algorithm, but it ’s close.

The probabilistic method was popularized by Erdős beginning in 1947, with

systematic study of random graphs initiated by Erdős–Rényi [1959]. There will
always be room for new developments, because random structures have different
properties under different probability models, and different models may be appro-
priate for different applications. As in the preceding sections, we aim to convey
the main techniques and do not attempt an exhaustive presentation of results.

“ALMOST ALWAYS”

We have proposed studying properties that “almost always” hold.

14.3.3. Definition. Given a sequence  n of probability spaces, let qn be the prob-

ability that property Q holds in  n . Property Q holds with high probabil-
ity if limn→∞ qn = 1.

For example,  n may be a probability distribution over graphs with vertex

set [n]. A special case of this is our most common model for random graphs, be-
cause it leads to the simplest computations.

14.3.4. Definition. Let p be a function of n. Define the probability space (n , p)

on graphs with vertex set [n] as follows: any two vertices form an edge
with probability p, independently. Each graph with m edges has probabil-
ity pm(1 − p)(2)−m .
n

The random variable G p denotes a graph drawn from (n , p) for some

n. The sequence (n , p)n∈ is known as the standard or binomial random
graph model, because the number of edges has a binomial distribution. It is
also called the Erdős–Rényi model. “The random graph” refers to (n , 12 ).

Computations are much, much simpler in the binomial model(labeled graphs)

than in a uniform distribution over isomorphism classes. Since algorithms op-
erate on graphs with specified vertex sets (rather than isomorphism classes) as
inputs, this model seems appropriate.
688 Chapter 14: The Probabilistic Method

In (n , 12 ), all graphs with vertex set [n] are equally likely. When property Q
holds with high probability in (n , 12 ), we say “almost every graph satisfies Q” .1
We often shorten “with high probability” to “whp”.
Because we often measure running times of algorithms in terms of the num-
ber of vertices and number of edges, we may prefer to control the number of edges.
This suggests making the n-vertex labeled graphs with m edges equally likely.

14.3.5. Definition. Let m be an integer-valued function of n. The probability

−1
space (n , m) assigns probability ( N
m
) to each graph with vertex set [n]
having m edges, where N = (n2). The random variable G m denotes a graph
drawn from (n , m) for some m in the uniform random graph model.

Applications seem more natural in the uniform model. For example, we ask
“in terms of n, how many edges are needed to make a graph almost surely con-
nected?” rather than “in terms of n, what edge probability makes a graph almost
surely connected?” However, calculations are harder in the uniform model.
Fortunately, the binomial model approximates the uniform model when n is
large and m = p(2n). The proof requires detailed study of the binomial distribu-
tion. A graph property Q is convex if G satisfies Q whenever F ⊆ G ⊆ H and
both F and H satisfy Q.

14.3.6. Theorem. (Bollobás [1985, pp. 34–35]) If Q is a convex property and

p(1 − p)(n2) → ∞ , then G p satisfies Q whp if and only if, for every fixed x,
whp G m satisfies Q, where m = ⌊ p(n2) + x[p(1 − p)(n2)]1/2 ⌋ .

Many properties of interest are convex; “having an even number of edges” is

not. Since Theorem 14.3.6 justifies studying the binomial model to draw conclu-
sions about the uniform model, we will study only (n , p). Theorem 14.3.6 also
motivates letting p depend on n; to study graphs with a linear number of edges,
p must decline like 1/n. Constant p yields dense graphs.
Proving (Q) → 1 is usually easier than computing (Q); this distinction is
important. Probabilistic analysis uses asymptotic statements; exact computation
of probabilities is difficult, unnecessary, and avoided wherever possible.
To compare growth rates of sequences, we use “Oh notation”, with the sets
O(º ), o(º ), (º ), and ( ) as defined in Chapter 0. We aim to discard lower-order
terms that do not affect whether lim n→∞ (Q) = 1. We only need that (¬ Q) is
bounded by something tending to 0. Many arguments are “sloppy” in this sense;
it does not matter how loose the bound is as long as it tends to 0. Experience
refines one ’s intuition about what can be discarded without getting into trouble.
1
In probability theory, “almost always” means “with probability 1 ”. When an event holds
with probability tending to 1 in a sequence of spaces, probabilists say it holds “asymp-
totically almost surely”. In discrete spaces, the distinction between “a.a.” or “a.s.” and
“a.a.s” is minor, and the extra word is cumbersome. Erdős and Rényi, and also Alon–
Spencer [2000], used “almost always” to mean “with probability tending to 1 ”. A younger
generation, interacting more with probabilists, uses “a.a.s.” to be more correct. We re-
treat to “with high probability”. Nevertheless, in a sequence of spaces such as (n , 12 ), we
continue to use the informal “almost every graph” , believing this to be clear and not in
conflict with the usage of “almost always” in probability.
Section 14.3: Moments and Thresholds 689

14.3.7. Theorem. (Gilbert [1959]) For fixed p, whp G p is connected.

Proof: We can make G disconnected by picking a vertex bipartition and forbid-
ding edges joining the parts. Occurrence of edges within the parts is irrelevant.
We bound the probability qn that G p is disconnected by summing ([S, S] = ∅)
over all S. Graphs with many components are counted many times. When | S| = ¾ ,
there are ¾(n − ¾) possible edges in [S, S]. Each has probability 1 − p of not appear-
ing, independently, so ([S, S] = ∅) = (1 − p)¾(n−¾) . Considering all S generates
n−1
each partition from each side, so qn ≤ 12 ∑¾=1 (¾n)(1 − p)¾(n−¾) .
⌊n/2⌋
With symmetry in ¾ and n − ¾ , we have qn ≤ ∑¾=1 (¾n)(1 − p)¾(n−¾) . We loosen
the bound to simplify it. Using (¾n) ≤ n¾ and (1 − p)n−¾ ≤ (1 − p)n/2 (for ¾ ≤ n/2)
⌊n/2⌋
yields qn < ∑¾=1 (n(1 − p)n/2)¾ . Let x = n(1 − p)n/2 . When n is large, x < 1, and
hence qn < 1−x x . In fact, x → 0 when p is constant, so our bound on qn approaches
0 as n → ∞.

Computations simplify further when we introduce counting variables

(those taking only nonnegative integer values) and use expectation. If X is a
counting variable and G p satisfies Q when X = 0, then (X) → 0 implies that
whp G p satisfies Q. This is a special case of the next lemma. An analogous state-
ment holds for continuous variables.

14.3.8. Lemma. (Markov’s Inequality) If X is a counting variable, then

(X ≥ t) ≤ (X)
t . Thus
(X) → 0 implies (X = 0) → 1.
Proof: Letting p¾ = (X = ¾), we have
(X) = ∑¾≥0 ¾ p¾ ≥ ∑¾≥t ¾ p¾ ≥ t ∑¾≥t p¾ = t(X ≥ t).

Suppose that X = 1 if G is disconnected and X = 0 otherwise. In Theorem

14.3.7, we proved directly that (X = 1) → 0 (when p is constant) to prove that
whp G p is connected. A different random variable X leads to a simpler proof and
stronger result.

14.3.9. Example. Indicator variables, linearity, and Markov’s Inequality. Let X =

∑ X i , where each X i is an indicator variable for a bad event, and property Q holds
when no bad events occur. If the bad events are easy to analyze, then the linear-
ity of expectation and the convenience of (X i) = (X i = 1) for indicator variables
simplifies the task of proving (X) → 0.
Consider Theorem 14.3.7. The probability that a nontrivial vertex partition
is a cut is at most (1 − p)n−1 , since at least n− 1 pairs of vertices cross the partition.
Let X be the number of cuts. There are 2 n−1 partitions, so (X) ≤ [2(1 − p)]n−1 .
If p > 12 , then (X) → 0 and the probability of having no cut tends to 1. Thus
whp G p is connected when p > 12 .
Other indicator variables yield stronger results. Let X count the pairs of ver-
tices having no common neighbor. When there are no such pairs, G is connected
and in fact has diameter 2. Proving (X) → 0 in (n , p) means that we expect
almost no bad pairs, and then Markov ’s Inequality states that whp G p has none.
690 Chapter 14: The Probabilistic Method

Here X is the sum of ( n2) indicator variables, say X i for the ith vertex pair
{x , y}, with X i = 1 when x and y have no common neighbor. Since n − 2 vertices
must fail, (X i = 1) = (1 − p2)n−2 and (X) = (n2)(1 − p2)n−2 . When p is constant,
(X) → 0, and whp G p has diameter 2 (we will see that much smaller p suffices).
We generalize further in the next theorem.

14.3.10. Theorem. (Blass–Harary [1979]) Fix s , t , p. Whp, G p satisfies the prop-

erty that for every choice of disjoint vertex sets S and T of sizes s and t, there
is a vertex v ∈
/ S ∪ T adjacent to all of S and none of T .
Proof: Let X be the number of bad pairs (S, T); we need only show that (X) →
0. For the ith way to choose disjoint S, T ⊆ V(G), define an indicator variable
X i with value 1 when there is no vertex v with S ⊆ N(v) and T ⊆ N(v). That is,
(S, T) is bad if every v ∈/ S ∪ T fails to have s specified adjacencies and t specified
nonadjacencies. Thus (X i = 1) = (1 − ps(1 − p)t)n− s−t . The number of variables
(choices of (S, T)) is given by a multinomial coefficient. Since X = ∑ X i , we obtain
n
(X) = ( )(1 − ps(1 − p)t)n− s−t .
s,t,n− s− t
For fixed s , t , p, the multinomial coefficient is a polynomial in n. It is bounded
by ns+t , while (X i) dies exponentially as n → ∞. The logarithm of the product
approaches −∞ , and thus (X) → 0.

v
•

S T

THRESHOLD FUNCTIONS

With constant edge probability, random graphs are dense, with many more
edges than needed to be connected. To improve Theorem 14.3.7 and Example
14.3.9, we want to make p(n) smaller while still having G p connected whp. A
closer look at the computation in Example 14.3.9 shows that we can still guaran-
tee (X) → 0 when p(n) is much smaller. To make the discussion precise, we need
another numerical lemma.

14.3.11. Lemma. If np2 → 0 as n → ∞ , then (1 − p)n ∼ e−np.

Proof: Since np2 → 0 requires p → 0, we can use the series expansion of ln(1 − p)
to write
(1 − p)n = en ln(1− p) = en(− p−O(p )) = e−npe−O(np ) .
2 2

Since np2 → 0, the second factor tends to 1, and we obtain (1 − p)n ∼ e−np.
Section 14.3: Moments and Thresholds 691

14.3.12. Example. Diameter 2, continued. When X counts the pairs having no

common neighbor, (X) = (n2)(1 − p2)n−2 (Example 14.3.9). If p tends to 0 but not
too quickly, then it may still be that (X) → 0.
Always 1 − q < e− q (Proposition 14.1.1). If nq2 → 0, then (1 − q)n ∼ e−nq
(Lemma 14.3.11), and constant powers of 1 − q tend to 1. When p(n) is small
enough that np4 → 0, we thus have (X) ∼ 12 n2 e−np . When p = ( c lnn n)1/2 , the
2

1 2− c
expression simplifies to (X) ∼ 2 n . If c > 2, then still (X) → 0.
By parametrizing the probability to simplify the asymptotic formula for the
expectation, we have shown that G p has diameter 2 whp for a much smaller edge
probability than before (and thus a much smaller number of edges in the uniform
model). This strengthens the result.

What happens for smaller edge probability? If c = 2 − in Example 14.3.12,

then (X) → ∞. This does not yet imply that whp diameter 2 fails, but it sug-
gests the notion of a threshold.

14.3.13. Definition. A monotone property is a graph property preserved by

adding edges. A threshold probability function for a monotone property
Q in (n , p) is a function t(n) such that when p(n) is sufficiently smaller than
t(n) whp G p fails Q, and when p(n) is sufficiently larger than t(n) whp G p sat-
isfies Q. The definition of threshold edge function for (n , m) is similar.

By Theorem 14.3.6, a threshold edge probability in (n , p) generally yields

a threshold number of edges for the same property in (n , m).

14.3.14. Remark. Definition 14.3.13 leaves “sufficiently” unspecified. The dis-

tinction between above and below the threshold may require p(n)/t(n) to tend to
0 or ∞ (see Theorem 14.3.18). In the candidate for threshold in Example 14.3.12,
the distinction is whether the multiplicative constant c exceeds 2. “Sharp”
thresholds depend on a constant in a lower-order term (see Theorem 14.3.17).
When our analysis of the model uses a counting variable X such that X = 0

implies Q, often the threshold depends on whether (X) → 0, and Markov ’s In-
 
equality does half the job, since (X) → 0 implies (Q) → 1. In this situation

(as in Example 14.3.12), the formula for (X) in terms of a parametrized expres-
sion for p(n) indicates how sharp the natural threshold is. We obtain candidates

for threshold functions by determining which p(n) yield (X) → 0. Often we ob-
 
tain p(n) so that (X) → 0 or (X) → ∞ depending on the value of a parameter
c. We do not aim in advance for a threshold of specified sharpness.

 
Although (X) → ∞ suggests that (X = 0) → 0, this is not always true.
  
For example, (X) → ∞ when (X = 0) = .5 and (X = n) = .5. To obtain
(X = 0) → 0, we must show that the probability is not dispersed like this.

14.3.15. Definition. The rth moment of X is the expectation of X r . The vari-

 
ance of X , written Var(X), is the quantity ((X − (X))2 ) . The standard
deviation of X is the square root of Var(X).

An important special case is the variance of a binomial random variable X .

With success probability p in n trials, Var(X) = np(1 − p) (see Example 14.4.3).
692 Chapter 14: The Probabilistic Method

14.3.16. Lemma. (Second Moment Method) If X is a counting variable with

− (X)2
(X) > 0, then (X = 0) ≤ (X )(X) . Thus (X = 0) → 0 when (X
2)

(X)2 → 1.
2
2

Proof: Applied to the variable (X − (X))2 and the value t2 , Markov ’s Inequality
yields ((X − (X))2 ≥ t2 ) ≤ ((X − (X))2 )/t2 . Written as ( | X − (X)| ≥ t) ≤
Var(X)/t2 , this is Chebyshev’s Inequality. Since
((X − (X))2 ) = ( X 2 − 2 X (X) + ((X))2 ) = (X 2) − ((X))2 ,
the inequality is ( | X − (X)| ≥ t) ≤ ( (X 2) − (X)2 )/t2 . Since X = 0 only when
| X − (X)| ≥ (X), setting t = (X) completes the proof.

Intuitively, when the standard deviation grows more slowly than the expec-
tation, all probability is pulled away from 0, and (X = 0) → 0.
A connected graph has no isolated vertices, so a connectedness threshold is no
smaller than a threshold for no isolated vertices. In fact, the thresholds are the
same (Bollobás–Thomason [1985]); later we will suggest intuitively why. Proving
the latter threshold is easier, because the indicator variables for isolated vertices
have very simple distributions. We need an approximation to (1 − p)n .

14.3.17. Theorem. In the binomial model, c lnn n with c = 1 is a threshold proba-

bility function for the disappearance of isolated vertices (for (G) ≥ 1). This
corresponds to 2c n ln n in the uniform model.
Proof: Let X be the number of isolated vertices, with X i indicating whether
vertex i is isolated. Now (X) = ∑ (X i) = n(1 − p)n−1 . We study the asymp-
√
totic behavior of (X). As long as p ∈ o(1/ n), we have np2 → 0 and can write
(1 − p)n ∼ e−np. Also (1 − p)−1 ∼ 1, so (X) ∼ ne−np.
To simplify further, set p = c lnn n to obtain ne−np = n1−c , where c is a function
√
of n. When c is constant, p ∈ o(1/ n), as needed to invoke Lemma 14.3.11. When
c > 1, we have (X) ∼ n1−c → 0, and we have proved one side of the threshold.
When c < 1, we have (X) → ∞ and use the Second Moment Method; we need
only show that (X 2) ∼ (X)2 . Computing (X 2) uses another helpful property
of indicator variables: X i2 = X i . Thus (X 2) = (X) + ∑ i
= j (X i X j ). Each X i X j
is an indicator variable, with value 1 only when vi and vj are both isolated, which
requires forbidding 2(n − 2) + 1 edges. Thus (X i X j ) = (1 − p)2n−3 . Again (1 − p)n ∼
e−np , so (X i X j ) ∼ e−2np , and
(X 2) ∼ (X) + n(n − 1)e−2np ∼ (X) + (X)2 .
Since (X) ∈ o((X)2) when (X) → ∞ , we obtain (X 2) ∼ (X)2 .

Next we derive a threshold function for the appearance of fixed subgraphs. A

graph is balanced if the average vertex degree in every induced subgraph is no
larger than the average degree of the entire graph. All regular graphs and all
trees are balanced.

14.3.18. Theorem. If H is a balanced graph with ¾ vertices and l edges, then

p = n−¾/l is a threshold probability function for the appearance of H as a
subgraph of G p .
Section 14.3: Moments and Thresholds 693

Proof: Let X count the copies of H that appear. There are n(¾) ways to map V(H)
into [n]. Each copy of H arises under A of these maps, where A is the number of
automorphisms of H , so there are n(¾)/A possible copies. Note that A depends on
H but not on n. Let X i be the indicator variable for the occurrence of Hi , the ith
copy of H , so X = ∑ X i . Since H has l edges, (X i = 1) = pl . Because ¾ is fixed,
(X) ∼ n¾ pl/A.
Setting p(n) = cn n−¾/l yields (X) ∼ cln/A. Hence (X) → 0 when cn → 0
and (X) → ∞ when cn → ∞. By the Second Moment Method, it remains only
to prove that (X 2) ∼ (X)2 when cn → ∞. Again (X 2) = (X) + ∑i
= j (X i X j ).
(Note that (X) ∈ o( (X)2) when (X) → ∞.)
The summands are not equal; (X i X j ) depends on Hi ∩ Hj . Group terms by
the intersection, H . Let EH denote the sum of the contributions to ∑ (X i X j )
from all pairs (i , j) with Hi ∩ Hj ∼
= H . When H has r vertices and s edges, the
number of edges needed to create Hi and Hj is 2l − s, so (X i X j ) = p2l− s .

Hi H Hj

The case r = s = 0 occurs when Hi ∩ Hj = ∅. The vertex sets for Hi and Hj

can then be chosen in ¾ !¾ !(nn!−2¾)! ways, and each can be placed on its vertex set in
¾ !/A ways. Since ¾ is constant, there are asymptotically n2¾/A2 disjoint choices of
Hi and Hj . Hence the contribution to (X 2) from ∅ is asympotic to (X)2 . The
proof is completed by showing that the total contribution from all other choices
of H has lower order.
When r
= 0, the vertices for pairs (i , j) such that Hi ∩ Hj ∼ = H can be chosen
in r!(¾−r)!(¾−r)!(n−2¾+r)! ways. Since ¾ and r are constant, this is asymptotic to n2¾−r .
n!

The number of ways to extend H into copies of H on these sets depends only on
H and H , not on n or p; call it H . Thus EH is asymptotic to H n2¾−r p2l− s .
Since 2s/r is the average degree of H , and H is balanced, we have r/s ≥ /l
(including when s = 0). Hence pnr/s ≥ pn¾/l → ∞ when cn → ∞. Thus n−r p− s →
0, and EH ∈ o( (X)2) for H
= ∅. Since the number of such subgraphs H is
constant (independent of n), we obtain (X 2) ∼ (X) + E∅ ∼ (X)2 .

This result generalizes for all H. Let d(H) = | E(H)| / | V(H)| ; the ratio is the
density of H. Let (H) = max F ⊆ H d(F); we call (H) the maximum density of
H. The density and maximum density are equal when H is balanced, and then
p = n−1/Ô(H) is the threshold for appearance of H. Every graph H has a balanced
subgraph F such that d(F) = (H). When pnÔ(H) → 0, whp G p has no copy of F ;
hence it also has no copy of H. In fact, always p = n−1/Ô(H) is a threshold function
for the appearance of H (Exercise 13).

14.3.19. Remark. Comments on Second Moment computations.

1. Try to simplify asymptotic expressions. For example, (¾n) ∼ n¾/ ! when
 ∈ o(n1/2), and (1 − p)n ∼ e−np when np2 → 0.
2. Analyzing the counting variable by introducing a parameter that gov-
erns whether the expectation tends to 0 or ∞ can determine the sharpness of the
threshold.
694 Chapter 14: The Probabilistic Method

3. The expression (X 2) ∼ (X)2 does not say that the difference approaches
0, only that the difference is in o( (X)2). The notation (X 2) → c makes sense
only when c is a constant.
4. Linearity of expectation is valid for sums of random variables, not prod-
ucts. When X = ∑i=1 X i and each X i is an indicator variable, (X 2) = (X) +
n

∑i
= j (X i X j ). When they are independent and identically distributed, (X i X j ) =
(X i) (X j ) and (X 2) = (X)2 + n(n − 1)(E(X i))2 . In general, analysis of (X i X j )
depends on how X i and X j interact.
5. When proving (X 2) ∼ (X)2 , isolating the terms that produce (X)2 can
simplify the treatment of the remainder, because it suffices to prove that the total
contribution of the rest has lower order.

CONVERGENCE OF MOMENTS

Sometimes we can say much more about the distribution of a counting vari-
able X . Beyond the second moment, suppose that we can determine the rth mo-
ment (X r) asymptotically. Usually, the sequence of moments determines the dis-
tribution (see Exercise 29). Determining the moments asymptotically determines
the asymptotic behavior of the distribution, and we can compute lim (X = ¾).
After developing the method, we present two applications. First is a sharp
threshold for the disappearance of isolated vertices. What happens “at ” the
threshold for a monotone property Q, when (Q) does not tend to 0 or 1? When
occurrence of Q is equivalent to X = 0 for some counting variable X , being at
the threshold can mean introducing a parametrized lower-order term in the edge
probability p(n) to study (X = 0) more precisely. We may obtain not only (Q)
in terms of that parameter, but in fact the entire asymptotic distribution of X .
The second application gives us a way to generate random d-regular graphs.
Although whp the minimum and maximum degree of G p are relatively close, the
random graph is decidedly not regular, so we need another method.
We will use two other types of moments.

14.3.20. Definition. The rth factorial moment Er (X) of a random variable X

is the expectation of the rth falling factorial X(r) . When X = ∑i=1 X i , the rth
m

binomial moment Sr(X) is ∑ (X i1 · · · X ir ), where the sum is taken over all
r-subsets of the index set [m].

14.3.21. Lemma. If X is a sum of m indicator variables, then Sr(X) = 1

r! Er
(X)
and
m
r
(X = ¾) = ∑(−1)r−¾ ( )Sr (X).
¾
r =¾

Proof: The binomial moment Sr(X) is itself the expectation of a sum of indi-
cator variables: Sr(X) = ∑| T |=r (X i = 1 for i ∈ T). The value of X j for j ∈
/
T is unimportant, so we allow all possibilities for such X j . Hence Sr(X) =
∑| T |=r ∑S⊇ T (AS), where AS is the event that X i = 1 if and only if i ∈ S.
Section 14.3: Moments and Thresholds 695

When | S| = j , the event AS appears (rj ) times in the double sum. Also
∑|S|= j (AS) = (X = j). Thus
m

∑ ( r )(X = ∑ j(r)(X =
j 1 1
Sr(X) = j) = j) = Er(X).
r! r!
j =r j

Finally, given the expression for Sr(X) in terms of (X = j), the claimed for-
mula for (X = ¾) follows from the Inclusion-Exclusion Formula for the number
of elements belonging to exactly ¾ of the events of the form X i = 1. Both sides of
the equality are normalized by the same amount. Alternatively, one can just sub-
stitute the formula for Sr(X) into the claimed expression to show that it equals
(X = ¾) (Exercise 30).
In applying Lemma 14.3.21, typically m → ∞ as n → ∞. We do not hope to
compute (X = ¾) for a given m; we only seek limm→∞ (X = ¾). Fortunately, we
can approximate the alternating sum by truncation.

14.3.22. Lemma. (Bonferroni Inequalities) If X is a sum of m indicator vari-

ables, and q ∈ , then
¾ +2q−1 ¾ +2q

∑ (−1)r−¾ (¾)Sr(X) ≤ (X = ¾) ≤ ∑ (−1)r−¾ (¾)Sr(X) .

r r
r =¾ r =¾


Proof: Letting Ni = (X = i), we have Sr = ∑ j =r (rj )Nj . Consider a truncation
m

of the sum N¾ = ∑r=¾ (−1)r−¾ (¾r )Sr up to the term r = t. Using the expression for
m

Sr , interchanging the order of summation, and using the Subcommittee Identity

−¾
(¾r )(rj ) = (¾j )(rj − ¾)
(Theorem 1.2.3(3)) yields
t t m
r r j
∑(−1)r−¾ ( )Sr¾
= ∑(−1)r−¾ (¾) ∑ ( r )Nj
r =¾ r =¾ j =r
m t m t
r j j j−¾
= ∑ Nj ∑(−1)r−¾ (¾)(r ) = ∑ Nj (¾) ∑(−1)r−¾ ( r − ¾ ) .
j =¾ r =¾ j =¾ r =¾

The inner sum has the form s(a , b) = ∑ l=0(−1)l (bl), where l = r − ¾ , a = t − ¾ ,
a

and b = j − ¾ . Observe that s(a , 0) = 1 and s(a , b) = 0 for 1 ≤ b ≤ a. For b ≥ a, it

follows by induction on b that s(a , b) = (−1)a (b−a1 ) (Exercise 31). Thus the inner
sum is 1 for j = ¾ , is 0 for ¾ < j ≤ t, and is (−1)t−¾ ( j −t−¾−
¾ ) for j > t. For the full
1

sum, this yields

t m
r j j −¾−1
∑(−1)r−¾ ( )Sr = Nj + (−1)t−¾
¾ ∑ Nj ( )(
¾ t−¾
).
r =¾ j = t+1

The sign of the error to obtain the truncation from Nj depends on the parity of
t − ¾ , as claimed.

14.3.23. Definition. A discrete random variable X has the Poisson distribu-


tion with parameter (and mean) Þ if (X = ¾) = e− Þ Þ¾/¾ !.
696 Chapter 14: The Probabilistic Method

14.3.24. Theorem. Let X be a sum of m indicator variables, where m → ∞. If

Sr(X) → r/r! for all r, then (X = ) → e−  ! (that is, X is asymptotically
¾

Poisson distributed with mean ). In particular, (X = 0) = e− .

Proof: Letting m → ∞ and using the asymptotic expression for each Sr (X) in the
formula we derived for (X = ) suggests that
r  −( )r− 
(X = ) →
r
∑(−1)r− () r! =  ! ∑ (r − )! =  ! e− .
r ≥ r ≥

However, the number of summands is unbounded, so we must be careful about

summing asymptotic formulas.
+2q−1 −Þ ¾
Fix  > 0. We have observed that ∑r= (−1)r− (r ) r! → e  ! as q → ∞. In
r

this series, choose q large enough to make the error at most . Since q is now
fixed, we can choose M large enough so that for all r <  + 2q simultaneously
−1
we have | Sr(X) − r/r!| < 2q+1 (r ) for m > M. The Bonferroni inequalities now
imply for m > M that (X = ) differs from ∑r= (−1)r− (r ) r! by at most .
+2q−1 r

#
Now ###(X = ) − e −  #
/ !### < 2  for m > M. Finally, let  → 0.
# #
The technique used in Theorem 14.3.24 was introduced by V. Brun in 1915;
in number theory it is known as Brun’s sieve. The probabilistic interpretation
was initiated by Bonferroni. In probabilistic combinatorics, some call this the
Method of Moments (Janson–Łuczak–Rucinski [2000, p. 140]) or the Conver-
gence of Moments Method. Earlier, Spencer called it the Poisson Paradigm
(see Alon–Spencer [1992]).
We can now sharpen the threshold for the disappearance of isolated vertices.

14.3.25. Remark. Proposing a sharp threshold. We had p = c lnnn as a threshold

because the probability of having no isolated vertices tends to 0 when c < 1 and
to 1 when c > 1. We are at the threshold when p = (1 + n) lnnn with n ∈ o(1).
To study the sharpness of a probability threshold for a property determined
by X = 0 when X is a counting function, examine the computation of (X).
A function p(n , x) such that (X) approaches a nonzero constant in terms of
x yields a “candidate” for a sharp threshold. We may then be able to prove
(X = 0) → e− by finding lim Er(X) for r > 1.
In counting isolated vertices, we have (X) ∼ n1−c when p = c lnn n . Setting
c = 1 +  with n = lnxn yields (X) → e− x . We now invoke Theorem 14.3.24. This
needs no familiarity with Poisson distributions. It simply yields (X = ) →
e− / ! when there is a constant such that Sr(X) → r/r! for r ∈ 0 .
If the lower order term is properly chosen, then should tend to 0 as x → ∞
(since (X = 0) → 1 above the threshold), and should tend to ∞ as x → −∞
(since (X = 0) → 0 below the threshold). Note that e− x has this property.

14.3.26. Theorem. If p = lnnn + nx , then the number X of isolated vertices in

G p is asymptotically Poisson distributed with expectation = e− x . That is,
(X = ) → e− / !. In particular, (X = 0) → e−e−x .
Proof: Since (X) ∼ n1−c when p = c lnn n , setting c = 1 + lnxn yields (X) → e− x .
By Theorem 14.3.24, it suffices to prove Sr(X) → e− r/r!; always S1 (X) = (X).
Section 14.3: Moments and Thresholds 697

Recall that X = ∑ X i , with a variable for each vertex. The term in Sr(X)
for a set of variables T is the expectation that the vertices corresponding to T
in G p have no incident edges. Hence Sr(X) = (nr)(1 − p)r(n−r)+r(r−1)/2 . For fixed r,
we use p → 0 to conclude that constant powers of (1 − p) tend to 1, obtaining
Sr(X) ∼ (n(1 − p)n)r/r!. For constant p, this tends to 0, but with p = (ln n + x)/n,
we have n(1 − p)n → e− x . Hence Sr(X) → Þr/r!.

Our second application allows us to study random d-regular graphs. A thor-

ough survey of models of random d-regular graphs appears in Wormald [1999b].

14.3.27. Definition. (Bender–Canfield [1978]; Bollobás [1980].) The pairing

model (or configuration model) generates a random d-regular multigraph.
Start with n groups of d vertices. Pair the nd vertices at random (we can it-
eratively pick any unmatched vertex and pair it with another unmatched
vertex chosen uniformly). Finally, merge each group into a single vertex.

14.3.28. Remark. The pairing model is useful because it has an asymptotically

constant probability p of producing a simple graph. This has two important con-
sequences. First, the probability of failing to obtain a simple graph in t tries is
(1 − p)t , so we can quickly generate a random d-regular graph.
Second, we can study d-regular graphs by studying the pairing model. If
an event Q happens whp in the pairing model, then it also happens whp in the
restriction of the model to simple graphs, since for fixed d they occupy an asymp-
totically constant portion of the space as n → ∞. That is, even if all the failures
of Q are for simple graphs, still for simple graphs we have (Q) → 1.

14.3.29. Theorem. (Bollobás [1980]) The probability that a d-regular multi-

graph G generated in the pairing model is simple tends to e−(d −1)/4 .
2

Proof: The probability that G is simple is the product of the probability that G
has no loop and the probability that G has no multiedge given that it has no loop.
Let X be the number of loops in G. The probability of a particular loop aris-
ing as an edge joining two specified vertices vi and vj in a specified group v is dn1−1 ,
since we can generate the graph by first picking the mate of vi . The number of
these variables is n(d2 ), so (X) = d−2 1 1−1/(dn)
1
. Thus (X) → d−2 1 .
Let Þ = d−2 1 . If Sr(X) → Þr/r!, then X is asymptotically Poisson distributed
with mean Þ , and the probability of having no loops is e−(d−1)/2 . Pick r of the indi-
r
cator variables for loops. The probability that all these loops occur is ∏i=1 dn−12i+1 .
¾
The number of choices of ¾ such pairs from distinct groups is n(¾)(d2 ) . There can
also be ¾ such pairs with a group contributing more than one loop, but there are
only O(n¾−1) such selections. Thus Sr(X) ∼ ( d−2 1 )¾ , as desired.
Given that no loops occur, let Y count the pairs of edges with the same end-
points that arise. When two vertices are selected from each of two groups, they
2
form a multiedge with probability 2 dn1−1 dn1−3 . There are (n2)(d2 ) such choices of
(d−1)2 (d−1)2
four vertices, so (Y) = ( d−2 1 )2 1−1/(dn)
1 1
1−3/(dn)
→ 4 . Letting ë = 4 , an ar-
gument like that above yields Sr(Y) → ë /r!. Thus (Y = 0) = e
r −(d−1)2 /4.
Finally, e−(d−1)/2 e−(d−1) /4 = e−(d −1)/4 .
2 2
698 Chapter 14: The Probabilistic Method

Theorem 14.3.29 yields the result of Bender–Canfield [1978] and Wormald

[1978] that for
√ fixed d, the number of d-regular graphs with vertex set [n] is
(1− d 2)/4(dn/e)dn/2
asymptotic
√ to 2e /(d!)n (Exercise 35). The formula holds as long
as d ≤ 2 log n − 1 (Bollob ás [1980]) and also for d ∈ o(n1/3) (McK ay [1985]).
McK ay–Wormald [1991] answered the question for d ∈ o(n1/2), showing that then
ln( is simple)) = − d2)/4 − d3/(12n) + O((d2)/n).
(G (1

GRAPH EVOLUTION
We can generate random graphs with vertex set [n] by iteratively adding an
edge that is equally likely to be any edge not already present. This yields the
same probability space as the uniform model when we have added m edges.
This procedure is called the random graph process, introduced in Erdős–
Rényi [1959]. The likely effect of a new edge on the present structure gives intu-
ition about the structure of random graphs, which is usually correct.
A stage of evolution is a range for m(n) (or p(n)) where the properties of the
typical graph do not change much. We describe such stages. Verification uses the
techniques we have developed, but the computations can be difficult.
14.3.30. Remark. Stages of evolution. With many vertices and few edges, each
new edge is likely to be isolated. Later, with small components, each new edge is
likely to join two components, thereby avoiding cycles (see Exercise 6).
Let t¾ (n) = n−¾/(¾−1) . If p/t¾ → ∞ but p/t¾+1 → 0, then whp every tree on
¾ vertices appears but none on ¾ + 1 vertices appears. This uses the fact that
if A1 , . . . , Ar each happen whp, and r is constant, then whp all of A1 , . . . , Ar
happen. For a single tree, Theorem 14.3.18 applies.
When p = c/n with 0 < c < 1, the number of cycles is asymptotically Pois-
son distributed with constant mean. With only a few cycles and all components
small, we still expect the next edge to join two components or create a cycle in an
acyclic component. Whp the largest component has about ln n vertices, and each
component has at most one cycle. Most vertices still lie in acyclic components.
As c reaches and passes 1, the structure of G p changes radically. There is a
double jump: the structure is significantly different for c < 1, c ∼ 1, and c > 1.
At pn = 1, the Second Moment Method guarantees that whp G p has a cycle. Also,
the order of the largest component jumps from log n to n2/3 . With c = 1 + O(n−2/3),
whp a giant component appears with a linear number of vertices, the other com-
ponents are small, and there is a cycle with three crossing chords (nonplanar).
Next, p grows to c lnn n . For c < 1, whp G p has isolated vertices. When c > 1,
they disappear. Edges added to a disconnected graph may go within a component
or connect two. When all components are small, whp new edges join components.
After a giant component emerges, new edges tend to lie within it or join it to an-
other component. The components most likely to be absorbed are the larger ones.
Therefore, as c passes through 1 the last remaining small components swallowed
by the giant component are isolated vertices, making the threshold for connect-
edness the same as the threshold for disappearance of isolated vertices.
With c > 1, suddenly whp G p also has a spanning cycle. Minimum degree ¾
(and the spanning cycle when ¾ = 2) has a threshold that involves a lower order
term: lnnn + (¾ − 1) ln ln n
n . This is closely related to the result in Exercise 24.
Section 14.3: Moments and Thresholds 699

The last stages of evolution are pn/ln n → ∞ with p = o(1), and then constant
p. When p = c lnn n with c → ∞ , increasing density makes evolutionary viewpoint
less valuable. We give less study to probability thresholds and concentrate on the
likely value of graph parameters. When p = 1/2, properties of the random graph
can lead to fast algorithms that whp solve a difficult problem.

14.3.31.* Example. Vertex degrees and isomorphism testing. Erdős–Rényi [1966]

proved that for pn/ln n → ∞ and fixed positive , whp G p has all vertex degrees
between (1 − )pn and (1 + )pn. One expects most degrees to be near pn, but
some will be farther away. For p ≤ 12 , Bollobás [1982] showed that the vertex of
maximum degree is unique in (n , p) whp if and only if pn/ln n → ∞.
For constant p, whp there are many isolated high degrees before the degrees
begin to bunch up. Bollobás [1981b] proved that when t ∈ o(n/ln n)1/4 , whp G p
has different degrees for its t vertices of highest degree, and this is sharp.
Babai–Erdős–Selkow [1980] used this in a quadratic-time algorithm to test
isomorphism that succeeds whp. It labels the vertices in a canonical order (usu-
ally), so that when it labels vertices as v1 , . . . , vn in one graph and w1 , . . . , wn in
another, the only possible isomorphism is the bijection mapping each vi to wi .
Let r = ⌊ 3 lg n⌋ . The canonical labeling sorts the vertex degrees and lets
v1 , . . . , vr be those of highest degree in order. Almost always (in (n , 12 )) these
are uniquely determined. Almost always the other vertices have distinct neigh-
borhoods within {v1 , . . . , vr }. This is tested by expressing the neighborhoods as
binary r-tuples and sorting them!
Almost all graphs pass both tests and have no nontrivial automorphisms.
Having canonically labeled two such graphs, we test isomorphism by comparing
adjacency matrices (see Exercise 41). Babai–Kučera [1979] presented a refine-
ment with smaller probability of rejection.
Deterministic algorithms take longer but give definite answers. For many
years, the best time for testing graph isomorphism for n-vertex graphs was
eO((n log n) ) (see Babai–K antor–Luks [1983]). A breakthrough by Babai [2015]
1/2

showed that it can be tested in “quasipolynomial” time: e(log n) .

O(1)

Similarly, some NP-hard problems are trivial for random graphs. Although
(G) ≤ (G) ≤ (G)+1 for every graph G (Vizing [1964]), deciding between these
values is NP-hard (Holyer [1981]). Vizing proved that (G) = (G) + 1 requires
at least three vertices of maximum degree. By the uniqueness of the vertex of
maximum degree in G 1/2 , Erdős–Wilson [1977] observed that whp (G) = (G).
Often the distribution of a graph parameter over (n , p) is nontrivial but
still is highly concentrated. Given a parameter  , we want to show (G p) ∼  (n)
whp. We often write (G p) = (1 + o(1)) (n), meaning that for any positive  ,
whp (G p) is between (1 − ) (n) and (1 + ) (n). Another way of expressing this
extends the notion of threshold function from Definition 14.3.13.

14.3.32. Definition. Given a sequence of probability spaces indexed by a param-

eter n, a threshold t for a parameter  is a function of n such that ( ≥ )
tends to 1 or 0 depending on whether  is above or below the threshold t. As
in Remark 14.3.14, “above or below” may mean that the ratio to t tends to ∞
or 0, or it may just mean they differ by a small constant.
700 Chapter 14: The Probabilistic Method

The exercises contain examples for various random structures. Here we study
connectivity, independence/clique number, and chromatic number in graphs.

14.3.33. Example. Connectivity of random graphs. For sparse graphs, the

thresholds for a fixed value of the connectivity and the minimum degree are the
same. When p is fixed, the expected number of common neighbors for two fixed
vertices is p2(n − 2). The techniques in the next section permit showing that
whp all pairs of vertices have nearly this many common neighbors, which makes
the connectivity linear in n (see Exercise 14.4.3). However, requiring ¾ common
neighbors is too restrictive for the best result; Bollobás [1981b] showed for con-
stant p that whp G p has connectivity equal to minimum degree.

What about clique number? For fixed ¾ , we have derived probability thresh-
olds for the appearance of a ¾-clique, but for constant p the clique number grows
with n. Determining the size of the largest clique in an input graph is NP-
complete, but for the random graph we can guess the correct value with high
probability without looking at the graph! Amazingly, for constant p, whp G p has
one of two possible values for the clique number (as a function of n). Indeed, for
each ¾ ∈  there is a range of n where the clique number is highly likely to be
¾ . The approach is to find bounds on the function r(n) such that whp G p has an
r-clique and whp has no (r + 1)-clique.

14.3.34. Theorem. (Matula [1972]) For constant p and fixed > 0, whp G p
has clique number between ⌊ d − ⌋ and ⌊ d + ⌋ , where d = 2 log1/p n −
2 log1/p log1/p n + 1 + 2 log1/p(e/2).
Proof: (sketch) Let b = 1/p. If X r is the number of r-cliques, then (X r) = (nr)p(2).
r

√
Since r! ∼ (r/e)r 2 r (Stirling ’s Formula),
(X r) ∼ (2 r)−1/2( en
r p
(r −1)/2)r
.
If r → ∞ and (enr−1 p(r−1)/2) ≤ 1, then (X r) → 0. To obtain such r, take
logarithms (base b) in the inequality and solve for r to find
r ≥ 2 logb n − 2 logb r + 1 + 2 logb e.
This is approximately equivalent to r ≥ d(n) as defined above. More precisely, if
r > d + , then whp G p has no clique of size r.
The lower bound uses the Second Moment Method, as in Theorem 14.3.18,
but the dependence of r on n makes the analysis more difficult. The expectation
of X r2 sums the probability of common occurrence for all ordered pairs of r-cliques.
This probability depends only on the number of common vertices, so
r
n r n − r 2(2r)−(¾2)
(X r2) = ( ) ∑ ( )( )p .
r  r−
¾ =0
We want to show that the term for  = 0 dominates. Hence we write
−1 −1
(X r2)/ (X r)2 =  n + n , where  n = (nr) (n−r r) and n = (nr) ∑r¾=1 (¾r )(rn−−¾r)b(2).
¾

We seek  n ∼ 1 and n → 0. When r ∼ 2 logb n, an asymptotic formula for (¾a)/(¾b)

yields  n ∼ e−r /(n−r) → 1. The discussion of n is harder; Palmer [1985, pp. 75–
2

80] presents further details.

Section 14.3: Moments and Thresholds 701

Finally, consider chromatic number for constant p. Since 1 − p is also con-

stant, the results on clique number apply. With high probability, no independent
set has size more than (1 + o(1))2 logb n, where b = 1/(1 − p). Hence whp (G p) ≥
(1 + o(1)) 2 logn n . Achieving this bound requires many disjoint independent sets
b
with near-maximum sizes. For a decade, the best result was an algorithmic guar-
antee of a coloring with at most twice the number of colors in the lower bound.
Bollobás [1988] proved that the lower bound is achievable by showing that
whp, in G p every set having at least n/(log b n)2 vertices contains a clique of order
at least 2 logb n − 5 logb logb n. This allows independent sets of near-maximum size
to be extracted until too few vertices remain to cause trouble; the remainder can
be given distinct colors.
We explore Bollobás’ approach in the next section. Here we present the ear-
lier result for its algorithmic interest; whp the greedy algorithm uses at most
(1 + )n/log b n colors. Thus it usually works as an approximation algorithm in the
same sense that the isomorphism algorithm in Example 14.3.31 usually works.
Garey–Johnson [1976] showed no fast algorithm uses at most twice the optimum
number of colors on every graph, unless P = NP. Bollobás’ proof does not yield
a fast algorithm for coloring almost every graph with an asymptotically optimal
number of colors; it is an existence proof only.

14.3.35. Theorem. (Grimmett–McDiarmid [1975]) With edge-probability p, set

b = 1/(1 − p). For > 0, whp G p satisfies
( 12 − )n/logb n ≤ (G p) ≤ (1 + )n/logb n.

Proof: For the lower bound, use independent sets as suggested above. For the
upper bound, let h = ⌊(1 + )n/log b n⌋ . We show that greedily coloring v1 , . . . , vn
in order whp uses at most h colors on G p . Among the n-vertex graphs using more
colors, let B m be the set such that vm is the first vertex to use color h + 1. We
prove ∑m=1 (B m) → 0 as n → ∞.
n

Given G, let G m = G[v1 , . . . , vm−1 ]. Before color h + 1 is used, color h must

be used, so for each G ∈ B m the greedy coloring of G m uses h colors. Let i be
the number of uses of color i in this coloring. To require use of color h + 1, vm+1
must have at least one neighbor of each color 1 , . . . , h. Given the numbers {i},
the probability of this is ∏i=1 [1 − (1 − p)¾i ].
h

Bollobás–Erdős [1976] simplified the subsequent computations involving this

bound by observing that it is maximized when 1 = · · · =  h (Exercise 40). Thus

∏i=1 [1 − (1 − p)¾i ] ≤ [1 − (1 − p)(m−1)/h] h < [1 − (1 − p)n/h] h .

h

Let bn = [1 − (1 − p)n/h ] h . Given G m , the probability that the full graph G belongs
to B m is at most bn . Since the bound holds for each G m , we conclude that (B m) <
bn . This holds for all m, so ∑m=1 (B m) < nbn .
n

Using 1 − x < e− x , we obtain nbn < ne− h(1− p) . Substituting h = cn/logb n

n/h

yields (1 − p)n/h
= n . The logarithm of the bound is log n − cn1−1/c/logb n. It
− 1/c

tends to −∞ for c > 1, so the bound on the probability that the greedy algorithm
uses more than h colors tends to 0.

14.3.36. Remark. Some properties known to hold for almost every graph occur
in no known examples! For the known lower bound on diagonal Ramsey numbers,
702 Chapter 14: The Probabilistic Method

there is still no construction of a sequence of graphs such that (G) < log√2(| V(G)|)
and (G) < log√2(| V(G)|), even though almost every graph has this property.
The point is that the range from a billion to infinity is much bigger than
from 1 to a billion. A statement that is true for almost every graph may not be
true for many graphs that are small enough to look at. The probabilistic method
gives good results when almost every graph has a desired property, but even then
“almost every” is only asymptotic.

EXERCISES 14.3

14.3.1. (−) Let H be a graph. For constant p, prove that whp G p contains H as an induced
subgraph.

14.3.2. (−) Generalization of Theorem 14.3.10.

(a) Fix ¾ , s , t , p. Prove that whp G p has the following property: for every choice of
disjoint vertex sets S and T of sizes s and t, there are at least ¾ vertices that are adjacent
to every vertex of S and to no vertex of T .
(b) Conclude that whp G p is ¾-connected.
(c) Apply the same argument to random tournaments: whp, for every choice of dis-
joint vertex sets S and T of sizes s and t, there are at least ¾ vertices with an edge to every
vertex of S and from every vertex of T .

14.3.3. (−) Let ⟨ ⟩ be any sequence tending to ∞ , fix p ∈ (0 , 1), and let b = 1/(1 − p). Prove
that whp the first log b(n n) vertices in [n] form a dominating set in the graph G p .

14.3.4. (−) Determine the smallest connected graph that is not balanced.

14.3.5. For fixed p, and  ∈ o(n/ ln n), prove that whp G p is -connected.

14.3.6. (♦) In (n , p) with p depending on n, prove that if pn → 0, then whp G p has no

cycles. Use a different random variable to prove that if pn → ∞ , then whp G p has a cycle.
(Comment: The result of Theorem 14.3.18 is stronger than this and is not permitted for
use here; the desired statements for cycles can be proved with simpler arguments.)

14.3.7. (♦) By Theorem 13.2.19 or Theorem 14.2.5, we know ex (n; K r,r) < cn2−1/r for some
constant c. Use the Deletion Method to prove ex (n; K r,r) > cn2−2/(r+1) for some constant c
when n is large. (Erdős–Spencer [1974, p. 61])

14.3.8. (♦) With p = 12 , whp G p has a perfect matching when n is even. Let v1 , . . . , v2¾ be
the vertices. For  rounds, iteratively match the least-indexed vertex that remains to a
remaining neighbor, if any exists; otherwise pair it with an arbitrary remaining vertex.
Delete these two vertices and continue. Prove that this “Method of Deferred Decisions”
finds a perfect matching with probability more than 1/3. (Molloy–Reed [2002, p. 20])

14.3.9. For p = 1/n and fixed  > 0, show that whp G p has no component with more than
(1 + )n/2 vertices. (Hint: Bound the probability of such a component by the probability
of another event whose probability tends to 0.)

14.3.10. Suppose that years have t days and that each person’s birthday is random among
them. Let (s , t) be the probability that no two among s given people have the same birth-
day. Prove that (s , t) → 0 as s → ∞ if t ∈ o(s2). (Comment: When t = cs2 , (s , t) tends
to a constant depending on c.) (Griggs [1998])
Exercises for Section 14.3 703

14.3.11. (♦) Find a small window around p = 1/2 such that above it whp G p has at least
2 (2)
1 n
edges, below it whp G p has at most 12 (2n) edges, and within it neither statement holds.

14.3.12. Derive a “local” probability threshold in the binomial model for the property “Ev-
ery edge belongs to a triangle” (the property is not monotone, but within an interval for p
there is a threshold). (Hint: Let X be the number of edges of G not in triangles.)

14.3.13. Extend the Second Moment argument of Theorem 14.3.18 to prove that n−1/Ô(H)
is a threshold function for the appearance of H as a subgraph of G p , where (H) is the
maximum density of H . (Bollobás [1981a], Ruc ínski–Vince [1985])

14.3.14. A random tournament T with vertex set [n] is generated by orienting each edge
i j as i → j or j → i independently with probability 12 . A king in a tournament is a vertex
such that every other vertex can be reached from it by a path of length at most 2. By
Proposition 5.2.18, every tournament has a king. Prove that whp in T every vertex is
a king. (Palmer [1985]) (Comment: Moon–Moser [1962b] proved that whp T is strongly
connected, with probability of failure between (2n − 1)/2 n−1 and (2n + 1)/2 n−1 .)

14.3.15. (♦) A -vertex tournament is transitive if it has a vertex ordering u1 , . . . , u¾

such that ui → uj if and only if i < j .
(a) Given a positive constant  , prove that the fraction of n-vertex tournaments having
no transitive subtournament with more than (1 + ) + 2 lg n vertices tends to 1. (Comment:
Every n-vertex tournament has a transitive subtournament with at least lg n vertices.)
(b) Use part (a) to prove R( , ) > 2(¾−2)/2 for Ramsey numbers. (Erdős–Hajnal)

14.3.16. For n-vertex tournaments, let Q¾ be the property that every set of  vertices has
a common successor. Prove that if  = lg n − (2 + ) lg lg n, where  is a positive constant ,
then whp a random n-vertex tournament satisfies Q¾ . (Erdős [1963])

14.3.17. (♦) With p = (1 − ) ln n/n, find a large m such that whp G p has at least m isolated
vertices. In particular, what m(n) results from Chebyshev’s Inequality?

14.3.18. For fixed p, find a threshold for the maximum  such that whp every  vertices
in G p have a common neighbor.

14.3.19. Let Q¾ be the following graph property (in (n , p)): for every choice of disjoint
vertex sets S, T of size  , there is an edge with endpoints in S and T .
(a) For constant  , obtain a candidate for a threshold probability for property Q¾ . Ex-
plain what is needed to prove that it is a threshold for occurrence of Q¾ . (It may be helpful
to write it as a threshold on 1 − p.)
(b) For  = c lg n, prove that almost every graph (edge probability 12 ) has property Q¾
if c > 2. (Comment: Thus under every vertex ordering, some edge will be “stretched” to
length at least n − 2 log n; see “bandwidth” in Section 16.2.)

14.3.20. (♦) Prove that a longest constant string in a list of n random heads and tails has
length (1 + o(1)) lg n. That is, for  > 0, the probability that a random list has at least
(1 + ) lg n consecutive identical flips tends to 0, but the probability that it has at least
(1 − ) lg n consecutive identical flips tends to 1.

14.3.21. A monotone sequence of vectors in d is a sequence for which the values in

each coordinate form a monotone sequence in (nondecreasing or nonincreasing). Con-
struct a sequence v1 , . . . , vn in d by letting each coordinate be a random permutation of
[n]. Prove that the expected length of the longest monotone subsequence is bounded by a
constant times n1/(d+1) . (Comment: By Theorem 10.1.15, any sequence of n vectors in d
d
has a monotone subsequence of length at least n1/2 , and this is sharp (Exercise 10.1.28).)
704 Chapter 14: The Probabilistic Method

14.3.22. (♦) A computer has n messages to send along a channel. Messages are dropped
with constant probability p, independently. The computer chooses the next message to
transmit uniformly at random from among the n messages. Determine a threshold for the
number of transmissions so that whp at least one copy of each message is received.
14.3.23. Each box of cereal contains one of n prizes, equally likely, independently. Obtain
a very accurate threshold for the number of boxes that should be opened to obtain at least
two copies of each prize. In particular, n ln n is not sufficiently accurate; adjusting a lower-
order term changes the situation from almost-certain failure to almost-certain success.
14.3.24. (+) Let º be a random function from [m] to [n]; all nm functions are equally likely.
Let Q be the property that each element of [n] occurs more than ¾ times in the image of
º . Prove that n ln n + cn ln ln n with c = ¾ is a threshold on m for Q.
(Hint: Prove that when p = o(1) and mp → ∞ and ¾ is constant , the probability of at
most ¾ successes in m coin flips with success probability p is asymptotic to the probability
of exactly ¾ successes. Comment: When ¾ = 0, this problem extends the Coupon Collector
Problem (Exercise 14.1.30) by narrowing the threshold.)
14.3.25. (♦) Suppose that each trial produces one of n coupons, uniformly and indepen-
dently. Prove that after n(ln n + x) trials, the number of coupons that have never been seen
has a Poisson distribution with mean e− x .
14.3.26. (♦) For a real constant c, determine the asymptotic probability that a graph
drawn from (n , c/n) has no triangle.
 
14.3.27. For fixed ¾ ∈ and x ∈ , consider (n , p) with p = 1 − ( ¾ lnnn+ x )1/¾ . Determine
the asymptotic expected number of dominating sets of size ¾ . What does this suggest about
the probability of having a dominating ¾-set , and what would be needed to prove it?
14.3.28. Obtain a sharp threshold for diameter at most 2. In particular, find p(n , x) such

that (diam (G p) ≤ 2) → e− Þ , where Þ is a function of the parameter x, and Þ → 0 as x →
∞ and Þ → ∞ as x → −∞. (Hint: Drop lower-order terms when discussing asymptotics of
the appropriate random variable.)
14.3.29. The moment generating function of a nonnegative integer-valued random

variable is the generating function ∑ an x n , where an = (X n). Show that if the moment

generating function (et X ) is finite for some interval of t around 0, then knowing all the
moments of a distribution of X is equivalent to knowing the distribution.
14.3.30. For X = ∑i=1 X i , let Sr(X) be the rth binomial moment of X (Definition 14.3.20).
m


Given Sr(X) = ∑ j =r ( rj ) (X = j) as in Lemma 14.3.21, use direct substitution and compu-
m


tation to prove (X = ¾) = ∑r=¾ (−1)r−¾ (¾r ) Sr(X).
m

⎧
⎪ 1 if b = 0 ,
⎪
14.3.31. Prove ∑ l=0(−1)l (bl) = ⎨ 0
a
if 0 < b ≤ a ,
⎪
⎪ a b−1
⎩ (−1) ( a ) if b ≥ a .
14.3.32. Let p = c/n for some constant c.
(a) Prove for s < n(20c2) that whp no set of s vertices in G p induces at least 2s edges.
(b) Prove that whp (G) ≤ ln n/ln ln n.
(c) Prove that whp the vertices of degree at least 32lnlnlnnn form an independent set.
14.3.33. (♦) Determine the asymptotic probability that a random d-regular graph gener-
ated using the pairing model has no triangle.
14.3.34. (♦) In the pairing model, one can generate the random matching sequentially,
picking at each step any desired point to initiate the next pair. Use this with d = 2 to prove
that in a random 2-regular n-vertex graph the expected number of cycles is asymptotic to
1
2
ln n. Compare with Theorem 3.1.20. (Alon–Prałat–Wormald [2008])
Exercises for Section 14.3 705

14.3.35. From Theorem

√ 14.3.29, prove that the number of d-regular graphs on vertex set
[n] is asymptotic to 2e(1 − d )/4 (dn/e)dn/2/(d!)n . (Bender–Canfield [1978], Wormald [1978])
2

14.3.36. (♦) In the d-regular pairing model on n vertices, prove that the number of ver-
tices in cycles of length at most log d−1 nn is o(n), where n is any sequence tending to ∞.
(Hint: Explore the graph from a given vertex u to get a bound on the expected number of
edges that can yield short cycles through u.)
14.3.37. Consider the model (n , p) with p constant. Let r = c log1/p n, where c is con-
stant; in this range, we can approximate (nr) by (ne/r)r . For c > 2, prove that whp G p has
no r-clique. State what is needed to prove that when c < 2, whp G p has an r-clique.
14.3.38. (♦) The strength of a theorem A ⇒ B in a probability space is  (A)
(B) . A theorem
with strength 0 is useless, since the hypothesis never holds; a theorem with strength 1
characterizes its conclusion. (Palmer [1985, pp. 84–85])
(a) Use the Second Moment Method to prove that almost every graph fails Ore’s Con-
dition (Corollary 7.3.7) for spanning cycles, so Ore’s Theorem has asymptotic strength 0.
(b) Prove asymptotic strength 1 for the Chvátal–Erdős Theorem (Theorem 7.3.14).
(Comment: Chvátal’s Condition(Theorem 7.3.11)cannot hold for both G and G (Kostochka–
West [2006]; Exercise 7.3.42). Thus Theorem 7.3.11 has strength at most 12 for each n. It is
stronger than Ore’s Theorem, but whether its asymptotic strength is nonzero is unknown.)
14.3.39. The boxicity of G is the least number of interval graphs whose intersection is G.
The interval number of G is the least t such that G is the intersection graph of subsets
of composed of at most t intervals. Prove that the interval number and boxicity of the
random graph (G1/2) are each at least n/(4 lg n). (Erdős–West [1985])
14.3.40. Suppose that 0 < p < 1 and that 1 , . . . , r are nonnegative integers summing
to m. Prove that ∏i=1 [1 − (1 − p)i ] ≤ [1 − (1 − p)m/r ]r .
r

14.3.41. Given the statements in Example 14.3.31 about the behavior of vertex degrees

in (n , 12 ), prove that the isomorphism algorithm there runs in time O(n2) for n-vertex
graphs and works whp. Assume that n numbers can be sorted using O(n log n) pairwise
comparisons. (Babai–Erdős–Selkow [1980])
14.3.42. (♦) Perfect matchings in bipartite graphs. Fix  > 0. Let G be a random subgraph
of K n ,n , with parts A and B and independent edge probability (1 + ) lnnn . Say S fails if
| N(S)| < | S|. By Hall’s Theorem, G has a perfect matching if and only if no set fails.
(a) Prove that if  < 0, then whp G has no perfect matching.
(b) For a minimal failing set S, prove | N(S)| = | S| − 1 and G[S ∪ N(S)] is connected.
(c) Prove that if G has no perfect matching, then A or B contains a failing set with at
most ⌈ n/2⌉ elements.
(d) If r, s ≥ 1, then K r,s has r s−1 sr−1 spanning trees. Use this, part (b), part (c), and
Markov’s Inequality to prove that if  > 0, then whp G has a perfect matching. (Hint: A
summation in the bound on the expected number of minimal failing sets can be bounded
by a geometric series.)
14.3.43. An X , Y -bigraph G with | X | = | Y | = n is an (n ,  ,  , d)-expander if G is regular
of degree d and | N(S)| ≥  | S| whenever | S| <  n. Expanders of constant degree permit
rapid widespread communication using only a linear number of edges.
(a) Consider an experiment in which d -subsets of [n] are chosen at random, and let

X be the size of the union. Prove that (X ≤ l) ≤ (nl)( nl )d .
(b) Use the probabilistic method to prove that if  < 1, then there is a constant
d such that , for all n sufficiently large, an (n ,  ,  , d)-expander exists. (Hint: Using a
suitable probability space on d-regular multigraphs, bound the probability that some set
fails the expansion property by a geometric series. Then choose d so that the sum will be
less than 1. Use (n) < (ne/) .)
706 Chapter 14: The Probabilistic Method

14.4. Concentration Inequalities

In the Second Moment Method to obtain threshold functions, we used the
second moment to show that the random variable is unlikely to be far from its
expectation. We derived the Second Moment Method using the simplest result
of this sort: Chebyshev ’s Inequality, which is obtained by applying Markov ’s In-
equality to the squared deviation from the mean.
The tighter the bound on the “tail probability” (the probability of being at
least a specified distance from the mean), the stronger the results that can be
proved. The Chernoff Bound strengthens Chebyshev ’s Inequality when applied
to binomial random variables.
A binomial variable is the result of a sequence of trials. Sequences of random
variables in which the expectation of the next variable is the value of the current
one are “martingales”. In terms of the amount by which successive variables in
a martingale can differ, Azuma’s Inequality provides a tight bound on tail prob-
ability for the final outcome. Intuitively, a lengthy probabilistic process displays
more consistent and predictable global behavior than individual steps do.
The computations required to apply the tail inequalities are often simpler
than computing moments, because moment computations are subsumed in the
development of these techniques. The theory develops paradigms that can be ap-
plied without repeating difficult computations. McDiarmid [1998] surveyed con-
centration inequalities and their applications in discrete mathematics. Molloy–
Reed [2002] discusses many of these results in the context of applications to graph
coloring. Dubhashi–Panconesi [2009] also develops the concentration inequalities
and applies them to the analysis of randomized algorithms.

CHEBYSHEV AND CHERNOFF BOUNDS

We proved Chebyshev ’s Inequality to develop the Second Moment Method

(Lemma 14.3.16). For a random variable X , Chebyshev ’s Inequality is ( | X −
(X)| ≥ t) ≤ Var(X)/t2 , where the variance Var(X) is defined as (X − (X)) and
equals (X 2) − (X)2 . We begin by applying Chebyshev ’s Inequality to a problem
in combinatorial number theory (see Alon–Spencer [1992] and Molloy [1998]).

14.4.1. Example. A set A consisting of ¾ positive integers has distinct sums if

the sums of its 2¾ subsets are distinct. By using powers of 2, one can construct
a subset of [n] having distinct sums that has size ⌊ lg n⌋ + 1. Can there be such
a subset that is much bigger? Erdős offered $300 for a determination of whether
the maximum size is lg n + O(1); the question remains open.
The sum of a ¾-set in [n] is less than ¾ n. If the sums are distinct, then 2¾ ≤
¾ n, which yields ¾ ≤ lg n + lg lg n + 1. This easy upper bound is not far from the
lower bound. Chebyshev ’s Inequality allows us to bring the upper bound halfway
to the lower bound.
Section 14.4: Concentration Inequalities 707

14.4.2. Theorem. Every subset of [n] having distinct sums has size at most
lg n + 1
2 lg lg n + O(1).
¾
Proof: Let {a1 , . . . , a¾ } be a subset of [n] with distinct sums. Let s = ∑i=1 ai . If
¾ is near the bound lg n + lg lg n + 1, then ¾ n is not much bigger than 2¾ , which
suggests that there must be both small sums and sums near s. Our aim is to
show to the contrary that when the sums are distinct, most of them lie in a much
smaller interval near the middle of [s]. This will improve the upper bound.
Since the 2¾ sums are distinct, we can choose uniformly from the sums by
choosing a random subset of a1 , . . . , a¾ . Using each with probability 12 , indepen-
dently, let X be the sum of the elements selected. By linearity, (X) = s/2.
To study the variance, write X as a linear combination of indicator variables:
¾
X = ∑i=1 ai X i . Since (X i2) = 12 and (X i X j ) = 14 ,
¾
1 2
(X 2) =
2 ∑ a2i + 4 ∑ ai aj .
i=1 i< j

¾ ¾
On the other hand, (X) = 1
2 ∑i=1 ai , so (X)2 = 14 ∑i=1 a2i + 24 ∑i< j ai aj . The
¾
difference is 14 ∑i=1 a2i . Since ai ≤ n for all i, we have Var(X) < 4 n ¾.
1 2
Using
√
t = 2 Var(X) in Chebyshev ’s Inequality yields
√
(| X − (X)| ≥ n ¾ ) < 14 .

Since the 2¾ possible values for X are equally likely, √

we conclude that at least
3/4 of the√possible sums lie in an interval of length 2n ¾ around s/2. That is,
4 2 ≤ 2n ¾ , yielding approximately ¾ ≤ lg n + 2 lg lg n.
3 ¾ 1

For binomial random variables, the Chernoff Bound is much tighter on the
probability of large deviations from the mean. This leads to results that cannot
be proved by Chebyshev ’s Inequality. First, Chebyshev for binomial variables:

14.4.3. Example. A random variable X has the binomial distribution Bin(n , p)

if (X = ¾) = (¾n)p¾ (1 − p)n−¾ . This distribution arises when X is the sum of n inde-
pendent Bernoulli random variables X 1 , . . . , X n , which are 0 , 1-random vari-
ables with (X i = 1) = p. Since they are independent, (X i X j ) = (X i)(X j ) =
p2 . Also X i2 = X i . Thus (X 2) = np + n(n − 1)p2 = np(np + 1 − p), and
Var(X) = (X 2) − (X)2 = np(np + 1 − p) − n2 p2 = np(1 − p).
Letting t =
√ The square root of the variance is the standard deviation.
√
ë Var(X), Chebyshev ’s Inequality yields (| X − (X)| ≥ ë Var(X)) ≤ ë−2 . The
√
number of heads is unlikely to be many multiples of n from the mean.

Chebyshev ’s Inequality came from applying Markov ’s Inequality to a trans-

formed variable, the squared deviation from the mean. This is also the method
for the Chernoff Bound, using the variable euX , where u is a parameter. Note
that (euX ) = ( ∑(uX)¾/¾ !) = ∑ (X ¾)u¾/¾ !. This is the exponential generat-
ing function for the moments of X , called the moment generating function.
708 Chapter 14: The Probabilistic Method

Using the exponential to improve the tail bound is due to Bernstein in the
1920s. Chernoff [1952] strengthened the analysis for binomial random variables.

14.4.4. Theorem. (Chernoff Bound; Chernoff [1952]) If the variable X has the
binomial distribution Bin(n , p) with 0 < p < 1 and t > 0, then
( X − np ≥ nt) ≤ e−2nt 2
and ( X − np ≤ −nt) ≤ e−2nt 2
.

Proof: It suffices to prove the upper tail bound. For the lower, we then apply the
bound on (X − n(1 − p) ≥ nt), where X = n − X .
Let q = 1 − p. If t ≥ q, then (X − np ≥ nt) = 0, so we may assume t < q. Let
m = n(p + t). For u > 0, Markov ’s Inequality yields
(X ≥ m) = (euX ≥ eum) ≤ (euX )/eum .
To obtain the best bound, we will pick u to minimize (euX )/eum . Note that X =
∑i=1 X i , where X1 , . . . , X n are independent 0 , 1-variables with (X i = 1) = p.
n

Using multiplicativity of the expectations of independent variables,

(euX ) = (∏euX i ) = ∏ (euX i ) = (q + peu)n .

(euX )/eum = ( qe+u(ppe+t) ) . With x = eu , we can choose any x to get a bound;

u n
Thus
(p+ t)q
calculus tells us that setting x = p(q− t)
will minimize the function. That is, we
(p+ t)q
set u = ln p(q− t)
. With this choice, careful manipulation (noting t < q) yields

q + peu p p+t q q−t

=( ) ( ) .
e u(p+ t) p+ t q−t
p+ t q− t
It now suffices to bound this quantity by e−2t . Let
2
º (t) = ln (( p+p t ) ( q−q t ) ).
p(q− t)
By computation, º (t) = ln (p+ t)q and º (t) = −1
(p+ t)(q− t) ≤ −4. Since º (0) = º (0) =
0, it follows that º (t) = ∫0 ∫0 º
t t
(s)ds ≤ −2t2 .

14.4.5. Remark. Other variations. With = (X), Theorem 14.4.4 states

( X ≥ + s) ≤ e−2s /n
2
and ( X ≤ − s) ≤ e−2s /n .
2

Other variations are more useful when the difference from the mean is larger
or when p is small. Another upper tail bound is
(X ≥ + s) ≤ e− s /(2
2
+ s) or (X ≥ (1 + ) ) ≤ e− /(2+). 2

this bound is (X − np ≥ nt) ≤ e−nt /(2p+t) ,

2
In the notation of Theorem 14.4.4,
improving Theorem 14.4.4 when 2p + t < 1/2. A stronger version reducing the
denominator from 2 + s to 2( + s/3) is proved in Janson–Łuczak–Rucinski [2000,
pp. 26–28] and Bonato–Prałat [2018]. See also Mitzenmacher–Upfal [2017].
A bound on the lower tail that improves Theorem 14.4.4 when p < 1/4 is
(X ≤ − s) ≤ e− s /(2
2 )
or (X − np ≤ −nt) ≤ e−nt /(2p). 2

For small deviations from the mean, combining alternative upper and lower
bounds yields  (| X − | ≥  ) ≤ 2e− /3 , valid when  ≤ 1.
2
Section 14.4: Concentration Inequalities 709

The Chernoff Bound is much stronger because Chebyshev only uses pairwise
independence, while Chernoff uses the full independence of the trials. Exercise 2
compares them. In essence, Chernoff uses all the moments.
Hoeffding [1963] extended the Chernoff Bound to more general sums of in-
dependent random variables. The crucial step in the proof of Theorem 14.4.4
required independence, but the computation did not essentially require that
X 1 , . . . , X n be integer-valued, identical, or of limited range.
The proof generalizes immediately when (X i = 1) = pi , with p = ∑ pi/n,
because the Arithmetic–Geometric Mean Inequality yields ∏(pi eu + 1 − pi) ≤
[∑(pi eu + 1 − pi)/n] = (peu + 1 − p)n . Extending to variables in the interval [0 , 1]
n

uses the convexity of the exponential function, and then arbitrary ranges can be
introduced by transforming variables.

14.4.6.* Theorem. (Chernoff–Hoeffding Bound) Let X1 , . . . , X n be indepen-

dent random variables. If 0 ≤ X i ≤ 1 and X = ∑i=1 X i , with p = (X)/n and
n

0 < t < 1, then

( X − np > nt) ≤ e−2nt 2
.

More generally, if ai ≤ X i ≤ bi and X = ∑i=1 X i , with = (X), then

n

(X − ≥ s) ≤ e−2s / ∑ (b −a ) .
2 2
¾ ¾

The same bounds apply to ( X − np < − nt) and ( X − < − s) .

Proof*: Again we want to apply Markov ’s Inequality to the moment generating
function, yielding (X ≥ m) ≤ (euX )/eum , where m = n(p + t).
In comparison to Theorem 14.4.4, the variables are now distinct and are
not 0 , 1-random variables. Let pi = (X i). Independence still yields (euX) =
∏ (euX i). For 0 ≤ x ≤ 1, we have eux ≤ (1 − x)e0u + xe1u , by convexity of the
exponential function. Hence the expectation of euX i is bounded by the expecta-
tion of 1 − X i + X i eu . Using this and the Arithmetic–Geometric Mean Inequality
∏i=1 i ≤ (∑i=1 i/n)n , we obtain
n n

n n

∏ (euX i) ≤ ∏(1 − pi + pi eu) ≤ (1 − p + peu)n .

i=1 i=1

With q = 1 − p, we now have (X ≥ n(p + t)) ≤ ( qe+u(ppe+t) ) , which is exactly the
u n

bound we had in the proof of Theorem 14.4.4. The argument to show that the
right side of this inequality is at most e−2nt is exactly the same as there.
2

Although the distribution need not be symmetric around the mean, letting
X i = 1 − X i and applying the bound on the upper tail for ∑ X i again yields the
bound on the lower tail for X .
The extension to arbitrary ranges is requested in Exercise 11.

As an application of the simple form of the Chernoff Bound, we consider the

conjecture of Hajós [1961] that every -chromatic graph contains a subdivision
of K ¾ (see Section 8.2). The conjecture fails badly, but Chebyshev ’s Inequality is
not strong enough to show this.
710 Chapter 14: The Probabilistic Method

14.4.7. Theorem. (Erdős–Fajtlowicz [1981]) In (n , 12 ), almost every graph has

chromatic number at least 2 lgn n and has no subdivision of K c√n , where c > 2.
Proof: Choose G from (n , 12 ). By Theorem 14.3.34, whp (G) < 2 lg n. (The ex-
pected number of stable sets of size ⌈ 2 lg n⌉ is less than 1n , so ( (G) ≥ 2 lg n) < 1n .)
Thus whp (G) > 2 lgn n .
Now consider a K r-subdivision H in an n-vertex graph. Let S be the set of r
branch vertices. At most n − r paths in H pass √ through vertices outside S, so S
induces at least (2r ) − (n − r) edges. When r = c n with c > 2, we have (2r ) − (n − r) >
( 12 + t)(2r ) with t = 12 − c22 . Thus S induces more than ( 12 + t)(2r ) edges.
The Chernoff Bound shows that this is highly unlikely. The number X of
edges induced by a given r-set has the distribution Bin((2r ) , 12 ). Applying Theorem
14.4.4 yields  (X ≥ ( 12 + t)(2r )) ≤ e−r(r−1)t .
2

We multiply by (nr) to consider all S and obtain an upper bound on the prob-
ability of a K r-subdivision. Since (nr) < nr = er ln n , the exponent in this factor is
√
bounded by c n ln n, which grows more slowly than r(r − 1)t2 , which is linear in
n. Hence the bound on the probability of having a K r-subdivision tends to 0.

In the proof of Theorem 14.4.7, Chebyshev ’s Inequality would not suffice. We

need an exponentially small bound on the tail probability because we multiply by
(nr), the number of experiments. For another consequence of the Chernoff Bound
that Chebyshev ’s Inequality is not strong enough to prove, see Exercise 12.
On the other hand, Chebyshev ’s Inequality is strong enough to prove that a
weaker statement than the Hajós Conjecture does hold with high probability.

14.4.8. Remark. Hadwiger’s Conjecture with high probability. A K ¾ -subdivision

contracts to K ¾ , by contracting only edges incident to vertices of degree 2. More
generally, a graph G has a K ¾ -minor if K ¾ arises from some subgraph of G by
contractions of any edges. Hadwiger conjectured that every -chromatic graph
has a K ¾ -minor, a statement weaker than the Haj ós Conjecture.
Mader [1968] proved the existence of a constant c such that if G has average
degree at least c ln  , then G has a K ¾ -minor. Kostochka [1982] and Thomason
[1984] improved this by showing that√ G has a K ¾ -minor when G has a subgraph
with minimum degree at least c ln  , and Thomason [2001] showed that the
best such c is about .319. These are difficult probabilistic results.
The random graph whp has average degree at least (1 − )n/2 (using Cheby-
Inequality on the number of edges). If  = n/lg n, then n/2 grows faster
shev ’s √
than  ln  . Hence almost every graph has a K ¾ -minor, but by Theorem 14.3.35
its chromatic number is less than  . Thus whp Hadwiger ’s Conjecture holds.

The Chernoff Bound also enables us to show that the easy lower bound on
the size of bipartite subgraphs proved in Exercise 14.1.27 is surprisingly sharp.
Finding the largest bipartite subgraph can be viewed as finding the largest edge
cut and is thus also called the max-cut problem; determining the maximum
size is NP-hard. Every graph has a cut with at least half of its edges. In Exercise
⌈n/2⌉
14.1.27, the Existence Method is used to improve the lower bound to m 2 ⌈n/2⌉−1 for
a balanced edge cut, meaning that the two part-sizes differ by at most 1.
Section 14.4: Concentration Inequalities 711

The largest cut in the complete graph K n has size ⌊ n2/4⌋ and is balanced. We
show next that the bound from Exercise 14.1.27 is essentially sharp by obtaining
graphs where every balanced cut has roughly the guaranteed size.

14.4.9. Theorem. Let d and be positive constants with d ≥ (12 ln 2)/ 2 and
< 1. If n > d and n is even, then there is an n-vertex graph with between
(1 − ) d(n2−1) and (1 + ) d(n2−1) edges such that every balanced edge cut has size
between (1 − ) dn 4 and
(1 + ) dn 4 .
Proof: We generate a random graph from (n , p), where p = dn . Let Y be the
number of edges in the graph, and X be the number of edges in a specified bal-
anced cut, so (Y) = d(n2−1) and (X) = dn 4 . Both variables are binomial with suc-
cess probability p. In order to study deviation from the mean, we want to bound
( | Z − (Z)| > (Z)) (for Z ∈ {X , Y }), so we set t = p in Theorem 14.4.4.
Using the simple form of the (two-tailed) Chernoff Bound (Theorem 14.4.4),

( | Y − (Y)|) > d(n−1)

2
) ≤ 2e−2(2)
n 2
d 2/n2
≈ 2e−
2
d2
.

By the two-tailed alternative form  (| X − | ≥  ) ≤ 2e−

2
/3
(Remark 14.4.5),
( | X − (X)| >  dn
4 ) ≤ 2e
− 2
dn/12
.
n
1 n
The number of balanced partitions is ( ),
2 n/2 which is asymptotic to √2 . It
2 n
suffices to show that the probabilities of the events for balanced cuts or number
of edges outside the desired ranges sum to less than 1. Since d ≥ (12 ln 2)/2 , we
have 2 nd/12 > n ln 2, so by the Union Bound the probability of some cut failing
tends to 0. Also 2 d2 is a constant, so the probability of a large deviation from
the expected number of edges is bounded by a constant. Hence for large n some
n-vertex graph satisfies the claim.

We next apply the Chernoff Bound to random sampling of multidimensional

data. Let P be a set of n points in d , such as profiles of n people along d linear
attributes. A query q specifies an interval in each dimension. The answer is the
subset of P lying in the resulting d-dimensional box; such a set of points from P is
a “range”. The family R of possible ranges from P defines a range space (P, R).
Using the Chernoff Bound, a relatively small random sample from the population
accurately describes the fraction of the population lying in any range. The key
is that the number of possible ranges is limited. We give a weaker version of the
result of Vapnik–Chervonenkis [1971], as presented in Phillips [2013].

14.4.10.* Theorem. Given a d-dimensional range space (P, R) and positive con-
stants  and  , let S randomly sample (d/2) ln(n/) points from P. With
probability at least 1 − 2  2d , it holds for all q ∈ R that
### q(P) q(S) ###
### − ## ≤  ,
### | P | |S| ####
where q(T) is the number of points in T returned by the query q.
Proof: Let  = ⌈(d/2) ln(n/)⌉ , and fix a query q. For the ith random choice si
of a point for S, let X i be the indicator variable for its membership in the box for
712 Chapter 14: The Probabilistic Method

q. Thus q(S) is the sum of ¾ identical Bernoulli random variables with success
probability q(P)/ | P | ; the expectation is ¾ q(P)/ | P | . By the Chernoff Bound,
# #
 (####### q(S) − |S| q(P) ###
| P | ###
#>¾ ) ≤ 2e−2¾ = 2  2d n−2d .
2

# #
To complete the result, we show that at most n2d distinct subsets of P can be
returned by queries. A box representing a query can shrink without changing the
answer until each endpoint of each of the d intervals defining it agrees in that
coordinate with some point of the answer set. Doing this identifies a set of at
most 2d points in n (in order to indicate the relevant dimensions, with repetition
allowed) that generates this box by specifying its projections on the axes. There
are fewer than n2d ways to specify this canonical box corresponding to a query, so
there are fewer than n2d sets in the range space. Hence with probability at least
1 − 2  2d all the ranges have sizes approximated within .

Vapnik–Chervonenkis [1971] proved that amazingly only O( −2 d ln(1/ ))

points are needed in the random sample to represent all the ranges this accu-
rately, independent of n. Talagrand [1995] further improved this by showing that
only O( −2(d + ln(1/))) points are needed. This says roughly that the number of
samples needed to guarantee accuracy within for one query also guarantees the
same accuracy for all queries.

MARTING ALES

In the classical random walk on a line, at each step there is probability p

of moving one unit left, probability p of moving one unit right, and probability
1 − 2p of not moving. Regardless of earlier history, the expected position after t
steps (given the earlier history) equals the actual position after t − 1 steps. This
is the defining property of a martingale.

14.4.11. Definition. A martingale is a sequence of random variables X 0 , . . . , X n

such that the expectation of X i given the values of X 0 , . . . , X i−1 equals X i−1 .

The expected position of the random walk after n steps is at the origin. The
exact position is the sum of n independent ±1-random variables, so after a simple
transformation the Chernoff Bound tells us that the walk is highly unlikely to be
very far from the origin, in terms of n.
A martingale X 1 , . . . , X n allows dependence of successive variables. How-
ever, if successive positions cannot differ by much, then the final position is
still highly concentrated around its expectation. When the technique applies,
it makes the detailed computation in the Second Moment Method unnecessary
and yields better bounds on tail probabilities. The work is accomplished by
Azuma’s Inequality, also called the Martingale Tail Inequality.√This states that
if | X i − X i−1 | ≤ 1, then the probability that X n − X 0 exceeds  n is bounded by
e−  /2 . We first prove two lemmas.
2
Section 14.4: Concentration Inequalities 713

14.4.12. Lemma. Let Z be a random variable such that (Z) = 0 and | Z | ≤ 1. If º

is a convex function on [−1 , 1], then (º (Z)) ≤ 12 [º (−1)+ º (1)]. In particular,
(e tZ) ≤ 12 (e t + e−t) for all t > 0.
Proof: Since º is convex and x = 1−2 x (−1) + 1+2 x (1), the convexity inequality yields
º (x) ≤ 1−2 x º (−1) + 1+2 x º (1) for −1 ≤ x ≤ 1. By linearity and (Z) = 0, we have
(º (Z)) ≤ 12 [º (−1) + º (1)].

We extend the notion of conditional probability (Definition 0.10) to random

variables. Recall that the conditional probability of event A given event B (when
(B)
= 0) is obtained by restricting the probability space to B, normalizing by
(B). Thus (A| B) = (A ∩ B)/ (B).
Given random variables X and Z, we write Z | X for the conditional ran-
dom variable “ Z given X ”. We view the slice of the space where X = a as a
probability space, after normalizing the resulting distribution for Z by (X = a).
Thus Z | X becomes an ordinary random variable on a given slice X = a. The con-
ditional expectation (Z | X) is a function of X , defined to be the expectation of
Z over the portion of the probability space in which X has the given value.
Thus (Z | X) is itself a random variable, and we can take its expectation.
Weighting (Z | X = a) by (X = a) and summing over a yields ((Z | X)). We
obtain the expectation for Z over the entire probability space. This removes the
effect of conditioning, so ((Z | X)) = (Z). We make this precise only for dis-
crete random variables.

14.4.13. Lemma. ( (Z | X)) = (Z).

Proof: Let pi , j = (X = i and Z = j), defined for all i such that (X = i)
= 0. Since
(Z | X = i) = j jpi , j / (X = i),

( (Z | X)) = ∑ (Z | X = i) (X = i) = ∑ ∑ jpi , j = (Z).

i i j

The proof for Azuma’s Inequality is about the same as that for the Chernoff
Bound, with additional care involving conditional expectation.

14.4.14. Theorem. (Azuma’s Inequality) √ If X−0ë2,/2. . . , X n is a martingale with

| i
X − X i−1 | ≤ 1, then ( X n − X 0 ≥ ë n) ≤ e .
√
Proof: By translation, we reduce to X 0 = 0. For t > 0, we have X n ≥ ë n if and
√ √ √
only if e tX n ≥ et ë n . Hence (X n ≥ ë n) = (etX n ≥ et ë n ). Markov ’s Inequal-
√ √
ity yields (etX n ≥ et ë n ) ≤ (e tX n)/e ë t n . We further simplify (and weaken) the
upper bound and then choose t to minimize it.
First we prove by induction on n that (e tX n) ≤ [ 12 (e t + e−t)] . By the induc-
n

tion hypothesis, it suffices to prove (etX n) = 12 (e t + e−t)(e tX n−1 ).

We introduce X n−1 to condition on it. Lemma 14.4.13 yields
(etX ) = (etX
n n−1
et(X n − X n−1)) =  ((etX n−1 et(X n − X n−1) | X n−1)) .
When conditioned on X n−1 , in the inner expectation X n−1 is constant. Hence we
can extract etX n−1 to obtain (e tX n) =  (etX n−1 (e tY | X n−1)), where Y = X n − X n−1 .
714 Chapter 14: The Probabilistic Method

Because {X n} is a martingale, (Y) = 0. By hypothesis, | Y | ≤ 1. Hence

Lemma 14.4.12 applies, yielding (e tY | X n−1) ≤ 12 (e t + e−t). This itself is now a
constant, so (e tX n) = 12 (e t + e−t) (e tX n−1 ). We weaken the bound to a more useful
form using
e t + e− t t2¾ t2¾
=∑ ≤ ∑ ¾ = et /2 .
2

2 (2 ¾)! 2 ¾!
Now we have
√ √ √
( X n ≥ ë n) = (etX n ≥ et ë ) ≤ ent /2− ë t
2
n n

for each t > 0. We obtain the best bound by minimizing√ over t. The exponent
√ is
quadratic; we minimize it by choosing t so that tn − ë n = 0, or t = ë/ n. The
resulting bound is e− ë /2 .
2

Azuma’s Inequality is one-sided; it bounds the probability that X n is much

larger than X 0 . The conditions are symmetric in sign, so using {− X i } yields the
same inequality for the other tail, where X n is much smaller than X 0 .

14.4.15. Example. The pragmatic gambler . A gambler can bet up to n times,

where n is fixed. Each √ time he bets, he wins or loses 1 with equal probability.
His goal is winning ë n, so he stops if he reaches that value. Letting X i be his
√
winnings after i games, we have X i = X i−1 if X i−1 ≥ ë n, and otherwise X i =
X i−1 ± 1, each with probability .5. Hence {X i } is a martingale that changes by
at most one at each step,√and Azuma’s Inequality applies. The probability that
the gambler will earn ë n is bounded by e− ë /2 . If ë = 1, then there may be a
2

reasonable chance of success, but if ë = 10, then there is little hope.

As mentioned earlier, the variable X n in this martingale is a simple transfor-
mation of a binomial random variable. The computation in the proof of Azuma’s
Inequality reduces to that for the Chernoff Bound, √ and in fact the bound we ob-
tain on the probability that the gambler earns ë n is the same as that provided
by the Chernoff Bound when p = 12 . However, the Chernoff Bound is better for
Bin(n , p) when p
= 12 .

For combinatorial applications, it is helpful to have a more general notion

of martingale to distinguish between an underlying random process and random
variables that result from it.

14.4.16. Definition. A sequence Y0 , . . . , Yn is a martingale with respect to

a sequence X 1 , . . . , X n of random variables if (1) each Y¾ is a function of
X 1 , . . . , X ¾ and (2) ( Y¾ | X 1 , . . . , X ¾−1 ) = Y¾−1 .

The martingales we discussed earlier are the special case Y¾ = X ¾ for all ¾ ,
where we drop “with respect to”.

14.4.17. Lemma. If Y is a martingale with respect to X 1 , . . . , X n , then

(1) ( Yi+ j | X 1 , . . . , X i ) = Yi for 1 ≤ i + j ≤ n.
(2) (Yi) = Y0 for 0 ≤ i ≤ n.
Section 14.4: Concentration Inequalities 715

14.4.18. Definition. A martingale Y with respect to X satisfies the Bounded

Differences Condition if there exist constants c1 , . . . , cn such that always
| Yi − Yi−1 | ≤ ci for all i.

Allowing the bounds on the differences of consecutive values to differ yields a

generalization of Azuma’s Inequality (Theorem 14.4.14). The resulting inequal-
ity is also called the Hoeffding–Azuma Inequality.

14.4.19. Theorem. (Martingale Tail Inequality; Hoeffding [1963], Azuma

[1967]) If a martingale Y with respect to X satisfies the Bounded Differences
2/
Condition for c , . . . , c , then ( Y − Y ≥ t) ≤ e−t (2 ∑ ci ) for t > 0.
2
1 n n 0

Applying the argument to − Y yields the same bound on ( Yn − Y0 ≤ − t) .

Since (Yn) = Y0 , this indeed is a bound on the probability of large deviation from
the mean. The proof is a relatively straightforward generalization of the proof of
Theorem 14.4.14 (see Exercise 19).
Definition 14.4.18 is tailored to the study of parameters on random struc-
tures. We “reveal” the structure a little at a time, aiming to show high concen-
tration of a parameter of the structure. For example, we may discover a ran-
dom graph one edge or vertex at a time. The pieces of the random structure are
X1 , . . . , X n and the parameter is º .

14.4.20. Definition. Let X denote the list (X 1 , . . . , X n), where X 1 , . . . , X n

are random variables. For a random variable º (X), the Doob process is
(Y0 , . . . , Yn), where Y0 = ( º (X)) and Yi = ( º (X) | X 1 , . . . , X i ) for i ∈ [n].

14.4.21. Proposition. A Doob process is a martingale with respect to its under-

lying variables X 1 , . . . , X n .
Proof: Let X(i) denote (X 1 , . . . , X i). We use the definitions of Yi and Yi−1 at the
beginning and end and Lemma 14.4.13 in the middle to “uncondition” on X i :

( Yi | X(i−1) ) = ( ( º (X) | X(i)) | X(i−1) )

= ( º (X) | X(i−1)) = Yi−1 .

14.4.22. Example. Discovering random graphs. In sampling from the model

(n , p), a graph is generated, but we may not discover the full graph at once. The
edge-exposure process looks at one pair of vertices at a time to learn whether
they are adjacent. The vertex-exposure process includes one vertex at a time to
learn its neighbors among the previous vertices. What we discover in one expo-
sure step does not depend on what we learned in previous steps. Random variables
X 1 , . . . , X n record the information discovered.
For a graph parameter º , the random variable º (X) is the value when the
full graph is known. In the Doob process, Y0 = (º (X)), and Yn is the actual value
on the graph produced by X 1 , . . . , X n . Along the way, Yi is the expectation of º
over the possible completions of the graph, given the knowledge of X 1 , . . . , X i . As
observed in Proposition 14.4.21, always Y is a martingale with respect to X .
716 Chapter 14: The Probabilistic Method

To apply the martingale tail inequalities, we need the Bounded Differences

Condition, which requires bounds on | Yi − Yi−1 |. The intuition is that if a small
bit of knowledge about the structure (say, X i) cannot change º (X) by much, then
the expectation of º (X) over the possible completions cannot change by much.
For example, let º be the chromatic number, and let X be the vertex-exposure
process. Thus X i tells which edges from vi to earlier vertices are present. When
we know everything about a graph G except X i , learning that information cannot
much affect Ò(G); either Ò(G) = Ò(G − vi) or Ò(G) = Ò(G − vi) + 1. The next lemma
will then yield | Yi − Yi−1 | ≤ 1.
The figure below suggests a general step in a Doob process. The large oval F0
is the set of all outcomes for X . For 1 ≤ j ≤ n, let F j be the subset to which we
are restricted by knowing X 1 , . . . , X i . Thus F i is a subset of F i−1 .
For example, consider coin flips. The sample points in F0 are lists of length
n, and F i is the knowledge of the first i flips. Thus | F i | = 2 n−i , and always F i is
one block in a partition of F i−1 into two sets of equal size.

F i−1

Fi

14.4.23. Lemma. Let F0 be the cartesian product space for independent exper-
iments with outcomes X 1 , . . . , X n . Let F i be the random event defined by
X 1 , . . . , X i . Let Y0 , . . . , Yn be the Doob process with respect to X for a ran-
dom variable º (X). Let A be the event defined by {X j : j
= i}. If for each such
A the values of º on A differ by at most 1, then | Yi − Yi−1 | ≤ 1 for all i.
Proof: Consider an instance of F i−1 , with Yi−1 = (º (X) | F i−1). Note that F i−1
is a cartesian product, having all choices for X i , . . . , X n , although X 1 , . . . , X i−1
are fixed in F i−1 . The value of X i determines a block in the partition of F i−1 ,
represented below by a row. Each column is an event A within F i−1 in which all
of X i+1 , . . . , X n are fixed; only X i varies. By hypothesis, in each column s the
minimum and maximum of º (ms and Ms , respectively), differ by at most 1.

Choices of A (X i+1 , . . . , X n fixed within column)

Choices
of F i
(or X i)

Because X i and X i+1 , . . . , X n are specified independently, the probability of

the outcome in row r and column s is qr ps , where qr is the probability that X i
yields this row and ps is the probability that X i+1 , . . . , X n yields this column.
The computation of Yi is expectation across one row:
Section 14.4: Concentration Inequalities 717

∑ ms p s ≤ (º (X) | F i) ≤ ∑ Ms ps ≤ 1 + ∑ ms ps .
Since these upper and lower bounds are independent of the row index, taking
the expectation over the entire grid to compute Yi−1 yields the same inequalities.
Hence Yi−1 and Yi lie in one interval of length 1 and differ by at most 1.

When the conditions of Lemma 14.4.23 hold, we conclude immediately that

the value of º (X) is highly concentrated around its mean.

14.4.24. Proposition. (Shamir–Spencer [1987]) In (n , p), the chromatic num-

ber is highly concentrated around its expectation. In particular,
√
( | 
(G) − ( (G))| ≥  n) ≤ 2e− ë /2 .
2

Proof: Suppose we discover the instance G p via the vertex martingale. At stage
i, we learn the edges from vi to the previous vertices; this is X i , and the outcomes
of these steps are independent. The event A in which all but X i are specified is
the subgraph G − vi of the random graph G plus the knowledge of edges from vi
to later vertices. Since (G − vi) ≤ (G) ≤ (G − vi) + 1, the value of  (X) differs
by at most 1 over all possibilities in A. The hypotheses of Lemma 14.4.23 hold,
and hence Theorem 14.4.19 applies. Using both tails, the claim follows.


Proposition 14.4.24 says nothing about the value of ( (G)). To approximate
this we again use martingales. With constant edge probability p, whp (G p) is
within 1 of 2 logb n− 2 logb logb n+ 1 + 2 log b(e/2), where b = 1/p (Theorem 14.3.34).
The same holds for (G p) using the base c = 1/(1 − p) for the logarithm. To show
that the chromatic number of G p is close to n/(2 log c n), Bollobás showed that one
can extract independent sets of near-maximum size until few vertices remain.

14.4.25. Theorem. (Bollobás [1988]) For p = 1 − 1/c (constant), whp every in-
duced subgraph of G p with order at least m has an independent set of size at
least r, where m = ⌈ n/log2c n⌉ and r = 2 logc n − 5 log c log c n.
Proof: (sketch) Let S be a set of m vertices. We bound the probability that S has
1+
no independent r-set by e−dm for some d , . This bounds the probability that
n − dm1 + 1+
some m-set has no independent r-set by (m )e , which is less than 2 n e−dm .
Since n = m1+o(1) , this bound goes to 0, and Markov ’s Inequality implies that whp
G p has no bad m-set.
It suffices to study the subgraph G induced by [m]. Let  (G) be the maximum
number of “pair-disjoint ” independent r-sets in G; any two of them share at most
one vertex. We will show that whp  (G) ≥ 1. To do this, we show that (1)  (G) is

highly concentrated around its mean, and (2) ( (G)) is unbounded.
We invoke Azuma’s Inequality for (1). Consider the edge-exposure martin-
gale. At each step, we learn whether one additional pair of vertices induces an

edge. We have Y0 = ( (X)) and Y(m) =  (X). The status of one edge slot changes
2

 (X) by at most one, so Lemma 14.4.23 applies, and ( Y − (Y) ≤ − ( m2 ) ) ≤

1/2

e− ë /2 . With  = (Y)/(m
2 1/2
2) ,

(Y = 0) = ( Y − (Y) ≤ −(Y)) ≤ e− (Y)2/(m2 − m)

.
718 Chapter 14: The Probabilistic Method

Hence it suffices to show that (º (X))/m → ∞.

To prove this, we consider another random variable ºˆ(X), the number of in-
dependent r-sets in G sharing no pair with any other independent r-set. Such sets
are pairwise pair-disjoint, so º (X) ≥ ºˆ(X). We introduced º (X) because the re-
striction martingale for ºˆ(X) does not satisfy #### Ŷi − Ŷi−1 #### ≤ 1. In the drawing of
# #
G below, for example, we have r = 4 and seek 4-cliques; if the last (dotted) edge is
present in G, then ºˆ(X) = 0, but if it is absent from G, then ºˆ(X) = 3.

• • •
• •
•
• •
•

It is easier to compute (ºˆ(X)) than (º (X)). Expressing ºˆ(X) as the sum

of (mr) indicator variables, we obtain (ºˆ(X)) as (m r ) times the probability that
[r] induces an independent set that is pair-disjoint from all other independent r-
sets. This is (1 − p)(2) times the conditional probability that [r] does not conflict
r

with other independent r-sets, given the event Z that [r] is in fact independent.
Let Y be the number of other independent r-sets overlapping [r] in at least two
elements. By Markov ’s Inequality, (Y | Z) → 0 implies (Y = 0 | Z) → 1. Since
each set counted shares at least two vertices with [r], we have

(Y | Z) = ∑i≥2 ,r−1 (ri)(mr−−ir)(1 − p)( )−( ).

r
2
i
2

From the expression for r in terms of m, this can be shown to approach 0 as

m → ∞. Hence (ºˆ(X)) is asymptotic to (m (2r)
r )(1 − p) . The expression for r in
terms of m yields (ºˆ(X)) ∈ (m5/3). Thus (Y)/m → ∞ , as desired.

14.4.26. Corollary. (Bollobás [1988], Matula–Kučera [1990]) For constant edge

probability p = 1 − 1/c, whp G p satisfies

(1 + )n/(2 log c n) ≤ (G p) ≤ (1 + )n/(2 log c n),

where = logc log c n/log c n and = 5 log c logc n/log c n.

Proof: The lower bound holds because whp G p has no independent set larger
than 2 logc n − 2 log c log c n. The upper bound follows from Theorem 14.4.25, be-
cause whp we can select independent sets of size 2 logc n − 5 log c logc n until only
n/(log c n)2 vertices remain. Since n/(log c n)2 ∈ o(n/log c n), we can complete the
coloring by using distinct new colors on the remaining vertices.
Section 14.4: Concentration Inequalities 719

BOUNDED DIFFERENCES (optional)

We remarked earlier that the constructive version of the Local Lemma was
based initially on viewing the underlying probability space as a product space over
independent variables. A similar viewpoint allows us to generalize the setting of
martingales, although the resulting tail inequalities may be weaker. We follow
the development in McDiarmid [1998], beginning by rewriting the Bounded Dif-
ferences Condition (Definition 14.4.18) in terms of a function.

14.4.27. Definition. A function º : n → satisfies the Bounded Differences

Condition with respect to the variables X if there exist constants c1 , . . . , cn
# #
such that #### ( º (X) | X 1 , . . . , X i ) − ( º (X) | X 1 , . . . , X i−1 ) #### ≤ ci for all i.

Applying the Martingale Tail Inequality (Theorem 14.4.19) to the Doob pro-
cess for a function satisfying this definition yields another tail inequality.

14.4.28. Corollary. If º : n →  satisfies the Bounded Differences Con-

dition with respect to the variables X using constants c1 , . . . , cn , then
(º (X) − (º (X)) ≥ t) ≤ e−t2 / (2 ∑ c2i ) for t > 0.
14.4.29. Definition. For c , u , v ∈ n , the c-Hamming distance d c(u , v) is de-
fined by d c(u , v) = ∑i∈ I ci , where I = {i ∈ [n]: ui
= vi}. The Hamming dis-
tance d H (u , v) is the case where all ci = 1. We extend Hamming distance by
letting the distance d H (x , A) from x to a finite set A be min{d H (x , y): y ∈ A}.

Note that the distance d c(u , v) does not depend on the value of ui − vi for the
coordinates i where the difference is nonzero, only on which coordinates differ.

14.4.30. Corollary. (Independent Bounded Differences Inequality – IBDI)

Let X = (X 1 , . . . , X n), where X 1 , . . . , X n are independent random variables.
If |º (u) − º (v)| ≤ d c(u , v) whenever u , v ∈ n , then

(º (X) − (º (X)) ≥ t) ≤ e−t / (2∑c )

2 2
i

for t > 0 (with the same bound on lower tails).

Proof: (sketch for the case where the variables X are discrete). It suffices to check
the Bounded Differences Condition with respect to X using c1 , . . . , cn ; Corollary
14.4.28 then yields the conclusion. For the details, see Exercise 23.

The condition “ | º (u) − º (v)| ≤ d c(u , v) for u , v ∈ n ” is called a Lipschitz

condition in analysis. In some contexts the metric d c(u , v) depends on |ui − vi | ,
but here it is only the sum of weights for the coordinates where ui
= vi . A function
º : n →  satisfying this condition is a c-Lipschitz function.
For a subset A of a product space  , let At = {x ∈  : d H (x , A) < t}. That is,
At “fattens” A by adding a boundary of thickness t (differing from points of A in
less than t coordinates). When all points in  are equally √ likely, the next result
implies that if A is a positive fraction of  and t > n, then At contains all but
an exponentially small part of  . McDiarmid [1998] contains a similar result.
720 Chapter 14: The Probabilistic Method

14.4.31. Theorem. Let X = (X 1 , . . . , X n), with X 1 , . . . , X n independent and

each X ¾ taking values in a probability space  ¾ . If A is a subset of the prod-
uct space ∏i  i , then for any t ∈ , 
(X ∈ A) (d H (X , A) ≥ t) ≤ e−t /(8n) .
2

Proof: Since A is fixed, we may let º (X) = d H (X , A). Let Þ = (º (X)). Since
d H (X , A) = 0 if and only if X ∈ A, the left side of the desired inequality can be
written as (º (X) − Þ ≤ − Þ) (º (X) − Þ ≥ t − Þ), the product of lower and upper
tail probabilities. We use a bound arising from the first factor when t ≤ 2 Þ and a
bound arising from the second factor when t ≥ 2 Þ.
Since º (X) changes by at most 1 when any one coordinate is changed, º sat-
isfies the condition “ | º (u) − º (v)| ≤ d c(u , v) whenever u , v ∈ n ” in Corollary
14.4.30 with each ci equal to 1. Hence we can apply the IBDI to both the lower
tail and the upper tail. When t ≤ 2 Þ ,
( º (X) − Þ ≤ − Þ) ≤ ( º (x) − Þ ≤ − t/2) ≤ e−t /(8n).
2

When t ≥ 2 Þ ,
( º (X) − Þ ≤ t − Þ) ≤ ( º (x) − Þ ≤ t/2) ≤ e−t /(8n) .
2

With probability at most 1 in the other tail, in each case the claim follows.

14.4.32. Application. A famous result of Harper [1966] concerns the isoperi-

metric problem on the hypercube. An isoperimetric problem in a metric space
asks to minimize the size of the boundary for a subset of specified size. Harper
proved that for A ⊆ 2[n] with | A| ≥ ∑¾=0 (¾n), the number of subsets of [n] with
r

r+t
Hamming distance at most t from A is at least ∑¾=0 (¾n). This result is sharp,
since equality holds when A consists of all the subsets with size at most r.
Theorem 14.4.31 leads quickly to an approximate version of Harper ’s result.
Let all subsets X of [n] be equally likely. Given a family A, by Theorem 14.4.31
the probability that a randomly chosen subset of [n] differs in at least t positions
from all sets in A is at most 1a e−t /(8n) , where a is the fraction of all subsets that
2

are in A. Thus At occupies a fraction at least 1 − 1a e−t /(8n) of all sets.

2

To compare with Harper ’s result, we must approximate the tail of the bi-
nomial distribution. Using the Chernoff Bound (X > n2 + nÌ) ≤ e−2nÌ as in
2

Theorem 14.4.4, when X is distributed as Bin(n , 12 ) and r ≥ n/2 we have (X ≥

r + t) ≤ e−2(r+t−n/2) /n . With | A| ≥ ∑¾=0 (¾n), we reduce r to n/2 to simplify the bound.
2 r

Harper ’s result then implies that At occupies a fraction at least 1 − 1a e−2t /n of all
2

sets, similar to the guarantee from Theorem 14.4.31.

Talagrand [1995] greatly strengthened Theorem 14.4.31 to an inequality of

a similar form, by introducing a distance function d T (X , A) that is at least as
large as d H (X , A) and making the upper bound much tighter and independent of
n. For a subset A of a product space  , let d T (x , A) be sup d c(x , A), where the
supremum is taken over all unit vectors c.

14.4.33. Theorem. (Talagrand’s Inequality) Under the same conditions as in

Theorem 14.4.31, (X ∈ A) (d T (X , A) ≥ t) ≤ e−t /4 .
2
Exercises for Section 14.4 721

The applications of Talagrand ’s Inequality yield tighter concentration around

the median, which also implies tighter concentration around the mean.
For example, let X be the length of a longest increasing
√ subsequence in a ran-
dom permutation of [n]. It is known that (X) = 2 n (see the discussion in Sec-
√
tion 16.3). Before Talagrand’s result, it was known that whp | X − (X)| < c n.
Talagrand ’s Inequality can be used to show that whp | X − (X)| < cn1/4 .
Further discussion of Talgrand ’s Inequality and its applications can be found
in McDiarmid [1998], Dubhashi–Panconesi [2009], and Alon–Spencer [2016].

EXERCISES 14.4

14.4.1. (−) An algorithm to compute an answer succeeds with probability p, where p = 43 .

How many times must the algorithm be run to ensure that with probability at least .99
the correct answer is produced at least half of the time?
14.4.2. (−) For a variable X with distribution
√ Bin(n , p), Chebyshev’s Inequality in Exam-
ple 14.4.3 yields  (| X − (X)| ≥ Var(X)) ≤ −2 . Determine the improved bound that
the simple Chernoff Bound gives for this probability, again independent of n.
14.4.3. (−) Use the Chernoff bound to show for constant p and positive  that whp G p has
connectivity at least (1 − )p2 n.

√ (♦) Use the Chernoff Bound for small p (from Remark 14.4.5) to prove that if
14.4.4.
p > c ln n/n for some constant c, then whp diam (G p) ≤ 2. What value of c suffices? (Com-
ment: Exercise 14.3.28 gives the actual threshold.)
14.4.5. (♦) Use the Chernoff√Bound to show that almost always the random graph has
minimum degree at least 2n − cn ln n for some appropriate constant c.
14.4.6. (♦) Use the Chernoff
√ Bound to show that almost always the random graph has
connectivity at least 4n − cn ln n for some appropriate constant c. (Comment: Bollobás
[1981b] showed that the connectivity almost always equals the minimum degree.)
14.4.7. (♦) The expected number of edges joining any two sets of size n/2 in the random n-
vertex graph is n2/8, and hence the expected maximum size of a bipartite subgraph is at
least n2/8. Use the Chernoff Bound and the Union Bound to prove that whp the random
graph has no bipartite subgraph with more than n2/8 + n3/2 edges. (Molloy–Reed [2002])
14.4.8. (♦) Place n balls into n boxes, uniformly and independently.
(a) Use the alternative form of the Chernoff Bound (Remark 14.4.5) to prove that whp
no box has more than 2 ln n balls, asymptotically.
(b) Use a more direct probability computation to prove that whp no box has more than
O( lnlnlnnn ) balls. (Comment: Also whp some box has ( lnlnlnn
n
) balls. When the tail is far from
the expectation the Chernoff Bound may not be strong enough for optimal results.)
14.4.9. A ranking of a tournament T is a linear ordering  of its vertices. A good ranking
agrees with T on many pairs. Let DT ( ) = a − b, where  and T agree on a vertex pairs
and disagree
√ on b pairs. Prove that there is an n-vertex tournament T such that DT ( ) ≤
2n3/2 ln n for every ranking  . (Comment: De la Vega proved further that for some n-
vertex tournament , every ranking agrees with at most o(n3/2) pairs.) (Erdős–Moon [1965])
14.4.10. Polling. The fraction of the population preferring A to B is p. In a poll of n people,
the fraction who prefer A is X . To be accurate, we want (| X − p| ≤  p) > 1 −  . Using the
simple Chernoff Bound, how large should n be in terms of  and  , assuming p > .25?
722 Chapter 14: The Probabilistic Method

14.4.11. Chernoff–Hoeffding Bound (Hoeffding [1963]) Let X 1 , . . . , X n be independent

random variables, and let X = ∑i=1 X i . Prove the more general version of Theorem 14.4.6:
n

≥ s) ≤ e−2s / ∑ (bi −ai) .

2 2
if ai ≤ X i ≤ bi for 1 ≤ i ≤ n, with = (X), then (X −
14.4.12. (♦) Let  color the vertices of a hypergraph using {1 , −1}. The discrepancy of
 is max e∈ E(H) ####∑v∈e  (v)####. Use the Chernoff Bound to prove that every hypergraph with
# # √
m edges and n vertices has a {1 , −1}-coloring with discrepancy at most 2n ln(2m). What
weaker bound on discrepancy can be proved using Chebyshev’s√ Inequality? (Comment: For
the case m = n, Spencer [1985] improved the bound to 5.32 n; see Alon–Spencer [1992].)

14.4.13. Let H be a hypergraph where every edge has size at least r and intersects at
most  other edges. Use the Chernoff Bound and the Local Lemma to prove that H has a
coloring with discrepancy at most  if  ≤ 18 e /6r . (Molloy–Reed [2002])
2

# #
14.4.14. The discrepancy of a signing  : E(K n) → {1 , −1} is max S⊆ V(K n) ####∑{u ,v}⊆ S  (uv)####.
√ # #
Prove that some signing has discrepancy at most ln 2(n3/2 + 12 n1/2). (J.-H. Kim)

14.4.15. In n tosses of a fair coin, let

√ Y be the number of heads minus the number of tails.
Compare the bounds on (| Y | >  n) using martingales and the simple Chernoff Bound.

14.4.16. Let  be a random function from [n] to [n]. Let Y be the number of elements of
[n] that are missing from the image of  . Prove (| Y − (Y)| ≥ t) ≤ e− t /(2n) .
2

14.4.17. (♦) Homogenizing triples. Let G0 be an n-vertex graph. At time t, to form G t from
G t−1 , first select a random triple S of vertices. If S induces j edges in G t−1 , then turn S
into a triangle with probability j/3 and into an independent set with probability 1 − j/3.
(a) Determine the probability p that the process ends by turning the graph into K n .
(b) For | E(G0)| = 12 (n2) , prove the expected number of steps is (n4). (Wormald [1999a])

14.4.18. (♦) Let Ln denote the maximum length of an increasing sublist of a random per-
√
mutation of [n]. Prove (| Ln − (Ln)| ≥  n) ≤ 2e−  /2 . (Comment: Logan–Shepp [1977]
2

√
and Versik–Kerov [1977] together proved (Ln) ∼ 2 n; see Section 16.3. Hence the result
here is a bit unsatisfying.
√ √
Frieze [1991] proved a family of stronger bounds, including
(| Ln − (Ln)| ≥ n) ≤ 2e− n . Bollobás–Brightwell [1992] generalized.)

14.4.19. (Martingale Tail Inequality) Let Y be a martingale with respect to X satisfy-

ing the Bounded Differences Condition for c1 , . . . , cn . Prove Theorem 14.4.19: For t > 0,
− t 2 / (2 ∑ c2i ) .
(Yn − Y0 ≥ t) ≤ e (Hoeffding [1963], Azuma [1967])

14.4.20. Generate two random binary lists of length n; the bits are chosen by unbiased
coin flips, independently. Let Yn be the length of a longest common subsequence in the two
lists (a common subsequence need not use the same positions in the two lists and need not
appear in consecutive positions). Prove (| Yn − (Yn)| ≥ ) ≤ 2e−  /8n .
2

14.4.21. Let X be the number of triangles in the random n-vertex graph. By linearity,
(X) = 18 (3n). Prove
(| X − (X)| >  n2 ) ≤ 2e−c  for some constant c.
2

14.4.22. (♦) Bin-packing. The numbers a1 , . . . , an are drawn uniformly and independently
from the interval [0 , 1]. They must be placed in bins, each having total capacity 1. Let X
√
be the number of bins needed. Use IBDI to prove that (| X − (X)| ≥  n) ≤ 2e−  /2 .
2

14.4.23. Complete the proof of Corollary 14.4.30 for discrete random variables.
Chapter 15

Linear Algebra
Some combinatorial problems have elegant solutions using algebraic tech-
niques. In this chapter we consider applications of linear algebra. We consider
the uses of polynomials and dimension in vector spaces, determinants and per-
manents of matrices, and eigenvalues of matrices associated with graphs. An
important topic that is too large in scope to include here is coding theory.

15.1. Dimension and Polynomials

The dimension of a vector space is an upper bound on the size of a linearly
independent set of vectors. This elementary statement in linear algebra is a pow-
erful tool for extremal problems in combinatorics. To apply the method we need a
vector space in which the desired set corresponds to linearly independent vectors.
Often we use a space of polynomials, viewing a polynomial as a linear combination
of monomials. Babai–Frankl [1992] and Matoušek [2010] provide many examples
of this and other aspects of linear algebra in combinatorics.

THE DIMENSION ARGUMENT

Two easy examples illustrate the elegance of dimensionality proofs and the
process of turning extremal problems into dimension problems.

15.1.1. Example. Eventown vs. Oddtown. A town with n people contains many
clubs, and every two clubs have an even number of common members. How many
clubs can there be if all clubs have even size? How many if all clubs have odd size?
When the clubs have even size (“Eventown”), we can form 2⌊n/2⌋ clubs by
grouping the residents into pairs and letting each club be a subset of these pairs.
In fact, there is no larger set of clubs (Exercise 4).
When the clubs have odd size (“Oddtown”), we can form n clubs by using
clubs of size 1, or clubs of size n − 1, for example. In fact, there are between
2 n(n+2)/8/(n!)2 and 2 n /n! nonisomorphic constructions of size n (Exercise 5). These
2

are much smaller than the Eventown constructions, but nevertheless we show
next that there cannot be more than n clubs.

723
724 Chapter 15: Linear Algebra

The incidence vector of a subset A of [n] is the binary n-tuple u such that
ui = 1 if i ∈ A and otherwise ui = 0. The simple observation that allows us to con-
vert problems about intersections of sets into algebraic problems is that if u and
v are the incidence vectors of subsets A and B of [n], then u · v = | A ∩ B| , where
u · v is the ordinary dot product of u and v: u · v = ∑i=1 ui vi . Since subscripts
n

indicate coordinates of vectors, we use superscripts to index distinct vectors.

15.1.2. Theorem. (Berlekamp [1969]) If is a family of odd-size subsets of [n]

whose pairwise intersections have even size, then | | ≤ n.
Proof: Let = { A1 , . . . , Am}. We view the corresponding incidence vectors
u1 , . . . , um as elements of the space 2n , which is n-dimensional. It suffices to
show that u1 , . . . , um are linearly independent, since every n-dimensional vector
space has at most n linearly independent vectors. Computations of dot products
of vectors in 2n are modulo 2.
The conditions on the sizes of the sets and their intersections require ui · ui ≡
1 for 1 ≤ i ≤ m and ui · u j ≡ 0 for i
= j . To prove linear independence, consider
an equation of dependence: ∑i=1 ci ui = 0. For the dot product of both sides with
m

u j , the values of ui · u j reduce the equation to cj u j · u j = 0, and then cj = 0.

The argument using the dot product generalizes. For u ∈ K n , where K is a

field, we can view u · x as a polynomial function of x; that is, u · x ∈ K [x1 , . . . , x n].
The polynomial has n variables, with degree 1 in each. The argument applies for
functions in a linear (vector) space.

15.1.3. Proposition. (The Diagonal Criterion) Let º1 , . . . , ºm be elements of

a linear space of functions. If v1 , . . . , vm exist such that ºi(v j ) is nonzero for
i = j and zero for i
= j , then º1 , . . . , ºm are linearly independent.
Proof: Consider c1 , . . . , cm such that ∑i=1 ci ºi is identically zero. Evaluating
m

∑i=1 ci ºi at v j yields cj ºj (v j) = 0 and hence cj = 0.

m

Our next application shows the method more fully.

15.1.4. Definition. A ¾-distance set is a set of points such that the distances
between points lie in a set of at most ¾ numbers.

For example, a one-distance set in n must lie at the vertices of a simplex, so

the size of a 1-distance set in n is at most n + 1. For a two-distance set, Exercise
9 requests a construction for a lower bound of (n+2 1 ). We prove an upper bound
that is not much larger.

15.1.5. Theorem. (Larman–Rogers–Seidel [1977]) Every two-distance set in n

has at most (n + 1)(n + 4)/2 points.

Proof: Let {v1 , . . . , vm} be a two-distance set, using distances c and d. To avoid
square roots, we compute with squared distances. Write  x − y2 for the square
of the distance between x and y; it equals ∑ j =1 (x j − yj )2 .
n

Define º1 , . . . , ºm by ºi(x) = ( x − vi 2 − c2)( x − vi 2 − d2). Note ºi(vi) =

c d
= 0, and ºi(v j ) = 0 for i
= j , since v j − vi  ∈ {c , d}. By the Diagonal
2 2

Criterion, º1 , . . . , ºm are linearly independent.

Section 15.1: Dimension and Polynomials 725

To obtain a good bound on m, we want to capture º1 , . . . , ºm within a small

subspace of [x1 , . . . , x n]. Written as a polynomial, we have
n n
ºi(x) = (∑(x¾ − v¾) − ci 2 2
)(∑(x¾ − v¾i )2 − d2 ).
¾ =1 ¾ =1

When expanded completely, the total degree in each monomial term is at most 4.
The polynomial is a linear combination of such monomials. The number of ways
to distribute total degree at most 4 over n variables, forming such monomials, is
less than n4 . Hence m < n4 .
To prove a better bound, we capture ºi in the span of fewer monomials. When
n 2
expanding the product, the terms with degree 4 are all generated by ( ∑¾=1 x2¾ ) .
Those with degree 3 can be grouped as multiples of x j ( ∑¾=1 x2¾ ) . Thus ºi is a
n

linear combination of polynomials of the following forms:

n n
2
(∑ x2¾ ) , x j (∑ x2¾ ) , x j x¾ , xj , 1,
¾ =1 ¾ =1

where j , ¾ ∈ [n]. The number of such polynomials is 1 + n + (n2) + n + n + 1, which

simplifies to (n + 1)(n + 4)/2.

15.1.6. Remark. The proof of Theorem 15.1.5 outlines the dimension argu-
ment to show that a set S has size at most m.
(1) Define polynomials associated with the elements of S.
(2) Show that the polynomials are linearly independent.
(3) Show that the polynomials are spanned by a set of size m.
Step 3 shows that the polynomials lie in a space of dimension at most m. Since
they are linearly independent, there are at most m of them.
The resulting bound can be improved by adding polynomials if the augmented
family is still linearly independent. Blokhuis [1984] did this to improve the bound
in Theorem 15.1.5 from (n + 1)(n + 4)/2 to (n + 1)(n + 2)/2. To the polynomi-
als º1 , . . . , ºm , he added the constant polynomial 1 and the linear polynomials
x1 , . . . , x n of degree 1. The full set is spanned by the same polynomials as before
and is still linearly independent, so the bound | S| ≤ m becomes | S| ≤ m − (n + 1)
(see Exercise 11).
We will apply this augmentation technique in Theorem 15.1.23.

Another criterion for linear independence holds more often than the diagonal
criterion; we will use it to study a measure of graph complexity.

15.1.7. Proposition. (The Triangular Criterion) Let º1 , . . . , ºm be functions

in a linear space. If v1 , . . . , vm exist such that ºi(v j ) is nonzero for i = j and
zero for i > j , then º1 , . . . , ºm are linearly independent.
m
Proof: Consider coefficients c1 , . . . , cm such that ∑i=1 ci ºi is the identically-zero
function. Evaluating at v1 yields c1 º1 (v1) = 0 and hence c1 = 0. Inductively, if
c1 = · · · = c j −1 = 0, then evaluating ∑i=1 ci ºi at v j yields cj ºj (v j ) = 0, because
m

the earlier terms have coefficient 0 and the later functions evaluate to 0. Hence
cj = 0 for 1 ≤ j ≤ m.
726 Chapter 15: Linear Algebra

15.1.8. Definition. A product representation of length d encodes the vertices

of a graph G as distinct d-tuples so that uv ∈ E(G) if and only if the codes
for u and v differ in every position. The product dimension pdim (G) of G
is the minimum length of a product representation.

15.1.9. Example. Product dimension of K n , K n , and K 1 + K n−1 . Trivially,

pdim (K n) = 1. Since every two vectors must agree somewhere but the vectors
must be distinct, pdim (K n) ≥ 2; assigning (0 , j) to vj suffices.
For K 1 + K n−1 , the vectors for the clique must be distinct in each coordi-
nate. The isolated vertex must agree with the others somewhere, but it cannot
agree with more than one in each coordinate. Hence at least n − 1 coordinates
are needed. This suffices, by using (1 , 2 , . . . , n − 1) for the isolated vertex and
(i , i , . . . , i) for the ith vertex of the clique.

Product dimension is also called Prague dimension in honor of the seminal

paper of Lovász–Nešet řil–Pultr [1980]. They proved for n ≥ 3 that pdim (G) ≤
n − 1 when G has n vertices (Exercise 12). Using linear algebra, they proved a
general lower bound whose sharpness is shown by the next construction.

15.1.10. Example. pdim ( n2 K 2) = ⌈ lg n⌉ . Given ¾ coordinates, the graph that

results from using all binary ¾-tuples as vertex encodings is 2¾−1 K 2 . Each vector
disagrees everywhere only with its complement, so the graph is a matching. If n
is not a power of 2, then we can discard some complementary pairs and still encode
2 K 2 . Thus ⌈ lg n⌉ is an upper bound. Equality follows from the next theorem.
n

15.1.11. Theorem. (Lovász–Nešet řil–Pultr [1980]) If V(G) has vertices u1 , . . . , ur

and v1 , . . . , vr such that ui vi ∈ E(G) for all i and ui vj ∈
/ E(G) for i < j , then
G has product dimension at least ⌈ lg r⌉ .
Proof: Given a representation in d coordinates, let x1 , . . . , x r and y1 , . . . , y r be
the encodings of the specified vertices: x i and y i differ in each coordinate, but
x i and y j agree in some coordinate when i < j . Thus ∏¾=1 (x¾i − y¾j ) is nonzero
d

when i = j and zero when i < j . Using this observation, we construct r linearly
independent vectors in 2 ; this yields r ≤ 2 d and hence d ≥ ⌈ lg r⌉ .
d

Expanding the product ∏¾=1 (w¾ − ¾) for two vectors w , ∈ d yields the
d

sum ∑S⊆[d] ∏i∈S wi ∏ j ∈S(− j). We view this as a dot product of vectors in 2 d-
dimensional space, with coordinates indexed by subsets of [d]. For w ∈ d , we
define two vectors w and ŵ in 2 by setting wS = ∏i∈S wi and ŵS = ∏i/∈S(−wi)
d

for each S ⊆ [d]. With this definition, ∏¾=1 (w¾ − ¾) = w · ˆ . The conditions on
d

x1 , . . . , x r and y1 , . . . , y r thus imply that xi · ŷ j is nonzero if i = j and zero if i < j .

The dot product with a fixed vector is a linear function: let i(ŷ) j = x i · ŷ j .
With ŷ1 , . . . , ŷr , these 1 , . . . , r satisfy the Triangular Criterion (indexed in re-
verse). Hence x1 , . . . , xr are independent, yielding r ≤ 2 d as desired.

Theorem 15.1.11 does not require the two lists to be disjoint, just that vi is
the first vertex in the second list adjacent to ui . When applied to n2 K 2 , both lists
have all n vertices; each list runs through both halves of the matching to prove
pdim n2 K 2 ≥ ⌈ lg n⌉ , and thus the encoding in Example 15.1.10 is optimal.
Section 15.1: Dimension and Polynomials 727

RESTRICTED INTERSECTIONS OF SETS (optional)

Restricting the intersections of sets in a family restricts | | . Recall that a

family of sets is an intersecting family if every two members have a nonempty
intersection. For n ≥ 2 ¾ , a special case of the Erdős–Ko–Rado Theorem (Theorem
11.2.18) states that the maximum size of an intersecting family of ¾-sets in [n] is
−1
(¾n− 1
). We can generalize by specifying all allowed sizes of intersections.


15.1.12. Definition. For L ⊆ 0 , an L-intersecting family of sets is a family
such that | A ∩ B| ∈ L for all A , B ∈ .

In this language, the Erdős–Ko–Rado Theorem uses L = {1 , . . . , ¾ − 1}. With

L = {0 , . . . , ¾ − 1}, trivially all ¾-sets can be included, yielding a family of size
(| Ln|). If we do not restrict to a ¾-uniform family, then we can include all sets of
size at most ¾ , yielding size ∑|i=|0 (ni). Frankl–Wilson [1981] proved that no L-
L

intersecting family is larger.

To prove the Frankl–Wilson Theorem, we use the Triangular Criterion and
introduce another method.

15.1.13. Remark. A polynomial is multilinear if each positive exponent in its

expression as a sum of monomials is 1. Given a polynomial º in n variables, the
multilinear reduction of º is the multilinear polynomial ºˆ whose monomials
have the same variables as corresponding monomials in º ; that is, each positive
exponent in the expression of º is reduced to 1. Because 0r = 0 and 1 r = 1 for

r ∈ , the values of º and ºˆ agree on {0 , 1}n . If º1 , . . . , ºm are linearly indepen-
dent because no linear combination of them is 0 on all of {0 , 1}n , then the same
holds for ºˆ1 , . . . , ºˆm . The idea is to bound m by bounding the number of linearly
independent multilinear polynomials.

15.1.14. Theorem. (Frankl–Wilson [1981]) If is an L-intersecting family of

subsets of [n], where | L| = s, then | | ≤ ∑i=0 (ni).
s

Proof: (Babai [1988]) Let = { A1 , . . . , Am}, indexed so | A1 | ≤ · · · ≤ | Am|. Let

L = {l1 , . . . , ls}. Define polynomials º1 , . . . , ºm by ºi(x) = ∏¾ : l¾ <| Ai |(x · vi − l¾),
where vi is the incidence vector of Ai . By construction, ºi(v j )
= 0 for j = i. Using
the indexing of by size, | A j ∩ Ai | < | Ai | for j < i, and hence ºi(v j ) = 0 for j < i.
By the Triangular Criterion, º1 , . . . , ºm are linearly independent.
Since v1 , . . . , vm ∈ {0 , 1}n , the computations are the same for the multilin-
ear reductions ºˆ1 , . . . , ºˆm , so those also are linearly independent. We prove the
desired bound by capturing them in a small space. Each ºi is the product of at
most s linear factors, so the total degree of each monomial in the expansion of ºi
is at most s. Hence the multilinear reduction of ºi is in the span of the products
s
of at most s distinct variables. There are ∑ i=0 (ni) such monomials.

The bound in Theorem 15.1.14 can be strengthened when the requirement

| i | L for Ai ∈ is added. For example, for odd-sized sets and L all even, the
A ∈
/
Oddtown theorem (Theorem 15.1.2) implies | | ≤ n. This is consistent with the
bound (| Ln|) when we view the intersection sizes as congruence classes modulo 2.
728 Chapter 15: Linear Algebra

Class 0 contains all intersection sizes, but the sizes of the sets are forbidden from
that class. See also Exercise 13.
A variation of Theorem 15.1.14 is proved in Lemma 16.2.40 (by a similar
method) and applied in Theorem 16.2.41 to solve a question asked by Borsuk.
Meanwhile, here we present a modular version of Theorem 15.1.14. The proof is
analogous, and the bound is the same.

15.1.15. Definition. For L ⊆ p with p prime, write t ∈ L (mod p) when t ≡

l (mod p) for some l ∈ L. A family ⊆ 2[n] is p-modular L-intersecting if
/ L (mod p) for A ∈ and | A ∩ B| ∈ L (mod p) for distinct A , B ∈ .
| A| ∈

15.1.16. Theorem. (Deza–Frankl–Singhi [1983]) If is a p-modular L-intersect-

ing family of subsets of [n], where | L| = s and p is prime, then | | ≤ ∑i=0 (ni).
s

Proof: (Alon–Babai–Suzuki [1991]) Let = { A1 , . . . , Am}, with vi the incidence

vector of Ai . Define polynomials º1 , . . . , ºm by ºi(x) = ∏ l∈ L(x · vi − l). Perform
computations over p , and write “ = ” instead of “ ≡ ”. Note that v j · vi = | Ai ∩ A j |.
Thus ºi(v j )
= 0 for i = j (since | Ai | ∈
/ L (mod p)), while ºi(v j ) = 0 for i
= j (since
| Ai ∩ Aj | ∈ L (mod p)). By the Diagonal Criterion, º1 , . . . , ºm are independent.
The bound now follows in the same way as in Theorem 15.1.14. The multi-
linear reductions ºˆ1 , . . . , ºˆm are linearly independent by the same criterion as
º1 , . . . , ºm , and they lie in a space of dimension ∑is=0 (ni).
s
It may seem that the bound ∑ i=0 (ni) is much larger than (ns), but actually it
is not much bigger when s is not too big.

15.1.17. Lemma. If n ≥ 2s and s = n/r, then

s
n n s n 1
∑ ( i ) ≤ ( s ) (1 + n − 2s + 1 ) < ( s ) (1 + r − 2 ) .
i=0

In particular, if s ≤ n/3, then ∑i=0 (ni) < 2(ns).

s

Proof: By factoring (ns) from each term and then enlarging (and extending) the
terms to obtain a geometric series,
s
n n s s(s − 1)
∑ ( i ) = ( s ) (1 + n − s + 1 + (n − s + 1)(n − s + 2) + · · ·)
i=0
n s s2
≤ ( ) (1 + + + · · ·)
s n − s + 1 (n − s + 1)2
n 1 n n− s+1 n s
=( ) =( ) = ( ) (1 + )
s 1 − n− s+1
s
s n − 2s + 1 s n − 2s + 1
n n/r n 1
= ( ) (1 + ) < ( ) (1 + ) .
s n − 2n/r + 1 s r−2

These results yield a constructive superpolynomial lower bound for the Ram-
sey number R(t , t), weaker than Erdős’ nonconstructive exponential bound, but
using explicit graphs. Coloring (t − 1)K t−1 red and K t−1 ,... ,t−1 blue yields R(t , t) >
Section 15.1: Dimension and Polynomials 729

(t − 1)2 . Nagy [1972] increased this to (t3 ) (Exercise 18). Frankl [1977] found
3

graphs proving R(t , t) > t (t) using -systems (sunflowers), where (t) → ∞.
Frankl–Wilson [1981] obtained similar results from p-modular L-intersecting
families. (A vertex set is homogeneous if it is a clique or independent set.)

15.1.18. Theorem. (Frankl–Wilson [1981]) Let p be a prime, and choose n > 2p2 .
Let G be the graph with vertex set ( p[n]
2 −1 ) defined by AB ∈ E(G) if and only if

| A ∩ B|
≡ −1 (mod p). The graph G has no homogeneous set with more than
2( p−n1 ) vertices. As a consequence, R(t , t) > t(1−) (t) , where (t) = 4 ln
ln t
ln t .
Proof: If A1 , . . . , Am is a clique in G, then it is a p-modular L-intersecting fam-
ily, where L = {0 , . . . , p − 2}, because | Ai | = p2 − 1 ≡ −1 (mod p), and Ai A j ∈ /
E(G) when | Ai ∩ A j | ≡ −1 (mod p). With | L| = p − 1, Theorem 15.1.16 yields
p−1
m ≤ ∑i=0 (ni) < 2( p−n1 ). If A1 , . . . , Am is an independent set, then | Ai ∩ A j | ∈
{p − 1 , 2p − 1 , . . . , p2 − p − 1}. Here p − 1 intersection sizes are allowed, so Theo-
p−1
rem 15.1.14 yields m ≤ ∑i=0 (ni) < 2( p−n1 ).
Fixing t, let p be the largest prime with 2( pp−1 ) < t, and let n = p3 . We have
3

(1−) (t)
proved R(t , t) > ( p2n−1). The choice of p yields p ∼ 2 ln ln t , and then p2 −1 > t
3
ln t
( p ) ,
where (t) = 4 ln
3 3
ln t
(p ) (p )
ln t . That is, we compare roughly p for t with p2 for the lower
bound on R(t , t). The logarithm of the latter is roughly p/2 times the logarithm
of the former, so roughly R(t , t) > t p/2 (Exercise 16 requests further details).

We present another application. The famous Hadwiger–Nelson problem

(Hadwiger [1945]) asks for the fewest colors in a coloring of n such that no two
points at distance 1 have the same color. For n = 2, the answer is in {4 , 5 , 6 , 7}
(Exercise 8.1.13). There is an easy general upper bound of nn/2 (Exercise 15).
Larman–Rogers
√ [1972] presented a quadratic lower bound and an upper bound of
(2 2 + o(1))n and conjectured an exponential lower bound. Frankl–Wilson [1981]
proved that; we obtain it from a corollary of Theorem 15.1.16.

15.1.19. Corollary. Let p be a prime, and let be a (2p − 1)-uniform family of

subsets of [4p − 1]. If no two members of have exactly p − 1 common ele-
−1
ments, then | | ≤ 2(4p
p−1
).
Proof: Let L = {0 , . . . , p− 2}. The family is p-modular L-intersecting since is
(2p − 1)-uniform with 2p − 1 ∈ / L (mod p), and the remaining possible intersection
sizes {p , . . . , 2p − 2} are congruent to elements of L. By | L| = p − 1, Theorem
15.1.16 and Lemma 15.1.17 yield the bound.

15.1.20. Theorem. (Frankl–Wilson [1981]) For large n, the chromatic number

of the unit-distance graph in n is greater than 1 .1397n .
Proof: The graph defined using distance d is isomorphic to the unit-distance
graph. Hence it suffices to prove the lower bound for a subgraph of the distance-d
graph. We use an appropriate d and a subgraph induced by some binary n-tuples
with ¾ nonzero positions.
The squared distance between two points in {0 , 1}n is the number of coordi-
nates where they differ. As incidence vectors of subsets A and B of n, they differ
730 Chapter 15: Linear Algebra

in | A B| places. If A and B have size ¾ , then the symmetric difference has size
2(¾ − | A ∩ B|). Hence forbidding one distance between the points is equivalent
to forbidding one intersection size for the sets. If ¾ = 2p − 1, then forbidding
intersection size p − 1 is equivalent to forbidding squared distance 2p.
Let p be the largest
√ prime such that 4p − 1 ≤ n; we use only 4p − 1 of the coor-
dinates. Let d = 2p. By Corollary 15.1.19, the maximum size of an independent
set in the subgraph of the distance-d graph induced by the incidence vectors of
−1
the (2p − 1)-sets in [4p − 1] is at most 2(4p
p−1
). Hence the chromatic number is at
4p−1 4p−1 4p 4p p!(3p)!
least (2p−1 )/2( p−1 ), which equals (2p)/( p ) and simplifies further to (2p)!(2p)! .
With Stirling ’s Formula
√ (Application 2.3.8), the ratio is approximately
c(33/4/2)4p , where c = 3/4. When m is large, there is a prime between m and
m − m7/12 (Huxley [1973]). Applying this with m = n/4 completes the proof, since
33/4/2 > 1 .1397. (Note: Exercise 17 improves the bound to about (1 .2)n .)

Frankl–Füredi [1984] conjectured that the bound in Theorem 15.1.14 on the

size of an L-intersecting family improves from ∑ i=0 (ns) to ∑i=0 (n−s 1 ) when L =
s s

[s]. Ramanan [1997] proved this. Snevily conjectured and proved a substantial
generalization, after earlier proving special cases (Snevily [1994, 1999]).

15.1.21. Theorem. (Snevily’s Theorem; Snevily [2003]) If L is a set of s

positive integers, and is an L-intersecting family of subsets of [n], then
| | ≤ ∑is=0 (n−i 1).

Fisher ’s Inequality (Theorem 13.1.15) states that n ≥ v when is a family of

n members of ([v] ¾ ) in
[v] (with ¾ < v) such that every two elements of [v] appear in
ë common blocks. The dual statement (transposing the incidence matrix) is that
an L-intersecting family of subsets of [n] has size at most n when L = { ë}. Since
n = (n−0 1 ) + (n−1 1), Fisher ’s Inequality follows from the case s = 1 of Snevily ’s
Theorem, which was proved by Majumdar [1953].
Instead of Snevily ’s Theorem, we present only a modular version that he
proved earlier. The proof uses a refinement of the dimension argument. To the
set of polynomials obtained from elements of  , we add more polynomials. If the
full set is still linearly independent in the same space, then the bound on  be-
comes the dimension of the space minus the number of extra polynomials. This
technique is also used in Exercise 11 to improve the bound on two-distance sets.

15.1.22. Lemma. Let C1 , . . . , Ct be a family of subsets of [n]. If polynomials

h1 , . . . , ht are defined on n by hj (x) = ∏r∈ Cj xr , then h1 , . . . , ht are linearly
independent on {0 , 1}n .
Proof: Let w j be the incidence vector of Cj , indexed so | C1 | ≤ · · · ≤ | Ct |. Note that
hj (w j) = 1. If i > j , then the indexing guarantees an element r ∈ Ci − Cj . Now hi
has x r as a factor, but the value in coordinate r of w j is 0, so hi(w j) = 0. By the
Triangular Criterion, {h1 , . . . , ht } is linearly independent.

15.1.23. Theorem. (Snevily [1994]) Let p be a prime number. If  is a p-modular

L-intersecting family of subsets of [n] and | L| = s, then || ≤ ∑i=0 (n−i 1).
s
Section 15.1: Dimension and Polynomials 731

Proof: Let = { A1 , . . . , Am}, indexed so that A1 , . . . , A q omit the element 1 and

A q+1 , . . . , Am contain it. Begin the proof as in Theorem 15.1.16, letting ºi(x) =
∏ l∈ L(x · vi − l), where vi is the incidence vector of Ai . Again the multilinear re-
ductions ºˆ1 , . . . , ºˆm are spanned by the ∑i=0 (ni) multilinear monomials of degree
s

at most s.
Let C1 , . . . , Ct be all the sets of size less than s in [n] that omit element 1,
indexed so | C1 | ≤ · · · ≤ | Ct |. Define polynomials hj and j by hj (x) = ∏r∈ Cj xr and
j (x) = (x1 − 1)hj (x). Each j has degree at most s and is multilinear and spanned
by the same set of ∑i=0 (ni) monomials as ˆ1 , . . . , ˆm . Since t = ∑i=1 (ni−−11), it suf-
s s

fices to prove that {ˆ1 , . . . , ˆm} ∪ { 1 , . . . , t} is linearly independent.

Let P = ∑i=1  i ˆi + ∑ j =1 j j , with each  i and j in p . Suppose that P
m t

is identically 0. Let Ai = Ai ∪ {1}, and let y i be the incidence vector of A i , for

1 ≤ i ≤ m. Each y i has 1 in the first coordinate, so the contribution of the second
sum to P(y i) is always 0.
Note that A j ∩ Ai = A j ∩ Ai if i ≤ j . This holds because A j = A j if j > q and
1∈ / Ai if i ≤ q. Thus i(y j ) = i(v j ) for i ≤ j . Since i(v j ) = 0 if and only if i
= j
and ˆi(x) = i(x) when x ∈ {0 , 1}n , evaluating P at ym , . . . , y1 successively yields
 m , . . . , 1 = 0.
By Lemma 15.1.22, h1 , . . . , ht are linearly independent. Multiplying them
all by x1 − 1 leaves 1 , . . . , t independent. Since  m , . . . , 1 = 0, making P iden-
tically 0 thus also requires 1 , . . . , t = 0. Hence there is no equation of linear
dependence for {ˆ1 , . . . , ˆm} ∪ { 1 , . . . , t }.

Taking p sufficiently large in Theorem 15.1.23 yields the bound in Theorem

15.1.21 for the special case of Theorem 15.1.21 where the sizes of members of the
intersecting family do not lie in L. Extensions of Theorem 15.1.23 to -wise
intersections appear in Grolmusz–Sudakov [2002] and Cao–Hwang–West [2007].
The top two entries in ∑i=0 (n−i 1 ) sum to (ns). We next show that this smaller
s

value tightens the upper bound on the size of an L-intersecting family when
we add the restriction that is a uniform family, even if 0 is allowed in L. The
original proof used linear algebra in a different way; combining the ideas we have
presented led to a shorter proof.

15.1.24. Theorem. (Ray-Chaudhuri–Wilson [1975]) If n ≥ 2s, and L is a set of s

nonnegative integers, then every L-intersecting -uniform family of subsets
of [n] has size at most (ns).
Proof: (Alon–Babai–Suzuki [1991]) We may assume  ∈ / L. Let v1 , . . . , vm be the
incidence vectors for the sets in the family. Let i(x) = ∏ l∈ L(x · vi − l). By the
Diagonal Criterion, 1 , . . . , m are linearly independent on {0 , 1}n . Each i has
degree s; the multilinear reduction yields polynomials ˆ1 , . . . , ˆm that are lin-
s
early independent on {0 , 1}n and spanned by the ∑i=0 (ni) multilinear monomials
with degree at most s.
As in Theorem 15.1.23, we add polynomials to this set. Let C1 , . . . , Ct be the
sets of size less than s in [n]. Define hj by hj (x) = ∏r∈ Cj xr . By Lemma 15.1.22,
h1 , . . . , ht are linearly independent over {0 , 1}n . Define j by j (x) = (x · 1 n −
)hj (x). As in Theorem 15.1.23, we have multiplied independent polynomials by
one linear factor, and the resulting polynomials are independent.
732 Chapter 15: Linear Algebra

Let P = ∑i=1 i i + ∑ j =1 j  j . Consider coefficients 1 , . . . ,  m and 1 , . . . , t

m t

such that P is identically 0. Since is -uniform, the contribution from the sec-
ond sum is 0 when evaluated at vi . Since i(v j ) = 0 when j
= i, evaluating P
at vi thus yields  i = 0. With each  i being 0, linear independence of  1 , . . . ,  t
implies also that each j is 0.
We conclude that { 1 , . . . , m} ∪ { 1 , . . . ,  t } is linearly independent. Again
we may take the multilinear reduction of  j since independence was established
by evaluation over {0 , 1}n . The degree of  j is at most s, so these vectors also lie
s
in the span of the ∑i=0 (ni) multilinear monomials with degree at most s. Since
s −1
t = ∑ i=0 (ni), we conclude m ≤ (ns).

The bound (ns) holds trivially with equality when L = {0 , . . . , s − 1} and  = s.

Is it still sharp for larger  ? Since (ns) ∼ ns/s!, the construction below is not so
much smaller than the upper bound when s and t are fixed and n is large.

15.1.25. Theorem. For n ≥ 2  2 ≥ 2s2 and L = {0 , . . . , s − 1}, some -uniform

L-intersecting family satisfies | | > (n/2 )s .
Proof: Let p be the largest prime bounded by n/ , so n/2  < p ≤ n/ . Fix a -set
A contained in p ; our family will consist of ps -sets within A × p . Given a
polynomial of degree less than s, let Aº = {(a , (a)): a ∈ A}. There are ps poly-
nomials over p with degree less than s, so this defines ps sets of size  . Distinct
polynomials of degree d over p agree on at most d points in p . Since  ≥ s, the
ps sets are distinct, and any two of them have fewer than s common elements.

COMBIN ATORIAL NULLSTELLENSATZ

The Combinatorial Nullstellensatz is a result about zeros of multivariable

polynomials over a field. Surprisingly easy to prove, it has many applications
in additive number theory, discrete geometry, and graph theory. Applying it is
sometimes called the polynomial method. The theorem was presented by Noga
Alon at a conference in 1995, but the proceedings did not appear until 1999. Alon
had already applied the theorem in at least five papers with eight coauthors from
1984 to 1996, proving new results and giving short proofs of old results. Many
others have since also used it. Tao [2014] surveyed applications in arithmetic
combinatorics and number theory; Clark [2014] explored further extensions.
We need a lemma that generalizes to n variables the familiar statement that
a nonzero polynomial of degree d in one variable takes the value 0 at most d times.
That statement is proved by induction on d, using the Euclidean algorithm to fac-
tor out x −  when  is a root. The discussion is valid in any field. We compute
with equalities rather than using congruence notation for finite fields. Also, su-
perscripts return to being exponents.

15.1.26. Lemma. Let be a polynomial in n variables x1 , . . . , x n , over a field K .

For each i, let the degree of as a polynomial in x i be at most d i , and let Si
be a set of d i + 1 distinct values in K . If (x1 , . . . , x n) = 0 for (x1 , . . . , x n) ∈
n
∏i=1 Si , then is identically 0.
Section 15.1: Dimension and Polynomials 733

Proof: We take the result in one variable as the basis for induction on n. For
n > 1, we collect terms to write º as a polynomial in x n . That is, º =
∑ j =n 0 ºj (x1 , . . . , x n−1)x n , where each ºj is a polynomial having degree at most d i in
d j

n−1
each x i . For (x1 , . . . , x n−1) ∈ ∏i=1 Si , evaluating º0 , . . . , ºd n yields a one-variable
polynomial in x n of degree at most d n . Furthermore, the hypothesis implies that
this polynomial is 0 for x n ∈ Sn .
By the basis step (n = 1), the one-variable polynomial we obtain for a fixed
(x1 , . . . , x n−1) ∈ ∏in=−11 Si is the zero polynomial. Thus each ºi is 0 at all values
n−1
in ∏i=1 Si . By the induction hypothesis, each ºi is identically zero. Thus the
coefficients of º are all zero, and º is identically zero.

We want that if the coefficient of a term ∏ x tii is nonzero in a polynomial º

of degree ∑ t i , and | Si | > t i for all i, then the polynomial is nonzero at some point
in ∏ Si . The lemma does not say this, since other terms may have degree larger
than t i in x i , for some i. Fortunately, it is not hard to overcome this technicality.
The degree of a polynomial is the maximum, over all monomials, of the sum
n
of the exponents on the variables. We extract the coefficient of a monomial ∏i=1 x tii
in a polynomial º (x1 , . . . , x n) using the coefficient operator [ ∏i=1 x i ] , which
n ti

for formal power series in one variable was used extensively in Chapter 3.
n
15.1.27. Theorem. (Combinatorial Nullstellensatz; Alon [1999]) If ∏i=1 x tii is
a monomial with nonzero coefficient in a polynomial º having degree ∑i=1 t i
n

over a field K , and S1 , . . . , Sn are sets with | Si | > t i for 1 ≤ i ≤ n, then

º (x)
= 0 for some x ∈ ∏ Si .
Proof: It suffices to prove the statement when | Si | = t i + 1 for each i. The idea is
to change º into another polynomial ºˆ that agrees with º on ∏ Si but has degree
at most t i as a polynomial in x i , for each i. Lemma 15.1.26 then implies that
ºˆ(x)
= 0 for some x ∈ ∏ Si . Since ºˆ agrees with º on ∏ Si , also º (x)
= 0.
For each index i, define a polynomial ½ i by ½ i(x) = ∏ s∈Si (x i − s); note that
½ i depends only on x i . It has degree t i + 1 in x i and degree 0 in other variables.
Expanding the product yields ½ i(x) = x tii +1 − hi(x), where hi is a polynomial with
degree at most t i in x i and degree 0 in other variables.
By definition, ½ i(x) = 0 for x ∈ ∏ Si , since x i ∈ Si in that case. Therefore,
t i +1
xi = hi(x) for all x ∈ ∏ Si . Thus we can replace each appearance of a variable
having too large an exponent with a polynomial having smaller degree in that
variable. By making such a replacement as long as the polynomial still has degree
greater than t i in some x i , we obtain ºˆ having degree at most t i in x i for each i.
We must also show [ ∏ x tii ] ºˆ(x)
= 0. Since no exponent is too large, we made
no change to that term. Also we did not introduce any terms that could cancel
it; since º has degree ∑ t i , any monomial containing a variable with too large an
exponent has some x j with exponent less than t j . Since the substitutions increase
no exponents, no substitution can introduce a contribution to [ ∏ x tii ] .

An early application was a short proof of the Cauchy–Davenport Theorem

of additive number theory, first proved by Cauchy in 1813 and by Davenport in
1935. Let A and B be subsets of n , with | A| = a and | B| = b. How many elements
of n must arise as x + y with x ∈ A and y ∈ B?
734 Chapter 15: Linear Algebra

Setting A = {0 , . . . , a − 1} and B = {0 , . . . , b − 1} shows that the number

can be as small as min{n , a + b − 1}. If a + b > n, then for any c ∈ n the sets A
and {c − y: y ∈ B} must intersect, and when an element x is in the intersection
we have x ∈ A and y ∈ B such that c = x + y. Hence the number of sums always
equals n if a + b > n. Also, when n = 2 ¾ , taking the “even” classes for both A and
B yields only even classes as sums, so here the sum can be as small as n/2 even
though a = b = n/2. This suggests restricting our attention to prime moduli.

15.1.28. Theorem. (Cauchy–Davenport Theorem) If p is prime, and A , B ⊆

p with | A| = a and | B| = b and a + b ≤ p, then the smallest possible size of
{x + y: x ∈ A , y ∈ B} is a + b − 1.
Proof: It suffices to prove the lower bound; equality holds in the construction
presented above. Suppose that there are fewer than a + b − 1 sums. Let C be a
set of size a + b − 2 in p that contains all the sums. Let º (x , y) = ∏c∈ C(x + y − c),
over p . We have a polynomial in two variables, and its degree is a + b − 2.
We claim that [x a−1 y b−1 ]º (x , y) = (a+a−b−1 2)
≡ 0 (mod p). Contributions to this
coefficient use x or y in each factor when expanding º , choosing x exactly a − 1
times and y exactly b − 1 times. The number of ways to do that, each contribut-
ing +1 to the coefficient, is (a+a−b−1 2). Finally, that binomial coefficient is nonzero
modulo p since a + b − 2 < p; there is no factor of p in the numerator and no other
way to introduce a factor of p.
Since | A| = a and | B| = b, the Combinatorial Nullstellensatz yields x ∈ A and
y ∈ B such that º (x , y)
= 0. This is a contradiction, since º was constructed to
be 0 at all such pairs (x , y).

This short proof illustrates the method for applying the Combinatorial Null-
stellensatz. Using the set of sums, we design º that is 0 at (x , y) when x ∈ A and
y ∈ B. If the set of sums is too small, then A × B is too big for º to be identically
0 there when the appropriate coefficient is nonzero.
When A = B, the lower bound in Theorem 15.1.28 is min{2 | A|− 1 , p}. Erdős–
Heilbronn [1964] conjectured that almost as much is forced even when ignoring
the contributions by adding elements to themselves. Given the ease of proving
this from the Combinatorial Nullstellensatz, it is remarkable that the problem
was open for 30 years. The original proof used exterior algebra and representa-
tion theory of the symmetric group.

15.1.29. Theorem. (Erdős–Heilbronn Conjecture; Dias da Silva & Hami-

doune [1994]) If A ⊆ p , where p is prime, and C is the set of sums of distinct
elements of A, then | C| ≥ min{2 | A| − 3 , p}.
Proof: (Alon–Nathanson–Rusza [1996]) Since there are only p classes, we may
assume 2a − 3 < p, where a = | A| . As in the proof of Theorem 15.1.28, we design
a polynomial º that is 0 at (x , y) when x + y ∈ C. The polynomial is the same as
before, except for including the factor (x − y) to ensure that º is 0 when x = y,
since 2x may not be in C. That is, let º (x , y) = (x − y) ∏c∈ C(x + y − c). Note that
deg(º ) = m + 1, where m = | C|.
We study the coefficient of x a−1 y m− a+2 . As before, contributions to the desired
coefficient use x or y in each factor. The contributions choosing x in the first factor
Section 15.1: Dimension and Polynomials 735

are positive, and those choosing − y are negative. Thus [ x a−1 y m− a+2 ] º (x , y) =
m− a+2
−2) − ( a−1 ) = [1 − a−1 ]( a−2). If m ≤ 2a − 4, then this coefficient is nonzero,
(am m m

and a > m − a + 2. Now the Combinatorial Nullstellensatz guarantees (x , y) ∈ A2

such that º (x , y)
= 0. These are distinct elements of A whose sum is not in C,
which is a contradiction. We conclude that m ≥ 2a − 3.

The theorem below extends Theorem 15.1.29 to restricted sums over many
variables. See Exercises 23–25 for the proof and applications.

15.1.30. Theorem. (Alon–Nathanson–Rusza [1996]) Let h be a polynomial in ¾

variables over p , where p is prime. Let A1 , . . . , A¾ be nonempty subsets of
¾ ¾
p , with ci = | Ai | − 1 for all i. Let m = ∑i=1 ci − deg(h). Let C = {∑i=1 ai : ai ∈
¾ ¾
Ai and h(a)
= 0}. If [ ∏i=1 x ci i ](∑i=1 x i)m h(x)
= 0, then | C| ≥ m + 1.

Our next consequence is also number-theoretic but has a geometric applica-

tion. It was conjectured by Artin in 1934, proved by Chevalley [1935], and ex-
tended in Chevalley–Warning [1935]. The proof uses Fermat ’s Little Theorem
(Application 1.3.10), which states that if p is a prime, then a p−1 ≡ 1 (mod p) for
every integer a not divisible by p.

15.1.31. Theorem. (Chevalley–Warning Theorem) Let P1 , . . . , Pm be polyno-

mials over p in n variables, where p is prime. If ∑i=1 deg(Pi) < n and the
m

polynomials have a common zero, then they have another common zero.
Proof: Let (c1 , . . . , cn) be a common zero. Let
m n
º (x) = ∏(1 − Pi(x)p−1) − ∏(1 − (x j − cj)p−1).
i=1 j =1

Note that º (c) = 1 − 1 = 0. If there is no other common zero, then for x ∈ pn − {c},
there exists i such that Pi(x)
≡ 0 (mod p). Also there exists j such that x j
= cj .
By Fermat ’s Little Theorem, Pi(x)p−1 ≡ 1 ≡ (x j − cj )p−1 (mod p). Hence º (x) = 0.
The degree of the first term in º is bounded by (p − 1) ∑i=1 deg(Pi), which is
m

less than (p − 1)n. The degree of the second term is (p − 1)n, and the coefficient
of ∏ x jp−1 in º is (−1)n+1 , which is nonzero modulo p. Since | p| > p − 1 and we
choose each x i from p , Theorem 15.1.27 guarantees x ∈ pn such that º (x)
= 0.
By this contradiction, the polynomials must have another common zero.

Chevalley proved the case m = 1 and Warning proved the general case; both
in fact obtained p common zeros. We apply Theorem 15.1.31 to determine the
transversal number of a special hypergraph. The vertex set is pn , and for each
hyperplane H in pn we make an edge consisting of all the points in H. These
points are the solutions to a · x = b, where a ∈ pn − {0} and b ∈ p are fixed. Every
edge has pn−1 vertices. The transversal number Ì( ) of a hypergraph is the
minimum size of a vertex set intersecting all the edges.

15.1.32. Theorem. (Jamison [1977], Brouwer–Schrijver [1978]) For prime p, the

transversal number of the hypergraph of all hyperplanes in pn is n(p − 1) + 1.
736 Chapter 15: Linear Algebra

Proof: First we produce a transversal of this size. Let B be the set of points in pn
having at most one nonzero coordinate; by construction B has the specified size.
To prove that B is a transversal, we use induction on n. For n = 1, each point is
a hyperplane, and indeed B = 1p .
For n > 1, hyperplanes of the form x n = c are hit by the point in B having c in
the last coordinate. For other hyperplanes, consider the fixed hyperplane H con-
sisting of {x ∈ pn : x n = 0}. The hyperplanes of the form x n = c include H and all
hyperplanes disjoint from H. The others intersect H in a hyperplane of pn−1 ob-
tained by dropping the last coordinate (0) from the points in the intersection. By
the induction hypothesis, these hyperplanes are hit by the points in B that have
0 in the last coordinate.
For the lower bound, let B be any transversal. By applying a translation in
each coordinate, we may assume 0 ∈ B. Let A = B − {0}. The set A intersects all
hyperplanes not containing 0. This means that for all x ∈ pn − {0}, the equation
x · y = 1 has a solution y ∈ A.
Let º (x) = ∏a∈ A(x · a − 1). Since x · y = 1 has a solution in A when x
= 0, we
have º (x) = 0 for x ∈ pn − {0} and º (0) = (−1)| A| . Given variables x i for i ∈ [n]
(j)

and j ∈ [p − 1], define a polynomial P by

p−1
P = (∑ j =1 º (x1 , . . . , x n )) − (−1)| A|(p − 1).
(j) (j)

Since º takes only the values 0 and (−1)| A| , the sum has magnitude p − 1 only
when each summand is nonzero, requiring each variable to have value 0. Hence
P has value 0 only when all n(p − 1) variables have value 0. The contrapositive
of the Chevalley–Warning Theorem (for m = 1) now yields n(p − 1) ≤ deg P =
deg º = | A| = | B| − 1.

We turn now to applications in graph theory. Berge and Sauer conjectured

that every 4-regular graph has a 3-regular subgraph; Tashkinov [1984] proved
this. The claim is false for multigraphs (consider a fat triangle with edges of
multiplicity 2), but it becomes true when there is at least one “extra” edge, as
seen by setting p = 3 in the next theorem. For convenience, when v is a vertex in
a graph, let Γ(v) denote the set of edges incident to v.

15.1.33. Theorem. (Alon–Friedland–K alai [1984]) If p is prime, then every loop-

less multigraph G with average degree greater than 2p − 2 and maximum
degree at most 2p − 1 contains a p-regular subgraph.
Proof: Suppose that G has n vertices and m edges. We want to design a poly-
nomial º over the field p such that when º (x)
= 0, the point x selects for us
a p-regular subgraph. Hence we introduce a variable x e for each edge e taking
value 0 (absent) or 1 (present). To apply the Combinatorial Nullstellensatz, we
will want a multilinear monomial term with a nonzero coefficient. Define º by
p−1
º (x) = ∏ [1 − ( ∑ x e ) ]− ∏ (1 − x e).
v∈ V(G) e∈Γ(v) e∈ E(G)

Each factor in the first term has degree p − 1, so the degree of the first term
is at most (p − 1)n. This quantity is less than m, since the average degree exceeds
Section 15.1: Dimension and Polynomials 737

2p − 2. Hence the degree is determined by the second term, which has degree m,
with [ ∏e∈ E(G) x e ] º (x) = (−1)m+1
= 0.
By the Combinatorial Nullstellensatz, º (x̂)
= 0 for some x̂ ∈ {0 , 1}m . Since
º (0) = 1 − 1 = 0, this occurs with x̂
= 0. Since x̂
= 0, the second term in º (x̂)
has a factor that is 0. Hence the first term in º (x̂) must be nonzero. By Fermat ’s
Little Theorem, this requires that ∑e∈Γ(v) x̂ e is a multiple of p for every vertex v.
Therefore, the degree of each vertex in the subgraph H of G with edge set
{e ∈ E(G): x̂ e = 1} is a multiple of p. Since (G) ≤ 2p − 1, the degree is always 0
or p. Since x̂
= 0, it cannot always be 0. Thus H has a nontrivial component, and
it is a p-regular subgraph of G.

With no bound on the maximum degree, more edges may be needed to force a
¾-regular subgraph. Pyber [1985] proved that an n-vertex graph having at least
32 ¾ 2 n ln n edges has a ¾-regular subgraph. The bound is not too far from optimal:
Pyber–Rödl–Szemerédi [1995] proved by probabilistic arguments that there are
graphs with at least (n log log n) edges that have no 3-regular subgraph (and
O(n log (G)) edges force a 3-regular subgraph).
Instead of fixing the degree at each nonisolated vertex, we can be more flexi-
ble. Specify for each v ∈ V(G) a bad set B(v) ⊆ {1 , . . . , d G(v)}. We seek a subgraph
H with d H (v) ∈ / B(v) for v ∈ V(H). Shirazi–Verstraëte [2008] gave an easy proof
from the Combinatorial Nullstellensatz that there is a nontrivial such subgraph
H when ∑v∈V(G) B(v) < | E(G)| (Exercise 27), and this inequality is sharp.
They also proved a conjecture of Addario-Berry et al. [2007] that allows 0 to
be in the forbidden sets. This was stated originally in terms of allowed degrees,
but it is a bit cleaner for forbidden degrees. Consider the design of the polynomial
º . We want the multivariate point x with º (x)
= 0 to select the desired subgraph
H. Hence we make a variable for each edge, and we restrict it to the values 0 and
1 to model whether the edge is used in H. For each vertex v, we design a factor
that is 0 when the constraint at v is violated.

15.1.34. Theorem. (Shirazi–Verstraëte [2008]) For each vertex v in a graph G,

specify a bad set B(v) ⊆ {0 , . . . , d G(v)}. If | B(v)| ≤ ⌊ d(v)/2⌋ for all v, then G
has a subgraph H with d H (v) ∈ / B(v) for all v.
Proof: Let Γ(v) denote the set of edges in G incident to vertex v. Introduce a
variable x e for each edge e in G, with value 0 or 1, and let x = (x1 , . . . , x| E(G)|).
Define a real-valued polynomial º by

º (x) = ∏ ∏(∑ x e − c).

v∈ V(G) c∈ B(v) e∈Γ(v)

The variables set to 1 yield a subgraph with degree ∑e∈Γ(v) x e at v. The factor for
v is 0 if and only if that degree is forbidden. We seek x ∈ {0 , 1}| E(G)| with º (x)
= 0.
Since º is a product of linear factors, deg(º ) ≤ ∑v∈V(G) | B(v)| . By the Com-
binatorial Nullstellensatz, it suffices to find a monomial with this degree having
nonzero coefficient, whose variables all have exponent at most one. Monomials in
the product arise by choosing, for each forbidden degree at each vertex, an edge
incident to that vertex. We must not choose a given edge from both endpoints.
738 Chapter 15: Linear Algebra

To avoid repeated selection, we orient G and pick for the monomial at v only
variables x e such that v is then the tail of e. If the orientation has at least ⌊ d(v)/2⌋
edges leaving each vertex v, then there are enough such edges to choose distinct
ones for the elements of B(v), since | B(v)| ≤ ⌊ d(v)/2⌋ . To form an orientation D
such that d+D(v) ≥ ⌊ d G(v)/2⌋ , simply add a vertex w adjacent to all vertices of odd
degree in G and orient by following an Eulerian circuit in each component.
We thus obtain a linear monomial. Every contribution to the coefficient of
a monomial with degree ∑v∈V(G) | B(v)| is positive, since obtaining that degree re-
quires selecting some x e (and not c) from each factor.
Finally, when x is the point with º (x)
= 0 guaranteed by the Combinatorial
Nullstellensatz, each factor must be nonzero, which means that the number of
edges selected at v (via x e = 1) does not lie in B(v).

Theorem 15.1.34 is sharp; the conclusion may fail when one bad set is a bit
too large. Let G = K 2r,2r with bipartition X , Y . If B(x) = {0 , . . . , r − 1} for x ∈
X and B(y) = {r + 1 , . . . , 2r} for y ∈ Y , then each B(v) has size d(v)/2, and a
subgraph is good if and only if it is r-regular. When r is added to one bad set,
there is no longer a good subgraph.

THE ALON–TARSI THEOREM

Our final application of the Combinatorial Nullstellensatz to graphs is one

of the most famous. The Alon–Tarsi Theorem uses a polynomial associated with
a graph G to obtain upper bounds on the list chromatic number Ò l(G). We first
state the result, which can be applied without knowing the algebraic background.

15.1.35. Definition. A circulation is a digraph D such that d+D(v) = d−D(v) for

all v ∈ V(D). The parity of a circulation D is the parity of | E(D)| . For any
digraph D , let diff(D ) denote the absolute difference between the number of
even spanning circulations and the number of odd spanning circulations.

In this context, what we call circulations have usually been called “Eulerian
subgraphs” (see Alon [1993]). We use “circulation” because there is no connected-
ness requirement and because the term is used analogously with network flows.

15.1.36. Theorem. (Alon–Tarsi Theorem; Alon–Tarsi [1992]) A graph G hav-

ing an orientation D such that diff(D)
= 0 is º -choosable, where º (v) =
1 + d+D(v) for each v ∈ V(D).

15.1.37. Example. Let G = Cn . Let D be a cyclic orientation of G. Here d+D(v) =

1 for all v, and the only spanning circulations in D are the trivial subgraph (no
edges) and D itself. If n is even, then diff(D) = 2, and G is 2-choosable. If n is
odd, then diff(D) = 0, and D gives us no information.
If we reverse one edge of D to obtain D , then the only circulation is the triv-
ial subgraph, and diff(D ) = 1. Since D has a vertex with outdegree 2, we learn
that C2¾+1 is 3-choosable. The theorem provides only upper bounds; we do not
learn that C2¾+1 is not 2-choosable.
Section 15.1: Dimension and Polynomials 739

An acyclic digraph contains only one circulation: the trivial subdigraph.

Thus Theorem 15.1.36 implies that G is ¾-choosable if G has an acyclic orien-
tation where every outdegree is less than ¾ . This again proves that ¾-degenerate
graphs are (¾ + 1)-choosable.

• • •
• •
• •
• •
• • • •

Before proving Theorem 15.1.36, we motivate it by describing some appli-

cations. A relatively easy application (Alon–Tarsi [1992]) is that every planar
bipartite graph is 3-choosable (Exercise 32).
An impressive application is the Cycle-plus-triangles Theorem. Consider
a 4-regular graph formed from C3m by adding m pairwise disjoint triangles. Du
and Hsu conjectured in 1986 that such graphs have independence number m, and
in 1987 Erdős conjectured more strongly that they are 3-colorable. Fleischner–
Stiebitz [1992] proved this by using Theorem 15.1.36 to prove the stronger result
that every such graph is 3-choosable. (Later, Sachs [1993] gave a direct combina-
torial proof that such graphs are 3-colorable.)
The analysis of circulations in the Cycle-plus-triangles Theorem is quite
lengthy. Instead we present an easier result that also illustrates the use of the
Alon–Tarsi Theorem. Recall that Cn2 is the graph defined on n vertices around a
circle by making each vertex adjacent to the four nearest vertices. We have a 4-
regular graph. If we can orient it with two edges in and two out at each vertex
such that the numbers of spanning circulations of even and odd size differ, then
we have 3-choosability. The result arose in the context of “total coloring ”.

15.1.38. Example. A total coloring of a graph colors both the vertices and the
edges so that no adjacent or incident objects have the same color. The total chro-
matic number is the number of colors needed; it is the chromatic number of
the total graph, obtained from a graph G by subdividing every edge and then
taking the square (adding edges joining vertices at distance 2 in the subdivision
graph). For a cycle, subdividing merely doubles the length, so total coloring of a
cycle corresponds to proper coloring of the square of a cycle twice as long.
Note that 3 | n is already needed for Cn2 to be 3-colorable; hence it is also
needed for 3-choosability. (Exercise 10.1.29 computes (Cn¾) in general.)

• •
• •

• •
• •
740 Chapter 15: Linear Algebra

15.1.39. Theorem. (Juvan–Mohar–Škrekovski [1998b]) Cn2 is 3-choosable if and

only if 3 | n. Consequently, Cm is 3-total-choosable if and only if 3 | m.
Proof: As noted above, 3 | n is necessary. For the converse, we seek a suitable
orientation of Cn2 with maximum outdegree 2. Orient every edge in the clockwise
direction as shown above; call this orientation D.
Let S be the set of spanning circulations in D. Let S− , S1 , S+ be the subsets
of S in which all vertices have outdegree at most 1, all equal 1, or all at least 1,
respectively. Obviously, S− ∩ S+ = S1 .
Furthermore, S = S− ∪ S+ . A circulation D in S − (S− ∪ S+) has a vertex
with outdegree 2 and a vertex with outdegree 0. If d+D (v) = 0 and w is the next
vertex following v along the circle, then at most one edge of D “crosses the gap”
between v and w. However, when D has at most one edge crossing a gap, also D
has at most one edge crossing the next gap. Hence we cannot reach a gap crossed
by at least two edges. This forbids existence of a vertex with outdegree 2.
Let t+ and t− be the numbers of even spanning circulations and odd spanning
circulations, respectively, so t+ + t− = | S| . To prove diff(D)
= 0, we prove that
|S| ≡ 2 (mod 4) and that t+ and t− are both even. The latter statement is easy. For
a circulation H , the remaining edges in D also form a circulation H . Although
H
= H , the numbers of edges in H and H have the same parity, since D has 2n
edges. Hence t+ and t− are even.
Note that S1 contains only the outside cycle and the set of “inner ” (length 2)
edges. Furthermore, complementation of edge sets matches S− − S1 with S+ − S1 ;
they have the same size. Hence it suffices to show that | S− − S1 | is even.
Other than the empty circulation, each circulation in S− − S1 has exactly one
cycle and has a vertex with outdegree 0. Recall that the adjusted Fibonacci num-
ber F̂ i is the number of 1 , 2-lists with sum i (Example 2.1.2). For a fixed vertex
v, a cycle that omits v corresponds to a 1 , 2-list with sum n − 2, while a cycle that
visits v corresponds to a 1 , 2-list with sum n. In the latter case, omit the one
list having no 2, since the corresponding circulation is in S1 . Replace it with the
empty circulation. Now | S− − S1 | = F̂ n−2 + F̂ n = F̂ n−1 + 2 F̂ n−2 . Since F̂0 = F̂1 = 1,
both are odd, and hence F̂ 2 is even. The parity pattern then repeats, with F̂ r
even if and only if r ≡ 2 (mod 3) (this is generalized in Exercise 2.1.22). Thus the
number of circulations is an odd multiple of 2 if and only if 3 | n.

The key idea here is used also in the proof of the Cycle-plus-triangles Theo-
rem and in other applications of Theorem 15.1.36 to 4-regular graphs: show that
the total number of circulations in the specified orientation is an odd multiple
of 2. Woodall–Prowse [2003] generalized the application; a special case of their
result is that  l(G) = (G) whenever G is a power of a cycle.
To prove the Alon–Tarsi Theorem (Theorem 15.1.36), we view colors as real
numbers. We associate with each vertex vi a variable x i and define a polynomial
that is nonzero precisely when the assigned numbers form a proper coloring.

15.1.40. Definition. Given a graph G with vertex set v1 , . . . , vn , let E (G) =

{(i , j): i < j and vi vj ∈ E(G)}. The graph polynomial pG of G is defined by
pG(x1 , . . . , x n) = ∏(i , j)∈ E (G)(x i − x j ).

Early applications of the graph polynomial include Petersen [1891], Scheim

Section 15.1: Dimension and Polynomials 741

[1974], Li–Li [1981]. Indeed, Petersen introduced graphs in order to study such
polynomials (see L ützen–Sabidussi–Toft [1992] for the history).
The relation of pG to orientations is seen by expanding the product.

15.1.41. Definition. Given a fixed ordering v1 , . . . , vn of the vertices of a graph

G, an edge vi vj in an orientation of G is decreasing if i > j . The parity of an
orientation is the parity of the number of decreasing edges.

15.1.42. Lemma. Let G be a graph with m edges and vertices v1 , . . . , vn , and let
(d1 , . . . , d n) be a list with sum m. Let S be the set of orientations D of G such
that d+D(vi) = d i for all i. In the graph polynomial pG , the coefficient of ∏ x di i
is the number of even minus the number of odd orientations in S.
Proof: The polynomial is homogeneous of degree m, since each factor is homoge-
neous of degree 1. Each contribution to the expansion is formed by selecting one
endpoint of each edge. This corresponds to an orientation by letting the selected
vertex be the source of the edge. The resulting contribution to the expansion is
(−1)t ∏ x di i , where d i is the outdegree of vi in the corresponding orientation and
t is the number of decreasing edges. For a given list d of outdegrees, the even
orientations count +1, and the odd orientations count −1.

In order to relate this coefficient to diff(D) in the statement of Theorem

15.1.36, we establish a bijection from the set of orientations with the same out-
degrees as D to the set of spanning circulations in D.

15.1.43. Lemma. For an orientation D of G with d i = d+D(vi) for each i, the abso-
lute value of the coefficient of ∏ x di i in pG is diff(D).
Proof: Again the vertex ordering v1 , . . . , vn is fixed. Let S be the set of orienta-
tions D of G such that d+D (v) = d+D(v) for all v ∈ V(G). For D ∈ S, let D ⊕ D be
the spanning subdigraph of D whose edges are the edges of D oriented oppositely
in D . Since d+D(v) = d+D (v) for all v, the subdigraph D ⊕ D of D is a circulation.
Reversing the orientation on the edges of any circulation in D does not change
any outdegree, so it yields a member of S. Thus the map taking D to D ⊕ D for
D ∈ S is a bijection from S to the set of circulations contained in D.
For each edge e ∈ D ⊕ D , the reverse edge occurs in D , so exactly one of
them is a decreasing edge. Therefore, D ⊕ D has an even number of edges if and
only if the numbers of decreasing edges in D and D have the same parity. That
is, the number of members of S with the same parity as D is the number of even
circulations in D, and the number with the opposite parity is the number of odd
circulations in D. Hence diff(D) equals the difference between the numbers of
even and odd circulations in S, and Lemma 15.1.42 applies.

Proof of Theorem 15.1.36. The Alon–Tarsi Theorem now follows from the Com-
binatorial Nullstellensatz. We are given a graph G and orientation D. The graph
polynomial pG is homogeneous with degree | E(G)|. If [ ∏ x di i ] pG is nonzero, then
the Combinatorial Nullstellensatz guarantees that pG(s)
= 0 for some s ∈ ∏ Si
when | Si | > d i for each i. By Lemmas 15.1.42–15.1.43, this coefficient is nonzero
precisely when there is an orientation D with outdegrees d1 , . . . , d n such that
diff(D)
= 0. We are given D with these properties.
742 Chapter 15: Linear Algebra

Schauz [2009, 2010] strengthened the Alon–Tarsi Theorem to obtain the

same upper bound on a coloring parameter p called “paint number ” that is al-
ways at least as large as  l . His proof is completely combinatorial, so the Combi-
natorial Nullstellensatz is no longer needed to prove the Alon–Tarsi Theorem.
After any proof, applying the Alon–Tarsi Theorem is a matter of finding a
suitable orientation. This leads to a new parameter that bounds the choosability.

15.1.44.* Definition. An Alon–Tarsi orientation or AT-orientation of a

graph G is an orientation D such that diff(D)
= 0. The Alon–Tarsi number
AT(G) is the least ¾ such that G has an AT-orientation D with +(D) < ¾ .

By the Alon–Tarsi Theorem, Ò l(G) ≤ AT(G). By Schauz’s result, the paint

number is between them. Duraj–Gutowski–Kozik [2016] showed that Òp(K n ,n)
exceeds Ò l(K n ,n) by (log log n). Hence AT(G) ≤ ¾ suffices for Ò l(G) ≤ ¾ but is not
necessary. On the other hand, there are many families where the bound on AT(G)
is sharp for Ò l(G) ≤ ¾ , and hence proving the existence of an AT-orientation with
the maximum outdegree at most ¾ is stronger than proving Ò l(G) ≤ ¾ .
For example, Zhu [2019] proved AT(G) ≤ 5 for every planar graph G (we could
say they are “5-AT-orientable”). In another direction, Cushing–K ierstead [2010]
proved that for any 4-uniform list assignment L of a planar graph G, there is a
matching M in G such that G − M is L-colorable. Grytczuk–Zhu [2020] strength-
ened this in several ways by guaranteeing a matching M such that AT(G − M) ≤ 4.
Thus one matching chosen in advance works for all L.
A modification of the proof yields an AT-orientation of G with outdegree at
most 4 at the tails of edges in M and at most 3 elsewhere, strengthening AT(G) ≤
5. Both Zhu [2019] and Grytczuk–Zhu [2020] use the approach of Thomassen
[1994b] (Theorem 9.3.5): inductively prove a stronger statement, using cases de-
pending on whether the outer cycle of the embedded graph has a chord.

15.1.45.* Theorem. (Grytczuk–Zhu [2020]) Every planar graph G contains a

matching M such that AT(G − M) ≤ 4.
Proof: We prove the following stronger statement. Given an edge e with end-
points v1 and v2 on the unbounded face of G, there exists a matching M in G that
contains e and an AT-orientation D of G − M such that d+D(v1) = d+D(v2) = 0, any
other vertex v on the unbounded face satisfies d+D(v) ≤ 2 − d M (v), and d+D(v) ≤ 3
for every internal vertex v.
A triangle has such a matching and orientation by letting M = {e} and ori-
enting the other two edges in to the endpoints of e. The only circulation is the
one with no edges. This provides a basis for induction on the number of vertices.
Exercise 34 implies that if a graph has an AT-orientation D, then every sub-
graph has an AT-orientation in which each vertex has outdegree at most its outde-
gree in D. Thus we may assume that G is 2-connected and every bounded face is a
triangle. Let C denote the cycle bounding the outer face, with vertices v1 , . . . , vn
in order. The induction step considers two cases.
Case 1: C has a chord vi vj with i < j. Let G 1 be the graph enclosed by the
cycle through v1 , . . . , vi and v j , . . . , vn , and let G 2 be the graph enclosed by the
cycle through vi , . . . , v j . Let e = vi vj . Both G 1 and G 2 are 2-connected, so the
induction hypothesis yields (M1 , D1) for (G 1 , e) and (M2 , D2) for (G 2 , e ). Let M =
Section 15.1: Dimension and Polynomials 743

M1 ∪ (M2 − {e }) and D = D1 ∪ D2 . Note that M is a matching in G and D is an

orientation of G − M.
Since d+D2 (vi) = d+D2 (vj ) = 0, each vertex v ∈ V(C) − {v1 , v2 } inherits d+D(v) ≤
2 − d M (v) from D1 or D2 , and similarly d+D(v) ≤ 3 for interior vertices.
Since d+D2 (vi) = d+D2 (vj ) = 0, every circulation in D is the union of a circu-
lation in D1 and a circulation in D2 . The parity of the circulation is odd if and
only if the parities of the two component circulations differ. Letting e(F) and
o(F) denote the numbers of even and odd circulations in a digraph F , we have
diff(F) = #####e(F) − o(F)#####. By the computation below, D is an AT-orientation.

e(D) − o(D) = e(D1)e(D2) + o(D1)o(D2) − e(D1)o(D2) − o(D1)e(D2)

= (e(D1) − o(D1))(e(D2) − o(D2))
= 0

Case 2: C has no chord. Let G = G − vn ; note that G is 2-connected, with e

on its outer cycle C . Let the neighbors of vn be v1 , u1 , . . . , u¾ , vn along C . All in-
terior vertices of G other than u1 , . . . , u¾ are interior vertices in G . Let (M , D )
be the matching and orientation obtained by applying the induction hypothesis
to (G , e). Let M = M and define a candidate orientation D for G − M by taking
D on G − M and orienting the new edges as vn v1 , vn vn−1 , and ui vn for 1 ≤ i ≤ ¾ .
In D, the only vertices that gain outdegree compared to D are vn and
u1 , . . . , u¾ . Since vn is not covered by M , we have d+D(vn) = 2 ≤ 2 − d M (vn), and
d+D(ui) = 1 + d+D (ui) ≤ 3 − d M (ui) ≤ 3, as needed.
Since d+D (v1) = 0, any circulation in D that is not contained in D uses one
edge of the form ui vn and the edge vn vn−1 . Let Si be the set of circulations in D
using ui and vn . If each Si = ∅, then diff(D) = diff(D )
= 0.
If Si
= ∅, then let Ci be a cycle through ui and vn . Let D ̂ i be the orientation
obtained from D by reversing Ci − vn . Let S ̂i be the set of circulations in D ̂ i ; they
do not visit vn , since D ̂ i does not contain vn . A circulation in Si uses some edges
of Ci (including ui vn and ui vn−1). Deleting those edges and adding the reverse of
the edges of Ci that were not used yields a circulation in S ̂i (it does not visit vn).
Similarly, for any circulation in S ̂i , deleting its edges used from the reverse of Ci
and adding instead the edges of Ci whose reverses were not used yields a circu-
lation in Si . Hence | Si | = | S ̂i |. The parity of the number of edges changes under
this map if and only if Ci has odd length. Hence |e(D ̂ i) − o(D̂ i)| is the difference
between the numbers of even and odd members of Si .
If this difference is 0 for all i, then diff(D) = diff(D )
= 0, and we are finished.
̂ i) − o(D
If |e(D ̂ i)| =

0, then obtain an orientation D ̂ of G − M from D by revers-
ing all edges of the cycle Ci except ui vn . This includes reversing vn vn−1 to become
vn−1 vn . In D ̂ there is no circulation through vn . Indeed, the circulations in D ̂ are
̂ ̂
simply the circulations in Di . We are in the case diff(Di)
= 0, so diff(D) = 0. ̂
In obtaining D ̂ from D, outdegree increases only for ui ; it now has two out-
neighbors on Ci . Since d+D (ui) ≤ 2 − d M (ui), we have d+̂(ui) ≤ 4 − d M (ui). Since
D
ui is now internal, this is acceptable if ui is covered by M . If not, then we add
the edge ui vn to obtain M from M and delete ui vn from D. ̂ Now ui has the needed
̂
outdegree in the orientation of G − M , and because D had no circulations using
the edge ui vn the resulting orientation is still an AT-orientation.
744 Chapter 15: Linear Algebra

EXERCISES 15.1

15.1.1. (−) Prove that the incidence vectors of the clubs in Oddtown (Example 15.1.1) are
linearly independent over the two-element field. (Berlekamp [1969])
15.1.2. (−) Form a digraph D by replacing every edge of the graph Cn with two oppositely
directed edges having the same endpoints. Count the circulations in D.
15.1.3. (♦) Let A1 , . . . , A m be a family of even subsets of [n] with odd-sized intersections.
(a) Prove that m ≤ n, with equality possible when n is odd.
(b) Prove that m ≤ n− 1 when n is even, with equality possible. (Babai–Frankl [1992])

15.1.4. (♦) Prove that Eventown (Example 15.1.1) has at most 2⌊n/2⌋ clubs. That is, [n]
contains at most 2⌊n/2⌋ even sets whose intersections all have even size. (Hint: Prove that
the span U of the set of incidence vectors of the clubs is contained in the subspace of vectors
orthogonal to all of U.) (Berlekamp [1969])
2 2
15.1.5. For even n, prove that there are between 2 n /4/n! and 2 n /n! sets of n odd-sized
subsets of [n] such that the intersection of any two has even size. (Hint: Let n = 2 ¾ . From
a ¾-by- ¾ binary matrix A, form an n-by-n binary matrix ( A+AI¾ A+A I ).) (M. Szegedy)
¾

15.1.6. For 0 ≤ t ≤ (n − 1)/2, construct a family of n − 2t odd-size subsets of [n], whose

pairwise intersections have even size, such that no odd-size subset can be added having
even intersection with all the current members of .
15.1.7. (♦) A town with n people has m sports clubs A1 , . . . , A m and m theater clubs
B1 , . . . , Bm . Prove that if | Ai ∩ Bi | is odd for every i, and | Ai ∩ Bj | is even whenever i > j ,
then m ≤ n, and this is sharp. (Babai–Frankl [1992])
15.1.8. Let A be a 2n-by-2n matrix with entries in {1 , 0 , −1} such that ai , j = 0 if and only
if i = j . Prove that A is nonsingular. Conclude that if w1 , . . . , w2n+1 are real weights such
that any 2n of them can be partitioned into two n-sets with the same total weight , then
all the weights are equal. (Babai–Frankl [1992])
15.1.9. By Theorem 15.1.5, a two-distance set in n has at most (n + 1)(n + 4)/2 points.
Construct a two-distance set of size (n+2 1 ) in n . (Hint: Start with such a set in n+1 .)

15.1.10. Prove that any two-distance set of points in the sphere S n−1 in n has size at most
n(n + 3)/2. (Hint: Follow the proof of Theorem 15.1.5, but confine the resulting indepen-
dent polynomials to a space of dimension n(n + 3)/2.) (Delsarte–Goethals–Seidel [1977])
15.1.11. (+) Improved bound on the size of a two-distance set {v1 , . . . , vm} in n .
(a) In n , the affine hull of vectors {v1 , . . . , vm} is {∑i=1 ci vi : ∑i=1 ci = 0}. Let B be
m m

the m-by-(n + 1) matrix whose ith row is vi plus 1m as column 0. Prove that if the affine
hull is all of n , then the columns of B are linearly independent.
(b) Prove that if the columns of a real m-by-p matrix B are linearly independent , then
B T B is nonsingular.
(c) Prove that adding {1 , x1 , . . . , x n} to the set of polynomials constructed from a two-
distance set in Theorem 15.1.5 yields a linearly independent set. Conclude that the maxi-
mum size of a two-distance set in n is at most ( n+2 2). (Blokhuis [1984])
15.1.12. An equivalence on G is a spanning subgraph whose components are complete.
(a) Prove that pdim (G) is the minimum number of equivalences on G whose union is
G and whose overall intersection is empty.
(b) Prove that pdim (G) ≤  (G) when  (G) > 1, with equality if G is triangle-free.
(c) Use Vizing ’s Theorem to prove pdim (G) ≤ n − 1 when | V(G)| = n ≥ 3.
Exercises for Section 15.1 745

15.1.13. (♦) L-intersecting families. (Alon–Babai)

(a) Let A be an m-by-m matrix of integers. Prove that if some prime power divides ev-
ery off-diagonal entry but no diagonal entry, then A is nonsingular.
(b) Conclude that if the greatest common divisor of the elements of L does not divide
¾ , then L-intersecting families of ¾-subsets of [n] have size at most n.
(c) Let q be a prime power. Let A1 , . . . , A m be subsets of [n] such that | Ai ∩ Aj | is di-
visible by q and | Ai | is not divisible by q whenever i , j ∈ [m]. Prove that m ≤ n. (Comment:
When ¾ = 1, this becomes an instance of the Deza–Frankl–Singhi Theorem.)

15.1.14. Given 1 , 2 ⊆ ([n] ¾ ) , let

( 2) denote the permutation of 2 obtained by applying
the permutation : [n] → [n].
(a) With all permutations of [n] equally likely, prove | 1 ∩ ( 2)| = | 1 | | 2 | /(¾n) .
(b) Suppose that S1 and S2 are disjoint , and suppose that 1 is S1 -intersecting and 2
is S2 -intersecting. Prove | 1 | · | 2 | ≤ (¾n). (M. Szegedy)

15.1.15. Prove that the points of 2 can be colored using asymptotically nn/2 colors so that
points at distance 1 have different colors.
3 3
15.1.16. Use Stirling ’s Formula to approximate ln ( p2p−1)/ ln ( pp−1) for large p.


15.1.17. Choose n , p ∈ with p prime and n > 2p. Let G n ,p be the graph whose vertices
√
are the incidence vectors of (2p − 1)-sets in [n], adjacent when their distance in n is 2p.
Prove that (G n ,p) ≥ (2pn−1 )/( p−n1 ). Improve the lower bound on the chromatic number of
the unit-distance graph in n by choosing p to maximize the lower bound on (G n ,p).

15.1.18. (♦) Color the edges of the complete graph with vertex set ([t3 ]) by making an edge
3

red if its endpoints have one common element and blue otherwise. Conclude R(t , t) > (t3 ) .
3

(Comment: The graph (t − 1)K t−1 gives only R(t , t) > (t − 1)2 .) (Nagy [1972])

15.1.19. Use Snevily’s Theorem (Theorem 15.1.21) to prove the Frankl–Wilson Theorem
(Theorem 15.1.14).

15.1.20. An odd representation of a graph G assigns each vertex a binary -tuple so

that vertices are adjacent if and only if the dot product of their -tuples is odd. Prove that
every n-vertex graph has an odd representation with  = n − 1. (Eaton–Grable [1996])

15.1.21. Cauchy–Davenport in higher dimensions. The Hopf–Stiefel function relative to

a prime p is a function r ◦ s of positive integers r and s, defined to be the smallest integer n
such that (¾n) is divisible by p whenever n − r <  < s. It can be computed recursively from
the base-p representations of r and s. Note that r ◦ s = r + s − 1 when r + s ≤ p + 1.
(a) Prove that if A and B are nonempty subsets of a finite vector space over p , with 
r = | A| and s = | B| , then | A + B| ≥ r ◦ s.
(b) Prove that part (a) is sharp for all r and s. (Hint: In one dimension, part (a) re-
duces to the Cauchy–Davenport Theorem. Prove sharpness by generalizing the sharpness
example for that theorem.) (Eliahou–Kervaire [1998])

 
15.1.22. (♦) Let 2m be the vector space of dimension m over 2 . Given nonempty subsets

A and B of 2m , let A + B = {a + b: a ∈ A , b ∈ B}. Use the Combinatorial Nullstellensatz to
prove that | A|+| B| > 2 n implies | A + B| ≥ 2 n . (Comment: The result was proved inductively
in Fon-Der-Flaass & Alekseyev [2012]. R. Chapman and T. Viteam independently found
proofs (unpublished) using the Combinatorial Nullstellensatz.)

15.1.23. Show that Theorem 15.1.29 is a special case of Theorem 15.1.30. Prove Theorem
15.1.30. (Alon–Nathanson–Rusza [1996])
746 Chapter 15: Linear Algebra

15.1.24. Let A and B be nonempty subsets of p , where p is prime. Prove that the number
of sums x + y such that x ∈ A, y ∈ B, and xy
= 1 is at least min{p , | A| + | B| − 3}. (Hint:
Use Theorem 15.1.30.) (Alon–Nathanson–Rusza [1995])
15.1.25. More from Theorem 15.1.30. (Alon–Nathanson–Rusza [1996]) Let p be a prime.
(a) For c1 , . . . , c¾ ∈ ∪ {0} with sum m + (¾2) , where m ≥ 0, prove
¾ ¾ m
m!
[∏ xci i ](∑ xi ) ∏ (xj − xi) = ¾ ∏ (cj − ci).
i =1 i =1 1 ≤ i< j ≤ ¾ ∏i=1 ci ! 1≤i< j ≤¾
(Hint: Use the Hook-Length Formula (Theorem 4.3.4) or give a direct proof.)

(b) Let A1 , . . . , A¾ be nonempty subsets of p . Let S(A1 , . . . , A¾) be the set of sums of
distinct elements a1 , . . . , a¾ such that ai ∈ Ai for all i. Prove that if | A1 | , . . . , | A¾ | are dis-
¾
tinct and sum to less than p + (¾+2 1 ) , then | S(A1 , . . . , A¾)| > ∑i=1 | Ai | − (¾+2 1 ) . (Hint: Use
part (a) and Theorem 15.1.30.)

(c) Let A1 , . . . , A¾ be nonempty subsets of p , indexed so that | A1 | ≥ · · · ≥ | A¾ | . Let
b1 = | A1 | , and let bi = min{bi−1 − 1 , | Ai |} for 2 ≤ i ≤ ¾ . Prove that if b¾ > 0, then
¾
| S(A1 , . . . , A¾)| ≥ min{p , ∑i=1 bi − (¾+2 1) + 1}.

(d) Conclude that if A is a nonempty subset of p , then the number of sums of s dis-
tinct elements of A is at least min{p , s | A| − s2 + 1}. (Comment: Theorem 15.1.29 is the
special case s = 2.) (Dias da Silva–Hamidoune [1994])
15.1.26. Prove that the minimum number of hyperplanes in n that do not contain the 
origin but together cover all other points of {0 , 1}n is n. (Alon–Füredi [1993])
15.1.27. (♦) For each vertex v in a graph G, specify B(v) ⊆ {1 , . . . , d G(v)}.
(a) Prove that if ∑v∈V(G) | B(v)| < | E(G)| , then G has a nontrivial subgraph H such
that d H (v) ∈
/ B(v) for all v ∈ V(G) (note that degree 0 is allowed, but not at all vertices).
(Shirazi–Verstraëte [2008])
(b) Show that part (a) is sharp by constructing infinitely many examples such that
∑v∈V(G) | B(v)| = | E(G)| and no such subgraph exists.
15.1.28. For odd prime p, let ¾ be an integer with 1 ≤ ¾ < p. Given a1 , . . . , a¾ ∈ p 

and distinct elements b1 , . . . , b¾ ∈ p , prove that for some permutation of [] the values
ai + b (i) are distinct modulo p. (Hint: Use the Vandermonde determinant in an application
of the Combinatorial Nullstellensatz.) (Alon [2000a])
15.1.29. Given a permutation , let d (i , j) = a − b, where i and j occupy positions a and b
in the word form, respectively. For i , j ∈ [] with i < j , specify a forbidden distance  (i , j).

Prove that there is a permutation ∈  such that d (i , j)
=  (i , j) for 1 ≤ i < j ≤  .

Conclude that for a1 , . . . , a ∈ n with 2  ≤ n + 1, there is a permutation ∈  such that 
the elements a (i) + i for 1 ≤ i ≤  are distinct modulo n. (K ézdy–Snevily [2002])
 
15.1.30. A Kakeya set in nq is a set K such that for all y ∈ nq , there exists b ∈ n such 

that b + ay ∈ K for all a ∈ . This means that K contains a line in every direction.

(a) Prove that if K ⊆ nq is a Kakeya set with | K | < (q+ nn−1 ) , then there is a nonzero

polynomial  ∈ q [x1 , . . . , x n] of degree less than q that is 0 at all points of K .
 
(b) Conclude that for all y ∈ nq , there exists b ∈ nq such that  (b + ay) = 0 for all

a ∈ . As a polynomial in a, conclude that  is identically 0, contradicting (a) and proving
that every Kakeya set has size at least (q+ nn−1 ) . (N. Alon and T. Tao; see Dvir [2017+])
15.1.31. Prove that K4 is 3-edge-choosable.
15.1.32. (♦) Hakimi [1965] (Corollary 6.1.5) proved that a graph G has an orientation in
which each vertex has outdegree at most d if and only if every subgraph H has at most
d | V(H)| edges. Conclude that planar bipartite graphs are 3-choosable. (Alon–Tarsi [1992])
Section 15.2: Matrices 747

15.1.33. For a plane graph G, let H be the hypergraph with vertex set V(G) whose edges
are the vertex sets of faces of G. Use the 5-choosability of G (Theorem 9.3.5) and the 3-
choosability of planar bipartite graphs (Exercise 15.1.32) to prove that H is 3-choosable.
(Comment: It is conjectured that H is 2-choosable.) (Ramamurthi [2001])
15.1.34. (♦) Let e be an edge in an AT-orientation D of a graph G. If e lies in no cycle in
D, then D − e is an AT-orientation of G − e.
(a) Prove that if e lies in a cycle C in D, then the orientation D obtained from D by
reversing the edges of C is an AT-orientation.
(b) Prove that D − e or D − e is an AT-orientation of G in which every vertex has
outdegree at most its outdegree in D (here e is the reverse of e).

15.2. Matrices
In this section we consider enumerative aspects of matrices. We use determi-
nants to count spanning trees in graphs, permanents to count perfect matchings
in bipartite graphs, and matrix inversion to study inversion formulas for func-
tions on partially ordered sets.

DETERMIN ANTS AND TREES

In Section 4.1, we used signed involutions to express certain enumeration

problems as determinant computations. Here we consider another enumeration
problem solved by a determinant, but this time the combinatorics is more closely
tied to the algebraic aspects of the determinant.
By Cayley ’s Formula (Theorem 1.3.4), there are nn−2 trees with vertex set
[n]. Viewing them as spanning trees of K n suggests the general problem of count-
ing the spanning trees in a graph G. Example 2.1.13 develops a recurrence for
this number Ì(G), but obtaining Ì(G) from the recurrence generally requires ex-
ponential work in the number of edges.
The Matrix Tree Theorem computes Ì(G) using a determinant. Determi-
nants of n by n matrices can be computed using fewer than n3 operations (for
large n), much faster than the recurrence. We delete loops before the computa-
tion because they appear in no spanning trees. Multiedges cause no difficulties.

15.2.1. Example. The Matrix Tree Theorem (Theorem 15.2.5) tells us to form
the diagonal matrix of vertex degrees, subtract the adjacency matrix, delete a
row and a column, and take the determinant. For the graph K 4 − e, the ver-
tex degrees are 3 , 3 , 2 , 2, so we form the matrix on the left below and take the
determinant of the matrix in the middle. We get the number of spanning trees.
⎛ 3 −1 −1 −1 ⎞
⎛ 3 −1 −1 ⎞
⎜ −1 3 −1 −1 ⎟
⎜
⎜ −1
⎟ → ⎜ −1 2 0 ⎟ → 8
−1 2 0 ⎟ ⎝ −1
⎝ −1 0 2 ⎠
−1 0 2 ⎠

To state and prove the theorem, we need some basic linear algebra.
748 Chapter 15: Linear Algebra

15.2.2. Definition. The cofactor Ai , j of position (i , j) in a square matrix A is

(−1)i+ j times the determinant of the matrix obtained by deleting row i and
column j . The adjugate matrix Adj A has in position (i , j) the cofactor of
position (j , i) in A.

15.2.3. Remark. Properties of the adjugate matrix. In linear algebra we obtain

Adj (AB) = Adj (A) Adj (B) and (Adj A)A = (det A)I.

15.2.4. Lemma. If every row of an n-by-n matrix A has sum 0, then the cofactors
of any row are all equal.
Proof: When every row has sum 0, the columns are dependent, so rank (A) < n. If
rank (A) < n − 1, then all cofactors are 0. Otherwise, rank (A) = n − 1 and det A =
0. The equation AAdj A = 0 now puts every column of Adj A into the nullspace of
A. Every row-sum of A being 0 means that 1n is in the nullspace. Since rank A =
n − 1, every vector in the nullspace is a multiple of 1n . Hence the columns of Adj A
are constant-valued, making the cofactors in each row of A equal.

We also need the Cauchy–Binet Formula: If C is the product of an n-by-

m matrix A and an m-by-n matrix B, then det C = ∑S⊆[m] det AS det BS , where
AS [BS ] is the n-by-n submatrix of A [B] with columns [rows] indexed by S. We
proved this combinatorially in Proposition 4.1.32; also it has a short proof in lin-
ear algebra using block multiplication of cleverly-defined matrices (Exercise 11).

15.2.5. Theorem. (Matrix Tree Theorem) Given a loopless multigraph G with

vertex set v1 , . . . , vn , let ai , j be the number of edges of the form vi vj . Let Q
be the matrix with entry (i , j) being − ai , j when i
= j and d(vi) when i = j .
If Q∗ is the matrix obtained by deleting any row s and column t of Q, then
Ì(G) = (−1)s+t det Q∗ .
Proof: By Lemma 15.2.4, we need only prove this when s = t.
Step 1. If D is an orientation of G, and M is the incidence matrix of D, then
Q = MM T . With edges e1 , . . . , e m , we put mi , j = 1 when vi is the tail of e j , mi , j =
−1 when vi is the head of e j , and mi , j = 0 otherwise. Entry i , j in MM T is the dot
product of rows i and j of M. When i
= j , the product counts −1 for every edge
of G joining the two vertices; when i = j , it counts 1 for every incident edge and
yields the degree.
a b c d e 1 c 2
1 ⎛ −1 1
• •
1 0 0 ⎞ ⎛ 3 −1 0 −2 ⎞
0 −1 −1 0 ⎟ −1 2 −1 0 ⎟
Q=⎜
2⎜ 0
M= ⎜ b a d ⎜ 0 −1 2 −1 ⎟
3 0 0 0 1 −1 ⎟
4 ⎝ 1 −1 0 0 1 ⎠ ⎝ −2 0 −1 3 ⎠
• •
4 e 3

Step 2. If B is an (n − 1)-by-(n − 1) submatrix of M, then det B = 0 if the corre-

sponding n − 1 edges contain a cycle, and det B = ±1 if they form a spanning tree of
G. If the edges corresponding to the columns contain a cycle C, then the columns
sum to the zero vector when weighted with +1 or −1 according as the directed
edge is followed forward or backward when following the cycle. This equation of
dependence yields det B = 0.
Section 15.2: Matrices 749

For the other case, we use induction on n. For n = 1, by definition a 0-by-0

matrix has determinant 1. For n > 1 let T be the spanning tree whose edges are
the columns of B. Since T has at least two leaves, B contains a row corresponding
to a leaf x of T . This row has only one nonzero entry in B. When computing the
determinant by expanding along that row, the only submatrix B given nonzero
weight in the expansion corresponds to the spanning subtree of G − x obtained by
deleting x and its incident edge from T . Since B is an (n − 2)-by-(n − 2) subma-
trix of the incidence matrix for an orientation of G − x, the induction hypothesis
implies that det B is ±1, and multiplying it by ±1 gives the same result for B.
Step 3. Computation of det Q∗ . Let M ∗ be the matrix obtained by deleting
row t of M , so Q∗ = M ∗(M ∗)T . We may assume m ≥ n − 1, else both sides have
determinant 0 and Ì(G) = 0. When we use the Cauchy–Binet Formula for det Q∗ ,
the submatrix AS is an (n − 1)-by-(n − 1) submatrix of M as discussed in Step 2,
and BS = AST . Hence the summation counts 1 = (±1)2 for each set of n − 1 edges
forming a spanning tree and 0 for each other set of n − 1 edges.

Rényi proved the Matrix Tree Theorem combinatorially by using generating

functions to enumerate trees. Tutte extended the theorem to multidigraphs. To
avoid awkwardness, we will use the term “digraph” in the remainder of this sec-
tion even though multiedges are allowed. Tutte ’s result yields the Matrix Tree
Theorem when the digraph is symmetric (with the same multiplicity for edges
xy and yx). Chaiken and K leitman observed that Rényi’s argument also yields
Tutte ’s result and generalized it considerably. First we state Tutte ’s result.

15.2.6. Definition. A branching (or out-tree) is an oriented rooted tree with

all edges directed away from the root. An arborescence is a union of disjoint
branchings. An in-tree is the reversal of an out-tree. For a digraph G, let
Q− = D− − A and Q+ = D+ − A , where D− and D+ are the diagonal matrices
of indegrees and outdegrees, and A i , j is the number of copies of vj vi in E(G).

15.2.7. Theorem. (Directed Matrix Tree Theorem; Tutte [1948]) With Q−

and Q+ defined as above, the number of spanning out-trees [in-trees] rooted
at vi is the value of any cofactor in the ith row of Q− [ith column of Q+].

15.2.8. Example. The digraph below has two out-trees rooted at 1 and two in-
trees rooted at 3. The determinants behave as claimed.
1• •3
⎛ 0 0 0⎞ ⎛ 2 0 0⎞
Q− = ⎜ −1 1 0⎟ Q+ = ⎜ −1 1 0⎟
⎝ −1 −1 2⎠ ⎝ −1 −1 0⎠
•
2
Since the earlier theorems hold for multigraphs, it is natural to generalize
using edge weights; non-edges are represented by weight 0. To generalize The-
orem 15.2.7, we let the weight of a tree or arborescence in a digraph D be the
product of its edge weights. For further generality and a combinatorial proof, we
use variables to encode the outdegree of each vertex. This leads to a determinan-
tal formula for the sum of the weights of arborescences with specified roots.
750 Chapter 15: Linear Algebra

15.2.9. Theorem.(Matrix Arborescence Theorem; Chaiken–K leitman [1978])

Given real weights ai , j , variables x1 , . . . , x n , and an arborescence A, let w A =
∏vj vi ∈ E(A) ai , j x j . For S ⊆ [n], let T(S) be the set of all arborescences on
{v1 , . . . , vn} whose roots are {vi : i ∈ S}. Define a matrix Q by qi , j = − ai , j x j
for i
= j and qi ,i = ∑ j
=i ai , j x j . If QS denotes the submatrix of Q obtained by
omitting all rows and columns indexed by S, then ∑ A∈ T(S) w A = det QS .

Proof: For S = ∅, the sum is empty and det Q∅ = det Q = 0, so the claim holds.
Let ¾ = | S| and V = {v1 , . . . , vn}. For nonempty S, we use induction on n − ¾ . If
¾ = n, then there is one arborescence (no edges) in the sum, with weight 1 (empty
product), and the 0-by-0 determinant by convention also equals 1.
Now suppose ¾ < n. Both expressions are polynomials in {x j } in which each
nonzero term has total degree n − ¾ ; we write det QS = ºS(x1 , . . . , x n) to empha-
size this. With total degree n − ¾ , each term has degree 0 in at least ¾ variables.
Let {x i : i ∈ S} be “root variables” and the others “non-root variables”. To com-
plete the proof, we prove (1) each term in both ∑ w A and ºS has degree 0 in some
non-root variable, and (2) the terms in which a particular non-root variable has
degree 0 agree in ∑ w A and ºS . The second claim uses the induction hypothesis.
Proof of Claim 1. In w A , the degree of a variable x j is the outdegree of vj in
A. When ¾ < n, there is a non-root leaf, and it has outdegree 0.
Now consider ºS . Deleting the rows and columns of Q indexed by S does not
remove the root variables; they remain in the diagonal terms. Setting the root
variables to 0 makes each row of QS sum to 0, so det QS and ºS have value 0 re-
gardless of the values of the non-root variables when the root variables are set to
0. Since a polynomial that is 0 at all points has every coefficient 0, each term in
ºS has a root variable with positive degree. Since each term has degree 0 in at
least ¾ variables, it has degree 0 in some non-root variable.
Proof of Claim 2. Consider terms having degree 0 in a fixed x t with t ∈ / S.
In ∑ w A , these arise from arborescences where vt is a leaf. Such an arborescence
A arises from an arborescence A on n − 1 vertices with roots in S by adding some
edge vj vt , yielding w A = w A at , j x j . Let T be the set of arborescences on V − {vt }
with root set S. Since j is arbitrary and we can start with any arborescence in
T , the sum of the terms omitting x t equals (∑ A ∈ T w A )(∑j
=t at , j x j ).
Now consider the terms of ºS that omit x t ; setting x t = 0 in ºS yields their
sum. The column for vt in QS becomes 0 except for the diagonal term M , which
is ∑j
=t at , j x j . Terms involving x t disappear from other entries in the matrix.
Expanding the determinant along column t yields M times the determinant of
the matrix Q S obtained from QS by deleting the tth row and column and setting
x t = 0. This is just the determinant defined for the smaller problem with ver-
tex set V − {vt } and the same root set S. By the induction hypothesis, det Q S =
∑ A ∈ T w A . Thus the terms omitting x t are the same in ∑ w A and in ºS .
0
• 0
• vj
• S• • •vt ∑j
=t at , j x j QS at
A • 0 xt = 0
• 0
Section 15.2: Matrices 751

15.2.10. Example. To obtain the Directed Matrix Tree Theorem from the Ma-
trix Arborescence Theorem, let ai , j be the number of edges from vj to vi , set each
variable to 1, and let S consist of one vertex.
Keeping the variables x j as indeterminates yields a generating function that
enumerates the arborescences rooted at S by their outdegrees. For the digraph
in Example 15.2.8, the matrix Q appears below. If S = {1}, then the determinant
is º (S) = x12 + x1 x2 ; one branching consists of two edges from v1 and one is the
path ⟨v1 , v2 , v3 ⟩. If S = {1 , 2}, then the determinant is º (S) = x1 + x 2 , and the
choices for the arborescence are the edge v1 v3 or the edge v2 v3 .

⎛ 0 0 ⎞
0
⎜ − x1 x1 ⎟
0
⎝ − x1 − x2 x1 + x 2 ⎠

Spanning in-trees in digraphs are closely related to Eulerian circuits. After

distinguishing a vertex v, we will obtain a spanning in-tree to v from each Eule-
rian circuit by a method that generates each in-tree the same number of times.
This enables us to count the circuits (as cyclic lists of edges).

15.2.11. Example. In the digraph below, the edges labeled in order from 1 form
an Eulerian circuit C starting along e from v. Each bold edge is the last edge in
C that departs from its tail. The bold edges form an in-tree to v.
11
• • 4
12 10 •
13
14
5 3
v• •
15
e 6 •
7
9 1 2
• •
8

15.2.12. Lemma. For any Eulerian circuit that begins from v along edge e in a
digraph G, the set T of edges along which vertices other than v are exited for
the last time is the edge set of an in-tree to v.
Proof: Each vertex other than v is left for the last time only once, so | E(T)| =
|V(G)| − 1. Each vertex other than v has outdegree 1 in T , and v has outdegree
0. A vertex other than v cannot be left for the last time until it is entered for the
last time, so T contains no cycle. These properties make T an in-tree to v.

We can reverse the process of obtaining an in-tree from an Eulerian circuit.

15.2.13. Algorithm. (Eulerian circuit in directed graph)

Input: An Eulerian digraph G and a spanning in-tree T rooted at v.
Step 1: For each u ∈ V(G), specify an ordering of the edges leaving u, such that
for u
= v the edge leaving u in T comes last.
Step 2: Starting at v, construct an Eulerian circuit by always leaving the cur-
rent vertex u along the next unused edge in the ordering given at u.
752 Chapter 15: Linear Algebra

15.2.14. Lemma. Algorithm 15.2.13 produces an Eulerian circuit.

Proof: Given the in-tree T to v, Algorithm 15.2.13 constructs a trail. It suffices
to show that the trail ends only at v and only after using all edges.
When we enter a vertex u other than v, the edge leaving u in T has not yet
been used, since d+(u) = d−(u). Thus there is still a way out of u, and hence the
trail can end only at v.
We end only after using all edges leaving v. Since d−(v) = d+(v), we then also
have used all edges entering v, including those in T . Since an edge of T is not
used until all other edges leaving its tail are used, and T contains a path from
each vertex to v, we cannot end at v until all edges are used at all vertices.

An Eulerian circuit is given by a cyclic ordering of edges; the starting edge

does not matter. Hence we may require that Algorithm 15.2.13 always uses a fixed
edge e as the first edge leaving v. This associates with each in-tree the same num-
ber of circuits. The authors’ initials from two relevant papers form an acronym by
which the theorem is known; the earlier paper solved the case where each d i = 2.

15.2.15. Theorem. (BEST Theorem; van Aardenne-Ehrenfest & de Bruijn

[1951], Tutte–Smith [1941]) An Eulerian digraph with d i = d+(vi) = d−(vi)
has exactly c ∏i(d i − 1)! Eulerian circuits, where c counts the in-trees to or
out-trees from any vertex.
Proof: Since we always start from v and along edge e, the flexibility of choosing
the rest of the orderings allows Algorithm 15.2.13 to generate ∏u∈V(G)(d+(u) − 1)!
different Eulerian circuits from each in-tree to v. The last out-edge is fixed by the
tree for vertices other than v (after starting along e), so the circuits generated
from different trees are different. From the same tree, we reach a difference in the
pairs of consecutive edges as soon as we reach a difference in the exit orderings.
Hence we have built c ∏i(d i − 1)! distinct Eulerian circuits, and Lemma
15.2.12 ensures that all Eulerian circuits have been found.

Note in this statement that c is independent of the choice of v. If G is Eule-

rian, then Q+ = Q− , and the row-sums and column-sums are all 0. Now Lemma
15.2.4 implies that the cofactors are all equal.

CYCLE SPACE AND BOND SPACE

The Matrix Tree Theorem uses the relationship between cycles and the inci-
dence matrix of an oriented graph. In Section 11.3, we used the incidence matrix
to represent cycle matroids over 2 . The incidence vectors of cycles formed the
null space of the matrix.

15.2.16. Definition. In a graph G with edges e1 , . . . , e m , the incidence vector

for a set F ⊆ E(G) has coordinates ai = 1 when e i ∈ F and ai = 0 when e i ∈
/ F.
The cycle space of G is the set C(G) of incidence vectors of even subgraphs
(those with all vertex degrees even). The bond space of G is the set B(G) of
incidence vectors of edge cuts.
Section 15.2: Matrices 753

We view members of C(G) as even subgraphs or as their incidence vectors; we

view members of B(G) as edge cuts or their incidence vectors.

15.2.17. Theorem. The cycle space and bond space of a connected graph G with
n vertices and m edges are binary vector spaces with dimensions m − n + 1
and n − 1, respectively.
Proof: We have vector spaces over 2 because the subsets are closed under binary
addition: the symmetric difference of two even subgraphs is an even subgraph,
and the symmetric difference of two edge cuts is an edge cut (Exercise 12).
Choose a spanning tree T . Each edge of E(G)− E(T) forms a unique cycle with
edges in T ; these are the fundamental cycles relative to T . These cycles are
linearly independent in C(G), since each has a nonzero coordinate outside E(T)
that is zero in all other incidence vectors in this set. Hence dim C(G) ≥ m − n + 1.
Choose n − 1 vertices in G. Let v1 , . . . , vn−1 be the incidence vectors for the
corresponding edge cuts. Let ∑ ci vi = 0 be an equation of dependence. Since each
coefficient is 0 or 1, we are just summing the vectors for a set S of vertices. The
resulting coordinate for an edge is 0 if and only if the edge has an even number
of endpoints in S. Edges in [S, S] appear once in the sum, and such coordinates
remain nonzero. Since G is connected, [S, S]
= ∅ unless S or S is empty. Hence
there is no nontrivial equation of dependence, the vectors are linearly indepen-
dent, and dim B(G) ≥ n − 1. We have also shown that [S, S] = ∑i∈S vi for every
edge cut [S, S], so dim B(G) = n − 1.
The n − 1 vectors v1 , . . . , vn−1 are rows in the incidence matrix M(G). By def-
inition, every even subgraph has an even number of edges at each vertex, so the
vectors in C(G) are those satisfying M(G)x = 0. Thus C(G) is the nullspace of
M(G), and we have shown that the matrix has rank n − 1. The famous Rank–
Nullity Theorem of linear algebra states that the dimension of the nullspace
is the number of columns minus the rank of the matrix (over any field), so
dim C(G) = m − n + 1. (Alternatively, one can show explicitly that the funda-
mental cycles span the cycle space.)

15.2.18. Definition. A 2-basis for a graph G is a basis for C(G) such that each
coordinate is nonzero in at most two members of the basis.

15.2.19. Theorem. (MacLane) A graph is planar if and only if it has a 2-basis.

Proof: (Tutte) The facial walks of the bounded faces in a planar embedding form
a 2-basis: each cycle is the sum of the bounding walks for the faces inside it.
To show that a nonplanar graph has no 2-basis, it suffices by Kuratowski’s
Theorem to prove (1) if G has a 2-basis and contains a subdivision of H , then H
has a 2-basis, and (2) neither K 5 nor K 3 ,3 has a 2-basis.
For e ∈ E(G), we obtain a 2-basis for G − e from a 2-basis for G. If e is a cut-
edge, then C(G − e) = C(G). Otherwise, dim C(G − e) = dim C(G) − 1. If e is
in one member of the 2-basis for G, delete that member. If e is in two members,
replace them by their sum. This leaves dim C(G) − 1 vectors, all 0 in coordinate e.
Deleting coordinate e yields independent elements of C(G − e) (dependence would
make the 2-basis for C(G) dependent).
754 Chapter 15: Linear Algebra

Also, a subdivision of a graph H has a 2-basis if and only if H has a 2-basis,

because the two edges incident to a vertex of degree 2 in a graph belong to pre-
cisely the same even subgraphs.
To show that K 3 ,3 and K 5 have no 2-basis, consider a 2-basis v1 , . . . , v¾ for a
connected n-vertex graph H with girth ½ , written as binary m-tuples with m =
| E(G)|. Let v0 = ∑¾i=1 vi . As a nonzero linear combination of basis elements, v0 is
a nonzero element of C(G). Recall that ¾ = m − n + 1.
Let w be the total number of nonzero entries in v0 , . . . , v¾ . Since even graphs
decompose into cycles, each nonzero vector has at least ½ such entries. Hence w ≥
½(¾ + 1). Each coordinate contributes at most two such entries, by the definition
of 2-basis, so w ≤ 2m.
Hence ½(m − n + 2) ≤ 2m. Solving for m yields the same bound m ≤ (n − 2) −2
that holds for planar graphs by Euler ’s Formula (Exercise 9.1.45). Since both K 5
and K 3 ,3 have too many edges to satisfy the bound, each cannot have a 2-basis.

The argument for Theorem 15.2.19 is analogous to that for Whitney ’s char-
acterization of planarity by matroid duality (Theorem 11.3.35). Also Theorem
15.2.19 can be used to prove Theorem 11.3.35; see Exercise 17.

PERMANENTS AND PL AN AR GRAPHS

When we omit the signs on the terms in the permutation expansion of a de-
terminant, the computation still has meaning.

15.2.20. Definition. Let A be a square matrix of order n, with entry ai , j in posi-

n
tion (i , j). The permanent of A, denoted by per A, is ∑  ∏i=1 ai , (i) , where
the sum runs over all permutations of [n].

Terms in the permanent are products of one entry from each row and column.
For the biadjacency matrix of a bipartite multigraph G, the permanent counts
the perfect matchings. Using row-reduction operations, determinants are easy
to compute, but no fast algorithm is known to compute the permanent. Counting
the perfect matchings in a bipartite graph belongs to the class of difficult count-
ing problems called #P (Valiant [1979]; see also Jerrum–Sinclair–Vigoda [2001]).
In contrast, determining whether there is at least one such matching is easy.
Computing per A from the definition takes superexponential time in n, as
does computing det A by the permutation expansion. Elementary row operations
can compute determinants quickly, but these transformations do not preserve the
permanent. When G is planar, however, the permanent of the biadjacency matrix
A can be found by computing the determinant of a matrix obtained by negating
some entries of A. At first the result is surprising, but the main idea is simple.
Nonzero terms in computing per A correspond to 1-factors in G. When we
negate some entries to form A and compute det A , the nonzero terms in the ex-
pansion still correspond to the 1-factors in G. The task is to cleverly negate some
entries so that all nonzero terms when computing det A have the same sign. This
would yield per A = |det A | . The proof is valid for multigraphs, but we phrase it
just for graphs. The method was presented in Vazirani–Yannakakis [1989].
Section 15.2: Matrices 755

15.2.21. Theorem. (Permanent-Determinant Method) Let M be a perfect

matching in an X , Y -bigraph G with biadjacency matrix A. A matrix A
obtained from A by negating some entries is coherent if for every M-
alternating cycle C, the number of edges of C whose positions are negative in
A has opposite parity from ¾ , where 2 ¾ is the length of C. If A is a coherent
signing of A, then per A = |det A |.
Proof: Index X and Y so that the edges of M are {x i yi : 1 ≤ i ≤ n}, where X =
{x1 , . . . , x n} and Y = {y1 , . . . , yn}; this permutes rows and columns but does not
change the permanent or the magnitude of the determinant.
Each perfect matching in G corresponds to a permutation matrix and makes
one contribution to per A or det A . Let M be a perfect matching in G. The en-
tries in A for M are {a i , (i)}ni=1 for a particular permutation of [n]. As discussed
above, it suffices to show that ∏i=1 a i ,i = sgn( ) ∏i=1 a i , (i).
n n

The edges in the symmetric difference M M form pairwise disjoint M-

alternating cycles. Let C be one such cycle, with length 2  . We may assume
an indexing so that the edges of M in C correspond to the first  positions along
the diagonal. Shared edges contribute the same diagonal elements as factors in
the product. Since we can make the same argument for each cycle, it suffices to
show that ∏i=1 a i ,i = sgn( ) ∏i=1 a i , (i) , where is the restriction of to [].
The edges of M on C correspond to off-diagonal positions in the upper-left
-by-  submatrix of A . The restriction of to [] is a -cycle. The sign of is
the product of the signs on the cycles. The sign of a -cycle is (−1) −1 , because it
takes  − 1 transpositions (or column exchanges) to bring the permutation to the
identity (or the entries to the diagonal).
On the other hand, since A is a coherent signing, the number of elements
in {a i ,i}i=1 ∪ {a i , (i)}i=1 that are negative has the same parity as  − 1. That is,
∏i=1 a i ,i ∏i=1 a i ,(i) = (−1) −1 . Since each entry is its own multiplicative inverse,
the desired equality holds.

The most famous use of the method is the computation by K asteleyn [1961] of
the number of perfect matchings in Pn Pn . K asteleyn [1963, 1967] extended the
method to all planar bipartite graphs (see also Percus [1969]). We need to show
that the biadjacency matrix of any planar bipartite graph has a coherent signing.
The argument is based on a special orientation that exists for every planar graph.

15.2.22. Lemma. (K asteleyn [1963]) Every planar graph G has an orientation

such that for every perfect matching M , every M-alternating cycle has an
odd number of edges in each direction (forward or backward).
Proof: Consider an embedding of G. Using induction on | E(G)| , we first build an
orientation such that, when traveling the boundary of any bounded face in the
clockwise direction, we traverse an odd number of edges in the forward sense. If
G is a tree, then there is nothing to prove.
Otherwise, some edge e along the unbounded face also lies along a bounded
face F . Let G be the orientation of G − e guaranteed by the induction hypothesis.
Precisely one choice for the orientation of e yields an odd number of forward edges
when F is traversed clockwise. This extends the desired orientation to G.
756 Chapter 15: Linear Algebra

Let C be an M-alternating cycle, and let t be the number of forward edges

in a clockwise traversal of C. For each face F interior to C, let q(F) denote the
number of forward edges in a clockwise traversal of F . Since each term in ∑ q(F)
is odd (summing over faces inside C), the parity of the sum is the same as the
number of faces inside. Also, the sum equals t plus the number of edges inside,
since each inside edge is forward for the clockwise traversal of exactly one of the
faces it bounds. Hence it suffices to show that the numbers of edges and faces
inside C have opposite parity.
Let H be the subgraph consisting of C and the vertices and edges inside C.
By Euler ’s Formula, n − m + p = 1, where n , m , p equal | V(H)| , | E(H)| , and the
number of faces inside C. Since C is M-alternating, the mates of all vertices of C
lie in C, and M pairs the points inside C. Hence both n and the number of edges
of H not inside C are even. Therefore, by Euler ’s Formula the number of edges
inside C has opposite parity from the number of faces inside.

Given an embedding, which can be found quickly, the orientation can be

found quickly. It remains to show that this orientation yields the desired signing
of A for application of Theorem 15.2.21.
For an oriented graph G, the adjacency matrix B has bi , j = 1 if vi → vj ,
bi , j = −1 if vj → vi , and otherwise bi , j = 0; note that B is skew-symmetric. If G
is an X , Y-bigraph, then the X , Y -portion of B (rows for X and columns for Y) is
a signing of the biadjacency matrix of G.
15.2.23. Theorem. (K asteleyn [1963]) If G is a planar X , Y -bigraph with biad-
jacency matrix A, then per A = |det A | , where A is the signing of A given
by the orientation from Lemma 15.2.22.
Proof: (Vazirani–Yannakakis [1989]) We show that A is a coherent signing of
A for some (actually, any) perfect matching M. By Lemma 15.2.22, every M-
alternating cycle C traverses an odd number of edges of the orientation forward
and an odd number backward. Along C, we alternate X-to-Y edges and Y-to- X
edges. Only those from X to Y affect A .
Let 2 ¾ be the length of C and view C in a fixed direction. The edges of C are
of four types; we introduce r and s to count these.
X to Y Y to X
backward r ¾−s
forward ¾−r s
The number of edges of C giving negative entries in A is r + s, since A records
the orientation when an edge is viewed from X to Y . The number of edges along
C traversed forward is ¾ − r + s. By Lemma 15.2.22, this number is odd. Hence
r + s has opposite parity from ¾ . Since this holds for every M-alternating cycle,
A is coherent, and Theorem 15.2.21 applies.

The permanent-determinant method of Theorem 15.2.21 is a special case of

the “Hafnian-Pfaffian method”, valid also for nonbipartite graphs. Indeed, the
original proof of Tutte ’s 1-Factor Theorem used these ideas. The method is de-
scribed in Lovász–Plummer [1986, pp. 315–329] and more generally in Kuperberg
[1994, 1998]. The combinatorial aspects are outlined in Exercises 25–26.
Section 15.2: Matrices 757

M ÖBIUS INVERSION (optional)

Many inversion formulas can be explained using functions on intervals in a

poset. These can be studied via linear algebra when the poset is encoded in a ma-
trix. We obtain a common generalization of Inclusion-Exclusion and many other
inversion formulas. Given a function ½ expressed in terms of another function ,
an inversion formula computes in terms of  . We recall classical examples.

15.2.24. Example. Summation and difference. When (n) = ∑i=1 (i) for n ∈ ,
n

we obtain from  by (1) = (1) and (n) = (n) − (n − 1) for n > 1. This is a
discrete version of the Fundamental Theorem of Calculus: just as differentiation
is the inverse of integration, so differencing is the inverse of summation.

15.2.25. Example. Inclusion-Exclusion. Let {A1 , . . . , An} be subsets of a uni-

verse U. In Section 4.1, we considered two functions on 2[n] . For S ⊆ [n], let
(S) = #####⋂i∈S Ai ##### , and let (S) be the number of elements x ∈ U such that x ∈ Ai
if and only if i ∈ S. An element is counted by (T) if it lies in all sets indexed by
T and perhaps in other sets. Thus (T) = ∑S⊇ T (S). The Inclusion-Exclusion
Formula (Theorem 4.1.3) states that (T) = ∑S⊇ T (−1)| S|−| T | (S).

15.2.26. Example. Möbius inversion in number theory. Let (n) be the number
of -ary sequences with period n. Every n-tuple generates a periodic sequence,
but (n)
=  n , since we must exclude those whose periods divide n. Letting (n)
be the number of -ary n-tuples, we have (n) =  n and (n) = ∑d|n (d). The
classical Möbius Inversion Formula of number theory inverts such summations:
(n) = ∑d|n ( dn )(d), (∗)

where (j) = (−1)¾ when j is the product of  distinct primes, and (j) = 0 when
j is divisible by the square of a prime, and (1) = (−1)0 = 1.

We can prove (∗) by substituting for (d) its definition in terms of and check-
ing that the resulting weighted sum of values of gives weight 1 to (n) and
weight 0 to all other values of . The Inclusion-Exclusion Formula can also be
proved this way, but we want to prove all such inversion formulas at once.
The functions in these examples are defined on posets: a chain, 2[n] , or a divi-
sor lattice. In general we want to invert (x) = ∑ y≤ x (y), where “ ≤ ” is the order
relation in a poset. We need functions on the intervals of a poset.

15.2.27. Definition. When x ≤ y in a poset, the interval [x , y] is { : x ≤  ≤ y}.

A poset P is locally finite if every interval is finite. Let Int (P) denote the set
of intervals in P. An incidence function on P is a function : Int (P) → .
Write (x , y) for ([x , y]), defined only when x ≤ y. The incidence algebra
A(P) (over ) is the set of incidence functions on P , endowed with operations
of sum +  , scalar multiple c , and convolution (or product)  :
( + )(x , y) = (x , y) + (x , y)
(c )(x , y) = c · (x , y)
( )(x , y) = ∑ (x , )( , y).
x≤ ≤ y
758 Chapter 15: Linear Algebra

Restricting to locally finite posets makes each convolution a finite sum. For
convenience, we further restrict to finite posets. The set of functions on a set
S forms a vector space with S as a canonical basis. Each function is then a lin-
ear combination of basis elements. Scalar multiplication and vector addition are
coordinate-wise, as in Definition 15.2.27. A vector space becomes an algebra when
endowed with a suitable multiplication rule for vectors; here we use convolution.

15.2.28. Remark. Incidence algebra as matrix multiplication. Given an ordering

x1 , . . . , x n of P such that x i <P x j implies i < j , we interpret A(P) another way.
Record an incidence function in a matrix A by putting ai , j = º (x i , x j) if x i ≤ P x j
and ai , j = 0 otherwise. Under the chosen ordering, A is upper triangular.
The product of such matrices A and B for incidence functions º and is also
upper triangular and is computed by convolutions: ci¾ = ∑j ai j bj¾ . If x i and x ¾
are incomparable, then there is no x j between them, and the convolution is 0.
For x i ≤ x¾ , the computation yields ( )(x i , x¾). Hence A(P) is isomorphic to the
algebra of upper triangular matrices such that ai , j = 0 when x i
≤ x j .

Multiplication in A(P) is not commutative. Nevertheless, A(P) has a two-

sided multiplicative identity corresponding to the identity matrix. This func-
tion, called  or 1, is defined by 1(x , y) =  x ,y (1 if x = y and 0 if x < y). Next we
seek multiplicative inverses in the incidence algebra.

15.2.29. Lemma. An incidence function  has an inverse if and only if  (x , x)

=
0 for all x ∈ P. If  has an inverse, then  −1 is unique and two-sided, and its
value on [x , y] depends only on the values of  on subintervals of [x , y].
Proof: A left inverse is a function such that  = 1. We show first that the
existence of a left inverse requires  (x , x)
= 0, and that in this case the equations
required by being a left inverse have a unique solution.
Convolution on [x , x] yields (x , x) (x , x) = 1, and thus  (x , x)
= 0 and
(x , x) =  (x , x)−1 . If  (x , x) is nonzero for each x, then the value of (x , y) for
x < y is determined inductively by its values on subintervals. Since  = 1 re-
quires ∑ x≤ ≤ y (x , ) ( , y) = 0 for x < y, we must have
(x , y) = − (y , y)−1 ∑ x≤ < y (x , ) ( , y).
This uniquely determines the left inverse of  .
Given that  = 1, one can verify inductively that ( )(x , y) =  x ,y , and thus
a left inverse is also a right inverse (Exercise 32). Arguments as above show
that the right inverse also is unique.

Since A(P) is isomorphic to an algebra of upper triangular matrices under

matrix multiplication, Lemma 15.2.29 also follows because the invertible trian-
gular matrices are those with nonzero determinant. Triangular matrices have
nonzero determinant precisely when the diagonal has no zero.
Various incidence functions measure structural aspects of P.

15.2.30. Definition. (Fundamental incidence functions).

1 x=y
(a) 1(x , y) = { ; the identity function.
0 x<y
Section 15.2: Matrices 759

1 x=y
(b) (x , y) = { ; the zeta function.
1 x<y
0 x=y
(c) (x , y) = { ; the “strictly less-than” function;  = − 1.
1 x<y
(d) (x , y) = { 1 x≺y ; the cover function.
0 otherwise

We can count structures of interest in a poset using these incidence functions

and an understanding of the meaning of multiplication. A weak x , y-chain of
length  is a list x0 , . . . , x ¾ such that x = x0 ≤ · · · ≤ x¾ = y; the chain is a strict
chain if x0 , . . . , x ¾ are distinct.

¾
15.2.31. Lemma. Multiplication in A(P) is associative, and ∏i=1 i is specified

by ∑ x= 0 ≤···≤ ¾ = y ∏i=1 i(i−1 , i). In particular,
(a)  (x , y) is the number of weak x , y-chains of length  ,
with 2(x , y) = |[x , y]| .
(b)  (x , y) is the number of strict x , y-chains of length  .
(c) (1 − )−1 (x , y) is the total number of strict x , y-chains.
(d)  (x , y) is the number of maximal x , y-chains of length  .
(e) (1 − )−1 (x , y) is the total number of maximal x , y-chains.

Proof: Any order for computing the product yields ∏i=1 i as an iterated sum-
mation. The distributive law pulls the summations to the front, expressing the
formula as one sum. The possible nonzero terms correspond to successive related
elements (not necessarily distinct) forming chains from the bottom to the top of
the interval. The sum is over all such weak chains of length  .
This yields (a)–(e). If each factor in the term for a weak chain is 1, accom-
plished by using , then the sum counts the weak chains. To forbid repetitions,
use  instead. Maximal chains are those where no additional element can be in-
serted; successive elements satisfy the cover relation, and we use .
To count all strict chains, form the series 1 +  + 2 + · · ·. The function 1
(that is, 0) counts chains of length 0; there are none if x
= y. A geometric series
of incidence functions behaves like a geometric series of numbers (Exercise 32).
We need 1 −  to be nonzero on every interval [x , x], which holds by definition.
Counting all maximal x , y-chains using ∑≥0  is analogous to this.

To solve the inversion problem, we need the inverse of .

15.2.32. Definition. The Möbius function of a poset P , denoted by  , is the

incidence function defined by

1 if x = y
(x , y) = { −
∑ x≤ < y (x , ) if x < y.

Definition 15.2.32 is just the requirement of Lemma 15.2.29 for an incidence

function to be the inverse of . The next theorem extends to posets in which every
down-set generated by one element is finite.
760 Chapter 15: Linear Algebra

15.2.33. Theorem. (Möbius Inversion Formula) Let º and be complex-

valued functions defined on a finite poset P. If  and satisfy (x) =
∑ y≤ x  (y) for all x ∈ P , then  (x) = ∑ y≤ x (y)(y , x) for all x ∈ P.
Proof 1: (Computation) Expand the right side of the desired formula, using the
recursive definition of (x , y) and the expression (y) = ∑ ≤ y  (). The proof is
completed by showing that this counts  (x) with weight 1 and other values of 
with weight 0 (Exercise 33).
Proof 2: (Incidence functions) Add a unique minimal element 0̂ to P , yielding
P . Define  (0̂ , y) =  (y), (0̂ , y) = (y), and  (x , y) = (x , y) = 0 for x
= 0̂
(also  (0̂ , 0̂) = (0̂ , 0̂) = 0). The statement (x) = ∑ y≤ x  (y) becomes (0̂ , x) =
∑0̂≤ y≤ x  (0̂ , y) · 1. This combines with the zero values on other intervals to yield
=   in the incidence algebra. Hence it must hold that  = −1 =  , and
the resulting convolution for the value of  (0̂ , x) is the desired formula, because
 P restricts to  P on Int(P).
Proof 3: (Matrix multiplication – no incidence functions) After indexing P by a
linear extension, the incidence algebra is an algebra of upper triangular matrices
under multiplication. View  and as row vectors, with a value for each element.
The statement (x) = ∑ y≤ x  (y)(y , x) is the statement that is obtained when
 is multiplied on the right by the matrix for  . Since the matrix for  is its
inverse,  can be retrieved by multiplying on the right by the matrix for .

Proofs 2 and 3 avoid detailed manipulation and make clear that the validity
of (x) = ∑ y≤ x  (y) for all x is not only sufficient but also necessary for  (x) =
∑ y≤ x (y)(y , x) to hold for all x.
Viewing  , as column vectors instead of row vectors or convolving with 
and  on the left instead of the right yields a dual form of the Möbius Inversion
Formula, which reduces to the usual Inclusion-Exclusion Formula when P = 2 n .

15.2.34. Theorem. (Dual Möbius Inversion Formula). Let  , be complex-

valued functions on a finite poset P. The functions  and satisfy (x) =
∑ y≥ x  (y) for all x ∈ P if and only if  (x) = ∑ y≥ x (x , y) (y) for all x ∈ P.

To apply the inversion formula, we need the Möbius function.


15.2.35. Example. For the chain formed by under the usual ordering, Defini-
tion 15.2.32 yields (j , j) = 1, (j , j + 1) = −1, and (j , ) = 0 if  > j + 1. Hence
Theorem 15.2.33 explains why differencing is inverse to summation.

To compute Möbius functions for posets such as 2[n] and the divisor poset, we
use a combining formula for Möbius functions on products.

15.2.36. Theorem. (Product Formula) The Möbius function for the product of
locally finite posets P and Q is given by
 P × Q((x , y) , (x , y )) =  P (x , x )  Q(y , y ).
Section 15.2: Matrices 761

Proof: Let I = [(x , y) ,(x , y )], and let IP = [x , x ] and IQ = [y , y ]. We use in-
duction on | I |. When | I | = 1, both sides equal 1. For | I | > 1, the definition of P × Q
yields ∑(x ,y )∈ I P × Q((x , y) , (x , y )) = 0. Using I = IP × IQ and the definitions
of P and Q , we also compute

∑ P (x , x ) Q(y , y ) = ∑ P (x , x ) ∑ Q(y , y ) = 0.
(x ,y )∈ I x ∈ IP y ∈ IQ

By the induction hypothesis, P × Q((x , y) , (x , y )) = P (x , x ) Q(y , y ) when

(x , y) ≤ (x , y ) < (x , y ). Canceling such terms from the two sums that equal 0
leaves the desired equality when (x , y ) = (x , y ).

15.2.37. Example. In the chain product M e with e = e1 , . . . , e n , the elements

are encoded as lists (a1 , . . . , an) with 0 ≤ ai ≤ e i − 1. In the ith factor, (j , j) = 1,
(j , j + 1) = −1, and (j , ) = 0 if  > j + 1, as in Example 15.2.35.
In 2 n the factor chains all have size 2. The product formula thus yields
2 n (T, S) = (−1)| S|−| T | . Hence the Inclusion-Exclusion Formula (Example 15.2.25)
is the special case of Theorem 15.2.34 for P = 2 n .
Now view M e as D(N), the poset of divisors of N . We have D(N)( , n) = (−1) j
if ¾ is the product of j distinct primes, or D(N)( , n) = 0 if ¾n is divisible by a
n

square. This special case is the classical Möbius function ( ¾n).

When formulas for both  and  in the Möbius Inversion Formula are known,
the formula can be used to find . Helped by the Polynomial Principle, this works
for the special lattices Ln(q) (Exercise 34) and Π n .

15.2.38. Example. Möbius function of the partition lattice Π n. Since  ≤  when

every block of  lies in a block of  , 0̂ consists of singletons and 1̂ has one block.
If  has  blocks, then the interval [ ,  ] is the product of partition lattices
Π n1 , . . . , Π n¾ , where ni counts the blocks of  contained in the ith block of  . By
the product rule, knowing (0̂ , 1̂) for each Π ni allows us to compute Π n ( , ).
Let ( ) count mappings from [n] to a set X of size x that are constant on
blocks of  . If  has b( ) blocks, then ( ) = xb( ). Let  ( ) count the mappings
whose inverse images are the blocks of  . Thus ( ) = ∑ ≥  (), and Möbius
inversion guarantees  ( ) = ∑ ≥ ( , )xb() .
To compute (0̂ , 1̂), consider  (0̂). This counts injective mappings, so  (0̂) =
x(x − 1) · · · (x − n + 1). Again this is a polynomial equality, since it holds for every
positive integer x. Since 1̂ is the only partition with 1 block, (0̂ , 1̂) is the only
contribution to the coefficient of the linear term, so (0̂ , 1̂) = (−1)n−1 (n − 1)!.

The Möbius function of a lattice yields much structural information; see

Aigner [1979] and Stanley [1986]. We briefly mention two related topics.

15.2.39. Remark. Whitney numbers. For a graded poset P with lower bound
0̂, the rank generating function ∑ W x has the Whitney numbers (rank sizes)
as coefficients; it sums (0̂ , a) over a ∈ P , weighted by x r(a) . The character-
istic polynomial ∑ w x sums (0̂ , a) over a ∈ P , weighted by xr(P)−r(a). The
coefficient w is the  th Whitney number of the first kind, while W is the  th
Whitney number of the second kind. For the partition lattice Π n , we obtain the
Stirling numbers: W (Π n) = Sn ,n− and wn−(Π n) = sn , .
762 Chapter 15: Linear Algebra

15.2.40. Remark. Some classical counting problems in terms of ¾ and n are

solved by a polynomial in  of degree n such that the answer to a related problem
is (−1)n (−). Stanley [1974] explored this combinatorial reciprocity.
For a bounded poset P , the zeta-polynomial is defined by ZP () =  ¾ (0̂ , 1̂).
Since  = 1 +  as incidence functions, ZP () = ∑ (¾i ) i(0̂ , 1̂), so ZP () is a polyno-
mial in  . Recall that  ¾ (0̂ , 1̂) counts the weak chains 0 ≤ · · · ≤ ¾ with (0 , ¾) =
(0̂ , 1̂), while i(0̂ , 1̂) counts the strict chains 0 < · · · < i with (0 , i) = (0̂ , 1̂).
Given an n-element poset P , let  P () be the number of order-preserving
maps from P to k, and let  P () be the number of strict order-preserving maps
from P to k. These are the order polynomial and strict order polynomial of
P , respectively. Each order-preserving map  : P → k corresponds to a weak chain
in J(P) by letting I¾ = {x ∈ P:  (x) ≤ }; we have ∅ = I0 ⊆ I1 ⊆ · · · ⊆ I¾ = P. The
bijection yields  P () = Z J(P)(). Thus  P() is a polynomial in  with degree n.
Applying Stanley ’s reciprocity theorem to Z J(P) yields  P () = (−1)n  P (−).
For example, when P = n, the order polynomial and strict order polyno-
mial count n-element multisets and subsets from  types, respectively. We saw in
Chapters 1 and 3 that there are (−1)n (−n¾ ) multisets.
Another example is the chromatic polynomial  G() of a graph G (Application
4.1.11). An acyclic orientation D and a -coloring  are strictly compatible if
x → y in D implies  (x) <  (y). Thus  G() is the number of strictly compatible
pairs. Changing the restriction  (x) <  (y) to  (x) ≤  (y) yields weakly compati-
ble pairs. Stanley ’s reciprocity theorem on an appropriate poset yields the num-
ber of weakly compatible pairs as (−1)n (G; −). As a special case, (−1)n (G; −1)
counts the acyclic orientations of G.

EXERCISES 15.2

15.2.1. (−) Use the Matrix Tree Theorem to find (G) for the graph below.
• •

• •

15.2.2. (−) In terms of the incidence functions in Lemma 15.2.31, compute

(a) The number of comparable pairs of elements in [x , y].
(b) The number of covering pairs in [x , y].
(c) The number of elements of height  in [x , y] (assume the interval is graded).
(d) The number of elements of co-height  in [x , y].
15.2.3. (−) Let P be the lattice of partitions of 6 with 111111 removed, so that 0̂ is the
partition 21111, and 1̂ is the partition 6.
(a) Compute (0̂ , x) for all x. Suppose that (x) = ∑ y≥ x  (y) and we know that (x) is
given by the rank of x in P. Compute  (0̂).
(b) Compute the zeta-polynomial of P. Use it to count the chains 0 ≤ · · · ≤ 10 such
that 0 = 0̂ and 10 = 1̂.
Exercises for Section 15.2 763

15.2.4. Use the Matrix Tree Theorem to prove Cayley’s Formula (Ì(K n) = nn−2).

15.2.5. Use the Matrix Tree Theorem to find Ì(K r,s). (Lovász [1979, p. 223])

15.2.6. (♦) The generalized book Bn ,¾ is the graph K ¾ K n−¾ . Compute Ì(Bn ,¾).

15.2.7. A matrix is totally unimodular if every square submatrix has determinant in

{0 , 1 , −1}. Prove that the incidence matrix of every loopless digraph is totally unimod-
ular. Prove that the incidence matrix of a graph is totally unimodular if and only if the
graph is bipartite (each column has two +1 ’s).

15.2.8. (♦) de Bruijn cycle (see Exercise 10.1.28). Prove that the following explicit al-
gorithm generates a binary de Bruijn cycle of length 2 n: start with n copies of 0. Sub-
sequently, append 1 if doing so does not repeat a previous string of length n; otherwise
append 0. (Martin [1934], Ungar [1950], and others)

15.2.9. (♦) Tarry’s Algorithm (Tarry [1895], as presented by D.G. Hoffman). A castle has
finitely many rooms and corridors. Each corridor has a door into a room at each end. Any
room can be reached from any other by traversing corridors and rooms. Initially, no doors
have marks. A robot starting in some room explores the castle using the following rules.
(1) After entering a corridor, traverse it into the room at the other end.
(2) Upon entering a room with all doors unmarked, mark I on the entry door.
(3) In a room, mark O on an unmarked door (if any exists), and use it.
(4) In a room with all doors marked, exit via a door marked I if one exists.
(5) In a room with all doors marked O, stop.
Prove that the robot stops only after using every corridor exactly twice, once in each direc-
tion. (Comment: All decisions are completely local; the robot sees nothing but the current
room or corridor. See also K önig [1936, pp. 35–56] and Fleischner [1983, 1991].)

15.2.10. Let M be the incidence matrix of an orientation of a tree T , and let B be an

(n − 1)-by-(n − 1) submatrix obtained by deleting the row of M for a fixed vertex v0 . The
columns of B correspond to the edges {e1 , . . . , e n−1 } of T . Let bi , j be the entry in B−1 in
the row for vertex vi and the column for edge ej . Let Pj be the unique path from vj to v0
in T . Prove that bi , j = 1 if ei is an odd edge of Pj , bi , j = −1 if ei is an even edge of Pj , and
otherwise bi , j = 0. (Branin [1959]; see also Bryant [1967])

15.2.11. Prove the Cauchy–Binet Formula det C = ∑S⊆[m] det AS det BS by considering
the matrix equation DE = F below.

Im 0 −I B −I B
( )( m )=( m )
A In A 0 0 AB

15.2.12. Given a graph G, prove combinatorially that the symmetric difference of any two
even subgraphs is an even subgraph, and the symmetric difference of any two edge cuts is
an edge cut. (Comment: This completes the proof of Theorem 15.2.17.)

15.2.13. (♦) Apply the results of this section to obtain the following statements about edge
cuts in graphs that earlier were proved combinatorially.
(a) A set of edges in a graph is an edge cut if and only if it has an even number of edges
in every cycle (Exercise 7.1.31).
(b) Every edge cut is a union of pairwise edge-disjoint bonds (Exercise 7.1.30).

15.2.14. (♦) Without using the bond space, show directly that the dimension of the cycle
space of a connected graph G is | E(G)|−| V(G)|+ 1. That is, show that the incidence vectors
of the cycles generated by adding one edge to a spanning tree T form a basis.
764 Chapter 15: Linear Algebra

15.2.15. Let G be a graph with n vertices, m edges, and ¾ components. Determine the
dimensions of the cycle space and the bond space of G.
15.2.16. Prove that the cycle space and bond space of a graph are the null space and row
space of its incidence matrix.
15.2.17. Prove Theorem 11.3.35 from Theorem 15.2.19.
15.2.18. (♦) Let W be a subspace of 2n . Prove that every coordinate that is not identically
0 in W is 1 in exactly half the elements of W . Conclude that the sum of the sizes of the
even subgraphs of a 2-edge-connected graph G with n vertices and m edges is n2 m− n .
15.2.19. A cycle C in a graph G is isometric if d C(x , y) = d G(x , y) for all x , y ∈ V(C).
Prove that the set of isometric cycles in G spans the cycle space of G. (see Naatz [2000])
15.2.20. Let Q be a nonempty family of subsets of [n]. Prove that the number of subsets
of Q using each element of [n] an even number of times equals the number using each
element of n an odd number of times. (Beckwith [2001])
15.2.21. (+) Light bulbs l1 , . . . , ln are controlled by switches s1 , . . . , sn . The ith switch
flips the status of the ith light and possibly others, but si flips the status of lj if and only
if sj flips the status of li . Initially all lights are off. Prove that it is possible to turn all the
lights on. (Lovász [1979], Peled [1992])
15.2.22. Prove that the number of 1-factors in the ¾-dimensional hypercube Q¾ is at least
exponential in ¾ . (Clark–George–Porter [1997], Lovász–Plummer [1986])
15.2.23. (♦) Show that Lemma 15.2.22 can generate the orientation of the grid below.
Use it and Theorem 15.2.23 to count the 1-factors in P4 P4 .

• • • •

• • • •

• • • •

• • • •

15.2.24. The Permanent Lemma, from the Combinatorial Nullstellensatz.

(a) Let A be an n-by-n matrix with nonzero permanent over a field K . Prove that for
any b ∈ K n and sets S1 , . . . , Sn of size 2 in K , there is a vector x ∈ ∏i=1 Si such that Ax
n

differs from b in every coordinate. (Alon–Tarsi [1989])

(b) Let p be a prime. Prove that every list of 2p − 1 members of p contains p terms
that sum to 0 modulo p. (Erdős–Ginzburg–Ziv [1961])
15.2.25. Two permutations are equivalent if they differ only in direction on cycles.
(a) Let B be a skew-symmetric matrix. Show that in the permutation expansion of
det B, the contribution of the terms from all the permutations containing odd cycles is 0.
(b) Let be a permutation in which all cycles have even length. Prove that when B is
skew-symmetric the sum of the contributions to det B over all permutations equivalent to
is (−1)r 2 r− s ∏i bi , (i) , where consists of r cycles, of which s have length 2.
(c) Given a 2n-by-2n matrix B and a partition P of [2n] into pairs {i1 , j1 } , . . . {in , j n},
let bP = (−1)n ∏r=1 bir , j r , The Pfaffian Pf B of a skew-symmetric matrix B is ∑ P bP , sum-
n

ming over all partitions of [2n] into pairs. Establish a bijection that associates a set of
terms in the computation of (Pf B)2 with an equivalence class of permutations of [n] whose
cycles all have even length. Conclude det B = (Pf B)2 . (Muir [1882, 1906])
Exercises for Section 15.2 765

15.2.26. For an oriented graph G with adjacency matrix B, prove (A)–(C) equivalent:
(A) |Pf B| is the number of 1-factors of G (see Exercise 15.2.25 for Pf B).
(B) For every perfect matching M in G, every M-alternating cycle traverses an odd
number of edges of G in each direction.
(C) For some perfect matching M in G, every M-alternating cycle traverses an odd
number of edges of G in each direction.
15.2.27. For m ≥ n/2, with n even, determine the multigraph with m edges and n vertices
that has the most perfect matchings. (Hajek–Narayanan [1994])
15.2.28. Let Mn be the graph obtained from the cycle C2n by adding edges joining opposite
vertices. Count the perfect matchings in Mn and in the graph Cn K 2 . (Hosoya–Harary
[1993], McSorley [1998]) (Comment: Esperet–Kardoš–King–Král’–Norine [2011] proved
that the minimum number of perfect matchings in 3-regular bridgeless graphs grows ex-
ponentially with the number of vertices.)
15.2.29. Let G n be the class of 3-regular connected graphs with n white vertices, n black
vertices, and no monochromatic cycles. In a good orientation of G ∈ G n , white vertices
have indegree 2 and black vertices have outdegree 2. (R.E. Prather)
(a) Prove that the numbers of purely white edges and purely black edges are equal, and
both color classes induce forests, with white trees W1 , . . . , W¾ and black trees B1 , . . . , B¾ .
(b) Let M(G) denote the ¾-by- ¾ matrix in which entry mi , j is the number of edges from
Wi to Bj . Prove that the number of good orientations of G is the permanent of M(G).
(c) Find the least n such that G n has a graph with no good orientation.
(d) For even n, construct a member of G n with 1 + 3n/2 good orientations.
15.2.30. A matrix is doubly stochastic if all entries are nonnegative and rows and
columns all sum to 1. Egorychev [1981] and Falikman [1981] proved the van der Waer-
den Conjecture that the permanent of such a matrix of order n is at least n!/nn , with
equality only when every entry is 1/n. Here are some initial steps.
(a) Prove that a nonnegative matrix of order n has permanent 0 if and only if it has a
¾-by-(n + 1 − ¾) submatrix of 0s for some ¾ .
(b) A matrix of order n is decomposable if it has a ¾-by-(n−¾)submatrix of 0s, for some
¾ . Prove that a doubly stochastic matrix minimizing the permanent is indecomposable.
(c) A doubly stochastic matrix has 1 as an eigenvalue. Prove that the multiplicity of
1 as an eigenvalue is 1 if and only if the matrix is indecomposable.
15.2.31. (+) van der Waerden Conjecture for matrices with all entries nonzero. Doubly
stochastic matrices of order n having the smallest permanent are called optimal.
(a) Prove that if A is an indecomposable matrix with nonnegative entries (Exercise
15.2.30), then A A T and A T A are also indecomposable.
(b) Let Ai , j be obtained from A by deleting row i and column j . Prove that if A is an
optimal matrix with ai , j > 0, then per Ai , j = per A. (Marcus–Newman [1959])
(c) Show that if per Ai , j = per A for every i , j , then replacing a column by an arbi-
trary vector with the same sum does not change the permanent.
(d) Conclude that the doubly stochastic matrix with all entries equal has the smallest
permanent among doubly stochastic matrices with all entries nonzero.
15.2.32. Properties of invertible incidence functions.
(a) If º is an invertible incidence function, prove that the left inverse of º computed
in Lemma 15.2.29 is also a right inverse.
(b) If 1 − º is an invertible incidence function, prove that (1 − º )−1 = ∑¾≥0 º ¾ .

15.2.33. Let º and ½ be functions on poset P such that ½(x) = ∑ y≤ x º (x) for all x ∈ P.
Using the recursive definition of the Möbius function, prove by algebraic manipulation
that ∑ y≤ x ½(y)Þ(y , x) = º (x). (Comment: This completes Proof 1 of Theorem 15.2.33.)
766 Chapter 15: Linear Algebra

15.2.34. Möbius function of the subspace lattice Ln(q)(subspaces of an n-dimensional vector

space V over a q-element field, ordered by inclusion).
(a) Reduce computing the Möbius function to computing (0̂ , 1̂) for all n.
(b) Consider linear functions from V to a vector space X of size x. Let  (U) count such
functions with nullspace U. Let (U) count such functions whose nullspace contains U;
thus (U) = ∑W ⊇U  (W). Prove that (U) = x n−dim U .
(c) Prove that  (∅) = (x − 1)(x − q) · · · (x − qn−1).
(d) Conclude that (0̂ , 1̂) = (−1)n q(2).
n

15.2.35. Let  and  be incidence functions on a poset P. Let A be the matrix with
rows and columns indexed by elements of P whose (x , y)-entry is ∑ ∈ P  ( , x)( , y). Prove
that det A = ∏ ∈ P  ( , )( , ). (Hint: Generalize the argument for Exercise 4.1.65.)
(Altinisik–Sagan–Tuglu [2005])
15.2.36. Let F be a family of finite sets closed under taking subsets (this is the definition
of a simplicial complex). For A ∈ F , let U(A) denote the family of all members of F that

contain A. Fix  ∈ . Suppose that ∑ B∈U(A)(−1)| B| = 0 whenever | A| <  . Prove that | F | is
divisible by 2  . (Stanley [2009])

15.3. Eigenvalues
There are important applications of eigenvalues of various combinatorial ma-
trices, such as transition matrices of Markov chains. We restrict our attention
to matrices arising from the adjacency matrix of a graph. The classical approach
uses the adjacency matrix itself; see Biggs [1974] and a more encyclopedic treat-
ment in Cvetković–Doob–Sachs [1979]. Other texts with further applications
and eigenvalues of other graph matrices include Bapat [2010], Brouwer–Haemers
[2012], and the broader Godsil–Royle [2001], which studies many algebraic top-
ics in graph theory. We consider first the classical eigenvalues and then briefly
discuss a modern alternative.

SPECTRA OF GRAPHS

We assume a basic acquaintance with eigenvalues of matrices and simply

state several elementary properties.
15.3.1. Definition. The eigenvalues of a graph are the eigenvalues of its adja-
cency matrix A. These are the roots ë 1 , . . . , ë n of the characteristic poly-
nomial á given by á(G; ë) = det(ë I − A) = ∏i=1 (ë − ë i). The spectrum is the
n

list of distinct eigenvalues with their respective multiplicities m1 , . . . , mt ;

we write Spec (G) = (më 11 ··· ët
···mt ).

15.3.2. Remark. Elementary properties of eigenvalues.

(0) The eigenvalues are the values ë for which the square matrix ë I − A is
singular, which is equivalent to det(ë I − A) = 0.
(1) ∑ ë i = Trace A. The trace is the sum of the diagonal elements and is the
negative of the coefficient of ën−1 in det(ë I − A). Since det(ë I − A) = ∏i=1 (ë − ë i),
n

this also equals ∑ ë i . For graphs, it is 0.

Section 15.3: Eigenvalues 767

(2) ∏ i = (−1)n (G; 0) = det A = ∑ sign( ) ∏ni=1 ai , (i) , where the sum runs
over permutations  of [n].
(3) For a symmetric real n-by-n matrix A and ∈ , the multiplicity of as
an eigenvalue of A is n − rank( I − A).
(4) Adding c to the diagonal shifts the eigenvalues by c, since  + c is a root
of det( I − (cI + A)) if and only if  is a root of det( I − A).

15.3.3. Example. The adjacency matrix of K n is J − I, where J is the matrix of

all 1s. Hence the eigenvalues of K n are 1 less than those of J. Since Spec J =
(1n n−01 ), we have Spec K n = ( n−11 n−−11).
The adjacency matrix of K m ,n has rank 2, so it has two nonzero eigenvalues
1 , 2 . The trace is 0, so 1 = − 2 ; call this constant b. Hence (K m ,n , ) =
n
− b2 n−2 . We compute b using (G; ) = det( I − A). Since appears only on
the diagonal, contributions in the permutation expansion to the coefficient of n−2
arise only from permutations that use n− 2 positions on the diagonal. The remain-
ing two positions are − ai , j and − aj,i for some i , j . All mn nonzero
√
contributions
√
of this form are negative. Hence b = mn, and Spec (K m ,n) = ( 1 m+0n−2 − mn
2 mn
1
).

We index the coefficients of  so that (G; ) = ∑i=0 ci n−i . Since (G; ) =

n

det( I − A), always c0 = 1 and c1 = −Trace A = 0.

The computation of c2 for K m ,n generalizes. Since contributions to c2 n−2 in-
volve n − 2 factors of from the diagonal, c2 is the sum of the principal 2-by-2
subdeterminants of − A. For a graph, the off-diagonal elements are both −1 when
vi vj ∈ E(G) and both 0 otherwise. Each permutation selecting these positions is
odd. Thus summing over all vertex pairs yields c2 = − | E(G)|. The other coeffi-
cients of (G; ) can also be described in terms of subgraphs of G (Exercise 7).
Eigenvalues yield characterizations of bipartite graphs.

15.3.4. Proposition. The (i , j)th entry of A¾ counts the vi , vj -walks of length  .

The eigenvalues of A¾ are ( i)¾ .
Proof: The statement about walks holds easily by induction on  (Exercise 1). For
the second statement, Ax = x implies A¾ x = ¾ x, by repeated multiplication.
Since x is any eigenvector, multiplicities of eigenvalues for A and A¾ are equal.

15.3.5. Lemma. If G is bipartite and is an eigenvalue of G with multiplicity

m, then − is also an eigenvalue with multiplicity m.
Proof: We may assume that the two parts have equal size, since adding isolated
vertices merely adds rows and columns of 0s to the adjacency matrix and adds
zeros to the list of eigenvalues. This allows us to write the adjacency matrix A(G)
in the form ( B0T B0), where B is square.
Let be an eigenvalue associated with eigenvector v. Write v in the form
( yx), partitioned via the bipartition of G. Now v = Av = ( B0T B0)( yx) = ( BBy
T x). Hence

By = x and B x = y.
T
− y)
Let v = (−xy). Compute Av = ( B(BT x
) = (−ëëyx) = − v . Hence v is an eigen-
vector of A with eigenvalue − . Furthermore, m independent eigenvectors for
yield m independent eigenvectors for − . Since the same argument holds for − ,
we conclude that − is an eigenvalue with the same multiplicity as .
768 Chapter 15: Linear Algebra

15.3.6. Theorem. The following are equivalent for a graph G.

(A) G is bipartite.
(B) The nonzero eigenvalues occur in pairs ë i , ë j such that ë i = − ë j .
(C) (G; ) or  (G; ) is a polynomial in 2 .
(D) ∑ 2t
i
−1 = 0 for every positive integer t.

Proof: We proved A ⇒B in the lemma.

B ⇔ C: We have ( −  i)( −  j ) = (2 − a) if and only if  j = −  i .
B ⇒ D: If  j = −  i , then 2t
j
−1 = − 2t−1 .
i
D ⇒ A: Since ∑ ¾i counts the closed -walks in the graph (from each vertex),
D forbids closed walks of odd length. This forbids odd cycles, since an odd cycle is
an odd closed walk, and hence G is bipartite.

Standard results from linear algebra are useful. For a real symmetric ma-
trix of order n, the Spectral Theorem guarantees real eigenvalues and n or-
thonormal eigenvectors. Also, x T Ax attains its maximum and minimum over
unit vectors at eigenvectors of A, where it equals the corresponding eigenvalue.
The Cayley–Hamilton Theorem states (A) = 0. Finally, the minimum poly-
nomial of a matrix A is the polynomial  of least degree such that (A) = 0; it is
given by (A) = ∏i( − i), where 1 , . . . , r are the distinct eigenvalues of A.

EIGENVALUES AND GRAPH PARAMETERS

We first apply the minimum polynomial.

15.3.7. Theorem. The diameter of a graph G is less than the number of distinct
eigenvalues of G.
Proof: The adjacency matrix A satisfies a polynomial of degree r if and only if
some linear combination of A0 , . . . , Ar is 0. Since the number of distinct eigenval-
ues is the degree of the minimum polynomial, we need only show that A0 , . . . , A¾
are linearly independent when  ≤ diam (G).
Finally, for  ≤ diam (G) choose vi , vj ∈ V(G) such that d(vi , vj ) =  . By
counting walks, we have A¾i , j
= 0 but Ati , j = 0 for t <  . Hence A¾ cannot be a
linear combination of the smaller powers.

Let  max(G) and  min(G) denote the largest and smallest eigenvalues.
15.3.8. Lemma. If G is an induced subgraph of G, then
 min(G) ≤  min(G ) ≤  max(G ) ≤  max(G) .
Proof: The adjacency matrix A of G is real and symmetric, so  min(A) ≤ x T Ax ≤
 max(A) for every unit vector x. Index V(G) to put the adjacency matrix A of
G in the upper left of A. Let  be the unit eigenvector of A such that A  =
 max(G ) . Let  be the unit vector in n obtained by appending zeros to  . Now
 max(G ) =  T A  =  T A  ≤  max(G).
Similarly,  min(G ) ≥  min(G).
Section 15.3: Eigenvalues 769

Lemma 15.3.8 is a special case of the Interlacing Theorem, which we state

without proof (it uses only linear algebra).

15.3.9. Theorem. Indexed in nonincreasing order, the eigenvalues 1 ,... , n of

a graph G and 1 , . . . , n−1 of G − x (for x ∈ V(G)) interlace as
1 ≥ 1 ≥ 2 ≥ · · · ≥ n−1 ≥ n.

15.3.10. Lemma. For an n-vertex graph G with m edges,

2m
(G) ≤ ≤ max (G) ≤ (G).
n
Proof: Let x be an eigenvector for eigenvalue , and let x j be the largest entry in
x. With A being the adjacency matrix, ≤ (G) follows from
x j = (Ax)j = ∑vi ∈ N(vj ) x i ≤ d(vj)x j ≤ (G)x j .

For the lower bound, evaluate

√ x T Ax by letting x be the unit vector with equal
coordinates, written as 1n/ n, where every entry of 1n is 1. Since the sum of the
entries in the adjacency matrix is twice | E(G)| ,
1T 1 1 2m
max ≥ √n A √n =
n n n ∑ ∑ ai j =
n
.

Lemma 15.3.10 enables us to improve the trivial bound (G) ≤ 1 + (G) given
by the greedy coloring algorithm. Replacing (G) with the average degree is too
small; K n + K 1 has chromatic number n and average degree less than n − 1. Since
max is always at least the average degree, 1 + max (G) has a chance to work and
cannot be much improved.

15.3.11. Theorem. (Wilf [1967]) Always (G) ≤ 1 + max(G).

Proof: Let  = (G). Let H be a minimal induced subgraph of G with chromatic

number  . As in Proposition 8.2.2, (H) ≥  − 1 (otherwise a proper coloring of
a subgraph would extend to H). Since H is an induced subgraph of G, Lemma
15.3.10 and then Lemma 15.3.8 yield

 ≤ 1 + (H) ≤ 1 + max(H) ≤1+ max (G) .

The relationship between max and the average degree 2m/n also aids in de-
veloping an eigenvalue proof of the upper bound in Mantel’s Theorem, the case of
Turán’s Theorem (Theorem 5.2.11) for triangle-free graphs. The essence of this
argument can be found in Bapat [2010, p. 74].

15.3.12. Theorem. Every n-vertex triangle-free graph has at most n2/4 edges.
Proof: Let G be an n-vertex graph with m edges, t triangles, adjacency matrix
A, and eigenvalues 1 , . . . , n in nonincreasing order.
The triangles in a graph are its 3-cycles, which are closed walks. Each cycle
of length 3 through vertex vi is counted twice in the ith diagonal entry of A3 ;
770 Chapter 15: Linear Algebra

hence Trace A3 counts each triangle six times. The trace of a matrix is the sum
of its eigenvalues, and the eigenvalues of A¾ are ë¾1 , . . . , ë¾n . Hence
6t = Trace A3 = ∑i=1 ë3i = ë31 + ∑ i=2 ë3i .
n n

We want to prove that the number of triangles is positive when m > n2/4.
For this it suffices to show that the assumption m > n2/4 implies ë31 > ####∑i=2 ë3i ####.
n
√ √ # #
If m > n2/4, then m > n/2, and hence ë 1 ≥ 2m/n > m. Since the ith
diagonal element of A2 counts the edges incident to vi ,
2m = Trace A2 = ∑i=1 ë2i > m + ∑i=2 ë2i .
n n

√
From this and ë 1 > m we obtain ë21 > m > ∑i=2 ë2i . Thus ë 1 is the only eigen-
n

value of largest absolute value. Now we use the triangle inequality and these
facts to compute
###∑n ë3 ### ≤ ∑ n ### ë3 ### = ∑n | ë i |3 < ∑n ë 1 | ë i |2 < ë 1 ë2 = ë3 .
## i=2 i ## i=2 ## i ## i=2 i=2 1 1

15.3.13.* Remark. Mantel’s Theorem is just the beginning. Recall that T n ,r de-
notes the r-partite Turán graph with n vertices (Definition 5.2.10), and tr(n) =
| E(T n ,r)|. Nikiforov [2007] proved that if G is an n-vertex graph not contain-
ing K r+1 , then ë max(G) ≤ ë max(T n ,r), with equality only when G = T n ,r . Using
Lemma 15.3.10 and the fact that | E(T n ,r)| = ⌊ n ë max(T n ,r)/2⌋ , Turán’s Theorem
that T n ,r is the unique such graph with the most edges is a corollary.
Nikiforov went further. Fixing s ∈ and c > 0, the Erdős–Stone Theorem
(Theorem 11.1.10) states that for sufficiently large n, every n-vertex graph with
at least tr(n) + cn2 edges contains a complete (r + 1)-partite graph with s vertices
in each part. Bollobás–Erdős [1973] showed that exponentially large n suffices, or
equivalently that one can guarantee s to be (log n).
Nikiforov [2009] proved a stronger version of these results using eigenvalues,
r −1
showing for r ≥ 3 that if s = ⌊(c/r r)r log n⌋ ≥ 1 and t = ⌈ n1−c ⌉ , then every n-
vertex graph with ë max ≥ (1 − 1/(r − 1) + c)n contains a complete r-partite graph
with r − 1 parts of size s and one part of size t. Note that | E(G)| ≥ tr−1 (n) + cn2/2
implies this condition on ë max .

Eigenvalues are also related to the number of complete bipartite graphs

needed to decompose a graph. For convenience, here we use biclique to mean
a complete bipartite graph. Because stars are bicliques, the number needed is
at most the vertex cover number Î(G). Graphs with special structure may have
more efficient decompositions. The spectrum provides a lower bound.

15.3.14. Theorem. For every graph G, the number of bicliques needed to par-
tition E(G) is at least max{p , q}, where p and q count (including repetition)
the positive and negative eigenvalues of the adjacency matrix A.
Proof: When G decomposes into subgraphs G 1 , . . . , G t , we may write A =
t
∑i=1 Bi , where Bi is the adjacency matrix of the spanning subgraph of G with
edge set E(G i). To write this algebraically, let {Ri , Si} be the bipartition of G i ,
and for U ⊆ V(G) let 1U denote the vector that is 1 in the positions for U and 0
elsewhere. Now A = ∑i=1 (1 Ri 1STi + 1Si 1 TRi).
t
Section 15.3: Eigenvalues 771

Let W0 be the orthogonal complement of the space spanned by 1 R1 , . . . , 1 Rt ;

that is, W0 = {x ∈ n : x T 1 Ri = 0 for 1 ≤ i ≤ t}. Let W+ be the space spanned
by the eigenvectors of A associated with positive eigenvalues. Thus x T Ax > 0 for
x ∈ W+ . In contrast, for x ∈ W0 ,
t t
x T Ax = ∑ x T 1 Ri 1STi x + ∑ x T 1Si 1 TRi x = 0.
i=1 i=1

Thus W0 ∩ W+ = {0}, which yields dim W0 + dim W+ ≤ n. Vectors in W0 satisfy t

linear equations, so dim W0 ≥ n − t. By definition, dim W+ = p. Thus t ≥ p. Using
instead of W+ the space associated with the negative eigenvalues yields t ≥ q.

Sylvester ’s Law of Inertia states that for any symmetric real matrix Q with p
positive eigenvalues and q negative eigenvalues, the number of products of linear
functionals needed to express x t Qx is at least max{p , q}. The special case above
for biclique decomposition is directly tailored to the graph application.
The biclique decomposition problem arose with a network addressing motiva-
tion in Graham–Pollak [1971], where the focus was especially on decomposing K n .
They attribute Theorem 15.3.14 to H.S. Witsenhausen. The proof given here is
from Gregory & Vander Meulen [1996] (see also Brouwer–Haemers [2012]). For
other discussions of the lower bound, see Tverberg [1982], Peck [1984], Reznick–
Tiwari–West [1985]. A conjecture of Erdős from the 1980s that a random n-vertex
graph decomposes into no fewer than n − (G) bicliques was disproved in Alon
[2015] and more strongly in Alon–Bohman–Huang [2017].

15.3.15. Example. Biclique decompositions. When G = K n , there are n − 1 neg-

ative eigenvalues. Using stars centered at the complement of an independent set
always yields an upper bound of n − (G), so the eigenvalue bound here is sharp.
Kratzke–Reznick–West [1988] studies other such graphs. For K n , no short purely
combinatorial proof is known that n − 1 bicliques are needed. If a smallest decom-
position must use a spanning biclique, then the rest follows by induction.
Next consider G = Cm Cn . There are simple formulas for the eigenvalues of
a cycle (Exercise 12) and for the behavior of eigenvalues under cartesian product
(Exercise 13). From these, C(2t+1)n Cn for odd n has (2t + 1)(n2 + 1)/2 positive
eigenvalues and (2t + 1)(n2 − 1)/2 negative eigenvalues (0 is not an eigenvalue).
Furthermore, this product decomposes into (2t + 1)(n2 + 1)/2 bicliques, using
(2t + 1)(n − 1)/2 copies of K 2 ,2 and (2t + 1)(n + 1)/2 stars (Kratzke–West). All bi-
cliques in Cm Cn are 4-cycles or stars; see C15 C5 below. Edges wrap from top to
bottom and right to left, and all grid points are vertices. The dots indicate centers
of stars in the decomposition; the circles indicate 4-cycles in the decomposition.

• • •
• • • • • •
• • • • • • • • •
• • • • • •
• • •
772 Chapter 15: Linear Algebra

REGUL AR AND STRONGLY REGUL AR GRAPHS

Like bipartite graphs, regular graphs have spectral characterizations. The

n-vector 1n with all coordinates 1 and the all-1 matrix J play important roles in
arguments involving eigenvalues of regular graphs.

15.3.16. Theorem. The eigenvalue of G with largest absolute value is (G) if

and only if some component of G is (G)-regular. The multiplicity of (G) as
an eigenvalue is the number of such components.
Proof: Let A be the adjacency matrix. Since the spectrum of a graph is the union
of the spectra of its components, we may assume that G is connected.
The ith entry of A1n is d(vi). When G is ¾-regular, A1n = ¾ 1n , so ¾ is an
eigenvalue with eigenvector 1n .
By Lemma 15.3.10, no eigenvalue exceeds (G). For the lower bound, let x be
an eigenvector for eigenvalue ë, with x j an entry of largest absolute value. Since
− x is also an eigenvector for ë, we may assume x j ≥ 0. We obtain | ë| ≤ (G) from
# #
| ë x j | = |(Ax)j | = ##### ∑ x i##### ≤ d(vj)x j ≤ (G)x j .
# #
vi ∈ N(vj )

Now suppose ë = (G). Removing the absolute value signs, equality requires
d(vj) = (G) and x i = x j for all vi ∈ N(vj). Iterating the argument reaches all
coordinates, since G is connected. Hence the eigenvalue associated with x equals
(G) only if G is regular and x is constant.

When G is connected and not regular, eigenvalues of largest absolute value

still have multiplicity 1. For non-regular graphs, we will see in the next subsec-
tion that the analogous results are much simpler when we modify the adjacency
matrix to take the vertex degrees into account. Meanwhile, we study the struc-
ture of regular graphs. Powers of the adjacency matrix yield another characteri-
zation of regular graphs.

15.3.17. Theorem. (Hoffman [1963]) A graph G with adjacency matrix A is reg-

ular and connected if and only if J is a linear combination of powers of A.
Proof: Sufficiency. If J can be so expressed, then for each (i , j) we have (A¾)i , j
=
0 for some nonnegative ¾ , which requires a vi , vj -walk of length ¾ . Hence G is
connected. For regularity, consider JA and AJ. The entry in position (i , j) of AJ
is d(vi) (constant on rows), and in JA it is d(vj) (constant on columns). Since J
is a linear combination of powers of A, each of which commutes with A, we have
JA = AJ. Thus the entry in position (i , j) is both d(vi) and d(vj), and G is regular.
Necessity. Let G be ¾-regular, so ¾ is an eigenvalue. Hence the minimum poly-
nomial Ú(G; ë) is (ë − ¾)½(ë) for some polynomial ½ . Since Ú(G; A) = 0, we have
A½(A) = ¾½(A). Hence each column of ½(A) is an eigenvector of A with eigen-
value ¾ . As shown in the proof of Theorem 15.3.16, an eigenvector of a connected
regular graph associated with eigenvalue (G) must have all coordinates equal.
However, ½(A) is a linear combination of powers of a symmetric matrix and must
itself be symmetric. Hence the columns are equal, and ½(A) is a multiple of J.
Section 15.3: Eigenvalues 773

The independence number is bounded in terms of the eigenvalues. Haemers

[1980] generalized this to all graphs, proving (G) ≤ ( (G)−2/ën )−  n, but we present
1 n
only the special case for -regular graphs, where (G) = (G) =  1 = .

15.3.18. Theorem. (A. Hoffman) If G is a -regular graph with n vertices, then

(G) ≤ −−nn n. Consequently, (G) ≥ (G)
n
≥ 1 + − n .
Proof: The eigenvectors for the adjacency matrix A and for A −  n I are the same,
and the eigenvalues shift by  n . In particular, (A −  n I)1n = ( −  n)1n . This
yields M1n = 0, where M = A −  n I − ( −  n)J. The other eigenvalues of A have
corresponding eigenvectors that are orthogonal to 1n and hence in the nullspace
of J. Thus the eigenvalues of M are the same as for A −  n I; they are nonnegative.
Let x be the incidence vector of an independent set of size s. Since x T Ax =
∑i , j x i ai , j x j , we have x T Ax = 0. Using the Spectral Theorem,
− n − n
0 ≤ x T Mx = x T Ax −  n x T Ix − x T Jx = −  n s − s2 ,
n n
which yields the desired inequality.

We study a special class with stronger regularity properties.

15.3.19. Definition. An n-vertex graph G is strongly regular with parameters

( ,  , ) if G is -regular, every two adjacent vertices have  common neigh-
bors, and every two nonadjacent vertices have  common neighbors. We then
say that G is ( ,  , )-strongly regular.

15.3.20. Example. The graph mK n is a strongly regular graph with param-

eters (n − 1 , n − 2 , 0). Its complement, the complete m-partite graph K n ,... ,n ,
is (mn − n , mn − 2n , mn − n)-strongly regular. Indeed, the complement of any
strongly regular graph is strongly regular (Exercise 23).
The Petersen graph is (3 , 0 , 1)-strongly regular. The cartesian product
K n K n is (2n − 2 , n − 2 , 2)-strongly regular. Strongly regular graphs with  = 
can be obtained from symmetric designs (13).

Counting arguments and then eigenvalue arguments greatly restrict the pos-
sibilities for the parameters ( ,  , ) of a strongly regular graph.

15.3.21. Proposition. If G is a ( ,  , )-strongly regular graph with n vertices,

then ( −  − 1) = (n − − 1).
Proof: We count induced copies of P3 having v as an endpoint. The middle vertex
w can be picked in ways. For each such w, the third vertex can be any neighbor
of w not adjacent to v. With v unavailable, there are −  − 1 ways to pick it. On
the other hand, the other endpoint can be picked in n− − 1 ways as a nonneighbor
of v, and each such choice has  common neighbors with v that can serve as w.

15.3.22. Remark. Every strongly regular graph with parameters  = 0 or  = − 1

or = n − 1 has the form mK +1 . By Proposition 15.3.21,  = − 1 if and only
if  = 0 or = n − 1. If  = − 1, then every neighbor of v is adjacent to every
other, which forces the claimed structure.
774 Chapter 15: Linear Algebra

Henceforth, we assume > 0 and  <  − 1. Next we compute the eigenvalues

in terms of the parameters.

15.3.23. Lemma. If G is ( ,  , )-strongly regular, then the eigenvalues of G are

√
 and 12 (  − ± ( − )2 + 4( − )).
Proof: Let A be the adjacency matrix. Since positions in A2 count walks of length
2, the i jth entry of A2 is  if i = j , is  if vi vj ∈ E(G), and is if vi vj ∈
/ E(G).
Since J − I − A is the adjacency matrix of G, we obtain A =  I +  A + (J − I − A).
2

Rearranging terms yields

A2 = ( − )I + ( − )A + J.
Since G is -regular, G has eigenvalue  with eigenvector 1. Let x be another
eigenvector, with eigenvalue  . Since x · 1 = 0, we have Jx = 0. Computing A2 x
thus produces 2 x = ( − )x + ( − ) x.
This yields the quadratic equation 2 −( − ) −( − ) = 0 for  . Its solutions
r and s are the only possible values for the eigenvalues other than  . The stated
values arise from the Quadratic Formula.

The argument above can also be traced in the opposite direction: a -regular
connected graph G is strongly regular with parameters  ,  , if and only if it has
exactly three eigenvalues  > r > s and these satisfy r + s =  − and rs = −( − ).
Next we obtain a further necessary condition for the parameters of a strongly
regular graph, again using that the multiplicities of the eigenvalues are integers.

15.3.24. Theorem. (Integrality Condition) If G is strongly regular with n ver-

tices and parameters ( ,  , ), then the following are nonnegative integers:

1 (n − 1)( − ) − 2 
(n − 1 ± √ ).
2 ( − )2 + 4( − )

Proof: We show that these numbers are the multiplicities a and b of the eigen-
values r and s in Lemma 15.3.23. Remark 15.3.22 describes all cases when = 0.
Hence we may assume > 0, and thus G is connected.
Because G is connected, eigenvalue  has multiplicity 1, and 1 + a + b = n.
Also the eigenvalues sum to 0, so  + ra + sb = 0. The solution to these two lin-
ear equations is a = − ¾+ s(n−1)
r− s and b = ¾+r(n−1)
r− s . Letting r and s be the numbers
computed in Lemma 15.3.23 yields the claimed formulas as the multiplicities.

15.3.25. Example. Classes of strongly regular graphs. Consider the two cases
(n − 1)( − ) = 2  and (n − 1)( − )
= 2  .
Excluding trivial values, in the first case =  + 1, since 0 < 2  < 2n − 2, and
hence  = (n − 1)/2. By Exercise 23, G and G are thus strongly regular with the
same parameters. Using =  + 1 and  = (n − 1)/2, Proposition 15.3.21 yields
n = 4 + 1. Furthermore, the eigenvalues r and s have the same multiplicity.
In the second case, rationality requires ( − )2 + 4( − ) = d2 for some d ∈
 , and d must divide (n − 1)( − ) − 2  . By Lemma 15.3.23, the eigenvalues are
integer multiples of 1/2. For ( , ) = (0 , 2), three such graphs are known, but
Section 15.3: Eigenvalues 775

it is not known whether the list is finite! The known examples, with parameters
(n , ¾ , ë , Þ), are C4 (4 , 2 , 0 , 2), the Clebsch graph (16 , 5 , 0 , 2) (see Exercise 24),
and the Gewirtz graph (56 , 10 , 0 , 2) (see Cameron–van Lint [1991], p. 43). Other
strongly regular graphs appear in Exercises 25–27.

The most famous application of the Integrality Condition is the nonexistence

of certain graphs. Recall that a graph with maximum degree ¾ and diameter d
has at most 1 + ¾ ∑i=1 d i−1 vertices (Proposition 5.4.12); graphs satisfying equal-
d

ity are Moore graphs. Damerell [1973] and Bannai–Ito [1973] independently
proved that no Moore graph exists with diameter at least 3. The Integrality Con-
dition eliminates almost all possibilities when d = 2.

15.3.26. Theorem. (Hoffman–Singleton [1960]) If a ¾-regular graph G with di-

ameter 2 has ¾ 2 + 1 vertices, then ¾ ∈ {2 , 3 , 7 , 57}.
Proof: Reaching ¾ 2 other vertices from any one vertex in two steps forbids trian-
gles and 4-cycles. Hence G is (¾ , 0 , 1)-strongly regular. √
From Theorem 15.3.24, we conclude that 12 ( ¾ 2 ± ¾(¾ − 2)/ 4¾ − 3) are non-
√
negative integers. Let t = 4¾ − 3. Since 4¾ − 3 is a positive integer, also t is a
positive integer, or t is irrational. If t is irrational, then ¾(¾ − 2) must vanish,
which is the case where G = C5 .
Hence t is a positive integer, and ¾ = (t2 + 3)/4. Since the multiplicities of
the eigenvalues are distinct nonnegative integers, their difference is an integer.
Its value is ¾(¾ − 2)/t, which equals (t4 − 2t2 − 15)/(16t). Letting u be this integer,
t4 − 2t2 − 16ut = 15. Since t divides the left side, it must also divide 15.
Hence t ∈ {1 , 3 , 5 , 15}, which leaves for ¾ the possibilities {1 , 3 , 7 , 57} (and
earlier 2). For ¾ = 1, the diameter is 1, and K 2 is a degenerate Moore graph. We
restrict to ¾ ∈ {2 , 3 , 7 , 57}. The first three give known graphs: C5 , the Petersen
graph, and the Hoffman–Singleton graph (Exercise 5.4.59), but it is unknown
whether there exists such a graph with degree 57.

Another classical application is the “Friendship Theorem”: at any party

where every two people have exactly one common acquaintance, some person
knows everyone (presumably the host). The resulting graph of the acquaintance
relation consists of some number of triangles sharing a vertex.

• •
• •
• •
•
• •
• •
• •
The symmetry of the condition suggests that G might be regular. The Inte-
grality Condition excludes this; the non-regular case is combinatorial.

15.3.27. Theorem. (Friendship Theorem; Higman (see Wilf [1971])) If G is a

graph in which any two distinct vertices have exactly one common neighbor,
then G has a vertex joined to all others.
776 Chapter 15: Linear Algebra

Proof: If G is regular, √ ë = Þ = 1. By Theorem

√ then it is strongly regular with
15.3.24, 2 − 1 ± ¾/ ¾ − 1) are integers. Hence ¾/ ¾ − 1 is an integer, which
1
(n
requires ¾ = 2. However, K 3 is the only 2-regular graph satisfying the condition,
and it does have vertices of degree n − 1.
Now suppose that G is not regular. We show first that v  w requires d(v) =
d(w). Since each vertex in N(v) does not equal w, it has a unique neighbor in N(w),
since w is in no 4-cycle. Similarly, each vertex of N(w) has a unique neighbor in
N(v). Hence N(v) and N(w) are joined by a matching, yielding | N(v)| = | N(w)|.
Since G is not regular, it has vertices v and w with d(w)
= d(v). By the pre-
ceding paragraph, vw ∈ E(G). Let u be their common neighbor. Since u cannot
have the same degree as both, we may assume d(u)
= d(v). If G has a vertex
x∈ / N(v), then d(x) = d(v), but this requires w , u ∈ N(x). This creates the 4-cycle
v , u , x , w. Hence d(v) = n − 1.

15.3.28. Remark. Such a combinatorial statement should have a short combina-

torial proof. An early simple proof was in Longyear–Parsons [1972], and another
found in 1975 was not published until Huneke [2002]; we present the latter.
In a ¾-regular n-vertex graph G where every two vertices have exactly one
common neighbor, exactly one walk of length 2 from v reaches each vertex of G,
and hence n = ¾(¾ − 1) + 1. Now count closed walks of length p from each vertex.
When p is prime, the total number of such walks is divisible by p, since each walk
is counted p times (repeated vertices cannot introduce periodicity).
Fixing a vertex v, let º (j) count the closed v , v-walks of length j . Of these,
¾º (j − 2) visit v two steps before the end. The total number of walks of length
j − 2 from v is ¾ j −2 , since G is ¾-regular. Each one not ending at v extends to a
v , v-walk of length j in exactly one way, yielding ¾ j −2 − º (j − 2) walks. Hence º (j)
equals ¾ j −2 + (¾ − 1)º (j − 2), with º (0) = 1.
Letting p be a prime divisor of ¾ − 1, we have º (p) ≡ 1 (mod p). Summing
over v yields nº (p) as the total number of closed walks of length p. Since n =
¾(¾ − 1) + 1 ≡ 1 (mod p), we have nº (p) ≡ 1 (mod p), but we observed that the total
number of these walks is divisible by p.

L APL ACIAN EIGENVALUES

Since the powers of the adjacency matrix count walks of fixed lengths, eigen-
values are related to distance and diameter. Eigenvalues of regular graphs are
also related to vertex degrees. Some of those results extend to non-regular graphs
(using essentially the same proofs) when the adjacency matrix is replaced with a
matrix that adjusts for the differences in vertex degrees. The “Laplacian matrix”
achieves this. A thorough treatment of the resulting “Lapacian eigenvalues”
appears in Chung [1997]; see also Grone–Merris–Sunder [1990], Mohar [1991,
1992], Mohar–Poljak [1993], Merris [1994, 1995], and Grone–Merris [1994].

15.3.29. Definition. The Laplacian matrix of a graph G is the matrix D − A,

where A is the adjacency matrix of G and D is the diagonal matrix whose
diagonal entry in row i is d G(vi). Let Q denote the Laplacian matrix of G.
The eigenvalues of Q form the Laplacian spectrum of G.
Section 15.3: Eigenvalues 777

Having used ë 1 , . . . , ë n (nonincreasing) for the ordinary eigenvalues, we use

Þ1 , . . . , Þn (nondecreasing) for the Laplacian eigenvalues. Since each row of Q
has sum 0, the constant vector 1n is a eigenvector, with eigenvalue 0. That is, 0
is for Laplacian spectra much as (G) is for ordinary spectra of regular graphs.

15.3.30. Theorem. For the Laplacian spectrum of an n-vertex graph G,

(a) The smallest Laplacian eigenvalue of G is 0, with multiplicity 1 if and
only if G is connected.
(b) If G is ¾-regular, then Þ is a Laplacian eigenvalue if and only if ¾ − Þ is
an ordinary eigenvalue of G, with the same multiplicity.
(c) If (0 , Þ2 , . . . , Þn) is the Laplacian spectrum of G (in nondecreasing order),
then (0 , n − Þn , . . . , n − Þ2 } is the Laplacian spectrum of G.
Proof: Exercise 29.

Any result involving the Laplacian eigenvalues of a ¾-regular graph can be

restated in terms of its ordinary eigenvalues by subtracting the Laplacian eigen-
values from ¾ , but many results are more natural using the Laplacian.
The Matrix Tree Theorem (Theorem 15.2.5) computes the number of span-
ning trees Ì(G) of a graph G as the common cofactor of each entry in the Laplacian
matrix. Hence it is not surprising that the Matrix Tree Theorem can be restated
in terms of the Laplacian eigenvalues. Recall the definition and properties of the
adjugate matrix from Definition 15.2.2 and Remark 15.2.3.

15.3.31. Lemma. For a graph G with Laplacian matrix Q, the number of span-
ning trees is det(J + Q)/n2 .
Proof: The Matrix Tree Theorem states Adj Q = Ì(G)J. Since the Laplacian of
K n is nI − J , Cayley ’s Formula yields Adj (nI − J) = nn−2 J. Since J 2 = nJ and
JQ = 0, we have (nI − J)(J + Q) = nQ. Thus
(nn−2 J)Adj (J + Q) = Adj (nI − J) Adj (J + Q) = Adj [(nI − J)(J + Q)]
= Adj (nQ) = nn−1 Adj Q.
Canceling nn−2 yields JAdj (J + Q) = nÌ(G)J. By Remark 15.2.3, multiply-
ing both sides on the right by (J + Q) yields J(det(J + Q)I) = nÌ(G)nJ. Both sides
are multiples of J, so the desired equality holds.

We now compute Ì(G) from the Laplacian eigenvalues.

15.3.32. Theorem. (Kelmans [1967]; see also Kelmans–Chelnokov [1974]) The

number of spanning trees in a graph whose Laplacian eigenvalues are
Þ1 , . . . , Þn (in nondecreasing order) is 1n ∏ni=2 Þi .
Proof: As mentioned in Remark 15.3.2(2), the determinant of a matrix is the
product of its eigenvalues. In light of Lemma 15.3.31, we seek the eigenvalues of
J + Q. Since (J + Q)1n = J1n + Q1n = n1n , we have n as an eigenvalue in place
of 0. Since J + Q is real and symmetric, we may take all other eigenvectors to be
orthogonal to 1n , just as they are for Q. Letting x be an eigenvector associated
with eigenvalue Þi of Q for i > 1, we have (J + Q)x = Jx + Qx = Qx = Þi x. Hence
the eigenvalues of J + Q are the eigenvalues of Q, except for replacing 0 with n.
By Lemma 15.3.31, we now have Ì(G) = det(J + Q)/n2 = 1n ∏i=2 Þi .
n
778 Chapter 15: Linear Algebra

15.3.33. Example. The graphs K n and K p ,p are regular, so by Theorem 15.3.30

the Laplacian eigenvalues are the vertex degree minus the ordinary eigenvalues.
For K n , all Laplacian eigenvalues other than the smallest equal n, and Theorem
15.3.32 reduces to Cayley ’s Formula, Ì(K n) = nn−2 . The Laplacian spectrum of
K p ,p has 0, one copy of 2p, and 2p − 2 copies of p, so Ì(K p ,p) = 2p
1
2p2p−1 = p2p−2 .

Other methods for counting spanning trees appear in Kelmans [1965/66] and
Hartsfield–Kelmans–Shen [1996].
The second smallest Laplacian eigenvalue has important structural aspects.
The larger it is, the more highly connected the graph is. We have seen that 2(G)
is nonzero if and only if G is connected; this is a special case of (G) ≥ 2(G)(Corol-
lary 15.3.36). Since it bounds the connectivity from below, the second smallest
Laplacian eigenvalue is called the algebraic connectivity.

15.3.34. Lemma. If G 1 and G 2 are edge-disjoint graphs with the same vertex
set, then 2(G 1) + 2(G 2) ≤ 2(G 1 ∪ G 2).
Proof: Let G = G 1 ∪ G 2 . With Q , Q1 , Q2 denoting the Laplacian matrices of the
three graphs, respectively, we have Q = Q1 + Q2 .
The Laplacian matrix is a real symmetric matrix. By the Spectral Theorem,
1 (G) is the minimum of x Qx over unit vectors. Its value is 0, using eigenvector
T

1. Since a real symmetric matrix has a full set of orthonormal eigenvectors, 2(G)
is the minimum of x T Qx over the set C of unit vectors orthogonal to 1.
Since x T (A + B)x = x T Ax + x T Bx and Q = Q1 + Q2 ,
2(G) = min(x T Q1 x + x T Q2 x) ≥ min x T Q1 x + min x T Q2 x = 2(G 1) + 2(G 2).
x∈ C x∈ C x∈ C

15.3.35. Theorem. If S is a set of vertices in a graph G such that | S| ≤ | V(G)|− 2,

then 2(G − S) ≥ 2(G) − | S| .
Proof: Let H = G − v, where v ∈ V(G). It suffices to show 2(H) ≥ 2(G) − 1.
Let G = H K 1 . Let Q and Q be the Laplacian matrices of H and G , respec-
tively. Let x be an eigenvector for Q that is associated with eigenvalue 2(H) and
satisfies x T 1 = 0. With G having n vertices, block matrix multiplication yields

x Q+I −1 ) ( x ) = (
Q ( ) = ( 2(H) + 1) ( x ) .
0 −1 T n−1 0 0

Thus 2(H) + 1 is a Laplacian eigenvalue of G , which yields 2(G ) ≤ 2(H) + 1.

Since all Laplacian eigenvalues are nonnegative, and G arises from G by adding
edges, Lemma 15.3.34 yields 2(G) ≤ 2(G ).

15.3.36. Corollary. (Fiedler [1973]) For a non-complete graph G, connectivity

is bounded below by algebraic connectivity. That is, (G) ≥ 2(G).

The bound fails for K n , since 2(K n) = n. It holds with equality for non-
complete cographs (Exercise 30). de Abreu [2007] surveyed results on 2(G).
Meanwhile, the algebraic connectivity also yields an important lower bound for
the density of edge cuts.
Section 15.3: Eigenvalues 779

15.3.37. Theorem. (Alon–Milman [1985]) Let G be an n-vertex graph. If S is a

nonempty proper subset of V(G), then
#### [S, S]#### Þ2(G)
# #
## ## ≥ n .
|S| ### S###
Proof: The claim is trivial if Þ2(G) = 0, so by Theorem 15.3.30 we may assume
that G is connected. With Laplacian Q, the terms of x T Qx include −2x i x j when
vi vj ∈ E(G) and x2i with coefficient d(vi). Hence

x T Qx = ∑ (x2i − 2x i x j + x2j ) = ∑ (x i − x j)2 .

vi vj ∈ E(G) vi vj ∈ E(G)

We choose x so that edges crossing the cut contribute positively. Let s = | S| , and
set x i = −(n − s) for i ∈ S and x i = s for i ∈ / S. The sum on the right becomes
# #
n2 #### [S, S]#### . Our choice of x yields ∑ x i = (n − s)s − s(n − s) = 0, so x · 1 = 0. Since
1 is an eigenvector for Þ1 (G), the Spectral Theorem yields xx TQx ≥ Þ2(G). Hence
T

x T Qx ≥ Þ2(G)x T x = Þ2(G) (s(n − s)2 + (n − s)s2 ) = Þ2(G)s(n − s)n.

Since x T Qx = n2 #### [S, S]#### , we have #### [S, S]#### ≥ Þ2(G)s(n − s)/n.
# # # #

15.3.38. Corollary. If G is an n-vertex graph with maximum degree ¾ , and S is

a nonempty vertex subset with | S| ≤ n/2, then at least (Þ2(G)/2 ¾) | S| vertices
of S have neighbors in S.
Proof: Theorem 15.3.37 yields ##### [S, S]##### ≥ Þ2(G) |S| ##### S #####/n. Each vertex of S re-
# #
ceives at most ¾ of these edges, so S has at least Þ2(G) |S| #### S####/(n¾) neighbors in
# #
S. Since #### S####/n ≥ 1/2, the claim follows.

When applied to ¾-regular graphs, these results can be stated using ordinary
eigenvalues by substituting ¾ − ë 2 for Þ2 . Corollary 15.3.38 states that an n-
vertex graph having maximum degree ¾ and second Laplacian eigenvalue Þ is an
(n , ¾ , c)-magnifier with c = Þ/(2 ¾), as defined below.

15.3.39.* Definition. An (n , ¾ , c)-magnifier is an n-vertex graph G such that

(G) ≤ ¾ and that #### N(S) ∩ S#### ≥ c · | S| for every S ⊆ V(G) with | S| ≤ n/2. An
# #
(n , ¾ , c)-expander is an X , Y-bigraph G with | X | = | Y | = n such that (G) ≤
¾ and that | N(S)| ≥ (1 + c(1 − | S| /n)) · | S| for every S ⊆ X with |S| ≤ n/2.

15.3.40.* Remark. Given an (n , ¾ , c)-magnifier G, construct an X , Y-bigraph

H by letting X and Y be copies of V(G), putting x i yj ∈ E(H) if and only if vi vj ∈
E(G), and adding x i yi for all i. The result is an (n , ¾ + 1 , c)-expander (Exercise
5). The expansion condition strengthens Hall’s Condition. Expanders appear in
the parallel sorting network of Ajtai–Komlós–Szemerédi [1983]. Walters [1996]
collects other definitions that have been used to measure expansion properties.
Probabilistic methods (Exercise 31) yield existence of large expanders with
bounded average degree (Pinsker [1973], Pippenger [1977], Chung [1978b]). Mar-
gulis [1973] built explicit examples algebraically (see also Gabber–Galil [1981]).
780 Chapter 15: Linear Algebra

Random graphs generally are good expanders, but measuring expansion is

hard. Tanner [1984] and Alon–Milman [1984, 1985] independently proved good
expansion properties for graphs whose two largest ordinary eigenvalues (or two
smallest Laplacian eigenvalues) differ greatly. Since eigenvalues are easy to ap-
proximate, this yields a good test for expansion.
For ¾-regular expanders, Alon–Milman [1984, 1985] showed c ≥ 23¾¾−−22 ëë22 .
Alon [1986] proved a partial converse: if a ¾-regular graph G is an (n , ¾ , c)-
magnifier, then the eigenvalue separation ¾ − ë 2 is at least 4+c2c2 . Explicit con-
2

structions of regular graphs are known with ë 1 − ë 2 nearly as large as possible.

√ second largest eigenvalue of a ¾-regular graph with diameter d is at least
The
2 ¾ − 1(1 −O(1/d))(see Nilli [1991]). Lubotzky–Phillips–Sarnak [1986] and Mar-
gulis [1988] constructed infinite families of (p + 1)-regular graphs, where
√ p is a
prime congruent to 1 mod 4, with second largest eigenvalue at most 2 p. See
Alon–Spencer [1992, pp. 119–125].

EXERCISES 15.3

15.3.1. (−) Let A¾ be the ¾ th power of the adjacency matrix of a graph G. Prove that the
value in position (i , j) of A¾ is the number of vi , vj -walks of length ¾ in G.
15.3.2. (−) Prove that if a graph G has s vertices with identical neighborhoods, then 0
is an eigenvalue of G with multiplicity at least s − 1. Prove that if G has s vertices with
identical closed neighborhoods, then −1 is an eigenvalue with multiplicity at least s − 1.
15.3.3. (−) Obtain the eigenvalues of G 2 in terms of the eigenvalues of G, where G 2 is
obtained from G by adding the edges xy such that d G(x , y) = 2.
15.3.4. (−) Let F be a cartesian product where any two nonadjacent vertices have exactly
two common neighbors. Prove that both factors are complete graphs.
15.3.5. (−) Let G be an (n , ¾ , c)-magnifier with vertices v1 , . . . , vn . Define an X , Y -bigraph
H with X = {x1 , . . . , x n} and Y = {y1 , . . . , yn} by xi yj ∈ E(H) if and only if i = j or vi vj ∈
E(G). Prove that H is an (n , ¾ + 1 , c)-expander.
15.3.6. Let ¾ be the number of -cycles in G. Let L¾ = ∑ ¾i and D¾ = ∑ d ¾i , summing
the  th powers of the eigenvalues and the vertex degrees, respectively. Obtain formulas
for 3 and 4 in terms of {L¾ } and {D¾ }.
n−1
15.3.7. (♦) Coefficients of the characteristic polynomial ∑i=0 ci n−i .
(a) Let H be the set of spanning subgraphs of a graph G such that every compo-
nent is an edge or a cycle. Let (H) and s(H) denote the number of components of H
and the number of components of H that are cycles, respectively. Prove that det A(G) =
∑ H∈H(−1)|V(H)|−¾(H) 2 s(H) . (Harary [1962b])
(b) Prove that ci = (−1)i ∑| S|=i det A(G[S]). (Hence c3 = −2t when G has t triangles.)
(c) Let Hi be the set of i-vertex subgraphs of a graph G whose components are edges
or cycles. Prove that ci = ∑ H ∈H (−1)¾(H) 2 s(H) . (Sachs [1967])
i

15.3.8. Write (G; ) as G . Let v and xy denote a fixed vertex and edge in G; let Z(v) and
Z(xy) denote the sets of cycles containing v or xy. Use Sach’s formula above to:
(a) Prove G = G −v − ∑u∈ N(v) G −v−u − 2 ∑ C∈ Z(v) G − V(C) .
(b) Prove G = G − x y − G − x− y − 2 ∑ C∈ Z(x y) G − V(C) .
(c) Obtain recurrences for the characteristic polynomials of paths and cycles.
Exercises for Section 15.3 781

15.3.9. Given that (G; x) = x8 − 24x6 − 64x5 − 48x4 , determine G.

15.3.10. (♦) Use Exercise 15.3.7 to prove that the coefficient of n−2¾ in the characteristic
polynomial of a tree is (−1)¾ m¾ , where m¾ is the number of matchings of size  . Construct
two nonisomorphic 8-vertex trees, both having characteristic polynomial 8 − 7 6 + 9 4 .
15.3.11. (+) Let T be a tree. Prove that (T) is the number of nonnegative eigenvalues of
T . (Cvetkovic–Doob–Sachs [1979, p. 233])
15.3.12. Prove that Cn has eigenvalues {2 cos(2  j/n): 0 ≤ j ≤ n − 1}.
15.3.13. Let  1 , . . . ,  m and 1 , . . . , n be the eigenvalues of G and H , respectively. Show
that the mn eigenvalues of the cartesian product G H are {  i + j : i ∈ [m] , j ∈ [n]}. Use
this to derive the spectrum of the -dimensional cube.
15.3.14.
√ Show that the eigenvalues of a graph with n vertices and m edges are bounded
by 2m(n − 1)/n.
15.3.15. For a graph G with adjacency matrix A, prove that the maximum of x T Ax over
vectors summing to 1 is 1 − 1/(G). Conclude Turán’s Theorem. (Motzkin–Straus [1965])
15.3.16. Prove that G is bipartite if G is connected and  max(G) = −  min(G).
15.3.17. Prove that a regular graph and its complement have the same eigenvectors. What
is the relationship between the corresponding eigenvalues?
15.3.18. Find the spectrum of the complete r-partite graph with m vertices in each part.
15.3.19. (♦) Let B1 , . . . , Bm be a biclique decomposition of K n with vertex set [n], where
B¾ has parts X ¾ and Y¾ . Let A¾ be the 0,1-matrix with 1 in position (i , j) if and only if
i ∈ X ¾ and j ∈ Y¾ . Let S = ∑¾=1 A¾ . Observe that S + S T = J − I. Prove that every n-by-n
m

matrix satisfying this equation has rank at least n − 1. Since the rank of the sum of two
matrices is at most the sum of their ranks, conclude rank S ≤ m and hence m ≥ n − 1.
15.3.20. (♦) For a graph G, a squashed cube embedding in dimension t encodes each
vertex as a vector in {0 , 1 , ∗}t such that d G(u , v) is the number of coordinates where one
of {u , v} has 0 and the other has 1. (Graham–Pollak [1972])
(a) Prove that every n-vertex graph G has a squashed cube embedding. (Comment:
Winkler [1983] proved qdim (G) ≤ n − 1, where qdim (G) is the least dimension needed.)
(b) Use a bijection between biclique decompositions and squashed-cube embeddings of
K n to show that qdim K n is the least size of a biclique decomposition.
15.3.21. (♦) Let G be a graph with eigenvalues  1 , . . . ,  n (in nonincreasing order).
(a) Let H be an induced subgraph of G with eigenvalues 1 , . . . , t . From Theorem
15.3.9 (Interlacing), conclude that i ≤  i for 1 ≤ i ≤ t and that  2 < 0 implies G = K n .
(b) Prove that (G) ≤ n − max{p , q}, where G has p positive and q negative eigenval-
ues (including repetition). (Comment: This approach proves the Erdős–Ko–Rado Theorem
from the eigenvalues of the Kneser graph K(n , ). See Godsil–Meagher [2016, 2.9–10].)
15.3.22. Given a real symmetric matrix partitioned as M = ( QPT QR) with P, R square, a
lemma in linear algebra yields  max(M) +  min(M) ≤  max(P) +  max(R).
(a) Let A be a real symmetric matrix partitioned into t2 submatrices Ai , j such that
each Ai ,i is square. Prove that  max(A) + (t − 1) min(A) ≤ ∑i=1  max(Ai ,i).
m

(b) Prove (G) ≥ 1 + − ë (G) when G is nontrivial. (Hoffman [1970])

ë max (G)
min
(c) Prove  max(G) + 3  min(G) ≤ 0 when G is planar (use the Four Color Theorem).

15.3.23. (♦) Let G be strongly regular with n vertices and parameters ( ,  , ). Prove
that G is strongly regular, and compute its parameters ( ,  ,  ).
782 Chapter 15: Linear Algebra

15.3.24. (♦) Let G be a triangle-free n-vertex graph in which any two nonadjacent ver-
tices have exactly two common neighbors.
(a) Prove that G is ¾-regular, where ¾ satisfies n = 1 + (¾+2 1 ) .
(b) Prove that ¾ is one more than the square of an integer not divisible by 4.
(c) Construct such graphs for ¾ ∈ {1 , 2 , 5}. Let H be the graph for ¾ = 5. Prove
that H − N[v] is the Petersen graph (for v ∈ V(H)) and that H is obtained from the 4-
dimensional hypercube by adding edges joining antipodal vertices. (Comment: A realiza-
tion for ¾ = 10 is known using combinatorial designs.)
15.3.25. Let G be a strongly regular graph containing K n+1 + K n ,n as an induced subgraph.
Use the Interlacing Theorem and the eigenvalue theorem for strongly regular graphs to
prove that G must have more than n2 vertices. (Comment: Vu proved this. Fon-Der-Flaass
[2002] proved that every graph with n vertices is an induced subgraph of a strongly regu-
lar graph with at most 4n2 vertices.)
15.3.26. (♦) Prove that the Petersen graph is strongly regular and determine its spec-
trum. Apply the spectrum to show that K10 does not decompose into three copies of the
Petersen graph. (Hint: Use the spectrum to prove that two copies of the Petersen matrix
have a common eigenvector other than the constant vector.) (Schwenk [1983])
15.3.27. The subconstituents of a graph G are the induced subgraphs of the form G[U]
where v ∈ V(G) and U = N(v) or U = N[v]. Vince [1989] defined G to be superregu-
lar if G has no vertices or if G is regular and every subconstituent of G is superregular.
Let S be the class consisting of {aK b : a , b ≥ 0} (disjoint unions of isomorphic cliques),
{K m K m : m ≥ 0}, C5 , and their complements.
(a) Prove that every graph in S is superregular and that every disconnected super-
regular graph is in S. (Comment: S is the set of all superregular graphs, but the full
inductive proof requires several pages. (Maddox [1996], West [1996])
(b) Prove that every superregular graph is strongly regular.
15.3.28. A ¾-regular connected graph of diameter d is distance-regular if there exist
c1 , . . . , cd and b0 , . . . , bd−1 such that whenever d(x , y) = i, the neighborhood of y contains
exactly ci vertices with distance i − 1 from x and bi vertices with distance i + 1 from x.
(a) Prove that the ¾-dimensional hypercube Q¾ is distance-regular.
(b) The Odd graph O¾ is the disjointness graph of the ¾-element subsets of [2 ¾ + 1]
(the Petersen graph is O2). Prove that O¾ is distance-regular.
(c) Given a distance-regular graph G, let Ai be the matrix with 1 in position (r, s) if
d(vi , vj ) = i and 0 otherwise. Note that ∑i=0 Ai = J. Prove that Ai is a polynomial of
d

degree i in the adjacency matrix, that (A − ¾ I)(∑ i=0 Ai) = 0, and that A0 , . . . , A d is a basis
d

for the space of polynomials in the adjacency matrix.

15.3.29. (♦) Prove the properties of the Laplacian stated in Theorem 15.3.30. Conclude
that the largest Laplacian eigenvalue is | V(G)| when G is disconnected. (Kelmans [1967])
15.3.30. (♦) Let G be a cograph (Exercise 8.3.58), so G and G are not both connected.
(a) Prove that the Laplacian eigenvalues of G are all integers. (Merris [1998])
(b) Prove that (G) = 2(G) (see Corollary 15.3.36). (Abrishami [2019])
15.3.31. (+) Expanders of linear size.
(a) Prove (| S| ≤ l) ≤ (nl)(l/n)¾t when S is the union of  random t-subsets of [n].
(b) For  < 1, prove that  ∈  exists such that , when n is large, there exists G ⊆
K n ,n with (G) ≤  such that | N(S)| ≥  | S| whenever | S| ≤  n. Conclude that (n ,  , c)-
expanders exist when n is large.
Chapter 16

Geometry and Topology

Many questions about placing objects in the plane with requirements on rel-
ative position are combinatorial in nature, such as the characterization of planar
graphs. Questions involving specific locations of points become more geometric.
We consider a few such questions and later discuss relationships between volume
computations and enumeration. We only scratch the surface of these topics. For
books on connections between combinatorics and geometry, see Pach–Agarwal
[1995], Penrose [2003], Felsner [2004], and Pach–Sharir [2009].

16.1. Graph Drawings

We have observed that every planar graph has a straight-line embedding
(Fáry ’s Theorem). For applications in computer science, we seek a straight-line
embedding with the vertices at the integer points in a small grid. The geometric
techniques involved in doing this also lead to a characterization of planar graphs
in terms of poset dimension. Subsequently, we consider drawings of nonplanar
graphs in the plane; the aim is to minimize the number of crossings of edges.

EMBEDDINGS ON GRIDS

Let G be a plane graph with n vertices, where n ≥ 3. Schnyder [1990] proved

that G has a straight-line embedding with its vertices at points of {0 , . . . , n − 2} ×
{0 , . . . , n − 2}. Furthermore, this embedding is computable in linear time from
an embedding of G. de Fraysseix–Pach–Pollack [1990] obtained embeddings in
an (n − 2)-by-(2n − 4) grid. Our lemmas follow Schnyder [1990], refining Schny-
der [1989]. Another approach to embeddings was pioneered by Tutte [1960] (the
“rubber band method”), extended in Ribó Mor–Rote–Schulz [2011].
Consider the triangle in space with corners at {(1 , 0 , 0) , (0 , 1 , 0) , (0 , 0 , 1)}.
The coordinates of any point x in the triangle sum to 1. The coordinates express
x as a convex combination of the three corners. Placing the triangle in any plane,
these coordinates become the barycentric coordinates of x in terms of the three
corners. The points with a fixed value of one coordinate lie on a line parallel to
the side where that coordinate is 0.

783
784 Chapter 16: Geometry and Topology

16.1.1. Definition. A barycentric representation of G is an injection º from

V(G) to 3 such that (1) the coordinates of º (v) are nonnegative and sum to
1 for all v ∈ V(G), and (2) if xy ∈ E(G) and ∈ V(G) − {x , y}, then some
coordinate of  ( ) is greater than the corresponding coordinates for  (x) and
 (y). For all v ∈ V(G), we write  (v) = (v1 , v2 , v3).

16.1.2. Lemma. If  is a barycentric representation of G, and  ,  ,  are non-

collinear points in the plane, then letting (v) = v1  + v2  + v3  yields a
straight-line embedding of G in the plane.
Proof: By definition the images are distinct. Suppose that xy and uv are edges
having four distinct endpoints. Since each vertex does not belong to the other set,
there exist indices i , j ,  , l ∈ {1 , 2 , 3} such that
x i > max{ui , vi}, u¾ > max{x¾ , y¾ },
yj > max{uj , vj }, vl > max{x l , y l }.
These inequalities require {i , j} ∩ { , l} = ∅, but there are only three possible
indices. Hence i = j or  = l. By symmetry, suppose i = j . Now there is a value
t with min{x i , yi} > t > max{ui , vi}, and the line of points with i-coordinate t
(parallel to the side of the triangle  not through the ith corner) separates (x)
and (y) from (u) and (v). Hence (x)(y) and (u)(v) do not intersect.
We must also consider incident edges. Let xy and y be embedded with one
containing the other, say with xy containing y . In this case lies on the segment
joining x and y, so  ( ) is a convex combination of  (x) and  (y), contradicting
the existence of a coordinate i such that i > max{x i , yi}.

When the coordinates are rational, multiplying by a constant m yields inte-

ger triples summing to m. Letting { ,  , } = {(1 , 0) , (0 , 1) , (0 , 0)}, we obtain a
straight-line plane embedding of G on integer points with coordinates in [0 , m].
Hence we seek a rational barycentric representation of G with small denomi-
nator. We restrict to triangulations, since a straight-line embedding of a triangu-
lation with vertices at integer points yields such an embedding of every subgraph.
Call the bounded faces cells. By Euler ’s Formula, an n-vertex triangulation has
2n − 4 regions and 2n − 5 cells. To obtain coordinates for vertex v, we will split
the triangulation into three regions meeting at v and count the cells in the three
regions. This will yield an embedding using integer coordinates in the range
[0 , 2n − 5]. A refinement to grid-length n − 2 counts vertices instead of faces.
The combinatorial structure used to obtain the desired partitions of the cells
is a special labeling of the angles of a triangulation. All arithmetic with labels
and colors in this discussion will be modulo 3. Since the boundary of each cell is a
3-cycle, we can refer to the cells as triangles and the incidences between vertices
and cells as angles (even without straight-line embeddings). The labeling idea
was extended to 3-connected planar maps in Felsner [2001] and Miller [2002].

16.1.3. Definition. A Schnyder labeling of a triangulation G is a labeling of

the angles in cells of G with the labels 1 , 2 , 3 such that
(1) Each cell has its angles labeled 1 , 2 , 3 in clockwise order.
(2) Each interior vertex has angles with each label, appearing in the clock-
wise order as all 1s then all 2s then all 3s.
Section 16.1: Graph Drawings 785

1 1 1

2• 3 2
•3
1 3 1 1

3 2 2
2•3
3 1 2
• 3 2 •

16.1.4. Remark. Edge-coloring of Schnyder labelings. A Schnyder labeling leads

to a natural labeling and orientation of the internal edges. Each internal ver-
tex has three incident edges where the two adjoining angles have different labels.
Letting {i , j , ¾} = {1 , 2 , 3}, give color i to the edge between the angles labeled j
and ¾ at vertex x, and orient this edge away from x.
An interior edge now has color i if and only if both angles at its head have
color i. Hence the edges at an internal vertex appear clockwise in this order: one
departing edge with color 1, all entering edges having color 3, one departing edge
with color 2, all entering edges having color 1, one departing edge with color 3,
all entering edges having color 2. A vertex of degree 3 has no entering edges.
Let T i be the set of internal edges colored i. We will show that T i is an in-
tree rooted at an external vertex. The paths leaving v in T1 , T2 , T3 will define
the regions where we count cells to obtain a barycentric representation.

•
j

y• i ¾ •x
i i j

¾
•

16.1.5. Lemma. In a Schnyder labeling, the external vertices can be named

w1 , w2 , w3 (clockwise) so that all angles at wi have label i.
Proof: The 3n − 9 internal edges are oriented by the procedure above. Since each
internal vertex is the tail of three of these edges, all internal edges incident to ex-
ternal vertices enter the external vertices. Since the angles at the head of an edge
have the same label, the labels at a fixed external vertex are the same. Since each
external edge belongs to a cell using two external vertices, the angles at distinct
external vertices have different labels, appearing in the order claimed.

In order to facilitate inductive proofs, we seek an internal edge e whose end-

points have only two common neighbors, which are the other vertices on the cells
bounded by e. Contracting e and converting each of the two resulting multiedges
to single edges yields a smaller simple triangulation.
786 Chapter 16: Geometry and Topology

16.1.6. Definition. An internal edge of a triangulation is contractible if its

endpoints have only two common neighbors.

16.1.7. Lemma. (K ampen [1976]) If a is an external vertex of a triangulation G

other than K 3 , then a has an internal neighbor u such that au is contractible.
Proof: Let x0 , x1 , . . . , x t be the neighbors of a in clockwise order, where x0 and
x t are the other external vertices. Since G is a triangulation, these vertices form
a path P. Among the chords of P , choose x i x j with i < j to minimize j − i; such
a chord exists because x0 and x t are adjacent. Let u be any x¾ with i < ¾ < j .
Planarity and the minimality of j − i guarantee that the only common neighbors
of a and u are x ¾−1 and x¾+1 .

16.1.8. Theorem. Every triangulation has a Schnyder labeling.

Proof: Let G be an n-vertex triangulation. We use induction on n; the claim
is trivial for n = 3. For n > 3, let a be an external vertex. By Lemma 16.1.7,
there is a contractible edge au. Let G be the smaller triangulation obtained by
contracting au into a. Let N(u) consist of a and x1 , . . . , x t in clockwise order.
By the induction hypothesis, G has a Schnyder labeling. By Lemma 16.1.5
and cyclic symmetry, we may assume that all angles at a have label 1. In G, we
use the same angle labels as in G , except for the new angles involving u. The
angles at u not bordered by au get label 1, and the two angles bordered by au get
labels 2 and 3, in clockwise order.

x1 x 2 x1 x 2
3• • 3• •
2 32 2 32
2
3 • x3 3 • x3
1
3 • 11 2 ↔ 2
u2 3 x4 3• x4
1 3 • 11
1 2 1 2
1 1
• 1 1
a a•
G G
The labeling of a cell in G is inherited from the labeling of G , with u replac-
ing a in the triangles at a having no external edges, and the two triangles involv-
ing ua labeled as shown above. The requirements of consecutivity and clockwise
order at internal vertices are also inherited from the labeling of G , except for
being explicitly enforced at u.

Properties of Schnyder labelings can now be proved inductively if we show

that every Schnyder labeling arises as in Theorem 16.1.8 from a labeling of K 3 .

16.1.9. Lemma. Given a Schnyder labeling of a triangulation G other than K 3 ,

let a be the external vertex wi with angles labeled i. There is a contractible
edge au such that the internal angles at u all have label i except for the two
angles in triangles involving au.
Proof: By symmetry, we may assume i = 1. Let y0 , . . . , yt be the neighbors of
a in clockwise order, where y0 and yt are the other external vertices. Since all
Section 16.1: Graph Drawings 787

angles at a have label 1, the angles at yr and yr+1 in their triangle with a have
the labels 2 and 3, respectively. Therefore, in the edge-coloring and orientation
of Remark 16.1.4, the edge yr yr+1 has color 2 if oriented from yr+1 to yr and color
3 if oriented from yr to yr+1 .
Since y0 and yt are external vertices, y0 y1 has color 2 and yt−1 yt has color
3. Hence there exists ¾ such that y¾−1 y¾ has color 2 and y¾ y¾+1 has color 3. The
angles between them (not involving a) have label 1, and the edges between them
have color 1 and enter y¾ . The only edges leaving y¾ go to a, y¾−1 , and y¾+1 .
If a and y¾ have a common neighbor outside {y¾−1 , y¾+1 }, then we have
shown that y¾ has color 1 and enters y¾ . However, a has color 1 and enters a.
Since no vertex has two exiting edges colored 1, there is no such vertex . Hence
y¾ is the desired vertex u.

y ¾ −1
• •
1
2 1 •
1 •
y¾ 3 •y¾+1
1
1
a•

16.1.10. Theorem. (Uniform Angle Lemma) Given a cycle C in a Schnyder la-

beling of G, and i ∈ {1 , 2 , 3}, there is an i-uniform vertex x i on C (that is, the
angles at x i inside C all have label i). Furthermore, such vertices x1 , x2 , x3
can be found occurring in clockwise order on C.
Proof: (Lovelace–K ündgen) Consider a Schnyder labeling of a triangulation G
with n vertices. We use induction on n. The claim is trivial if C visits all three
external vertices, which includes the base case n = 3. Otherwise, let a be an ex-
ternal vertex not in C. By symmetry, we may assume a = w1 . By Lemma 16.1.9,
a has a neighbor u such that contracting au into a yields a Schnyder labeling of
a smaller triangulation G from which the labeling arises as in Theorem 16.1.8.
If u ∈/ V(C), then C has the same interior in G and G , and the uniform
vertices for C from applying the induction hypothesis to G work also for C in G.
If C visits u, then C becomes a cycle C through a in G , and u behaves like
y¾ in the figure above. In particular, since C avoids a, the vertex u is 1-uniform
in C, just as a is 1-uniform in C . The rest of the interior remains the same, so
the triple of vertices obtained by applying the induction hypothesis to C in G
works also for C in G, with a replaced by u if a is the 1-uniform vertex of the
triple found in C .

16.1.11. Theorem. In the edge-coloring and orientation for a Schnyder labeling,

the edges of color i form a tree T i of paths directed to the external vertex wi .
Also, for each internal vertex v, the paths from v to w1 , w2 , w3 in T1 , T2 , T3 ,
respectively, pairwise share only v.
Proof: Each internal vertex has outdegree 1 in T i ; also wi is a sink in T i , and
the other external vertices have no incident edges of color i in T i . It thus suffices
788 Chapter 16: Geometry and Topology

to show that as a digraph T i has no cycle. This follows from the Uniform Angle
Lemma. A cycle in color i would have an inside angle of label i at the head of each
edge and an inside angle of another color at the tail of each edge, so it would have
no uniform angle in any label.
For the second statement, let {i , j , ¾} = {1 , 2 , 3}. If T i and T j each contain
v , u-paths Pi and P j , then Pi ∪ P j is a cycle C in G. However, along Pi inside C
is an angle of color i at the head of each edge and an angle of another color at the
tail of each edge. The same holds for P j with color j . Hence the only candidates
for uniform vertices are v and u, but the Uniform Angle Lemma requires three
such vertices on C.

Let Pi(v) denote the path from v to the root in the tree T i . Since no two of
P1 (v), P2(v), and P3(v) have a common internal vertex, they establish three re-
gions associated with v, as on the left below. Let Ri(v) denote the region (with
boundary) enclosed by Pi+1 (v), Pi+2 (v), and the external edge not containing wi .
As shown on the right, regions for distinct vertices can be ordered by inclusion;
we prove this and use it to obtain the barycentric representation.

w1
• •

1
1
v
v 3 • 2
R2(v) • R3(v) • •
3 • 2
x y
3 R (v) 2 u
1
w3• •w2 • •

16.1.12. Lemma. If u and v are distinct internal vertices in a Schnyder labeling,

and u ∈ Ri(v), then Ri(u) is properly contained in Ri(v).
Proof: By symmetry, we may assume i = 1. We prove first that P2(u) is confined
to R1 (v). Since T2 is an in-tree, when P2 (u) intersects P2(v) it thereafter follows
P2(v). Hence if P2(u) is not confined to R1(v), then some edge e of T2 departs R1 (v)
from some vertex w of P3(v). This requires the edge with label 2 that leaves w to
be clockwise after the edge with label 3 leaving w and before an edge with label
3 entering w (see figure above), which contradicts Remark 16.1.4. A symmetric
argument shows that P3(u) is confined to R1(v).
With these paths confined to R1 (v), the region bounded by P2 (u), P3(u), and
the external edge w2 w3 is contained in R1(v). Since u
= v, the two regions do not
have the same boundary, so R1 (u) omits some cell contained in R1 (v).

16.1.13. Theorem. Consider a Schnyder labeling of an n-vertex triangulation G.

For each internal vertex v of G and each i ∈ {1 , 2 , 3}, let vi be the number of
cells in Ri(v). For the external vertex v at which the angles have label j , let
vj = 2n − 5, and let the other coordinates of v be 0. Assigning to each vertex
v the triple 2n1−5 (v1 , v2 , v3) yields a barycentric representation of G.
Section 16.1: Graph Drawings 789

Proof: By construction, the coordinates of a vertex are nonnegative and sum to 1.

We must also check that if xy ∈ E(G) and ∈ / {x , y}, then there exists ∈ {1 , 2 , 3}
such that ¾ > max{x¾ , y¾ }. This holds when  is the external vertex w¾ since all
other vertices have less 1 in coordinate . If  is internal, then x , y ∈ R¾ () for
some , since the edge xy cannot cross any P j (). Again ¾ > max{x ¾ , y¾ }, since
any internal vertex other than  in R¾ () has smaller value than  in coordinate
, and the external vertices in R¾ () have 0 in that coordinate.

Theorem 16.1.13 completes the proof that every planar graph with n vertices
has a straight-line embedding with the vertices at grid points in the triangle with
corners (2n − 5 , 0), (0 , 2n − 5), and (0 , 0).
Using a similar approach, the proof can be modified to obtain a more compact
embedding using only the points in a grid with side-length n − 1. The trick is to
count vertices instead of regions, but we must be more careful about the order
relation, and we need a more general notion of barycentric representation.
Let v i be the number of vertices in Ri(v) − Pi−1 (v), so v 1 + v 2 + v 3 = n − 1.
When v is the jth external vertex, Rj (v) contains all the vertices, P j (v) contains
one vertex, and P j +1 (v) contains two vertices. Hence we use the triples (n− 2 , 1 , 0),
(0 , n − 2 , 1), and (1 , 0 , n − 2) for these vertices. Note that again the coordinates
sum to n − 1.
Below we list the steps of converting these triples to a straight-line embed-
ding in the grid with side-length n − 1, leaving verification of the details as exer-
cises. The arguments are like those using the count of cells.

16.1.14. Definition. In the lexicographic order on vectors, x < y if x has

smaller value than y in the first coordinate where the vectors differ. A weak
barycentric representation of G is an injection  : V(G) → 3 such that
(1) the coordinates of  (v) = (v1 , v2 , v3) are nonnegative and sum to 1 for
all v ∈ V(G), and (2) if xy ∈ E(G) and  ∈ V(G) − {x , y}, then there ex-
ists ∈ {1 , 2 , 3} such that (x¾ , x¾+1) and (y¾ , y¾+1) are less than (¾ , ¾+1) in
lexicographic order (with indices taken modulo 3).

16.1.15. Lemma. If  is a weak barycentric representation of G, with  (v) =

(v1 , v2 , v3) for all v ∈ V(G), and  ,  ,  are noncollinear points in the plane,
then placing each vertex v at the point (v) given by v1  + v2  + v3  yields a
straight-line embedding of G in the plane.

16.1.16. Lemma. If u and v are distinct vertices in a Schnyder labeling and u ∈

Ri(v), then (u i , u i+1) < (v i , v i+1) in lexicographic order.

16.1.17. Theorem. Given a Schnyder labeling of an n-vertex triangulation G,

the mapping that sends v to n−1 1 (v 1 , v 2 , v 3) is a weak barycentric representa-
tion of G, and the map that sends v to (v 1 , v 2) is a straight-line embedding of
G on the grid with side-length n − 2.

16.1.18. Example. For K 4 and for the octahedron labeled as at the start of this
subsection, the resulting embeddings appear below, with the image of the jth
external vertex labeled as j .
790 Chapter 16: Geometry and Topology

2• .

. . .

2• . . • • .

. • •1 . . • . •1
. • . . • . .
3 3

Finally, these results about Schnyder labelings also yield a characterization

of planar graphs in terms of dimension of posets.

16.1.19. Definition. The incidence poset P(G) of a graph G is the bipartite

poset under containment whose minimal elements are the vertices of G and
whose maximal elements are the pairs of adjacent vertices.

For an n-vertex graph G, the poset P(G) is a subposet of 2 n using only 1-sets
and 2-sets. Thus dim P(G) ≤ lg lg n + 12 lg lg lg n + O(1) (Corollary 12.3.35).

16.1.20. Theorem. (Schnyder [1989]) A graph G is planar if and only if

dim P(G) ≤ 3.
Proof: By Lemma 12.3.28, dim P(G) is the minimum number of linear order-
ings of V(G) such that for each vertex x and edge y with x ∈ / {y , }, both y and
appear before x in some ordering. Theorem 16.1.13 states that when G is pla-
nar, three linear orderings obtained by extending the coordinate rankings defined
there have the desired property.
Conversely, suppose that three such linear orderings exist. Assigning each
vertex the triple of its heights on them produces a barycentric representation,
by Definition 16.1.1. By Lemma 16.1.2, a barycentric representation of a graph
yields an embedding of it in the plane.

For further work on dimension of posets associated with planar graphs,

see Brightwell–Trotter [1993, 1997], Felsner [2003], Felsner–Li–Trotter [2010],
Barrera-Cruz–Haxell [2011], and Felsner [2014].

CROSSING NUMBER

Every drawing of a nonplanar graph in the plane has edge crossings. A nat-
ural objective is to minimize the number of crossings. There are several related
parameters, but for now we focus on the most common model. Recall that a cross-
ing is a common internal point of two edges.
Section 16.1: Graph Drawings 791

16.1.21. Definition. A drawing of G is proper if no vertex is an interior point on

an edge, no two edges are tangent, no two incident edges cross, and no point
is an interior point of more than two edges. The crossing number cr(G) of
a graph G is the minimum number of crossings in a proper drawing of G.

The notation cr(G) is the modern choice; other notations have been used but
are now less prevalent. We observed in Section 9.1 that all graphs have proper
drawings in the plane; we henceforth consider only proper drawings.
Crossing numbers of small graphs can often be found by extracting maximal
planar subgraphs.

16.1.22. Example. cr(K 6) = 3 and cr(K 3 ,2 ,2) = 2. If H with ¾ edges is a maximal

plane subgraph of a drawing of G in the plane, then every edge of G not in H
crosses some edge of H , so the drawing has at least | E(G)| − ¾ crossings. If G has
n vertices, then ¾ ≤ 3n − 6. If also G has no triangles, then ¾ ≤ 2n − 4.
Planar 6-vertex graphs have at most 12 edges, but K 6 has 15. Hence cr(K 6) ≥
3, and the drawing on the left below proves equality.
Since K 3 ,2 ,2 has 16 edges, and planar 7-vertex graphs have at most 15 edges,
cr(K 3 ,2 ,2) ≥ 1. The drawing on the right below has two crossings. To strengthen
the lower bound, observe that K 3 ,2 ,2 contains K 3 ,4 . Since K 3 ,4 is triangle-free,
its planar subgraphs have at most 2 · 7 − 4 edges, which equals 10, and hence
cr(K 3 ,4) ≥ 2. Every drawing of K 3 ,2 ,2 contains a drawing of K 3 ,4 , so cr(K 3 ,2 ,2) ≥
cr(K 3 ,4) ≥ 2.

•
• •
•
• • •
• • • •
• •

16.1.23. Proposition. Let G be an n-vertex graph with m edges. If ¾ is the

maximum number of edges in a planar subgraph of G, then cr(G) ≥ m − ¾ .
Furthermore, cr(G) ≥ m
2¾ − 2 .
2
m

Proof: Given a drawing of G in the plane, let H be a maximal subdrawing with-

out crossings. Every edge not in H crosses at least one edge in H ; otherwise, it
could be added to H. Since H has at most ¾ edges, we have at least m − ¾ crossings
between edges of H and edges of G − E(H).
After discarding E(H), we have at least m − ¾ edges remaining. The same
argument yields at least (m − ¾) − ¾ crossings in the drawing of the remaining
graph. Iterating the argument yields at least ∑ i=1 (m − i¾) crossings, where t =
t

⌊ m/¾ ⌋ . The value of the sum is mt − ¾ t(t + 1)/2.

We now write m = t ¾ + r, where 0 ≤ r ≤ ¾ − 1. Letting t = (m − r)/¾ in the
r(¾ −r)
value of the sum, we simplify to obtain cr(G) ≥ m 2¾ − 2 + 2¾ .
2
m
792 Chapter 16: Geometry and Topology

The bound m − ¾ in Proposition 16.1.23 is useful when G is sparse. The com-

putation can be iterated, but for dense graphs the result is too weak. Consider
K n , for example. Proposition 16.1.23 yields cr(K n) ≥ 24 n + O(n2), but the growth
1 3

of cr(K n) is quartic, not cubic. The crossing number cannot exceed (4n), since we
can place the vertices on the circumference of a circle and draw chords. In this
drawing of K n , each set of four vertices contributes one crossing.
In fact, this is the worst straight-line drawing, since four vertices never pro-
duce more than one crossing in such a drawing. How much can be saved by a
better drawing? We give a recursive lower bound.

16.1.24. Lemma. If a graph G contains a copies of a subgraph H , and each cross-

ing in a drawing of G appears in at most b copies of H , then cr(G) ≥ ab cr(H).
Proof: Consider a drawing of G with cr(G) crossings. Each copy of H in G con-
tributes at least cr(H) crossings to the drawing. Each crossing in G is counted at
most b times. Hence bcr(G) ≥ acr(H).

16.1.25. Example. We apply Lemma 16.1.24 with G = K n and H = K n−1 . Now

a = n and b = n − 4, so cr(K n) ≥ n−n 4 cr(K n−1). Dividing by (4n) yields
cr(K n) cr(K n−1)
≥ n−1 .
(n4) ( 4 )

Since always cr(K n)/(n4) < 1, the recursive inequality implies that cr(K n)/(n4) has
a limit as n → ∞. However, the value of that limit is unknown.

16.1.26. Theorem. (Guy [1972]) 1 4

80 n + O(n3) ≤ cr(K n) ≤ 1 4
64 n + O(n3).
Proof: Lemma 16.1.24 and Example 16.1.25 yield ≥ =
cr(K n) 15 (n4) 120
1
n4 + O(n3)
for n ≥ 5. The denominator of the quartic term can be improved from 120 to 80
by using copies of K 6 ,n−6 , which has crossing number 6⌊ n−2 6 ⌋ ⌊ n−2 7 ⌋ (Exercise 15b).
A better drawing lowers the upper bound from (4n) to 64 n + O(n3). Consider
1 4

n = 2 ¾ . View the sphere as a tin can. Place ¾ vertices on the top of the can and ¾
vertices on the bottom, drawing chords on the top and bottom for those ¾-cliques.
The edges from top to bottom fall into ¾ natural classes. The “class number ”
is the circular separation between the top and bottom endpoints, ranging from
⌈ −¾2+1 ⌉ to ⌈ ¾−2 1 ⌉ . We draw these edges to wind around the can as little as possible
in passing from top to bottom, so edges in the same class don’t cross. We now
twist the can to make the class displacements run from 1 to ¾ . This makes the
crossings easier to count but does not change the pairs of edges that cross.
• • •
x• y •
• • •

• • •
• •w
• • •
a
Crossings on the side of the can involve two top vertices and two bottom ver-
tices. For top vertices x , y and bottom vertices , w, where x has smaller posi-
tive displacement than xw, edges xw and y cross if and only if the displacements
Section 16.1: Graph Drawings 793

from x to y , , w are distinct positive values in increasing order. (This holds for
x , y , , w as shown, but not for x , y , , a, since ya winds around the can.) Hence
there are ¾ (¾3) crossings on the side of the can, so
cr(K n) ≤ 2(¾4) + ¾ (¾3) = 1 4
64 n + O(n3).

16.1.27. Example. For cr(K m ,n), a naive drawing puts the m vertices of one part
on one side of a channel and the n vertices of the other part on the other side,
with all edges drawn straight across. This produces (n2)(m )
2 crossings, but we can
do much better. Place the vertices along two perpendicular axes. Put the ver-
tices of the m-vertex part along the x-axis, with ⌈ m/2⌉ vertices on the positive
side and ⌊ m/2⌋ on the negative side. Similarly split the n-vertex part between
the positive and negative portions of the y-axis. Summing the four types of cross-
ings generated by adding straight-line segments for the edges yields cr(K m ,n) ≤
⌊m2
⌋ ⌊ m2−1 ⌋ ⌊ n2 ⌋ ⌊ n−2 1 ⌋ (Zarankiewicz [1954]).
The construction is conjectured optimal (Guy [1969] tells the history). K leit-
man [1970] proved this for min{n , m} ≤ 6. Aided by a computer search, Woodall
[1993] brought the smallest unknown cases to K 7 ,11 and K 9 ,9 . Using K leitman’s
result, Guy [1970] showed cr(K m ,n) ≥ m(m5−1) ⌊ n2 ⌋ ⌊ n−2 1 ⌋ (Exercise 15), not far from
the upper bound.

For dense graphs, the inductive argument of Theorem 16.1.26 yields a gen-
eral lower bound for crossing number conjectured in Erdős–Guy [1973]. There
is also an elegant probabilistic proof, which we present here. Stronger results
appear in Pach–Tóth [1997].

16.1.28. Theorem. (Ajtai-Chvátal-Newborn-Szemerédi [1982], Leighton [1983])

Let G be an n-vertex graph with m edges. If m ≥ 4n, then cr(G) ≥ 64
1
m3/n2 .
Proof: Consider a drawing of G with cr(G) crossings. Take a random induced
sub-drawing H by including each vertex independently with probability p. With
n and m denoting the order and size of H , we have (n ) = pn and (m ) = p2 m.
Let Y be the number of edge crossings in H. Since all four endpoints of two edges
must appear for a crossing in G to appear in H , we have (Y) = p4 cr(G).
Always Y ≥ m − (3n − 6), by the argument in Proposition 16.1.23. Lin-
earity of expectation yields p4 cr(G) = (Y) ≥ p2 m − 3pn, which simplifies to
3n + p3 cr(G) − pm > 0. Let p = 4n/m, which is allowable since m ≥ 4n. Now
3n + 64n3/m3 cr(G) > 4n, which yields the desired bound.

16.1.29. Example. The order of magnitude for the lower bound in Theorem
16.1.28 is best possible. Consider G = 2m n2
K 2m/n , where 2m is a multiple of n.
The total number of vertices is n, and the total number of edges is asymptotic to
( ) , which equals m. Since cr(K r) ≤ 64
n2 1 2m 2
r , we have cr(G) ≤ 2m ( ) = 18 m
n2 1 2m 4 3
1 4
2m 2 n 64 n n2
.
This is within a constant factor of the lower bound from Theorem 16.1.28.

We present an application of Theorem 16.1.28 in discrete geometry.

16.1.30. Example. Unit distances among n points. Erdős [1946] asked how many
unit distances can occur among n points in the plane. In a unit grid, n points
794 Chapter 16: Geometry and Topology
√
can achieve about 2n − O( n) unit distances. The points of a finer grid within an
appropriate distance from the origin can achieve about n1+c/log log n unit distances
(Erdős). This growth is superlinear but slower than n1+ for all positive .
Erdős also proved an upper bound of O(n3/2). Let m count the pairs at distance
1 in a given set of n points. Let G be the graph with those pairs forming edges.
Since two circles of radius 1 intersect in at most two points, G does not contain
K 2 ,3 , so any two points have at most two common neighbors. Since each vertex v is
a common neighbor for its (d(v) 2
) pairs of neighbors, ∑ (d(v)
2
) ≤ 2(n2). Since 2m/n is
the average vertex degree, convexity yields ∑ (d(v) 2
) ≥ n(2m/n
2
). Thus (2m/n
2
) ≤ n − 1,
which yields the upper bound of O(n 3/2). (This
is the special case for , t) = (3 , 2)
(s
of the argument given for the upper bound on ex (n; K 2 ,3) in Theorem 13.2.19.)

Using the Regularity Lemma, J ózsa–Szemerédi [1975] improved the upper

bound to o(n3/2). By number-theoretic arguments about incidences between lines
and points, Spencer–Szemerédi–Trotter [1984] improved it to O(n4/3). Using The-
orem 16.1.28, Székely gave an elegant short proof. A similar argument yields the
Szemerédi–Trotter [1983] result that n points and m lines in the plane generate
at most O(n2/3 m2/3) point /line incidences (Exercise 23).

16.1.31. Theorem. (Spencer–Szemerédi–Trotter [1984]) Among a set of n points

in the plane, at most 4n4/3 pairs of points have distance 1.
Proof: (Székely [1997]) The claim is obvious for n ≤ 3; suppose n > 3. By moving
points or pairs of points without reducing the number of pairs at distance 1, we
can ensure that each point is involved in a unit pair and that no two points have
distance 1 only from each other. If any point still is involved in only one unit pair,
we can rotate it around its mate until it has distance 1 from another point. Thus
we may assume that every point is involved in at least two such pairs.
Let P be an optimal n-point configuration, with q unit distance pairs. We
obtain a graph from P , not by using the unit pairs as edges, but by drawing a
unit circle around each point. If a point in P has distance 1 from ¾ other points
in P , then they partition the circle into ¾ arcs. Each pair of points at distance 1
generates one arc moving clockwise on the circle around each of the two points;
altogether we obtain 2q arcs. These are the edges of a loopless multigraph G.
Since two points can appear on two (but not three) unit circles, G may have
edges of multiplicity 2 but no larger multiplicity. We delete one copy of each du-
plicated edge to obtain a simple graph G with at least q edges. We may assume
q ≥ 4n; otherwise the bound already holds.
Because these arcs lie on n circles, they cannot produce many crossings; any
two circles cross at most twice. Thus our layout of G has at most 2(n2) crossings.
By Theorem 16.1.28, G has at least 64 q /n2 crossings. Together, q ≤ 4n4/3 .
1 3

16.1.32. Example. Products of cycles. After cr(K n) and cr(K m ,n), the question of
computing cr(Cm Cn) arises. For m ≤ n, a natural drawing proves an upper bound
of (m − 2)n, which Harary–K ainen–Schwenk [1973] conjectured to be optimal.
This has been proved for m = 3 (Ringeisen–Beineke [1978]), m = 4 (Dean–Richter
[1995]), m = 5 (Richter–Thomassen [1995] and K lešč–Richter–Stobert [1996],
and m ∈ {6 , 7} (Anderson–Richter–Rodney [1996, 1997]). Juarez–Salazar [2001]
gave a short proof that cr(Cm Cn) ≥ (m − 2)n/2.
Section 16.1: Graph Drawings 795

16.1.33. Example. Variants on crossing number. In a drawing with the fewest

crossings, any two edges cross at most once, since switching the routing of two
edges between two points where they cross reduces the number of crossings. How-
ever, it may be possible to reduce the number of pairs of edges that cross by allow-
ing edges to cross more than once, using a drawing with crossings at more than
cr(G) points. The minimum number of pairs of edges that cross (among all draw-
ings of G) is the pair-crossing number pair-cr(G). Another variation is the
odd-crossing number odd-cr(G), the minimum number of pairs of edges that
cross an odd number of times. Since two edges that cross an odd number of times
cross at least once, and two edges that cross contribute a crossing,
odd-cr(G) ≤ pair-cr(G) ≤ cr(G).
It was long unknown whether there exists a graph on which these three param-
eters are not all equal; Pelsmajer–Schaefer–Stefanovic [2006] proved that odd-cr
and cr are not the same parameter. Pach–Tóth [2000a, 2000b] discuss this prob-
lem and other variants on crossing numbers.
Determining the crossing number is NP-hard (Garey–Johnson [1983]), even
when restricted to 3-regular graphs (Hliněný [2006]). Cabello–Mohar [2013]
proved that it is also NP-hard for graphs obtained from a planar graph by adding
one edge. Their arguments also give a geometric proof of Hliněný ’s result.

Another alternative is to restrict the allowable drawings. In particular, given

our discussion of straight-line embeddings for planar graphs, it is natural to ask
how many crossings are forced in a straight-line drawing of a nonplanar graph.

16.1.34. Definition. The linear crossing number cr1 (G) is the minimum num-
ber of crossings in a drawing of G whose edges are drawn as straight-line
segments. More generally, the t-linear crossing number cr t(G) is the min-
imum number of crossings when the edges are drawn as unions of at most t
line segments, allowing t − 1 “bends”.

The linear crossing number has historically been called the rectilinear
crossing number, but the t-linear generalization and the permissibility of arbi-
trary slopes (there is no relation to rectangles or axes) suggest dropping “recti”.
Guy [1972] proved that cr1 (K 8) > cr(K 8). The smallest graph for which this
occurs seems to be K 2 ,2 ,2 ,2 . Bienstock–Dean [1992, 1993] proved many startling
results. For graphs with crossing number at most 3, the crossing number and lin-
ear crossing number are equal. For ¾ ≥ 4, there are graphs with crossing number
¾ and linear crossing number arbitrarily high, but cr1(G) ≤ c (G)cr(G)2 . Also,
allowing one bend in the edges permits a better bound: cr 2(G) ≤ 2cr(G)2 . It is not
known whether these bounds are sharp. Bienstock [1991] proved that there is no
fixed t such that cr t(G) = cr(G) for all G.

16.1.35. Theorem. (Bienstock–Dean [1993]) For m ≥ ¾ ≥ 4, graphs exist with

crossing number ¾ and linear crossing number at least m.
Proof: (sketch). We construct an explicit graph. Begin by forming a graph H
as follows: add to K 2 ,4 a matching on the larger side, then subdivide each of the
eight original edges. There are essentially two ways to embed H in the plane
without crossings, shown on the left below.
796 Chapter 16: Geometry and Topology

• • •
•• •• •••• ••••
• • • • • • • • • • • •
•• •• •••• ••••
• • •
Form H by adding to H eight more edges, joining each vertex in the original
part of size 2 in the bipartition to the four neighbors of the other such vertex in
H. Call these chords of H . As shown above, H has crossing number at most 4.
We complete the construction by replacing each edge of H in H with a copy
of K 2 ,m , forming G. Since the crossings in our drawing of H do not involve edges
of H , we also have cr(G) ≤ 4 (in fact, equality holds).
If cr1 (G) < m, then each copy of K 2 ,m has a “clean” (uncrossed) path joining
its “endpoints”. Extracting this path yields a drawing of H in which the edges of
H are drawn using two segments, and the edges of H − H are drawn as straight
segments. The contradiction arises from a technical argument showing that this
yields a planar embedding of H that is not of the forms on the left above.

16.1.36. Theorem. (Bienstock–Dean [1993]) cr 2(G) < 2[cr(G)]2 for any graph G.
Proof: Let D be a drawing of G in the plane with cr(G) crossings. Obtain G from
D by deleting all edges involved in crossings; G is a plane graph. From G , form
a graph H by placing a vertex inside each face of G that lost edges. Let vC be
the vertex placed inside a face bounded by a cycle C, and make vC adjacent to the
vertices of C. Note that H is planar.
By Fáry ’s Theorem, H has a straight-line drawing D . Deleting vC from D
still leaves C as the boundary of a face. Finiteness of C allows us to assume that
vC does not lie on a segment joining vertices of C.
To embed G, we replace each deleted edge xy that was removed from the face
bounded by C in D. We add it to D − {vC } as the union of two segments close to
xvC and yvC , meeting near vC at a point on the bisector of the angle made by
xvC and yvC . We may need to adjust how far is from vC .
Let t C be the number of crossings inside C removed from the original drawing
of G; they involved at most 2t C edges, now replaced. We want any two such edges
to cross at most once, forming at most (2t2C ) crossings by replacing the edges.
One segment of x y may cross both segments of x y if x and y are on the
x , y-path along . To avoid this, we make close to vC when the x , y-path along
is long, farther from vC when the path is short. The narrower wedge is then
captured inside the wider wedge, avoiding crossings. We can make this precise
by placing the points near vC using a linear extension of the containment poset
of the paths along C. When the endpoints alternate along C, as in (x , x , y , y ),
there is only one crossing, behaving like the original crossing edges.
We have a 2-linear drawing of G. Since the original drawing was optimal,
∑C t C ≤ cr(G). Hence cr 2(G) ≤ ∑ (2t2C ) < 2 ∑ t2C ≤ 2[cr(G)]2 .
Exercises for Section 16.1 797

EXERCISES 16.1

16.1.1. (−) Prove that each point in a triangular region has a unique expression as a convex
combination of the vertices of the triangle (convex combinations are linear combinations
where the coefficients are nonnegative and sum to 1).
16.1.2. Find a Schnyder labeling and small grid embedding of the icosahedron.
16.1.3. Prove Lemmas 16.1.15–16.1.16 and Theorem 16.1.17 to complete the straight-line
embeddability of planar graphs in small grids.
16.1.4. Compute the crossing numbers for K4 ,4 , K 2 ,2 ,2 ,2 , and the Petersen graph.
16.1.5. Determine cr(K1 ,2 ,2 ,2) and use it to compute cr(K 2 ,2 ,2 ,2 ).
16.1.6. Prove that K3 ,2 ,2 has no planar subgraph with 15 edges. Use this to give another
proof that cr(K3 ,2 ,2) ≥ 2.
16.1.7. Prove that the crossing number of the graph below is at most 5.

• •
•
• •
• •

16.1.8. (♦) Let Mn be the graph obtained from the cycle Cn by adding chords between ver-
tices that are opposite (if n is even) or nearly opposite (if n is odd). The graph is 3-regular
if n is even, 4-regular if n is odd. Determine cr(Mn). (Guy–Harary [1967])
16.1.9. (♦) The graph Pn¾ has vertex set [n] and edge set {i j : |i − j | ≤ ¾}. Use a planar
embedding of Pn3 to prove cr(Pn4) = n − 4. (Harary–Kainen [1993])
16.1.10. Determine the crossing number of the graph below.

• • • • • •

• • • • • •

16.1.11. (♦) For each positive integer ¾ , show constructively that some graph embeddable
on the torus requires at least ¾ crossings when drawn in the plane.
16.1.12. (♦) A graph is 1-planar if it can be drawn in the plane with each edge involved
in at most one crossing. Prove that a 1-planar graph with n vertices has at most 4n − 8
edges, for n ≥ 2. The bound is sharp for n ≥ 12; prove sharpness for multiples of 4 at least
8. (Pach–Tóth [1997], Albertson–Mohar [2006], Nagasawa–Noguchi–Suzuki [2018])
16.1.13. (♦) Use the Four Color Theorem to prove that every 2-edge-connected 3-regular
graph with crossing number 1 is 3-edge-colorable. (Jaeger [1980])
16.1.14. (♦) Suppose that n is odd. Prove that in all proper drawings of K n , the parity of
the number of edge crossings is the same. Conclude that cr(K n) is even when n is congruent
to 1 or 3 modulo 8 and is odd when n is congruent to 5 or 7 modulo 8.
798 Chapter 16: Geometry and Topology

16.1.15. (♦) Use cr(K6 ,n) = 6⌊ n2 ⌋ ⌊ n−2 1 ⌋ (Kleitman [1970]) to prove the following.
(a) cr(K m ,n) ≥ m m5−1 ⌊ n2 ⌋ ⌊ n−2 1 ⌋ . (Guy [1970])
(b) cr(Kp) ≥ 80 p + O(p3).
1 4

m−1 n−1
16.1.16. It is conjectured that cr(K m ,n) = ⌊ m
2 ⌋ ⌊ 2 ⌋ ⌊ 2⌋ ⌊ 2 ⌋
n
. Suppose that this conjec-
ture holds for K m ,n and that m is odd. Prove that the conjecture then holds also for K m+1 ,n .
(Kleitman [1970])

16.1.17. Suppose that m and n are odd. Prove that in all proper drawings of K m ,n , the
parity of the number of edge crossings is the same. Conclude that cr(K m ,n) is odd when
m − 3 and n − 3 are divisible by 4 and even otherwise.

16.1.18. Use Exercise 16.1.17 and Kleitman’s computation of cr(K6 ,n) to prove cr(K7 ,7) ∈
{77 , 79 , 81}. (Comment: Woodall [1993] proved cr(K7 ,7) = 81.)

16.1.19. (♦) Harary–Kainen–Schwenk [1973] conjectured cr(Cm Cn) = (m − 2)n for m ≤ n.

This is now known for m ≤ 7, along with cr(K4 Cn) = 3n. Find plane drawings to establish
the upper bounds. (Beineke–Ringeisen [1980]) (Comment:

16.1.20. Prove cr(C3 C3) ≥ 2. (Hint: Find three subdivisions of K3 ,3 that together use
each edge exactly twice.)

16.1.21. Let Q¾ denote the ¾-dimensional hypercube. Prove cr(Q4) ≤ 8. (Comment: Faria–
de Figueiredo–S ýkora–Vrt ’o [2008] proved cr(Q¾ ) ≤ 32 4 − ⌊ ¾ 2+1 ⌋ 2 ¾−2 , conjectured to hold
5 ¾ 2

with equality by Erdős–Guy [1973]. S ýkora–Vrt ’o [1993] proved cr(Q¾) ≥ 20 4 + O(¾ 2 2 ¾).)
1 ¾

16.1.22. (♦) Let º (n) = cr(K n ,n ,n). Prove that n3(n − 1)/6 ≤ º (n) ≤ (9/16)n4 + O(n3). (Hint:
For the lower bound, use Lemma 16.1.24 after proving cr(K3 ,3 ,2) ≥ 5 and cr(K3 ,3 ,3) ≥ 9.
For the upper bound, embed K l ,m ,n on a tetrahedron or generalize a drawing of K n .)

√ (♦) In the plane, let P be a set of n points, and let L be a set of m lines, with
16.1.23.
m ≥ n. Each occurrence of a point in P on a line in L is an incidence. Prove that there
are fewer than 4n2/3 m2/3 incidences. (Szemerédi–Trotter [1983]) (Hint: Use the technique
of Theorem 16.1.31. Comment: This short proof is due to Székely [1997].)

16.2. Combinatorial Topology

Like the probabilistic method, combinatorial topology provides non-construc-

tive existence proofs of natural combinatorial statements. There is a further
analogy. Probabilistic arguments can be “derandomized” to obtain constructive
procedures. Similarly, the topological theorems generally have discrete algorith-
mic proofs, which allow these conclusions also to be made constructive. We will
not study this aspect of the topic.
The most famous instance of the topological method in combinatorics may be
the use of the Borsuk–Ulam Theorem to prove Kneser ’s Conjecture on the chro-
matic number of Kneser graphs. We will derive the topological theorems, but it is
important to note that typically one can apply them knowing only the statements.
We start with easy applications of a simple geometric lemma that turns out
to be an intimate part of the topological development.
Section 16.2: Combinatorial Topology 799

SPERNER’S LEMMA AND BANDWIDTH

Sperner ’s Lemma in the plane involves only parity arguments, but its induc-
tive higher-dimensional generalization to simplices takes us toward the applica-
tions of simplices considered later.

16.2.1. Definition. A simplicial subdivision of a large triangle T is a parti-

tion of T into triangular cells such that every intersection of two cells is a
common edge or corner. We call the corners of cells nodes. A proper label-
ing of a simplicial subdivision of T assigns labels from {0 , 1 , 2} to the nodes,
avoiding label i on the ith edge of T (i ∈ {0 , 1 , 2}). A completely labeled
cell has all three labels.

In a proper labeling, each label appears at one corner of T , and label i avoids
the edge of T joining the corners not labeled i. The figure below illustrates a
simplicial subdivision and the graph we will obtain from it to prove that it has a
completely labeled cell.
1

0 1 2 • •
• • •
0 2 • •
1
1
0 1 • •• • • • •
1 2 • • • • •
0 • • • •
0 0 2 0 2

16.2.2. Theorem. (Sperner ’s Lemma; Sperner [1928]) Every properly labeled

simplicial subdivision contains a completely labeled cell.
Proof: We prove the stronger result that the number of completely labeled cells
is odd. We define a graph G that is a subgraph of the dual of the triangulation.
We introduce a vertex for each cell, plus one for the outside region. Vertices of G
are adjacent if their regions share a boundary segment with labels 0 and 1. The
graph on the right above results from the proper labeling on the left.
A completely labeled cell becomes a vertex of degree 1 in G. A cell with no 0
or no 1 becomes a vertex of degree 0. The other cells are labeled 0 , 0 , 1 or 0 , 1 , 1
and become vertices of degree 2. Hence the desired cells are the only bounded
cells that become vertices of odd degree.
The vertex v for the outside region also has odd degree. In moving from the
0-corner to the 1-corner along the edge of T avoiding label 2, we cross an edge of
G involving v for each switch between label 0 and label 1. Since we start with 0
and end with 1, we switch an odd number of times.
Since the number of vertices of odd degree in every graph is even, the number
of vertices other than v having odd degree is odd, so there are an odd number of
completely labeled cells.

We apply Sperner ’s Lemma to linear embeddings of triangular grids.

800 Chapter 16: Geometry and Topology

16.2.3. Definition. When the vertices of a graph G are numbered with distinct
integers, the dilation is the maximum difference between integers assigned
to adjacent vertices. The bandwidth B(G) of G is the minimum dilation of
a numbering of G.

Dilation is minimized when the numbering has no gaps, but gaps can be con-
venient (Exercise 8). One motivation for studying bandwidth is to minimize the
delay between adjacent vertices when the vertices must be processed in a linear
order. The name originates from sparse matrix computations; the bandwidth of
a matrix is the maximum of |i − j | for a nonzero position (i , j). Computation of
bandwidth is NP-hard even for trees with maximum degree 3 (Garey–Graham–
Johnson–Knuth [1978]). We present two simple lower bounds.

16.2.4. Lemma. (Local density bound) Always B(G) ≥ max H ⊆ G |V(H)|−1

diam H
Proof: Every numbering of G also numbers each subgraph of G. For every ¾-
vertex subgraph H , the two numbers farthest apart differ by at least ¾ − 1. By
the Pigeonhole Principle, some step on a path between them has dilation at least
¾ − 1 divided by the distance between them.

16.2.5. Definition. For a set S of vertices in a graph G, the boundary S is the

set of vertices outside S having neighbors in S.

16.2.6. Lemma. (Boundary bound; Harper [1966]) For a graph G,

B(G) ≥ max¾ min{| S| : | S| = }.

Proof: For each value of  , some set S of  vertices must be the first  vertices in
an optimal numbering of G. The bandwidth of G must be at least | S| , because
the vertex among S that has the highest label has an edge of dilation at least
| S| to a neighbor in S.

Computation of the bound is nontrivial. For the hypercube Q¾ , the value is

n−1
∑ (⌊i/2⌋
i=0
i
), studied in Harper [1966]. Nevertheless, one can use the idea of the
bound without computing it. To prove B(G) ≥ l it suffices to show that an optimal
numbering (or every numbering) has some initial segment S with | S| ≥ l.

16.2.7. Proposition. If v1 , . . . , vn is a numbering of the vertices of a graph G

that is optimal for bandwidth, then
B(G) ≥ max{| {v1 , . . . , v¾ }| : 1 ≤  ≤ n}.

16.2.8. Theorem. (Chvátalová [1975]) B(Pm Pn) = min{m , n}.

Proof: Numbering the vertices in successive ranks along the short direction
yields maximum difference min{m , n}. If m ≥ n ≥ 2, then Pn Pn is a subgraph
of Pm Pn . Hence it suffices to prove that every numbering of Pn Pn has some
initial segment S with at least n neighbors outside S.
Given an ordering v1 , . . . , vn , let S be the smallest initial segment that has
n − 1 positions from a single row or column. By symmetry, we may assume this is
Section 16.2: Combinatorial Topology 801

row r, with position (r, s) being the only element of row r omitted by S. In every
column other than s, there is an element of S, but S contains no column, so each
such column contains an element of S. Finally, position (r, s) itself lies in S,
since it completes row r. Thus | S| ≥ n, and Proposition 16.2.7 applies.
√
The local density bound for Pn Pn is only about (2 − 2)n (Exercise 3). The
boundary bound for a graph may be smaller than the bound in Proposition 16.2.7,
but for Pn Pn they are equal (Exercise 4).

16.2.9. Example. Bandwidth of special graphs. An argument like that in Theo-

rem 16.2.8 shows that B(Cn Cn) = 2n − 1 (Exercise 5; Li–Tao–Shen [1981]). Also
B(K n K n) = ⌊(n2 + n − 1)/2⌋ (Balogh–Csirik [2004]).
The triangular grid T l consists of vertices (i , j , ) such that i , j ,  are non-
negative integers summing to l, with two vertices adjacent when two correspond-
ing coordinates differ by 1. Below we show T4 . Numbering the vertices by rows
produces an upper bound of l + 1 for B(T l). This is optimal, but the local density
bound is only about l/2.

•
• •
• • •
• • • •
• • • • •

We will use Sperner ’s Lemma to prove that B(T l) = l + 1.

16.2.10. Lemma. (Hochberg–McDiarmid–Saks [1995]) Let G be a plane graph

whose bounded faces are triangles and whose unbounded face is a cycle parti-
tioned into three paths. For any 2-coloring of the vertices, exactly one of the
two subgraphs induced by a color class has a connector, meaning a compo-
nent that intersects each of the three bounding paths.
Proof: Call the three bounding paths “sides”, as in a triangle; a connector con-
tains a vertex of each side (not necessarily distinct). If neither color contains a
connector, then label each vertex v with the least index of a side not reachable
from v via vertices of its color. Vertices on the ith side cannot have label i, so the
labeling is proper.
By Sperner ’s Lemma, some cell is completely labeled. Since it has three cor-
ners and only two colors, two corners have the same color. Since they are adjacent,
they can reach the same vertices via paths in their color, and the least side un-
reachable from them cannot be different. This contradiction implies that at least
one color contains a connector.
A connector in one color splits the other vertices into sets that cannot reach
all three sides. Hence connectors cannot exist in both colors.
802 Chapter 16: Geometry and Topology

Because the vertices of T l correspond to the nonnegative integer triples

(i , j , ¾) with i + j + ¾ = l, and vertices are adjacent in T l precisely when they
agree in one coordinate and differ by 1 in the other two, the sum of the distances
from any vertex to the three sides is exactly l. Hence the result B(T l) = l + 1 is
immediate from the following theorem.

16.2.11. Theorem. (Hochberg–McDiarmid–Saks [1995]) Let G be a plane graph

whose bounded faces are triangles and whose unbounded face is a cycle parti-
tioned into three paths. If l is the minimum over v ∈ V(G) of the sum of the
distances from v to each of the three paths, then B(G) ≥ l + 1.
Proof: Let the vertices be v1 , . . . , vn in an optimal numbering. Let t be the
largest index such that the induced subgraph G[v1 , . . . , vt ] contains no connec-
tor. Let R = {v1 , . . . , vt }, let S = R, and let T = V(G) − R − S. By construction,
vt+1 ∈ S. Since R ∪ {vt+1 } contains a connector, R ∪ S contains a connector and
T does not. Since no edge joins R and T , and R contains no connector, also R ∪ T
contains no connector. Lemma 16.2.10 now implies that S contains a connector.
A connector contains paths from each vertex to each bounding path. By hy-
pothesis, the sum of the lengths of these paths from any vertex is at least l. For
some vertex v in S, the paths are disjoint; hence | S| ≥ l + 1. Since S = R, using
this argument on each vertex numbering yields B(G) ≥ | R| = | S| ≥ l + 1.

EQUIVALENT TOPOLOGICAL LEMMAS

Sperner ’s Lemma and Lemma 16.2.10 have higher-dimensional analogues.

These statements and several others including the Brouwer Fixed-Point Theorem
are equivalent. Our chain of implications was developed by Chris Hartman.
Since its proof is so easy, we begin with Sperner ’s Lemma. The statement in
d dimensions is the same as in two dimensions, but we must define precisely the
d-dimensional analogues of the terms.

16.2.12. Definition. A d-dimensional simplex in n is the set of convex combi-

nations of d + 1 points (its corners) not in one affine (d − 1)-dimensional space.
A simplicial subdivision of a region in n decomposes it into cells that are
n-dimensional simplices, with the intersection of any two cells being a lower-
dimensional simplex. The 0-dimensional cells are called vertices.
For a simplex S with corners w0 , . . . , wd , a facet is a (d − 1)-dimensional
simplex generated by d corners of S. A proper labeling of a simplicial sub-
division of S is a coloring of the vertices using {0 , . . . , d} such that color i
does not appear on the facet of S that omits wi . A completely labeled cell
is one having all d + 1 colors.

In a proper labeling, wi must have color i, for 0 ≤ i ≤ d. The proof in Theorem

16.2.2 is a special case of the general induction step.

16.2.13. Theorem. (Sperner ’s Lemma; Sperner [1928]) Every proper label-

ing of a simplicial subdivision of a d-dimensional simplex T has a completely
labeled cell.
Section 16.2: Combinatorial Topology 803

Proof: We use induction on d, proving the stronger statement that the number
of completely labeled cells is odd. For d = 0 the claim is trivial. For d > 0, form a
graph G having a vertex for each cell inside T and one vertex w for the “outside”.
Join two vertices by an edge in G if their intersection is a (d − 1)-dimensional
simplex whose corners have every label except d.
Every completely labeled cell has degree 1 in G. Every other cell has degree
0 or 2, since having d labels without label d requires a repeated label, and a facet
generating an edge is formed by taking either vertex having the repeated label
plus all the remaining vertices (if they have distinct colors).
The vertex w has neighbors in G only via the facet T omitting the corner of
T with color d, since no other facet has vertices with all labels but d. Restrict-
ing the subdivision to T yields a properly labeled simplicial subdivision in d − 1
dimensions. Its completely labeled cells yield the edges incident to w in G.
By the induction hypothesis, w has odd degree in G. Since the number of
vertices of odd degree is even and all other vertices of odd degree in G correspond
to completely labeled cells, the number of completely labeled cells is odd.

We note one application: Aharoni–Haxell [2000] used Sperner ’s Lemma to

prove a hypergraph generalization of Hall’s Theorem, applied in Aharoni [2001].
We use Sperner ’s Lemma to obtain a natural generalization of Lemma 16.2.10.

16.2.14. Theorem. (Connector Lemma; Hochberg–McDiarmid–Saks [1995])

In a d-coloring of the vertices of a simplicial subdivision of a d-dimensional
simplex, some color has a component touching all (d − 1)-dimensional facets.
Proof: If there is no such component, then coloring each vertex with the least
index of a facet that it cannot reach in its own color (along 1-dimensional faces)
yields a proper labeling. By Sperner ’s Lemma, there is a completely labeled cell.
By the Pigeonhole Principle, two vertices of this cell have the same color. Since
they are joined by an edge of the subdivision, they cannot have distinct labels.
This contradiction implies that the desired component exists.

16.2.15. Example. The game of Hex. In two dimensions, n-by-n Hex is played on
a grid of n2 hexagons, shown below on the left for n = 4. Two players alternately
seize hexagonal cells. One player wants to seize cells forming a path connecting
the upper-left border and the lower-right border; the other wants a path connect-
ing the upper-right border and the lower-left border.
The game was invented by the Danish poet and engineer Piet Hein in 1942
and reinvented by John Nash in 1948 as a graduate student at Princeton. Our
account comes from Nasar [1998, Chapter 6]. At Princeton the game was known
as “Nash” or “ John”. David Gale suggested the name “Hex” under which Parker
Brothers later marketed the game with n = 11. Nash proved that, unlike 2-
dimensional Tic-Tac-Toe, every game of Hex has a winner. Gale [1979] published
the equivalence in d between the extension of this result and the Brouwer Fixed-
Point Theorem. Tanaka [2007] showed that the Hex game theorem is also equiv-
alent to Arrow’s Impossibility Theorem of social choice theory.
The (weak) dual of the Hex board, for n = 4 on the right below, is the square
grid graph plus diagonal edges in one direction. In d dimensions, it is the cubical
grid plus edges within cubelets for the comparability graph of 2 n .
804 Chapter 16: Geometry and Topology

16.2.16. Theorem. (Hex Theorem; Gale [1979]) In a d-coloring of a simplicial

subdivision of a d-dimensional cube such that the points on the opposite faces
in each direction i have color i (but points on more than one face may have
any color), there is for some i a path in color i joining the faces in direction i.
Proof: Add d vertices, one of each color; the new vertex of color i forms a “pyra-
mid” over one face of the cube in direction i. The new points also form a simplex
S with the common point of those d faces. The result can be viewed as a d-colored
simplicial subdivision of a d-dimensional simplex, as illustrated below for d = 2.
The vertices of the outer simplex are the d added points plus the common point
of the d faces over which the pyramids were not built.
By the Connector Lemma, in some color c there is a component C that touches
all (d−1)-dimensional facets. One facet is the (d−1)-dimensional simplex S formed
by the d added points. In another, all points with color c lie in the face of the cube
opposite the face attached to the point of S with color c. Since C reaches from S
to that facet, C contains a path in color c joining the opposite faces of color c.

1 2
S

2 2
2 2
2

16.2.17. Theorem. (Pouzet’s Lemma; see Zaks [1995]) Let A be a cartesian

product of d integer intervals. Points of A are considered adjacent if they
differ by 0 or 1 in every coordinate. Let º assign to each point in A a unit
vector parallel to some axis (positive or negative). If x + º (x) ∈ A for each
x ∈ A, then there exist adjacent points x and y such that º (x) = −º (y).
Proof: Using A as the vertex set, form a simplicial subdivision of the region en-
compassed by A such that all 1-dimensional simplices join vertices deemed adja-
cent in A. Color each vertex x with the direction of º (x) (positive and negative
yield the same color) to form a d-coloring of the vertices of the resulting graph.
Associate color i with the extreme faces in coordinate i. By the Hex Theorem,
for some color i a path P in color i joins the extreme faces in color i, say with end-
points u and v. For a point x with color i in an extreme face in coordinate i, there
is only one choice for º (x), since x + º (x) ∈ A. In particular, º (u) = −º (v). Since
there are only two choices for º along P , somewhere along P adjacent vertices x
and y satisfy º (x) = −º (y).
Section 16.2: Combinatorial Topology 805

A set in d is compact if it is closed and bounded. We need the property

that every sequence in a compact set has a convergent subsequence. Two sets in
d
are homeomorphic if one can be mapped onto the other by a continuous bijec-
tion with a continuous inverse; such a function preserves compactness. Brouwer ’s
Theorem is often stated for the unit ball or simplex, but it holds equivalently for
all sets homeomorphic to these, and we use the unit cube in the proof. For ex-
tensions of Brouwer ’s Theorem and its relation to Hex, see Bj örner–Matoušek–
Ziegler [2017].

16.2.18. Theorem. (Brouwer Fixed-Point Theorem; Hadamard [1910],

Brouwer [1911]) Let S be a set in d that is homeomorphic to the unit cube.
Every continuous function º from S to itself has a fixed point.
Proof: Let Q be the unit cube in d . Let h: S → Q be a continuous bijection
with a continuous inverse. Now hº h−1 is a continuous function from Q to itself.
If x = hº h−1 (x), then h−1 (x) = º (h−1 (x)). Hence we reduce to the case S = Q.
Define by letting (x) =  (x) − x. If  has no fixed point, then is never
0. Since Q is compact, | (x)| is bounded away from 0 (otherwise some sequence
converges to a point in Q where is 0). Choose  so that | (x)| >  for x ∈ Q.
Place a grid of points on Q using n steps in each direction. At each vertex x
of the grid, choose a unit vector v in a coordinate direction closest to the vector
(x). The distance between
√ (x) and the hyperplane perpendicular to v through
the origin is at least / d. Hence the distance between√ the values of at vectors
assigned opposite coordinate vectors is at least 2 / d.
Since  is continuous, also is continuous. Therefore, choosing n large
enough yields a grid A within Q such that the distance between values of √at
gridpoints that are equal or consecutive in each coordinate is less than 2 / d.
(Actually, this uses uniform continuity of continuous functions on compact sets.)
Scale A so that steps of length 1/n in coordinate directions become unit steps.
The selection of unit coordinate vectors yields a labeling of the gridpoints as speci-
fied in Pouzet ’s Lemma. Hence some adjacent vertices choose opposite directions.
This contradicts the choice of n and implies that  has a fixed point.

16.2.19. Theorem. (Sperner ’s Lemma; Sperner [1928]) Every proper labeling

of a simplicial subdivision of a simplex T has a completely labeled cell.
Proof: (from Brouwer ’s Theorem). Given a proper labeling of a simplicial subdi-
vision of a d-dimensional simplex T , we define a continuous mapping from T to
itself. Map each vertex with label j into the corner of T with label j + 1 (mod d).
Extend the map into each cell by mapping a convex combination of the vertices
of a cell into the same combination of their images. This defines a continuous
function  on T . By Brouwer ’s Theorem,  has a fixed point.
By construction, vertices of the subdivision are not fixed points. Any other
point is a convex combination of vertices. Let R be the minimal simplex contain-
ing a point x in T . If the vertices of R do not have distinct labels, then  (x) lies
in a simplex of smaller dimension, and then  (x)
= x.
If the vertices of R have distinct labels, and S is that set of labels, then R is
mapped into a simplex U spanned by the corners of T having labels in { j + 1: j ∈
S}. If there is no completely labeled cell, then some label in S is not among those
806 Chapter 16: Geometry and Topology

corners. Since the labeling is proper, points with that missing label are forbidden
/ U , so again º (x)
= x.
from U. Hence some corner of R is not in U , which yields x ∈
Since Brouwer ’s Theorem requires a fixed point, it cannot happen that there
is no completely labeled cell.

By this cycle of implications, the statements of these five theorems are equiv-
alent. Since we proved Sperner ’s Lemma (Theorem 16.2.13), they are all true.

THE BORSUK–UL AM THEOREM

The Borsuk–Ulam Theorem is one of the most famous theorems of algebraic

topology. Some call it just Borsuk’s Theorem; Borsuk’s paper credited Ulam for
conjecturing it. A popular expression of the 2-dimensional case is that at any
moment, on the earth’s surface there are antipodal locations with the same tem-
perature and wind speed.
The theorem has many proofs, equivalent versions, and applications. In his
book Using the Borsuk–Ulam Theorem, Matoušek [2003, p.25] notes “hundreds of
papers with various new proofs, variations of old proofs, generalizations, and ap-
plications have appeared; the most comprehensive survey known to me, Steinlein
[1985], lists nearly 500 items in the bibliography.”
The Borsuk–Ulam Theorem has many proofs using algebraic topology. In-
stead, we obtain it from Tucker ’s Combinatorial Lemma. We then obtain several
equivalent forms of the theorem and give combinatorial applications.

16.2.20. Definition. Let B̂n = {x ∈ n : ∑ | x i | ≤ 1} and Ŝ n−1 = {x ∈ n : ∑ | x i | = 1},

so Ŝ n−1 is the boundary of B̂n . The set B̂n has 2n extreme points (vertices)
obtained by setting one coordinate to ±1. An orthant in n is a simplex
having one corner at the origin and at most n other corners reached by unit
vectors along distinct coordinate axes. The octahedral subdivision is the
simplicial subdivision of B̂n whose simplices are the 2 n full-dimensional or-
thants. A special subdivision of B̂n is a simplicial subdivision that refines
the octahedral subdivision and is centrally symmetric on Ŝ n−1 .
1

2 1

−2 1 1

−2 2
1 1 1
2
1 2

2
−1 2

−1
Section 16.2: Combinatorial Topology 807

The special subdivision of B̂2 shown above is adapted from Matoušek [2003,
pp. 37–40], which also presents the proof we give for Tucker ’s Lemma. “Refining ”
the octahedral subdivision means that each orthant is a union of cells in the spe-
cial subdivision. Labels from {±1 , . . . , ± n} are given to the vertices, with each
vertex in Ŝ n−1 given the negation of the label on the antipodal vertex. We say that
labels i and −i are complementary and that an edge (1-dimensional simplex) with
complementary labels is a complementary edge. We seek a complementary edge.
Our proof of Tucker ’s Lemma uses the same tool as Sperner ’s Lemma: we
only need that the number of odd-degree vertices in a special graph is even. In
both cases, the resulting proof is algorithmic. That is, following a path in the
special graph leads to the desired configuration. Tucker [1946] published a 2-
dimensional version. A proof for n dimensions via simplicial topology appears in
Lefschetz [1949].

16.2.21. Theorem. (Tucker ’s Combinatorial Lemma) Let T be a special sub-

division of B̂n . If the vertices of T are labeled using {±1 , . . . , ± n} so that
antipodal points of Ŝ n−1 have complementary labels, then T contains a com-
plementary edge.
Proof: (Freund–Todd [1981]) Let sgn( ) denote the sign (0, +1, or −1) of a real
number ; for x ∈ n , let sgn(x) denote the vector of signs. Sign vectors of points
interior to full-dimensional orthants have no zeros, and such sign vectors deter-
mine the 2 n orthants. We consider lower-dimensional orthants by letting C(s) be
the closure of the set of points in B̂n with sign vector s.
Say that a simplex in the special subdivision T requires label i or label −i
if the interior points in the smallest orthant containing have +1 or −1, respec-
tively, in coordinate i of their common sign vector. Also say that is fully labeled
if all labels thus required appear on vertices of . Always the 0-dimensional sim-
plex O at the origin is fully labeled, since it has no required label. If a simplex
contained in Ŝ n−1 is fully labeled, then so is its antipodal simplex.
Next define a graph G whose vertices are the fully labeled simplices in T .
Let fully labeled simplices and  be adjacent if they both lie in Ŝ n−1 and are
antipodal or if one is a facet of the other (obtained by deleting one corner) and the
smaller has all labels required of the larger.
It suffices to prove the following claims about vertex degrees in G, since they
imply that the component of G containing O also contains a vertex having a com-
plementary edge. That component will be a path from O to a simplex containing
a complementary edge.
(1) O has degree 1.
(2) A fully labeled simplex containing a complementary edge has degree 1.
(3) All other fully labeled simplices have degree 2.
(In the example below Definition 16.2.20, the path from O to the desired sim-
plex is as shown below. The path has length 9, taking four steps in the top row
of the figure. Each step changes the dimension by 1, up or down, except when
the adjacency of antipodal simplices in the boundary is used. Such a move occurs
here in the next to last step. The horizontal direction is coordinate 1; the vertical
is coordinate 2. The graph has another component that is an 8-cycle.)
808 Chapter 16: Geometry and Topology

• • 2 2
1 1 1 1 1 1

1 1 −2 −2
1
1 1

2 2 2 −1 −1

Now we prove the degree claims. Let ¾ be the number of labels required on a
fully labeled simplex , and let C be the smallest orthant containing . Since
is fully labeled, C has dimension  , and has dimension  − 1 or  .
Suppose first that has dimension  − 1. Since it has  vertices and requires
labels for  distinct indices, has no complementary edge. If
⊂ Ŝ n−1 , then is a
facet of exactly two -dimensional simplices in C. They are fully labeled because
they have the same smallest orthant as and already is fully labeled; this gives
two neighbors in G. Although a facet of may be fully labeled (this can happen
when the facet lies in a smaller orthant), it cannot be adjacent to , because it
has only  − 1 vertices and thus cannot have all  labels required by . If ⊂ Ŝ n−1 ,
then by the same argument is a facet of one neighbor that is a -dimensional
simplex in C, but the antipodal simplex to is also fully labeled and a neighbor.
In both cases, has degree 2.
Now suppose that has dimension  ; it has  + 1 vertices. Since only re-
quires  labels, the  + 1 labels on it may (a) have a duplicate label, (b) have a
complementary pair, or (c) have an unrequired label ± j with s j = 0. In case (a),
has two fully labeled facets with the  required labels, in the same smallest or-
thant, and any larger fully labeled simplex containing and adjacent to would
force another label on ; hence has degree 2. In case (b), contains a complemen-
tary edge; since larger neighbors are forbidden by the same argument as above,
and only one facet has the required labels, has degree 1.
In case (c), form a new sign vector t from the sign vector s associated with
by setting the jth coordinate to the sign of the extra label on . Since s j = 0,
orthant C(s) is contained in orthant C(t), and is a facet of a fully labeled simplex
in C(t) that requires  + 1 labels. The other ( + 1)-dimensional simplex of which
is a facet has the opposite sign in the jth coordinate, so even if it is fully labeled
it is not adjacent to . Meanwhile, as in the paragraph above, has exactly one
fully labeled facet that has the labels required by , so has degree 2. The one
exception is that when = O there is no smaller facet, so O has degree 1.

The central symmetry of the subdivision on the boundary is crucial for giv-
ing degree 2 to simplices there. No such symmetry is needed inside, since the
adjacency conditions are local. Refinement of the octahedral subdivision allows
the proof to use orthants of all dimensions, starting the path from the origin.
Tucker ’s Combinatorial Lemma sounds somewhat like Pouzet ’s Lemma (The-
orem 16.2.17). A short proof of Tucker ’s Lemma from Pouzet ’s Lemma would
complete a cycle of implications including the Borsuk–Ulam Theorem with only
Section 16.2: Combinatorial Topology 809

Sperner ’s Lemma proved directly. Tucker ’s Lemma holds also without requiring
refinement of the octahedral subdivision (Exercise 12, using Borsuk’s Theorem).
To obtain Borsuk’s Theorem from Tucker ’s Lemma any sequence of successively
refined subdivisions suffices, so we only need the restricted version proved above.
There are many versions of the Borsuk–Ulam Theorem. We start with one
well suited for the use of Tucker ’s Lemma and one used in our first application.

16.2.22. Definition. Let S n−1 denote the (n − 1)-dimensional sphere (in n), con-
sisting of the points having Euclidean distance 1 from the origin; the ball Bn
consists of those with distance at most 1. A function º on S n is antipodal if
º (− x) = −º (x) for all x ∈ S n .
16.2.23. Theorem. (Borsuk–Ulam Theorem; Borsuk [1933]) (A) If a continu-
ous function º : Bn → n is antipodal on the boundary S n−1 , then º (x) = 0
for some x ∈ Bn . (B) If a continuous function º : S n → n is antipodal on S n ,
then there exists x ∈ S n with º (x) = 0.
Proof: (Tucker [1946]; see Freund–Todd [1981]) Consider (A). To facilitate appli-
cation of Tucker ’s Lemma, we may use compositions with continuous functions as
in Theorem 16.2.18 (preserving antipodality) to assume instead that º is defined
on B̂n and antipodal on its boundary, which we have called Ŝ n−1 .
Suppose that º (x) is never zero. For any special subdivision T of B̂n , label
each vertex x with the least index i such that ºi(x) has the largest magnitude
among the coordinates of º (x), and let the sign of the label be the sign of ºi(x).
Since º (− x) = −º (x) when x ∈ Ŝ n−1 , antipodal points have complementary labels.
The subdivision and labeling satisfy the hypotheses of Theorem 16.2.21, and
hence there is a complementary edge. Successively refining the triangulation so
that the maximum length of edges tends to 0 yields a sequence of positive ends
and a sequence of negative ends of complementary edges. Since B̂n is compact,
the sequence of positive ends has a convergent subsequence. Since the distance
between points in the two sequences tends to 0, the corresponding points in the
sequence of negative ends converge to the same point x∗ , which lies in B̂n . It is ap-
proached through points on which the largest nonzero coordinate in º is positive
and through points on which it is negative. By continuity, º (x∗) = 0.
(A)⇒(B): let º : S n → n be continuous and antipodal. Note that S n ⊆ n+1 ,
so points in S n have n + 1 coordinates. For x ∈ Bn ⊆ n , define x ∈ S n by letting
x i = x i for 1 ≤ i ≤ n and x n+1 = (1 − ∑i=1 x2i )1/2 . Define ½ : Bn → n by ½(x) =
n

º (x ). Since º is continuous, also ½ is continuous. Furthermore, if x ∈ S n−1 , then

x ∈ S n and (− x) = −(x ). Hence ½ is antipodal on S n−1 , since º is antipodal on
S n . We conclude by the first statement that ½(x) = 0 for some x ∈ Bn , and hence
x in S n has the desired property: º (x ) = 0. (Also (B)⇒(A); Exercise 14.)

16.2.24. Application. Splitting of necklaces. Consider a necklace, opened at the

clasp, having 2n beads chosen from ¾ colors and arranged in some order with 2ai
beads of color i. Two jewel thieves want to split the beads so that each receives
half of each color, doing this with the fewest possible cuts. When the beads of
each color appear consecutively, there must be a cut within each color, so ¾ cuts
may be needed. Do ¾ cuts always suffice? The question arose in the context of
VLSI circuit design (Bhatt–Leiserson [1981]).
810 Chapter 16: Geometry and Topology

Goldberg–West [1985] answered affirmatively using induction on ¾ , simpli-

cial topology, and a continuous analogue. Alon–West [1986] noted that a suitable
continuous analogue was proved 20 years earlier from Borsuk’s Theorem.
The continuous analogue allows a ¾-coloring of the unit interval, where the
set of points with a single color is measurable. Without generalizing to measur-
able sets, just suppose that the density of each color is an integrable function of
position on the interval and that the densities sum to 1 at each point. The to-
tal amount of each color is the integral of its density function. A bisection is a
set of points y0 , . . . , yr+1 with 0 = y0 ≤ · · · ≤ yr+1 = 1 such that the union of the
intervals [y2i , y2i+1 ] for all i captures half the amount of each color. We seek a
bisection with r = ¾ , corresponding to using at most ¾ distinct cuts.
The continuous result implies the discrete result, by modeling a 2n-bead neck-
lace using 2n equal subintervals where the i interval has density 1 in the color of
the ith bead and density 0 in the others. A bisection yields the desired split un-
less cuts occur in the middle of beads. If there is such an internal cut, then there
must be another internal cut in the same color, and these two cuts can be moved
toward the edges of the subintervals to reduce the number of bad cuts. Similarly,
one can capture the ceiling or floor of half the number of beads, as desired, when
a color has an odd number of beads. For circular necklaces, see Exercise 13.

The next theorem yields bisections with at most ¾ cuts. Giving the positively
signed intervals to one thief and the negatively signed intervals to the other gives
each half the measure of each color. When consecutive signs are equal, fewer in-
tervals (and cuts) can be used.

16.2.25. Theorem. (Hobby–Rice [1965]) Given integrable densities 1 , . . . , ¾

on [0 , 1], there are points y0 , . . . , y¾+1 with 0 = y0 ≤ · · · ≤ y¾+1 = 1 and signs
0 , . . . , ¾ ∈ {−1 , +1} such that ∑¾j =0  j ∫ y j +1 i(t)dt = 0 for 1 ≤ i ≤  .
y
j

Proof: From the  densities, we define a function  : S ¾ → ¾ . For x ∈ S ¾ , let

y0 = 0 and y j = ∑i=1 x2i for 1 ≤ j ≤  + 1. Note that y¾+1 = 1, since x ∈ S ¾ ; we
j

use y1 , . . . , y¾ as break points on the interval. For 1 ≤ i ≤  , let  j = sign(x j ) and

i(x) = ∑¾j =0  j ∫ y j +1 i(t)dt. Let  (x) = (1 (x) , . . . , ¾ (x)).
y
j
The densities are integrable, so  is continuous. For  (− x), subintervals are
the same as for  (x), but the coefficients change signs; hence  (− x) = − (x). By
the Borsuk–Ulam Theorem, there exists x ∈ S ¾ with  (x) = 0. The corresponding
values y1 , . . . , y¾ and 0 , . . . , ¾ are the desired break points and signs.

Alon [1987] generalized the Hobby–Rice Theorem to simultaneous splitting

of t continuous measures, using a generalization of the Borsuk–Ulam Theorem
in Bárány–Shlosman–Szűcs [1981]. With this, he proved the conjecture of Gold-
berg and West that an open necklace with  colors of beads can be cut in (t − 1)
places so that the resulting segments can be distributed to t thieves with each
thief receiving the same amount of each color.
Our second application is more subtle and more important and needs another
version of the Borsuk–Ulam Theorem. The original paper (Borsuk [1933]) gave
three variants, and Lyusternik–Shnirel’man [1930] earlier gave another. We
have proved Statement A below in Theorem 16.2.23; we need Statement E. We
Section 16.2: Combinatorial Topology 811

only need to work our way down to E; but we include the return trip for complete-
ness. See Exercises 14–16 for other equivalent statements. We mostly follow the
presentation of Matoušek [2003].

16.2.26. Theorem. The following statements are equivalent.

(A) For all continuous antipodal º : S n → n , there is x ∈ S n with º (x) = 0.
(B) For continuous º : S n → n , there is x ∈ S n with º (x) = º (− x).
(C) There is no continuous antipodal º : S n → S n−1 .
(D) (Lyusternik–Shnirel’man Theorem) In a cover of S n by n + 1 closed
sets, at least one contains antipodal points.
(E) In any cover of S n by n + 1 sets such that each of the first n is open or
closed, at least one set contains antipodal points.
Proof: A ⇒ B. Apply A to defined by (x) =  (x) −  (− x).
B ⇒ E. Let U1 , . . . , Un+1 be such a cover, and suppose no Ui contains an-
tipodal points. Define  : S n → n by letting i(x) be the distance from x to Ui ,
written d(x , Ui), for 1 ≤ i ≤ n (the distance from a point to a set is the infimum
of distances to points in the set, so this is well-defined also when Ui is open). Note
that  is continuous in x.
By Statement B, there exists x ∈ S n with  (x) =  (− x). Since Un+1 does not
contain antipodal points, x or − x belongs to one of the other sets, say U¾ . By sym-
metry, we may assume x ∈ U¾ , and hence d(x , U¾) = 0. Hence also d(− x , U¾) = 0.
If U¾ is closed, then d(− x , U¾) = 0 yields − x ∈ U¾ , and U¾ contains antipodal
points. If U¾ is open, then − x lies in U ¾ , the closure of U¾ , which is the unique
smallest closed set containing U¾ . The complement of an open set is a closed set,
so S n − (−U¾) is a closed set containing U¾ , since U¾ does not contain antipodal
points. We obtain U ¾ ⊆ S n −(−U¾). Hence − x ∈ S n −(−U¾), which yields − x ∈ / −U¾
and x ∈ / U¾ , a contradiction.
E ⇒ D. Statement D is just a special case of statement E.
D ⇒ C. Let  be a simplex in n containing the origin in its interior. Pro-
jecting the facets of  onto S n−1 yields closed sets F1 , . . . , F n+1 that cover S n−1
without any of them containing antipodal points. If  : S n → S n−1 is antipodal,
then let G i =  −1 (F i) for 1 ≤ i ≤ n + 1. If G j contains antipodal points, then 
maps them to antipodal points in F j , which do not exist.
C ⇒ A. If  : S n → n avoids 0, then letting (x) = ºº (x) (x) contradicts C.

The special case of Theorem 16.2.26E in which all n + 1 sets are open is also
called the Lyusternik–Shnirel’man Theorem. The case in which each set is open
or closed (slightly weaker than E) was shown by Greene [2002]. Aigner–Ziegler
[2014, Chapter 42] observed that no condition is needed on the last set.

KNESER CONJECTURE AND G ALE’S LEMMA

The Kneser graph K(n , ) (also written as KG(n , ) or KG n , ¾) is the graph

with vertex set ([n]
¾ ) whose edges are the pairs of disjoint -sets. Note that the
special case K(5 , 2) is yet another appearance of the Petersen graph. Although
not originally posed in graph-theoretic language, the famous conjecture of Kneser
[1955] is that (K(n , )) = n − 2  + 2 when n ≥ 2  . The upper bound is easy.
812 Chapter 16: Geometry and Topology

16.2.27. Proposition. (K(n , )) ≤ n − 2 + 2 when n ≥ 2 .

Proof: We cover V(K(n , )) with n− 2 + 2 independent sets. For 1 ≤ i ≤ n− 2 + 1,
the -sets containing i pairwise intersect and hence form an independent set. The
uncovered vertices are the -sets contained in {n − 2 + 2 , . . . , n}. This set has
size 2 − 1, so its -sets pairwise intersect and hence form an independent set.

16.2.28. Remark. The Kneser graph is of interest as an explicit graph with

chromatic number much larger than the typical lower bounds. The clique num-
ber is only ⌊ n/ ⌋ . For example, when n = 3 − 1 the graph is triangle-free but
has chromatic number + 1.
In addition, K(n , ) has chromatic number much larger than its fractional
chromatic number, where the fractional chromatic number º (G) is the in-
fimum of a/b such that every vertex can be covered b times using a total of a
independent sets. Always º (G) ≤ (G), since (G) independent sets can cover
every vertex once.
Since independent sets have size at most (G), an n-vertex graph G requires
at least bn/(G) independent sets to cover each vertex b times. Hence º (G) ≥
n/(G). For a vertex-transitive graph like the Kneser graph, equality holds; in
fact, º (K(n , )) = n/ (Exercise 17). Usually n/(G) is a better lower bound than
(G), but for the Kneser graph they are essentially equal.

16.2.29. Theorem. (Kneser Conjecture) (K(n , )) = n − 2 + 2 when n ≥ 2 .

Proof: (Greene [2002]) It suffices to prove (K(n , )) ≥ n − 2 + 2. Let d = n −
2 + 1. Let X be a set of n points in S d with no more than d on any hyperplane
through the origin (this holds for “general position” in d+1).
Consider a proper d-coloring  of K(n , ). We define subsets of S d correspond-
ing to color classes under  . For x ∈ S d and 1 ≤ i ≤ d, put x ∈ Ui if some -tuple
with color i is contained in the open hemisphere centered at x. Because we use
open hemispheres, the sets U1 , . . . , Ud are open. To complete a cover of S d , define
Ud+1 = S d − ⋃i=1 Ui . By Theorem 16.2.26E, some Ui contains antipodal points.
d

If Ui contains antipodal points, where i ≤ d, then the open hemispheres

around those points are disjoint, and hence we have given color i to disjoint -
sets, violating the choice of  as a proper coloring of K(n , ).
Hence Ud+1 contains antipodal points x and − x. Since these points were not
put in another set, the open hemispheres around x and − x each contain at most
− 1 points from the set X . This leaves at least n − 2 + 2 points lying in the
hyperplane separating the two hemispheres. However, n − 2 + 2 = d + 1, so this
contradicts the choice of X . Hence K(n , ) has no proper d-coloring.

16.2.30.* Remark. Lovász [1978] proved the Kneser Conjecture using topolog-
ical methods (homotopy). Bárány [1978] gave a short proof from Gale ’s Lemma
(Theorem 16.2.33) and the open set version of the Lyusternik–Shnirel’man Theo-
rem (see Exercise 19). These proofs appear in the book of Matoušek [2003], along
with proofs of the Kneser Conjecture by Dol’nikov [1981] and Sarkaria [1990]
using other variants of the Borsuk–Ulam Theorem.
Matoušek [2004] found a combinatorial argument for Kneser ’s Conjecture
using Tucker ’s Lemma without the Borsuk–Ulam Theorem, thereby avoiding
Section 16.2: Combinatorial Topology 813

topological statements. Ziegler [2002] developed a modular analogue of Tucker ’s

Lemma, extending Matoušek’s approach to give combinatorial proofs of related
theorems (see also Matoušek–Ziegler [2004]). Although we reached Greene ’s
proof by proving Borsuk’s Theorem from Tucker ’s Lemma, our proof is not fully
combinatorial because the proof of Theorem 16.2.26E from Theorem 16.2.26B
uses topological properties of open and closed sets.
Exercise 18 presents an application of Theorem 16.2.29.

Schrijver [1978] showed that a small subgraph of K(n , ¾) already has the full
chromatic number n − 2 ¾ + 2. For the proof we need to develop several tools.

16.2.31. Definition. The moment curve in d

is { (t): t ∈ }, where (t) =
(t , t2 , . . . , t d).

16.2.32. Lemma. Every hyperplane in d intersects the moment curve in at

most d points. Furthermore, if a hyperplane contains d points of , then
crosses it at each intersection point.
Proof: A hyperplane H is the set of solutions to a single linear equation:
∑i=1 ai x i = b, where a is a nonzero d-tuple. When (t) ∈ H , we have ∑i=1 ai ti =
d d

b. Hence the intersection points are zeros of the polynomial p defined by p(t) =
(∑di=1 ai t i) − b. Since p has degree d, there are at most d points of intersection.
Furthermore, if there are d distinct points of intersection, then p has d sim-
ple zeros, and the derivative of p is nonzero at all those points. Thus p changes
sign at each of those points, which means that crosses H at each of them.

The moment curve has many applications in discrete geometry. Here we use
it to prove a stronger version of Gale’s Lemma; the original result of Gale [1956]
is the weaker version of the next result obtained by omitting the requirement of
cyclically nonconsecutive indices.

16.2.33. Theorem. (Schrijver [1978]) Given d ≥ 0 and  ≥ 1, there is a set

x1 , . . . , x 2¾+d in S d such that every open hemisphere contains a -set in
x1 , . . . , x 2¾+d having no two elements with cyclically consecutive indices.
Proof: (Matoušek–Ziegler [2004]) Open hemispheres correspond to open half-
spaces whose boundaries are hyperplanes through the origin. Using projection to
S d via lines through the origin, it thus suffices to find points x1 , . . . , x 2¾+d in d+1
(with distinct projections on S d) such that every open half-space whose boundary
contains the origin contains a -set in {x1 , . . . , x 2¾+d } whose indices are cyclically
nonconsecutive.
Let H1 be the hyperplane in d+1 defined by x1 = 1; points in H1 have distinct
projections on S d . We move the moment curve into H1 ; let ¯(t) = (1 , t , t2 , . . . , t d).
For any hyperplane H in d+1 not parallel to H1 , dropping the first coordinate of
the points in H ∩ H1 yields a hyperplane in d . Hence Lemma 16.2.32 applies
to the intersections with ¯ of any hyperplane through the origin not parallel to
H1 . Furthermore, there is a one-to-one correspondence between hyperplanes in
d
and hyperplanes through the origin in d+1 not parallel to H1 .
Take any 2  + d points w1 , . . . , w2¾+d along ¯ , indexed in order of appear-
ance along ¯ . Let W = {w1 , . . . , w2¾+d }. Let H be a hyperplane through the
814 Chapter 16: Geometry and Topology

origin in d+1 , bounding open half-spaces H + and H − . Let x i = (−1)i wi and X =

{x1 , . . . , x 2¾+d }. We claim that each of H + and H − has at least ¾ points from X ,
and that their indices are cyclically nonconsecutive within [2 ¾ + d].
By Lemma 16.2.32, H contains at most d points from ¯ . If H contains j
points of ¯ , with j < d, then these points plus the origin determine a subspace J
of dimension j + 1, and we can rotate H around J (always containing J , like a
plane rotating around a line in 3-space) until the moving hyperplane acquires an
additional point of W . Since we did not pass through any point of W , no point of
W moved from one side of H to the other. Hence this process allows us to assume
that H contains d points of W .
Let W0 = H ∩ W and W1 = W − W0 . Let the parity of a point in W be the
parity of its subscript, so each point of W1 is even or odd. Color w ∈ W1 black if
exactly one of {w is even, w ∈ H + } holds; otherwise color w white.

w1 H+ w11

H−

We claim that black and white points alternate along ¯ . Let j be the number
of times that ¯ crosses H between w and the next point w of W1 . The points of
crossing all lie in W0 , and by Lemma 16.2.32 the curve crosses through H at each
such point. Hence w and w are on the same side of H if and only if j is even, but
they have the same parity in W if and only if j is odd. Hence exactly one of the
two properties changes, which gives w and w opposite colors.
If wi is odd and lies in H − , then x i is in H + . Thus points of X that lie in H +
correspond to the white points in W1 , and the points of X that lie in H − corre-
spond to the black points in W1 . Since the colors alternate on the 2 points of
W1 along ¯ and 2  is even, there are  points of each color, and no two cyclically
consecutive points in the indexing lie on the same side of H.

In [5], the 2-sets having no cyclically consecutive elements are the pairs of
the form {i , i + 2}, computed modulo 5. In the Petersen graph, which is K(5 , 2),
these vertices induce C5 ; already (C5) = 3. Schrijver proved that one can always
restrict K(n , ) to such sets without reducing the chromatic number.

16.2.34. Corollary. (Schrijver [1978]) The subgraph of K(n , ) induced by the -

sets with no cyclically consecutive elements has chromatic number n − 2  + 2.
Proof: Let d = n − 2  , and let X be an n-set in S d as guaranteed by Theorem
16.2.33, indexed in order. Let G be the specified subgraph of K(n , ); vertices
of G correspond to -sets in X . If G has a proper (d + 1)-coloring  , then define
A1 , . . . , Ad+1 ⊆ S d by putting x in Ai if some vertex of color i in G corresponds to
a -set contained in the open hemisphere of S d centered at x.
By Theorem 16.2.33, A1 , . . . , Ad+1 are open sets covering S d . By the
Lyusternik–Shnirel’man Theorem (open set version), some Ai contains antipo-
dal points, so  cannot be a proper coloring.
Section 16.2: Combinatorial Topology 815

HAM SANDWICHES AND BISECTIONS

The Ham Sandwich Theorem is a well-known application of the Borsuk–Ulam

Theorem. The name arises from a special case: the ham, cheese, and bread of a
ham sandwich can be simultaneously bisected by a single knife cut. The theorem
was conjectured by Steinhaus and proved by Banach (see Steinlein [1985]).
The theorem holds for general measures, which are beyond our scope to de-
fine carefully. Hence we restrict to the case where each object is a nonnegative
integrable function; we call it a mass. Integrating over any full-dimensional vol-
ume gives a portion of the mass; integrating over a lower-dimensional set like a
hyperplane gives 0. For a mass Þ , we write Þ(S) for the integral over a volume S.

16.2.35. Theorem. (Ham Sandwich Theorem) Given masses Þ1 , . . . , Þd in d

,
some hyperplane cuts d into half-spaces that contain half of each mass.
Proof: (sketch) A half-space H + in d is specified by constants u0 , . . . , ud (not
all zero) such that H + = {x ∈ d : ∑i=1 ui x i < u0}. Furthermore, the coefficient
d

vector can be normalized so that (u0 , . . . , ud) ∈ S d . For u ∈ S d , let H +(u) denote
the half-space specified by this correspondence.
Define º : S d → d by ºi(u) = Þi(H +(u)). Since H +(−u) = {x ∈ d : ∑i=1 ui x i >
d

u0}, the half-spaces associated with u and −u are disjoint and omit only the hy-
perplane defined by ∑i=1 ui x i = u0 . If º is continuous, then the Borsuk–Ulam
d

Theorem (Theorem 16.2.26B) provides u ∈ S d such that º (u) = º (−u). This can-
not occur when u = (1 , 0 , . . . , 0), since then H +(u) = d and H +(−u) = ∅. Thus
º (u) = º (−u) provides a hyperplane that bisects each mass.
We omit the argument that º is continuous. In more general versions of the
theorem this is a matter of some delicacy, but for the integrable functions we use
as masses it is intuitive.

16.2.36. Remark. Equipartitioning. Using the Ham Sandwich Theorem, it is

easy to show that any mass distribution in the plane can be partitioned into four
equal parts by two lines (Exercise 23). It is possible (but not easy) to partition a
distribution in 3 into eight equal parts by three planes (for partitioning point
sets, see Hadwiger [1966] and Yao–Dobkin–Edelsbrunner–Paterson [1989]).
It is not generally possible to split a set in d into 2 d equal parts by d hyper-
planes when d ≥ 5 (Exercise 23). The question remains open when d = 4 (see Exer-
cise 24 for points along the moment curve). What is now called the Gr ünbaum–
Hadwiger–Ramos problem asks for the smallest dimension d such that for
any m masses in d there are ¾ hyperplanes that cut each mass into 2¾ equal
pieces. The Ham Sandwich Theorem settles the case ¾ = 1, where the answer is
m. Ramos [1996] proved the general lower bound ⌈ 2 ¾−1 m⌉ , which is conjectured
¾

to be the correct answer. See Blagojević–Frick–Haase–Ziegler [2016, 2018] for a

summary and analysis of known results.

Other interesting applications arise when the Ham Sandwich Theorem is ap-
plied to bisect finite sets of points. Recall that a set of points in general position
in d has no more than d points on any hyperplane.
816 Chapter 16: Geometry and Topology

16.2.37. Theorem. If A1 , . . . , Ad are disjoint finite sets whose union is in gen-

eral position in d , then there is a hyperplane H in d such that each open
half-space with boundary H contains exactly ⌊ 12 | Ai |⌋ points from each Ai .
Proof: For each i, proceed as follows. If | Ai | is odd, then let Bi = Ai ; otherwise,
delete any one point from Ai to obtain Bi . Form B i by replacing each point in Bi
with a ball of radius centered at that point, where is chosen small enough that
each hyperplane intersects at most d of the balls over all i.
Theorem 16.2.35 provides a hyperplane H that bisects all of B1 , . . . , B d .
Since each | Bi | is odd, H intersects a ball from each B i , and hence it intersects
exactly one ball from each. Furthermore, H passes through the center of each of
those balls, and hence H bisects each Bi .
Since H contains one point from each Bi , when | Ai | is even the missing point
cannot lie on H. Returning the missing point thus gives one of the half-spaces
half of Ai . Let S and T be the sets of points on H from the sets Ai of even size
and odd size, respectively. Move each point of S a small distance into the half-
space containing half of the corresponding set, turning S into a new set S . Now
S ∪ T has size d and determines a unique hyperplane H . If is small enough,
then moving from H to H does not pass the hyperplane through any points in
⋃i=1 Ai . Now the points of S are on the side of H that was deficient for each such
d

i, and H bisects all of the sets.

From this the necklace result follows.

16.2.38. Corollary. Every open necklace with  colors of beads can be cut using
at most  cuts, so that the union of odd-indexed intervals captures half the
beads of each color, with each color having an odd number of beads rounded
up or down as specified arbitrarily.
Proof: Place the beads of the necklace in order as points along the moment curve.
Let H be a hyperplane as guaranteed by Theorem 16.2.37, bisecting the  point
sets. The cuts of the necklace correspond to where H crosses the moment curve.
No color with an even number of beads has a point on H , but each color with an
odd number of beads does. The cut made at each such bead can be put before
or after the bead depending on whether the half of that color in the odd-indexed
intervals should be rounded up or down.

Given n red points and n blue points in 2 , one can find a pairing so that the
resulting n segments are disjoint: just take a pairing to minimize the total length
of the segments. Like the necklace problem, this has a higher-dimensional gener-
alization that is nontrivial but follows easily from the Ham Sandwich Theorem.

16.2.39. Theorem. (Akiyama–Alon [1989]) If A1 , . . . , Ad are sets of n points

in general position in d , then the union splits into n sets, each having one
point from each Ai , whose convex hulls are disjoint.
Proof: We use induction on n; for n = 1, the union is a single set. For n > 1,
let H be a hyperplane as guaranteed by Theorem 16.2.37. If n is odd, then the
points on H provide one of the sets. Since a half-space contains the convex hull of
all its points, it suffices to apply the induction hypothesis separately to each open
half-space bounded by H.
Section 16.2: Combinatorial Topology 817

BORSUK’S CONJECTURE

The Lyusternik–Shnirel’man version of the Borsuk–Ulam Theorem states

that any covering of the sphere S d−1 in d by d closed sets puts some pair of an-
tipodal points into the same set. The distance between antipodal points is 2. Since
S d−1 is a set of diameter 2 in d , one way to interpret the theorem is that S d−1
has no partition into d sets that have diameter smaller than the full set.
Borsuk [1933] asked whether every bounded set in d (size at least 2) decom-
poses into d + 1 pieces having smaller diameter than the full set. The problem
intrigued many people who expected a “yes” answer, so it became known as Bor-
suk’s Conjecture, although Borsuk never conjectured an answer. The set of
d + 1 vertices of a regular simplex already needs d + 1 pieces.
Borsuk obtained a positive answer when d = 2. Eggleston [1955] proved it for
d = 3 (see also Heppes–Revész [1956]). For general d, Hadwiger [1946] proved it
for bodies with smooth boundaries, Riesling [1971] for centrally symmetric bod-
ies, and Dekster [1995] for bodies of revolution.
Let º (d) be the least ¾ such that every bounded set in d decomposes into
¾ pieces of smaller diameter. It is known that √º (4) ≤ 9. Lassak [1982] proved
º (d) ≤ cd , and Schramm [1988] proved º (d) ≤ ( 3/2 + o(1))d .
In a 3-page paper, K ahn–K alai [1993] answered the original question neg-
atively. They produced a finite set that, when d = 1326, requires √ more than
d + 1 pieces of smaller diameter, and in general yields º (d) ≥ 1 .1 . In just two
d

pages, Nilli [1994] (pseudonym of Alon) reduced the counterexample to d = 946.

Further improvements
√ obtained counterexamples with d = 561 and º (d) ≥
(1 .2255 + o(1)) d) (Raı̆gorodskiı̆ [1997]), d = 321 (Pikhurko [2002]), d = 298
(Hinrichs–Richter [2003]), and d = 65 (Bondarenko [2014]). The last shows the
existence of a two-distance set (Definition 15.1.4) of 416 points in 65 that cannot
be partitioned into 83 sets of smaller diameter.
The idea of K ahn and K alai is as follows. Create a set of n-tuples such that
2
only small subsets avoid having orthogonal pairs. Use these to form points in n
so that the diameter of the set is achieved precisely by the pairs of points arising
from orthogonal pairs in the original set. Since sets of smaller diameter must
avoid orthogonal pairs, many sets will be needed.
Recall that a family of sets is L-intersecting if | A ∩ B| ∈ L for A , B ∈ .
The lemma about the sets in n uses a variant of the Frankl–Wilson Theorem
(Theorem 15.1.14), which states that every L-intersecting family of subsets of
[n] has size at most ∑|iL=|0 (ni). For n-tuples in {−1 , 1}n , dot product plays a role
analogous to intersection for n-tuples in {0 , 1}n .

16.2.40. Lemma. Let n = 4p, where p is an odd prime. Let V be the set of
n-tuples in {−1 , 1}n such that the first coordinate is positive and the total
number of positive coordinates is even. If U is a subset of V containing no
p−1
two orthogonal vectors, then | U | ≤ ∑i=0 (ni).
Proof: Consider u , v ∈ U. Since u and v agree in the first coordinate, u · v
= − n.
If u
= v, then u · v
= n. Since u , v ∈ U , they are not orthogonal, so u · v
= 0.
We claim u · v ≡ 0 (mod 4). Let A and B be the subsets of [n] consisting of the
positions where u and v are positive, respectively. Now u · v = 4p − 2 | A B| , where
818 Chapter 16: Geometry and Topology

 denotes symmetric difference. Since | A B| = | A| + | B| − 2 | A ∩ B| , and | A| and

| B| are even, the claim follows. Now, since p is odd, u · v
≡ 0 (mod p) when u = v.
To bound | U | , we define polynomials over p . For u ∈ U , define ºu by ºu(x) =
p−1
∏i=1 (u · x − i). If v = u, then u · v = 4p, so ºu(v)
≡ 0 (mod p). On the other hand,
ºu(v) ≡ 0 (mod p) when v
= u, since u · v
= 0 (mod p) makes one factor in the
definition of ºu divisible by p.
Now we capture |U | linearly independent polynomials in a space of dimension
p−1
∑i=0 (ni). Each polynomial ºu is a product of p − 1 linear factors and hence has
total degree at most p − 1 in each monomial. Define a new polynomial ºˆu from ºu
by reducing the exponents modulo 2 in each monomial. Since x 2 j = 1 and x2 j +1 =
x when x ∈ {−1 , 1}, we have ºˆu(x) = ºu(x) for all x ∈ {−1 , 1}n .
Therefore, {ºˆu : u ∈ U} are polynomials such that ºu(v)
≡ 0 (mod p) if and
only if v = u, for u , v ∈ U. By the “Diagonal Criterion” (Section 15.1), these
polynomials are linearly independent in the space of all polynomials in n variables
over p . That is, if ∑u∈U cu ºˆu(v) = 0 for all v ∈ U , then all coefficients must be 0.
Finally, each polynomial ºˆu is a linear combination of monomials that are
p−1
products of at most p − 1 variables, each with degree 1. There are ∑i=0 (ni) such
p−1
monomials. Hence | U | ≤ ∑i=0 (ni).

16.2.41. Theorem. (K ahn–K alai [1993]) If all bounded sets√ in

d
decompose into
º (d) sets with smaller diameter, then º (d) ≥ (1 .203) for large enough d.
d

Proof: (Nilli [1994]) Choose n = 4p, where p is a prime, and define V ⊆ n as

in Lemma 16.2.40. For v ∈ V , define p(v) ∈ n by letting the (i , j)-coordinate of
2

p(v) be vi vj . Let P = {p(v): v ∈ V}.

For u , v ∈ V , we have
p(u) · p(v) = ∑i=1 ∑ j =1 ui uj vi vj = ∑i=1 ui vi ∑ j =1 uj vj = (u · v)2 ≥ 0.
n n n n

Furthermore, each p(v) has length n. Since V does contain orthogonal pairs of
vectors and p(u) · p(v) ≥ 0, it follows that the distance between p(u) and p(v) equals
the diameter of P if and only if u · v = 0.
Let m(n) = ∑i=0 (ni). By Lemma 16.2.40, each subset of P with no orthogonal
n/4

pairs has size at most m(n). Hence the number of sets needed to partition P into
pieces with smaller diameter is at least | P | /m(n), which equals 2 n−2/m(n).
For a slight further improvement, note that P is spanned by a set of (n2) vec-
√
tors (Exercise 26). With d = (n2), we have n ≈
√
2d, and º (d) ≥ 2 n−2/m(n) >
for large d. In the computations, we use ∑ (ni) < 32 (n/4
n/4
(1 .203) d n
) (see The-
√ i=0
orem 14.4.4) and (n/4) ≈ 8/(3 n)( 33/4
n 4 )n (by
Stirling ’s Formula). Thus  (d) ≥
√ √
c(33/4/2)n ≈ (1 .136)n . With n ≈ 2d, we then raise 1 .136 to the 2 power.

EXERCISES 16.2

16.2.1. (−) Compute the bandwidths of Pn , K n , and Cn .

16.2.2. (−) Compute the bandwidth of the complete multipartite graph K n1 ,... ,n¾ with n1 ≥
· · · ≥ n¾ and ∑ ni = n. (Eitner [1979])
Exercises for Section 16.2 819

16.2.3. Prove that the local density lower bound for B(Pn Pn) is about n/2.

16.2.4. Let S be a set of ¾ vertices in the grid Pn Pn .

(a) Prove that | S| is minimized for | S| =  by a set occupying a Ferrers diagram; that
is, the first ai vertices in row i, with a1 ≥ · · · ≥ an and ∑i=1 ai =  .
n

n+1
(b) For each  such that (2) <  < ( 2 ) , prove that | S| ≥ n whenever S ⊆ V(G) with
n

=
| |  . Thus the boundary bound is n. (Chvatalová [1975])
S

16.2.5. (♦) Prove that B(Cn Cn) = 2n − 1 when n is odd. (Hint: For the lower bound, use
the technique of Theorem 16.2.8. Comment: The answer is also 2n − 1 when n is even, but
the argument for the upper bound is messier then.) (Li–Tao–Shen [1981])

16.2.6. Prove that every tree with  leaves is the union of ⌈ /2⌉ pairwise intersecting
paths, and use this to prove that the bandwidth of a tree with  leaves is at most ⌈ /2⌉ .
(Ando–Kaneko–Gervacio [1996])

16.2.7. (♦) Define a graph on the infinite grid 2 by making two vertices adjacent if they
are consecutive along a line with slope in {0 , ∞ , 1 , −1} (thus the graph is 8-regular). Prove
that for any set S having vertices from r rows and s columns, the boundary of S is at least
2r + 2s + 4. (Hint: The lower bound is immediate when S is an r-by-s rectangle. Comment:
This lemma was applied in Balogh–Kaul [2007] to random geometric graphs.)

16.2.8. (+) Bandwidth of caterpillars. Let G be a caterpillar, and let  be an integer such
that ⌈ |V(H)|−1
diam H
⌉ ≤  for all H ⊆ G. Prove that B(G) ≤  . (Hint: Prove that G has a number-
ing  in which  (v) is a multiple of  whenever v is on the spine and | (u) −  (v)| ≤  for
all uv ∈ E(G). (Sysło–Zak [1982], Miller [1981])

16.2.9. Let G be a graph with order n and bandwidth b.

(a) For e ∈ G, prove that B(G + e) ≤ 2b.
(b) Prove that if n ≥ 6b, then B(G + e) can be as large as 2b.
(Comment: The maximum of B(G + e) is b + 1 if n ≤ 3b + 4 and is ⌈ (n − 1)/3⌉ if 3b + 5 ≤
n ≤ 6b − 2.) (Wang–Yao–West [1995])

16.2.10. Prove Brouwer ’s Theorem (Theorem 16.2.18) from Sperner ’s Lemma.

16.2.11. Prove Brouwer ’s Theorem (Theorem 16.2.18) from the Borsuk–Ulam Theorem.

16.2.12. Prove the general form of Tucker ’s Combinatorial Lemma (Theorem 16.2.21)
from the Borsuk–Ulam Theorem (Theorem 16.2.23). That is, if T is a simplicial subdi-
vision of Bn that is antipodally symmetric on the boundary, and the vertices are labeled
from {±1 , . . . , ± n} so that labels on antipodal points of the boundary are complementary,
then there is a complementary edge.

16.2.13. Consider a circular necklace with beads of  colors. To capture half the beads of
each color from such a necklace, an even number of cuts must be made. Show that 2⌈ /2⌉
cuts suffice, and that when  is odd the first cut can be made between any two beads.

16.2.14. Prove that in Theorem 16.2.23, the second statement implies the first. That is,
if every continuous antipodal function from S n to n is somewhere zero, prove that every
continuous function from Bn to n that is antipodal on S n−1 is somewhere zero.

16.2.15. Prove that the following statement is equivalent to Theorem 16.2.26C: There is
no continuous  : Bn → S n−1 that is antipodal on S n−1 .

16.2.16. For  ∈ with 0 <  < 2, the Borsuk graph B(n + 1 , ) has vertex set S n , with
points adjacent when separated by distance at least  . Prove that the Borsuk–Ulam The-
orem is equivalent to the statement that (B(n + 1 , )) ≥ n + 2 for all n and  .
820 Chapter 16: Geometry and Topology

16.2.17. Prove the following statements.

(a) If G is a vertex-transitive graph, then º (G) = | V(G)| /(G).
(b) The Kneser graph is vertex-transitive, and º (K(n , )) = n/ .

16.2.18. (♦) Let P be the inclusion poset on the -sets and (n − )-sets in [n]. Apply The-
orem 16.2.29 to prove dim P ≥ n − 2  + 2 for n > 2  . (Hint: Given a realizer L1 , . . . , Lt ,
consider the unforced pairs consisting of a -set and its complement.) (Füredi [1994])

16.2.19. Kneser from Gale. For d ≥ 0 and  ≥ 1, Gale [1956] proved that there is a set X
of 2  + d points in S d such that every open hemisphere contains at least  points of X . Use
such a set X to give a proof of Theorem 16.2.29. (Hint: Given a proper (n − 2  + 1)-coloring
of K(n , ), use X to define appriopriate open sets A1 , . . . , A d+1 in S d and apply the open
set version of the Lyusternik–Shnirel’man Theorem.) (Bárány [1978])

16.2.20. Given a family of sets, the generalized Kneser graph K( ) has a vertex set ,
with vertices adjacent when the corresponding sets are disjoint.
(a) Prove that every graph is a generalized Kneser graph.
(b) Construct families of sets such that is the edge set of a hypergraph with chro-
matic number 2, but (K( )) is arbitrarily large.

16.2.21. (♦) Let G be the generalized Kneser graph of a family , defined in Exercise
16.2.20. Let c( ) be the minimum size of a set Y ⊆ ⋃( ) such that the hypergraph whose
edges are the members of disjoint from Y is 2-colorable. Prove (G) ≥ c( ). Show
c(([n]
¾ )) = n − 2  + 2. (Hint: Use the Lyusternik–Schirelman Theorem.)

16.2.22. (♦) The Schrijver graph S(n , ) is the subgraph of the Kneser graph K(n , ) in-
duced by the vertices that are -sets of [n] whose elements are nonconsecutive, viewed
cyclically. Recall that (S(n , )) = n − 2  + 2.
(a) Schrijver proved that S(n , ) is vertex-color-critical (deleting any vertex reduces
the chromatic number). Prove that not all Schrijver graphs are edge-critical.
(b) Prove that S(n , ) is guaranteed to be regular only when  = 2.

16.2.23. (♦) Consider a mass distribution in d (no mass at points).

(a) For d = 2, use the Ham Sandwich Theorem to show that the distribution can be
split into four parts of equal measure by two lines. (Comment: It also holds for d = 3 that
a distribution can be split into eight equal parts by three planes.)
(b) For d = 5, use the moment curve to show that it is not always possible to split a
distribution into 32 equal parts using five hyperplanes. (Avis [1984])

16.2.24. Construct a spanning cycle in the 4-dimensional hypercube Q4 using four edges
in each direction. Use this to show that any 16 distinct points on the moment curve in 4
can be separated using 4 hyperplanes. (Comment: Robinson–Cohn [1981] showed that Qd
has a spanning cycle with 2 d/d edges in each direction if and only if d is a power of 2.)

16.2.25. Bisection of a point set A by a hyperplane is defined by requiring ⌊ | A| /2⌋ points

of A in each open half-space bounded by the hyperplane. Show that if the definition of
bisection of A is modified by letting each half-space count “one-half ” for each point on the
hyperplane and requiring each half-space to capture | A| /2, then there are pairs of point-
sets in 2 that cannot both be bisected by one line.

16.2.26. Prove that there is a set S of (n2) vectors in n such that every point in the
2

set P in Theorem 16.2.41 is a linear combination of vectors in S. (Comment: Although

(v2 , . . . , vn−1) determines p(v), this dependence is not linear. For example, on the moment
curve the first coordinate determines the rest , but any n points on the moment curve in
n
are independent , by the Vandermonde determinant.)
Section 16.3: Volumes and Containment 821

16.3. Volumes and Containment

In this section we consider combinatorial aspects of volume and position in
high-dimensional spaces. We discuss monotone sublists of random permutations,
sorting algorithms starting from partial information about an underlying linear
order, and geometric representations of partial orders. The common theme is con-
verting problems of counting permutations into problems of computing volumes.

MONOTONE SUBSEQUENCES

Ulam asked for the expected maximum length of a monotone sublist in a ran-
√ permutation of [n]. The Erdős–Szekeres Theorem guarantees always at least
dom √
n. Hammersley [1972] proved that the expected maximum is asymptotic to c n
for some c. Logan–Shepp [1977] proved c ≥ 2; Versik–Kerov [1977] proved c ≤ 2.
Pilpel [1990] provided a combinatorial proof of the upper bound. √
Here we show only that the expected maximum length is between (1 − 1/e) n
√
and e n, but we extend to higher dimensions. The proof of the lower bound follows
ideas in Winkler [1985] and Bollobás–Winkler [1988]. Recall from Chapter 10
that a d-dimensional permutation has a permutation of [n] in each coordinate;
we can view it as a d-tuple of permutations of [n]. Also, a list of d-dimensional
vectors is monotone if in each coordinate the list of values is monotone; some
coordinates may be strictly increasing and some weakly decreasing.

16.3.1. Theorem. For large n, the expected maximum length º (n) of a monotone
sublist of a d-dimensional permutation of [n] satisfies
(1 − 1/e)n1/(d+1) < º (n) < en1/(d+1) .

Proof: For the upper bound, we consider only d = 1, leaving the generalization
as Exercise 3. For the lower bound, we treat all dimensions.
Upper bound. Let X be the random variable counting monotone sublists
of length t. Each such list occurs in some t positions. For a given set of t posi-
tions, the probability that they form a monotone list is 2/t!. By the linearity of
expectation, (X) ≤ (nt) t!2 ≤ 2nt/(t!)2 . Using Stirling ’s Formula to approximate
√
t!, the bound becomes (ne2/t2)t/( t). In particular, if t ≥ e n, then (X) → 0. By
√
Markov ’s Inequality, whp there is no monotone sublist with length at least e n,
so the expected maximum length is less than that.
Lower bound. Instead of studying random d-permutations of [n], we study
n random points in a unit cube in d + 1 dimensions. Each coordinate of each point
is drawn uniformly from [0 , 1]. Equal values occur with probability 0, so each co-
ordinate gives an ordering. The first d coordinates give permutations by ordering
the indices according to the last coordinate. A chain in the product then means
822 Chapter 16: Geometry and Topology

points having the values in each coordinate in the same order. This allows us to
ignore the order of the generated points; we just look for a chain in the set.
When generating the permutations from n random points, each set of per-
mutations has the same probability. The jth coordinate gives the permutation
such that 0 ≤ a −1(1) < · · · < a −1(n) ≤ 1, where ai is the jth coordinate of the point
whose last coordinate is ith smallest. The n-tuples in the jth coordinate that map
to form an n-dimensional simplex in [0 , 1]n with volume independent of . (We
will revisit this simplex in Definition 16.3.7.) The choices of n points in the cube
mapping to a particular d-permutation form a cartesian product of d such sim-
plices. Thus when we generate a random element of ([0 , 1]d+1 )n , the volume (the
probability) that maps to any d-permutation is the same.
We claim that n random points in [0 , 1]d+1 expect to include a chain of the de-
sired size among chains of a restricted type. For large n, consider ⌊ n1/(d+1)⌋ small
cubes along the main diagonal in the unit cube; each has side length 1/⌊ n1/(d+1)⌋
and volume at least 1/n. The upper corner of each cube is the lower corner of the
next. If the n chosen random points hit X of these small cubes, then picking a
point from each of those cubes yields a chain of size X .
It thus suffices to show that (1 − 1/e)n1/(d+1) is a lower bound on (X). We
have X = ∑ X i , where X i = 1 if and only if the ith small cube is nonempty. It
is empty only if the n independent points all miss it, which has probability at
most (1 − 1/n)n since the small cube has volume at least 1/n. This probability is
bounded by 1/e, so (X i = 1) > 1 − 1/e. By the linearity of expectation, (X) =
⌊ n1/(d+1)⌋ (X i) > n1/(d+1)(1 − 1/e), where we have ignored lower-order terms.

Generating n points in a (d + 1)-dimensional cube is a way to generate random

(d + 1)-dimensional posets; Theorem 16.3.1 gives bounds for the expected height.

BAL ANCED COMPARISONS

We return now to partial orders. We view a poset P as partial information

about an underlying linear order on its elements, as in Section 12.4. Sorting is
the problem of finding the true underlying linear order, using comparisons be-
tween elements. When P consists of disjoint chains, sorting is called merging.
Given no prior information, P is an antichain, and the answer is one of n!
linear orders. In general, the answer is a linear extension of P. Each comparison
splits the remaining possible orders into two sets, and the result is consistent
with one of them. Since an adversary can choose which set remains, in the worst
case any algorithm uses at least lg e(P) comparisons, where we write e(P) for the
number of linear extensions of P.
The value lg e(P) is called the information-theoretic lower bound. The
bound is achieved if every poset has a comparison that splits the remaining ex-
tensions into two equal-sized sets. Fortunately, we do not need perfect balance to
obtain performance within a constant factor of the lower bound.

16.3.2. Definition. For x , y ∈ P , the probability of the event x < y, written

(x < y), is the fraction of the linear extensions of P putting x below y. A
pair (x , y) of elements is -balanced if ≤ (x < y) ≤ 1 − .
Section 16.3: Volumes and Containment 823

If every poset contains a -balanced pair, then sorting from initial informa-
tion P takes at most log1/(1− ) e(P) comparisons. The closer is to 12 , the better.

16.3.3. Conjecture. ( 13 , 23 -Conjecture; K islicyn [1968], Fredman∼1975, Linial

[1984]) Every poset has a 13 -balanced pair.

16.3.4. Example. The poset 1 + 2 has just three linear extensions, so (x < y) ∈
{0 , 13 , 32 , 1} for all (x , y). Thus Conjecture 16.3.3 is best possible.

16.3.5. Remark. Conjecture 16.3.3 remains open, but K ahn–Saks [1984] proved
that = 113 3
suffices: in any finite poset, 11 < (x < y) < 11
8
for some pair (x , y).
Brightwell–Felsner–Trotter [1995] improved the guarantee to ≥ 12 − √1 for
2 5
finite posets and many infinite posets, using the machinery of K ahn–Saks plus
the Ahlswede–Daykin Inequality (see also Felsner–Trotter [1993]).
There is reason to think that the conjecture might fail on a semiorder (Defini-
tion 12.4.1). Brightwell [1988] constructed infinite semiorders where one cannot
guarantee a -balanced pair with > 12 − √1 (Exercise 8), but Brightwell [1989]
2 5
proved the 13 , 23 -conjecture for finite semiorders (Exercise 9). Brightwell–Wright
[1992] proved it for posets in which every element is incomparable to at most five
others, Trotter–Gehrlein–Fishburn [1992] for bipartite posets, and Olson–Sagan
[2018] for some other classes. Brightwell [1999] provides a survey.
The K ahn–Saks result implies that sorting can be completed using about
2.2 · lg e(P) comparisons. The proof does not produce a balanced comparison in
polynomial time, but K ahn–K im [1995] later addressed this using entropy (Sec-
tion 11.2), which is easy to compute, unlike e(P) (see Brightwell–Winkler [1991]).
Fredman [1976] presented an algorithm using lg e(P) + 2 | P | comparisons.
K ahn and Saks proposed a more detailed conjecture. Let w be the largest
such that every poset of width w has a -balanced pair. They conjectured that
w increases with w, approaching 2 from below. Komlós [1990] used functional
1

analysis to prove this for bipartite posets. It is known that 2 = 13 and 3 > 13 .
The proof below that 2 = 13 is easy; this case includes merging of two lists.

16.3.6. Theorem. (Linial [1984]) Posets of width 2 have 13 -balanced comparisons.

Proof: A poset P of width 2 is covered by two chains, by Dilworth’s Theorem;

let them be x1 , . . . , x n and y1 , . . . , ym . We may assume x1  y1 (otherwise P has a
unique minimum and we reduce the problem) and (x1 < y1) < 12 . In terms of i,
(x1 < yi) is strictly increasing. Let j be the largest index with (x1 < yj) < 12 .
We have j < m unless P is a chain, because there is at most one extension with
x1 > ym . If (x1 < yj) ≥ 13 , then we have a 13 -balanced comparison. Otherwise,
1
2 ≤ (x1 < yj +1) = (x1 < yj) + (yj < x1 < yj +1) ≤ 2(x1 < yj) < 23 ,
where the inequality (yj < x1 < yj +1) ≤ (x1 < yj) holds because x1  yj allows x1
and yj to be switched on any extension with x1 covering y j .
824 Chapter 16: Geometry and Topology

The numerical aspects of the K ahn–Saks [1984] 11 3 8

, 11 -result are quite diffi-
cult. Instead, we present the later and easier argument of K ahn–Linial [1991]
1
that there is always a 2e -balanced pair; since 2e1
< 11
3
, this is weaker. The idea
is to consider the average height of an element over all linear extensions. It is
easy to show that for some pair of elements this differs by less than 1. The task
1
then is to show that any such pair is a 2e -balanced pair. We use a geometric object
associated with a linear extension.

16.3.7. Definition. For a poset Q with elements q1 , . . . , qn , the order poly-

tope O(Q) is the set of vectors (º (q1) , . . . , º (qn)) in n such that º is an
order-preserving function from Q to [0 , 1]. For a linear extension of Q, the
canonical simplex ( ) in O(Q) is the set of vectors (x1 , . . . , x n) such that
0 ≤ x −1(1) ≤ · · · ≤ x −1(n) ≤ 1. The centroid of an n-dimensional region R is
the vector c(R) whose ith coordinate is (∫ R x i dV)/(∫ R dV).

16.3.8. Lemma. (Stanley [1981]) Let Q be a poset with elements q1 , . . . , qn , or-

der polytope O(Q), and canonical simplex ( ) for each linear extension .
(a) O(Q) is the union of ( ) over all linear extensions of Q.
(b) ( ) has volume n! 1
, independent of .
(qi)
(c) c( ( ))i = n+1 , where c( ( )) is the centroid of ( ).
(d) Vol (O(Q)) = e(P)n! and c(O(Q))i = n+1 , where hi = e(P) ∑
hi 1 (qi).
Proof: (a) A vector in n satisfying all inequalities imposed by pairs in Q has its
coordinates in some order, which must correspond to a linear extension of Q.
(b) The canonical simplex is seen to be a simplex using the variables y0 , . . . , yn ,
where yi = x −1(i+1) − x −1(i) , with x and extended by (0) = 0, (n + 1) = n + 1,
x0 = 0, and x n+1 = 1. The n-dimensional volume of a “cone” is Bt/n, where B is
the (n − 1)-dimensional volume of the base and t is the height from base to apex.
The formula for these simplices follows by induction on n, since the height is 1.
(c) This can be evaluated by induction on n (Exercise 4).
(d) The canonical simplices for distinct linear extensions intersect only at
their boundaries (belonging to both requires equality among the coordinates for
which the corresponding elements of Q are ordered differently in the two exten-
sions). Hence the volume of the union O(Q) is the sum of the individual volumes.
Note that hi is the average height of qi on linear extensions. The second
equality is obtained from part (c) by averaging over the linear extensions, since
all canonical simplices have the same volume.

In this setting, the probability of an event A is the fraction of Vol (O(Q)) in

the canonical simplices for the extensions where A occurs. The event qi < qj is
described by the inequality  (qi) <  (qj ), and each canonical simplex lies on one
side of the hyperplane  (qi) =  (qj ). The K ahn–Linial proof studies the volume
of this subset of the order polytope using a theorem from convex geometry.
Given a convex body K in n , let H − and H + be two parallel hyperplanes that
are both tangent to K , and let H be the hyperplane halfway between them. Let
x̂ be a point in H , and let u be the vector from x̂ to a point in H + along a line per-
pendicular to H. For  ∈ [−1 , 1], the cross-section K ë of K is the intersection
of K with the hyperplane through x̂ +  u parallel to H.
Section 16.3: Volumes and Containment 825

H+
H
H−

16.3.9. Theorem. (Brunn–Minkowski Theorem) Let K be a convex body in n

with cross-sections K ë for ∈ [−1 , 1]. The function r defined by

r( ) = [Vol n−1(K ë)/Vn−1 ]1/(n−1)
is concave on [−1 , 1], where Vn−1 is a normalizing constant equal to the vol-
ume of the (n − 1)-dimensional unit ball.

We are viewing  (x) as the height of x on a linear extension  of Q, from 1

to | Q|. Let h(x) denote the average of  (x) over all linear extensions. Deleting a
unique minimum or maximum does not change the probabilities for pairs of other
elements. Hence we may assume 1 < h(x) < | Q| for x ∈ Q. By the Pigeonhole
Principle, some pair (x , y) satisfies | h(y) − h(x)| < 1. This requires x below y in
some extensions and x above y in others. We show that each case occurs in at least
1
the fraction 2e of the extensions.

16.3.10. Theorem. (K ahn–Linial [1991]) If in a poset Q the expected heights of

q and q satisfy h(q) − h(q ) > −1, then (q > q ) > 2e
1
.
Proof: We write q1 , . . . , qn for the elements of Q and take x , y ∈ n . Given qi  qj
and h(qi)− h(qj ) > −1, we prove (qi > qj ) > 2e 1
. Using Lemma 16.3.8, we rephrase
this in terms of the order polytope; let K = O(Q). With K + = {x ∈ K : x i − x j ≥ 0},
(K +)
we want to prove Vol Vol(K)
> 2e
1
when h(qi) − h(qj ) > −1.
We express K + using cross-sections of K . Let l be the line in n generated by
the vector w with 1/2 in coordinate i and −1/2 in coordinate j and 0 elsewhere.
The projection of x on l is w when x i − x j = . The cross-section of K in the
hyperplane through w perpendicular to l is thus {x ∈ K : x i − x j = }, which we
write as K ë . Note that K ë
= ∅ only when ∈ [−1 , 1]. Also, c(K)i − c(K)j > n−+11 .
We prove the desired inequality for any convex body K satisfying these two facts.
We use Theorem 16.3.9 to transform K into a “double cone” where computa-
tions will simplify. We first “symmetrize” K around the line l. Letting r( ) be
defined as in Theorem 16.3.9, we replace K ë with the (n − 1)-dimensional ball Bë
of radius r( ) centered at w and contained in the same hyperplane as K ë . Let
B = ⋃−1≤ ë≤1 Bë . Theorem 16.3.9 implies that B is convex. Also Vol (B) = Vol (K)
and Vol (B+) = Vol (K +), where B+ = ⋃ ë≥0 Bë . Furthermore, c(B)i − c(B)j =
c(K)i − c(K)j > n−+11 . Hence it suffices to prove the inequality for B+ and B.
The symmetrization reduces the problem to two dimensions, parallel and per-
pendicular to l. These become coordinates y1 and y2 , with B being the “solid of
revolution” of the curve y2 = r(y1) around the line l, which is now the y1 axis (see
figure below). In the (y1 , y2)-plane, let Q = (0 , r(0)), and choose P on the positive
y1 axis (at (u , 0)) so that the volume obtained by revolving PQ equals Vol (B+).
826 Chapter 16: Geometry and Topology

Since r(y1) is concave, u ≥ 1. Let S = (−1 , 0), and choose R on the line PQ so that
the double cone D obtained by revolving the triangle PRS has volume Vol (B).

R
(0 , r(0)) = Q

O
S = (−1 , 0) ( , 0) (1 , 0) (u , 0) = P

In both the left and the right half-planes, the volume under D is shifted to
t
the right from the volume under B. That is, ∫−1 [r(y1) − s(y1)]dy1 ≥ 0, where
s(y1) is the height of the piecewise linear curve SRP at y1 . This follows from the
concavity of r. Hence c D(y1) ≥ c B(y1). The problem thus reduces to showing the
following for the double cone:
If D is an n-dimensional double cone with apexes (−1 , 0 , . . . , 0) and
(u , 0 , . . . , 0) and centroid ( , 0 , . . . , 0) such that u ≥ 1 and ≥ −1/(n+ 1),
+)
then Vol(D
Vol (D)
> 2e1 .
Let C1 and C2 be the cones comprising D, with common base containing R, apexes
(−1 , 0 , . . . , 0) and (u , 0 , . . . , 0), and heights h1 , h2 with h1 + h2 = u + 1. Because
(C2) (C2)
C1 and C2 have the same base, we have Vol Vol (C1 )
= hh12 , or Vol
Vol(D)
= uh+21 . The volume of
an n-dimensional cone from a point is proportional to the nth power of its height,
so we can rewrite the ratio as
Vol (D+) Vol (D+) Vol (C2) u n h2
= · =( ) .
Vol (D) Vol (C2) Vol (D) h2 u + 1
Because this formula is minimized by maximizing h2 , which is equivalent to min-
imizing when Vol (D) is fixed, we may assume = n−+11 .
The distance from the apex to the centroid of an n-dimensional right cone is
n
n+1 times the height. Hence the centroids of C1 and C2 have y1 -coordinates −1 +
h1 n+n1 and u − h2 n+n 1 . Taking the weighted average of the centroids in proportion
to their volume yields the centroid of D, so

−1 1 h1 n h2 n
= [h1 (−1 + ) + h2 (u − )] .
n+1 u+1 n+1 n+1
Using h1 + h2 = u + 1 and solving for h2 yields h2 = nun
−1 . Substituting this
in our expression for the volume ratio and using u ≥ 1 yields
Vol (D+) n − 1 n−1 u 1
=( ) > .
Vol (D) n u+1 2e

The K ahn–Linial proof was inspired by the proof of Mityagin’s Theorem,

which says that the fraction of the volume on one side of a hyperplane through the
centroid of an n-dimensional convex body is at least ( n+n 1 )n (this is sharp for a sim-
plex and exceeds e−1). In the proof above, the hyperplane x i = x j passes through
Section 16.3: Volumes and Containment 827

the centroid of the order polytope if h(qi) = h(qj ). Hence when h(qi) = h(qj ) we
can conclude e−1 < (qi < qj ) < 1 − e−1 by Mityagin’s Theorem.
The K ahn–Saks [1984] proof uses the Alexandrov–Fenchel Inequalities for
mixed volumes. We only present a simpler application of these inequalities by
Stanley that motivated the K ahn–Saks approach.

16.3.11. Theorem. (Alexandrov–Fenchel Inequalities) Let K 0 , K 1 be two con-

vex sets in n , and let K ë = {(1 − )x + y: x ∈ K 0 , y ∈ K 1 } for 0 ≤ ≤ 1.
If the dimension of these sets is d, then the volume of K ë is a homogeneous
polynomial in of degree d given by
d
d
Vol (K ë) = ∑ ( )ai i(1 − )d−i ,
i
i=0
where the “mixed volumes” ai form a log-concave sequence.

Here “Inequalities” refers to a2i ≥ ai−1 ai+1 , the log-concavity of ⟨a⟩. The name
“mixed volume” for ai arises from a0 = Vol (K 0) and ad = Vol (K 1). A shorthand
notation for K ë is K ë = (1 − )K 0 + K 1 . Stanley applied Theorem 16.3.11 to ap-
propriate subsets of the order polytope to study the height distribution. Since
here we view a linear extension of an n-element poset as a map into [n], we say
“position  ” rather than “height  ”.

16.3.12. Theorem. (Stanley [1981]) Given a poset Q and an element x ∈ Q, the

number of extensions of Q in which x has position  is log-concave in  .
Proof: Let A = { ∈ O(Q): y ≤ x ⇒  (y) = 0}, and let B = { ∈ O(Q): y ≥ x ⇒
 (y) = 1}. The mixture K ë = (1 − )A + B is
K ë = { ∈ O(Q − x): (y ≤ x ⇒  (y) ≤ ) and (y ≥ x ⇒  (y) ≥ )} .
Note that K ë is (n − 1)-dimensional, since  (x) = for  ∈ K ë .
Theorem 16.3.11 gives one expression for Vol (K ë); we get another from the
decomposition of O(Q) into canonical simplices. Let  be a linear extension of
Q in which x has position  , and let S = K ë ∩ ( ). The vectors in S are the
order-preserving functions  respecting  and satisfying  (x) = . Since these
coordinates respect a fixed order and bound, the restriction of S to the first  − 1
elements in  is a ( − 1)-dimensional simplex of side-length . Similarly, the
restriction to the final n −  elements in  is an (n − )-dimensional simplex of
side-length 1 − . The vectors in S are all pairings of these, so S is the product of
n− ¾
the two simplices. By Lemma 16.3.8 its (n − 1)-dimensional volume is (¾ë−1)! (1(n−−ë)¾)! .
¾ −1

Since K ë is composed of its intersections with each of the canonical n-

dimensional simplices corresponding to extensions, we have
n ¾ −1 (1 − )n−¾
Vol (K ë) = ∑ e¾(x) ,
( − 1)! (n − )!
¾ =1
where e¾ (x) is the number of linear extensions of Q in which x has position  . On
n−1
the other hand, Vol (K ë) = ∑¾=0 (n−¾ 1)a¾ ¾ (1 − )n−1−¾ . Hence e¾ (x)/(n − 1)! = a¾−1 ,
and e¾ is a log-concave sequence in  .
828 Chapter 16: Geometry and Topology

CONTAINMENT ORDERS

Every family of sets, under inclusion, forms a poset. We have studied such
posets for subsets of a finite set, for down-sets in a poset, etc. Specifying a family
from which the sets must be drawn yields a “containment class” of posets.

16.3.13. Definition. The containment order of sets S1 , . . . , Sn is the poset P

on a fixed set x1 , . . . , x n given by x i < x j if and only if Si ⊂ Sj . The sets form
a containment representation of P. When the sets are constrained to lie
in a family , the resulting posets are -containment posets.

16.3.14. Example. When is the set of discs in the plane, the -containment
posets are the circle orders. When is the set of planar “wedges” separated
from the rest of the plane by two rays with a common endpoint, the -containment
posets are the angle orders. Allowing the set of polygonal regions bounded by a
path of ¾ segments yields the ¾-gon orders. We may also use boxes in t dimen-
sions, unions of t intervals, etc. When we must use finite sets, we are embedding
P in 2 t , where t is the size of the union of the sets, and we seek to minimize t.

In a ¾-dimensional poset, ¾ “criteria” specify the order relation. Intuition

suggests that we all ¾-dimensional posets should be -containment posets if and
only if sets in are specified with ¾ “degrees of freedom”.

16.3.15. Example. When is the set of ¾-dimensional boxes having one corner
at the origin and another with all coordinates positive, the -containment posets
are precisely the ¾-dimensional posets.
Angles are specified by four numbers: the coordinates of the center and the
angles of the rays. As expected, every 4-dimensional poset is an angle order (Ex-
ercise 12; Fishburn–Trotter [1985]), but some 5-dimensional posets are not. A
¾-gon is specified by 2 ¾ numbers, and every 2 ¾-dimensional poset is a ¾-gon order
(Exercise 16; Sidney–Sidney–Urrutia [1988]).
Circles in the plane are specified by three numbers: the radius and position of
the center. Some 4-dimensional posets are not circle orders. Hurlbert [1988] and
Scheinerman–Wierman [1988] proved that some infinite 3-dimensional posets are
not circle orders, leaving the finite case open. Finally, Felsner–Fishburn–Trotter
[1999] proved the existence of finite 3-dimensional posets that are not contain-
ment orders of spheres in any dimension; hence they are not circle orders.

16.3.16. Example. Containment posets of high dimension. Recall from Example

12.3.6 the “standard example” Sn , which is the subposet of 2[n] consisting of the
sets of size 1 and their complements. We show that Sn is both a circle order and
an angle order, as suggested in the figure below.

◦
◦ ◦

◦ ◦
◦
Section 16.3: Volumes and Containment 829

To express Sn as a circle order, assign the 1-set i a tiny disc Ci so that the
centers of these are equally spaced on the unit circle, and assign each (n − 1)-set
i a large disc so that these centers are equally spaced on a large circle centered
at the origin. Arrange them so that Ci does not contain the unit circle, and the
segment joining the centers of Ci and Ci contains the origin. Choose the radii so
that Ci does not contain Ci but contains all Cj with j
= i.
To express Sn as an angle order, assign the 1-sets thin wedges pointing down
from corners on a horizontal line, and assign the (n − 1)-sets reflex-angle wedges
with corners directly below the corners of their complements. Such a wedge con-
tains all the narrow wedges except the one for the complementary set.

Our objective now is a general argument to show that when has ¾ degrees
of freedom, not all (¾ + 1)-dimensional posets are -containment posets. At the
time, many instances of this question for particular were open. To prove the
result, we need a lower bound on the number of n-element posets of a fixed di-
mension and an upper bound on the number of distinct n-element posets that are
-containment posets. When the dimension is too large, the former exceeds the
latter. Here we take logarithms with respect to any base, since only the ratio of
the logarithms matters.

16.3.17. Theorem. (Alon–Scheinerman [1988]) If p(n , ¾) counts the posets on [n]

p(n ,¾)
n¾ log n = 1.
with dimension at most ¾ , then limn→∞ log
Proof: Every poset with dimension at most ¾ is the intersection of ¾ linear ex-
tensions. We obtain an upper bound by choosing ¾ arbitrary linear orders. Hence
p(n , ¾) ≤ (n!
¾ ) ≤ n /¾ !, which yields log p ≤ n¾ log n.
n¾

For the lower bound, we construct nearly this many distinct posets. A poset
P has dimension at most ¾ if there exists º : P → ¾ such that x ≤ y if and only
if º (x)t ≤ º (y)t for all x , y , t. We construct point-sets {º (r) ∈ ¾ : 1 ≤ r ≤ n} that
yield distinct posets on [n].
Let m be a parameter, which we will later set to n/log n. Define m¾ boxes in
¾ as follows: for 1 ≤ j ≤ ¾ and 1 ≤ i ≤ m, let B
i , j be the product of [i , i + 2 ]
1

in coordinate j and [0 , m + 1] in all other coordinates. For fixed j , we have m

“parallel” boxes; call this set of boxes Mj . The intersection of a family of boxes
consisting of one box from each of M1 , . . . , M¾ is a cube with side-length 12 ; in this
way we form m¾ small cubes.
We use these boxes to specify º (1) , . . . , º (n) in ¾ . Each box Bi , j is deter-
mined by two points in ¾ . For the first 2m¾ points in [n], let º (1) , . . . , º (2m¾)
be the determining corners of the boxes {Bi , j }, in a fixed order. Thus the subposet
induced by the elements 1 , . . . , 2m¾ in [n] is the same for each poset we construct.
For 2m¾ + 1 ≤ r ≤ n, let º (r) be a point inside one of the m¾ small cubes. There
are (m¾)n−2¾m ways of making these choices. They determine distinct subposets of
relations between [2m¾] and [n] − [2m¾], because r is between the points assigned
to the determining corners of the box Bi , j if and only if the small cube containing
º (r) is contained in Bi , j . Hence p(n , ¾) ≥ (m¾)n−2¾m . Setting m = n/log n yields
log p ≥ n(¾ − o(1)) log n.

Next we define “degrees of freedom” more precisely.

830 Chapter 16: Geometry and Topology

16.3.18. Definition. A family of sets is ¾-generated if the sets of are spec-

ified by an injection º : → ¾ and there is a finite list of polynomials {pj }
in 2 ¾ variables such that the containment S ⊂ T for S, T ∈ depends only
on the signs of pj (º (S) , º (T)).

16.3.19. Example. Discs are 3-generated. Let be the set of discs in the plane.
For each disc, let º give the radius r and the coordinates (x , y) of the center. For
discs C1 and C2 , we then have C1 ⊆ C2 if and only if r1 − r2 ≤ 0 and (x1 − x2)2 +
(y1 − y2)2 − (r1 − r2)2 ≤ 0. In this example, only one sign pattern corresponds to
containment, but in general there may be more than one.

We show next that when is ¾-generated, the number of -containment

posets on [n] grows more slowly than the number of (¾ + 1)-dimensional posets,
and hence not every (¾ + 1)-dimensional poset is a -containment poset. This is
an existence proof and does not construct a forbidden (¾ + 1)-dimensional poset.
For example, it implies only that some 4-dimensional poset is not representable
as a circle order. In fact, 2[4] − {∅ , [4]} is such a poset (Exercise 15).

16.3.20. Theorem. (Alon–Scheinerman [1988]) If is a ¾-generated family of

sets, then some (¾ + 1)-dimensional poset is not a -containment poset.
Proof: Each -containment poset on [n] is specified by n¾ real numbers, ¾ for each
i ∈ [n] to specify the set Si ∈ that represents i. Let p1 , . . . , pt be the polynomi-
als used to test containment for a pair of sets in . With n(n − 1) ordered pairs of
sets, we have tn(n − 1) signs whose pattern specifies the poset. Each polynomial
depends on only 2 ¾ variables, but there are n¾ variables altogether.
Distinct posets must have distinct signs for some pj applied to some pair
(Si , Sj). If we could have the full set of 3tn(n−1) sign patterns, then we might repre-
sent all ¾-dimensional posets, since the logarithm of 3tn(n−1) grows quadratically
with n, faster than n¾ log n. However, the limited number of variables restricts
the number of possible sign patterns.
The desired restriction is the result of Warren’s Theorem. Given r polyno-
mials in l variables with maximum degree d, Warren [1968] proved that the num-
ber of plus/minus sign patterns is bounded by (4edr/l)l if r ≥ l and e is the base of
the natural logarithms. This can easily be extended to bound the plus/zero/minus
patterns by (8edr/l)l (Exercise 20).
For our application, set r = tn(n − 1) and l = n¾ . Since t , ¾ , d are constants
independent of n, the values of r and l yield (cn)n¾ as an upper bound on the
number of sign patterns, for some constant c. This also bounds the number of -
containment posets on [n]. The logarithm of the bound is asymptotic to n¾ log n,
so by Theorem 16.3.17 not all (¾ + 1)-dimensional posets are representable.

Given a ¾-generated family , we can now ask whether every finite ¾-

dimensional poset is a -containment order. As a “no” example, the family
of vertical rays in the upper half-plane is 2-generated, but all -containment
posets are disjoint unions of chains, omitting most 2-dimensional posets (Alon–
Scheinerman [1988]). We have noted that the answer is “yes” for ¾-dimensional
boxes, angles, and t-gons, but it is “no” for discs. In general, the question is hard.
Exercises for Section 16.3 831

EXERCISES 16.3

16.3.1. (−) Show that if is the set of ¾-dimensional axis-parallel boxes, then the -
containment posets are precisely the 2 ¾-dimensional posets. (Dushnik–Miller [1941] for
¾ = 1; Golumbic–Scheinerman [1989] for all ¾)

16.3.2. Give a formula for the number of linear extensions of m × n. (Hint: Model the
linear extensions using objects counted in Part I of this book.)

16.3.3. (♦) Prove that the expected maximum length of a monotone sublist of a d-
dimensional permutation of [n] is at most en1/(d+1) . (Hint: See Theorem 16.3.1.)

16.3.4. Let Q be a poset with n elements. Suppose that element qi ∈ Q has height ¾ on a
linear extension . Prove that c( ( ))i = n+¾ 1 (statement (c) of Lemma 16.3.8).

16.3.5. Merging many chains. Let P be a disjoint union of  chains with sizes n1 , . . . , n¾ .
(a) Determine the information-theoretic lower bound for the number of comparisons
needed to sort from the partial information P.
(b) Construct algorithms that come close to this lower bound when all but one chain
have size 1 and when all the chains have the same size. (Saks [1985])

16.3.6. Consider arbitrary sets of linear orders equally likely to be the “true underlying
linear order ”. Balanced comparisons are no longer guaranteed. Fix  > 0. Show that there
is a set of linear orders on some finite set admitting no -balanced comparison, even when
the orders are a subset of the linear extensions of one poset. (Hint: Use orderings obtained
from a single ordering by transposing two consecutive elements.)

16.3.7. Kahn–Saks [1984] proved that an incomparable pair is 11 3

-balanced when its ex-
pected heights differ by less than 1. For sharpness, Saks [1982] constructed the poset
below having a pair (x , y) such that h(y) − h(x) = 1 and (y < x) = 113
. Explain how the
1
proof of Theorem 16.3.6 finds a 3 -balanced pair in this poset.
•d
•c
• •b
e• •a

16.3.8. (♦) Define a poset Q on {xi : i ∈ } by xi < xj if and only if i ≤ j − 2. Let Qn and Q n
be the subposets consisting of {x1 , . . . , x n} and {x− n+1 , . . . , x0 , . . . , x n}, respectively.
(a) Determine the number of linear extensions of Qn . Compute limn→∞ (x1 > x2).
(b) Within Q n , let n = (x0 > x1). Compute limn→∞ n . What does this say about
balanced pairs in Q? (Brightwell [1988])

16.3.9. (+) In a poset P , say that an element  is good for an ordered incomparable pair
(x , y) if  covers x and is incomparable to y or if  is covered by y and is incomparable to x.
(a) Prove that if 1 > (x < y) > 23 and every element  ∈ P − {x , y} satisfies ( < x) ≥
2
3
or (y < ) ≥ 23 , then at least two elements are good for (x , y).
(b) Recall that a semiorder (Definition 12.4.1) is a poset representable by a function
 such that x < y if and only if  (y) −  (x) > 1. Prove that if a semiorder P has no 13 -
balanced pair, then (x < y) > 23 whenever  (x) <  (y).
(c) Prove that every finite semiorder has a 13 -balanced pair. (Brightwell [1989])
832 Chapter 16: Geometry and Topology

16.3.10. Show that every poset is a containment poset of substars of a star. (Golumbic–
Scheinerman [1989])

16.3.11. Interval orders as containment orders. Let be the family of point-deleted real
intervals unbounded below; that is, a set A(a , b) ∈ is determined by two numbers a , b
with a < b by setting A(a , b) = {x ≤ b: x
= a}. Prove that the class of interval orders is
properly contained in the set of -containment orders.

16.3.12. Prove that every interval order and every 4-dimensional poset is an angle order.
(Fishburn–Trotter [1985])

16.3.13. Prove that every bipartite poset is a sphere containment order for spheres in
some dimension. (Scheinerman [1993])

16.3.14. Prove that every 3-dimensional poset is the containment poset of some set of equi-
lateral triangles with the same orientation. (Santoro–Urrutia [1987])

16.3.15. (♦) 2[4] − {∅ , [4]} is not a circle order. A function representation of P assigns each
i ∈ P a continuous function ºi : [0 , 1] → such that i ≤ j in P if and only if ºi(x) ≤ ºj (x)
for all x ∈ [0 , 1]. The crossing number of P is the minimum over all function diagrams of
the maximum number of crossings of two curves.
(a) Consider a function representation of the standard example Sn . Prove that there
are n different points x1 , . . . , x n such that ºi(xi) < ºi(xi) for 1 ≤ i ≤ n.
(b) For n even, let T n be the subposet of 2[n] consisting of the elements of size
1 , n/2 , n − 1. Prove that the crossing number of T n is n − 1.
(c) Prove that the crossing number of a circle order is at most 2. (Hint: Show first that
any finite circle order has a circle representation in which all the circles have a common
interior point. Use such a representation to construct a suitable function representation.)
(Sidney–Sidney–Urrutia [1988])

16.3.16. For ¾ ≥ 3, prove that every 2 ¾-dimensional poset is the containment order of a
family of ¾-gons, where ¾-gon is a polygonal region bounded by a closed curve consisting of
¾ segments. (Sidney–Sidney–Urrutia [1988])

16.3.17. Use Theorem 16.3.20 to prove that there are (2 ¾ + 1)-dimensional posets that are
not containment orders of ¾-gons, where a ¾-gon is the region enclosed by a simple polygon
with ¾ sides. (Hint: Show that if the segment from (x , y) to (x , y ) is not vertical, then it

intersects the segment from (a , b) to (a , b ) if and only if ( ba −−yx − yx −− xy ) ( ba−− yx − yx −− xy ) ≤ 0.)

16.3.18. (♦) Interval number of posets. The interval inclusion number of a poset P , written
i(P), is the least t such that P is a containment poset of a family of sets in which each set
is the union of at most t real intervals. (Madej–West [1991])
(a) Prove that i(Sn) = 2, where Sn is the standard example of dimension n.
(b) Prove that i(P) ≤ ⌈ r/2⌉ when dim P = r.
(c) Use Theorem 16.3.20 to show that part (b) is sharp for large r.
(d) Prove that | i(P) − i(P ∗)| ≤ 1, where P ∗ is the dual of P. (Comment: The difference
can be 1; i(P) is not a comparability invariant.)

16.3.19. Prove that the poset obtained from the n-dimensional subset lattice 2[n] by delet-
ing the top and bottom element has interval inclusion number ⌈ n/2⌉ . (Madej–West [1991])

16.3.20. Use Warren’s Theorem (see Theorem 16.3.20) to show that r polynomials in l
variables have at most (8edr/l)l plus/zero/minus sign patterns (Alon–Scheinerman [1988])
Hints to Selected Exercises

1.1.20. Consider the number of runs of odd entries.

1.1.26. Guess an extremal permutation and show that pushing an arbitrary per-
mutation toward it does not decrease the displacement.
1.1.31. Sum the absolute differences of consecutive numbers, in two ways.
1.1.33. Use induction on ¾ .
1.2.18. For the algebraic proof, use the generalized form of Theorem 1.2.3(5).
1.2.22. For part (a), use induction on n. For part (b), first generalize part (a),
using induction on m.
1.2.23. Do (b) by induction on m, using (a) to reduce the number of indices.
1.2.26. Count the triples (x , A , B) such that A , B ⊆ [n] and x ∈ A ∩ B.
1.2.27. Express each sum as a product.
1.2.39. On the left, group the pairs according to the choice of A. On the right,
group the pairs according to the choice of B ∩ N .
1.2.49. Group the steps in pairs, and encode paths as words of length n in an
alphabet of size 4.
1.3.10. Every tree with at least two vertices has a leaf.
1.3.14. Specify two vertices. Delete an edge leaving one on the path to the other.
1.3.16. The term for ¾ = 0 may need to be treated separately.
1.3.18. Use Corollary 1.3.7 and the Multinomial Formula.
1.3.26. Use induction on d to add the next term in the sum.
1.3.29. Count the pairs consisting of an ordered tree and one of its leaves, or
apply the Narayana numbers.
1.3.32. Change the first peak or valley at even height into the other type.
1.3.38. Consider cyclic arrangements of a + 1 1s and b 0s.
1.3.39. Convert the lists to binary lists, applying Theorem 1.3.17 when n is odd
and Theorem 1.3.20 when n is even.
1.3.46. Use the idea of Theorem 1.3.17.

833
834 Hints to Selected Exercises

⌊n/2⌋
2.1.26. Prove ∑i=1 F̂ n+1−2i < F̂ n for n ∈ . 
2.1.29. For length at least 2, solve the problem first for d = 1.
2.1.39. Interpret am ,n as counting strings in three types of steps, one vertical
and two horizontal.
2.1.42. Consider the argument for the derangement recurrence.
2.1.43. Count the permutations satisfying the increase condition. Eliminate the
non-derangements by considering the smallest and largest fixed points.
2.1.46. Both sides count the permutations of [n + ¾] with no fixed points in [n].
2.1.54. Use induction.
2.1.56. With bn = an/n, prove that bn+1 = bn + Cn(pq)n .
2.1.57. Find recurrences for both sequences.
2.1.63. One can introduce auxiliary sequences, obtain a system of recurrences,
and turn them into a single recurrence for ⟨a⟩, or obtain a single recurrence of
unbounded order and manipulate it to eliminate summations.
2.1.68. Use row and column operations and then induction.
2.2.27. Manipulate the expression for the sum so that the Binomial Formula can
be used to simplify it.
2.2.34. Divide (1 − x) and (1 − y) into the denominator and use the remainder
to express the denominator using (1 − x) and (1 − y) as factors. Expanding the
remaining factor leads to an infinite sum for the generating function. From the
¾ th term in the sum, extract its coefficient of x m y n .
2.2.37. With corners x0 , . . . , x n+1 , let x ¾+1 be the least indexed vertex such that
x¾+1 x n+1 is a chord; note that there may be no such chord.
2.2.39. Develop a system of recurrences for tilings of several types of boards.
2.2.40. Show that ⌊ n⌋ satisfies the recurrence that defines  .
2.3.16. For part (a), consider the possible locations of the first ball.
2.3.18. For part (c), give a recurrence for  in terms of which seat is filled last,
and use this to prove  (n) = (n) by induction.

3.1.25. For part (a), consider first {1 , . . . , r}.

3.1.26. Show that the cycle lengths in the canonical cycle representation, from
back to front, follow the same process as the production of (Y1 , . . . , Yn).
3.1.28. Reduce to the probability that a random element of Ën has every element
in a cycle with one of {1 , . . . , }.
3.1.30. Consider canonical cycle representations.
3.1.33. Use ∑n≥0 c( , n)x n = x(¾) to find the generating function ∑n≥0 F n()x n .
3.1.37. Prove that they satisfy the same recurrence.
3.2.5. For (a), be careful about small n. For (b), consider choosing committees.
3.2.19. Begin with a double sum, but evaluate it using two convolutions.
3.2.23. Letting An(x) denote the desired sum, find a recurrence in n for An(x).
3.2.27. Count the words ending in 0 and the words ending in 1 separately.
Hints to Selected Exercises 835

3.2.34. For part (a), use induction. For part (b), use part (a). For part (c), use
induction and integration without part (a).
3.2.36. First find the generating function, then use partial fractions.
3.2.40. Compare with Example 3.2.17.
3.3.15. For part (a), see Example 3.3.7.
3.3.16. Use the recurrence to obtain a differential equation for the EGF.
3.3.18. Use the correspondence between trees and n-ary lists.
3.3.22. Give a bijective argument using one Stirling number.
3.3.32. For part (a), begin with ∑ j =1 S(n , j)¾(j) = ¾ n from Theorem 3.3.13.
n

3.3.37. For part (a), consider the position of element n in a permutation of [n].
3.3.38. Express the even graphs with vertex set [n] in terms of general graphs
with vertex set [n − 1].
3.3.43. For part (a), modify the Exponential Formula to require an even number
of component structures.
3.3.48. Generalize the proof of Lemma 3.3.33.
3.3.53. Write the left side as a function of y.
3.4.17. Prove that they satisfy the same recurrence.
3.4.21. Interpret both sides using partitions, and then show bijectively that the
coefficients are the same.
3.4.34. Count permutations of [n].
3.4.37. For each multiset, choose a canonical path associated with it such that
the peak heights occur in nonincreasing order.
3.4.43. Interpret the two multisets using paths along the boundary of the Fer-
rers diagrams of the partitions.
3.4.45. Build an infinite sum in which each factorization of n contributes 1/n2 .
3.4.46. Reduce the problem to showing that negative entries can be eliminated.
3.4.49. Express the answer in terms of the greatest integer ¾ such that (¾2) < n.

4.1.15. Group the fractions j/n for j ∈ [n] by their denominators in lowest terms.
4.1.27. When is a number both an rth power and an sth power?
4.1.30. Count the n-sets in [2n] whose intersection with [n] is [t].
4.1.31. Group [2n] into pairs and consider the subsets of a certain size.
4.1.32. Count a certain type of subsets from a family of ¾ sets of size n.
4.1.33. For the inclusion-exclusion part, count a particular family of 0 , 1-lists
with no consecutive 1s.
4.1.44. View the squares as 1s in a permutation matrix, and interpret the color
of a square using parity.
4.1.46. Show that the left side is the inclusion-exclusion formula to count a set
counted more directly by the right side.
836 Hints to Selected Exercises

4.1.49. Expand the factor (n − ¾)n and interchange the order of summation to in-
troduce Stirling numbers, which eliminates most terms. For one nontrivial term
1
that remains, use ¾ −1 = ∫0 x¾−1 .
4.1.52. Consider the board B obtained from B by adding x columns of n squares
to the left of B.
4.1.57. Say that a partition has Property i if 2i is a part or i is a repeated part.
4.1.64. In summing the weights of paths from x i to y j , consider how the paths
arrive at yj .
4.1.65. The identity n = ∑d|n (d) of Exercise 4.1.15 is needed.
4.1.68. Use induction on the number of rows.
4.1.69. For part (a), interpret terms in the product of determinant expansions as
pairs of matchings in a weighted bipartite graph.
4.2.12. Encode caterpillars as binary lists or compositions of integers.
4.2.15. Color a crown with pl positions. Use induction on l.
4.2.18. Use the pattern inventory or Burnside ’s Lemma with inclusion-exclusion.
4.2.27. Among the graphs with vertex set [n], compare those with even size and
odd size left fixed by a given permutation of [n].
4.2.30. In part (b), the answer depends on the parity of d.
4.3.5. How many comparisons are made when an element is processed, in terms
of where the shape grows?
4.3.12. Using planarization is shorter than using the bumping procedure.
4.3.13. In part (b), show that both sides satisfy the same recurrence and initial
conditions; in particular,  () = 0 when  i+1 =  i + 1.
4.3.16. For part (a), consider the location of element n.
4.3.21. For necessity in part (a), use induction on n − i − j . Part (a) is needed to
prove in part (b) that  is surjective.

5.1.14. For the impossibility part, consider the usage of the cross edges (copies of
the factor K 2) in the two cycles.
5.1.25. When n is divisible by 4, generalize the structure of P4 . When n − 1 is
divisible by 4, add one vertex to the graph constructed for n − 1.
5.1.26. Consider the pairs of consecutive vertices and pairs of opposite vertices
along a cycle in  .
5.1.34. How many independent sets of size 4 or 3 contain a given vertex?
5.1.38. Consider a maximal path. Remember that a construction is needed to
prove that the bound cannot be improved.
5.1.43. How many coordinates can change along a cycle of length 2r?
5.1.45. There may be no 6-cycle. Begin with an edge.
5.1.46. For the last part, use induced subgraphs with n − 1 and n − 2 vertices.
5.2.12. Define an appropriate graph and use parity of the degrees.
5.2.13. Define a graph to model the movements, and use the even parity of the
number of vertices with odd degree.
Hints to Selected Exercises 837

5.2.15. When n is even, prove that the number of paths of length 2 starting any
vertex is odd. Counted in another way, use this to prove that some vertex has odd
degree. By studying the triangles at such a vertex, obtain a contradiction.
5.2.16. Use induction on n. In the induction step, be careful to verify the condi-
tions required to apply the induction hypothesis.
5.2.18. Use induction on ¾ .
5.2.20. Follow the method of Theorem 5.2.6.
5.2.27. A short inductive proof uses the statement in the proof of Theorem 5.2.9
about having a bipartite subgraph capturing half the degree at each vertex.
5.2.28. Show that the underlying graph has no triangles.
5.2.29. In part (a), be careful about the base step.
5.2.31. Can G contain K 4 ?
5.2.34. Show that in every orientation D not having the desired property, there
is a path from some vertex x with d+D(x) > ¾ to some vertex y with d+D(y) < ¾ .
5.2.41. What does switching two consecutive vertices in the order do to S?
5.3.28. For the edge vi vi+1 in an odd cycle [v1 , . . . , v¾ ] of G, what happens when
the edge is not oriented from vi to vi+1 ?
5.3.36. Use induction on the number of vertices.
5.3.37. Consider a component that omits some vertex of maximum degree.
5.3.38. Consider a maximal path.
5.3.39. For part (b), use part (a) and the hypothesis on minimum degree to obtain
upper and lower bounds on the number of edges joining V(C) and V(G) − V(C).
For part (c), construct a graph with vertices grouped into sets of size ¾/2.
5.3.41. Consider a realization with the fewest components.
5.3.42. Use induction or show that there is such a realization among the realiza-
tions whose underlying graphs have the fewest edges.
5.3.53. Delete two vertices to apply induction.
5.3.56. For part (a), think about for what kind of graphs it is easy to obtain such
an orientation.
5.3.57. For the first part, use an Eulerian circuit in a related even graph.
5.3.60. For part (a), build a list of trails starting at u. For part (b), consider all
cycles through v.
5.4.18. Use Exercise 5.4.17 and induction.
5.4.19. For part (a), show that among all the ways to map the vertices of G onto
the vertices of H , at least one maps no edge of G onto an edge of H.
5.4.20. Use induction on ¾ . In the induction step, delete one vertex from each
component of the subgraph of G induced by the vertices with degree less than ¾ .
5.4.27. Reduce to connected graphs and add edges to a spanning tree.
5.4.29. For part (a), use induction or explicit construction.
5.4.33. For part (a), build a graph H whose vertices are decompositions of G into
two subgraphs each having | V(G)|− 1 edges, such that the desired decompositions
into trees with the right degrees are the odd-degree vertices. The even-degree
838 Hints to Selected Exercises

vertices are decompositions (Ŝ , T̂) where the degrees agree with S and T except
at two vertices: Ŝ isolates a fixed leaf w of S and has one cycle through the one
other vertex whose degree does not agree with S, and T̂ is a spanning tree.
5.4.39. Let x ∈ n be a vertex such that d G(0 , x) ≥ 3. Consider the maximum
size of an independent set in the graph with vertex set n in which vertices are
adjacent if and only if they differ by x.
5.4.41. Use induction on the number of leaves or the number of vertices.
5.4.56. Use induction on n. One can prove the stronger statement that an n-
vertex tree other than a star occurs in edge-disjoint copies such that each non-leaf
vertex appears at distinct vertices in the two copies. Another proof treats stars
with one edge subdivided as a special case and then proves the claim for other
trees from the induction hypothesis.
5.4.58. Given vertices x and y as in the hint, form G from G − x − y by adding
¾ disjoint edges joining NG(x) to NG(y).

6.1.16. For the second part, generalize the example of C6 when ¾ = 2.

6.1.17. Restrict to a subset of X that is covered by some maximum matching.
6.1.19–20. Look for a perfect matching in an appropriate bipartite graph.
6.1.24. Given n positions and m winning sets, form an X , Y -bigraph with X =
{x1 , . . . , x n} and Y = {w1 , . . . , wm} ∪ {w1 , . . . , w m} having edges x i wj and x i w j
for each instance of position x i in winning set Wj .
6.1.26. Consider Theorem 6.1.10.
6.1.29. Reduce to the case (G) < n. Note that it is not possible to prove that
(G) ≥ 2 (G) always holds, since that is false when 2 (G) > n.

6.1.31. For part (a), reduce the problem to the case where r = d − 1 and the
vertices of degree d form an independent set, and then apply Hall’s Theorem.
6.1.38. Translate the statements “v belongs to some smallest vertex cover ” and
“v is covered by every maximum matching ” into notation.
6.1.42. When G has incident edges e1 and e2 , let S1 and S2 be minimal ver-
tex covers of G − e1 and G − e2 . Add to S1 ∩ S2 an appropriate vertex cover of
G [(S1  S2) ∪ {x}] to contradict (G ) = (G).
6.2.8. Use inclusion-exclusion.
6.2.18. For sufficiency, consider a subgraph with the most vertices among those
whose components are nontrivial stars with at most m edges.
6.2.22. If (G) < ⌊n/2⌋, then every maximum matching leaves at least two ver-
tices uncovered.
6.2.31. For part (b), consider the edges joining S and S, where S is the vertex
set of a maximum matching. For part (c), use induction on s and the idea of 2-
switches (Definition 5.2.7); remember that multiedges are allowed.
6.2.41. For part (a), consider a maximal Tutte set S in G − {x , y}, and apply
Lemma 6.2.7 to S ∪ {x , y} in G. For part (b), use induction on n.
6.2.42. Design an auxiliary graph H in which the  -factors of G are the vertices
of odd degree. The other vertices are spanning subgraphs with the desired num-
ber of edges, but they have degree  (w) − 1 at a fixed vertex w and degree  (u) + 1
at some other vertex u, otherwise agreeing with  .
Hints to Selected Exercises 839

6.3.7. How do the various edges contribute to the sum of the degrees of the ver-
tices covered by M? Use that relationship to obtain an upper bound on m.
6.3.24. Suppose that the first such occurrence is a rejecting x even though xa is
a pair in some stable matching M , and a rejects x for y. Let b be the mate of y in
M. What can be deduced about the preferences of these people?

7.1.4. Induction permits a simple proof.

7.1.19. For part (a), argue that some if | S| ≤ ¾ , then some component of G − S
contains a vertex of degree at least d n−¾ ; consider vertices in other components of
G − S. For part (b), build a graph G with a separating ¾-set S such that G − S
has a component with n − 1 − d n−¾ vertices.
7.1.20. Consider a minimal subgraph H such that H has at least 2 ¾ − 1 vertices
and more than (2 ¾ − 3)(| V(H)| − ¾ + 1) edges. Show that H is ¾-connected.
7.1.27. In part (a), G H has four cyclically overlapping copies of K r+1 K r+1 .
7.1.32. Use induction on n.
7.2.20. Use Menger ’s Theorem.
7.2.26. For vertices u and v in a ¾-edge-connected graph G, consider a family of
¾ edge-disjoint u , v-paths with minimum total length.
7.2.27. Make sure that the construction has connectivity ¾ and diameter d.
7.2.30. Define an appropriate digraph from the input graph.
7.2.32. Consider a longest cycle.
7.2.34. Among all the edges and all the ¾-cuts containing them, choose edge xy
and ¾-cut S so that the resulting subgraph G − S has a smallest possible compo-
nent, H. Prove that v is -contractible, where ∈ S and v ∈ N( ) ∩ V(H).
7.2.41. One proof chooses P so that G − V(P) has a component of largest order.
7.2.42. Consider a cycle in G − u whose deletion leaves u in a largest component.
7.2.48. Consider the blocks of G − v.
7.2.50. Prove that the cycle C contains at least two such ears.
7.2.53. After Alice chooses u and v, Bob produces a special ear decomposition and
a special vertex numbering (in fact, an s , t-numbering with s = u and t = v) and
uses them to orient at most 2n − 3 edges in a way that will win.
7.3.16. Find an appropriate set of vertices in the corresponding graph to violate
the necessary condition.
7.3.21. Prove both statements in a single induction.
7.3.23. Generalize the construction of the bipartite graph in Example 7.3.3.
7.3.28. For part (a), consider the second graph in Example 7.3.3. For part (b),
prove that when G is a tree a stronger statement holds: for any specified edge xy
of G, there is a spanning x , y-path in G 3 − xy.
7.3.32. For part (b), start with n ∈ {5 , 6} and generalize.
7.3.34. Given that Ore ’s Condition fails for G, determine the structure of G.
7.3.36. For part (a), delete a vertex with large degree. For part (b), show that if
G fails Chvátal’s Condition, then G has at least n − 2 edges.
840 Hints to Selected Exercises

7.3.43. Consider the location of the vertices of a largest independent set A rela-
tive to a smallest separating set S.
7.3.44. Proving the two claims together permits a simpler proof.
7.3.45. In part (b), transform the graph to apply a known result.
7.3.47. Apply Ore ’s Theorem.
7.3.48. Apply Chvátal’s Theorem to a modified graph.
7.3.49. Use induction on ¾ , making an argument like that of Ore ’s Theorem.
7.3.54. Consider the degree sum for the vertices in a largest independent set,
and use the Chvátal–Erdős Theorem.
7.3.56. Use induction on ¾ , after choosing an appropriate cycle in G.
7.3.61. Reduce to trees and use induction on | S| .

8.1.13. For the upper bound, present an explicit coloring by regions, with atten-
tion to the boundaries.
8.1.17. For the upper bound, cover the vertices with n − 2 ¾ + 2 independent sets.
8.1.19. Use induction on | V(G)|.
8.1.27. Use large neighborhoods and induction on r.
8.1.29. Consider a smallest-last ordering.
8.1.30. Prove that an optimal coloring of G has many color classes of size 1.
8.1.36. Obtain an independent n-set in G K r from a proper r-coloring of G and
vice versa.
8.1.40. For part (a), use induction on n or degeneracy.
8.2.17. For Ò(G) = 4, consider a shortest odd cycle.
8.2.19. When G − xy is ¾-colorable and G is not, obtain a cycle through xy with
length congruent to 1 modulo ¾ for each cyclic permutation of the ¾ colors.
8.2.27. For part (a), consider a partition V1 , . . . , Vt minimizing ∑i | E(G i)|
Di .
8.2.28. Modify G by adding vertices to make a graph G where an argument
similar to that for part (a) of the previous problem yields the desired result on G.
8.2.30. Construct a special list assignment with list sizes equal to degree.
8.2.33. For part (a), choose a bad assignment L with smallest union, and consider
a maximal X such that | L(X)| < | X | . For part (b), the union of the bad lists should
have size 2 ¾ ; force many colors to appear by using the same two disjoint lists on
each part of size 2. The construction should restrict to the bad 2-uniform list
assignment for K 4 ,2 .
8.2.34. Reduce to the case where the lists on each part are disjoint, and apply
Hall’s Theorem to choose a proper coloring from the lists.
8.2.43. Delete one leaf from each nontrivial component of F to obtain F . Let R
be the set of neighbors of the deleted vertices. Map R onto an m-set X ⊆ V(G)
chosen to minimize | E(G[X])| . Extend X to a copy of F . Use Hall’s Theorem to
match X into the remaining vertices.
8.2.46. When an optimal coloring º has a color class of size 1, use ½ to make an
alteration in º .
Hints to Selected Exercises 841

8.3.6. Use a graph transformation.

8.3.18. If  (G) > (G), then G has a minimal subgraph G such that  (G ) >
(G). Obtain a proper (G)-edge-coloring of G from such a coloring of G − xy,
by switching colors along a 2-edge-colored path to make a color available for xy.
8.3.22. In part (b), use part (a) to restrict attention to overfull subgraphs with
at most half the vertices.
8.3.26. In part (a), first color the subgraph obtained by deleting all copies of
edges with multiplicity more than (G)/2. In part (b), for an uncolored edge xy,
consider an edge y having a color that is missing at x. Show that the coloring
extends if has a common missing color with y or x.
8.3.27. Use a graph transformation.
8.3.28. For part (a), use Eulerian circuits. For part (c), consider an optimal
( (G) + 1)-edge-coloring of a graph G, and build a color fan at a vertex whose
incident colors are not distinct.
8.3.30. Apply Puleo ’s result with  = maxv∈V(G)(d G(v) + (v)) to obtain S = V(G).
8.3.32. Given a proper edge-coloring with q colors that satisfies the intersection
condition, compute upper and lower bounds on the number of ordered pairs of
edges having the same color.
8.3.34. Start with a proper vertex coloring and a proper edge-coloring that use
the fewest colors, sharing two colors. Uncolor the edges that have the common
colors and show that they can be recolored to complete a total coloring.
8.3.35. For part (b), view K r,s as a composition of K m ,n and K ¾ ,¾ , where m and n
are relatively prime.
8.3.36. For sufficiency, use a parity edge-coloring and a spanning tree of G to
map the vertices of G into vertices of Q¾ .
8.3.47. Use induction on the number of vertices.
8.3.58. Prove that the complement of a connected P4 -free graph is disconnected.
(This problem does not allow using that P4 -free graphs are perfectly orderable.)
8.3.63. Reduce to proving (G ) = (G ) and (G ) = (G ) when G and G are
perfect, respectively, where G is obtained from G by duplicating one vertex x.
The case of (G ) = (G ) where x is in no maximum stable set requires looking
at a smallest clique cover of G.
8.3.69. For part (b), there is an example with six vertices.

9.1.15. Show that G ∗ can be drawn so that each edge of G ∗ intersects only the
edge corresponding to it in G.
9.1.23. Show that a separating 2-set must induce two edges.
9.1.26. The solution is easier to write when the characterization used for j-vertex
trees is “connected and j − 1 edges” rather than “connected and acyclic”.
9.1.30. Use induction or duality.
9.1.33. Prove that the complement of any 7-vertex maximal outerplanar graph
is not outerplanar.
9.1.34. Show that the edge set decomposes into triangles.
842 Hints to Selected Exercises

9.1.40. Add a point at infinity or a circle that encloses all intersection points.
9.1.41. Find a simple formula in terms of n for the number of crossings of chords.
9.1.48. Use Euler ’s Formula or induction on the number of facial triangles.
9.2.12,14b,16. Use Kuratowski’s Theorem.
√
9.2.22. Use the corollary that guarantees a (c n , 1/2)-separation (where c =
√ √
(2 2)/(1 − 2/3)).
9.3.16. Find a way to use Grinberg ’s Theorem.
9.3.17. Consider distance from a fixed vertex.
9.3.18. Consider distance from a fixed vertex.
9.3.23. For the first part, use an idea from the proof of Tait ’s Theorem (Theorem
9.3.1). The proof using the Four Color Theorem is easy.
9.3.25. For part (c), build ⌊n/3⌋ areas that must be watched by different guards.
9.3.27. For the induction step in part (c), delete the edges of a face neighboring
the unbounded face.
9.3.31. Prove that a counterexample with fewest vertices has no 4-face, by show-
ing that some two vertices of a 4-face can be merged.
9.3.34. It may help first to construct such a graph with 114 vertices and then
one with 86 vertices.
9.3.38. Given a separating cycle C with length at most 4, prove that the C-lobes
of G have proper 4-colorings that agree on C. In particular, when C has length 4,
show that a C-lobe has a proper 4-coloring where two specified opposite vertices
of C receive distinct colors.
9.3.39. Begin by showing that w4 and w6 can be assumed not to be adjacent out-
side the ring.
9.3.40. Force two vertices on the unbounded face to have the same coloring in
any proper 3-coloring, and then make them adjacent.
9.3.42. Consider complete bipartite graphs.
9.3.43. Use degree charging, and let each vertex needing charge take what it
needs equally from its neighbors.
9.3.44. In part (a), design a discharging argument without knowing b, and de-
termine how small b must be to make it work. Part (b) can be done with face
charging or with balanced charging. This exercise completes the bottom line in
Remark 9.3.23 in the same way that the first and third lines were completed.
9.3.51. Show that edges of weight at most 2t + 1 are reducible. When there is
no such edge, Lemma 9.3.30 guarantees a 4-cycle through two 3-vertices. Use a
decomposition into t linear forests for the graph G obtained by deleting those
two 3-vertices to obtain such a decomposition for G.
9.3.57. Use vertex charging, with 5-vertices taking 12 from incident 4+-faces and
the remaining needed charge from neighbors along edges shared by two triangles.
9.3.58. Use vertex charging; 6-vertices that give charge to 5-neighbors will need
charge from 7+-neighbors.
9.3.64. Use degree charging. Let vertices of degree at most 4 take what they
need equally from their neighbors, but let a vertex of degree 5 or 6 take 14 from
Hints to Selected Exercises 843

each 6+-neighbor. In checking that vertex v ends happy, consider cases depending
on d(v), and let j be the least degree among the neighbors of v.
9.3.68. For part (c), use the existence of planar graphs that are not 4-choosable.

10.1.10. Find a set S of size (m+ 1)n in [2 m n] partitioned into n divisibility chains.
10.1.14. Consider the players with positive scores.
10.1.18. Compare the number of nonempty subsets of ¾ elements with the largest
sum of ¾ elements.
10.1.19. Use induction, deleting a largest element of S.
10.1.23. Reduce the problem to the case m = 2n.
10.1.25. Use a scheme in which 90 members each get only one key.
10.1.26. Break the sequence into segments whose length is one more than the
size of the alphabet.
10.1.31. For odd n, use a decomposition into (n − 1)/2 isomorphic subgraphs con-
sisting of a 4-cycle plus pendant edges. Use different constructions for n = 4¾ − 1
and n = 4¾ + 1. In the latter case, use one central vertex and distribute the others
evenly over two concentric circles, as (n − 1)/2 pairs.
10.1.36. Generalize the argument of Exercise 10.1.35 by fixing positions that do
not end increasing lists of length ¾ − 1.
10.2.8. Define a weighted transitive tournament with the points as vertices,
where edges point to the right and weights correspond to slopes.
10.2.10. 2-color triples of points so that m points whose triples all have the same
color will form a convex m-gon. Possible color criteria include convex/concave
shape, indices clockwise or not, and parity of the number of points inside.
10.2.11. Use induction. Consider the neighbors or nonneighbors of one vertex.
10.2.21. Be careful about parity.
10.2.28. Obtain a bow-tie with monochromatic triangles, plus complementary
monochromatic 5-cycles, and then use symmetry.
10.2.30. The answer depends on the relationship between n and (G).
10.2.32. For the first part, consider the average of (x) + d(y) over edges xy in a
subgraph G of K s ,t .
10.2.34. Prove a more general statement by induction, or prove that if ∑v∈ X (d(v)s )>
(t − 1)(ns) for an X , Y-bigraph G contained in K n ,n , then G contains K s ,t .
10.2.37. For the upper bound, use induction on m. After finding a red copy of
Pm−1 in a given coloring, consider blue paths that alternate between vertices out-
side this path and vertices inside it.
10.2.39. For part (a), determine the colors of the edges at displacement 2, 3, and
4 on the monochromatic (2  + 1)-cycle. For part (b), consider the edges with even
displacement on the monochromatic 2 -cycle.
10.3.4. Take n ≥ R¾ (3; 2), as defined in Definition 10.2.10, and consider the proof
of Theorem 10.3.1. One need not use all the nonempty subsets of [n].
10.3.9. Use van der Waerden’s Theorem on the indices corresponding to a long
arithmetic progression in ⟨a⟩.
844 Hints to Selected Exercises

10.3.11. Consider a monochromatic ¾-set or a rainbow ¾-set.

10.3.13. After finding red points a1 , . . . , an whose centroid is a lattice point, if
is blue consider replacing aj by a point bj so that the new set has centroid aj .
10.3.14. Given a coloring of E(K n), consider a monochromatic connected sub-
graph with the most vertices.

/ E(G) and xu , xv
= 0, then  (x ) ≥  (x) for
11.1.8. For the first part of (a), if uv ∈

some x such that x u xv = 0.
11.1.12. Solve it first for (G), and then use that to give the answer for (G).
11.1.15. Reduce part (b) to Mantel’s Theorem.
11.1.33. Use induction on n. Consider the larger family of graphs in which adding
any edge increases the number of t-cliques, regardless of whether the graph al-
ready contains a t-clique.
11.1.37. For part (b), first find a vertex in A having at least  n neighbors in each
of B and C. For part (c), prove that if G has (1/8 + )n2 edges, then G contains
K 4 or an independent set of size at least (2/cM)n for some appropriate constant
c, where M is the bound on the number of classes when the Regularity Lemma
is applied with the arguments  and l, where  = /6 and l = 1/ . From the
resulting partition with  classes, where 1/ ≤  ≤ M , delete edges except for
those in  -regular pairs with density more than 2  . Now applying parts (a) and
(b) to the remaining graph shows that with no K 4 and no large independent set,
| E(G)| < (1/8 + )n2 .
11.1.39. For part (a), first show that  (G) > n − (G). Then show that a non-
perfect maximum matching admits an augmenting path of length 5.
11.1.40. In part (a), the path needs to be grown in two phases. First obtain a path
with length at least (d − 3)2m − 2K , and then show that a path at least this long
can be modified to obtain a longer path until length (1 −  − d− )2m is reached.
11.1.41. Let P be a longest path in G having an even number of vertices. Let
2t = | V(P)| . If t ≤ (1 − d− )n, then find a detour to obtain a longer such path. If
t > (1 − d− )n, then find a cycle of length 2t through V(P) and then a longer path.
11.1.42. For part (a), use induction on  . In the induction step, consider sep-
arately the -tuples where the first  − 1 elements have at most (d − )¾−1 | Y |
common neighbors in Y and those having more such common neighbors.
11.1.43. For part (c), reduce to when every vertex is in at least two edges and no
two edges share two vertices. The link graph of v has as its edges the pairs that
with v form edges of H. Apply part (b) to the union of all the link graphs.
11.2.11. The -sets in F must avoid the shadows of higher elements.
11.2.15. Use shift operators to compress the family.
11.2.35. Count the elements used in a sunflower of size s.
11.2.39. Express a random subset of [n] in terms of the sizes of its intersections
with the members of a distinguishing family.
11.2.41. Let F consist of congruence classes modulo  .
11.3.33. For part (c), use induction on | X  Y |.
Hints to Selected Exercises 845

11.3.35. For part (b), use property (J) and induction on | I1 − I2 | to obtain the
augmentation.
11.3.40. For part (a), use transitivity of dependence or properties of cocircuits.
For part (c), using induction on | X 1 |.
11.3.42. For equality, define a special chain using the weights of the elements.
11.3.53. Devise a strategy using deletions and contractions.
11.3.56. Use the set-counting formula that appears in the proof of the Matroid
Union Theorem.

12.1.9. Consider orientations of the ten 4-cycles.

12.1.13. Finding a poset having exactly one largest chain and exactly one largest
antichain will show that nine of the statements are false if that chain and an-
tichain are disjoint. One more of the remaining seven statements is also false.
12.1.16. For part (b), when F has no union-free family of size m, how many ele-
ments can the inclusion poset on F have at height h?
12.2.8. Consider (0 , i , i) < · · · < (0 , i , n − i) < · · · (0 , 2i , n) < · · · < (n − 2i , n , n).
12.2.31. Rewrite (b) so it is twice the needed inequality in (c).
12.2.32. Compare ½ ¾+1 (r, s + 1)/½ ¾ (r, s) with ½ ¾+1 (r, s)/½ ¾ (r, s − 1) and with ½ ¾+1 (r +
1 , s + 1)/½ ¾ (r + 1 , s) by clearing fractions and using log-concavity.
12.2.33. After using the binomial recurrence to express ºn+1 in terms of ºn , ex-
press the coefficients of [ºn(x)]2 − ºn−1 (x)ºn+1 (x) as two sums, and show that the
coefficients are nonnegative.
12.3.8. Forbid a 2-realizer or consider the edges of size 2 in the hypergraph of
alternating cycles.
12.3.23. Embed P by letting ºi(x) = j if the least element greater than or equal
to x on chain i is the jth from the bottom, ºi(x) = ¾ if no such j exists.
12.3.24. For the upper bound, form a directed graph with an edge for each ex-
tension in a minimal realizer.
12.3.32. For part (b), use an incidence matrix to establish a bijection between
n-tuples of subsets of [r] that are (n − m)-scrambling and r-tuples of vertices (as
binary vectors) that break all m-cubes in the n-dimensional hypercube.
12.4.31. Use the definition of distributivity, the isomorphism to J(P(L)), the
embedding of L in a subset lattice, or the forbidden sublattice characterization.
12.4.44. Express X − X as X ∧ X , where X is the family of complements of
members of X in the union of all the sets, and compute | X ∧ X | and | X ∨ X |.

13.1.10. After normalizing all the top rows to be the same, only one choice re-
mains for each other position of the square being added. Show that it works.
13.1.14. Given Latin squares L1 and L2 , let á(s) = (s3 − L1 (s1 , s2) , s4 − L2(s1 , s2)).
13.2.4. Prove that in a complete orthogonal family, any pair of positions not in
the same row or column have the same value in exactly one of the squares.
13.2.10. View K 2 ,t as two vertices with t common neighbors.
846 Hints to Selected Exercises

13.2.13. For part (b), partition [n] into sets X1 , . . . , X ¾ of size m, and form pair-
wise edge-disjoint graphs H1 , . . . , H¾ such that each Hi uses q + 1 of the sets
X 1 , . . . , X ¾ as its vertex set.
13.2.18. After obtaining the orbits under a purported multiplier, consider the
pairs generating the difference 28.
13.3.6. The operation table of an idempotent quasigroup is a Latin square with
element i in position i on the diagonal. For even n, it cannot be commutative, but
it can be built from an idempotent commutative quasigroup of order n − 1.
13.3.23. Sum the absolute edge differences (mod 2) in two different ways.
13.3.25. Partition the vertices into three copies of 3 , so the graph has twelve
types of edges. Define two 6-cycles that together use one of each type of edge.

14.1.18. Generalize a construction of non- ¾-choosable bipartite graphs.

14.1.19. What are the possible seats for the last passenger?
14.1.24. This can be done using induction or linearity.
14.1.28. Generalize the Caro–Wei Theorem.
14.1.30. Part (a) can be solved by computing a sum or by a shorter argument. For
part (b), apply part (a).
14.1.32. Somewhere use 1 − x ≤ e− x to simplify a bound.
14.1.34. For part (a), consider the more general situation where the walk starts
at a given spot among the n + 1 possible positions. Sum the resulting equations.
14.1.35. In order to reach u last, the walk must first visit a neighbor of u and
then visit the other neighbor of u before visiting u.
14.1.36. Let t i be the 0 , 1-variable that is 1 if and only if the ith head is immedi-
ately preceded by a tail.
14.1.39. Establish a one-to-one correspondence between the selectable sets and
the subsets of Q forming chains in the product order on 2 .
14.1.42. Choose B at random and use linearity of expectation.
14.1.43. For a sufficiently large prime p congruent to 2 modulo 3, multiply A by

a random element of p to capture many elements in a sum-free subset consisting

of more than 1/3 of p .
14.1.45. Define a suitable event for each point.
14.2.12. Use the Dependent Random Choice Lemma.
14.2.13. Use the Dependent Random Choice Lemma. In coloring K N with N =
23¾ , the more plentiful color has at least 2−7/3 N 2 edges. To contain Q¾ , we want
2¾−1 vertices whose ¾-sets have 2¾ common neighbors in this color.
14.2.14. Show that the experiment produces a graph with no large independent
set and not many triangles.
14.2.19. Prove first that if there is a solution, then there is a solution where the
two weights are equal.
14.2.20. Use the Local Lemma.
14.2.21. Choosing the endpoints of an edge joining members of S is a bad event.
Hints to Selected Exercises 847

14.3.1. Follow Theorem 14.3.18.

14.3.6. What forces an n-vertex graph to have a cycle? Recall that the variance of
a binomial random variable with m trials and success probability p is mp(1 − p).
14.3.7. Consider G p with p chosen later as a suitable function of n. How many
edges remain when the copies of K r,r are broken?
14.3.8. Bound the probability that the pair on the ith round fails to form an edge.
14.3.11. Consider the probability in the tail on the other side of 12 (n2) from the
expected number of edges.
14.3.12. Design p with a constant parameter c so that the expression for the ex-
pected number of edges on no triangle simplifies.
14.3.13. For the Second Moment computation, group the pairs of potential copies
of H according to whether or not they share an edge.
14.3.39. Otherwise, the number of possible representations is a vanishing frac-
tion of the total number of graphs.
14.4.6. Show that whp any two vertices in G 1/2 have many common neighbors.
14.4.7. Obtain an exponentially small bound on the probability of having this
many edges across a given bipartition.
14.4.12. For√a random coloring º , let X be the number of edges e such that
## #
###∑v∈e º (v)#### > 2n ln(2m).
14.4.18. Generate a random permutation by choosing a random sequence in the
interval [0 , 1].

15.1.3. For part (a), one approach applies Theorem 15.1.2 after adding a new res-
ident. For part (b), let M be the incidence matrix of the family. Show that MM T
is nonsingular over 2 if m = n (when n is even), and hence the incidence vectors
are linearly independent over 2 . However, show that they can’t be independent.
15.1.6. Use a family of disjoint sets.
15.1.8. Compare det A with det J − I, and then model the partitioning condition
as a matrix equation.
15.1.11. For part (c), consider an equation of dependence for the set consisting
of the polynomials º1 , . . . , ºm from Theorem 15.1.5 plus {1 , x1 , . . . , x n}. Take
partial derivatives 2/ x i x j to obtain equations that the coefficients must sat-
isfy. Describe these equations using the matrix B formed as in part (a) from the
row vectors v1 , . . . , vm}. Use parts (a) and (b) to show that the coefficients on the
added polynomials must be 0.
15.1.15. Make use of a cubical grid.
15.1.26. Use the Combinatorial Nullstellensatz.
15.1.31. The claim can be proved using case analysis or the Alon–Tarsi Theorem.
15.2.9. Prove that the process ends in the initial room, that all incident corridors
of a reached room are followed both ways, and that every room is reached.
15.2.18. Use induction on  + n.
15.2.19. Use induction on the length of C.
15.2.21. Express the desired situation as a matrix equation over 2 .
848 Hints to Selected Exercises

15.2.31. For part (b), Use the Birkhoff–von Neumann Theorem (Theorem 6.1.8)
to argue that optimal matrices are interior points in the space of n-by-n matrices,
and then apply Lagrange multipliers.
15.2.36. Although the statement involves the inclusion relation on a family of
sets, incidence functions are not needed for the solution.
15.3.9. Use Exercise 15.3.7(b) to interpret the coefficients.
15.3.11. Involve the vertex cover number, the fact that the spectrum of a bipar-
tite graph is symmetric around 0, and Exercise 15.3.9.
15.3.13. Construct eigenvectors for the product from eigenvectors for the factors.
15.3.18. Use the complement.
15.3.29. Recall that a matrix M is positive semidefinite, meaning that all eigen-
values are nonnegative, if and only if x T Mx ≥ 0 for all real vectors x. Recall also
that a real symmetric matrix has an orthonormal basis of eigenvectors.
15.3.30. Use the properties in Theorem 15.3.30, proved in Exercise 8.3.58. For
part (b), prove that the largest Laplacian eigenvalue of the complement of a con-
nected cograph G equals the number of vertices in the largest component of G.
15.3.31. Generate subgraphs of K n ,n as the union of ¾ random perfect matchings.

16.1.11. A single toroidal family with arbitrarily large crossing number suffices.
16.1.12. Consider the graph obtained from a maximal such graph by deleting one
edge from each crossing. Argue that edges remaining from the crossings do not
lie on the same face in the resulting plane graph.
16.1.13. Consider proper 3-edge-colorings of the 3-regular plane graph in which
the crossing is replaced with a 4-cycle through four new vertices.
16.1.14. Consider what happens when a vertex moves across an edge. For the
second part, consider a drawing where the crossings are easy to count.
16.1.16. Consider the copies of K m ,n in a drawing of K m+1 ,n .
16.1.17. Consider what happens when a vertex moves across an edge. For the
second part, consider a drawing where the crossings are easy to count.
16.2.6. Choose suitable paths with maximum total length.
16.2.9. For part (b), add a suitable troublesome edge to an appropriate graph.
16.2.10. Use the fact that every infinite sequence of points in a closed and
bounded set has a convergent subsequence.
16.2.13. Use the Hobby–Rice Theorem.
16.3.8. For e(Qn), group the extensions by which element is at the top. For part
(b), compare the limit with 1/3 and with other incomparable pairs.
16.3.9. For part (a), begin by mapping extensions having x before y but no el-
ement between them that is good for (x , y) injectively into the set of extensions
with y before x. Conclude that many extensions with x before y have some ele-
ment between them that is good for (x , y).
16.3.11. For the second part, consider 2 + 2.
16.3.14. Use a realizer to specify triangles all containing the origin.
16.3.19. Prove the lower bound by induction on n.
References
The final item in each entry is the number of the citing page in the text.

Abbott H.L., Lower bounds for some Ramsey numbers. Discr. Math. 2 (1972), 289–293. [459]
Abbott H.L., D. Hanson, and N. Sauer, Intersection theorems for systems of sets. J. Combin. Th. A 12
(1972), 381–389. [504]
Abbott H.L. and B. Zhou, On small faces in 4-critical planar graphs. Ars Combin. 32 (1991), 203–207.
[414]
Abeledo H. and G. Isaak, A characterization of graphs that ensure the existence of a stable matching.
Math. Soc. Sci. 22 (1991), 93–96. [288]
Aberth O., On the sum of graphs. Rev. Fr. Rech. Opér. 33 (1964), 353–358. [343]
Abrham J. and A. Kotzig, Construction of planar Eulerian multigraphs. In Proc. 10th Southeastern Intl.
Conf. Combin. Graph Th. Comput. (Boca Raton), Congr. Numer. 23/24 (1979), 123–130. [388]
Abrishami T., A combinatorial analysis of the eigenvalues of the Laplacian matrices of cographs (Johns
Hopkins Univ., 2019). Masters Thesis. [782]
Acharya B.D. and M. Las Vergnas, Hypergraphs with cyclomatic number zero, triangulated graphs, and
an inequality. J. Combin. Th. B 33 (1982), 52–56. [374]
Achlioptas D. and F. Iliopoulos, Random walks that find perfect objects and the Lovász local lemma. J.
ACM 63 (2016), Art. 22, 29. [680]
Addario-Berry L., K. Dalal, C.J.H. McDiarmid, B.A. Reed, and A.G. Thomason, Vertex-colouring edge-
weightings. Combinatorica 27 (2007), 1–12. [737]
Adrian C., Problem E3459. Amer. Math. Monthly 98 (1991), 754. Solution 100 (1993), 593–594. [511]
Aharoni R., Ryser’s conjecture for tripartite 3-graphs. Combinatorica 21 (2001), 1–4. [803]
Aharoni R. and P. Haxell, Hall’s theorem for hypergraphs. J. Graph Theory 35 (2000), 83–88. [803]
Ahlswede R. and D.E. Daykin, An inequality for the weights of two families of sets, their unions and
intersections. Z. Wahrsch. Verw. Gebiete 43 (1978), 183–185. [597]
Ahlswede R. and L.H. Khachatrian, The complete intersection theorem for systems of finite sets. Euro-
pean J. Combin. 18 (1997), 125–136. [500, 511]
Ahlswede R. and L.H. Khachatrian, Katona’s intersection theorem: four proofs. Combinatorica 25 (2005),
105–110. [497]
Ahuja R.K., T.L. Magnanti, and J.B. Orlin, Network Flows (Prentice Hall, 1993). [246]
Aigner M., Lexicographic matching in Boolean algebras. J. Combin. Th. B 14 (1973), 187–194. [556]
Aigner M., Combinatorial Theory (Springer-Verlag, 1979). [123, 520, 529, 538, 607, 761]
Aigner M., Graphentheorie, Eine Entwicklung aus dem 4-Farben problem (B.G. Teubner Verlagsge-
sellschaft, 1984). (English transl. BCS Assoc., 1987). [399]
Aigner M., Turán’s graph theorem. Amer. Math. Monthly 102 (1995), 808–816. [476]
Aigner M., Lattice paths and determinants. In Computational discrete mathematics, Lect. Notes Comp.
Sci. 2122 (Springer, 2001), 1–12. [165, 168, 170, 176, 177]
Aigner M. and G.M. Ziegler, Proofs from The Book (Springer-Verlag, 1999). Also 2001, 2004, 2009, 2014.
[37, 476, 811]
849
850 References

Ajtai M., V. Chvátal, M.M. Newborn, and E. Szemerédi, Crossing-free subgraphs. In Theory and Practice
of Combinatorics, Ann. Discr. Math. 12 (North-Holland, 1982), 9–12. [793]
Ajtai M., J. Komlós, and E. Szemerédi, A note on Ramsey numbers. J. Combin. Th. A 29 (1980), 354–
360. [226, 451, 663, 664]
Ajtai M., J. Komlós, and E. Szemerédi, Sorting in c log n parallel steps. Combinatorica 3 (1983), 1–19.
[779]
Ajtai M. and E. Szemerédi, Sets of lattice points that form no squares. Stud. Sci. Math. Hungar. 9 (1974),
9–11 (1975). [492]
Akin E. and M. Davis, Bulgarian solitaire. Amer. Math. Monthly 92 (1985), 237–250. [143, 144, 152]
Akiyama J. and N. Alon, Disjoint simplices and geometric hypergraphs. In Combinatorial Mathematics:
Proceedings of the Third International Conference (New York, 1985), Ann. New York Acad. Sci.
555 (New York Acad. Sci., 1989), 1–3. [816]
Akiyama J., H. Era, S.V. Gervacio, and M. Watanabe, Path chromatic numbers of graphs. J. Graph
Theory 13 (1989), 569–575. [418]
Akiyama J., G. Exoo, and F. Harary, Covering and packing in graphs. III. Cyclic and acyclic invariants.
Math. Slovaca 30 (1980), 405–417. [412]
Akiyama J., G. Exoo, and F. Harary, Covering and packing in graphs. IV. Linear arboricity. Networks 11
(1981), 69–72. [677]
Akiyama J. and M. Kano, Factors and factorizations of graphs—a survey. J. Graph Th. 9 (1985), 1–42.
[268]
Akiyama J. and M. Watanabe, Maximum induced forests of planar graphs. Graphs Combin. 3 (1987),
201–202. [419]
Aksionov V.A. and L.S. Mel nikov, Essay on the theme: the three-color problem. In Combinatorics (Proc.
5th Hung. Colloq., Keszthely, 1976), I, Colloq. Math. Soc. János Bolyai 18 (North-Holland, 1978),
23–34. [414]
Albert M.H., M. Elder, A. Rechnitzer, P. Westcott, and M. Zabrocki, On the Stanley–Wilf limit of 4231-
avoiding permutations and a conjecture of Arratia. Adv. Appl. Math. 36 (2006), 96–105. [433]
Albertson M.O., A lower bound for the independence number of a planar graph. J. Combin. Th. B 20
(1976), 84–93. [423]
Albertson M.O., You can’t paint yourself into a corner. J. Combin. Th. B 73 (1998), 189–194. [356]
Albertson M.O. and D.M. Berman, A conjecture on planar graphs. In Graph Theory and Related Topics
(J.A. Bondy and U.S.R. Murty, eds.) (Academic Press, 1979), 357. [419]
Albertson M.O. and B. Mohar, Coloring vertices and faces of locally planar graphs. Graphs Combin. 22
(2006), 289–295. [797]
Alexanderson G.L. and J.E. Wetzel, Dissections of a plane oval. Amer. Math. Monthly 84 (1977), 442–
449. [388]
Alkan E., Problem 10473. Amer. Math. Monthly 102 (1995), 745–746. Solution 104 (1997), 371. [80]
Alladi K. and V.E. Hoggatt, Jr., Compositions with ones and twos. Fibonacci Quart. 13 (1975), 233–239.
[59]
Alon N., On the number of subgraphs of prescribed type of graphs with a given number of edges. Israel
J. Math. 38 (1981), 116–130. [510]
Alon N., Eigenvalues and expanders. Combinatorica 6 (1986), 83–96. [780]
Alon N., Splitting necklaces. Advances in Math. 63 (1987), 246–253. [810]
Alon N., The linear arboricity of graphs. Israel J. Math. 62 (1988), 311–325. [412, 677, 678]
Alon N., Transversal numbers of uniform hypergraphs. Graphs Combin. 6 (1990), 1–4. [666, 670]
Alon N., Choice numbers of graphs: a probabilistic approach. Combin. Probab. Comput. 1 (1992), 107–
114. [356]
Alon N., Restricted colorings of graphs. In Surveys in Combinatorics, London Math. Soc. Lect. Notes 187
(Cambridge Univ. Press, 1993), 1–33. [661, 738]
Alon N., Probabilistic methods in coloring and decomposition problems. Discr. Math. 127 (1994), 31–46.
[678]
Alon N., Disjoint directed cycles. J. Combin. Th. B 68 (1996), 167–178. [663, 685]
References 851

Alon N., Combinatorial Nullstellensatz. Combin. Probab. Comput. 8 (1999), 7–29. [733]
Alon N., Additive Latin transversals. Israel J. Math. 117 (2000a), 125–130. [746]
Alon N., Degrees and choice numbers. Random Structures & Algorithms 16 (2000b), 364–358. [661]
Alon N., Bipartite decomposition of random graphs. J. Combin. Th. B 113 (2015), 220–235. [771]
Alon N., L. Babai, and H. Suzuki, Multilinear polynomials and Frankl–Ray-Chaudhuri–Wilson type
intersection theorems. J. Combin. Th. A 58 (1991), 165–180. [728, 731]
Alon N., T. Bohman, R. Holzman, and D.J. Kleitman, On partitions of discrete boxes. Discr. Math. 257
(2002), 255–258. [665]
Alon N., T. Bohman, and H. Huang, More on the bipartite decomposition of random graphs. J. Graph
Th. 84 (2017), 45–52. [771]
Alon N. and R.B. Boppana, The monotone circuit complexity of Boolean functions. Combinatorica 7
(1987), 1–22. [504]
Alon N., P. Frankl, and L. Lovász, The chromatic number of Kneser hypergraphs. Trans. Amer. Math.
Soc. 298 (1986), 359–370. [503]
Alon N. and S. Friedland, The maximum number of perfect matchings in graphs with a given degree
sequence. Electron. J. Combin. 15 (2008), Note 13, 2. [510]
Alon N., S. Friedland, and G. Kalai, Regular subgraphs of almost regular graphs. J. Combin. Theory B
37 (1984), 79–91. [736]
Alon N. and Z. Füredi, Covering the cube by affine hyperplanes. European J. Combin. 14 (1993), 79–83.
[746]
Alon N., Z. Füredi, and M. Katchalski, Separating pairs of points by standard boxes. European J. Combin.
6 (1985), 205–210. [442]
Alon N., A. Kostochka, B. Reiniger, D.B. West, and X. Zhu, Coloring, sparseness and girth. Israel J.
Math. 214 (2016), 315–331. [434, 442]
Alon N., M. Krivelevich, and B. Sudakov, Turán numbers of bipartite graphs and related Ramsey-type
questions. 12 (2003), 477–494. [672, 673, 684]
Alon N., C.J.H. McDiarmid, and M. Molloy, Edge-disjoint cycles in regular directed graphs. J. Graph
Theory 22 (1996), 231–237. [685]
Alon N. and V.D. Milman, Eigenvalues, expanders and superconcentrators. In Proc. 25th IEEE Symp.
Found. Comp. Sci. (IEEE, 1984), 320–322. [779, 780]
Alon N. and V.D. Milman, ë 1 , isoperimetric inequalities for graphs, and superconcentrators. J. Combin.
Th. B 38 (1985), 73–88. [778, 779]
Alon N., M.B. Nathanson, and I. Rusza, Adding distinct congruence classes modulo a prime. Amer. Math.
Monthly 102 (1995), 250–255. [746]
Alon N., M.B. Nathanson, and I. Rusza, The polynomial method and restricted sums of congruence
classes. J. Number Theory 56 (1996), 404–417. [734, 735, 745, 746]
Alon N., P. Prałat, and N. Wormald, Cleaning regular graphs with brushes. SIAM J. Discrete Math. 23
(2008/09), 233–250. [704]
Alon N. and P. Pudlák, Constructive lower bounds for off-diagonal Ramsey numbers. Israel J. Math. 122
(2001), 243–251. [504]
Alon N., L. Rónyai, and T. Szabó, Norm-graphs: variations and applications. J. Combin. Th. B 76 (1999),
280–290. [633]
Alon N. and E.R. Scheinerman, Degrees of freedom versus dimension for containment orders. Order 5
(1988), 11–16. [829, 830, 832]
Alon N., P.D. Seymour, and R. Thomas, Planar separators. SIAM J. Discr. Math. 7 (1994), 184–193.
[396, 397]
Alon N., A. Shpilka, and C. Umans, On sunflowers and matrix multiplication. Comput. Complexity 22
(2013), 219–243. [504]
Alon N. and J. Spencer, The Probabilistic Method (Wiley, 1992). Also 2000, 2008, 2016.
[504, 657, 662, 663, 666, 669, 672–674, 688, 696, 706, 721, 722, 780]
Alon N. and M. Tarsi, A nowhere-zero point in linear mappings. Combinatorica 9 (1989), 393–5. [764]
852 References

Alon N. and M. Tarsi, Colorings and orientations of graphs. Combinatorica 12 (1992), 125–134.
[738, 739, 746]
Alon N. and D.B. West, The Borsuk–Ulam theorem and bisection of necklaces. Proc. Amer. Math. Soc.
98 (1986), 623–628. [809]
Alon N. and N. Wormald, High degree graphs contain large-star factors. In Fete of combinatorics and
computer science, Bolyai Soc. Math. Stud. 20 (János Bolyai Math. Soc., 2010), 9–21. [671]
Alspach B., Research problem 3. Discr. Math. 36 (1981), 333. [646]
Alspach B. and H. Gavlas, Cycle decompositions of K n and K n − I. J. Combin. Th. B 81 (2001), 77–99.
[646]
Alspach B. and R. Häggkvist, Some observations on the Oberwolfach problem. J. Graph Th. 9 (1985),
177–187. [647]
Alspach B., P.J. Schellenberg, D.R. Stinson, and D. Wagner, The Oberwolfach problem and factors of
uniform odd length cycles. J. Combin. Th. A 52 (1989), 20–43. [647]
Altinisik E., B.E. Sagan, and N. Tuglu, GCD matrices, posets, and nonintersecting paths. Linear Mul-
tilinear Algebra 53 (2005), 75–84. [177, 766]
Amahashi A. and M. Kano, On factors with given components. Discr. Math. 42 (1982), 1–6. [273]
Andersen L.D., On edge-colourings of graphs. Math. Scand. 40 (1977), 161–175. [360]
Andersen L.D., The strong chromatic index of a cubic graph is at most 10. 108 (1992), 231–252. [373]
Anderson I., Perfect matchings of a graph. J. Combin. Th. B 10 (1971), 183–186. [264]
Anderson I., Sufficient conditions for matchings. Proc. Edinburgh Math. Soc. (2) 18 (1973), 129–136.
[275]
Anderson I., Intersection theorems and a lemma of Kleitman. Discr. Math. 16 (1976), 181–185.
[497, 561, 595, 607]
Anderson I., Combinatorial Designs: Construction Methods (Ellis Horwood, 1990). [654]
Anderson I., Combinatorial designs and tournaments, Oxford Lecture Series in Mathematics and its
Applications 6 (Oxford Univ. Press, 1997). [609, 612, 649]
Anderson M., R.B. Richter, and P. Rodney, The crossing number of C6 × C6 . In Proc. 27th Southeastern
Intl. Conf. Combin. Graph Th. Comput. (Boca Raton), Congr. Numer. 118 (1996), 97–107. [794]
Anderson M., R.B. Richter, and P. Rodney, The crossing number of C7 × C7 . In Proc. 28th Southeastern
Intl. Conf. Combin. Graph Th. Comput. (Boca Raton), Congr. Numer. 125 (1997), 97–117. [794]
Ando K., A. Kaneko, and S.V. Gervacio, The bandwidth of a tree with ¾ leaves is at most ⌈¾/2⌉. Discr.
Math. 150 (1996), 403–406. [249, 819]
André D., Sur les permutations alternées. J. Math. Pures et Appliquées 7 (1881), 167–184. [208]
André D., Théorème sur les formes quadratiques. Bull. Soc. Math. France 15 (1887), 188–192. [40]
Andreae T., On the unit interval number of a graph. Discrete Appl. Math. 22 (1988), 1–7. [366]
Andrews G. and E. Deutsch, Problem 11908. Amer. Math. Monthly 123 (2016), 504. [148]
Andrews G.E., The theory of partitions (Addison-Wesley, 1976). [140]
Andrews G.E., Problem 10627. Amer. Math. Monthly 104 (1997a), 974. Solution 107 (2000), 86. [150]
Andrews G.E., Problem 10628. Amer. Math. Monthly 104 (1997b), 974. Solution 107 (2000), 87. [151]
Andrews G.E. and P. Paule, Solution to problem E3376. Amer. Math. Monthly 99 (1992), 63–65. Proposed
97 (1990), 240. [66]
Andrews P. and E.T. Wang, Problem E3260. Amer. Math. Monthly 95 (1988), 350. Solution 97 (1990),
74–75. [174]
Appel K. and W. Haken, Every planar map is four colorable. Bull. Amer. Math. Soc. 82 (1976), 711–712.
[402]
Appel K., W. Haken, and J. Koch, Every planar map is four colorable. Part II: Reducibility. Illinois J.
Math. 21 (1977), 491–567. [406]
Arnautov V.I., Estimation of the exterior stability number of a graph by means of the minimal degree of
the vertices (Russian). Prikl. Mat. i Programmirovanie 11 (1974), 3–8, 126. [670]
Arratia R., On the Stanley–Wilf conjecture for the number of permutations avoiding a given pattern.
Electron. J. Combin. 6 (1999), Note, N1, 4 pp. (electronic). [432, 433]
References 853

Arrow K.J., Social Choice and Individual Values, Cowles Commission Monograph No. 12 (Wiley & Sons;
Chapman & Hall, 1951). [585]
Ash P., Problem 11520. Amer. Math. Monthly 117 (2010), 649. Solution 119 (2012), 701. [34]
Avis D., Nonpartitionable point sets. Inform. Process. Lett. 19 (1984), 125–129. [820]
Axenovich M., Z. Füredi, and D. Mubayi, On generalized Ramsey theory: the bipartite case. J. Combin.
Th. B 79 (2000), 66–86. [686]
Ayel J., Hamiltonian cycles in particular ¾-partite graphs. J. Combin. Th. B 32 (1982), 223–228. [331]
Azuma K., Weighted sums of certain dependent random variables. Tôhoku Math. J. (2) 19 (1967), 357–
367. [715, 722]
Babai L., A short proof of the nonuniform Ray-Chaudhuri–Wilson inequality. Combinatorica 8 (1988),
133–135. [727]
Babai L., Graph isomorphism in quasipolynomial time (2015). (arXiv:1512.03547). [699]
Babai L., P. Erdős, and S.M. Selkow, Random graph isomorphisms. SIAM J. Comput. 9 (1980), 628–635.
[699, 705]
Babai L. and P. Frankl, Linear algebra methods in combinatorics with applications to geometry and
computer science (Univ. of Chicago, 1992). [723, 744]
Babai L., W.M. Kantor, and E.M. Luks, Computational complexity and the classification of finite simple
groups. In Proc. 24th IEEE Symp. Found. Comp. Sci. (1983), 162–171. [699]
Babai L. and L. Kučera, Canonical labelling of graphs in linear average time. In Proc. 20th IEEE Symp.
Found. Comp. Sci. (IEEE, 1979), 39–46. [699]
Bäbler F., über die Zerlegung regulärer Streckenkomplexe ungerader Ordnung (German). Comment.
Math. Helv. 10 (1938), 275–287. [267]
Bäbler F., Über eine spezielle Klasse Euler’scher Graphen. Comment. Math. Helv. 27 (1953), 81–100.
[250]
Bach E., Exact analysis of a priority queue algorithm for random variate generation. In Proc. 5th ACM-
SIAM Symp. Discrete Algorithms (Arlington, 1994) (ACM, 1994), 48–56. [105]
Bacher R., Problem 10891. Amer. Math. Monthly 108 (2001), 668. Solution 110 (2003), 439–440. [66]
Bachmann P., Die Analytische Zahlentheorie (German) (Teubner, 1894). [10]
Backelin J., J. West, and G. Xin, Wilf-equivalence for singleton classes. Adv. in Appl. Math. 38 (2007),
133–148. [442]
Bagga J.S. and B.N. Varma, Hamiltonian properties in bipartite graphs. Bull. Inst. Combin. Appl. 26
(1999), 71–85. [325]
Baker K.A., Dimension, join-independence, and breadth in partially ordered sets (1961). Mimeographed
notes. [571]
Baker K.A., A generalization of Sperner’s lemma. J. Combinatorial Theory 6 (1969), 224–225. [561]
Balko M., J. Cibulka, D. Král’, and J. Kynčl, Ramsey numbers of ordered graphs 49 (2015), 419–424.
[457]
Ball S. and V. Pepe, Asymptotic improvements to the lower bound of certain bipartite Turán numbers.
Combin. Probab. Comput. 21 (2012), 323–329. [633]
Ball S. and V. Pepe, Forbidden subgraphs in the norm graph. Discr. Math. 339 (2016), 1206–1211.
[633]
Ball W.W.R., Mathematical Recreations and Essays (McMillan, 1892). [70]
Balof B. and K.P. Bogart, Simple inductive proofs of the Fishburn and Mirkin theorem and the Scott–
Suppes theorem. Order 20 (2003), 49–51. [586]
Balogh J. and J.A. Csirik, Index assignment for two-channel quantization. IEEE Trans. Inform. Theory
50 (2004), 2737–2751. [801]
Balogh J. and H. Kaul, A threshold for random geometric graphs with a hamiltonian cycle (2007). Un-
published manuscript. [819]
Balogh J., M. Kochol, A. Pluhár, and X. Yu, Covering planar graphs with forests. J. Combin. Th. B 94
(2005), 147–158. [421, 424]
Banderier C. and S.R. Schwer, Why Delannoy numbers? J. Statist. Plann. Inference 135 (2005), 40–54.
[28, 41]
854 References

Bandlow J., An elementary proof of the hook formula. Electron. J. Combin. 15 (2008), Research paper
45, 14. [190]
Bang C.M. and H. Sharp, Jr., Score vectors of tournaments. J. Combin. Th. B 26 (1979), 81–84. [261]
Bang S.-J., Problem 10490. Amer. Math. Monthly 102 (1995), 930. Solution 106 (1999), 588–589. [32]
Bang-Jensen J., S. Bessy, and S. Thomassé, Disjoint 3-cycles in tournaments: a proof of the Bermond–
Thomassen conjecture for tournaments. J. Graph Theory 75 (2014), 284–302. [685]
Bannai E. and T. Ito, On finite Moore graphs. J. Fac. Sci. Univ. Tokyo Sect. IA Math. 20 (1973), 191–
208. [244, 775]
Bapat R.B., Graphs and matrices, Universitext (Springer, 2010). [766, 769]
Bárány I., A short proof of Kneser’s conjecture. J. Combin. Theory A 25 (1978), 325–326. [812, 820]
Bárány I., S.B. Shlosman, and A. Szűcs, On a topological generalization of a theorem of Tverberg. J.
London Math. Soc. (2) 23 (1981), 158–164. [810]
Barát J. and C. Thomassen, Claw-decompositions and Tutte-orientations. J. Graph Theory 52 (2006),
135–146. [238]
Barber B., D. Kühn, A. Lo, R. Montgomery, and D. Osthus, Fractional clique decompositions of dense
graphs and hypergraphs. J. Combin. Th. B 127 (2017), 148–186. [640]
Barber B., D. Kühn, A. Lo, and D. Osthus, Edge-decompositions of graphs with high minimum degree.
Adv. Math. 288 (2016), 337–385. [640]
Barra M., Editorial comment to solution of problem 10663. Amer. Math. Monthly 107 (2000), 370. Pro-
posed 105 (1998), 464. [62]
Barrera-Cruz F. and P. Haxell, A note on Schnyder’s theorem. Order 28 (2011), 221–226. [790]
Basavaraju M., L.S. Chandran, and M. Kummini, d-regular graphs of acyclic chromatic index at least
d + 2. J. Graph Theory 63 (2010), 226–230. [329]
Basin S.L. and V.E. Hoggatt, Jr., A primer on the Fibonacci sequence, part II. Fibonacci Quart. 1 (1963),
61–68. [61]
Batagelj V. and T. Pisanski, Hamiltonian cycles in the Cartesian product of a tree and a cycle. Discr.
Math. 38 (1982), 311–312. [330]
Bauer D., Regular K n -free graphs. J. Combin. Th. B 35 (1983), 193–200. [490]
Bauer D., H.J. Broersma, and E.F. Schmeichel, Toughness in graphs—a survey. Graphs Combin. 22
(2006), 1–35. [318]
Bauer D., H.J. Broersma, and H.J. Veldman, Not every 2-tough graph is Hamiltonian. In Proc. 5th
Twente Workshop on Graphs and Combin. Optimization (Enschede, 1997), 99 (2000), 317–321.
[318]
Bauer D. and E.F. Schmeichel, Hamiltonian degree conditions which imply a graph is pancyclic. J. Com-
bin. Th. B 48 (1990), 111–116. [326]
Baumert L.D., Cyclic difference sets, Lect. Notes Math. 182 (Springer-Verlag, 1971). [637]
Baumert L.D., S.W. Golomb, and M. Hall, Jr., Discovery of an Hadamard matrix of order 92. Bull. Amer.
Math. Soc. 68 (1962), 237–238. [622]
Bean D.R., Effective coloration. J. Symbolic Logic 41 (1976), 469–480. [342]
Beck J., On 3-chromatic hypergraphs. Discr. Math. 24 (1978), 127–137. [659, 673]
Beck J., On size Ramsey number of paths, trees, and circuits. I. J. Graph Th. 7 (1983), 115–129. [455]
Beck J., An algorithmic approach to the Lovász local lemma. I. Random Structures Algorithms 2 (1991),
343–365. [680]
Beckwith D., Problem 10669. Amer. Math. Monthly 105 (1998), 559. Solution 107 (2000), 568–9. [152]
Beckwith D., Problem 10809. Amer. Math. Monthly 107 (2000), 566. Solution 109 (2002), 477–8. [149]
Beckwith D., Problem 10865. Amer. Math. Monthly 108(2001),371. Solution 109 (2002), 859–60. [764]
Beckwith D., Problem 11183. Amer. Math. Monthly 112 (2005), 839. Solution 114 (2007), 551–2. [148]
Beckwith D., Problem 11212. Amer. Math. Monthly 113(2006a),268. Solution 115 (2008), 366–7. [173]
Beckwith D., Problem 11249. Amer. Math. Monthly 113(2006b),760. Solution 115 (2008), 859–60. [49]
Beckwith D., Problem 11343. Amer. Math. Monthly 115 (2008), 166. Solution 116 (2009), 944–5. [32]
Beckwith D., Problem 11464. Amer. Math. Monthly 116 (2009), 845. Solution 118 (2011), 750. [149]
Beckwith D., Problem 11583. Amer. Math. Monthly 118 (2011), 558. Solution 120 (2013), 756–7. [149]
References 855

Beckwith D., Problem 11754. Amer. Math. Monthly 121 (2014), 170. Solution 123 (2016), 298. [59]
Behrend F.A., On sets of integers which contain no three terms in arithmetical progression. Proc. Nat.
Acad. Sci. U. S. A. 32 (1946), 331–332. [484]
Behzad M., Graphs and their chromatic numbers (Michigan State U., 1965). Ph.D. Thesis. [373, 654]
Behzad M. and S.E. Mahmoodian, On topological invariants of the product of graphs. Canad. Math. Bull.
12 (1969), 157–166. [329, 343, 398]
Beineke L.W. and R.D. Ringeisen, On the crossing numbers of products of cycles and graphs of order
four. J. Graph Theory 4 (1980), 145–155. [798]
Beineke L.W. and A.J. Schwenk, On a bipartite form of the Ramsey problem. In Proc. 5th British Combin.
Conf. (Univ. Aberdeen, 1975), Congr. Numer. 15 (Utilitas Math., 1976), 17–22. [620, 621, 624]
Belck H.B., Reguläre Faktoren von Graphen. J. Reine Angew. Math. 188 (1950), 228–252. [268, 276]
Belevitch V., Theory of 2n-terminal networks with applications to conference telephony. Electr. Commun.
27 (1950), 231–244. [621]
Bender E.A. and E.R. Canfield, The asymptotic number of labeled graphs with given degree sequences.
J. Combin. Th. A 24 (1978), 296–307. [697, 704]
Bender E.A. and J.R. Goldman, Enumerative uses of generating functions. Indiana Univ. Math. J. 20
(1971), 753–765. [127]
Bender E.A. and D.E. Knuth, Enumeration of plane partitions. J. Combin. Th. A 13 (1972), 40–54.
[207]
Bender E.A., F. Kochman, and D.B. West, Adding up to powers. Amer. Math. Monthly 97 (1990), 139–
143. [62]
Benevides F.S. and J. Skokan, The 3-colored Ramsey number of even cycles. J. Combin. Th. B 99 (2009),
690–708. [461]
Benhocine A. and A.P. Wojda, The Geng–Hua Fan conditions for pancyclic or Hamilton-connected
graphs. J. Combin. Th. B 42 (1987), 167–180. [326]
Benjamin A.T., G.M. Levin, K. Mahlburg, and J.J. Quinn, Random approaches to Fibonacci identities.
Amer. Math. Monthly 107 (2000), 511–516. [664]
Benjamin A.T. and J.J. Quinn, The Fibonacci numbers—exposed more discretely. Math. Mag. 76 (2003),
182–192. [53]
Bennett G., Problem 10216. Amer. Math. Monthly 99 (1992), 363. Solution 101 (1994), 917. [152]
Bentz H.J., Proof of the Bulgarian Solitaire conjectures. Ars Combin. 23 (1987), 151–170. [144, 152]
Benzer S., On the topology of the genetic fine structure. Proc. Nat. Acad. Sci. USA 45 (1959), 1607–1620.
[337]
Berge C., Two theorems in graph theory. Proc. Nat. Acad. Sci. U.S.A. 43 (1957), 842–844. [277]
Berge C., Sur le couplage maximum d’un graphe. C.R. Acad. Sci. Paris 247 (1958), 258–259. [266]
Berge C., Les problèmes de coloration en théorie des graphes. Publ. Inst. Statist. Univ. Paris 9 (1960),
123–160. [366, 367]
Berge C., Perfect graphs. In Six papers on Graph Theory (Indian Stat. Institute, 1963), 1–21. [370]
Berge C., Une propriété des graphes ¾ -stables-critiques. In Combinatorial Structures and Their Appli-
cations (R. Guy, H. Hanani, N.W. Sauer, and J. Schönheim, eds.) (Gordon and Breach, 1970), 7–
11. [222, 263]
Berge C., Graphs and Hypergraphs (North-Holland, 1973). (translation and revision of Graphes et Hy-
pergraphes, Dunod, 1970). [267, 333, 335, 667]
Berge C., Nombres de coloration de l’hypergraphe h-parti complet. Ann. Mat. Pura Appl. (4) 103 (1975),
3–9. [501]
Berge C., A theorem related to the Chvátal conjecture. In Proc. 5th British Combin. Conf. (Univ. Ab-
erdeen, 1975), 15 (Utilitas Math., 1976), 35–40. [501]
Berge C., ¾ -optimal partitions of a directed graph. European J. Combin. 3 (1982), 97–101. [549]
Berge C., Path-partitions in directed graphs and posets. In Graphs and order (Banff, Alta., 1984), NATO
Adv. Sci. Inst. C Math. Phys. Sci. 147 (Reidel, 1985), 447–464. [549]
Berge C. and V. Chvátal, Topics on Perfect Graphs, Ann. Discr. Math. 21 (North-Holland, 1984). [370]
856 References

Berge C. and J.L. Ramı́rez-Alfonsı́n, Origins and genesis. In Perfect graphs, Wiley-Intersci. Ser. Discr.
Math. Optim. (Wiley, 2001), 1–12. [368]
Berger E. and I.B-A. Hartman, Proof of Berge’s strong path partition conjecture for ¾ = 2. European J.
Combin. 29 (2008), 179–192. [549]
Berger E. and I.B-A. Hartman, A unified approach to known and unknown cases of Berge’s conjecture.
J. Graph Th. 71 (2012), 317–330. [549]
Berlekamp E.R., On subsets with intersections of even cardinality. Canad. Math. Bull. 12 (1969), 471–
474. [724, 744]
Berman K.A., Parity results on connected º -factors. Discr. Math. 59 (1986), 1–8. [241, 249]
Bermond J.-C., On Hamiltonian walks. In Proc. Fifth Brit. Combin. Conf. (C.St.J.A. Nash-Williams and
J. Sheehan, eds.) (Utilitas Math., 1976), 41–51. [323]
Bermond J.-C. and C. Thomassen, Cycles in digraphs—a survey. J. Graph Theory 5 (1981), 1–43.
[316, 685]
Bernardi C., On a theorem about vertex colorings of graphs. Discr. Math. 64 (1987), 95–96. [355]
Bernshteyn A., The local cut lemma. European J. Combin. 63 (2017), 95–114. [680]
Bertrand J., Solution d’un problème. Comptes Rendus Acad. Sci., Paris 105 (1887), 369. [39]
Beth T., D. Jungnickel, and H. Lenz, Design Theory (Cambridge Univ. Press, 1986). [609]
Beutelspacher A. and W. Brestovansky, Generalized Schur numbers. In Combinatorial theory (Schloss
Rauischholzhausen, 1982), Lecture Notes in Math. 969 (Springer, 1982), 30–38. [473]
Bey C. and J.R. Griggs, Problem 10932. Amer. Math. Monthly 109 (2002), 298. Solution 111 (2004), 262–
263. [511]
Bhanu K.S. and M.N. Deshpande, Problem 11503. Amer. Math. Monthly 117 (2010), 458. Solution 119
(2012), 349–350. [669]
Bhasker J., T. Samad, and D.B. West, Size, chromatic number, and connectivity. Graphs Combin. 10
(1994), 209–213. [353]
Bhatt S.N. and C.E. Leiserson, Minimizing the longest edge in a VLSI layout (1981). MIT memo. [809]
Bialostocki A. and J. Schönheim, On some Turán and Ramsey numbers for C4 . In Graph theory and
combinatorics (Cambridge, 1983) (Academic Press, 1984), 29–33. [490]
Biedl T., E.D. Demaine, C.A. Duncan, R. Fleischer, and S.G. Kobourov, Tight bounds on maximal and
maximum matchings. Discr. Math. 285 (2004), 7–15. [286]
Bienstock D., Some provably hard crossing number problems. Discrete Comput. Geom. 6 (1991), 443–
459. [795]
Bienstock D. and N. Dean, New results on rectilinear crossing numbers and plane embeddings. J. Graph
Theory 16 (1992), 389–398. [795]
Bienstock D. and N. Dean, Bounds for rectilinear crossing numbers. J. Graph Theory 17 (1993), 333–
348. [795, 796]
Biggs N., Algebraic graph theory (Cambridge Univ. Press, 1974). [766]
Biggs N.L., E.K. Lloyd, and R.J. Wilson, Graph theory: 1736–1936 (Clarendon Press, 1976). [37, 46]
Binet J.P., Mémoire sur l’intégration des équations linéaires aux différences finies, d’un ordre quel-
conque, à coefficients variables. Comptes Rendus hebdomadaires des séances de l’Académie des
Sciences 17 (1843), 559–567. [67]
Birkhoff G., The reducibility of maps. Amer. J. Math. 35 (1913), 114–128. [405, 421]
Birkhoff G., On the structure of abstract algebras. Proc. Cambridge Phil. Soc 31 (1935), 433–454.
[594]
Birkhoff G., Tres observaciones sobre el algebra lineal. Rev. Univ. Nac. Tucumán, Series A 5 (1946), 147–
151. [257]
Birkhoff G., Lattice theory, Third edition. American Mathematical Society Colloquium Publications, Vol.
XXV (AMS, 1967). [607]
Biró C., P. Hamburger, A. Pór, and W.T. Trotter, Forcing posets with large dimension to contain large
standard examples. Graphs Combin. 32 (2016), 861–880. [580]
Bixby R.E., Matroids and operations research. In Advanced techniques in practice of operations research
(H.J. Greenberg, F.H. Murphy, and S.H. Shaw, eds.) (North-Holland, 1981), 333–458. [520]
References 857

Björner A., J. Matoušek, and G.M. Ziegler, Using Brouwer’s fixed point theorem, A Journey through Discr.
Mathematics. A Tribute to Jiřı́Matoušek (Springer, to appear). (arXiv:1409.7890). [804]
Blagojević P.V.M., F. Frick, A. Haase, and G.M. Ziegler, Hyperplane mass partitions via relative equi-
variant obstruction theory. Doc. Math. 21 (2016), 735–771. [815]
Blagojević P.V.M., F. Frick, A. Haase, and G.M. Ziegler, Topology of the Grünbaum–Hadwiger–Ramos
hyperplane mass partition problem. Trans. Amer. Math. Soc. 370 (2018), 6795–6824. [815]
Blass A. and F. Harary, Properties of almost all graphs and complexes. J. Graph Th. 3 (1979), 225–240.
[690]
Blokhuis A., A new upper bound for the cardinality of 2-distance sets in Euclidean space. In Convexity
and graph theory (Jerusalem, 1981), North-Holland Math. Stud. 87 (1984), 65–66. [725, 744]
Bloom D.M., Problem 10921. Amer. Math. Monthly 109 (2002), 201. Solution 110 (2003), 958–9. [33]
Bloome L., P. Johnson, and N. Saritzky, Problem 11625. Amer. Math. Monthly 119 (2012), 162. Solution
121 (2014), 273. [344]
Boesch F. and R. Tindell, Robbins’s theorem for mixed multigraphs. Amer. Math. Monthly 87 (1980),
716–719. [237]
Bogart K.P., Introductory combinatorics, 2nd ed. (Harcourt Brace Jovanovich, 1990). Also 1983, 2000.
[147]
Bogart K.P., An obvious proof of Fishburn’s interval order theorem. Discr. Math. 118 (1993), 239–242.
[586]
Bogart K.P. and W.T. Trotter, Maximal dimensional partially ordered sets. II. Characterization of 2n-
element posets with dimension n. Discr. Math. 5 (1973), 33–43. [583]
Bogart K.P. and D.B. West, A short proof that “proper = unit”. Discr. Math. 201 (1999), 21–23. [587]
Bognár J., J. Mogyoródi, A. Prékopa, A. Rényi, and D. Szász, Problem Book on Probability (Hungarian)
(Tankönyvkiadó, 1970). [105]
Bohman T., The triangle-free process. Adv. Math. 221 (2009), 1653–1677. [682, 685]
Bohman T. and P. Keevash, Dynamic concentration of the triangle-free process. In 7th European Conf.
Combin. Graph Th. Appl., CRM Series 16 (Ed. Norm., 2013), 489–495. [682]
Bollobás B., On generalized graphs. Acta Math. Acad. Sci. Hungar. 16 (1965), 447–452. [568, 668]
Bollobás B., Graphs with given diameter and maximal valency and with a minimal number of edges.
In Combinatorial mathematics and its applications (Proc. Conf., Oxford, 1969) (Academic Press,
1971), 25–37. [244]
Bollobás B., Sperner systems consisting of pairs of complementary subsets. J. Combin. Th. A 15 (1973),
363–366. [511]
Bollobás B., On complete subgraphs of different orders. Math. Proc. Cambridge Philos. Soc. 79 (1976a),
19–24. [238]
Bollobás B., Relations between sets of complete subgraphs. In Proc. 5th British Combin. Conf. (Univ.
Aberdeen, 1975), Congr. Numer. 15 (Utilitas Math., 1976b), 79–84. [478]
Bollobás B., Extremal graph theory, London Mathematical Society Monographs 11 (Academic Press,
1978). [215, 272, 311, 316, 335, 478, 632]
Bollobás B., A probabilistic proof of an asymptotic formula for the number of labelled regular graphs.
European J. Combin. 1 (1980), 311–316. [697]
Bollobás B., Threshold functions for small subgraphs. Math. Proc. Camb. Phil. Soc 90 (1981a), 197–206.
[703]
Bollobás B., Degree sequences of random graphs. Discr. Math. 33 (1981b), 1–19. [699, 700, 721]
Bollobás B., Vertices of given degree in a random graph. J. Graph Th. 6 (1982), 147–155. [699]
Bollobás B., Random Graphs (Academic Press, 1985). Also 2001. [657, 688]
Bollobás B., Combinatorics: Set systems, hypergraphs, families of vectors and combinatorial probability
(Cambridge Univ. Press, 1986). [499]
Bollobás B., The chromatic number of random graphs. Combinatorica 8 (1988), 49–55. [701, 717, 718]
Bollobás B., Modern graph theory, Graduate Texts in Mathematics 184 (Springer-Verlag, 1998). [488]
Bollobás B. and G.R. Brightwell, The height of a random partial order: concentration of measure. Ann.
Appl. Probab. 2 (1992), 1009–1018. [722]
858 References

Bollobás B. and P. Erdős, On the structure of edge graphs. Bull. London Math. Soc. 5 (1973), 317–321.
[482, 770]
Bollobás B. and P. Erdős, Cliques in random graphs. Math. Proc. Camb. Phil. Soc. 80 (1976), 419–427.
[701]
Bollobás B., P. Erdős, and M. Simonovits, On the structure of edge graphs. II. J. London Math. Soc. (2)
12 (1976), 219–224. [482]
Bollobás B. and A.J. Harris, List-colourings of graphs. Graphs Combin. 1 (1985), 115–127. [362, 373]
Bollobás B. and Y. Kohayakawa, An extension of the Erdős–Stone theorem. Combinatorica 14 (1994),
279–286. [482]
Bollobás B. and E. Szemerédi, Girth of sparse graphs. J. Graph Theory 39 (2002), 194–200. [250]
Bollobás B. and A. Thomason, Random graphs of small order. In Random graphs ’83 (Poznań, 1983),
North-Holland Math. Stud. 118 (North-Holland, 1985), 47–97. [692]
Bollobás B. and A.G. Thomason, On the girth of Hamiltonian weakly pancyclic graphs. J. Graph Theory
26 (1997), 165–173. [250]
Bollobás B. and A.G. Thomason, Proof of a conjecture of Mader, Erdős and Hajnal on topological complete
subgraphs. Europ. J. Combin. 19 (1998), 883–887. [351]
Bollobás B. and P. Winkler, The longest chain among random points in Euclidean space. Proc. Amer.
Math. Soc. 103 (1988), 347–353. [821]
Bóna M., Problem E3378. Amer. Math. Monthly 97 (1991), 340. Solution 99 (1992), 65-66. [474]
Bóna M., Permutations avoiding certain patterns: the case of length 4 and some generalizations. Discr.
Math. 175 (1997), 55–67. [432]
Bóna M., The solution of a conjecture of Stanley and Wilf for all layered patterns. J. Combin. Th. A 85
(1999), 96–104. [432]
Bóna M., A walk through combinatorics (World Scientific Publishing Co., Inc., River Edge, NJ, 2002).
[440]
Bóna M., Combinatorics of permutations, Discr. Mathematics and its Applications (Boca Raton) (Chap-
man & Hall/CRC, 2004). [101, 102, 106, 133, 158, 432, 442]
Bonamy M., Planar graphs with ≥ 8 are ( + 1)-edge-choosable. SIAM J. Discr. Math. 29 (2015), 1735–
1763. [413]
Bonamy M., T. Perrett, and L. Postle, Colouring graphs with sparse neighbourhoods: Bounds and appli-
cations (2018). (arXiv:1810.06704). [685]
Bonato A. and P. Prałat, Graph searching games and probabilistic methods, Discr. Mathematics and its
Applications (Boca Raton) (CRC Press, 2018). [708]
Bondarenko A., On Borsuk’s conjecture for two-distance sets. Discrete Comput. Geom. 51 (2014), 509–
515. [817]
Bondy J.A., Properties of graphs with constraints on degrees. Stud. Sci. Math. Hung. 4 (1969), 473–475.
[296]
Bondy J.A., Pancyclic graphs. I. J. Combin. Th. B 11 (1971a), 80–84. [326, 330, 460]
Bondy J.A., Large cycles in graphs. Discr. Math. 1 (1971b), 121–132. [323]
Bondy J.A., Induced subsets. J. Combin. Th. B 12 (1972a), 201–202. [251]
Bondy J.A., Variation on the Hamiltonian theme. Canad. Math. Bull. 15 (1972b), 57–62. [332]
Bondy J.A., A remark on two sufficient conditions for Hamilton cycles. Discr. Math. 22 (1978a), 191–193.
[316]
Bondy J.A., Hamilton cycles in graphs and digraphs. In Proc. 9th Southeastern Intl. Conf. Combin. Graph
Th. Comput. (Boca Raton), 21 (Utilitas Math., 1978b), 3–28. [332]
Bondy J.A., Longest paths and cycles in graphs of high degree, Res. Rep. No. CORR (Univ. Waterloo,
1980). [326]
Bondy J.A., Integrity in graph theory. In The theory and applications of graphs (Kalamazoo, Mich., 1980)
(Wiley, 1981), 117–125. [327]
Bondy J.A., Basic graph theory: paths and circuits. In Handbook of combinatorics, Vol. 1, 2 (Elsevier,
1995), 3–110. [316]
Bondy J.A. and V. Chvátal, A method in graph theory. Discr. Math. 15 (1976), 111–136. [320]
References 859

Bondy J.A. and M. Kouider, Hamiltonian cycles in regular 2-connected graphs. J. Combin. Th. B 44
(1988), 177–186. [326]
Bondy J.A. and S.C. Locke, Largest bipartite subgraphs in triangle-free graphs with maximum degree
three. J. Graph Theory 10 (1986), 477–504. [227]
Bondy J.A. and L. Lovász, Cycles through specified vertices of a graph. Combinatorica 1 (1981), 117–
140. [302]
Bondy J.A. and U.S.R. Murty, Graph Theory with Applications (North-Holland, 1976). Also 2007.
[215, 228, 330, 489, 639]
Bondy J.A. and U.S.R. Murty, Graph theory, Grad. Texts Math. 244 (Springer, 2008). [215]
Bondy J.A. and M. Simonovits, Cycles of even length in graphs. J. Combin. Th. B 16 (1974), 97–105.
[479, 684]
Bonnice W.E., On convex polygons determined by a finite planar set. Amer. Math. Monthly 81 (1974),
749–752. [458]
Booth K.S. and G.S. Luecker, Testing for the consecutive ones property, interval graphs, and graph
planarity using PQ-tree algorithms. J. Comp. Syst. Sci. 13 (1976), 335–379. [393]
Borchardt C.W., Ueber eine der Interpolation entsprechende Darstellung der Eliminations-Resultante.
J. Reine Angew. Math. 57 (1860), 111–121. [37]
Borg P., On Chvátal’s conjecture and a conjecture on families of signed sets. European J. Combin. 32
(2011), 140–145. [503]
Borg P. and K. Meagher, The Katona cycle proof of the Erdős–Ko–Rado theorem and its possibilities. J.
Algebraic Combin. 43 (2016), 915–939. [500]
Borodin O.V., Criterion of chromaticity of a degree prescription (in Russian). In Abstracts of IV All-Union
Conf. on Theoretical Cybernetics (Novosibirsk) (1977), 127–128. Details in “Problems of colour-
ing and of covering the vertex set of a graph by induced subgraphs,” Ph.D. Thesis (in Russian),
Novosibirsk, 1979. [348]
Borodin O.V., Solution of the Ringel problem on vertex-face coloring of planar graphs and coloring of 1-
planar graphs. Metody Diskret. Analiz. (1984), 12–26, 108. [416]
Borodin O.V., Solving the Kotzig and Grünbaum problems on the separability of a cycle in planar graphs.
Mat. Zametki 46 (1989a), 9–12, 103. [423]
Borodin O.V., On the total coloring of planar graphs. J. Reine Angew. Math. 394 (1989b), 180–185.
[411]
Borodin O.V., A generalization of Kotzig’s theorem and prescribed edge coloring of planar graphs. Mat.
Zametki 48 (1990), 22–28, 160. [412, 413, 423]
Borodin O.V., A new proof of the 6 color theorem. J. Graph Theory 19 (1995), 507–521. [416]
Borodin O.V., Structural properties of plane graphs without adjacent triangles and an application to 3-
colorings. J. Graph Theory 21 (1996a), 183–186. [414, 415]
Borodin O.V., Structural theorem on plane graphs with application to the entire coloring number. J.
Graph Theory 23 (1996b), 233–239. [423]
Borodin O.V., Colorings of plane graphs: A survey. Discr. Math. 313 (2013), 517–539. [407]
Borodin O.V., A.N. Glebov, A. Raspaud, and M.R. Salavatipour, Planar graphs without cycles of length
from 4 to 7 are 3-colorable. J. Combin. Th. B 93 (2005), 303–311. [414]
Borodin O.V. and A.O. Ivanova, 2-distance ( + 2)-coloring of planar graphs with girth six and ≥ 18.
Discr. Math. 309 (2009), 6496–6502. [412]
Borodin O.V. and A.V. Kostochka, On an upper bound of the graph’s chromatic number depending on
the graph’s degree and density. J. Combin. Th. B 23 (1977), 247–250. [355]
Borodin O.V., A.V. Kostochka, and D.R. Woodall, Total colorings of planar graphs with large maximum
degree. J. Graph Theory 26 (1997), 53–59. [364, 365, 413, 414]
Borodin O.V. and D.P. Sanders, On light edges and triangles in planar graphs of minimum degree five.
Math. Nachr. 170 (1994), 19–24. [416, 423]
Borozan V., G.J. Chang, N. Cohen, S. Fujita, N. Narayanan, R. Naserasr, and P. Valicov, From edge-
coloring to strong edge-coloring. Electron. J. Combin. 22 (2015). [373]
860 References

Borsuk K., Drei Sätz über die n-dimensionale euklidische Sphäre (German). Fundamenta Mathematicae
20 (1933), 177–190. [809, 810, 817]
Borůvka O., O jistém problému minimálnı́m (Czech). Práce mor. přı́rodevěd. spol. v Brně 3 (1926), 37–
58. [246, 519]
Bosák J., Hamiltonian lines in cubic graphs. In Theory of Graphs (Internat. Sympos., Rome, 1966) (Gor-
don and Breach; Dunod, 1967), 35–46. [319, 331, 418]
Bose R.C., On the construction of balanced incomplete block designs. Ann. Eugenics 9 (1939), 353–399.
[610, 615, 627, 641–643, 655]
Bose R.C. and S.S. Shrikhande, On the falsity of Euler’s conjecture about the non-existence of two or-
thogonal Latin squares of order 4t + 2. Proc. Nat. Acad. Sci. U.S.A. 45 (1959), 734–737. [651]
Bose R.C., S.S. Shrikhande, and E.T. Parker, Further results on the construction of mutually orthogonal
Latin squares and the falsity of Euler’s conjecture. Canad. J. Math. 12 (1960), 189–203.
[612, 652–654]
Bouchet A. and J.L. Fouquet, Trois types de décompositions d’un graphe en chaı̂nes. In Combinatorial
mathematics (Marseille-Luminy, 1981), North-Holland Math. Stud. 75 (North-Holland, 1983),
131–141. [273]
Boyer J. and W. Myrvold, Stop minding your P’s and Q’s: a simplified O(n) planar embedding algorithm.
In Proc. 10th ACM-SIAM Symp. Discrete Algs. (Baltimore) (Assoc. Comput. Mach., 1999), 140–
146. [393]
Brandstädt A., V.B. Le, and J.P. Spinrad, Graph classes: a survey, SIAM Monographs on Discr. Mathe-
matics and Applications (Society for Industrial and Applied Mathematics (SIAM), 1999).
[370]
Brandt A., M. Ferrara, M. Kumbhat, S. Loeb, D. Stolee, and M. Yancey, I,F-partitions of sparse graphs.
European J. Combin. 57 (2016), 1–12. [422]
Brandt J., Cycles of partitions. Proc. Amer. Math. Soc. 85 (1982), 483–486. [143–145]
Brandt S., Subtrees and subforests of graphs. J. Combin. Th. B 61 (1994), 63–70. [262, 356]
Brandt S., Expanding graphs and Ramsey numbers (1996). Bielefeld preprint server, No. A 96-24.
[454]
Branin F.H., Jr., The relation between Kron’s method and the classical methods of network analysis. In
IRE Wescon Convention Record, II (1959), 3–28. [763]
Brègman L.M., Certain properties of nonnegative matrices and their permanents. Dokl. Akad. Nauk
SSSR 211 (1973), 27–30. [510]
Brightwell G.R., Linear extensions of infinite posets. Discr. Math. 70 (1988), 113–136. [823, 831]
Brightwell G.R., Semiorders and the 13 – 23 conjecture. Order 5 (1989), 369–380. [823, 831]
Brightwell G.R., On the complexity of diagram testing. Order 10 (1993), 297–303. [543]
Brightwell G.R., Balanced pairs in partial orders. Discr. Math. 201 (1999), 25–52. [823]
Brightwell G.R., S. Felsner, and W.T. Trotter, Balancing pairs and the cross product conjecture. Order
12 (1995), 327–349. [823]
Brightwell G.R. and E.R. Scheinerman, Representations of planar graphs. SIAM J. Discr. Math. 6 (1993),
214–229. [394]
Brightwell G.R. and W.T. Trotter, The order dimension of convex polytopes. SIAM J. Discr. Math. 6
(1993), 230–245. [790]
Brightwell G.R. and W.T. Trotter, The order dimension of planar maps. SIAM J. Discr. Math. 10 (1997),
515–528. [790]
Brightwell G.R. and W.T. Trotter, A combinatorial approach to correlation inequalities. Discr. Math. 257
(2002), 311–327. [600]
Brightwell G.R. and P. Winkler, Counting linear extensions. Order 8 (1991), 225–242. [823]
Brightwell G.R. and C. Wright, The 1/3–/3 conjecture for 5-thin posets. SIAM J. Discr. Math. 5 (1992),
467–474. [823]
Broersma H., Z. Ryjáček, and I. Schiermeyer, Closure concepts: a survey. Graphs Combin. 16 (2000), 17–
48. [320]
References 861

Broersma H. and H. Tuinstra, Independence trees and Hamilton cycles. J. Graph Th. 29 (1998), 227–
237. [327]
Broersma H.J., On some intriguing problems in Hamiltonian graph theory—a survey. Discr. Math. 251
(2002), 47–69. [318]
Broersma H.J., F.V. Fomin, P.A. Golovach, and G.J. Woeginger, Backbone colorings for graphs: tree and
path backbones. J. Graph Theory 55 (2007), 137–152. [342]
Broline D.M., Renumbering of the faces of dice. Math. Mag. 52 (1979), 312–314. [103]
Broline D.M. and D.E. Loeb, The combinatorics of Mancala-type games: Ayo, Tchoukaillon, and 1/ .
UMAP J. 16 (1995), 21–36. [144]
Brooks R.L., On colouring the nodes of a network. Proc. Cambridge Phil. Soc. 37 (1941), 194–197.
[336]
Brouwer A.E. and W.H. Haemers, Spectra of graphs, Universitext (Springer, 2012). [766, 771]
Brouwer A.E. and A. Schrijver, The blocking number of an affine space. J. Combin. Th. A 24 (1978), 251–
253. [735]
Brouwer L.E.J., Über Abbildung von Mannigfaltigkeiten. Math. Ann. 71 (1911), 97–115. [805]
Brown J.L., Note on complete sequences of integers. Amer. Math. Monthly 68 (1961), 557–560. [441]
Brown T.C., P. Erdős, F.R.K. Chung, and R.L. Graham, Quantitative forms of a theorem of Hilbert. J.
Combin. Th. A 38 (1985), 210–216. [473]
Brown W.G., On graphs that do not contain a Thomsen graph. Canad. Math. Bull. 9 (1966), 281–285.
[632, 633]
Brown W.G. and F. Harary, Extremal digraphs (1970), 135–198. [228]
Brualdi R.A., Comments on bases in dependence structures. Bull. Austral. Math. Soc. 1 (1969), 161–167.
[538]
Brualdi R.A., Introductory combinatorics (5th ed.) (Pearson Prentice Hall, 2010). First ed. North–
Holland, 1977. [174]
Brualdi R.A. and S. Kirkland, Aztec diamonds and digraphs, and Hankel determinants of Schröder
numbers. J. Combin. Th. B 94 (2005), 334–351. [34]
Bruck R.H. and H.J. Ryser, The nonexistence of certain finite projective planes. Canadian J. Math. 1
(1949), 88–93. [616, 627]
Bruhn H., P. Charbit, and J.A. Telle, The graph formulation of the union-closed sets conjecture. In The
Seventh European Conference on Combinatorics, Graph Theory and Applications, CRM Series 16
(Ed. Norm., 2013), 73–78. [512]
Bruhn H. and O. Schaudt, The journey of the union-closed sets conjecture. Graphs Combin. 31 (2015),
2043–2074. [503]
Bryant D. and P. Danziger, On bipartite 2-factorizations of K n − I and the Oberwolfach problem. J.
Graph Th. 68 (2011), 22–37. [647, 648]
Bryant D., D. Horsley, and W. Pettersson, Cycle decompositions V: Complete graphs into cycles of arbi-
trary lengths. Proc. Lond. Math. Soc. (3) 108 (2014), 1153–1192. [646]
Bryant D. and C.A. Rodger, Cycle decompositions. In The CRC Handbook of Combinatorial Designs (2nd
ed.) (Colbourn, Charles J. and Dinitz, Jeffrey H., eds.) (CRC Press, 2007), 373–382. [647]
Bryant P.R., Graph theory applied to electrical networks. In Graph Theory and Theoretical Physics (Aca-
demic Press, 1967), 111–137. [763]
Brylawski T.H., Some properties of basic families of subsets. Discr. Math. 6 (1973), 333–341. [538]
Brylawski T.H., Appendix of matroid cryptomorphisms. In Theory of matroids (N. White, ed.), Encyc.
Math. Appl. 26 (Cambridge Univ. Press, 1986), 298–316. [528, 529]
Bucić M., An improved bound for disjoint directed cycles. Discr. Math. 341 (2018), 2231–2236. [685]
Bucić M., S. Letzter, and B. Sudakov, Three colour bipartite Ramsey number of cycles and paths (2018).
(arXiv:1803.03689). [460]
Buckley F. and M. Lewinter, A Friendly Introduction to Graph Theory (Prentice Hall, 2002). [215]
Bukh B. and Z. Jiang, A bound on the number of edges in graphs without an even cycle. Combin. Probab.
Comput. 26 (2017), 1–15. [479]
862 References

Bunde D.P., E.W. Chambers, D.W. Cranston, K.G. Milans, and D.B. West, Pebbling and optimal pebbling
in graphs. J. Graph Theory 57 (2008), 215–238. [668]
Buneman P., A characterization of rigid circuit graphs. Discr. Math. 9 (1974), 205–212. [373]
Buneman P. and L. Levy, The Towers of Hanoi problem. Info. Process. Lett. 10 (1980), 243–244. [80]
Buratti M., Rotational ¾ -cycle systems of order v < 3¾ ; another proof of the existence of odd cycle systems.
J. Combin. Des. 11 (2003), 433–441. [646]
Burns D. and S. Schuster, Embedding (p , p − 1) graphs in their complements. Israel J. Math. 30 (1978),
313–320. [251]
Burr S.A., Generalized Ramsey theory for graphs—a survey. In Graphs and Combinatorics, Proc. Capital
Conf. (Washington, 1973), Lect. Notes Math. 486 (Springer, 1974), 52–75. [459, 460]
Burr S.A., Ramsey numbers involving graphs with long suspended paths. J. London Math. Soc. (2) 24
(1981), 405–413. [459]
Burr S.A. and P. Erdős, On the magnitude of generalized Ramsey numbers for graphs. In Infinite and
finite sets (Colloq., Keszthely, 1973), I, Colloq. Math. Soc. János Bolyai 10 (North-Holland, 1975),
215–240. [484]
Burr S.A. and P. Erdős, Extremal Ramsey theory for graphs. Utilitas Math. 9 (1976), 247–258. [460]
Burr S.A. and P. Erdős, Generalizations of a Ramsey-theoretic result of Chvátal. J. Graph Th. 7 (1983),
39–51. [454]
Burr S.A., P. Erdős, and L. Lovász, On graphs of Ramsey type. Ars Combinatoria 1 (1976), 167–190.
[456, 457, 461]
Burr S.A., P. Erdős, and J. Spencer, Ramsey theorems for multiple copies of graphs. Trans. Amer. Math.
Soc. 209 (1975), 87–99. [454]
Burr S.A. and J.A. Roberts, On Ramsey numbers for stars. Utilitas Math. 4 (1973), 217–220. [459]
Buršteı̆n M.I., An upper bound for the chromatic number of hypergraphs. Sakharth. SSR Mecn. Akad.
Moambe 75 (1974), 37–40. [418, 419]
Burungale A., Problem 11262. Amer. Math. Monthly 113 (2006), 940. Solution 115 (2008), 862. [238]
Butler S. and J. Mao, Problem 11265. Amer. Math. Monthly 114 (2007), 77. Solution 116 (2009), 181.
[237]
Butterfield J., T. Grauman, W.B. Kinnersley, K.G. Milans, C. Stocker, and D.B. West, On-line Ramsey
theory for bounded degree graphs. Electron. J. Combin. 18 (2011), Paper 136, 17. [457]
Cabello S. and B. Mohar, Adding one edge to planar graphs makes crossing number and 1-planarity
hard. SIAM J. Comput. 42 (2013), 1803–1829. [795]
Callan D., Problem 10643. Amer. Math. Monthly 105 (1998), 175. Solution 107 (2000), 278–279. [63]
Callan D., Solution to problem 10596. Amer. Math. Monthly 106 (1999), 367. Proposed 104 (1997), 456.
[174]
Callan D., Solution to problem 10878. Amer. Math. Monthly 110 (2003a), 342–343. Proposed 108 (2001),
470. [35]
Callan D., Solution to problem 10894. Amer. Math. Monthly 110 (2003b), 443–444. Proposed 108 (2001),
770. [64]
Callan D., Problem 11013. Amer. Math. Monthly 110 (2003c), 438. Solution 112 (2005), 184. [81]
Callan D., Problem 11091. Amer. Math. Monthly 111 (2004), 534. Solution 113 (2006), 462–463. [49]
Callan D., Problem 11362. Amer. Math. Monthly 105 (2008), 461. Solution 117 (2010), 187. [59]
Callan D., Problem 11567. Amer. Math. Monthly 118 (2011), 371. Solution 120 (2013), 370. [134]
Callan D. and E. Deutsch, The run transform. Discr. Math. 312 (2012), 2927–2937. [151]
Cameron K. and J. Edmonds, Some graphic uses of an even number of odd nodes. Ann. Inst. Fourier
(Grenoble) 49 (1999), 815–827. [241, 242, 249, 276]
Cameron K.B., Polyhedral and Algorithmic Ramifications of Antichains (ProQuest, 1982). Ph.D. Thesis,
Univ. Waterloo. [549]
Cameron K.B., On ¾ -optimum dipath partitions and partial ¾ -colourings of acyclic digraphs. European
J. Combin. 7 (1986), 115–118. [549]
Cameron P.J. and J.H. van Lint, Designs, graphs, codes and their links, London Math. Soc. Student Texts
22 (Cambridge Univ Press, 1991). [609, 774]
References 863

Cameron P.J. and I.M. Wanless, Covering radius for sets of permutations. Discr. Math. 293 (2005), 91–
109. [261]
Camion P., Chemins et circuits hamiltoniens des graphes complets. C. R. Acad. Sci. Paris 249 (1959),
2151–2152. [334]
Campbell C. and W. Staton, On extremal regular graphs with given odd girth. In Proc. 22nd Southeastern
Intl. Conf. Combin. Graph Th. Comput. (Baton Rouge), Congr. Num. 81 (1991), 157–159. [238]
Canfield E.R., On a problem of Rota. Adv. in Math. 29 (1978), 1–10. [567]
Cannings C. and J. Haigh, Montreal solitaire. J. Combin. Th. A 60 (1992), 50–66. [144]
Cantor D.G. and W.H. Mills, Determination of a subset from certain combinatorial properties. Canadian
J. Math. 18 (1966), 42–48. [513]
Cao W., K.W. Hwang, and D.B. West, Improved bounds on families under ¾ -wise set-intersection con-
straints. Graphs Combin. 23 (2007), 381–386. [731]
Carlitz L., Eulerian numbers and polynomials. Math. Mag. 32 (1958/1959), 247–260. [101]
Carlitz L., D.P. Roselle, and R.A. Scoville, Some remarks on ballot-type sequences of positive integers.
J. Combinatorial Theory A 11 (1971), 258–271. [177]
Caro N. and C. Pohoata, Solution to problem 11403 (solved independently). Amer. Math. Monthly 118
(2011), 276–277. Proposed 115 (2008), 949. [174]
Caro Y., New results on the independence number. Tech. Rep. 05-79, Tel-Aviv University (1979).
[263, 662]
Caro Y., Zero-sum problems—a survey. Discr. Math. 152 (1996), 93–113. [461]
Caro Y., I. Krasikov, and Y. Roditty, On the largest tree of given maximum degree in a connected graph.
J. Graph Th. 15 (1991), 7–13. [327]
Caro Y. and Y. Roditty, Connected colorings of graphs. Ars Combin. 70 (2004), 191–196. [236]
Caro Y. and Z. Tuza, Improved lower bounds on ¾ -independence. J. Graph Theory 15 (1991), 99–107.
[263]
Cartwright D.P. and F. Harary, Structural balance: a generalization of Heider’s theory. Psychological
Review 63 (1956), 277–293. [237]
Catalan E., Note sur une équation aux différences finies. J. Math. Pures Appl. 3 (1838), 508–516. [41]
Catlin P.A., A bound on the chromatic number of a graph. Discr. Math. 22 (1978), 81–83. [355]
Catlin P.A., Hajós’ graph-coloring conjecture: variations and counterexamples. J. Combin. Th. B 26
(1979), 268–274. [349, 356]
Cayley A., A theorem on trees. Quart. J. Math. 23 (1889), 376–378. [36, 37, 47]
Chaiken S. and D.J. Kleitman, Matrix tree theorems. J. Combin. Th. A 24 (1978), 377–381. [749]
Chandrasekharan K., Arithmetical functions, Die Grundlehren der mathematischen Wissenschaften,
Band 167 (Springer-Verlag, 1970). [140]
Chang W.I. and E.L. Lawler, Edge coloring of hypergraphs and a conjecture of Erdős, Faber, Lovász.
Combinatorica 8 (1988), 293–295. [439]
Chappell G.G., Polyunsaturated posets and graphs and the Greene–Kleitman theorem. Discr. Math. 257
(2002), 329–340. [551]
Charalambides C.A., Enumerative combinatorics, CRC Press Series on Discr. Mathematics and its Ap-
plications (Chapman & Hall/CRC, 2002). [101]
Chartrand G., D. Geller, and S. Hedetniemi, Graphs with forbidden subgraphs. J. Combinatorial Theory
B 10 (1971), 12–41. [418]
Chartrand G. and F. Harary, Planar permutation graphs. Ann. Inst. H. Poincaré Sect. B (N.S.) 3 (1967),
433–438. [398]
Chartrand G. and F. Harary, Graphs with prescribed connectivities. In Theory of Graphs (Tihany, 1966)
(P. Erdős and G. Katona, eds.) (Academic Press, 1968), 61–63. [295]
Chartrand G., S.F. Kapoor, L.M. Lesniak, and S. Schuster, Near 1-factors in graphs. In Proc. 2nd West
Coast Conf. Combin. Graph Th. Comput. (Eugene, OR, 1983), Congr. Numer. 41 (1984), 131–147.
[273]
Chartrand G. and H.V. Kronk, The point-arboricity of planar graphs. J. London Math. Soc. 44 (1969),
612–616. [343]
864 References

Chartrand G. and L.M. Lesniak, Graphs and Digraphs (2nd ed.) (Wadsworth, 1986). Also 1996, 2005,
2011. [215, 250, 315, 377]
Chartrand G., A.D. Polimeni, and M.J. Stewart, The existence of 1-factors in line graphs, sqaures, and
total graphs. Indagationes Math. 35 (1973), 228–232. [272]
Chein M., Graphe régulièrement décomposable. Rev. Fran.̧ Info. Rech. Opér. 2 (1968), 27–42. [312]
Chen G., R.J. Gould, and X. Yu, Graph connectivity after path removal. Combinatorica 23 (2003), 185–
203. [314]
Chen G., M.S. Jacobson, A.E. Kézdy, and J. Lehel, Tough enough chordal graphs are Hamiltonian. Net-
works 31 (1998), 29–38. [318]
Chen G., J. Lehel, M.S. Jacobson, and W.E. Shreve, Note on graphs without repeated cycle lengths. J.
Graph Theory 29 (1998), 11–15. [249]
Chen G., R.H. Schelp, and B. Wei, Monochromatic-rainbow Ramsey numbers (2001). Presented 14th
Cumberland Conference. [473]
Chen Z.H. and H.J. Lai, The higher-order edge toughness of a graph and truncated uniformly dense
matroids. J. Combin. Math. Combin. Comput. 22 (1996), 157–160. [539]
Cherkashin D.D. and J. Kozik, A note on random greedy coloring of uniform hypergraphs. Random
Structures Algorithms 47 (2015), 407–413. [660, 673]
Chernoff H., A measure of asymptotic efficiency for tests of a hypothesis based on the sum of observations.
Ann. Math. Statistics 23 (1952), 493–507. [707]
Chetwynd A.G. and A.J.W. Hilton, Partial edge-colourings of complete graphs or of graphs which are
nearly complete. In Graph theory and combinatorics (Cambridge, 1983) (Academic Press, London,
1984), 81–97. [359]
Chetwynd A.G. and A.J.W. Hilton, Star multigraphs with three vertices of maximum degree. Math. Proc.
Cambridge Math. Soc. 100 (1986), 303–317. [359]
Chetwynd A.G. and A.J.W. Hilton, 1-factorizing regular graphs of high degree—an improved bound. In
Graph Theory and Combinatorics (Cambridge, 1988), Discr. Math. 75 (1989), 103–112. [359]
Chevalley C., Démonstration d’une hypothèse de M. Artin. Abh. Math. Sem. Univ. Hamburg 11 (1935),
73–75. [735]
Chevalley H. and E. Warning, Bemerkung zur vorstehenden Arbeit. Abh. Math. Sem. Univ. Hamburg
11 (1935), 76–83. [735]
Chiba S. and T. Yamashita, Degree conditions for the existence of vertex-disjoint cycles and paths: a
survey. Graphs Combin. 34 (2018), 1–83. [326]
Chiue W.S. and B.S. Shieh, On connectivity of the Cartesian product of two graphs. Appl. Math. Comput.
102 (1999), 129–137. [292]
Choi J.O., L. Özkahya, and D.B. West, Degree-splittability of multigraphs and caterpillars. In Proc. 41st
Southeastern Intl. Conf. Combin. Graph Th. Comput. (Boca Raton), Congr. Numer. 202 (2010),
137–147. [239]
Chow T., A short proof of the rook reciprocity theorem. Electron. J. Combin. 3 (1996), R10. [175]
Chowla S. and H.J. Ryser, Combinatorial problems. Canadian J. Math. 2 (1950), 93–99. [616]
Chudnovsky M., N. Robertson, P.D. Seymour, and R. Thomas, The strong perfect graph theorem. Ann.
of Math. (2) 164 (2006), 51–229. [370]
Chung F.R.K., On partitions of graphs into trees. Discr. Math. 23 (1978a), 23–30. [249]
Chung F.R.K., On concentrators, superconcentrators, generalizers and nonblocking networks. Bell Syst.
Tech. J. (1978b), 1765–1777. [779]
Chung F.R.K., Spectral graph theory, CBMS Conf. Series 92 (Amer. Math. Soc., 1997). [776]
Chung F.R.K. and R.L. Graham, On multicolor Ramsey numbers for complete bipartite graphs. J. Com-
bin. Th. B 18 (1975), 164–169. [460]
Chung F.R.K. and R.L. Graham, Erdős on graphs (A K Peters, 1998). [341, 446]
Chung F.R.K., R.L. Graham, P. Frankl, and J.B. Shearer, Some intersection theorems for ordered sets
and graphs. J. Combin. Th. A 43 (1986), 23–37. [508, 509]
Chung F.R.K. and C.M. Grinstead, A survey of bounds for classical Ramsey numbers. J. Graph Th. 7
(1983), 25–37. [451]
References 865

Chung F.R.K. and L. Lu, An upper bound for the Turán number t3(n , 4). J. Combin. Theory A 87 (1999),
381–389. [491]
Chung K.L. and W. Feller, On fluctuations in coin-tossing. Proc. Nat. Acad. Sci. U. S. A. 35 (1949), 605–
608. [50]
Chung M.-S. and D.B. West, Large P4 -free graphs with bounded degree. J. Graph Th. 17 (1993), 109–
116. [490]
Chvátal V., Planarity of graphs with given degrees of vertices. Nieuw Arch. Wisk. (3) 17 (1969), 47–60.
[389]
Chvátal V., The smallest triangle-free 4-chromatic 4-regular graph. J. Combin. Th. 9 (1970), 93–94.
[354]
Chvátal V., On Hamilton’s ideals. J. Combin. Th. B 12 (1972), 163–168. [320, 326, 332]
Chvátal V., Tough graphs and Hamiltonian circuits. Discr. Math. 2 (1973), 215–223. [318]
Chvátal V., Intersecting families of edges in hypergraphs having the hereditary property. In Hypergraph
Seminar (Proc. First Working Sem., Ohio State Univ., 1972; dedicated to Arnold Ross), Lecture
Notes in Math. 411 (Springer, 1974), 61–66. [458, 501, 502]
Chvátal V., A combinatorial theorem in plane geometry. J. Combin. Th. B 18 (1975), 39–41. [419]
Chvátal V., Tree-complete graph Ramsey numbers. J. Graph Th. 1 (1977), 93. [453]
Chvátal V., Star-cutsets and perfect graphs. J. Combin. Th. B 39 (1985b), 138–154. [375]
Chvátal V. and P. Erdős, A note on Hamiltonian circuits. Discr. Math. 2 (1972), 111–113. [321]
Chvátal V. and F. Harary, Generalized Ramsey theory for graphs, III. Small Off-diagonal Numbers. Pac.
J. Math. 41 (1972), 335–345. [454]
Chvátal V. and F. Harary, Generalized Ramsey theory for graphs, I. Diagonal numbers. Period. Math.
Hungar. 3 (1973), 115–124. [666]
Chvátal V. and J. Komlós, Some combinatorial theorems on monotonicity. Canad. Math. Bull. 14 (1971),
151–157. [458]
Chvátal V. and L. Lovász, Every directed graph has a semi-kernel. In Hypergraph Seminar (Columbus,
1972), Lect. Notes Math. 411 (Springer, 1974), 175. [229]
Chvátal V., V. Rödl, E. Szemerédi, and W.T. Trotter, The Ramsey numbers of a graph with bounded
maximum degree. J. Combin. Th. B 34 (1983), 239–243. [455, 483]
Chvátalová J., Optimal labelling of a product of two paths. Discr. Math. 11 (1975), 249–253. [800, 818]
Cigler J., Some remarks on Catalan families. European J. Combin. 8 (1987), 261–267. [50]
Clapham C.R.J., Hamiltonian arcs in self-complementary graphs. Discr. Math. 8 (1974), 251–255.
[332]
Clark D.S. and J.T. Lewis, Avoiding-sequences with minimum sum. Discrete Appl. Math. 22 (1989), 103–
108. [429]
Clark D.S. and J.T. Lewis, Circular avoiding sequences with prescribed sum. Discrete Appl. Math. 43
(1993), 27–36. [429]
Clark L.H., J.C. George, and T.D. Porter, On the number of 1-factors in the n-cube. In Proc. 28th South-
eastern Intl. Conf. Combin. Graph Th. Comput. (Boca Raton), Congr. Numer. 127 (1997), 67–69.
[764]
Clark P.L., The Combinatorial Nullstellensätze revisited. Electron. J. Combin. 21 (2014), Paper 4.15, 17.
[732]
Clements G.F., A minimization problem concerning subsets of a finite set. Discr. Math. 4 (1973), 123–
128. [511]
Clements G.F. and B. Lindström, A generalization of a combinatorial theorem of Macaulay. J. Combi-
natorial Theory 7 (1969), 230–238. [496]
Cohen N. and F. Havet, Planar graphs with maximum degree ≥ 9 are ( + 1)-edge-choosable—a short
proof. Discr. Math. 310 (2010), 3049–3051. [413]
Cohen-Addad V., M. Hebdige, D. Král’, Z. Li, and E. Salgado, Steinberg’s conjecture is false. J. Combin.
Th. B 122 (2017), 452–456. [414]
Colbourn C.J. and J.H. Dinitz (eds.), The CRC handbook of combinatorial designs, CRC Press Series on
Discr. Math. Appl. (CRC Press, 1996). Also 2007. [609, 646]
866 References

Colbourn C.J. and A. Rosa, Triple systems, Oxford Mathematical Monographs (Clarendon Press, 1999).
[641]
Coleman M., An answer to a question by Wilf on packing distinct patterns in a permutation. Electron.
J. Combin. 11 (2004), Note 8, 4 pp. [434, 442]
Comtet L., Analyse combinatoire. Tomes I, II, Collection SUP: “Le Mathématicien”, 4 5 (Presses Univer-
sitaires de France, 1970). [65]
Comtet L., Advanced combinatorics (D. Reidel Publishing Co., 1974). [65]
Conlon D., A new upper bound for diagonal Ramsey numbers. Ann. of Math. (2) 170 (2009a), 941–960.
[450]
Conlon D., On-line Ramsey numbers. SIAM J. Discr. Math. 23 (2009b/10), 1954–1963. [457]
Conlon D. and J. Fox, Graph removal lemmas. In Surveys in combinatorics 2013, London Math. Soc.
Lecture Note Ser. 409 (Cambridge Univ. Press, 2013), 1–49. [485]
Conlon D., J. Fox, C. Lee, and B. Sudakov, Ordered Ramsey numbers. J. Combin. Th. B 122 (2017), 353–
383. [457]
Conlon D., J. Fox, and B. Sudakov, Hypergraph Ramsey numbers. J. Amer. Math. Soc. 23 (2010), 247–
266. [451, 459, 608]
Conlon D., J. Fox, and B. Sudakov, On two problems in graph Ramsey theory. Combinatorica 32 (2012),
513–535. [484]
Conlon D., J. Fox, and B. Sudakov, An improved bound for the stepping-up lemma. Discrete Appl. Math.
161 (2013), 1191–1196. [451, 453]
Conlon D., J. Fox, and B. Sudakov, Cycle packing. Rand. Struct. Alg. 45 (2014), 608–626. [334]
Conlon D., J. Fox, and B. Sudakov, Recent developments in graph Ramsey theory. In Surveys in combi-
natorics 2015, London Math. Soc. Lecture Note Ser. 424 (Cambridge Univ. Press, 2015), 49–118.
[425]
Cook S., The complexity of theorem proving procedures. In Proc. 3rd ACM Symposium on Theory of
Computing (1971), 151–158. [11]
Corrádi K., Problem at Schweitzer competition. Mat. Lapo 20 (1969), 159–162. [511]
Corradi K. and A. Hajnal, On the maximal number of independent circuits in a graph. Acta Math. Acad.
Sci. Hungar. 14 (1963), 423–439. [490]
Cox D.A. and U. Thieu, Problem 11862. Amer. Math. Monthly 122 (2015), 802. Solution 124 (2017), 473.
[173]
Cranston D.W., Strong edge-coloring of graphs with maximum degree 4 using 22 colors. Discr. Math.
306 (2006), 2772–2778. [373]
Cranston D.W., Edge-choosability and total-choosability of planar graphs with no adjacent 3-cycles. Dis-
cuss. Math. Graph Theory 29 (2009), 163–178. [413, 423]
Cranston D.W., R. Erman, and R. Škrekovski, Choosability of the square of a planar graph with maxi-
mum degree four. Australas. J. Combin. 59 (2014), 86–97. [424]
Cranston D.W., S. Jahanbekam, and D.B. West, The 1 , 2 , 3-conjecture and 1 , 2-conjecture for sparse
graphs. Discuss. Math. Graph Theory 34 (2014), 769–799. [422]
Cranston D.W. and S.J. Kim, List-coloring the square of a subcubic graph. J. Graph Theory 57 (2008),
65–87. [424]
Cranston D.W., S.J. Kim, and G. Yu, Injective colorings of sparse graphs. Discr. Math. 310 (2010), 2965–
2973. [423, 424]
Cranston D.W. and L. Rabern, Brooks’ theorem and beyond. J. Graph Th. 80 (2015), 199–225. [336]
Cranston D.W. and D.B. West, Problem 11712. Amer. Math. Monthly 120 (2013), 569. Solution 122 (2015),
505–506. [152]
Cranston D.W. and D.B. West, An introduction to the discharging method via graph coloring. Discr.
Math. 340 (2017), 766–793. [407, 411]
Cranston D.W. and G. Yu, Linear choosability of sparse graphs. Discr. Math. 311 (2011), 1910–1917.
[423]
Crapo H.H. and G.C. Rota, On the Foundations of Combinatorial Theory: Combinatorial Geometries
(MIT Press, 1970). [520]
References 867

Cruse A.B., A note on i-factors in certain regular multigraphs. Discr. Math. 18 (1977), 213–216. [274]
Cull P., Tours of graphs, digraphs, and sequential machines. IEEE Trans. Comp. C29 (1980), 50–54.
[237]
Cull P. and E.F. Ecklund, On the Towers of Hanoi and generalized Towers of Hanoi problems. In Proc.
13th Southeastern Intl. Conf. Combin. Graph Th. comput. (Boca Raton), Congr. Numer. 35 (1982),
229–238. [80]
Curran S., O. Lee, and X. Yu, Finding four independent trees. SIAM J. Comput. 35 (2006), 1023–1058.
[315]
Cushing W. and H.A. Kierstead, Planar graphs are 1-relaxed, 4-choosable. European J. Combin. 31
(2010), 1385–1397. [742]
Cvetković D.M., M. Doob, and H. Sachs, Spectra of graphs: Theory and applications, Pure and Applied
Mathematics 87 (Academic Press, 1979). Also 1982, 1995. [766, 781]
Czipzer J., Solution to problem 127 (Hungarian). Mat. Lapok 14 (1963), 373–374. [216]
Dályay P.P., Problem 11631. Amer. Math. Monthly 119 (2012), 247–8. Solution 121 (2014), 367. [623]
Dályay P.P., Problem 11897. Amer. Math. Monthly 123 (2016), 297. Solution 125 (2018), 86. [115]
Damerell R.M., On Moore graphs. Proc. Cambridge Philos. Soc. 74 (1973), 227–236. [244, 775]
David K., Solution to problem 11144. Amer. Math. Monthly 114 (2007), 262–264. Proposed 112 (2005),
274. [47]
Daykin D.E., Erdős–Ko–Rado from Kruskal–Katona. J. Combin. Th. A 17 (1974a), 254–255. [500]
Daykin D.E., A simple proof of the Kruskal–Katona theorem. J. Combin. Th. A 17 (1974b), 252–253.
[496]
Daykin D.E., A lattice is distributive if and only if | A | | B |≤| A ⋁ B | | A ⋀ B |. Nanta Math. 10 (1977),
58–60. [596, 608]
Daykin D.E., J. Godfrey, and A.J.W. Hilton, Existence theorems for Sperner families. J. Combinatorial
Theory A 17 (1974), 245–251. [511]
Daykin D.E., A.J.W. Hilton, and D. Miklós, Pairings from down-sets and up-sets in distributive lattices.
J. Combin. Th. A 34 (1983), 215–230. [501, 512]
Daykin D.E., D.J. Kleitman, and D.B. West, The number of meets between two subsets of a lattice. J.
Combin. Th. A 26 (1979), 135–156. [596]
Daykin D.E. and L. Lovász, The number of values of a Boolean function. J. London Math. Soc. (2) 12
(1976), 225–230. [497, 607, 608]
de Abreu N.M.M., Old and new results on algebraic connectivity of graphs. Linear Algebra Appl. 423
(2007), 53–73. [778]
de Bruijn N.G., A combinatorial problem. Nederl. Akad. Wetensch., Proc. 49 (1946), 758–764; Indaga-
tiones Math. 8, 461–467 (1946). [239]
de Bruijn N.G. and P. Erdős, A colour problem for infinite graphs and a problem in the theory of relations.
Nederl. Akad. Wetensch. Proc. A. 54 = Indagationes Math. 13 (1951), 369–373. [468]
de Bruijn N.G. and T. van Aardenne-Ehrenfest, Circuits and trees in oriented linear graphs. Simon
Stevin 28 (1951), 203–217. [752]
de Bruijn N.G., E. van Tengbergen, and D. Kruyswijk, On the set of divisors of a number. Nieuw Arch.
Wisk. (2) 23 (1951), 191–193. [554]
de Fraysseix H., J. Pach, and R. Pollack, How to draw a planar graph on a grid. Combinatorica 10 (1990),
41–51. [783]
de Grey A., The chromatic number of the plane is at least 5. Geombinatorics 28 (2018), 18–31. (Also
arXiv:1804.02385) [342]
de Moivre A., The Doctrine of Chances (Pearson, 1718). [153]
de Moivre A., Miscellanea Analytica (London, 1730). [67]
de Werra D., Equitable colorations of graphs. Rev. Française Informat. Recherche Opérationnelle 5
(1971), 3–8. [372]
Dean A.M. and R.B. Richter, The crossing number of C4 × C4 . J. Graph Theory 19 (1995), 125–129.
[794]
868 References

Dean N., Contractible Edges and Conjectures about Path and Cycle Numbers (ProQuest, 1987). Ph.D.
Thesis, Vanderbilt University. [249]
Dean R.A. and G. Keller, Natural partial orders. Canad. J. Math. 20 (1968), 535–554. [587, 604]
DeBiasio L., A. Gyárfás, R.A. Krueger, M. Ruszinkó, and G.N. Sárközy, Monochromatic bal-
anced componenets, matchings, and paths in multicolored complete bipartite graphs (2018).
(arXiv:1804.04195). [460]
DeBiasio L. and A. Lo, Spanning trees with few branch vertices (2017). (arXiv:1709.04937). [327, 334]
Dekster B.V., The Borsuk conjecture holds for bodies of revolution. J. Geom. 52 (1995), 64–73. [817]
Delannoy H., Emploi de l’échiquier pour la résolution de divers problèmes de probabilité. Assoc. Franç.
Paris 18 (1889), 43–52. [28, 29]
Delcourt M. and L. Postle, On the list coloring version of Reed’s conjecture. Elec. Notes Discr. Math. 61
(2017), 343–349. [685]
Delsarte P., J.M. Goethals, and J.J. Seidel, Spherical codes and designs. Geometriae Dedicata 6 (1977),
363–388. [744]
Demoucron G., Y. Malgrange, and R. Pertuiset, Graphes planaires: reconnaissance et construction des
représentations planaires topologiques. Rev. Française Recherche Opérationnelle 8 (1964), 33–
47. [393]
Dénes J. and A.D. Keedwell, Latin squares and their applications (Academic Press, 1974). Also 2015
(Elsevier/North-Holland). [610, 623]
Dénes J. and A.D. Keedwell, Latin squares, Annals Discr. Math. 46 (North-Holland, 1991). [610]
Derbyshire J., Prime obsession (Joseph Henry Press, 2003). [10]
DeSario R., Problem 10931. Amer. Math. Monthly 109 (2002), 298. Solution 111 (2004), 169–70. [442]
Descartes B., A three colour problem. Eureka (1947). Solution 1948. [434]
Deshpande B. and M.N. Deshpande, Problem 11350. Amer. Math. Monthly 115 (2008), 262. Solution 117
(2010), 89. [668]
Deshpande B. and M.N. Deshpande, Problem 11500. Amer. Math. Monthly 117 (2010), 371. Solution 119
(2012), 348. [92]
Deshpande M.N. and K. Laghate, Problem 11042. Amer. Math. Monthly 110 (2003), 843–842. Solution
112 (2005), 574. [104]
Deshpande M.N. and R.M. Welukar, Problem 11033. Amer. Math. Monthly 110 (2003), 742. Solution 112
(2005), 470–472. [115]
DeTemple D. and J.M. Robertson, The equivalence of Euler’s and Pick’s theorems. Math. Teacher 67
(1974), 222–226. [389]
Deuber W., Partitionen und lineare Gleichungssysteme. Math. Z. 133 (1973), 109–123. [462]
Deutsch E., Problem 10649. Amer. Math. Monthly 105 (1998), 271. Solution 107 (2000), 279–280. [62]
Deutsch E., Problem 10751. Amer. Math. Monthly 106 (1999), 686. Solution 108 (2001), 872–873. [49]
Deutsch E., Problem 10795. Amer. Math. Monthly 107 (2000), 367. Solution 108 (2001), 980. [49]
Deutsch E., Problem 10877. Amer. Math. Monthly 108 (2001), 470. Solution 110 (2003), 245–246. [82]
Deutsch E., Problem 10902. Amer. Math. Monthly 108 (2001a), 871. Solution 110 (2003), 639–640.
[22, 62]
Deutsch E., Problem 11071. Amer. Math. Monthly 111 (2004a), 259. Solution 113 (2006), 460–1. [50]
Deutsch E., Problem 11108. Amer. Math. Monthly 111 (2004b), 725. Solution 113 (2006), 466–7. [59]
Deutsch E., Problem 11150. Amer. Math. Monthly 112 (2005), 367. Solution 114 (2007), 264–265.
[48, 164]
Deutsch E., Problem 11237. Amer. Math. Monthly 113 (2006), 655. Solution 115 (2008), 666–7. [149]
Deutsch E., Problem 11373. Amer. Math. Monthly 115 (2008), 568. Solution 117 (2010), 462. [105]
Deutsch E., Problem 11424. Amer. Math. Monthly 116 (2009), 277. Solution 118 (2011), 376. [21]
Deza M. and P. Frankl, Problem. In Combinatorics (1976), Coll. Math. Soc. J. Bolyai 18 (North-Holland,
1978), 1193. [511]
Deza M., P. Frankl, and N.M. Singhi, On functions of strength t. Combinatorica 3 (1983), 331–339.
[728]
References 869

Dias da Silva J.A. and Y.O. Hamidoune, Cyclic spaces for Grassman derivatives and additive theory.
Bull. London Math. Soc. 26 (1994), 140–146. [734, 746]
Dı́az-Barrero J.L., Problem 11164. Amer. Math. Monthly 112 (2005), 568. Solution 114 (2007), 364–365.
[32]
Dickson L.E., College Algebra (Wiley & Sons, 1902). [48]
Dickson L.E., Lower limit for the number of sets of solutions of x e + y e + e ≡ 0 (mod p). J. Reine Angew.
Math. 135 (1909), 181–188. [473]
Diestel R., Graph theory, Graduate Texts in Mathematics 173 (Springer-Verlag, 1997). Also 2000, 2006,
2010, 2016. [215, 377, 488]
Dijkstra E.W., A note on two problems in connexion with graphs. Numer. Math. 1 (1959), 269–271.
[246, 252]
Dilcher K., Some q-series identities related to divisor functions. Discr. Math. 145 (1995), 83–93. [32]
Dilworth R.P., A decomposition theorem for partially ordered sets. Ann. of Math. (2) 51 (1950), 161–166.
[546, 574, 607]
Dilworth R.P., Some combinatorial problems on partially ordered sets. In Proc. Sympos. Appl. Math. 10
(Amer. Math. Soc., 1960), 85–90. [606]
Dinitz J.H. and D.R. Stinson, A brief introduction to design theory. In Contemporary design theory,
Wiley-Intersci. Ser. Discr. Math. Optim. (Wiley, 1992), 1–12. [609]
Dirac G.A., On the Colouring of Graphs: Combinatorial Topology of Linear Complexes (Univ. London,
1951). Ph.D. Thesis. [349]
Dirac G.A., A property of 4-chromatic graphs and some remarks on critical graphs. J. London Math.
Soc. 27 (1952a), 85–92. [349]
Dirac G.A., Some theorems on abstract graphs. Proc. London Math. Soc. (3) 2 (1952b), 69–81.
[320, 323, 349]
Dirac G.A., The structure of  -chromatic graphs. Fund. Math. 40 (1953), 42–55. [345]
Dirac G.A., A theorem of R. L. Brooks and a conjecture of H. Hadwiger. Proc. London Math. Soc. (3) 7
(1957), 161–195. [355]
Dirac G.A., In abstrakten Graphen vorhandene vollständige 4-graphen und ihre Unterteilungen. Math.
Nachr. 22 (1960), 61–85. [301, 302]
Dirac G.A., On rigid circuit graphs. Abh. Math. Sem. Univ. Hamburg 25 (1961), 71–76. [367]
Dirac G.A., Extension of Turán’s theorem on graphs. Acta Math. Acad. Sci. Hungar. 14 (1963), 417–422.
[490]
Dirac G.A., Homomorphism theorems for graphs. Math. Ann. 153 (1964), 69–80. [351, 353]
Dirac G.A., Minimally 2-connected graphs. J. Reine Angew. Math. 228 (1967), 204–216. [315, 316]
Dirac G.A., B.A. Sørensen, and B. Toft, An extremal result for graphs with an application to their colour-
ings. J. Reine Angew. Math. 268/269 (1974), 216–221. [345]
Djidjev H.N., A linear algorithm for partitioning graphs. C. R. Acad. Bulgare Sci. 35 (1982), 1053–1056.
[397]
Djidjev H.N., On the constants of separator theorems. C. R. Acad. Bulgare Sci. 40 (1987), 31–34. [397]
Djidjev H.N. and S.M. Venkatesan, Reduced constants for simple cycle graph separation. Acta Inform.
34 (1997), 231–243. [397]
Djokovic D., Problem E3465. Amer. Math. Monthly 98 (1991), 852. Solution 100 (1993), 800. [32]
Dmitriev I.G., Weakly cyclic graphs with integral chromatic spectra. Metody Diskret. Analiz. (1980), 3–
7, 100. [374]
m
Dobiński G., Summirung der Reihe ∑ nn! für m = 1 , 2 , 3 , 4 , 5 , . . .. Grunert’s Archiv 61 (1877), 333–336.
[134]
Dodgson C.L., Condensation of determinants. Proc. Roy. Soc. A 15 (1866), 150–155. [170, 178]
Dol’nikov V.L., Transversals of families of sets. In Studies in the theory of functions of several real vari-
ables (Russian) (Yaroslav. Gos. Univ., 1981), 30–36, 109. [812]
Došlić T. and D. Rautenbach, Factor-critical graphs with the minimum number of near-perfect match-
ings. Discr. Math. 338 (2015), 2318–2319. [275]
Doster D., Problem E3391. Amer. Math. Monthly 97 (1990), 528. Solution 98 (1991), 860–861. [82]
870 References

Doster D., Problem 10403. Amer. Math. Monthly 101 (1994), 792. Solution 104 (1997), 368. [91]
Doubilet P., G.C. Rota, and R.P. Stanley, On the foundations of combinatorial theory. VI. The idea of
generating function. In Proceedings of the Sixth Berkeley Symposium on Mathematical Statistics
and Probability (Univ. California, Berkeley, Calif., 1970/1971), Vol. II: Probability theory (Univ.
California Press, 1972), 267–318. [127]
Dross F., Fractional triangle decompositions in graphs with large minimum degree. SIAM J. Discr. Math.
30 (2016), 36–42. [640]
Dubhashi D.P. and A. Panconesi, Concentration of measure for the analysis of randomized algorithms
(Cambridge Univ. Press, 2009). [706, 721]
Dudek A. and P. Prałat, An alternative proof of the linearity of the size-Ramsey number of paths. Combin.
Probab. Comput. 24 (2015), 551–555. [455]
Dudek A. and J.R. Schmitt, On the size and structure of graphs with a constant number of 1-factors.
Discr. Math. 312 (2012), 1807–1811. [276]
Duffus D., H. Lefmann, and V. Rödl, Shift graphs and lower bounds on Ramsey numbers r¾ (l; r). Discr.
Math. 137 (1995), 177–187. [601]
Dujmović V., G. Joret, J. Kozik, and D.R. Wood, Nonrepetitive colouring via entropy compression. Com-
binatorica 36 (2016), 661–686. [680]
Duraj L., G. Gutowski, and J. Kozik, Chip games and paintability. Electron. J. Combin. 23 (2016), Paper
3.3, 12. [742]
Dushnik B., Concerning a certain set of arrangements. Proc. Amer. Math. Soc. 1 (1950), 788–796.
[576, 577]
Dushnik B. and E.W. Miller, Partially ordered sets. Amer. J. Math. 63 (1941), 600–610. [569, 570, 831]
Dvir Z., On the size of Kakeya sets in finite fields (to appear). [746]
Dvořák Z., K.I. Kawarabayashi, and R. Thomas, Three-coloring triangle-free planar graphs in linear
time. ACM Trans. Algorithms 7 (2011), Art. 41, 14. [414]
Dvořák Z. and L. Postle, Correspondence coloring and its application to list-coloring planar graphs with-
out cycles of lengths 4 to 8. J. Combin. Th. B 129 (2018), 38–54. [356, 414]
Dvoretzky A. and T.S. Motzkin, A problem of arrangements. Duke Math. J. 14 (1947), 305–313. [42]
Dzhumadiĺdaeva A.A., Problem 11406. Amer. Math. Monthly 116 (2009), 82. Solution 117 (2010), 935.
[115]
Dziobek O., Eine Formel der Substitutionstheorie. Sitzungsber. Berl. Math. G. 17 (1917), 64–67. [46]
Eaton N., Ramsey numbers for sparse graphs. Discr. Math. 185 (1998), 63–75. [484]
Eaton N. and D.A. Grable, Set intersection representations for almost all graphs. J. Graph Theory 23
(1996), 309–320. [745]
Eckhoff J. and G. Wegner, Über einen Satz von Kruskal. Period. Math. Hungar. 6 (1975), 137–142.
[496]
Edelman P. and C. Greene, Balanced tableaux. Adv. in Math. 63 (1987), 42–99. [208]
Edmonds J., Existence of ¾ -edge connected ordinary graphs with prescribed degrees. J. Res. Nat. Bur.
Standards Sect. B 68B (1964), 73–74. [297]
Edmonds J., Minimum partition of a matroid into independent subsets. J. Res. Nat. Bur. Standards Sect.
B 69B (1965a), 67–72. [251, 520, 535]
Edmonds J., Lehman’s switching game and a theorem of Tutte and Nash-Williams. J. Res. Nat. Bur.
Standards Sect. B 69B (1965b), 73–77. [251, 520, 535]
Edmonds J., Paths, trees, and flowers. Canad. J. Math. 17 (1965c), 449–467. [264, 275, 278]
Edmonds J., Maximum matchings and a polyhedron with 0,1-vertices. J. Res. Nat. Bur. Standards 69B
(1965d), 125–130. [278, 282]
Edmonds J., Submodular functions, matroids and certain polyhedra. In Combinatorial Structures and
Their Applications (Calgary, 1969) (Gordon and Breach, 1970), 69–87. [529]
Edmonds J., Matroid intersection. In Discrete Optimization I (P.L. Hammer, E.L. Johnson, and B.H.
Korte, eds.), Ann. Discr. Math. 4 (1979), 39–49. [532]
Edmonds J. and D.R. Fulkerson, Transversals and matroid partition. J. Res. Nat. Bur. Standards Sect.
B 69B (1965), 147–153. [517, 533]
References 871

Edwards K., D.P. Sanders, P. Seymour, and R. Thomas, Three-edge-colouring doublecross cubic graphs.
J. Combin. Th. B 119 (2016), 66–95. [416]
Egawa Y., H. Matsuda, T. Yamashita, and K. Yoshimoto, On a spanning tree with specified leaves. Graphs
Combin. 24 (2008), 13–18. [327]
Eğecioğlu Ö. and J.B. Remmel, Bijections for Cayley trees, spanning trees, and their q-analogues. J.
Combin. Th. A 42 (1986), 15–30. [36]
Egerváry E., On combinatorial properties of matrices (Hungarian with German summary). Mat. Lapok
38 (1931), 16–28. [258, 530]
Eggleston H.G., Covering a three-dimensional set with sets of smaller diameter. J. London Math. Soc.
30 (1955), 11–24. [817]
Egorychev G.P., The solution of van der Waerden’s problem for permanents. Adv. in Math. 42 (1981),
299–305. [256, 765]
Eitner P.G., The bandwidth of the complete multipartite graph (1979). Presentation at Toledo Sympo-
sium on Applications of Graph Theory. [818]
Eliahou S. and M. Kervaire, Sumsets in vector spaces over finite fields. J. Number Theory 71 (1998), 12–
39. [745]
Elkies N., G. Kuperberg, M. Larsen, and J.G. Propp, Alternating-sign matrices and domino tilings. I. J.
Algebraic Combin. 1 (1992), 111–132. [34]
Ellingham M. and P. Salehi Nowbandegani, The Chvátal–Erdős condition for prism-Hamiltonicity
(2018). Unpublished preprint. [334]
Ellingham M.N., Spanning paths, cycles, trees and walks for graphs on surfaces. Congr. Numer. 115
(1996), 55–90. [326]
Ellingham M.N. and J.D. Horton, Non-Hamiltonian 3-connected cubic bipartite graphs. J. Combin. Th.
B 34 (1983), 350–353. [330]
Ellingham M.N. and X. Zha, Toughness, trees, and walks. J. Graph Th. 33 (2000), 125–137. [327]
Ellis D., Y. Filmus, and E. Friedgut, Triangle-intersecting families of graphs. J. Eur. Math. Soc. (JEMS)
14 (2012), 841–885. [509]
Enchev O., Problem 10390. Amer. Math. Monthly 101 (1994), 574. Solution 104 (1997), 367–368. [261]
Engel A., Problem-Solving Strategies (Springer, 1998). [226]
Engel K., Sperner theory, Encyclopedia of Mathematics and its Applications 65 (Cambridge Univ. Press,
1997). [493, 496, 511]
Enomoto H., B. Jackson, P. Katerinis, and A. Saito, Toughness and the existence of ¾ -factors. J. Graph
Th. 9 (1985), 87–95. [318]
Enomoto H., K. Ohba, K. Ota, and J. Sakamoto, Choice number of some complete multi-partite graphs.
Discr. Math. 244 (2002), 55–66. [355]
Enomoto H. and B. Péroche, The linear arboricity of some regular graphs. J. Graph Theory 8 (1984),
309–324. [412]
Entringer R.C., A short proof of Rubin’s block theorem. In Cycles in graphs (Burnaby, B.C., 1982), North-
Holland Math. Stud. 115 (North-Holland, 1985), 367–368. [347]
Entringer R.C. and E.F. Schmeichel, Edge conditions and cycle structure in bipartite graphs. Ars Com-
bin. 26 (1988), 229–232. [330]
Erdős P., Problem 3739. Amer. Math. Monthly 42 (1935), 396. Solution 44 (1937), 120. [426]
Erdős P., On a lemma of Littlewood and Offord. Bull. Amer. Math. Soc. 51 (1945), 898–902. [561]
Erdős P., On sets of distances of n points. Amer. Math. Monthly 53 (1946), 248–250. [793]
Erdős P., Some remarks on the theory of graphs. Bull. Amer. Math. Soc. 53 (1947), 292–294. [450, 658]
Erdős P., Some remarks on set theory. Proc. Amer. Math. Soc. 1 (1950), 127–141. [468]
Erdős P., Some theorems on graphs. Riveon Lematematika 9 (1955), 13–17. [478]
Erdős P., Graph theory and probability. Can. J. Math. 11 (1959), 34–38. [434, 671]
Erdős P., Graph theory and probability, II. Canad. J. Math. 13 (1961), 346–352. [341, 664, 682]
Erdős P., Remarks on a paper of Pósa. Magyar Tud. Akad. Mat. Kut. Int. Közl. 7 (1962a), 227–229.
[332]
872 References

Erdős P., On the number of complete subgraphs contained in certain graphs. Magy. Tud. Akad. Mat.
Kutató Int. Közl. 7 (1962b), 459–464. [476, 478]
Erdős P., On a combinatorial problem. Nord. Mat. Tidskr. 11 (1963), 5–10. [659, 703]
Erdős P., Extremal problems in graph theory. In Theory of Graphs and Its Applications (Academic Press,
1964), 29–36. [489, 659]
Erdős P., Extremal problems in number theory. In Proc. Sympos. Pure Math., Vol. VIII (AMS, 1965),
181–189. [669]
Erdős P., Problem 2. In Theory of Graphs (Tihany, 1966) (P. Erdős and G. Katona, eds.) (Academic Press,
1968). [354]
Erdős P., On the graph theorem of Turán. Mat. Lapok 21 (1970), 249–251 (1971). [223]
Erdős P., Problems and results on finite and infinite combinatorial analysis. In Infinite and finite sets
(Colloq., Keszthely, 1973), I, Colloq. Math. Soc. János Bolyai 10 (North-Holland, 1975), 403–424.
[504]
Erdős P., Problems and results in graph theory and combinatorial analysis. In Proceedings of the Fifth
British Combinatorial Conference (Aberdeen, 1975), Congr. Numer. 15 (Utilitas Math., 1976),
169–192. [438]
Erdős P., Problem E3284. Amer. Math. Monthly 95 (1988a), 762. Solution 97 (1990), 248–249. [490]
Erdős P., Problem E3255. Amer. Math. Monthly 95 (1988b), 259. Solution 97 (1990), 848–849. [490]
Erdős P., Some of my favourite unsolved problems. In A tribute to Paul Erdős (Cambridge Univ. Press,
1990), 467–478. [497]
Erdős P. and S. Fajtlowicz, On the conjecture of Hajós. Combinatorica 1 (1981), 141–143. [709]
Erdős P., R.J. Faudree, C.C. Rousseau, and R.H. Schelp, The size Ramsey number. Period. Math. Hungar.
9 (1978), 145–161. [455]
Erdős P., Z. Füredi, R.J. Gould, and D.S. Gunderson, Extremal graphs for intersecting triangles. J.
Combin. Th. B 64 (1995), 89–100. [491]
Erdős P. and T. Gallai, On maximal paths and circuits of graphs. Acta Math. Acad. Sci. Hung. 10 (1959),
337–356. [322, 334, 460]
Erdős P. and T. Gallai, Graphs with prescribed degrees of vertices (Hungarian). Mat. Lapok 11 (1960),
264–274. [269, 276]
Erdős P. and T. Gallai, On the minimal number of vertices representing the edges of a graph. Publ. Math.
Inst. Hung. Acad. Sci. 6 (1961), 181–203. [273, 354]
Erdős P., A. Ginzburg, and A. Ziv, Theorem in the additive number theory. Bull. Res. Council Israel 10
(1961). [764]
Erdős P., A.W. Goodman, and L. Pósa, The representation of graphs by set intersections. Canad. J. Math.
18 (1966), 106–112. [238]
Erdős P., R.L. Graham, and E. Szemeredi, On sparse graphs with dense long paths. In Computers and
mathematics with applications (Pergamon, 1976), 365–369. [394]
Erdős P. and R.K. Guy, Crossing number problems. Amer. Math. Monthly 80 (1973), 52–58. [793, 798]
Erdős P. and A. Gyárfás, A variant of the classical Ramsey problem. Combinatorica 17 (1997), 459–467.
[686]
Erdős P. and A. Hajnal, On chromatic numbers of graphs and set systems. Acta Math. Acad. Sci. Hung.
17 (1966), 61–99. [343]
Erdős P., A. Hajnal, and J.W. Moon, A problem in graph theory. Amer. Math. Monthly 71 (1964), 1107–
1110. [491]
Erdős P. and H. Heilbronn, On the addition of residue classes mod p. Acta Arith. 9 (1964), 149–159.
[734]
Erdős P., M. Herzog, and J. Schönheim, An extremal problem on the set of noncoprime divisors of a
number. Israel J. Math. 8 (1970), 408–412. [512]
Erdős P., C. Ko, and R. Rado, Intersection theorems for systems of finite sets. Quart. J. Math. Oxford (2)
12 (1961), 313–320. [498]
Erdős P. and J. Komlós, On a problem of Moser. In Combinatorial theory and its applications, I (Proc.
Colloq., Balatonfüred, 1969) (North-Holland, 1970), 365–367. [551]
References 873

Erdős P. and L. Lovász, Problems and results on 3-chromatic hypergraphs and some related questions.
In Infinite and finite sets (Colloq., Keszthely, 1973), II, Colloq. Math. Soc. János Bolyai 10 (North-
Holland, 1975), 609–627. [675, 679]
Erdős P. and J.W. Moon, On sets of consistent arcs in a tournament. Canad. Math. Bull. 8 (1965), 269–
271. [721]
Erdős P. and L. Pósa, On the maximal number of disjoint circuits of a graph. Publ. Math. Debrecen 9
(1962), 3–12. [273, 490]
Erdős P. and R. Rado, A combinatorial theorem. J. London Math. Soc. 25 (1950), 249–255. [470]
Erdős P. and R. Rado, Combinatorial theorems on classifications of subsets of a given set. Proc. London
Math. Soc. (3) 2 (1952), 417–439. [451, 608]
Erdős P. and R. Rado, A partition calculus in set theory. Bull. Amer. Math. Soc. 62 (1956), 427–489.
[456]
Erdős P. and R. Rado, Intersection theorems for systems of sets. J. London Math. Soc. 35 (1960), 85–90.
[503, 504]
Erdős P. and A. Rényi, On random graphs, I. Publ. Math. Debrecen 6 (1959), 290–297. [687, 698]
Erdős P. and A. Rényi, On a problem in the theory of graphs. Magyar Tud. Akad. Mat. Kutató Int. Közl.
7 (1962), 623–641. [252]
Erdős P. and A. Rényi, On two problems of information theory. Magyar Tud. Akad. Mat. Kutató Int. Közl.
8 (1963), 229–243. [513]
Erdős P. and A. Rényi, On the existence of a factor of degree one of a connected random graph. Acta
Math. Acad. Sci. Hung. 17 (1966), 359–368. [699]
Erdős P., A. Rényi, and V.T. Sós, On a problem of graph theory. Studia Sci. Math. Hungar. 1 (1966), 215–
235. [244, 632]
Erdős P., A.L. Rubin, and H. Taylor, Choosability in graphs. In Proc. West Coast Conf. Combin. Graph Th.
Comput. (Arcata), Congr. Numer. 26 (Utilitas Math., 1979), 125–157. [346, 348, 356, 660, 667]
Erdős P. and H. Sachs, Reguläre Graphen gegebener Taillenweite mit minimaler Knotenzahl. Wiss. Z.
Martin-Luther-Univ. Halle-Wittenberg Math.-Natur. Reihe 12 (1963), 251–257. [219, 251, 274]
Erdős P. and S. Shelah, Separability properties of almost-disjoint families of sets. Israel J. Math. 12
(1972), 207–214. [551]
Erdős P. and M. Simonovits, A limit theorem in graph theory. Studia Sci. Math. Hungar. 1 (1966), 51–
57. [479]
Erdős P., M. Simonovits, and V.T. Sós, Anti-Ramsey theorems 10 (1975), 633–643. Colloq. Math. Soc.
János Bolyai. [457]
Erdős P. and J. Spencer, Probabilistic methods in combinatorics (Academic Press, 1974). [702]
Erdős P. and A.H. Stone, On the structure of linear graphs. Bull. Amer. Math. Soc. 52 (1946), 1087–1091.
[479, 491]
Erdős P. and G. Szekeres, A combinatorial problem in geometry. Composito Math 2 (1935), 464–470.
[344, 430, 446]
Erdős P. and G. Szekeres, On some extremum problems in elementary geometry. Ann. Univ. Sci. Bu-
dapest. Eötvös Sect. Math. 3–4 (1960/1961), 53–62. [446, 458]
Erdős P. and E. Szemerédi, Combinatorial properties of systems of sets. J. Combin. Th. A 24 (1978),
308–313. [504]
Erdős P. and P. Turán, On Some Sequences of Integers. J. London Math. Soc. S1-11 (1936), 261. [484]
Erdős P. and D.B. West, A note on the interval number of a graph. Discr. Math. 55 (1985), 129–133.
[705]
Erdős P. and R.J. Wilson, On the chromatic index of almost all graphs. J. Combin. Th. B 23 (1977), 255–
257. [699]
Erdős P.L. and L.A. Székely, Applications of antilexicographic order. I. An enumerative theory of trees.
Adv. in Appl. Math. 10 (1989), 488–496. [47, 135]
Eršov A.P. and G.I. Kožuhin, Estimates of the chromatic number of connected graphs (Russian). Dokl.
Akad. Nauk. SSSR 142 (1962), 270–273. [353]
874 References

Esperet L., F. Kardoš, A.D. King, D. Král’, and S. Norine, Exponentially many perfect matchings in cubic
graphs. Adv. Math. 227 (2011), 1646–1664. [765]
Esperet L. and A. Parreau, Acyclic edge-coloring using entropy compression. European J. Combin. 34
(2013), 1019–1027. [680]
Etienne G., Tableaux de Young et solitaire bulgare. J. Combin. Th. A 58 (1991), 181–197. [144, 152]
Euler L., Solutio problematis ad geometriam situs pertinentis (Latin). Comment. Acad. Sci. U. Petrop 8
(1736), 128–140. [234]
Euler L., Introductio in analysin infinitorum (Latin) (Culture et civilisation (Bruxelles), 1748). Reprinted
in 2000 (Sociedad Andaluza de Educación Matemática “Thales”). [141, 148, 151]
Euler L., Demonstratio nonnullarum insignium proprietatum, quibus solida hedris planis inclusa sunt
praedita (Latin). Novi Commentarii acad. sci. Petropolitanae 4 (1758), 140–160. [383]
Evans A.B., G. Isaak, and D.A. Narayan, Representations of graphs modulo n. Discr. Math. 223 (2000),
109–123. [624]
Everman D., A.E. Danese, and K. Venkannayah, Problem e1396. Amer. Math. Monthly 67 (1960), 63–64.
Solution 67 (1960), 694. [61]
Faà di Bruno F., Sullo sviluppo delle funzioni (Italian). Ann. di Sci. Math. e Fisiche 6 (1855), 479–480.
[127]
Faà di Bruno F., Note sur une nouvelle formule de dalcul differentiel (French). Quart. J. Pure and Appl.
Math. 1 (1857), 359–360. [127]
Faigle U., Matroids in combinatorial optimization. In Combinatorial Geometries (N. White, ed.) (Cam-
bridge Univ. Press, 1987), 161–210. [532]
Falgas-Ravry V. and E.R. Vaughan, Applications of the semi-definite method to the Turán density prob-
lem for 3-graphs. Combin. Probab. Comput. 22 (2013), 21–54. [475]
Falikman D.I., Proof of the van der Waerden conjecture on the permanent of a doubly stochastic matrix.
Mat. Zametki 29 (1981), 931–938, 957. [256, 765]
Fan G., New sufficient conditions for cycles in graphs. J. Combin. Th. B 37 (1984), 221–227. [324–326]
Farber M. and R.E. Jamison, Convexity in graphs and hypergraphs. SIAM J. Algeb. Disc. Meth. 7 (1986),
433–444. [374]
Faria L., C.M.H. de Figueiredo, O. Sýkora, and I. Vrt’o, An improved upper bound on the crossing number
of the hypercube. J. Graph Theory 59 (2008), 145–161. [798]
Fáry I., On the straight line representations of planar graphs. Acta Sci. Math. 11 (1948), 229–233.
[393]
Farzad B., M. Mahdian, E.S. Mahmoodian, A. Saberi, and B. Sadri, Forced orientation of graphs. Bull.
Iranian Math. Soc. 32 (2006), 79–89, 98. [237]
Fasenmyer M.C., Some Generalized Hypergeometric Polynomials (ProQuest, 1946). Ph.D. Thesis, Univ.
Michigan. [90]
Faudree R.J., Forbidden subgraphs and Hamiltonian properties: a survey. Congr. Numer. 116 (1996),
33–52. [316]
Faudree R.J., Survey of results on ¾ -ordered graphs. Discr. Math. 229 (2001), 73–87. [316]
Faudree R.J., R.J. Gould, M.S. Jacobson, and L.M. Lesniak, Neighborhood unions and a generalization
of Dirac’s theorem. Discr. Math. 105 (1992), 61–71. [327]
Faudree R.J., R.J. Gould, M.S. Jacobson, and R.H. Schelp, Neighborhood unions and Hamiltonian prop-
erties in graphs. J. Combin. Th. B 47 (1989), 1–9. [327]
Faudree R.J., A. Gyárfás, R.H. Schelp, and Z. Tuza, Induced matchings in bipartite graphs. Discr. Math.
78 (1989), 83–87. [262]
Faudree R.J. and R.H. Schelp, Various length paths in graphs. In Theory and applications of graphs
(Proc. Internat. Conf., Western Mich. Univ., Kalamazoo, Mich., 1976), Lecture Notes in Math. 642
(Springer, 1978), 160–173. [454]
Faudree R.J. and R.H. Schelp, A survey of results on the size Ramsey number. In Paul Erdős and his
mathematics, II (Budapest, 1999), Bolyai Soc. Math. Stud. 11 (János Bolyai Math. Soc., 2002),
291–309. [455]
References 875

Favaron O., F. Genest, and M. Kouider, Regular path decompositions of odd regular graphs. J. Graph
Theory 63 (2010), 114–128. [273]
Feige U., M.M. Halldórsson, G. Kortsarz, and A. Srinivasan, Approximating the domatic number. SIAM
J. Comput. 32 (2002), 172–195. [684]
Feller W., An introduction to probability theory and its applications, I, Third edition (Wiley & Sons,
1968). [40, 668]
Felsner S., Convex drawings of planar graphs and the order dimension of 3-polytopes. Order 18 (2001),
19–37. [784]
Felsner S., Geodesic embeddings and planar graphs. Order 20 (2003), 135–150. [790]
Felsner S., Geometric graphs and arrangements, Advanced Lectures in Mathematics (Friedr. Vieweg &
Sohn, 2004). [783]
Felsner S., The order dimension of planar maps revisited. SIAM J. Discr. Math. 28 (2014), 1093–1101.
[790]
Felsner S., P.C. Fishburn, and W.T. Trotter, Finite three-dimensional partial orders which are not sphere
orders. Discr. Math. 201 (1999), 101–132. [828]
Felsner S., C.M. Li, and W.T. Trotter, Adjacency posets of planar graphs. Discr. Math. 310 (2010), 1097–
1104. [790]
Felsner S. and W.T. Trotter, Balancing pairs in partially ordered sets. In Combinatorics, Paul Erdős is
eighty, Vol. 1, Bolyai Soc. Math. Stud. (János Bolyai Math. Soc., 1993), 145–157. [823]
Feng W., Bounds on maximum b-matchings. Discr. Math. 309 (2009), 4162–4165. [286]
Ferguson T. and C. Melolidakis, Problem E3061. Amer. Math. Monthly 91 (1984), 580. Solution 94 (1987),
189. [669]
Fiedler M., Algebraic connectivity of graphs. Czech. Math. J. 23 (1973), 298–305. [778]
Fielder D.C., Fibonacci numbers in tree counts for sector and related graphs. Fibonacci Quart. 12 (1974),
355–359. [65]
Figaj A. and T. Łuczak, The Ramsey number for a triple of long even cycles. J. Combin. Th. B 97 (2007),
584–596. [492]
Finck H.-J., On the chromatic numbers of a graph and its complement. In Theory of Graphs (Tihany,
1966) (P. Erdős and G. Katona, eds.) (Academic Press, 1968), 99–113. [344]
Finck H.-J. and H. Sachs, Über eine von H. S. Wilf angegebene Schranke für die chromatische Zahl
endlicher Graphen. Math. Nachr. 39 (1969), 373–386. [338]
Fink J., Perfect matchings extend to Hamilton cycles in hypercubes. J. Combin. Th. B 97 (2007), 1074–
1076. [330]
Fink J., Matching graphs of hypercubes and complete bipartite graphs. European J. Combin. 30 (2009),
1624–1629. [330]
Fiorini S., Some remarks on a paper by Vizing on critical graphs (“Critical graphs with a given chro-
matic class” (Russian), Diskret. Analiz. No. 5 (1965), 9-17). Math. Proc. Cambridge Philos. Soc.
77 (1975), 475–483. [414]
Fiorini S. and R.J. Wilson, Edge-colourings of graphs (Pitman; distrib. Fearon-Pitman, 1977). [361]
Firke F.A., P.M. Kosek, E.D. Nash, and J. Williford, Extremal graphs without 4-cycles. J. Combin. Th.
B 103 (2013), 327–336. [630]
Fishburn P. and B. Monjardet, Norbert Wiener on the theory of measurement (1914, 1915, 1921). J.
Math. Psych. 36 (1992), 165–184. [586]
Fishburn P.C., Intransitive indifference with unequal indifference intervals. J. Mathematical Psychology
7 (1970), 144–149. [586]
Fishburn P.C., Interval lengths for interval orders: a minimization problem. Discr. Math. 47 (1983), 63–
82. [604]
Fishburn P.C., A correlational inequality for linear extensions of a poset. Order 1 (1984a), 127–137.
[600]
Fishburn P.C., Numbers of lengths for representations of interval orders. In Progress in combinatorial
optimization (Waterloo, 1982) (Academic Press, 1984b), 131–146. [587]
876 References

Fishburn P.C., Interval orders and interval graphs, Wiley-Interscience Series in Discr. Mathematics (Wi-
ley & Sons, 1985). [587]
Fishburn P.C. and R.L. Graham, Classes of interval graphs under expanding length restrictions. J.
Graph Theory 9 (1985), 459–472. [604]
Fishburn P.C. and W.T. Trotter, Angle orders. Order 1 (1985), 333–343. [828, 832]
Fisher D.C., K.L. Collins, and L.B. Krompart, Problem 10406. Amer. Math. Monthly 101 (1994), 793.
Solution 104 (1997), 572–573. [418]
Fisher D.C., K. Fraughnaugh, L. Langley, and D.B. West, The number of dependent arcs in an acyclic
orientation. J. Combin. Th. B 71 (1997), 73–78. [550]
Fisher R.A., An examination of the different possible solutions of a problem in incomplete blocks. Ann.
Eugenics 10 (1940), 52–75. [614]
Fisk S., A short proof of Chvátal’s watchman theorem. J. Combin. Th. B 24 (1978), 374. [419]
Fiz Pontiveros G., S. Griffiths, and R. Morris, The triangle-free process and r(3 , ¾) (2013).
(arXiv:1302.6279). [682]
Fleischner H., The square of every two-connected graph is hamiltonian. J. Combin. Th. B 16 (1974), 29–
34. [331]
Fleischner H., Eulerian graphs. In Selected Topics in Graph Theory, 2 (L.W. Beineke and R.J. Wilson,
eds.) (Academic Press, 1983), 17–54. [763]
Fleischner H. and M. Stiebitz, A solution to a colouring problem of P. Erdős. Discr. Math. 101 (1992),
39–48. [739]
Floyd R.W., Problem E3399. Amer. Math. Monthly 97 (1990), 611–612. [257]
Foata D. and J. Riordan, Mappings of acyclic and parking functions. Aequationes Math. 10 (1974), 10–
22. [50]
Foata D. and M.P. Schützenberger, Théorie géométrique des polynômes eulériens, Lecture Notes in Math-
ematics, Vol. 138 (Springer-Verlag, 1970). [99, 101, 106, 127, 177]
Folkman J., Graphs with monochromatic complete subgraphs in every edge coloring. SIAM J. Appl.
Math. 18 (1970), 19–24. [455, 466]
Fon-Der-Flaass D.G., New prolific constructions of strongly regular graphs. Adv. Geom. 2 (2002), 301–
306. [649, 782]
Ford G.W. and G.E. Uhlenbeck, Combinatorial problems in the theory of graphs. I and III. Proc. Nat.
Acad. Sci. U.S.A. 42 (1956), 122–128, 529–535. [135]
Ford K., Problem 10383. Amer. Math. Monthly 101 (1994), 473. Solution 104 (1997), 457. [669]
Ford L.R., Jr. and D.R. Fulkerson, Maximal flow through a network. Canad. J. Math. 8 (1956), 399–404.
[300]
Ford L.R., Jr. and D.R. Fulkerson, Network flows and systems of representatives. Canad. J. Math. 10
(1958), 78–85. [302, 531]
Fortuin C.M., P.W. Kasteleyn, and J. Ginibre, Correlation inequalities on some partially ordered sets.
Comm. Math. Phys. 22 (1971), 89–103. [596]
Foulkes H.O., Enumeration of permutations with prescribed up-down and inversion sequences. Discr.
Math. 15 (1976), 235–252. [197]
Fouquet J.L. and J.-L. Jolivet, In “problémes”. In Problémes combinatoires et théorie des graphes (Orsay,
1976), Colloq. Internat. CNRS, 260 (1978), 437–443. [334]
Fournier J.-C., Colorations des arêtes d’un graphe. In Colloque Théorie des Graphes (Bruxelles, 1973),
Cahiers Ctr. Étud. Rech. Opér 15 (1973), 311–314. [372]
Fox J., A new proof of the graph removal lemma. Ann. of Math. (2) 174 (2011), 561–579. [485]
Fox J., C. Lee, and B. Sudakov, Maximum union-free subfamilies. Israel J. Math. 191 (2012), 959–971.
[551]
Fox J., J. Pach, B. Sudakov, and A. Suk, Erdős–Szekeres-type theorems for monotone paths and convex
bodies. Proc. Lond. Math. Soc. (3) 105 (2012), 953–982. [457, 601]
Fox J. and B. Sudakov, Density theorems for bipartite graphs and related Ramsey-type results. Combi-
natorica 29 (2009), 153–196. [673, 684]
References 877

Fox J. and B. Sudakov, Dependent random choice. Random Structures Algorithms 38 (2011), 68–99.
[672, 673, 684]
Frame J.S., G.de B. Robinson, and R.M. Thrall, The hook graphs of the symmetric groups. Canadian J.
Math. 6 (1954), 316–324. [190]
Frank A., On the orientation of graphs. J. Combin. Th. B 28 (1980a), 251–261. [310]
Frank A., On chain and antichain families of a partially ordered set. J. Combin. Th. B 29 (1980b), 176–
184. [548]
Frank A., Applications of submodular functions. In Surveys in Combinatorics, 1993 (K. Walker, ed.),
Lond. Math. Soc. Lect. Notes 187 (Cambridge Univ. Press, 1993), 85–136. [307, 310]
Frank A. and E. Tardos, Matroids from crossing families. In Finite and infinite sets (Eger, 1981), I–II,
Colloq. Math. Soc. János Bolyai 37 (North-Holland, 1984), 295–304. [310]
Frankl P., Families of finite sets satisfying an intersection condition. Bull. Austral. Math. Soc. 15 (1976),
73–79. [500]
Frankl P., A constructive lower bound for some Ramsey numbers. Ars Combinatoria 3 (1977), 297–302.
[728]
Frankl P., The Erdős–Ko–Rado theorem is true for n = c¾ t. In Combinatorics (Proc. 5th Hung. Colloq.,
Keszthely, 1976), I, Colloq. Math. Soc. János Bolyai 18 (North-Holland, 1978), 365–375.
[500, 511]
Frankl P., A new short proof for the Kruskal–Katona theorem. Discr. Math. 48 (1984), 327–329.
[495, 496, 510]
Frankl P., The shifting technique in extremal set theory. In Surveys in combinatorics 1987 (New Cross,
1987), London Math. Soc. Lecture Note Ser. 123 (Cambridge Univ. Press, 1987), 81–110.
[493, 499]
Frankl P., Constructive Ramsey bounds and intersection theorems for sets. In Mathematics of Ramsey
theory, Algorithms Combin. 5 (Springer, 1990), 53–56. [450]
Frankl P., Extremal set systems. In Handbook of combinatorics, Vol. 1, 2 (Elsevier Sci. B. V., Amsterdam,
1995), 1293–1329. [497]
Frankl P. and M. Deza, On the maximum number of permutations with given maximal or minimal
distance. J. Combin. Th. A 22 (1977), 352–360. [510]
Frankl P. and Z. Füredi, Families of finite sets with missing intersections. In Finite and infinite sets
(Eger, 1981), I–II, Colloq. Math. Soc. János Bolyai 37 (North-Holland, 1984), 305–318. [730]
Frankl P. and Z. Füredi, A new short proof of the EKR theorem. J. Combin. Th. A 119 (2012), 1388–
1390. [500]
Frankl P. and R.L. Graham, Old and new proofs of the Erdős–Ko–Rado theorem. Sichuan Daxue Xuebao
26 (1989), 112–122. [500]
Frankl P. and V. Rödl, Extremal problems on set systems. Random Structures Algorithms 20 (2002),
131–164. [488]
Frankl P. and N. Tokushige, Invitation to intersection problems for finite sets. J. Combin. Th. A 144
(2016), 157–211. [493]
Frankl P. and R.M. Wilson, Intersection theorems with geometric consequences. Combinatorica 1 (1981),
357–368. [459, 727–729]
Franklin F., On Newton’s Method of Approximation. Amer. J. Math. 4 (1881), 275–276. [151]
Franklin P., The Four Color Problem. Amer. J. Math. 44 (1922), 225–236. [406]
Franzblau D.S. and D. Zeilberger, A bijective proof of the hook-length formula. J. Algorithms 3 (1982),
317–343. [192]
Fraughnaugh (Jones) K., Minimum independence graphs with maximum degree four. In Graphs and
applications (Boulder, 1982), Wiley-Intersci. Publ. (Wiley, 1985), 221–230. [420]
Fredman M.L., How good is the information theory bound in sorting? Theoret. Comput. Sci. 1 (1976),
355–361. [823]
Fredman M.L. and D.E. Knuth, Recurrence relations based on minimization. J. Math. Anal. Appl. 48
(1974), 534–559. [92]
878 References

Freund R.M. and M.J. Todd, A constructive proof of Tucker’s combinatorial lemma. J. Combin. Th. A 30
(1981), 321–325. [807, 809]
Friedgut E. and J. Kahn, On the number of copies of one hypergraph in another. Israel J. Math. 105
(1998), 251–256. [510]
Friedgut E., J. Kahn, G. Kalai, and N. Keller, Chvátal’s conjecture and correlation inequalities. J. Com-
bin. Th. A 156 (2018), 22–43. [503]
Frieze A., On the length of the longest monotone subsequence in a random permutation. Ann. Appl.
Probab. 1 (1991), 301–305. [722]
Frieze A. and M. Karoński, Introduction to Random Graphs (Cambridge, 2016). [657]
Fritsch R. and G. Fritsch, The Four-Color Theorem (Springer, 1998). (published in German by F.A. Brock-
haus, 1994). [399]
Frobenius G., Ueber die Congruenz nach einem aus zwei endlichen Gruppen gebildeten Doppelmodul.
J. Reine Angew. Math. 101 (1887), 273–299. [181]
Frobenius G., über die Charaktere der symmetriscen Gruppe (German). Sitzungberichte der Königlich
Preussischen Akademie der Wissenschaften zu Berlin (1900), 516–534. [189, 190, 207]
Frobenius G., Über zerlegbare Determinanten. Sitzungsber. König. Preuss. Adad. Wiss. XVIII (1917),
274–277. [256]
Fronček D., Locally linear graphs. Math. Slovaca 39 (1989), 3–6. [684]
Fujita S., C. Magnant, and K. Ozeki, Rainbow generalizations of Ramsey theory: a survey. Graphs Com-
bin. 26 (2010), 1–30. [457]
Fulkerson D.R., Note on Dilworth’s decomposition theorem for partially ordered sets. Proc. Amer. Math.
Soc. 7 (1956), 701–702. [547]
Fulkerson D.R. and O.A. Gross, Incidence matrices and interval graphs. Pac. J. Math. 15 (1965), 835–
855. [370]
Fulkerson D.R., A.J. Hoffman, and M.H. McAndrew, Some properties of graphs with multiple edges.
Canad. J. Math. 17 (1965), 166–177. [222]
Fulton W., Young tableaux, London Mathematical Society Student Texts 35 (Cambridge Univ. Press,
1997). [189]
Füredi Z., Graphs without quadrilaterals. J. Combin. Th. B 34 (1983), 187–190. [630]
Füredi Z., Turán type problems. In Surveys in combinatorics, 1991 (Guildford, 1991), London Math. Soc.
Lecture Note Ser. 166 (Cambridge Univ. Press, 1991), 253–300. [475, 672]
Füredi Z., The order dimension of two levels of the Boolean lattice. Order 11 (1994), 15–28.
[582, 584, 819]
Füredi Z., On the number of edges of quadrilateral-free graphs. J. Combin. Th. B 68 (1996a), 1–6.
[630]
Füredi Z., New asymptotics for bipartite Turán numbers. J. Combin. Th. A 75 (1996b), 141–144. [633]
Füredi Z., An upper bound on Zarankiewicz’ problem. Combin. Probab. Comput. 5 (1996c), 29–33.
[632, 639]
Füredi Z., A proof on the stability of extremal graphs, Workshop on Extremal Graphs and Hypergraphs
(Carnegie-Mellon) (2007). [491]
Füredi Z., A proof of the stability of extremal graphs, Simonovits’ stability from Szemerédi’s regularity.
J. Combin. Th. B 115 (2015), 66–71. [228]
Füredi Z., J.R. Griggs, and D.J. Kleitman, A minimal cutset of the Boolean lattice with almost all mem-
bers. Graphs Combin. 5 (1989), 327–332. [550]
Füredi Z. and D.S. Gunderson, Extremal numbers for odd cycles. Combin. Probab. Comput. 24 (2015),
641–645. [479]
Füredi Z. and P. Hajnal, Davenport–Schinzel theory of matrices. Discr. Math. 103 (1992), 233–251.
[432, 442]
Füredi Z., P. Hajnal, V. Rödl, and W.T. Trotter, Interval orders and shift graphs. In Sets, graphs and
numbers (Budapest, 1991), Colloq. Math. Soc. János Bolyai 60 (North-Holland, 1992), 297–313.
[581, 604]
References 879

Füredi Z. and J. Kahn, On the dimensions of ordered sets of bounded degree. Order 3 (1986), 15–20.
[576, 578]
Füredi Z. and G.O.H. Katona, 2-Bases of quadruples. Combin. Probab. Comput. 15 (2006), 131–141.
[490]
Füredi Z., A.V. Kostochka, R. Škrekovski, M. Stiebitz, and D.B. West, Nordhaus–Gaddum-type theorems
for decompositions into many parts. J. Graph Theory 50 (2005), 273–292. [239, 639]
Füredi Z. and M. Simonovits, The history of degenerate (bipartite) extremal graph problems. In Erdős
centennial, Bolyai Soc. Math. Stud. 25 (János Bolyai Math. Soc., 2013), 169–264. [479]
Fuss N., Solutio quaestionis, quot modis polygonum n laterum in polygona m laterum, per diagonales
resolvi queat. Nova Acta Academiae Sci. Petropolitanae 9 (1791), 243–251. [42]
Gabber O. and Z. Galil, Explicit construction of linear-sized superconcentrators. J. Comput. Systems Sci.
22 (1981), 407–420. [779]
Gabow H.N., An efficient implementation of Edmonds’ algorithm for maximum matchings on graphs. J.
Assoc. Comput. Mach. 23 (1975), 221–234. [282]
Gabow H.N., Data structures for weighted matching and nearest common ancestors with linking. In
Proc. 1st ACM-SIAM Symp. Disc. Algs (San Francisco) (SIAM, 1990), 434–443. [282]
Gabow H.N., Z. Galil, T. Spencer, and R.E. Tarjan, Efficient algorithms for finding minimum spanning
trees in undirected and directed graphs. Combinatorica 6 (1986), 109–122. [246]
Gabow H.N. and R.E. Tarjan, Faster scaling algorithms for general graph matching problems. Tech.
Rep. CU-CS-432-89, Dept. Comp. Sci., Univ. Colorado–Boulder (1989). [282]
Gale D., Neighboring vertices on a convex polyhedron. In Linear inequalities and related system, Annals
of Mathematics Studies, no. 38 (Princeton Univ. Press, 1956), 255–263. [813, 820]
Gale D., The game of Hex and the Brouwer fixed-point theorem. Amer. Math. Monthly 86 (1979), 818–
827. [803, 804]
Gale D. and L.S. Shapley, College admissions and the stability of marriage. Amer. Math. Monthly 69
(1962), 9–15. [285, 288]
Gallai T., On factorisation of graphs. Acta Math. Acad. Sci. Hungar. 1 (1950), 133–153. [264]
Gallai T., Über extreme Punkt- und Kantenmengen. Ann. Univ. Sci. Budapest, Eötvös Sect. Math. 2
(1959), 133–138. [260, 263]
Gallai T., Kritische Graphen I. Magyar Tud. Akad. Mat. Kut. Int. Közl. 8 (1963a), 165–192. [355]
Gallai T., Neuer Beweis eines Tutte’schen Satzes. Magyar Tud. Akad. Mat. Kut. Int. Közl. 8 (1963b),
135–139. [275]
Gallai T., Kritische Graphen II. Magyar Tud. Akad. Mat. Kut. Int. Közl. 8 (1963c), 373–395. [357]
Gallai T., On directed paths and circuits. In Theory of Graphs (Tihany, 1966) (P. Erdős and G. Katona,
eds.) (Academic Press, 1968), 115–118. [339]
Gallai T. and A.N. Milgram, Verallgemeinerung eines graphentheoretischen Satzes von Rédei. Acta Sci.
Math. Szeged 21 (1960), 181–186. [548]
Gallian J.A., A dynamic survey of graph labeling. Electron. J. Combin. 5 (1998), Dynamic Survey 6, 43.
[646]
Gallian J.A. and D.J. Rusin, Cyclotomic polynomials and nonstandard dice. Discr. Math. 27 (1979), 245–
259. [103]
Galperin G. and H. Gauchman, Problem 11103. Amer. Math. Monthly 111 (2004), 724. Solution 113
(2006), 465–466. [61]
Galvin D., Three tutorial lectures on entropy and counting (2014). (arXiv:1406.7872). [504, 506, 509]
Galvin F., Problem 6034. Amer. Math. Monthly 82 (1975), 592. Solution 84 (1977), 224. [474]
Galvin F., The list chromatic index of a bipartite multigraph. J. Combin. Th. B 63 (1995), 153–158.
[363, 364]
Galvin F., Problem 10701. Amer. Math. Monthly 105 (1998), 956. Solution 108 (2001), 79–80. [355]
Galvin F., Problem 10761. Amer. Math. Monthly 106 (1999), 864. Solution 108 (2001), 773–774. [228]
Ganter B. and L. Teirlinck, A combinatorial lemma. Math. Z. 154 (1977), 153–156. [441]
Gardner M., Mathematical games (a.k.a. Bulgarian Solitaire and other seemingly endless tasks). Sci.
Amer. 249 (1983), 8–13. [143]
880 References

Garey M.R., R.L. Graham, D.S. Johnson, and D.E. Knuth, Complexity results for bandwidth minimiza-
tion. SIAM J. Appl. Math. 34 (1978), 477–495. [800]
Garey M.R. and D.S. Johnson, The complexity of near-optimal graph coloring. J. Assoc. Comp. Mach. 23
(1976), 43–49. [701]
Garey M.R. and D.S. Johnson, Computers and intractability (W. H. Freeman and Co., 1979). A guide to
the theory of NP-completeness, A Series of Books in the Mathematical Sciences. [11]
Garey M.R. and D.S. Johnson, Crossing number is NP-complete. SIAM J. Algebraic Discrete Methods 4
(1983), 312–316. [795]
Gargano L., M. Hammar, P. Hell, L. Stacho, and U. Vaccaro, Spanning spiders and light-splitting
switches. Discr. Math. 285 (2004), 83–95. [327, 334]
Gaskell R.W., M.S. Klamkin, and P. Watson, Triangulations and Pick’s theorem. Math. Mag. 49 (1976),
35–37. [389]
Gasparyan G.S., Minimal imperfect graphs: a simple approach. Combinatorica 16 (1996), 209–212.
[369]
Gavril F., The intersection graphs of subtrees in trees are exactly the chordal graphs. J. Combin. Th. B
16 (1974), 47–56. [373]
Gazit H. and G.L. Miller, Planar separators and the Euclidean norm. In Algorithms (Tokyo, 1990), Lec-
ture Notes in Comput. Sci. 450 (Springer, Berlin, 1990), 338–347. [397]
GCHQ Problem Solving Group, Solution to problem 11883. Amer. Math. Monthly 125 (2018), 82–83.
Proposed 123 (2016), 97. [149]
Geetha J., N. Narayanan, and K. Somasundaram, Total colorings-A survey (2018). (arXiv:1812.05833v1).
[373]
George J.C., 1-Factorizations of tensor products of graphs (ProQuest, 1991). Ph.D. Thesis, Univ. Illinois
at Urbana–Champaign. [273, 371]
Georges J.P., Non-Hamiltonian bicubic graphs. J. Combin. Th. B 46 (1989), 121–124. [317]
Gerbner D. and B. Patkós, Extremal Finite Set Theory (CRC Press, 2019). [493]
Gerencsér L. and A. Gyárfás, On Ramsey-type problems. Ann. Univ. Sci. Budapest. Eötvös Sect. Math.
10 (1967), 167–170. [460]
Gerke S. and A. Steger, The sparse regularity lemma and its applications. In Surveys in combinatorics
2005, London Math. Soc. Lecture Note Ser. 327 (Cambridge Univ. Press, 2005), 227–258. [488]
Gessel I.M., Generating Functions and Enumeration of Sequences (ProQuest LLC, 1977). Ph.D. Thesis,
Massachusetts Inst. Tech. [102]
Gessel I.M., Problem 10357. Amer. Math. Monthly 101 (1994), 75. Solution 104 (1997), 177–178. [115]
Gessel I.M., Problem 10424. Amer. Math. Monthly 102 (1995), 70. Solution 104 (1997), 466–467. [117]
Gessel I.M., The Smith College diploma problem. Amer. Math. Monthly 108 (2001), 55–57. [106]
Gessel I.M., Lagrange inversion. J. Combin. Th. A 144 (2016), 212–249. [129]
Gessel I.M. and R.P. Stanley, Stirling polynomials. J. Combin. Th. A 24 (1978), 24–33. [102]
Gessel I.M. and G. Viennot, Binomial determinants, paths, and hook length formulae. Adv. in Math. 58
(1985), 300–321. [165, 167, 168]
Gessel I.M. and G. Viennot, Determinants, paths, and plane partitions (1989). Available online at
http://contscience.xavierviennot.org/xavier/articles_files/determinant_89.pdf
[169]
Getz M. and D. Jones, Problem 11005. Amer. Math. Monthly 110 (2003), 340. Solution 112 (2005), 89–
90. [61]
Ghouila-Houri A., Une condition suffisante d’existence d’un circuit Hamiltonien. C. R. Adac. Sci. Paris
251 (1960), 495–497. [334]
Ghouila-Houri A., Caractérisation des graphes non orientés dont on peut orienter les arětes de manière
à obtenir le graphe d’une relation d’ordre. C. R. Acad. Sci. Paris 254 (1962), 1370–1371. [550]
Gilbert E.N., Lattice theoretic properties of frontal switching functions. J. Math. Physics 33 (1954), 57–
67. [564]
Gilbert E.N., Random graphs. Ann. Math. Stat. 30 (1959), 1141–1144. [688]
References 881

Gilmore P.C. and A.J. Hoffman, A characterization of comparability graphs and of interval graphs.
Canad. J. Math. 16 (1964), 539–548. [374, 549]
Glaisher J., A theorem in partitions. Messenger of Math. 12 (1883), 158–170. [142, 148]
Glock S., F. Joos, J. Kim, D. Kühn, and D. Osthus, Resolution of the Oberwolfach problem (2018).
(arXiv:1806.04644). [647]
Glock S., D. Kühn, A. Lo, and D. Osthus, The existence of designs via iterative absorption (2016).
(arXiv:1611.06827). [640]
Goddard W., Acyclic colorings of planar graphs. Discr. Math. 91 (1991), 91–94. [418]
Godsil C. and K. Meagher, Erdős–Ko–Rado theorems: algebraic approaches, Cambridge Studies in Ad-
vanced Mathematics 149 (Cambridge Univ. Press, 2016). [500, 781]
Godsil C. and G. Royle, Algebraic graph theory, Graduate Texts in Mathematics 207 (Springer-Verlag,
2001). [766]
Goldberg C.H. and D.B. West, Bisection of circle colorings. SIAM J. Algebraic Discrete Methods 6 (1985),
93–106. [809]
Goldberg M.K., Structure of multigraphs with restrictions on the chromatic class (Russian). Metody
Diskret. Analiz. 30 (1977), 3–12. [360]
Goldman J.R., J.T. Joichi, and D.E. White, Rook theory. I. Rook equivalence of Ferrers boards. Proc.
Amer. Math. Soc. 52 (1975), 485–492. [175]
Golomb S.W. and L.R. Welch, Perfect codes in the Lee metric and the packing of polyominoes. SIAM J.
Appl. Math. 18 (1970), 302–317. [29]
Golovina L.I. and I.M. Yaglom, Introduction in Geometry, Topics in mathematics (Heath, 1963). [419]
Golumbic M.C., Trivially perfect graphs. Discr. Math. 24 (1978), 105–107. [366]
Golumbic M.C., Algorithmic Graph Theory and Perfect Graphs (Academic Press, 1980, 2004). [370]
Golumbic M.C., D. Rotem, and J. Urrutia, Comparability graphs and intersection graphs. Discr. Math.
43 (1983), 37–46. [582]
Golumbic M.C. and E.R. Scheinerman, Containment graphs, posets, and related classes of graphs. In
Combinatorial Mathematics: Proceedings of the Third International Conference (New York, 1985),
Ann. New York Acad. Sci. 555 (New York Acad. Sci., 1989), 192–204. [831, 832]
Gonçalves D., Covering planar graphs with forests, one having bounded maximum degree. J. Combin.
Th. B 99 (2009), 314–322. [421]
Gondran M. and M. Minoux, Graphs and algorithms, Wiley-Interscience Series in Discr. Mathematics
(Wiley & Sons, 1984). [287]
Good I.J., Normal recurring decimals. J. Lond. Math. Soc. 21 (1946), 167–169. [239]
Goodman A.W., On sets of acquaintances and strangers at any party. Amer. Math. Monthly 66 (1959),
778–783. [477]
Goodman S. and S. Hedetniemi, Sufficient conditions for a graph to be Hamiltonian. J. Combin. Th. B
16 (1974), 175–180. [330]
Gosper R.W., Jr., Decision procedure for indefinite hypergeometric summation. Proc. Nat. Acad. Sci.
U.S.A. 75 (1978), 40–42. [89]
Gottschalk W.H., Choice functions and Tychonoff’s theorem. Proc. Am. Math. Soc. 2 (1951), 172. [468]
Gould H.W., Explicit formulas for Bernoulli numbers. Amer. Math. Monthly 79 (1972), 44–51.
[23, 30, 65]
Gould R.J., Graph Theory (Benjamin/Cummings, 1988). Also 2012. [215, 393]
Gould R.J., Updating the Hamiltonian problem—a survey. J. Graph Theory 15 (1991), 121–157. [316]
Gould R.J., Advances on the Hamiltonian problem—a survey. Graphs Combin. 19 (2003), 7–52. [316]
Gould R.J., Recent advances on the Hamiltonian problem: Survey III. Graphs Combin. 30 (2014), 1–46.
[316]
Goulden I.P. and L.G. Serrano, Maintaining the spirit of the reflection principle when the boundary has
arbitrary integer slope. J. Combin. Th. A 104 (2003), 317–326. [42]
Govorčin J. and R. Škrekovski, On the connectivity of Cartesian product of graphs. Ars Math. Contemp.
7 (2014), 293–297. [297]
882 References

Gowers W.T., A new proof of Szemerédi’s theorem for arithmetic progressions of length four. Geom. Funct.
Anal. 8 (1998), 529–551. [672]
Gowers W.T., A new proof of Szemerédi’s theorem. Geom. Funct. Anal. 11 (2001), 465–588. [466, 484]
Graham N., R.C. Entringer, and L.A. Székely, New tricks for old trees: maps and the pigeonhole principle.
Amer. Math. Monthly 101 (1994), 664–667. [428, 441]
Graham N. and F. Harary, Changing and unchanging the diameter of a hypercube. Discrete Appl. Math.
37-38 (1992), 265–274. [428, 565]
Graham N., F. Harary, M. Livingston, and Q.F. Stout, Subcube fault-tolerance in hypercubes. Inform.
and Comput. 102 (1993), 280–314. [584]
Graham R.L., Linear extensions of partial orders and the FKG inequality. In Ordered sets (Banff, Alta.,
1981), NATO Adv. Study Inst. C: Math. Phys. Sci. 83 (Reidel, 1982), 213–236. [597]
Graham R.L. and S. Butler, Rudiments of Ramsey theory (2nd ed.), CBMS Regional Conference Series in
Mathematics 123 (AMS, 2015). [425]
Graham R.L. and L.H. Harper, Some results on matching in bipartite graphs. SIAM J. Appl. Math. 17
(1969), 1017–1022. [559, 560]
Graham R.L. and D.J. Kleitman, Increasing paths in edge ordered graphs. Period. Math. Hungar. 3
(1973), 141–148. [431, 442]
Graham R.L., D.E. Knuth, and O. Patashnik, Concrete mathematics (Addison-Wesley, 1989). 2nd edition
1994. [24, 33, 41, 61, 89, 101]
Graham R.L. and H.O. Pollak, On the addressing problem for loop switching. Bell Sys. Tech. J. 50 (1971),
2495–2519. [771]
Graham R.L. and H.O. Pollak, On embedding graphs in squashed cubes. In Graph Th. and Appl. (Proc.
Conf., Western Michigan Univ., Mich., 1972; dedicated to the memory of J. W. T. Youngs), Lect.
Notes Math. 303 (Springer, 1972), 99–110. [781]
Graham R.L., V. Rödl, and A. Ruciński, On graphs with linear Ramsey numbers. J. Graph Theory 35
(2000), 176–192. [484]
Graham R.L., V. Rödl, and A. Ruciński, On bipartite graphs with linear Ramsey numbers. Combinatorica
21 (2001), 199–209. [484, 673]
Graham R.L. and B.L. Rothschild, A short proof of van der Waerden’s theorem on arithmetic progres-
sions. Proc. Amer. Math. Soc. 42 (1974), 385–386. [463, 464]
Graham R.L., B.L. Rothschild, and J. Spencer, Ramsey Theory (Wiley, 1980). Also 1990.
[425, 452, 463, 465, 470]
Graver J.E. and J. Yackel, Some graph theoretic results associated with Ramsey’s Theorem. J. Combin.
Th. 4 (1968), 125–175. [449, 451]
Gravier S. and F. Maffray, Graphs whose choice number is equal to their chromatic number. J. Graph
Theory 27 (1998), 87–97. [356]
Greene C., A multiple exchange property for bases. Proc. Amer. Math. Soc. 39 (1973), 45–50. [538]
Greene C., An extension of Schensted’s theorem. Advances in Math. 14 (1974), 254–265. [205]
Greene C., G.O.H. Katona, and D.J. Kleitman, Extensions of the Erdős–Ko–Rado theorem. Studies in
Appl. Math. 55 (1976), 1–8. [511]
Greene C. and D.J. Kleitman, The structure of Sperner ¾ -families. J. Combin. Th. A 20 (1976a), 41–68.
[548, 551]
Greene C. and D.J. Kleitman, Strong versions of Sperner’s theorem. J. Combin. Th. A 20 (1976b), 80–
88. [555, 565]
Greene C. and D.J. Kleitman, Proof techniques in the theory of finite sets. In Studies in combinatorics,
MAA Stud. Math. 17 (Math. Assoc. Amer., 1978), 22–79. [496, 561, 607]
Greene C. and T.L. Magnanti, Some abstract pivot algorithms. SIAM J. Appl. Math. 29 (1975), 530–539.
[538]
Greene C., A. Nijenhuis, and H.S. Wilf, A probabilistic proof of a formula for the number of Young
tableaux of a given shape. Adv. in Math. 31 (1979), 104–109. [190]
Greene J.E., A new short proof of Kneser’s conjecture. Amer. Math. Monthly 109 (2002), 918–920.
[811, 812]
References 883

Greenwell D.L. and H.V. Kronk, Uniquely line colorable graphs. Canad. Math. Bull. 16 (1973), 525–529.
[329]
Greenwood R.E. and A.M. Gleason, Combinatorial relations and chromatic graphs. Canad. J. Math. 7
(1955), 1–7. [449]
Gregory D.A. and K.N. Vander Meulen, Sharp bounds for decompositions of graphs into complete r-
partite subgraphs. J. Graph Theory 21 (1996), 393–400. [771]
Griggs J.R., Sufficient conditions for a symmetric chain order. SIAM J. Appl. Math. 32 (1977), 807–809.
[557, 561]
Griggs J.R., Collections of subsets with the Sperner property. Trans. AMS 269 (1982), 575–591. [568]
Griggs J.R., Problem 10665. Amer. Math. Monthly 105 (1998), 464. Solution 107 (2000), 653–4. [702]
Griggs J.R. and C.C. Ho, The cycling of partitions and compositions under repeated shifts. Adv. in Appl.
Math. 21 (1998), 205–227. [144, 145, 152]
Griggs J.R., J. Stahl, and W.T. Trotter, A Sperner theorem on unrelated chains of subsets. J. Combin.
Th. A 36 (1984), 124–127. [568]
Grimmett G.R. and C.J.H. McDiarmid, On colouring random graphs. Math. Proc. Camb. Phil. Soc. 77
(1975), 313–324. [701]
Grimmett G.R. and D.R. Stirzaker, Probability and random processes, 2nd ed. (Oxford Univ. Press, 1992).
Also 1982, 2001. [666]
Grinberg E.J., Plane homogeneous graphs of degree three without hamiltonian circuits. Latvian Math.
Yearbook 5 (1968), 51–58. [400]
Grinstead C.M. and S.M. Roberts, On the Ramsey numbers R(3 , 8) and R(3 , 9). J. Combin. Th. B 33
(1982), 27–51. [449]
Grolmusz V. and B. Sudakov, On ¾ -wise set-intersections and ¾ -wise Hamming-distances. J. Combin.
Th. A 99 (2002), 180–190. [731]
Grone R. and R. Merris, The Laplacian spectrum of a graph. II. SIAM J. Discr. Math. 7 (1994), 221–229.
[776]
Grone R., R. Merris, and V.S. Sunder, The Laplacian spectrum of a graph. SIAM J. Matrix Anal. Appl.
11 (1990), 218–238. [776]
Gross J.L. and T.W. Tucker, Topological graph theory, Wiley-Interscience Series in Discr. Mathematics
and Optimization (Wiley & Sons, 1987). [377]
Gross J.L. and J. Yellen, Graph Theory (CRC Press, 1999). [377]
Grötzsch H., Ein Dreifarbensatz für dreikreisfreie Netze auf der Kugel. Wiss. Z. Martin-Luther-U., Halle-
Wittenberg, Math.-Nat. Reihe 8 (1959), 109–120. [414]
Grünbaum B., Acyclic colorings of planar graphs. Israel J. Math. 14 (1973), 390–408. [423]
Grünbaum B. and T.S. Motzkin, The number of hexagons and the simplicity of geodesics on certain
polyhedra. Canad. J. Math. 15 (1963), 744–751. [387, 389]
Grytczuk J., J. Kozik, and P. Micek, New approach to nonrepetitive sequences. Random Structures Al-
gorithms 42 (2013), 214–225. [680]
Grytczuk J. and X. Zhu, The Alon–Tarsi number of a planar graph minus a matching (2020).
(arXiv:1811.12012). [403, 742]
Guan M., Graphic programming using odd and even points. Chinese Math. 1 (1962), 273–277. [287]
Guldan F., The linear arboricity of 10-regular graphs. Math. Slovaca 36 (1986), 225–228. [412]
Gunderson D.S. and V. Rödl, Extremal problems for affine cubes of integers. Combin. Probab. Comput.
7 (1998), 65–79. [473]
Guo S. and Z.W. Sun, Determination of the two-color Rado number for a1 x1 +· · ·+ am x m = x0 . J. Combin.
Th. A 115 (2008), 345–353. [473]
Gupta H., Combinatorial proof of a theorem on partitions into an even or odd number of parts. J. Combin.
Th. A 21 (1976), 100–103. [176]
Gupta R.P., The chromatic index and the degree of a graph (Abstract 66T-429). Notices Amer. Math. Soc.
13 (1966), 719. [358, 359, 372]
Gurgel M.A. and Y. Wakabayashi, On ¾-leaf-connected graphs. J.Combin. Th.B 41 (1986), 1–16. [327]
884 References

Gusfield D. and R.W. Irving, The Stable Marriage Problem: Structure and Algorithms (MIT Press, 1989).
[286]
Gustavsson T., Decompositions of Large Graphs and Digraphs with High Minimum Degree (Univ. of
Stockholm, 1991). Ph.D. Thesis. [640]
Gutin G., Finding a longest path in a complete multipartite digraph. SIAM J. Discr. Math. 6 (1993),
270–273. [334]
Gutner S., The complexity of planar graph choosability. Discr. Math. 159 (1996), 119–130. [402, 420]
Guy R.K., A problem of Zarankiewicz. In Theory of Graphs (Proc. Colloq., Tihany, 1966) (Academic Press,
1968), 119–150. [632]
Guy R.K., The decline and fall of Zarankiewicz’s Theorem. In Proof Techniques in Graph Theory (F.
Harary, ed.) (Academic Press, 1969), 63–69. [793]
Guy R.K., Sequences associated with a problem of Turán and other problems. In Combin. Conf. Bala-
tonfüred, 1969 (North-Holland, 1970), 553–569. [793, 797]
Guy R.K., Crossing numbers of graphs. In Graph Theory and Applications (Y. Alavi et al, ed.), Lect. Notes
Math. 303 (Springer, 1972), 111–124. [792, 795]
Guy R.K., The strong law of small numbers. Amer. Math. Monthly 95 (1988), 697–712. [105]
Guy R.K., Catwalks, sandsteps, and Pascal pyramids. J. Integer Seq. 3 (2000), Article 00.1.6 (electronic).
[64]
Guy R.K. and F. Harary, On the Möbius ladders. Canad. Math. Bull. 10 (1967), 493–496. [797]
Gyárfás A., On Ramsey covering-numbers. In Infinite and finite sets (Colloq., Keszthely, 1973), II, Colloq.
Math. Soc. János Bolyai 10 (North-Holland, 1975), 801–816. [351]
Gyárfás A., Particiófedések és lefogóhalmazok hipergráfokban. Tanulmányok-MTA Számitástechn. Au-
tomat. Kutató Int. Budapest (1977), 62. [460]
Gyárfás A., Vertex coverings by monochromatic paths and cycles. J. Graph Th. 7 (1983), 131–135.
[460]
Gyárfás A., Problems from the world surrounding perfect graphs. Tanulmányok—MTA Számitástech.
Automat. Kutató Int. Budapest (1985), 53. [351, 352]
Gyárfás A., Large monochromatic components in edge colorings of graphs: a survey. In Ramsey theory,
Progr. Math. 285 (Birkhäuser/Springer, 2011), 77–96. [460]
Gyárfás A., Vertex covers by monochromatic pieces—a survey of results and problems. Discr. Math. 339
(2016), 1970–1977. [460]
Gyárfás A., J. Lehel, J. Nešetřil, V. Rödl, R.H. Schelp, and Z. Tuza, Local ¾ -colorings of graphs and
hypergraphs. J. Combin. Th. B 43 (1987), 127–139. [472]
Gyárfás A. and G. Simonyi, Edge colorings of complete graphs without tricolored triangles. J. Graph
Theory 46 (2004), 211–216. [474]
Gyárfás A., E. Szemerédi, and Z. Tuza, Induced subtrees in graphs of large chromatic number. Discr.
Math. 30 (1980), 235–244. [352, 357]
Győri E., On division of graphs to connected subgraphs. In Combinatorics (Proc. 5th Hung. Colloq.,
Keszthely, 1976), I, Colloq. Math. Soc. János Bolyai 18 (North-Holland, 1978), 485–494. [312]
Habib M., C.J.H. McDiarmid, J. Ramirez-Alfonsin, and B.A. Reed (eds.), Probabilistic methods for algo-
rithmic discrete mathematics, Algorithms and Combin. 16 (Springer-Verlag, 1998). [657]
Hadamard J., Résolution d’une question relative aux déterminants (French). Bull. Sci. Math. 17 (1893),
240–246. [618, 619]
Hadamard J., Note sur quelques applications de l’indice de Kronecker in Jules Tannery. In Introduction
à la théorie des fonctions d’une variable 2 (Hermann & Fils, 1910), 737–477. [805]
Hadwiger H., Über eine Klassifikation der Streckenkomplexe. Vierteljschr. Naturforsch. Ges. Zürich 88
(1943), 133–142. [349]
Hadwiger H., Ueberdeckung des Euklidischen Raumes durch kongruente Mengen. Portugaliae Math. 4
(1945), 238–242. [729]
Hadwiger H., Mitteilung betreffend meine Note: Überdeckung einer Menge durch Mengen kleineren
Durchmessers. Comment. Math. Helv. 19 (1946), 72–73. [817]
Hadwiger H., Simultane Vierteilung zweier Körper. Arch. Math. (Basel) 17 (1966), 274–278. [815]
References 885

Haemers W.H., Eigenvalue techniques in design and graph theory, Mathematical Centre Tracts 121
(Mathematisch Centrum, 1980). [772]
Hagelstein P. and D. Herden, Problem 12071. Amer. Math. Monthly 125 (2018), 851. Solution 127 (2020),
464. [461]
Hagelstein P., D. Herden, and D. Young, Ramsey-type theorems for sets satisfying a geometric regularity
condition. J. Math. Anal. Appl. 447 (2017), 951–956. [461]
Häggkvist R., A lemma on cycle decompositions. In Cycles in graphs (Burnaby, B.C., 1982), North-
Holland Math. Stud. 115 (North-Holland, 1985), 227–232. [647, 648]
Häggkvist R. and J.C.M. Janssen, New bounds on the list-chromatic index of the complete graph and
other simple graphs. Combin. Probab. Comput. 6 (1997), 295–313. [362]
Häggkvist R. and C. Thomassen, Circuits through specified edges. Discr. Math. 41 (1982), 29–34.
[326]
Hajek B. and B. Narayanan, Multigraphs with the most edge covers. Inst. Math. Appl. Preprint Series
(1994). [765]
Hajós G., Über eine Konstruktion nicht n-färbbarer Graphen. Wiss. Z. Martin-Luther-Univ. Halle-
Wittenberg Math.-Nat. Reihe 10 (1961), 116–117. [355, 709]
Hakimi S.L., On the realizability of a set of integers as degrees of the vertices of a graph. SIAM J. Appl.
Math. 10 (1962), 496–506. [221, 238]
Hakimi S.L., On the degrees of the vertices of a directed graph. J. Franklin Inst. 279 (1965), 290–308.
[255, 261, 746]
Hakimi S.L. and E.F. Schmeichel, Improved bounds for the chromatic number of a graph. J. Graph
Theory 47 (2004), 217–225. [354]
Hales A.W. and R.I. Jewett, Regularity and positional games. Trans. Amer. Math. Soc. 106 (1963), 222–
229. [467, 474]
Halin R., A theorem on n-connected graphs. J. Combin. Th. 7 (1969), 150–4. [310]
Hall M., Distinct representatives of subsets. Bull. Amer. Math. Soc. 54 (1948), 922. [254, 255, 262]
Hall M. and H.J. Ryser, Cyclic incidence matrices. Canadian J. Math. 3 (1951), 495–502. [637]
Hall M., Jr., Cyclic projective planes. Duke Math. J. 14 (1947), 1079–1090. [637]
Hall P., On representatives of subsets. J. Lond. Math. Soc. 10 (1935), 26–30. [254]
Halmos P.R. and H.E. Vaughan, The marriage problem. Amer. J. Math 72 (1950), 214–215. [254]
Hammersley J.M., A few seedlings of research. In Proceedings of the Sixth Berkeley Symposium on Math-
ematical Statistics and Probability (Univ. California, Berkeley, Calif., 1970/1971), Vol. I: Theory
of statistics (Univ. California Press, 1972), 345–394. [821]
Hanani H., Über wesentlich unplättbare Kurven im drei-dimensionalen Raume (German). Fund. Math.
23 (1934), 135–142. As Chaim Chojnacki. [398]
Hanani H., The existence and construction of balanced incomplete block designs. Ann. Math. Statist. 32
(1961), 361–386. [656]
Hanani H., On balanced incomplete block designs with blocks having five elements. J. Combin. Th. A
12 (1972), 184–201. [641]
Hanani H., Balanced incomplete block designs and related designs. Discr. Math. 11 (1975), 255–369.
[639]
Hanlon P., Counting interval graphs. Trans. Amer. Math. Soc. 272 (1982), 383–426. [587]
Hansel G., Sur le nombre des fonctions booléennes monotones de n variables. C. R. Acad. Sci. Paris Sér.
A-B 262 (1966), A1088–A1090. [557]
Hansel G., Complexes et décompositions binomiales. J. Combinatorial Theory A 12 (1972), 167–183.
[496]
Hansen H.M., Scheduling with minimum waiting periods (Danish) (Odense Univ., 1992). Master Thesis.
[373]
Hanson D., C.O.M. Loten, and B. Toft, On interval colourings of bi-regular bipartite graphs. Ars Combin.
50 (1998), 23–32. [273, 373]
Harary F., On the notion of balance of a signed graph. Michigan Math. J. 2 (1953–54), 143–146. [237]
886 References

Harary F., The number of linear, directed, rooted, and connected graphs. Trans. Amer. Math. Soc. 78
(1955a), 445–463. [224]
Harary F., Note on the Pólya and Otter formulas for enumerating trees. Michigan Math. J. 3 (1955b),
109–112. [184]
Harary F., The maximum connectivity of a graph. Proc. Nat. Acad. Sci. U.S.A. 48 (1962a), 1142–1146.
[290]
Harary F., The determinant of the adjacency matrix of a graph. SIAM Review 4 (1962b), 202–210.
[780]
Harary F., Graph Theory (Addison-Wesley, 1969). [215]
Harary F., Reviews: The Polya Picture Album–Encounters of a Mathematician. Amer. Math. Monthly 96
(1989), 750–753. [224]
Harary F., D.F. Hsu, and Z. Miller, The biparticity of a graph. J. Graph Th. 1 (1977), 131–133. [343]
Harary F. and P.C. Kainen, The cube of a path is maximal planar. Bull. Inst. Combin. Appl. 7 (1993),
55–56. [797]
Harary F., P.C. Kainen, and A.J. Schwenk, Toroidal graphs with arbitrarily high crossing numbers.
Nanta Math. 6 (1973), 58–67. [794, 798]
Harary F. and P. Kovács, The smallest graphs with prescribed odd and even girth. Caribbean J. Math.
1 (1982), 24–26. [216]
Harary F. and C.St.J.A. Nash-Williams, On eulerian and hamiltonian graphs and line graphs. Canad.
Math. Bull. 8 (1965), 701–709. [329]
Harary F. and E.M. Palmer, Graphical enumeration (Academic Press, 1973). [135]
Harary F. and G. Prins, The block-cutpoint-tree of a graph. Publ. Math. Debrecen 13 (1966), 103–107.
[297]
Harary F. and A.J. Schwenk, The number of caterpillars. Discr. Math. 6 (1973), 359–365. [187]
Hardy G.H. and S. Ramanujan, Asymptotic formulæ in combinatory analysis. Proc. London Math. Soc.
(2) 17 (1918), 75–115. [140]
Harper L.H., Optimal numberings and isoperimetric problems on graphs. J. Combin. Th. 1 (1966), 385–
393. [496, 720, 800]
Harper L.H., The morphology of partially ordered sets. J. Combinatorial Theory A 17 (1974), 44–58.
[563]
Hartke S.G., D. Stolee, D.B. West, and M. Yancey, Extremal graphs with a given number of perfect
matchings. J. Graph Theory 73 (2013), 449–468. [276]
Hartman I.B-A., Berge’s conjecture on directed path partitions—a survey. Discr. Math. 306 (2006), 2498–
2514. [549]
Hartsfield N., A.K. Kelmans, and Y.Q. Shen, On the Laplacian polynomial of a K -cube extension. In
Proc. 27th Southeastern Intl. Conf. Combin. Graph Th. Comput. (Baton Rouge), Congr. Numer.
119 (1996), 73–77. [778]
Harzheim E., Remarks on Dilworth’s decomposition theorem. Ars Combin. 16 (1983), 27–31. [551]
Hasse M., Zur algebraischen Begründung der Graphentheorie. I. Math. Nachr. 28 (1964/1965), 275–290.
[339]
Havel I. and J. Morávek, b-valuations of graphs. Czechoslovak Math. J. 22 (1972), 338–351. [373]
Havel V., A remark on the existence of finite graphs (Czech). Časopis Pěst. Mat 80 (1955), 477–480.
[221]
Havet F. and J.S. Sereni, Improper choosability of graphs and maximum average degree. J. Graph Theory
52 (2006), 181–199. [412]
Haxell P.E., Partitioning complete bipartite graphs by monochromatic cycles. J. Combin. Th. B 69 (1997),
210–218. [492]
Haxell P.E., A note on vertex list colouring. Combin. Probab. Comput. 10 (2001), 345–347. [677]
Haxell P.E., Y. Kohayakawa, and T. Łuczak, The induced size-Ramsey number of cycles. Combin. Probab.
Comput. 4 (1995), 217–239. [455]
Hayes P.J., A note on the Towers of Hanoi problem. Computer J. 20 (1977), 282–285. [80]
Hayward R.B., Weakly triangulated graphs. J. Combin. Th. B 39 (1985), 200–208. [375]
References 887

Heawood P.J., Map-colour theorem. Q. J. Math. 24 (1890), 332–339. [399, 402]

Heawood P.J., On the four-colour map theorem. Q. J. Math. 29 (1898), 270–85. [419]
Hebbare S.P.R., Graceful cycles. Utilitas Math. 10 (1976), 307–317. [656]
Hedayat A.S., N.J.A. Sloane, and J. Stufken, Orthogonal arrays, Springer Series in Statistics (Springer-
Verlag, 1999). [651]
Hedetniemi S., On partitioning planar graphs. Canad. Math. Bull. 11 (1969), 203–210. [417]
Heffter L., Ueber gemeinsame Vielfache linearer Differentialausdrücke und lineare Differentialgle-
ichungen derselben Klasse. J. Reine Angew. Math. (1896), 157–166. [645]
Hell P. and J. Nešetřil, Graphs and homomorphisms, Oxford Lecture Series in Mathematics and its
Applications 28 (Oxford Univ. Press, 2004). [457]
Henning M.A. and A. Yeo, Tight lower bounds on the size of a maximum matching in a regular graph.
Graphs Combin. 23 (2007), 647–657. [274]
Henning M.A. and A. Yeo, Hypergraphs with large transversal number. Discr. Math. 313 (2013a), 959–
966. [263]
Henning M.A. and A. Yeo, Total domination in graphs, Springer Monographs in Mathematics (Springer,
2013b). [638]
Heppes A. and P. Révész, A splitting problem of Borsuk. Mat. Lapok 7 (1956), 108–111. [817]
Herrendörfer G. and D. Rasch, Complete block designs. II. Analysis of partially balanced designs. Bio-
metrical J. 19 (1977), 455–461. [655]
Heule M.J., Computing small unit-distance graphs with chromatic number 5. Geombinatorics 28 (2018),
32–50. (Also arXiv:1805.12181) [342]
Hierholzer C., Über die Möglichkeit, einen Linienzug ohne Wiederholung und ohne Unterbrechnung zu
umfahren. Math. Ann. 6 (1873), 30–32. [234]
Hilbert D., Ueber die Irreducibilität ganzer rationaler Functionen mit ganzzahligen Coefficienten. J.
Reine Angew. Math. 110 (1892), 104–129. [461, 473]
Hill C., Solution to problem 10992. Amer. Math. Monthly 111 (2004), 827–829. Proposed 110 (2003), 155.
[34]
Hilton A.J.W., The number of spanning trees of labeled wheels, fans and baskets. In Combinatorics (Proc.
Conf. Combin. Math., Math. Inst., Oxford, 1972) (Inst. Math. Appl., 1972), 203–206. [65]
Hilton A.J.W., A theorem on finite sets. Quart. J. Math. Oxford (2) 27 (1976), 33–36. [497, 607]
Hilton A.J.W., A simple proof of the Kruskal–Katona theorem and of some associated binomial inequal-
ities. Period. Math. Hungar. 10 (1979), 25–30. [496]
Hilton A.J.W., Two conjectures on edge colouring. Discr. Math. 74 (1989), 61–64. [359]
Hilton A.J.W. and E.C. Milner, Some intersection theorems for systems of finite sets. Quart. J. Math.
Oxford Ser. (2) 18 (1967), 369–384. [499, 511]
Hilton P. and J. Pederson, Catalan numbers, their generalization, and their uses. Math. Intel. 13 (1991),
64–75. [41]
Hind H., M. Molloy, and B.A. Reed, Colouring a graph frugally. Combinatorica 17 (1997), 469–482.
[681]
Hind H., M. Molloy, and B.A. Reed, Total coloring with + poly(log ) colors. SIAM J. Comput. 28 (1999),
816–821. [373, 681]
Hindman N., Ultrafilters and combinatorial number theory. In Number theory, Carbondale 1979 (Proc.
Southern Illinois Conf., Carbondale, 1979), Lecture Notes in Math. 751 (Springer, 1979), 119–
184. [466]
Hindman N., On a conjecture of Erdős, Faber, and Lovász about n-colorings. Canad. J. Math. 33 (1981),
563–570. [439]
Hinrichs A. and C. Richter, New sets with large Borsuk numbers. Discr. Math. 270 (2003), 137–147.
[817]
Hiraguchi T., On the dimension of partially ordered sets. Sci. Rep. Kanazawa Univ. 1 (1951), 77–94.
[571, 574, 583]
Hiraguchi T., On the dimension of orders. Sci. Reports Kanazawa Univ. 4 (1955), 1–20.
[571, 574, 575, 583]
888 References

Hladký J., D. Král’, and U. Schauz, Brooks’ theorem via the Alon–Tarsi theorem. Discr. Math. 310 (2010),
3426–3428. [347]
Hliněný P., Crossing number is hard for cubic graphs. J. Combin. Th. B 96 (2006), 455–471. [795]
Hoare A.H.M., An involution of blocks in the partitions of n. Amer. Math. Monthly 93 (1986), 475–476.
[152]
Hobby C.R. and J.R. Rice, A moment problem in L1 approximation. Proc. Amer. Math. Soc. 16 (1965),
665–670. [810]
Hochberg R., C.J.H. McDiarmid, and M.E. Saks, On the bandwidth of triangulated triangles. In 14th
Brit. Combin. Conf. (Keele, 1993) 138 (Discr. Math., 1995), 261–5. [801–803]
Hoeffding W., Probability inequalities for sums of bounded random variables. J. Amer. Statist. Assoc. 58
(1963), 13–30. [708, 715, 722]
Hoffman A.J., On the polynomial of a graph. Amer. Math. Monthly 70 (1963), 30–36. [772]
Hoffman A.J., On eigenvalues and colorings of graphs. In Graph Theory and Its Applications (B. Harries,
ed.) (Academic Press, 1970), 79–91. [781]
Hoffman A.J., Extending Greene’s theorem to directed graphs. J. Combin. Theory A 34 (1983), 102–107.
[549]
Hoffman A.J. and R.R. Singleton, On Moore graphs with diameters 2 and 3. IBM J. Res. Develop. 4
(1960), 497–504. [244, 775]
Hoffman D.G. and P.J. Schellenberg, The existence of C¾ -factorizations of K 2n − F. Discr. Math. 97 (1991),
243–250. [647]
Holland F., Problem 11798. Amer. Math. Monthly 121 (2014), 798. Solution 124 (2017), 370. [23, 117]
Holley R., Remarks on the FKG inequalities. Comm. Math. Phys. 36 (1974), 227–231. [597]
Hollingsworth S., Packing trees into complete bipartite graphs. Discr. Math. 313 (2013), 945–948.
[248]
Holton D. and R.E.L. Aldred, Planar graphs, regular graphs, bipartite graphs and Hamiltonicity. Aus-
tralas. J. Combin. 20 (1999), 111–131. [326]
Holton D.A. and J. Sheehan, The Petersen Graph (Cambridge Univ. Press, 1993). [213]
Holyer I., The NP-completeness of edge-coloring. SIAM J. Comput. 10 (1981), 718–720. [359, 699]
Holzmann C.A. and F. Harary, On the tree graph of a matroid. SIAM J. Appl. Math. 22 (1972), 187–193.
[540]
Hook J. and G. Isaak, Star-critical Ramsey numbers. Discrete Appl. Math. 159 (2011), 328–334. [459]
Hopcroft J.E. and R.M. Karp, An n5/2 algorithm for maximum matchings in bipartite graphs. SIAM J.
Comput. 2 (1973), 225–231. [283, 284]
Hopcroft J.E. and R.E. Tarjan, Efficient planarity testing. J. Assoc. Comput. Mach. 21 (1974), 549–568.
[393]
Horadam A.F., A generalized Fibonacci sequence. Amer. Math. Monthly 68 (1961), 455–459. [61]
Horák P., Q. He, and W.T. Trotter, Induced matchings in cubic graphs. J. Graph Th. 17 (1993), 151–160.
[373]
Horák P. and Z. Tuza, A coloring problem related to the Erdős–Faber–Lovász conjecture. J. Combin. Th.
B 50 (1990), 321–322. [344]
Horton J.D., On two-factors of bipartite regular graphs. Discr. Math. 41 (1982), 35–41. [317]
Hosoya H. and F. Harary, On the matching properties of three fence graphs. In Appl. Graph Th. and
Discr. Math. Chem. (Saskatoon, 1991) 12 (1993), 211–218. [765]
Hoşten S. and W.D. Morris, Jr., The order dimension of the complete graph. Discr. Math. 201 (1999),
133–139. [580]
Howard F.T., The number of multinomial coefficients divisible by a fixed power of a prime. Pacific J.
Math. 50 (1974), 99–108. [48]
Hsieh W.N. and D.J. Kleitman, Normalized matching in direct products of partial orders. Studies in
Appl. Math. 52 (1973), 285–289. [563]
Huang D. and D. Scully, Problem 11052. Amer. Math. Monthly 110 (2003), 957. Solution 113 (2006), 183.
[152]
Huneke C., The friendship theorem. Amer. Math. Monthly 109 (2002), 192–194. [776]
References 889

Hurlbert G.H., A short proof that n3 is not a circle containment order. J. Order 5 (1988), 235–237.
[828]
Hutchinson J.P., Problem 10478. Amer. Math. Monthly 102 (1995), 746. Solution 105 (1998), 274–275.
[419]
Hutchinson J.P., Three- and four-coloring nearly triangulated surfaces. In Proc. 32nd Southeastern Intl.
Conf. Combin. Graph Th. Comput. (Boca Raton), 150 (2001), 129–143. [419]
Hutchinson J.P., Arc- and circle-visibility graphs. Australas. J. Combin. 25 (2002), 241–262. [394]
Hutchinson J.P. and P.B. Trow, Some pigeonhole principle results extended. Amer. Math. Monthly 87
(1980), 648–651. [429]
Huxley M., The difference between consecutive primes. In Analytic Number Theory (St. Louis Univ.,
1972), Proc. Sympos. Pure Math 24 (Amer. Math. Soc., 1973), 141–145. [730]
Igusa K., Solution of the Bulgarian solitaire conjecture. Math. Mag. 58 (1985), 259–271. [144, 152]
Ionin Y., Solution to problem 11678. Amer. Math. Monthly 121 (2014), 952–953. Proposed 119 (2012),
880. [178]
Isaacs R., Infinite families of nontrivial trivalent graphs which are not Tait colorable. Amer. Math.
Monthly 82 (1975), 221–239. [372]
Ishigami Y., Proof of a conjecture of Bollobás and Kohayakawa on the Erdős–Stone theorem. J. Combin.
Th. B 85 (2002), 222–254. [482]
Itai A. and M. Rodeh, Covering a graph by circuits. In Automata, Langs. and Prog. (Udine, 1978), Lect.
Notes Comp. Sci. 62 (Springer-Verlag, 1978), 289–299. [249]
Jackson B., Hamilton cycles in regular 2-connected graphs. J. Combin. Th. B 29 (1980), 27–46. [326]
Jackson B., Hamilton cycles in 7-connected line graphs (1989). Unpublished preprint. [326]
Jackson B. and O. Ordaz, Chvátal–Erdős conditions for paths and cycles in graphs and digraphs. A
survey. Discr. Math. 84 (1990), 241–254. [321]
Jacobi C.G.J., De resolutione aequationum per series infinitas. J. Reine Angew. Math. 6 (1830), 257–
286. [129]
Jacobson M.S., F.R. McMorris, and H.M. Mulder, Tolerance intersection graphs. In Graph Theory, Com-
binatorics, and Applications (Kalamazoo, 1988) (Y. Alavi, G. Chartrand, O.R. Oellerman and A.J.
Schwenk, eds.) (Wiley, 1991), 705–724. [374]
Jaeger F., Tait’s theorem for graphs with crossing number at most one. Ars Combin. 9 (1980), 283–287.
[797]
Jahanbekam S., J. Kim, S. O, and D.B. West, On r-dynamic coloring of graphs. Discrete Appl. Math. 206
(2016), 65–72. [668]
Jahanbekam S. and D.B. West, New lower bounds for matching numbers of general and bipartite graphs.
Congr. Numer. 218 (2013), 57–59. [262, 273]
Jahanbekam S. and D.B. West, Anti-Ramsey problems for t edge-disjoint rainbow spanning subgraphs:
cycles, matchings, or trees. J. Graph Th. 82 (2016), 75–89. [333]
Jamison R.E., Covering finite fields with cosets of subspaces. J. Combin. Th. A 22 (1977), 253–266.
[735]
Jamison R.E., Orientable edge colorings of graphs. Discrete Appl. Math. 159 (2011), 595–604. [472]
Jamison R.E., T. Jiang, and A.C.H. Ling, Constrained Ramsey numbers of graphs. J. Graph Theory 42
(2003), 1–16. [473]
Jamison R.E. and D.B. West, On pattern Ramsey numbers of graphs. Graphs Combin. 20 (2004), 333–
339. [472]
Janson S., T. Łuczak, and A. Ruciński, Random Graphs (Wiley-Interscience, 2000). [657, 696, 708]
Janssen J.C.M., The Dinitz problem solved for rectangles. Bull. Amer. Math. Soc. (N.S.) 29 (1993), 243–
249. [363]
Jarnı́k V., O jistém problému minimálnim. Acta Societatis Scientiarum Natur. Moravicae 6 (1930), 57–
63. [246, 252]
Jendrol’ S., A short proof of Kotzig’s theorem on minimal edge weights of convex 3-polytopes (1999), 35–
38. [411]
890 References

Jendrol’ S. and H.-J. Voss, Light subgraphs of graphs embedded in the plane—a survey. Discr. Math.
313 (2013), 406–421. [407]
Jensen T.R. and B. Toft, Graph coloring problems, Wiley-Interscience Series in Discr. Mathematics and
Optimization (Wiley & Sons, 1995). [335]
Jenssen M. and J. Skokan, Exact Ramsey numbers of odd cycles via nonlinear optimisation (2016).
(arXiv:1608.05705). [461]
Jerrum M., A. Sinclair, and E. Vigoda, A polynomial-time approximation algorithm for the permanent of
a matrix with non-negative entries. In Proceedings of the Thirty-Third Annual ACM Symposium
on Theory of Computing (ACM, 2001), 712–721. [754]
Jiang T., K.G. Milans, and D.B. West, Degree Ramsey numbers for cycles and blowups of trees. European
J. Combin. 34 (2013), 414–423. [685]
Jiang T. and D. Mubayi, New upper bounds for a canonical Ramsey problem. Combinatorica 20 (2000),
141–146. [471]
Jichang S. and D.J. Kleitman, Superantichains in the lattice of partitions of a set. Studies in applied
mathematics 71 (1984), 207–241. [567]
Johannson K.R., Variations on a theorem by van der Waerden (Univ. Manitoba, 2007). Mas-
ters Thesis. Available online at http://mspace.lib.umanitoba.ca/bitstream/handle/1993/321/
thesismain.pdf [466]
Johnson D.M., A.L. Dulmage, and N.S. Mendelsohn, Orthomorphisms of groups and orthogonal latin
squares. I. Canad. J. Math. 13 (1961), 356–372. [654]
Johnson N.L. and S. Kotz, Urn models and their application (Wiley, 1977). [106]
Johnson R.T. and T.P. Vaughan, On union-closed families. I. J. Combin. Th. A 84 (1998), 242–249.
[503]
Jonsson J. and J.G. Propp, Problem 11298. Amer. Math. Monthly 114 (2007), 547. Solution 116 (2009),
371–372. [387]
Jordan C., Sur les assemblages de lignes. J. Reine Angew. Math. 70 (1869), 185–190. [243, 251]
Jordan C., Questions de probabilités. Bull. Soc. Math. France 1 (1872), 256–258. [669]
Jorza A., Problem 10856. Amer. Math. Monthly 108 (2001), 172. Solution 113 (2006), 180–183. [388]
Joyal A., Une théorie combinatoire des séries formelles. Adv. in Math. 42 (1981), 1–82. [37, 127]
Józsa S. and E. Szemerédi, The number of unit distances on the plane. In Infinite and finite sets (Colloq.,
Keszthely, 1973), II, Colloq. Math. Soc. Jáanos Bolyai 10 (North-Holland, 1975), 939–950.
[794]
Juarez H.A. and G. Salazar, Drawings of Cm × Cn with one disjoint family. II. J. Combin. Th. B 82 (2001),
161–165. [794]
Jukna S., Extremal combinatorics (2nd ed.), Texts in Theoretical Computer Science. An EATCS Series
(Springer, 2011). [493]
Juvan M., B. Mohar, and R. Škrekovski, On list edge-colorings of subcubic graphs. Discr. Math. 187
(1998a), 137–149. [412]
Juvan M., B. Mohar, and R. Škrekovski, List total colourings of graphs. Combin. Probab. Comput. 7
(1998b), 181–188. [60, 739]
Juvan M., B. Mohar, and R. Škrekovski, Graphs of degree 4 are 5-edge-choosable. J. Graph Theory 32
(1999), 250–264. [412]
Kabela A. and T. Kaiser, 10-tough chordal graphs are Hamiltonian. J. Combin. Th. B 122 (2017), 417–
427. [318]
Kahn J., Asymptotically good list-colorings. J. Combin. Th. A 73 (1996), 1–59. [362]
Kahn J., On some hypergraph problems of Paul Erdős and the asymptotics of matchings, covers and
colorings. In The mathematics of Paul Erdős, I, Algorithms Combin. 13 (Springer, 1997), 345–
371. [439]
Kahn J. and G. Kalai, A counterexample to Borsuk’s conjecture. Bull. Amer. Math. Soc. (N.S.) 29 (1993),
60–62. [817, 818]
Kahn J. and J.H. Kim, Entropy and sorting. J. Comput. System Sci. 51 (1995), 390–399. [823]
References 891

Kahn J. and N. Linial, Balancing extensions via Brunn–Minkowski. Combinatorica 11 (1991), 363–368.
[824, 825]
Kahn J. and M.E. Saks, Balancing poset extensions. Order 1 (1984), 113–126. [823, 824, 826, 831]
Kainen P.C., A generalization of the 5-color theorem. Proc. Amer. Math. Soc. 45 (1974), 450–453. [419]
Kaiser T., E. Rollová, and R. Lukot’ka, Nowhere-zero flows in signed graphs: A survey. Lect. Notes. Sem-
inar. Interdiscip. Matematica 14 (2017), 85–104. [237]
Kaiser T., Z. Ryjáček, D. Král’, M. Rosenfeld, and H.J. Voss, Hamilton cycles in prisms. J. Graph Th. 56
(2007), 249–269. [326]
Kalbfleisch J.D., J.G. Kalbfleisch, and R. Stanton, A combinatorial problem on convex n-gons. In
Proc. Louisiana Conf. on Combinatorics, Graph Theory and Computing (Louisiana State Univ.,
1970)(RC Mullin, KB Reid, and DP Roselle, eds.) (1970), 180–188. [458]
Kalbfleisch J.G., Upper bounds for some Ramsey numbers. J. Combin. Th. 2 (1967), 35–42. [449]
Kamat V.M., Erdős–Ko–Rado Theorems: New Generalizations, Stability Analysis and Chvatal’s Conjec-
ture (ProQuest LLC, 2011). Thesis (Ph.D.)–Arizona State University. [503]
Kampen G.R., Orienting planar graphs. Discr. Math. 14 (1976), 337–341. [786]
Kano M. and X. Li, Monochromatic and heterochromatic subgraphs in edge-colored graphs—a survey.
Graphs Combin. 24 (2008), 237–263. [457]
Kantrowitz M., Problem E3130. Amer. Math. Monthly 93 (1986), 131. Solution 95 (1988), 555–6. [61]
Kapoor S.F., A.D. Polimeni, and C.E. Wall, Degree sets for graphs. Fund. Math. 95 (1977), 189–194.
[227]
Karaganis J.J., On the cube of a graph. Canad. Math. Bull. 11 (1968), 295–296. [331]
Karaivanov B. and T.S. Vassilev, Solution to problem 11798. Amer. Math. Monthly 124 (2017). Proposed
121 (2014). [35]
Karger D.R., P.N. Klein, and R.E. Tarjan, A randomized linear-time algorithm to find minimum spanning
trees. J. Assoc. Comput. Mach. 42 (1995), 321–328. [246]
Karp R.M., Reducibility among combinatorial problems. In Complexity of computer computations (Proc.
Sympos., IBM Thomas J. Watson Res. Center, Yorktown Heights, N.Y., 1972) (Plenum, 1972), 85–
103. [11]
Kárteszi F., Piani finiti ciclici come risoluzioni di un certo problema di minimo. (Italian). Boll. Un. Mat.
Ital. (3) 15 (1960), 522–528. [639]
Kasteleyn P.W., The statistics of dimers on a lattice, I. The number of dimer arrangements on a quadratic
lattice. Physica 27 (1961), 1209–1225. [755]
Kasteleyn P.W., Dimer statistics and phase transitions. J. Math. Phys. 4 (1963), 287–293. [755, 756]
Kasteleyn P.W., Graph theory and crystal physics. In Graph Theory and Theoretical Physics (Academic
Press, 1967), 43–110. [755]
Kászonyi L. and Z. Tuza, Saturated graphs with minimal number of edges. J. Graph Theory 10 (1986),
203–210. [491]
Katerinis P., Some conditions for the existence of º -factors. J. Graph Theory 9 (1985), 513–521. [276]
Katona G.O.H., Intersection theorems for systems of finite sets. Acta Math. Acad. Sci. Hungar. 15 (1964),
329–337. [497, 511]
Katona G.O.H., On a conjecture of Erdős and a stronger form of Sperner’s theorem. Studia Sci. Math.
Hungar. 1 (1966), 59–63. [565]
Katona G.O.H., A theorem of finite sets. In Theory of graphs (Proc. Colloq., Tihany, 1966) (Academic
Press, 1968), 187–207. [495, 496]
Katona G.O.H., A generalization of some generalizations of Sperner’s theorem. J. Combinatorial Theory
B 12 (1972a), 72–81. [554]
Katona G.O.H., A simple proof of the Erdős–Chao Ko–Rado theorem. J. Combin. Th. B 13 (1972b), 183–
184. [498]
Kauers M., Problem 11545. Amer. Math. Monthly 118 (2011), 84. Solution 119 (2012), 885–886. [174]
Kawarabayashi K.i., A survey on Hamiltonian cycles. In Proc. Workshop on Graph. Th. and Related
Topics (Sendai, 1999), 7 (2001), 25–39. [316]
892 References

Kearnes K.A. and E.W. Kiss, Finite algebras of finite complexity. Discr. Math. 207 (1999), 89–135.
[665]
Keedwell A.D. and J. Dénes, Latin squares and their applications (Elsevier/North-Holland, 2015), second
edn. [610]
Keevash P., Shadows and intersections: stability and new proofs. Adv. Math. 218 (2008), 1685–1703.
[496]
Keevash P., Hypergraph Turán problems. In Surveys in combinatorics 2011, London Math. Soc. Lecture
Note Ser. 392 (Cambridge Univ. Press, 2011), 83–139. [475]
Keevash P., The existence of designs (2014). (arXiv:1401.3665). [640]
Keevash P., Counting designs. J. Eur. Math. Soc. (JEMS) 20 (2018), 903–927. [640]
Kellogg A., Problem 10585. Amer. Math. Monthly 104 (1997), 361. Solution 106 (1999), 170–171. [177]
Kelly D., The 3-irreducible partially ordered sets. Canad. J. Math. 29 (1977), 367–383. [573]
Kelly D., On the dimension of partially ordered sets. Discr. Math. 35 (1981), 135–156. [571, 606]
Kelly D. and W.T. Trotter, Dimension theory for ordered sets. In Ordered sets (Banff, Alta., 1981), NATO
Adv. Study Inst. C: Math. Phys. Sci. 83 (Reidel, 1982), 171–211. [572, 573, 583]
Kelly J.B. and L.M. Kelly, Paths and circuits in critical graphs. Amer. J. Math. 76 (1954), 786–792.
[434]
Kelmans A.K., The number of trees in a graph, I and II. Automat. Remote Control 26/27 (1965/66), 2118–
2129, 233–241. [778]
Kelmans A.K., The properties of the characteristic polynomial of a graph (Russian). Cybernetics 4 (Izdat.
“Énergija”, 1967), 27–41. [777, 782]
Kelmans A.K., A new planarity criterion for 3-connected graphs. J. Graph Th. 5 (1981), 259–267.
[393]
Kelmans A.K., A strengthening of the Kuratowski planarity criterion for 3-connected graphs. Discr.
Math. 51 (1984), 215–220. [393]
Kelmans A.K., On Hamiltonicity of {claw, net}-free graphs. Discr. Math. 306 (2006), 2755–2761. [297]
Kelmans A.K. and V.M. Chelnokov, A certain polynomial of a graph and graphs with an extremal number
of trees. J. Combin. Th. B 16 (1974), 197–214. [777]
Kempe A.B., On the geographical problem of four colours. Amer. J. Math. 2 (1879), 193–200. [399]
Kempe A.B., A memoir on the theory of mathematical form. In Philosophical Transactions of the Royal
Society of London 177 (1886), 1–70. [213]
Kendall M.G. and B.B. Smith, On the method of paired comparisons. Biometrika 31 (1940), 324–345.
[228]
Kerimov A., Problem 11454. Amer. Math. Monthly 116 (2009), 746. Solution 118 (2011), 659. [315]
Keselman G., Solution to problem 11274. Amer. Math. Monthly 115 (2008). Proposed 114 (2007), 165.
[116]
Kézdy A.E. and H.St.C. Snevily, Distinct sums modulo n and tree embeddings. Combin. Probab. Comput.
11 (2002), 35–42. [746]
Khan M.A., Problem E3451. Amer. Math. Monthly 98 (1991), 645. Solution 100 (1993), 303. [133]
Kharaghani H. and B. Tayfeh-Rezaie, A Hadamard matrix of order 428. J. Combin. Des. 13 (2005), 435–
440. [622]
Khare C.B., Problem E3315. Amer. Math. Monthly 96 (1989), 253. Solution 98 (1991), 366–367. [440]
Kierstead H.A., Long stars specify -bounded classes. In Sets, graphs and numbers (Budapest, 1991),
Colloq. Math. Soc. János Bolyai 60 (North-Holland, 1992), 421–428. [352]
Kierstead H.A., On the choosability of complete multipartite graphs with part size three. Discr. Math.
211 (2000), 255–259. [355, 356]
Kierstead H.A. and G. Konjevod, Coloring number and on-line Ramsey theory for graphs and hyper-
graphs. Combinatorica 29 (2009), 49–64. [457]
Kierstead H.A. and A.V. Kostochka, Efficient graph packing via game colouring. Combin. Probab. Com-
put. 18 (2009), 765–774. [248]
Kierstead H.A. and S.G. Penrice, Recent results on a conjecture of Gyárfás. In Proc. 21st Southeastern
Intl. Conf. Combin. Graph Th. Comput. (Boca Raton), 79 (1990), 182–186. [352]
References 893

Kierstead H.A. and S.G. Penrice, Radius two trees specify -bounded classes. J. Graph Th. 18 (1994),
119–129. [352]
Kierstead H.A., A. Salmon, and R. Wang, On the choice number of complete multipartite graphs with
part size four (2014). (arXiv:1407.3817v1). [356]
Kierstead H.A. and W.T. Trotter, Explicit matchings in the middle levels of the Boolean lattice. Order 5
(1988), 163–171. [50]
Kierstead H.A. and W.T. Trotter, A note on removable pairs. In Graph theory, combinatorics, and appli-
cations. Vol. 2 (Kalamazoo, MI, 1988), Wiley-Intersci. Publ. (Wiley, 1991), 739–742. [584]
Kierstead H.A. and Y. Zhu, Radius three trees in graphs with large chromatic number. SIAM J. Discr.
Math. 17 (2004), 571–581. [352]
Kim J.H., The Ramsey number R(3 , t) has order of magnitude t2/log t. Random Structures Algorithms
7 (1995), 173–207. [451, 664, 682]
Kim S.J. and W.J. Park, List dynamic coloring of sparse graphs. In Combinatorial optimization and
applications, Lecture Notes in Comput. Sci. 6831 (Springer, 2011), 156–162. [424]
Kimble R.J., Jr, Extremal Problems in Dimension–Theory for Partially-Ordered Sets (ProQuest, 1973).
Ph.D. Thesis, Massachusetts Inst. Tech. [573]
Kimble R.J., Jr. and A.J. Schwenk, On universal caterpillars. In The theory and applications of graphs
(Kalamazoo, 1980) (Wiley, 1981), 437–447. [187]
King A.D. and B.A. Reed, A short proof that  can be bounded away from + 1 toward . J. Graph
Th. 81 (2016), 30–34. [685]
Kirchhoff G., über die Auflösung der Gleichungen, auf welche man bei der Untersuchung der linearen
Verteilung galvanischer Ströme geführt wird. Ann. Phys. Chem. 72 (1847), 497–508. [37]
Kirdar M.S. and T.H.R. Skyrme, On an identity relating to partitions and repetitions of parts. Canad.
J. Math. 34 (1982), 194–195. [152]
Kirkman T.P., On a problem in cominations. Cambridge and Dublin Math. J. 2 (1847), 191–204.
[612, 641]
Kirkman T.P. In The Lady’s and Gentleman’s Diary, (J. Greenhill, 1850) (1850). [649]
Kislicyn S.S., Finite partially ordered sets and their corresponding permutation sets. Mat. Zametki 4
(1968), 511–518. [823]
Klavžar S., U. Milutinović, and C. Petr, On the Frame–Stewart algorithm for the multi-peg Tower of
Hanoi problem. Discrete Appl. Math. 120 (2002), 141–157. [70]
Klavžar S. and M. Petkovšek, Problem E3281. Amer. Math. Monthly 95 (1988), 655. Solution 97 (1990),
924–925. [606]
Klazar M., The Füredi–Hajnal conjecture implies the Stanley–Wilf conjecture. In Formal power series
and algebraic combinatorics (Moscow, 2000) (Springer, 2000), 250–255. [432]
Kleitman D., On Dedekind’s problem: The number of monotone Boolean functions. Proc. Amer. Math.
Soc. 21 (1969), 677–682. [509, 558]
Kleitman D., Review of “On a problem of Moser”. MathSciNet Review (1973). MR0297582 (45 #6636)
[551]
Kleitman D.J., On a lemma of Littlewood and Offord on the distribution of certain sums. Math. Z. 90
(1965), 251–259. [565]
Kleitman D.J., Families of non-disjoint subsets. J. Combin. Th. 1 (1966), 153–155. [595, 608]
Kleitman D.J., On families of subsets of a finite set containing no two disjoint sets and their union. J.
Combinatorial Theory 5 (1968), 235–237. [500]
Kleitman D.J., The crossing number of K 5 ,n. J. Combin. Th. 9 (1970), 315–323. [566, 793, 797]
Kleitman D.J., On an extremal property of antichains in partial orders. The LYM property and some of
its implications and applications. In Combinatorics (Proc. NATO Advanced Study Inst., Breuke-
len, 1974), Part 2: Graph theory; foundations, partitions and combinatorial geometry (Math. Cen-
trum, 1974), 77–90. Math. Centre Tracts, No. 56. [559]
Kleitman D.J., A note on some subset identities. Studies in Appl. Math. 54 (1975), 289–292. [40]
894 References

Kleitman D.J., Extremal hypergraph problems. In Surveys in combinatorics (Proc. Seventh British Com-
binatorial Conf., 1979), London Math. Soc. Lecture Note Ser. 38 (Cambridge Univ. Press, 1979),
44–65. [501]
Kleitman D.J., Extremal problems on hypergraphs. In Extremal problems for finite sets (Visegrád, 1991),
Bolyai Soc. Math. Stud. 3 (János Bolyai Math. Soc., 1994), 355–374. [493]
Kleitman D.J. and T.L. Magnanti, On the number of latent subsets of intersecting collections. J. Combin.
Th. A 16 (1974), 215–220. [501]
Kleitman D.J. and G. Markowsky, On Dedekind’s problem: the number of isotone Boolean functions. II.
Trans. Amer. Math. Soc. 213 (1975), 373–390. [558, 580]
Kleitman D.J., J. Shearer, and D. Sturtevant, Intersections of ¾ -element sets. Combinatorica 1 (1981),
381–384. [507]
Klešč M., R.B. Richter, and I. Stobert, The crossing number of C5 × Cn . J. Graph Theory 22 (1996), 239–
243. [794]
Klotz W., A constructive proof of Kuratowski’s Theorem. Ars Combinatoria 28 (1989), 51–54. [393]
Klove T., Problem 10460. Amer. Math. Monthly 102 (1995), 553. Solution 105 (1998), 69. [22]
Kneser M., Aufgabe 300 (German). Jahresber Deutsch. Math.-Verein 58 (1955), 27. [251, 811]
Knuth D., Problem 12055. Amer. Math. Monthly 125 (2018), 660. Solution 127 (2020), 186. [152]
Knuth D.E., The Art of Computer Programming, Vol. 1: Fundamental Algorithms (Addison-Wesley,
1968). Also 1973, 1997. [99, 101, 122]
Knuth D.E., Permutations, matrices, and generalized Young tableaux. Pacific J. Math. 34 (1970), 709–
727. [193, 195, 198, 208]
Knuth D.E., The Art of Computer Programming. Vol. 3: Sorting and searching (Addison-Wesley, 1973).
[50, 101, 190, 193]
Knuth D.E., Big omicron and big omega and big theta. ACM SIGACT News 8 (1976a), 18–24. [10]
Knuth D.E., Mariages Stables (Les Presses de l’Univ. de Montréal, 1976b). English trans. AMS, 1997.
[286]
Knuth D.E., Problem E3463. Amer. Math. Monthly 98 (1991), 852. Solution 100 (1993), 693–4. [441]
Knuth D.E., Problem 10298. Amer. Math. Monthly 100 (1993), 400. Solution 103 (1996), 80–81. [133]
Knuth D.E., Problem 10546. Amer. Math. Monthly 103 (1996), 695. Solution 105 (1998), 867–8. [176]
Knuth D.E., Problem 10720. Amer. Math. Monthly 106 (1999), 264. Solution 110 (2003), 60–61. [669]
Knuth D.E., Problem 11151. Amer. Math. Monthly 112 (2005), 367. Solution 114 (2007), 265–266. [81]
Knuth D.E., Problem 11274. Amer. Math. Monthly 114 (2007), 165. Solution 116 (2009), 548–549.
[35, 116]
Knuth D.E., Problem 11452. Amer. Math. Monthly 116 (2009), 648. Solution 118 (2011), 657. [65]
Knuth D.E., Problem 11985. Amer. Math. Monthly 124 (2017), 275. Solution 126 (2019), March. [567]
Knuth D.E. and T.J. Buckholtz, Computation of tangent, Euler, and Bernoulli numbers. Math. Comp.
21 (1967), 663–688. [65]
Knuth D.E. and J. McCarthy, Problem E3429. Amer. Math. Monthly 98 (1991), 264. Solution 99 (1992),
684. [667]
Koebe P., Contaktprobleme der konformen Abbildung (German). Berichte Über die Verhandlungen der
Sächsischen Akad. Wissen. Leipzig, Math.-Phys. Klasse 88 (1936), 141–164. [394]
Kohayakawa Y., Szemerédi’s regularity lemma for sparse graphs. In Foundations of computational math-
ematics (Rio de Janeiro, 1997) (Springer, 1997), 216–230. [488]
Kohayakawa Y. and V. Rödl, Szemerédi’s regularity lemma and quasi-randomness. In Recent advances
in algorithms and combinatorics, CMS Books Math./Ouvrages Math. SMC 11 (Springer, 2003),
289–351. [488]
Kohayakawa Y., V. Rödl, M. Schacht, and E. Szemerédi, Sparse partition universal graphs for graphs of
bounded degree. Adv. Math. 226 (2011), 5041–5065. [455]
Köhler E., über das Oberwolfacher Problem. In Beiträge zur geometrischen Algebra (Proc. Sympos.,
Duisburg, 1976), Lehrbücher Monograph. Geb. Exakten Wissensch., Math. Reihe 21 (Birkhäuser,
Basel, 1977), 189–201. [647]
References 895

Kollár J., L. Rónyai, and T. Szabó, Norm-graphs and bipartite Turán numbers. Combinatorica 16 (1996),
399–406. [633, 672]
Komlós J., A strange pigeonhole principle. Order 7 (1990), 107–113. [823]
Komlós J., G.N. Sárközy, and E. Szemerédi, Blow-up lemma. Combinatorica 17(1997),109–123. [481]
Komlós J., A. Shokoufandeh, M. Simonovits, and E. Szemerédi, The regularity lemma and its applica-
tions in graph theory. In Theoretical aspects of computer science (Tehran, 2000), Lect. Notes Comp.
Sci. 2292 (Springer, 2002), 84–112. [488]
Komlós J. and M. Simonovits, Szemerédi’s regularity lemma and its applications in graph theory. In
Combinatorics, Paul Erdős is eighty, Vol. 2 (Keszthely, 1993), Bolyai Soc. Math. Stud. 2 (János
Bolyai Math. Soc., 1996), 295–352. [488, 492]
Komlós J. and E. Szemerédi, Topological cliques in graphs II. Combin. Probab. Comput. 5 (1996), 79–
90. [351]
Komm H., On the dimension of partially ordered sets. Amer. J. Math. 70 (1948), 507–520. [571]
Konheim A.G. and B. Weiss, An occupancy discipline and applications. SIAM J. Applied Math. 14 (1966),
1266–1274. [50]
König D., Über Graphen und ihre Anwendung auf Determinantentheorie und Mengenlehre. Math. Ann.
77 (1916), 453–465. [256, 259, 359, 371]
König D., Graphen und Matrizen. Mat. Lapok 38 (1931), 116–119. [258, 530]
König D., Theorie der endlichen und unendlichen Graphen (Akademische Verlagsgesellschaft, 1936).
Also Chelsea, 1950; Teubner, 1986. [763]
Kopylov G.N., Maximal paths and cycles in a graph. Dokl. Akad. Nauk SSSR 234 (1977), 19–21. [323]
Korshunov A.D., The number of monotone Boolean functions. Problemy Kibernet. (1981), 5–108, 272.
[558]
Koshy T., Fibonacci and Lucas numbers with applications, Pure and Applied Mathematics (New York)
(Wiley-Interscience, 2001). [53]
Kostochka A.V., The total coloring of a multigraph with maximal degree 4. Discr. Math. 17 (1977), 161–
163. [373]
Kostochka A.V., The minimum Hadwiger number for graphs with a given mean degree of vertices.
Metody Diskret. Analiz. (1982), 37–58. [710]
Kostochka A.V., The total chromatic number of any multigraph with maximum degree five is at most
seven. Discr. Math. 162 (1996), 199–214. [373]
Kostochka A.V., Extremal problems on -systems. In Numbers, information and complexity (Bielefeld,
1998) (Kluwer Acad. Publ., 2000), 143–150. [504]
Kostochka A.V., A new tool for proving Vizing’s theorem. Discr. Math. 326 (2014), 1–3. [360]
Kostochka A.V., M.J. Pelsmajer, and D.B. West, A list analogue of equitable coloring. J. Graph Theory
44 (2003), 166–177. [374]
Kostochka A.V. and V. Rödl, On graphs with small Ramsey numbers. J. Graph Th. 37 (2001), 198–204.
[484]
Kostochka A.V. and V. Rödl, On graphs with small Ramsey numbers. II. Combinatorica 24 (2004), 389–
401. [672]
Kostochka A.V. and N. Sheikh, On the induced Ramsey number IR(P3 , H). In Topics in discrete mathe-
matics, Algorithms Combin. 26 (Springer, 2006), 155–167. [460]
Kostochka A.V. and D.B. West, Chvátal’s condition cannot hold for both a graph and its complement.
Discuss. Math. Graph Theory 26 (2006), 73–76. [332, 705]
Kostochka A.V. and M. Yancey, Ore’s conjecture for ¾ = 4 and Grötzsch’s theorem. Combinatorica 34
(2014), 323–329. Generalized in “Ore’s conjecture on color-critical graphs is almost true,” J. Com-
bin. Th. B (2014). [355, 414, 420]
Kotzig A., Contribution to the theory of Eulerian polyhedra. Mat.-Fyz. Časopis. Slovensk. Akad. Vied 5
(1955), 101–113. [411]
Kotzig A., Aus der Theorie der endlichen regulären Graphen dritten und vierten Grades. Časopis Pěst.
Mat. 82 (1957), 76–92. [273]
896 References

Kotzig A., On the theory of finite graphs with a linear factor. I, II, III. Mat.-Fyz. Časopis. Slovensk. Akad.
Vied. 9/10 (1959), 73–91, 136–159; 205–215. [264, 276]
Kotzig A., On even regular graphs of the third degree. Mat.-Fyz. C̆asopis Sloven. Akad. Vied 16 (1966a),
72–75. [227]
Kotzig A., 1-factorizations of Cartesian products of regular graphs. J. Graph Th. 3(1979),23–34. [371]
Kotzig A. and A. Rosa, Nearly Kirkman systems. In Proc. 5th Southeastern Intl. Conf. Combin. Graph
Th. Comput. (Boca Raton), Congr. Numer. 10 (1974), 607–614. [647]
Kouider M. and P.D. Vestergaard, Connected factors in graphs—a survey. Graphs Combin. 21 (2005),
1–26. [327]
Kouril M., Computing the van der Waerden number W(3 , 4) = 293. Integers 12 (2012), Paper No. A46,
13. [465]
Kouril M. and J.L. Paul, The van der Waerden number W(2 , 6) is 1132. Experiment. Math. 17 (2008),
53–61. [465]
Kovář P., M. Kubesa, and M. Meszka, Factorizations of complete graphs into brooms. Discr. Math. 312
(2012), 1084–1093. [239]
Kövari T., V.T. Sós, and P. Turán, On a problem of K. Zarankiewicz. Colloquium Math. 3 (1954), 50–57.
[631, 672]
Kraitchik M., The Gambler’s Ruin. Mathematical Recreations (W. W. Norton, 1942), 140. [66]
Kratochvı́ l J., Z. Tuza, and M. Voigt, Brooks-type theorems for choosability with separation. J. Graph
Theory 27 (1998), 43–49. [403]
Krattenthaler C., Bijective proofs of the hook formulas for the number of standard Young tableaux,
ordinary and shifted. Electron. J. Combin. 2 (1995), Research Paper 13. [192]
Krattenthaler C., The enumeration of lattice paths with respect to their number of turns. In Ad-
vances in combinatorial methods and applications to probability and statistics, Stat. Ind. Tech-
nol. (Birkhäuser Boston, 1997), 29–58. [50]
Krattenthaler C., Advanced determinant calculus. Sém. Lothar. Combin. 42 (1999), Art. B42q, 67 pp.
(electronic). The Andrews Festschrift (Maratea, 1998). [170]
Kratzke T., B. Reznick, and D.B. West, Eigensharp graphs: decomposition into complete bipartite sub-
graphs. Trans. Amer. Math. Soc. 308 (1988), 637–653. [771]
Kriz I., A hypergraph-free construction of highly chromatic graphs without short cycles. Combinatorica
9 (1989), 227–229. [341, 434]
Król M., On a sufficient and necessary condition of 3-colorableness for the planar graphs. I, II. Prace
Nauk. Inst. Mat. Fiz. Teoret. Politechn. Wrocław. Ser. Studia i Materiały (1972), 37–40. [419]
Kruskal J.B., On the shortest spanning subtree of a graph and the traveling salesman problem. Proc.
Am. Math. Soc. 7 (1956), 48–50. [245]
Kruskal J.B., The number of simplices in a complex. In Mathematical optimization techniques (Univ.
California Press, 1963), 251–278. [495, 496]
Kruskal J.B., Jr., Monotonic subsequences. Proc. Amer. Math. Soc. 4 (1953), 264–274. [431]
Kubale M. (ed.), Graph colorings, Contemporary Mathematics 352 (AMS, 2004). [335]
Kuczma M.S., A multi-well problem for phase transformations. In The mathematics of finite elements
and applications, X, MAFELAP 1999 (Uxbridge) (Elsevier, 2000), 271–282. [24]
Kühn D. and D. Osthus, Embedding large subgraphs into dense graphs. In Surveys in combinatorics
2009, London Math. Soc. Lecture Note Ser. 365 (Cambridge Univ. Press, 2009), 137–167. [488]
Kühn D. and D. Osthus, Hamilton cycles in graphs and hypergraphs: an extremal perspective. In Proc.
of the Int. Congr. Math. IV (Seoul, 2014) (Kyung Moon Sa, 2014), 381–406. [316]
Kuhn H.W., The Hungarian method for the assignment problem. Naval Research Logistics Quarterly 2
(1955), 83–97. [280]
Kündgen A., Art galleries with interior walls. Discrete Comput. Geom. 22 (1999), 249–258. [419]
Kündgen A. and R. Ramamurthi, Coloring face-hypergraphs of graphs on surfaces. J. Combin. Th. B 85
(2002), 307–337. [287]
Kuo E.H., Applications of graphical condensation for enumerating matchings and tilings. Theoret. Com-
put. Sci. 319 (2004), 29–57. [34]
References 897

Kuperberg G., Symmetries of plane partitions and the permanent-determinant method. J. Combin. Th.
A 68 (1994), 115–151. [756]
Kuperberg G., An exploration of the permanent-determinant method. Electron. J. Combin. 5 (1998),
Research Paper 46. [756]
Kuperberg G., Kasteleyn cokernels. Electron. J. Combin. 9 (2002), Research Paper 29, 30. [34]
Kupka J., Problem E3402. Amer. Math. Monthly 97 (1990), 612. Solution 99 (1992), 367. [49]
Kuratowski K., Sur le problème des courbes gauches en topologie. Fund. Math. 15 (1930), 271–283.
[390]
Labelle G., Une nouvelle démonstration combinatoire des formules d’inversion de Lagrange. Adv. in
Math. 42 (1981), 217–247. [129]
Lagrange J.L., Nouvelle méthode pour résoudre des équations littérales par le moyen des séries. Mém.
Acad. Roy. des Sci. et Belles-Lettres de Berlin 24 (1770). [129]
Lam C.W.H., L. Thiel, and S. Swiercz, The nonexistence of finite projective planes of order 10. Canad.
J. Math. 41 (1989), 1117–1123. [612]
Lamé G., Note sur la limite du nombre des divisions dans la recherche du plus grand commun diviseur
entre deux nombres entiers. C. R. Acad. Sci. Paris 19 (1844), 867–870. [67]
Landau E., Handbuch der Lehre von der Verteilung der Primzahlen (German) (Teubner, 1909). [10]
Landau H.G., On dominance relations and the structure of animal societies, III: The condition for score
structure. Bull. Math. Biophys. 15 (1953), 143–148. [225, 228, 261]
Lander E.S., Symmetric designs and self-dual codes. J. London Math. Soc. (2) 24 (1981), 193–204.
[623]
Lander E.S., Symmetric designs: an algebraic approach, London Mathematical Society Lecture Note Se-
ries 74 (Cambridge Univ. Press, 1983). [637]
Landman B.M. and A. Robertson, Ramsey theory on the integers (2nd ed.), Student Mathematical Library
73 (AMS, 2014). [425]
Larman D.G., J. Matoušek, J. Pach, and J. Törőcsik, A Ramsey-type result for convex sets. Bull. London
Math. Soc. 26 (1994), 132–136. [461]
Larman D.G. and C.A. Rogers, The realization of distances within sets in Euclidean space. Mathematika
19 (1972), 1–24. [729]
Larman D.G., C.A. Rogers, and J.J. Seidel, On two-distance sets in Euclidean space. Bull. London Math.
Soc. 9 (1977), 261–267. [724]
Las Vergnas M., Sur l’existence de cycles hamiltoniens dans un graphe. C. R. Acad. Sci. Paris Sér. A-B
270 (1970), A1361–A1364. [333]
Las Vergnas M., A note on matchings in graphs. Cahiers Centre Études Recherche Opér. 17 (1975), 257–
260. [272]
Lassak M., An estimate concerning Borsuk partition problem. Bull. Acad. Polon. Sci. Sér. Sci. Math. 30
(1982), 449–451 (1983). [817]
Lavallée I., Note sur le problème des Tours de Hanoı̈. Acta Math. Vietnam. 7 (1982), 131–137 (1984).
[80]
Lawler E.L., Combinatorial Optimization: Networks and Matroids (Holt, Rinehart, and Winston, 1976).
[282, 532]
Lawrence J., Covering the vertex set of a graph with subgraphs of smaller degree. Discr. Math. 21 (1978),
61–68. [355]
Lawrence S.L., Cycle-star Ramsey numbers. Notices Amer. Math. Soc. 20 (1973), A–420 (Notice #73T–
157). [460]
Laywine C.F. and G.L. Mullen, Discrete mathematics using Latin squares, Wiley-Interscience Series in
Discr. Mathematics and Optimization (Wiley & Sons, 1998). [623, 656]
Lazarson T., Independence functions in algebra (U. London, 1957). Thesis. [537]
Lebesgue H., Quelques conséquences simples de la formule d’Euler. J. Math. Pures Appl. 19 (1940), 27–
43. [422]
Lederberg J., Systematics of organic molecules, graph topology and Hamiltonian circuits (Instrumenta-
tion Res. Lab. Rept.). Tech. Rep. 1040, Stanford Univ. (1966). [418]
898 References

Lee C., Lecture notes for Extremal Combinatorics (Spring 2015 course at mit) (2015). Available online
at http://math.mit.edu/˜cb_lee/18.318/materials.html. [488, 672]
Lee C., Ramsey numbers of degenerate graphs (2017). (arXiv:1505.04773). [484]
Lefmann H., V. Rödl, and R. Thomas, Monochromatic vs multicolored paths. Graphs Combin. 8 (1992),
323–332. [471]
Lefschetz S., Introduction to Topology, Princeton Mathematical Series, vol. 11 (Princeton Univ. Press,
1949). [807]
Lehman A., A solution of the Shannon switching game. J. Soc. Indust. Appl. Math. 12 (1964), 687–725.
[528, 539]
Lehmer D.H., Lacunary recurrence formulas for the numbers of Bernoulli and Euler. Ann. of Math. (2)
36 (1935), 637–649. [65]
Leighton F.T., Complexity Issues in VLSI: optimal layouts for the shuffle-exchange graph and other net-
works, Foundations of Computing (MIT Press, 1983). [793]
Lekkerkerker C.G., Representation of natural numbers as a sum of Fibonacci numbers. Simon Stevin
29 (1952), 190–195. [61]
LeSaulnier T., Finding a repeated difference (solution to problem 11084). Amer. Math. Monthly 113
(2006), 371–372. Proposed 111 (2004), 440. [441]
Lesniak L.M., Hamiltonicity in some special classes of graphs. Congr. Numer. 116 (1996), 53–70.
[316]
Leuck D.H., Solution to problem E3057. Amer. Math. Monthly 94 (1987), 187–188. Proposed 91 (1984),
515. [175]
Li H., Generalizations of Dirac’s theorem in Hamiltonian graph theory—a survey. Discr. Math. 313
(2013), 2034–2053. [316]
Li Q., M.Q. Tao, and Y.Q. Shen, The bandwidth of the discrete tori Cm × Cn. J. China Univ. Sci. Tech.
11 (1981), 1–16. [801, 819]
Li S.Y.R. and W.C.W. Li, Independence numbers of graphs and generators of ideals. Combinatorica 1
(1981), 55–61. [740]
Li X.L., The connectivity of path graphs. In Combinatorics, graph theory, algorithms and applications
(Beijing, 1993) (World Sci. Publ., 1994), 187–192. [312]
Lick D.R., Characterizations of n-connected and n-line-connected graphs. J. Combin. Th. B 14 (1973),
122–124. [312]
Liggett T.M., Extensions of the Erdős–Ko–Rado theorem and a statistical application. J. Combinatorial
Th. A 23 (1977), 15–21. [511]
Lih K.W., C.Y. Lin, and L.D. Tong, On an interpolation property of outerplanar graphs. Discrete Appl.
Math. 154 (2006), 166–172. [287]
Lindner C.C. and C.A. Rodger, Design Theory (CRC Press, 1997). Also 2009. [609, 641]
Lindström B., On a combinatorial problem in number theory. Canad. Math. Bull. 8 (1965), 477–490.
[513]
Lindström B., On the vector representations of induced matroids. Bull. London Math. Soc. 5 (1973), 85–
90. [165, 167, 169]
Lindström B., A partition of L(3 , n) into saturated symmetric chains. European J. Combin. 1 (1980),
61–63. [557]
Linial N., A lower bound for the circumference of a graph. Discr. Math. 15 (1976), 297–300. [323]
Linial N., Extending the Greene–Kleitman theorem to directed graphs. J. Combin. Th. A 30 (1981), 331–
334. [549]
Linial N., A new derivation of the counting formula for Young tableaux. J. Combin. Theory A 33 (1982),
340–342. [207]
Linial N., The information-theoretic bound is good for merging. SIAM J. Comput. 13 (1984), 795–801.
[823]
Liouville B., Sur la connectivité des produits de graphes. C. R. Acad. Sci. Paris Sér. A-B 286 (1978),
A363–A365. [292]
Lipski W., Jr., On strings containing all subsets as substrings. Discr. Math. 21 (1978), 253–259. [566]
References 899

Lipton R.J. and R.E. Tarjan, A separator theorem for planar graphs. SIAM J. Appl. Math. 36 (1979),
177–189. [394–396, 399]
Lipton R.J. and R.E. Tarjan, Applications of a planar separator theorem. SIAM J. Comput. 9 (1980),
615–627. [399]
Littlewood D.E., The Theory of Group Characters and Matrix Representations of Groups (Oxford Univ.
Press, 1940). Also 1950. [207]
Liu B., The theorem on partition of connected graph and its applications in graphical enumeration. J.
South China Normal Univ. (Natural Science) 1 (1985), 51–56. [236]
Liu C.L., Topics in combinatorial mathematics (MAA, 1972). MAA Summer Seminar lect. notes. [17]
Liu H., R. Morris, and N. Prince, Highly connected monochromatic subgraphs of multicolored graphs.
J. Graph Th. 61 (2009), 22–44. [460]
Liu J. and H. Zhou, Maximum induced matchings in graphs. Discr. Math. 170 (1997), 277–281. [262]
Locke S.C., Problem 10447. Amer. Math. Monthly 102 (1995), 360. Solution 104 (1997), 976. [229]
Locke S.C., Problem 10892. Amer. Math. Monthly 108 (2001), 668. Solution 110 (2003), 440–1. [330]
Logan B.F. and L.A. Shepp, A variational problem for random Young tableaux. Advances in Math. 26
(1977), 206–222. [722, 821]
Long C.T., On the Moessner theorem on integral powers. Am. Math. Monthly 73 (1966), 846–851. [62]
Long C.T., Strike it out: Add it up. Math. Gazette 66 (1982), 273–277. [62]
Long C.T., A note on Moessner’s process. Fibonacci Quart. 24 (1986), 349–355. [62]
Longyear J.Q. and T.D. Parsons, The friendship theorem. Nederl. Akad. Wetensch. Proc. A 75=Indag.
Math. 34 (1972), 257–262. [776]
Loomis L.H. and H. Whitney, An inequality related to the isoperimetric inequality. Bull. Amer. Math.
Soc 55 (1949), 961–962. [513]
Lossers O.P. and R.S. Pinkham, Noncrossing trees, solution I to problem e3170. Amer. Math. Monthly
96 (1989), 359–361. Proposed by Howard University Group, 93 (1986), 650. [129]
Lovász L., On graphs not containing independent circuits. Mat. Lapok 16 (1965), 289–299. [272]
Lovász L., On decomposition of graphs. Studia Sci. Math. Hungar. 1 (1966), 237–238. [249, 262, 355]
Lovász L., On covering of graphs. In Theory of Graphs (Tihany, 1966) (P. Erdős and G. Katona, eds.)
(Academic Press, 1968a), 231–236. [232]
Lovász L., On chromatic number of finite set-systems. Acta Math. Acad. Sci. Hung. 19 (1968b), 59–67.
[434, 436]
Lovász L., Normal hypergraphs and the perfect graph conjecture. Discr. Math. 2 (1972a), 253–267.
[368, 369, 375, 376]
Lovász L., A characterization of perfect graphs. J. Combin. Th. B 13 (1972b), 95–98. [369]
Lovász L., On the structure of factorizable graphs. Acta Math. Hungarica 23 (1972c), 179–195. [276]
Lovász L., On the sieve formula. Mat. Lapok 23 (1972d), 53–69 (1973). [477]
Lovász L., Problem 5. Period. Math. Hungar. 4 (1974), 82. [308, 310]
Lovász L., Three short proofs in graph theory. J. Combin. Th. B 19 (1975), 269–271. [263, 264]
Lovász L., A homology theory for spanning trees of a graph. Acta Math. Acad. Sci. Hungar. 30 (1977),
241–251. [312]
Lovász L., Kneser’s conjecture, chromatic number, and homotopy. J. Combin. Th. A 25 (1978), 319–324.
[342, 532, 812]
Lovász L., Combinatorial problems and exercises (North-Holland, 1979).
[47, 104, 495, 496, 510, 762, 764]
Lovász L., Combinatorial problems and exercises, 2nd ed. (North-Holland, 1993). [419, 491]
Lovász L., J. Nešetřil, and A. Pultr, On a product dimension of graphs. J. Combin. Th. B 28 (1980), 47–
67. [511, 726]
Lovász L. and M.D. Plummer, Matching Theory, Ann. Discrete Math. 29 (Akademiai Kiado and North-
Holland, 1986). [262, 275, 756, 764]
Lovász L. and M. Simonovits, On the number of complete subgraphs of a graph. In Proc. 5th British
Combin. Conf. (Univ. Aberdeen, 1975), 15 (Utilitas Math., 1976), 431–441. [478]
900 References

Lovász L. and M. Simonovits, On the number of complete subgraphs of a graph. II. In Studies in pure
mathematics (Birkhäuser, 1983), 459–495. [478]
Lovász L. and P. Winkler, A note on the last new vertex visited by a random walk. J. Graph Theory 17
(1993), 593–596. [668]
Lu X., A Chvátal–Erdős type condition for Hamiltonian graphs. J. Graph Th. 18 (1994), 791–800.
[322]
Lu X., D.W. Wang, and C.K. Wong, On avoidable and unavoidable claws. Discr. Math. 184 (1998), 259–
265. [566]
Lubell D., A short proof of Sperner’s lemma. J. Combinatorial Theory 1 (1966), 299. [558]
Lubell D., Problem 10992. Amer. Math. Monthly 110 (2003), 155. Solution 111 (2004), 827–829. [34]
Lubotzky A., R. Phillips, and P. Sarnak, Explicit expanders and the Ramanujan conjectures. In Proc.
18th ACM Symp. Th. Comp. (Assoc. Comput. Mach., 1986), 240–246. [780]
Lubotzky A., R. Phillips, and P. Sarnak, Ramanujan graphs. Combinatorica 8 (1988), 261–277.
[341, 434]
Lucas F.E.A., Sur les congruences des nombres eulériens et les coefficients différentiels des functions
trigonométriques suivant un module premier. Bull. Soc. Math. France 6 (1878), 49–54. [47]
Lucas F.E.A., Note sur les intersections de trois quadriques. Bull. Soc. Math. France 19 (1891), 118–119.
[162]
Lucas F.E.A., Sur les polygones inscrits dans les coniques. Bull. Soc. Math. France 20 (1892), 33–34.
[238]
Luce R.D., Semiorders and a theory of utility discrimination. Econometrica 24 (1956), 178–191. [585]
Luo R. and C.Q. Zhang, Edge coloring of graphs with small average degrees. Discr. Math. 275 (2004),
207–218. [423]
Lützen J., G. Sabidussi, and B. Toft, Julius Petersen 1839–1910. A biography. Discr. Math. 100 (1992),
5–82. [740]
Lyusternik L.A. and L. Shnirel’man, Topological Methods in the Calculus of Variations (Russian)
(Moscow, 1930). [810]
Ma M. and D.B. West, Problem 11731. Amer. Math. Monthly 120 (2013), 755. Solution 122 (2015), 807.
[371]
Mabry R., Bipartite graphs and the Four-color Theorem. Bull. ICA 14 (1995), 119–112. [417]
MacLane S., Some interpretations of abstract linear dependence in terms of projective geometry. Amer.
J. Math. 58 (1936), 236–240. [514, 529]
MacMahon P.A., Combinatory analysis, I and II (Cambridge Univ. Press, 1916). Reprinted Chelsea, 1960;
Dover, 2004. [45, 149, 169, 178, 207]
MacNeish H.F., Euler squares. Ann. of Math. (2) 23 (1922), 221–227. [611]
Maddox R., The superregular graphs (solution to Problem 6617). Amer. Math. Monthly 103 (1996), 600–
603. Proposed 96 (1989), 942. [782]
Madej T. and D.B. West, The interval inclusion number of a partially ordered set. Discr. Math. 88 (1991),
259–277. [832]
Mader W., Homomorphieeigenschaften und mittlere Kantendichte von Graphen. Math. Ann. 174 (1967),
265–268. [350]
Mader W., Homomorphiesätze für Graphen. Math. Ann. 178 (1968), 154–168. [710]
Mader W., Minimale n-fach kantenzusammenhängende Graphen. Math. Ann. 191 (1971), 21–28.
[309]
Mader W., Existenz n-fach zusammenhängender Teilgraphen in Graphen genügend grosser Kanten-
dichte. Abh. Math. Sem. Univ. Hamburg 37 (1972), 86–97. [296, 310, 311, 315, 316]
Mader W., 1-Faktoren von Graphen. Math. Ann. 201 (1973), 269–282. [264]
Mader W., A reduction method for edge-connectivity in graphs. Ann. Discr. Math. 3 (1978), 145–164.
[308, 310]
Mader W., Zur Struktur minimal n-fach zusammenhängender Graphen. Abh. Math. Sem. Univ. Ham-
burg 49 (1979), 49–69. [311, 316]
References 901

Mader W., On vertices of degree n in minimally n-connected graphs and digraphs. In Combinatorics,
Paul Erdős is eighty (Keszthely, 1993), Bolyai Soc. Math. Stud. (János Bolyai Math. Soc., 1996),
423–449. [311]
Mader W., 3n − 5 edges do force a subdivision of K 5 . Combinatorica 18 (1998), 569–595. [351, 353]
Mahmoodian S.E., On edge-colorability of Cartesian products of graphs. Canad. Math. Bull. 24 (1981),
107–108. [371]
Maillet E., Contributions à la théorie des groupes. TIM (9) 6 (1894), 258–280. [623]
Majumdar K.N., On some theorems in combinatorics relating to incomplete block designs. Ann. Math.
Statistics 24 (1953), 377–389. [730]
Mann H.B., On orthogonal Latin squares. Bull. Amer. Math. Soc. 50 (1950), 249–257. [623]
Mann H.B. and H.J. Ryser, Systems of distinct representatives. Amer. Math. Monthly 60 (1953), 397–
401. [254]
Mantel W., Problem 28, soln. by H. Gouwentak, W. Mantel, J. Teixeira de Mattes, F. Schuh and W.A.
Wythoff. Wiskundige Opgaven 10 (1907), 60–61. [224, 475]
Marcus A. and G. Tardos, Excluded permutation matrices and the Stanley–Wilf conjecture. J. Combin.
Th. A 107 (2004), 153–160. [432, 433]
Marcus M. and M. Newman, On the minimum of the permanent of a doubly stochastic matrix. Duke
Math. J. 26 (1959), 61–72. [765]
Marcus M. and R. Ree, Diagonals of doubly stochastic matrices. Quart. J. Math. 2 (1959), 295–302.
[257]
Margulis G.A., Explicit constructions of concentrators. Problems of Information Transmission 9 (1973),
325–332. [779]
Margulis G.A., Explicit group-theoretic constructions of combinatorial schemes and their applications
in the construction of expanders and concentrators (Russian). Problems of Information Trans-
mission 24 (1988), 39–46. [780]
Marica J., Orthogonal families of sets. Canad. Math. Bull. 14 (1971), 573. [512]
Martin M.H., A problem in arrangements. Bull. Amer. Math. Soc. 40 (1934), 859–864. [763]
Martin R., Notes on Extremal Graph Theory (2012), 101. Available online at
http://orion.math.iastate.edu/rymartin/ISU608EGT/EGTbook.pdf. [488]
Martinov N., Uncontractible 4-connected graphs. J. Graph Th. 6 (1982), 343–344. [313]
Massey J.L., On the fractional weight of distinct binary n-tuples. IEEE Trans. Information Theory IT-
20 (1974), 131. [507]
Matoušek J., Using the Borsuk–Ulam theorem, Universitext (Springer-Verlag, 2003). [806, 810, 812]
Matoušek J., A combinatorial proof of Kneser’s conjecture. Combinatorica 24 (2004), 163–170. [812]
Matoušek J., Thirty-three miniatures, Student Mathematical Library 53 (AMS, 2010). [723]
Matoušek J. and G.M. Ziegler, Topological lower bounds for the chromatic number: a hierarchy. Jahres-
ber. Deutsch. Math.-Verein. 106 (2004), 71–90. [812, 813]
Matsko V.J., D.B. West, and J.E. Wetzel, Trifold arrangements and cevian dissections. J. Geom. 72
(2001), 115–127. [565]
Matthews M.M. and D.P. Sumner, Hamiltonian results in K1 ,3 -free graphs. J. Graph Th. 8 (1984), 139–
146. [326, 330]
Matula D.W., The cohesive strength of graphs. In The Many Facets of Graph Theory (Proc. Conf., Western
Mich. Univ., Kalamazoo, Mich., 1968) (Springer, 1969), 215–221. [353]
Matula D.W., The employee party problem. Notices Amer. Math. Soc. 19 (1972), A–382. [700]
Matula D.W., An extension of Brooks’ Theorem. Tech. Rep. 69, Center for Numerical Analysis, University
of Texas–Austin (1973). [354]
Matula D.W., Ramsey theory for graph connectivity. J. Graph Theory 7 (1983), 95–103. [297]
Matula D.W. and L. Kučera, An expose-and-merge algorithm and the chromatic number of a random
graph. In Random graphs ’87 (Poznań, 1987) (Wiley, 1990), 175–187. [718]
Maunsell F.G., A note on Tutte’s paper “The factorization of linear graphs.”. J. London Math. Soc. 27
(1952), 127–128. [264]
Maurer R., E2404. Amer. Math. Monthly 80 (1973), 316. Proposed 81 (1974), 287. [106]
902 References

Maurer S., The king chicken theorems. Math. Mag. 53 (1980), 67–80. [228]
Maurer S. and I. Rabinovitch, Large minimal realizers of a partial order. Proc. Amer. Math. Soc. 66
(1977), 211–216. [584]
Maurer S., I. Rabinovitch, and W.T. Trotter, Large minimal realizers of a partial order II. Discr. Math.
31 (1980), 297–314. [228]
McConnell R.M. and J.P. Spinrad, Modular decomposition and transitive orientation. Discr. Math. 201
(1999), 189–241. [573]
McDiarmid C.J.H., The solution of a timetabling problem. J. Inst. Math. Appl. 9 (1972), 23–34. [372]
McDiarmid C.J.H., Concentration. In Probabilistic methods for algorithmic discrete mathematics, Algo-
rithms Combin. 16 (Springer, 1998), 195–248. [706, 719, 721]
McFarland R. and H.B. Mann, On multipliers of difference sets. Canad. J. Math. 17 (1965), 541–542.
[637]
McKay B.D., Asymptotics for symmetric 0-1 matrices with prescribed row sums. Ars Combin. 19 (1985),
15–25. [697]
McKay B.D. and S.P. Radziszowski, The first classical Ramsey number for hypergraphs is computed. In
2nd Symp. Disc. Alg. (San Francisco), ACM-SIAM (1991), 304–308. [449]
McKay B.D. and S.P. Radziszowski, R(4 , 5) = 25. J. Graph Th. 19 (1995), 309–322. [449]
McKay B.D. and N.C. Wormald, Asymptotic enumeration by degree sequence of graphs with degrees
o(n1/2). Combinatorica 11 (1991), 369–382. [697]
McKay B.D. and K.M. Zhang, The value of the Ramsey number R(3 , 8). J. Graph Th. 16 (1992), 99–105.
[449]
McKee T.A., Recharacterizing Eulerian: intimations of new duality. Discr. Math. 51 (1984), 327–242.
[239]
McKee T.A., How chordal graphs work. Bull. ICA 9 (1993), 27–39. [374]
McSorley J.P., Counting structures in the Möbius ladder. Discr. Math. 184 (1998), 137–164. [765]
Melham R.S. and A.G. Shannon, A generalization of the Catalan identity and some consequences. Fi-
bonacci Quart. 33 (1995), 82–84. [60]
Menger K., Zur allgemeinen Kurventheorie. Fund. Math. 10 (1927), 95–115. [298, 300]
Merca M., Problem 11767. Amer. Math. Monthly 121 (2014a), 267. Solution 123 (2016), 505–506. [150]
Merca M., Problem 11772. Amer. Math. Monthly 121 (2014b), 366. Solution 123 (2016), 614. [148]
Merris R., Laplacian matrices of graphs: a survey. 197/198 (1994), 143–176. [776]
Merris R., A survey of graph Laplacians. Linear and Multilinear Algebra 39 (1995), 19–31. [776]
Merris R., Laplacian graph eigenvectors. Linear Algebra Appl. 278 (1998), 221–236. [782]
Meshalkin L.D., Generalization of Sperner’s theorem on the number of subsets of a finite set (Russian).
Th. Prob. Appl. 8 (1963), 203–204. [558]
Meyniel H., On the perfect graph conjecture. Discr. Math. 16 (1976), 339–342. [375]
Miao L. and Q. Sun, On the size of critical graphs with maximum degree 8. Discr. Math. 310 (2010),
2215–2218. [414]
√
Micali S. and V.V. Vazirani, An O( | V |·| E|) algorithm for finding maximum matching in general graphs.
In Proc. 21th IEEE Symp. Found. Comp. Sci. (Assoc. Comput. Mach., 1980), 17–27. [284]
Mihók P., On vertex partition numbers of graphs. In Graphs and Other Combin. Topics (Prague, 1982),
Teubner-Texte Math. 59 (Teubner, 1983), 183–8. [418]
Mikola M., The Lucas number as the number of spanning trees. Práce Štúd. Vysokej Školy Doprav. Žiline
Sér. Mat.-Fyz. 2 (1980), 69–77. [65]
Milans K.G., D. Stolee, and D.B. West, Ordered Ramsey theory and track representations of graphs. J.
Comb. 6 (2015), 445–456. [602, 608]
Miller E., Planar graphs as minimal resolutions of trivariate monomial ideals. Doc. Math. 7 (2002), 43–
90. [784]
Miller Z., The bandwidth of caterpillar graphs. In Proc. 12th Southeastern Intl. Conf. Combin. Graph
Th. Comput. (Baton Rouge), Congr. Numer. 33 (1981), 235–252. [819]
Mills G., A quintessential proof of van der Waerden’s theorem on arithmetic progressions. Discr. Math.
47 (1983), 117–120. [463]
References 903

Minty G.J., A theorem on n-coloring the points of a linear graph. Amer. Math. Monthly 69 (1962), 623–
624. [338]
Minty G.J., On the axiomatic foundations of the theories of directed linear graphs, electrical networks
and network-programming. J. Math. Mech. 15 (1966), 485–520. [539]
Mirkin B.G., Description of some relations on the set of real-line intervals. J. Mathematical Psychology
9 (1972), 243–252. [586]
Mirsky L., Transversal theory. Mathematics in Science and Engineering 75 (Academic Press, 1971).
[547]
Mirsky L. and H. Perfect, Applications of the notion of independence to combinatorial analysis. J. Com-
bin. Th. 2 (1967), 327–357. [517]
Mirzakhani M., A small non-4-choosable planar graph. Bull. Inst. Combin. Appl. 17 (1996), 15–18.
[402, 421]
Mitas J., Minimal representation of semiorders with intervals of same length. In Orders, algorithms,
and applications (Lyon, 1994), Lect. Notes Comp. Sci. 831 (Springer, 1994), 162–175. [587]
Mitchem J. and E.F. Schmeichel, Pancyclic and bipancyclic graphs—a survey. In Graphs and applica-
tions (Boulder, Colo., 1982), Wiley-Intersci. Publ. (Wiley, 1985), 271–278. [326]
Mitzenmacher M. and E. Upfal, Probability and computing (Cambridge University Press, Cambridge,
2017), second edn. [708]
Moessner A., Eine Bemerkung über die Potenzen der natürlichen Zahlen. S.-B. Math.-Nat. Kl. Bayer.
Akad. Wiss. 1951 (1951), 29 (1952). [62]
Moews D., Optimally pebbling hypercubes and powers. Discr. Math. 190 (1998), 271–276. [668]
Mohanty S.G., Lattice path counting and applications (Academic Press, 1979). [50]
Mohar B., The Laplacian spectrum of graphs. In Graph theory, combinatorics, and applications. Vol. 2
(Kalamazoo, MI, 1988), Wiley-Intersci. Publ. (Wiley, 1991), 871–898. [776]
Mohar B., Laplace eigenvalues of graphs—a survey. 109 (1992), 171–183. [776]
Mohar B. and S. Poljak, Eigenvalues in combinatorial optimization. In Combinatorial and graph-
theoretical problems in linear algebra (Minneapolis, MN, 1991), IMA Vol. Math. Appl. 50
(Springer, 1993), 107–151. [776]
Mohar B. and C. Thomassen, Graphs on surfaces, Johns Hopkins Studies in the Mathematical Sciences
(Johns Hopkins Univ. Press, 2001). [377]
Möhring R.H., Algorithmic aspects of comparability graphs and interval graphs. In Graphs and order
(Banff, 1984), NATO Adv. Sci. Inst. C Math. Phys. Sci. 147 (Reidel, 1985), 41–101. [573]
Molloy M., The probabilistic method. In Probabilistic methods for algorithmic discrete mathematics, Al-
gorithms Combin. 16 (Springer, 1998), 1–35. [706]
Molloy M. and B.A. Reed, A bound on the strong chromatic index of a graph. J. Combin. Th. B 69 (1997),
103–109. [373]
Molloy M. and B.A. Reed, Near-optimal list colorings. In Proceedings of the Ninth International Confer-
ence “Random Structures and Algorithms” (Poznan, 1999), 17 (2000), 376–402. [362]
Molloy M. and B.A. Reed, Graph colouring and the probabilistic method, Algorithms and Combinatorics
23 (Springer-Verlag, 2002). [335, 657, 668, 674, 677, 681, 686, 702, 706, 721, 722]
Montágh B., A simple proof and a generalization of an old result of Chung and Feller. Discr. Math. 87
(1991), 105–108. [50]
Montellano-Ballesteros J.J. and V. Neumann-Lara, An anti-Ramsey theorem on cycles. Graphs Combin.
21 (2005), 343–354. [457]
Montgomery B., Dynamic coloring of graphs (ProQuest LLC, 2001). Ph.D. Thesis, West Virginia Univer-
sity. [424]
Montgomery R., A. Pokrovskiy, and B. Sudakov, A proof of Ringel’s conjecture (2020).
(arXiv:2001.02665v2). [646]
Montmort P.R., Essay d’Analyse sur les Jeux de Hazard (Paris, 1708). [162]
Moon J.W., On the diameter of a graph. Michigan Math. J. 12 (1965a), 349–351. [251]
Moon J.W., On a problem of Ore. Math. Gaz. 49 (1965b), 40–41. [333]
904 References

Moon J.W., On the number of complete subgraphs of a graph. Canad. Math. Bull. 8 (1965c), 831–834.
[489]
Moon J.W., On subtournaments of a tournament. Canad. Math. Bull. 9 (1966), 297–301. [334]
Moon J.W., Various proofs of Cayley’s formula for counting trees. In A seminar on Graph Theory (Holt,
Rinehart and Winston, 1967), 70–78. [37, 47]
Moon J.W., Counting Labeled Trees (Canad. Math. Congress, 1970). [37]
Moon J.W. and L. Moser, On a problem of Turán. Magyar Tud. Akad. Mat. Kutató Int. Közl. 7 (1962a),
283–286. [477]
Moon J.W. and L. Moser, Almost all tournaments are irreducible. Canad. Math. Bull. 5 (1962b), 61–65.
[703]
Moon J.W. and L. Moser, On Hamiltonian bipartite graphs. Israel J. Math. 1 (1963), 163–165.
[325, 332]
Moore B.R. and D.B. West, Cycles in color-critical graphs (2019). (arXiv:1912.03754v2). [354]
Moore E.H., Concerning triple systems. Math. Ann. 43 (1893), 271–285. [641]
Moore E.H., Tactical memoranda I–III. Am. J. Math. 18 (1896), 264–303. [610, 611]
Morris W. and V. Soltan, The Erdős–Szekeres problem. In Open problems in mathematics (Springer,
[Cham], 2016), 351–375. [446]
Morris W.D., Jr. and V. Soltan, The Erdős–Szekeres problem on points in convex position—a survey.
Bull. Amer. Math. Soc. (N.S.) 37 (2000), 437–458. [446]
Moser L., Problem 4300. Amer. Math. Monthly 55 (1948), 369. Solution 57 (1950), 47. [441]
Moser L., Problem B-6: Some reflections. Fibonacci Quarterly 1 (1963), 75–76. [59]
Moser L., The second moment method in combinatorial analysis. In Combin. Struct. Appl. (Proc. Calgary
Internat. Conf., 1969) (Gordon and Breach, 1970), 283–284. [513]
Moser R.A., A constructive proof of the Lovász local lemma. In STOC’09—Proceedings of the 2009 ACM
International Symposium on Theory of Computing (ACM, 2009), 343–350. [680]
Moser R.A. and G. Tardos, A constructive proof of the general Lovász local lemma. J. ACM 57 (2010),
Art. 11, 15. [680]
Moshkovitz G. and A. Shapira, Ramsey theory, integer partitions and a new proof of the Erdős–Szekeres
theorem. Adv. Math. 262 (2014), 1107–1129. [602, 603]
Motwani R. and P. Raghavan, Randomized algorithms (Cambridge University Press, 1995). [657]
Motzkin T.S. and E.G. Straus, Maxima for graphs and a new proof of a theorem of Turán. Canad. J.
Math. 17 (1965), 533–540. [489, 781]
Mozhan N.N., Twice critical graphs with chromatic number five. Metody Diskret. Analiz. (1987), 50–59,
73. [354]
Mubayi D., Generalizing the Ramsey problem through diameter. Electron. J. Combin. 9 (2002), Research
Paper 42, 10. [460]
Mubayi D., Counting substructures I: color critical graphs. Adv. Math. 225 (2010), 2731–2740. [478]
Mubayi D., Variants of the Erdős–Szekeres and Erdős–Hajnal Ramsey problems. European J. Combin.
62 (2017), 197–205. [457]
Mubayi D. and A. Suk, Off-diagonal hypergraph Ramsey numbers. J. Combin. Th. B 125 (2017), 168–
177. [451]
Mubayi D. and J. Verstraëte, A survey of Turán problems for expansions. In Recent trends in combina-
torics, IMA Vol. Math. Appl. 159 (Springer, [Cham], 2016), 117–143. [500]
Muir T., A Treatise on the Theory of Determinants (Macmillan, 1882). Also 1906. [764]
Mulder H.M., Julius Petersen’s theory of regular graphs. Discr. Math. 100 (1992), 157–175. [253, 274]
Mullin R. and G.C. Rota, On the foundations of combinatorial theory. III. Theory of binomial enumera-
tion. In Graph Theory and its Applications (Proc. Advanced Sem., Madison, WI, 1969) (Academic
Press, 1970), 167–213. [123]
Munkres J., Algorithms for the assignment and transportation problems. J. Soc. Indust. Appl. Math. 5
(1957), 32–38. [280]
Mycielski J., Sur le coloriage des graphes. Coll. Math. 3 (1955), 161–162. [340]
References 905

Myers B.R., On spanning trees, weighted compositions, Fibonacci numbers, and resistor networks. SIAM
Rev. 17 (1975), 465–474. [65]
Myers B.R. and R. Liu, A lower bound on the chromatic number of a graph. Networks 1 (1972), 273–277.
[354]
Naatz M., A connectivity lemma. In 6th International Conference on Graph Theory (Marseille, 2000),
Electron. Notes Discr. Math. 5 (Elsevier, 2000), 4. [312, 764]
Nagasawa T., K. Noguchi, and Y. Suzuki, Optimal 1-embedded graphs on the projective plane which
triangulate other surfaces. J. Nonlinear Convex Anal. 19 (2018), 1759–1770. [797]
Nagy Z., A certain constructive estimate of the Ramsey number. Mat. Lapok 23 (1972), 301–302.
[459, 728, 745]
Nakasawa T., Zur Axiomatik der linearen Abhängigkeit. I, II, III (in German). Sci. Rep. Tokyo Bunrika
Daigaku Sect. A 2 (1935), 129–149. II and III in 3 (1936), 17–41 and 77–90. [514]
Nakayama A. and B. Péroche, Linear arboricity of digraphs. Networks 17 (1987), 39–53. [678]
Narayana T.V., A combinatorial problem and its application to probability theory. I. J. Indian Soc. Agric.
Statist. 7 (1955), 169–178. [45]
Nasar S., A beautiful mind (Simon & Schuster, 1998). Also 2001. [803]
Nash-Williams C.St.J.A., On orientations, connectivity and odd-vertex-pairings in finite graphs. Canad.
J. Math. 12 (1960), 555–567. [307, 309]
Nash-Williams C.St.J.A., Edge-disjoint spanning trees in finite graphs. J. Lond. Math. Soc. 36 (1961),
445–450. [251, 535, 540]
Nash-Williams C.St.J.A., Decomposition of finite graphs into forests. J. Lond. Math. Soc. 39 (1964), 12.
[251, 535]
Nash-Williams C.St.J.A., An application of matroids to graph theory. In Theory of Graphs, Intl. Sympos.,
Rome (Dunod, 1966), 263–265. [533]
Nash-Williams C.St.J.A., An unsolved problem concerning decomposition of graphs into triangles. Com-
binatorial Theory and Its Applications III (1970), 1179–1183. [238, 640]
Nash-Williams C.St.J.A., Connected detachments of graphs and generalized Euler trails. J. London
Math. Soc. (2) 31 (1985), 17–29. [540]
Naslund E. and W.F. Sawin, Upper bounds for sunflower-free sets (2016). (arXiv:1606.09575). [504]
Nemhauser G.L. and L.A. Wolsey, Integer and combinatorial optimization (Wiley, 1988). [520]
Nešetřil J., A combinatorial classic—sparse graphs with high chromatic number. In Erdős Centennial,
Bolyai Soc. Math. Stud. 25 (János Bolyai Math. Soc., Budapest, 2013), 383–407. [434]
Nešetřil J. and V. Rödl, The Ramsey property for graphs with forbidden complete subgraphs. J. Combin.
Th. B 20 (1976), 243–249. [455]
Nešetřil J. and V. Rödl, A short proof of the existence of highly chromatic hypergraphs without short
cycles. J. Combin. Th. B 27 (1979), 225–227. [434, 436, 437]
Nešetřil J. and V. Rödl, Complexity of diagrams. Order 3 (1987), 321–330. Corrigendum: Order 10 (1993),
393. [543]
Nešetřil J. and V. Rödl (eds.), Mathematics of Ramsey theory, Algorithms and Combinatorics 5 (Springer-
Verlag, 1990). [425]
Netto E., Lehrbuch der Combinatorik (German) (Teubner, 1901). Also 1927. [41]
Neumann-Lara V. and E. Rivera-Campo, Spanning trees with bounded degrees. Combinatorica 11
(1991), 55–61. [327]
Nicoaescu L.I., Problem E3157. Amer. Math. Monthly 93 (1986), 482. Solution 95 (1988), 557. [22]
Niessen T. and B. Randerath, Regular factors of simple regular graphs and factor-spectra. Discr. Math.
185 (1998), 89–103. [273]
Niessen T. and L. Volkmann, Class 1 conditions depending on the minimum degree and the number of
vertices of maximum degree. J. Graph Th. 14 (1990), 225–246. [359]
Nikiforov V., The cycle-complete graph Ramsey numbers. Combin. Probab. Comput. 14 (2005), 349–370.
[454]
Nikiforov V., Bounds on graph eigenvalues. II. Linear Algebra Appl. 427 (2007), 183–189. [770]
906 References

Nikiforov V., Graphs with many r-cliques have large complete r-partite subgraphs. Bull. Lond. Math.
Soc. 40 (2008), 23–25. [482]
Nikiforov V., A spectral Erdős–Stone–Bollobás theorem. Combin. Probab. Comput. 18 (2009), 455–458.
[770]
Nikiforov V., The number of cliques in graphs of given order and size. Trans. Amer. Math. Soc. 363 (2011),
1599–1618. [478]
Nikšić F., Are surjections [n] → [¾] more common than injections [¾] → [n]? Available online at
www.mathoverflow.net/questions/268544 (2017). Solution by M. Wildon, July 2, 2017. [133]
Nilli A., On the second eigenvalue of a graph. Discr. Math. 91 (1991), 207–210. [780]
Nilli A., On Borsuk’s problem. In Jerusalem combinatorics ’93, Contemp. Math. 178 (AMS, 1994), 209–
210. [817, 818]
Nishimura H. and S. Kuroda, A Lost Mathematician, Takeo Nakasawa: The Forgotten Father of Matroid
Theory, Mathematics and Statistics (Mirkhäuser, 2009). [514]
Niu Y.Y. and B.W. Zhu, Connectivities of Cartesian products of graphs. In Combinatorics, graph theory,
algorithms and applications (Beijing, 1993) (World Sci. Publ., 1994), 301–305. [292]
Niven I., Formal power series. Amer. Math. Monthly 76 (1969), 871–889. [107]
Noel J.A., B.A. Reed, and H. Wu, A proof of a conjecture of Ohba. J. Graph Theory 79 (2015), 86–102.
[355]
Noel J.A., D.B. West, H. Wu, and X. Zhu, Beyond Ohba’s conjecture: a bound on the choice number of ¾ -
chromatic graphs with n vertices. European J. Combin. 43 (2015), 295–305. [356]
Nordhaus E.A. and J.W. Gaddum, On complementary graphs. Amer. Math. Monthly 63 (1956), 175–177.
[344]
Norman R.Z. and M. Rabin, Algorithm for a minimal cover of a graph. Proc. Amer. Math. Soc. 10 (1959),
315–319. [263]
Novelli J.C., I. Pak, and A.V. Stoyanovskii, A direct bijective proof of the hook-length formula. Discr.
Math. Theor. Comput. Sci. 1 (1997), 53–67. [192]
Nyblom M., Problem 11117. Amer. Math. Monthly 111 (2004), 915. Solution 113 (2006), 762. [172]
O S. and D.B. West, Balloons, cut-edges, matchings, and total domination in regular graphs of odd de-
gree. J. Graph Theory 64 (2010), 116–131. [237, 274]
O S., D.B. West, and H. Wu, Longest cycles in ¾ -connected graphs with given independence number. J.
Combin. Th. B 101 (2011), 480–485. [334]
O’Hara K.M., Unimodality of Gaussian coefficients: a constructive proof. J. Combin. Th. A 53 (1990),
29–52. [557]
Ohba K., On chromatic-choosable graphs. J. Graph Theory 40 (2002), 130–135. [355]
Ohtsuka H. and R. Tauraso, Problem 12016. Amer. Math. Monthly 125 (2018), 81. Solution 126 (2019),
665. [33]
Olmsted C., Problem E3175. Amer. Math. Monthly 93 (1986), 732. Solution 96 (1989), 59–60. [176]
Olson E.J. and B.E. Sagan, On the 1/3–2/3 conjecture. Order 35 (2018), 581–596. [823]
Ore O., A problem regarding the tracing of graphs. Elemente der Math. 6 (1951), 49–53. [250]
Ore O., Graphs and matching theorems. Duke Math. J. 22 (1955), 625–639. [258, 531]
Ore O., Graphs and subgraphs. II. Trans. Amer. Math. Soc. 93 (1959), 185–204. [332]
Ore O., Note on Hamilton circuits. Amer. Math. Monthly 67 (1960), 55. [319]
Ore O., Arc coverings of graphs. Ann. Mat. Pura Appl. 55 (1961), 315–321. [321]
Ore O., Theory of graphs, Amer. Math. Soc. Colloquium Publications 38 (Amer. Math. Soc., 1962).
[260, 571]
Ore O., Hamiltonian connected graphs. J. Math. Pures Appl. 42 (1963), 21–7. [332, 333]
Ore O., The four-colour problem (Academic Press, 1967). [355, 360, 387, 399, 420]
Ore O., Diameters in graphs. J. Combinatorial Theory 5 (1968), 75–81. [249, 359]
O’Rourke J., Art gallery theorems and algorithms, International Series of Monographs on Computer Sci-
ence (Oxford Univ. Press, 1987). [394]
Owens A.B., On the planarity of regular incidence sequences. J. Combinatorial Theory B 11 (1971), 201–
212. [389]
References 907

Oxley J.G., Matroid theory, Oxford Science Publications (Oxford Univ. Press, 1992). 2nd edition 2011.
[514, 525]
Ozeki K. and T. Yamashita, Spanning trees: a survey. Graphs Combin. 27 (2011), 1–26. [327, 334]
Paasche I., Eine Verallgemeinerung des Moessnerschen Satzes. Compositio Math. 12 (1956), 263–270.
[62]
Pach J. and P.K. Agarwal, Combinatorial geometry, Wiley-Interscience Ser. Discr. Math. Optim. (Wiley &
Sons, 1995). [783]
Pach J. and M. Sharir, Combinatorial geometry and its algorithmic applications, Mathematical Surveys
and Monographs 152 (AMS, 2009). [783]
Pach J. and G. Tardos, Forbidden paths and cycles in ordered graphs and matrices. Israel J. Math. 155
(2006), 359–380. [491]
Pach J. and G. Tóth, Graphs drawn with few crossings per edge. Combinatorica 17 (1997), 427–439.
[793, 797]
Pach J. and G. Tóth, Thirteen problems on crossing numbers. Geombinatorics 9 (2000a), 194–207.
[795]
Pach J. and G. Tóth, Which crossing number is it anyway? J. Combin. Th. B 80 (2000b), 225–246.
[795]
Pak I., Partition bijections, a survey. Ramanujan J. 12 (2006), 5–75. [141, 142, 152]
Palacios J.L., Problem 11672. Amer. Math. Monthly 119 (2012), 800. Solutions by P. Condon and R.
Strong 121 (2014), 948–949. [66]
Palacios J.L. and D.P. Sandell, Problem 6665. Amer. Math. Monthly 98 (1991), 655. Solution 100 (1993),
405–407. [668]
Paley R.E.A.C., On orthogonal matrices. J. Math. Phys. 12 (1933), 311–320. [621]
Palmer C., Extremal Combinatorics (2015). Lecture notes. [484, 488]
Palmer E.M., Graphical Evolution: An Introduction to the Theory of Random Graphs (Wiley, 1985).
[657, 700, 703, 705]
Palumbı́ny D., On decompositions of complete graphs into factors with equal diameters. Boll. Un. Mat.
Ital.(4) 7 (1973), 420–428. [249]
Papadimitriou C.H. and K. Steiglitz, Combinatorial Optimization: Algorithms and Complexity (Prentice
Hall, 1982). Also Dover, 1998. [520]
Parker E.T., Orthogonal latin squares. Proc. Nat. Acad. Sci. U.S.A. 45 (1959), 859–862. [652]
Paulraja P., A characterization of Hamiltonian prisms. J. Graph Th. 17 (1993), 161–171. [326]
Payan C., Sur le nombre d’absorption d’un graphe simple. In Colloque sur la Théorie des Graphes (Paris,
1974), Cahiers Centre Études Recherche Opér. 17 (1975), 307–317. [670]
Peart P. and W.J. Woan, A bijective proof of the Delannoy recurrence. In Proc. 33rd Southeastern Intl.
Conf. Combin. Graph Th. Comput. (Boca Raton), Congr. Numer. 158 (2002), 29–33. [64]
Peck G.W., A new proof of a theorem of Graham and Pollak. Discr. Math. 49 (1984), 327–328. [771]
Peck G.W., Noncrossing trees, solution II to problem e3170. Amer. Math. Monthly 96 (1989), 359–361.
Proposed by Howard University Group, 93 (1986), 650. [42]
Peled U., Problem 10197. Amer. Math. Monthly 99 (1992), 162. Solution 100 (1993), 806–807. [764]
Pelsmajer M.J., M. Schaefer, and D. Štefanovič, Odd crossing number is not crossing number. In Graph
drawing, Lect. Notes Comp. Sci. 3843 (Springer, 2006), 386–396. [795]
Peltesohn R., Eine Lösung der beiden Heffterschen Differenzenprobleme. Compositio Math. 6 (1939),
251–257. [645]
Penaud J.G., Une propriété de bicoloration des hypergraphes planaires. In Colloque sur la Théorie des
Graphes (Paris, 1974), Cahiers Centre Études Recherche Opér. 17 (1975), 345–349. [418, 419]
Penrose M., Random geometric graphs, Oxford Studies in Probability 5 (Oxford Univ. Press, 2003).
[783]
Percus J.K., Combinatorial methods, Notes recorded by Ora Engelberg Percus (Courant Inst. Math. Sci.,
New York Univ., 1969). [755]
Perkovic L. and B.A. Reed, Edge coloring regular graphs of high degree. In Graphs Combinatorics (Mar-
seille, 1995) 165/166 (Discr. Math., 1997), 567–578. [359]
908 References

Perles M.A., A proof of Dilworth’s decomposition theorem for partially ordered sets. Israel J. Math. 1
(1963), 105–107. [546]
Perron O., Beweis des Moessnerschen Satzes. S.-B. Math.-Nat. Kl. Bayer. Akad. Wiss. 1951 (1951), 31–
34 (1952). [62]
Perz S. and S. Rolewicz, Norms and perfect graphs. Z. Oper. Res. 34 (1990), 13–27. [369]
Petersen J., Die Theorie der regulären Graphen. Acta Math. 15 (1891), 193–220.
[213, 267, 268, 273, 274, 277, 740]
Petersen T.K., Eulerian numbers, Birkhäuser Advanced Texts: Basler Lehrbücher. (Birkhäuser/Springer,
2015). [101]
Petkovsek M., Finding closed-form solutions of difference equations by symbolic methods (ProQuest LLC,
1991). Ph.D. Thesis, Carnegie Mellon Univ. [90]
Petkovšek M., H.S. Wilf, and D. Zeilberger, A = B (A K Peters, 1996). [87, 89, 90]
Phillips J., Chernoff–Hoeffding inequality and applications (2013). (arXiv:1209.63946). [711]
Pikhurko O., Borsuk’s conjecture fails in dimensions 321 and 322 (2002). (arXiv:math/0202112). [817]
Pikhurko O., A note on the Turán function of even cycles. Proc. Amer. Math. Soc. 140 (2012), 3687–3692.
[479]
Pilpel S., Descending subsequences of random permutations. J. Combin. Th. A 53 (1990), 96–116.
[821]
Pinsker M., On the complexity of a concentrator. 7th International Teletraffic Conference (1973), 318/1–
318/4. [779]
Piotrowski W.L., Untersuchungen über das Oberwolfacher Problem (1979). Unpublished manuscript.
[647]
Piotrowski W.-L., The solution of the bipartite analogue of the Oberwolfach problem. Discr. Math. 97
(1991), 339–356. [647]
Pippenger N., Superconcentrators. SIAM J. Comput. 6 (1977), 298–304. [779]
Pippenger N., Entropy and enumeration of Boolean functions. IEEE Trans. Inform. Theory 45 (1999),
2096–2100. [509]
Pité E., Problem 11957. Amer. Math. Monthly 124 (2017), 179. Solution 125 (2018), November. [133]
Pitman J., Coalescent random forests. J. Combin. Th. A 85 (1999), 165–193. [37]
Plantholt M., Overfull conjecture for graphs with high minimum degree. J. Graph Theory 47 (2004), 73–
80. [359]
Plaza A. and S. Falcón, Problem 11920. Amer. Math. Monthly 123 (2016), 296. Solution 125 (2018), 374–
375. [80]
Plesnevič P.S. and V.G. Vizing, On the problem of the minimal coloring of the vertices of a graph (Rus-
sian). Sibirsk. Mat. Zh. 6 (1965), 234–236. [343]
Plesnı́k J., Connectivity of regular graphs and the existence of 1-factors. Mat. Časopis Sloven. Akad.
Vied 22 (1972), 310–318. [267]
Plesnı́k J., Critical graphs of given diameter. Acta Fac. Rerum Natur. Univ. Comenian. Math. 30 (1975),
71–93. [294]
Plesnı́k J. and Š. Znám, On equality of edge-connectivity and minimum degree of a graph. Arch. Math.
(Brno) 25 (1989), 19–25. [296]
Plummer M.D., On minimal blocks. Trans. Amer. Math. Soc. 134 (1968), 85–94. [315, 316]
Plummer M.D., On the cyclic connectivity of planar graphs. In Graph theory and applications (Proc.
Conf., Western Michigan Univ., 1972; dedicated to the memory of J. W. T. Youngs), Lecture Notes
in Math. 303 (Springer, 1972), 235–242. [297]
Plummer M.D., Problem. In Infinite and finite sets (A. Haynal, R. Rado, and V.T. Sós, eds.) 10 (North-
Holland, 1975), 1549–1550. [401]
Plummer M.D., Matching theory—a sampler: from Dénes König to the present. Discr. Math. 100 (1992),
177–219. [253]
Plummer M.D., Graph factors and factorization: 1985–2003: a survey. Discr. Math. 307 (2007), 791–821.
[268]
References 909

Pnueli A., A. Lempel, and S. Even, Transitive orientation of graphs and identification of permutation
graphs. Canad. J. Math. 23 (1971), 160–175. [343]
Pólya G., Kombinatorische Anzahlbestimmungen für Gruppen, Graphen und chemische Verbindungen.
Acta Math. 68 (1937), 145–254. [37, 128, 184]
Poonen B., Union-closed families. J. Combin. Th. A 59 (1992), 253–268. [503]
Popadić M.S., On the number of antichains of finite power sets. Mat. Vesnik 7 (22) (1970), 199–203.
[550]
Popescu C., Problem 10770. Amer. Math. Monthly 106 (1999), 963. Solution 108 (2001), 979–80. [172]
Pósa L., On the circuits of finite graphs. Magyar Tud. Akad. Mat. Kutató Int. Közl. 8 (1963), 355–361
(1964). [326, 333]
Pournin L., The diameter of associahedra. Adv. Math. 259 (2014), 13–42. [387]
Pratt R., Problem 11573. Amer. Math. Monthly 118 (2011), 463. Solution 120 (2013), 372. [22]
Preen J., A census of all 5-regular planar graphs with diameter 3. Ars Combin. 106 (2012), 129–135.
[386]
Prim R.C., Shortest connection networks and some generalizations. Bell Syst. Tech. J. 36 (1957), 1389–
1401. [246, 252]
Proctor R.A., Solution of two difficult combinatorial problems with linear algebra. Amer. Math. Monthly
89 (1982), 721–734. [557]
Proctor R.A., Let’s expand Rota’s twelvefold way for counting partitions! (2006). (arXiv:math/0606404).
[147]
Prodinger H., A correspondence between ordered trees and noncrossing partitions. Discr. Math. 46
(1983), 205–206. [49, 64]
Prömel H.J., Ramsey theory for discrete structures (Springer, 2013). [425]
Propp J.G., Problem 10679. Amer. Math. Monthly 105 (1998), 666. Solution 107 (2000), 374. [511]
Prowse A. and D.R. Woodall, Choosability of powers of circuits. Graphs Combin. 19 (2003), 137–144.
[740]
Prüfer H., Neuer beweis eines satzes über permutationen (German). Arch. Math. Phys. 27 (1918), 742–
744. [37]
Pudaite P.R., Problem 10801. Amer. Math. Monthly 107 (2000), 368. Solution 109 (2002), 394–5. [66]
Puleo G.J., Maximal ¾ -edge-colorable subgraphs, Vizing’s theorem, and Tuza’s conjecture. Discr. Math.
340 (2017), 1573–1580. [372]
Pyber L., An extension of a Frankl–Füredi theorem. Discr. Math. 52 (1984), 253–268. [249]
Pyber L., Regular subgraphs of dense graphs. Combinatorica 5 (1985), 347–349. [737]
Pyber L., V. Rödl, and E. Szemerédi, Dense graphs without 3-regular subgraphs. J. Combin. Th. B 63
(1995), 41–54. [737]
Pyke R., The supremum and infimum of the Poisson process. Ann. Math. Statist. 30 (1959), 568–576.
[50]
Pym J.S., The linking of sets in graphs. J. London Math. Soc. 44 (1969), 542–550. [299]
Qiao P. and X. Zhan, The largest graphs with given order and diameter: A simple proof. Graphs and
Combinatorics 36 (2019), 1715–1716. [249]
Rabinovitch I., The Dimension-Theory of Semiorders and Interval-Orders (ProQuest LLC, 1973). Ph.D.
Thesis, Dartmouth College. [604]
Rabinovitch I., The Scott–Suppes theorem on semiorders. J. Mathematical Psychology 15 (1977), 209–
212. [604]
Rabinovitch I., The dimension of semiorders. J. Combin. Th. A 25 (1978), 50–61. [574, 583, 604]
Rademacher H., On the partition function p(n). Proc. London Math. Soc. 43 (1937), 241–254. [140]
Radhakrishnan J., An entropy proof of Brégman’s theorem. J. Combin. Th. A 77 (1997), 161–164.
[510]
Radhakrishnan J., Entropy and counting. In Computational mathematics, modeling and algorithms (J.
C. Misra, ed.) (Narosa, 2003), 146–148. [504, 508]
Radhakrishnan J. and A. Srinivasan, Improved bounds and algorithms for hypergraph 2-coloring. Ran-
dom Structures Algorithms 16 (2000), 4–32. [660, 673]
910 References

Rado R., Studien zur Kombinatorik. Math. Z. 36 (1933), 424–470. [462, 467]
Rado R., Axiomatic treatment of rank in infinite sets. Canadian J. Math. 1 (1949), 337–343. [468]
Rado R., Note on independence functions. Proc. Lond. Math. Soc. 7 (1957), 300–320. [519]
Rado R., Some partition theorems. In Combinatorial Theory and its Applications, III (Proc. Colloq., 1969)
(North-Holland, 1970), 929–936. [466]
Radoux C., Nombres de Catalan généralisés. Bull. Belg. Math. Soc. Simon Stevin 4 (1997), 289–292.
[177]
Radziszowski S.P., Small Ramsey numbers. Electronic J. Combin. (1994), Dynamic Survey 1. [449]
Raı̆gorodskiı̆ A.M., On the dimension in Borsuk’s problem (Russian). Uspekhi Mat. Nauk 52 (1997), 181–
182. [817]
Ramalingam G. and C. PanduRangan, A unified approach to domination problems on interval graphs.
Inform. Process. Lett. 27 (1988), 271–274. [374]
Ramamurthi R., Coloring problems on graphs and hypergraphs (ProQuest, 2001). Ph.D. Thesis, Univ.
Illinois, Urbana-Champaign. [746]
Ramanan G.V., Proof of a conjecture of Frankl and Füredi. J. Combin. Th. A 79 (1997), 53–67. [730]
Ramı́rez-Alfonsı́n J.L. and B.A. Reed (eds.), Perfect graphs, Wiley-Interscience Series in Discr. Mathe-
matics and Optimization (Wiley & Sons, 2001). [370]
Ramos E.A., Equipartition of mass distributions by hyperplanes. Discr. Comput. Geom. 15 (1996), 147–
167. [815]
Ramras M., Minimum cutsets in hypercubes. Discr. Math. 289 (2004), 193–198. [295]
Ramsey F.P., On a Problem of Formal Logic. Proc. Lond. Math. Soc. 30 (1930), 264–286.
[443, 444, 468]
Raney G.N., Functional composition patterns and power series reversion. Trans. Amer. Math. Soc. 94
(1960), 441–451. [50, 129]
Ray-Chaudhuri D.K., Characterization of line graphs. J. Combin. Th. 3 (1967), 201–214. [370]
Ray-Chaudhuri D.K. and R.M. Wilson, Solution of Kirkman’s schoolgirl problem (1971), 187–203.
[647, 649]
Ray-Chaudhuri D.K. and R.M. Wilson, On t-designs. Osaka J. Math. 12 (1975), 737–744. [731]
Raynaud H., Sur le circuit hamiltonien bi-coloré dans les graphes orientés. Period. Math. Hungar. 3
(1973), 289–297. [460]
Razborov A.A., Lower bounds on the monotone complexity of some Boolean functions. Dokl. Akad. Nauk
SSSR 281 (1985), 798–801. [504]
Razborov A.A., Flag algebras. J. Symbolic Logic 72 (2007), 1239–1282. [475]
Razborov A.A., On the minimal density of triangles in graphs. Combin. Probab. Comput. 17 (2008), 603–
618. [478]
Razborov A.A., On 3-hypergraphs with forbidden 4-vertex configurations. SIAM J. Discr. Math. 24
(2010), 946–963. [475]
Razborov A.A., On the Fon-der-Flaass interpretation of extremal examples for Turán’s (3 , 4)-problem.
Tr. Mat. Inst. Steklova 274 (2011), 269–290. [491]
Read R.C., Review of “die chromatischen Polynome unterringfreier Graphen”. MathSciNet Review
(1975). MR0354428 (50 #6906) [374]
Rédei L., Ein kombinatorischer Satz. Acta Litt. Szeged 7 (1934), 39–43. [334]
Redfield J.H., The Theory of Group-Reduced Distributions. Amer. J. Math. 49 (1927), 433–455. [184]
Reed B.A., ,  , and . J. Graph Th. 27 (1998), 177–212. [339, 685]
Reed B.A., A strengthening of Brooks’ Theorem. J. Combin. Th. B 76 (1999), 136–149. [677]
Reed B.A. and B. Sudakov, Asymptotically the list colouring constants are 1. J. Combin. Th. B 86 (2002),
27–37. [355, 677]
Rees D., Note on a paper by I.J. Good. J. Lond. Math. Soc. 21 (1946), 169–172. [239]
Regev A., Asymptotic values for degrees associated with strips of Young diagrams. Adv. in Math. 41
(1981), 115–136. [433]
Reiher C., The clique density theorem. Ann. of Math. (2) 184 (2016), 683–707. [478]
References 911

Reiman I., Über ein Problem von K. Zarankiewicz. Acta. Math. Acad. Sci. Hungar. 9 (1958), 269–273.
[632]
Remmel J.B., Bijective proofs of formulae for the number of standard Young tableaux. Linear and Mul-
tilinear Algebra 11 (1982), 45–100. [192]
Renault M., Lost (and found) in translation: André’s actual method and its application to the generalized
ballot problem. Amer. Math. Monthly 115 (2008), 358–363. [40]
Rényi A., Some remarks on the theory of trees. Magyar Tud. Akad. Mat. Kut. Int. Közl. 4 (1959), 73–85.
[133]
Rényi A., Théorie des éléments saillants d’une suite d’observations (French). Ann. Fac. Sci. Univ.
Clermont-Ferrand No. 8 (1962), 7–13. [99]
Rényi A., On the enumeration of trees. In Combinatorial Structures and their Applications (Proc. Calgary
Internat. Conf., 1969) (Gordon and Breach, 1970), 355–360. [37]
Reuter K., On the dimension of the Cartesian product of relations and orders. Order 6 (1989a), 277–293.
[572]
Reuter K., Removing critical pairs. Order 6 (1989b), 107–118. [584]
Rey J.G., Problem 10615. Amer. Math. Monthly 104 (1997), 767. Solution 106 (1999), 692–693. [47]
Reznick B., P. Tiwari, and D.B. West, Decomposition of product graphs into complete bipartite sub-
graphs. Discr. Math. 57 (1985), 179–183. [771]
Ribó Mor A., G. Rote, and A. Schulz, Small grid embeddings of 3-polytopes. Discrete Comput. Geom. 45
(2011), 65–87. [783]
Richter R.B., Problem 10330. Amer. Math. Monthly 100 (1993), 796. Solution 103 (1996), 700–1. [489]
Richter R.B. and C. Thomassen, Intersections of curve systems and the crossing number of C5 × C5 .
Discrete Comput. Geom. 13 (1995), 149–159. [794]
Riddell R.J., Contributions to the theory of condensation (Univ. Michigan, Ann Arbor, 1951). Ph.D. Thesis.
[135]
Riddell R.J., Jr. and G.E. Uhlenbeck, On the theory of the virial development of the equation of state of
mono-atomic gases. J. Chem. Phys. 21 (1953), 2056–2064. [127]
Riesling A.S., Boruk’s problem in three-dimensional spaces of constant curvature. Ukr. Geom. Sborni 11
(1971), 78–83. [817]
Rieß W., Zwei Optimierungsprobleme auf Ordnungen. Arbeitsber. Inst. Math. Masch. Datenverarb. (In-
form.) 11 (1978), 59. [557]
Riguet J., Les relations de Ferrers. C. R. Acad. Sci. Paris 232 (1951), 1729–1730. [605]
Ringeisen R.D. and L.W. Beineke, The crossing number of C3 × Cn. J. Combin. Th. B 24 (1978), 134–
136. [794]
Ringel G., Selbstkomplementare graphen. Arch. Math. 14 (1963), 354–358. [218]
Ringel G., Problem 25. In Theory of Graphs and Its Applications (Smolenice, 1963) (Czech. Acad. Sci.,
1964), 162. [646, 656]
Riordan J., An introduction to combinatorial analysis (Wiley; Chapman & Hall, 1958). [153, 161]
Riordan J., Inverse relations and combinatorial identities. Amer. Math. Monthly 71 (1964), 485–498.
[133]
Riordan J., Combinatorial identities (Wiley & Sons, 1968). Also 1979. [47, 65]
Robbins H., A theorem on graphs, with an application to a problem in traffic control. Amer. Math. Monthly
46 (1939), 281–283. [306]
Roberts F.S., Graph Theory and Its Applications to the Problems of Society, CBMS-NSF Monograph 29
(SIAM Publications, 1978). [337]
Robertson N., D. Sanders, P. Seymour, and R. Thomas, The four-colour theorem. J. Combin. Th. B 70
(1997), 2–44. [406, 416, 417]
Robertson N., P. Seymour, and R. Thomas, Excluded minors in cubic graphs. J. Combin. Th. B 138
(2019), 219–285. [416]
Robertson N., P.D. Seymour, and R. Thomas, Hadwiger’s conjecture for K6 -free graphs. Combinatorica
13 (1993), 279–361. [349]
912 References

Robertson N., P.D. Seymour, and R. Thomas, Tutte’s edge-colouring conjecture. J. Combin. Th. B 70
(1997), 166–183. [416]
Robinson G.de B., On the representations of the symmetric group. Amer. J. Math. 60 (1938), 745–760.
[193, 195]
Robinson J.P. and M. Cohn, Counting sequences. IEEE Trans. Comput. 30 (1981), 17–23. [820]
Rödl V. and M. Schacht, Generalizations of the removal lemma. Combinatorica 29 (2009), 467–501.
[485]
Rödl V. and M. Schacht, Regularity lemmas for graphs. In Fete of combinatorics and computer science,
Bolyai Soc. Math. Stud. 20 (János Bolyai Math. Soc., 2010), 287–325. [488]
Rödl V. and J. Skokan, Regularity lemma for ¾ -uniform hypergraphs. Random Structures Algorithms 25
(2004), 1–42. [488]
Rödl V. and E. Szemerédi, On size Ramsey numbers of graphs with bounded degree. Combinatorica 20
(2000), 257–262. [455]
Roman S.M., On a Problem of Zarankiewicz Concerning Matrices of Zeros and Ones (ProQuest, 1975).
Ph.D. Thesis, Univ. Washington. [632]
Roman S.M., The umbral calculus, Pure and Applied Mathematics 111 (Academic Press, 1984). [123]
Roman S.M. and G.C. Rota, The umbral calculus. Advances in Math. 27 (1978), 95–188. [123]
Rosa A., On certain valuations of the vertices of a graph. In Theory of Graphs (Rome, 1966), Intl. Symp.
(Gordon and Breach; Dunod, 1967), 349–355. [646, 656]
Rosenfeld M., On the total coloring of certain graphs. Israel J. Math. 9 (1971), 396–402. [373]
Rota G.C., On the foundations of combinatorial theory I. Z. Wahrsch. Verw. Gebiete 2 (1964), 340–368.
[520, 529]
Roth K.F., On certain sets of integers. J. London Math. Soc. 28 (1953), 104–109. [484, 485]
Rotman J.J., Problem E3462. Amer. Math. Monthly 98 (1991), 755. Solution 100 (1993), 594. [238]
Rousseau C.C. and J. Sheehan, A class of Ramsey problems involving trees. J. London Math. Soc. (2) 18
(1978), 392–396. [459]
Roussel F., I. Rusu, and H. Thuillier, The strong perfect graph conjecture: 40 years of attempts, and its
resolution. Discr. Math. 309 (2009), 6092–6113. [370]
Roy B., Nombre chromatique et plus longs chemins d’un graphe. Rev. Française Automat. Informat.
Recherche Opérationelle sér. Rouge 1 (1967), 127–132. [339]
Royle G., Graphs where every two vertices have odd number of mutual neighbours (2010).
Available online at https://mathoverflow.net/questions/17809/graphs-where-every-
two-vertices-have-odd-number-of-mutual-neighbours. [226]
Rubel L.A., Problem 6565. Amer. Math. Monthly 95 (1988), 60. Solution 97 (1990), 80–81. [473]
Ruciński A. and A. Vince, Balanced graphs and the problem of subgraphs of random graphs. In Proc. 16th
Southeastern Intl. Conf. Combin. Graph Th. Comput. (Boca Raton), 49 (1985), 181–190. [703]
Rupp C.A., Problem 3468. Amer. Math. Monthly 37 (1930), 552. Solution 38 (1931), 355. [178]
Ruzsa I.Z. and E. Szemerédi, Triple systems with no six points carrying three triangles. In Combinatorics
(Proc. Fifth Hungarian Colloq., Keszthely, 1976), II, Colloq. Math. Soc. János Bolyai 18 (North-
Holland, 1978), 939–945. [485, 492, 684]
Ryjáček Z., On a closure concept in claw-free graphs. J. Combin. Th. B 70 (1997), 217–224. [320, 326]
Rymer N.W., Projects, problems and patience. Math. Gazette 63 (1979), 1–7. [66]
Ryser H.J., Combinatorial properties of matrices of zeros and ones. Canad. J. Math. 9 (1957), 371–377.
[227]
Ryser H.J., Combinatorial mathematics, The Carus Mathematical Monographs, No. 14 (Math. Assoc.
Amer.; distributed by Wiley, 1963). [314]
Ryser H.J., Matrices of zeros and ones in combinatorial mathematics. In Recent Advances in Matrix
Theory (Madison, 1963) (Univ. Wisc. Press, 1964), 103–124. [228]
Saaty T.L. and P.C. Kainen, The four-color problem (McGraw-Hill, 1977). [399, 418]
Sabidussi G., Graphs with given group and given graph-theoretical properties. Canad. J. Math. 9 (1957),
515–525. [343]
References 913

Sachs H., Über Teiler, Faktoren und charakteristische Polynome von Graphen II. Wiss. Z. Techn.
Hochsch. Ilmenau 13 (1967), 405–412. [780]
Sachs H., Finite graphs (Investigations and generalizations concerning the construction of finite graphs
having given chromatic number and no triangles). In Recent Progress in Combinatorics (Proc.
3rd Waterloo Conf. Combin., 1968) (Academic Press, 1969), 175–184. [339]
Sachs H., Elementary proof of the cycle-plus-triangles theorem. In Combinatorics, Paul Erdős is eighty,
Vol. 1, Bolyai Soc. Math. Stud. (János Bolyai Math. Soc., 1993), 347–359. [739]
Sachs H. and H. Zernitz, Remark on the dimer problem. Discrete Appl. Math. 51 (1994), 171–179. [34]
Sagan B., On selecting a random shifted Young tableau. J. Algorithms 1 (1980), 213–234. [207]
Sagan B.E., The symmetric group, The Wadsworth & Brooks/Cole Mathematics Series (Wadsworth &
Brooks, 1991). [189]
Saito A., Chvátal–Erdős theorem: old theorem with new aspects. In Computational geometry and graph
theory, Lecture Notes in Comput. Sci. 4535 (Springer, Berlin, 2008), 191–200. [321]
Šajna M., Cycle decompositions. III. Complete graphs and fixed length cycles. J. Combin. Des. 10 (2002),
27–78. [646]
Saks M.E., A short proof of the existence of ¾ -saturated partitions of partially ordered sets. Adv. in Math.
33 (1979), 207–211. [548]
Saks M.E., Duality Properties of Finite Set Systems (ProQuest, 1980). Ph.D. Thesis, Mass. Inst. Tech.
[549]
Saks M.E., A class of perfect graphs associated with planar rectilinear regions. SIAM J. Algebraic Dis-
crete Methods 3 (1982), 330–342. [831]
Saks M.E., The information theoretic bound for problems on ordered sets and graphs. In Graphs and
order (Banff, Alta., 1984), NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci. 147 (Reidel, 1985), 137–
168. [831]
Salamon G. and G. Wiener, Leaves of spanning trees and vulnerability. In 5th Hungarian–Japanese
Sym. Discr. Math. and Appl. (2007), 225–235. [334]
Sanders D.P. and Y. Zhao, A note on the three color problem. Graphs Combin. 11 (1995), 91–94.
[414, 415]
Sanders D.P. and Y. Zhao, Planar graphs of maximum degree seven are class I. J. Combin. Th. B 83
(2001), 201–212. [422]
Sanders D.P. and Y. Zhao, On the size of edge chromatic critical graphs. J. Combin. Th. B 86 (2002),
408–412. [414]
Sanders P.R., The central automorphisms of a finite group. J. London Math. Soc. 44 (1969), 225–228.
[466]
Sands A.D., On generalised Catalan numbers. Discr. Math. 21 (1978), 219–221. [50]
Santoro N. and J. Urrutia, Angle orders, regular n-gon orders and the crossing number. Order 4 (1987),
209–220. [832]
Sarkaria K.S., A generalized Kneser conjecture. J. Combin. Theory B 49 (1990), 236–240. [812]
Sarvate D.G. and J.-C. Renaud, On the union-closed sets conjecture. Ars Combin. 27 (1989), 149–153.
[503, 512]
Sauer N., A generalization of a theorem of Turán. J. Combinatorial Th. B 10 (1971), 109–112. [476]
Sauer N. and J. Spencer, Edge disjoint placement of graphs. J. Combin. Theory B 25 (1978), 295–302.
[248]
Sauvé L., On chromatic graphs. Amer. Math. Monthly 68 (1961), 107–111. [477]
Savage C.D. and C.Q. Zhang, The connectivity of acyclic orientation graphs. Discr. Math. 184 (1998),
281–287. [312]
Scarpis U., Sui determinanti di valore massimo (Italian). Rendiconti della R. Istituto Lombardo di Sci.
e Lettere 31 (1898), 1441–1446. [622]
Schaal D., On generalized Schur numbers. In Proc. 24th Southeastern Intl. Conf. Combin. Graph Th.
Comput. (Boca Raton), 98 (1993), 178–187. [473]
Schäuble M., Bemerkungen zur Kounstruktion dreikreisfreier ¾ -chromatischer Graphen. Wiss.
Zeitschrift TH Ilmenau 15 (1969), 59–63. [354]
914 References

Schauz U., Mr. Paint and Mrs. Correct. Electron. J. Combin. 16 (2009), Research Paper 77, 18. [741]
Scheim D.E., The number of edge 3-colorings of a planar cubic graph as a permanent. Discr. Math. 8
(1974), 377–382. [740]
Scheinerman E.R., A note on graphs and sphere orders. J. Graph Theory 17 (1993), 283–289. [832]
Scheinerman E.R. and J.C. Wierman, On circle containment orders. Order 4 (1988), 315–318. [828]
Schensted C., Longest increasing and decreasing subsequences. Canad. J. Math. 13 (1961), 179–191.
[193, 195, 198, 206]
Schläfli L., Theorie der vielfachen Kontinuität (German, completed 1852) (George & Co., 1901). Reprinted
in Collected mathematical works Vol. I (Verlag Birkhäuser, 1949). [48]
Schmeichel E.F. and S.L. Hakimi, Pancyclic graphs and a conjecture of Bondy and Chvátal. J. Combin.
Th. B 17 (1974), 22–34. [326]
Schmeichel E.F. and S.L. Hakimi, A cycle structure theorem for Hamiltonian graphs. J. Combin. Theory
B 45 (1988), 99–107. [326]
Schmidt F.W., Problem 10285. Amer. Math. Monthly 100 (1993), 185. [188]
Schmidt F.W., Problem 10364. Amer. Math. Monthly 101 (1994), 177. Solution 104 (1997), 179. [105]
Schmidt F.W., Problem 10481. Amer. Math. Monthly 102 (1995), 840. Solution 104 (1997), 877–878.
[134]
Schmidt F.W., Problem 10629. Amer. Math. Monthly 104 (1997), 974. Solution 107 (2000), 87–8. [152]
Schmuland B., Solution to problem 11590. Amer. Math. Monthly 120 (2013), 760–761. Proposed by Bibak
118 (2011), 653. [176]
Schnyder W., Planar graphs and poset dimension. Order 5 (1989), 323–343. [783, 790]
Schnyder W., Embedding planar graphs on the grid. In Proc. 1st Annual. ACM–SIAM Symp. Discrete.
Algorithms, SODA (SIAM, 1990), 138–148. [394, 783]
Schönberger T., Ein Beweis des Petersenschen Graphensatzes. Acta Scientia Mathematica Szeged 7
(1934), 51–57. [267]
Schönheim J., On a problem of Daykin concerning intersecting families of sets. In Combinatorics (Proc.
British Combinatorial Conf. Univ. Coll. Wales, 1973), London Math. Soc. Lecture Note Ser. 13
(Cambridge Univ. Press, 1974), 139–140. [497, 607]
Schönheim J., Hereditary systems and Chvátal’s conjecture. In Proc. 5th British Combin. Conf. (Ab-
erdeen), Congr. Numer. 15 (Utilitas Math., 1976), 537–539. [512]
Schramm O., Illuminating sets of constant width. Mathematika 35 (1988), 180–189. [817]
Schrijver A., Vertex-critical subgraphs of Kneser graphs. Nieuw Arch. Wisk. (3) 26 (1978), 454–461.
[813, 814]
Schrijver A., Combinatorial optimization. Polyhedra and efficiency. Vol. A, B, C, Algorithms and Combi-
natorics 24 (Springer-Verlag, 2003). [277, 520, 533, 540]
Schröder E., Vier combinatorische Probleme. Zeit. fur Math. 15 (1870), 361–376. [47, 64, 81, 135]
Schur I., Über die Kongruenz x m + y m ≡ m(mod p). Jber. Deutsch. Math.-Verein. 25 (1916), 114–116.
[462, 473]
Schützenberger M.P., A non-existence theorem for an infinite family of symmetrical block designs. Ann.
Eugenics 14 (1949), 286–287. [616]
Schützenberger M.P., A characteristic property of certain polynomials of E.F. Moore and C.E. Shannon.
In RLE Quarterly Progress Report 55 (1959), 117–118. [496]
Schützenberger M.P., Quelques remarques sur une construction de Schensted. Math. Scand. 12 (1963),
117–128. [197, 198, 206, 208]
Schützenberger M.P., Sur un théorème de G. de B. Robinson. C. R. Acad. Sci. Paris Sér. A-B 272 (1971),
A420–A421. [129]
Schützenberger M.P., La correspondance de Robinson. In Combinatoire et représentation du groupe
symétrique (Strasbourg, 1976), Lect. Notes Math. 579 (Springer, 1977), 59–113. [203]
Schwenk A.J., Problem 6434. Amer. Math. Monthly 90 (1983), 403. Solution 94 (1987), 885–887. [782]
Schwenk A.J., Problem E3143. Amer. Math. Monthly 93 (1986), 299. Solution 95 (1988), 352. [206]
Schwer S.R., S-arrangements avec répétitions. C. R. Math. Acad. Sci. Paris 334 (2002), 261–266. [31]
References 915

Scott A., Szemerédi’s regularity lemma for matrices and sparse graphs. Combin. Probab. Comput. 20
(2011), 455–466. [488]
Scott A.D., Induced trees in graphs of large chromatic number. J. Graph Th. 24 (1997), 297–311. [352]
Scott D. and P. Suppes, Foundational aspects of theories of measurement. J. Symb. Logic 23 (1958), 113–
128. [587]
Seberry J., A computer listing of Hadamard matrices. In Combinatorial mathematics (Proc. Internat.
Conf. Combinatorial Theory, Australian Nat. Univ., 1977), Lect. Notes Math. 686 (Springer, 1978),
275–281. [622]
Sedláček J., On the skeletons of a graph or digraph. In Proc. Calgary Intl. Conf. Combin. Structures and
Their Applications (Univ. Calgary, 1969) (Gordon & Breach, 1970), 387–391. [65]
Segner J.A., Enumeratio modorum quibus figurae planae (Latin). Novi Comment. Acad. Sci. Imp.
Petropol. 7 (1759), 203–210. [41]
Seidenberg A., A simple proof of a theorem of Erdős and Szekeres. J. London Math. Soc. 34 (1959), 352.
[430]
Seinsche D., On a property of the class of n-colorable graphs. J. Combin. Th. B 16 (1974), 191–193.
[375, 490]
Sekanina M., On an ordering of the set of vertices of a connected graph. Spisy Přı́rod. Fak. Univ. Brno
1960 (1960), 137–141. [331]
Servedio R. and Y.N. Yeh, A bijective proof on circular compositions. Bull. Inst. Math. Acad. Sinica 23
(1995), 283–293. [144]
Seymour P., Hadwiger’s conjecture. In Open problems in mathematics (Springer, 2016), 417–437.
[349]
Seymour P.D., On incomparable collections of sets. Mathematika 20 (1973), 208–209. [497, 607, 608]
Seymour P.D., A short proof of the matroid intersection theorem (1976). Unpublished note. [530]
Seymour P.D., Disjoint paths in graphs. Discr. Math. 29 (1980), 293–309. [314]
Shamir E. and J. Spencer, Sharp concentration of the chromatic number on random graphs G n ,p. Com-
binatorica 7 (1987), 121–129. [717]
Shannon A.G., Fibonacci and Lucas numbers and the complexity of a graph. Fibonacci Quart. 16 (1978),
1–4. [65]
Shannon C.E., A theorem on coloring the lines of a network. J. Math. Phys. 28 (1949), 148–151.
[358, 360, 365, 372]
Shapira A., Extremal Graph Theory (2016). Lecture notes scribed by Guy Rutenberg. Available online
at http://www.math.tau.ac.il/˜asafico/ext-graph-theory/notes.pdf. [488]
Shapiro L.W., A short proof of an identity of Touchard’s concerning Catalan numbers. J. Combin. Th. A
20 (1976), 375–376. [82]
Shapiro L.W., Problem 10753. Amer. Math. Monthly 106 (1999), 777. Solution 108 (2001), 873–4. [48]
Shapiro L.W. and W. Hamilton, The Catalan numbers visit the world series. Math. Mag. 66 (1993), 20–
22. [65]
Shapiro L.W. and D. Rogers, Problem E3343. Amer. Math. Monthly 96 (1989), 734–735. Solution 97
(1991), 368. [64]
Shearer J.B., A simple counterexample to a conjecture of Rota. Discr. Math. 28 (1979), 327–330. [567]
Shearer J.B., A note on the independence number of triangle-free graphs. Discr. Math. 46 (1983), 83–87.
[663, 682]
Shearer J.B., On a problem of Spencer. Combinatorica 5 (1985), 241–245. [675]
Shearer J.B., On the independence number of sparse graphs. Random Structures Algorithms 7 (1995),
269–271. [663]
Shearer J.B. and D.J. Kleitman, Probabilities of independent choices being ordered. Stud. Appl. Math.
60 (1979), 271–276. [566]
Shelah S., Primitive recursive bounds for van der Waerden numbers. J. Amer. Math. Soc. 1 (1988), 683–
697. [466]
Shepp L.A., The FKG inequality and some monotonicity properties of partial orders. SIAM J. Algebraic
Discrete Methods 1 (1980), 295–299. [608]
916 References

Shepp L.A., The X YZ conjecture and the FKG inequality. Ann. Probab. 10 (1982), 824–827. [599]
Shih W.K. and W.L. Hsu, A new planarity test. Theoret. Comput. Sci. 223 (1999), 179–191. [393]
Shirazi H. and J. Verstraëte, A note on polynomials and º -factors of graphs. Electron. J. Combin. 15
(2008), Note 22, 5. [737, 746]
Shrikhande S.S., The impossibility of certain symmetrical balanced incomplete block designs. Ann.
Math. Statistics 21 (1950), 106–111. [616]
Shrikhande S.S., A note on mutually orthogonal Latin squares. Sankhyā Ser. A 23 (1961), 115–116.
[623]
Sidney J.B., S.J. Sidney, and J. Urrutia, Circle orders, n-gon orders and the crossing number. Order 5
(1988), 1–10. [828, 832]
Silwal S., Problem 12039. Amer. Math. Monthly 125 (2018), 371. Solution 127 (2020), 86. [226]
Simion R. and F.W. Schmidt, Restricted permutations. European J. Combin. 6 (1985), 383–406. [442]
Simonovits M., A method for solving extremal problems in graph theory, stability problems. In Theory
of Graphs (Proc. Colloq., Tihany, 1966) (Academic Press, 1968), 279–319. [479]
Simonovits M., How to solve a Turán type extremal graph problem? (linear decomposition). In Contem-
porary trends in discrete mathematics (Štiřı́n Castle, 1997), DIMACS Ser. Discr. Math. Theoret.
Comput. Sci. 49 (Amer. Math. Soc., 1999), 283–305. [490]
Sinclair A., Lecture notes for Randomness and Computation (Spring 2018) (U California–Berkeley, 2018).
Available at https://people.eecs.berkeley.edu/˜sinclair/cs271/s18.html. [680]
Singleton R.R., On minimal graphs of maximum even girth. J. Combinatorial Theory 1 (1966), 306–332.
[639]
Singmaster D., Problem E2897. Amer. Math. Monthly 88 (1981), 537. Solution by J.W. Grossman 90
(1983), 287–288. [388]
Singmaster D., Reviews: After Math. Puzzles and Brainteasers // The Chicken from Minsk // New Math-
ematical Diversions. Amer. Math. Monthly 105 (1998), 579–587. [70]
Skolem T., Some remarks on the triple systems of Steiner. Math. Scand. 6 (1958), 273–280. [641, 644]
Sleator D.D., R.E. Tarjan, and W.P. Thurston, Rotation distance, triangulations, and hyperbolic geome-
try. J. Amer. Math. Soc. 1 (1988), 647–681. [387]
Slivnik T., Short proof of Galvin’s theorem on the list-chromatic index of a bipartite multigraph. Combin.
Probab. Comput. 5 (1996), 91–94. [364]
Sloane N.J.A., Hamiltonian cycles in a graph of degree 4. J. Combinatorial Theory 6 (1969), 311–312.
[318]
Smith H.J.S., On the value of a certain arithmetical determinant. J. Reine Angew. Math. 251 (1876),
100–109. [177]
Smolenskii E.A., A method for the linear recording of graphs (English translation). U.S.S.R. Comput.
Math. and Math. Phys. 2 (1962), 396–397. [250]
Smoot N., Solution to problem 11908. Amer. Math. Monthly 125 (2018). Proposed 123 (2015), 504.
[148]
Snevily H.St.C., Combinatorics of finite sets (ProQuest LLC, 1991). Ph.D. Thesis, Univ. Illinois, Urbana–
Champaign. [502]
Snevily H.St.C., On generalizations of the de Bruijn–Erdős theorem. J. Combin. Th. A 68 (1994), 232–
238. [730]
Snevily H.St.C., A generalization of Fisher’s inequality. J. Combin. Th. A 85 (1999), 120–125. [730]
Snevily H.St.C., A sharp bound for the number of sets that pairwise intersect at ¾ positive values. Com-
binatorica 23 (2003), 527–533. [730]
Snevily H.St.C. and D.B. West, The bricklayer problem and the strong cycle lemma. Amer. Math. Monthly
105 (1998), 131–143. [50]
Soderberg S. and H.S. Shapiro, A Combinatory Detection Problem. Amer. Math. Monthly 70 (1963),
1066–1070. [513]
Sofair I., Problem 11775. Amer. Math. Monthly 121 (2014), 455. Solution 123 (2016), 508. [174]
Soifer A., The Mathematical Coloring Book: Mathematics of Coloring and the Colorful Life of its Creators
(Springer, 2008). [342]
References 917

Soifer A., Ramsey Theory: Yesterday, Today, and Tomorrow 285 (Birkhauser, 2011). [425]
Soifer A., Progress in my favorite open problem of mathematics, chromatic number of the plane: An
étude in five movements. Geombinatorics 28 (2019), 206–210. [342]
Solymosi J., Note on a generalization of Roth’s theorem. In Discrete and Computational Geometry: The
Goodman–pollack Festschrift (B. Aronov, S. Basu, J. Pach, and M. Sharir, eds.), Algorithms and
Combinatorics 25 (Springer, Berlin, 2003), 825–827. [492]
Sondow J., Problem 11026. Amer. Math. Monthly 110 (2003), 636. Solution 112 (2005), 367–369. [32]
Soneoka T., H. Nakada, M. Imase, and C. Peyrat, Sufficient conditions for maximally connected dense
graphs. Discr. Math. 63 (1987), 53–66. [296]
Sorel J., Problem 11899. Amer. Math. Monthly 123 (2016), 297. Solution 125 (2018), 88. [33]
Sós V.T., P. Erdős, and W.G. Brown, On the existence of triangulated spheres in 3-graphs, and related
problems. Period. Math. Hungar. 3 (1973), 221–228. [684]
Špacapan S., Connectivity of Cartesian products of graphs. Appl. Math. Lett. 21 (2008), 682–685.
[292, 297]
Spencer J., Minimal scrambling sets of simple orders. Acta Math. Acad. Sci. Hungar. 22 (1971/72), 349–
353. [576, 579, 580, 584]
Spencer J., A generalized Rota conjecture for partitions. Studies in Appl. Math. 53 (1974), 239–241.
[567]
Spencer J., Ramsey’s theorem—a new lower bound. J. Combin. Th. A 18 (1975), 108–115. [675]
Spencer J., Asymptotic lower bounds for Ramsey functions. Discr. Math. 20 (1977), 69–76. [682, 685]
Spencer J., Six standard deviations suffice. Trans. Amer. Math. Soc. 289 (1985), 679–706. [722]
Spencer J., Ten lectures on the probabilistic method, CBMS-NSF Regional Conference Series in Applied
Mathematics 52 (SIAM, 1987). [682]
Spencer J., E. Szemerédi, and W. Trotter, Unit distances in the Euclidean plane. In Graph theory and
combinatorics (Cambridge, 1983) (Academic Press, London, 1984), 293–303. [794]
Sperner E., Neuer Beweis für die Invarianz der Dimensionszahl und des Gebietes. Hamburger Abhand.
6 (1928), 265–272. [496, 558, 561, 799, 802, 805]
Spinrad J.P., Two-dimensional partial orders (ProQuest LLC, 1982). Ph.D. Thesis, Princeton Univ.
[573]
Spitzer F., A combinatorial lemma and its application to probability theory. Trans. Amer. Math. Soc. 82
(1956), 323–339. [50]
Stanley R.P., Problem E2315. Amer. Math. Monthly 78 (1971), 904. Solution 79 (1972), 908–910.
[81, 565]
Stanley R.P., Ordered structures and partitions (AMS, 1972). [207]
Stanley R.P., Combinatorial reciprocity theorems. Advances in Math. 14 (1974), 194–253. [761]
Stanley R.P., Generating functions. In Studies in Combinatorics, MAA Stud. Math. 17 (Math. Assoc.
America, 1978), 100–141. [127, 132, 135]
Stanley R.P., Two combinatorial applications of the Aleksandrov–Fenchel inequalities. J. Combin. Th.
A 31 (1981), 56–65. [824, 827]
Stanley R.P., Some aspects of groups acting on finite posets. J. Combin. Th. A 32 (1982), 132–161.
[557, 606]
Stanley R.P., On the number of reduced decompositions of elements of Coxeter groups. European J.
Combin. 5 (1984), 359–372. [208]
Stanley R.P., Enumerative combinatorics. Vol. I, The Wadsworth & Brooks/Cole Mathematics Series
(Wadsworth & Brooks, 1986). [48, 76, 93, 145, 150, 177, 761]
Stanley R.P., Log-concave and unimodal sequences in algebra, combinatorics, and geometry. In Graph
theory and its applications: East and West (Jinan, 1986), Ann. New York Acad. Sci. 576 (New
York Acad. Sci., 1989), 500–535. [564]
Stanley R.P., Problem 10199. Amer. Math. Monthly 99 (1992), 162. Solution 101 (1994), 278–279. [61]
Stanley R.P., Enumerative combinatorics. Vol. 2, Cambridge Studies in Advanced Mathematics 62 (Cam-
bridge Univ. Press, 1999). [42, 127, 129, 135, 136, 189]
Stanley R.P., Problem 11453. Amer. Math. Monthly 116 (2009), 746. Solution 118 (2011), 658–9. [766]
918 References

Stanley R.P., A survey of alternating permutations. In Combinatorics and graphs, Contemp. Math. 531
(AMS, 2010), 165–196. [207]
Stanley R.P., Problem 11610. Amer. Math. Monthly 118 (2011), 937. Solution 120 (2013), 943. [116]
Stanley R.P., Problem 11762. Amer. Math. Monthly 121 (2014), 266. Solution 123 (2016). [150, 206]
Stanley R.P., Catalan Numbers (Cambridge Univ. Press, 2015a). [41, 42]
Stanley R.P., Problem 11838. Amer. Math. Monthly 122 (2015b), 500. [261]
Stanley R.P. and J.M. Steele, Problem E3344. Amer. Math. Monthly 96 (1989), 734. Solution 98 (1991),
649. [669]
Stanton D. and D.E. White, Constructive combinatorics, Undergraduate Texts in Mathematics (Springer-
Verlag, 1986). [100]
Stathopoulos D., Problem 11668. Amer. Math. Monthly 119 (2012), 700. Solution 121 (2014), 743–744.
[62, 135]
Steele J.M., Variations on the monotone subsequence theme of Erdős and Szekeres. In Discrete proba-
bility and algorithms (Minneapolis, 1993), IMA Vol. Math. Appl. 72 (Springer, 1995), 111–131.
[430]
Stein S.K., Convex maps. Proc. Amer. Math. Soc. 2 (1951), 464–466. [393]
Stein S.K., b-sets and coloring problems. Bull. Amer. Math. Soc. 76 (1970), 805–806. [387]
Steiner J., Einige Gesetze über die Theilung der Ebene und des Raumes. J. Reine Angew. Math. 1 (1826),
349–364. [48]
Steiner J., Combinatorische aufgabe. J. Reine Angew. Math. 45 (1853), 181–182. [641]
Steinlein H., Borsuk’s antipodal theorem and its generalizations and applications: a survey. In Topolog-
ical methods in nonlinear analysis, Sém. Math. Sup. 95 (Presses Univ. Montréal, 1985), 166–235.
[806, 815]
Stevens W.L., Solution to a geometrical problem in probability. Ann. Eugenics 9 (1939), 315–320.
[610]
Stiebitz M., K 5 is the only double-critical 5-chromatic graph. Discr. Math. 64 (1987), 91–93. [354]
Stiebitz M. and M. Voigt, List-colourings. In Topics in chromatic graph theory, Encyclopedia Math. Appl.
156 (Cambridge Univ. Press, 2015), 114–136. [346]
Stinson D.R., A short proof of the nonexistence of a pair of orthogonal Latin squares of order six. J.
Combin. Th. A 36 (1984), 373–376. [612]
Stinson D.R., Combinatorial designs (Springer-Verlag, 2004). [609]
Stockmeyer P.K., Solution to problem 10663. Amer. Math. Monthly 107 (2000), 370. Proposed 105 (1998),
464. [92]
Stong R., Solution to problem 10892. Amer. Math. Monthly 110 (2003), 440–441. Proposed 108 (2001),
668. [330]
Stong R., Solution to problem 11086. Amer. Math. Monthly 113 (2006), 372. Proposed 111 (2004), 440.
[343]
Stong R., Solution to problem 1192. Amer. Math. Monthly 114 (2007), 839–840. Proposed 112 (2005),
930. [623]
Stong R., Solution to problem 11931. Amer. Math. Monthly 125 (2018). Proposed 123 (2016), 831.
[461]
Strehl V., Binomial identities—combinatorial and algorithmic aspects. Discr. Math. 136 (1994), 309–
346. [31]
Subba Rao K., Some properties of Fibonacci numbers. Amer. Math. Monthly 60 (1953), 680–684. [60]
Sudakov B., Recent developments in extremal combinatorics: Ramsey and Turán type problems. In Proc.
Inter. Congr. Math IV (Hindustan Book Agency, 2010), 2579–2606. [425]
Suk A., On the Erdős–Szekeres convex polygon problem (2016). (arXiv:1604.08657v1). [446]
Sulanke R.A., Bijective recurrences concerning Schröder paths. Electron. J. Combin. 5 (1998), Research
Paper 47, 11. [64]
Sulanke R.A., Problem 10894. Amer. Math. Monthly 108 (2001), 770. Solution 110 (2003), 443–444.
[62]
References 919

Sulanke R.A., Objects counted by the central Delannoy numbers. J. Integer Seq. 6 (2003), Article 03.1.5,
19 pages. [28, 29, 34]
Sumner D.P., Graphs with 1-factors. Proc. Amer. Math. Soc. 42 (1974), 8–12. [272]
Sumner D.P., 1-factors and antifactor sets. J. London Math. Soc. (2) 13 (1976), 351–359. [273]
Sumner D.P., Subtrees of a graph and the chromatic number. In The theory and applications of graphs
(Kalamazoo, Mich., 1980) (Wiley, 1981), 557–576. [351]
Sved M., Counting and recounting: the aftermath. Math. Intelligencer 6 (1984), 44–45. [40]
Sýkora O. and I. Vrt’o, On crossing numbers of hypercubes and cube connected cycles. BIT 33 (1993),
232–237. [798]
Sylvester J.J., On the change of systems of independent variables. Quart. J. Math. 1 (1857), 42–56.
[37]
Sylvester J.J., Thoughts on orthogonal matrices, simultaneous sign-successions, and tessellated pave-
ments in two or more colours, with applications to Newton’s rule, ornamental tile-work, and the
theory of numbers. Phil. Mag. 34 (1867), 461–475. [618]
Sylvester J.J. and F. Franklin, A Constructive Theory of Partitions, Arranged in Three Acts, an Interact
and an Exodion. Amer. J. Math. 5 (1882), 251–330. [141, 142]
Sysło M.M. and J. Zak, The bandwidth problem: critical subgraphs and the solution for caterpillars. In
Bonn Workshop on Combin. Opt. (Bonn, 1980) (North-Holland, 1982), 281–286. [819]
Szegedy M., The Lovász local lemma—a survey. In Computer science—theory and applications, Lecture
Notes in Comput. Sci. 7913 (Springer, 2013), 1–11. [674]
Székely L.A., Crossing numbers and hard Erdős problems in discrete geometry. Combin. Probab. Com-
put. 6 (1997), 353–358. [794, 798]
Szekeres G. and L. Peters, Computer solution to the 17-point Erdős–Szekeres problem. ANZIAM J. 48
(2006), 151–164. [458]
Szekeres G. and H.S. Wilf, An inequality for the chromatic number of a graph. J. Combin. Th. 4 (1968),
1–3. [338]
Szele T., Combinatorial investigations concerning complete directed graphs (Hungarian). Mat. Fiz.
Lapok 50 (1943), 223–236. [666]
Szemerédi E., On sets of integers containing no four elements in arithmetic progression. In Number
Theory (János Bolyai Math. Soc., Debrecen, 1968) (North-Holland, 1970), 197–204. [484]
Szemerédi E., On graphs containing no complete subgraph with 4 vertices. Mat. Lapok 23 (1972), 113–
116. [492]
Szemerédi E., On sets of integers containing no ¾ elements in arithmetic progression. Acta Arith. 27
(1975), 199–245. [480, 484]
Szemerédi E., Regular partitions of graphs. In Problémes combinatoires et théorie des graphes (Orsay,
1976), Colloq. Internat. CNRS, 260 (1978), 399–401. [480]
Szemerédi E., Arithmetic progressions, different regularity lemmas and removal lemmas. Commun.
Math. Stat. 3 (2015), 315–328. [480, 484]
Szemerédi E. and W.T. Trotter, Extremal problems in discrete geometry. Combinatorica 3 (1983), 381–
392. [794, 798]
Szpilrajn E., Sur l’extension de l’ordre partiel (French). Fundamenta Mathematicae 16 (1930), 386–389.
[569]
’t Woord A.N., Solution to problem 10490. Amer. Math. Monthly 106 (1999), 589. Proposed 102 (1995),
930. [32]
Tagiuri A., On some recurrent sequences with positive integer terms (Italian). Periodico di Mat. (2) 3
(1900), 1–12. [61]
Tait P.G., Remarks on the colourings of maps. Proc. R. Soc. Edinburgh 10 (1880), 729. [399, 400]
Talagrand M., Concentration of measure and isoperimetric inequalities in product spaces. Inst. Hautes
Études Sci. Publ. Math. (1995), 73–205. [712, 720]
Tamassia R. and I.G. Tollis, A unified approach to visibility representations of planar graphs. Discrete
Comput. Geom. 1 (1986), 321–341. [394]
920 References

Tanaka Y., Equivalence of the HEX game theorem and the Arrow impossibility theorem. Appl. Math.
Comput. 186 (2007), 509–515. [803]
Tanner R.M., Explicit construction of concentrators from generalized n-gons. SIAM J. Algebr. Discrete
Meth. 5 (1984), 287–293. [779]
Tao T., Moser’s entropy compression argument (blog post) (2009). Available at http://terrytao.
wordpress.com/2009/08/05/mosers-entropy-compression-argument. [680]
Tao T., Algebraic combinatorial geometry: the polynomial method in arithmetic combinatorics, incidence
combinatorics, and number theory. EMS Surv. Math. Sci. 1 (2014), 1–46. [732]
Tao T. and V. Vu, Additive combinatorics, Cambridge Studies in Advanced Mathematics 105 (Cambridge
Univ. Press, 2006). [488]
Tarjan R.E., A simple version of Karzanov’s blocking flow algorithm. Oper. Res. Letters 2 (1984), 265–
268. [246]
Tarry G., Le problème des labyrinthes (French). Nouv. Ann. Math. 14 (1895), 187–190. [763]
Tarry G., Le problème de 36 officiers. Compte Rendu de l’Assoc. Français Avanc. Sci. Naturel 1 (1900),
122–123. [612]
Tarry G., Le problème de 36 officiers. Compte Rendu de l’Assoc. Français Avanc. Sci. Naturel 2 (1901),
170–203. [612]
Tashkinov V.A., 3-regular subgraphs of 4-regular graphs. Mat. Zametki 36 (1984), 239–259. [736]
Tator C., On the Dimension of Ordered Sets (Technische Hochschule Darmstadt, 1983). Thesis.
[575, 583]
Tauraso R., Problem 11241. Amer. Math. Monthly 113 (2006), 656. Solution 115 (2008), 858–859. [81]
Taylor H., Problem E3448. Amer. Math. Monthly 98 (1991), 553. Solution 100 (1993), 298–300. [91]
Thomas R. and P. Wollan, An improved linear edge bound for graph linkages. European J. Combin. 26
(2005), 309–324. [314]
Thomason A.G., Hamiltonian cycles and uniquely edge colourable graphs. Ann. Discr. Math. 3 (1978),
259–268. [242, 249, 318, 331, 332]
Thomason A.G., An extremal function for contractions of graphs. Math. Proc. Cambridge Philos. Soc. 95
(1984), 261–265. [710]
Thomason A.G., The extremal function for complete minors. J. Combin. Th. B 81 (2001), 318–338.
[710]
Thomassen C., Planarity and duality of finite and infinite graphs. J. Combin. Th. B 29 (1980a), 244–
271. [304, 392]
Thomassen C., 2-linked graphs. European J. Combin. 1 (1980b), 371–378. [314]
Thomassen C., Kuratowski’s Theorem. J. Graph Th. 5 (1981a), 225–241. [392]
Thomassen C., Nonseparating cycles in ¾ -connected graphs. J. Graph Theory 5 (1981b), 351–354.
[305, 313]
Thomassen C., A remark on the factor theorems of Lovász and Tutte. J. Graph Theory 5 (1981c), 441–
442. [262]
Thomassen C., A theorem on paths in planar graphs. J. Graph Th. 7 (1983), 169–176. [334, 401]
Thomassen C., A refinement of Kuratowski’s Theorem. J. Combin. Th. B 37 (1984), 245–253. [393]
Thomassen C., Interval representations of planar graphs. J. Combin. Th. B 40 (1986), 9–20. [394]
Thomassen C., Paths, circuits and subdivisions. In Selected topics in graph theory, 3 (Academic Press,
1988), 97–131. [333, 350]
Thomassen C., Grötzsch’s 3-Color Theorem. J. Combin. Th. B 62 (1994a), 268–279. [414]
Thomassen C., Every planar graph is 5-choosable. J. Combin. Th. B 62 (1994b), 180–181. [402, 742]
Thomassen C., 3-list-coloring planar graphs of girth 5. J. Combin. Th. B 64 (1995a), 101–107. [403]
Thomassen C., Decomposing a planar graph into degenerate graphs. J. Combin. Th. B 65 (1995b), 305–
314. [419]
Thomassen C., A short list color proof of Grötzsch’s theorem. J. Combin. Th. B 88 (2003), 189–192.
[414]
Thrall R.M., A combinatorial problem. Michigan Math. J. 1 (1952), 81–88. [192, 207]
References 921

Thurston W.P., Three-dimensional geometry and topology. Vol. 1, Princeton Mathematical Series 35
(Princeton Univ. Press, 1997). Also 1991. [394]
Tian F., A short proof of a theorem about the circumference of a graph. J. Combin. Th. B 45 (1988), 373–
375. [325]
Tian F., A short proof of Fan’s theorem. Discr. Math. 286 (2004), 285–286. [325]
Tian F. and R.H. Shi, A new class of pancyclic graphs. J. Systems Sci. Math. Sci. 6 (1986), 258–262.
[326]
Timmons C., Star coloring high girth planar graphs. Electron. J. Combin. 15 (2008), Research Paper
124, 17. [422]
Toft B., On the maximal number of edges of critical ¾ -chromatic graphs. Studia Sci. Math. Hungar. 5
(1970), 461–470. [354]
Toft B., Colouring, stable sets and perfect graphs. In Handbook of combinatorics, Vol. 1, 2 (Elsevier,
1995), 233–288. [335]
Toida S., Properties of an Euler graph. J. Franklin Inst. 295 (1973), 343–5. [239]
Tomescu I., Sur le problème du coloriage des graphes généralisés. C. R. Acad. Sci. Paris Sér. A-B 267
(1968), A250–A252. [667]
Tomescu I., Problem E3188. Amer. Math. Monthly 94 (1987), 72. Solution 95 (1988), 876–877. [512]
Tomescu I., Problem E3409. Amer. Math. Monthly 97 (1990), 916. Solution 99 (1991), 860–861. [355]
Touchard J., Sur les cycles des substitutions. Acta Math. 70 (1939), 243–297. [127]
Tracy P., Problem 10811. Amer. Math. Monthly 107 (2000), 566. Solution 109 (2002), 478. [249]
Trotignon N., Perfect graphs. In Topics in chromatic graph theory, Encyclopedia Math. Appl. 156 (Cam-
bridge Univ. Press, 2015), 137–160. [370]
Trotter W.T., Dimension of the crown Sn¾ . Discr. Math. 8 (1974a), 85–103. [584]
Trotter W.T., Irreducible posets with large height exist. J. Combin. Th. A 17 (1974b), 337–344. [583]
Trotter W.T., Inequalities in dimension theory for posets. Proc. Amer. Math. Soc. 47 (1975), 311–316.
[573, 583]
Trotter W.T., A generalization of Hiraguchi’s: inequality for posets. J. Combin. Th. A 20 (1976), 114–
123. [583]
Trotter W.T., The dimension of the Cartesian product of partial orders. Discr. Math. 53 (1985), 255–263.
[572]
Trotter W.T., W.G. Gehrlein, and P.C. Fishburn, Balance theorems for height-2 posets. Order 9 (1992),
43–53. [823]
Trotter W.T. and J.I. Moore, Jr., Some theorems on graphs and posets. Discr. Math. 15 (1976a), 79–84.
[583]
Trotter W.T. and J.I. Moore, Jr., Characterization problems for graphs, partially ordered sets, lattices,
and families of sets. Discr. Math. 16 (1976b), 361–381. [573]
Tsai M.T. and D.B. West, A new proof of 3-colorability of Eulerian triangulations. Ars Math. Contemp. 4
(2011), 73–77. [419]
Tsai S.F., Problem 11987. Amer. Math. Monthly 124 (2017), 563. Solution 126 (2019), March. [551]
Tucker A.W., Some topological properties of disk and sphere. In Proc. First Canadian Math. Congress,
Montreal, 1945 (Univ. Toronto Press, 1946), 285–309. [807, 809]
Turán P., Eine Extremalaufgabe aus der Graphentheorie. Mat. Fiz Lapook 48 (1941), 436–452.
[223, 475, 476, 489, 491]
Tutte W.T., On Hamiltonian circuits. J. Lond. Math. Soc. 21 (1946), 98–101. [242, 401]
Tutte W.T., The factorization of linear graphs. J. Lond. Math. Soc. 22 (1947), 107–111. [264]
Tutte W.T., The dissection of equilateral triangles into equilateral triangles. Proc. Cambridge Philos.
Soc. 44 (1948), 463–482. [749]
Tutte W.T., The factors of graphs. Canad. J. Math. 4 (1952), 314–328. [268, 270, 276]
Tutte W.T., A short proof of the factor theorem for finite graphs. Canad. J. Math. 6 (1954), 347–352.
[269, 270, 276]
Tutte W.T., A theorem on planar graphs. Trans. Amer. Math. Soc. 82 (1956), 99–116. [401]
Tutte W.T., A homotopy theorem for matroids, I, II. Trans. Amer. Math. Soc. 88 (1958), 144–174. [398]
922 References

Tutte W.T., Convex representations of graphs. Proc. Lond. Math. Soc. 10 (1960), 304–320. [392, 783]
Tutte W.T., A theory of 3-connected graphs. Indag. Math. 23 (1961), 441–55. [304, 305, 314, 535]
Tutte W.T., On the problem of decomposing a graph into n connected factors. J. Lond. Math. Soc. 36
(1961a), 221–230. [251]
Tutte W.T., How to draw a graph. Proc. London Math. Soc. (3) 13 (1963), 743–767. [398]
Tutte W.T., Connectivity in Graphs (Toronto Univ. Press, 1966). [314]
Tutte W.T., Introduction to the Theory of Matroids (Amer. Elsevier, 1970). [398, 520]
Tutte W.T., On the 2-factors of bicubic graphs. Discr. Math. 1 (1971), 203–208. [317, 330]
Tutte W.T., The subgraph problem. Ann. Discr. Math. 3 (1978), 289–295. [370]
Tutte W.T. and C.A.B. Smith, On Unicursal Paths in a Network of Degree 4. Amer. Math. Monthly 48
(1941), 233–237. [752]
Tuza Z., Graph coloring in linear time. J. Combin. Th. B 55 (1992), 236–243. [339, 343]
Tverberg H., A proof of the Jordan Curve Theorem. Bull. Lond. Math. Soc. 12 (1980), 34–38. [378]
Tverberg H., On the decomposition of K n into complete bipartite subraphs. J. Graph Th. 6 (1982), 493–
494. [771]
Ungar P., Problem 4385. Amer. Math. Monthly 57 (1950), 189. Solution 58 (1951), 573. [763]
Ungar P., Problem E3052. Amer. Math. Monthly 91 (1984), 438. Solution 94 (1987), 185–186. [175]
Valencia-Pabon M. and J.C. Vera, On the diameter of Kneser graphs. Discr. Math. 305 (2005), 383–385.
[251]
Valiant L.G., The complexity of enumeration and reliability problems. SIAM J. Comput. 8 (1979), 410–
421. [754]
van den Heuvel J. and S. McGuinness, Coloring the square of a planar graph. J. Graph Theory 42 (2003),
110–124. [424]
van der Waerden B.L., Beweis einer Baudetschen Vermutung (in German). Niew. Arch. Wisk. 15 (1927),
212–216. [463, 464]
van der Waerden B.L., Moderne Algebra Vol. 1 (2nd ed.) (Springer-Verlag, 1937). (Many editions.)
[514, 520]
van Lint J.H., Combinatorial Theory Seminar, Eindhoven University of Technology, Lect. Notes in Math.
382 (Springer-Verlag, 1974). [139]
van Lint J.H. and R.M. Wilson, A course in combinatorics (Cambridge Univ. Press, 1992).
[609, 622, 639]
Vanden Eynden C., Problem E3435. Amer. Math. Monthly 98 (1991), 365. Solution 99 (1992), 881–882.
[92]
Vandenbussche J. and D.B. West, Matching extendability in hypercubes. SIAM J. Discr. Math. 23 (2009),
1539–1547. [260]
Vandermonde A.-T., Mémoire sur des irrationanelles de différents ordres avec une application au cercle
(French). Acad. des Sci. (1772). [26]
Vapnik V.N. and A.Y. Chervonenkis, Theory of uniform convergence of frequencies of events to their
probabilities and problems of search for an optimal solution from empirical data. Avtomat. i Tele-
meh. (1971), 42–53. [711, 712]
Vassilev M. and K. Atanassov, On Delanoy numbers. Annuaire Univ. Sofia Fac. Math. Inform. 81 (1994),
153–162. [29]
Vazirani V.V., A theory of alternating paths and blossoms for proving correctness of the O(| V 1/2 || E|)
general graph matching algorithm. Combinatorica 14 (1994), 71–91. [284]
Vazirani V.V. and M. Yannakakis, Pfaffian orientations, 0-1 permanents, and even cycles in directed
graphs. Discrete Appl. Math. 25 (1989), 179–190. [754, 756]
Veršik A.M. and S.V. Kerov, Asymptotic behavior of the Plancherel measure of the symmetric group and
the limit form of Young tableaux. Dokl. Akad. Nauk SSSR 233 (1977), 1024–1027. [722, 821]
Verstraëte J., On arithmetic progressions of cycle lengths in graphs. Combin. Probab. Comput. 9 (2000),
369–373. [479]
References 923

Viennot G., Une forme géométrique de la correspondance de Robinson–Schensted. In Combinatoire et

représentation du groupe symétrique (Actes Table Ronde CNRS, Univ. Louis-Pasteur Strasbourg,
1976), Lect. Notes Math. 579 (Springer, 1977), 29–58. [199, 200]
Vijayaditya N., On total chromatic number of a graph. J. London Math. Soc. (2) 3 (1971), 405–408.
[373]
Vince A., Problem E1771. Amer. Math. Monthly 72 (1965), 316. Solution 73 (1966), 543. [441]
Vince A., Problem 6617. Amer. Math. Monthly 96 (1989), 642. [782]
Vitaver L.M., Determination of minimal coloring of vertices of a graph by means of Boolean powers of
the incidence matrix (Russian). Dokl. Akad. Nauk. SSSR 147 (1962), 758–759. [339]
Vizing V.G., The Cartesian product of graphs. Vyč. Sis. 9 (1963), 30–43. [343]
Vizing V.G., On an estimate of the chromatic class of a p-graph. Diskret. Analiz. 3 (1964), 25–30.
[358, 359, 699]
Vizing V.G., Critical graphs with a given chromatic class (Russian). Diskret. Analiz. 5 (1965), 9–17.
[362, 422]
Vizing V.G., Some unsolved problems in graph theory. Uspekhi Mat. Nauk (Russian Math. Surveys) 23
(1968), 9117–134. [414, 423]
Vizing V.G., Coloring the vertices of a graph in prescribed colors (Russian). Diskret. Analiz. 29 (1976),
3–10. [346, 347, 355, 356]
Vizing V.G. and M.K. Goldberg, The length of a circuit of a strongly connected graph. Kibernetika (Kiev)
(1969), 79–82. [237]
Voigt M., List colourings of planar graphs. Discr. Math. 120 (1993), 215–219. [402]
Voigt M. and B. Wirth, On 3-colorable non-4-choosable planar graphs. J. Graph Th. 24 (1997), 233–235.
[402]
Volkmann L., Bemerkungen zum p-fachen Kantenzusammenhang von Graphen. An. Univ. Bucureşti
Mat. 37 (1988), 75–79. [296]
Volkmann L., Graphen und Digraphen (Springer-Verlag, 1991). [215]
Volkmann L., Regular graphs, regular factors, and the impact of Petersen’s theorems. Jahresber.
Deutsch. Math.-Verein. 97 (1995), 19–42. [268]
Volkmann L., Fundamente der Graphentheorie (Springer-Verlag, 1996). [215]
Voloshin V.I., Properties of triangulated graphs (Russian). In Oper. Research and Progr. (B. A.
Shcherbakov, ed.) (Shtiintsa, 1982), 24–32. [374]
Voloshin V.I., Problem 10976. Amer. Math. Monthly 109 (2002), 855. Solution 111 (2004), 444–445.
[174, 374]
Voloshin V.I. and I.M. Gorgos, Some properties of 1-simply connected hypergraphs and their applica-
tions. Mat. Issled. (1982), 30–33, 191. [374]
von Neumann J., A certain zero-sum two-person game equivalent to the optimal assignment problem.
In Contributions to the theory of games, vol. 2, Annals of Mathematics Studies, no. 28 (Princeton
Univ. Press, 1953), 5–12. [257]
von Staudt K.G.C., Geometrie de Lage, Verlag von Bauer and Rapse 25 (Julius Merz, 1847). [387]
Wagner K., Bemerkungen zum Vierfarbenproblem. Jber. Deutsch. Math. Verein. 46 (1936), 21–22.
[393]
Wagner K., Über eine Eigenschaft der ebenen Komplexe. Math. Ann. 114 (1937), 570–590.
[390, 398, 527]
Wagon S., A bound on the chromatic number of graphs without certain induced subgraphs. J. Combin.
Th. B 29 (1980), 245–246. [343]
Wagon S., Fourteen proofs of a result about tiling a rectangle. Amer. Math. Monthly 94 (1987), 601–617.
[221]
Wagon S. and P. Zielinski, Problem 12082. Amer. Math. Monthly 125 (2018), 945. Solution 127 (2020),
571. [655]
Wallis W.D., Combinatorial designs, Monographs and Textbooks in Pure and Applied Mathematics 118
(Marcel Dekker, 1988). Also 2007. [609]
924 References

Wallis W.D. (ed.), Computational and constructive design theory, Mathematics and its Applications 368
(Kluwer Acad., 1996). [609]
Wallis W.D., A beginner’s guide to graph theory (Birkhäuser Boston, 2000). [215]
Wallis W.D. (ed.), Designs 2002, Mathematics and its Applications 563 (Kluwer Acad., 2003). [609]
Walsh T.R., The towers of Hanoi revisited: moving the rings by counting the moves. Inform. Process. Lett.
15 (1982), 64–67. [80]
Walsh T.R., Iteration strikes back—at the cyclic Towers of Hanoi. Inform. Process. Lett. 16 (1983), 91–
93. [80]
Walter J.R., Representations of rigid cycle graphs (Wayne State Univ., 1972). Ph.D. Thesis. [373]
Walters I.C., Jr., The ever expanding expander coefficients. Bull. Inst. Combin. Appl. (1996), 97. [779]
Wang D.L. and D.J. Kleitman, On the existence of n-connected graphs with prescribed degrees (n ≥ 2).
Networks 3 (1973), 225–239. [226]
Wang D.L. and P. Wang, Some results about the Chvátal conjecture. Discr. Math. 24 (1978), 95–101.
[501]
Wang J.F., D.B. West, and B. Yao, Maximum bandwidth under edge addition. J. Graph Theory 20 (1995),
87–90. [819]
Wang W. and K.W. Lih, Choosability and edge choosability of planar graphs without five cycles. Appl.
Math. Lett. 15 (2002), 561–565. [413]
Wang Y., Notes on Chvátal’s conjecture. Discr. Math. 247 (2002), 255–259. [503]
Wardlaw W.P., Problem E3358. Amer. Math. Monthly 96 (1989), 928. Solution 98 (1991), 650.
[116, 173]
Warren H.E., Lower bounds for approximation by nonlinear manifolds. Trans. Amer. Math. Soc. 133
(1968), 167–178. [830]
Watkins M.E., A lower bound for the number of vertices of a graph. Amer. Math. Monthly 74 (1967), 297.
[313]
Watkins M.E., On the existence of certain disjoint arcs in graphs. Duke Math. J. 35 (1968), 231–246.
[314]
Weaver W., Questions, Discussions, and Notes: Lewis Carroll and a Geometrical Paradox. Amer. Math.
Monthly 45 (1938), 234–236. [60]
Wei V.K., A Lower Bound on the Stability Number of a Simple Graph. Tech. Rep. TM 81-11217-9, Bell
Laboratories (1981). [263, 662]
Weinstein J.M., On the number of disjoint edges in a graph. Canad. J. Math. 15 (1963), 106–111.
[263]
Welsh D.J.A., Matroid Theory (Academic Press, 1976). [514, 520, 532]
Welsh D.J.A., Colouring problems and matroids. In Surveys in combinatorics (Proc. 7th British Combi.
Conf., 1979), London Math. Soc. Lecture Note Ser. 38 (Cambridge Univ. Press, 1979), 229–257.
[538, 539]
Welsh D.J.A. and M.B. Powell, An upper bound for the chromatic number of a graph and its application
to timetabling problems. Computer J. 10 (1967), 85–87. [337]
Wendel J.G., A problem in geometric probability. Math. Scand. 11 (1962), 109–111. [669]
Wernicke P., Über den kartographischen Vierfarbensatz. Math. Ann. 58 (1904), 413–426. [406]
West D.B., Pairs of adjacent Hamiltonian circuits with small intersection. Stud. Appl. Math. 59 (1978),
245–248. [331]
West D.B., A symmetric chain decomposition of L(4 , n). European J. Combin. 1 (1980), 379–383.
[557, 565]
West D.B., Extremal problems in partially ordered sets. In Ordered sets (Banff, Alta., 1981), NATO Adv.
Study Inst. C: Math. Phys. Sci. 83 (Reidel, 1982a), 473–521. [493]
West D.B., The number of complete subgraphs in graphs with non-majorizable degree sequences. In
Proc. Silver Jubilee Conf. (Waterloo, 1982), Progress in Graph Theory (Academic Press, 1982b),
509–521. [491]
West D.B., “Poly-unsaturated” posets: the Greene–Kleitman theorem is best possible. J. Combin. Th. A
41 (1986), 105–116. [551]
References 925

West D.B., Introduction to graph theory (Prentice Hall, 1996). Also 2001. [215]
West D.B., The superregular graphs. J. Graph Th. 23 (1996), 289–295. [782]
West D.B., Short proofs for interval digraphs. Discr. Math. 178 (1998), 287–292. [605]
West D.B., Introduction to graph theory, 2nd ed. (Prentice Hall, 2001). Also 1996. [539]
West D.B., A short proof of the Berge–Tutte formula and the Gallai–Edmonds structure theorem. Euro-
pean J. Combin. 32 (2011), 674–676. [275]
West D.B., L.H. Harper, and D.E. Daykin, Some remarks on normalized matching. J. Combin. Th. A 35
(1983), 301–308. [564, 567]
West D.B. and D.J. Kleitman, Skew chain orders and sets of rectangles. Discr. Math. 27 (1979), 99–102.
[565]
West D.B., W.T. Trotter, G.W. Peck, and P. Shor, Regressions and monotone chains: a Ramsey-type ex-
tremal problem for partial orders. Combinatorica 4 (1984), 117–119. [442]
West D.B. and W.H. Wiedemann, Problem E3290. Amer. Math. Monthly 95 (1988), 872. Solution 97
(1990), 428–429. [176]
West Don, Solution to problem E2404. Amer. Math. Monthly 81 (1974), 287. Proposed 80 (1973), 316.
[106]
White D.E. and S.G. Williamson, Recursive matching algorithms and linear orders on the subset lattice.
J. Combinatorial Theory A 23 (1977), 117–127. [566]
Whiting P.D. and J.A. Hillier, A method for finding the shortest route through a road network. Operations
Research Quart. 11 (1960), 37–40. [246]
Whitney H., A theorem on graphs. Ann. of Math. (2) 32 (1931), 378–390. [401, 418]
Whitney H., Congruent graphs and the connectivity of graphs. Amer. J. Math. 54 (1932a), 150–168.
[293, 300, 306, 314]
Whitney H., A logical expansion in Mathematics. Bull. Amer. Math. Soc. 38 (1932b), 572–579. [158]
Whitney H., Non-separable and planar graphs. Trans. Amer. Math. Soc. 34 (1932c), 339–362. [526]
Whitney H., 2-isomorphic graphs. Amer. J. Math. 55 (1933), 245–254. [380, 398]
Whitney H., On the abstract properties of linear dependence. Amer. J. Math. 57 (1935), 509–533.
[514, 520, 523, 536]
Whitworth W.A., Choice and Chance (2nd ed.) (Deighton, Bell & Co., 1897). First edition 1867.
[33, 40, 46, 115]
Wiener N., A contribution to the theory of relative position. Proc. Cambridge Philos. Soc. 17 (1914), 441–
449. [586, 605]
Wilf H.S., The eigenvalues of a graph and its chromatic number. J. Lond. Math. Soc. 42 (1967), 330–
332. [769]
Wilf H.S., The friendship theorem. In Combinatorial Mathematics and Its Applications (Oxford, 1969)
(Academic Press, 1971), 307–309. [775]
Wilf H.S., generatingfunctionology (Academic Press, 1990). [87, 112, 117, 129, 135, 136, 140]
Wilf H.S., Problem 10578. Amer. Math. Monthly 104 (1997), 270. Solution 106 (1999), 169. [91]
Will T.G. and H. Hulett, Parsimonious multigraphs. SIAM J. Discr. Math. 18 (2004), 241–245. [238]
Williamson J., Hadamard’s determinant theorem and the sum of four squares. Duke Math. J. 11 (1944),
65–81. [622]
Wilson R.M., Constructions and uses of pairwise balanced designs. In Combin. (Proc. NATO Advanced
Study Inst., 1974) Part 1: Theory of designs, finite geometry and coding theory, Centre Tracts 55
(Math. Centrum, 1974), 18–41. [641]
Wilson R.M., An existence theory for pairwise balanced designs, II–III. J. Combin. Th. A 18 (1975), 71–
79. [640]
Wilson R.M., The exact bound in the Erdős–Ko–Rado theorem. Combinatorica 4 (1984),247–257.
[500]
Win S., Existenz von Gerüsten mit vorgeschriebenem Maximalgrad in Graphen. Abh. Math. Sem. Univ.
Hamburg 43 (1975), 263–267. [327, 328]
Win S., On a conjecture of Las Vergnas concerning certain spanning trees in graphs. Resultate Math. 2
(1979), 215–224. [327]
926 References

Win Z., On the windy postman problem on Eulerian graphs. Math. Programming 44 (1989), 97–112.
[327]
Winkler P., Random orders. Order 1 (1985), 317–331. [821]
Winkler P.M., Average height in a partially ordered set. Discr. Math. 39 (1982), 337–341. [608]
Winkler P.M., Proof of the squashed cube conjecture. Combinatorica 3 (1983), 135–139. [781]
Winkler P.M., Correlation and order. In Combinatorics and ordered sets (Arcata, 1985), Contemp. Math.
57 (Amer. Math. Soc., 1986), 151–174. [600]
Winkler P.M., Mathematical puzzles: a connoisseur’s collection (A K Peters, 2004). [221, 667]
Winkler P.M., Puzzled: Solutions and sources. Commun. ACM 51 (2008), 118–118. [431]
Wismath S.K., Characterizing bar line-of-sight graphs. In Proc. 1st Symp. Comput. Geo. (Baltimore,
1985) (ACM, 1985), 147–152. [394]
Witt E., Ein kombinatorischer Satz der Elementargeometrie (in German). Math. Nachr. 6 (1952), 261–
262. [467]
Wood D., The towers of Brahma and Hanoi revisited. J. Recreational Math. 14 (1981/82), 17–24. [80]
Woodall D.R., Sufficient conditions for circuits in graphs. Proc. Lond. Math. Soc. 24 (1972), 739–755.
[323]
Woodall D.R., An exchange theorem for bases of matroids. J. Combin. Th. B 16 (1974), 227–228. [538]
Woodall D.R., Cyclic-order graphs and Zarankiewicz’s crossing-number conjecture. J. Graph Th. 17
(1993), 657–671. [793, 798]
Woodall D.R., The average degree of an edge-chromatic critical graph. II. J. Graph Th. 56 (2007), 194–
218. [414]
Woodall D.R., The average degree of a multigraph critical with respect to edge or total choosability. Discr.
Math. 310 (2010), 1167–1171. [414]
Woolhouse W.S.B., Prize question 1733. Lady’s and Gentleman’s Diary (1844). [641]
Wormald N.C., Some problems in the enumeration of labelled graphs (Univ. Newcastle, 1978). Ph.D.
Thesis. [697, 704]
Wormald N.C., The differential equation method for random graph processes and greedy algorithms. In
Lectures on Approximation and Randomized Algorithms (M. Karonski and H.J. Proemel, eds.)
(PWN, Warsaw, 1999a), 73–155. [722]
Wormald N.C., Models of random regular graphs. In Surveys in combinatorics, 1999 (Canterbury), Lon-
don Math. Soc. Lecture Note Ser. 267 (Cambridge Univ. Press, 1999b), 239–298. [697]
Worpitzky J., Studien über die bernoullischen und eulerschen Zahlen (German). J. Reine Angewandte
94 (1883), 203–232. [101]
Wu J.L., On the linear arboricity of planar graphs. J. Graph Theory 31 (1999), 129–134. [412, 422]
Wu J.L. and Y.W. Wu, The linear arboricity of planar graphs of maximum degree seven is four. J. Graph
Theory 58 (2008), 210–220. [412]
Xiao S., Solution to problem 11509. Amer. Math. Monthly 119 (2012), 430. Proposed 117 (2010), 558.
[24]
Xiao Y. and H. Zhao, New method for counting the number of spanning trees in a two-tree network. Phys.
A 392 (2013), 4576–4583. [65]
Yaglom A.M. and I.M. Yaglom, Challenging mathematical problems with elementary solutions. Vol. I:
Combinatorial analysis and probability theory, Translated by James McCawley, Jr.; revised and
edited by Basil Gordon (Holden-Day, 1964). [24]
Yamamoto K., Logarithmic order of free distributive lattice. J. Math. Soc. Japan 6 (1954), 343–353.
[558]
Yamashita T., Degree sum conditions and Chvátal–Erdős conditions for cycles. In The COE Seminar on
Mathematical Sciences 2004, Sem. Math. Sci. 31 (Keio Univ., 2004), 185–199. [316]
Yan C.H., Parking functions. In Handbook of enumerative combinatorics, Discr. Math. Appl. (Boca Raton)
(CRC Press, 2015), 835–893. [50]
Yannakakis M., The complexity of the partial order dimension problem. SIAM J. Algebraic Discrete
Methods 3 (1982), 351–358. [573]
Yao A.C.C., Should tables be sorted? J. Assoc. Comput. Mach. 28 (1981), 615–628. [447, 448]
References 927

Yao F.F., D.P. Dobkin, H. Edelsbrunner, and M.S. Paterson, Partitioning space for range queries. SIAM
J. Comput. 18 (1989), 371–384. [815]
Yap H.P., Total colourings of graphs. Bull. London Math. Soc. 21 (1989), 159–163. [373, 654]
Yates F., A new method of arranging field trials involving a large number of varieties. J. Agric. Sci. Comb.
26 (1936), 424–455. [612]
Yeh Y.N., A remarkable endofunction involving compositions. Stud. Appl. Math. 95 (1995), 419–432.
[144]
Young A., Quantitative substitutional analysis I. Proc. London Math. Soc. 33 (1901), 97–146. [189]
Young A., Quantitative substitutional analysis II. Proc. London Math. Soc. 35 (1902), 361–397.
[190, 207]
Yu X., Problem 10575. Amer. Math. Monthly 104 (1997), 168–169. Solution 106 (1999), 266–268. [134]
Zaker M., On lower bounds for the chromatic number in terms of vertex degree. Discr. Math. 311 (2011a),
1365–1370. [341]
Zaker M., Bounds for chromatic number in terms of even-girth and booksize. Discr. Math. 311 (2011b),
197–204. [342]
Zaks J., Towards a simpler proof of the Brouwer fixed point theorem. Geombinatorics 5 (1995), 35–37.
[804]
Zamani R. and D.B. West, Spanning cycles through specified edges in bipartite graphs. J. Graph Theory
71 (2012), 1–17. [326]
Zarankiewicz K., On a problem of P. Turán concerning graphs. Fund. Math. 41 (1954), 137–145. [793]
Zaslavsky T., Glossary and bibliography of signed and gain graphs and allied areas. Electron. J. Combin.
5 (1998), Dynamic Surveys 8, 124 and 9, 41. [237]
Zeckendorf E., A generalized Fibonacci numeration. Fibonacci Quart. 10 (1972), 365–372. [61]
Zehavi A. and A. Itai, Three tree-paths. J. Graph Theory 13 (1989), 175–188. [315]
Zeilberger D., Sister Celine’s technique and its generalizations. J. Math. Anal. Appl. 85 (1982), 114–145.
[90]
Zeilberger D., A short hook-lengths bijection inspired by the Greene–Nijenhuis–Wilf proof. Discr. Math.
51 (1984), 101–108. [192]
Zeilberger D., Kathy O’Hara’s constructive proof of the unimodality of the Gaussian polynomials. Amer.
Math. Monthly 96 (1989), 590–602. [557]
Zeilberger D., A fast algorithm for proving terminating hypergeometric identities. Discr. Math. 80 (1990),
207–211. [90]
Zeilberger D., The method of creative telescoping. J. Symbolic Comput. 11 (1991), 195–204. [90]
Zeilberger D., Reverend Charles to the aid of Major Percy and Fields medalist Enrico. Amer. Math.
Monthly 103 (1996), 501–502. [178]
Zeilberger D., Dodgson’s determinant-evaluation rule proved by two-timing men and women. Electron.
J. Combin. 4 (1997), Research Paper 22. [178]
Zhan S., On Hamiltonian line graphs and connectivity. Discr. Math. 89 (1991), 89–95. [326]
Zhang K.M. and Z.M. Song, Cycles in digraphs—a survey. Nanjing Daxue Xuebao Ziran Kexue Ban 27
(1991), 188–215. [316]
Zhang L., Every planar graph with maximum degree 7 is of class 1. Graphs Combin. 16 (2000), 467–
495. [422]
Zhang L. and B. Wu, Edge choosability of planar graphs without small cycles. Discr. Math. 283 (2004),
289–293. [413]
Zhou B., A note on the Erdős–Sós conjecture. Acta Math. Sci. (English Ed.) 4 (1984), 287–289. [248]
Zhou L., Problem 11187. Amer. Math. Monthly 112 (2005), 929. Solution 114 (2007), 554. [66]
Zhu Q.C., The structure of -critical graphs with | V(G)|− 2 (G) = 3. In Graph theory and its applications:
East and West (Jinan, 1986), Ann. New York Acad. Sci. 576 (1989), 716–722. [458]
Zhu X., The fractional version of Hedetniemi’s conjecture is true. European J. Combin. 32 (2011), 1168–
1175. [461]
Zhu X., The Alon–Tarsi number of planar graphs. J. Combin. Th. B 134 (2019), 354–358. [742]
928 References

Zhu Y.J., Z.H. Liu, and Z.G. Yu, An improvement of Jackson’s result on Hamilton cycles in 2-connected
regular graphs. In Cycles in Graphs (Burnaby, 1982) (B. Alspach and C. Godsil, eds.) (North-
Holland, 1985), 237–247. [326]
Ziegler G.M., Generalized Kneser coloring theorems with combinatorial proofs. Invent. Math. 147 (2002),
671–691. [812]
Zito J., The structure and maximum number of maximum independent sets in trees. J. Graph Theory
15 (1991), 207–221. [264]
Znám Š., Two improvements of a result concerning a problem of K. Zarankiewicz. Colloq. Math. 13
(1965), 255–258. [631]
Zykov A.A., On some properties of linear complexes (Russian). Mat. Sbornik 24 (1949), 163–188.
[354, 476]
Author Index

Agarwal, P. K. 783 Arnautov, V. I. 670 Basin, S. L. 61

Aigner, M. 37, 165, 168, 170, Arratia, R. 432–3 Batagelj, V. 330
176–7, 476, 811 Arrow, K. J. 585 Bauer, D. 318, 326, 490
Ajtai, M. 226, 451, 492, 663–4, Ash, P. 34 Baumert, L. D. 622, 637
779, 793 Atanassov, K. 29 Bean, D. R. 342
Akin, E. 143–4, 152 Avis, D. 820 Beck, J. 455, 659, 673, 680
Akiyama, J. 268, 412, 418–9, Axenovich, M. 686 Beckwith, D. 32, 49, 59, 148–9,
677, 816 Ayel, J. 331 152, 173, 764
Aksionov, V. A. 414 Azuma, K. 715, 722 Behrend, F. A. 484
Albert, M. H. 433 Behzad, M. 329, 343, 373, 398,
Albertson, M. O. 356, 419, 423, 654
Babai, L. 699, 705, 723, 727–8,
797 Beineke, L. W. 620–1, 624, 794,
731, 744
Aldred, R. E. L. 326 798
Bäbler, F. 250, 267
Alexanderson, G. L. 388 Belck, H.-B. 268, 276
Bach, E. 105
Alkan, E. 80 Belevitch, V. 621
Bacher, R. 66
Alladi, K. 59 Bender, E. A. 62, 127, 207, 697,
Bachmann, P. 10
Alon, N. 356, 396–7, 412, 434, 704
Backelin, J. 442
442, 503–4, 510, 633, 657, Benevides, F. S. 461
Bagga, J. S. 325
661–6, 669–74, 677–8, 684–5, Benhocine, A. 326
Baker, K. A. 561, 571
688, 696, 704, 706, 721–2, Benjamin, A. T. 53, 664
Balko, M. 457
728, 731, 733–6, 738–9, Bennett, G. 152
Ball, S. 633
745–6, 764, 771, 778–80, Bentz, H.-J. 144, 152
Ball, W. W. R. 70
809–10, 816, 829–30, 832 Benzer, S. 337
Balof, B. 586
Alspach, B. 646–7 Berge, C. 222, 263, 266–7, 277,
Balogh, J. 421, 424, 801, 819
Altinisik, E. 177, 766 333, 335, 366–8, 370, 501,
Banderier, C. 28, 41
Amahashi, A. 273 549, 667
Bandlow, J. 190
Andersen, L. D. 360, 373 Berger, E. 549
Bang, C. M. 261
Anderson, I. 264, 275, 497, 561, Berlekamp, E. R. 724, 744
Bang-Jensen, J. 685
595, 607, 609, 612, 649, 654 Berman, D. M. 419
Bang, S.-J. 32
Anderson, M. 794 Berman, K. A. 241, 249
Bannai, E. 244, 775
Ando, K. 249, 819 Bermond, J.-C. 316, 323, 685
Bapat, R. B. 766, 769
André, D. 40, 208 Bernardi, C. 355
Bárány, I. 810, 812, 820
Andreae, T. 366 Bernshteyn, A. 680
Barát, J. 238
Andrews, G. 148 Bertrand, J. 39
Barber, B. 640
Andrews, G. E. 66, 140, 150–1 Bessy, S. 685
Barra, M. 62
Andrews, P. 174 Beth, T. 609
Barrera-Cruz, F. 790
Appel, K. 402, 406 Beutelspacher, A. 473
Basavaraju, M. 329
929
930 Author Index

Bey, C. 511 Bouchet, A. 273 Cao, W. 731

Bhanu, K. S. 669 Boyer, J. 393 Carlitz, L. 101, 177
Bhasker, J. 353 Brandstädt, A. 370 Caro, N. 174
Bhatt, S. N. 809 Brandt, A. 422 Caro, Y. 236, 263, 327, 461, 662
Bialostocki, A. 490 Brandt, J. 143–5 Cartwright, D. P. 237
Biedl, T. 286 Brandt, S. 262, 356, 454 Catalan, E. 41
Bienstock, D. 795–6 Branin, F. H., Jr. 763 Catlin, P. A. 349, 355–6
Biggs, N. 766 Brègman, L. M. 510 Cayley, A. 36–7, 47
Biggs, N. L. 37, 46 Brestovansky, W. 473 Chaiken, S. 749
Binet, J. P. 67 Brightwell, G. R. 394, 543, 600, Chambers, E. W. 668
Birkhoff, G. 257, 405, 421, 594, 722, 790, 823, 831 Chandran, L. S. 329
607 Broersma, H. 320, 327 Chandrasekharan, K. 140
Biró, C. 580 Broersma, H. J. 318, 342 Chang, G. J. 373
Bixby, R. E. 521 Broline, D. M. 103, 144 Chang, W. I. 439
Björner, A. 804 Brooks, R. L. 336 Chappell, G. G. 551
Blagojević, P. V. M. 815 Brouwer, A. E. 735, 766, 771 Charalambides, C. A. 101
Blass, A. 690 Brouwer, L. E. J. 805 Charbit, P. 512
Blokhuis, A. 725, 744 Brown, J. L. 441 Chartrand, G. 215, 250, 272–3,
Bloom, D. M. 33 Brown, T. C. 473 295, 315, 343, 377, 398, 418
Bloome, L. 344 Brown, W. G. 228, 632–3, 684 Chein, M. 312
Boesch, F. 237 Brualdi, R. A. 34, 174, 539 Chelnokov, V. M. 777
Bogart, K. P. 147, 583, 586–7 Bruck, R. H. 616, 627 Chen, G. 249, 314, 318, 473
Bognár, J. 105 Bruhn, H. 503, 512 Chen, Z.-H. 540
Bohman, T. 665, 682, 685, 771 Bryant, D. 646–8 Cherkashin, D. D. 660, 673
Bollobás, B. 215, 238, 244, 250, Bryant, P. R. 763 Chernoff, H. 707
272, 311, 316, 335, 351, 362, Brylawski, T. H. 529–30, 539 Chervonenkis, A. Y. 711–2
373, 478, 482, 488, 499, 511, Bucić, M. 460, 685 Chetwynd, A. G. 359
568, 632, 657, 668, 688, 692, Buckholtz, T. J. 65 Chevalley, C. 735
697, 699–701, 703, 717–8, Buckley, F. 215 Chevalley, H. 735
721–2, 770, 821 Bukh, B. 479 Chiba, S. 326
Bóna, M. 101–2, 106, 133, 158, Bunde, D. P. 668 Chiue, W.-S. 292
432, 440, 442, 474 Buneman, P. 80, 373 Choi, J. O. 239
Bonamy, M. 413, 685 Buratti, M. 646 Chow, T. 175
Bonato, A. 708 Burns, D. 251 Chowla, S. 616
Bondarenko, A. 817 Burr, S. A. 454, 456–7, 459–61, Chudnovsky, M. 370
Bondy, J. A. 215, 227–8, 251, 484 Chung, F. R. K. 249, 341, 446,
296, 302, 316, 320, 323, Buršteı̆n, M. I. 418–9 451, 460, 473, 491, 508–9,
326–7, 330, 332, 460, 479, Burungale, A. 238 776, 779
489, 639, 684 Butler, S. 237, 425 Chung, K. L. 50
Bonnice, W. E. 458 Butterfield, J. 457 Chung, M.-S. 490
Booth, K. S. 393 Chvátal, V. 229, 318, 320–1,
Boppana, R. B. 504 Cabello, S. 795 326, 332, 354, 370, 375, 389,
Borchardt, C. W. 37 Callan, D. 35, 49, 59, 63–4, 81, 419, 453–5, 458, 483, 501–2,
Borg, P. 500, 503 134, 151, 174 666, 793
Borodin, O. V. 348, 355, 364–5, Cameron, K. 241–2, 249, 276 Chvátalová, J. 800, 818
407, 411–6, 423 Cameron, K. B. 549 Cibulka, J. 457
Borozan, V. 373 Cameron, P. J. 261, 609, 774 Cigler, J. 50
Borsuk, K. 809–10, 817 Camion, P. 334 Clapham, C. R. J. 332
Borůvka, O. 246, 520 Campbell, C. 238 Clark, D. S. 429
Bosák, J. 319, 331, 418 Canfield, E. R. 567, 697, 704 Clark, L. H. 764
Bose, R. C. 610, 612, 615, 627, Cannings, C. 144 Clark, P. L. 732
641–3, 651–5 Cantor, D. G. 514 Clements, G. F. 496, 511
Author Index 931

Cohen-Addad, V. 414 Descartes, B. 434 Edmonds, J. 241–2, 249, 251,

Cohen, N. 373, 413 Deshpande, B. 92, 668 264, 275–6, 278, 282, 297,
Cohn, M. 820 Deshpande, M. N. 92, 104, 115, 518, 521, 530, 533–4, 536
Colbourn, C. J. 609, 641, 646 668–9 Edwards, K. 416
Coleman, M. 434, 442 DeTemple, D. 389 Egawa, Y. 327
Collins, K. L. 418 Deuber, W. 462 Eğecioğlu, Ö. 36
Comtet, L. 65 Deutsch, E. 21–2, 48–50, 59, 62, Egerváry, E. 258, 531
Conlon, D. 334, 425, 450–1, 453, 82, 105, 148–9, 151, 164 Eggleston, H. G. 817
457, 459, 484–5, 608 Deza, M. 510–1, 728 Egorychev, G. P. 256, 765
Cook, S. 11 Dias da Silva, J. A. 734, 746 Eitner, P. G. 818
Corradi, K. 490 Dı́az-Barrero, J. L. 32 Elder, M. 433
Cox, D. A. 173 Dickson, L. E. 48, 473 Eliahou, S. 745
Cranston, D. W. 152, 336, 373, Diestel, R. 215, 377, 488 Elkies, N. 34
407, 411, 413, 422–4, 668 Dijkstra, E. W. 246, 252 Ellingham, M. 334
Crapo, H. H. 521 Dilcher, K. 32 Ellingham, M. N. 326–7, 330
Cruse, A. B. 274 Dilworth, R. P. 546, 574, 606–7 Ellis, D. 509
Csirik, J. A. 801 Dinitz, J. H. 609, 646 Enchev, O. 261
Cull, P. 80, 237 Dirac, G. A. 301–2, 315–6, 320, Engel, A. 226
Curran, S. 315 323, 345, 349, 351, 353, 355, Engel, K. 493, 496, 511
Cushing, W. 742 367, 490 Enomoto, H. 318, 355, 412
Cvetković, D. M. 766, 781 Djidjev, H. N. 397 Entringer, R. C. 330, 347, 428,
Czipzer, J. 216 Djokovic, D. 32 441
Dmitriev, I. G. 374 Era, H. 418
Dályay, P. P. 115, 623 Dobiński, G. 134 Erdős, P. 219, 223, 238, 244,
Damerell, R. M. 244, 775 Dobkin, D. P. 815 251–2, 269, 273–6, 321–2,
Danese, A. E. 61 Dodgson, C. L. 170, 178 332–4, 341–8, 354–6, 394,
Danziger, P. 647–8 Dol’nikov, V. L. 812 426, 430, 434, 438, 446,
David, K. 47 Doob, M. 766, 781 450–1, 454–61, 468–70, 473,
Davis, M. 143–4, 152 Došlić, T. 275 476–9, 482–4, 489–91, 497–8,
Daykin, D. E. 496–7, 500–1, Doster, D. 82, 91 503–4, 512–4, 551, 561, 608,
511–2, 564, 567, 596, 607–8 Doubilet, P. 127 632, 658–60, 664, 667–71,
de Abreu, N. M. M. 778 Dross, F. 640 675, 679, 682–7, 698–9,
de Bruijn, N. G. 239, 468, 554, Dubhashi, D. P. 706, 721 701–5, 709, 721, 734, 764,
752 Dudek, A. 276, 455 770, 793, 798
de Figueiredo, C. M. H. 798 Duffus, D. 601 Erdős, P. L. 47, 135
de Fraysseix, H. 783 Dujmović, V. 680 Erman, R. 424
de Grey, A. 342 Dulmage, A. L. 654 Eršov, A. P. 353
de Moivre, A. 67, 153 Duncan, C. A. 286 Esperet, L. 680, 765
de Werra, D. 372 Duraj, L. 742 Etienne, G. 144, 152
Dean, A. M. 794 Dushnik, B. 569–70, 576–7, 831 Euler, L. 141, 148, 151, 234, 383
Dean, N. 249, 795–6 Dvir, Z. 746 Evans, A. B. 624
Dean, R. A. 587, 604 Dvořák, Z. 356, 414 Even, S. 343
DeBiasio, L. 327, 334, 460 Dvoretzky, A. 42 Everman, D. 61
Dekster, B. V. 817 Dzhumadiĺdaeva, A. A. 115 Exoo, G. 412, 677
Delannoy, H. 28–9 Dziobek, O. 46
Delcourt, M. 685 Faà di Bruno, F. 127
Delsarte, P. 744 Faigle, U. 533
Demaine, E. D. 286 Eaton, N. 484, 745 Fajtlowicz, S. 709
Demoucron, G. 393 Eckhoff, J. 496 Falcón, S. 80
Dénes, J. 610, 623 Ecklund, E. F. 80 Falgas-Ravry, V. 475
Derbyshire, J. 10 Edelman, P. 208 Falikman, D. I. 256, 765
DeSario, R. 442 Edelsbrunner, H. 815 Fan, G. 324–6
932 Author Index

Farber, M. 374 Franklin, F. 141–2, 151 George, J. C. 273, 371, 764

Faria, L. 798 Franklin, P. 406 Georges, J. P. 317
Fáry, I. 393 Franzblau, D. S. 192 Gerbner, D. 493
Farzad, B. 237 Fraughnaugh (Jones), K. 420 Gerencsér, L. 460
Fasenmyer, M. C. 90 Fraughnaugh, K. 550 Gerke, S. 488
Faudree, R. J. 262, 316, 327, Fredman, M. L. 92, 823 Gervacio, S. V. 249, 418, 819
454–5 Freund, R. M. 807, 809 Gessel, I. M. 102, 106, 115, 117,
Favaron, O. 273 Frick, F. 815 129, 165, 167–9
Feige, U. 684 Friedgut, E. 503, 509–10 Getz, M. 61
Feller, W. 40, 50, 668 Friedland, S. 510, 736 Ghouila-Houri, A. 334, 550
Felsner, S. 783–4, 790, 823, 828 Frieze, A. 657, 722 Gilbert, E. N. 564, 688
Feng, W. 286 Fritsch, G. 399 Gilmore, P. C. 374, 549
Ferguson, T. 669 Fritsch, R. 399 Ginibre, J. 596
Ferrara, M. 422 Frobenius, G. 181, 189–90, 207, Ginzburg, A. 764
Fiedler, M. 778 256 Glaisher, J. 142, 148
Fielder, D. C. 65 Fronček, D. 684 Gleason, A. M. 449
Figaj, A. 492 Fujita, S. 373, 457 Glebov, A. N. 414
Filmus, Y. 509 Fulkerson, D. R. 222, 300, 302, Glock, S. 640, 647
Finck, H.-J. 338, 344 370, 518, 532, 534, 547 Goddard, W. 418
Fink, J. 330 Fulton, W. 189 Godfrey, J. 511
Fiorini, S. 361, 414 Füredi, Z. 442, 479, 730 Godsil, C. 500, 766, 781
Firke, F. A. 630 Fuss, N. 42 Goethals, J. M. 744
Fishburn, P. 586 Goldberg, C. H. 809
Fishburn, P. C. 586–7, 600, 604, Gabber, O. 779 Goldberg, M. K. 237, 360
823, 828, 832 Gabow, H. N. 246, 282 Goldman, J. R. 127, 175
Fisher, D. C. 418, 550 Gaddum, J. W. 344 Golomb, S. W. 29, 622
Fisher, R. A. 614 Gale, D. 285, 288, 803–4, 813, Golovach, P. A. 342
Fisk, S. 419 820 Golovina, L. I. 419
Fiz Pontiveros, G. 682 Galil, Z. 246, 779 Golumbic, M. C. 366, 370, 582,
Fleischer, R. 286 Gallai, T. 260, 263–4, 269, 273, 831–2
Fleischner, H. 331, 739, 763 275–6, 322, 334, 339, 354–5, Gonçalves, D. 421
Floyd, R. W. 257 357, 460, 548 Gondran, M. 287
Foata, D. 50, 99, 101, 106, 127, Gallian, J. A. 103, 646 Good, I. J. 239
177 Galperin, G. 61 Goodman, A. W. 238, 477
Folkman, J. 455, 466 Galvin, D. 504, 506, 509 Goodman, S. 330
Fomin, F. V. 342 Galvin, F. 228, 355, 363–4, 474 Gorgos, I. M. 374
Fon-Der-Flaass, D. G. 649, 782 Ganter, B. 441 Gosper, R. W., Jr. 89
Ford, G. W. 135 Gardner, M. 143 Gottschalk, W. H. 468
Ford, K. 669 Garey, M. R. 11, 701, 795, 800 Gould, H. W. 23, 30, 65
Ford, L. R., Jr. 300, 302, 532 Gargano, L. 327, 334 Gould, R. J. 215, 314, 316, 327,
Fortuin, C. M. 596 Gaskell, R. W. 389 393, 491
Foulkes, H. O. 197 Gasparyan, G. S. 369 Goulden, I. P. 42
Fouquet, J. L. 334 Gauchman, H. 61 Govorčin, J. 297
Fournier, J.-C. 372 Gavlas, H. 646 Gowers, W. T. 466, 484, 672
Fox, J. 334, 425, 451, 453, 457, Gavril, F. 373 Grable, D. A. 745
459, 484–5, 551, 601, 608, Gazit, H. 397 Graham, N. 428, 441, 565, 584
672–3, 684 GCHQ Problem Solving Group Graham, R. L. 24, 33, 41, 61, 89,
Frame, J. S. 190 149 101, 341, 394, 425, 431, 442,
Frank, A. 307, 310, 548 Geetha, J. 373 446, 452, 460, 463–5, 470,
Frankl, P. 450, 459, 488, 493, Gehrlein, W. G. 823 473, 484, 500, 508–9, 559–60,
495–7, 499–500, 503, 508–11, Geller, D. 418 597, 604, 673, 771, 781, 800
723, 727–30, 744 Genest, F. 273 Grauman, T. 457
Author Index 933

Graver, J. E. 449, 451 Haken, W. 402, 406 Herden, D. 461

Gravier, S. 356 Hakimi, S. L. 221, 238, 255, 261, Herrendörfer, G. 655
Greene, C. 190, 205, 208, 496, 326, 354, 746 Herzog, M. 512
511, 539, 548, 551, 555, 561, Hales, A. W. 467, 474 Heule, M. J. 342
565, 607 Halin, R. 310 Hierholzer, C. 234
Greene, J. E. 811–2 Hall, M. 254–5, 262, 637 Hilbert, D. 461, 473
Greenwell, D. L. 329 Hall, M., Jr. 622, 637 Hill, C. 34
Greenwood, R. E. 449 Hall, P. 254 Hillier, J. A. 246
Gregory, D. A. 771 Halldórsson, M. M. 684 Hilton, A. J. W. 65, 359, 496–7,
Griffiths, S. 682 Halmos, P. R. 254 499, 501, 511–2, 607
Griggs, J. R. 144–5, 152, 511, Hamburger, P. 580 Hilton, P. 41
550, 557, 561, 568, 702 Hamidoune, Y. O. 734, 746 Hind, H. 373, 681
Grimmett, G. R. 666, 701 Hamilton, W. 65 Hindman, N. 439, 466
Grinberg, E. J. 400 Hammar, M. 327, 334 Hinrichs, A. 817
Grinstead, C. M. 449, 451 Hammersley, J. M. 821 Hiraguchi, T. 571, 574–5, 583
Grolmusz, V. 731 Hanani, H. 398, 639, 641, 656 Hladký, J. 347
Grone, R. 776 Hanlon, P. 587 Hliněný, P. 795
Gross, J. L. 377 Hansel, G. 496, 557 Ho, C.-C. 144–5, 152
Gross, O. A. 370 Hansen, H. M. 373 Hoare, A. H. M. 152
Grötzsch, H. 414 Hanson, D. 273, 373 Hobby, C. R. 810
Grünbaum, B. 387, 389, 423 Harary, F. 135, 184, 187, 215–6, Hochberg, R. 801–3
Grytczuk, J. 403, 680, 742 224, 228, 237, 290, 295, 297, Hoeffding, W. 708, 715, 722
Guan, M. 287 329, 343, 398, 412, 428, 454, Hoffman, A. J. 222, 244, 374,
Guldan, F. 412 541, 565, 584, 666, 677, 690, 549, 772, 775, 781
Gunderson, D. S. 473, 479, 491 765, 780, 794, 797–8 Hoffman, D. G. 647
Guo, S. 473 Hardy, G. H. 140 Hoggatt, V. E., Jr. 59, 61
Gupta, H. 176 Harper, L. H. 496, 559–60, Holland, F. 23, 117
Gupta, R. P. 358–9, 372 563–4, 567, 720, 800 Holley, R. 597
Gurgel, M. A. 327 Harris, A. J. 362, 373 Hollingsworth, S. 248
Gusfield, D. 286 Hartke, S. G. 276 Holton, D. 326
Gustavsson, T. 640 Hartman, I.B-A. 549 Holton, D. A. 213
Gutin, G. 334 Hartsfield, N. 778 Holyer, I. 359, 699
Gutner, S. 402, 420 Harzheim, E. 551 Holzman, R. 665
Gutowski, G. 742 Hasse, M. 339 Holzmann, C. A. 541
Guy, R. K. 64, 105, 632, 792–3, Havel, I. 373 Hook, J. 459
795, 797–8 Havel, V. 221 Hopcroft, J. E. 283–4, 393
Gyárfás, A. 262, 351–2, 357, Havet, F. 412–3 Horadam, A. F. 61
460, 472, 474, 686 Haxell, P. 790 Horák, P. 344, 373
Győri, E. 312 Haxell, P. E. 455, 492, 677 Horsley, D. 646
Hayes, P. J. 80 Horton, J. D. 317, 330
Haase, A. 815 Hayward, R. B. 375 Hosoya, H. 765
Habib, M. 657 He, Q. 373 Hoşten, S. 580
Hadamard, J. 618–9, 805 Heawood, P. J. 399, 402, 419 Howard, F. T. 48
Hadwiger, H. 349, 729, 815, 817 Hebbare, S. P. R. 656 Hsieh, W. N. 563
Haemers, W. H. 766, 771–2 Hebdige, M. 414 Hsu, D. F. 343
Hagelstein, P. 461 Hedayat, A. S. 651 Hsu, W.-L. 393
Häggkvist, R. 326 Hedetniemi, S. 330, 417–8 Huang, D. 152
Haigh, J. 144 Heffter, L. 645 Huang, H. 771
Hajek, B. 765 Heilbronn, H. 734 Hulett, H. 238
Hajnal, A. 343, 490–1 Hell, P. 327, 334, 457 Huneke, C. 776
Hajnal, P. 432, 442, 581, 604 Henning, M. A. 263, 274, 638 Hurlbert, G. H. 828
Hajós, G. 355, 709 Heppes, A. 817 Hutchinson, J. P. 394, 419, 429
934 Author Index

Huxley, M. 730 Juvan, M. 60, 412, 739 Khan, M. A. 133

Hwang, K.-W. 731 Kharaghani, H. 622
Kabela, A. 318 Khare, C. B. 440
Igusa, K. 144, 152 Kahn, J. 362, 439, 503, 510, Kierstead, H. A. 50, 248, 352,
Imase, M. 296 576, 578, 817–8, 823–6, 831 355–6, 457, 584, 742
Ionin, Y. 178 Kainen, P. C. 399, 418–9, 794, Kim, J. 647, 668
Irving, R. W. 286 797–8 Kim, J. H. 451, 664, 682, 823
Isaacs, R. 372 Kaiser, T. 237, 318, 326 Kim, S.-J. 423–4
Isaak, G. 459, 624 Kalai, G. 503, 736, 817–8 Kimble, R. J., Jr. 187
Ishigami, Y. 482 Kalbfleisch, J. D. 458 King, A. D. 685, 765
Itai, A. 249, 315 Kalbfleisch, J. G. 449, 458 Kinnersley, W. B. 457
Ito, T. 244, 775 Kamat, V. M. 503 Kirchhoff, G. 37
Ivanova, A. O. 412 Kampen, G. R. 786 Kirdar, M. S. 152
Kaneko, A. 249, 819 Kirkland, S. 34
Jackson, B. 318, 321, 326 Kano, M. 268, 273, 457 Kirkman, T. P. 612, 641, 649
Jacobi, C. G. J. 129 Kantor, W. M. 699 Kislicyn, S. S. 823
Jacobson, M. S. 249, 318, 327, Kantrowitz, M. 61 Kiss, E. W. 665
374 Kapoor, S. F. 227, 273 Klamkin, M. S. 389
Jaeger, F. 797 Karaganis, J. J. 331 Klavžar, S. 70, 606
Jahanbekam, S. 262, 273, 333, Karaivanov, B. 35 Klazar, M. 432
422, 668 Kardoš, F. 765 Klein, P. N. 246
Jamison, R. E. 374, 472–3, 735 Karger, D. R. 246 Kleitman, D. 509, 551, 558
Janson, S. 657, 696, 708 Karoński, M. 657 Kleitman, D. J. 40, 226, 431,
Janssen, J. C. M. 362–3 Karp, R. M. 11, 283–4 442, 493, 496, 500–1, 507,
Jarnı́k, V. 246, 252 Kárteszi, F. 639 511, 548, 550–1, 555, 558–9,
Jendrol’, S. 407, 411 Kasteleyn, P. W. 596, 755–6 561, 563, 565–7, 580, 595–6,
Jensen, T. R. 335 Kászonyi, L. 491 607–8, 665, 749, 793, 797
Jenssen, M. 461 Katchalski, M. 442 Klešč, M. 794
Jerrum, M. 754 Katerinis, P. 276, 318 Klotz, W. 393
Jewett, R. I. 467, 474 Katona, G. O. H. 490, 495–8, Klove, T. 22
Jiang, T. 471, 473, 685 511, 554, 565 Kneser, M. 251, 811
Jiang, Z. 479 Kauers, M. 174 Knuth, D. 152
Jichang, S. 567 Kaul, H. 819 Knuth, D. E. 10, 24, 33, 35, 41,
Johannson, K. R. 466 Kawarabayashi, K.-i. 316 50, 61, 65, 81, 89, 92, 99, 101,
Johnson, D. M. 654 Kearnes, K. A. 665 116, 122, 133, 176, 190, 193,
Johnson, D. S. 11, 701, 795, 800 Keedwell, A. D. 610, 623 195, 198, 207–8, 286, 441,
Johnson, N. L. 106 Keevash, P. 475, 496, 640, 682 567, 667, 669, 800
Johnson, P. 344 Keller, G. 587, 604 Ko, C. 498
Johnson, R. T. 503 Keller, N. 503 Kobourov, S. G. 286
Joichi, J. T. 175 Kellogg, A. 177 Koch, J. 406
Jolivet, J.-L. 334 Kelly, D. 571–3, 583, 606 Kochman, F. 62
Jones, D. 61 Kelly, J. B. 434 Kochol, M. 421, 424
Jonsson, J. 387 Kelly, L. M. 434 Koebe, P. 394
Joos, F. 647 Kelmans, A. K. 297, 393, 777–8, Kohayakawa, Y. 455, 482, 488
Jordan, C. 243, 251, 669 782 Köhler, E. 647
Joret, G. 680 Kempe, A. B. 213, 399 Kollár, J. 633, 672
Jorza, A. 388 Kendall, M. G. 228 Komlós, J. 226, 351, 451, 458,
Joyal, A. 37, 127 Kerimov, A. 315 481, 488, 492, 551, 663–4,
Józsa, S. 794 Kerov, S. V. 722, 821 779, 823
Juarez, H. A. 794 Kervaire, M. 745 Komm, H. 571
Jukna, S. 493 Keselman, G. 116 Konheim, A. G. 50
Jungnickel, D. 609 Kézdy, A. E. 318, 746
Author Index 935

König, D. 256, 258–9, 359, 371, Kynčl, J. 457 Li, X. 457

531, 763 Li, X. L. 312
Konjevod, G. 457 Labelle, G. 129 Li, Z. 414
Kopylov, G. N. 323 Laghate, K. 104 Lick, D. R. 312
Korshunov, A. D. 558 Lagrange, J. L. 129 Liggett, T. M. 512
Kortsarz, G. 684 Lai, H.-J. 540 Lih, K.-W. 287, 413
Kosek, P. M. 630 Lam, C. W. H. 612 Lin, C.-Y. 287
Koshy, T. 53 Lamé, G. 67 Lindner, C. C. 609, 641
Kostochka, A. 434, 442 Landau, E. 10 Lindström, B. 165, 167, 169,
Kostochka, A. V. 239, 248, 332, Landau, H. G. 225, 228, 261 496, 557
355, 360, 364–5, 373–4, Lander, E. S. 623, 637 Ling, A. C. H. 473
413–4, 420, 460, 484, 504, Landman, B. M. 425 Linial, N. 207, 323, 549, 823–5
639, 672, 705, 710 Langley, L. 550 Liouville, B. 292
Kotz, S. 106 Larman, D. G. 461, 724, 729 Lipski, W., Jr. 566
Kotzig, A. 227, 264, 273, 276, Larsen, M. 34 Lipton, R. J. 394–6, 399
371, 411, 647 Las Vergnas, M. 272, 333 Littlewood, D. E. 207
Kouider, M. 273, 326–7 Lassak, M. 817 Liu, B. 236
Kouril, M. 465 Lavallée, I. 80 Liu, C. L. 17
Kovács, P. 216 Lawler, E. L. 282, 439, 533 Liu, H. 460
Kovář, P. 239 Lawrence, J. 355 Liu, J. 262
Kövari, T. 631, 672 Lawrence, S. L. 460 Liu, R. 354
Kozik, J. 660, 673, 680, 742 Laywine, C. F. 623, 656 Liu, Z. H. 326
Kožuhin, G. I. 353 Lazarson, T. 538 Livingston, M. 584
Kraitchik, M. 66 Le, V. B. 370 Lloyd, E. K. 37, 46
Král’, D. 326, 347, 414, 457, 765 Lebesgue, H. 422 Lo, A. 327, 334, 640
Krasikov, I. 327 Lederberg, J. 418 Locke, S. C. 227, 229, 330
Kratochvı́ l, J. 403 Lee, C. 457, 484, 488, 551, 672 Loeb, D. E. 144
Krattenthaler, C. 50, 170, 192 Lee, O. 315 Loeb, S. 422
Kratzke, T. 771 Lefmann, H. 471, 601 Logan, B. F. 722, 821
Krivelevich, M. 672–3, 684 Lefschetz, S. 807 Long, C. T. 62
Kriz, I. 341, 434 Lehel, J. 249, 318, 472 Longyear, J. Q. 776
Król, M. 419 Lehman, A. 529, 540 Loomis, L. H. 514
Krompart, L. B. 418 Lehmer, D. H. 65 Lossers, O. P. 129
Kronk, H. V. 329, 343 Leighton, F. T. 793 Loten, C. O. M. 273, 373
Krueger, R. A. 460 Leiserson, C. E. 809 Lovász, L. 47, 104, 229, 232,
Kruskal, J. B. 245, 495–6 Lekkerkerker, C. G. 61 249, 262–4, 272, 275–6, 302,
Kruskal, J. B., Jr. 431 Lempel, A. 343 308, 310, 312, 342, 355,
Kruyswijk, D. 554 Lenz, H. 609 368–9, 375–6, 419, 434, 436,
Kubale, M. 335 LeSaulnier, T. 441 456–7, 461, 477–8, 491,
Kubesa, M. 239 Lesniak, L. M. 215, 250, 273, 495–7, 503, 510–1, 533,
Kučera, L. 699, 718 315–6, 327, 377 607–8, 668, 675, 679, 726,
Kuczma, M. S. 24 Letzter, S. 460 756, 762, 764, 812
Kühn, D. 316, 488, 640, 647 Leuck, D. H. 175 Lu, L. 491
Kuhn, H. W. 280 Levin, G. M. 664 Lu, X. 322, 566
Kumbhat, M. 422 Levy, L. 80 Lubell, D. 34, 558
Kummini, M. 329 Lewinter, M. 215 Lubotzky, A. 341, 434, 780
Kündgen, A. 287 Lewis, J. T. 429 Lucas, F. E. A. 47, 162, 238
Kuo, E. H. 34 Li, C. M. 790 Luce, R. D. 585
Kuperberg, G. 34, 756 Li, H. 316 Łuczak, T. 455, 657, 696, 708
Kupka, J. 49 Li, Q. 801, 819 Luecker, G. S. 393
Kuratowski, K. 390 Li, S.-Y. R. 740 Lukot’ka, R. 237
Kuroda, S. 515 Li, W. C. W. 740 Luks, E. M. 699
936 Author Index

Luo, R. 423 McDiarmid, C. J. H. 372, 657, Molloy, M. 335, 362, 373, 657,
Lützen, J. 740 685, 701, 706, 719, 721, 801–3 668, 674, 677, 681, 685–6,
Lyusternik, L. A. 810 McFarland, R. 637 702, 706, 721–2
McGuinness, S. 424 Monjardet, B. 586
McKay, B. D. 449, 697 Montágh, B. 50
Ma, M. 371 McKee, T. A. 239, 374 Montellano-Ballesteros, J. J.
Mabry, R. 417 McMorris, F. R. 374 457
MacLane, S. 515, 530 McSorley, J. P. 765 Montgomery, B. 424
MacMahon, P. A. 45, 149, 169, Meagher, K. 500, 781 Montgomery, R. 640, 646
178, 207 Melham, R. S. 60 Montmort, P. R. 162
MacNeish, H. F. 611 Mel nikov, L. S. 414 Moon, J. W. 37, 47, 251, 325,
Maddox, R. 782 Melolidakis, C. 669 332–4, 477, 489, 491, 703, 721
Madej, T. 832 Mendelsohn, N. S. 654 Moore, B. R. 354
Mader, W. 264, 296, 308–11, Menger, K. 298, 300 Moore, E. H. 610–1, 641
315–6, 350–1, 353, 710 Merca, M. 148, 150 Moore, J. I., Jr. 573, 583
Maffray, F. 356 Merris, R. 776, 782 Morávek, J. 373
Magnant, C. 457 Meshalkin, L. D. 558 Morris, R. 460, 682
Magnanti, T. L. 501, 539 Meszka, M. 239 Morris, W. 446
Mahdian, M. 237 Meyniel, H. 375 Morris, W. D., Jr. 446, 580
Mahlburg, K. 664 Miao, L. 414 Moser, L. 59, 325, 332, 441, 477,
Mahmoodian, E. S. 237 Micali, S. 284 514, 703
Mahmoodian, S. E. 329, 343, Micek, P. 680 Moser, R. A. 680
371, 398 Mihók, P. 418 Moshkovitz, G. 602–3
Maillet, E. 623 Miklós, D. 501, 512 Motwani, R. 657
Majumdar, K. N. 730 Mikola, M. 65 Motzkin, T. S. 42, 387, 389, 489,
Malgrange, Y. 393 Milans, K. G. 457, 602, 608, 668, 781
Mann, H. B. 254, 623, 637 685 Mozhan, N. N. 354
Mantel, W. 224, 475 Milgram, A. N. 548 Mubayi, D. 451, 457, 460, 471,
Mao, J. 237 Miller, E. 784 478, 500, 686
Marcus, A. 432–3 Miller, E. W. 569–70, 831 Muir, T. 764
Marcus, M. 257, 765 Miller, G. L. 397 Mulder, H. M. 253, 274, 374
Margulis, G. A. 779–80 Miller, Z. 343, 819 Mullen, G. L. 623, 656
Marica, J. 512 Mills, G. 463 Mullin, R. 123
Markowsky, G. 558, 580 Mills, W. H. 514 Munkres, J. 280
Martin, M. H. 763 Milman, V. D. 778–80 Murty, U. S. R. 215, 228, 330,
Martin, R. 488 Milner, E. C. 499, 511 489, 639
Martinov, N. 313 Milutinović, U. 70 Mycielski, J. 340
Massey, J. L. 507 Minoux, M. 287 Myers, B. R. 65, 354
Matoušek, J. 723, 804, 806, 810, Minty, G. J. 338, 540 Myrvold, W. 393
812–3 Mirkin, B. G. 586
Matoušek, J. 461 Mirsky, L. 518, 547 Naatz, M. 312, 764
Matsko, V. J. 565 Mirzakhani, M. 402, 421 Nagasawa, T. 797
Matsuda, H. 327 Mitas, J. 587 Nagy, Z. 459, 728, 745
Matthews, M. M. 326, 330 Mitchem, J. 326 Nakada, H. 296
Matula, D. W. 297, 353–4, 700, Mitzenmacher, M. 708 Nakasawa, T. 515
718 Moessner, A. 62 Nakayama, A. 678
Maunsell, F. G. 264 Moews, D. 668 Narayan, D. A. 624
Maurer, R. 106 Mogyoródi, J. 105 Narayana, T. V. 45
Maurer, S. 228, 584 Mohanty, S. G. 50 Narayanan, B. 765
McAndrew, M. H. 222 Mohar, B. 60, 377, 412, 739, 776, Narayanan, N. 373
McCarthy, J. 667 795, 797 Nasar, S. 803
McConnell, R. M. 573 Möhring, R. H. 573 Naserasr, R. 373
Author Index 937

Nash, E. D. 630 Palmer, C. 484, 488 Piotrowski, W. L. 647

Nash-Williams, C. St. J. A. 541 Palmer, E. M. 135, 657, 700, Piotrowski, W.-L. 647
Naslund, E. 504 703, 705 Pippenger, N. 509, 779
Nathanson, M. B. 734–5, 745–6 Palumbı́ny, D. 249 Pisanski, T. 330
Nemhauser, G. L. 521 Panconesi, A. 706, 721 Pité, E. 133
Nešetřil, J. 425, 434, 436–7, 455, PanduRangan, C. 374 Pitman, J. 37
457, 472, 511, 543, 726 Papadimitriou, C. H. 521 Plantholt, M. 359
Netto, E. 41 Park, W.-J. 424 Plaza, A. 80
Neumann-Lara, V. 327, 457 Parker, E. T. 612, 652–4 Plesnevič, P. S. 343
Newborn, M. M. 793 Parreau, A. 680 Plesnı́k, J. 267, 294, 296
Newman, M. 765 Parsons, T. D. 776 Pluhár, A. 421, 424
Nicoaescu, L. I. 22 Patashnik, O. 24, 33, 41, 61, 89, Plummer, M. D. 253, 262, 268,
Niessen, T. 273, 359 101 275, 297, 315–6, 401, 756, 764
Nijenhuis, A. 190 Paterson, M. S. 815 Pnueli, A. 343
Nikiforov, V. 454, 478, 482, 770 Patkós, B. 493 Pohoata, C. 174
Nikšić, F. 133 Paul, J. L. 465 Pokrovskiy, A. 646
Nilli, A. 780, 817–8 Paule, P. 66 Polimeni, A. D. 227, 272
Nishimura, H. 515 Paulraja, P. 326 Poljak, S. 776
Niu, Y. Y. 292 Payan, C. 670 Pollack, R. 783
Niven, I. 107 Peart, P. 64 Pollak, H. O. 771, 781
Noel, J. A. 355–6 Peck, G. W. 42, 442, 771 Pólya, G. 37, 128, 184
Noguchi, K. 797 Pederson, J. 41 Poonen, B. 503
Nordhaus, E. A. 344 Peled, U. 764 Popadić, M. S. 550
Norine, S. 765 Pelsmajer, M. J. 374, 795 Popescu, C. 172
Norman, R. Z. 263 Peltesohn, R. 645 Pór, A. 580
Novelli, J.-C. 192 Penaud, J. G. 418–9 Porter, T. D. 764
Nyblom, M. 172 Penrice, S. G. 352 Pósa, L. 238, 273, 326, 333, 490
Penrose, M. 783 Postle, L. 356, 414, 685
O, S. 237, 274, 334, 668 Pepe, V. 633 Pournin, L. 387
O’Hara, K. M. 557 Percus, J. K. 755 Powell, M. B. 337
Ohba, K. 355 Perfect, H. 518 Prałat, P. 455, 704, 708
Ohtsuka, H. 33 Perkovic, L. 359 Pratt, R. 22
Olmsted, C. 176 Perles, M. A. 546 Preen, J. 386
Olson, E. J. 823 Péroche, B. 412, 678 Prékopa, A. 105
Ordaz, O. 321 Perrett, T. 685 Prim, R. C. 246, 252
Ore, O. 249–50, 258, 260, 319, Perron, O. 62 Prince, N. 460
321, 332–3, 355, 359–60, 387, Pertuiset, R. 393 Prins, G. 297
399, 420, 532, 571 Perz, S. 369 Proctor, R. A. 147, 557
O’Rourke, J. 394 Peters, L. 458 Prodinger, H. 49, 64
Osthus, D. 316, 488, 640, 647 Petersen, J. 213, 267–8, 273–4, Prömel, H. J. 425
Ota, K. 355 277, 740 Propp, J. G. 34, 387, 511
Owens, A. B. 389 Petersen, T. K. 101 Prowse, A. 740
Oxley, J. G. 515, 526 Petkovšek, M. 87, 89–90, 606 Prüfer, H. 37
Ozeki, K. 327, 334, 457 Petr, C. 70 Pudaite, P. R. 66
Özkahya, L. 239 Pettersson, W. 646 Pudlák, P. 504
Peyrat, C. 296 Puleo, G. J. 372
Paasche, I. 62 Phillips, J. 711 Pultr, A. 511, 726
Pach, J. 457, 461, 491, 601, 783, Phillips, R. 341, 434, 780 Pyber, L. 249, 737
793, 795, 797 Pikhurko, O. 479, 817 Pyke, R. 50
Pak, I. 141–2, 152, 192 Pilpel, S. 821 Pym, J. S. 299
Palacios, J. L. 66, 668 Pinkham, R. S. 129
Paley, R. E. A. C. 621 Pinsker, M. 779 Qiao, P. 249
938 Author Index

Quinn, J. J. 53, 664 Rényi, A. 37, 99, 105, 133, 244, Rothschild, B. L. 425, 452,
252, 514, 632, 687, 698–9 463–5, 470
Reuter, K. 572, 584 Rotman, J. J. 238
Rabern, L. 336 Révész, P. 817 Rousseau, C. C. 455, 459
Rabin, M. 263 Rey, J. G. 47 Roussel, F. 370
Rabinovitch, I. 228, 574, 583–4, Reznick, B. 771 Roy, B. 339
604 Ribó Mor, A. 783 Royle, G. 226, 766
Rademacher, H. 140 Rice, J. R. 810 Rubel, L. A. 473
Radhakrishnan, J. 504, 508, Richter, C. 817 Rubin, A. L. 346, 348, 356, 660,
510, 660, 673 Richter, R. B. 489, 794 667
Rado, R. 451, 456, 462, 466–8, Riddell, R. J. 135 Ruciński, A. 484, 657, 673, 696,
470, 498, 503–4, 520, 608 Riddell, R. J., Jr. 127 703, 708
Radoux, C. 177 Riesling, A. S. 817 Rupp, C. A. 178
Radziszowski, S. P. 449 Rieß, W. 557 Rusin, D. J. 103
Raghavan, P. 657 Riguet, J. 605 Rusu, I. 370
Raı̆gorodskiı̆, A. M. 817 Ringeisen, R. D. 794, 798 Rusza, I. 734–5, 745–6
Ramalingam, G. 374 Ringel, G. 218, 646, 656 Ruszinkó, M. 460
Ramamurthi, R. 287, 746 Riordan, J. 47, 50, 65, 133, 153, Ruzsa, I. Z. 485, 492, 684
Ramanan, G. V. 730 161 Ryjáček, Z. 320, 326
Ramanujan, S. 140 Rivera-Campo, E. 327 Ryjáček, Z. 320
Ramirez-Alfonsin, J. 657 Robbins, H. 306 Rymer, N. W. 66
Ramı́rez-Alfonsı́n, J. L. 368, 370 Roberts, F. S. 337 Ryser, H. J. 227–8, 254, 314,
Ramos, E. A. 815 Roberts, J. A. 459 616, 627, 637
Ramras, M. 295 Roberts, S. M. 449
Ramsey, F. P. 443–4, 468 Robertson, A. 425 Saaty, T. L. 399, 418
Randerath, B. 273 Robertson, J. M. 389 Saberi, A. 237
Raney, G. N. 50, 129 Robertson, N. 349, 370, 406, Sabidussi, G. 343, 740
Rasch, D. 655 416–7 Sachs, H. 34, 219, 251, 274,
Raspaud, A. 414 Robinson, G. de. B. 190, 193, 338–9, 739, 766, 780–1
Rautenbach, D. 275 195 Sadri, B. 237
Ray-Chaudhuri, D. K. 370, 647, Robinson, J. P. 820 Sagan, B. 207
649, 731 Rodeh, M. 249 Sagan, B. E. 177, 189, 766, 823
Raynaud, H. 460 Rodger, C. A. 609, 641, 647 Saito, A. 318, 321
Razborov, A. A. 475, 478, 491, Roditty, Y. 236, 327 Šajna, M. 646
504 Rödl, V. 425, 434, 436–7, 455, Sakamoto, J. 355
Read, R. C. 374 471–3, 483–4, 488, 543, 581, Saks, M. E. 548–9, 801–3,
Rechnitzer, A. 433 604, 672, 737 823–4, 826, 831
Rédei, L. 334 Rodney, P. 794 Salamon, G. 334
Redfield, J. H. 184 Rogers, C. A. 724, 729 Salavatipour, M. R. 414
Ree, R. 257 Rogers, D. 64 Salazar, G. 794
Reed, B. A. 335, 339, 355, 359, Rolewicz, S. 369 Salehi Nowbandegani, P. 334
362, 370, 373, 657, 668, 674, Rollová, E. 237 Salgado, E. 414
677, 681, 685–6, 702, 706, Roman, S. M. 123, 632 Salmon, A. 356
721–2 Rónyai, L. 633, 672 Samad, T. 353
Rees, D. 239 Rosa, A. 641, 646–7, 656 Sandell, D. P. 668
Regev, A. 433 Roselle, D. P. 177 Sanders, D. 406, 416–7
Reiher, C. 478 Rosenfeld, M. 326, 373 Sanders, D. P. 414–6, 422–3
Reiman, I. 632 Rota, G. C. 521, 530 Sanders, P. R. 466
Reiniger, B. 434, 442 Rote, G. 783 Sands, A. D. 50
Remmel, J. B. 36, 192 Rotem, D. 582 Santoro, N. 832
Renaud, J.-C. 503, 512 Roth, K. F. 484–5 Saritzky, N. 344
Renault, M. 40 Sarkaria, K. S. 812
Author Index 939

Sárközy, G. N. 481 Seidel, J. J. 724, 744 Skokan, J. 461, 488

Sarnak, P. 341, 434, 780 Seidenberg, A. 430 Skolem, T. 641, 644
Sarvate, D. G. 503, 512 Seinsche, D. 375, 490 Škrekovski, R. 60, 239, 412, 424,
Sauer, N. 248, 476 Sekanina, M. 331 639, 739
Sauvé, L. 477 Selkow, S. M. 699, 705 Skyrme, T. H. R. 152
Savage, C. D. 312 Sereni, J.-S. 412 Sleator, D. D. 387
Sawin, W. F. 504 Serrano, L. G. 42 Slivnik, T. 364
Scarpis, U. 622 Servedio, R. 144 Sloane, N. J. A. 318, 651
Schaal, D. 473 Seymour, P. 349, 406, 416–7 Smith, B. B. 228
Schacht, M. 455, 485, 488 Seymour, P. D. 314, 349, 370, Smith, C. A. B. 752
Schaefer, M. 795 396–7, 416, 497, 531, 607–8 Smith, H. J. S. 177
Schäuble, M. 354 Shamir, E. 717 Smolenskii, E. A. 250
Schaudt, O. 503 Shannon, A. G. 60, 65 Smoot, N. 148
Schauz, U. 347, 741 Shannon, C. E. 358, 360, 365, Snevily, H. St. C. 50, 502, 730,
Scheim, D. E. 740 372 746
Scheinerman, E. R. 394, 828–32 Shapira, A. 488, 602–3 Soderberg, S. 514
Schellenberg, P. J. 647 Shapiro, H. S. 514 Sofair, I. 174
Schelp, R. H. 262, 327, 454–5, Shapiro, L. W. 48, 64–5, 82 Soifer, A. 342, 425
472–3 Shapley, L. S. 285, 288 Soltan, V. 446
Schensted, C. 193, 195, 198, 206 Sharir, M. 783 Solymosi, J. 492
Schiermeyer, I. 320 Sharp, H., Jr. 261 Somasundaram, K. 373
Schläfli, L. 48 Shearer, J. 507 Sondow, J. 32
Schmeichel, E. F. 318, 326, 330, Shearer, J. B. 508–9, 566–7, Soneoka, T. 296
354 663, 675, 682 Song, Z. M. 316
Schmidt, F. W. 105, 134, 152, Sheehan, J. 213, 459 Sorel, J. 33
188, 442 Sheikh, N. 460 Sørensen, B. A. 345
Schmitt, J. R. 276 Shelah, S. 466, 551 Sós, V. T. 244, 457, 631–2, 672,
Schmuland, B. 176 Shen, Y. Q. 801, 819 684
Schnyder, W. 394, 783, 790 Shepp, L. A. 599, 608, 722, 821 Špacapan, S. 292, 297
Schönberger, T. 267 Shi, R. H. 326 Spencer, J. 248, 425, 452, 454,
Schönheim, J. 490, 497, 512, 607 Shieh, B.-S. 292 463, 465, 470, 504, 567, 576,
Schramm, O. 817 Shih, W.-K. 393 579–80, 584, 657, 662–3, 666,
Schrijver, A. 277, 521, 534, 541, Shirazi, H. 737, 746 669, 672–5, 682, 685, 688,
735, 813–4 Shlosman, S. B. 810 696, 702, 706, 717, 721–2,
Schröder, E. 47, 64, 81, 135 Shnirel’man, L. 810 780, 794
Schulz, A. 783 Shokoufandeh, A. 488 Spencer, T. 246
Schur, I. 462, 473 Shor, P. 442 Sperner, E. 496, 558, 561, 799,
Schuster, S. 251, 273 Shpilka, A. 504 802, 805
Schützenberger, M.-P. 99, 101, Shreve, W. E. 249 Spinrad, J. P. 370, 573
106, 127, 129, 177, 197–8, Shrikhande, S. S. 612, 616, 623, Spitzer, F. 50
203, 206, 208, 496, 616 651–4 Srinivasan, A. 660, 673, 684
Schwenk, A. J. 187, 206, 620–1, Sidney, J. B. 828, 832 Stacho, L. 327, 334
624, 782, 794, 798 Sidney, S. J. 828, 832 Stahl, J. 568
Schwer, S. R. 28, 31, 41 Silwal, S. 226 Stanley, R. P. 41–2, 48, 61, 76,
Scott, A. 488 Simion, R. 442 81, 93, 102, 116, 127, 129,
Scott, A. D. 352 Simonovits, M. 457, 478–9, 482, 132, 135–6, 145, 150, 177,
Scott, D. 587 488, 490, 492, 684 189, 206–8, 261, 557, 564–5,
Scoville, R. A. 177 Simonyi, G. 474 606, 669, 761, 766, 824, 827
Scully, D. 152 Sinclair, A. 680, 754 Stanton, D. 100
Seberry, J. 622 Singhi, N. M. 728 Stanton, R. 458
Sedláček, J. 65 Singleton, R. R. 244, 639, 775 Stathopoulos, D. 62, 135
Segner, J. A. 41 Singmaster, D. 70, 388 Staton, W. 238
940 Author Index

Steele, J. M. 430, 669 Szele, T. 666 Toft, B. 273, 335, 345, 354, 373,
Štefanovič, D. 795 Szemerédi, E. 226, 250, 351–2, 740
Steger, A. 488 357, 451, 455, 480–1, 483–5, Toida, S. 239
Steiglitz, K. 521 488, 492, 504, 663–4, 684, Tokushige, N. 493
Stein, S. K. 387, 393 737, 779, 793–4, 798 Tollis, I. G. 394
Steiner, J. 48, 641 Szpilrajn, E. 569 Tomescu, I. 355, 512, 667
Steinlein, H. 806, 815 Szűcs, A. 810 Tong, L.-D. 287
Stevens, W. L. 610 Törőcsik, J. 461
Stewart, M. J. 272 Tóth, G. 793, 795, 797
Stiebitz, M. 239, 346, 354, 639, ’t Woord, A. N. 32 Touchard, J. 127
739 Tagiuri, A. 61 Tracy, P. 249
Stinson, D. R. 609, 612, 647 Tait, P. G. 399–400 Trotignon, N. 370
Stirzaker, D. R. 666 Talagrand, M. 712, 720 Trotter, W. 794
Stobert, I. 794 Tamassia, R. 394 Trotter, W. T. 50, 228, 373, 442,
Stocker, C. 457 Tanaka, Y. 803 455, 483, 568, 572–3, 580–1,
Stockmeyer, P. K. 92 Tanner, R. M. 779 583–4, 600, 604, 790, 794,
Stolee, D. 276, 422, 602, 608 Tao, M. Q. 801, 819 798, 823, 828, 832
Stone, A. H. 479, 491 Tao, T. 488, 680, 732 Trow, P. B. 429
Stong, R. 330, 343, 461, 623 Tardos, E. 310 Tsai, M.-T. 419
Stout, Q. F. 584 Tardos, G. 432–3, 491, 680 Tsai, S.-F. 551
Stoyanovskii, A. V. 192 Tarjan, R. E. 246, 282, 387, Tucker, A. W. 807, 809
Straus, E. G. 489, 781 393–6, 399 Tucker, T. W. 377
Strehl, V. 31 Tarry, G. 612, 763 Tuglu, N. 177, 766
Stufken, J. 651 Tarsi, M. 738–9, 746, 764 Tuinstra, H. 327
Sturtevant, D. 507 Tashkinov, V. A. 736 Turán, P. 223, 475–6, 484, 489,
Subba Rao, K. 60 Tator, C. 575, 583 491, 631, 672
Sudakov, B. 334, 355, 425, 451, Tauraso, R. 33, 81 Tutte, W. T. 242, 251, 264,
453, 457, 459–60, 484, 551, Tayfeh-Rezaie, B. 622 268–70, 276, 304–5, 314, 317,
601, 608, 646, 672–3, 677, Taylor, H. 91, 346, 348, 356, 330, 370, 392, 398, 401, 521,
684, 731 660, 667 536, 749, 752, 783
Suk, A. 446, 451, 457, 601 Teirlinck, L. 441 Tuza, Z. 262–3, 339, 343–4, 352,
Sulanke, R. A. 28–9, 34, 62, 64 Telle, J. A. 512 357, 403, 472, 491
Sumner, D. P. 272–3, 326, 330, Thiel, L. 612 Tverberg, H. 378, 771
351 Thieu, U. 173
Sun, Q. 414 Thomas, R. 314, 349, 370, Uhlenbeck, G. E. 127, 135
Sun, Z.-W. 473 396–7, 406, 414, 416–7, 471 Umans, C. 504
Sunder, V. S. 776 Thomason, A. 692 Ungar, P. 175, 763
Suppes, P. 587 Thomason, A. G. 242, 249–50, Upfal, E. 708
Suzuki, H. 728, 731 318, 331–2, 351, 710 Urrutia, J. 582, 828, 832
Suzuki, Y. 797 Thomassé, S. 685
Sved, M. 40 Thomassen, C. 238, 262, 304–5, Vaccaro, U. 327, 334
Swiercz, S. 612 313–4, 316, 326, 333–4, 350, Valencia-Pabon, M. 251
Sýkora, O. 798 377, 392–4, 401–3, 414, 419, Valiant, L. G. 754
Sylvester, J. J. 37, 141–2, 618 685, 742, 794 Valicov, P. 373
Sysło, M. M. 819 Thrall, R. M. 190, 192, 207 van Aardenne-Ehrenfest, T. 752
Szabó, T. 633, 672 Thuillier, H. 370 van den Heuvel, J. 424
Szász, D. 105 Thurston, W. P. 387, 394 van der Waerden, B. L. 463–4,
Szegedy, M. 674 Tian, F. 325–6 515, 521
Székely, L. A. 47, 135, 428, 441, Timmons, C. 422 van Lint, J. H. 139, 609, 622,
794, 798 Tindell, R. 237 639, 774
Szekeres, G. 338, 344, 430, 446, Tiwari, P. 771 van Tengbergen, E. 554
458 Todd, M. J. 807, 809 Vanden Eynden, C. 92
Author Index 941

Vandenbussche, J. 260 Wang, W. 413 Wilson, R. M. 459, 500, 609, 622,

Vander Meulen, K. N. 771 Wang, Y. 503 639–41, 647, 649, 727–9, 731
Vandermonde, A.-T. 26 Wanless, I. M. 261 Win, S. 327–8
Vapnik, V. N. 711–2 Wardlaw, W. P. 116, 173 Win, Z. 327
Varma, B. N. 325 Warning, E. 735 Winkler, P. 668, 821, 823
Vassilev, M. 29 Warren, H. E. 830 Winkler, P. M. 221, 431, 600,
Vassilev, T. S. 35 Watanabe, M. 418–9 608, 667, 781
Vaughan, E. R. 475 Watkins, M. E. 313–4 Wirth, B. 402
Vaughan, H. E. 254 Watson, P. 389 Wismath, S. K. 394
Vaughan, T. P. 503 Weaver, W. 60 Witt, E. 467
Vazirani, V. V. 284, 754, 756 Wegner, G. 496 Woan, W.-J. 64
Veldman, H. J. 318 Wei, B. 473 Woeginger, G. J. 342
Venkannayah, K. 61 Wei, V. K. 263, 662 Wojda, A. P. 326
Venkatesan, S. M. 397 Weinstein, J. M. 263 Wollan, P. 314
Vera, J.-C. 251 Weiss, B. 50 Wolsey, L. A. 521
Veršik, A. M. 722, 821 Welch, L. R. 29 Wong, C. K. 566
Verstraëte, J. 479, 500 Welsh, D. J. A. 337, 515, 521, Wood, D. 80
Vestergaard, P. D. 327 533, 539–40 Wood, D. R. 680
Viennot, G. 165, 167–9, 199–200 Welukar, R. M. 115 Woodall, D. R. 323, 364–5,
Vigoda, E. 754 Wendel, J. G. 669 413–4, 539, 740, 793, 798
Vijayaditya, N. 373 Wernicke, P. 406 Woolhouse, W. S. B. 641
Vince, A. 441, 703, 782 West, D. B. 50, 62, 152, 176, Wormald, N. 671, 704
Vitaver, L. M. 339 215, 237, 239, 260, 262, Wormald, N. C. 697, 704, 722
Vizing, V. G. 237, 343, 346–7, 273–6, 326, 331–4, 353–4, Worpitzky, J. 101
355–6, 358–9, 362, 414, 356, 371, 374, 407, 411, 419, Wright, C. 823
422–3, 699 422, 434, 442, 457, 472, Wu, B. 413
Voigt, M. 346, 402–3 490–1, 493, 540, 550–1, 557, Wu, H. 334, 355–6
Volkmann, L. 215, 268, 296, 359 564–5, 567, 587, 596, 602, Wu, J.-L. 412, 422
Voloshin, V. I. 174, 374 605, 608, 639, 668, 685, 705, Wu, Y.-W. 412
von Neumann, J. 257 731, 771, 782, 809, 819, 832
von Staudt, K. G. C. 387 West, Don. 106 Xiao, S. 24
Voss, H.-J. 326 West, J. 442 Xiao, Y. 65
Voss, H.-J. 407 Westcott, P. 433 Xin, G. 442
Vrt’o, I. 798 Wetzel, J. E. 388, 565
Škrekovski, R. 297 White, D. E. 100, 175, 566 Yackel, J. 449, 451
Vu, V. 488 Whiting, P. D. 246 Yaglom, A. M. 24
Whitney, H. 158, 293, 300, 306, Yaglom, I. M. 24, 419
Wagner, D. 647 314, 380, 398, 401, 418, Yamamoto, K. 558
Wagner, K. 390, 393, 398, 528 514–5, 521, 524, 527, 537 Yamashita, T. 316, 326–7, 334
Wagon, S. 221, 343, 655 Whitworth, W. A. 33, 40, 46, 115 Yan, C. H. 50
Wakabayashi, Y. 327 Wiedemann, W. H. 176 Yancey, M. 276, 355, 414, 420,
Wall, C. E. 227 Wiener, G. 334 422
Wallis, W. D. 215, 609 Wiener, N. 586, 605 Yannakakis, M. 573, 754, 756
Walsh, T. R. 80 Wierman, J. C. 828 Yao, A. C. C. 447–8
Walter, J. R. 373 Wilf, H. S. 87, 89–91, 112, 117, Yao, B. 819
Walters, I. C., Jr. 779 129, 135–6, 140, 190, 338, Yao, F. F. 815
Wang, D. L. 226, 501 769, 775 Yap, H. P. 373, 654
Wang, D.-W. 566 Will, T. G. 238 Yates, F. 612
Wang, E. T. 174 Williamson, J. 622 Yeh, Y. N. 144
Wang, J. F. 819 Williamson, S. G. 566 Yellen, J. 377
Wang, P. 501 Williford, J. 630 Yeo, A. 263, 274, 638
Wang, R. 356 Wilson, R. J. 37, 46, 361, 699 Yoshimoto, K. 327
942 Author Index

Young, A. 189–90, 207 Zehavi, A. 315 Zhou, L. 66

Young, D. 461 Zeilberger, D. 87, 89–90, 178, Zhu, B. W. 292
Yu, G. 423–4 192, 557 Zhu, Q. C. 458
Yu, X. 134, 314–5, 421, 424 Zernitz, H. 34 Zhu, X. 356, 403, 434, 442, 461,
Yu, Z. G. 326 Zha, X. 327 742
Zhan, S. 326 Zhu, Y. 352
Zabrocki, M. 433 Zhan, X. 249 Zhu, Y. J. 326
Zak, J. 819 Zhang, C.-Q. 312, 423 Ziegler, G. M. 37, 476, 804,
Zaker, M. 341–2 Zhang, K. M. 316, 449 811–3, 815
Zaks, J. 804 Zhang, L. 413, 422 Zielinski, P. 655
Zamani, R. 326 Zhao, H. 65 Zito, J. 264
Zarankiewicz, K. 793 Zhao, Y. 414–5, 422 Ziv, A. 764
Zaslavsky, T. 237 Zhou, B. 248 Znám, Š. 296, 631
Zeckendorf, E. 61 Zhou, H. 262 Zykov, A. A. 354, 476
Glossary of Notation

Relations, operators, positional notation

{} - set
∩ , ∪ , ∈ ,  - intersection, union, membership, symmetric difference
⌊x⌋ , ⌈x⌉ - floor, ceiling (nearest integer at most or at least x)
[ ] - {1 , . . . , }
-object - “object ” with parameter value
[a , b] - {a , . . . , b}, if a , b are integers
[x , y] - { : x ≤  ≤ y}, if x , y are real numbers or poset elements
[x¾ ] - coefficient of x ¾ in a formal power series in x
[S, T] - cut consisting of all edges from S to T
⟨v1 , . . . , vn⟩ - path with vertices v1 , . . . , vn in order
[v1 , . . . , vn] - cycle with vertices v1 , . . . , vn in order
n
n! - ∏i=1 i
⌊(n−1)/2⌋
n!! - ∏i=0 (n − 2i)
n(r) - n(n − 1) · · · (n − r + 1)
n(r) - n(n + 1) · · · (n + r − 1)
d | n - d divides n
d  n - d does not divide n
|S| - size of set S
#{i: i ∈ S} - size of described set
#(G) - number of components of a graph
A T - transpose of a matrix A
S - complement of set S (within a given universe)
G - complement of graph G
G ∗ , P ∗ , M ∗ - dual (of a planar graph, poset, or matroid)
↔ - adjacency relation on vertices of graph
→ - succession relation on nodes (vertices) of digraph
 - incomparability relation on elements of poset
≺ ,  - cover relation on elements of poset
an ∼ bn - asymptotic to (ratio approaching 1)
  n -  partitioning the integer n, with  1 ≥ · · · ≥  ¾ and ∑  i = n

943
944 Glossary of Notation

G − v, G − S, G − e, G − F - deletion of vertex, vertex set, edge, edge set

G · e , M · e - contraction of edge or matroid element
G − - graph formed by deleting one edge from an edge-transitive graph G
G + - graph formed by adding one edge to a graph G with edge-transitive G
M | F , M.F - restriction or contraction of a matroid to the elements F
0 , 1 - unique lower and upper bounds of a poset
1n - column vector of n 1s
2 , n - two-element, n-element chains in posets
2 n - subset poset of order n
2 E - subsets of E, ordered by inclusion
(n¾ ) - binomial coefficient counting ¾-subsets of the set [n]
(¾1 ,...n ,¾m) - multinomial coefficient
(S¾ ) - the collection of ¾-subsets of set S
G + H - disjoint union of graphs or posets
mG - disjoint union of m copies of graph or poset
G H - join (disjoint union plus all edges between)
G[S] - subgraph of G induced by S
G[H1 , . . . , Hn], G[H] - lexicographic product, composition (graphs or posets)
G H - Cartesian product of graphs
G × H - strong product
G ∗ H - weak product, tensor product
P × Q - Cartesian product of sets or posets
P ¾ - Cartesian product of ¾ copies of poset P
G ¾ - graph on V(G) with u ↔ v if d G(u , v) ≤ ¾
P¾ - ¾ th rank of a ranked poset P
x ∨ y - least upper bound in poset (join)
x ∧ y - greatest lower bound in poset (meet)
X − , X + - down-set and up-set generated by set X in poset

Usage of Roman alphabet

A(G) - adjacency matrix of a graph G

A(P) - incidence algebra of a poset P
An - n-element antichain
A(n , ¾) - Eulerian number counting permutations of [n] with ¾ runs
A(x) , B(x) - typical generating functions with formal variable x
AP - arithmetic progression
box G - boxicity of a graph G
B(G) - bandwidth of a graph G, block graph of G
Bn - Bell number counting partitions of [n]
B(M) - bases of a matroid M
Bin (n , p) - binomial distribution (n trials, success probability p)
BIBD - balanced incomplete block design
c(G) - circumference of a graph G
c(n , ¾) - number of permutations of [n] with ¾ cycles
ch(G) - choosability (list chromatic number) of a graph G
Glossary of Notation 945

cr(G) - crossing number of a graph G

cr(G) - (recti)linear crossing number of a graph G
cr ¾ (G) - crossing number when each edge is at most ¾ segments
- complex numbers
Cn - cycle of length n, nth Catalan number
C(P) - set of critical (unforced) incomparable pairs of a poset P
C(M) - circuits of a matroid M
d(v) , d G(v) - degree, valence (of a vertex v in G)
d(G) - average vertex degree in a graph G
d(u , v) , d G(u , v) - distance between u and v (in a graph G)
d1 , . . . , d n - degree list
d¾ (P) - size of maximum ¾-family in a poset P
d̂¾ (P) - size of maximum ¾-cofamily in a poset P
d m ,n - Delannoy number counting up/right /diagonal paths from (0 , 0) to (m , n)
def(S) - deficiency of a vertex set S (given by o(G − S) − | S|)
df(S) - defect of a set S in a bipartite graph (given by | S| − | N(S)|)
det - determinant of a matrix
diam (G) - diameter of a graph G
dim (P) - order dimension of a poset P
¾ (m) - numerical shadow of -binomial expansion of an integer m
(F) - shadow of a set-family F
D[S] - down-set (ideal) of poset, generated by set S, often D[x]
Dn - number of derangements of [n]
D N - divisors of N , ordered by divisibility
e - base of natural logarithm, 2.71828...
e(P) - number of linear extensions of a poset P
e¾ (x) - number of linear extensions with x at height 
ex (n; H) - maximum number of edges in n-vertex graph not containing H
E - typical set of elements of a matroid
E(G) - set of edges in a (hyper)graph G
(X) - expectation of random variable X
EGF - exponential generating function
¾ - rank parameters of family F in a ranked poset
ë - number of Young tableaux with shape 
 (T) - in inclusion-exclusion, #elements whose properties are indexed by T
F n - nth classical Fibonacci number
F̂ n - nth adjusted Fibonacci number (equal to F n+1)
q - q-element field
F , G , H - typical graphs, digraphs, or hypergraphs
(G) - girth
G p - random graph in Model A (edge-probability model)
G m - random graph in Model B (fixed-size model)
(n , p) - random graph model with edge-probability p
G(P) - comparability graph of a poset P
h(P) - height of a poset P
h(x) - height of poset element x, or expected height of element
n
Hn - nth harmonic number, ∑ i=1 1/i
H¾ ,n - -connected Harary graph with n vertices
946 Glossary of Notation

i(G) - interval number of a graph G

I(M) - independent sets of a matroid M
I(P) - set of incomparable pairs of poset P
Int(P) - family of intervals in a poset P
I(x) - set of elements incomparable to x
I - identity matrix (diagonal matrix with 1s on diagonal),
J - square matrix with every entry equal to 1
J(P) - lattice of order ideals of a poset P , ordered by inclusion
K n - complete graph of order n
K m ,n - complete bipartite graph with parts of sizes m and n
K S,T - complete bipartite graph with parts S and T
K(n , ¾) - Kneser graph with vertex set ([n] ¾)
l(P) - length of a poset P (length of longest chain)
lg - logarithm in base 2
ln - natural logarithm
L(G) - line graph of a graph G
L(v) - list of colors at a vertex in a list assignment
L(m , n) - poset of m-tuples a with 0 ≤ a1 ≤ · · · ≤ am , componentwise ordered
Ln(q) - lattice of subspaces of finite vector space of dimension n over G F(q)
m - often the number of edges in a graph
m¾ (C) , m¾(a) - ¾-norm of chain partition or sequence [= ∑ i min{¾ , ai}]
M - a matching, a matroid, etc.
M e - chain product e1 × · · · × e n
M(G) - cycle matroid of a graph G
Mad(G) - maximum average degree among subgraphs of a graph G
MOLS(n , ¾) - family of ¾ pairwise (mutually) othogonal Latin squares of order n
M5 - the 5-element modular non-distributive lattice
n - typically the number of vertices or elements in a (hyper)graph or poset
N5 - the 5-element non-modular non-distributive lattice
N¾ (P) - size of ¾ th rank of a poset P
N(x) , NG(x) - (open) neighborhood of x in a graph G
N[x] - N(x) ∪ {x} (closed neighborhood)
N(S) - ∪ x∈S N(x)
N +(x) - out-neighborhood (successor set) of x in a digraph
N −(x) - in-neighborhood (predecessor set) of x in a digraph
N(m) - maximum number of pairwise orthogonal Latin squares of order m
 - set of natural numbers
 
0 - ∪ {0}
o(G) - number of odd components
o( (n)) - functions whose ratio to (n) approach 0
O( (n)) - functions bounded (for large n) by a constant multiple of (n)
O¾ - Odd graph (disjointness graph on -sets and ( + 1)-sets in [2  + 1])
O(Q) - order polytope of a poset Q
OGF - ordinary generating function
OA - orthogonal array
p(n , ) - threshold probability function
per - permanent of a matrix
(A) - probability of event A
Glossary of Notation 947

Pn - path with n vertices

P(n , ¾) - generalized Petersen graph
PIE - inclusion-exclusion principle
P(L) - poset of join-irreducibles of a lattice L
PBD - pairwise balanced design
P( ) - P-symbol of a permutation
Q( ) - Q-symbol of a permutation
- rational numbers
Q¾ - hypercube of dimension  (as a graph)
r(M), r(P) - rank of a matroid or poset
r(x), r(X) - rank of an element in a poset, rank of a set in a matroid
r M - rank function of matroid M
 - real numbers
R B(x) - rook polynomial of the board B
R(p1 , . . . , p¾ ; r) - Ramsey number for -coloring r-sets
R¾ (p; r) - Ramsey number for -coloring r-sets with common threshold p
R(p , q) - Ramsey number for 2-coloring 2-sets
R(G 1 , . . . , G ¾) - graph Ramsey number
 n - symmetric group on n elements (set of permutations of [n])
S½ - surface with  handles
S(n , ) , s(n , ) - Stirling numbers of second and first kinds
SDR - system of distinct representatives
STS(v) - Steiner triple system of order v
T - typical tree or tournament
T n ,r - Turán graph (equipartite complete r-partite n-vertex graph)
U[S] - up-set in poset generated by set S, typically U[x]
V(G) - set of vertices of a graph or hypergraph G
w(l , ) - van der Waerden number (guaranteeing l-term AP in -coloring)
w(P) - width of a poset P
W¾ (P) - size of  th rank of a poset (Whitney numbers)
X - typical random variable
X + , X − - positive and negative parts in a signed involution
X , Y - typical bipartition of a graph into two independent sets
Y - Young lattice of all partitions of integers, ordered componentwise
 - integers

Usage of Greek alphabet

(G) - maximum size of independent set in a graph G

(G) - maximum size of a matching in G (pairwise independent edges)
(G) - minimum number of vertices covering all edges in G
(G) - minimum number of edges covering all vertices in G
(G) - genus, domination number of a graph G
i , j - Kronecker delta (1 if i = j , 0 if i = j)
(G) - minimum vertex degree of a graph G
(G) - maximum vertex degree of a graph G
- a simplex
948 Glossary of Notation

¾ (P) - d ¾ (P) − d ¾ −1 (P) for a poset P

ˆ ¾ (P) - d̂¾(P) − d̂¾−1 (P) for a poset P
(v) - eccentricity of a vertex v
,  ,  ,  ,  - incidence functions on posets
(G) - minimum number of cliques to cover all vertices in a graph G
(G) - connectivity of a graph G
 (G) - edge-connectivity of a graph G
(x , y) - local connectivity for vertices in a graph
 (x , y) - local edge-connectivity for vertices in a graph
(x , y) - maximum number of independent x , y-paths
 (x , y) - maximum number of edge-disjoint x , y-paths
 1 , . . . ,  n - eigenvalues (of adjacency matrix)
Λ n - partitions of integer n, ordered by refinement
(n) - Möbius function of an integer n
(x , y) - Möbius function on interval [x , y]
1 , . . . , n - Laplacian eigenvalues of a graph
 - circumference of circle with unit diameter, 3.1415926...
 ,  , - typical permutations
Π n - lattice of partitions of an n-set, ordered by refinement

- density
 (X) - span of a set X in a matroid
- universe of sets for intersection classes, surface
∑ - summation
(G) - number of spanning trees in a graph G
(G) - minimum partition into complete bipartite subgraphs
 (G) - arboricity of a graph G
 - a mapping, often an isomorphism, coloring, or homomorphism
(m) - Euler totient function (numbers in [m] relatively prime to m)
(G) - chromatic number of a graph G
(G) - edge-chromatic number of G
(G) - total chromatic number of G
(G) - list-chromatic number of G
(G) - list-edge-chromatic number of G
(G) - clique number of a graph G
n - unbounded sequence
(A) - intersection graph of collection A
(∑) - intersection class for subsets of ∑
( (n)) - functions at least a constant multiple of  (n) (for large n)
Subject Index
A page number in italics indicates a definition. A single listing in italics may
indicate the definition for a concept so prevalent (such as “graph”) that it would not
be productive to list its occurrences. An item that appears on few pages may have
none italicized for the definition.
Page ranges in bold indicate material such as the proof of a major result or the
main treatment of the concept; this may also include a definition. Pages ranges may
also include isolated pages where the term does not appear.
Parenthetic clarifiers act as subheadings. Terms consisting of a word modified by
prefatory notation are alphabetized according to the root word.

absorption property 516, algebraic connectivity 778 Arithmetic–Geometric Mean

520–2, 536, 538 Alon–Tarsi number 742 Inequality 709
abstract dual 526–7 Alon–Tarsi orientation 742 arithmetic m-grid 474
achromatic number 342 Alon–Tarsi Theorem arithmetic progression 2, 61,
acyclic coloring 423 738–42, 746 463–7, 473–4, 479, 484–6,
acyclic graph 239–40 alphabet 17–20, 80, 104, 568, 685
acyclically ¾ -choosable 423 118–21, 132, 336, 441, Art Gallery Problem 394,
adjacency 210 566, 667 419
adjacency matrix 211, 216, alteration method 670, 673 ascent (in permutation) 65,
224–5, 234–5, 256, 747, alternating cycle 413, 423, 101, 106, 178, 197
756, 766–82 572–3, 583–4, 755–6, 765 aspect (of hereditary system)
adjacency relation 3–4, 218 alternating path 277–8, 516, 520–1, 528–30
adjacent 4, 210 282–3, 287, 518 Assignment Problem
adjugate matrix 748, 777 alternating permutation 279–82, 287–8
adjusted Fibonacci numbers 177–8, 207 associativity 179, 545, 588,
53, 59–61, 67, 104, 112, angle orders 828–9, 832 614, 759
115, 442, 664–5, 740 anti-Ramsey theory 457 Asymmetric Local Lemma
admissible transposition 208 antichain 496–500, 509, 511, 681
advance 197, 201–2 515, 520, 524, 543–64, asymptotic 9–10, 82–7,
affine m-cube 473 569, 575, 581, 583–4, 589, 657–60, 693–705, etc.
affine hull 744 593–4, 603–6, 666, 668 asymptotic series 86
affine plane 628, 649–50, 655 antipodal 782, 806–12, 819 AT-orientation 742–3, 747
affine space 69, 802 antisymmetric 8, 225, 541, augmentation property
Ahlswede–Daykin 621–2 518–22, 525, 529, 536–9
Inequality 597–9, 608, 823 arborescence 749–51 augmenting path 277–84,
Alexandrov–Fenchel arboricity 412, 421, 535, 677 287, 519
Inequalities 827 augmenting trail 287
949
950 Subject Index

automorphism (of graph) base (in vector space) 122–3, 78, 86, 88–9, 92, 111, 132,
212–4, 216, 218–9, 250, 568 156, 166, 178, 658, 734
459, 565, 666, 693, 699 Bell number 126, 134, 147 binomial convolution
automorphism (of poset) Berge cycle 435 118–20, 123–6, 132
565, 606 Berge’s Theorem 501, 512 binomial distribution 687,
average distance 251 Berge–Tutte Formula 707–10, 714, 720–1
-avoiding permutation 265–6, 272–4, 277  -binomial expansion 494–5
431–3, 442 Bernoulli number 123, 133 binomial inversion 124, 132
Aztec diamond 34 Bernoulli random variables rth binomial moment 694–6,
Azuma’s Inequality 706, 707, 712 704
712–7, 722 Bertrand’s Ballot Problem binomial random graph
39, 46, 49, 189 model 687–8, 692
BEST Theorem 752 Binomial Theorem 18, 26–7,
backbone chromatic number
biadjacency matrix 256–7, 33, 77, 85, 119–23, 157,
342
754–6 161, 185, 674
balanced charging 409–11,
biclique 770–1, 781 binomial type 123
413, 415, 422, 424
X , Y -bigraph 211, 216, 227, biorder 605
balanced edge cut 710–1
253–68, 278–9, 287, 332, bipartite graph 211, 231,
balanced graph 692–3, 702,
359, 364–5, 480, 486–7, 253–97, 359, 364–5,
822–4, 831
492, 517, 522, 566–7, 631, 368–9, 517–8, 529, 531,
balanced incomplete block
672–3, 755–6, 779–80 547, 660, 667, 746–7,
design (BIBD) 612–8
bigraphic 227 754–6, 767–8
balanced signed graph 237
bijection 2–4, 8, 15–6, 25, 36, bipartite poset 547, 575–80,
balanced tableau 208
41–50, 63–4, 99, 170, 177, 584, 790, 832
 -balanced pair 822–4, 831
196, 212, 442, 526–7, 538, bipartite Ramsey number
ballot list 41–5, 48–9, 56, 63,
545, 626, 741, 764, 781 620, 624
605
Bijection Principle 15 bipartite tournament 334
q-ballot list 42, 50, 64, 128
bijective proof 15, 22, 29–30, bipartition 211, 223, 231,
ballot path 41, 45, 49, 55,
55–6, 63, 66, 105–6, 115, 280, 329, 426, 456, 770
166
132–5, 141–3, 148, 192–3 biplane 616
Ballot Problem 39, 46, 49,
binary n-tuple 2, 25–6, 53, Birkhoff diamond 405, 421
189–90
58–9, 62, 236, 426, 507, Birkhoff–von Neumann
bandwidth 703, 799–801,
555, 580, 619, 699, 722, Theorem 257, 261
818–9
724, 726, 729, 745, 754 bisection 810, 820
bar visibility graph 394
binary list 15, 21, 25, 39, 41, block (in a design) 9, 612–5,
barred permutation 102, 158
93, 116, 121, 505, 513, 669 622–6, 634, 640–1, 645–6,
X , Y -barrier 299, 395
binary matrix 2, 431–4, 510, 649–56, 730
barycenter 251
744 block (in a graph) 136,
barycentric coordinates 783
binary maze 669 294–7, 314, 341–2, 348,
barycentric representation
binary relation 2, 8, 224, 355, 606
784–5, 788–90
542, 605 block (in a matrix) 432–3,
base exchange graph 541
binary search 84, 91, 447 462, 611, 618, 763, 778
base exchange property 516,
binary tree 43–4, 47–9, 135, block (in a partition) 3, 46,
520–4, 537, 539, 541
440 49, 56, 64, 121–2, 125–7,
base (in ideal of sets) 501–2,
Binet’s Formula 68, 664 133–4, 145–7, 157, 175,
512
binomial coefficient 18, 194, 488, 522, 567–8, 588,
base (in matroid) 515–6,
24–35, 40, 56, 59, 62, 74, 605, 607, 678, 716, 761
520–6, 530, 536–41
Subject Index 951

block design 9, 609, 612–9, Brooks’ Theorem 337, 342, Cayley’s Formula 36–7,
623 347–8, 355, 371 46–7, 58, 127–8, 136, 747,
block structure 463–7 broom 239, 352 763, 777–8
block-cutpoint graph 295, Brouwer Fixed-Point Cayley–Hamilton Theorem
297 Theorem 802–6, 819 768
blow-up graph 269, 330 Bruck–Chowla–Ryser cell 154, 191–3, 196, 578,
Blow-up Lemma 481 Theorem 616–7, 626 784–9, 799–807
blow-up poset 560 Brun’s sieve 696 center 233, 239, 243, 250–1,
board of forbidden positions Brunn–Minkowski Theorem 441, 502, 620, 771, 828–30
160–3, 171–2, 175–6 825 central Delannoy number
bond 293, 297, 380–1, 386, Bulgarian Solitaire 143 28, 34, 52, 62, 64, 117
395, 525–7, 763 bumping procedure centroid 251, 440, 474,
bond matroid 525–7 195–200, 203, 206, 208 824–7
bond space 752–3, 763–4 Burnside’s Lemma 179–85, chain 538, 543–71, 574,
Bondy–Chvátal Theorem 187 583–92, 595–7, 600–8,
320 757–62, 821–3, 831
Bondy’s Lemma 323–5 chain decomposition 546–8,
cactus 295
Bonferroni Inequalities 553–7, 561–2, 565–7, 606
(¾ , ½)-cage 219
695–6 chain rule 125, 130, 506–8
canonical basis 758
book 763 character 622
canonical coloring 443, 470
Boolean algebra 545 characteristic equation
canonical cycle represen-
Borsuk’s Conjecture 817–8 67–80, 173
tation 99–100, 104–6
Borsuk–Ulam Theorem 798, characteristic function 17,
Canonical Ramsey Theorem
806–12, 815, 817, 819 597–600
461, 470–2, 474
Bose Construction 642–3 characteristic polynomial
canonical simplex 824, 827
bottleneck 252 67–71, 74, 79, 81, 761,
cardinality 2–3
boundary 11, 379–93, 398–9, 766, 780–1
Caro–Wei Theorem 662–3,
402, 411, 415, 719–20, characteristic root 67–72,
684
755, 784, 788, 796, 800–1, 76, 79, 81, 84
cartesian product 2, 7, 14,
806–9, 813, 816, 819 charge 406–16, 421–4
214–6, 224, 244, 292, 326,
boundary bound 800 Chebyshev’s Inequality
329, 334, 344, 359, 372,
¾ -bounded 327–8, 334 595–7, 692, 703, 706–10,
394, 541, 545, 665, 716,
Ò-bounded 351–2 721–2
771, 773, 780–1, 804, 822
Bounded Differences Chernoff Bound 706–14,
Cassini’s Identity 60–1
Condition 714–6, 719, 722 720–2
Catalan numbers 41–6, 48,
bounded poset 571–2, 762 Chernoff–Hoeffding Bound
52, 55–6, 63–4, 77, 115–6,
boxicity 705 709, 722
136, 177, 206, 432, 587
bracketing structure 555–8, Chevalley–Warning
Catalan recurrence 55–6,
565–6 Theorem 735–6
63–4, 77, 136
branch vertex 327, 334, 356, child 43–4, 47, 136, 177, 443
Catalan’s Identity 61
390, 392, 710 Chinese Postman Problem
caterpillar 187, 654, 819
¾ -branched 327 287
Cauchy–Binet Formula 168,
branching 749, 751 Chinese Remainder
748–9, 763
Breadth-First Search (BFS) Theorem 441
Cauchy–Davenport Theorem
247, 252, 336 choice number 346–7, 355–6,
733–4, 745
bridge 231, 233, 765 660–1
952 Subject Index

choosability 346–8, 362, 365, circuit (in graph) 233–6, ¾ -colorable/¾ -coloring 182,
403, 407–8, 414, 417, 424, 239, 250, 268, 307, 389, 335–46, 352–7, 369,
660, 739, 742, 747 419, 678, 738, 751–2 399–407, 414–22, 434–9,
(¾ , d)-choosable 403–4 circuit (in matroid) 514–29, 443–4, 448–9, 453–69,
º -choosable 347, 356, 363, 536–9, 657 473–4, 573, 581–2, 601–3,
365, 738 circulation 738–44 638, 658–62, 666–7,
¾ -choosable 346–7, 353–6, circumference 322–5, 332 673–5, 684–6, 801–4,
363–5, 402–3, 414, 417, Class 1 358–9, 369–73, 812–4, 820
420–4, 623, 660, 667, 677, 422–3 2-colored bipartite graph
738–40, 746–7 Class 2 358–9, 372 135
chordal graph 366–7, 370, claw 212, 297, 330 ¾ -colored graph 104
374–6, 386 clique 210, 336, 357, 366–71, coloring 158–9, 335
chordless cycle 366–7, 374 374–6, 444, 459, 685, 700, H-coloring 422, 457
chord (geometric) 21, 24, 48, 726, 729 L-coloring 346–7, 355–6,
79, 82, 128, 387–8, 792 ¾ -clique 210, 337, 375, 363–5, 403, 407–8, 661,
chord (in graph) 213, 216, 449–50, 476–8, 483, 667, 677
315–6, 348, 366–7, 377, 489–90, 504 coloring number 338
382, 393, 396, 401, 403, clique covering 368 column matroid 516–7, 537
411, 415, 550, 698, 742–3, clique number 336, 343, 368, column-canonical
786, 796–7 374, 685, 687, 700 permutation 208
¾ -chromatic 335, 340–4, clique Ramsey number 455 column-strict tableau 189,
349–57, 405, 421, 434–7, clique tree 374 192–6, 203–5, 207
442, 451, 461, 478–9, 682, closed ear 305–7 Columns Condition 462–3,
684, 709–10 closed form 89 474
chromatic index 357–62 closed set (geometric) 811, Combinatorial
chromatic number 335–57, 813, 817 Nullstellensatz 732–8,
366–70, 402, 421, 434–8, closed set (in matroid) 741–2, 745–6, 764
442–3, 451, 454, 550, 528–9, 534–9, 595 combinatorial proof 15,
572–3, 581, 671–2, 684–5, closure (geometric) 807, 811 22–3, 26, 31–4, 40, 48,
701, 709–10, 716–7, closure (Hamiltonian) 60–6, 96, 102, 109, 114–7,
729–30, 738–9, 745, 769, 320–1, 325, 332–3 132–4, 156, 173, 176, 208
812–4, 820 closure operator 529 combinatorial reciprocity
chromatic polynomial 159, cobases 523–4 762
174, 374, 762 cocircuit 523–4, 537–9 Combinatorial Shearer’s
chromatic Ramsey number cocycle matroid 524 Lemma 508–9
455–7, 461 coefficient operator 73, 96–7, Committee-Chair Identity
Chvátal’s Condition 320–1, 110, 733 26, 40, 42
326, 332–3, 705 cofactor 748–9, 752, 777 common system of distinct
Chvátal’s Conjecture 318, cograph 375, 778, 782 representatives (CSDR)
501–3, 512 coherent signing 755–6 302–3, 314, 532, 562
Chvátal’s Theorem 320–1, colex ordering 493–5 compact set 805, 809
332–3 co-lexicographic 493 Compactness Principle
Chvátal–Erdős Theorem color fan 360–1 468–9
321–2, 327, 332, 334, 705 ¾ -color Ramsey number 459, comparability digraph
circle order 828–30, 832 461, 601 542–3, 548, 570
color-critical 344–5, 352–4,
820
Subject Index 953

comparability graph 374–5, configuration (planar convex combination 256–7,

542–5, 547, 549–50, 570, coloring) 404–8, 421 261, 512, 783–4, 797, 802,
575, 582–3, 803 configuration (discharging) 805
comparable 541–4, 568, 590, 411–7, 423–4 convex embedding 392–3,
762 (6 , 3)-configuration 684 397–8
compatible orientation 762 configuration model 697 convex function 505, 620,
compatible vertices 315 conflict graph 398 630–1, 663, 672, 681, 709,
complement 2, 175, 210 congruence class 3, 38, 103, 713, 794
complement reducible graph 142, 149, 338, 427, 441, convex n-gon 21, 43–4, 48,
375 477, 606, 613, 635–6, 56, 79, 82, 165, 387–8,
complete d-ary tree 440, 641–2, 646, 727 446, 458
1 , 2 -Conjecture 823
443, 667 2 3 convex property 688
complete bipartite graph conjugate partition 141–3, convex region 388, 446, 816,
211–2, 236–7, 334, 346, 150, 152, 192, 208 824–5
375, 547, 631, 770 connected graph 4–6, 124–7, convolution 26, 31, 94, 108,
complete family (of Latin 135–6, 187–8, 229–52 110–1, 117–20, 123–6,
squares) 610, 623–4, ¾ -connected 289–306, 311–7, 132, 563–4, 757–60
627–8 323–6, 329–34, 347–8, copy (as a subgraph) 212
complete r-uniform 353, 386–402, 411, 415–8, cost 279–82, 394, 399, 447
hypergraph 435, 475, 491, 422, 425, 702, 742–3, 784 countable 2, 6–7, 469, 662
601, 608, 613, 616 connected set 229–30 Counting by Cases 14
completely labeled cell 799, connection relation 4–5, Counting by Stages 14
801–6 229–30 counting sequence 51–2
component (of graph) 4, 229 connectivity 289–92, counting variable 662,
component structure 125–7, 295–304, 313, 322, 326–7, 689–94
188 700, 721, 778 Coupon Collector 668, 704
composition (of an integer) connector 801–2 cover (weighted) 279–82
19–21, 23, 30, 45, 53, 60, Connector Lemma 803–4 cover diagram 542–4, 549,
81, 137, 146, 173–4, 606 containment poset 553, 552, 559, 565
composition (of functions) 3, 588–9, 606, 828–32 cover digraph 542–3, 550
98, 125, 127, 130–1, containment relation 3, 8 cover graph 542–5, 549–50
179–80, 212, 427, 809 containment representation cover relation 542, 544, 550,
composition (of graphs / 828 605, 759
posets) 376, 459, 583 contractible edge 304–5, covering set 281
Compositional Formula 127 313, 785–6 -critical 264
compound event 675 contraction 304, 313, 349,  -critical 344–6, 349–55,
Condensation Method 170, 381, 390, 392, 398, 524–7, 402, 415, 420
178 540, 710 p-critical 369–70, 375–6
conditional entropy 506, 508 Contraction Lemma 304,  -critical 568
conditional expectation 713 393 jth critical class 464
conditional probability 7, contractive mapping 442 critical graph 361
191, 506, 679, 713, 718 Convergence of Moments p-critical graph 369, 375
conditional random variable Method 694–6 critical pair 573
713 converges (formal power cross-section 824–5
conference matrix 621–2, series) 107, 109, 125, 140 crossing 377–81, 393, 405,
624 420, 663, 698, 779, 790–8
crossing number 790–8, 832
954 Subject Index

crown 22, 185, 584 cyclic rotation/shift 41–2, 45, deletion method 670–2, 682,
cryptomorphism 529 48, 50, 145, 178, 185, 634, 684–5, 702
CSDR 302–3, 313, 532 643–4 deletion operator 208
cube 182–3, 214, 219, 248, density 371, 407, 479–87,
251–2, 260, 291, 296, 492, 693, 703, 778–9,
d’Ocagne’s Identity 60–1
329–30, 358, 373, 380, 800–1, 810, 819
decision problems 10
385, 388, 428, 461, 467, dependent edge 550
decomposable matrix 765
473–4, 566, 584, 668, 684, dependent event 679, 685
decomposition (of graph)
720, 764, 781–2, 798, 800, dependent set 514–5, 527,
232–9, 249–51, 275,
803–5, 820–2, 829 538
305–7, 312, 315, 318–9,
cubic graph 210 Dependent Random Choice
389, 407, 411–2, 422, 534,
curve 377–81 672–3, 684
639–40, 646–50, 677–8,
cut-edge 231, 237–41, 249, dependent vectors 748, 753
771, 781
252, 267, 272–6, 295, 305, derangement 46, 52, 54–5,
decomposition (of poset) 546,
373, 379–80, 523, 753 59, 62–3, 82–3, 123–4,
548, 553–7, 561–2, 565–7,
cut-vertex 231–2, 240, 276, 134–5, 154–5, 159–62,
574, 606
289, 294–5, 297, 315, 371, 174–5
deconditioning 506–8
382, 391, 421 derivative (for series) 107–8,
Dedekind’s Problem 557–8,
cutset 296, 375–6, 550 116, 125, 130
603
¾ -cycle 135, 184, 186, 213, descent (in permutation) 65,
Defect Formula 258, 262,
218–20, 252, 323–4, 372, 101, 148, 172, 197, 206,
265, 531, 540
377, 411–5, 420–3, 457, 667
deficiency 258, 265–7, 275,
460, 630, 646–7, 684–5, descent (in poset) 602
567
755, 780 descent set 101, 172
degeneracy 338, 347, 639,
cycle (in graph) 4–6, 209, (v , ¾ , ë)-design 612–29, 634,
667
231–4, 316–48, etc. 639–41, 645–6, 649–50,
¾ -degenerate graph 338,
cycle (in digraph) 36–7, 60, 653–6
343, 347–8, 373, 402, 417,
144–5, 152, 225–45, 389 deterministic algorithm 10
419, 421, 484, 677, 739
cycle (in hypergraph) 435–7 de Bruijn cycle 239, 763
degree (of vertex) 5
cycle (in permutation) 8, Diagonal Criterion 724, 728,
degree charging 407–10,
54–7, 62, 66, 98–100, 731, 818
414, 423
104–6, 109, 112, 122, 126, diagonal Ramsey number
degree list 221–2, 238–9,
135, 147, 172, 175, 181–6 449, 465, 504, 675, 685,
249, 269, 295, 297, 320–1,
cycle index 181–7 701
332, 387, 398, 663
Cycle Lemma 42, 49, 425 diagram (of poset) 542–4,
degree Ramsey number 455
cycle matroid 515–8, 523–9, 549, 552, 559, 565
degree sequence 221
536–40 diameter 243–5, 248–52,
degree set 227
cycle representation 99–100, 275, 294, 296, 313, 386–7,
degree-choosable 347–8,
104–6 395, 399, 428, 441, 566,
355, 415
cycle space 752–3, 764 628, 630, 689–91, 704,
Degree-Density Lemma 484
cycle structure 100, 182, 768, 775–6, 780, 782,
Degree-Sum Formula 220,
185–6 817–8
223, 233, 290, 311, 380,
cycle-factor 260 difference family 639
385, 476–7, 527, 630
Cycle-plus-triangles difference set 460, 634–9,
Delannoy number 28–31,
Theorem 739–40 645
33–4, 52, 57, 62, 64, 81,
cyclic edge-connectivity 297 difference triple 645
113, 117
cyclic group 185 digraph 8, 165–9, 224–6, etc.
Subject Index 955

dihedral group 186 distance (between points) dual graph 379–89, 395,
Dijkstra’s Algorithm 246–7 342, 440, 490, 619–20, 400, 406, 409, 419, 513,
dilation 800 633, 639, 719–20, 724, 524, 799, 803
Dilworth decomposition 546, 729–30, 744–5, 793–4, dual hypergraph 438–9
548, 551, 554, 556, 566, 805, 809, 811, 817–9, 826 dual ideal 545
574 distance (in graph) 189, dual matroid 523–7, 537–40
Dilworth’s Theorem 430, 242–52, 287, 314, 325, Dual Möbius Inversion
546–8, 551–3, 564, 580, 336, 374, 389, 399, 422, Formula 760
606–7, 823 428, 442–3, 583, 667–8, dual poset 543–4, 549, 553,
dimension (of poset) 545, 739, 782, 800–2 557, 560, 562, 567, 591–2,
568–84, 607, 783, 790, ¾ -distance set 724, 730, 744, 607, 832
828–32 817 dual problem 259, 272, 276,
dimension (of space) 9, 29, distance-regular 782 279–83, 287–8, 298–9,
39, 48–9, 64, 68–9, 76, 96, distributive lattice 591–600, 366, 368
170, 261, 388, 431, 464–7, 606–8 Durfee square 143
474, 520, 568, 606, 633, divide-and-conquer 84, 394 Dushnik’s Theorem 576–7,
711, 723–30, 744–5, 753, F-divisible 640 584
763–6, 814–8, 821–32 Dixon’s Identity 88 Dyck n-path 45, 48, 64, 151
¾ -dimension 584 Dobiński’s Formula 134 dynamic coloring 424, 668
dimension argument 723–30 dodecahedron 316, 328, 385,
d-dimensional permutation 416
ear decomposition 305–7,
821, 831 domain 2, 17, 99, 190, 444
312, 315, 389
d-dimensional simplex 799, dominating set 639, 670–1,
eccentricity 243, 250
802–8, 822, 827 684, 702, 704
edge 4
Dirac’s Condition 320, 327, Doob process 715–6, 719
edge cover 259–60, 263
330 dot product 9, 614, 618–9,
edge cut 293–8, 306–7, 345,
Dirac’s Theorem 320, 324 724–6, 745, 748, 817
372, 380–1, 395, 710–1,
direct sum 533 double jump 698
752–3, 763, 778–9
Directed Matrix Tree double shift graph 581
edge set 4
Theorem 749–50 double-partition 150
edge-choosability 362–5,
Dirichlet Drawer Principle double-star 249, 460
407–8, 413–4, 746
425 doubly stochastic matrix
edge-chromatic number
Discharging Method 407–17 256–7, 261, 765
357–62, 365, 370–3, 414,
disconnecting set 293 down-set 501, 544–6, 557,
420, 423, 439
discrepancy 722 564, 569, 581–2, 589,
¾ -edge-colorable/coloring
discrete probability space 593–608, 759, 828
357–65, 371–3, 400, 407,
6–8, 657, 662 DP-coloring 356
416–7, 420, 422, 685, 797
discrete random variable 7, drawing 4, 8, 211–3, 225,
edge-coloring (hypergraph)
111, 662, 695, 713, 722 377–85, 398, 421, 542–4,
438–9, 602–3
disjoint ¾ -split 305 663, 718, 791–8
L-edge-coloring 362–5, 413
disjoint union 215 dual augmentation property
¾ -edge-connected 267, 293,
disjoint-path system 539
296–8, 301, 305–15, 329,
166–70, 172 dual base exchange 520–1,
345, 355, 372, 386, 400,
disjointness graph/relation 538–9
416–7, 420, 536, 540, 797
3, 213–4, 218–9, 782 Dual Degree-Sum Formula
edge-connectivity 293–301,
displacement 22, 176, 186, 380
417
792
956 Subject Index

edge-exposure process 715, Erdős–Stone Theorem 700, 706–9, 712–7, 721,

718 478–82, 491, 770 793, 821–2
edge-threshold weighting Erdős–Szekeres Theorem exponential enumerator
376 206, 344, 430–1, 550, 580, 118–9
edge-transitive 212, 214, 601, 821 Exponential Formula 124–8,
218–9 establishes (incomparable 135, 172, 188
EGF 93, 118–29, 132–6, 175 pair) 569 exponential generating
eigenvalue 80–1, 616, 723, Euler Conjecture 612, 642, function 118
765–82 648, 651–4 exponential polynomial 134
EKR(¾ , t)-family 497–9, 511 Euler totient 155, 172, 177, extended binomial
elimination property 185 coefficient 26, 31, 74
517–21, 528, 536–9 Euler’s Formula 383–8, 397, Extended Binomial Theorem
elongation 539 401–2, 409, 537, 754–6, 27, 77, 85
embedding 377–85, 389–94, 784 external vertex 402–3, 785–9
397–9, 510, 571, 582, 589, Euler’s Identity 143, 150 extremal number 479
753–6, 781, 783–4, Euler’s Pentagonal Number extremality 223
789–90, 795–7 Theorem 151
Embedding Lemma 480–5 Eulerian circuit/graph
face 182–4, 187–8, 378–403,
embeds (for posets) 543–4, 233–6, 239, 250, 268, 297,
409–17, 421–3, 523, 742,
571, 582–3, 589, 607, 828 307, 312, 381, 386–7, 389,
747, 753–6, 784, 796,
¾ -ended 327 417, 419, 656, 678, 738,
801–4
endpoints 4–5, 8, etc. 751–2
j-face 401, 406, 410–5,
entropy 493, 504–10, 513, Eulerian number 101–3,
422–3
823 105–6, 111, 122, 133, 158
face charging 409–10, 415
Entropy Compression Eulerian polynomial 102,
facet 802–4, 807–8, 811
Method 680 135
factor (in graph) 268–78, 678
enumerator by w 93–5, 103, Eulerian trail 234
factor (in product) 18, 68–9,
118–9 even cycle 209, 237, 277, 295,
95–100, 108–13, 119–21,
equality subgraph 280–2 346–8, 363, 413, 415, 518
137–9, 184, 553, 591,
equipartition 480, 485, even graph 135, 178, 188,
727–31, 734–8, 741, 755,
815–6 232, 234, 237–8, 249, 295,
759, 761, 818
equitable coloring 372, 420 297, 754
1-factor 264–76, 314,
equivalence class 3, 4, 65, event 6–7, 16, 24, 171, 191,
329–30, 358–9, 647–8,
137, 145, 179–85, 212, 506, 595, 599–601, 608,
650, 678, 754–6, 764–5
230, 234, 236, 632, 764 657–62, 665, 668–9,
º -factor 268–71, 276
equivalence relation 3, 4, 8, 674–86, 689, 694–7, 702,
¾ -factor 264, 268, 274, 318,
145, 178–80, 212, 230, 632 711, 713, 716–8, 822, 824
358, 371, 372–3, 400, 419,
erasure 227 excedance 106
647–8, 764
Erdős Principle 20, 162 excess matrix 280–2
factor-critical 275
Erdős–Faber–Lovász exhaustive regular covering
factorial 17–8, 24, 33, 54,
Conjecture 438 568
82–6, 106, 118–28, 132–6,
Erdős–Heilbronn Conjecture expander 705, 779–80, 782
159–62, 190–2, 206–7
734 Expansion Lemma 301, 313
factorial moment 694
Erdős–Ko–Rado Theorem expectation/expected value
factorial polynomial 175
497–501, 512, 584, 727, 7–8, 104, 111, 162, 479,
factorization 89, 139, 152,
781 505, 561, 563, 596, 657,
428, 553, 593–4, 627, 647
Erdős–Rényi model 687 662–6, 670–2, 689–97,
1-factorization 358–9, 371
Subject Index 957

falling factorial 17, 22, 69, 412, 414, 419–23, 515–9, Gaussian polynomial 152,
122–3, 170, 176, 694 524, 535, 567, 677–8 568
family 2 formal Laurent series GCD matrix 177
¾ -family 548, 549, 551, 130–1, 136 general solution (for
553–4, 561, 565, 567 formal power series 73, 77, recurrence) 67–72, 91
x , U-fan 301–2, 350 93–7, 107–8, 114, 116, generalized Catalan
Fan Lemma 301–2, 350 118, 125, 129–31, 136, numbers 42
Fan’s Theorem 324–5 140, 553, 733 generalized partition
Fano matroid 537 Four Color Theorem 316, matroid 534
Fano plane 435, 613–6, 349, 377, 399–401, 404–6, generalized Petersen graph
623–8, 634, 638, 641 414–23, 781, 797 373
Fáry’s Theorem 393, 398, Four Function Inequality generalized Young tableau
783, 796 597 189
Fermat’s Little Theorem fractional chromatic number ¾ -generated family 830
38–9, 48, 428, 441, 735–7 812 generating function
Ferrers diagram 141–5, fragment 398–9 (ordinary) 14, 73–82,
148–52, 175, 177, 185, H-free (graph) 223, 297, 93–118, 137–9, 142, 146,
189–90, 557, 604, 819 330, 342, 351–2, 375–6, 148–9, 161–2, 169, 172–3,
Ferrers relation/digraph 605 489–91, 631–3 178, 182–4, 431, 553, 557,
Fibonacci numbers 52, 53–4, free matroid 522 564, 567, 704, 707, 709,
58, 60–2, 65, 67–8, 80, Friendship Theorem 775–6 751, 761
104, 112, 115, 150, 434, frugal coloring 681 generating function method
442, 664, 740 fullerene 387 66, 72–81, 105, 124, 132
FILL (in tableaux) 202–4 functional digraph 8, 36–8, generating series 93–4
filter 466, 545 54, 166, 185, 224 geometric distribution 111
first-fit coloring 336 fundamental cycle 753 Geometric random variable
Fisher’s Inequality 614–5, fundamental set of circuits 668
641, 655, 730 537 Ghouilà-Houri graph 550
Five Color Theorem 402, Füredi–Hajnal Conjecture girth 213, 218–20, 237, 244,
404, 419 432–3 248–52, 275, 296, 312,
fixed point 8, 54, 59, 63, 98, Fuss–Catalan numbers 42, 332–3, 357, 388, 403,
105–6, 123, 135, 144, 152, 128 409–10, 414, 417, 422–4,
154–5, 159–67, 174, 178, 434–7, 442–3, 527, 550,
206, 250, 631, 666, 805–6 639, 671–2, 754
Gale’s Lemma 811–3, 820
FKG Inequality 596–600, Glaisher’s Theorem 149
Gale–Shapley Proposal
608 global discharging 413
Algorithm 285, 288
flat (in matroid) 529 Golden Ratio 68
Gallai coloring 474
flower snark 372 Gosper’s Algorithm 89–90
Gallai’s Theorem 260, 263,
Folkman’s Theorem 466–7, Grötzsch graph 340, 550
355, 540
474 Grötzsch’s Theorem 414, 420
Gallai–Edmonds Theorem
Ford–Fulkerson CSDR Grünbaum–Hadwiger–
268, 275
Theorem 302–3, 313, 530, Ramos problem 815
Gallai–Roy Theorem 339,
532, 562 graceful labeling 22, 646,
342, 344
forest 47, 128, 136, 239, 654, 656
Gallai–Witt Theorem 467
247–51, 262, 297, 311, Graceful Tree Conjecture
gambler problems 66, 116,
338, 351, 357, 381, 382–3, 646
714
958 Subject Index

graded poset 552–4, 558–62, Hall’s Condition / Theorem heterochromatic 470

565–8, 583, 589, 594, 254–66, 273, 302, 313, Hetyei’s Theorem 276
606–7, 761 365, 404, 551, 559–61, Hex (game) 803–4
Graeco–Latin square 612, 705, 779, 803 Hilbert’s Cube Lemma 473
656 Ham Sandwich Theorem Hiraguchi’s Inequality 575
(d , r, ½)-graph 443 815–6, 820 Hoeffding–Azuma
graph 4 Hamiltonian closure,318–9, Inequality 715, 722
graph polynomial 740–1 323, 330–1 Hoffman–Singleton graph
graphic list 221–2, 227, 276, Hamiltonian cycle 316–34, 244, 252, 775
295, 297 400–1 hole (in graph) 366
graphic matroid 515, 522–7, Hamiltonian decomposition i-homogeneous 444–5, 453,
536–7 318–9, 332 470
graphical design 645–6 Hamiltonian path 321, 329, homogeneous recurrence
greedy algorithm 245, 252, 334, 401, 418, 537, 666 67–71, 79
513, 519–22, 529, 536, Hamiltonian-connected homogeneous set 444–5,
538, 556, 566, 671, 326–7, 333, 401 461, 466, 468–70, 658, 729
greedy coloring 336–43, 345, Hamming distance 619, homomorphism 457, 461
357, 362, 371, 375, 677, 719–20 homothetic 467
701, 769 Hanani–Tutte Theorem 398 Hook-Length Formula
Greene numbers 201, 203–5 Handshaking Lemma 220 189–92, 207–8, 432, 746
Greene’s Theorem 201–6 Hankel matrix 177 Hopf–Stiefel function 745
Greene–Kleitman Theorem Happy End Theorem 446 Hungarian Algorithm
548–52, 589 Harary graph 290–1, 313 280–3, 286
grid 48, 61, 79, 166, 173, harmonic number 91 hypercube 214, see “cube”
261, 363, 394, 440, 464, Hasse diagram 542 hypergeometric series 89–90
467, 474, 511, 665, 764, Hat-Check Problem 54 hypergraph 9, 434–40, 442,
771, 783–4, 789, 793–4, Havel–Hakimi Theorem 221 451–2, 468–9, 475, 491–3,
797, 800–5, 819 head (of coin) 7, 31, 46, 59, 496–8, 572–3, 601, 608,
grid graph 394 66, 86, 95, 97, 103–4, 111, 613–6, 638–41, 659–60,
Grinberg graph 400, 418 135, 668–9, 703, 707, 722 666–7, 673–4, 684, 686,
Grinberg’s Theorem 401, 418 head (of edge) 8, 166, 224–5, 722, 735, 803, 820
group 98, 153, 179–90, 193, 361, 370, 389, 522, 748, hyperplane 48, 523, 529,
212, 427, 441, 459, 606, 785, 788 539, 656, 735–6, 746, 805,
623, 734 head partition matroid 522 812–6, 820, 824–6
Heaviside method 75, 138–9
Heawood graph 219, 274,
Hadamard matrix 618–24 icosahedron 351, 385, 387,
424, 638
Hadwiger Conjecture 349, 416, 420, 797
height (in lattice path) 26,
710 ideal map 593–4, 607
45, 48, 151, 177, 226,
Hadwiger–Nelson Problem ideal of sets 501–3, 510, 512,
height (in poset) 544, 551–2,
729–30 514, 519, 522, 533–4, 545
561–2, 571, 582, 588, 608,
Hajós Conjecture 349, idempotence property 528,
762, 790, 822–7, 831
709–10 539
height (in tower) 452–3, 484,
Hajós construction 355 idempotent quasigroup
488, 602–3, 673
Hales–Jewett Theorem 463, 642–4, 655
hereditary class 366–8, 394,
467, 474 image 2
415
half-idempotent 643–4 in-neighborhood 224
hereditary system 514–40
in-tree 749–52, 785, 788
Subject Index 959

incidence algebra 757–60 independent positions (in intersection graph 337,

incidence bigraph 628, 638 grid) 160, 257, 272 342–3, 367–8, 374, 583,
incidence function 757–62, independent set (graph) 174, 608, 705
765 210–1, 228–9, 259–64, interval (in poset) 590, 608,
incidence graph 253, 517, 322, 327–9, 335–6, 340–4, 722, 757–62
531, 628–32 363–4, 368, 375–6, 419, interval (of numbers) 173,
incidence matrix (design) 9, 435, 449–51, 492, 544, 547, 565, 604, 608, 653, 707,
614–9, 624–5, 628, 645–6 662–4, 671, 678, 685, 701, 722, 804, 810, 816, 832
incidence matrix ([di]graph) 704, 717–8, 729–30, 812 interval chromatic number
211, 216, 224, 388, 537, independent set (matroid) 491
748–9, 752–3, 763–4 514–40 interval coloring 373
incidence matrix Independent Bounded interval graph 337, 366–7,
(hypergraph) 9, 438–9 Differences Inequality 374, 586, 608, 705
incidence matrix (relation) (IBDI) 719–22 interval number 608, 705,
3, 9 indexed by (in OGF/EGF) 832
incidence poset 790 93–8, 103–4, 113–4, interval order 585–7, 594,
incidence relation 3–4, 211, 118–21, 125–6, 132, 135, 604–5, 832
378, 614, 624–6, 629, 631 137, 142, 184, 187–8, 553 interval representation 366,
incidence vector 3, 15, 21, indicator variable 662, 374, 585–7, 604
493, 545, 556, 724, 670–1, 689–96, 707, 711, inversion (in permutation)
727–31, 744–5, 752–3, 718 98, 103–4, 109, 163, 167,
763, 773 induced circuit property 520, 208, 432, 567
incident 4–5, 211 538 inversion formulas 102, 124,
inclusion-exclusion 122, Infinite Ramsey Theorem 127–33, 136, 157, 163, 747,
153–64, 171–6, 477, 695, 468, 471 757, 760–1
757, 760–1 information-theoretic lower Inversion Principle 111, 155
incomparability graph 586 bound 822, 831 involution 126, 133, 163–8,
incomparable 541, 544, inhomogeneous term 51, 176, 178, 193–4, 206, 629
568–75, 583–7, 593, 70–5, 84, 173 irreducible 573, 582, 584,
603–4, 608, 758, 823, 831 injective function 2 592–4, 604–7
incorporation property 528, injective coloring 424 irredundant representation
532, 539 integer simplex 342, 371 593–4
increasing path/trail 431, integer triangle 142, 149 irreflexive 3–4, 224
472, 601 Integrality Condition 774–5 isometric subgraph 764
increasing tree 106 Interlacing Theorem 769, isomorphism 4, 178, 212,
indegree 104, 224, 370, 678, 782 216–8, 225, 526, 544–5,
686, 749, 765 internal vertex 5, 229, 284, 589, 594, 607, 699, 705
independence number 259, 298, 301, 305, 311–2, 324, isomorphism class 178,
260, 334, 336, 356, 368, 348–9, 383, 396, 403, 411, 183–4, 187–9, 212, 218,
458, 663, 685, 739, 773 422, 742, 785–9 248, 250, 549, 605, 687
independent edges 253 intersecting family 497–502,
independent 509–12, 608, 727–32, 817
J(P) 589, 592, 602–3 605,
events/variables 7, 65–6, L-intersecting family
607, 762
111, 116, 190, 512, 584, 727–32, 745, 817
Jensen’s Inequality 505–8
595, 658–722 t-intersecting family
Jeu de Taquin 201–3, 208
independent paths 298–303, 497–500, 511, 513
join (in posets) 588–600,
313–5
605–8
960 Subject Index

join (of graphs) 213, 336 Lagrange’s Theorem 616–7 line graph 300, 314, 316,
join-irreducible 592–3, Laplacian eigenvalues / 326, 329, 357, 362–3, 368,
604–7 matrix / spectrum 776–82 371, 608, 678
joint random variable 505 Las Vergnas’ Theorem 333 linear arboricity 412, 677
Jordan Curve Theorem Latin square 9, 609–12, t-linear crossing number
378–83, 400, 402 622–4, 627–8, 642–3, 648, 795–6
651–2, 656 linear extension 568–84,
lattice (poset) 529, 571, 595, 599–602, 604, 608,
König’s Other Theorem
588–608, 757, 761–2, 766, 760, 822–7, 831
259–60, 263, 368–9, 540
832 linear forest 412, 422, 677–8
König–Egerváry Theorem
lattice ball 29–30, 33, 57 linear hypergraph 438
258–9, 262–3, 273, 278,
lattice path 24–8, 30, 39–42, linear independence 68–9,
280, 282, 287, 302, 313,
49, 55, 61, 64, 104, 115, 76, 80, 176, 516, 519, 538,
368–9, 530–1, 547
165–8, 170, 172, 176 614, 723–32, 744, 753,
Kakeya set 746
lattice point 25, 57, 169, 176, 767–8, 818
Kempe chain 404, 406, 421
221, 389, 426, 440, 464, linear matroid 516, 519, 528,
kernel/kernel-perfect 363–4
474 532
king (in tournament) 225,
lattice polygon 389 linear recurrence 51–2,
228, 703
lattice walk 25, 49, 64, 117 67–9, 74, 79, 84
Kirkman Schoolgirls
layered block structure linearity of expectation 7–8,
Problem 649
464–5 104, 162, 662–72, 689,
Kirkman triple system 650
leaf 5–6, 240–2, etc. 694, 707, 713, 722, 793,
Kleitman’s Inequality 595,
leaf block 295, 297, 348 821–2
597, 608
 -leaf-connected 327 X , Y -link 299–300, 395
Kneser Conjecture 342, 798,
left subtree 43–4, 47, 135,  -linked 314
811–2
177 Lipschitz condition 719
Kneser graph 251, 342, 781,
length 2, 17, 25, 41, 209, etc. n-ary list 2, 26, 667
798, 811–2, 820
level (in poset) 553, 556, 0 , 1-list 2, 40, 51, 173
knight’s tour 330
561–2, 567 1 , 2-list 53–4, 59, 104, 112,
Knuth-equivalent 208
lexical 472–3 664–5, 740
Kotzig’s Theorem 371, 411
lexicographic order 192, list assignment 346–8,
Kruskal’s Algorithm 245,
195–6, 198, 201, 493, 512, 355–6, 362–3, 403, 421,
252
556, 566, 789 661, 677, 686, 742
Kruskal–Katona Theorem
lexicographic product 459, list chromatic number
495–6, 500, 510
583 346–8, 362, 685, 738
Kuratowski subgraph 391–3
lg 87, 447, 505–9, 580, etc. List Color Conjecture 362–3,
Kuratowski’s Theorem
light edge 407–13, 416 373, 412
390–3, 397–9, 526–7, 753
line (geometric) 30, 48, 52–3, list  -colorable 346
59, 72–3, 79, 199–201, list edge-chromatic number
Λ n 606 342–3, 388, 427, 565, 613, 362–5
L(m , n) 557, 562, 567, 589, 794, 798, 813, 815, 820 S-lobe 349–50, 353, 367, 375,
605 line (in projective plane) 9, 391, 399
labeled structure 119–20, 624–30, 635, 638, 649–50, Local Cut Lemma 680
125, 128, 132, 188 653 local density bound 800
-labeling 656 line digraph 300 Local Lemma 674–86, 719,
Lagrange Inversion Formula 2-line form 99 722
127–31, 136 local Ramsey number 473
Subject Index 961

locally finite 590, 757 Matroid Partition 537 minimal element (in poset)
locally linear graph 684 Matroid Union 533–5, 543–7, 551, 565, 571,
log-concave 563–4, 567–8, 539–40 576–7, 582–4, 588–93,
827 max-cut problem 710 606, 760, 790
log-supermodular 596–600 max-flow min-cut 303 minimal imperfect graph
lollipop graph 331 maximal 4 369–70
Lovász Local Lemma 674, maximal antichain 544, 551 minimal LYM order 567
679 maximal chain 544, 547, minimal nonplanar graph
lower extension 574, 605 550, 552–3, 558–61, 565, 390–1
Lu’s Theorem 322 759 minimal vertex separator
Lucas numbers 53, 61, 65 maximal element (in poset) 375
LYM inequality / order / 512, 543–6, 551–2, 557, minimal separating set 304,
property 558–65, 567–8 571, 576–7, 582–3, 586, 322, 367, 375
Lyusternik–Shnirel’man 588–93, 606, 790 minimax spanning tree 252
Theorem 811–4, 817, 820 maximal outerplanar graph minimum polynomial (of
382, 387–9 matrix) 768, 772
maximal planar graph minimum weighted cover
Mader’s Theorem 311, 316
384–7, 400, 419 279–82
(n , ¾ , c)-magnifier 779–80
maximally connected 289, minor 525, 527, 539, 710
Markov’s Inequality 671,
291 Minty’s Theorem 338–9
689–92, 705–9, 713,
maximum antichain 497, Mityagin’s Theorem 826–7
717–8, 821
544, 546–51, 553, 558, mixed difference system 655
Marriage Theorem 256, 262,
568, 583, 606 Möbius function 759–62, 765
268, 359, 678
maximum average degree Möbius Inversion Formula
martingale 706, 712–9, 722
408–9, 414, 417, 421–4 163, 757, 760–1
matching 253–88, 312–4,
maximum density 693, 703 Möbius ladder 235
329, 356–8, 364–6, 434,
maximum weighted p-modular L-intersecting
442, 460, 492, 510, 517–8,
matching 279–83, 287–8 family 728–30
522, 529–31, 547, 559–63,
meet (in poset) 588–600, modular lattice 594–5, 607
566–7, 645–9, 668, 702–5,
605–8, Moessner’s Process 62, 105
726, 742–3, 754–6, 765
meet-irreducible 592, 606–7 moment (of variable) 691,
º -matching 287
membership function 17 694, 696, 704–9
matching number 253
Menger’s Theorem 298–303, moment curve 813, 815–6,
0 , 1-matrix 2–3, 160, 211,
313, 546, 560 820
224, 432–3, 442, 542, 570,
Method of Deferred moment generating function
604, 620, 781
Decisions 702 704, 707, 709
Matrix Arborescence
Method of Moments 696 monotone 430, 458, 821
Theorem 750
Meyniel graph 375 monotone (Boolean)
Matrix Tree Theorem
min-max relation 259, functions 557, 596,
747–52, 762–3, 777
264–5, 273, 298–9, 303, 599–600
matroid 9, 251, 513–40, 595,
357, 366, 368, 513, monotone list/sequence 206,
656, 754
529–33, 538, 546–7 430–1, 442, 452–3, 703,
Matroid Covering/Packing
minimal 4 821–2, 831
536
minimal ¾ -connected graph monotone path 601–3, 608
Matroid Intersection
310–1, 315–6 monotone property 691, 694,
529–34, 537, 540
minimal ¾ -edge-connected 703
Matroid Matching 532
multigraph 309–10 monotone tournament 458
Matroid Parity 532
962 Subject Index

Moore bound 244 nondeterministic 10 orbit 3, 8, 54, 180, 187, 635,

Moore graph 244, 252, 775 nontrivial graph 232, 297, 639, 645
Moore–MacNeish Theorem 329, 345, 349, 371 order (of group element) 441
611, 648, 653–4 ¾ -norm 548–9 order (of graph) 4, 209
multidigraph 8, 749 norm graph 633–4 order (of Latin square) 609
multiedge 233–4, 300, 304, normal bipartite poset order (of projective plane)
308, 357–9, 364, 379, 381, 575–6 626
384, 516, 697, 747, 785 normal plane map 411–2 order (of recurrence) 51
multigraph 8, 233–5, etc. normalized Hadamard order ideal 545
multilinear polynomial 727 matrix 619–20 order polynomial 762
multilinear reduction 727, normalized matching order polytope 824–7
731–2 property 559–63, 567 order relation 8, 541–4, 557,
multinomial coefficient 37–9, NP-complete 11, 316, 336, 568, 570, 577, 582, 585,
48, 147, 637, 690 344, 359, 435, 532, 537, 591, 600, 604, 606, 757,
Multinomial Theorem 38 543, 573, 700 789, 828
multiplier (of difference set) NP-hard 11, 699, 710, 795, order-preserving 469, 472,
634–9 800 491, 528–9, 557, 565,
multiset 2, 19, 27, 30, 94–7, nullspace 52, 388, 748, 753, 596–601, 762, 824, 827
104, 114, 120–1, 156, etc. 766, 773 order-reversing 596–9
Mutual Independence s , t-numbering 315 ordered graph 491
Principle 676–8, 681–2 ordered hypergraph 601–4,
mutually independent 608
Oberwolfach Problem 647,
675–83, 686 ordered partition 121, 133
656
Mycielski’s Construction ordered Ramsey theory 457
octahedral subdivision
340, 345, 354, 434 ordered tree 43–5, 48–50,
806–9
63–4, 136, 206
octahedron 385, 789
¾ -gon order 828
Narayana number 45, 49, 64 odd 163, 229
ordinary enumerator 118
Nash-Williams’ Orientation odd component 264–7, 272–3
Ordinary Exponential
Theorem 309–10 odd cycle 209–12, 230–2,
Formula 188
near-triangulation 395, 402 237–8, 264, 278, 335–7,
ordinary generating
necklace 22, 178, 181–8, 342–4, 348, 355, 370, 479,
function 93
313, 809–10, 816, 819 522, 764, 768
Ore’s Defect Formula 258,
negative binomial 27, 74, Odd graph 219, 782
262, 265, 531, 540
108 odd representation 745
Ore’s Condition 319–20,
neighborhood 210, 254 odd walk 230–1, 550, 768
326–7, 330, 332, 705
Neighborhood Local Lemma odd-crossing number 795
Ore’s Lemma 319, 332
681–3, 686 OGF 93–6, 110–20, 123–4,
Ore’s Theorem 319–20,
net (a graph) 297 128–9, 136–7, 141–3,
323–4, 327, 332, 705
nonadjacency relation 210 148–9, 161, 182–9
orientation 225, 228, 237–9,
noncrossing circuit 388, 419 on-line Ramsey theory 457
255–6, 261, 306–10,
noncrossing pairing 43, 48, One-Point Removal Theorem
338–9, 344, 363–4, 374,
63, 82 574–5
389, 404, 428, 542, 550,
noncrossing partition 43, 49, one-to-one correspondence 2,
570, 678, 738–43, 746–9,
64, 15, 206, 545, 605, 813
755–6, 762–5, 785–7, 832
noncrossing tree 128–9 open set 378, 811–4, 820
Orientation Theorem 309–10
nondecreasing sequence 201, ¾ -optimal partition 549
204–5, 595
Subject Index 963

oriented graph 225–6, 237, parity subgraph 237, 249 path (in digraph) 166–70,
752, 756, 765 parking function 50 225, 229, 339, etc.
orthant 806–8 2-part Sperner property 565 path (in graph) 4–6, 209–12,
orthogonal array 651–3, 656 partial ¾ -coloring 549 232–8, 277–84, 323–34
orthogonal chain partitions partial partition 134 path (in integer lattice)
566 partial tableau 201–3, 208 24–30, 33–4, 39–45,
orthogonal vectors 618, 621, partially ordered set 48–50, 55, 59, 64, 104,
744, 771, 773, 777–8, 541–608, 821–32 115–7, 151, 172, 199, etc.
817–8 particular solution 70–5 u , v-path 4–6, 229–31,
orthogonal Latin squares ¾ -partite graph 211, 223–4, 239–42, 246–7, 298–303,
609–12, 622–4, 627–8, 226, 228, 331, 335, 355–6, 314–6, etc.
638, 651–6 454, 456, 475–9, 489, 491, X , Y -path 299–300, 313
out-neighborhood 224 513, 655, 663, 770, 773, path matrix 166–8
out-tree 247, 749, 752 781 path system 165–70
outcome 6–7 N-partite hypergraph 435–7 pattern (in permutation)
outdegree 224–5, 228–9, partition (into chains) 206, 431–4, 442
235, 238, 255, 261, 363–5, 546–57, 560, 566, 574 pattern inventory 181–8
370, 404, 678, 686, 738–51 partition (of edge set) 232–5 pattern Ramsey number
outerplanar graph 382–3, partition (of integer) 137–52, 472–4
387–9, 398–9, 419 175–6, 185–90, 202, 207, paw 330, 489
outstanding line 199–201 557, 606, 647–8, 762 peak 45, 48, 151, 177, 452–3
overfull subgraph 359, 371 partition (of set) 3, 14, 49, pebbling move 668
56, 64, 121–2, 125–7, Pell sequence 62, 81, 92
133–4, 157, 175, 251, 343, Pentagonal Number
packing (of two graphs) 248
354–5, 419, 443–4, 461–5, Theorem 151
pair group 183–4, 186
468–72, 480–8, 492, 501, perfect elimination ordering
[¾]-pair-covering 624
568, 649, 815, 817–8 366
pair-crossing number 794–5
I , F-partition 422 perfect graph 366–71, 374–6
pairing model 697, 704–5
,  -partition 480–3, 487–8 Perfect Graph Theorem
pairwise balanced design
523, 530–8, 368–70, 547
650–3, 655
partition lattice 567, 588, perfect matching 253, 256,
Paley matrix 621–2
590, 605, 607, 761 260–88, 314, 329–30, 333,
pancyclic 326, 330
partition matroid 522, 356, 358, 365, 492, 510,
parallel classes 628, 649–50,
529–37 560, 567, 645, 702, 705,
654–5
parts (of bi/  -partite graph) 754–6, 765
parallel elements 516–7
135, 211–2, 221, 223–4, perfectly orderable graph
parent 43, 443
256, 269, 280, 295, 364, 375
parity (applications) 231,
456–7, 475–6, 491, 517, permanent 510, 754–6,
330, 379, 382, 400, 426,
531, 607, 655, 770 764–5
465, 755–6, 797–9, 814
parts (of composition) 19–20, Permanent-Determinant
parity (of circulation)
23, 30, 45, 53, 60, 81, 137 Method 755–6
738–43
parts (of integer partition) permutation 2, 22, 54, 57,
parity (of closed walk) 231
137–52, 175–6, 186–7, 62–6, 98–106, 109, 112,
parity (of dissection) 165
192, 206–7, 606, 647 123–6, 135, 158–86,
parity (of permutation)
Pascal’s Formula 25, 27, 31, 192–200, 205–8, 261,
163–4
39, 56–7, 78, 98, 450 431–4, 442, 576–80,
parity edge-coloring 373
Pascal’s Triangle 24–5
Parity Lemma 265–7, 276
964 Subject Index

666–8, 754–5, 764, 821–2, polynomial-time 10–1, 298, product representation 726
etc. 360, 550, 606, 687 projective plane 9, 244–5,
permutation graph 343 polyunsaturated 552 479, 612, 616, 624–35,
permutation matrix 160, poset 541–608, 757–62, 638–9, 645, 649–50, 653
196, 256–7, 432–3, 755 765–6, 790, 820, 822–32 proper coloring (graph)
permutation statistics positional game 261, 474 158–9, 174, 335–48,
98–102 positive lattice walk 49, 64, 399–424, 581, 661, 677,
Petersen graph 213–4, 117 684–6, 739–40, 812–4, etc.
218–20, 226–7, 236, 238, positive semidefinite 134 proper coloring (hypergraph)
244, 272, 318, 326, positively correlated 435–40, 659–60
329–30, 335, 353, 358–9, 595–601, 608 proper dissection 665–6
373, 397, 416, 490, 773, Pouzet’s Lemma 804–5, 808 proper edge-coloring
775, 782, 797, 814, etc. Prüfer code 46 357–65, 371–3, 438–40
Petersen’s 2-Factor Theorem Prague dimension 726 proper labeling 799, 802–5
268, 274, 372 predecessor 54, 225, 324, 15-puzzle 202
Pfaffian 756, 764 605 Pym’s Theorem 299–300,
PGT 368–70, 375–6 Prim’s Algorithm 252 395, 546
Pick’s Theorem 389 prime factors 21, 155, 172,
PIE 155–61, 164, 172–4 428, 553, 592–3, 626
quasigroup 642–4, 655
Pigeonhole Principle 16, principal submatrix 134, 767
query 447, 711–2
302, 339, 425–34, 440–6, Principle of Counting Two
quota 426, 444–5, 448, 451,
449, 451, 456, 466, 468, Ways 14
468
504, 513, 577, 602, 651, prism 326, 334
659, 800, 803, 825 prism-Hamiltonian 326
pigeonhole property 7, probability 6–8, 16, 20–4, radius 29, 243, 248, 250,
662–3, 670–1 39–40, 46, 54, 65–6, 83, 352, 794, 825, 828, 830
planar graph 316, 351, 86, 99, 103–6, 111, 116, Rado Selection Principle 468
377–424, 523–7, 537, 739, 135, 171–4, 189–92, Rado’s Theorem 462–3, 466,
742–3, 746–7, 753–6, 781, 505–7, 512–3, 578, 584, 473–4
783–4, 789–91, 795–7 595–6, 657–722, 793, rainbow 470–4
1-planar graph 797 821–4 Ramsey family (of patterns)
Planar Separator Theorem probability generating 472
394–7, 399 function -Ramsey number 455–7
planarization 199–201 product (of formal power Ramsey number 444–5,
plane multigraph 378–87, series) 94–7 448–62, 472–4, 483–4,
401, 411 product (of graphs) 214–6, 504, 601–2, 620, 624, 663,
Poisson distribution 695–7, 244–5, 273, 292, 319, 326, 673, 675, 682–5, 703, 728
704 329, 334, 359, 372, 376, Ramsey Theorems 468–74
Poisson Paradigm 696 394, 459, 773, 780–1, 794 Ramsey’s Theorem 443–9,
polarity 629–32 product (of posets) 545, 453, 461–4, 468–9, 474–5,
Pólya’s Theorem 184–5, 188 553–5, 559, 562–8, 583, 483–4, 601, 604, 658
Pólya’s Urn 106 588, 590–2, 595–6, 600–1, Ramsey–Turán problem 492
polychromatic 470 607–8, 760–1 random graph 338, 671–722
polygonal curves 378, 385 product dimension 624, 726 random variable 7–8, 50,
polynomial method 732 Product Formula (Möbius 111, 505–6, 512–3, 561,
Polynomial Principle 16, 18, functions) 760–1 608, 662–72, 687–722, 821
22, 26, 34, 100–2, 122, 761 Product Principle 14–8, 94 range space 711–2
Subject Index 965

rank (in matrix) 369, 388, regular polyhedron 188, 385 247, 347, 399, 443, 749–51,
614, 748, 753, 767, 781 -regular pair/partition 785–8
rank (in matroid) 514–40, 480–8, 492 Roth’s Theorem 484–6
656 regular system (of row-canonical permutation
rank (in poset) 552–68, 576, equations) 462 208
583, 585, 588, 594–5, Regularity Lemma 461, RSK Correspondence
603–7, 761–2 478–89, 492, 673, 794 193–201, 206
rank generating function relation 2–3 Rubin’s Block Theorem 348
552–3, 557, 564, 567, 761 removable pair 575, 583 run (in list/subset) 21, 45,
Rank–Nullity Theorem 388, representable matroid 516 49, 59, 174, 669
753 residue 130–1, 140, 622 run (in permutation) 101–2,
rank-symmetric / unimodal resolution class 649 105–6, 135, 147, 158, 432,
553–4, 557, 561, 564, 567 resolvable cycle systems 647 442, 667
ranking 133, 285–8, 585–6, resolvable design 648–50
604, 721 retreat (in permutation) 197
sample point 7–8, 662, 716
rational function 74–5, 88–9 reverse canonical
SATISFIABILITY (SAT) 11,
realizer 569–84, 820 representation 105
667
rectilinear crossing number reverse plane partition 189
 -saturated partition 548,
795 rhombicosidodecahedron
551–2
recurrence relation 51–92, 388
F-saturated graph 491–2
101–4, 115, 127–9, 132–6, right subtree 43–4, 47, 135,
saturates (in matching) 253
146–9, 161–3, 173–6, 177
Schensted’s Algorithm 195
189–91, 207–8, 399, 433, right-left sequence 197
Schensted’s Theorem 198,
452, 462, 747, 780 rigid circuit graph 366
201, 430, 432
redressé 203 ring (in configurations)
Schnyder labeling 784–90,
reduced adjacency matrix 404–6, 421, 520
797
256, 631 Ringel–Kotzig Conjecture
Schröder n-path 59
reduced graph 480–4, 492 646
Schröder number 64
reducible configuration rising antichain 511
Schröder–Bernstein
404–17, 421–3 rising factorial 17, 23,
Theorem 24
reduction 11, 203 122–3, 147
Schur’s Theorem 462
reduction (of tableau) 202–6 Robbins’ Theorem 306–7
Scott–Suppes Theorem 587
refinement (of matroid) 538 Robinson–Schensted
 -scrambling 578–80, 584
refinement (of partition) Correspondence (see RSK)
SDR 253–4, 302, 517
486–8, 588, 606, 784, 195
secant-and-tangent numbers
808–9 Rogers–Ramanujan
65
reflection principle 40 partition 151
Second Moment Method
reflexive 3–4, 8, 180, 541–2 rook polynomial 160–1, 163,
692–3, 698, 700, 703,
region 6, 24, 48, 52–3, 59, 171–5
705–6, 712
72–3, 79, 154, 165, 220–1, root (characteristic) 67–76,
self-complementary graph
378–401, 418, 566, 784–9, 79, 81, 84, 766–7
218, 232, 370, 388, 477,
797, 799, 802–4, 824, 832 rooted forest 128, 136, 278,
490
regular covering 559–68 750–1
self-conjugate partition 143
regular digraph 678, 686 rooted graph 43, 136
self-dual poset 543–5, 549,
regular graph 9, 210 rooted labeled structure 128
553, 557, 560, 567, 592
regular polygon 22, 169, 179, rooted tree 43–4, 47, 106,
semiantichain 565
385 128, 135, 177, 189, 206,
966 Subject Index

semimodular lattice 595, simple hypergraph 438–9 spectrum 766, 770, 772,
607 simple word 17–8, 120, 146 776–8, 781–2
semiorder 585–7, 594, 605, simplicial complex 766 Sperner property 553–4,
823, 831 simplicial elimination 557–61, 567–8, 606
separating set 289–92, ordering 367, 373 Sperner’s Lemma 799–809,
295–304, 310, 318, 322, simplicial subdivision 799, 819
367, 375, 387, 391–2, 802–6, 819 Sperner’s Theorem 497, 511,
396–7 simplicial vertex 65, 367, 374 564, 666, 668
x , y-separating set 298, 303 sink 224, 561, 787 split (of necklace) 809–10
(m , )-separation 394, 396, Six Color Conjecture 417 split (of poset) 547
399 size 2 split graph 376
x , y-separator 375 size Ramsey number 455–6 squashed cube embedding
( , )-separator 394 skeleton 199–200 781
sequence 2 skew chain order 565 stable matching 284–8
 -sequence (in tableaux) skew tableau 202–5, 207–8 stable set 210, 368–9, 375–6,
201–5 skew-overlap graph 399 536, 710
 -set 17 skew-symmetric matrix 756, stack-sortable 43, 63
set system 493 764 stage of evolution 698–9
shade 493, 497–8, 510, skipless chain 553–6, 562, staircase tableau 152, 208
559–60 565 standard deviation 691–2,
shadow 200, 493–7, 500, Skolem Construction 644 707
510, 559, 562, 567 Small Pot Lemma 355 standard example 570, 573,
t-shadow 493, 500 smallest-last ordering 338 576, 584, 605, 828, 832
Shannon switching game Snake Oil method 112–3, Stanley–Wilf Conjecture
539 117, 176 432–4
shape (of tableau) 189–96, snark 372 star (graph) 212, 227, 233,
202–8 Snevily’s Theorem 730, 745 243, 248–9, 251, 262,
Shapiro n-paths 64  -soluble 276 351–2, 368, 510, 832
Shearer’s Lemma 508–10, source 104, 224, 228, 548, t-star (of sets) 497–503, 511
513 561, 741 star  -coloring 422
SHIFT (in tableaux) 202–5, span function 527–9, 537, star property 501, 512
208 539 star-cutset 375
shift graph 581, 602, 604 spanning set 523, 525, 527, Star-Cutset Lemma 375–6
shift operator 495–6, 498–9, 539–40 Steinberg’s 3-Color
512 spanning subgraph 5, 159, Conjecture 414, 421
shifted tableau 192, 207 210, 239, 249, 263–4, 268, Steiner system 640
Shortcut Lemma 308–10 326, 334, 370, 384, 396, Steiner triple system (STS)
u , v-shortcut 308–10 420, 516, 518, 537–8, 566, 641–5, 655
shuffle 206, 442 605, 657, 744, 770, 780 Steinitz exchange 528–9
Sicherman dice 103 spanning tree 5, 46, 58, 65, Stepping-Up Lemma 452–3
signed graph 237, 297 239–52, 266, 297, 327–8, Stirling Inversion 132
signed involution 163–8, 334, 342–3, 347, 374, 399, Stirling number 121–2,
176, 178 428, 474, 515, 523–4, 535, 132–3, 146, 174–7, 567,
signing 297, 755–6 540, 747–53, 763, 777–8 761
signless Stirling number 122 special subdivision 806–9 Stirling permutation 134
simple graph 234, 359, 362, Spectral Theorem 768, 773,
527, 697, 794 778–9
Subject Index 967

Stirling’s Formula 85–6, subdivision (simplicial) 799, Symmetric Local Lemma

91–2, 139, 507, 580, 658–9, 802–9 674–81, 685–6
700, 730, 745, 818, 821 F-subdivision 349–51, 353, -system 503, 728
Strehl’s Identity 31 356, 390–3, 710 system of distinct
strength (of theorem) 705 ¾ -subdivision 684 representatives (SDR) 253
strict chain 759, 762 subgraph 4–5 Szemerédi’s Theorem 484
strict digraph 334 subhypergraph 468, 601
strict order polynomial 762 sublattice 589–94, 597, 600,
tabular representation 99
strict order relation 542 605–6
tail (of coin) 7, 31, 46, 66, 86,
strict Sperner property 568 submodularity 308–9, 520–1,
95, 103, 703
strictly compatible 762 530, 533–40, 595, 607
tail (of distribution) 706–22
strong absorption property subposet 543
tail (of edge) 8, 36, 166,
520–1, 536, 538 subspace 68–9, 76, 520, 528,
224–5, 255–6, 361, 370,
strong component 229–30, 568, 606, 725, 744, 764,
389, 522, 738, 742, 748,
236 766, 814
751–2, 785, 788
strong dependence property Substitution Lemma 376
tail partition matroid 522
528 substitution method 82–4,
Tait coloring 400
strong digraph 229, 237, 91
Tait’s Theorem 400, 417
298, 306–7 successor 54, 225, 363–5,
Talagrand’s Inequality
strong edge-coloring 373 410, 605, 686, 703
720–1
strong elimination property suitable set (for poset)
tangency graph 394
528, 539 576–80
tensor product 273
strong Hall property 567 sum (of posets) 545
term (hypergeometric) 89–90
strong orientation 306–7 Sum Principle 14–5, 18
term ratio 89
Strong Perfect Graph sum-free set 669
ternary list 2, 23, 121
Theorem 370 Summation Identity 26, 33,
tetrahedron 187–8, 385, 798
strong Sperner property 66, 122
(6,3)-Theorem 492
553–4, 559, 561, 567, 606 sunflower (set system)
thread 312, 411, 422–3
strongly connected digraph 503–4, 510, 512–3, 729
threshold edge function 691
229, 237, 298, 306, 389, superposition principle 72
threshold graph 376
669, 703 superregular 782
threshold probability
strongly connected vertices surjective function 2
function 691–2
230 2-switch 222, 227, 230, 297,
tiling 34, 54, 66, 81, 169–70,
strongly perfect 376 371
427, 665
strongly regular 649, 773–6, P-symbol/Q-symbol
topological ordering 569
781–2 195–203, 206, 208
total coloring 373, 642, 655,
subadditivity (of entropy) symmetric chain (posets)
681, 739
507–10 553–7, 561–2, 565–7, 603,
Total Coloring Conjecture
Subcommittee Identity 26, 606
373, 655
110, 695 symmetric design 615–8,
total dominating set 639
subconstituents 782 624, 626, 634, 646, 773
total graph 739
subcube 219, 329, 584 symmetric difference 277,
totally unimodular 763
subdivision (graph) 305, 286, 512, 518, 540, 730,
Touchard polynomial 134
311, 349–53, 356, 390–1, 753, 755, 763, 817–8
Touchard’s Formula 82
398, 416, 527, 709–10, symmetric group 98, 186–9,
toughness 318, 327, 329–30
739, 753–4, 798 606, 734
968 Subject Index

tournament 225–9, 261, 334, 434, 451, 479, 485, unavoidable set 404–7, 413,
342, 374, 441, 458, 666–7, 489–90, 663, 682, 685, 416, 423
686, 702–3, 721 744, 769, 782, 791, 812 underlying graph 225, 234,
tower function 452–3, 459, Triangle-Removal Lemma 238–9, 443, 540
465–6, 484, 488, 602–4, 485–6, 492 unforced pair 573, 576, 584,
673 triangular chord 550 606, 620
Tower of Hanoi 70, 80, 84 Triangular Criterion 725–7, Uniform Angle Lemma
trace 766–7, 770 730 787–8
track number 608 triangular grid 173, 801 uniform distribution 99, 505
trail 234–5, 239, 250, 287, triangular matrix 758, 760 uniform family/hypergraph
313, 315, 329, 431, 442, triangular number 144, 151 9, 435–7, 442, 452, 475,
752 triangulated graph 366 489–500, 503–4, 510, 568,
transitive closure 543, 572 triangulation 43–4, 48, 56, 601, 608, 613, 616, 640–1,
transitive graph 212, 214, 384, 387–8, 395–6, 659–62, 666–7, 673–4,
218–9, 386, 812, 819 402–12, 416–23, 784–9, 684, 686, 727–32, 787
transitive orientation 226, 799, 809 uniform list assignment
374, 542, 550, 570 tripartite 460, 667 355–6, 667, 686, 742
transitive relation 3–4, 8, triple system 613, 640–5, uniformly at random 7, 91,
180, 230, 517, 541–3, 549, 649–50 507, 509, 561, 665, 668–9,
transitive tournament 374, truncation (of matroid) 540 677–8, 681, 685–7, 697,
703 Tucker’s Combinatorial 704, 707, 721–2, 821
transitivity of dependence Lemma 806–9, 812–3, 819 uniform matroid 522, 533,
528, 539 n-tuple 2 537
translate 512, 634–5, 638–9, Turán graph 223, 228, uniform random graph
655 475–8, 489–90, 770 model 688, 691–2, 698
transversal (of hypergraph) Turán number 479, 672 uniformity property 520–1,
504, 638, 735–6 Turán problem 478, 493, 524–5, 536
transversal (of matrix) 496, 498, 630 union (of graphs) 4
279–81, 287–8, 518–9, Turán’s Theorem 223–4, Union Bound 658–60, 671,
523, 532–3, 537, 541, 623, 228, 475–6, 478, 480, 483, 674, 685, 711, 721
638, 657, 735–6 489–91, 663, 769–70, 781 union-closed family 503, 512
transversal (of Latin square) Tutte graph 401 Union-Closed Sets
623, 638, 657, 735–6 Tutte set 265–7, 272–5 Conjecture 503, 512
transversal design 656 Tutte’s Condition 265–7, 269 union-free family 551
transversal matroid 517–8, Tutte’s 1-Factor Theorem unipathic graph 228
522, 531–2, 536, 540 264–75, 314, 756 unit-distance graph 342,
Traveling Salesman Tutte’s º -Factor Theorem 729, 745
Problem (TSP) 316 270–1, 276 universal search list 669
tree 5–6, 35–8, 239–52, etc. Tutte’s 3-Edge-Coloring universal subset list 566
¾ -tree 65, 375 Conjecture 416 unlabeled graph 212
triangle 142, 212, etc. Two Squares Theorem 428, unstable pair 284
Triangle-Counting Lemma 627 up-down permutation 65
485 Two-Point Removal up-down sequence 197, 203,
triangle-free 223–4, 227, Conjecture 575 207
238, 296, 305, 313, 316, two-step method 670 up-set 544–6, 595, 608
326, 334, 339–43, 351–4, upper extension 574
375, 384, 414, 417, 420,
ultrafilter 466
Subject Index 969

Vajda’s Identity 61 vertex-transitive 212, 214, weight (of light subgraphs)

valley 45, 177–8, 453 218–9, 386, 812, 820 407–13, 417, 421–4
van der Waerden Conjecture Vizing’s Adjacency Lemma Weighted Neighborhood
765 362, 414, 423, 699 Local Lemma 686
van der Waerden number Vizing’s Planar Graph Weights Problem 24
465, 685 Conjecture 422 Well-Ordering Principle
van der Waerden’s Theorem Vizing’s Theorem 359–61, 469–70
463–7, 474, 484, 685 365, 370–2, 425, 438–9, wheel 305
Vandermonde Convolution 744 Whitney number 761
26, 31, 110 width 544, 546, 549, 552–3,
Vandermonde determinant 573–4, 607, 823
Wagner’s Theorem 391, 398,
69, 207, 746, 820 Wilson’s Theorem 441,
527
Vandermonde matrix 80 640–1
walk (in graphs) 230–1
Vandermonde’s Theorem 26, with high probability (whp)
2-walk 326
33, 92 687–705, 710, 717–8, 721,
u , v-walk 230–1, 235, 338–9,
0 , 1-variable 507, 662, 666, 821
767, 772, 776, 780
708 word 16–8, 28, 37–42, 48, 56,
Wallis product 86, 91
variance 104, 691–2, 706–7, 80, 118–22, 132, 145–6,
Warren’s Theorem 830, 832
721 505, 511, 513, 566,
weak absorption property
varieties (in designs) 613, 619–20, 667
520–1, 536
625 ¾ -word 16–8
weak barycentric
vector matroid 516, 520, word form (permutation) 2,
representation 789
537–8 18, 98–102, 105–6, 158,
weak chain 759, 762
Venn diagam 154, 530, 578, 160, 163–4, 168, 174, 197,
weak dual 382–3
591 201, 205, 208, 236, 431,
weak ear decomposition
vertex 4 510, 746
306–7
vertex charging 409–12, Worpitzky’s Identity 101–5
weak elimination property
416, 422–3 WZ method 87–90, 92
517–21, 536, 538–9
vertex cover 258–60, 263, WZ pair 88
weak excedance 106
273, 276–82, 287, 298,
weak order 567, 585
366, 368, 492, 531, 547,
weakly Ò-bounded 351–2 XYZ Inequality 595,
568, 770
weakly chordal 375 599–601, 608
vertex cut 289, 296, 318,
weakly compatible pair 762
349, 353, 391
weight (in binary lists) 260,
vertex deletion 221 Young diagram 141
494, 619
vertex multiplication 376 Young lattice 606
weight (in enumerator) 93,
vertex set 4 Young tableau 189–98,
162, 184–5, 188
vertex ¾ -split 291, 301, 305 206–8, 430
weight (in graphs) 234, 237,
n-vertex graph 13
245–8, 251–2, 273,
vertex-color-critical 352–3,
279–83, 286–8, 374, 376, Zarankiewicz’s Problem
820
399, 458, 477, 510, 631–3, 639
vertex-exposure process
519–22, 536, 538, 560–1 zero-sum Ramsey number
715–6
weight (in path systems) 461
165–9, 176–7 zeta-polynomial 762