LINEAR ALGEBRA METHODS

IN COMBINATORICS

László Babai and Péter Frankl

Version 2.1∗
March 2020

—————
∗
Slight update of Version 2, 1992.

——————————————————————— 1

c László Babai and Péter Frankl. 1988, 1992, 2020.
Preface

Due perhaps to a recognition of the wide applicability of their elementary concepts and
techniques, both combinatorics and linear algebra have gained increased representation in
college mathematics curricula in recent decades.
The combinatorial nature of the determinant expansion (and the related difficulty in
teaching it) may hint at the plausibility of some link between the two areas. A more profound
connection, the use of determinants in combinatorial enumeration goes back at least to the
work of Kirchhoff in the middle of the 19th century on counting spanning trees in an electrical
network.
It is much less known, however, that quite apart from the theory of determinants, the
elements of the theory of linear spaces has found striking applications to the theory of families
of finite sets. With a mere knowledge of the concept of linear independence, unexpected
connections can be made between algebra and combinatorics, thus greatly enhancing the
impact of each subject on the student’s perception of beauty and sense of coherence in
mathematics. If these adjectives seem inflated, the reader is kindly invited to open the first
chapter of the book, read the first page to the point where the first result is stated (“No
more than 32 clubs can be formed in Oddtown”), and try to prove it before reading on. (The
effect would, of course, be magnified if the title of this volume did not give away where to
look for clues.)
What we have said so far may suggest that the best place to present this material is a
mathematics enhancement program for motivated high school students. While we contend
that parts of the first four chapters could well support such a program, the techniques
presented also provide powerful research tools in combinatorics and related areas such as
combinatorial geometry and theoretical models of computation.
A striking example from geometry is the a disproof in the early 1990s of Borsuk’s then over
half-century old, much studied conjecture on decomposing n-dimensional solids of a given
diameter into pieces of smaller diameter. What Borsuk conjectured was that n + 1 pieces
always suffice. This conjecture was widely believed to be true; it was verified for various
classes of solids, including centrally symmetrical ones, and those with a smooth boundary.
The disproof by Kahn and Kalai (1992), to be presented as Theorem 5.23 in Chapter 5 was
stunning both for its force and for its simplicity. It did not just beat the conjectured bound

c László Babai and Péter Frankl. 1988, 1992, 2020.

i
by a trifle: it produced √
an infinite family where the minimum number of pieces grew as an
exponential function of n. Yet the proof took only a page to describe, with reference to a
combinatorial result which occupies a central place in this book.1
Rather than presenting as many results as possible, we have concentrated on developing
techniques and showing different methods to yield different proofs and a variety of gener-
alizations to a small set of focal results. The eclectic collection of exercises serves to add
both in depth and in breadth to the scope of the book. Many exercises are accompanied
with “Hints”; and full solutions are given in an appendix (“Answers to exercises”) to those
exercises marked with a diamond (♦). Asteriscs indicate the degree of difficulty.
Most results are motivated by applications. Discrete geometry is a prime target. Appli-
cations to the theory of computing are prominent in several sections (computational learning
theory (Section 7.4), communication complexity theory (Chap. 10.1)). But the theory of
computing plays a more subtle role in motivating many of the concepts, even though this
may often not be obvious. A brief survey at the end makes some of these connection explicit
(Section 10.2). One problem area of cardinal importance to the theory of computing is
the problem of finding explicit constructions for combinatorial and geometric objects whose
existence is known through probabilistic arguments. Such problems tend to be notoriously
hard; in the precious few successful attempts on record, methods of algebra and number
theory have been the winners. In this volume, an explicit Ramsey graph construction (Sec-
tions 4.2, 5.7) serves as simple illustration of the phenomenon. Some of the much more
complex examples known to be directly relevant to the theory of computing are mentioned
briefly, along with a number of open problems in this area (Section 10.2).
Having said this, naturally, the prime application area of the methods presented remains
combinatorics, especially the theory of extremal set systems. We have made an effort to
motivate each combinatorial application area and to give some idea about the alternative
(non-linear-algebra) approaches to the same area.
We have done our best to make all material accessible to undergraduates with some ex-
posure to linear algebra (determinants, matrix multiplication) and a degree of mathematical
maturity, the only prerequisites to starting on this book. Although the notion of fields and
their characteristic are used throughout, the reader will lose little by taking the term “field of
characteristic p” to be a synonym of the domain {0, 1, . . . , p − 1} with operations performed
modulo p where p is a prime number; and the term “field of characteristic zero” to mean the
domain Q of rational numbers or R, the real numbers.
Algebraic techniques not normally covered in standard courses (such as affine subspaces,
orthogonality in spaces over finite fields, the exterior algebra, subspaces in general position)
are introduced in full detail. An occasional review of the relevant chapters of a text on
abstract algebra or the elements of number theory might be helpful; the review sections of
Chapter 2 are specifically intended to guide such recollection.
In an effort to keep prerequisites to a minimum, the size of the volume manageable, and
1
Sections 5.4 and 5.6 together give a complete and self-contained proof of this surprising result.

ii ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
to minimize the overlap with existing texts and monographs, we have omitted major areas
that would fit the title of the book. The most painful omission is that of spectral (eigenvalue)
methods. However, excellent expositions of some of these methods are easily accessible for
the more advanced reader (see e. g., in increasing order of demand on the reader, Chapter 11
of Lovász (1979c), Biggs (1974), Godsil (1989), Brouwer, Cohen, Neumaier (1989), Bannai,
Ito (1984)).
The list of major relevant areas omitted or hardly touched upon in the present volume
includes

– spectral techniques in the study of highly regular structures (strongly regular graphs,
association schemes, designs, finite geometries);

– algebraic theory of error correcting codes;

– spectral techniques in the study of properties of graphs such as connectivity and ex-
pansion, diameter, independent sets, chromatic number;

– finite Markov chains;

– matroids: a combinatorial model of linear independence;

– linear programming and combinatorial optimization;

– lattices and the “geometry of numbers”;

– applications to the design of efficient algorithms;

– applications to lower bound, simulation, and randomization techniques in various mod-

els of computation and communication.

While this may look like too long a list to ignore, we feel that the modest material we do
cover is in areas in the most pressing need of exposition.
The intended audience of the text includes undergraduates, graduate students and re-
searchers working in discrete mathematics, discrete geometry, the theory of computing, ap-
plications of algebra, as well as open-minded mathematicians irrespective of their specific
area of interest. The text can be used as course material in several ways. First of all, it can
be the text for a one-semester graduate course, providing techniques for research level open
problems. Some other recipes for classroom use:

• Use Chapters 1–5 embedded in an introductory course on combinatorics.

• Use Chapters 5–7 embedded in an advanced course on combinatorics.

• Present parts of Chapters 1 and 4 as entertaining applications in an introductory course

on linear algebra.

——————————————————————— iii

c László Babai and Péter Frankl. 1988, 1992, 2020.
 • Use Chapter 6 to introduce and motivate wedge products.

• Select material from Chapters 3, 5, and 6 to show delightful applications of the elements
of abstract algebra.

• Present the full proof of the Kahn–Kalai Theorem (Sections 5.4–5.6) in just two classes
in a “Topics” course.

• Advise students to review their basic abstract and linear algebra along the outline
given in Chapter 2 and parts of Chapter 3. Note that, while these two chapters are in-
troductory in character, substantial results appear already in Chapter 3. These include
Gale’s Theorem on how to distribute points on a sphere evenly, and the resultant proof
by Bárány of Kneser’s Conjecture on the chromatic number of certain graphs. A puz-
zling computational problem, with Jacob Schwartz’s amusing Monte Carlo algorithm
to solve it, appears in Section 3.1.4.

The material can be used in conjunction with other texts on combinatorics. A combina-
tion with one of the following is especially recommended:

Ian Anderson, Combinatorics of finite sets, Oxford University Press, 1987;

Béla Bollobás, Combinatorics, Cambridge University Press 1986;

or in an advanced seminar with Chapters 8, 9, and 13 of

László Lovász, Combinatorial Problems and Exercises, North–Holland 1979.

Most of the combinations suggested above have been tried out in a variety of course
settings on audiences of widely varying backgrounds. The presentation of the material is
based on classroom experience gathered by the authors at Eötvös University, Budapest; The
University of Chicago; Tokai University, Hiratsuka; and the University of Tokyo. We are
indebted to these institutions for the opportunity to experiment with the material.
Our thanks are due to all the friends including a number of our past or current students
who read parts of the manuscript and helped improve it; we are most indebted to Bob
Beals, Collette Coullard, Katalin Friedl, Robert Freud, Albert Goodman, Barry Guiduli,
Péter Hajnal, Zsolt Hátsági, Penny Haxell, Anna Lubiw, Ákos Seress, Mario Szegedy, László
Székely, Norihide Tokushige, and Máté Wierdl for their generous support. In addition to
the institutions mentioned above, special thanks are due to the Institute für Ökonometrie
und Operations Research, Universität Bonn, where one of us spent a month in the quiet
atmosphere of an office inside a library while writing version number zero in 1986. We are
grateful to CNRS, Paris and to AT&T Bell Laboratories for making possible several extended
visits to the U.S. for one of us; such visits were helpful in facilitating communication between
the coauthors. Thanks are due to Richard Carnes of the University of Chicago for making his

iv ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
TEXnical skills available for the first stage of this project which resulted in a widely circulated
preliminary version produced in Summer 1988. The Department of Computer Science of the
University of Chicago has most generously supported the project and provided the technical
facilities.
Last but not least, we wish to express our deep gratitude to our teachers and mentors at
Eötvös University, Budapest, and the Mathematical Institute of the Hungarian Academy of
Sciences, many of whom have been among the foremost creators of combinatorial theory in
the past decades. While listing them all would be impossible, we should especially mention
Pál Erdős, everybody’s “Pali bácsi” [paw-lee but-chee] who took interest in us at our epsilon
age; and our mentors, Pál Turán, Vera T. Sós, András Hajnal, Gyula O. H. Katona, Lajos
Pósa, László Lovász. We feel fortunate to have grown up in the mathematical environment
they helped create.
We wish to thank the many readers of earlier versions of the manuscript for their helpful
comments. Several problems contributed by readers are now included among the exercises.
The first version was circulated from 1988, and significant update followed in 1992. The
process of a new round of updates started in March 2020.

Chicago, IL László Babai

Tokyo Péter Frankl
September 1992, March 2020

——————————————————————— v

c László Babai and Péter Frankl. 1988, 1992, 2020.
vi ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
Contents

Preface i

Notation and terminology xiii

1 Warm-up 1
1.1 Counting clubs in Oddtown . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Point sets in Rn with only two distances . . . . . . . . . . . . . . . . . . . . 7
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Two solutions to a jigsaw puzzle? . . . . . . . . . . . . . . . . . . . . . . . . 13
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.4 Addressing into the squashed cube . . . . . . . . . . . . . . . . . . . . . . . 21
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.5 Beauty is rare . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2 Basic linear algebra and combinatorics 31

2.1 A guide to basic abstract algebra . . . . . . . . . . . . . . . . . . . . . . . . 31
2.1.1 Fundamental structures . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.1.2 Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.1.3 Linear spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.1.4 Criteria of linear independence . . . . . . . . . . . . . . . . . . . . . 37
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.2 Affine subspaces, linear equations, rank . . . . . . . . . . . . . . . . . . . . . 43
2.2.1 Inequalities for subspaces . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.2.2 Linear maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.2.3 Matrices, rank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.2.4 Systems of linear equations. Affine subspaces . . . . . . . . . . . . . 46
2.2.5 Projective spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.2.6 Extending the field . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.3 Orthogonality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.3.1 Inner product spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.3.2 Eventown revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

vii
CONTENTS
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.4 Graphs and set systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.4.1 Notation, terminology . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.4.2 Chromatic number and short cycles . . . . . . . . . . . . . . . . . . . 58
2.4.3 Block designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

3 “General position” arguments 61

3.1 Configurations in general position . . . . . . . . . . . . . . . . . . . . . . . . 61
3.1.1 Points in general position. The moment curve. . . . . . . . . . . . . . 61
3.1.2 Subspace in general position w.r.t. a family of subspaces . . . . . . . 62
3.1.3 Linear maps in general position . . . . . . . . . . . . . . . . . . . . . 65
3.1.4 Checking identities: a Monte Carlo algorithm . . . . . . . . . . . . . 66
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.2 Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.2.1 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.2.2 Helly’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.2.3 A polytope with many faces . . . . . . . . . . . . . . . . . . . . . . . 73
3.2.4 Distributing points on the sphere . . . . . . . . . . . . . . . . . . . . 76
3.2.5 Borsuk’s and Kneser’s graphs . . . . . . . . . . . . . . . . . . . . . . 77
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.2.6 Linear and statistical independence . . . . . . . . . . . . . . . . . . . 80

4 Set systems with restricted intersections 81

4.1 When all intersections are equal size . . . . . . . . . . . . . . . . . . . . . . 81
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.2 Ramsey theory – a constructive lower bound . . . . . . . . . . . . . . . . . . 85
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.3 Restricted intersections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.4 Extremal set theory: the classics . . . . . . . . . . . . . . . . . . . . . . . . . 95

5 Spaces of polynomials 99
5.1 Helly-type theorems for finite sets . . . . . . . . . . . . . . . . . . . . . . . . 99
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.2 Resultants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.3 The Prague dimension of graphs . . . . . . . . . . . . . . . . . . . . . . . . . 104
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.4 Sets with few intersection sizes mod p . . . . . . . . . . . . . . . . . . . . . . 108
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.5 Geometric application: unit distance is hard to miss . . . . . . . . . . . . . . 112
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.6 Reducing the diameter of bodies: Borsuk’s conjecture disproved . . . . . . . 116
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

viii ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CONTENTS
5.7 Constructive Ramsey graphs via intersection theorems . . . . . . . . . . . . 119
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.8 Geometric application: any distance is hard to miss . . . . . . . . . . . . . . 121
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.9 Prime power moduli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.10 The nonuniform RW Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 126
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.11 The Ray-Chaudhuri – Wilson Theorem . . . . . . . . . . . . . . . . . . . . . 128
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.12 A modular Ray-Chaudhuri – Wilson Theorem . . . . . . . . . . . . . . . . . 130
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

6 Tensor product methods 135

6.1 Wedge products — a concrete introduction . . . . . . . . . . . . . . . . . . . 136
6.1.1 The Laplace expansion of determinants . . . . . . . . . . . . . . . . . 136
6.1.2 Alternating k-linear functions . . . . . . . . . . . . . . . . . . . . . . 136
6.1.3 Exterior powers of Fn . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.2 Bollobás–type theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
6.3 Symmetric products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.4 The Shannon capacity of a graph . . . . . . . . . . . . . . . . . . . . . . . . 147

7 A class of higher incidence matrices: the inclusion matrix 149

7.1 The inclusion matrix; s-independent families . . . . . . . . . . . . . . . . . . 149
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
7.2 Extended inclusion matrices. The Nonuniform RW Theorem revisited . . . . 151
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.3 Inclusion matrices of uniform families . . . . . . . . . . . . . . . . . . . . . . 154
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
7.4 Linear dependencies among the rows of inclusion matrices and the Vapnik–
Chervonenkis dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
7.5 Shadows of s-independent families . . . . . . . . . . . . . . . . . . . . . . . . 165
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

8 Applications of inclusion matrices 169

8.1 The edge-reconstruction problem . . . . . . . . . . . . . . . . . . . . . . . . 169
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
8.2 Chromatic critical graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
8.3 Partially ordered sets, unimodal sequences, and the Sperner property . . . . 174

——————————————————————— ix

c László Babai and Péter Frankl. 1988, 1992, 2020.
 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

9 Partially ordered sets 179

9.1 Geometric semilattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
9.1.1 Matroids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
9.1.2 Geometric lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
9.1.3 RW-type theorems for semilattices . . . . . . . . . . . . . . . . . . . 179
9.2 Incidence matrices of full rank . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
9.3 The Möbius function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
9.3.1 Möbius inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
9.3.2 The Möbius function in geometric lattices . . . . . . . . . . . . . . . 182
9.3.3 Whitney number inequalities . . . . . . . . . . . . . . . . . . . . . . . 184
9.3.4 The VC dimension revisited: shattered elements in a poset . . . . . . 185

10 Applications to the Theory of Computing 189

10.1 Communication complexity theory . . . . . . . . . . . . . . . . . . . . . . . 189
10.2 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

Answers to the exercises 191

A.1 Chapter 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
A.2 Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
A.3 Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
A.4 Chapter 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
A.5 Chapter 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
A.6 Chapter 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
A.7 Chapter 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
A.8 Chapter 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
A.9 Chapter 9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

Bibliography 213

Index 212

x ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
List of Figures

1.1 Clubs in Oddtown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Squaring a regular triangle: Dudeney’s 4-piece jigsaw puzzle . . . . . . . . . 21
1.3 Three minimal decompositions of K4 . . . . . . . . . . . . . . . . . . . . . . 22
1.4 Squashing the cube. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.5 The complete graph K4 as a squashed 3-cube. . . . . . . . . . . . . . . . . . 24
1.6 Petersen’s graph. (Which one is it?) . . . . . . . . . . . . . . . . . . . . . . . 26

2.1 The multiplication table of F4 . . . . . . . . . . . . . . . . . . . . . . . . . . 40

2.2 (a) The Fano plane. (b) The anti-Fano configuration. . . . . . . . . . . . . 52

A.1 Four Mod-4-town clubs which are dependent mod 2. . . . . . . . . . . . . . . 193

A.2 Squaring a rectangle by dissection. . . . . . . . . . . . . . . . . . . . . . . . 197
A.3 A good cut for the pastry chef. . . . . . . . . . . . . . . . . . . . . . . . . . 198
A.4 The Fano plane, coordinatized . . . . . . . . . . . . . . . . . . . . . . . . . 201
A.5 Six vectors in general position in F34 . . . . . . . . . . . . . . . . . . . . . . 202
A.6 Three colors don’t suffice. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

xi
xii ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
Notation and terminology

Sets.
|X| denotes the cardinality of a set X, i. e., its size (the number of its elements). A k-set
is a set of k elements.
[n] = {1, 2, . . . , n}.  
2X is the set of all subsets of the set X. If |X| = n then 2X  = 2n . 

X  X 
 k  = nk .



k
denotes the set of all k-subsets of the set X (k ≥ 0) If |X| = n then

 P
X

 X  k n


≤k
denotes the set of all subsets of X of size ≤ k. If |X| = n then  ≤k  = i=0 i .
n
 n
We call the set {0, 1} := (1 , . . . , n ) : i ∈ {0, 1} the n-cube. It has 2 points.
The incidence vector of the set A ⊆ [n] is (α1 , . . . , αn ) ∈ {0, 1}n , where
(
1 if i ∈ A ;
αi =
0 if i 6∈ A .

Strings

S n denotes the set of ordered n-tuples from the set S. In this context we sometimes refer
to S as the alphabet; the members of S n are strings of length n over S.
The Hamming distance of two strings x = (x1 , . . . , xn ) ∈ S n and y = (y1 , . . . , yn ) ∈ S n is
defined as the number of places i where xi 6= yi . In particular, the Hamming distance of the
incidence vectors of two sets A, B is the cardinality of their symmetric difference. If a symbol
“0” belongs to S, the weight of x ∈ S n is the Hamming-distance of x and (0, . . . , 0) ∈ S n . If
S is a subset of an abelian group then the Hamming distance of x, y ∈ S n is the weight of
x − y.

Families of sets
A set system or a family of sets (often simply a family) is a set of sets. The sets belonging
to the family are its members. When using the notation F = {A1 , . . . , Am }, the Ai are
automatically assumed to be distinct unless the opposite is explicitly stated.
A set system F over a set X is a family of subsets of X. We sometimes refer to X as the
universe of F.
A set system F is k-uniform if its members are k-sets. F is uniform if it is k-uniform for
some k.

xiii
NOTATION AND TERMINOLOGY
Let L be a set of integers. A set system F is L-intersecting if |E ∩ F | ∈ L for any two
distinct E, F ∈ F. (For a modular version of this concept, see Def. 5.14.)

Graphs
A graph G = (V, E) is a pair consisting of a set V of vertices and a 2-uniform set system
E whose members are the edges. An edge is thus an unordered pair of vertices. The edge
{u, v} is said to join the vertices u and v. The vertices u and v are said to be adjacent in
the graph G if the pair {u, v} is one of the edges of G. The vertices adjacent to vertex u are
the neighbors of u. Their number is the degree of u. G is a regular graph if all vertices have
the same degree.
The maximum number of edges a graph with n vertices can have is n2 . If all the n2

edges are present in G then G is said to be complete. The complete graph on n vertices is
denoted by Kn .
The complement of the graph G = (V, E) is another graph G on the same vertex set,
V


having complementary edge-set: G = (V, E) where E = 2 \ E.
The adjacency matrix of G = ([n], E) is an n × n (0, 1)-matrix A = (αij ) where αij = 1
if vertices i and j are adjacent; αij = 0 otherwise. Note that A = AT and the diagonal
elements of A are αii = 0.
Note that the adjacency matrix of the complete graph Kn is Jn − In .
A bipartite graph has its vertex set split into two disjoint subsets and all edges must go
between the two parts only. (No edge is allowed to join two vertices in the same part.) If
the two parts have r and s vertices, resp., then there are at most rs edges in the bipartite
graph. If all the rs edges are present in G then G is a complete bipartite graph. The complete
bipartitie graph with r and s vertices in the two parts, resp., is denoted by Kr,s .
A cycle of length k in a graph is a sequence of k ≥ 3 distinct vertices (a1 , . . . , ak ) such
that each ai is adjacent to ai+1 (i = 1, . . . , k − 1), and ak is adjacent to a1 . The girth of a
graph is the length of its shortest cyle. (Graphs with no cycles are called forests; they have
infinite girth.) A Hamilton cycle is a cycle that passes through all points of the graph.
More terminology for graphs and set systems can be found in Section 2.4.1. For the
elements of graph theory we refer the reader to Bollobás (1979) or Bondy-Murty (1976). For
more advanced problem–solvers, Lovász (1979c) is an invaluable source.

Fields, rings, matrices

Z – set of integers
Q – set of rational numbers
R – set of real numbers
C – set of complex numbers
Fq – field of q elements (same as GF (q); q is a prime power). If q = p is a prime, then Fp
can be viewed as the set {0, 1, . . . , p − 1} with addition and multiplication performed modulo
p.
F – any field
F[x] – ring polynomials in one indeterminate over F
F[x1 , . . . , xn ] – ring of polynomials in n indeterminates over F

xiv ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 Dk×n – set of k × n matrices over the domain D
Dn = Dn×1 – set of column vectors of length n over D. Occasionally we use this notation
for row vectors for typographic convenience when no danger of confusion arises
In – n × n identity matrix
×P
Jn – n p n all-ones matrix (every entry is 1)
n 2 n
kxk = i=1 αi – Euclidean norm of a vector x = (α1 , . . . , αn ) ∈ R
kxk∞ = maxni=1 αi – maximum norm

Multivariate polynomials
monomial – a product of variables with a scalar coefficient, e. g., 3x32 x4 x27
degree of monomial – sum of the exponents
polynomial of degree k – maximum degree of the monomials appearing in its full expan-
sion. The degree of the zero polynomial is −∞
homogeneous polynomial of degree k – a sum of monomials of degree k
monic monomial – monomial with coefficient 1, e. g., x32 x4 x27
multilinear polynomial – the degree in each variable is ≤ 1, e. g., 3x1 x4 − 5x2 x4 x7
More algebra terminology can be found in Sections 2.1 and 2.2.

Comparison of orders of magnitude1 of two sequences {an } and {bn } of real numbers:
an = O(bn ) (“big-oh of bn ”) means |an | ≤ Cbn for some constant C and all sufficiently
large n;
an = Ω(bn ) means bn = O(an ), i.e., an ≥ c |bn | for some constant c > 0 and all sufficiently
large n;
an = Θ(bn ) means an = O(bn ) and an = Ω(bn ) simultaneously, i.e., 0 ≤ c1 bn ≤ an ≤ c2 bn
for some positive constants c1 , c2 and all sufficiently large n;
an = o(bn ) (“little-oh of bn ”) means limn→∞ an /bn = 0.

1
The story of the big-O notation is told by D.E. Knuth in “Big Omicron and Big Omega and Big Theta”,
SIGACT NEWS 8:2 (1976), pp. 18–24.

——————————————————————— xv

c László Babai and Péter Frankl. 1988, 1992, 2020.
Chapter 1

Warm-up

1.1 Counting clubs in Oddtown

Eventown has 32 inhabitants. They have a habit of forming clubs, small and large; in fact,
that seems to be their principal activity. Clubs cannot be formed arbitrarily; the city passed
a rule requiring that
(i) each club have an even number of members,
(ii) each pair of clubs share an even number of members,
(iii) no two clubs are allowed to have identical membership.
The citizens of Eventown wish to form as many clubs as possible under the “Eventown rules.”
They will not refrain from registering the “Empty club” which has no members. (Zero is an
even number, they argue.)
The question is, how many clubs can there be in Eventown.
A simple strategy will allow them to have quite a few. Assume for simplicity that everyone
in town is married and just insist that spouses belong to precisely the same clubs. How many
clubs can be formed observing this single rule? Each club is determined by a sequence of 16
binary decisions: will the couple #1, #2, . . ., #16 belong to the club? There are 216 = 65, 536
such decision sequences resulting in the same number of clubs. Quite a large number of clubs
for such a small town.
The unmanageable number of clubs would rapidly undermine law and order. To prevent
this, heavy-handed legislation is being introduced in the city council. The proposed new law
would replace the word even by odd in rule (i) as well as in the name of the city.
Observe that this change will eliminate the need for rule (iii). Thus the new law (“Odd-
town rules”) would read:
(a) Each club shall have an odd number of members.
(b) Each pair of clubs shall share an even number of members.

c László Babai and Péter Frankl. 1988, 1992, 2020.

1
CHAPTER 1. WARM-UP
City legislators hope by this seemingly minute change to drastically reduce the number of
clubs. Just how drastically, we are going to find out.
It is still possible to have 32 clubs in Oddtown. For instance, each individual could form
a one-member club. (Even if such solitary clubs were not acceptable, there are other ways
of forming 32 clubs. For instance, each club could have 31 members, everybody being a
member of all but one of the 32 clubs. A more imaginative plan would result in 31 clubs
of 7 members each and one additional club having 31 members — but in order to describe
this configuration, we have to use finite projective planes, to be introduced in Section 2.4.3
(Cf. Exercise 1.1.13 below.) In fact, there is a tremendous number of ways to form 32 clubs,
as shown in Exercise 1.1.14.) The real surprise might be, that, although 32 clubs can be

Figure 1.1: Clubs in Oddtown

formed in a large number of different ways, there is no way of forming 33 or more clubs
under the new rules.
Even more unexpectedly, it is basic linear algebra that provides the tools for a remarkably
elegant proof of this fact. We conclude this section by describing the proof, thus illustrating
a fundamental method in the theory of extremal combinatorial configurations. Most of the
present book is devoted to applications and variants of this technique which we call the linear
algebra bound method.
The idea is to associate vectors in 32-dimensional space with each club and to prove that
the vectors obtained are linearly independent. This implies that there are at most 32 of
them.
Suppose we have m clubs, C1 , . . . , Cm . The incidence vector vi of Ci is a (0,1) vector
with 32 entries; the j th entry of vi is 1 or 0 according to whether or not citizen #j belongs
to Ci . (For simplicity we assume that Oddtown citizens are numbered 1 through 32.)
In the space of vectors with 32 entries, we introduce the standard inner product operation
by setting

x · y = x1 y1 + x2 y2 + · · · + x32 y32 ,

where x = (x1 , x2 , . . . , x32 ), and y = (y1 , y2 , . . . , y32 ).

2 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.1. COUNTING CLUBS IN ODDTOWN
A moment’s reflection will convince the reader that the inner product of the incidence
vectors vi and vj is precisely
vi · vj = |Ci ∩ Cj | . (1)
The “Oddtown Rules” (a) and (b) can thus be rephrased in an algebraic form:
(
odd, if i = j ;
vi · vj = (2)
even, if i 6= j .

Further simplification will result if we agree to work over the field F2 of two elements. This
means our domain of “numbers” will consist of 0 and 1 only, with the arithmetic operations
performed modulo 2 (thus 1 + 1 = 0). With this convention, rules (a) and (b) take the form
(
1 if i = j ;
vi · vj = (3)
0 if i 6= j .

We shall prove that under these conditions, the vectors v1 , . . . v32 must be linearly in-
dependent over the coefficient domain F2 . This, incidentally, is a stronger statement, than
saying they are independent over the domain of real (or rational) numbers. (Why? Cf.
Exercises 1.1.16 and 1.1.25.)
Indeed, let us consider a linear relation over F2 among the vi :

λ1 v1 + λ2 v2 + · · · + λ32 v32 = 0. (4)

(Here, λi ∈ F2 .) We must prove that all coefficients are zero. In order to see, for instance,
that λ1 must vanish, let us compute the inner product of each side of equation (4) by v1 :

λ1 v1 · v1 + λ2 v2 · v1 + · · · + λ32 v32 · v1 = 0. (5)

Taking equation (3) into account, equation (5) simply reduces to λ1 = 0. An analogous
argument proves λi = 0 for each i (1 ≤ i ≤ 32). This completes the proof that the vi are
linearly independent, and consequently m ≤ 32.
Of course, the number 32 played no role in the argument. What we proved in essence is
the following.
Theorem 1.1. Assume that the clubs in a town of n citizens satisfy the rules (a) and (b).
Then the incidence vectors of the clubs are linearly independent over the field F2 .
We state the immediate consequence we have been after.
Corollary 1.2 (“Oddtown Theorem”). In a town of n citizens, no more than n clubs
can be formed under rules (a) and (b).
This result seems to first have appeared in a note by E. Berlekamp (1969). More gems from
the same collection will be on display in Section 2.3.2 where the Eventown Rules will be
examined in detail. (See especially Exercise 2.3.11.) We give another formulation of the
proof of Theorem 1.1, using the rank inequality

rk (AB) ≤ min{rk A, rk B}, (6)

——————————————————————— 3

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 1. WARM-UP
where A and B are matrices over an arbitrary field, and the number of columns of A equals
the number of rows of B.
We think of the vi as the rows of an m × n matrix M , the incidence matrix of the system
of clubs. We claim that the rows of M are linearly independent over F2 .
In other words, we claim that the rank of M over F2 is rk2 M = m.
Consider the m × m matrix A = M M T , where M T denotes the transpose of M . We call
A the intersection matrix of the system {Ci : i = 1, . . . , m}, because, as the reader readily
verifies, the entry of A in position (i, j) is precisely |Ci ∩ Cj |. (This is just a restatement of
equation (1).)
By equation (6), rk2 A ≤ rk2 M . So it will suffice to prove that rk2 A = m. But this is now
immediate since by rules (a) and (b), A is the m × m identity matrix. (This is just another
way of stating equation (3).) With this we have concluded our second proof of the Oddtown
Theorem.

In both proofs, we have made use of the elements of linear algebra over the field F2 . With
slight modifications, each proof can be carried out directly over the more familiar domain
Q of rational numbers (cf. Exercises 1.1.1 and 1.1.2.). While this approach does not yield
the full power of Theorem 1.1 (linear independence over the rationals rather than over F2
will only follow) this is perfectly sufficient for proving the Oddtown Theorem. Although the
“mod 2” approach seems simpler, the “rational” solution is better suited for generalizations
(cf. Exercises 1.1.23 and 1.1.25 as well as Section 7.3). This is not a general rule, though:
Exercises 1.1.5, 1.1.28, and the “Eventown problems” 1.1.9 and 1.1.10 demonstrate cases
when “mod 2” is the only approach available.

Exercises
♦Ex. 1.1.1 (Oddtown Theorem, third proof). Give a direct proof that under rules (a) and (b), the
(0,1)-vectors vi are linearly independent over Q, the field of rational numbers. Use the idea of the
first proof given in the main text.

♦Ex. 1.1.2 (Oddtown Theorem, fourth proof). Solve the previous exercise by adapting the second
proof given in the main text to the field of rational or real numbers.

♦Ex. 1.1.3. Let us switch “odd” and “even” in rules (a) and (b). Under the new “Reverse
Oddtown Rules,” clubs must be even and their pairwise intersections must be odd. Prove that no
more than n clubs can be formed under the new rules in a town of n citizens.

♦Ex. 1.1.4. Let Jm denote the m × m matrix with 1’s in every cell, and let Im be the m × m
identity matrix. Consider the matrix Jm − Im (zeros in the diagonal, ones elsewhere).

(a) Compute the determinant of Jm − Im .

(b) Show that the rank of Jm − Im over F2 is m if m is even.

(c) Show that the rank of Jm − Im over F2 is m − 1 if m is odd.

4 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.1. COUNTING CLUBS IN ODDTOWN
Hint. (a) The result is (−1)m−1 (m − 1).
♦Ex. 1.1.5. What is the maximum number of clubs that can be formed under the “Reverse
Oddtown Rules” stated in Exercise 1.1.3?
Hint (“Reverse Oddtown Theorem”). The answer is n if n is odd and n − 1 if n is even.
♦Ex. 1.1.6. Let A be a 2n × 2n matrix with zeros in the diagonal and ±1 everywhere else. Prove
that A is nonsingular (over R).

Ex. 1.1.7 (Bipartite Oddtown Theorem). Suppose there are m red clubs R1 , . . . , Rm and m blue
clubs B1 , . . . , Bm in a town of n citizens. Assume that these clubs satisfy the following rules:
(α) |Ri ∩ Bi | is odd for every i;

(β) |Ri ∩ Bj | is even for every i 6= j.

Prove that m ≤ n.

Ex. 1.1.8 (Skew Oddtown Theorem). Weaken assumption (β) of the preceding exercise to
(γ) |Ri ∩ Bj | is even for 1 ≤ i < j ≤ m.
Prove that (α) and (γ) still imply m ≤ n.

Ex. 1.1.9 (Eventown Theorem). Prove that no more than 2bn/2c clubs can be formed under
Eventown Rules (i), (ii), and (iii), where bxc (“floor of x”) denotes the greatest integer not exceeding
x.
(See Section 2.3.2 for the solution. If n is even, this upper bound remains valid even if we drop
condition (i), and it increases by only 1 if n is odd (Berlekamp–Graver, see Exercise 2.3.11).)

Ex. 1.1.10.* Suppose Eventown is living under its rules (i), (ii), and (iii), and has fewer than
2bn/2c clubs. Prove that there is room for a new club without violating the law.
(See Section 2.3.2 for the solution.)
♦Ex. 1.1.11. Prove that the analogous statement for the Oddtown Rules (a) and (b) is false: it is
possible to form fewer than n clubs under these rules so that no more clubs can be added. In fact,
for any integer t between 0 and (n − 1)/2 there is a way of forming a maximal system of precisely
n − 2t clubs.
♦Ex. 1.1.12. Prove that for n ≥ 7, there exist nonisomorphic extremal solutions to the Eventown
problem. (An extremal solution is a system of maximum possible number of clubs, i. e., m = 2bn/2c .
Two systems of clubs are isomorphic if one can be transformed into the other by renaming the
citizens.)
Hint. Construct a small system first, and use Ex. 1.1.10.
♦Ex. 1.1.13. For now, let Oddtown again have n = 32 citizens. Construct a set of 32 clubs
obeying rules (a) and (b) with the parameters of the “more imaginative plan” indicated in the
main text, i. e., one of the clubs should have 31 members, all the others 7 members each. Each pair
of 7-member clubs will share 2 members; and 6 of the 7 members of each small club will belong to
the large club.

——————————————————————— 5

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 1. WARM-UP
Hint. Use projective planes (Sec. 2.4.3).

♦Ex. 1.1.14 (M. Szegedy, 1988). Show that if n is even then there exist at least 2n(n+2)/8 /(n!)2
nonisomorphic extremal solutions to the Oddtown problem. (This is a very large number: for large
2 2
n, it is greater than 2n /9 . (Prove!)) Prove that there are no more than 2n /n! extremal solutions.

* * *

The next sequence of problems investigates the connections between linear independence
over various fields.
♦Ex. 1.1.15. Let v1 , . . . , vm be vectors with n rational entries. Prove that the vi are linearly
independent over Q if and only if they are linearly independent over R.

Ex. 1.1.16. Let v1 , . . . , vm be vectors with n (0,1)-entries. Prove that if these vectors are linearly
independent over Fp for some prime number p then they are linearly independent over Q.

♦Ex. 1.1.17. Given an integer d, exhibit a square (0, 1)-matrix with determinant d.

♦Ex. 1.1.18. Construct a set of (0, 1)-vectors, linearly independent over Q and F3 but linearly
dependent over F2 and F5 .

Ex. 1.1.19. Prove: if a set of (0,1)-vectors is linearly independent over Q, then they are linarly
independent over Fp for every sufficiently large prime p.

* * *

Ex. 1.1.20. Give mod 3, mod 5, etc. analogues of the Oddtown Theorem. Generalize each of the
four proofs given.
Hint. See the next problem for prime numbers s. When adapting the proofs given in the main
text, use the finite field Fs in place of F2 .

* * *

For a positive integer s, define the set R(s) of “mod-s-town” rules as follows:

Rule R(s) Let A1 , . . . , Am be subsets of a set of n elements. Assume the sizes of the
Ai are not divisible by s, but the sizes of their pairwise intersections are
divisible by s.

6 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.2. TWO-DISTANCE SETS
Ex. 1.1.21. Prove that R(s) implies m ≤ n when s is a prime number. (This is identical with the
preceding exercise.)

Ex. 1.1.22. Does Rule R(4) imply m ≤ n ?

Hint. Yes, it does. (See the next two exercises.) The solutions to Exercises 1.1.1 and 1.1.2 can be
adapted. But the proofs given in the main text do not generalize, as indicated by Exercise 1.1.25.

Ex. 1.1.23 (Mod-pk -town Theorem (Babai–Frankl, 1980)). Prove that rule R(pk ) implies
m ≤ n for all prime powers pk . Adapt the proof given in Exercise 1.1.1.
Hint. Proceed as in the solution of Exercise 1.1.1. Obtain a contradiction by showing that each
coefficient must be divisible by p.

Ex. 1.1.24. Solve the previous problem along the lines of Exercise 1.1.2.
Hint. Now the diagonal entries of the intersection matrix are not divisible by pk , all other entries
are. Prove that the determinant of such a matrix is nonzero. (See Proposition 4.14 in Section 4.3.)

Note that this proof yields linear independence of the incidence vectors over R while the previous
proof yields linear independence over Q only. These two results, however, are equivalent, as shown
by Ex. 1.1.15.

♦Ex. 1.1.25. Construct a set of clubs satisfying Rule R(4) (“Mod-4-town Rules,” see Exer-
cise 1.1.20) such that their incidence vectors be linearly dependent over F2 . (Note that according
to Exercise 1.1.23, they will necessarily be linearly independent over Q !) Try to make the town as
small as possible.

Ex. 1.1.26. State the bipartite and skew versions of the “Mod-pk -town Theorem” (Exercise 1.1.23.)
(The corresponding versions of the Oddtown Theorem are stated in Exercises 1.1.7 and 1.1.8). Prove
the bipartite version of the “Mod-pk -town Theorem” and the skew version of the “Mod-p-town The-
orem.” (We don’t know whether the skew version holds for prime powers. Where does the proof
break down?)

♦Ex. 1.1.27. Prove: Rule R(6) implies m ≤ 2n. More generally, for every s, there exists a constant
c(s) such that R(s) implies m ≤ c(s)n.
Hint. Let c(s) be the number of different primes, dividing s. (Note that this value of c(s) may not
be best possible. In fact, we don’t know whether R(s) implies m ≤ n when s is not a prime power.
This question is open even for s = 6.)

Ex. 1.1.28* (M. Szegedy, 1988). Prove: For n 6= 3, rule R(6) implies m ≤ 2n − 2 log2 n.
(See Section 2.3.2, Ex. 2.3.12 for the solution.)

1.2 Point sets in Rn with only two distances

We illustrate on a problem from geometry the unexpected ways in which the linear algebra
bound can be employed.
Let a1 , . . . , am be points in the n-dimensional Euclidean space Rn .

——————————————————————— 7

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 1. WARM-UP
If the pairwise distances of the ai are all equal then clearly m ≤ n + 1, the extreme case
being the set of vertices of a regular simplex.
Assume now that the pairwise distances between the ai take two values. Such a set is
called a two-distance set. What is the maximum number of points in a two-distance set in
Rn ?
Let m = m(n) denote this maximum. Although we shall not be able to determine the
exact value of m(n), we shall find good estimates.
We shall see that m(n) is approximately n2 /2 (for large n). More precisely, we shall prove
the following bounds:
Theorem 1.3. The maximum cardinality m(n) of a two-distance set in Rn satisfies the
inequalities
n(n + 1)/2 ≤ m(n) ≤ (n + 1)(n + 4)/2.
Note that the ratio of the two bounds tends to 1 as n → ∞.

We leave the lower bound proof as Exercise 1.2.3. For the upper bound proof, let us
assume that the two distances occurring between the ai are δ1 and δ2 . Using the notation
1/2
kxk = ( nk=1 x2k ) for the Euclidean norm of x = (x1 , . . . , xn ) ∈ Rn , the distance between
P
two points x, y ∈ Rn is kx − yk. It is therefore natural to consider the polynomial
F (x, y) := kx − yk2 − δ12 kx − yk2 − δ22



(7)
in 2n real variables: x, y ∈ Rn . This polynomial puts our two-distance condition in a simple
algebraic form: (
(δ1 δ2 )2 6= 0 if i = j ;
F (ai , aj ) = (8)
0 6 j.
if i =
Substituting ai for y we obtain the polynomial fi (x) := F (x, ai ) in n variables x =
(x1 , . . . , xn ).
We claim that the polynomials f1 , . . . , fm are linearly independent over R.
In order to see this, assume that some linear combination of the fi is (identically) zero:
λ1 f1 (x) + · · · + λm fm (x) = 0. (9)
Substituting aj for x we obtain that λj fj (aj ) = 0, consequently all coefficients must
vanish, thus proving the claim.
On the other hand, all polynomials fi can be represented as linear combinations of the
following polynomials:
n
!2 n
!
X X
x2k , x2k · xj , xi xj , xi , 1. (10)
k=1 k=1

(The range of i and j is from 1 to n.) The number of polynomials listed is 1 + n + n(n +
1)/2 + n + 1 = (n + 1)(n + 4)/2. Therefore, all the fi belong to a linear space of dimension
≤ (n + 1)(n + 4)/2. As they are linearly independent, their number cannot exceed the
dimension of this space.

8 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.2. TWO-DISTANCE SETS
The result is due to D. G. Larman, C. A. Rogers and J. J. Seidel (1977); the beautiful
trick seems to have first been used by T. H. Koornwinder (1976).
We should point out that this result is closely related to extremal problems of the kind
treated in the Oddtown section (Section 1.1). Exercises 1.2.3 and 1.2.13–1.2.16 provide good
examples. We shall devote the entire Chapter 5 to developing combinatorial applications of
the polynomial space method, illustrated here.
The upper bound in Theorem 1.3 can be restated, with a generalization in mind, as
   
(n + 1)(n + 4) n+2 n+1
m(n) ≤ = + . (11)
2 2 1
Bannai and Bannai (1981) proved that this bound is never tight (they improved the right
hand side by 1). A more significant improvement was achieved by A. Blokhuis (1981) who
showed that the first term on the right hand side of (11) is sufficient (Exercise 1.2.27).
Blokhuis’s trick adds substantial extra power to the linear algebra bound technique, as we
shall demonstrate in Sections 5.11 and 5.12.
If all vectors of a 2-distance set have unit length, the set is called a spherical 2-distance
set. A slight modification of the argument presented yields the following bound ms (n) for
the maximum size of a spherical 2-distance set in Rn :

ms (n) ≤ n(n + 3)/2 (12)

(Exercise 1.2.8). What makes this inequality particularly interesting is that it is tight in
at least three different dimensions: n = 2, 6, 22, and, in these cases, equality is achieved by
quite remarkable geometric configurations. (See Exercise 1.2.10 for more details.)
A natural further generalization is to consider s-distance sets. This direction is discussed
in the notes after Exercise 1.2.21.

Exercises
Ex. 1.2.1. Let f (x) = α0 + α1 x + · · · + αn xn be a nonzero polynomial in one variable (n ≥ 0, αn 6=
0). Prove that f (x) has a nonzero multiple in which all the exponents are prime numbers. (For
instance, such a multiple of f (x) = x2 −x+5 is the polynomial x5 +4x3 +5x2 = (x3 +x2 )(x2 −x+5).)

Ex. 1.2.2. Prove: if all distances between m points in Rn are equal, then indeed m ≤ n + 1.
♦Ex. 1.2.3. Prove the lower bound m(n) ≥ n(n + 1)/2.
Ex. 1.2.4. Prove: m(2) = 5.
Ex. 1.2.5. Let Ω denote an arbitrary set and F a field; G(x, y) a function Ω × Ω → F for
some field F (x, y ∈ Ω). For b1 , . . . , bm ∈ Ω, define the functions gi : Ω → F by gi (x) := G(x, bi )
(i = 1, . . . , m). Prove: if the m×m matrix (G(bi , bj ))m
i,j=1 is nonsingular (has nonzero determinant)
then the functions g1 , . . . , gm are linearly independent (over F).
Hint. Any linear relation between the functions gi would imply the same linear relation between
the rows of the matrix (G(bi , bj )).

——————————————————————— 9

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 1. WARM-UP
♦Ex. 1.2.6. An m × m real matrix A = (aij ) is diagonally dominated if aii > j6=i |aij | for
P
i = 1, . . . , m (each diagonal entry is greater than the absolute sum of the rest of its row). Prove:
in this case, det A 6= 0.

Ex. 1.2.7. Prove the same upper bound ma (n) ≤ (n + 1)(n + 4)/2 if there are two approximate
distances, i. e., each distance is nearly equal to one of two given numbers.
Hint. Use the two preceding exercises.
♦Ex. 1.2.8 (Spherical two-distance sets: Delsarte–Goethals–Seidel (1977)). The (n − 1)-dimen-
sional unit sphere is defined as the set SS n−1 = {x ∈ Rn : kxk = 1}. A spherical 2-distance set is a
2-distance subset of SS n−1 . Prove: if ms (n) denotes the maximum number of points in such a set
then
n(n + 1)/2 ≤ ms (n) ≤ n(n + 3)/2.

Ex. 1.2.9** (Spherical s-distance sets: Delsarte–Goethals–Seidel, 1977). The maximum

cardinality of ms (n, s) spherical s distance sets in SS n−1 ⊂ Rn is
     
n+1 n+s−1 n+s−2
≤ ms (n, s) ≤ + .
s s s−1

Comment. The lower bound is easy. In the case s = 2, the bounds revert to the the spherical
two-distance bound of the previous exercise.
♦Ex. 1.2.10 (Gosset polytopes). (a) The 7-dimensional Gosset polytope has 56 vertices, given
in R8as (1, 1, 1, 1, 1, 1, −3, −3), (−1, −1, −1, −1, −1, −1, 3, 3), and all points obtained from these
by permuting the coordinates. (a1) Verify the stated number of vertices. (a2) Why is this a 7-
dimensional object? (a3) Show that all the 56 points belong to a sphere. (a4) Determine the set
of 56


2 = 1540 pairwise distances between these points. (a5) Find 28 among these 56 points that
form a 2-distance set. (Note that in dimension 7, Exercise 1.2.8 would allow up to 35 points in a
two-distance set.) (b) (The 6-dimensional Gosset polytope.) Find a hyperplane which contains 27
vertices of the 7-dimensional Gosset polytope that form a 2-distance set. Observe that this is the
maximum number allowed for a spherical 2-distance set in dimension 6 (Exercise 1.2.8).
Note. The 7-dimensional Gosset polytope itself is a face of the 8-dimensional Gosset polytope
which has 240 vertices and is related to the nonassociative “Cayley numbers.” These and other
remarkable configurations were found by T. Gosset, an unemployed lawyer, who found recreation
in work in geometry. He introduced the notion of semiregular polytopes at the turn of the century
and discovered some intriguing examples. These are polytopes with regular but not necessarily
congruent faces, and with a vertex-transitive group of symmetries (i. e., all vertices are equivalent
under symmetry transformations). Some references: T. Gosset (1900), H. S. M. Coxeter (1973),
H. S. M. Coxeter (1927), H. S. M. Coxeter (1968), D. G. Larman and C. A. Rogers (1972), P.
Delsarte, J. M. Goethals and J.J. Seidel (1977), L. A. Székely and N. C. Wormald (1989),

Ex. 1.2.11. Prove the msa (n) ≤ n(n + 3)/2 bound for spherical approximate 2-distance sets. (Cf.
Exercise 1.2.7.)
♦Ex. 1.2.12. Deduce Theorem 1.3 from the preceding exercise.
♦Ex. 1.2.13. Let A1 , . . . , Am be subsets of a set of n elements. Assume that their pairwise

10 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.2. TWO-DISTANCE SETS
symmetric differences have only two sizes. Prove that m ≤ n(n + 3)/2.
♦Ex. 1.2.14. (Continued.) Improve the upper bound given in the preceding exercise to m ≤
n(n+1)
1+ 2 .

Ex. 1.2.15. (Continued.) Find m = 1+ n(n−1)

2 subsets of an n-set with only two sizes of symmetric
differences.

Ex. 1.2.16. Let A1 , . . . , Am be k-subsets of a set of n elements. Assume that their pairwise
intersections have only two sizes. Prove that m ≤ 1 + n(n + 1)/2.
Hint. This is a particular case of Exercise 1.2.14. (Why?)
Stronger results on this problem will follow in Chapter 5.

Ex. 1.2.17. Let A, B ⊂ R3 such that all distances between members of A and members of B are
equal. Assume |A| ≤ |B|. Prove: |A| ≤ 2.
♦Ex. 1.2.18. (Continued.) Does there exist a finite bound on the size of the smaller of two sets
with analogous properties in R4 ?

Ex. 1.2.19. State and prove a “bipartite version” of Theorem 1.3 (in the sense of Exercise 1.1.7).
Hint. The preceding exercise serves as a warning: the first idea might be wrong. The right setting
is this. Let a1 , . . . , am , b1 , . . . , bm ∈ Rn satisfy the following condition:

• There exist two reals δ1 , δ2 such that the distance between ai and bj belongs to {δ1 , δ2 } if
and only if i 6= j.

Under this condition, the same upper bound m ≤ (n + 1)(n + 4)/2 follows.

Ex. 1.2.20. State and prove a “skew” version of Theorem 1.3 (in the sense of Exercise 1.1.8).
Hint. Adapt the solution of the preceding exercise with no assumption on the distances for i > j.
Obtain the same upper bound.
♦Ex. 1.2.21 (s-distance sets). Let m(n, s) denote the maximum number of points in Rn such that
their pairwise distances take at most s values. Prove:
   
n+1 n+s+1
≤ m(n, s) ≤ .
s s

Notes. Most of these results are not best possible. With a more accurate estimate on the
dimension
ofn+s−1
the space containing all the fi , one can improve the upper bound on m(n, s)
to n+s


s
+ s−1
(Delsarte–Goethals–Seidel, 1977). For s = 2, this is precisely the
Larman–Rogers–Seidel bound. Bannai and Bannai (1981) proved that this bound is never
tight. Subsequently A. Blokhuis, then a student in
Eindhoven (Netherlands), significantly
improved the Larman–Rogers–Seidel bound, to n+2 2
. (Exercise 1.2.27.) He then went on
to generalizing his result to s-distance sets and obtained the appealing inequality m(n, s) ≤
n+s


s
(Blokhuis, 1982, 1984). With Blokhuis’s first result in their hand, Bannai, Bannai, and
Stanton (1983) simultaneously arrived at the same general result. Although the improvement
may seem slight, the continuing quest for tight bounds has a very strong motivating factor:

——————————————————————— 11

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 1. WARM-UP
one expects that a tight bound might lead to extremal configurations of unique beauty (as is
the case for spherical 2-distance sets; see Exercises 1.2.8–1.2.10). Section 1.5 demonstrates
a particularly impressive case in point.
♦Ex. 1.2.22* (L. Danzer, B. Grünbaum, 1962). Let S be a finite subset of Rn . A support
hyperplane of S is a hyperplane which contains at least one point of S and does not separate any
two points of S. We say that S is nondegenerate if aff(S) = Rn . Let us call S ⊂ Rn a Klee-set, if
it is nondegenerate and there is a pair of parallel support planes through every pair of points in S.
Verify Victor Klee’s conjecture (1960): If S ⊆ Rn is a Klee-set then |S| ≤ 2n .
Ex. 1.2.23* (B. Csikós, 1981). Let us consider an arbitrary norm on Rn . Prove: if all pairwise
distances (defined by the given norm) between m points in Rn are equal then m ≤ 2n .
Hint. Use the preceding exercise.
Note that equality holds for the vertex set of the n-cube {1, −1}n in maximum norm.
It is not known whether equality can hold in any norm other than the maximum norm and
its affine equivalents (i. e., norms where the unit ball is a parallelopiped). For Lp -norms for any
real number p, 1 ≤ p < ∞, we conjecture that m ≤ nc holds for some absolute constant c. More
generally, this may be true for any norm whereP the unit ball has smooth boundary. (The Lp -norm
of x = (x1 , . . . , xn ) ∈ Rn is defined as kxkp = ( |xi |p )1/p .)
Ex. 1.2.24. (Continued.) Prove: for every even integer p ≥ 2, there exists a constant c(p) such
that any one-distance set with respect to the Lp -norm in Rn has at most nc(p) points.
Hint. If the single distance occurring is δ, consider the polynomial

F (x, y) := kx − ykpp − δ p .

Proceed as in the main text.

* * *

The following two routine linear algebra exercises build some background for the proof of Blokhuis’s
Theorem (Exercise 1.2.27).
♦Ex. 1.2.25. The hull of a set S = {v1 , . . . , vm } ⊂ Rn consists of all linear combinations
P m Paffine
m
where i=1 λi = 0. Let A be the m × n matrix of which the rows are v1 , . . . , vm . Add
i=1 λi vi ,
a column of all ones to the matrix to obtain the m × (n + 1) matrix B. Prove: if the affine hull of
S is Rn then the columns of B are linearly independent.
♦Ex. 1.2.26. Prove: if the columns of the m×k matrix B with real entries are linearly independent
then the k × k matrix B T B is nonsingular, i. e., det(B T B) 6= 0. (B T is the transpose of B).
♦Ex. 1.2.27* (Blokhuis, 1981). Prove: the maximum cardinality m(n) of a two-distance set in
Rn satisfies the inequality  
n+2
m(n) ≤ . (13)
2
Hint. With the notation of the proof of Theorem 1.3, consider the functions f1 , . . . , fm . Instead of
trying to show that they actually belong to a smaller space than the one spanned by the functions

12 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.3. A JIGSAW PUZZLE
(10), show that that space can accommodate more linearly independent functions. Augment the set
{f1 , . . . , fm } by the n + 1 additional functions 1, x1 , . . . , xn . Prove that this collection of m + n + 1
functions is still linearly independent. They all belong to the space of dimension (n + 1)(n + 4)/2
spanned by the polynomials listed in (10). Therefore m + n ≤ (n + 1)(n + 4)/2, which is the same
as inequality (13).

1.3 Two solutions to a jigsaw puzzle?

Can you cut up a regular triangle by straight cuts and put the pieces together to form a
square?
As we see in Figure 1.2 on p. 21, this can be done, and indeed a mere four pieces suffice. (A
solution in 5 pieces is not too hard to come by, following the general recipe of Ex. 1.3.3. Try,
before you turn the page to the figure with the magical 4-piece solution by H. E. Dudeney.)1

Now can you do the same in space, dissect the regular tetrahedron by plane cuts, and
put the pieces together to form a cube?
This question of elementary geometry was the essence of “Problem 3,” proposed by
German mathematician David Hilbert (1862–1943) in his historic address on August 8, 1900
to an audience of more than two hundred mathematicians who had convened in Paris, home of
that year’s Exposition Universelle, for the Second International Congress of Mathematicians.2
Perhaps the last of mathematicians (as opposed to logicians, ring theorists, geometers,
topologists you could have met half a century ago, or specialists of the mixed Hodge theory of
singular varieties3 you may admire these days), Hilbert celebrated the threshold to the 20th
century by a lecture entitled simply “Mathematische Probleme,” a collection of problems
“from the discussion of which an advancement of science may be expected.” The problems,
numbering 23 in the published version, encompassed a broad range of areas of mathematics,
from set theory to the calculus of variations.
As Hilbert pondered the idea of the lecture half a year before the Congress, his colleague
Hermann Minkowski boldly predicted that “With such a choice of subject you could have
people talking about your lecture decades later.”4
Now, more than a century later, Hilbert’s powerful vision is attested to by the volumes
of work around these problems, volumes that have profoundly transformed our view of
1
Henry Ernest Dudeney, famed as England’s foremost inventor of puzzles, discovered the four-piece
solution in 1902. For many more dissection puzzles, see Eves (1963), Chap. 5.
2
There were two such congresses before the “second.” The first Congress took place in Zürich in 1897,
with 208 participants. Prior to that, perhaps to be numbered zero, a “World Congress of Mathematicians
and Astronomers” was held in Chicago in 1893, after the conclusion of the first academic year at the newly
founded University of Chicago. (The occasion was the Chicago World’s Columbian Exposition, celebrating
the 400th anniversary of Columbus’s first voyage to the New World.) Forty-five mathematicians attended
this “World Congress,” including only four from the Old World, a fact that may explain the subsequent
careless numbering at a time of overwhelming European dominance of the mathematical scene.
3
Mathematics Subject Classification (1991) code: 32S35.
4
Emphasis added.

——————————————————————— 13

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 1. WARM-UP
mathematics.
Hilbert emphasized in his lecture that a particularly significant way in which an old
problem may be “solved” is a rigorous proof of the impossibility of the solution “in the
sense contemplated,” citing as milestones in the development of mathematics the proofs of
impossibility of proving the axiom of parallels, squaring the circle, solving equations of the
fifth degree by radicals.
Today, we might quote such further examples as the impossibility of deciding Georg
Cantor’s “Continuum Hypothesis” on the basis of the usual axioms of set theory (Paul Cohen,
1961), proving (within the formal system of arithmetic) that there is no contradiction hidden
in the axioms of arithmetic (Kurt Gödel, 1931), devising a general “method” (algorithm) to
decide the solvability of an equation in integers (Yuri V. Matiyasevich, 1970). These were
the rigorous negative solutions to three of Hilbert’s problems, quite unexpected perhaps to
Hilbert himself.5
One problem where Hilbert did suggest the impossibility of a solution was Problem 3,
also the problem that came to be solved first. The ingenious negative solution was given
by Max Dehn, Hilbert’s student, in the same year. Subsequently, Dehn’s proof was greatly
simplified. Below we follow the version of Hadwiger and Boltyanski.
Let us call two polyhedra in R3 equidissectible if one can dissect each of them by a finite
number of plane cuts so that the resulting two sets of smaller polyhedra can be paired off into
congruent pairs. In other words, one can cut up one of them and then reassemble the pieces
to obtain the other. (Do these two phrases really mean the same thing? See Exercise 1.3.6.)
Theorem 1.4 (M. Dehn, 1900). The regular simplex and the cube (of the same volume)
are not equidissectible.
Before proving Dehn’s result, we should mention that in the plane, nothing like this could
happen. Indeed, any two polygons of equal area are equidissectible. This result, obtained
independently by Farkas Bolyai6 (1832) and P. Gerwien7 (1833), was also alluded to in
Hilbert’s speech. Euclid based his discussion of the area of triangles and parallelograms
on simple dissection arguments. The Bolyai–Gerwien Theorem brought these ideas to a
definitive conclusion. “A splendid result for high school students to enjoy” (Kaplansky) (see
Exercise 1.3.3).
In contrast, Euclid’s treatise of the volume of solids is not based on the elementary
method of dissections but on a limiting process (“the method of exhaustion”), containing
5
In this paragraph we have indicated the author and date of the concluding step for each problem only.
The impossibility of disproving the Continuum Hypothesis was demonstrated by Gödel in 1940; Cohen added
the impossibilityof proving it. Matiyasevich proved an intricate relation for the Fibonacci numbers; that this
was all that was needed to complete the solution of the Tenth Problem came mainly from previous work by
Martin Davis and Julia Robinson (see Davis, Matiyasevich, Robinson (1976)).
6
Farkas (Wolfgang) Bolyai (1775–1856), Hungarian mathematician, professor at the Reformed College
of Maros-Vásárhely (Tirgu Mureş), Transylvania, was a Göttingen–educated geometer and former college
classmate of Gauss. His major pursuit was the elusive proof of the “axiom of parallels,” until his son, János
Bolyai (1802–1860), an army officer and a mathematical genius without formal training, shattered his father’s
hopes by announcing, at the age of 21, the invention of non-euclidean (hyperbolic) geometry (independently
discovered by N. I. Lobachevsky).
7
P. Gerwien was a German officer and mathematical amateur.

14 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.3. A JIGSAW PUZZLE
the germ of the idea of integration. Gauss called this situation “regrettable”; Dehn’s result
demonstrates that it is inevitable.

For the proof of Theorem 1.4 we require the following facts from linear algebra. Let V
be a linear space over Q. A linear function over V is a map f : V → Q with the property
that
f (a + b) = f (a) + f (b) for all a, b ∈ V. (14)
Fact. If v, w ∈ V are linearly independent then there exists a linear function f such that
f (v) = 0 and f (w) = 1.
We also require the following elementary result. Let α = arccos(1/3).
Proposition 1.5. α/π is irrational.
We defer the easy proof to the exercises (Ex. 1.3.10).
A dihedral angle is the angle subtended by two half-planes with a common bounding line
(the “spine”). We note that α is the dihedral angle (measured in radians) at each edge of a
regular tetrahedron. The dihedral angle at the edges of the cube is π/2.
Now assume for a contradiction that a regular tetrahedron and a cube are equidissectible.
Let β1 , . . . , βm be all the dihedral angles that occur at edges of the smaller polyhedra obtained
in the course of dissection. Let V denote the set of all linear combinations of the βi with
rational coefficients. Then V is a (finite dimensional) linear space over Q. By Proposition 1.5,
α and π are linearly independent over Q. Therefore, by the Fact stated above, there exists a
linear function f : V → Q such that f (π) = 0 and f (α) = 1. Note that f (π/2) = 0 follows.
Let us now consider, for each polytope P arising in the dissection process, the so-called
Dehn invariant of P with regard to f :
X
Φ(P ) = |ei |f (γi ), (15)
where the summation extends over all edges ei of P ; |ei | denotes the length of ei , and γi is
the dihedral angle of P at ei .
Lemma 1.6. The Dehn invariant is additive.
What this means is that if a polyhedron is cut up into pieces, the Φ-values of the pieces
add up to the Φ-value of the whole.
It suffices to prove this for a single cut P = P1 ∪ P2 , where the two pieces are cut apart
along a plane S and have disjoint interiors (P1 ∩ P2 ⊂ §). We have to show that

Φ(P ) = Φ(P1 ) + Φ(P2 ). (16)

Let us expand each term in eqn. (16) according to eqn. (15). Let us examine what happens to
the terms on the left hand side of the resulting equation. The terms corresponding to edges
not cut by S show up intact on the right hand side. If S cuts across an edge ei , it divides ei
into two pieces both still attached to the same dihedral angle γi , so the the corresponding
two terms on the right hand side add up to |ei |f (γi ). If S cuts into the “spine” (the edge ei
lies in the plane S), then S splits the dihedral angle γi and leaves ei unaltered. This time
the additivity of f guarantees that the balance of the two sides of eqn. (16) is maintained.

——————————————————————— 15

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 1. WARM-UP
Finally, we have to consider the contribution to the right hand side of eqn. (16) made by
the new edges arising along S but not appearing in P . Let e be such an edge (common to P1
and P2 ), and γi the corresponding dihedral angle in Pi . Since the new edge arose along the
intersection the plane S with a face of P , we have γ1 + γ2 = π. Therefore the contribution
of this edge to the right hand side of (16) is

|e|f (γ1 ) + |e|f (γ2 ) = |e|f (γ1 + γ2 ) = |e|f (π) = 0.

This completes the proof of the additivity lemma. The following corollary is immediate.
Corollary 1.7. If two polyhedra are equidissectible then their Dehn invariants are equal.
Indeed, the Dehn invariant of each is the sum of the Dehn invariants of the pieces.
However, the Dehn invariant of the cube is 0 (because f (π/2) = 0), while the Dehn
invariant of the regular tetrahedron is not (because f (α) 6= 0). This completes the proof of
Dehn’s Theorem.
For more on Hilbert’s problems, we refer to Aleksandrov (1969), Browder (1976), Ka-
plansky (1977). A master of elementary presentation, V. G. Boltyanski gives a wonderful
introduction into the modern theory of equidissectibility of polyhedra, accessible to the un-
dergraduate, in his book “Hilbert’s Third Problem” (1978).8
In addition to some of the exercises here, nice problems of elementary geometry also
appear among the exercises of Section 7.4.
In conclusion, we should mention the problem of general (set-theoretic) equidecom-
posability. Two subsets A, B ⊆ Rn are equidecomposable if there exist partitions A =
A1 ∪˙ · · · ∪A
˙ m and B = B1 ∪˙ · · · ∪B
˙ m such that Ai and Bi are isometric. (The dotted cups
indicate disjoint union.) We do not put any restriction on the geometric nature of the parts
Ai ; they are not even required to be Lebesgue measurable.
This concept gives rise to astonishing paradoxes, beginning with the Banach-Tarski para-
dox that says that the three dimensional ball is equidecomposable to two copies of itself.
No such thing can happen in the plane: if two plane figures are equidecomposable, they
must have equal area. (“Figure” here can mean any Lebesgue measurable set; polygons and
discs are the simplest examples.) The reason is that a Banach measure exists in the plane:
one can assign a nonnegative number (including ∞) β(A) to every set A ⊆ R2 such that
(1) if A = B ∪C ˙ then β(A) = β(B) + β(C); (2) if A has an “area” in the usual sense (or
more generally, it is Lebesgue measurable), then β(A) is equal to the area; (3) β is invariant
under isometries. (Prove, using the Banach–Tarski paradox, that no Banach measure exists
in R3 !)
Curiously, it turns out that it is the group theoretic structure of the isometry groups that
is responsible for this difference between R2 and R3 : the group of the isometries of the plane
is solvable, while the rotation group of R3 includes a free group on two generators. Groups
for which such paradoxical decompositions do not exist, are called amenable, a concept that
8
A piece of mathematical culture seldom encountered in the West is revealed in Boltyanski’s preface to
the English edition: “Mathematics is my hobby as well as my profession. I am especially fond of geometry
. . .. I first came across equidecomposability theory in a small booklet by Prof. V. F. Kagan, given to me as
a prize at the Sixth Moscow High School Mathematical Olympiad in 1940.”

16 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.3. A JIGSAW PUZZLE
has turned out to be of interest to a variety of areas of mathematics, including mathematical
logic, analysis, topology, and of course geometry and group theory.
Paradoxical decompositions are explored in Stan Wagon’s nice book (1978). Perhaps the
most splendid paradox of all does not appear in the book; it came with the brilliant proof
by M. Laczkovich (1990) that

SQUARING THE CIRCLE IS POSSIBLE!

Not in the sense of the ancient Greeks, of course.9 Laczkovich solved Tarski’s version of
the problem, proving that the square and the circle of the same area are equidecomposable
in the set theoretic sense defined above. In fact, they are translation-equidecomposable in
the obvious sense of this term.10

Exercises

♦Ex. 1.3.1. We are given a set of 13 weights. Assume that if we remove any one of them, the
remaining 12 weights can be divided into two sets of 6, with equal weight. Prove that all weights
must be equal.
Hint 1. First prove the result under the assumption that all weights are integers. Then extend to
rational weights. As a further extension, imagine that the weights are not numbers, but vectors
from a linear space over Q. Finally, reduce the original problem (with real numbers as weights) to
this auxiliary result.
Hint 2. For a direct proof, use Ex. 1.1.6.
♦Ex. 1.3.2. Let X be a finite set. An additive set-function on X is a function µ : 2X → R
(it associates a “measure” with every subset of X) such that if A ⊆ X and A = B ∪C ˙ then
µ(A) = β(B) + β(C).
Characterize those set systems F ⊂ 2X with the property that every function µ0 : F → R
extends to some additive set-function on X.
Hint. The necessary and sufficient condition is that the incidence vectors of the members of F are
linearly independent.
♦Ex. 1.3.3 (Bolyai–Gerwien Theorem). Prove that polygons of equal area are equidissectible.
Hint. Let P be a polygon of unit area; the goal is to turn it into a unit square. First cut it into
triangles; then turn each triangle into a rectangle. The hardest part is to show that rectangles of
equal area are equidissectible. (Why does this suffice now?) First show how to change the shape of
an oblong rectangle to one that is closer to a square: the ratio of the sides is less than 2. Finally,
turn this rectangle into a square.
♦Ex. 1.3.4* (Hilbert’s Third Problem). What Hilbert actually asked for was to exhibit two tri-
9
The Greeks asked for a construction, using a straight edge and a compass, of a square with area equal to
the area of a given circle. The impossibility of such a construction was demonstrated by F. Lindemann who
showed in 1882 that π is transcendental (it does not satisfy any algebraic equation with integer coefficients).
For more on this subject, see the literature on Hilbert’s Seventh Problem.
10
An insightful review of Laczkovich’s remarkable work was given by Gardner and Wagon (1989).

——————————————————————— 17

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 1. WARM-UP
angular pyramids (tetrahedra) with equal bases and equal altitudes, which are not equidissectible.

Hint. Take a unit cube. Call one face ABCD, and let A0 B 0 C 0 D0 be the oppoite face, labeled so that
A and A0 are adjacent, etc. Consider the two tetrahedra G = ABCB 0 and H = ABCC 0 . These
have equal base and equal height. Prove that H (“Hill’s tetrahedron”) is equidissectible with a
cube (M. J. M. Hill, 1895); while G isn’t.
♦Ex. 1.3.5 (Problem of the absent-minded pastry chef (Moscow Mathematics Olympiad problem)).
How can you put a triangular piece of cake into a box of which the shape is the mirror image of
the cake’s?
Hint. Cut it up into pieces (as economically as possible) and put the pieces together to fit into the
box. Make sure you don’t turn the icing side upside down.
♦Ex. 1.3.6 (Gerling, 1844). (a) Let P be a polyhedron in R3 and Q a mirror image of P . Prove
that one can dissect P by plane cuts into a finite number of pieces and put the pieces together to
obtain Q. (b) Infer that the two definitions of equidecomposability, given before the statement of
Theorem 1.4 are equivalent.
Hint. Let us call a polyhedron symmetrical if it has a plane of symmetry. Dissect P into symmetrical
polyhedra.

Ex. 1.3.7. Call two polygons translation-equidissectible if one can cut up one of them, and then
rearrange the pieces by applying a translation to each, so that the other polygon will result.
Prove that parallelograms of equal area are translation-equidissectible.
♦Ex. 1.3.8. Prove that a triangle and a square are never translation-equidissectible.
Hint. Introduce an additive function, invariant under translations.

Ex. 1.3.9 (Hadwiger–Glur, 1951). Prove: a convex polygon P and a square of the same area are
translation-equidissectible if and only if P is centrally symmetric.
Hint. See the solution of the preceding exercise. For full details, see Boltyanski (1978), pp. 78–86.

* * *

Hilbert’s Seventh Problem was “The irrationality and transcendence of certain numbers.”
In the set of elementary exercises to follow, we can but barely scratch the surface of this
theory, replete with questions of extreme difficulty.
♦Ex. 1.3.10. Let α = arccos(1/3). Show that α/π is irrational.
Hint. Suppose the contrary. Then there exist positive integers k, ` such that 2kπ = `α. Prove that
cos(`α) is of the form s/3` , where s is an integer not divisible by 3. This will contradict the fact
that cos(2kπ) = 0.

Ex. 1.3.11. Prove: the number arccos(1/n)/π is irrational for every positive integer n ≥ 3.
√
Ex. 1.3.12. Prove: the number arccos(1/ 3)/π is irrational.
Ex. 1.3.13. Prove that e (the base of the natural logarithm) is irrational.

18 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.3. A JIGSAW PUZZLE
P∞
Hint. Using the alternating series 1/e = i=0 (−1)i /i!,
prove for every n there exists an integer N
such that
0 < |n!/e − N | < 1/n.

♦Ex. 1.3.14.** Prove that π is irrational.

Hint. Here is an outline of a devilish one-page proof due to Ivan Niven (1947). The proof uses
elementary calculus only.
Assume π = a/b, the quotient of positive integers. For a positive integer n, consider the
polynomials
xn (a − bx)n
f (x) = (17)
n!
and
F (x) = f (x) − f (2) (x) + f (4) (x) − · · · + (−1)n f (2n) (x). (18)
Verify that
1. all derivatives f (j) (x) take integral values at x = 0;

2. infer that the same holds at x = π = a/b;

3. {F 0 (x) sin x − F (x) cos x} = f (x) sin x.

d
dx
Rπ
Calculate the integral rn = 0 f (x) sin x. Show that rn is an integer. On the other hand, prove
that 0 < rn < 1 for all sufficiently large n.

* * *

Ex. 1.3.15. Find the next few terms in this sequence: 1, 1, 2, 3, 5, 8, 13, . . . (Fibonacci numbers).
In subsequent exercises we denote them by uk , starting with u1 = u2 = 1. It is natural to add
u0 = 0 to the sequence.)
♦Ex. 1.3.16. Let us say that an infinite sequence a0 , a1 , . . . of complex numbers is a Fibonacci-type
sequence if ak = ak−1 + ak−2 for all k ≥ 2. Find all geometric progressions which are Fibonacci
type.

Ex. 1.3.17. Show that the Fibonacci-type sequences form a 2-dimensional linear space over C.
(The operations over sequences are termwise addition and termwise multiplication by a scalar.)
Hint. Show that the two sequences beginning with 0, 1 and 1, 0 form a basis.

Ex. 1.3.18. Show that the two Fibonacci-type

√ geometric progressions φk1 and φk2 are linearly
independent over C, where φ1,2 = (1 ± 5)/2.
Hint. It suffices to consider the first two terms (k = 0, 1) of each sequence.

Ex. 1.3.19. Prove the explicit formula for the Fibonacci number uk :
√ !k √ !k
 
1  (1 + 5) (1 − 5) 
uk = √ + . (19)
5 2 2

——————————————————————— 19

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 1. WARM-UP
Hint. By the previous exercises, the two Fibonacci-type geometric progressions form a basis in the
space of Fibonacci-type sequences. Therefore uk can be written as uk = α1 φk1 + α2 φk2 , where the
coefficients αi do not depend on k. Use the first two terms (k = 0, 1) to determine α1 , α2 .
√
k / 5, where φ =
Ex. 1.3.20
√ . Prove that uk (k ≥ 0) is the integer nearest to the number φ
(1 + 5)/2 ≈ 1.618034 is the golden ratio. Show that the rounding error alternates signs.
Estimate the error when k = 10.

Ex. 1.3.21. Prove: uk+1 uk−1 − u2k = ±1.

Hint. 1. Use the explicit formula. 2. For a more elegant solution, do not use the explicit
formula. Proceed by induction on k.

Ex. 1.3.22. (a) Prove: limn→∞ uk−1 /uk is the golden ratio. (b) Prove the same without using the
explicit formula.
Hint. (b) First assume the limit exists, and prove it can only be the golden ratio. Then prove that
the limit exists, using Exercise 1.3.21 and the easy fact that u2k > 2k .

Ex. 1.3.23. Prove: (a) uk and uk−1 are relatively prime. (b) If k|` (k divides `) then uk |u` . (c)
If the g.c.d. of k and ` is d then the g.c.d. of uk and u` is ud .

Ex. 1.3.24. Prove: u3k = 5u3k + 3 · (−1)k uk .

Ex. 1.3.25. For a positive integer m, consider the sequence of residues of of uk modulo m. Prove
that this sequence is periodic. Give an upper bound on the length of the period.
Hint. Show that the period is not longer than m2 − 1.
♦Ex. 1.3.26. Find an explicit formula for the terms of the following sequence: a0 = 0, a1 = 0,
a2 = 10, ak = 2ak−1 − ak−2 + 2ak−3 .

* * *

Ex. 1.3.27 (Cauchy’s functional equation). Let f : R → R be an unknown function satisfying

f (x + y) = f (x) + f (y) for all x, y ∈ R. (20)

(We call this a “functional equation” because what we are looking for is not a number (a root) but
rather a whole function f .) The trivial solutions to this equation are the linear functions f (x) = cx.
Prove: if f is a continuous function satisfying eqn. (20) then f is trivial.

Ex. 1.3.28. Prove that if a solution f of Cauchy’s functional equation is continuous at some
point x0 ∈ R then f is trivial.

Ex. 1.3.29. Prove that if a solution f of Cauchy’s functional equation is bounded in some interval
(a, b) ⊂ R then f is trivial.

Ex. 1.3.30. Prove that if a solution f of Cauchy’s functional equation is Lebesgue measurable in
some interval (a, b) ⊂ R then f is trivial.

20 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.4. THE SQUASHED CUBE
So if a solution of Cauchy’s equation is nontrivial, it must be quite pathological. Yet,
such solutions do exist (at least if we are willing to accept the “Axiom of Choice”).
Ex. 1.3.31 (G. Hamel, 1905). Prove that there exists a nontrivial solution to Cauchy’s
functional equation.
Hint. View R as an (infinite dimensional) linear space over Q. Then this space has a basis H
(a “Hamel basis”). This fact is a consequence of “Zorn’s Lemma,” an equivalent of the Axiom
of Choice. Zorn’s Lemma allows us to consider a maximal linearly independent set; such a set is
necessarily a basis, i. e., every real number can be written uniquely as a linear combination with
rational coefficients of (finitely many) members of this basis (cf. Exx. 2.1.25–2.1.28).
Now define a linear function f : R → R by assigning arbitrary values to the members of H, and
extending to all of R by linearity.
Remark. Note that the great degree of freedom in this construction shows more than the mere
existence of nontrivial solutions: it demonstrates that such solutions abound, their number is
2continuum ; whereas the cardinality of the set of all continuous functions is only continuum. It also
shows that if α1 , α2 , . . . are linearly indepedent real numbers (over Q), and β1 , β2 , . . . are arbitrary
reals, then Cauchy’s equation has a solution satisfying f (αi ) = βi for all i simultaneously.
An explicit family of linearly independent real numbers will be constructed in Ex. 2.1.41.

Figure 1.2: Squaring a regular triangle: Dudeney’s 4-piece jigsaw puzzle

1.4 Addressing into the squashed cube

In this section we consider a “graph-theoretic jig-saw puzzle.” The big picture is the complete
graph Kn ; the pieces must be complete bipartite. The edge-sets of the pieces are not allowed
to overlap.
In other words, i. e., we want to represent the set of n2 edges of Kn as the disjoint

union of the edge sets of a family of complete bipartite graphs. We shall refer to such a
decomposition as a bipartite decomposition of Kn .
The question is, what is the minimum number of parts in such a decomposition.

——————————————————————— 21

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 1. WARM-UP
A trivial solution is to view every edge as a K1,1 . This requires m = n2 pieces; we should

be able to do much better than this.

Indeed, it is easy to show that m = n − 1 pieces suffice. Just pick a vertex and join it to
the remaining n − 1 vertices. This is a complete bipartite graph K1, n − 1. The remaining
edges form a complete graph on n − 1 vertices so we can continue in the same fashion. We
end up splitting the edge set as E(Kn ) = E(K1,n−1 )∪E(K ˙ ˙
1,n−2 ) · · · ∪E(K 1,1 ). This is not
the only solution out of n − 1 pieces; indeed such solutions abound (see Ex. 1.4.5).

Figure 1.3: Three minimal decompositions of K4

Can one further reduce the number of pieces? If we drop the condition that the pieces
must be disjoint, we do get a drastic reduction: a mere dlog2 ne bipartite graphs suffice to
cover Kn .
Yet under the original conditions, n − 1 is the minimum number, as shown by an elegant
and surprising argument found by R. L. Graham and H. O. Pollak of Bell Laboratories in
1972.
Theorem 1.8 (Graham–Pollak, 1972). If the edge set of the complete graph on n
vertices is the disjoint union of the edge sets of m complete bipartite graphs then m ≥ n − 1.
Proof. Suppose the complete graph on the vertex set {1, . . . , n} has been decomposed into
the disjoint union of the complete bipartite graphs B1 , . . . , Bm . Let (Xk , Yk ) be the two parts
of the vertex set of Bk . Note that Xk and Yk are disjoint subsets of the set of vertices.
Let us associate with each Bk an n × n matrix Ak in the following way. The entry in
row i, column j will be 1 if i ∈ Xk and j ∈ Yk ; zero P otherwise.
It is clear that each Ak has rank 1. Let S = m k=1 Ak . For every i 6= j, precisely one of
T
(i, j) and (j, i) is represented in S. Therefore S + S = Jn − In where Jn is the n × n all-ones
matrix and In is the identity matrix.
We claim that if a matrix S with real entries satisfies this equation then its rank is
≥ n − 1. From this, the inequality m ≥ n − 1 follows because of the subadditivity of the
rank function (rk(A + B) ≤ rk(A) + rk(B), where A, B are n × n matrices).
For a contradiction assume that the rank of S is ≤ n − 2. In this case there exists a
nontrivial solution xT = (x1 , . . . , xn ) ∈ Rn to the system of homogeneous linear equations
n
X
Sx = 0, xi = 0.
i=1

For such an x we have Jx = 0 and therefore S T x = −x. Consequently, −kxk2 = xT S T x =

xT Sx = 0, a contradiction.

22 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.4. THE SQUASHED CUBE
* * *

Mathematical arguments of such grace don’t require further justification. Yet it will
be instructive to review how the two mathematicians at Bell Laboratories came across this
problem in the course of setting up and analysing a mathematical model for a problem in
electrical engineering.11
In the early seventies J. R. Pierce proposed a communication network consisting of one-
way loops connected at various points to one another. In order for a message to find its way
from a point on one loop to its destination on some other loop, some device was needed to
tell it “where to get off.”
Graham and Pollak suggested that each loop be treated as a vertex of a graph, and
labelled with a string of symbols from the alphabet {0, 1, ∗} in such a way that a modified
Hamming distance of the strings correspond to distances in the graph. Put another way,
Graham and Pollak’s addressing scheme allows a shortest path between two vertices of a
graph to be found without looking beyond the immediate neighborhood of each successive
vertex in the path.
The modification of the Hamming distance comes from treating 00 ∗00 as a Jolly Joker: it
can mean any symbol, therefore its distance from any symbol is zero.
Formally, let S denote the 3-element alphabet {0, 1, ∗}. Define the distance d(x, y) be-
tween these symbols by (
1 if {x, y} = {0, 1} ;
d(x, y) = (21)
0 if x or y is ∗ .
(Note that this does not even resemble a metric: different things can be at distance 0, the
triangle inequality is violated.) Extend this definition to strings of length m over S by adding
the coordinatewise distances: for x = (x1 , . . . , xm ) and y = (y1 , . . . , ym ) (x, y ∈ S m ), set
m
X
d(x, y) = d(xi , yi ). (22)
i=1

Now an addressing for an graph G = (V, E) is a distance-preserving map f : V → S m for

some m. Since the distances within S m are finite, G must be connected. Conversely, Graham
and Pollak showed that such an addressing exists for every connected graph (cf. Ex. 1.4.9).
Their addressing scheme required strings of length m = diam(G) · (n − 1), where diam(G) is
the diameter of the graph G (greatest pairwise distance between its vertices) and n = |V | is
the number fo vertices. For the sake of efficiency it is desirable to have the lengths of the
addresses as small as possible. For reasons to be explained shortly, we call this minimum
length the squashed-cube dimension of the graph G.
It was conjectured by Graham and Pollak (1972) that m = n − 1 is always sufficient.
This conjecture was confirmed a decade later by P. Winkler (1983).
To show that m = n − 1 is best possible in general, Graham and Pollak made a further
step of translation, showing that the squashed-cube dimension of the complete graph Kn is
11
The following paragraphs borrow some passages from P. Winkler’s lucid description (1983).

——————————————————————— 23

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 1. WARM-UP
precisely the minimum number of complete bipartite graphs that add up to Kn . We leave it
as Exercise 1.4.7 to show this equivalence.
Theorem 1.8 is thus a disguise for the equivalent statement that the squashed-cube di-
mension of Kn is n − 1.
Finally, we should explain the title of this section.
The 2n vertices of the unit cube are represented by the (0, 1)-strings of length n. How
can we represent an edge of this cube? An edge joins two vertices which differ in a single
coordinate, e. g. (1, 1, 0, 0) and (1, 1, 1, 0) (for n = 4). Let us write (1, 1, ∗, 0) to indicate that
the third coordinate is not determined; this is a succinct notation for this edge. Let’s now
try to describe a 2-dimensional face, e. g., the face spanned by the adjacent parallel edges
(0, 1, ∗, 0) and (1, 1, ∗, 0). It is natural to indicate this edge by (∗, 1, ∗, 0), representing the
four vertices obtained by substituting 0’s and 1’s in all possible ways for the ∗’s.
Likewise, any string of length n over the alphabet S = {0, 1, ∗} represents a collection of
vertices of the n-cube: if k of the entries are ∗’s, then we obtain a k-dimensional face of the
cube (with 2k vertices).

Figure 1.4: Squashing the cube.

Now let us squash a face F of the cube. What this means is that we collapse all vertices
of F to a single point which will be adjacent to all the previous neighbors of F .
We can also do this simultaneously with several disjoint faces; the graphs so obtained
are the squashed cubes. Now an addressing f : V → S m that preserves distances (the
distance in the graph G = (V, E) vs. the distance defined by eqn. (22)) really means a
distance-preserving map of G into a squashed cube, hence the title of the section.

Figure 1.5: The complete graph K4 as a squashed 3-cube.

For more about isometric embeddings of this and other types, we refer to Graham (1988)
and Winkler (1987).

Exercises
♦Ex. 1.4.1 (Color–critical set systems). Assume there are m clubs in a town of n inhabitants, and
each club has at least two members. We want to distribute red and blue hats among the citizens
(one hat per person) such that the hat-checkers of each club may enjoy the sight of both colors.

24 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.4. THE SQUASHED CUBE
(a) (Erdős-Hajnal, 1961) Prove: if each club has at least k members and the are no more than
2k−1 clubs then such color assignments always exist. (b) Let us call the situation critical if no
appropriate color assignment exist, but it will exist as soon as any one of the clubs closes shop.
Prove that for odd n ≥ 3, critical systems of m = n clubs exist. (c)* (P. Seymour, 1974). Prove
that every critical system of clubs satisfies m ≥ n0 where n0 is the number of citizens belonging to
at least one club.
Hint. (a) Assign the colors by flipping coins for each citizen. Show that the expected number of
clubs where the hat of each member is the same color is ≤ 1. (Cf. Erdős-Spencer (1974), Chap. 4.)
(b) Form two-member clubs. (c) Assuming m > n, use the fact that the columns of the incidence
matrix (cf. Section 1.1) of the system of clubs are linearly dependent.
♦Ex. 1.4.2. Prove Theorem 1.8 using Sylvester’s “Law of Inertia.” (This is the way the original
proof went.)
Hint. Associate a productPof two linear forms with each bipartite constituent such that these
quadratic forms add up to i<j xi xj . Represent each such product as the difference of two squares.
♦Ex. 1.4.3 (Odd cover problem). What happens if instead of requiring that the complete bipartite
graphs be disjoint, we ask that every edge of the complete graph be covered an odd number of
times? Prove that m ≥ (n − 1)/2 in this case.
Remark. The smallest possible value of m is not known.
Ex. 1.4.4. If we drop the condition that the bipartite graphs be disjoint, prove that dlog2 ne
complete bipartite graphs suffice to cover the complete graph. Prove that this many are necessary,
too.
♦Ex. 1.4.5. Prove: there are more than 2n−4 essentially different (nonisomorphic) decompositions
of the complete graph on n vertices into n − 1 disjoint complete bipartite graphs.
♦Ex. 1.4.6. Prove that the squashed cube dimension of Kn is ≤ n − 1.
♦Ex. 1.4.7 (Graham–Pollak, 1972). Prove the equivalence of Theorem 1.8 on bipartite decom-
positions of Kn and the statement that the squashed cube dimension of Kn is n − 1.
Hint. An addressing of Kn into a squashed m-cube means a function f : [n] → S m such that for
1 ≤ i < j ≤ n,
d(f (i), f (j)) = 1, (23)
where the distance function d is defined by eqn. (22). Show that such addressings are in 1–1-
correspondence with the bipartite decompositions of Kn .
Ex. 1.4.8. Show that not every connected graph is isomorphic to a squashed cube.
Hint. Two examples: the path of length two, the pentagon. (Construct isometric embeddings of
these graphs into squashed cubes!)
Ex. 1.4.9 (Graham–Pollak, 1972). Show that every connected graph has an isometric embedding
in a squashed cube.
Hint. Try the m-dimensional cube where m is the sum of pairwise distances between the pairs of
vertices of the graph.
Ex. 1.4.10** (P. Winkler, 1983). If G is a connected graph on n vertices then G has an isometric
embedding in a squashed (n − 1)-cube.

——————————————————————— 25

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 1. WARM-UP
1.5 Beauty is rare
This section illustrates the role eigenvalues play in the study of graphs displaying a high
degree of regularity. This is one of those important areas of the applications of linear algebra
to combinatorics which this volume will not treat in greater detail. Correspondingly, we
have allowed ourselves to use a little more matrix theory background than promised in the
Preface. All the later chapters will introduce or review the algebra they use in fair detail.
For basic graph theoretic terminology we refer to Section 2.4.1.
Recall that the degree of a vertex is the number of its neighbors. A graph is regular if
every vertex has the same degree. The girth of a graph is the length of its shortest cycle.
Five is a magic number, so let us consider regular graphs of girth five. What is the
minimum possible number of vertices such a graph can have, if the degree of each vertex
is r?
Take a vertex u. It has r neighbors. Each of those has r − 1 additional neighbors. This
is 1 + r + r(r − 1) = r2 + 1 vertices so far. They must all be distinct for otherwise a cycle of
length ≤ 4 would arise.
It is a natural question to ask: do we need even more vertices? Let’s see. For r = 2, we
have r2 + 1 = 5, and there is the pentagon, a single cycle of length 5. For r = 3 we would
have r2 + 1 = 10 vertices. This is again possible, as the reader may verify after a bit of
experimentation. The resulting graph is unique, and is called Petersen’s graph. This is one
of the most remarkable small objects in mathematics.

Figure 1.6: Petersen’s graph. (Which one is it?)

One may not immediately realize that it has as many as 120 symmetries. (A symmetry
or automorphism is an adjacency preserving permutation of the vertex set. The pentagon
has 10 automorphisms.) The best way to see this large degree of symmetry is the following.
Label 10 points by the ten 2-subsets of a set of five elements, and join two of them if they
correspond to disjoint 2-sets. The resulting graph is Petersen’s, and clearly all the 120
permutations of the 5-set give rise to symmetries of the graph.
Among the many remarkable properties of Petersen’s graph let us remark that although
all of its vertices are equivalent under automorphisms, it has no Hamilton cycle. Only four
connected graphs with this property are known.
If this graph turns out to have such unique properties, it may be a good idea to continue
the search along these lines. If r = 4, we would expect a graph with 17 vertices. It may take

26 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.5. BEAUTY IS RARE
quite a bit of trial and error before we would be able to convince the reader this candidate
for the beauty contest of girth-5-graphs does not exist.
Degree 5, number of vertices: 26, the next no-show. Degree 6, no better. Maybe Pe-
tersen’s graph just stays alone, in the company of the pentagon (of which it contains 12
copies).
Not quite so: in 1960, A. J. Hoffman and R. R. Singleton managed to construct a graph
of girth 5 and degree 7 with 72 + 1 = 50 vertices! This graph is again quite a miracle; it has
lots of symmetry, and consists of a large number of copies of Petersen’s graph glued together.
So perhaps there are more individuals to this species, we just need more patience, and
find clever tricks to glue many copies of the Hoffman-Singleton graph together in some neat
fashion.
While the reader contemplates this thought, let us deviate a bit from the subject. Suppose
a positive integer s satisfies the equation
ax4 + bx3 + cx2 + dx − 15 = 0, (24)
where the coefficients a, b, c, d are integers. What can we say about s?
We can say for sure that it must be 1, 3, 5, or 15. Indeed, move the number 15 to the
other side of the equation; it becomes clear, that 15 is a multiple of s.
We shall see shortly that this fact is largely responsible for the absence of further examples
of the kind of graphs we have been looking for, with the possible exception of degree 57 (3250
vertices).
Theorem 1.9 (Hoffman–Singleton, 1960). If a regular graph of degree r and girth 5
has r2 + 1 vertices, then r ∈ {2, 3, 7, 57}.
The case r = 57 is still undecided, although computers have been used to aid the search.
We start the proof with observing that such a graph G is more regular than one would
immediately suspect. Let us calculate the number of common neighbors of two vertices, u
and v. If v is a neighbor of u then they have no common neighbor since that would give rise
to a triangle. If v is not a neighbor of u then by the argument that showed that G must have
at least r2 + 1 vertices we also see that u and v must have precisely one common neighbor.
We want to put this information into matrix language.
First we associate an n × n (0, 1)-matrix with every graph.
This matrix serves to record the adjacency relation and is called the adjacency matrix.
If the vertex set of the graph G is [n] then the entries of the adjacency matrix A = (αij ) are
defined by (
1 if i and j are adjacent ;
αij = (25)
0 otherwise.
In particular, the diagonal entries of A are zero, and A is a symmetric matrix: A = AT .
It is easy to see that the entries of the matrix A2 = (βij ) count common neighbors: βij is
the number of common neighbors of vertices i and j. In particular, the diagonal entry βii is
the degree of i. Let A denote the adjacency matrix of the complement of the graph A. Then
In + A + A = Jn , (26)

——————————————————————— 27

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 1. WARM-UP
where In and Jn are the n × n identity matrix and all-ones matrix, resp. (n = r2 + 1 is the
number of vertices.)
Recall that in our graph, adjacent vertices have no common neighbors, nonadjacent
vertices have one common neighbor. In matrix form, we have

A2 = rI + A. (27)

Taking equation (26) into account, we obtain

A2 + A − (r − 1)In = Jn . (28)

A is symmetric and therefore it has an orthogonal basis consisting of eigenvectors (Principal

Axis Theorem). Let f = (1, . . . , 1)T denote a column vector of length n of all ones. It is easy
to see that for a regular graph of degree r, we have Af = rf . Hence f is an eigenvector, with
corresponding eigenvalue r. Having picked f , we may now focus on the vectors e, orthogonal
to f , i. e., f T e = 0. Then Jn e = 0, too. Assume now that e is such an eigenvector: Ae = λe.
Multiplying each side of equation (28) from the right by e we obtain

λ2 e + λe − (r − 1)e = 0; (29)

and therefore
λ2 + λ − (r − 1) = 0. (30)
This equation has two roots:
1 √
λ1,2 = (−1 ± 4r − 3). (31)
2
Let mi be the multiplicity of the eigenvalue λi (i = 1, 2). The sum of the multiplicities is n,
so, not forgetting the eigenvector f , we have

1 + m1 + m2 = n = r2 + 1. (32)

It is known that the sum of the eigenvalues is the trace of the matrix (i. e., the sum of the
diagonal elements). This gives us another equation:

r + m1 λ1 + m2 λ2 = 0. (33)

Substituting the expression (31) for the λi , we obtain

2r − (m1 + m2 ) + (m1 − m2 )s = 0, (34)

√
where s = 4r − 3. Using equation (32) this changes to

2r − r2 + (m1 − m2 )s = 0. (35)

Now s is the square root of a positive integer. So either it is a positive integer itself, or it
is irrational. In the latter case its coefficient m1 − m2 must vanish, and we are left with the
equation 2r − r2 = 0. Since r 6= 0, we infer that r = 2 which is the case of the pentagon.

28 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 1.5. BEAUTY IS RARE
Henceforth we may assume s is a positive integer. We can express r through s: r =
(s + 3)/4. Let us plug this into equation (35). After expanding r2 and multiplying by −16,
2

we obtain
s4 − 2s2 − 16(m1 − m2 )s − 15 = 0. (36)
So s satisfies an equation of the form (24). Therefore its only possible values are 1, 3, 5, and
15. From these we obtain the respective values r = (s2 + 3)/4 = 1, 3, 7, 57. We discard the
value 1 because r ≥ 2; the rest is the list of possibilities in addition to r = 2.
Beautiful graphs are rare. And so are gems like this proof.

——————————————————————— 29

c László Babai and Péter Frankl. 1988, 1992, 2020.
30 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
Chapter 2

Basic linear algebra and combinatorics

In this chapter we briefly review some of the basic structures of linear algebra and combina-
torics. The terminology we adopt is mostly standard. The reader with firm background in
these areas may skip most but not all of this chapter. The last subsection of each section is
recommended reading even for the better versed.

2.1 A guide to basic abstract algebra

This brief survey is directed toward the reader who has had some encounter with abstract
algebra but may use a little reminder. Those not in the need of such reminders are advised
to skip this section except for the last paragraphs of each subsection. The most standard
parts have been set in smaller type.

2.1.1 Fundamental structures

We start with the definitions of mod m congruences, groups, rings, and fields. No deep under-
standing of these concepts will be required since most of the time we shall work within the familiar
domains of real numbers or mod p residue classes.
As customary, Z will denote the domain of integers (after the German word “Zahl,” meaning
“number”). For a, b ∈ Z we say that a is a divisor of b, denoted by a|b, if b = ax for some x ∈ Z.
Note in particular that 0|b if and only if b = 0. For a, b, m ∈ Z, we say that a is congruent to
b modulo m, denoted by a ≡ b (mod m), if m | b − a. (When is a ≡ b (mod 1)? What does
congruence mod 0 mean?) If a ≡ b (mod m) and c ≡ d (mod m) then

a±c≡b±d and ac ≡ bd (mod m). (1)

For m > 0, Z splits into m residue classes mod m, the residue class of i ∈ Z being {k ∈ Z : k ≡ i
(mod m}). Two residue classes either coincide or are disjoint. Zm denotes the set of residue classes
mod m. Any member of a residue class is said to be a representative of that class. With a slight
abuse of notation, we shall use representatives to denote their residue classes. This will turn familiar
congruence relations such as “4 · 5 ≡ −1 (mod 7)” into statements such as “4 · 5 = −1 over Z7 .”

c László Babai and Péter Frankl. 1988, 1992, 2020.

31
CHAPTER 2. THE BASICS
Equations (1) guarantee that addition, subtraction, and multiplication of residue classes yield
the same result no matter which representatives are used.
A group is a nonempty set G together with an operation G × G → G called multiplication
((a, b) 7→ ab) or addition ((a, b) 7→ (a + b), depending on the context. The group operation is
required to be

(i) associative: (ab)c = a(bc) for every a, b, c ∈ G (or (a + b) + c = a + (b + c) in additive

notation);

(ii) G must have an identity element 1 ∈ G satisfying 1a = a1 = a for every a ∈ G (or zero
element 0 in the additive case, such that 0 + a = a + 0 = a for every a ∈ G);

(iii) every a ∈ G must have an inverse a−1 such that aa−1 = a−1 a = 1 (or an additive inverse −a
such that a + (−a) = (−a) + a = 0 when additive notation is used).
The group G is Abelian if it satisfies the commutative law:

(iv) ab = ba for every a, b ∈ G (or a + b = b + a).

Additive notation will be used for Abelian groups only, but most of our multiplicative groups
will also be Abelian. For Abelian groups we use the notation a/b = ab−1 and a − b = a + (−b),
resp. For n ∈ Z one can define an for a ∈ G in a multiplicative group and na in an additive group
by repeated multiplication (addition, resp.) and inversion if n < 0. The order of an element a ∈ G
is the smallest positive n such that an = 1 (na = 0, resp.); if no such n exists, we say that a has
infinite order. The order of a group is its cardinality. Lagrange’s Theorem asserts that for a finite
group G, the order of each element divides the order of G. When applied to the multiplicative
group of nonzero residue classes modulo a prime p, this result turns into Fermat’s (little) Theorem:
if a ∈ Zp , a 6= 0, then ap−1 = 1. (The equality is that of residue classes rather than of integers.)
A ring R is a nonempty set endowed with two operations called addition and multiplication.
With respect to addition, R is required to form an Abelian group. Multiplication must be associative
and the two operations have to satisfy the distributive laws:

(v) a(b + c) = ab + ac and (a + b)c = ac + bc (a, b, c ∈ R).

A ring is commutative if multiplication in the ring satisfies (iv).

It easily follows from the distributive laws that 0a = a0 = 0 for any a ∈ R. The element a ∈ R
is called a (left) zero-divisor if a 6= 0 and ax = 0 for some x 6= 0, x ∈ R. A commutative ring
without zero-divisors is called an integral domain, the foremost example of such rings being Z, the
ring of integers.
A ring with identity element has an element 1 6= 0 satisfying (ii). If R is such a ring, let
R denote the set of invertible elements of R. It is easy to see that R× forms a group under
×

multiplication.
A map ϕ : R → S of rings is a homomorphism if ϕ(a + b) = ϕ(a) + ϕ(b) and ϕ(ab) = ϕ(a)ϕ(b)
for every a, b ∈ R. It easily follows that ϕ(0) = 0. An isomorphism is an invertible (i. e., one-to-one,
onto) homomorphism. The rings R and S are isomorphic if an isomorphism exists between them.
A field F is a commutative ring with identity element such that F× = F \ {0}. In other words,
every nonzero element in F must have a multiplicative inverse. It follows that a field is an integral
domain. Of the rings Zm , precisely those with m = p a prime are fields (the others have zero
divisors, or, for m = 1, have only one element). To emphasize this fact, we use the notation

32 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.1. BASIC ABSTRACT ALGEBRA
Fp = Zp for prime numbers p. There exist finite fields other than these, one of each prime power
order. The unique field of order q is denoted by Fq (q a prime power). (Of course, these fields are
unique up to isomorphism only.) Fq should not be confused with the ring Zq which is not a field
unless q is a prime. Some of the exercises indicate how fields of prime power order are constructed.
For most of this book, however, familiarity with such fields is not required.
In an integral domain R of order ≥ 2, every nonzero element has the same additive order, and
this order is either infinite or some prime number p. We say that R has characteristic zero in the
former case and characteristic p in the latter. The ring Z and the fields Q, R, C (rational, real,
complex numbers, resp.) have characteristic zero; the fields Fp and Fq where q is a power of the
prime p have characteristic p. Nonzero characteristic is often referred to as finite characteristic,
as if 0 were infinity. (The way we have defined zero characteristic makes this terminology quite
natural.)
For a set Ω and a ring R, let RΩ denote the set of Ω → R functions. Under the pointwise
operations
(f + g)(ω) := f (ω) + g(ω); (f g)(ω) := f (ω)g(ω) (ω ∈ Ω),
the set RΩ forms a ring. This ring is commutative if R is, but for |Ω| ≥ 2 it will contain zero-divisors
even if R is a field.

2.1.2 Polynomials
For a commutative ring R with identity we can form the ring R[x] of formal polynomials over R.
The members of this ring are formal expressions of the form f (x) = a0 + a1 x + · · · + an xn where
ai ∈ R, and the symbol x is called an indeterminate. Two such expressions are regarded equal if
their corresponding coefficients are equal, allowing for a difference in the number of leading zero
coefficients (those on the right end). The degree of the polynomial f is n if an 6= 0. We say that
the degree of the zero polynomial (all coefficients 0) is −∞.
R[x] is a ring with identity under the natural operations. If R is an integral domain then so is
R[x]; in particular, the polynomial rings over fields have no zero divisors. For the degrees we have

deg(f + g) ≤ max{deg(f ), deg(g)}, and

deg(f g) = deg(f ) + deg(g) (2)

(the latter over integral domains only). (Here we use the natural convention −∞ + n = −∞.)
If R is an integral domain of characteristic p (p a prime or zero) then the characteristic of R[x]
is also p. In particular, for a prime p, Fp [x] is an infinite integral domain of finite characteristic.
For every a ∈ R, one can substitute a for the indeterminate x in the expression f (x) and
evaluate the result f (a) ∈ R. We say that a ∈ R is a root or zero of f ∈ R[x] if f (a) = 0. In this
case f (x) can be written as f (x) = (x − a)g(x) for some g ∈ R[x]. If R is an integral domain, it
follows that the number of roots of f does not exceed its degree (unless f is the zero polynomial).
For every a ∈ R, the map f 7→ f (a) is a homomorphism R[x] 7→ R. (Verify!) Substitution
associates a function R → R (a “polynomial function”) with every formal polynomial and thus
gives rise to another ring homomorphism: R[x] → RR . (The indeterminate x thus turns into a
variable.)
For Ω ⊆ R, the restrictions of the polynomial functions R → R to Ω are the polynomial functions
over Ω.

——————————————————————— 33

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 2. THE BASICS
We shallQoften consider polynomial functions over finite domains. If Ω ⊆ R is finite then the
polynomial ω∈Ω (x − ω) vanishes over Ω. Consequently the same polynomial function over Ω is
represented by infinitely many formal polynomials. When speaking of the degree of a polynomial
function f over Ω, we shall mean the smallest degree of the polynomials representing f . In this
context, (2) is replaced by

deg(f + g) ≤ max{deg(f ), deg(g)}, and

deg(f g) ≤ deg(f ) + deg(g) (3)

We note that such ambiguity cannot occur over infinite domains, though: if f, g ∈ R[x] and
f (a) = g(a) for infinitely many values of a ∈ R then f = g. (This follows from what was said about
the number of roots in the preceding paragraph.)
In the text we shall not make an effort to pedantically distinguish between (formal) polynomials
and polynomial functions, indeterminates and variables: the choice will always be clear from the
context.
As we have already seen in Section 1.2, one of our key tools will be polynomials in several
indeterminates. We continue to assume that R is a commutative ring with identity. Let {x1 , . . . , xn }
be a set of indeterminates. A monic monomial of degree Ps k over R is the product of k not necessarily
r1 r2 rs
distinct indeterminates, f = xi1 xi2 · · · xis , where j=1 rj = k. A monomial is a monic monomial
times a nonzero coefficient from R. The monomial αf is said to belong to the monic monomial f .
For k = 0, the product (without the coefficient) is 1. A polynomial f ∈ R[x1 , . . . , xn ] is a
finite sum of monomials. Operations are performed in the natural way (the indeterminates are
assumed to commute), thus turning the set R[x1 , . . . , xn ] into a commutative ring with identity. If
combining monomials belonging to the same monic monomial into a single term (possibly zero),
the representation of a polynomial f as a (possibly empty) sum of monomials becomes unique. The
monomials involved in this unique representation are called the expansion terms of f . The degree
of f is the maximum degree of its expansion terms. For instance, deg(3x51 x23 + 2x1 x2 x3 − 6x71 ) is 7
over Q but only 3 over F3 . The zero polynomial has degree −∞.
Again, if R is an integral domain, equation (2) remains in force. Consequently, R[x1 , . . . , xn ],
too, is an integral domain.
The polynomial f ∈ R[x1 , . . . , xn ] is homogeneous of degree k if each of its expansion terms has
degree k. According to this definition, the zero polynomial is homogeneous of degree k for every
k ≥ 0.
It is clear that every polynomial can uniquely be written as the sum of homogeneous polynomials
of different degrees. The homogeneous component of degree k of f is the sum of the expansion terms
of degree k of f .
We call a monomial multilinear if it is the product of distinct indeterminates, times a
coefficient. Sums of multilinear monomials are multilinear polynomials. Such polynomials
frequently occur in combinatorial applications. (See, e.g., Exercise 1.2.14, Theorems 5.12,
5.34.) As an example we note that the determinant, as a polynomial of n2 indeterminates, is
multilinear and homogeneous of degree n. In combinatorics, one often restricts the domain of
polynomial functions f : Rn → R to the n-cube Ω = {0, 1}n ⊂ Rn (the set of (0, 1)-vectors).
On this domain, x2i = xi for every variable xi and thus every polynomial of degree k can be
represented by a multilinear polynomial of degree ≤ k. We shall refer to this observation by
saying that every polynomial in RΩ is multilinear.

34 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.1. BASIC ABSTRACT ALGEBRA
2.1.3 Linear spaces
A linear space over the field F is an additive Abelian group V together with an operation F×V → V
of “multiplication by scalars.” The elements of V are called vectors, the elements of F scalars. The
product of λ ∈ F and v ∈ V is denoted by λv ∈ V . The following additional requirements link the
two operations:

(vi) (λµ)v = λ(µv) (λ, µ ∈ F; v ∈ V );

(vii) λ(v + w) = λv + λw (λ ∈ F; v, w ∈ V );

(viii) (λ + µ)v = λv + µv (λ, µ ∈ F; v ∈ V );

(ix) 1v = v where 1 ∈ F is the identity and v ∈ V .

It easily follows that λv = 0 if and only if either λ = 0 ∈ F or v = 0 ∈ V .

Standard examples of linear spaces over F include the following sets, each with the straightfor-
ward (componentwise, pointwise) multiplication by scalars:

– the set Fn of n-tuples over F(written as column vectors, i. e., n×1 matrices (cf. Section 2.2.3);
for typographical convenience, as below, we sometimes write the elements of Fn as row
vectors)
Fn = {(α1 , . . . , αn ) : αi ∈ F};

– the set Fk×n of k × n matrices over F, i. e., tables of the form

 
α11 α12 . . . α1n
α21 α22 . . . α2n 
A = (αij )k,n = (αij ∈ F); (4)
 
i,j=1  .. .. .. .. 
 . . . . 
αk1 αk2 . . . αkn

– the set FΩ of functions Ω → F for an arbitrary set Ω;

– the set of polynomials F[x1 , . . . , xn ];

– any extension field G of F. (This means F is a subfield of G, i. e., a subset of G closed under
the operations of G. The operations on F must be the same as those in G, restricted to F.
For instance, Q is a subfield of R, but F2 is not, because 1 + 1 = 0 in F2 while 1 + 1 6= 0 in
R.)

Let V and W be linear spaces over the same field F. A map ϕ : V → W is linear , if ϕ(a + b) =
ϕ(a) + ϕ(b) and ϕ(λa) = λϕ(a) for every a, b ∈ V and λ ∈ F. It easily follows that ϕ(0) = 0. Linear
maps are also called homomorphisms of the linear spaces. An isomorphism is an invertible (i. e.,
one-to-one, onto) linear map. The linear spaces V and W are isomorphic, if an isomorphism exists
between them.
A linear combination of the vectors v1 , . . . , vm is a vector of the form λ1 v1 + · · · + λm vm . By
a linear combination of an infinite family of vectors we mean a linear combination of some finite
subfamily. A subspace of V is a nonempty subset W , closed under linear combinations. We use

——————————————————————— 35

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 2. THE BASICS
W ≤ V to denote this. The set of all linear combinations of a set of vectors is a subspace and
is called the span of this set. The vectors v1 , . . . , vm are said to generate their span, denoted by
span{v1 , . . . , vm }. A set of generators of V is a set of which V is the span. The span of the empty
set consists of the zero vector. (Empty sum = 0.) A space is said to have finite dimension if it has
a finite set of generators.
A linear combination is trivial if all of its coefficients are zero. A linear relation among the
vectors v1 , . . . , vm is a linear combination that gives the zero vector:

λ1 v1 + · · · + λm vm = 0. (5)

This relation is nontrivial if the linear combination on the left hand side is nontrivial. The vectors
v1 , . . . , vm are linearly independent if no nontrivial linear relation exists between them. Otherwise
they are linearly dependent. An infinite family is linearly independent if each finite subfamily is.
A subfamily of an independent family is independent. An independent family cannot include the
zero vector, and no two members of an independent family are equal. A vector w depends on the
vectors v1 , . . . , vm if w ∈ span{v1 , . . . , vm }. A family is independent precisely if none of its members
depends on the rest.
If v1 , . . . , vm are linearly independent, it is easy to see that every member of their span can be
written as their linear combination in a unique way. This observation has the following consequence.

Proposition 2.1. If v1 , . . . , vm are vectors over some finite field Fq , then |span{v1 , . . . , vm }| ≤ q m .
Here, equality holds if and only if the vi are linearly independent.
A linearly independent set of generators is a basis. Every linearly independent set can be
extended to a basis, and the vectors one adds to this end may be required to belong to a given
set of generators. (For spaces of infinite dimension, this result requires some trick from set theory
such as Zorn’s lemma. Even though we do consider infinite dimensional spaces (such as spaces of
polynomials), all our actual work will be done in their finite dimensional subspaces given by explicit
finite sets of generators.
Let u1 , . . . , un be a basis of V . With every vector w = ni=1 λi vi , associate the n-tuple of its
P
coordinates with respect to this basis: [w] = (λ1 , . . . , λn ). The map w 7→ [w] is then an isomorphism
V → Fn .
The first fundamental result of linear algebra asserts that m linearly independent vectors cannot
be generated from fewer vectors:

Theorem 2.2 (The linear algebra bound). If v1 , . . . , vm are linearly independent vectors and
each of them belongs to the span of the vectors u1 , . . . , uk , then m ≤ k.
It follows immediately that all bases have equal cardinality. This number is the dimension of
the space. Theorem 2.2 is usually proved through Steinitz’s Exchange Principle. The reader may
note that over finite fields, Theorem 2.2 immediately follows from Proposition 2.1.

We observe that dim(Fn ) = n, and, as stated above, every n-dimensional space over F is isomorphic
to Fn . An immediate consequence of the linear algebra bound is the invariance of the value of n in
the definition of Fn .

Corollary 2.3 (Dimension invariance in linear algebra). If the spaces Fn and Fm are
isomorphic, then m = n.
However basic, no five-line proof of this fact is known unless F is finite.

36 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.1. BASIC ABSTRACT ALGEBRA
2.1.4 Criteria of linear independence
In order to apply the linear algebra bound in a variety of situations, we need sufficient
conditions for linear independence in concrete spaces. Here is a particularly useful one, for
the function space FΩ , where F is a field and Ω an arbitrary set.
Proposition 2.4 (Diagonal Criterion). For i = 1, . . . , m, let fi : Ω → F be functions
and ai ∈ Ω elements such that
(
6= 0, if i = j ;
fi (aj ) (6)
= 0, if i 6= j .
, fm are linearly independent members of the space FΩ .
Then f1 , . . .P
Proof. Let m i=1 λi fi be a linear relation between the fi . Substitute aj for the variable on
each side. By condition (6), all but the j th term vanish, and what remains is λj fj (aj ) = 0.
This, again by (6), implies λj = 0. Since this must hold for every j, the linear relation under
consideration was trivial.
The following slight modification of the Diagonal Criterion, which we may call its skew
version in the spirit of some of the exercises in Chapter 1, often leads to remarkable conse-
quences. (See, e.g., Theorem 5.6 and the subsequent comments.) The only difference in the
statement of the result occurs in the second line of condition (6): half the pairs (i, j) are not
covered by the corresponding condition (7).
Proposition 2.5 (Triangular Criterion). For i = 1, . . . , m, let fi : Ω → F be functions
and ai ∈ Ω elements such that
(
6= 0, if i = j ;
fi (aj ) (7)
= 0, if i < j .
Then f1 , . . . , fm are linearly independent members of the space FΩ .
Proof. For a contradiction, assume there exists a nontrivial linear relation m
P
i=1 λi fi between
the fi . Let i0 be the smallest i such that λi 6= 0. Substitute ai0 for the variable on each
side. By condition (6), all but the ith 0 term vanish, and what remains is λi0 fi0 (ai0 ) = 0. This,
again by (6), implies λi0 = 0. This contradicts the choice of i0 . The proof is complete.
The determinant of a matrix (over an arbitrary field) tells whether or not the rows are
linearly independent.
Theorem 2.6. The rows of an n × n matrix A over F are linearly independent (in the space
Fn ) if and only if det(A) 6= 0.
If this is the case, the matrix A is called nonsingular.
Theorem 2.6 implies a further generalization of our linear independence criteria for func-
tion spaces.
Proposition 2.7 (Determinant Criterion). For i = 1, . . . , m, let fi : Ω → F be functions
and ai ∈ Ω elements such that the m × m matrix A = (fi (aj ))m i,j=1 is nonsingular. Then
Ω
f1 , . . . , fm are linearly independent members of the space F .

——————————————————————— 37

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 2. THE BASICS
Proof. Assume there exists a nontrivial linear relation m
P
i=1 λi fi between the fi . Substituting
aj for every j, we obtain a nontrivial linear relation (with the same coefficients) between the
rows of A. Hence by Theorem 2.6, A is singular.

Propositions 2.4 and 2.5 are particular cases of the Determinant Criterion. Indeed, the
matrix A in those cases is diagonal and triangular, resp., with no zero entries in the diago-
nal. Therefore in both cases, det(A) is the product of the diagonal elements and therefore
det(A) 6= 0.
There is a simple further generalization of Propositions2.4 and 2.5 which does not follow
from the Determinant Criterion. For this, let T be a linear space over F and Ω an arbitrary
set. The set T Ω of Ω → T functions is a linear space over F under the natural operations.
Proposition 2.8 (Triangular Criterion, version 2). Let T be a linear space over F and
Ω an arbitrary set. For i = 1, . . . , m, let fi : Ω → T be functions and ai ∈ Ω elements such
that (
6= 0, if i = j ;
fi (aj ) (8)
= 0, if i < j .
Then f1 , . . . , fm are linearly independent members of the space T Ω .
The proof is identical with that of Proposition 2.5.

There is yet another version of the Triangular Criterion worth stating because of its
frequent occurrence in applications.
Let W and T be linear spaces over F and Ω an arbitrary set. We say that a function
f : W × Ω → T is linear in the first variable if
f (λu + µv, ω) = λf (u, ω) + µf (v, ω)
for every u, v ∈ W, ω ∈ Ω. (9)

Proposition 2.9 (Triangular Criterion, version 3). Let W, T be linear spaces over F
and Ω an arbitrary set. Let f : W × Ω → T be a function, linear in the first variable. For
i = 1, . . . , m, let wi ∈ W and ai ∈ Ω be such that
(
6= 0, if i = j ;
f (wi , aj ) (10)
= 0, if i < j .
Then w1 , . . . , wm are linearly independent.
Proof. Define the functions fi : Ω → T by setting fi (ω) = f (wi , ω) (i = 1, . . . , m; ω ∈ Ω).
By version 2 of the Triangular Criterion, the functions fi are linearly independent members
of
Pm the space T Ω . This implies the linear independence of the wi . Indeed, a linear relation
i=1 λi wi = 0 would imply the same relation for the fi :
m
X Xm
λi fi (ω) = f ( λi wi , ω) = 0,
i=1 i=1

using the condition that f is linear in the first variable.

38 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.1. BASIC ABSTRACT ALGEBRA
Exercises
Ex. 2.1.1. State versions 2 and 3 of the Diagonal Criterion.
♦Ex. 2.1.2. Determine the dimension (over the field F) of the space of homogeneous polynomials
of degree k in n indeterminates.
Hint. The answer is n+k−1


k .

♦Ex. 2.1.3. Determine the dimension (over the field F) of the space of multilinear polynomials in
n indeterminates.
Hint. The answer is 2n .

Ex. 2.1.4. Determine the dimension (over the field F) of the space of multilinear homogeneous
polynomials of degree k in n indeterminates.
Hint. The answer is nk .

Ex. 2.1.5. Consider the set of real functions {α cos(x + β) : α, β ∈ R}. Prove that this set is a
subspace of RR . Calculate its dimension.
Hint. Show that {sin x, cos x} form a basis.

* * *

Exercises 2.1.6–2.1.23 are standard basic algebra which the more advanced readers may
want to skip.
Ex. 2.1.6. (a) Let R be a commutative ring, |R| ≥ 2, and assume that for every a, b ∈ R, if a 6= 0
then the equation ax = b has a solution in R. Prove: R is a field. (b) Prove: every finite integral
domain of order ≥ 2 is a field. (c) Deduce that Zp is a field.

Ex. 2.1.7. Prove that the characteristic of integral domains, as defined in Section 2.1.1, exists,
and is either 0 or a prime number.

Ex. 2.1.8. (a) Prove that every integral domain of characteristic 0 with identity element (6= 0)
contains Z. (b) Prove that every integral domain of characteristic p with at least two elements
contains Zp . (c) Q is a subfield of every field of characteristic 0.

Ex. 2.1.9. (a) Prove: if a finite field of order q is a subfield of a finite field of order r, then r = q t
for some positive integer t. (b) Prove: every finite field has prime power order.
Hint. (a) t is the dimension of the larger field over the smaller. (b) Combine part (a) with part (b)
of the preceding exercise.

Ex. 2.1.10. (a) Prove: if q is a power of the prime p and R is an integral domain of characteristic
p then (a + b)q = aq + bq for every a, b ∈ R. (b) Deduce Fermat’s Theorem (ap ≡ a (mod p)).
Hint. Prove the case q = p first. Expand the left hand side by
the binomial theorem. Verify and
p
use the fact that for 1 ≤ k ≤ p − 1, the binomial coefficient k is divisible by p.

——————————————————————— 39

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 2. THE BASICS
Ex. 2.1.11. Construct a ring Fp [i] the way the complex numbers are constructed from the reals:
Fp [i] = {a + bi : a, b ∈ Fp }; this is a two-dimensional space over Fp ; define multiplication in the
natural way, observing the rule i2 = −1. When is the ring Fp [i] a field?
Hint. Precisely if no integer x satisfies x2 ≡ −1 (mod p) (−1 is a quadratic nonresidue mod p).
This happens if and only if p ≡ −1 (mod 4).

Ex. 2.1.12. Construct F4 .

Hint. Follow the example of the preceding exercise. Start from F2 , construct the 2-dimensional
space {a + bω : a, b ∈ F2 }. Define multiplication observing the rule ω 2 + ω + 1 = 0. Here is
the resulting multiplication table. (Addition of coefficients is performed in F2 , so for instance
ω + (1 + ω) = 1.)

F4 0 1 ω 1+ω
0 0 0 0 0
1 0 1 ω 1+ω
ω 0 ω 1+ω 1
1+ω 0 1+ω 1 ω

Figure 2.1: The multiplication table of F4

Ex. 2.1.13. (a) Prove that Lagrange’s Theorem, as stated in the main text, is equivalent to the
statement that for a finite group G of order n, every element g ∈ G satisfies g n = 1.
(b) Prove Lagrange’s Theorem for Abelian groups.
Q Q
Hint. (b) Verify: a∈G a = a∈G (ga).

Ex. 2.1.14. Use Lagrange’s Theorem to prove: in a finite field of order q, every element α
satisfies the equation αq = α.
Pn i
Ex. 2.1.15 . Define the formal derivative of the polynomial f (x) = i=1 αi x ∈ R[x] to be
0
P n i−1 0
f (x) = i=1 iαi x , setting x = 1. (R is an arbitrary commutative ring with identity. (a) Prove
the following identities: (f ± g)0 = f 0 ± g 0 ; (αf )0 = αf 0 ; (f g)0 = f 0 g + f g 0 ; and the chain rule:
f (g(x))0 = f 0 (g(x))g 0 (x). (b) Prove: if f, g ∈ R[x] and f 2 divides g then f divides g 0 . (c) Deduce:
if F is a field, g ∈ F[x], g.c.d.(g,g’)=1, then g is square free, i. e., it has no multiple roots in any
extension field G ⊇ F.
Hint. (a) Observe that in the ring R[x, y], we have f (x + y) = f (x) + yf 0 (x) + y 2 h(x, y) for some
h ∈ R[x, y]. Deduce the stated identities from this. (b) Use the product rule. (c) The multiplicity
of the root α ∈ G is the largest k such that (x − α)k divides g. Use Euclid’s algorithm to show that
the g.c.d. will not change if we switch to an extension field.
A field F is algebraically closed if every polynomial f ∈ F[x] of degree ≥ 1 has a root in F. It is
known that every field is a subfield of an algebraically closed field.

Ex. 2.1.16. Let F be an algebraically closed field of characteristic p. Prove that F contains
precisely one subfield of order pn for every n ≥ 1.
Hint. Let q = pn . The preceding exercise suggests that we have to consider the set S = {α ∈ F :
αq = α}. Prove that this set indeed is a field, i. e., it is closed under the four arithmetic operations.

40 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.1. BASIC ABSTRACT ALGEBRA
Prove that the polynomial f (x) = xq
− x is a product of linear factors over F. Prove that there
is no repetition among the roots of these factors, i. e., (x − α)2 does not divide f for any α ∈ F.
(To this end, compute (formally) the derivative of f , and, substituting α, obtain a contradiction.)
Conclude that S has precisely q elements.

Ex. 2.1.17. Let F be a field and f ∈ F[x] a polynomial of degree n. Define congruences and residue
classes mod f in the ring F[x]. Define the ring operations on the residue classes (“computing with
polynomials modulo f ”). Show that they form a ring. This ring is denoted by F[x]/(f ). Prove that
it is a linear space of dimension n over F.
A polynomial f ∈ F[x] is irreducible over F if it is not a constant polynomial (i. e., deg f ≥ 1)
and it cannot be written as f = gh for any g, h ∈ F[x] with deg g, deg h ≥ 1.

Ex. 2.1.18. (a) Prove that the ring F[x]/(f ) is an integral domain if and only if f is either a
constant or irreducible over F.
(b) Prove: F[x]/(f ) is a field if and only if f is irreducible.

Ex. 2.1.19. Let q = pn , p a prime. Show that in order to construct a field of order q, it suffices
to find an irreducible polynomial of degree n over Fp .

Ex. 2.1.20. Prove that the following polynomials are irreducible over F2 : (a) x2 + x + 1; (b)
x3 + x + 1; (c) x4 + x + 1; (d) x4 + x3 + x2 + x + 1.

Ex. 2.1.21. Prove that the following polynomials are reducible over F2 : (a) x2 + 1; (b) x6 + x4 +
x2 + 1; (c) x5 + x4 + 1.

* * *

Ex. 2.1.22. Construct an infinite field of finite characteristic.

Hint. See the next exercise.

Ex. 2.1.23. (a) Consider the formal fractions of the form f /g : f, g ∈ F[x], g 6= 0. View two
such fractions f1 /g1 and f2 /g2 equal if f1 g2 = f2 g1 . Prove that under this equivalence relation,
the equivalence classes form a field. This field is denoted by F(x) and is (confusingly) called the
field of rational functions over F. (The elements of F(x) are not functions; it is possible, that the
denominator of a fraction vanishes if we attempt to substitute values from F.) (b) Deduce that
every field is a subfield of an infinite field.
1
Ex. 2.1.24. Prove: in the space of rational functions over R, the set { x−α : α ∈ R} is linearly
independent. Prove the same over any field.
For the following five problems, we assume familiarity with the notions of countable and uncountable
infinite cardinals and Zorn’s Lemma.

Ex. 2.1.25. Prove, using Zorn’s Lemma, that every (infinite dimensional) linear space has a basis.
Ex. 2.1.26. Prove (for infinite dimensional spaces), that all bases have equal cardinality.
Ex. 2.1.27. Prove: the dimension of the polynomial space R[x] is countably infinite.

——————————————————————— 41

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 2. THE BASICS
Ex. 2.1.28. Prove: the dimension of the space of rational functions R(x) is uncountable.
Ex. 2.1.29. Recall that R is a linear space over Q. Prove that its dimension is (a) infinite; (b)
uncountable.

* * *

The following sequence of exercises provides an introduction to Algebraic Number Theory. The
problems link the notions of fields, polynomials, and linear spaces and cover classical material. The
reader not familiar with these results may be well advised to spend some time working them out,
although the rest of the book will not rely on their factual contents.
Recall that if F is a subfield of a field G, then G is a vector space over F. The dimension of G
over F is denoted by (G : F).
√ √
Ex. 2.1.30. Prove that the real numbers 1, 2, 3 are linearly independent over Q.
√ √ √
Ex. 2.1.31. Prove that the real numbers 1, 2, 3, 6 are linearly independent over Q.
Ex. 2.1.32. Construct a field G such that Q ⊂ G ⊂ R and (a) (G : Q) = 2; (b) (G : Q) = 3.
Ex. 2.1.33. Prove: if F ⊆ G ⊆ H are fields then
(H : F) = (H : G)(G : F).

Let F ⊆ G be a field extension. An element α ∈ G is algebraic over F if f (α) = 0 for some nonzero
polynomial f ∈ F[x]. Such a polynomial of minimal degree and leading coefficient 1 is called the
minimum polynomial of α over F. (Prove: this polynomial is unique.) The degree of α over F is
degF (α) := deg(mα ) where mα ∈ F[x] is the minimum polynomial. If α is not algebraic then it is
transcendental .

Ex. 2.1.34. Prove: if (G : F) < ∞ then every element of G is algebraic over F.

Let F ⊆ G and α ∈ G. The ring extension of F by α is the set F[α] = {f (α) : f ∈ F[x]}.

Ex. 2.1.35. Prove: F[α] is a linear space over F. If α is transcendental over F then dim F[α] = ∞.
If α is algebraic over F then dim F[α] = degF (α).

Ex. 2.1.36. Prove: if α is algebraic over F then the ring F[α] is a field.
Ex. 2.1.37. Prove: if (G : F) = k and α ∈ G then degF (α)| k.
Ex. 2.1.38. Let F ⊂ G be a field extension. Let A denote the set of those α ∈ G which are
algebraic over F. Prove: A is a field. In particular, the set of algebraic numbers (i. e., those
complex numbers which are algebraic over Q) form a field.
Hint. In order to prove that A is closed under arithmetic operations, take α, β ∈ A, and consider
the sequence of field extensions F ⊆ F[α] ⊆ F[α][β]. Prove that (F[α][β] : F) < ∞. Conclude that
α ± β, αβ, αβ −1 are algebraic over F.

Ex. 2.1.39. Using the notation of the previous exercise, assume some β ∈ G is algebraic over A.
Prove, that β ∈ A.

42 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.2. AFFINE SPACES
Hint. In order to show that β is algebraic over F, let its minimal polynomial over A be f (x) =
α0 + α1 x + · · · + αn xn ∈ A[x]. Consider the following tower of fields:

F ⊆ F[α0 ] ⊆ F[α0 ][α1 ] ⊆ · · · ⊆ F[α0 ][α1 ] · · · [αn ] ⊆ B,

where B = F [α0 ][α1 ] · · · [αn ][β]. Like in the previous exercise, prove that (B : F) < ∞.

Ex. 2.1.40.** Let G be a subfield of C such that (C : G)< ∞. Prove that (C : G) is either 1 or
2.
♦Ex. 2.1.41* (A. S. Besicovitch, 1940). √ √ √
Prove that the numbers 1, p1 , p2 , . . . , pk are linearly
independent over Q, where the pi are distinct prime numbers.
Hint. An integer is square free if it is not divisible by the square of any prime. Let p1 , . . . , pk be
pairwise relatively prime square free positive integers. Prove, by induction on k, that
√ √
Lk := Q[ p1 , . . . , pk ] = 2k . (11)

Observe that this yields the stronger statement that the set of the products of each of the 2k subsets
of the pi is linearly independent (since this set spans Lk as a linear space over Q).

Ex. 2.1.42.* Assume Q is a subfield of F and F contains the square roots of k distinct prime
numbers. Prove: (F : Q) ≥ 2k .

Ex. 2.1.43. Prove:

√ if F is the smallest extension field of Q containing the square roots of the first
3
k primes then 2∈
/ F.

Ex. 2.1.44 (M. Wierdl, B. Bajnok, 1989). Let 1 < n1 < n2 < . . . be a sequence of integers such
that log n1 , log n2 , . . . are linearly independent over Q. Prove that limk→∞ nk /k = ∞.
Hint. Prove that if pk is the kth prime number then nk ≥ pk .

Ex. 2.1.45 (M. Wierdl, 1989). Let α ∈ R. Prove that for almost every α, the set {1α , 2α , 3α , . . .}
is linearly independent over Q. In this statement, “almost all” is defined to mean that the set of
exceptional values of α is countable.

2.2 Affine subspaces, linear equations, rank

2.2.1 Inequalities for subspaces
For two subsets U, V ⊆ W , we use the notation

U + V = {u + v : u ∈ U, v ∈ V }. (12)

If U and V are subspaces then we have

U + V = span{U, V }. (13)

The codimension of a subspace U ≤ W is the minimum number of vectors that together

with U generate W . This quantity is denoted by codimW (U ) or simply by codim(U ) if the

——————————————————————— 43

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 2. THE BASICS
universe W is clear from the context. (Warning: often it is not.) If W has finite dimension
n then
codim(U ) = n − dim(U ). (14)
For subspaces U, V ≤ W , the modular identities hold :

dim(U + V ) + dim(U ∩ V ) = dim(U ) + dim(V ); (15)

codimW (U + V ) + codimW (U ∩ V ) =
codimW (U ) + codimW (V ); (16)

codimU +V U = codimV (U ∩ V ). (17)

The following inequalities are immediate consequences:

dim(U + V ) ≤ dim(U ) + dim(V ); (18)

codimW (U ∩ V ) ≤ codimW (U ) + codimW (V ); (19)

codimV (U ∩ V ) ≤ codimW (U ). (20)
If dim(U ) = r, codimW (V ) = t, then inequality (20) can be rewritten as

dim(U ∩ V ) ≥ max{r − t, 0}. (21)

What this inequality says is that intersecting with a subspace of codimension t cannot reduce
the dimension of a subspace by more than t.
The rank and corank of a set S ⊆ W of vectors are defined as

rk(S) = dim (span(S))

corankW (S) = codimW (span(S)) . (22)

The rank is equal to the maximum number of linearly independent vectors from S.
For subsets S, T ⊆ W , equation 15 implies the submodular inequality

rk(S ∪ T ) + rk(S ∩ T ) ≤ rk(S) + rk(T ). (23)

2.2.2 Linear maps

For the linear spaces W , T over the field F, linear maps ϕ : W → T have been defined in
Section 2.1.3. For any basis e1 , . . . , en of W and every collection of vectors t1 , . . . , tn ∈ T
there exists a unique linear map ϕ : W → T such that ϕ(ei ) = ti (i = 1, . . . , n).
The set of images under ϕ is the set

im(ϕ) = {ϕ(w) ∈ T : w ∈ W }; (24)

44 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.2. AFFINE SPACES
the preimage of zero is the kernel

ker(ϕ) = {w ∈ W : ϕ(w) = 0}. (25)

The set im(ϕ) is a subspace of T ; and ker(ϕ) is a subspace of W . The following fundamental
relation links their dimensions:

dim ker(ϕ) + dim im(ϕ) = n, (26)

where n = dim W . The dimension of im(ϕ) is called the rank of ϕ, denoted by rk ϕ.

Equation 26 can be strengthened to

rk ϕ := dim im(ϕ) = codim ker(ϕ). (27)

(If W has finite dimension, (26) and (27) are equivalent.)

The next observation shows that equation (27) is the only relation the kernel and the
image must satisfy.
Proposition 2.10. Given a subspace U ≤ W , and a space T such that dim T = codimW U ,
there exists a linear map ϕ : W → T such that ker(ϕ) = U and im(ϕ) = T .
Proof. (Although we formulate the proof for finite dimension, it can be translated to any
space with virtually no change.) Let e1 , . . . , ek be a basis Pnof U . Extend this set to a basis
of W . Let T0 = span{ek+1 , . . . , en }. For w = i=1 αi ei ∈ W , define ϕ0 (w) to be
e1 , . . . , enP
ϕ0 (w) = ni=k+1 αi ei ∈ T0 . This map is linear; ker(ϕ0 ) = U ; and im(ϕ0 ) = T0 . Combine ϕ0
with an isomorphism T0 → T to obtain the desired map ϕ : W → T .

2.2.3 Matrices, rank

Let R be a commutative ring with identity. (Most often, R will be either a field, or the ring
of polynomials in one or several variables over a field.) Let A = (αij )k,n
i,j=1 ∈ R
k×n
be the
n n×1
k × n matrix over R shown in (4) (now αij ∈ R). We set R = R (column vectors); e. g.,
 
α1
 α2 
v =  ..  ∈ Rn .
 
 . 
αn

For convenience, sometimes we use Rn for R1×n , the set of row vectors as well.
The transpose of this matrix, AT = (βij )n,k
i,j=1 ∈ R
n×k
is the n × k matrix with βij = αji .
In particular, the transpose of the column vector v is the row vector v T = (α1 , α2 , . . . , αn ) ∈
R1×n .
Submatrices are called minors. An k 0 × n0 minor of A is obtained by selecting k 0 rows
and n0 columns of A and arranging the k 0 n0 entries in the natural way into a k 0 × n0 matrix.
Matrix operations are performed in the usual manner over any commutative ring.

——————————————————————— 45

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 2. THE BASICS
The column space of a k × n matrix A (as in (4)) over a field F is the subspace of Fk
spanned by the columns of A. The column rank of A is the rank of its set of columns, i. e.,
the dimension of its column space. A is said to have full column rank if this quantity is equal
to n, the number of columns, i. e., if the columns are linearly independent. The row space,
row rank, and full row-rank are defined analogously. One of the first nontrivial results in
matrix theory asserts that the row and column ranks are equal; this common value is the
rank of A, denoted by rk A or rkF A if the field needs to be specified. Another way of stating
this result is that
rk A = rk AT , (28)
where rk stands for column rank. A k × n matrix A has full rank if

rk A = min{k, n}. (29)

It is important to note that the rank is not sensitive to field extensions.

Lemma 2.11 (Rank Insensitivity Lemma). If F is a subfield of G and A ∈ Fk×n , then

rkF A = rkG A. (30)

In particular, for matrices A over Z, the rank rkF depends on the characteristic of F only. If
charF = p (prime or zero), we can write rkp A for rkF A.
Equations (28) and (30) are immediate consequences of the determinant characterization
of the rank: if A has a nonsingular r × r minor but every (r + 1) × (r + 1) minor is singular
then rk(A) = r. (Recall that a square matrix is singular if its determinant is zero.)
The following inequalities hold for the rank:

rk(A) − rk(B) ≤ rk(A + B) ≤ rk(A) + rk(B); (31)

rk(AB) ≤ min{rk A, rk B}. (32)

Inequality (31) is a consequence of inequality (18). Inequality (32) follows from the obser-
vation that the column space of AB is a subspace of the column space of A; and the row
space of AB is a subspace of the row space of B.
Let us define the corank of a matrix A ∈ Fk×n as the codimension of its column space,
i. e.,
corank(A) = k − rk A (A ∈ Fk×n ). (33)
We leave it to the reader to verify that

corank(AB) ≤ corank(A) + corank(B). (34)

2.2.4 Systems of linear equations. Affine subspaces

If x1 , . . . , xn denote indeterminates, x = (x1 , . . . , xn )T is the column vector they form, and
b ∈ Fk , then the matrix equation
Ax = b (35)

46 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.2. AFFINE SPACES
is a concise form of writing a system of k linear equations in the variables x1 , . . . , xn :

αi1 x1 + αi2 x2 + · · · + αin xn = βi (i = 1, . . . , k).

Let a1 , . . . , an denote the columns of A. Observe that

Ax = x1 a1 + x2 a2 + · · · + xn an . (36)

It follows that for A ∈ Fk×n , the set {Ax : x ∈ Fn } is the column space of A. The system
(35) of linear equations is thus solvable if and only if b ∈ span{a1 , . . . , ak }, or equivalently,
if and only if
rk A = rk [A|b], (37)
where [A|b] denotes the k × (n + 1) matrix obtained by adding the column b to A.
The matrix A ∈ Fk×n determines an Fn → Fk linear map (also denoted by A), defined
by x 7→ Ax for x ∈ Fn . After fixing bases in W and T , thereby identifying them with Fn
and Fk , resp. (n = dim(W ) and k = dim(T )), every linear map ϕ : W → T takes the form
x 7→ Ax for some A ∈ Fk×n .
The image of the map x 7→ Ax being the column space of A, we have the relation
rk A = dim(im(A)), in agreement with the leftmost part of equation (27). The system
Ax = b is solvable if and only if b ∈ im(A).
If b = 0, the system (35) is called a system of homogeneous linear equations. The set of
solutions of Ax = 0 is clearly a subspace: it is the kernel of the map A. By equation (26),
its dimension is n − rk A. We summarize this result.
Proposition 2.12. For A ∈ Fk×n , the set of solutions of the system of linear equations
Ax = 0 is a subspace of dimension n − rk A of the space Fn .
Remark 2.13. It follows from Proposition 2.10 that every subspace of dimension k of F n
coincides with the set of solutions of some system of (n − k) homogeneous linear equations.

Definition
Pm 2.14. An affine combination
Pm of the vectors v1 , . . . , vm ∈ W is a linear combina-
tion i=1 λi vi (λi ∈ F) where i=1 λi = 1.

An affine closed set in W is a subset closed under affine combinations. We allow this set
to be empty. The affine hull of a set S ⊆ W is the set of affine combinations of all finite
subsets of S. This set is denoted by aff(S); it is affine closed. Note that while the span of
the empty set is the zero subspace, the affine hull of the empty set is empty.
Specializing the notation for the sum of two subsets, introduced at the beginning of
Section 2.2.1, for S ⊆ W and u ∈ W we use the notation S + u = {s + u : s ∈ S}. We call
S + u the translate of S by u.
Proposition 2.15. (a) The translate of a subspace is affine closed.
(b) An affine closed set is a subspace if and only if it contains the origin (i. e., the zero
vector).

——————————————————————— 47

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 2. THE BASICS
(c) Every affine closed set is either empty or the translate of a subspace. This subspace
is uniquely associated with the set.
(d) The set of solutions of the system of linear equations Ax = b is affine closed. It is
either empty or of the form U + x0 where U = ker(A) is the solution space of the system
Ax = 0 of homogeneous linear equations; and x0 is an arbitrary solution of Ax0 = b.
The span of a set is also called its linear hull . The notions of affine and linear hulls are close
relatives.
Proposition 2.16. (a) The span of the set S ⊆ W is the affine hull of S ∪ {0}.
(b) The affine hull of a set S ⊆ W is empty if S is empty; otherwise for any u ∈ S,
aff(S) = u + span(S − u). (Here S − u means the translate of S by −u.)
Affine closed sets are also called affine subspaces. Subspaces in the sense used so far may be
called linear subspaces if there is need to emphasize the contrast. The dimension of an affine
subspace is the dimension of its corresponding linear subspace (of which it is a translate).
The dimension of the empty set is −1. Hence the dimension of the solution set of the system
Ax = b of equations is either −1 or n − rk A. We summarize the qualitative information we
obtained about the solution sets.
Proposition 2.17. (a) The system Ax = b of linear equations is solvable if and only if
rk A = rk [A|b].
(b) If solvable, the set of solutions of the system Ax = b is an affine subspace of dimension
n − rk A in Fn .
One more bit on affine vs. linear subspaces.
Proposition 2.18. The intersection of a family of affine subspaces is either empty or equal
to a translate of the intersection of their corresponding linear subspaces.
Definition 2.19. The vectors v1 , . . . , vm ∈ W are affine independent if for every
λ1 , . . . , λm ∈ F, the two conditions
m
X m
X
λi vi = 0 and λi = 0
i=1 i=1

imply λ1 = · · · = λm = 0.

The vectors v1 , . . . , vm ∈ W are affine independent precisely if none of them belongs to

the affine hull of the others. Note that any single vector (including the zero vector) is affine
independent and affine closed at the same time.
Proposition 2.20. (a) The vectors v1 , . . . , vm ∈ W are linearly independent if and only if
the family 0, v1 , . . . , vm is affine independent.
(b) The vectors v1 , . . . , vm ∈ W are affine independent if and only if v2 − v1 , . . . , vm − v1 ∈
W are linearly independent.
Definition 2.21. An affine basis of an affine subspace U is an affine independent set S such
that aff(S) = U .

48 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.2. AFFINE SPACES

Various properties of affine bases follow from Proposition 2.20. We state one.
Proposition 2.22. Every affine basis of the affine subspace U has 1 + dim(U ) elements.

Corollary 2.23. If U1 , . . . , Um are affine subspaces, then

m
X


dim aff{U1 , . . . , Um } ≤ (m − 1) + dim(Ui ).
i=1

Remark 2.24. Affine spaces, just as linear spaces, can be treated axiomatically. In an
affine space, affine linear combinations are the basic operation. Rather than stating a set
of natural axioms they are postulated to satisfy, we make some statements that give a full
characterization of affine spaces via linear spaces. A linear space (and any of its affine
subspaces) is an affine space; and conversely, a nonempty affine space can be transformed
into
Pm a linear space by declaring any of its points to be the origin. (Then a linear combination
i=1 λi vi will be defined by
Pturning it into an affine combination by adding the origin (“zero
vector”) with coefficient − m i=1 λi .) Affine spaces can thus be thought of as just being linear
spaces from an affine viewpoint. So if speaking of subspaces of an affine space, we mean its
affine subspaces; independence refers to affine independence; there is no specified origin, etc.

2.2.5 Projective spaces

Let W be a linear space of dimension n over F. Let W × = W \ {0}. Let ρ be the equivalence
relation on W × defined by

xρy if and only if span{x} = span{y}. (38)

The equivalence classes (“rays”) are the points of the projective space W . We define the
dimension of W to be n − 1, and for W = Fn , we denote W by P n−1 F. Let us denote the
equivalence class of x ∈ W × by [x]. Clearly,
 n−1  q n − 1
P Fq  = . (39)
q−1

We define the subspaces of W to be the sets of the form U for all subspaces of W of
dimension ≥ 1. A hyperplane is a subspace of codimension 1, i. e., dimension n − 2 in our
case. The one-dimensional subspaces of W are called lines. (Note that they correspond to
2-dimensional subspaces of W .) A set of points is collinear if they belong to the same line.
Each line of P n−1 Fq has q + 1 points (q + 1 = (q 2 − 1)/(q − 1)).
The span of a subset S ⊆ W is the subspace corresponding to the span (in W ) of
representatives of each equivalence class [x] ∈ S. A minimal set spanning a subspace is its
basis. A basis of size k spans a subspace of dimension k − 1 (like in the affine case).

——————————————————————— 49

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 2. THE BASICS
A set of k ≥ n points is in general position (cf. Section 3.1) if no n of them belong to a
hyperplane.
Isomorphisms W → T of the linear spaces W and T induce projective transformations
W → T . (Note that a projective transformation is always a bijection by definition. The
reason is that 0 ∈ W does not correspond to anything in W . The Fundamental Theorem of
Projective Geometry asserts that any two (n + 1)-tuples in general position are equivalent
under projective transformations.

Theorem 2.25 (Fundamental Theorem of Projective Geometry). Let W and T be

two n − 1-dimensional projective spaces. Let w1 , . . . wn+1 ∈ W be in general position, and
similarly, t1 , . . . , tn+1 ∈ T . Then there exists a projective transformation σ : W → T such
that σ(wi ) = ti for all i.
Proof. Let wi = [xi ] and ti = [yi ] (xi ∈ W , yi ∈ T ). Then x1 , . . . , xn form a basis of W ;
therefore there exists a linear isomorphism ϕ : W → T such that ϕ(xi ) = yi for i = 1, . . . , n.
We may thus assume W = T and xi = yi for i = 1, . . . n. Now we need a linear transformation
ψ such that ψ(xi ) ∈ [xi ] for i = 1, . . . , n (so thePcorresponding projective
Pn transformation fixes
n
these wi ), and ψ(xn+1 ) = yn+1 . Let xn+1 = i=1 αi xi and yn+1 = i=1 βi xi . The general
position condition implies that none of the αi and βi are zero. Let us now consider the
transformation ψ defined by ψ(xi ) = (βi /αi )xi for i = 1, . . . , n. (This defines ψ uniquely.)
It is easy to check that ψ(xn+1 ) = yn+1 indeed.
Sometimes we declare a hyperlane H of W to be the hyperplane at infinity. The set
W \ H can be identified with an (n − 1)-dimensional affine space in the following way. Using
an appropriate projective transformation, we may identify W with P n−1 F and H with the
subspace corresponding to the equation xn = 0. Now, with every [x] ∈ P n−1 F \ H we
associate the unique vector x0 ∈ Fn−1 such that (x0 , 1) ∈ [x]. (To obtain x0 , we normalize x
by dividing it by its last coordinate; and then omit the last coordinate.) Note that for each
subspace U of W , the set U \ H becomes an affine subspace of Fn−1 , and all affine subspaces
arise (uniquely) in this way.

2.2.6 Extending the field

In the next chapter we shall introduce the notion of “general position,” a technique we
shall often rely on but which only works over large fields. We shall normally overcome this
difficulty by extending the field to an infinite one. In this context it is significant to formalize
that our operations on subspaces are insensitive, in several ways, to field extensions. We
have noted the corresponding fact about matrix rank (Lemma 2.11); everything else will be
a direct consequence.

Let F be a subfield of G. Then Fn is a subset (not a subspace) of Gn . For S ⊆ Fn , we use

spanG (S) and affG (S) to denote the linear and affine hulls, resp., of S in Gn .
First we observe that a set of vectors in Fn is linearly independent in Fn if and only if
it is linearly independent in Gn . This is essentially a restatement of Lemma 2.11. As an

50 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.2. AFFINE SPACES
immediate corollary, we have the following.
Proposition 2.26. (a) If U is a subspace of Fn , then
dimG spanG U = dimF U ; (40)
(b) If U is an affine subspace of Fn , then
dimG affG U = dimF U. (41)
An affine subspace can be specified as the affine hull of a set. A dual specification is by
a system of linear equations. We shall see the same insensitivity there.
Lemma 2.27. Let A ∈ Fk×n and b ∈ Fk . Let S(F) and S(G) denote the solution sets of
the system Ax = b of linear equations in Fn and in Gn , resp. Then
S(G) = affG S(F). (42)
Proof. Part (a) of Proposition 2.17 characterizes the solvability of the system Ax = b in
terms of matrix rank, which by Lemma 2.11, does not change when switching from F to G.
So we may now assume that neither side of equation (42) is empty. S(G) being affine closed,
it is clear that the left hand side of (42) contains the right hand side. Therefore it suffices
to see that their dimensions are equal. But this is immediate by part (b) of Propositon 2.17
combined with Lemma 2.11.
Corollary 2.28. Let U1 , . . . , Um be affine subspaces of Fn . Then
m
! m
\ \
affG Ui = (affG (Ui )) . (43)
i=1 i=1

Note that for subsets containing 0, the affine and linear hulls are equal. Therefore equa-
tion (43) remains valid if the Ui are subspaces and “aff” is replaced by “span” in the state-
ment.
Proof. Let Ui be the solution set of a system Ai x = bi of ki P linear equations. (Cf. Remark
2.13.) Let us combine these to a single system Ax = b of m i=1 ki equations. The set of
solutions of the large system is the intersection of the Ui . The solution set over G is then,
by the previous Lemma, the left hand side of equation (43). On the other hand, the set of
solutions of Axi = bi over G is affG Ui . Consequently the set of solutions of Ax = b over G is
the right hand side of equation (43).
The following consequence will be used in the field extension arguments in Chapters 5 and
6.
Lemma 2.29 (Field Extension Lemma). If U1 , . . . , Um are affine subspaces of Fn , then
m
! m
!
\ \
dimG affG (Ui ) = dimF Ui . (44)
i=1 i=1

Again, we note that the Lemma includes as a special case the situation when the Ui are
linear subspaces and “aff” is replaced by “span.”
Proof. Combine Corollary 2.28 and Proposition 2.26.

——————————————————————— 51

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 2. THE BASICS
Exercises
♦Ex. 2.2.1. Let A be a k × n matrix over a field F. When is it true that the system of linear
equations Ax = b is solvable for every b ∈ Fk ?
Hint. The necessary and sufficient condition is that rk(A) = k, i. e., the rows of A are linearly
independent.

Ex. 2.2.2. Prove the subspace inequalities and identities (12)–(23).

Ex. 2.2.3. Prove Propositions 2.12, 2.15, 2.17, and inequality (34).
♦Ex. 2.2.4. Prove the rank inequalities (31) and (32).
♦Ex. 2.2.5. Interpret the statement that Figure 2.2 (a) represents the 2-dimensional projective
space over F2 , called the Fano plane.

Figure 2.2: (a) The Fano plane. (b) The anti-Fano configuration.

♦Ex. 2.2.6. Let S be a subset consisting of seven points of a 2-dimensional projective space over
the field F. Consider the family of those triples of S which are collinear. Prove: (a) If these triples
form the seven lines of the Fano plane (Figure 2.2 (a)) then the characteristic of F is 2. (b) If these
triples form the six lines of the anti-Fano configuration (Figure 2.2 (b)) then the characteristic of
F is 6= 2.
Hint. Use the Fundamental Theorem of Projective Geometry to move 4 points in general position
into the positions with coordinates (1, 0, 0), (0, 1, 0), (0, 0, 1), and (1, 1, 1). Determine the coordi-
nates of the remaining points, making no reference to the “circle” in the figure. Draw the conclusion
based on whether or not the 3 points on the “circle” should be collinear.

2.3 Orthogonality
2.3.1 Inner product spaces
A bilinear form over a linear space W is a map β : W × W → F such that β is linear in each
variable:
β(λu + µv, w) = λβ(u, w) + µβ(v, w);
and
β(w, λu + µv) = λβ(w, u) + µβ(w, v)

52 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.3. ORTHOGONALITY
for every λ, µ ∈ F and u, v, w ∈ W . The form β is symmetric if

β(u, v) = β(v, u)

for every u, v ∈ W .
For W = Fn , every bilinear form can be written as

β(u, v) = uT Bv

for some n × n matrix B ∈ Fn×n , uniquely associated with β; and every map (u, v) 7→ uT Bv
is a bilinear form. The form uT Bv is symmetric if and only if the matrix B is symmetric:
B = BT .
If β is a symmetric bilinear form over W , we call the pair (W, β) an inner product space,
with β the inner product.
The most common example is the standard inner product over Fn , defined by the identity
matrix:
Xn
T T
u · v = u In v = u v = ξi ηi ,
i=1

where u = (ξ1 , . . . , ξn )T and v = (η1 , . . . , ηn )T (u, v ∈ Fn ).

For the rest of this section, β will denote an arbitrary inner product over the n-dimensional
space W which we may identify with Fn . Let B be the associated symmetric matrix.
The vectors u, v are called perpendicular (or orthogonal ) if their inner product is zero.
The notation for this is u ⊥ v. This is clearly a symmetric relation. For a subset S ⊆ W ,
we define the perpendicular space of S as

S ⊥ = {v ∈ W : β(u, v) = 0 for every u ∈ S}.

The symbol ⊥ is pronounced “perp.” Two subsets S, T ⊆ W are called perpendicular (S ⊥ T )

if u ⊥ v for every u ∈ S, v ∈ T . This is clearly equivalent to the relation S ⊆ T ⊥ .
It is easy to see that S ⊥ is a subspace for any subset S ⊆ W . It follows that

S ⊥ = (span(S))⊥ . (45)

Another easy observation:

if S ⊆ T ⊆ W then T ⊥ ≤ S ⊥ ≤ W.

It follows that
S ⊆ S ⊥⊥
for every S ⊆ W .
A nonzero vector w ∈ W is isotropic if w ⊥ w. Otherwise w is anisotropic. A subspace
U ≤ W is isotropic if it contains an isotropic vector; U is totally isotropic if U ⊥ U , i. e.,
every pair of vectors in U is perpendicular. (Equivalently, U ≤ U ⊥ .)

——————————————————————— 53

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 2. THE BASICS
The radical of a subspace U is its intersection with its perp:

rad(U ) = U ∩ U ⊥ .

The subspace U is singular if rad(U ) 6= 0; and nonsingular otherwise. We call the inner
product singular or nonsingular according to whether or not W itself is singular.
Proposition 2.30. (a) For every subspace U ≤ W ,

dim(U ) + dim(U ⊥ ) ≥ n (n = dim(W )).

(b) If W is nonsingular then for every subspace U ≤ W ,

dim(U ) + dim(U ⊥ ) = n.

(c) The space W is nonsingular if and only if the matrix B is nonsingular.

Proof. Let u1 , . . . , uk be a basis of U . Then x ∈ W belongs to U ⊥ precisely if x satisfies the
system of k homogeneous linear equations

uTi Bx = 0 (i = 1, . . . , k). (46)

The rank of this system is at most k, therefore its solution space U ⊥ has dimension ≥ (n−k).
If B is nonsingular, then the vectors u1 B, . . . , uk B are linearly independent and therefore
the rank of the system (46) is exactly k, and so we have dim(U ⊥ ) = n − k. It follows that
if B is nonsingular then so is W (we just saw that dim(rad(W )) = dim(W ⊥ ) = n − n = 0
in this case). On the other hand, if B is singular, then Bu = 0 for some u ∈ W , u 6= 0.
Clearly, this u belongs to rad(W ). This completes the proof.
Indeed, we note that rad(W ) = ker B.
Corollary 2.31. In a nonsingular inner product space of dimension n, every totally isotropic
subspace has dimension ≤ bn/2c.
Proof. Combine the two facts that dim(U ) + dim(U ⊥ ) = n and U ≤ U ⊥ .

2.3.2 Eventown revisited

Recall that a family F = {F1 , . . . , Fm } of distinct subsets of the set [n] = {1, . . . , n} satisfies
the Eventown Rules (Section 1.1) if |Fi ∩ Fj | is even for every i, j = 1, . . . , m (including the
cases i = j). The first question considered in this book was the maximum possible value of
m for a given n. We saw in Section 1.1 that the value m = 2bn/2c is easily attained (“married
couples” solution). Now we are in the position to show, as a corollary of the results of the
previous section, that this number is best possible.
Theorem 2.32 (Eventown Theorem). If a family of m subsets of a set of n elements
satisfies the Eventown Rules then m ≤ 2bn/2c .

54 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.3. ORTHOGONALITY
A stronger version of this result, due to E. R. Berlekamp and J. E. Graver, appears among
the exercises (Ex. 2.3.11).
Proof. Fn2 , together with the standard inner product, is a nonsingular inner product space.
Let S denote the set of incidence vectors of the Fi ; this is a subset of Fn2 . The Eventown
Rules translate to the condition S ⊥ S, or, in other words, S ⊆ S ⊥ . Let U = span(S). Then,
by equation (45), U ≤ U ⊥ = S ⊥ , that is, the subspace U is totally isotropic. Therefore, by
Corollary 2.31, dim(U ) ≤ bn/2c. We conclude that |S| ≤ |U | ≤ 2bn/2c .
One difference between Oddtown and Eventown is that in Eventown, an exponentially large
number of clubs are admissible. Another important difference between the nature of the
two sets of rules is that in Eventown, every maximal system of clubs is maximum. In other
words, if no club can be added to the existing system, then we already have the maximum
possible number of clubs. (This maximum number has just been established.) We shall
prove this assertion, which the reader should contrast with the Oddtown situation discussed
Exercise 1.1.11.
In view of the observations made in the proof of the Eventown Theorem, we see that the
incidence vectors of a maximal system of Eventown clubs form a maximal totally isotropic
subspace of Fn2 . All we have to prove is that such a subspace necessarily has dimension
bn/2c.
Theorem 2.33. Every maximal totally isotropic subspace of Fn2 has dimension bn/2c.
Here and elsewhere, if we do not specify the inner product, we mean the standard one.
Proof. Let U be a totally isotropic subspace of Fn2 , i. e., U ≤ U ⊥ . Assume dim(U ) ≤ (n−2)/2.
This means that dim(U ⊥ ) ≥ 2 + dim(U ). We have to prove that U is not maximal. This will
be accomplished as soon as we find an isotropic vector w ∈ U ⊥ , w ∈ / U . Indeed, for such a
w, the subspace span{U, w} is totally isotropic and properly includes U .
Let u, v ∈ U ⊥ be linearly independent modulo U . By this we mean that no nontrivial
linear combination of u and v belongs to U . If either u or v is isotropic, we are done. If
neither of them is, then u · u = v · v = 1 (remember that we are working over the tiny field
F2 ). We conclude that (u + v) · (u + v) = u · u + v · v = 1 + 1 = 0, the vector u + v is isotropic
and does not belong to U .
Corollary 2.34. Every maximal Eventown club system is maximum.

Exercises

Ex. 2.3.1. Prove: for n ≥ 7, there exist nonisomorphic extremal Eventown club systems.
Hint. Extremal means the number of clubs is maximum possible. Isomorphism means renaming
the citizens. Use Corollary 2.34 See the hint to Exercise 1.1.12.

Ex. 2.3.2. If every nonzero vector in U is isotropic then U is totally isotropic, assuming char F 6= 2.

Ex. 2.3.3. Disprove the previous exercise for characteristic 2.

——————————————————————— 55

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 2. THE BASICS
Ex. 2.3.4. Prove: every inner product space over F (whether singular or not) has an orthogo-
6 2. This, too, fails in
nal basis (the vectors of the basis are pairwise perpendicular), if charF =
characteristic 2, in every dimension ≥ 2.

Ex. 2.3.5. For what primes is F2p isotropic (i. e., contains an isotropic vector)?
Hint. For p = 2 and for p ≡ 1 (mod 4).

Ex. 2.3.6. Prove: if (V, β) is an inner product space of finite dimension over Fp , p an odd prime,
then (V, β) is isometric with some subspace of Fnp for sufficiently large n. (An isometry is an
isomorphism that preserves inner products. The inner product over Fnp is the standard one.)

Ex. 2.3.7. Prove: F3p is isotropic for every p.

Ex. 2.3.8. Prove: every 3-dimensional inner product space over Fp is isotropic.
Hint. Write β(u, u) in terms of the coordinates of u with respect to an orthogonal basis.

Ex. 2.3.9. Prove that every maximal totally isotropic subspace of Fnp has dimension bn/2c or
b(n − 1)/2c.

Ex. 2.3.10. Prove that every maximal totally isotropic subspace of Cn has dimension bn/2c.

Ex. 2.3.11 (Strong Eventown Theorem (E. R. Berlekamp, 1969; J. E. Graver, 1975.)). Consider
the following weaker version of Eventown: the pairwise intersections of the clubs must be even, but
the clubs themselves may be even or odd. Prove: still, the number of clubs is at most 2bn/2c + 
where  = 0 if n is even and  = 1 if n is odd.
Hint. Let S ⊆ Fn2 be the set of incidence vectors of the clubs; then S = S0 ∪ S1 , where S0
corresponds to the even, and S1 to the odd clubs. The condition is that the vectors in S are
pairwise perpendicular. Let Ui = span(Si ), ni = dim(Ui ). The set S1 is linearly independent by
the Oddtown Theorem; therefore m = |S| ≤ n1 + 2n0 . Prove that U1 is nonsingular. (Prove that
no nontrivial linear combination of S1 is perpendicular to all vectors in S1 .) Deduce from this that
U0 ∩ U1 = 0, therefore dim(U0 + U1 ) = n0 + n1 .(Since U0 + U1 ≤ U0⊥ , it follows that 2n0 + n1 ≤ n.
0 n even
Consequently, m ≤ n1 + 2b(n−n1 )/2c ≤ 2bn/2c +
1 n odd.

♦Ex. 2.3.12 (M. Szegedy, 1988). Recall the Mod-6-town Rules R(6): club sizes are not divisible
by 6, but their pairwise intersections are. For n 6= 3, prove the upper bound m ≤ 2n − 2 log2 n on
the number of clubs.

2.4 Graphs and set systems

2.4.1 Notation, terminology
Some of the basic notation for graphs and set systems is given in the “Notation” section on
p. xiii.
A walk of length k in a graph G is a sequence v0 , . . . , vk of vertices such that vi−1 and
vi are adjacent for every i. A walk without repeated vertices is a path. A closed walk has

56 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.4. GRAPHS AND SET SYSTEMS
v0 = vk . If it has no other repeated vertices, it is a k-cycle. A triangle in G is a 3-cycle.
The girth of a graph is the length of its shortest cycle. The odd girth is the length of the
shortest odd cycle. Cycle-free graphs have infinite girth; bipartite graphs have infinite odd
girth. Graphs of girth ≥ 4 are triangle-free.
The distance between two vertices is the length of the shortest walk (necessarily a path)
between them. If no such walk exists, the distance is infinite. G is connected if its vertices
are pairwise at finite distance. The diameter of a graph is the maximum distance between
pairs of vertices. (Take the supremum if G is infinite.)
A cycle-free graph is a forest. A connected forest is a tree. A graph without odd cycles
is bipartite.
A subgraph is obtained by deleting edges and vertices. An induced subgraph is obtained
by deleting vertices (and their incident edges) only, so an induced subgraph is determined
by its set of vertices. For W ⊆ V , we use G[W ] to denote
the subgraph of G induced on the
W
vertex set W . The edge set of G[W ] therefore is E ∩ 2 .
An empty graph has no edges: (V, ∅). An independent set in G is a subset W ⊆ V which
induces an empty subgraph. The size of the largest independent set in G is denoted by α(G).
This is the same as the size of the largest complete graph in G.
A subset is homogeneous if it induces either an empty or a complete subgraph.
A legal coloring of G is an assignment of “colors” to each vertex such that adjacent vertices
receive different colors. In other words, this is a partition of the vertex set into independent
sets. The minimum number of colors required for that is the chromatic number , denoted by
χ(G). Bipartitite graphs are precisely those with chromatic number ≤ 2.
Lower bounds on the chromatic number are of great importance and are usually hard to
come by. The following simple observation is often helpful.
Proposition 2.35. Let G be a graph with n vertices. The following relation holds between
the chromatic number χ(G) and the independence number α(G):

χ(G) ≥ n/α(G). (47)

Indeed, every color class in a legal coloring is an independent set, so χ(G) sets each of size
≤ α(G) add up to the set of vertices of G.
The notions of independent set and chromatic number extend to set systems in general.
For a set system F over the universe X, the subset Y ⊆ X is called independent if Y does
not contain any member of F, i. e., F ∩ 2Y = ∅. A legal coloring of F is an assignment of
colors to X such that no member of F be monochromatic (one-colored). In other words, we
again have a partition into independent sets. The minimum number of colors needed is the
chromatic number of F. (Exercise 1.4.1 concerns 2-chromatic set systems in hat-checkers’
terminology.)
A cover (or transversal ) of F is a set T ⊆ X which intersects every member of F. This is
equivalent to saying that X \ T is independent. The minimum size of a cover is the covering
number and is denoted by τ (F). Clearly, α(F) + τ (F) = n.
We mention an alternate terminology for set systems that is frequently used. Because
of their intimate conceptual relation to graphs, a set system is often called a hypergraph.

——————————————————————— 57

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 2. THE BASICS
Its members are called edges; the elements of its universe are vertices. The reader should
be aware of this terminology. We shall avoid its use because in our experience, the term
“hypergraph” to many students suggests some complex generalized concept. In our view,
few notions are as primitive in mathematics as that of a family of sets.

2.4.2 Chromatic number and short cycles

One of the most intriguing problems in graph theory is the construction of graphs with large
chromatic number plus some additional constraint, such as the absence of short cycles. Any
graph containing a complete subgraph on k vertices has chromatic number at least k, but
the presence of a Kk is far from necessary in order to push the chromatic number up.
Odd cycles have chromatic number 3, but the construction of a triangle-free graph of
chromatic number 4 takes some effort. (The smallest such graph has 11 vertices. Find it!)
In Section 3.2.5 we shall see some interseting examples of triangle-free graphs with arbitrarily
large (finite) chromatic number. In fact, the graphs we construct will avoid short odd cycles
up to any prescribed length.
Avoiding even cycles is, surprisingly, even more difficult and will not be considered in
this volume. While some constructions of graphs with large chromatic number and without
short odd cycles (not the ones to be described here) easily generalize to infinite chromatic
numbers, this is not the case for even cycles. One of the many fascinating results of Erdős
and Hajnal (1966) asserts that every graph with uncountably infinite chromatic number
must contain a 4-cycle (and in fact a complete bipartite subgraph for Km,ℵ1 for every finite
m). On the other hand, Erdős (1962) proved that for any k ≥ 2 and g ≥ 3, there exist
finite graphs of chromatic number k and girth g. (Recall: girth = length of shortest cycle.)
Erdős’s proof is nonconstructive. He actually proves the stronger result that for any c, g > 0
and there exist graphs G without cycles of length ≤ g and with the property that α(G) ≤ cn
where n is the number of vertices of G. The chromatic number of such a graph is ≥ 1/c
by inequality 47. (See Erdős–Spencer, 1974.) Known elementary constructions for the same
problem yield graphs with an enormous number of vertices (Lovász (1968), Křı́ž (1989)).
Recently, very deep results in number theory combined with eigenvalue methods in graph
theory have been invoked with success to explicitly construct relatively small graphs with
large chromatic number and girth (Margulis (1988), and Lubotzky–Phillips–Sarnak (1987)).

2.4.3 Block designs

TO BE WRITTEN

Definitions: BIBD, STS, finite projective plane

basic parameters of finite projective planes
incidence matrix of projective plane nonsingular
Thm: cycle structure of automorphism on points and lines the same
Thm: no involution without absolute point
Cor.: Friendship Theorem

58 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 2.4. GRAPHS AND SET SYSTEMS
State: Bruck–Ryser Theorem

——————————————————————— 59

c László Babai and Péter Frankl. 1988, 1992, 2020.
60 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
Chapter 3

“General position” arguments

3.1 Configurations in general position

The concept of “general position” frequently occurs in geometry. Informally it refers to
arrangements of objects where only things that must coincide do. (Opposite of Murphy’s
Law.) For instance, three lines in the Euclidean plane may be concurrent (pass through a
common point); but that means a certain algebraic equation is satisfied by the coefficients of
the equations defining the lines. (Their 3×3 determinant vanishes.) This equation not being
an identity, we may regard its fulfillment accidental. Accidental relations are not permitted
to hold between objects in general position; in particular, the three lines in general position
are not allowed to be concurrent. Nor are any two of them allowed to be parallel, since that
would mean another algebraic coincidence. Indeed, if three lines in the plane are chosen “at
random” (whatever this means), we expect them not to be concurrent, and also, no two of
them to be parallel.
The reader might interject: no pair should be perpendicular either. Indeed. And no pair
at 60 degrees. Or at any specific angle. Any number of prespecified conditions of this sort
will be met. But of course it is impossible to meet all of them at once.
We shall only encounter a few cases of this vaguely defined concept, and will give the exact
definition in each case. The reader may find the definitions somewhat arbitrary; indeed, it is
only one or two kinds of relation that we shall rule out in each case. All we can promise is that
we make judicious choices; the use of the resulting rigorous concepts of “general position”
will be amply demonstrated in later chapters. Especially Section 5.1 and Chapter 6 abound
with “general position” arguments.

3.1.1 Points in general position. The moment curve.

Let us start with a set S of points in the linear space W of dimension n.
Definition 3.1. We say that S ⊆ W is in general position, if any n of the elements of S are
linearly independent, where n = dim(W ).

61
CHAPTER 3. GENERAL POSITION
The problem of determining the maximum number of points in general position in a space
over a finite field will be addressed in the Exercises. Although there are many ways to select
|F| points in general position in Fn , a particularly elegant and explicit choice is provided
(over any field, whether finite or infinite) by a set called the moment curve. This curve is
defined as the range of the F → Fn function
mn (α) = (1, α, α2 , . . . , αn−1 ) ∈ Fn . (1)

Definition 3.2. The moment curve1 in Fn is the set

Mn = {mn (α) = (1, α, α2 , . . . , αn−1 ) : α ∈ F} (2)

Proposition 3.3. The points of the moment curve are in general position.
In addition to its role in “general position” arguments scattered all over this volume, the
moment curve over R will be central to some major results: the proof of Kneser’s Con-
jecture in graph theory (Theorem 3.32) and the statement of the Upper Bound Theorem
(Theorem 3.27).
The moment curve and some of its remarkable geometric properties (cf. Section 3.2)
were discovered by C. Carathéodory in 1907 and then forgotten, until D. Gale, unaware of
Carathéodory’s work, rediscovered it (1956).


Proof. For n distinct elements α1 , . . . , αn ∈ F, consider the determinant det m(α1 ), m(α2 ), . . . , m(αn ) .
This is a Vandermonde determinant; its value is
Y
(αj − αi ) 6= 0.
1≤i<j≤n

Therefore the rows m(αi ) are linearly independent.

This property of the moment curve allows us to construct a large set in general position if
F is large. In applications where the field may be too small, we can usually get around this
problem using the Field Extension Lemma (Lemma 2.29) and the fact that every field can be
extended to an infinite one (Exercise 2.1.23 (b)). This principle is illustrated by the proofs
of Theorems 5.7, 6.14, and 6.15.

3.1.2 Subspace in general position w.r.t. a family of subspaces

In the next setting, a family of subspaces is given, and we want a subspace of prescribed
dimension to be in general position with respect to the family.
Definition 3.4. Let W be an n-dimensional space over F and for i = 1, . . . , m, let Ui be
a subspace of W ; dim(Ui ) = ri . A subspace V ≤ W of codimension t (i. e., of dimension
This definition deviates slightly from common usage. The moment curve is usually defined by m0n (α) =
1

(α, α2 , . . . , αn ). This curve is congruent (isometric) to Mn+1 . Note that Mn+1 lies in the hyperplane x0 = 1
so indeed it is an n-dimensional curve.

62 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 3.1. CONFIGURATIONS IN GENERAL POSITION
s = n − t) is in general position with respect to the Ui if

dim(Ui ∩ V ) = max{ri − t, 0} for i = 1, . . . , m.

Recall that by inequality (21), max{ri − t, 0} is the minimum possible dimension of Ui ∩ V .

So what we require is that the intersection of V and each Ui be as small as it conceivably
can get.

Our objective is to prove that given the Ui and the number s ≤ n, such a subspace V
always exists if F is infinite or at least large enough (greater than some function of n and
m).
Theorem 3.5 (Subspace in General Position). Let W be an n-dimensional space over
the field F, and U1 , . . . , Um subspaces of W . Further, let s be an integer, 0 ≤ s ≤ n. Then
there exists a subspace V ≤ W of dimension s, in general position with respect to the Ui , if
the field F has order > s(m + 1).
Remark 3.6. Note that, in particular, a subspace of given dimension and in general position
with respect to any (finite) number of subspaces will exist if F is infinite.
The proof will be based on the observation that elements randomly selected from a large
finite set are unlikely to satisfy a given polynomial equation unless the polynomial is zero.
Lemma 3.7 (Sparse Zeros Lemma) (J. T. Schwartz, 1980). Let f ∈ F[x1 , . . . , xn ] be
a nonzero polynomial of degree d and Ω ⊆ F a finite set, |Ω| = N . Let Z(f, Ω) denote the
set of roots of f from Ωn :

Z(f, Ω) = {(α1 , . . . , αn ) : f (α1 , . . . , αn ) = 0, αi ∈ Ω}.

Then
|Z(f, Ω)| ≤ dN n−1 . (3)

Remark 3.8. The cardinality of Ωn is N n . So the Sparse Zeros Lemma can be reworded in
probabilistic terms as follows.
If we pick an element (α1 , . . . , αn ) ∈ Ωn at random, the probability that we found a root
of f is at most d/N .
For infinite fields we can make this probability arbitrarily small by selecting a sufficiently
large subset for Ω.
Proof. We proceed by induction on n, the number of variables.
For n=1, the number of roots of f does not exceed its degree. This means |Z(f, Ω)| ≤ d,
proving this case of inequality (3).
Let now n ≥ 2. Write f in terms of the powers of xn :

f (x1 , . . . , xn ) = g0 + g1 xn + g2 x2n + · · · + gk xkn , (4)

——————————————————————— 63

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 3. GENERAL POSITION
where gi ∈ F[x1 , . . . , xn−1 ], deg gi ≤ d−i, and gk is not the zero polynomial. Now (α1 , . . . , αn )
may belong to Z(f, Ω) for two reasons.
It is possible that gk (α1 , . . . , αn−1 ) = 0. (This alone will not put (α1 , . . . , αn ) in Z(f, Ω),
so our upper bound may be somewhat generous.) As deg gk ≤ d−k, the number of roots of gk
in Ωn−1 is ≤ (d−k)N n−2 by the inductive hypothesis. Therefore the number of corresponding
n-tuples (α1 , . . . , αn ) is ≤ (d − k)N n−1 .
It is also possible that gk (α1 , . . . , αn−1 ) 6= 0. We bound the number of such (n − 1)-tuples
simply by N n−1 (this again seems quite liberal). Having fixed α1 , . . . , αn−1 this way, the
number of possible choices for αn is at most k, since αn must now be a root of the nonzero
polynomial (4) of degree k. Hence this case results in at most kN n−1 roots of f .
The two upper bounds add up to a total of dN n−1 , completing the induction step.
We now wish to formalize the statement that matrices over large domains normally have full
rank.
Proposition 3.9. Let u1 , . . . , ur ∈ Fn be linearly independent row vectors. Then there
exists a nonzero polynomial f ∈ F[x1 , . . . , x(k−r)n ] of degree min{n − r, k − r} such that the
following holds for any matrix A ∈ Fk×n . If the first r rows of A are u1 , . . . , ur then A has
full rank, unless the entries in the remaining k − r rows satisfy f .
Proof. For A to have full rank, it is necessary and sufficient that at least one of the maximal
square minors be nonsingular. Since u1 , . . . , ur are linearly independent, it is possible to
continue filling in A so as to make it have full rank. This means at least one of these minors
can be nonsingular. Let f be the determinant of this minor, viewed as a polynomial in the
entries of the last k − r rows. When f does not vanish, A has full rank.
Next, we link nonminimal intersection to algebraic equations.
Proposition 3.10. Let U be a subspace of W = Fn , dim(U ) = r, and 0 ≤ s ≤ n. Then
there exists a nonzero polynomial f ∈ F[x1 , . . . , xsn ], deg(f ) = min{n − r, s}, with the
following property. Let v1 , . . . , vs ∈ W be linearly independent, and V = span{v1 , . . . , vs }.
Then either dim(U ∩ V ) = max{r + s − n, 0}, or or the sn entries of the vi satisfy f .
Proof. Let u1 , . . . , ur be a basis of U . Fix the ui and consider the (r + s) × n matrix A whose
rows are the ui and the vj . We know from the modular identity (15) that dim(U ∩ V ) will be
minimal, i. e., equal to max{r + s − n, 0}, precisely when dim(U + V ) = rk A is maximal, i. e.,
A has full rank. By the previous Proposition, this will surely be the case unless a certain
polynomial f of degree min{n − r, s} vanishes.
Proposition 3.11. Let U1 , . . . , Um ≤ W = Fn be as in Theorem 3.5. Then there exists
a nonzero polynomial g ∈ F[x1 , . . . , xsn ] of degree ≤ s(m + 1) with the following property.
Let v1 , . . . , vs ∈ W and V = span{v1 , . . . , vs }. Then either dim(V ) = s and V is in general
position with respect to the Ui , or the sn components of the vi satisfy g.
Proof. In order for V to have dimension s, the vectors v1 , . . . , vs must be linearly independent.
This will be true unless certain polynomial f0 of degree s (the determinant of an arbitrary
s × s minor of the s × n matrix whose rows are the vi ) vanishes. If dim V = s then by
the previous Proposition, dim(Ui ∩ V ) will be minimum unless some fixed polynomial fi of
degree min{n − ri , s} ≤ s vanishes. Set g = f0 f1 . . . fm to conclude the proof.

64 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 3.1. CONFIGURATIONS IN GENERAL POSITION
Now the proof of Theorem 3.5 consists of simply combining Proposition 3.11 and the Sparse
Zeros Lemma (Lemma 3.7). Any subspace generated by s vectors whose sn components do
not satisfy certain nonzero equation of degree ≤ s(m + 1) will have dimension s and will be
in general position with respect to the given subspaces. In order for such values for the ns
variables to exist, it suffices that the field have order greater than s(m + 1), according to
Lemma 3.7.

3.1.3 Linear maps in general position

Let W and T be linear spaces over F, n = dim(W ), and t = dim(T ). Although it will not
be an assumption, the situation we have in mind is that t < n. Therefore any linear map
ϕ : W → T will have to identify certain pairs of points. Given a family of affine subspaces
in W , we wish to achieve that ϕ reduce their dimensions by as little as possible.
Definition 3.12. Let W and T be linear spaces over F, n = dim(W ), and t = dim(T ).
Let Ui (i = 1, . . . , m) be a family of affine subspaces of W ; dim(Ui ) = ri . The linear map
ϕ : W → T is in general position with respect to the Ui if

dim(ϕ(Ui )) = min{ri , t} for i = 1, . . . , m.

Obviously, this is the best we can hope for: the dimension of ϕ(Ui ) cannot be greater than
the dimension of either Ui or T .
Theorem 3.13 (Linear Map in General Position). Given W, T, Ui (i = 1, . . . , m)
as above, a linear map ϕ : W → T in general position exists, if the field F has order
> (n − t)(m + 1).
Proof. The theorem is trivially true if n ≤ t: any injective map ϕ will be in general position.
So we may assume t < n.
We may assume all the Ui are nonempty. For a nonempty affine subspace S, let (S)0
denote the corresponding linear subspace (of which it is a translate). It is clear that



ϕ(S) 0 = ϕ (S)0 (5)

for any affine subspace S of W and any linear map ϕ : W → T . Therefore it suffices to
prove the result with the subspaces (Ui )0 in the place of Ui . For simplicity we may assume
Ui = (Ui )0 .
Now let V ≤ W be a subspace of codimension t, in general position with respect to the
Ui . (Theorem 3.5 assures the existence of such V .) Take a linear map ϕ : W → T such
that ker(ϕ) = V . Since dim(T ) = codimW (V ), such a map exists (Proposition 2.10). Now,
according to Definition 3.4, dim(Ui ∩ V ) = max{ri − t, 0}. Hence

dim(ϕ(Ui )) = ri − dim(Ui ∩ ker ϕ)

= ri − max{ri − t, 0} = min{t, ri },

——————————————————————— 65

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 3. GENERAL POSITION
as required.

We illustrate the use of this result on two examples. Both of them will have applications in
later chapters.
Corollary 3.14. Let W and T be linear spaces over the infinite field F; dim T = t. Let
Ui and Vi be subspaces of W ; assume dim(Ui + Vi ) ≤ t. Then there exists a linear map
ϕ : W → T such that for every i,

dim(ϕ(Ui )) = dim(Ui ); dim(ϕ(Vi )) = dim(Vi );

dim(ϕ(Ui ) ∩ ϕ(Vi )) = dim(Ui ∩ Vi ).

Proof. Let ϕ : W → T be in general position with respect to the subspaces Ui + Vi . Since
the dimension of these subspaces is ≤ t, according to Definition 3.12, dim(ϕ(Ui + Vi )) =
dim(Ui + Vi ). This means ϕ is injective on Ui + Vi . Hence it preserves the dimensions of
each subspace of Ui + Vi .

It is clear from the proof that the condition that the field be infinite can be relaxed to
|F| > (n − t)(m + 1), where n = dim W .
Corollary 3.15. Let W and T be linear spaces over the infinite field F; dim T = t ≥ 1. For
i = 1, . . . , m, let Ui ≤ W be a subspace of dimension ≤ (t − 1), and let Bi ⊆ W be a finite
subset of W . Then there exists a linear map ϕ : W → T such that for each i
(i) |ϕ(Bi )| = |Bi |;

(ii) dim(ϕ(Ui )) = dim(Ui ); and

(iii) |ϕ(Ui ) ∩ ϕ(Bi )| = |Ui ∩ Bi |.

Proof. Let B = m
S
i=1 Bi . Let ϕ : W → T be in general position with respect to the following
finite collection of subspaces: span{Ui , a}, for i = 1, . . . , m and a ∈ B; and aff{a, b}, for
a, b ∈ B.
Since all affine subspaces listed have dimensions ≤ t, the map ϕ will be injective on each.
The 1-dimensional affine subspaces listed guarantee that ϕ will be injective on B. This
implies (i) and shows that in order to verify (iii), we only have to see that a ∈ / Ui implies
ϕ(a) ∈ / ϕ(Ui ) (a ∈ B). Both this latter requirement and (ii) follow from the injectivity of ϕ
on the subspace span{Ui , a}.

3.1.4 Checking identities: a Monte Carlo algorithm

In this section we describe an entertaining algorithmic application of the Sparse Zeros Lemma
(Lemma 3.7), due to Jacob T. Schwartz.
Imagine a polynomial f ∈ Z[x1 , . . . , xn ], given in the form of a k × k determinant f =
det(gij ), where the gij are homogeneous linear polynomials of the n indeterminates. How
can one determine whether or not f is the zero polynomial?

66 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 3.1. CONFIGURATIONS IN GENERAL POSITION
Well, this doesn’t look hard. Indeed, where is the problem? Let’s just expand the
determinant and see if everything cancels.
On second thought, we may realize with disappointment, what a hopeless task this would
be. Think just of a diagonal matrix. Then, the determinant has a single expansion term,
the product of the diagonal entries. (So we can tell if it is zero without any multiplications,
but this is not the point. If the recipe is to expand the determinant, let’s see what that
costs.) This product is a homogeneous polynomial
of degree k in the n indeterminates, and
will expand to a linear combination of the n+k−1 k
monic monomials (cf. Exercise 2.1.2). In
the general case, intermediate results may blow up at a similar rate before anything would
cancel.
For n, k in the range of 200, this number is more than 10100 . No one can expect even an
idealized computer to perform this many operations within 14 billion years (the estimated
age of the Universe). (The computer would have to perform an astronomical number of
operations within a fraction of the time it takes for light to travel the length of the diameter
of an atom.)
How, then, can we answer this simple question? No reasonable deterministic algorithm
has yet been found. So we shall not be able to tell for sure if f is zero.
But we could at least try to substitute values for the variables. For any particular
substitution, one can compute the value of the determinant reasonably quickly (by Gaussian
elimination). Now if the value computed is not zero then we know the answer: f 6= 0. But
what happens if we get zero? Well, we just hit a root of f , let’s try again. Another root?
Try once more. The problem is, no matter how many roots we find, f still might be nonzero.
Maybe our system was wrong. Or we just had bad luck.
Speaking of luck, suppose we select the values we assign to the xi from the set Ω =
{1, . . . , N }. If we select the assigment at random, what is the chance that we get a root of
f ? (Assuming, of course, that f 6= 0.)
The degree of f is k. By the Sparse Zeros Lemma (Lemma 3.7), the probability that
a random n-tuple α ∈ Ωn is a root is less than k/N . For this estimate to yield anything,
it is advisable to make N greater than k. Let us set, for instance, N = 10k. Assuming
f 6= 0, what is the probability that a random substitution will not discover this? At most
0.1. What is the chance that none of 100 independent random substitutions will? Less than
10−100 .
So if a sequence of 100 random substitutions did not prove that f 6= 0, we can be pretty
certain that actually f = 0. Not 100%, though. But close enough to bet on it. If we were
to lose the bet, we would know that an experiment that had two logically possible outcomes
ended with the one that had probability < 10−100 .
This procedure raises puzzling thoughts about the nature of mathematical proof. Here is
this polynomial f . We know beyond reasonable (or even utterly unreasonable) doubt, that
f = 0. We have a proof. A court would accept it. A mathematician might not; he may insist
on a formal proof. Suppose we are able to furnish a formal proof a few thousand pages long
(like the proof of the “Enormous Theorem,”1 the Classification of Finite Simple Groups).

1
See D. Gorenstein, The Enormous Theorem, Scientific American 253, December 1985, pp. 104–115.

——————————————————————— 67

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 3. GENERAL POSITION
Would such a proof be more convincing?
Leaving these thoughts to the reader for further contemplation, let us formulate what we
have just proved.
First of all, we give a fairly general definition of how a polynomial can be represented.
Definition 3.16. A straight line program in the ring R = Z[x1 , . . . , xn ] of multivariate
polynomials with integer coefficients is a sequence of polynomials f1 , . . . , fm ∈ R such that
each fi is either

(a) ±1, or

(b) one of the indeterminates, or

(c) a sum of the form fi = fj + fk for some j, k < i, or

(d) a product of the form fi = fj fk for some j, k < i.

We say that the polynomials fi are computed by this straight line program; m is the length
of the program.
With every step of the straight line program we associate an estimated degree edeg(i),
defined inductively as follows.
If fi = ±1, then edeg(i) = 0. If fi is an indeterminate, then edeg(i) = 1. If fi = fj + fk
(j, k < i), then edeg(i) = max{edeg(j), edeg(k)}. Finally, if fi = fj fk (j, k < i) then
edeg(i) = edeg(j) + edeg(k).
The estimated degree of the straight line program is max1≤i≤m edeg(i).

We now define the algorithmic concept we have illustrated above.

Definition 3.17. Let π be a predicate which associates truth values (0 or 1) with strings over
a finite alphabet. The objective of a decision procedure is to determine π(x), given the input
string x. A Monte Carlo decision procedure makes random choices along the way and for
every input x and prescribed parameter value t arrives at a decision “π(x) = ε” (ε ∈ {0, 1}).
If π(x) = 0, the procedure must yield “ε = 0” (no error in this direction allowed). If, however,
π(x) = 1, the procedure is still allowed to output (erroneously) “ε = 0” with probability
≤ 2−t . (We call 2−t the error tolerance). So if the output is “ = 1,” we know for sure that
π(x) = 1; but if the output is “ = 0,” all we can say is that either π(x) = 0 or we were
out of luck. How much out of luck depends on our choice of t; e. g. if we choose t = 20, our
chance of getting a wrong answer is less than one in a million.

Theorem 3.18 (Polynomial Identity Test) (J. T. Schwartz, 1980). Let f ∈ Z[x1 ,
. . . , xn ] be given by a straight line program of length m and estimated degree d. There exists
a Monte Carlo procedure to decide whether or not f 6= 0. If the error tolerance is 2−t , the
procedure will perform mt arithmetic operations on integers with O(d(m + log d)) binary
digits.

68 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 3.1. CONFIGURATIONS IN GENERAL POSITION
Proof. Let N = 2d, Ω = {1, . . . , N }. Let us select α1 , . . . , αt ∈ Ωn at random and evaluate
f (αi ) (i = 1, . . . , t). If f (αi ) 6= 0 for some i, declare “f 6= 0”; otherwise declare “f = 0.”
Since the actual degree of the polynomial f (which we may never find out) is clearly ≤ d, the
probability of wrong decision is ≤ 2−t . The growth rate of the number of digits can easily
be kept track of.
Strictly speaking, the result as stated does not cover the determinant problem we started
with. The reason is that the most common method for determinant expansion (Gaussian
elimination) employs all the four arithmetic operations and therefore cannot be performed
within the polynomial ring Z[x1 , . . . , xn ].
There are two answers to this objection. One is that we can define straight line programs
in fields (fi is allowed to be fj /fk , where j, k < i and fk 6= 0). Divisions contribute to
the estimated degree just as multiplications do. We can then perform Gaussian elimination
within the field Q(x1 , . . . , xn ) of rational functions over Q. We use Schwartz’s algorithm to
find nonzero pivots to avoid division by zero. If we find none, we declare the determinant
zero.
The second answer, given by V. Strassen (1973), is a general result, which says, roughly,
that divisions can be avoided. More specifically, if a polynomial f ∈ Z[x1 , . . . , xn ] can be
computed by a straight line program in the field of rational functions, then it can also be
computed by one in the polynomial ring, without significant loss of efficiency.
But the simplest answer is to extend Theorem 3.18 to procedures other than straight line
programs. There is, however, no way to get around to limiting the blowup of the numbers
computed. When evaluating the polinomial at a particular input, the partial results should
not be allowed to be either too large or too small. (The number n! is not too large: it
n
only takes O(n log n) digits to write down. But 1 + 22 requires an exponential number of
digits. We can get such monstrous numbers by a straight line program of length n + 2 if we
don’t watch the estimated degree! (How?)) It is not evident at all (but true) that Gaussian
elimination will not produce too large partial results (Edmonds, 1967).

Exercises
For the following sequence of problems, let g(n, q) denote the maximum number of points in general
position in Fnq where q is a prime power; n ≥ 2.
♦Ex. 3.1.1. Prove: g(n, q) ≥ n + 1.
♦Ex. 3.1.2. Prove: g(n, q) ≥ q + 1.
Hint. Add one point to the moment curve.
Ex. 3.1.3. Prove: g(n + 1, q) ≤ 1 + g(n, q).
Hint. Assume we have a g(n + 1, q) × (n + 1) matrix A over Fq whose rows are in general position
in Fn+1
q . Performing an isomorphism Fn+1
q → Fn+1
q if necessary (change of basis), we may assume
that the top n + 1 rows of A form the identity matrix In+1 . Verify that deletion of the first row
and column results in a matrix whose rows are again in general position (now in Fnq ).
Ex. 3.1.4. Prove: g(2, q) = q + 1. Conclude: g(n, q) ≤ q + n − 1.

——————————————————————— 69

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 3. GENERAL POSITION
♦Ex. 3.1.5. Give a direct proof of the inequality g(n, q) ≤ q + n − 1.
Hint. Consider a subspace spanned by n − 2 of the vectors.

Ex. 3.1.6. Prove: The product of the nonzero elements of Fq is −1.

elements of Fq are roots of the polynomial xq − x (Exercise 2.1.14),
Hint. Using the fact that all Q
prove the identity xq−1 − 1 = α∈F× q
(x − α). Compare the constant terms. (Why is the result right
for even q?)

Ex. 3.1.7. Prove: For odd q, g(3, q) = q + 1 and therefore g(n, q) ≤ q + n − 2 (n ≥ 3).
Hint. Assume for a contradiction, that A is a (q + 2) × 3 matrix over Fq whose rows are in general
position. As we have done earlier, we may assume that the first 3 rows of A form the identity matrix
I3 . Let B = (βij ) denote the remaining (q − 1) × 3 matrix. Clearly no entry of B is zero. Without
loss of generality we may therefore assume that βi1 = 1 for every i (i = 1, . . . , q − 1). It follows
(verify!) that neither the second nor the third column of B can have equal entries. Therefore both
the second and the third columns contain all nonzero elements of Fq in some order. Moreover, the
ratios βi2 /βi3 must also all be different (why?) and therefore represent all nonzero elements of Fq in
some order. Prove, using the previous exercise, that this leads to the contradiction 1 = −1. (This
is not a contradiction if charFq = 2, i. e., q is even.)

♦Ex. 3.1.8. Prove: g(3, 4) = 6.

Ex. 3.1.9. Prove: For even q, g(3, q) = q + 2.

Hint. g(3, q) ≤ q + 3 − 1 = q + 2 by Exercise 3.1.4. To show this bound is tight, take, using the
notation of the solution of Exercise 3.1.7, the matrix B where βi1 = 1 (i = 1, . . . , q − 1); the values
2 (i = 1, . . . , q − 1).
of βi2 are the nonzero elements of Fq in some order, and βi3 = βi2

♦Ex. 3.1.10. Prove: for n ≥ q, g(n, q) = n + 1.

♦Ex. 3.1.11. Let W and T be linear spaces over the infinite field F; dim T = t. Let Ui and Vi be
nonempty affine subspaces of W such that dim(Ui ) + dim(Vi ) ≤ t, and dim(Ui ) + dim(Vi ) ≤ t − 1
if 0 ∈
/ Ui ∩ Vi (i = 1, . . . , m). Prove: there exists a linear map ϕ : W → T such that for every i, ϕ
is injective on aff{Ui , Vi }.

Ex. 3.1.12. Define what it should mean that N subspaces of prescibed dimensions d1 , . . . , dN in Fn
are in general position. Show that such subspaces exist, assuming F is infinite or at least |F| ≥ nN .
Hint. One possible requirement is that their pairwise intersection is as small as possible, or equiva-
lently, that their pairwise sums be as large as possible (i. e., min{n, di + dj }). A stronger condition
is to require that for all subsets I ⊆ [N ],
X n X o
dim Ui = min n, di .
i∈I i∈I

Prove that the subspaces spanned by disjoint subsets of the moment curve satisfy this P stronger
condition. So in order for this to work over a possibly finite field F, it suffices if |F| ≥ N
i=1 di .

Ex. 3.1.13. Over an infinite field, when is it not possible to fill a partially filled n × n matrix to
full rank ?

70 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 3.2. CONVEXITY

Note. We don’t know the answer. Here is one obstacle: if there exists a k × ` minor, completely
filled, of rank < k+`−n, then clearly, the matrix cannot be completed to full rank. – Is the presence
of such an obstacle not only sufficient but also necessary? At least in a special case, the answer is
yes. The special case: all entries filled must be zeros and algebraically independent transcendentals
(distinct variables). In this case, the positive answer is a restatement of the König–Hall Theorem
characterizing bipartite graphs without perfect matchings (cf. Lovász (1979c), Problem 7.4).

3.2 Convexity
We have seen in the preceding chapter that a number of geometric concepts, including some
“metric” ones such as orthogonality, have their analogs in spaces over arbitrary fields. There
is, however, a wealth of material that depends on the ordering of the real numbers and
involves inequalities rather than just equations. One of the most important of these ordering
related concepts is convexity. This being the subject of the present section, our base field
will be R throughout.

3.2.1 Terminology
We start with introducing terminology and listing a number of plausible basic facts, which,
however, are often not so easy to prove. For the proofs, see e.g., B. Grünbaum (1967), or V.
Chvátal (1983).
Let W be a linear space over R, typically Rn . We shall think of Rn as an inner product
T
space with respect to the √ inner product u · v = u v. The length or Euclidean norm
√ standard
of a vector v is kvk = v · v = v v.T

Definition 3.19. A convex combination of the vectors v1 , . . . , vm ∈ W is a linear combina-

tion m
P Pm
λ v
i=1 i i (λi ∈ R) where λ
i=1 i = 1 and λi ≥ 0.

A convex set is a subset of W , closed under convex combinations. The convex hull of a
subset S ⊆ W is the set of all convex combinations of the finite subsets of S; it is denoted by
conv(S). This set is always convex. S is convex if and only if S = conv(S). The intersection
of convex sets is convex. The convex hull of a pair of points is called the straight line
segment connecting the two points. A set S is convex if and only if it contains the straight
line segment connecting each pair of points of S.
Note that every convex combination is an affine combination; therefore conv(S) ⊆ aff(S).
It follows that affine subspaces are convex.
The dimension of a convex set is the dimension of its affine hull. A convex subset C of
W is full-dimensional if aff(C) = W .
The convex hull of a finite set of points is called a polytope. The convex hull of a set of
n+1 affine independent points is an n-dimensional simplex. A 1-dimensional simplex is a line
segment; a 2-dimensional simplex is a triangle (together with its interior); a 3-dimensional

——————————————————————— 71

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 3. GENERAL POSITION
simplex is a tetrahedron. A regular simplex is the convex hull of a set of points pairwise at
equal distance.
A hyperplane is an affine subspace of codimension 1. Hyperplanes in W = Rn are defined
by a single equation aT x = β, where a ∈ Rn , a 6= 0 is a normal vector of the hyperplane,
and β ∈ R.
A hyperplane divides the space into two halfspaces defined by the inequalities aT x ≥ β
and aT x ≤ β; their intersection is the hyperplane. The hyperplane is their common boundary.
Every convex set is the intersection of the halfspaces containing it.
Every polytope is the intersection of a finite number of halfspaces. Conversely, if C is
the intersection of a finite number of halfspaces and C is bounded then C is a polytope.
Let P ⊂ Rn be a hyperplane and H one of the halfspaces bounded by P . Assume C ⊆ H
is a convex set. Then C ∩ P = C ∩ H 0 is a face of C, where H 0 is the other halfspace bounded
by P . In addition, the empty set and C itself are also called faces of C. If dim(C) = d
then all faces of C other than itself have dimension ≤ d − 1. The faces are convex sets. The
zero-dimensional faces are the vertices, the one-dimensional faces are the edges of C. The
faces of dimension (d − 1) are the facets of C. If C is full-dimensional and F = C ∩ P is a
facet of C then F uniquely determines the corresponding halfspace: P = aff(F ) and H is
the unique halfspace bounded by P and containing C.
Not every convex set has facets. Polytopes do, and every polytope is the intersection of
the halfspaces defining its facets. Each face of a polytope is again a polytope. For a finite
set S ⊂ Rn , every face of the convex hull of S is the convex hull of some subset of S. In
particular, each vertex of conv(S) is a member of S. A subset S 0 ⊆ S determines a face if
conv(S 0 ) is a face and S 0 is its vertex set. Obviously, not every subset of the set of vertices
determines a a face. (Think of the cube, for instance.) But, as we shall soon see, a polytope
may have more faces than one would expect.

3.2.2 Helly’s Theorem

The dimension of a linear space over the reals has a characterization in terms of intersection
properties of convex sets. The result, Helly’s Theorem, has a long history of analogues in
combinatorics, some of which we shall encounter in Chapter 5.
Theorem 3.20 (Helly’s Theorem). If C1 , . . . , Cm ⊆ Rn are convex sets such that any
n + 1 of them intersect then all of them intersect.
The result is tight (the quantity n + 1 cannot be reduced), as shown by the facets of a
full-dimensional simplex. (See Exercise 3.2.1.)
We shall deduce Helly’s Theorem from the following easy lemma.
Lemma 3.21 (J. Radon, 1921). Let S ⊂ Rn be a set of m ≥ n + 2 points in Rn . Then
S has two disjoint subsets S1 and S2 whose convex hulls intersect.
This, again, is tight: if S is an affine basis of Rn and so |S| = n + 1, then for any two
disjoint subsets S1 , S2 ⊆ S, even their affine hulls are disjoint.
Proof. Since |S| > n + 1, the set S is affine dependent. This means there exists a nontrivial
linear relation with zero-sum coefficients among the elements of S. Let S1 and S2 consist of

72 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 3.2. CONVEXITY
those elements with positive and negative coefficients, resp., in this combination. Separating
the two subsets of terms we obtain a relation
X X
λu u = µv v; (6)
u∈S1 v∈S2
P P
where u∈S1 λu = v∈S2 µv , λu , µv > 0, and S1 , S2 6= ∅. Thus dividing each side of equation
(6) by the sum of their coefficients we obtain a point in conv(S1 ) ∩ conv(S2 ).
T
Proof of Helly’s Theorem. Assume first that m = n + 2. Let ai be a point in j6=i Cj . Set
S = {a1 , . . . , an+2 }. Now, by the preceding lemma, S has two disjoint subsets S1 and S2
with intersecting convex hulls; let w ∈ conv(S1 ) ∩ conv(S2 ). We claim that w belongs to all
the Ci . Indeed, pick some ai ; it belongs to at most one of S1 and S2 . Suppose, say, ai ∈ / S1 .
But then S1 ⊆ Ci , therefore conv(S1 ) ⊆ Ci , hence w ∈ Ci . This concludes the proof for
m = n + 2.
The general case now follows by induction on m. For m ≤ n + 1, there is nothing to
prove. Assume m ≥ n + 3. By the particular case just proved, every n + 2 of the Ci intersect.
It follows, that every n + 1 of the sets C1 , . . . , Cm−2 , Cm−1 ∩ Cm intersect. But then, by the
induction hypothesis, all intersect.

3.2.3 A polytope with many faces

Let S be a set of n points in Rd . Let us consider the polytope M = conv(S).
How many faces of dimension k can M have?
If F is such a face then F = conv(S 0 ) for some S 0 ⊆ S. Let T ⊆ S 0 be an affine basis of
aff(F ) = aff(S 0 ). Then |T | = k + 1. Observe that T uniquely determines F : F = aff(T ) ∩ M .
Therefore, the number of possible choices of T is an upper bound on the number of faces of
dimension k. We state the conclusion.
Proposition
3.22. The number of faces of dimension k of a polytope with n vertices is
n
at most k+1 . Moreover, if this upper bound is attained, then all k-dimensional faces are
simplices and every set of (k + 1) vertices determines a face of dimension k.
We leave the verification of the second part of the Proposition to the reader.

How far are these bounds from best possible? They are tight if n ≤ d + 1 (take a simplex;
cf. Exercise 3.2.3). On the other hand, no n-dimensional faces exist if n ≥ d + 2. But even
the estimates for the low dimensional faces seem far too generous if n is large compared
to
d. What would it mean, for instance, that the number of edges (1-dimensional faces) is n2 ?
Every pair of vertices is connected by an edge. This seems quite absurd. Indeed, in the
plane (d = 2), M is a convex polygon, and the number of its edges is n, the same as the
number of vertices. In 3-space, one can prove that the number of edges is at most 3n − 6
(Exercise 3.2.4). It would be natural to expect a linear upper
bound in dimension 4 as well.
It may be a little surprising that in 4-space, the bound n2 is actually tight, for every

n.
n
And, as soon as the dimension d of the space is greater than 2k + 1, the bound k+1 also
becomes tight, for every n.

——————————————————————— 73

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 3. GENERAL POSITION
n


A polytope with n vertices and k+1 k-dimensional faces will be called (k +1)–neighborly.
Being k-neighborly means every k points determine a (k − 1)-dimensional face. For k ≤ n it
follows that a k-neighborly polytope is also l-neighborly for every l ≤ k.
Theorem 3.23. For every d, n ≥ 1 there exists a bd/2c–neighborly polytope M ⊂ Rd with
n vertices.
The moment curve, introduced in Section 3.1, gives the clue to the solution. Recall that
the points of the moment curve in Rd+1 are

md+1 (α) = (1, α, α2 , . . . , αd ) (α ∈ R).

In order to define the cyclic polytopes, we omit the first coordinate1 in md+1 and consider
the following curve in dimension d:

m0d (α) = (α, α2 , . . . , αd ) (α ∈ R).

Let us choose any n ≥ d + 1 distinct real numbers α1 , . . . , αn . The cyclic polytope M (d, n)
with n vertices in Rd is the convex hull conv{m0d (α1 ), . . . , m0d (αn )}. (We shall have to prove
that this indeed has n vertices. This is part of the next Theorem.)
Theorem 3.24 (Carathéodory, 1907; Gale, 1956). The cyclic polytope M (d, n) ⊂ Rd
has n vertices and is bd/2c–neighborly.
(For a brief history of the discovery and the rediscovery of this result, see Section 7.4
(p. 127) of Grünbaum (1967) .)
For the proof of the theorem we need a lemma which will be used again, for another
purpose, in the next two sections. We shall use the term “linear hyperplane” to indicate a
hyperplane through the origin, i. e., a subspace of codimension 1. (The term “hyperplane”
in general refers to affine hyperplanes.)
Lemma 3.25 (Moment Curve Kissing Lemma). Let d ≥ 1, 0 ≤ k ≤ d/2, and
α1 , . . . , αk ∈ R. Then there exists a linear hyperplane P such that the moment curve
md+1 ⊂ Rd+1 lies entirely on one side of P , and, of all points of the moment curve, P
contains precisely md+1 (αi ), i = 1, . . . , k.
Proof. The linear hyperplane P will be defined by the homogeneous linear equation cx = 0
for some row vector c = (γ0 , γ1 , . . . , γd ) ∈ Rd+1 , c 6= 0. We shall have to select c such that

(i) c · m(ξ)T > 0 for every ξ ∈ R, ξ ∈

/ {α1 , . . . , αk };

(ii) c · m(αi )T = 0 for i = 1, . . . , k.

Let f be the polynomial of degree k whose roots are the αi :

k
Y
f (ξ) = (ξ − αi ).
i=1

1
See the footnote to Definition 3.2.

74 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 3.2. CONVEXITY
2
Define the γi to be the coefficients of the polynomial f :

2
γ0 + γ1 ξ + · · · + γd ξ d = f (ξ) .

(This makes sense because 2k ≤ d.) With this choice of the γi we have

2
c · m(ξ)T = f (ξ) for every ξ ∈ R

from which both requirements follow immediately.

Before proceeding to the proof of Theorem 3.24, we restate the essence of the Lemma in
matrix language.
Corollary 3.26. Let 1 ≤ d ≤ n − 1 and 0 ≤ k ≤ d/2. Then there exists a matrix
A ∈ Rn×(d+1) with the following properties. The rows of A are in general position (every
d + 1 of them are linearly independent). Moreover, for every k-subset I ⊂ [n] there exists a
vector c ∈ Rd+1 such that, letting cAT = (β1 , . . . , βn ), we have βi = 0 if i ∈ I and βi > 0 if
i∈/ I.
Indeed, let A be the matrix with rows md+1 (α1 ), . . . , md+1 (αn ) for any n distinct reals αi ;
and let c be the normal vector of the linear hyperplane corresponding to the set {md+1 (αi ) :
i ∈ I}.
Proof of Theorem 3.24. Let m0d (α1 ), . . . , m0d (αk ) be arbitrary k ≤ d/2 points from the set
which was used in the definition of the cyclic polytope M (d, n) ⊂ Rd . (We have carefully
avoided calling these points “vertices of M (d, n)”; this fact will be a consequence of what
we prove below.) We claim that there exists a hyperplane P 0 such that M (d, n) lies entirely
on one side of P 0 , and the intersection M (d, n) ∩ P 0 is conv{m0d (α1 ), . . . , m0d (αk )}. For k = 1
this proves that the αi are indeed vertices; and it proves k–neighborliness for every k ≤ d/2.
In order to verify the claim, we just have to find a hyperplane P 0 such that m0d (αi ) ∈ P 0
for i = 1, . . . , k, and all the remaining m0d (αj ) (j = k + 1, . . . , n) lie strictly on one side of
P 0.
This is accomplished by taking the linear hyperplane P ≤ Rd+1 constructed in the pre-
vious lemma, intersecting it with the hyperplane x0 = 1, and omitting the first component
of every vector. (As above, we use the notation (x0 , . . . , xd ) for the points in Rd+1 .)
Our conclusion is that the d-dimensional cyclic polytope has as many faces as conceivable in
every dimension ≤ bd/2c − 1. What happens in higher dimensions? For k > d/2, the cyclic
polytope is no longer k-neighborly, but can any other polytope with the same number of
vertices beat the cyclic polytope? No polytope can beat it for any k: the cyclic polytope is
the winner simultaneously in all dimensions. This was Theodore Motzkin’s celebrated Upper
Bound Conjecture which he stated in 1957. The conjecture was confirmed 13 years later by
P. McMullen (1970).
Theorem 3.27 (Upper Bound Theorem) (McMullen, 1970). No d-dimensional
polytope with n vertices has more k-dimensional faces than M (d, n), for any d, n, k (−1 ≤
k ≤ d ≤ n − 1).

——————————————————————— 75

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 3. GENERAL POSITION
Richard Stanley (1975) has shown in one of the most impressive combinatorial applications
of commutative algebra that the result remains valid among all simplicial complexes home-
omorphic to the sphere SS d−1 . A relatively simple elementary proof of an intermediate
generalization of McMullen’s result (to “shellable complexes”) was given by N. Alon and G.
Kalai (1985). At the heart of their proof is a Helly-type extremal theorem on finite sets, one
of the central results to be discussed in this book (Theorem 5.6, see also Theorem 6.12).

3.2.4 Distributing points on the sphere

In this section we shall find out, following Gale (1956), how to distribute 2k +d points “fairly
evenly” on a sphere. An application of this result to the chromatic theory of graphs will
follow in Section 2.4.
It is easy to select 2k+1 points on the circle (the 1-sphere) such that every open semicircle
contain at least k of them: just take the vertices of a regular (2k + 1)-gon. But how does
one select 2k + 2 points on the 2-sphere (the ordinary sphere) with the property that every
open hemisphere contain at least k of them? No such regular arrangements exist. And why
2k + 2? Any 2 points will be on a great circle, so we need at least 2k + 2 points in order to
have k in the interior of each hemisphere bounded by this great circle. D. Gale discovered,
that 2k + 2 points suffice, and the analogous statement is true in every dimension.
The r-sphere SS r ⊂ Rr+1 is defined as the set of vectors of unit length in Rr+1 :

SS r = {x ∈ Rr+1 : kxk = 1.}

(This set is the boundary of the ball B r+1 .) An open hemisphere is defined as the part of
SS r lying strictly on one side of a hyperplane across the origin:

{x ∈ SS r : aT x > 0}

for some a ∈ Rr+1 , a 6= 0.

Theorem 3.28 (Gale, 1956). For every m, r ≥ 0 there exists an arrangement of 2m + r
points on the r-sphere such that every open hemisphere contains at least m of them.
Proof. Let n = 2m + r. We have to find nonzero vectors v1 , . . . , vn ∈ Rr+1 such that for
every nonzero vector x ∈ Rr+1 , at least m of the inequalities viT x > 0 (i = 1, . . . , n) hold.
Indeed, given such vi , dividing each by its length we obtain an appropriate set of points on
the sphere.
We construct the vi in the following somewhat mysterious way. Let d = 2m − 2. First we
take an n × (d + 1) matrix A with the properties guaranteed by Corollary 3.26. This matrix
has full column rank. Let U denote its column space. Let B denote an n × (n − d − 1)
matrix whose columns form a basis of U ⊥ (in R(n−d−1) ). Call the rows of B v1 , . . . , vn . Note
that now n − d − 1 = r + 1, so the vi certainly belong to the right space (Rr+1 ). We must
prove they are “evenly distributed in every direction” in the sense of the first paragraph of
this proof.
First we note that B, too, has full column rank, so Bx 6= 0 for any nonzero x ∈ Rr+1 .
We have to prove that for such an x, the number of positive entries in Bx is at least m.

76 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 3.2. CONVEXITY
T
Assume for a contradiction that z = (ζ1 , . . . , ζn ) := Bx has at most m − 1 positive
entries; let I ⊂ [n] be the corresponding index set. So, k := |I| ≤ m − 1 = d/2; ζi > 0 if
i ∈ I; and ζi ≤ 0 if i ∈/ I. Let c ∈ Rd+1 be a row vector with the property guaranteed by
Corollary 3.26 for this particular set I. Setting b = (β1 , . . . , βn ) := cAT , this property is that
βi = 0 if i ∈ I and βi > 0 if i ∈
/ I.
Let us now observe that the perpendicularity of the columns of A and B is expressed by
the equation AT B = 0. Consequently bz = cAT Bx = 0. But
X X
bz = βi ζi + βi ζi .
i∈I i∈I
/

Each term in the first sum is zero (βi = 0); and each term in the second sum is the product
of a positve βi and a nonpositive ζi . We conclude that each term is zero, therefore ζi = 0 for
i∈/ I.
But this is too many zeros in z. Indeed, we have AT z = AT Bx = 0, which means
that a linear combination of k rows of A, with coefficients ζi (i ∈ I) is zero. But the
rows of A are in general position, and now we are talking about a linear relation among
k ≤ m − 1 = d/2 < d + 1 of the rows. The conclusion is that z = 0. This, however,
contradicts z = Bx 6= 0, a fact recorded at the start of the proof.
We remark that the fact that the rows of the matrix A are in general position is not crucial
for this proof (Exercise 3.2.5) and indeed Corollary 3.26 would suffice for the proof even if
this condition is omitted.

3.2.5 Borsuk’s and Kneser’s graphs

The diameter of a set S ⊂ Rd is defined as supx,y∈S kx − yk.
In an influential paper published in 1933, Polish geometer and topologist Karol Borsuk
considered the problem of decomposing d-dimensional bodies of finite diameter into subsets
of strictly smaller diameter. His question was to find the minimum number of pieces required.
He conjectured that d + 1 pieces always suffice, a problem to cause a great deal of headache
to generations of geometers, until refuted very recently by combinatorialists. We shall tell
this story in detail in Section 5.6.
It is clear that fewer than d + 1 pieces will not suffice in general; just take the vertices
of the regular simplex. A much less obvious example is the sphere. It is easy to see that
the unit sphere SS d−1 can be dissected into d + 1 pieces, each of diameter strictly less than
2. (Verify!) A difficult topological argument of K. Borsuk (1933) demonstrates that this is
best possible. A simpler proof was found by B. Weiss (1989).
Theorem 3.29 (Borsuk’s Theorem). If we partition the unit sphere SS d−1 ⊂ Rd into
≤ d sets, then the diameter of at least one of the pieces is 2.
Two points of the unit sphere x, y ∈ SS d−1 ⊂ Rd are antipodal if x + y = 0. Given  > 0,
the two points are -nearly antipodal if kx + yk < .
Another way of stating Borsuk’s Theorem is that no matter how we partition the unit
sphere into ≤ d sets, at least one of these sets will contain -nearly antipodal points for every

——————————————————————— 77

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 3. GENERAL POSITION
 > 0.
Erdős and Hajnal (1966) observed that this result can be restated in terms of the chro-
matic number of a graph. They define Borsuk’s graph B(d, ) to be an infinite graph with
vertex set SS d−1 ; two points are adjacent if they are -nearly antipodal.
It is easy to see that for small enough , B(d, ) has a (legal) coloring with no more than
d + 1 colors (Exercise 3.2.8). Borsuk’s Theorem is equivalent to saying that for any  > 0,
this number of colors is best possible.
Theorem 3.30 (Borsuk’s Theorem restated). For any  > 0, the chromatic number of
Borsuk’s graph B(d, ) is at least d + 1.
Hence, for sufficiently small , χ(B(d, )) = d + 1. We shall not prove this theorem here.
We stated it for the sake of an interesting application.
Observe that for small , the odd girth (length of the shortest odd cycle) of B(d, )
becomes large (Exercise 3.2.9).
If the reader is somewhat disturbed by the fact that Borsuk’s graph is infinite, we should
mention another result: every infinite graph of finite chromatic number k has a finite sub-
graph of chromatic number k.
This is a common particular case of Gödel’s Compactness Theorem in first order logic
as well as of Tikhonov’s Theorem that the product of infinitely many compact spaces is
compact. The paper of de Bruijn and Erdős (1951) was the first one to explicitly state this
fact.
In 1955, M. Kneser proposed another interesting family of graphs without short odd
cycles and with suspected large chromatic number.
Definition 3.31. For n ≥ 2m + 1, the vertex set of Kneser’s graph K(n, m) is [n]


m
, the
[n]


set of m-subsets of [n] = {1, . . . , n}. Two vertices A, B ∈ m are adjacent if A ∩ B = ∅.

Observe that Petersen’s graph, familiar from Section 1.5, is one of the Kneser graphs,
K(5, 2).
Setting n = 2m + r (r ≥ 1), it is easy to see that Kneser’s graph has a (legal) coloring
with r +2 colors. Kneser conjectured that this was the precise chromatic number in all cases.
Erdős and Hajnal (1966) point out some similarities between Kneser’s and Borsuk’s
graphs. If we think of “distance” between k-sets being the size of their symmetric difference
(this is called Hamming distance) then adjacency in Kneser’s graph corresponds to “antipo-
dality” (largest Hamming distance). If r is small, it is easy to see that the odd girth of
Kneser’s graph is large (Exercise 3.2.11).
This intuition turned out to be fully justified. In 1978, Lovász devised a clever topological
construction, which associates high dimensional topological spaces with every graph such
that adjacency would correspond to a notion of antipodality in Lovász’s space; and showed
how to “find” a copy of Borsuk’s sphere inside the space corresponding to Kneser’s graph.
From this he inferred that the chromatic number of Kneser’s graph was no less than Borsuk’s,
thus proving Kneser’s Conjecture.
Theorem 3.32 (Kneser’s Conjecture) (Lovász, 1978). The chromatic number of

78 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 3.2. CONVEXITY
Kneser’s graph K(2m + r, m) is r + 2.
Within weeks of Lovász’s announcement, I. Bárány noticed that a much simpler proof
of Kneser’s Conjecture follows from Gale’s Theorem, combined with Borsuk’s. This is the
proof we are now able to present.
Proof. (I. Bárány, 1978) We only have to prove that χ(K(2m + r, m) ≥ r + 2.
Let G ⊂ SS r be a Gale–set of 2m + r points, i. e., a set with the property that every
open hemisphere contains at least m points from G (Theorem 3.28). It is easy to show that
for some δ > 0, this property√remains valid even if we exclude the “rim” of width δ of each
hemisphere: within distance 2 − δ of every point in SS r , there will be
√ at least m points of
G. (The distance of the perimeter of a hemisphere from its center is 2.)
Suppose we have a legal coloring of K(2m + r, m) with k colors. From this, we construct
a legal k-coloring of Borsuk’s graph B(r + 1, ) for some small but positive . This implies
k ≥ r + 2 by Borsuk’s Theorem.
Let us label the points of G by {1, . . . , 2m + r}. We can √
thus think of the vertices of
K(2m+r, m) as k-subsets of G. Let x ∈ SS r . Within distance 2−δ of x there is a k-subset
A of G. This set, viewed as a vertex of Kneser’s graph, has just been assigned a color; let
us give x the same color.
We have to see that adjacent points in Borsuk’s graph did not receive the same color.
Suppose x, y ∈ SS r got the same color. This means their corresponding k-sets √ A and B
must intersect. If z ∈ A ∩ B, this means that both x and y are within distance 2 − δ of
z, and therefore they cannot be near antipodes. (The value of  implicit in this sentence is
approximately 2δ, so a choice  = δ will certainly be good for small enough δ.)

Exercises
♦Ex. 3.2.1. Explain how the set of facets of a full-dimensional simplex demonstrates that Helly’s
Theorem is tight.

Ex. 3.2.2. Prove that the n-dimensional ball

B n = {x ∈ Rn : kxk ≤ 1}
√
is convex. Find its faces. (Here, kxk = xT x is the Euclidean norm.)

Ex. 3.2.3. Let S be a finite subset of Rn ; set m = |S|. Prove: (a) conv(S) has at most 2m faces.
(b) If it has 2m faces then S is affine independent and therefore conv(S) is a simplex.
Hint. (a) follows from the fact that faces of M := conv(S) are convex hulls of subsets of S. (b)
The condition implies that M has a chain of faces ∅ = F0 ⊂ F1 ⊂ · · · ⊂ Fm = M . It follows that
dim(M ) ≥ m − 1; therefore S is affine independent.

Ex. 3.2.4. Let M be a convex polyhedron (i. e., polytope in R3 ) with n vertices. Prove that M
has at most 3n − 6 edges. Which polyhedra attain this bound?
Hint. Use Euler’s Formula: If a convex polyhedron has n vertices, m edges, and f facets, then
n − m + f = 2. Equality will occur when all facets are triangles.

——————————————————————— 79

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 3. GENERAL POSITION
Ex. 3.2.5. Finish the proof of Theorem 3.28 without using the property that the rows of A are in
general position.
Hint. All we needed was that no k ≤ m − 1 = d/2 rows of A have a linear relation with positive
coefficients. Actually, this cannot happen with any number of rows. Indeed, assume AT z = 0 for
some nonnegative z. Apply Corollary 3.26 with I = ∅ to obtain c0 such that all entries of c0 AT be
positive. Now c0 AT z = 0 is a sum of nonnegative numbers, therefore z = 0.
Remark. The significance of this observation is that it allows a direct conversion each way between
“Gale-sets” on the sphere (Theorem 3.28) and highly neighborly polytopes (Theorem 3.24).

* * *

Ex. 3.2.6. Prove that Borsuk’s Theorem is equivalent to the following statement.
If SS d−1 = F1 ∪ . . . ∪ Ft , where each Fi is a closed set in Rd of diameter less than 2, then
t ≥ d + 1.
Hint. A set and its (topological) closure have the same diameter.
♦Ex. 3.2.7 (Squashing the sphere) (Borsuk, 1933). Prove that Borsuk’s Theorem follows from the
following statement (also proved by Borsuk).
If f : SS d−1 → Rd−1 is a continuous map then there exist antipodal points x, y ∈ SS d−1 such
that f (x) = f (y).
Ex. 3.2.8. Prove that for small enough  > 0, Borsuk’s graph B(d, ) can indeed be colored by
d + 1 colors.
Hint. Do it for small dimensions. Project an inscribed simplex onto the sphere; color the image of
each facet by a separate color. How small does  have to be?
Ex. 3.2.9. Prove: the odd girth of B(d, ) is

1 + 2dπ/(2 arcsin )e.

Note that for small  this quantity has the growth rate of Θ(1/).
Ex. 3.2.10. Prove: Kneser’s graph K(2m + r, m) has a legal coloring with r + 2 colors.
Hint. Use “greedy coloring”: color as many k-sets by color 1 as you can. (Choose a point, and
assign color 1 to all sets containing this point.) Proceed in this fashion until there are only 2m − 1
points left. The remaining set can all receive a single color.
Ex. 3.2.11. Prove: the odd girth of Kneser’s graph K(2m + r, m) is 1 + 2dm/re.
Ex. 3.2.12** (A. Schrijver, 1978). Take a cycle of length 2m + r. Let W denote the set of
independent sets of size m in this graph. View this set as a subset of the vertex set of Kneser’s
graph K(2m + r, m). Prove: the subgraph of Kneser’s graph induced on W has the same chromatic
number, r + 2, as the entire graph.

3.2.6 Linear and statistical independence

TO BE WRITTEN

80 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
Chapter 4

Set systems with restricted

intersections

4.1 When all intersections are equal size

How many subsets of a set of cardinality n can pairwise share the same number of elements?

Two distinct lines of thought converged on this problem in the late 1940s, giving birth
to both the method and the concept treated in a large part of this book.
Statistician R. A. Fisher, while working on the “design of experiments,” made the great
discovery. In a short section of a long paper in the British Annals of Eugenics in 1940 he
proved the surprising fact that in a BIBD, (balanced incomplete block design), the number
of blocks is never less than the number of points. Fisher’s proof uses a clever counting
argument (second moment), based on the regularity conditions satisfied by a block design.
It was Indian-born mathematician R. C. Bose who recognized a few years later that the
validity of Fisher’s inequality extends to far more general circumstances. He demonstrated
that, if every pair of sets in a uniform family has equal intersection size, then the number of
sets does not exceed the number of points. (How does this imply Fisher’s original inequality?
See Exercise 4.1.2.) Even more significantly, Bose’s seminal two-page note, published in 1949
in the Annals of Mathematical Statistics, introduced the technique which we call the “linear
algebra bound” method, abundantly employed in this volume.
The affiliation listed on Bose’s paper is the Institute of Statistics, University of North
Carolina. Before taking up residence in the U.S. in 1948, Bose worked at the Indian Statistical
Institute in Calcutta. One of the most influential combinatorialists of the decades to come,
Bose was forced to become a statistician by the lack of employment chances in mathematics
in his native country. A pure mathematician hardly in disguise, he reared generations of
combinatorialists. His students at Chapel Hill included D. K. Ray-Chaudhuri, a name that
together with his student R. M. Wilson (so, maybe a grand student of Bose?) will appear
several dozen times on these pages for their far reaching extension of Bose’s method.

c László Babai and Péter Frankl. 1988, 1992, 2020.

81
CHAPTER 4. RESTRICTED INTERSECTIONS
Apparently at the same time as Bose was about to change our view on Fisher’s inequality,
P. Erdős arrived at another variation of the same problem. Motivated by a graph theory
result of P. Turán (1941) and some combinatorial number theory results of his own, Erdős
began to outline the scope of what has since become extremal set theory. Erdős’s numerous
results and innumerable questions have practically created the entire field. One of the first
problems Erdős raised in this direction and solved in a joint paper with N. G. de Bruijn
(1948) was this: Maximally how many subsets of an n-set can have pairwise precisely one
common element ?
They found that the answer was n and characterized the extremal set systems as sun-
flowers or possibly degenerate finite projective planes.
Note that the de Bruijn–Erdős Theorem does not make any uniformity assumption.
The following result, which it seems quite right to call the Nonuniform Fisher Inequality,
subsumes both this and Bose’s settings.
Theorem 4.1 (Nonuniform Fisher Inequality). Let C1 , . . . , Cm be distinct subsets of
a set of n elements such that for every i 6= j, | Ci ∩ Cj |= λ where 1 ≤ λ < n. Then m ≤ n.
The proof of this result is an adaptation of Bose’s, found by Majumdar (1953) and
rediscovered by Isbell (1959). The combinatorial method of de Bruijn and Erdős does not
seem to generalize to the case λ ≥ 2. Nor does their result, the characterization of the set
systems attaining the upper bound (m = n) have an analog for λ ≥ 2. (Some information
on those extremal systems is given in Exercises 4.1.7 through 4.1.10 below.)
Just as for the “Oddtown Theorem” (Corollary 1.2), we prove that under the conditions
of Theorem 4.1, the incidence vectors of the sets Ci are linearly independent over the reals.
The trick in proving this, however, is quite different; modular arguments don’t seem to work.
First we separate the case when one of the sets has λ elements. Then all the other sets
contain this one and are disjoint otherwise. It follows that m ≤ n + 1 − λ ≤ n.
def
Henceforth we may assume that all numbers γi = |Ci | − λ are positive.
Let M be the incidence matrix of the set system. Our intersection condition is summa-
rized in the matrix equation
A = M M T = λJ + C (1)
where J is the m × m all-ones matrix and C is the diagonal matrix C = diag(γ1 , . . . , γm ).
What we have to prove is that the rank of A is m. (Then, as in Section 1.1, m = rk A ≤
rk M ≤ n follows.) In the uniform case, it is easy to directly compute the determinant of
A and check that it is not zero. (This is the way Bose’s proof went; cf. Exercise 4.1.3.) It
is actually possible to compute the determinant in the general case as well (Exercise 4.1.4),
but there is a much more elegant and conceptually interesting way to finish the proof. All
it takes is to recall a familiar definition from linear algebra.
A symmetric m × m matrix B with real entries is positive semidefinite if for any x ∈ Rm ,
the quadratic form xBxT is nonnegative. If, in addition, the only case when xBxT vanishes
is when x itself is zero then B is positive definite. Obviously, a positive definite matrix must
have full rank ; otherwise a nontrivial solution of the homogeneous system of linear equations
BxT = 0 makes the quadratic form vanish. Moreover, it is immediate from the definition
that the sum of a positive definite and a positive semidefinite matrix is positive definite.

82 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 4.1. ALL INTERSECTIONS EQUAL
Therefore, in order to complete the proof of the Nonuniform Fisher Inequality, we just have
to observe that λJ is positive semidefinite and C is positive definite.
m
PmLetPxm= (x1 , . . . , xm ) ∈ R . For theT generic m × m matrix U = (µij ), we have xU xT =
2 T 2 2
i=1 j=1 µij xi xj . In particular, xλJx = λ(x1 + · · · + xm ) and xCx = γ1 x1 + · · · + γm xm ,
justifying both claims.

Exercises

Ex. 4.1.1 (J. A. Bondy, 1972). Let A1 , . . . , An be n distinct subsets of a set X of n elements.
Prove that for some x ∈ X, all the sets Ai \ {x} are distinct. (a) Give a combinatorial proof. (b)
Give a linear algebra proof. (c) Prove that the conclusion remains valid if |X| ≥ n but becomes
false if |X| ≤ n − 1.
Hint. (a) Find two combinatorial solutions in Lovász (1979c), Ch. 13, Probl. 13.10. (b) Let X = [n].
Let vi = (αi1 , . . . , αin ) ∈ {0, 1}n be the incidence vector of Ai , and M = (αij )ni,j=1 be the incidence
matrix. The condition is that all rows of M are different. We have to prove that this remains true
after deletion of an appropriate column of M . Case 1: det(M ) = 0. In this case some column is
linearly dependent on the others. Prove that deleting this column from M leaves no equal rows.
Case 2: det(M ) 6= 0. Let Ak be smallest among the Ai . Expand det(M ) by the k th row. Conclude
that for some j, the term αkj det(Mkj ) 6= 0, where Mkj is the (n − 1) × (n − 1) minor obtained by
deleting the k th row and the j th column from M . In particular, αkj = 1 and no two rows of Mkj
are identical. Prove that deleting column j from M leaves no equal rows.

Ex. 4.1.2. How does Fisher’s inequality follow from Bose’s?

Hint. Apply Bose’s result to the dual of Fisher’s block design: view the transpose of the incidence
matrix of the block design as the incidence matrix of a set system, thus switching the roles of points
and blocks.

Ex. 4.1.3. Reproduce Bose’s proof: compute det(λJm + γIm ).

Hint. The result is (γ + mλ)γ m−1 . Under our assumptions γ, λ ≥ 1, this quantity is not zero.

Ex. 4.1.4. Following Majumdar (1953) and Isbell (1959), finish the proof of Theorem 4.1 by
explicitly computing the determinant of the m × m matrix A = λJm + C.
Hint. Extend A to an (m + 1) × (m + 1) matrix by adding a first row of all ones and a first column
of all zeros except the 1 in the top left entry. (This is an example of the often used “bordering
trick.”)
 
1 1 1 ... 1
 0 λ + γ1 λ ... λ 
 
0 λ λ + γ2 . . . λ 
 
 .. .. .. .. .. 
. . . . . 
0 λ λ . . . λ + γm .

——————————————————————— 83

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 4. RESTRICTED INTERSECTIONS
Subtract λ times the first row from each row. This will produce a lot of zeros.
 
1 1 1 ... 1
−λ γ1 0 . . . 0 
 
−λ 0 γ2 . . . 0 
 
 .. .. .. . . .. 
 . . . . . 
−λ 0 0 . . . γm .

Using the diagonal entries, kill the first column to create an upper triangular matrix. Compute the
product of the diagonal. The result is
  
1 1 1
γ1 γ2 . . . γm 1 + λ + + ··· + . (2)
γ1 γ2 γm

♦Ex. 4.1.5. Give a direct proof of the linear independence of the incidence vectors of the Ci ,
straight from the definition of linear independence.

Ex. 4.1.6. Reproduce the de Bruijn–Erdős proof (λ = 1). Prove the following lemma first.
Lemma. Let F be a family of m sets over the universe [n]. The degree deg(x) of a point x ∈ [n]
is the number of members of F containing x. Suppose deg(x) < m and |E| < n for every x ∈ [n]
and E ∈ F. Assume further that for every x ∈
/ E, deg(x) ≤ |E|. Then m ≤ n.
Hint to the lemma. For a contradiction, suppose m > n. Take the sum of the inequalities
deg(x) |E|
<
m − deg(x) n − |E|

for all pairs (x, E) such that x ∈

/ E.
Remark. Observe the similarity between the expressions occurring in the solutions of the last three
exercises.

* * *

In the sequence of problems below we consider the structure of the set systems satisfying
the conditions of the Nonuniform Fisher Inequality with m = n. If such systems do exist for
a given value of n and λ (such as for λ = 1 and every n), they are “extremal” in the sense
that no other system with the same n and λ has more members.
Ex. 4.1.7. Prove the de Bruijn–Erdős characterization of the extremal cases for λ = 1.
Hint. After the results mentioned in Section 2.4.3, we only have to prove that in an extremal set
system, there exists a set incident to each pair of points.

Ex. 4.1.8.* Prove for all λ ≥ 1, if m = n holds then there exists a set incident to each pair of
points. (Note that for λ ≥ 2 this does not lead to a characterization of the cases with m = n.)
References. Ryser (1968), Woodall (1970), Seress (1989). For a simple direct proof, see Babai
(1987).

84 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 4.2. CONSTRUCTIVE RAMSEY THEORY
Ex. 4.1.9.* Prove, for any λ: if a set system satisfying m = n is uniform then it is regular as well
and consequently it is a symmetric design. (Cf. Section 2.4.3.)
Reference. Ryser (1950).
Ex. 4.1.10.** Prove that in a nonuniform extremal set system (“λ-design”), the sets will have two
different sizes only. Construct examples of such families (“point-complemented symmetric block
designs”).
Hint. The point-complemented symmetric block designs are obtained from a symmetric block
design B by selecting a block B0 and replacing all the other blocks B by the symmetric difference
B ⊕ B0 .
References. Ryser (1968), Woodall (1970).
J. H. Ryser (1968) states what has become known as the λ-design conjecture: All nonuniform
extremal set systems are point-complemented symmetric block designs.
References. Bridges (1977), Seress (1989).

4.2 Ramsey theory – a constructive lower bound

Informally speaking, the subject of Ramsey theory is to demonstrate that any sufficiently
large configuration, no matter how irregular it may seem, contains certain prescribed pat-
terns. Upper and lower Ramsey bounds describe how large the configuration has to be to
ensure the occurrence of the given pattern. The configurations in question usually arise by
partitioning (coloring) a familiar object into a given number of classes (the colors). The
patterns we are looking for are usually required to be monochromatic, i. e., to belong entirely
to one of the color classes.
For instance, if we color the Euclidean plane by 2 colors then at least one of the color
classes will contain the three vertices of a right triangle with given sidelengths. (We shall
see more of what is called Euclidean Ramsey Theory in Section 7.2.)
A difficult early result of this kind is van der Waerden’s Theorem which we can only state
here. For proofs and generalizations we refer to the book by Graham–Rothschild–Spencer
(1980) .
Theorem 4.2 (B. L. van der Waerden, 1927). If we color the set of natural numbers by
a finite number of colors, one of the color classes will contain arbitrarily long finite arithmetic
progressions.
A typical first exercise in Ramsey Theory is the following.
Proposition 4.3. If we color the 15 edges of a complete graph on 6 vertices red and blue
in any way, a monochromatic triangle will necessarily arise.
(A monochromatic triangle is a set of three vertices such that the three edges between
them have the same color; see Exercise 4.2.2.)
A game is based on this simple result. It is called the Ramsey game and is played by two
players, Red and Blue. They take turns to draw lines in red and blue, respectively, between
6 given points on a piece of paper. The game terminates in at most 15 moves (in each move,
a new pair must be connected). Whoever first completes a triangle of his own color loses.

——————————————————————— 85

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 4. RESTRICTED INTERSECTIONS
The result of Proposition 3.3 guarantees that no ties are possible. With just 5 points this
would not be the case.
If we color the edges of the complete graph Kn by 3 colors, n ≥ 17 will guarantee
a monochromatic triangle (Exercise 3.2.3). For any k there exists a (smallest) number
n = Rk (3) such that if we color Kn by k colors then a monochromatic triangle must arise.
The exact value of Rk (3) is not known for k ≥ 4. (A bound, published by I. Schur in 1916,
is given in Exercise 4.2.5. This was the first result in “Ramsey Theory,” predating Ramsey’s
paper by 14 years.)
The pattern we may be looking for may be other than a triangle. We may even look for
a different pattern in one color than in the others. Let R(s, t) denote the smallest n such
that coloring Kn red and blue in any way there must either be a red Ks or a blue Kt . For
instance, the Ramsey game is based on the result that R(3, 3) = 6.
Frank Plumpton Ramsey, the brilliant British logician, economist, and philosopher, pub-
lished in 1930 (shortly before his tragically early death at 28) a proof that all the numbers
R(s, t) (and much more general “Ramsey numbers,” cf. Exercise 4.2.6) exist. This result
and its infinite version had a profound effect on a number of branches of mathematics.
A few years later Pál Erdős learned about Ramsey’s Theorem and embraced it with great
enthusiasm.1 Erdős was 20 at the time. This theorem became one of his obsessions, an ever-
recurring theme in his innumerable conjectures and proofs. These conjectures and proofs
are largely responsible for the creation of fields like combinatorial set theory, the theory of
large cardinals, combinatorial geometry and combinatorial number theory.
In order to solve a combinatorial problem of Erdős and Eszter (Esther) Klein in plane
geometry (Exercise 4.2.8) and unaware of Ramsey’s work, György (George) Szekeres, then an
undergraduate in chemical engineering at the Technical University in Budapest, rediscovered
Ramsey’s Theorem (and subsequently married Eszter Klein, hence the name “happy end
problem”). Erdős and Szekeres were the first, in 1935 to derive explicit upper bounds for
Ramsey numbers.
For the special case of R(s, t) they proved
Theorem 4.4 (Erdős–Szekeres, 1935). No matter how we color the edges of a complete
graph on s+t−2


s−1
vertices red and blue, there will either be a complete red subgraph on s
vertices, or a complete blue subgraph on t vertices. In other words,
 
s+t−2
R(s, t) ≤ .
s−1

(See Exercise 4.2.7.)

In particular,  
2t − 2
R2 (t) := R(t, t) ≤ < 4t .
t−1
1
A warm and lively account of how this happened, in the course of weekly excursions of a small circle
of gifted students escaping to the hillside under the shadow of an increasingly menacing social climate
in Budapest, 1933, was given by György Szekeres in a foreword to Erdős’s selected combinatorial works
(“Reminiscences,” in: P. Erdős, The Art of Counting (J. Spencer, ed.), M.I.T. Press 1973, pp. xix–xxii.)

86 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 4.2. CONSTRUCTIVE RAMSEY THEORY
In spite of the simplicity of the proof of this bound and the considerable effort spent in search
for improvements, it is still an open question whether or not 4t could be replaced by (4 − )t
on the right hand side for some positive
√ . It is known that (even for large t) the number 4
cannot be replaced by anything < 2: according to another result of Erdős,
Theorem 4.5 (Erdős, 1947). There exists a two-coloring of the edges of the complete
graph on N = b2t/2 c vertices such that there will be no monochromatic complete subgraph
on t vertices. In other words,
R2 (t) > 2t/2 .
(See Exercise 4.2.9.)
The surprising aspect of Erdős’s proof is that it does not construct any such coloring; it
merely proves that such colorings exist. In fact, they exist in abundance: if we color Kv at
random, flipping a coin for each edge to decide its color, we have excellent chances to obtain
a two-coloring with no Kt in either color. (The probability that this will be the case tends
to 1 as n → ∞.) This paper of Erdős, modestly entitled “Some remarks on the theory of
graphs,” marks the advent of the probabilistic method, now recognized as one of the most
powerful tools in combinatorics and related fields (combinatorial number theory, theory of
computing, etc.).
Erdős’s lower bound has not been improved significantly over √ four decades. It is still an
t
open problem whether or not R2 (t) ≥ c for some constant c > 2 (cf. Spencer (1977)).
A perhaps even more intriguing problem is to match the probabilistic lower bound by
explicit construction. Questions of this kind (replacing probabilistic existence proofs with
explicit constructions) have particular importance to the theory of computing. In striking
contrast to the simplicity of some of the probabilistic arguments, the constructive proofs as
a rule require involved tools, mostly from algebra and number theory. And, very often, the
constructive results don’t come anywhere near the power of the existence proofs. The case of
R2 (t) illustrates this point. For many years, only an easy (t − 1)2 constructive lower bound
was known (Exercise 4.2.10). Then H. L. Abbott (1972) found a curious metaconstruction, by
giving a nonconstructive proof that for every fixed k and sufficiently large t, a constructive
lower bound of tk exists. We expand on this philosophically interesting idea in Exercise
4.2.11.
The first bona fide progress came with an Ω(t3 ) lower bound of Zsigmond Nagy in 1972.

Theorem 4.6 (Zs. Nagy, 1972). For v = 3t , there exists a constructive two-coloring of

the complete graph on v vertices without a Kt+1 in either color.

This result is particularly significant because, as we shall see presently, it brings extremal
set theory to bear on the subject of Ramsey bounds. This connection has eventually led
to constructive superpolynomial lower bounds, to be discussed in Section 5.7. (The term
“superpolynomial” refers to a function that grows faster than any polynomial of the variable,
i. e. a function of the form f (t) = tg(t) , where limt→∞ g(t) = ∞.

Proof. Let us identify the set of vertices of Kv with the set of triples from X = {1, . . . , t}.
Let us join two such triples by a red edge if they have precisely one element in common, blue

——————————————————————— 87

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 4. RESTRICTED INTERSECTIONS
otherwise.
Suppose m triples are pairwise joined by the same color. We have to prove that m ≤ t.
If the color is red, all the pairs have precisely one element in common, so m ≤ t follows
from Theorem 4.1.
If the color is blue, all pairwise intersections must have size 0 or 2 so m ≤ t follows from
the Oddtown Theorem (Corollary oddtownth).

Exercises
♦Ex. 4.2.1. Prove the following handy estimates for binomial coefficients (1 ≤ s ≤ n; e is the
base of the natural logarithm.)
 n s n  en s
≤ < . (3)
s s s

Ex. 4.2.2. (a) Prove Proposition 4.3. (b) Show that 5 points do not suffice.
Remark. This problem was proposed at a high school mathematics contest in Hungary, ** year (cf.
Hungarian Problem Book ***)

Ex. 4.2.3. The sheriff of Brave New County closely monitors the correspondences between county
residents. If he finds that three of them correspond with each other on the same subject, he arrests
them on charges of conspiracy. - There are three current hot topics in science, and the 17 scientists
living in the county eagerly communicate with each other on these topics: each corresponds with
every other on one of the three topics. Prove: no matter how they choose the topics of their
correspondence, they won’t be able to escape the sheriff’s wrath. Hint. Note that this is the
3-color version of Proposition 4.3: we take the complete graph K17 and color its edges with three
colors (the three “topics of correspondence”). We need to show that there will necessarily be a
monochromatic triangle.
Remark. This problem (about the correspondence of 17 scientists but without the sheriff) was
proposed at the ***th International Mathematics Olympiade, city, year **.

Ex. 4.2.4.* Show that 16 scientists can evade the sheriff: color the edges of K17 by three colors so
that no monochromatic triangle arises.
Hint. (Requires some abstract algebra not covered in Chapter 2.) We have to color the 16


2 pairs
of a set of 16 elements by colors called 0,1,2 such that no monochromatic triangle will arise.
Let the 16 points be the elements of F16 , the field of order 16. The multiplicative group F× 16 of
this field is cyclic of order 15 and therefore it has an onto homomorphism ϕ : F× 16 → Z3 . Let the
color of the pair {i, j} (i, j ∈ F16 , i 6= j) be ϕ(i + j).
Let H be the kernel of ϕ. H is cyclic of order 5. Reduce the problem of verifying that this
coloring has no monochromatic triangles to the following question:
If g is a generator of H then g does not satisfy any equation of the form 1 + g a + g b = 0 where
1 ≤ a < b ≤ 4.
In order to prove this, it suffices to note that since g 5 = 1 and g 6= 1, g satisfies the polynomial
x4 + x3 + x2 + x + 1, which is irreducible over F2 (Exercise 2.1.20).

88 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 4.2. CONSTRUCTIVE RAMSEY THEORY
Ex. 4.2.5 (I. Schur, 1916). Prove that in the k-color version of Proposition 4.3 1 + bek!c
vertices suffice to guarantee a monochromatic triangle. (Here e = 2.71828 . . . is the base of natural
logarithms.) In other words, Rk (3) < 1 + ek!.
Note that by Exercises4.2.2 and 4.2.4, this bound is tight for k = 2 and 3.
Hint. Proceed by induction on k. Prove the recurrence


Rk (3) − 1 ≤ 1 + k Rk−1 (3) − 1 . (4)

Ex. 4.2.6 (Ramsey’s Theorem—general case). Given r, k, t ≥ 1, there exists an integer Rkr (t)
N


such that if we color the r r-subsets of an N -element set by k colors in any way, then there will
be a monochromatic subset of size t, provided N ≥ Rkr (t). (A subset is monochromatic if all of its
r-tuples have the same color.)
Remark. A substantial portion of Ramsey’s original paper has been reprinted with comments along
with a fascinating biographical sketch in Graham–Rothschild–Spencer, Ramsey Theory, Wiley 1980,
Section 1.7.

Ex. 4.2.7. Prove the Erdős–Szekeres upper bound (Theorem 4.4)

Hint. The result is straightforward for s ≤ 2 and, by symmetry, for t ≤ 2. Assume s, t ≥ 3
and proceed by induction on s + t. Pick a vertex v and split the remaining vertices into two
classes according to the color joining them to v. Examining each class together with v, deduce the
inequality
R(s, t) ≤ R(s − 1, t) + R(s, t − 1). (5)
Apply the induction hypothesis to the terms on the right hand side.

Ex. 4.2.8. (a) (Esther Klein’s problem) Prove: among any 5 points in the plane in general
position (no three on a line), there are 4 which form the vertex set of a convex quadrilateral.
(b) Prove: for any k there exists n such that among any n points in the plane in general position
there exist k which form the vertex set of a convex k-gon. (Erdős–Szekeres, 1935).
Hint. Combine part (a) with Ramsey’s Theorem for quadruples.
(c) ***

Ex. 4.2.9. Prove Erdős’s lower bound (Theorem 4.5)

Hint. Let us fix a set V of N vertices. The number of red-and-blue colorings of the edges of KN
N
is 2( 2 ) . For a t-subset A ⊂ V , count the number of colorings that make the subgraph induced by
N t
A monochromatic. The result is 21+( 2 )−(2) . Adding up these numbers for all A ∈ Vt , conclude

from this that if  

N 1−( t )
2 2 < 1, (6)
t
then there exists a coloring without monochromatic Kt . Verify, using elementary estimates for the
N
binomial coefficients such as t < N t /t!, that inequality (6) does indeed hold for N = b2t/2 c.

Remark. Imagine that we are selecting the coloring by flipping a coin for each pair: heads = blue,
tails = red. Prove that the probability that a monochromatic Kt subgraph will occur is less than
the quantity on the left hand side of (6) (Indeed, the left hand side of (6) is precisely the expected

——————————————————————— 89

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 4. RESTRICTED INTERSECTIONS
number of monochromatic Kt subgraphs.) Show that for the given values of N , the left hand side
of (6) goes to zero as t → ∞. A random coloring thus has an excellent chance of being “good.”
♦Ex. 4.2.10. Prove the easy constructive lower bound: demonstrate the inequality R2 (t) > (t−1)2
by an explicitly constructed two-coloring.

Ex. 4.2.11 (H. L. Abbott, 1972). Prove that for every fixed k and all sufficiently large t there
exists an explicit construction of two-colorings of the complete graphs KN , N = tk , such that for
every sufficiently large t, the coloring obtained has no monochromatic Kt subgraph.
Hint. Let V1 and V2 be two sets. Let ϕi : V2i → {red,blue} denote a two-coloring of the complete

graph on the vertex set Vi . Define the lexicographic product of the two colorings as a coloring ψ =
ϕ1 ∗ ϕ2 of the complete graph on the vertex set V1 × V2 by setting, for (v1 , v2 ) 6= (w1 , w2 ) ∈ V1 × V2 ,
(
ϕ1 ({v1 , w1 }) if v1 6= w1 ;
ψ({(v1 , v2 ), (w1 , w2 )}) =
ϕ2 ({v2 , w2 }) if v1 = w1 .

Prove that, if the largest blue subset of Vi under ϕi has size ti then the largest blue subset of V1 ×V2
under ϕ1 ∗ ϕ2 has size t1 t2 . (Same for red.)
Now assume we have some value tk and a two-coloring ϕk of the complete graph on tkk vertices
with no monochromatic complete subgraph on tk vertices. Define the mth lexicographic power ϕm k
of ϕk . This is a coloring of the complete graph on tkm k vertices, without monochromatic complete
subgraphs on tm k vertices, an infinite sequence easily constructed from the single starting object.
Now comes the trick. How do we get hold of that starting object, one for each k? By Erdős’s
estimate, for each fixed k, such a coloring of the complete graph on tk vertices exists, as soon as
tk < 2t/2 , i. e., surely for some t < 3k log2 k. We thus know that for every k, a starter exists (never
mind how to find it; a finite search suffices anyway), with the consequence that for every k, a fully
explicit construction of an infinite family of appropriate colorings also exists.

Ex. 4.2.12. Prove that in Nagy’s coloring (Theorem 4.6)

(a) if t ≡ 2 or 3 (mod 4) then there is no blue Kr for r > t − 2;

(b) if t > 7 then there is no red Kr for r > (t − 1)/2.

Ex. 4.2.13* (M. Deza, 1973). Prove: if every pair of members in a k-uniform family F shares λ
k2
points, then either |F| ≤ − k + 1, or F is a sunflower, i. e., all the pairwise intersections are the
same λ-element set. Show that equality is attainable when a projective plane of order k − 1 exists.
Reference. Cf. Lovász (1979c), Exercise 13.17.

Ex. 4.2.14. Derive the result 4.2.12 (b) from Deza’s Theorem (preceding exercise).
[t]


Ex. 4.2.15. In Nagy’s coloring of the complete graph on vertex set V = (Theorem 4.6), two
3
triples A, B are joined by a red edge if |A ∩ B| = 1. This rule implies that the largest blue clique
had size ≤ t. (a) Prove: if we change the rule to |A ∩ B| = 0 for red edges, the largest blue cliques
will have size Θ(t2 ). (b) Prove the same conclusion assuming red edges are defined by the rule
|A ∩ B| = 0. (So Nagy made the only possible choice.)
♦Ex. 4.2.16 (Bipartite Ramsey Theorem). Given t, there exists a smallest integer n = BR2 (t)

90 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 4.3. RESTRICTED INTERSECTIONS
such that if we 2-color the edges of Kn,n , then there is a monochromatic Kt,t as a subgraph. Show
that BR2 (t) < t4t .
Hint. Show the stronger statement that if n = t4t /2, then any 2-coloring of K2t−1,n must have a
monochromatic Kt,t as a subgraph.

♦Ex. 4.2.17 (Lower bound for BP2 (t)). Show that BP2 (t) > 2t/2 .
Hint. Follow the counting method described in Ex. 4.2.9.

4.3 Restricted intersections

In previous sections we have seen a number of examples of extremal problems for families of
sets satisfying certain intersection conditions. There is a simply stated common generaliza-
tion to most of these problems.
Definition 4.7. Let L be a set of nonnegative integers. The family F is L-intersecting, if
|E ∩ F | ∈ L for every pair E, F of distinct members of F.

Problem 4.8(Restricted Intersection Problem — uniform case). Let L be a set of

nonnegative integers and k ≥ 1. What is the maximum number of members in a k-uniform
L-intersecting family of subsets of a set of n elements?
In the nonuniform version of this problem we omit the parameter k.
Problem 4.9(Restricted Intersection Problem — nonuniform case). Let L be a
set of nonnegative integers. What is the maximum number of members in an L-intersecting
family of subsets of a set of n elements?
No general answer to these problems has been found or is being conjectured but a number
of appealing partial results are known.
It is of particular interest to determine the rate of growth of the extrema for fixed k and
L while n → ∞.
The classical result of the subject was obtained in 1969 by Bose’s former student D. K.
Ray-Chaudhuri, and R. M. Wilson, then a recent graduate from Ohio State University,
the first one in a row of at least 17 advisees of Ray-Chaudhuri. Their surprisingly general
inequality gave a great impetus to further study of the restricted intersection problem. (The
impact could have been accelerated had Wilson been a little quicker in writing up the proof.
It finally appeared in Osaka J. Math. in 1975.) Even more significantly, the method of Ray-
Chaudhuri and Wilson opened up new horizons for the applications of the linear algebra
bound method. We can state only part of the result here; the full version will follow in
Chapter 7. Another proof of the part stated below appears in Chapter 5.
Theorem 4.10 (Ray-Chaudhuri – Wilson Theorem). Let L be a set of s integers and

——————————————————————— 91

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 4. RESTRICTED INTERSECTIONS
F an L-intersecting k-uniform family of subsets of a set of n elements, where s ≤ k. Then
 
n
|F| ≤ . (7)
s

This is clearly best possible as long as the answer is to be a function of the parameters
n and s only. Indeed, the set of s-subsets of an n-set forms a uniform {0, 1, . . . , s − 1}-
n


intersecting family of cardinality s .
Of course, stronger results are to be expected if more information on the set L is taken
into account. For instance, if k is odd and all numbers in L are even, then |F| ≤ n by the
Oddtown Theorem (Corollary 1.2).
But even when we focus on the set L := [s − 1] = {0, 1, . . . , s − 1}, there is something
unsettling about the extremal system just discussed because its members are so small that
the L-intersection condition puts no constraint on them. One might wonder if even for this
particular choice of L, larger values of k might force considerably stronger upper bounds.
Let us fix s and k arbitrarily (s ≤ k) and let n → ∞. Then the rate of growth of
the Ray-Chaudhuri–Wilson bound is Θ(ns ) (proportional to ns ). Our next objective is to
show that this rate of growth can actually be achieved for any given s and k (s ≤ k) while
L = [s − 1].
Theorem 4.11. For every k ≥ s ≥ 1 and n ≥ 2k 2 there exists a k-uniform family F of size
> (n/2k)s on n points such that |E ∩ F | ≤ s − 1 for any two distinct sets E, F ∈ F.
It is worth comparing this lower bound with the Ray-Chaudhuri–Wilson upper bound.
The ratio is  n s .n  s s
> , (8)
2k s 2ek
a constant for fixed s and k. (We used the binomial coefficient estimate of Exercise 4.2.1.)

Proof. Let p be the greatest prime ≤ n/k; this way n/(2k) < p ≤ n/k. Fix a k-subset A of
Fp . (k ≤ p because n ≥ 2k 2 .) Let X be an n-set containing
 A
× Fp .
For a function f : A → Fp , the graph G(f ) = ξ, f (ξ) : ξ ∈ A is a k-subset of X.
Our set system will consist of the graphs of the polynomials of degree ≤ s − 1 over Fp ,
restricted to A. It is easy to see that for two different polynomials of degree ≤ s − 1, their
graphs will have at most s − 1 points in common. The number of polynomials in question is
ps > (n/2k)s , thus confirming the Theorem.

It is an intriguing problem to determine what conditions on L and k force a linear (i. e.,
O(n)) upper bound on the size of F. A number of sufficient conditions were given among
the exercises following the Oddtown section (Section 1.1).
Although no complete answer to this question is known, it turns out that the situation
is remarkably simple if 0 ∈ L and k is large enough compared to the numbers in L. The
threshold between nonlinear and linear rates of growth is then determined by the sole crite-
rion whether or not the g.c.d. of the numbers in L divides k. Moreover, if it does, the rate
of growth of the extremum is at least quadratic.

92 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 4.3. RESTRICTED INTERSECTIONS
In order to see this, let us first consider the case when 0 ∈ L and k is a linear combination
of the numbers in L with nonnegative integer coefficients:

L = {l1 , . . . , ls }; l1 = 0; (9)
Xs
k = ai li (ai ∈ Z, ai ≥ 0). (10)
i=2

We claim that under these conditions, the rate of growth of the extremal L-intersecting
k-uniform families is at least quadratic.
Indeed, in order to exhibit such families of size Ω(n2 ), we just have to slightly modify the
proof of Theorem 4.11, performed for the case s = 2.
Claim 4.12. Let k ≥ 2 and n ≥ 2k 2 . If conditions (9) and (10) hold then there exists a
k-uniform L-intersecting family of size > (n/2k)2 on n points.
Proof. It follows from (9) and (10) that we can write k as the sum of exactly k not necessarily
distinct terms from L:
Xk
k= lij . (11)
j=1

Now apply the construction given in the proof of Theorem 4.11 to the case s = 2. Recall
that in this case the members of the family F obtained are graphs of linear functions A → Fp
where A = {α1 , . . . , αk } ⊆ Fp . Now replace each point (αj , β) ∈ X (β ∈ Fp ) by a set of
size lij (disjoint sets for different points). Equation (11) guarantees that this change will not
affect the sizes of either X, or of any member of F. On the other hand, intersection sizes
are under control in the new family: if E, F ∈ F and E and F are disjoint, then so are the
corresponding new sets; alternatively, if they had precisely one common element, say (αj , β),
then the corresponding new sets intersect in precisely lij elements.

Now we just have to summarize our results to obtain the threshold theorem indicated
above.
Theorem 4.13 (Linear Threshold Theorem). (a) Suppose that the greatest common
divisor of the numbers in L does not divide k. Then an L-intersecting k-uniform family of
subsets of an n-set has at most n members.
2
(b) Assume 0 ∈ L, |L| = s, and k ≥ |L| lmax . Suppose that the greatest common divisor
of the numbers in L divides k. Then for any n ≥ 2k 2 , there exist a k-uniform L-intersecting
family of size ≥ (n/2k)2 on n points.
Here, lmax denotes the largest number in L.
For the proof we need two observations.
Proposition 4.14. Let A be an m × m matrix with integer entries. If some prime power
q = pα divides each off-diagonal entry but it does not divide any of the diagonal entries then
A is nonsingular.

——————————————————————— 93

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 4. RESTRICTED INTERSECTIONS
Proof. The product of the diagonal elements of A is divisible by a lower power of p than any
one of the remaining m! − 1 expansion terms of det A.
Proposition 4.15. Let L = {l1 , . . . , ls } be a set of integers, l1 < l2 < · · · < ls . Assume that
the g.c.d. of the li divides the integer k, and k ≥ sls2 . Then k can be represented as a linear
combination of the li with nonnegative integer coefficients.
Proof.
Ps Since the g.c.d. of the li divides k, k is an integral linear combination of the li :
k = i=1 ai li , ai ∈ Z. Choose such a representation with the sum of the negative ai as small
in absolute value as possible. We claim that none of the ai is negative.
For a contradiction, assume aj < 0. Then
X
ai li > k ≥ sls2 .
i6=j

This implies that ar lr > ls2 for some r 6= j, hence (ar − lj )lr > ls2 − lj lr >
P0. Now, setting
s
bj = aj + lr , br = ar − lj > 0, and bi = ai for i 6= j, r, we still have k = i=1 bi li . On the
other hand the only coefficient that decreased, ar , remained positive, and one of the negative
coefficients, aj , increased. This contradicts the choice of the ai .

Proof of Theorem 4.13. (a) There exists a prime power q = pα dividing each li but not
dividing k. The result is therefore a particular case of the “Mod-p-town Theorem,” stated
as Exercise 1.1.23. Since no proof was furnished with that exercise (albeit hints leading to
two different solutions were), let us work out at least one solution here. An application
of Proposition 4.14 to the intersection matrix A of the family proves that under the given
conditions, the incidence vectors of the members of the family are linearly independent (over
R), completing the proof of part (a).
For part (b), we just have to observe that under the conditions stated, Proposition 4.15
implies that the conditions of Claim 4.12 are satisfied. The conclusion of the Claim is the
conclusion of the Theorem.

Exercises
[n]

Ex. 4.3.1. Let F1 , F2 ⊆ k be k-uniform families and let F2σ denote the image of F2 under the
permutation
σ of the universe [n]. Prove that the average, over all σ, of the quantity |F1 ∩ F2σ |, is
|F1 | · |F2 |/ nk .
Hint. Let A ∈ F2 . The probability that Aσ ∈ F1 for random σ is |F1 |/ nk . (Verify.) This is the

contribution of A to the average size of the intersection of F1 and F2 . Add up these contributions
for all A ∈ F2 .

♦Ex. 4.3.2 (M. Szegedy, 1990). Let F1 , F2 ⊆ [n]

k be k-uniform families. Let L1 and L2 be
disjoint sets of integers and assume that Fi is Li -intersecting (i = 1, 2). Prove:
 
n
|F1 | · |F2 | ≤ . (12)
k

94 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 4.4. EXTREMAL SET THEORY
Ex. 4.3.3. Prove that the condition k ≥ sls2 in Proposition 4.15 can be relaxed to k ≥ (s −
1)(ls ls−1 − 1).

Ex. 4.3.4. Deduce from the preceding exercise that in Theorem 4.13 (b), the condition k ≥ sls2
can be relaxed to k ≥ (s − 1)(ls ls−1 − 1).

Ex. 4.3.5. What would be the “skew version” (in the spirit of some exercises after Section 1.1)
of Proposition 4.14? Prove it when q is a prime; and disprove it for every proper prime power.

Ex. 4.3.6. State and prove the bipartite version of Theorem 4.13 (a).

Ex. 4.3.7. Prove or disprove the skew version of Theorem 4.13 (a).
Note: The authors don’t know the answer to this problem.

Ex. 4.3.8. Prove: if k = 36, L = {2, 8, 11, 17, 20, 23, 32} then m ≤ n + 1.
Hint. Add one point.

Ex. 4.3.9. Prove: if k = 107, L = {2, 27, 52, 77, 102} then m ≤ n + 1.
Hint. Prove that the mod 5 rank of the intersection matrix is at least m − 1.

Ex. 4.3.10. Prove: for p a fixed prime and L = {0, 1, p, p2 }, k = p3 there exist L-intersecting
k-uniform families with m = Ω(n3 ) members.
Hint. Consider the 3-dimensional affine subspaces in the affine spaces over Fp .

♦Ex. 4.3.11. Prove the same for L = {0, 1, p + 1, p2 + p + 1} and k = p3 + p2 + p + 1.

Ex. 4.3.12. Prove, using the Ray-Chaudhuri–Wilson Theorem: if k = 90, L = {0, 60, 66} then
n


m≤ 2 .
Hint. Prove that among the members of the family, the relation of having nonempty intersection
is an equivalence relation. Handle each equivalence class separately.

Ex. 4.3.13. Assume we have the parameters of the preceding exercise. Prove that for every
sufficiently large n, there exist k-uniform L-intersecting families of size > 10−5 n2 on n points.
Hint. Ignore the admissibility of empty intersection. Designate a set of 60 elements to be a subset
of each member of the family to be constructed. Apply Theorem 4.13 (b) to the rest.

4.4 Extremal set theory: the classics

The subject of extremal set theory is to find the maximum cardinality of a set system
satisfying certain assumptions. In this brief section, we state three classical results of the
subject. Proofs can be found, e.g., in Bollobás (1986) or Lovász (1979c).
A set system F is called a Sperner family if no member of the family is a subset of
another. Another term for Sperner families is antichain (cf. Section
8.3). For instance,
n
the set of all k-subsets of [n] is a Sperner-family of cardinality k . The largest of these

——————————————————————— 95

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 4. RESTRICTED INTERSECTIONS
is obtained when k = bn/2c or k = dn/2e. Sperner families do not need to be uniform.
Nevertheless, E. Sperner proved in 1928 that no Sperner family can beat the uniform ones.
Theorem 4.16 (Sperner’s Theorem). If F is a Sperner family of subsets of a set of n
elements, then  
n
|F| ≤ . (13)
bn/2c
Only uniform families can attain the upper bound.
A chain is a family of sets of the form A1 ⊂ A2 ⊂ · · · ⊂ Am . One of the several
known
n
proofs of Sperner’s Theorem employs the idea that 2[n] can be decomposed into bn/2c chains.
Clearly, a Sperner family contains at most one member from each chain, hence the result. A
generalization of this proof appears in Section 8.3.
Another approach proves the following stronger result, called LYM-inequality after its
authors D. Lubell (1966), K. Yamamoto (1954), and L.D. Meshalkin (1963). The result is
also a special case of Bollobás (1965), to be stated as Theorem 5.5 in Section 5.1.
Theorem 4.17 (LYM inequality). If F is a Sperner family of subsets of a set of n
elements, then
X 1
n

≤ 1. (14)
A∈F |A|

Sperner’s Theorem is an immediate consequence. (Why?)

Proof. (Lubell’s Permutation Method, 1966) Let [n] be the universe of F. With every set
A ⊆ [n] let us associate the set P (A) of those permutations (i1 , . . . , in ) of the set [n] in
which the elements of A form the initial segment (i1 , . . . , i|A| ) (in any order). The number
of such permutations is |A|!(n − |A|)!. Observe now that the Sperner condition is equivalent
to saying that the sets P (E), E ∈ F, are pairwise disjoint. It follows that
X
|A|!(n − |A|)! ≤ n!.
E∈F

Dividing by n! we obtain the LYM inequality.

The second classical result we should mention concerns uniform families with no disjoint
pairs of members. One easy way of obtaining such a family is to fix a point and take all the
sets containing it. The fact that we cannot do better for k ≤ n/2 is what Erdős, Ko, and
Rado proved in 1961.
Theorem 4.18 (Erdős–Ko–Rado Theorem). If F is a k-uniform family of subsets of
a set of n elements (k ≤ n/2) and every pair of members of F intersect, then
 
n−1
|F| ≤ . (15)
k−1
There are many extremal systems for k = n/2 (take one member of every complementary
pair of (n/2)-subsets), but for k < n/2, the upper bound is attained only for families which
share a common point.

96 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 4.4. EXTREMAL SET THEORY
The elegant proof we present next is due to G. O. H. Katona (1972).
First we make a simple observation.
Proposition 4.19. Let C be a cycle of length n and H be a family of paths of length
k ≤ n/2 in C. Assume every pair of members of H has an edge in common. Then |H| ≤ k.
We leave the easy proof to the reader. We turn to the proof of the Erdős–Ko–Rado
Theorem.
Proof. (Katona’s Cyclic Permutation Method, 1972.) Think of the n elements being guests
at a dinner party. At the party, everybody is seated around a big round table with n seats.
This means n! possible seating arrangements.
The guests like to form clubs of k members each; and they have already exhausted all
n


possibilities. So there are k clubs. Think of the members of F as being red clubs; all the
other k-sets are blue clubs. Each club requests that all of its members be seated contiguously
so they could pass notes to each other without the danger of being intercepted by a non-
member.
Of course, very few of these requests can be honored at a time: in each particular seating
arrangement there are exactly n contiguous intervals of length k. How many of these are
red? No more than k, according to the Proposition. (We used the condition that red clubs
intersect now.)
The gallant host therefore invites everybody to a succession of n! parties, and tries out
a different seating each time. It is clear by symmetry that the request of each club will be
honored at exactly the same number of parties. On the other hand, at each party, at most
a k/n fraction of the requests honored come from red clubs. Therefore the number of red
clubs is at most a k/n fraction of the total number of clubs:
   
k n n−1
|F| ≤ = .
n k k−1
The third result we state in this section asserts that in a sufficiently large k-uniform
family, some highly regular configurations, called sunflowers, must occur, regardless of the
size of the universe.
A family F = {A1 , . . . , Am } is a sunflower with m petals if
m
\
Ai ∩ Aj = At (16)
t=1

for every i 6= j, (1 ≤ i, j ≤ m). The common intersection of the members of a sunflower

form its kernel. Note that a family of disjoint sets is a sunflower (with empty kernel).
Theorem 4.20 (Sunflower Theorem, Erdős–Rado, 1960). If F is a k-uniform set
system with more than k!(s − 1)k members, then F contains a sunflower with s petals.
Proof. We proceed by induction on k. For k = 1, we have more than (s − 1) points (disjoint
1-sets), so any s of them form a sunflower with s petals.
Now let k ≥ 2. Let T = {A1 , . . . , Ar } be a maximal family of pairwise disjoint members
of F. If r ≥ s, these sets form a sunflower with r ≥ s petals, and we are done.

——————————————————————— 97

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 4. RESTRICTED INTERSECTIONS
Assume r < s, and let B = ri=1 Ai . Then |B| ≤ k(s − 1). By the maximality of the
S
family T , every member of F intersects B. Therefore there exists a point x ∈ B, contained
in at least
|F| k!(s − 1)k
> = (k − 1)!(s − 1)k−1
|B| k(s − 1)
members of F. Let us delete x from these sets and consider the (k − 1)-uniform family

F(x) := {E \ {x} : E ∈ F, x ∈ E}. (17)

By the induction hypothesis, this family contains a sunflower G with s petals. Adding x to
each member of G we obtain a subfamily of F which forms a sunflower with s petals.
It is a major open problem (Erdős offers considerable monetary reward to the first solver)
whether or not there exists a positive integer C such that every k-uniform family with C k
members necessarily contains a sunflower with three petals.
The use of sunflowers has proved to be one of the most powerful methods in extremal set
theory. At times it competes with linear algebra methods, in other cases the two approaches
complement one another.
A particularly useful tool of the sunflower technique is Deza’s Theorem, stated in Exercise
4.2.13.

98 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
Chapter 5

Spaces of polynomials

5.1 Helly-type theorems for finite sets

Helly’s Theorem (Section 3.2.) asserts that, if a finite family of convex subsets of Rn has the
property that each set of ≤ n + 1 members of the family intersect then all of them intersect.

It is natural to ask if objects other than convex sets obey Helly-type laws. For example,
it is straightforward to prove that, if each set of at most three edges of a graph intersects,
then all edges intersect. Even the following generalization is not difficult to verify.
If each family of ≤ r + 1 members of an r-uniform set system intersect, then all members
intersect. (See Exercise 5.1.2.)
Erdős, Hajnal, and Moon generalized the easy observation about graphs in a different direc-
tion. A set S ⊆ V of vertices is said to cover a set F ⊆ E of edges of the graph G = (V, E)
if every edge in F has at least one of its endpoints in S.
Theorem 5.1 (Erdős–Hajnal–Moon, 1964). If each family of at most s+2


2
edges of a
graph can be covered by s vertices, then all edges can.
The complete graph on s + 2 vertices shows that this bound is best possible. The question
was, how to generalize the result to r-uniform families. The conjecture was easy enough
r+s


to formulate: all the numbers given so far suggest the formula r (which is the same as
r+s


s
). This indeed is the correct answer, as shown in a 1965 paper by Béla Bollobás, then
an Eötvös University undergraduate, whom Erdős, always eager to see promising “epsilons,”
had introduced to graph theory at 14.
If each family of at most r+s


Theorem 5.2 (B. Bollobás, 1965). r
members of an
r-uniform set system can be covered by s points then all members can.
This result is best possible, as shown by the family of all r-tuples of a set of r + s points.
Before proceeding to the proof, we should put the result in a different context. In the
early 60’s, Tibor Gallai (soon afterwards to become László Lovász’s most influential teacher)
initiated an important new approach to the study of such parameters of a graph as matching

c László Babai and Péter Frankl. 1988, 1992, 2020.

99
CHAPTER 5. SPACES OF POLYNOMIALS
number, chromatic number, covering number, independence number. He introduced the
notion of critical graphs: graphs for which the parameter in question has a given value but
if we remove any edge, the value of the parameter changes. Clearly, every graph contains
a subgraph critical with respect to the given value of the parameter (just delete edges one
by one as long as possible without changing the value). The structure of critical graphs is
therefore of particular importance.
Chromatic critical graphs will be the subject of Section 8.2. In the present section we
consider τ -critical graphs and set systems. Recall that the covering number τ (G) of a graph
G is the minimum number of vertices incident with all edges. In other words, this is the
size of the smallest subset S of the vertex set V such that the set V \ S is independent. A
graph is τ -critical if the removal of any edge reduces this quantity, i. e., it increases α(G),
the maximum size of independent sets.
The theory of τ -critical graphs has given birth to a beautiful classification theorem
(Lovász 1978b); Lovász makes use of a geometric generalization of Bollobás’s theorem in
the proof. We shall give details of this result in Chapter *** (The reader should note the
dynamics of how a theorem on set systems arose as a generalization of a graph-theoretic
result (Erdős –Hajnal–Moon, Theorem 5.1) and subsequently was applied back to graphs to
yield the definitive result of the subject.)
A set system F is thus τ -critical if the removal of any of its members decreases the value
of τ (F), the covering number.
Another way of stating Bollobás’s Theorem is this (cf. Exercise 5.1.7).
Theorem 5.3.
Let F be an r-uniform τ -critical set system with τ (F) = s + 1. Then
m := |F| ≤ r+s r
.
Let F = {A1 , . . . , Am }. For each i, the removal of Ai results in a family possessing
an s-element cover. Let Bi denote such a covering set. Bi does not intersect Ai because
τ (F) > s, but it does intersect all the other Ai . Thus, Theorem 5.3 will follow from this,
slightly stronger result.
Theorem 5.4 (Bollobás’s Theorem — uniform version). Let A1 , . . . , Am be r-element
sets and B1 , . . . , Bm be s-element sets such that

(a) Ai and Bi are disjoint for i = 1, . . . , m;

(b) Ai and Bj intersect whenever i 6= j (1 ≤ i, j≤ m).

Then  
r+s
m≤ .
r
The importance of this result is reflected, inter alia, by the list of proofs published (Bol-
lobás (1965), Jaeger–Payan (1971), Katona (1974), Lovász (1977)). It was Lovász, who,
in an entirely novel way, made the linear algebra bound bear on the subject. Using tensor
products, he gave two different proofs of Bollobás’s theorem along with several generaliza-
tions. We shall elaborate on Lovász’s method in Chapter 6. Below we give a variant of

100 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.1. HELLY-TYPE THEOREMS
one of Lovász’s proofs which can be stated without referring to the machinery of multilinear
algebra.

Proof. Let V be the union of all the sets Ai and Bi . Let us associate vectors p(v) =
(p0 (v), p1 (v), . . . , pr (v)) ∈ Rr+1 with each v ∈ V such that the set of vectors obtained is in
general position, i. e., any r + 1 of them are linearly independent. (We may, for instance,
select them from the moment curve, cf. Section 3.1, Proposition 3.3.) With every set W ⊂ V
we associate a polynomial fW (x) in the r + 1 variables x = (x0 , x1 , . . . , xr ) in the following
way. Let Y
fW (x) = (p0 (v)x0 + p1 (v)x1 + · · · + pr (v)xr ). (1)
v∈W

This is a homogeneous polynomial of degree |W | . Clearly,


6= 0, if x is orthogonal to none

fW (x) of the p(v), v ∈ W ; (2)

= 0, otherwise.


Let fi (x) stand for fBi (x); so fi is homogeneous of degree s.

The vectors corresponding to the elements of Aj generate a subspace of dimension r;
let aj be a nonzero vector, orthogonal to this subspace. Because the vectors chosen are in
general position, aj is orthogonal to p(v) precisely if v ∈ Aj .
By (1) we thus conclude that fi (aj ) = 0 precisely if Aj and Bi intersect, i. e., if i 6= j :
(
6= 0, if i = j ;
fi (aj ) (3)
= 0, if i 6= j .

From this it follows by the Diagonal Criterion (Proposition 2.4) that the polynomials fi
are linearly independent. Therefore their number, m, is not greater than the dimension of

the space of homogeneous polynomials of degree s in r + 1 variables, i. e., m ≤ (r+1)+s−1
s
according to Exercise 2.1.2.

This proof does not yield the even stronger result that Bollobás obtained for not neces-
sarily uniform families.
Theorem 5.5 (Bollobás’s Theorem — nonuniform version) (Bollobás, 1965)).
Let A1 , . . . , Am and B1 , . . . , Bm be finite sets satisfying conditions (a) and (b) of Theorem 5.4.
Then m
X 1
|Ai |+|Bi |

≤ 1.
i=1 |Ai |

(For two proofs, see Exercise 5.1.9.) Observe that in the case |Ai | = r, |Bi | = s, this
result turns into the uniform version.
To see that the nonuniform version does have added substance, we note, that Sperner’s
Theorem (Section 4.4) is an immediate corollary to Theorem 5.5 (see Exercise 5.1.10).

——————————————————————— 101

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 5. SPACES OF POLYNOMIALS
Nevertheless, there are several advantages to the linear algebra proof. It allows general-
izations and extensions that do not, at the moment, admit any direct combinatorial proof.
We shall spend most of Chapter 6 on such results. There is one important consequence that
we can present here without delay: the skew version of the theorem.
Theorem 5.6 (Bollobás’s Theorem — skew version). The conclusion of Theorem 5.4
remains valid if we weaken assumption (b) to

(b0 ) Ai and Bj intersect whenever i < j, (1 ≤ i, j ≤ m).

The proof is identical with the proof of Theorem 5.4 we just saw, except that instead of
(3) we now have (
6= 0, if i = j ;
fi (aj ) (4)
= 0, if i < j .
This suffices to guarantee linear independence of the fi by the Triangular Criterion (Propo-
sition 2.5).

The conclusion of the nonuniform version (Theorem 5.5) does not remain valid under these
circumstances (Exercise 5.1.1.)

Theorem 5.6 isn’t Bollobás’s. It was conjectured by him and by others, but until Lovász’s
paper (1977b) no appropriate technique was known. We shall discuss another linear algebra
proof, appearing in the same paper of Lovász, in Section 6.2.
An application of the skew version of Bollobás’s Theorem to automata theory was found
by J. E. Pin (1981).
In another application, the skew version of Bollobás’s Theorem is a key ingredient in a
relatively simple proof of an extension of the Upper Bound Theorem (Theorem 3.27) given
by Alon and Kalai (1985).
Section 3.2.3.
A proof similar to the one above yields the following Bollobás–type theorem. As it takes
no extra effort to prove, we immediately state the result in its skew version.
Theorem 5.7 (Bollobás’s Theorem for sets vs. subspaces) (Lovász 1977b). Let
U1 , . . . , Um be r-dimensional subspaces and B1 , . . . , Bm subsets of cardinality s in a linear
space W over the field F. Assume that

(a) Bi and Ui are disjoint for i = 1, . . . , s;

(b) Bi and Uj intersect whenever i < j (1 ≤ i, j ≤ m).

Then  
r+s
m≤ .
r
Proof. We may assume F is infinite. (Actually, there is some subtlety to this point; see
Exercise 5.1.12. We need this assumption because of the “general position” argument below.

102 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.1. HELLY-TYPE THEOREMS
For the combinatorial applications, one can usually take F = R and not worry about finite
fields.)
Let n = dim W ; clearly, n ≥ r + 1. After performing a linear map ϕ : W → Fr+1 in
general position, we may assume that n = r + 1, thus the Ui become hyperplanes. (Cf.
Exercise 5.1.13.) Select a nonzero vector ai , orthogonal to Ui for each i. Define, as in the
proof of Theorem 5.4, fi (x) := fBi (x) where the right hand side is defined by equation (1).
Now the alternative (3) holds again. As in the proof of Theorem 5.6, a reference to the
Triangular Criterion (Proposition 2.5) concludes the proof.

Exercises
Ex. 5.1.1 (Counterexample to nonuniform skew version of Bollobás’s Theorem). Given N > 1,
construct two set systems A1 , . . . , Am and B1 , . . . , Bm such that

(a) Ai and Bi are disjoint for every i (1 ≤ i ≤ m);

(b0 ) Ai and Bj intersect whenever i < j (1 ≤ i, j ≤ m),

but
m
X 1
|Ai |+|Bi |

≥ N.
i=1 |Ai |

Hint. Let the Ai be all subsets of [N], arranged in decreasing order of their size. Let Bi = [N ] \ Ai .
The sum on the left hand side will be N + 1.

Ex. 5.1.2. Give a simple direct proof of the analogue of Helly’s Theorem for r-uniform set systems
stated at the beginning of this chapter.

Ex. 5.1.3. Prove that any collection of subtrees of a tree satisfies the one-dimensional Helly
Theorem: if each pair of subtrees in the collection have a vertex in common then all of them do.

Ex. 5.1.4. State and prove a generalization of the one-dimensional Helly Theorem for s-point
covers.
Hint. If each set of s + 1 members of a finite collection of intervals on a line can be covered by s
points then all of them can.

Ex. 5.1.5. Prove the same for subtrees of a tree.

Hint. Proceed by induction on the number of vertices. Delete an endpoint for induction.
♦Ex. 5.1.6. Prove that for convex sets in dimensions ≥ 2, no generalization of Helly’s Theorem
to s-point covers exists.
Hint. Construct an arbitrarily large number of convex sets in the plane such that they cannot be
covered by 2 points but removing any one of the sets, the remaining sets can.

Ex. 5.1.7. Prove that Theorems 5.2 and 5.3 are equivalent.
♦Ex. 5.1.8. Prove that the bound in Theorem 5.4 is tight.

——————————————————————— 103

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 5. SPACES OF POLYNOMIALS
♦Ex. 5.1.9.* (a) Prove Bollobás’s original (nonuniform) result (Theorem 5.5). (b) Prove that if
equality holds in Theorem 5.5 then both families {Ai } and {Bi } are uniform and they form the
system described in the previous exercise.

Ex. 5.1.10. Deduce Sperner’s Theorem (Section 4.4) from the Nonuniform Bollobás Theorem.
Ex. 5.1.11. Deduce Theorem 5.4 from Theorem 5.7.
Ex. 5.1.12. Justify the assumption in the proof of Theorem 5.7 that the field F is infinite.
Hint. Use the Field Extension Lemma (Lemma 2.29).
♦Ex. 5.1.13. What does a linear map in general position mean in the proof of Theorem 5.7?

5.2 Resultants
TO BE WRITTEN Blokhuis’s proof of the uniform skew Bollobas Theorem via resultants

5.3 The Prague dimension of graphs

Various notions of dimension play a role in much of mathematics (algebra, topology, and all
the fields they are applied to). This is the case in combinatorics, too, wherever algebraic
or topological structures are associated with combinatorial objects. (This book provides an
ample supply of examples from linear algebra; for topology, ring theory, homology let us
refer, e.g., to Björner (1979,1981), Kalai (1984), Lovász (1979a), Stanley (1980, 1983).
Interesting notions of dimension of graphs arise naturally in purely graph theoretic con-
text as well.
One possible approach to defining the dimension of a class of objects is to single out
a subclass as the simplest, “one-dimensional” objects; define multiplication of the objects;
define embeddings; and finally define the dimension of an object as the smallest k such that
the object can be embedded in the product of k one-dimensional objects.
For graphs, such a program was carried out in Prague in the late 70s, much in the spirit
of the work of Dushnik and Miller (1941) on a notion of dimension for partially ordered sets.
The first two papers which introduced and investigated the new concept were by J. Nešetřil
and V. Rödl (1978) and J. Nešetřil and A. Pultr (1977).
Their idea was to choose the complete graphs to be one-dimensional and “embeddability”
to refer to induced subgraphs. They had to define how graphs should be multiplied. Out of
a multitude of possible choices (cf. Harary, 1969), the following seemed most appropriate.
Definition 5.8. The product of the graphs (V1 , E1 ) and (V2 , E2 ) is a graph with vertex set
V1 × V2 ; two vertices (v1 , v2 ) and (w1 , w2 ) are adjacent in the product graph if each pair
{vi , wi } (i = 1, 2) is adjacent in the corresponding graph. In particular, vi and wi must be
distinct.

104 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.3. PRAGUE DIMENSION
We denote the product of the graphs G1 and G2 simply by G1 G2 . This operation is also
called the categorical product of graphs, because, with a natural notion of homomorphisms
of graphs, the class of graphs becomes a category in which the operation just defined is the
product. (See Exercise 5.3.12.)
Definition 5.9. The Prague dimension of the graph G is the minimum number d such that
G is an induced subgraph of the product of d complete graphs.

It is clear that this dimension function is monotone: if G1 is an induced subgraph of G2

then dim (G1 ) ≤ dim (G2 ). It is easy to see that every finite graph has a (finite) dimension;
Exercises 5.3.5–5.3.8 provide upper bounds. What we need in addition is a technique to
prove lower bounds. Observing that the 1-dimensional graphs are precisely the complete
graphs, it would be particularly desirable if we could show that the product of d complete
graphs has dimension equal to d. (The definition only guarantees that the dimension of
such a product is ≤ d.) The problem is analogous to the problem in topological dimension
theory: everything that is topologically embeddable in Rn has dimension ≤ n; but we have to
prove (and this is not an easy exercise) that Rn does not embed in Rn−1 . (See, for example,
Pontriagin (1952), or Spanier (1977).) The answer, luckily, is affirmative (both in topology
and for graphs). We present the charming proof for graphs.
Theorem 5.10 (Dimension invariance for graphs) (Lovász–Nešetřil–Pultr, 1980).
The dimension of the product of d (nontrivial) complete graphs is d.
Nontrivial means having more than one vertex. By definition, such a product graph has
dimension ≤ d so we have to prove it is ≥ d at the same time. By monotonicity, it suffices
to prove this for products of two-point graphs.
Lemma 5.11. The product of d copies of K2 is the disjoint union of 2d−1 copies of K2 .
We leave the proof to the reader.
We have to prove that the graph described in the Lemma has dimension ≥ d. We shall
prove a stronger result.
Theorem 5.12. Assume that the graph G = (V, E) has two (not necessarily disjoint) sets
of m vertices each, S = {s1 , . . . , sm } and T = {t1 , . . . , tm }, such that

(a) si and ti are adjacent for every i (i = 1, . . . , m);

(b) si and tj are not adjacent for i < j (1 ≤ i, j ≤ m).

Then the dimension of G is ≥ log2 m.

This quite general lower bound is very helpful in estimating the dimension for some
classes of graphs. (See Exercise 5.3.11.) The Dimension Invariance Theorem will be a direct
consequence.

Proof. Let dim(G) = d. By definition this means that we can associate vectors p(v) =
(p1 (v), . . . , pd (v)) ∈ Rd with each vertex v ∈ V such that two vertices v, w ∈ V are adjacent

——————————————————————— 105

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 5. SPACES OF POLYNOMIALS
precisely if pi (v) 6= pi (w) for i = 1, . . . , d. It is therefore natural to associate the polynomial
in d variables
d
Y
fv (x1 , . . . , xd ) = (xi − pi (v))
i=1

with each v ∈ V . The vertices v and w will be adjacent precisely if fv (p(w)) 6= 0.

We can thus summarize conditions (a) and (b) in the statement that
(
6= 0, if i = j ;
fsi (p(tj ))
= 0, if i < j .

By the Triangular Criterion (Proposition 2.5), this implies that the polynomials associated
with the vertices in S are linearly independent.
Each polynomial fv (x1 , . . . , xd ) is multilinear, i. e., linear in all the d variables, and
is therefore a linear combination of the 2d monic multilinear monomials (cf. p. xiii and
Ex. 5.4.3). We conclude that m, the number of linearly independent polynomials we found
in this space, must be ≤ 2d .

In order to derive Theorem 5.10 from Theorem 5.12 (through Lemma 5.11), let G be the
disjoint union of 2d−1 copies of K2 . Let S = T be the vertex set of G, numbered in such a
way that si and ti be adjacent for i = 1, . . . , m = 2d . Then condition (b) in Theorem 5.12
will be automatically fulfilled. Therefore dim(G) ≥ log2 m = d.
Remark. Another combinatorial notion of dimension, related to isometric embeddings in the
squashed cube and introduced by Graham and Pollak (1972), was discussed in Section 1.4.
There is another, more geometrical, philosophy of graph dimension that has lately been
gaining increasing popularity. The basic idea is that one defines, in some uniform geometrical
fashion, a graph on the Euclidean space Rn for every n, and calls the smallest n for which the
graph G is a subgraph of the graph on Rn the dimension of the graph (if such an n exists).
One example is the unit-distance graph: pair of points at unit distance are adjacent
(Erdős–Harary –Tutte, 1965; Erdős–Simonovits 1980). (We shall consider the chromatic
number of the unit-distance graphs in Section 5.5.) Distance threshold graphs (adjacency
corresponds to distance > λ) on spheres have been considered among others by Alspach and
Rosenfeld (1977), Larman (1978), Maehara (1984), Rödl (1984), Reiterman–Rödl, Šinajova
(1989). A variant of distance threshold is the contact dimension (edge corresponds to distance
= λ, absence of edge to distance > λ, distances < λ are not permitted) (Frankl–Maehara,
1988).

Exercises
Ex. 5.3.1. Prove Lemma 5.11.
Ex. 5.3.2. Let χ(G) denote the chromatic number of the graph G. Prove: for any two graphs
G, H,
χ(GH) ≤ min{χ(G), χ(H)}.

106 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.3. PRAGUE DIMENSION
Ex. 5.3.3. Prove, that, for G = H, equality holds in 5.3.2.
Hint. Prove that G is a subgraph of GG.
Remark. It is conjectured that in 5.3.2, equality holds for all finite graphs G, H. This is trivially
true if one of the graphs has chromatic number 3. (Why?) The conjecture has been confirmed for
chromatic number 4 by El-Zahar and Sauer (1985). For infinite graphs with uncountable chromatic
number, counterexamples have been constructed by A. Hajnal (1985).

The following sequence of exercises (5.3.4–5.3.11) is adopted from Lovász–Nešetřil–Pultr (1980).

Ex. 5.3.4. Prove that the dimension of the empty graph (no edges) on n ≥ 2 vertices is 2.
Ex. 5.3.5. Prove that the dimension of a graph G = (V, E) is the smallest integer d such that
there exists a one-to-one map f : V → Zd such that v, w ∈ V are adjacent precisely if none of the
coordinates of f (v) − f (w) is zero.
♦Ex. 5.3.6. Prove that every finite graph has a dimension.
Hint. Prove that the dimension of G is not greater than 1 + the number of edges in the complement
of G.

Ex. 5.3.7. Let us say that a coloring of the vertex set V of G covers the pair {a, b} (a, b ∈ V ) if a
and b receive the same color. In a good coloring, this cannot happen if a and b are adjacent. Prove
that dim G is the minimum number of good colorings that cover all the nonadjacent pairs (plus 1
if this number is ≤ 1).

Ex. 5.3.8. The edge chromatic number of a graph is the minimum number of colors needed in
order to color the edges of a graph such that edges sharing a vertex receive different colors. Prove:
(a) dim(G) ≤ the edge chromatic number of the complement of G (plus 1 if this number is ≤ 1).
(b) Prove: if G has n ≥ 2 vertices then dim(G) ≤ n. (c) If the complement of G has no triangles
then we have equality in part (a).
Hint to part (b). Vizing’s Theorem states that the edge chromatic number of the graph is never
greater than 1 + the maximum degree of the graph.

Ex. 5.3.9. Prove that there is no upper bound on the dimension of the disjoint union of two
graphs G and H in terms of dim(G) and dim(H).
Hint. Prove: the dimension of the graph Kn + K1 (the complete graph on n vertices + an isolated
vertex) is n.

Ex. 5.3.10. Let G ∗ H denote the graph obtained by taking the disjoint union of G and H and
joining each vertex of G to each vertex of H. Determine dim(G ∗ H), given dim(G) and dim(H).
Hint. Prove: dim(G ∗ H) = max{dim(G), dim(H)}.

Ex. 5.3.11. Determine or estimate the dimensions of paths and cycles.

Hint. Use Theorem 5.12. The results for n ≥ 3 vertices: the dimension of the path of length n − 1
is dlog2 (n − 1)e; the dimension of the cycle of length n = 2k is 1 + dlog2 (k − 1)e; and the dimension
of the cycle of length n = 2k + 1 is between 1 + dlog2 ke and 2 + dlog2 ke.

Ex. 5.3.12. Define graph homomorphisms such that the operation introduced in Definition 5.8
becomes the product in the resulting category.

——————————————————————— 107

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 5. SPACES OF POLYNOMIALS
Ex. 5.3.13 (I. Křı́ž, 1984). Let G be a graph of dimension d. Prove: (a) When removing an
edge, the dimension increases by at most 1. (b) When adding an edge, the dimension will be ≤ 2d.
(c)∗ For every d ≥ 6, construct a graph G of dimension d such that by adding a suitable edge, the
dimension will be ≥ d + 2.

5.4 Sets with few intersection sizes mod p

Let F be a family of m subsets of a set of n elements. Let further L be a set of s nonnegative
integers. Recall that F is an L-intersecting family, if |A ∩ B| ∈ L for every pair of distinct
members A, B of F. The first major result on extremal L-intersecting families was the
Ray-Chaudhuri–Wilson Theorem (RW
Theorem, for short), asserting that a uniform L-
intersecting family has no more than ns members:
 
n
m≤ . (5)
s
This theorem assumes that the number of different intersection sizes is bounded by s, pre-
sumably a number, small compared to n. By contrast, the Oddtown Theorem liberally
allowed n/2 different intersection sizes, and still forced a very strong upper bound: m ≤ n.
The key constraint in the Oddtown Theorem is that the intersection sizes belong to one
residue class mod 2, the sizes of the sets to another.
An extension of the RW Theorem in this direction was accomplished by Frankl and
Wilson (1981). Their result (Theorem 7.15) states that the same conclusion (inequality (5))
follows under the considerably weaker condition that the intersection sizes belong to at most
s residue classes mod p, assuming that k, the size of the members of the family, does not
belong to these residue classes.
This result and other related modular extensions of the RW Theorem have turned out
to provide powerful tools for geometric and combinatorial problems.
The original proofs of these results used the machinery of higher incidence matrices, to
be discussed in Chapter 7. Fortunately, most of the results now admit conceptually simpler
proofs which we describe later in this chapter (see Theorem 5.35).
A slightly weaker upper bound of the form
     
n n n
m≤ + + ··· + (6)
s s−1 0
is often easier to prove; yet for most applications, this weaker version is perfectly sufficient
(see Prop. 5.13).
We shall prove the tight upper bounds of the form (5) in due course. However, the order
in which we discuss the RW-type results will follow neither the logical nor the chronological
order. Instead, we start with the result that is the easiest to prove (Theorem 5.15). While
the upper bound we obtain is of the weaker type (eqn. (6)), this small compromise will allow
us to illustrate the wealth of applications immediately in the subsequent sections, before
moving on to the more complex proofs.

108 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.4. FEW INTERSECTION SIZES
To support the claim that we lose little by using the upper bound (6) in place of (5), let
us examine the contribution of the tail of the

sum in (6). We shall observe that when s is
substantially smaller than n/2, the term ns determines the order of magnitude of the sum.
Proposition 5.13. For n ≥ 2s we have
         
n n n n s
+ + ··· + < · 1+ . (7)
s s−1 0 s n − 2s + 1

We leave the easy proof as Ex. 5.4.1.

Note that if s ≤ n/` then the right hand side is less than ns · 1 + 1



`−2
. For example,
for s ≤ n/4 we obtain the inequality
       
n n n n
+ + ··· + <2· . (8)
s s−1 0 s

The easy-to-prove yet widely applicable result to be proved below is a nonuniform modular
variant of the RW Theorem. Some modular terminology will come in handy in stating the
result.
Definition 5.14. For a set L ⊂ Z and integers r, t, we shall say that

t ∈ L (mod r), (9)

if t ≡ ` (mod r) for some ` ∈ L. The negation of this statement will be written as t 6∈ L

(mod r).
A set system F is L-intersecting modr if |A ∩ B| ∈ L (mod r) for any two distinct sets
E, F ∈ F.

Theorem 5.15 (Nonuniform modular RW Theorem) (Deza–Frankl–Singhi, 1983).

Let p be a prime number and L a set of s integers. Assume F = {A1 , . . . , Am } is a family

of subsets of a set of n elements such that

(a) |Ai | ∈
/ L (mod p) (1 ≤ i ≤ m);

(b) |Ai ∩ Aj | ∈ L (mod p) (1 ≤ j < i ≤ m).

Then inequality (6) holds.

The original proof of this result is based on higher incidence matrices and will be given
in Chap. 7 as Ex. 7.4.15.
The simple proof we shall present here, found by Alon, Babai, Suzuki (1991), will be
modeled after the prototype of “polynomial space” proofs, the proof of the two-distance set
bound described in Section 1.2. In fact, it might be instructive for the reader to review that
proof before proceeding.

——————————————————————— 109

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 5. SPACES OF POLYNOMIALS
We begin with a simple observation. Recall that a polynomial is multilinear if it has
degree ≤ 1 in each variable. Every multilinear polynomial of degree ≤ s is a linear
combination of monic multilinear monomials (products of distinct variables) of degree ≤ s.
Proposition 5.16 (Multilinearization). Let F be a field and Ω = {0, 1}n ⊆ Fn . If f
is a polynomial of degree ≤ s in n variables over F then there exists a (unique) multilinear
polynomial fe of degree ≤ s in the same variables such that

f (x) = fe(x) f or every x ∈ Ω. (10)

Indeed, just expand f and use the identity x2i = xi , valid over Ω.

Proof of Theorem 5.15. (Alon–Babai–Suzuki, 1991). In analogy with equation (7) in section
1.2, we introduce a polynomial F (x, y) in 2n variables; this time x, y ∈ Fnp . We set
Y
F (x, y) = (x · y − `) (11)
`∈L

where x · y = ni=1 xi yi is the standard inner product in Fnp . Now consider the n-variable
P
polynomials fi (x) := F (x, vi ), where vi ∈ Fnp is the incidence vector of the set Ai (i =
1, . . . , m). It is clear from the conditions that for 1 ≤ i, j ≤ m,
(
6= 0, if i = j ;
fi (aj ) (12)
= 0, if i 6= j .

By Proposition 5.16, these equations remain valid if we replace fi by the corresponding

multilinear polynomials fei . We conclude by the Diagonal Criterion (Prop. 2.4) that fe1 , . . . , fem
are linearly independent over Fp .
On the other hand, all the fei arePmultilinear polynomials of degree ≤ s and therefore
s n


they belong to a space of dimension k=0 k (Ex. 5.4.2).
As an immediate corollary, we can deduce a slightly weaker form of the Ray-Chaudhuri–
Wilson Theorem. We state it here for the sake of an application in constructive Ramsey
theory to follow in Section 5.7.
Corollary 5.17. Let L be a set of s integers and F an L-intersecting k-uniform family of
subsets of a set of n elements. Then
     
n n n
|F| = m ≤ + + ··· + .
s s−1 0

Proof. Since F is k-uniform, all pairwise intersections have ≤ (k − 1) elements, so we may

assume k 6∈ L. Now choose a prime number p > k and apply Theorem 5.15.
We remark that the same upper bound holds without the uniformity assumption, as we
shall see shortly (Theorem 5.34).

110 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.4. FEW INTERSECTION SIZES
A particularly significant consequence of this result is an upper bound of the form (2−c)n
on the size of certain families required to omit only a single intersection size.
Corollary 5.18 (Omitted Intersection Theorem). Let p be a prime number and F
a (2p − 1)-uniform family of subsets of a set of 4p − 1 elements. If no two members of F
intersect in precisely p − 1 elements, then
 
4p − 1
|F| ≤ 2 · < 1.75484p−1 . (13)
p−1

Proof. Set L = {0, . . . , p − 2}. It is clear that F satisfies the assumptions of Theorem 5.15.
The conclusion follows by inequality (8). The last inequality holds when p is sufficiently
large. The verification is left as Ex. 5.4.4.
We note that the factor of 2 on the right hand side can be omitted, using Theorem 5.37
in place of Theorem 5.15.
We mention that more recently a much stronger form of the Omitted Intersection Theo-
rem was proved, confirming an old conjecture of Pál Erdős. The result makes no assumptions
on the number theoretic nature of the sizes of the sets vs. the sizes of their intersections.
Theorem 5.19 (Omitted Intersection Theorem) (Frankl–Rödl, 1987). Let m(n, t)
denote the maximum number of subsets of a set of n elements such that no two of the sets
intersect in exactly t elements. Then

(a) m(n, bn/4c) < 1.99n .

(b) For every δ > 0 there exists  > 0 such that if δn ≤ t ≤ (1 − δ)n then

m(n, t) < (2 − )n .

We should mention that taking all subsets of size greater than (n + t)/2 we obtain a
rather large family F such that |E ∩ F | > t for every E, F ∈ F; in particular, intersections
of size t do not occur. Specifically for t = bn/4c, we obtain

m(n, bn/4c) > 1.9378n , (14)

showing that the upper bound of 1.99n is not that far from best possible.

Exercises
♦Ex. 5.4.1. Verify Proposition 5.13.
♦Ex. 5.4.2. Let F be a field and M (n, s) the space of multilinear polynomials of degree ≤ s in n
P s n


variables over F. Prove that dim M (n, s) = k=0 k .
Hint. The monic multilinear monomials (see p. xiii) form a basis of M (n, s). Count the monic
multilinear polynomials of each degree ≤ s.

——————————————————————— 111

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 5. SPACES OF POLYNOMIALS
Ex. 5.4.3. Let F be a field and M (n) the space of all multilinear polynomials in n variables over
F. Prove that dim M (n) = 2n .
Hint. M (n) is the same space as M (n, n) in the previous exercise.

Ex. 5.4.4. For 0 < α < 1, let

1
H(α)
e = . (15)
αα (1 − α)1−α
(H(α) = log2 H(α)
e = −α log2 α − (1 − α) log2 (1 − α) is the “entropy function.”) Derive the
following asymptotic estimate for binomial coefficients not too close to either tail. Assume αn is
an integer.  
n 1 + o(1) 1 n
=p · √ · (H(α))
e . (16)
αn 2πα(1 − α) n
Here, the o(1) notation indicates a quantity that tends to zero as n → ∞ (cf. p. xiii), assuming
γ < α < 1 − γ for some constant γ (which does not depend on n).
√
Hint. Use Stirling’s formula: n! = (n/e)n 2πn(1 + o(1)).

5.5 Geometric application: unit distance is hard to

miss
With Europe fallen to madness all around her, Switzerland was an unlikely small island
of reason during the dismal years of WW2. Among the mathematicians of that fortunate
country, a new school of “combinatorial geometers” sprouted around Hugo Hadwiger, cited
in Section 1.3 for his work on equidissectibility. In one of his earliest papers, Hadwiger (1944)
proposed the following dissection-type problem:1
What is the minimum number c(n) such that Rn can be divided into c(n) subsets Rn =
S1 ∪ . . . ∪ Sc(n) such that no pair of points within the same Si is at unit distance?
This problem lends itself naturally to rephrasing in graph theory language.
Definition 5.20. The distance-δ graph in Rn has the (infinite) set Rn as its vertex set; two
points are adjacent if their (Euclidean) distance is δ. The unit-distance graph corresponds
to δ = 1.
Now the number c(n) Hadwiger asks us to determine is the chromatic number of the
unit-distance graph.
This problem is wide open even in the plane. All that is known is that 7 colors suffice
but 3 don’t. (See Exercises 5.5.1, 5.5.2.)
For general n, we first observe that the chromatic number is finite. Indeed, it is easy to
give an nn/2 bound (Exercise 5.5.3).
Lower bounds are harder to come by. The simplex with unit sidelength shows that we
need at least n + 1 colors. The idea of Exercise 5.5.2 can be adapted to improve this lower
bound to n + 2. The first nonlinear lower bound (Ω(n2 )) was given by Larman and Rogers
1
Hadwiger’s actual problem was slightly different and will be stated in Section 5.8.

112 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.5. THE UNIT DISTANCE GRAPH
√
(1972). At the same time they gave an upper bound of (2 2+o(1))n (cf. Exercise 5.5.4), and
conjectured that the true rate of growth of this function is exponential. This was confirmed
by Frankl and Wilson in 1981, as a rather direct consequence of their modular version of the
RW Theorem (Theorem 7.15). Here we shall deduce the result from Theorem 5.15.
Theorem 5.21 (Frankl–Wilson, 1981). For large n, the chromatic number of the
n n
unit-distance graph on R is greater than 1.2 .
Proof. First we observe that the unit-distance graph and the distance-δ graph on Rn are
isomorphic for any δ > 0, therefore their chromatic number is the same (= c(n)).
The idea, then, is to show that the distance-δ graph of some subset S of the unit cube
Ω = {0, 1}n has exponentially large chromatic number for some δ > 0. (δ will depend on n.)
Each subset S ⊆ Ω corresponds to a set system H ⊆ 2[n] . We write S = S(H) to indicate
the inverse of this correspondence: S(H) consists of the incidence vectors of the members of
H.
Let d(A, B) denote the (Euclidean) distance of the incidence vectors of the sets A, B ∈ H.
Clearly, d(A, B)2 is the size of the symmetric difference of A and B. Assume henceforth that
H is k-uniform. Then
d(A, B)2 = 2(k − |A ∩ B|). (17)
This means that for k-uniform families, the distances are determined by the intersection
sizes. Avoiding a particular distance amounts to avoiding a particular intersection size. This
is how the results of the previous section become relevant.
With a direct application of the “Omitted Intersection Theorem” (Cor. 5.18) in mind,
let us assume for now that n = 4p − 1 for some prime p, and set k = 2p − 1. Moreover,
√ the
intersection size to be avoided should be p − 1; this corresponds to distance δ = 2p.
With this choice of parameters, the graph Gp we examine has vertex set Hp = [4p−1]


2p−1
;
and two sets A, B ∈ Hp are adjacent if |A ∩ B| = p − 1.
We shall prove an exponential lower bound on the chromatic number χ(Gp ). Our strategy
is to prove an upper bound on α(Gp ), the size of the largest independent set. Actually, that
job has already been done: in the language of the graph Gp , Cor. 5.18 translates directly into
the assertion that  
4p − 1
α(Gp ) ≤ 2 · . (18)
p−1
Indeed, by the definition of adjacency in Gp , no two members of F intersect in precisely p − 1
elements; hence Cor. 5.18 applies.
Our final step is an application of Prop. 2.35:
4p−1


2p−1
χ(Gp ) ≥ |Hp | /α(Gp ) ≥ 4p−1

> 1.13974p−1 . (19)
p−1

√
We thus conclude that the chromatic number of the distance- 2p graph of the set S(Hp ) ⊂
R4p−1 is greater than 1.13974p−1 (if p is large enough). It follows for Hadwiger’s c(n) function
that c(n) > 1.1397n for all sufficiently large n of the form n = 4p − 1.

——————————————————————— 113

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 5. SPACES OF POLYNOMIALS
A routine step, using the density of the prime numbers, extends this result to proving
that c(n) > 1.139n for all sufficiently large n (Ex. 5.5.5).
While this suffices to justify the claim of exponential growth, it falls somewhat short of
the stated lower bound 1.2n . To achieve that bound, one has to optimize the parameters of
this proof (Ex. 5.5.7).
It is natural to ask Hadwiger’s question for the rational n-space Qn . Of course, the
same upper bound for the chromatic number of the unit-distance graph remains valid. But
the lower bound argument was based on a lower bound for the chromatic number of the
unit-distance graph on the point set √12p Ω, a set of points with irrational coordinates and
therefore not in Qn .
This difficulty disappears if Qn contains an isometric copy of √12p Ω. One can show that
this is indeed the case when n is divisible by 4 (Ex. 5.8.3). The conclusion is:
Theorem 5.22 (Babai, 1992). For large n, the chromatic number of the unit-distance
graph on Qn is greater than 1.2n .
(See Ex. 5.5.18 for the details.) This result is a slight improvement over a 1.15n lower
bound obtained by Frankl and Wilson (1981) using a different argument which we shall
present in Sec. 5.9. They extended their modular
√ RW theorem to prime power moduli and
replaced p by an odd power of 2. This way 2p became a rational number and the proof
above went through. The reason their bound became slightly weaker is that the odd powers
of 2 do not populate the set of integers as densely as prime numbers do.

Exercises
Ex. 5.5.1. Prove: c(2) ≤ 7. In other words, color the plane with 7 colors such that points at unit
distance receive different colors.
Hint. Color the regions of a hexagonal grid with seven colors appropriately.

♦Ex. 5.5.2. Prove: c(2) ≥ 4. In other words, if we color the plane by three colors, there will be a
pair of points of the same color at unit distance.
Hint. Show that the unit-distance graph in the plane contains a subgraph with 7 points which is
not 3-colorable.

Ex. 5.5.3. Prove: for the n-space, nn/2 colors suffice.

Hint. Divide the unit cube into small cubes of unit diameter.

♦Ex. 5.5.4.* Prove: a simply exponential number (C n for some constant C) of colors suffices for
the n-space for every n.
Hint. Color Rn by 9n colors. Use a sphere packing argument: pick a maximal set of points at
distance ≥ 1/2 apart; color these points so that no two of them at distance ≤ 2 receive the same
color.

♦Ex. 5.5.5. Extend the lower bound c(4p − 1) > 1.1394p−1 to a bound c(n) > 1.139n , valid for all
sufficiently large n.

114 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.5. THE UNIT DISTANCE GRAPH
Ex. 5.5.6. Let Ω(n, k) ⊂ Rn
denote the set of incidence vectors of the k-subsets of [n]. Prove: for
√
any prime p < n/2, the chromatic number of the distance- 2p graph on Ω(n, 2p − 1) is at least
  
n n
. (20)
2p − 1 p−1

♦Ex. 5.5.7. In order to improve the lower bound on the chromatic number of the unit-distance
graph in Rn , maximize, for fixed n, the quantity (20).
Hint. Use the “entropy function” estimate for the binomial coefficients (Ex. 5.4.4).

* * *

The following sequence of exercises serves to clarify questions regarding distances and simi-
larity ratios in Qn and culminates in the proof of Theorem 5.22.
√
♦Ex. 5.5.8. Prove that no pair of points in Q2 is at distance 3.
Hint. Show that the equation x2 + y 2 = 3z 2 has no nonzero integer solutions.
√ √
♦Ex. 5.5.9. Prove that no pair of points in Q3 is at distance 7. (Note that distance 3 does
occur in Q3 .)
Hint. Show that the equation x2 + y 2 + z 2 = 7w2 has no nonzero integer solutions.

Ex. 5.5.10. Show that a number r > 0 occurs as a distance in Q2 if and only if r2 is rational and
all primes of the form 4k − 1 occur with an even exponent in the prime power decomposition of r2 .
Hint. Consult a text on number theory regarding the characterization of sums of squares of two
integers (e. g., Hardy–Wright (1979), Chap. XX).

Ex. 5.5.11. Every positive rational number can be written uniquely as 4` a/b where ` ∈ Z, a, b are
relatively prime positive integers, b is odd, and a is not divisible by 4.
Show that a number r > 0 occurs as a distance in Q3 if and only if r2 is rational and, writing
it as r2 = 4` a/b as above, ab 6≡ 7 (mod 8).
Hint. The necessity follows along the lines of Ex. 5.5.9. For the sufficiency, use the hard part of
Gauss’s theorem that a positive integer k is a sum of squares of three integers if and only if k is of
the form 4` u, where u is not divisible by 4 and u 6≡ 7 (mod 8). (See e. g., Flath (1989), Chap. 5.)

Ex. 5.5.12. Show for every n ≥ 4 that a number r > 0 occurs as a distance in Qn if and only if
r2 is rational.
Hint. Use Lagrange’s theorem that every positive integer is the sum of the squares of four integers
(see e. g., Hardy–Wright (1979), Chap. XX).
♦Ex. 5.5.13. Let r > 0 occur as a distance in Q2 . Show that rQ2 is isometric to Q2 . In other
words, there exists a similarity transformation Q2 → Q2 which stretches every distance by a factor
of r.
♦Ex.
p 5.5.14. Let α, β, γ, δ ∈ Q. Construct a 4 × 4 matrix over Q such that each row has norm
α2 + β2 + γ2 + δ 2 , and the rows are pairwise orthogonal.

——————————————————————— 115

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 5. SPACES OF POLYNOMIALS
♦Ex. 5.5.15. Let r > 0 be the square root of a rational number. Show that rQ4 is isometric to Q4 .
In other words, there exists a similarity transformation Q4 → Q4 which stretches every distance by
a factor of r.
♦Ex. 5.5.16. Show that the result of the preceding exercise holds in any dimension, divisible by
4.
♦Ex. 5.5.17. Show that the statement analogous to Exercises 5.5.13 and 5.5.15 fails to hold in
odd dimensions ≥ 3. More specifically, if n is odd, r2 is rational but r is irrational, then rQn and
Qn are not isometric.
Hint. Show that a necessary and sufficient condition for rQn to be isometric to Qn is that there
exists an n × n matrix A over Q such that the norm of each row of A is r and the rows of A are
orthogonal.
♦Ex. 5.5.18. Prove the lower bound 1.2n for the chromatic number of the unit-distance graph in
Qn (Theorem 5.22).
Hint. Read the paragraph before Theorem 5.22.

5.6 Reducing the diameter of bodies: Borsuk’s conjec-

ture disproved
Dead at the age of 60. Died after no apparent signs of illness, unexpectedly, of grave com-
binatorial causes.
The news of the demise of Borsuk’s venerable conjecture (1933) spread like brushfire
among combinatorialists in Summer 1992.
The disproof, found by Jeff Kahn (Rutgers) and Gil Kalai (Hebrew University), was the
hot topic between lectures at conferences (the result came too late to be included on the
regular programs). Countless copies of the manuscript traveled over electronic networks,
silently crossing oceans and continents at lightening speed. The authors of this book found
out about the result in more conventional ways. One of us heard it from Kahn himself
while examining Gabi Bollobás’s remarkable sculptures at the reception at a meeting in
Cambridge, England. By then, in Tokyo, the other author had learned about it in a telephone
conversation with a friend in New Jersey.
What Kahn communicated in a few minutes and without the benefit of paper or black-
board was not just the news of the result but also the complete proof. Remarkably, Borsuk’s
geometric conjecture was disproved in just a few lines, relying on the Frankl–Wilson Theorem
(Theorem 7.15), a modular version of the RW theorem. Below we deduce it from the result
of the previous section (Theorem 5.21), which in turn was a fairly immediate consequence
of Theorem 5.15, another modular variant of the RW Theorem.
Borsuk conjectured that every set of diameter 1 in Rd can be partitioned into d + 1 sets
of smaller diameter.
The conjecture was widely held to be true. Considerable effort was spent, with a measure
of success, on proving it under additional assumptions. It was verified for centrally symmetric

116 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.6. BORSUK’S CONJECTURE
bodies, for bodies with smooth surface, and for all bodies in dimensions 3 and 2. (See
Ex. 5.6.2) However, the general solution remained elusive. Boltyanski and Gohberg (1985,
p. 31) point out “the sharp contrast between the extreme simplicity of the statement of the
problem, and the huge difficulties in its solution, which seem at present to be completely
insurmountable.”
Yet the conjecture was not merely disproved by an accidental counterexample. The
disproof was devastating. What Kahn√ and Kalai showed was that some bodies needed to be
split into exponentially many (1.2 d ) pieces if the diameter of each piece was to be reduced.
One wonders how a widely known conjecture that fails so badly could stand for so many
decades. Although the subject area has been called “combinatorial geometry” for a long
time (the Russian original of the book by Boltyanski and Gohberg (1985) appeared in 1965),
the depth of its combinatorial structure has not been recognized until recently.
The result also indicates the futility of low dimensional attacks: the lowest dimension in
which the proof provides a counterexample is d = 1, 325.
Let us state the result.
Theorem 5.23 (Kahn–Kalai, 1992). Let f (d) denote the minimum number such that
every set of√ diameter 1 in Rd can be partitioned into f (d) pieces of smaller diameter. Then
f (d) > 1.2 d .
In the other direction,
p it is known that f (d) < 2d (in fact, for every  > 0 and sufficiently
large d, f (d) < ( (3/2) + )d (Schramm (1988)) (cf. Ex. 5.6.1). These upper bounds are
much more interesting now, in the light of the Kahn–Kalai lower bound. Indeed, it seems
the true order of magnitude of f (d) might be close to an exponential function of the form
C d for some constant C > 1.
Kahn and Kalai introduce their miniature proof with a wonderful quotation, the wisdom
of which the reader should bear in mind.
“However contracted, that definition is the result of expanded meditation.”
— Moby Dick

Let us see the fruit of that “expanded meditation.” We shall construct a finite set which
cannot be split into a small number of subsets of smaller diameter.
The proof will be based on the results of the preceding section which the reader is advised
to review.
Here, too, the idea is to consider a subset S of the unit cube Ω = {0, 1}d . Once again,
such a subset corresponds to a set system F ⊆ 2[d] , and we write S = S(F) ⊂ Rd to denote
the set of incidence vectors of the members of F. If F is `-uniform, as we shall assume
henceforth, we obtain (eqn. (17)) that d(A, B)2 = 2(` − |A ∩ B|) for any A, B ∈ F, where
d(A, B) denotes the distance of the incidence vectors of the sets A, B. Thus the maximum
distance occurs when |A ∩ B| is minimal. Let
µ(F) = min{|A ∩ B| : A, B ∈ F}. (21)
To conclude the translation of the geometric problem into combinatorial, we observe that a
partition of S(F) into sets of smaller diameter means a partition of F as F = F1 ∪ · · · ∪ Ft

——————————————————————— 117

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 5. SPACES OF POLYNOMIALS
such that µ(Fj ) > µ(F) for every j. Let g(F) denote the smallest t for which this is possible.
Then clearly f (d) ≥ g(F) for any uniform set system F ⊂ 2[d] .
Another way to think of this problem is to associate a graph G with the set system F
in the following way: the members of F are the vertices of G, and A, B ∈ F are adjacent if
|A ∩ B| = µ(F). Then clearly g(F) is the chromatic number χ(G).
Having worded the problem this way, the analogy with the preceding section is obvious.
Both there and here, we consider graphs, represented by set systems. The vertices of the
graph correspond to the members of the set system; adjacency is defined by some parameter
of the intersection. Such representations appear elsewhere in this book (explicit Ramsey
graph constructions, see Theorems 4.6, 5.24). But the connection between the previous
section and this one is much more intimate: here we consider two different representations
of the same graph. We shall construct the set system F such that the graph G it represents
according to our “minimum intersection size” adjacency rule will be isomorphic to the graph
Gp of the preceding section (represented there by a set system with the “intersection size
= p − 1” adjacency rule). Therefore the strong lower bound on the chromatic number of Gp
(eqn. (19)) will apply to our graph G.
Assume for now that d is of the form 4p−1


2
for some prime number p. We set n = 4p − 1,
[n]


k = 2p − 1, and Hp = k , as in the previous section.
Let, moreover, X = [n] ; so |X| = n2 = d. Our set system F will be defined over the



2
universe X, i. e., F ⊂ 2X .
We shall associate a set Φ(A) ⊂ 2X with each A ∈ Hp . The set system to beat Borsuk’s
conjecture will be
F = {Φ(A) : A ∈ Hp }. (22)
Remember: our goal is to make the correspondence A 7→ Φ(A) such that

|A ∩ B| = p − 1 if and only if |Φ(A) ∩ Φ(B)| = µ(F) (23)

for all A, B ∈ Hp . This will establish that Φ is an isomorphism between Gp and G, as desired.
Here is the simple construction: Φ(A) will be the set of those pairs of elements from [n]
which are split by A. Formally,

Φ(A) = {{x, y} : x ∈ A, y ∈ [n] \ A}. (24)

Clearly Φ(A) ⊂ X and the set system defined by eqn. (22) is `-uniform with ` = k(n − k).
The correspondence A 7→ Φ(A) being one-to-one, all we need to verify is that it preserves
adjacency (eqn. (23)).
To this end, assume |A ∩ B| = r (A, B ∈ Hp ). It is easy to see that

|Φ(A) ∩ Φ(B)| = r(n − 2k + r) + (k − r)2

= 2(r − (k − (n/4)))2 − 2(2k − (n/2))2 + k 2 .

The minimum of this expression is attained when r is as close to k − (n/4) = p − (3/4) as

possible, i. e., when r = p − 1. This completes the proof of the G ∼
= Gp isomorphism.

118 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.7. CONSTRUCTIVE RAMSEY GRAPHS
By inequality (19) we conclude that
4p−1


2p−1
g(F) = χ(G) = χ(Gp ) ≥ > 1.13974p−1 = 1.1397n . (25)
· 4p−1


2 p−1
√
(The last inequality holds when p is sufficiently large.) Since n > 2d, we obtain (for
sufficiently large d) √ √
f (d) ≥ g(F) > 1.1397 2d > 1.203 d , (26)
4p−1


completing the proof of Theorem 5.23 for all dimensions d of the form d = 2 = (4p −
1)(2p − 1) where p is a prime. The extension to all dimensions, using the Prime Number
Theorem, is analogous to the corresponding argument at the end of the proof of Theorem 5.21
(Ex. 5.5.5).

Exercises

Ex. 5.6.1.
Ex. 5.6.2. Prove Borsuk’s conjecture for bodies with smooth boundary.
(See Boltyanski–Gohberg (1985) for the solution.)

5.7 Constructive Ramsey graphs via intersection the-

orems
In Chapter 4, we gave a brief introduction to a problem in Ramsey Theory. Having colored
the edges of a complete graph on N vertices red and blue, we were looking for the largest
possible monochromatic complete subgraph.
In this section we shall use a slightly different language to describe the same problem,
but the change is not essential.
A graph G on a given vertex set V can be identified with a two-coloring of the complete
graph on V : color the edges of the graph red, and the remaining pairs blue.
Now a red complete subgraph will correspond to a complete subgraph of G; and a blue
subgraph to an independent set in G. Let us call a subset of V homogeneous if it induces
either a complete subgraph or an independent set. We wish to maximize the number N of
vertices in a graph without homogeneous t-sets of vertices. (In the notation of Section 4.2,
this number is R2 (t) − 1.) Our concern is the rate of growth of this number as a function of
t.
We know from Erdős’s nonconstructive proof that N can be exponentially large (essen-
tially 2t/2 ) without forcing homogeneous t-subsets. The question is, how close we are able
to come to this by explicit construction. The construction by Zs. Nagy given in Section 4.2

——————————————————————— 119

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 5. SPACES OF POLYNOMIALS
shows that elementary intersection theorems can be invoked to construct such Ramsey graphs
of nontrivial (ct3 ) size. This approach was generalized by Frankl (1977) to construct super-
polynomial size (i. e., tω(t) , where ω(t) → ∞) Ramsey graphs. His argument employed the
theory of sunflowers (cf. Section 4.4). Subsequently a simpler proof, resulting in essentially
the same rate of growth, was found by Frankl and Wilson (1981), using their modular version
of the RW Theorem. We present a slight variation of this proof, based on Theorem 5.15.

Construction. Let p be a prime number and n > 2p2 . Set

 
[n]
V = ,
p2 − 1

the set of all subsets of [n] of cardinality p2 − 1. We construct a graph G(n, p) on the vertex
set V . Let us join two vertices A, B ∈ V if

|A ∩ B| 6≡ −1 (mod p).

Observe that for p = 2, this graph becomes precisely the graph of Nagy, described in the
proof of Theorem 4.6.
Theorem 5.24. The graph G(n, p) has p2n−1 vertices and no homogeneous subgraph on


n


more than 2 · p−1 vertices.
2
Proof. Let k = p − 1. Assume F = {E1 , . . . , Em } is the vertex set of a complete subgraph
of G(n, p). This means |Ei ∩ Ej | 6≡ −1 (mod p) for any pair i 6= j (1 ≤ i, j ≤ m). In other
words, F is a k-uniform set system satisfying the conditions of Theorem 5.15 with s = p − 1
and L = {0, 1, . . . , p − 2}. The conclusion (using Prop. 5.13, eqn. (8)) is that

n
m≤2· , (27)
p−1

as desired.
Assume now that F is the vertex set of an empty subgraph of G(n, p). This means that
for every i 6= j, |Ei ∩ Ej | ∈ L where L = {p − 1, 2p − 1, . . . , p2 − p − 1}. We conclude by
Corollary 5.17 that inequality (27) holds again.
We note that the factor of 2 can be omitted in the upper bound by citing the stronger
upper bound (5), valid for uniform set systems. However, the gain is asymptotically negligi-
ble.
Corollary 5.25. Let ω(t) = ln t/(4 ln ln t). For every  > 0 and sufficiently large t, one can
construct an explicit graph on more than

t(1−)ω(t) (28)

vertices and with no homogeneous subgraphs of size t.

120 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.8. MISSING A DISTANCE
Let us stress again that the emphasis is on explicit constructibility. The mere existence of
much larger (2t/2 ) Ramsey graphs is known (see Section 4.2).
3
Proof. We just have to select
p and n appropriately. Let n = p . Select p to be the
p3
largest prime such that p−1 < t. For every δ > 0 and large enough t, this prime will be
between (1 ± δ) ln t/(2 ln ln t), using the Prime Number Theorem and elementary estimates
of the binomial coefficients. Easy calculation confirms the stated estimate for the number of
vertices.

It is an open problem, how to improve further the rate of growth of the function ω(t) in the
exponent.

Exercises
Ex. 5.7.1. Work out the calculation indicated at the end of the proof of Corollary 5.25.

5.8 Geometric application: any distance is hard to miss

A set S ⊂ Rn is said to miss distance δ if δ does not occur among the pairwise distances of
the points in S.
In Section 5.5, we considered Hadwiger’s problem splitting the space into subsets each of
which misses the unit distance.
Actually, Hadwiger’s question as well as the partial results and the conjecture of Larman
and Rogers concerned a slightly different problem. What they asked for is bounds for the
smallest integer m(n) such that the space Rn can be partitioned into m(n) subsets, each of
which misses some distance.
In the problem discussed in Section 5.5, each class of the partition was required to miss
the same distance (distance 1); in Hadwiger’s actual problem, each class is required to miss
some distance. Accordingly, a coloring of the unit-distance graph certainly gives an upper
bound on m(n), but in order to obtain a lower bound on m(n), an additional idea is needed.
This idea was provided by Larman and Rogers.
Lemma 5.26 (Larman–Rogers, 1972). Assume Rn has a finite subset S such that the
unit-distance graph on S has no independent set of size greater than t. Then

m(n) ≥ |S| /t. (29)

This is a considerable sharpening of the idea of using Prop. 2.35 to obtain a chromatic
number lower bound. In view of inequality (18) and its subsequent use to derive the chromatic
number lower bound in Section 5.5, the Larman–Rogers Lemma shows that precisely the
same lower bound is valid for Hadwiger’s function m(n):
Theorem 5.27. For large n, the space Rn cannot be partitioned into fewer than 1.2n
subsets, each missing a distance.

——————————————————————— 121

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 5. SPACES OF POLYNOMIALS
We begin the proof of the Larman–Rogers Lemma with a simple general position argu-
ment. (Cf. Section 3.1.) This time the objects in general position will be rotations of the
space; the nature of the coincidence they will be supposed to avoid is stated next.
n
Lemma 5.28 (General rotations). Let S1 , . . . , SN be Q finite subsets of R , n ≥ 2. Then
there exist isometric copies Ri of the Si such that the |Si | sums constituting the set
N
X
R1 + · · · + RN = vi : vi ∈ Ri
i=1

are all distinct.

Proof. Assume first that N = 2. Let Di = {u − v : u, v ∈ Si u 6= v}. Distance-preserving
linear maps Rn → Rn are called orthogonal transformations. We shall find an orthogonal
transformation ϕ such that D1 ∩ ϕ(D2 ) = ∅. Then R1 := S1 , R2 := ϕ(S2 ) will clearly be an
appropriate choice.
First we observe that if n = 2 then almost any rotation of the plane about the origin will
be all right; there are only a finite number of wrong angles.
If n = 3, we proceed as follows. Take a line U through the origin that does not intersect
D2 . Project D1 and D2 perpendicularly to the plane U ⊥ . Rotate the image of D2 about the
origin by an appropriate angle such as to become disjoint from the image of D1 . Apply the
same rotation about U to D2 itself. We leave it to the reader to generalize this process to
any dimension n ≥ 3. (Exercise 5.8.1.)
Assuming now that we are done for N = 2 and arbitrary n ≥ 2, an easy induction on N
completes the proof.
Proof of Lemma 5.26. We call a finite set S ⊂ Rn (δ, t)-critical if the distance-δ graph (see
Def. 5.20) on S has no independent set of size greater than t. The lemma says that if there
exists a (1, t)-critical set S ⊂ Rn then m(n) ≥ |S| /t.
Let us first note that any isometric copy of a (δ, t)-critical set is again (δ, t)-critical. Also,
for any γ > 0, the set γS := {γv : v ∈ S} is (γδ, t)-critical.
Suppose now Rn = Y1 ∪ · · · ∪ YN and Yi misses the distance αi . We have to prove:

N ≥ |S| /t. (30)

Let us set Si = αi S. Then Si is (αi , t)-critical. Now the trick is to construct a finite set
C ⊂ Rn which can, for every i, be split into the disjoint union of isometric copies of Si . Once
such a set C is found, we are done, because Yi (the set that misses distance αi ) contains at
most t points from each isometric copy of Si and therefore at most a t/ |Si | = t/ |S| fraction
of the points of C. As the N sets C ∩ Yi add up to C, and each of them contains at most a
t/ |S| fraction of C, there must be at least |S| /t of them, proving inequality (30).
The construction of C is not difficult. We apply Lemma 5.28 to the system S1 , . . . , SN .
Now we take the set C = R1 + · · · + RN , which, according to the proposition, consists of
|S|N distinct sums of the form v1 + · · · + vN (vi ∈ Ri ).
We claim that for each i, the set C is, as required, the disjoint
S union of isometric copies
of Si . Let Bi = R1 + · · · + Ri−1 + Ri+1 + · · · + RN . Now C = v∈Bi (Ri + v), and, according

122 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.9. PRIME POWER MODULI
N −1
to the definition of C, these |Bi | = |S| translates Ri + v are pairwise disjoint. They all
are isometric to Ri and therefore to Si , proving the claim.
These arguments can be extended (nontrivially) to the rational space Qn .
Let D(n) denote the set of distances occurring among pairs of points in Qn . These sets
have been determined, for each n, in Exercises 5.5.8–5.5.12. For n ≥ 4, the set D(n) consists
of all r > 0 such that r2 ∈ Q.
We shall say that a set S ⊂ Qn misses a distance if ∃r ∈ D(n) such that r does not occur
among the pairwise distances in S.
Let mQ (n) denotes the smallest integer such that Qn can be partitioned into mQ (n)
subsets, each missing a distance.
Theorem 5.29 (Babai, 1992). For large n, mQ (n) ≥ 1.2n . In other words, the space Qn
cannot be partitioned into fewer than 1.2n subsets, each missing a distance.
The proof follows from Theorem 5.15 along the lines of the proof above. We leave the
somewhat delicate details as Exercises 5.8.3–5.8.5.

Exercises
♦Ex. 5.8.1. Complete the proof of “General rotations” lemma (Lemma 5.28) by settling the case
N = 2 for arbitrary n ≥ 3.
Hint. Use a “general position” argument.
♦Ex. 5.8.2. Prove that there are infinitely many orthogonal transformations of Q2 .
Hint. We have to construct infinitely many 2 × 2 rational matrices A such that AT A = I2 .
Ex. 5.8.3. Prove the “General rotations” lemma (Lemma 5.28) over Q.
Hint. Verify that the proof given in the text and in Ex. 5.8.1 goes through over Q. For the
two-dimensional case, use Ex. 5.8.2.
♦Ex. 5.8.4. Prove the following rational version of the Larman–Rogers Lemma (Lemma 5.26).
Let r be the square root of a rational number and assume n is divisible by 4. Assume Qn has a
finite subset S such that the distance-r graph on S has no independent set of size greater than t.
Then
mQ (n) ≥ |S|/t. (31)

Hint. Follow, mutatis mutandis, the proof of Lemma 5.26 given in the main text.
Ex. 5.8.5. Prove Theorem 5.29.
Hint. Combine the proof of Theorem 5.22 (Ex. 5.5.18) with Ex. 5.8.4.

5.9 Prime power moduli

An important open question is the extension of Theorem 5.15 to composite moduli. The
O(ns ) upper bound is no longer valid in general (even if the family is uniform; see Exer-

——————————————————————— 123

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 5. SPACES OF POLYNOMIALS
cises 5.9.3–5.9.5). Two conjectured cases when it may hold are stated at the end of Sec-
tion 7.3.
Here we shall consider prime power moduli q in the special case s = q − 1, i. e., when the
intersections are allowed to occupy all residue classes but that of k, the residue class of the

Pb(q−1)/2c n
sizes of the sets. For this case we have the non-uniform style upper bound i=0 q−1−2i
n


although conceivably the q−1 term on the right hand side might suffice. It does for uniform
families, as we shall see in Section 7.3, Theorem 7.18. Here we give the weaker upper bound
but drop the uniformity condition. A geometric application (coloring the unit-distance graph
of the rational space) will follow in the next section.
Theorem 5.30. Let k be an integer and q = pα a prime power. Assume F = {A1 , . . . , Am }
is a family of subsets of a set of n elements such that
(a) |Ai | ≡ k (mod q) for i = 1, . . . , m;

(b) |Ai ∩ Aj | 6≡ k (mod q) for i 6= j (1 ≤ i, j ≤ m).

Then      
n n n
m≤ + + + ··· . (32)
q−1 q−3 q−5
Since the integers mod q don’t form a field, we shall somehow have to transform the
alternatives given in the Theorem to statements with respect to a prime modulus. The
following observation will help us achieve this and pick the right functions.
r−1


Proposition 5.31. Let q = pα , p a prime. For any integer r, the binomial coefficient q−1
is divisible by p precisely if r is not divisible by q.
We leave the proof as Exercise 5.9.2.

Proof of the Theorem. Let us consider the polynomials

 
x · vi − k − 1
fi (x) = (33)
q−1

in n real variables x = (x1 , . . . , xn ) ∈ Rn (i = 1, . . . , m). Observe that by Proposition 5.31,

the integer  
|Ai ∩ Aj | − k − 1
fi (vj ) = (34)
q−1
will be divisible by p if and only if i 6= j.
Notice also, that this situation will not change, if we replace the variable xn by x0n =
k − x1 − · · · − xn−1 , because by condition (a), we have

x0n ≡ xn (mod q)

if (x1 , . . . , xn ) is one of the vi . Let gi (x) = fi (x1 , . . . , xn−1 , x0n ). This is a function of the
variables x1 , . . . , xn−1 and we only indicate the nth coordinate for convenience. We thus
found that the m × m matrix G = (gi (vj )) is congruent mod p to a diagonal matrix with no

124 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.9. PRIME POWER MODULI
zeros (mod p) in the diagonal. It follows that det G 6≡ 0 (mod p) and therefore det G 6= 0.
Let us now replace the gi by the corresponding multilinear polynomials gei according to
Prop. 5.16, defined uniquely by the condition that gi and gei agree on the (n − 1)-cube
Ω = {0, 1}n−1 ⊂ Rn−1 . Since the vectors formed by the first n−1 coordinates of the vi belong
to Ω, we have G = (e gi (vj )). Hence we infer by the Determinant Criterion (Proposition 2.7)
that the multilinear polynomials gei are linearly independent.
On the other hand, the polynomials gei have degree ≤ q − 1. As before, we conclude (with
reference to Ex. 5.4.2) that

q−1   b(q−1)/2c  
X n−1 X n
m≤ = .
k=0
k t=0
n − 2t

* * *

As an application, we again consider Hadwiger’s problem of coloring the unit-distance

graph (Section 5.5) restricted to the rational space Qn .
Recall that we used Ω(n, k) to denote the set of incidence vectors of the k-subsets of [n];
so Ω(n, k) is a subset of the unit cube Ω = {0, 1}n ⊂ Qn .
√
Lemma 5.32. For any prime power q < n/2, the chromatic number of the distance- 2q
graph on Ω(n, 2q − 1) is at least
   
n n
n . (35)
2q − 1 q−1
Pq−1 n

n


Proof. The denominator we obtain using Theorem 5.30 is i=0 i
<n q−1
.

In fact, the factor of n can be omitted from the denominator by using Theorem 7.18
(Section 7.3) in place of Theorem 5.30. The order of magnitude of the quantity (35) is,
however, exponential (with the right choice of q), and so a factor of n has no influence on
the asymptotic result.
We can guarantee is that an odd power of 2 exists within any interval of the form (`, 4`].
Selecting ` = 0.06n will guarantee a lower bound of 1.15n on the quantity (35) for q the
odd power of 2 in the interval (0.06n, 0.24n]. This proves the exponential growth rate of the
chromatic number of the unit-distance graph in the rational space; and this is the way the
first proof of this fact went.
Corollary 5.33 (Frankl–Wilson, 1981). For large n, the chromatic number of the
n n
unit-distance graph on Q is greater than 1.15 .
Although this lower bound is slightly weaker than the bound 1.2n obtained in Theo-
rem 5.22, we have included the proof since it illustrates the role prime power moduli can
play in geometric applications.

——————————————————————— 125

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 5. SPACES OF POLYNOMIALS
Exercises
Ex. 5.9.1. Replace the prime number p by an arbitrary prime power q in the Construction. Prove
that Theorem 5.24 remains valid.
Hint. Use Theorem 5.30 rather than Theorem 5.15 in the proof.
♦Ex. 5.9.2. Prove Proposition 5.31.
♦Ex. 5.9.3. Let us consider the polynomial
r  
X x
f (x) = αi ,
i
i=0

where the αi are nonnegative integers. For a set L of nonnegative integers, set f (L) = {f (`) : ` ∈ L}.
Given a k-uniform L-intersecting family F on n points, construct an f (k)-uniform f (L)-intersecting
family G on f (n) points such that |F| = |G|.
♦Ex. 5.9.4. Prove that Theorem 5.15 will be false if the prime number p is replaced by p = 6; even
the order of magnitude of the right hand side will be wrong. Make your counterexamples uniform.
Hint. Let L = {3, 4, 5, 6, 7, 8, 9, 10} and f (x) = x2 . Construct a large 11-uniform L-intersecting

family on n points and
apply the construction of the previous exercise to obtain a 55-uniform family
of the same size on n2 points.
♦Ex. 5.9.5. Prove that Theorem 5.15 will be false if the prime number p is replaced by q = p2 ,
where p ≥ 7; even the order of magnitude of the right hand side will be wrong. Make your
counterexamples uniform.
Hint. Let L = {2, 3, . . . , q − 2} and f (x) = x2 . Construct a large (q − 1)-uniform L-intersecting
family on n points and apply the construction above (Exercise 5.9.3) to obtain a (q − 1)2 -uniform
family of the same size on n2 points.

5.10 The nonuniform RW Theorem

A slight modification of the proof of Theorem 5.15 yields a nonuniform version of the RW
Theorem. The original proof of this result (Frankl–Wilson, 1981) uses the method of higher
incidence matrices and will be reproduced in Chapter 7.
Theorem 5.34 (Nonuniform RW Theorem) (Frankl–Wilson, 1981)). Let L be a
set of s integers and F an L-intersecting family of subsets of a set of n elements. Then
     
n n n
|F| ≤ + + ··· + . (36)
s s−1 0
This result is best possible in terms of the parameters n and s, as shown by the family
of all subsets of size ≤ s of a set of n elements.

Proof (Babai, 1988). Let L = {`1 , . . . , `s } and F = {A1 , . . . , Am }, where Ai ⊆ [n] and
|A1 | ≤ · · · ≤ |Am |. With each set Ai we associate its incidence vector vi ∈ Rn . As in earlier

126 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.10. THE NONUNIFORM RW THEOREM
Pn n
sections, we use x · y = i=1 xi yi to denote the standard inner product of x, y ∈ R . Clearly,
vi · vj = |Ai ∩ Aj |.
For i = 1, . . . , m, let us define the polynomial fi in n variables as follows:
Y
fi (x) = (vi · x − lk ) (x ∈ Rn ). (37)
k
lk <|Ai |

Observe that (
6= 0, if j = i ;
fi (aj ) (38)
= 0, if j < i .
The result now follows in the same way as the proof of Theorem 5.15 has followed
from eqn. (12) via multilinearization (Prop. 5.16). (This time, we have to use the Triangle
Criterion (Prop. 2.5).

Exercises
♦Ex. 5.10.1 (Stronger Nonuniform Modular RW Theorem) (Alon–Babai–Suzuki, 1991).
Let p be a prime number and L1 , . . . , Lm sets of integers, |Li | = s. Assume F = {A1 , . . . , Am } is a
family of subsets of a set of n elements such that
(a) |Ai | ∈
/ Li (mod p) (1 ≤ i ≤ m);

(b) |Ai ∩ Aj | ∈ Li (mod p) (1 ≤ j < i ≤ m).

Then      
n n n
m≤ + + ··· + . (39)
s s−1 0

Show that this result is a common generalization of Theorems 5.15 and 5.34.
Hint. Combine the proofs of Theorems 5.15 and 5.34.
♦Ex. 5.10.2. In the proof of Theorem 5.34, the definition of the fi seems more complicated than
necessary. Why not just define, following the proof of Theorem 5.15, a function in 2n variables by
s
Y
F (x, y) = (x · y − lk ),
k=1

and set
fi (x) = F (x, vi ).
What will go wrong? What is the significance of ordering the Ai by size?
n


Ex. 5.10.3. Prove: if |L| = 2, then a uniform L-intersecting family has m ≤ 2 + 1 members.
(This number exceeds the s = 2 case of the RW upper bound by only 1.)
Hint. Use the (now correct) approach suggested
Pn by the previous exercise. Restrict the domain to
the intersection of Ω with the hyperplane k=1 xk = r where r = |Ai |. This will eliminate one

——————————————————————— 127

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 5. SPACES OF POLYNOMIALS
variable.
n−1

The
n−1

result
n−1 n

the same bound as in Theorem 5.34, with n − 1 in the place of n:

is thus
2 + 1 + 0 = 2 + 1.

Ex. 5.10.4. Prove: if |L| = s, then the number of members of a uniform L-intersecting family is
     
n n n
m≤ + + + ··· . (40)
s s−2 s−4

Hint. The proof isPidentical
with that of the preceding exercise, observing that the stated upper
bound is equal to sk=0 n−1 k .
Remark. The RW inequality eliminates all but the first term here, so the rest may be regarded
as error in the upper bound. Observe that the relative error is quite small; it is of the order of
O((s/n)2 ) as long as s < cn for some constant c < 1/2.

5.11 The Ray-Chaudhuri – Wilson Theorem

Let us state the result again.
Theorem 5.35 (D. K. Ray-Chaudhuri – R. M. Wilson, 1975). Let L be a set of s
integers and F an L-intersecting k-uniform family of subsets of a set of n elements. Then
 
n
|F| ≤ .
s

The proof we describe here is an extension of the idea presented in the preceding section,
along the lines of Blokhuis’s improvement of the bound for 2-distance sets discussed in
Section 1.2 (Blokhuis 1981, 1984).
Let Ω = {0, 1}n be the n-cube. We consider the space RΩ consisting of all functions
f : Ω → R. If f is defined by a polynomial of degree t then f , as a function on Ω, can be
replaced by a multilinear polynomial of degree ≤ t (Prop. 5.16).
The domain Ω can be identified with the set of subsets of [n] so if I ⊆ [n] and f ∈ RΩ
we write f (I) for f (vI ) where vI is the incidence vector of I. Moreover, we index the monic
multilinear monomials by the sets of their variables:
Y
xI := xi .
i∈I

In particular, x∅ = 1. Observe that for J ⊆ [n],

(
1 if I ⊆ J ;
xI (J) = (41)
0 otherwise.

We begin with a lemma.

Lemma 5.36. Let f ∈ RΩ . Assume f (I) 6= 0 for any |I| ≤ r. Then the set {xI f : |I| ≤
r} ⊆ RΩ is linearly independent.

128 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.11. RAY-CHAUDHURI – WILSON THEOREM
Proof. Let us arrange all subsets of [n] in a linear order, denoted ≺, such that J ≺ I implies
|J| ≤ |I|, J 6= I. By equation (41) we see that for every I, J ⊆ [n], if |I| , |J| ≤ r, then
(
f (I) 6= 0 if J = I ;
xI (J)f (J) =
0 if J ≺ I .

By the Triangular Criterion, linear independence of the xI f follows.

Proof of Theorem 5.35 (Alon–Babai–Suzuki, 1991) We use the notation introduced in the
first paragraph of the proof of Theorem 5.34 in the preceding section. For i = 1, . . . , m, we
define the functions fi ∈ RΩ as follows:
s
Y
fi (x) = (vi · x − lk ) (x ∈ Ω). (42)
k=1

Observe that (
6= 0 if j = i ;
fi (Aj ) (43)
=0 6 i.
if j =
Now we claim more than just the P linear independence of the functions fi . Even the fi
n

together with all the functions xI x
j=1 j − k for I ⊆ [n], |I| ≤ s − 1 remain linearly
independent.
For a proof of this claim, assume
m
X X n
X 
λi fi + µ I xI xj − k = 0 (44)
i=1 |I|≤s−1 j=1

for some λi , µI ∈ R. Substituting Ai , all terms in the second sum vanish because |Ai | = k,
and by (43), only the term with subscript i remains of the Pn first sum. We infer that λi = 0
for every i and therefore (44) is a relation among the xI ( j=1 xj − k). By Lemma 5.36, this
relation must be trivial.P
s−1 n

We thus found m + i=0 i
linearly independent functions, all of which are represented
P of
degree ≤ s. The space of such (now always multilinear)
by polynomials
polynomials has
dimension si=0 ni , forcing m not to be greater than the difference, ns .

Exercises
Ex. 5.11.1 (Stronger Nonuniform RW Theorem) (Alon–Babai–Suzuki, 1991). Prove the following
version of the RW Theorem. Observe that it includes both the uniform and the nonuniform RW
Theorems as particular cases.
Let K = {k1 , . . . , kr } and L = {l1 , . . . , ls } be two sets of nonnegative integers and assume ki > s−r.
Let F be an L-intersecting family of subsets of a set of n elements. Assume the size of every set in
F is a number from K. Then
     
n n n
|F| ≤ + + ··· + .
s s−1 s−r+1

——————————————————————— 129

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 5. SPACES OF POLYNOMIALS

Hint. Proceed as in the proof above, replacing the factor ( xj − k) by f = ri=1 ( xj − ki ), and
P Q P
using the auxiliary functions xI f only for |I| ≤ s − r.

5.12 A modular Ray-Chaudhuri – Wilson Theorem

Modular versions of the RW Theorem are powerful tools in applications to geometric and
combinatorial problems, as we have seen in previous sections.
Although for the applications given, upper bounds of the form si=1 ni were
P

sufficient,
n


the chronologically first such result came with the tight and more appealing s upper bound
(Frankl–Wilson, 1981). Below we give a slight generalization of that result.
Theorem 5.37 (Modular RW Theorem). Let p be a prime number and L a set of
s ≤ p − 1 integers. Let k be an integer, k ∈
/ L (mod p). Assume s + k ≤ n.
Let F be a family of subsets of a set of n elements such that

(i) |E| ≡ k (mod p) for E ∈ F;

(ii) |E ∩ F | ∈ L (mod p) for E, F ∈ F, E 6= F .

Then  
n
|F| ≤ . (45)
s
The uniform case of this theorem (all members of F have size k rather than ≡ k (mod p) was
the main result of Frankl–Wilson (1981). Their original proof, based on higher incidence
matrices, will be presented in Chapter 7. We shall now prove the stated nonuniform extension
(Alon–Babai–Suzuki, 1991) along the lines of the proof of the RW Theorem given in the
previous section.
At this point the reader should review that proof and try to mimic it under the new
conditions. Of course we have to replace R by Fp , and RΩ by FΩ p.
We should notice that only one thing can go wrong: when we get to the application of
Lemma P 5.36, the condition f (I) 6= 0 may be violated for some I ⊂ [n], |I| ≤ s − 1. Here,
f (x) = nj=1 xj − k, so f (I) = |I| − k.
Observe that in an important special case, this trouble does not occur: f (I) 6= 0 for any
I, |I| ≤ s − 1, unless k ∈ [s − 1] (mod p). We conclude that

if k 6∈ [s − 1] (mod p), the proof of Theorem 5.37 is complete. (46)

We should point out that this particular case covers all the applications given in previous
sections; in all those cases (in Ramsey theory as well as in geometry) we had k ≡ −1 (mod p)
and s = p − 1 (L = {0, 1, . . . , p − 2}). It is no surprise that this “easy” case includes the RW
Theorem as well (just select a prime p > n).

130 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.12. MODULAR RW THEOREM
For the general case of Theorem 5.37 we have to work a bit harder. We need a lemma,
considerably stronger than Lemma 5.36. We can no longer assume that our multiplier f
does not vanish on small sets.
Recall that now we work in the function space FΩ n
p , where, as before, Ω = {0, 1} is the
n-cube. We use the notation introduced before Lemma 5.36.
Pn
Lemma 5.38. Let p be a prime; Ω = {0, 1}n . Define f ∈ FΩ p by f (x) = i=1 xi − k.
Assume 0 ≤ s, k ≤ p − 1 and s + k ≤ n. Then the set {xI f : |I| ≤ s − 1} ⊆ FΩ p is linearly
independent (over Fp ).
Given this lemma, the proof of Theorem 5.37 proceeds along the lines of the previous section.
Like there, we claim that the set of functions fi ∈ FΩ (i = 1, . . . , m = |F|) defined by equation
Pn
p
(42) together with the functions xI j=1 xj −k (I ⊆ [n], |I| ≤ s−1) is linearly independent.
The reason is the same as there; an appeal to Lemma 5.38 concludes the proof.
We devote the rest of this section to the proof of Lemma 5.38.
First we have to introduce the basic concepts of “Moebius inversion over the Boolean
lattice.” The Boolean lattice over the set X is the set 2X of subsets of X, viewed as a
partially ordered set (the partial order being set inclusion).
Let F be a field and α : 2X → F a function. The zeta transform of α is a function
β : 2X → F defined by X
β(Y ) = α(W ). (47)
W ⊆Y

for all Y ⊆ X. The Moebius transform of a function γ : 2X → F is a function δ : 2X → F

defined by X
δ(W ) = (−1)|W | (−1)|T | γ(T ). (48)
T ⊆W

for all W ⊆ X. These two operations are inverses of one another:

Proposition 5.39 (Moebius inversion). If β is the zeta-transform of α then α is the
Moebius transform of β.
We leave the easy proof as Ex. 5.12.4.
The following relation between α and its zeta-transform β is easy to verify (Ex. 5.12.5).
Proposition 5.40. For any pair of sets W ⊆ Y ⊆ X, we have
X X
(−1)|T | β(T ) = (−1)|Y | α(U ). (49)
W ⊆T ⊆Y Y \W ⊆U ⊆Y

Now the following statement is immediate (see Ex. 5.12.6).

Corollary 5.41. For any integer s, the following are equivalent for a function α : 2X → F
and its zeta-transform β:

(i) α(U ) = 0 whenever |U | ≥ s (U ⊆ X).

|T |
P
(ii) W ⊆T ⊆Y (−1) β(T ) = 0 whenever |Y \ W | ≥ s (W ⊆ Y ⊆ X).

——————————————————————— 131

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 5. SPACES OF POLYNOMIALS
Definition 5.42. We shall say that a set H = {h1 , . . . , hm } ⊆ [n] has a gap of size ≥ g
(where the hi are arranged in increasing order), if either h1 ≥ g − 1, or n − hm ≥ g − 1, or
hi+1 − hi ≥ g for some i (1 ≤ i ≤ m − 1).

We are now ready to formulate a statement from which Lemma 5.38 will readily follow.
Lemma 5.43. Let |X| = n. Let α : 2X → F be a function and β its zeta-transform. Let
H ⊆ {0, 1, . . . , n} be a set of integers and s an integer, 0 ≤ s ≤ n. Let us make the following
assumptions:

(a) For U ⊆ X, we have α(U ) = 0 whenever |U | ≥ s.

(b) For T ⊆ X, we have β(T ) = 0 whenever |T | 6∈ H.

(c) H has a gap ≥ s + 1.

Then α = β = 0.
Proof. Let H = {h1 , . . . , hm } (1 ≤ h1 < h2 < . . . < hm ≤ n). We proceed by induction
on m. If m = 0 then β = 0 by definition, hence its Moebius transform, α, also vanishes.
Assume now m ≥ 1.
Let us add h0 = −1 and hm+1 = n + 1 to H; and let hi+1 − hi ≥ s + 1 be a gap as
required. Let us temporarily assume that i 6= 0.
Consider any pair of sets W ⊆ Y ⊆ X, |W | = hi , |Y | = hi + s. (Observe that hi + s ≤ n.)
By Cor. 5.41, we have X
(−1)|T | β(T ) = 0. (50)
W ⊆T ⊆Y

Because of the gap in H, the only possibly nonvanishing term on the left hand side cor-
responds to T = W ; therefore this term, too, must vanish. We conclude that β(W ) = 0
whenever |W | = hi , thus eliminating a member of H. This completes the induction step in
the case i 6= 0.
If i = 0, we take Y to have cardinality h1 and its subset W to have cardinality h1 − s.
(Observe that h1 − s ≥ 0.) Now the same argument as before shows that β(Y ) = 0, thus
eliminating h1 from H and thereby completing the proof.
The proof of Lemma 5.38 now follows. Let H = {h ∈ Z : 0 ≤ h ≤ n, h ≡ k (mod p)}.
(These are the admissible sizes of members of the family F in Theorem 5.37.) Observe that
under the conditions of Lemma 5.38, H has a gap of size ≥ s + 1 since k ∈ H, k + i 6∈ H for
i = 1, . . . , s (because s ≤ p − 1), and k + s ≤ n.
Set X = [n]. Assume that a linear dependence relation exists among the functions
{xI f : |I| ≤ s − 1}. Let α(I) denote the coefficient of xi f in this relation; set α(I) = 0 for
all |I| ≥ s (I ⊆ [n]). So the linear relation takes the following from:
X
α(I)xI f = 0. (51)
I⊆[n]

132 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 5.12. MODULAR RW THEOREM
Note that by the definition of f , f (J) = 0 precisely if |J| ≡ k (mod p), i. e., if |J| ∈ H. Now
let J ⊆ [x], |J| 6∈ H; therefore f (J) 6= 0. Substituting J for the variable x into eqn. (51), we
can divide by f (J) and we obtain
X X
β(J) = α(I) = α(I)xJ = 0 (52)
I⊆J I⊆[n]

so β(J) = 0 whenever |J| 6∈ H. We have thus verified that all conditions of Lemma 5.43
are met hence its conclusion applies. The conclusion is that α = β = 0; in particular, all
coefficients of our purported linear relation are zero. This completes the proof of Lemma 5.38
and thereby the proof of Theorem 5.37.

Exercises
Ex. 5.12.1. Prove the following generalization of Lemma 5.38.
Let p be a prime; Ω = {0, 1}n . Let K ⊆ {0, 1, . . . , p − 1} be a set of r integers. Assume
r(s − r + 1) ≤ p − 1 and s + kr ≤ n, where kr is the largest element in K. Define the polynomial
f ∈ FΩ
p in n variables by
Y
f (x) = f (x1 , . . . , xn ) = (x1 + . . . + xn − k). (53)
k∈K

Then the set of polynomials {xI f : |I| ≤ s − r} is linearly independent over Fp .

Hint. Let H = {h ∈ Z : 0 ≤ h ≤ n, h ∈ K (mod p)}. Show that under the given conditions,
H ⊆ [n] has a gap ≥ s − r + 2. Apply Lemma 5.43 exactly the way it was used to conclude the
proof of Lemma 5.38.
Ex. 5.12.2 (Alon–Babai–Suzuki, 1991). Prove the following generalization of Theorem 5.37.
Let p be a prime number and K, L two disjoint subsets of {0, 1, . . . , p − 1}. Let |K| = r, |L| = s.
Assume r(s − r + 1) ≤ p − 1 and s + kr ≤ n, where kr is the largest element of K.
Let F be a family of subsets of a set of n elements such that
(i) |E| ∈ K (mod p) for E ∈ F;

(ii) |E ∩ F | ∈ L (mod p) for E, F ∈ F, E 6= F .

Then      
n n n
|F| ≤ + + ... + . (54)
s s−1 s−r+1

Hint. Very little change is required compared to the proof of Theorem 5.37. Use the preceding
exercise in the place of Lemma 5.38.
Remarks. 1. Compare the result with Ex. 5.11.1. 2. For r ≥ 2, we don’t know if equality can be
attained. (The set of subsets of sizes s, s − 1, . . . , s − r + 1 does not qualify because of the condition
K ∩ L = ∅.) 3. The condition r(s − r + 1) ≤ p − 1 seems too restrictive when r ≥ 2. Perhaps this
condition can be dropped entirely. (Note that r + s ≤ p will still hold because K ∩ L = ∅.)

——————————————————————— 133

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 5. SPACES OF POLYNOMIALS
* * *

The remaining exercises verify statements made in the main text.

♦Ex. 5.12.3. Prove: for arbitrary subsets W, Y ⊆ X,
(
X
|Y \T | 1, if W = Y ;
(−1) = (55)
W ⊆T ⊆Y
0, 6 Y.
if W =

♦Ex. 5.12.4. Verify the Moebius inversion formula (Prop. 5.39).

♦Ex. 5.12.5. Prove Prop. 5.40.

Ex. 5.12.6. Deduce Corollary 5.41.

Hint. The direction (i)⇒(ii) is immediate from Prop. 5.40. We note that only this direction was
used in the proofs in the main text.
For the other direction, prove α(Y ) = 0 for all |Y | ≥ s by induction on |Y |, starting from
|Y | = s.

134 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
Chapter 6

Tensor product methods

Those not familiar with the elements of multilinear algebra might be puzzled to learn that
exterior powers have turned out to be helpful in solving mathematical problems. Indeed, the
exterior algebra, a key element in differential geometry since the 1930’s, was largely ignored
by the world of established mathematicians for nearly half a century after its invention by
Prussian school teacher Hermann Grassmann (1809–1877).2 While teaching in a high school
at the Baltic port city of Stettin (Szczecin), Grassmann published two editions of his principal
work “Die Ausdehnungslehre” (1844, and completely revised, 1862). He introduced a calculus
whereby subspaces can be treated as vectors, having length corresponding to the volume of
a basis selected in the subspace but apart from that, independent of the choice of bases;
their linear combinations can be formed, which may or (more often) may not correspond to
subspaces . . . . A world of obscure thoughts (and his style did not help a lot in making them
seem clear), which could not compete with the clarity of daily discoveries in contemporary
matrix theory, where the objects were just tables of numbers and the operations a formal
game. (We should admit that this is the point of view we largely adopt, too.) It was only
through the work of H. Poincaré and E. Cartan near the end of the last century that the
significance of Grassmann’s deep geometric vision became understood and began to occupy
the central role it plays in mathematics today.
Or at least in “continuous mathematics.” In spite of its successful career in the last half
century, the fact that exterior powers were helpful in extremal set theory, a mathematical
area as purely “discrete” as any could be, surprised even the informed when László Lovász
came out with this new insight at the 6th British Combinatorial Conference in 1977. Lovász
(1977) introduced two tensor product methods: one using symmetric products, the other
using alternating products (wedge products). The former can be translated into polynomial
space arguments, as we have already done in Chapter 5. In this chapter we shall present the
wedge product method which so far has resisted attempts at more elementary presentation.
We should mention that G. Kalai (1984), unaware of Lovász’s work, independently discovered
the same method and found a multitude of applications (see also Kalai (1986), Alon–Kalai

c László Babai and Péter Frankl. 1988, 1992, 2020.
2
For a brief and passionate summary of the life and mathematical work of this extraordinary man, see J.
Dieudonné, The tragedy of Grassmann, Linear and Multilinear Algebra 8 (1979), 1–14.

135
CHAPTER 6. TENSOR PRODUCT METHODS
(1985)).
The first section of this chapter provides a hands-on introduction to wedge products over
finite dimensional spaces, avoiding the trip to infinite dimension common to synthetic treat-
ments of the subject. We make no attempt to please the soul of Grassmann as represented by
Dieudonné (1979) with a coordinate-free route; quite on the contrary, we try to build on the
only prerequisite we have declared in the preface: the reader’s knowledge of determinants.
The main text contains material sufficient for all combinatorial applications known to
the authors. The construction of the exterior algebra is completed in the Exercises.

6.1 Wedge products — a concrete introduction

6.1.1 The Laplace expansion of determinants
Let [n] = {1, . . . , n}. For an n × n matrix A = (αij )i,j∈[n] over the field F and subsets
I, J ⊆ [n], let I AJ denote the |I| × |J| minor

I AJ = (αij )i∈I, j∈J . (1)

For I ⊆ [n], let I¯ = [n] \ I and ΣI = i∈I i. With this notation, Laplace’s identity asserts
P
that for any I ⊆ [n],
X
det(A) = (−1)ΣI+ΣJ det(I AJ ) det(I¯AJ¯). (2)
J⊆[n]
|J|=|I|

n


(There are k
terms on the right hand side where k = |I|.)
The proof of this formula is not difficult. Viewing the αij as indeterminates, both sides
are homogeneous polynomials of degree n with the exact same n! expansion terms; only the
correctness of the signs needs to be verified. We leave the details to the reader.

6.1.2 Alternating k-linear functions

Let W1 , . . . , Wk , and T be linear spaces over the field F. A function f : W1 × · · · × Wk → T
is k-linear if it is linear in each of the k variables, i. e.,

f (w1 , . . . , wi−1 , λui + µvi , wi+1 , . . . , wk ) =

λf (w1 , . . . , wi−1 , ui , wi+1 , . . . , wk )
+ µf (w1 , . . . , wi−1 , vi , wi+1 , . . . , wk ) (3)

for every i ∈ [k], λ, µ ∈ F and wj ∈ Wj , ui , vi ∈ Wi . Henceforth we shall assume

W1 = · · · = Wk = W and use the notation W k = W × · · · × W (k times). We say that a
k-linear function f : W k → T alternates if

wi = wj (i 6= j) implies f (w1 , . . . , wk ) = 0. (4)

136 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 6.1. WEDGE PRODUCTS
The most important example of alternating k-linear functions is the determinant, as a func-
tion of its rows:  
w1
 .. 
f (w1 , . . . , wn ) = det  .  (wi ∈ Fn ). (5)
wn
(In this case, W = Fn , T = F, and k = n.) For typographical reasons, we shall write

det(w1 , . . . , wn ) (6)

for the right hand side of (5).

We shall see that several of the familiar properties of the determinant are shared by
every alternating k-linear function. Throughout this subsection, f : W k → T denotes an
alternating k-linear function over F.
First we observe (Exercise 6.1.1), that f changes sign if we interchange any two of the
variables:
f (. . . , wi , . . . , wj , . . .) = −f (. . . , wj , . . . , wi , . . .). (7)
This is the property this class of functions owes its name to.
Next we note the effect of replacing a variable by a linear combination of the variables.
The following is immediate from (3) and (4).
Proposition 6.1. For λ1 , . . . , λk ∈ F, w1 , . . . , wk ∈ W , we have
k
X
f (w1 , . . . , wi−1 , λj wj , wi+1 , . . . , wk )
j=1
= λi f (w1 , . . . , wk ). (8)

Two more familiar properties are special cases of this.

Proposition 6.2. (a) If any of the wi is zero then f (w1 , . . . , wk ) = 0.
(b) The value of f will not change if we add to one of the variables a linear combination
of the others.
A generalization of properties (4) and 6.2(a) is now a corollary.
Corollary 6.3. If w1 , . . . , wk ∈ W are linearly dependent then f (w1 , . . . , wk ) =P
0.
For the proof, choose i such that λi 6= 0 in some nontrivial linear relation kj=1 λj wj .
Apply Propositions 6.1 and 6.2(a).

A fundamental property of alternating k-linear functions is, that, up to a scalar factor,

their value depends on the span of their arguments only.
Proposition 6.4. Let u1 , . . . , uk , v1 , . . . , vk ∈ W . If u1 , . . . , uk ∈ span{v1 , . . . , vk }, then

f (u1 , . . . , uk ) = λf (v1 , . . . , vk )

for some λ ∈ F.

——————————————————————— 137

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 6. TENSOR PRODUCT METHODS
Proof. We can express each of the ui as a linear combination of the vj . Expanding by
linearity in each component, we find that f (u1 , . . . , uk ) is a linear combination of vectors of
the form f (vi1 , . . . , vik ) (i1 , . . . , ik ∈ [k]). Of these k k vectors, those whose subscripts are not
all different, vanish by (4). The value of each of the remaining k! vectors is ±f (v1 , . . . , vk ),
the sign depending on the parity of the permutation (i1 , . . . , ik ).
Corollary 6.5. Let u1 , . . . , uk , v1 , . . . , vk ∈ W . If span{u1 , . . . , uk } = span{v1 , . . . , vk }, then

f (u1 , . . . , uk ) = λf (v1 , . . . , vk )

for some λ ∈ F, λ 6= 0.
Proof. By Proposition 6.4, we are done unless λ = 0. In any case, it follows from Proposi-
tion 6.4 that f (v1 , . . . , vk ) = 0 implies f (u1 , . . . , uk ) = 0. The converse follows by symmetry.
Hence, if λ = 0, then f (v1 , . . . , vk ) = f (u1 , . . . , uk ) = 0 and λ can be replaced by λ = 1.

The following is a slight generalization of Proposition 6.4.

Proposition 6.6. Let u1 , . . . , uk , v1 , . . . , vm ∈ W . If u1 , . . . , uk ∈ span{v1 , . . . , vm }, then
f (u1 , . . . , uk ) ∈
span{f (vi1 , . . . , vik ) : 1 ≤ i1 < . . . < ik ≤ m}. (9)

The proof goes along the lines of the proof of Proposition 6.4 and is left to the reader.
Corollary 6.7. If dim W = n then
 
n
dim(span{f (w1 , . . . , wk ) : wi ∈ W }) ≤ . (10)
k

a basis {v1 , . . . , vn } of W . The number of generators on the right hand side of

Proof. Fix
n
(9) is k . (We note
that for k > n, (10) is vacuously satisfied: f is identically zero by
Corollary 6.3, and nk = 0 by definition.)

A further easy but important corollary is this.

Corollary 6.8. Let v1 , . . . , vn be a basis of W . If two alternating k-linear functions g, h :
W k → T agree on the set {(vi1 , . . . , vik ) : 1 ≤ i1 < · · · < ik ≤ n}, then they are identical.
Proof. Let f = g − h. Then the right hand side of (9) (with m = n) consists of the zero
vector alone. Therefore, by Proposition 6.6,f is identically zero.

6.1.3 Exterior powers of Fn

It is most significant, that for all nonnegative values of n and k, equality can be achieved in
(10). The construction of such a “maximal” alternating k-linear function for every k is the
subject of this subsection. By Corollary 6.3, we may assume k ≤ n.

138 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 6.1. WEDGE PRODUCTS
n
n ( )
Let W = F , and, for 0 ≤ k ≤ n, let Tk = F k . Recall that a bilinear function is a
k-linear function with k=2.
Theorem 6.9. For k = 0, . . . , n there exist alternating k-linear functions fk : W k → Tk and
bilinear functions βk : Tk × Tn−k → F such that for any w1 , . . . , wn ∈ W ,

βk (fk (w1 , . . . , wk ), fn−k (wk+1 , . . . , wn ))

= det(w1 , . . . , wn ). (11)

Proof. We shall construct a specific alternating k-linear function fk : W k → Tk called the

wedge product and denoted
k
^
fk (w1 , . . . , wk ) = w1 ∧ . . . ∧ wk = wi . (12)
i=1

Correspondingly, we shall use the notation

Vk
Tk = W

and call this space the k th exterior power of W . We record the fact that by definition,
 
Vk n
dim W = . (13)
k

First of all, we index the entries of the vectors in Tk by k-subsets of [n]. Let [n]


k
denote the
set of all k-subsets of [n]. A typical element of Tk thus has the form

(αJ : J ∈ [n]


k
). (14)

For w1 , . . . , wk ∈ W , let A = A(w1 , . . . , wk ) denote the k × n matrix whose rows are

w1 , . . . , wk . We define the wedge product of the wi as follows.

w1 ∧ . . . ∧ wk = (det(AJ ) : J ∈ [n]


k
). (15)

Here AJ denotes the k × k minor [k] AJ , consisting of those columns of A indexed by the
elements of J.
It is an immediate consequence of the corresponding properties of the k × k determinants
that fk is k-linear and alternating.
Let x = (ξJ : J ∈ [n] [n]



k
) ∈ Tk and y = (ηK : K ∈ n−k ∈ Tn−k ). On these arguments, we
define the bilinear function βk by the following formula.
X
βk (x, y) = (−1)k(k+1)/2 (−1)ΣJ ξJ ηJ¯. (16)
[n]
J∈( k )

In order to verify (11) we just notice that substituting the components defined by equa-
tion (15) into (16) we obtain precisely the the Laplace expansion (2) of the determinant
det(w1 , . . . , wn ), with the set [k] playing the role of I in (2).

——————————————————————— 139

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 6. TENSOR PRODUCT METHODS
We note that by definition, for w1 , . . . , wn ∈ W = Fn ,

w1 ∧ . . . ∧ wn = det(w1 , . . . , wn ). (17)

It is customary to use the same wedge symbol to indicate the bilinear function βk :

x ∧ y := βk (x, y) for x ∈ Tk , y ∈ Tn−k . (18)

With this notation, equation (11) takes the following pleasing form:

(w1 ∧ . . . ∧ wk ) ∧ (wk+1 ∧ . . . ∧ wn ) = w1 ∧ . . . ∧ wn . (19)

Equation (19) explains why notation (18) is unlikely to lead to confusion. An important
generalization of the bilinear functions βk and of equation (19) is given in Exercise 6.1.5.
Corollary 6.10. The vectors w1 , . . . , wk are linearly independent if and only if w1 ∧. . .∧wk 6=
0.
Proof. The “if” part follows from Corollary 6.3. For the “only if” part, assume w1 , . . . , wk
are linearly independent and extend them to a basis w1 , . . . , wn . Now

0 6= det(w1 , . . . , wn ) = (w1 ∧ . . . ∧ wk ) ∧ (wk+1 ∧ . . . ∧ wn ).

Therefore the first term on the right hand side cannot be zero.

We close this section by drawing the reader’s attention to an important consequence of

Corollary 6.5.
For a k-dimensional subspace T ≤ W let us define ∧T ∈ k W by selecting a basis
V
t1 , . . . , tk of T and setting
∧ T := t1 ∧ . . . ∧ tk . (20)
Although the right hand side depends on the arbitrary choice of the basis t1 , . . . , tk , according
to Corollary 6.5 it can only vary by a nonzero scalar factor. The k th exterior power of W thus
allows us to view k-dimensional subspaces of W as vectors, and perform linear operations
on them. In particular, the question of linear independence of a family of k-dimensional
subspaces is now well defined, since the ambiguity in the scalar factors has no effect on
linear independence. In the next section, variants of Bollobás’s Theorem (Section 5.1) will
be proved by showing that subspaces associated in a certain way with finite sets satisfying
a Bollobás–type condition are linearly independent in the sense just defined.
The next lemma will help us link wedge products to intersection conditions.
Lemma 6.11. Let U and V be subspaces of W , and assume dim U +dim V = n (n = dim W ).
Then (∧U ) ∧ (∧V ) = 0 if and only if U ∩ V 6= 0.
(The last 0 denotes the subspace consisting of the zero vector alone.)
Proof. Let us combine the bases of U and V . It is clear that the system of n vectors obtained
will be linearly independent (and therefore a basis of W ) if and only if U ∩ V = 0. Hence
the Lemma follows by equations (19) and (17).

140 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 6.1. WEDGE PRODUCTS
The condition in this Lemma that U and V be of complementary dimension is unnecessary
except that otherwise the product (∧U ) ∧ (∧V ) is as yet undefined. This gap will be filled
by the bilinear functions βrs , to be defined in Exercise 6.1.5. The subsequent exercises give
the corresponding generalization of Lemma 6.11 and show how all the exterior powers of Fn
combine to an associative algebra of dimension 2n , the Grassmann algebra over Fn .

Exercises
♦Ex. 6.1.1. Show that for k-linear functions, equation (7) follows from condition (4).
♦Ex. 6.1.2. Show that for k-linear functions, (4) follows from (7) unless char F = 2.
♦Ex. 6.1.3. Show that for k-linear functions, (4) does not follow from (7) if char F = 2.
Pk
Ex. 6.1.4. Let u1 , . . . , uk , v1 , . . . , vk ∈ W and assume ui = j=1 αij . Prove that the value of λ
computed in the proof of Proposition 6.4 is

λ = det(αij )ki,j=1 .

Hint. Read the proof of Proposition 6.4 carefully.

Vr Vs Vr+s
Ex. 6.1.5. For r, s ≥ 0, construct bilinear functions βrs : W× W → W such that for
w1 , . . . , wr+s ∈ W ,

βrs ((w1 ∧ . . . ∧ wr ), (wr+1 ∧ . . . ∧ wr+s )) = w1 ∧ . . . ∧ wr+s . (21)

In the spirit of equation (19) we use the wedge symbol to denote the function βrs , thus equation
(21) will read
(w1 ∧ . . . ∧ wr ) ∧ (wr+1 ∧ . . . ∧ wr+s ) = w1 ∧ . . . ∧ wr+s . (22)
Vr Vs
Hint. First define βrs on a basis of W × {a basis of W } in the unique way required by
equation (21). Then extend βrs by bilinearity. (There is a unique way to do this, cf. Section 2.2.)
Finally show, that the two functions βrs ((w1 ∧ . . . ∧ wr ), (wr+1 ∧ . . . ∧ wr+s )) and w1 ∧ . . . ∧ wr+s
agree. To this end, first fix w1 , . . . , wr such that their wedge product be a member of the basis of
V r
W and show, using Corollary 6.8, that the resulting two alternating s-linear functions (of the
variables wr+1 , . . . , wr+s ) agree. Then fix wr+1 , . . . , wr+s arbitrarily, and show, using Corollary 6.8
again, that the resulting two alternating r-linear functions (of the variables w1 , . . . , wr ) agree.

Ex. 6.1.6. Prove that in Lemma 6.11, the condition that U and V have complementary dimensions
can be omitted.

Ex. 6.1.7. For r1 , . . . , rk ≥ 0, prove that there exists a k-linear function

Vr1 V rk Vr1 +···+rk
γ: W × ··· × W → W

with the following property. Let w1 , . . . , wr1 +···+rk ∈ W , z1 = w1 ∧ . . . ∧ wr1 , z2 = wr1 +1 ∧ . . . ∧

wr1 +r2 , . . . , zk = wr1 +···+rk−1 +1 ∧ . . . ∧ wr1 +···+rk . For any w1 , . . . , wr1 +···+rk ∈ W , γ should satisfy
the equation
γ(z1 , . . . , zk ) = w1 ∧ . . . ∧ wr1 +···+rk . (23)

——————————————————————— 141

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 6. TENSOR PRODUCT METHODS
This result justifies an extension of the notation (18): we write

z1 ∧ . . . ∧ zk := γ(z1 , . . . , zk ), (24)

and obtain, using the notation introduced above, the following generalization of equation (19):

z1 ∧ . . . ∧ zk = w1 ∧ . . . ∧ wr1 +···+rk . (25)

Hint. The result immediately follows from Exercise 6.1.5, by induction on k. For k = 2 it is actually
identical with that exercise, so assume k ≥ 3. Let γk denote the function we seek. Set

γk (z1 , . . . , zk ) = γk−1 (z1 , . . . , zk−1 ) ∧ zk ,

where the ∧ on the right hand side stands for the βrs function (r = r1 + · · · + rk−1 ; s = rk ), in the
sense of equation (22).
Remark. This exercise completes the construction of the exterior algebra over W = Fn . This algebra
consists of the 2n -dimensional linear space defined as the direct sum of the exterior powers k W
V
(k ≥ 0), together with the bilinear “product” operation ∧, now defined over the entire exterior
algebra. Equation (25) guarantees associativity and consistency of the wedge notation with no
parentheses.

Ex. 6.1.8 (Universality of wedge products). Let f : W k → TV be an arbitrary alternating k-

linear function. Prove: there exists a unique linear map ϕ : k W → T such that for every
w1 , . . . , w k ∈ W ,
f (w1 , . . . , wk ) = ϕ(w1 ∧ . . . ∧ wk ). (26)

6.2 Bollobás–type theorems

First we present Lovász’s strikingly elegant wedge product proof for the uniform version of
Bollobás’s Theorem (Theorem 5.4). As it takes no extra effort, we immediately prove the
skew version (Theorem 5.6). Let us recall the result.
Theorem 6.12 (Bollobás’s Theorem — skew version). Let A1 , . . . , Am be r-element
sets and B1 , . . . , Bm be s-element sets such that

(a) Ai and Bi are disjoint for i = 1, . . . , m;

(b) Ai and Bj intersect whenever i < j (1 ≤ i, j≤ m).

Then m ≤ r+s


r
.
Before proceeding to the proof, let us formulate, following Z. Füredi, a generalization of
this problem. It is natural to ask what happens if the alternative is not whether or not the
sets intersect trivially, but whether or not their intersections exceed a given threshold.
Theorem 6.13 (Bollobás’s Theorem – threshold version (Z. Füredi, 1984). Let
A1 , . . . , Am be r-element sets and B1 , . . . , Bm be s-element sets such that for some t ≥ 0,

142 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 6.2. BOLLOBÁS–TYPE THEOREMS
(a) |Ai ∩ Bi | ≤ t for i = 1, . . . , m;

(b) |Ai ∩ Bj | ≥ t + 1 for 1 ≤ i < j ≤ m.

Then  
r + s − 2t
m≤ .
r−t
Like Bollobás’s Theorem, this result is tight, too. To see this, take two disjoint sets C
and D of respective sizes t and r + s − 2t. Let Ai be the union of C and any (r − t)-subset
of D; and Bi the union of C and the complementary subset of D.
Füredi’s proof is based on Lovász’s. Let us now see how that proof works.
Proof of Theorem 6.12 (L. Lovász (1977)). Let X be a finite set containing all the Ai and
Bj . Let W = Rr+s and associate vectors wi ∈ W with each i ∈ X such that the family
of vectors {wi : i ∈ X} is in general position, i. e., every r + s of these vectors is linearly
independent. (Take, for instance, points of the moment curve. See Section 2.2.)
With every subset I ⊆ X we associate the wedge product
^
wi ∈ |I| W
V
wI = (27)
i∈I

where the terms in the product are taken in any order. (The order can only make a difference
in sign.)
Let now A, B ⊂ X, |A| = r, |B| = s. We claim that
(
6= 0 if A ∩ B = ∅ ;
wA ∧ wB (28)
= 0 if A ∩ B 6= ∅ .

Indeed, by equation (19), in the first case we have the wedge product of r + s linearly
independent vectors; in the second, there is repetition among the r + s terms of the wedge
product considered, hence the terms are linearly dependent and thus their product is zero.
A combination of equation (27) with the conditions of the Theorem shows that
(
6= 0 if i = j ;
wAi ∧ wBj (29)
= 0 if i < j .

By the 3rd version of the Triangular Criterion (Proposition

Vr r+s 2.9),r+s
this

implies that wA1 , . . . , wAm
are linearly independent. Consequently m ≤ dim R = r .

This proof admits a generalization to subspaces of a linear space that does not seem amenable
to the methods of Chapter 5.
Theorem 6.14 (Bollobás’s Theorem for subspaces) (L. Lovász, 1977). Let
U1 , . . . , Um be r-dimensional and V1 , . . . , Vm be s-dimensional subspaces of a linear space
W over the field F. Assume that

——————————————————————— 143

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 6. TENSOR PRODUCT METHODS
(a) Ui ∩ Vi = 0 for i = 1, . . . , m;

(b) Ui ∩ Vj 6= 0 whenever i < j (1 ≤ i, j ≤ m).

Then  
r+s
m≤ .
r
(Here 0 denotes the subspace consisting of the zero vector alone.)
Proof. We may assume that the field F is infinite. (Why? See Exercise 6.2.3.) We may
clearly assume that W has finite dimension; let n = dim W . It follows from condition (a)
that n ≥ r + s.
First we prove the result under the assumption n = r + s. With the notation of equation
(20), let
r
^
ui = ∧Ui ∈ W and (30)
s
^
vi = ∧Vi ∈ W (i = 1, . . . , m).

Combining Lemma 6.11 with the conditions of the Theorem, we obtain that
(
6= 0 if i = j ;
ui ∧ vj (31)
=0 if i < j .

Again by the 3rd version of the Triangular Criterion (Proposition

Vr r+s 2.9), this

implies that
r+s
u1 , . . . , um are linearly independent. Consequently m ≤ dim R = r .
We are now left with the task of reducing the general case to the case r + s = n. This is
done by a “general position” argument. What we need this time is a linear map ϕ : W → W0
such that

(i) dim W0 = r + s;

(ii) ker ϕ ∩ span{Ui , Vj } = 0 for every i, j (1 ≤ i ≤ j≤ m).

Indeed, for i ≤ j, the restriction of such a map ϕ to the subspace span{Ui , Vj } is an iso-
morphism. It follows that dim ϕ(Ui ) = r, dim ϕ(Vj ) = s, and conditions (a) and (b) of the
Theorem hold for the subspaces ϕ(Ui ) and ϕ(Vj ) of W0 . Now, in view of condition (i), the

situation is the one considered in the first part of this proof, hence the inequality m ≤ r+s
r
follows.
Since dim span{Ui , Vj } ≤ r + s for every i, j, a map ϕ : W → W0 in general position with
respect to the finite set of subspaces span{Ui , Vj } will satisfy (ii). Such a map ϕ exists by
the results of Section 3.1.3 (Theorem 3.13), because the field F is infinite.

144 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 6.2. BOLLOBÁS–TYPE THEOREMS
In order to prove his combinatorial result (Theorem 6.13), Füredi had to make a detour
and generalize Lovász’s subspace result first. He obtained the following threshold version of
Theorem 6.14.
Theorem 6.15 (Bollobás’s Theorem for subspaces — threshold version) (Z. Füredi,
1984). Let U1 , . . . , Um be r-dimensional and V1 , . . . , Vm be s-dimensional subspaces of a
linear space W over the field F. Assume that for some t ≥ 0,
(a) dim(Ui ∩ Vi ) ≤ t for i = 1, . . . , m;
(b) dim(Ui ∩ Vj ) ≥ t + 1 for (1 ≤ i < j ≤ m).
Then  
r + s − 2t
m≤ .
r−t
Proof. We again use a “general position” argument, this time in order to reduce Theorem 6.15
to Theorem 6.14. As usual with such arguments, they only work over sufficiently large fields.
So let us assume F is infinite, leaving the justification of this move, as before, to Exercise 6.2.2.
Let dim W = n. Let W0 be a subspace of codimension t (i. e., dimension n − t). We
observe that
(α) dim(Ui ∩ Vj ∩ W0 ) ≥ 1 for 1 ≤ i < j ≤ m.
On the other hand, if we choose W0 to be in general position with respect to the subspaces
Ui , Vi , and Ui ∩ Vi (i = 1, . . . , m) (Theorem 3.5), then intersecting these subspaces with W0
will reduce their dimensions by precisely t (or to zero, whichever is greater). It follows that
(β) dim(Ui ∩ Vi ∩ W0 ) = 0 for i = 1, . . . , m; and
(γ) dim(Ui ∩ W0 ) = r − t, dim(Vi ∩ W0 ) = s − t.
Observe now that (γ), (β), and (α) guarantee that for the subspaces Ui ∩ W0 and Vi ∩ W0
of W0 , the conditions of the previous theorem
hold, with r − t and s − t in the roles of r and
s. The conclusion therefore is m ≤ (r−t)+(s−t)
r−t
, as stated.
Füredi’s combinatorial result, Theorem 6.13, is now immediate.
Proof of Theorem 6.13. Let X be a finite set containing all the Ai and Bj . Let |X| = n, and
associate a member ex of a fixed basis of W = Rn with each x ∈ X. With each subset S ⊆ X
associate the subspace W (S) = span{ex : x ∈ S}. Now the subspaces W (Ai ) and W (Bi )
satisfy the conditions of Theorem 6.15 and the corresponding bound on m follows.
We close this section by calling the reader’s attention to a comparison of three Bollobás–
type results: the skew version of Bollobás’s original result (sets vs. sets, Theorem 5.6),
and Lovász’s two variants: sets vs. subspaces (Theorem 5.7), and subspaces vs. subspaces
(Theorem 6.14). We have used the polynomial space method to prove the “sets vs. subspaces”
result, and the wedge product method to prove the “subspaces vs. subspaces” result. Apart
from insignificant differences in presentation, these are the only known proofs in each case.
The basic “sets vs. sets” result follows from each of the two variants; this is why it has two
proofs.

——————————————————————— 145

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 6. TENSOR PRODUCT METHODS
Exercises
♦Ex. 6.2.1 (Affine and projective dimensions of graphs) (Pudlák–Rödl, 1992)). An affine represen-
tation of a graph in an affine space W assigns an affine subspace Ui to each vertex ui of the graph
such that Ui and Uj are disjoint if and only if ui and uj are not adjacent. The smallest dimension
of an affine space over a given field F where such an assignment is possible is the affine dimension
over F of G, denoted by adim(G).
An projective representation of a graph in a linear space W assigns a linear subspace Ui to each
vertex ui of the graph such that Ui ∩ Uj = {0} if and only if ui and uj are not adjacent. The
smallest dimension of a linear space over a given field F where such an assignment is possible is the
projective dimension over F of G, denoted by pdim(G).
Find a class of graphs with bounded affine dimensions but unbounded projective dimensions
over R.
Hint. Consider the complement of a perfect matching: 2n vertices u1 , . . . , u2 n with edges {u2`−1 , u2` }
(` = 1, . . . , n).

Ex. 6.2.2 (Threshold version of the “sets vs. subspaces” problem). Let U1 , . . . , Um be r-
dimensional subspaces; B1 , . . . , Bm s-subsets of W . Assume
(
≤t if i = j ;
|Ui ∩ Bj | (32)
≥ t + 1 if i 6= j

(Füredi, 1984.) (a) Prove:  

r+s−t
m≤ . (33)
r
(b) Show that the inequality one might expect in analogy to Theorem 6.13 is false:
 
r + s − 2t
m≤ (34)
r−t

does not follow from the conditions. Give infinitely many counterexamples with r = 2 and demon-
strate that inequality (34) fails badly (by the orders of magnitude). (c) Show that inequality (33)
is not tight.
Hint. (a) Let Ci = Bi \ Ui . Then |Ci | ≥ s − t and the conditions of Theorem 5.7 hold with s − t
in the role of s. (b) Rather than constructing planes in R3 through the origin, draw lines in the
plane x3 = 1. In this plane, construct a set S of k points (k large) and a set L of Ω(k 2 ) lines such
that each line pass through exactly three points of S. (Draw a 3 × (k/3) grid.) Let Ui be the span
(in R3 ) of the ith line from L (Ui is a plane), and let Bi = S \ Ui . This way s := |Bi | = |S| − 3, and
the conditions stated
in the exercise are fulfilled with r = 2, t = 1. Now m = Ω(s2 ), as
opposed
r+s−2t s+1
to the bound r−t = s, given by (34). (c) For r = 2, t = 1, improve the bound 2 , given by
(33), to 1 + 2s .

Ex. 6.2.3. Justify the assumption made in the proofs of Theorems 6.14 and 6.15 that the field F
is infinite.
Hint. Use the Field Extension Lemma of Section 2.2 2.2 (Lemma 2.29).
ADD: Lovász’s “k-tree” inequality.

146 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 6.3. SYMMETRIC PRODUCTS
6.3 Symmetric products
TO BE WRITTEN

6.4 The Shannon capacity of a graph

TO BE WRITTEN
Lovász’s method indicated
Haemers’s method in detail: succeeds for Petersen, Kneser K(n, 2)
R. M. Wilson’s proof for all Kneser graphs mentioned; sharp Erdős–Ko–Rado deduced

******** tentative partial text ********

Let G = (V, E) be a graph with V = [n].
Definition 6.16. We say that n × n matrix A = (aij ) fits G if aii 6= 0, and aij = 0 if ij 6∈ E.

Theorem 6.17. If a matrix A (over any field) fits a graph G, then Θ(G) ≤ rkA.
Proof. Since A⊗k fits Gk , A⊗k has a diagonal matrix, of size α(Gk ), with non-zero diagonal
entries, as a submatrix. Hence rk(A⊗k ) ≥ α(Gk ). On the other hand it is known that
rk(A⊗k ) = (rkA)k . Thus we have
p p
Θ(G) = sup k α(Gk ) ≤ sup k rk(A⊗k ) = rkA.
k k

For a graph G, Haemers introduced the following number.

Definition 6.18.
R(G) := min{rkA : A fits G}.

Theorem 6.17 gives us Θ(G) ≤ R(G).

Definition 6.19. Let
p be a prime not dividing k. We define a graph G(n, k, p) = (V, E)
as follows. Let V = [n]
k
, and let two vertices x and y be adjacent iff |x ∩ y| 6≡ 0 (mod p).

Observation 6.20. Let G := G(n, k, p). Then R(G) ≤ n.

Proof. Let M be the n × nk incidence matrix of G. Let A = M T M . For x, y ∈ V , the

(x, y)-entry of A is |x ∩ y|. Thus, in the field GF (p), the matrix A fits G, and rkA ≤ n.
Observation 6.21. Let p be a prime and m a positive integer. Set n := (p + 2)m, k := p + 1
and G := G(n, k, p). Then α(G) = R(G) = n.
Proof. Partition [n] into m classes, X1 , . . . , Xm , of size p + 2 each. Define S := {x ∈ V :
x ⊂ Xi for some i}. Then |S| = n. For x, y ∈ S, we have
(
p if x, y ∈ Xi for some i
|x ∩ y| =
0 otherwise

——————————————————————— 147

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 6. TENSOR PRODUCT METHODS
Note that in all cases, |x ∩ y| ≡ 0 (mod p). This means S is an independent set.
Thus we have n ≤ α(G) ≤ Θ(G) ≤ R(G) ≤ n.
Remark 6.22. Let G := G(n, 3, 2) and assume n is divisible by 4. The Lovász bound of G
is known:
n(2n2 − 15n + 22)
θ(G) = .
3(3n − 14)
Using Observation 6.21, we have θ(G) > n = R(G) ≥ Θ(G) if n ≥ 9, showing that the
Lovász θ function does not equal the Shannon capacity; in fact (Haemers)

θ(G)
sup = ∞.
G Θ(G)

148 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
Chapter 7

A class of higher incidence matrices:

the inclusion matrix

7.1 The inclusion matrix; s-independent families

A point either belongs to a set or not; this simple relation is recorded by the entries of the
incidence matrix. It may come as no surprise to the reader that we shall also be interested
in how the pairs, triples, etc., of points relate to the members of a set system. It was quite
a surprise, though, at the time when Dijen K. Ray-Chaudhuri and Richard M. Wilson made
the conceptual leap and demonstrated its power.
A variety of second, third and higher order incidence matrices arise, expressing relations
such as containment, disjointness, size or parity of intersection. Some families of such
matrices will be discussed in Section ** In the present chapter we focus on the inclusion
matrix, the most important class of higher incidence matrices.
Let F and T be two families of sets over the universe X of n points. We define the
(F, T )-inclusion matrix I(F, T ) to be an |F| × |T | (0, 1)-matrix whose rows and columns
are labeled by the members of F and T , resp.
The entry µI (E, T ) in position (E, T ) (E ∈ F, T ∈ T ) will be 1 or 0 according to whether
or not T ⊆ E.
(
1 if T ⊆ E
µI (E, T ) = (1)
0 if T 6⊆ E .

Of particular interest will be the case when T is the complete s-uniform family Xs . For

0 ≤ s ≤ n, we set
I(F, s) = I F, Xs .



(2)
We call the |F| × ns matrix I(F, s) the s-inclusion matrix of F, or the inclusion matrix of

order s of F. The rows of I(F, s) are labeled by the members of F and the columns by the
s-subsets T of X. (Just as in the case of the incidence matrix, the order in which we list

c László Babai and Péter Frankl. 1988, 1992, 2020.

149
CHAPTER 7. A CLASS OF HIGHER INCIDENCE MATRICES
the members of F and of Xs is immaterial as long as we always use the same order.) As

always, we set m = |F|.

The 0-inclusion matrix I(F, 0) is simply a column of 1’s of height m. The ordinary
incidence matrix of F is the 1-inclusion matrix I(F, 1).
We call F s-independent if the rows of I(F, s) are linearly independent, i. e., the s-
inclusion matrix has full row-rank. The significance of this circumstance is clear:
Proposition 7.1. If the family F is s-independent then
 
n
|F| ≤ . (3)
s
Proof. Indeed, ns is the number of columns of the s-inclusion matrix.

The way we have used the incidence matrix to exploit intersection conditions suggests that
here, again, the first thing to compute may be the product of an inclusion matrix and its
transpose. The result is appealing and signals the power of our new-found tool.
Proposition 7.2. The general entry µAs (E, F ) of the m×m matrix As (F) = I(F, s)I(F, s)T
is  
|E ∩ F |
µAs (E, F ) = (E, F ∈ F) (4)
s
We call As (F) the s-intersection matrix or the intersection matrix of order s of the family
F.
Proof. By definition, µAs (E, F ) counts the number of those T ∈ Xs satisfying T ⊆ E and

T ⊆ F , i. e., T ⊆ E ∩ F .
The inclusion matrices of complete uniform families will play a particularly important role.
For i ≥ j, let In (i, j) denote the j-inclusion matrix of the complete i-uniform family Xi .

In (i, j) = I Xi , j = I Xi , Xj .






(5)
The matrices In (i, j) are the coefficients in the following inclusion-exclusion type lemma.
This lemma establishes a useful duality between F and F c , the system of complements of
edges of F.
Lemma 7.3. For every family F,
s
X
(−1)j I(F, j)In (s, j)T = I(F c , s). (6)
j=0

X


Proof. Let ν(E, T ) denote the (E, T )-entry of the left hand side (E ∈ F, T ∈ s
). By
Proposition 7.2 we have
s  
j |E ∩ T |
X
ν(E, T ) = (−1) . (7)
j=0
j

The right hand side of equation (7) is 1 if E ∩ T = ∅ and (1 − 1)|E∩T | = 0 otherwise. These
two cases correspond to whether or not T ⊆ X \ E.

150 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 7.2. EXTENDED INCLUSION MATRICES
Exercises
Ex. 7.1.1. The full version of the Ray-Chaudhuri–Wilson Theorem (Theorem 7.13) asserts that

any k-uniform L-intersecting family is s-independent, where s = |L|. (The inequality m ≤ ns
is then immediate by Proposition 7.1.) Prove this result for k ≡ −1 (mod p), s = p − 1, and
L = {0, 1, . . . , p − 2}, where p is a prime number.
Hint. Solve the modular version, stated in the next exercise.

Ex. 7.1.2. Frankl–Wilson’s modular version of the RW Theorem (Theorem 7.15) asserts that
any k-uniform family which is L-intersecting mod p is s-independent,
n

assuming k ∈ / L (mod p),
where s = |L| and p is a prime number. (The inequality |F| ≤ s is then again immediate by
Proposition 7.1.)Prove this result for s = p − 1, assuming k ≡ −1 (mod p).
r


Hint. Prove that p−1 is divisible by p if and only if r 6≡ −1 (mod p). Use this to show that the
(p − 1)-intersection matrix of F is nonsingular (diagonal) over Fp and therefore nonsingular over Q.
Conclude that the (p − 1)-inclusion matrix of F has full row rank.

7.2 Extended inclusion matrices. The Nonuniform

RW Theorem revisited
In this section, we shall work over the field Q of rational numbers. (Any other field of
characteristic zero would also be suitable.)
Let us recall that the binomial coefficient xs is defined as a polynomial of degree s over

any field of characteristic zero by the formula

 
x 1
= x(x − 1) . . . (x − s + 1) (8)
s s!

provided s is a nonnegative integer. Now let f (x) be an
arbitrary
polynomial of degree

x x x


s ≥ 0. Being linearly independent, the polynomials 0 , 1 , . . . , s form a basis of the
space of polynomials of degree ≤ s and therefore f (x) is uniquely expressible as their linear
combination:      
x x x
f (x) = α0 + α1 + · · · + αs (αi ∈ Q). (9)
0 1 s
Equations (4) and (9) lead to a remarkable consequence.
Let us combine the inclusion matrices of orders s, s − 1, . . . , 0 to form the s
∗ -inclusion
matrix, also called the extended inclusion matrix of order s. This is an m × ( ns + s−1 n


+
n


· · · + 0 ) matrix:

I ∗ (F, s) = I(F, s) | I(F, s − 1) | · · · | I(F, 0) .

 
(10)
X

Ss
= t=0 Xt , we can restate this definition as


Using the notation ≤s

I ∗ (F, s) = I(F, X


≤s
). (11)

——————————————————————— 151

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 7. A CLASS OF HIGHER INCIDENCE MATRICES
We call the family F s∗ -independent if the rows of I ∗ (F, s) are linearly independent. An s-
independent family is clearly s∗ -independent, but the converse does not hold (Exercise 7.2.3).
The key observation of the next section will be that the converse does hold for uniform
families (Corollary 7.12).
Let us for the time being stay with not necessarily uniform families. In analogy with (3),
we note the trivial upper bound on the number of members in an s∗ -independent family.
Proposition 7.4. If F is s∗ -independent then
     
n n n
|F| ≤ + + ··· + . (12)
s s−1 0

Proof. The right hand side is the number of columns of the s∗ -inclusion matrix I ∗ (F, s).

Let f (x) be a polynomial of degree s. Consider the m × m matrix A(F, f ) whose general
entry µA (E, F ) is
µA (E, F ) = f (|E ∩ F |) (E, F ∈ F). (13)
We call A(F, f ) the f -intersection matrix of F. With this notation, the s-intersection matrix
x


becomes As (F) = A(F, s ) (Proposition 7.2).
Proposition 7.5. The f -intersection matrix of F can be written as
s
X
A(F, f ) = αi I(F, i)I(F, i)T (14)
i=1

where the coefficients αi are defined by equation (9).

Proof. Immediate from (4).
Lemma 7.6. If f (x) is a polynomial of degree s over a field of characteristic zero then the
column space of I ∗ (F, s) contains the column space of A(F, f ). Consequently,

rk A(F, f ) ≤ rk I ∗ (F, s). (15)

In particular, if A(F, f
) is non-singular then
F is s∗ -independent.

Proof. Let f (x) = α0 x0 + α1 x1 + · · · + αs xs as in (9). Let Aj := Aj (F) = I(F, j)I(F, j)T .

By Proposition 7.5 we have

A(F, f ) = α0 A0 + α1 A1 + · · · + αs As . (16)

For 0 ≤ j ≤ s, the column space of I ∗ (F, s) contains that of I(F, j) by definition; and the
latter contains the column space of Aj . Thus the column space of I ∗ (F, s) contains that of
any linear combination of A0 , A1 , . . . , As . The rank-inequality (15) is now immediate. In
particular, if rk A(F, f ) = |F| then I ∗ (F, s), too, must have full row-rank.

A slight generalization of Lemma 7.6 follows from the lemma itself.

152 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 7.2. EXTENDED INCLUSION MATRICES
Let F = {E1 , . . . , Em } and let f1 (x), . . . , fm (x) be polynomials. Let us consider the
m × m matrix A(F; f1 , . . . , fm ) whose entry in position (i, j) is
µA (i, j) = fj (|Ei ∩ Ej |). (17)
We call this the (f1 , . . . , fm )-intersection matrix of F. The conclusions are the same as in
Lemma 7.6.
Lemma 7.7. If f1 (x), . . . , fm (x) are polynomials of degree ≤ s over a field of characteristic
zero then
rk A(F; f1 , . . . , fm ) ≤ rk I ∗ (F, s). (18)
In particular, if A(F; f1 , . . . , fm ) is non-singular then F is s∗ -independent.
Proof. Again, we wish to prove that the column space of A(F; f1 , . . . , fm ) is contained in
the column space of I ∗ (F, s). This, however, follows immediately from Lemma 7.6, since the
j th column of A(F; f1 , . . . , fm ) coincides with the j th column of A(F, fj ).

The nonuniform version of the RW Theorem was discussed in Section 5.10 (Theorem 5.34),
with an omission we did not warn the reader about. We are now ready to state and prove
the full result.
Let L be a set of nonnegative integers. Recall that F is said to be L-intersecting if
|E ∩ F | ∈ L for each pair of distinct members E, F ∈ F.
Theorem 7.8 (Nonuniform RW Theorem, unabridged) (Frankl–Wilson, 1981).
Let L be a set of s integers and F an L-intersecting family. Then F is s∗ -independent.
Consequently,      
n n n
|F| ≤ + + ··· + . (19)
s s−1 0
The novelty in the statement of the result (compared to Theorem 5.34) is the fact of s∗ -
independence. This seemingly technical addition will lead to an essentially immediate proof
of the (uniform) RW Theorem in the next section.
Proof. Let L = {l1 , . . . , ls } and F = {E1 , . . . , Em } where |E1 | ≤ |E2 | ≤ · · · ≤ |Em |. Since
we are trying to apply Lemma 7.7, it is natural to consider polynomials fi whose roots are
the lk . (This will create a lot of zeros in the (f1 , . . . , fm )-intersection matrix. We have to be
a little careful not to create too many; specifically, to avoid putting zeros in the diagonal.)
For each j, we consider the polynomial
Y
fj (x) = (x − lk ). (20)
lk <|Ej |

Let us examine the (f1 , . . . , fm )-intersection matrix A of F. Note that for i < j we have
|Ei | ≤ |Ej | and therefore |Ei ∩ Ej | < |Ej |. Consequently, the entry of A in position (i, j)
is fj (|Ei ∩ Ej |) = 0 for i < j. On the other hand, the j th diagonal entry is not zero:
fj (|Ej |) 6= 0 because of the careful choice of the roots of fj . Summarizing, the matrix A is
a lower triangular matrix with no zeros in the diagonal. Consequently A is nonsingular and
therefore, by Lemma 7.7, F is s∗ -independent and in particular satisfies inequality (12).

——————————————————————— 153

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 7. A CLASS OF HIGHER INCIDENCE MATRICES
Exercises
Ex. 7.2.1. Prove that for t ≥ s, s∗ -independence implies t∗ -independence.
Ex. 7.2.2. Does s-independence imply t-independence for t ≥ s?
Hint. It doesn’t. For instance, if t > max{|E| : E ∈ F } then I(F, t) = 0, so F cannot be t-
independent. But t-independence will not follow even if t < min{|E| : E ∈ F}. On the other
hand, for k-uniform families, if s ≤ t ≤ k, then s-independence does imply t-independence (see
Exercise 7.3.2).
Ex. 7.2.3. Prove that for s ≥ 2, s∗ -independence does not imply s-independence, even for set
systems whose members are large compared to s. — Contrast this with the uniform case (Corol-
lary 7.12).
Ex. 7.2.4. Assume that |E| > lk for every E ∈ F, lk ∈ L. Prove the Nonuniform RW Theorem
for this case directly from Lemma 7.6, avoiding the use of several polynomials.
Hint. Use the f -intersection matrix for the polynomial
Y
f (x) = (x − lk ). (21)
lk ∈L

Remark. In the next section, we derive the RW Theorem from Theorem 7.8. Actually the special
case considered in this exercise will suffice.

7.3 Inclusion matrices of uniform families

This is a “harvest section.” We are ready to prove important results with no effort. The
results of this section include the complete Ray-Chaudhuri–Wilson Theorem, its mod p
generalization, an attempt at a mod pα version, and Gottlieb’s Theorem stating that the
incidence matrices of complete uniform families have full rank.
Throughout this section we assume that F is a k-uniform family of m subsets of a universe
X of n points. We continue to work over the field Q (or any other field of characteristic
zero).
Recall that In (i, j) denotes the j-inclusion matrix of the complete i-uniform family Xi .

The following simple observation provides the key to strong results for uniform families.
Proposition 7.9. If F is a k-uniform family on n points and 0 ≤ j ≤ i ≤ k then
 
k−j
I(F, i)I(i, j) = I(F, j). (22)
i−j
Proof. The (E, T )-entry of the left hand side (E ∈ F,
|T | = j) is the number of i-sets
containing T and contained in E. This number is k−j
i−j
or 0 depending on whether or not
T ⊆ E.
Corollary 7.10. For 0 ≤ j ≤ i ≤ k ≤ n,
 
k−j
In (k, i)In (i, j) = In (k, j). (23)
i−j

154 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 7.3. INCLUSION MATRICES OF UNIFORM FAMILIES
Proof. Apply Proposition 7.9 to the complete k-uniform family F = Xk .

Corollary 7.11. If F is a k-uniform family on n points and 0 ≤ j ≤ i ≤ k then the column

space of I(F, i) contains the column space of I(F, j). Consequently

rk I ∗ (F, s) = rk I(F, s) (24)

for every s ≤ k.
It follows that for uniform families, s∗ -independence implies s-independence. We shall soon
see some of the far-reaching consequences of this fact.
Corollary 7.12. If, for some s ≤ k, the k-uniform family F is s∗ -independent then it is
s-independent as well and consequently
 
n
|F| ≤ . (25)
s

The machinery is now ready for major conclusions to be drawn in just a few lines. First of
all, the full version of the RW Theorem follows immediately.
Theorem 7.13 (RW Theorem, unabridged) (Ray-Chaudhuri–Wilson, 1975). Let
L be a set of s integers and F an L-intersecting k-uniform family where s ≤ k. Then F is
s-independent. Consequently,  
n
|F| ≤ . (26)
s
The novelty compared to the statement of the same result in previous chapters is the fact of
s-independence.
Proof. We know from the nonuniform version of the theorem (Theorem 7.8) that F is
s∗ -independent. Therefore, F is s-independent by Corollary 7.12.

Extensions enabling a wide range of applications will follow. We start with a key lemma
which simultaneously illuminates the roles of polynomials, modularity and inclusion matrices
in handling set systems satisfying intersection conditions.
Lemma 7.14 (Frankl–Wilson, 1981). Let F be a k-uniform family, f (x) an integer-
valued polynomial of degree s ≤ k and p a prime number. Assume that

(i) f (k) 6≡ 0 (mod p);

(ii) f (|E ∩ F |) ≡ 0 (mod p) for each pair E, F ∈ F, E 6= F .

Then F is s-independent. Consequently,

 
n
|F| ≤ .
s

——————————————————————— 155

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 7. A CLASS OF HIGHER INCIDENCE MATRICES
Note that f (x) is required to take integer values at integral values of x but the coefficients
of f do not have to be integers. We can thus look at a value of the polynomial mod p but
we cannot reduce the polynomial itself
mod p. (Powers of p may occur in the denominators
of the coefficients: the polynomial xs is a typical example.) Rational and modular linear
algebra are intertwined in the proof. (This is not the first instance of such interaction. The
fourth proof of the Oddtown Theorem (Exercise 1.1.2) and the second proof of the “Mod-pk -
town Theorem” (Exercise 1.1.24) provided simple examples. A similar procedure was used
in the proof of a mod pα version of the RW Theorem (Theorem 5.36) and in Exercises 7.1.1
and 7.1.2.)
Ps x


Proof. As in (9), we write f (x) as f (x) = i=0 αi i
. (Clearly, the coefficients αi are
rational. Actually, they are integers (Exercise 7.3.3), but we do not require this fact.) Let us
consider the f -intersection matrix A = A(F, f ), defined by (14). By conditions (i) and (ii),
the diagonal entries of A are not divisible by p, all the others are. Therefore A has nonzero
determinant mod p and thus det A 6= 0. From Lemma 7.6 we infer that F is s∗ -independent
and therefore s-independent by Corollary 7.12.

Now, a modular extension of the unabridged RW Theorem follows as a simple corollary.

Recall from Definition 5.14 that for a set L of integers and integers r, t we say that

t∈L (mod r)

if t ≡ l (mod r) for some l ∈ L.

Theorem 7.15 (Frankl–Wilson, 1981). Let L be a set of s integers and p a prime
number. Assume F is a k-uniform family of subsets of a set of n elements such that

(i) k ∈
/ L (mod p);

(ii) |E ∩ F | ∈ L (mod p) for E, F ∈ F, E 6= F .

Then F is s-independent. Consequently

 
n
|F| ≤ .
s

Although this result is weaker than the more recent Theorem 5.37 in that it requires F to be
uniform, the fact of s-independence is an important extra. It is an open problem whether or
not s-independence follows under the weaker hypothesis of Theorem 5.37 where the condition
of k-uniformity is replaced by the condition that for E ∈ F, we have |E| ∈/ L (mod p).
We note that the condition that p is a prime cannot be removed in general (Exer-
cises 5.9.3–5.9.5). An extension of the case s = p−1 to prime power moduli will be presented
later in this section (Theorem 7.18).
Q
Proof. Set f (x) = li ∈L (x − li ) and apply Lemma 7.14.

156 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 7.3. INCLUSION MATRICES OF UNIFORM FAMILIES
It is clear that this result includes the RW Theorem. Another important result, with appli-
cations to the theory of partially ordered sets (Section 8.3) follows next. Recall that in this
section, we work in characteristic zero.
Theorem 7.16 (D. H. Gottlieb, 1966). For 0 ≤ j ≤ i ≤ n, the matrix In (i, j) has full
rank:    
n n
rk In (i, j) = min , . (27)
i j
Proof. First consider the case i+j ≥ n, i. e., ni ≤ nj . Then any two distinct i-sets intersect

in i − 1, i − 2, . . . , or i − (n − i) = 2i − n elements. The number of different intersection sizes

is therefore n − i ≤ j. Hence the RW Theorem tells us that the complete i-uniform family
is j-independent, i. e., In (i, j) has full rank.
we obtain The case i + j < n reduces to the preceding case via the identity In (i, j)T =
In (n − j, n − i).

It is a much more difficult problem to determine the mod p rank of the matrices In (i, j).
The mod 2 case was solved by Linial and Rothschild (1981). For general p, Richard M.
Wilson found the solution, a theory rather than a theorem. We hope he will publish it some
day. Here we only describe the simplest special case, with an application to chromatic graph
theory, to follow in Section 8.2, in mind.
Proposition 7.17. The mod 2 rank of In (i, i − 1) is n−1


i−1
.
n
− n−1 = n−1




So the “defect,” compared to the rank in characteristic zero, is i−1 i−1 i−2
. These
formulae suggest a proof of combinatorial nature.
Proof. Let F = Xi be the complete i-uniform family on |X| = n points.

Fix some a0 ∈ X and consider the subfamily F0 formed by those members of F containing
a0 . For E ∈ F0 , clearly E is the unique member of F0 containing the (i − 1)-set E \ {a0 }.
n−1


Consequently the rows corresponding to F0 contain an i−1 -dimensional identity matrix
and are therefore linearly independent over any field. This proves that for any p,
 
n−1
rkp In (i, i − 1) ≥ . (28)
i−1
To prove that equality holds for p = 2, we show that for each E ∈ F \ F0 , the rows
corresponding to F0 ∪{E} are dependent mod 2. Indeed, consider the (i+1)-set F = E∪{a0 }.
Then all (i − 1)-subsets of F are contained in exactly two i-subsets of F and therefore the
sum of the rows corresponding to the i-subsets of F is zero mod 2.

It is to be lamented that no modular extension of the RW Theorem exists for general moduli.
Exercises 5.9.3–5.9.5 give counterexamples to the immediate extension modulo 6 and modulo
powers of primes ≥ 7. There are, however, cases when a straight extension is still a possibility.
Conjecture C(r). Let F be a k-uniform family over a universe of n points. Let r ≥ 2 and
suppose that
|E ∩ F | 6≡ k (mod r)

——————————————————————— 157

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 7. A CLASS OF HIGHER INCIDENCE MATRICES
for any pair of distinct members E, F ∈ F. Then F is (r − 1)-independent and consequently
 
n
|F| ≤ .
r−1

Conjecture D(r). Let F be a k-uniform family over a universe of n points. Let r ≥ 2 and
let L be a set of two integers, not congruent to k mod r. Assume

|E ∩ F | ∈ L (mod r)

for any pair of distinct members E, F ∈ F. Then F is 2-independent and consequently

 
n
|F| ≤ .
2

Both conjectures are correct when r is a prime power. We conclude this section with a
proof of the first one; a geometric application of that result was described in Section 5.5. A
proof of of the second conjecture for prime powers is an exercise (7.3.8).
Theorem 7.18 (Frankl–Wilson, 1981). For prime powers q, Conjecture C(q) is valid.
α
For the proof, we need the following observation
(Proposition 5.31). Let q = p , p a prime.
r−1
For any integer r, the binomial coefficient q−1 is divisible by p precisely if r is not divisible
by q.
Proof of Theorem 7.18. Consider the polynomial f (x) = x−k−1


q−1
. Let A = A(F, f ) be the
f -intersection matrix of F (see equation (14)). By the just quoted Proposition 5.31 and
the conditions of the Theorem, the diagonal entries of A are not divisible by p while the
off-diagonal entries are. Hence A is nonsingular mod p and therefore nonsingular over Q. It
follows by Lemma 7.6 that F is (q − 1)∗ -independent and therefore (q − 1)-independent by
Corollary 7.12.

Exercises
Ex. 7.3.1. Prove that the 2-intersection matrix of a Steiner triple system is nonsingular.
Hint. Prove that every Steiner triple system is 2-independent.

Ex. 7.3.2. Prove: for k-uniform families, s-independence implies t-independence, provided s ≤
t ≤ k.
Hint. Use Corollary 7.11 with t in the role of s.

Ex. 7.3.3. Characterize the integral valued polynomials, i. e., polynomials f (x) which take integer
values for integral x.
x


Hint. They are precisely the integral linear combinations of the polynomials s , s ∈ Z, s ≥ 0. Use
Lagrange interpolation or induction on the degree for the proof.)

158 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 7.4. VC DIMENSION
Ex. 7.3.4 (R. R. Hall (1971), I. Z. Ruzsa, (1972)). Characterize the conguence preserving
polynomials, i. e., the polynomials f (x) such that for any a, b, n ∈ Z, n ≥ 1, if a ≡ b (mod n) then
f (a) ≡ f (b) (mod n).
Hint. They are precisely the integral linear combinations of the polynomials As xs where As is the

least common multiple of the integers 1, 2, . . . , s.

Ex. 7.3.5. Give a counterexample to equation (7.11) in every finite characteristic.

Ex. 7.3.6. Prove that the first

statement of Corollary 7.11 remains valid in characteristic p
k−i
provided p does not divide j−i . In particular, equation (7.11) remains valid for p > k.

Ex. 7.3.7 (The case s = 1 works for arbitrary modulus). (a) Let F be a k-uniform family of
subsets of [n] and let r ≥ 2. Assume |E ∩ F | ≡ l (mod r) for every pair of distinct members
E, F ∈ F, and l 6≡ k (mod r). Prove: |F| ≤ n. (b) Prove: without the uniformity condition, if we
only assume that the size of each set is ≡ k (mod r), we still have |F| ≤ n + 1.
Hint. (a) Prove that F is 1-independent. For this, it suffices to prove 1∗ -independence. Set
f (x) = x − l and consider the matrix A(F, f ). The diagonal entries are ≡ k − l (mod r), the
off-diagonal entries are divisible by r. Take a prime power q that divides r but does not divide
k − l, and use q to show, as in the second proof of the “Mod-pk -town Theorem (Exercise 1.1.24),
that such a matrix is nonsingular. (Use Proposition 4.14.) (b) Prove 1∗ -independence as in part
(a).

Ex. 7.3.8. Prove that conjecture D(q) holds for every prime power q. (In other words, the case
s = 2 works for prime power moduli.)
Hint. Let q = pα , L = {a, b}. Consider the polynomial f (x) = (x − a)(x − b). Show that the
diagonal entries of the matrix A(F, f ) are divisible by lower powers of p than are the off-diagonal
entries. Use Proposition 4.14 to show that A(F, f ) is nonsingular (over Q).

7.4 Linear dependencies among the rows of inclusion

matrices and the Vapnik–Chervonenkis dimension
The purpose of this section is to investigate the coefficients in linear relations between the
rows of (extended) inclusion matrices. We derive structural consequences of such dependen-
cies, yielding further criteria of s∗ -independence. One of these results (Theorem 7.26) has
an application to the theory of edge-reconstruction of graphs (Section 8.1). Another result
(Exercises 7.4.15) provides one more modular version of the RW Theorem.
But the main accomplishment in this section will be the foundation of a new concept
of fundamental importance, the Vapnik–Chervonenkis dimension (VC dimension) of a set-
system.
Let F ⊆ 2X be a set system over the (finite or infinite) universe X.
Definition 7.19. The trace of F in A ⊆ X is the set TA (F) = {A ∩ F : F ∈ F}. We say
that A is shattered by F if for every B ⊆ A, there exists a F ∈ F such that B = A ∩ F . In
other words, A is shattered if the trace of F in A is the entire power set 2A .

——————————————————————— 159

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 7. A CLASS OF HIGHER INCIDENCE MATRICES
Definition 7.20. The Vapnik–Chervonenkis dimension (VC dimension) of a set system F
is the maximum size of a set shattered by F (or infinite, if no such maximum exists).

Definition 7.21. The shatter function of F is the function σF : 2X → Z defined by

σF (A) = |TA (F)| . (29)

Applications ranging from probability theory to computational learning theory are largely
based on the following fundamental theorem.
Theorem 7.22 (Shatter Function Theorem). If the VC dimension of the set system
F ⊆ 2X is s < ∞ then for any finite set A ⊆ X, the shatter function is bounded by
s  
X |A|
σF (A) ≤ (30)
i=0
i

In particular, if F ⊆ 2X for a finite set X, |X| = n, then

s  
X n
|F| ≤ . (31)
i=0
i

This bound is clearly tight, as the family of all sets of size ≤ s demonstrates.
The Shatter Function Theorem was discovered independently by three sets of authors in
remarkable simultaneity: Perles–Shelah (1972), N. Sauer (1972), and Vapnik–Chervonenkis
(1971). No less remarkable is the range of contexts in which the result arose (logic, set
theory, probability theory).
In many applications, F will consist geometric objects (half-spaces, balls, plane polygonal
regions) and will have finite VC dimension. The exercises provide examples of obtaining
upper bounds on the VC dimensions of some families of geometric objects and illustrate the
geometric consequences of the Shatter Function Theorem.
For applications to computational learning theory we refer to Blumer et al. (1989),
Maass–Turán (1990), Littlestone (1987), and the references in those papers. Some of the
exercises illustrate how the VC dimension of “concept classes” yields lower bounds on the
complexity of learning a concept. While in the simple minded model discussed there, these
lower bounds are usually not optimal, in more sophisticated models defined in the context
of Valiant’s “PAC” model,1 the relationship is more intimate and both the lower and the
upper bounds are closely related to the VC dimension (see Blumer et al.).

1
In Valiant’s model of learnability, we attempt to learn a “target concept” (e. g., “elephant”) based on
seeing a sequence of randomly drawn samples (say, animals) arriving with a fixed probability distribution.
(The teacher reveals, which of the sample animals are elephants.) It is expected that after a while, we would
Probably be able to make an Approximately Correct guess of what the concept is (classify most newly arriving
animals as elephants or non-elephants; the new arrivals are from the same fixed but unknown probability
distribution).

160 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 7.4. VC DIMENSION
* * *

Let us now return to the basic context of this section: the study of linear dependencies
among the rows of inclusion matrices.
It will be useful for some of the applications to consider the modular version of the notions
of s- and s∗ -independence.
In this section, let p be either 0 or a prime number. As before, Fp denotes the field of
p elements for p a prime number. Let us define F0 to be Q, the field of rational numbers.
For s ≥ 1 we say that F is s-independent in characteristic p if the rows of the s-incidence
matrix I(F, s) are linearly independent over Fp . We define s∗ -independence in characteristic
p analogously, referring to the extended inclusion matrix I ∗ (F, s). For p = 0, these concepts
coincide with our previous notions of s-independence (s∗ -independence), resp.
Observation 7.23. If F is s-independent (s∗ -independent, resp.) in characteristic p for
some p, then F is s-independent (s∗ -independent, resp.) (in characteristic 0).

If F is s∗ -dependent (i. e., not s∗ -independent) in characteristic p then there exist coeffi-
cients γ(E) ∈ Fp (E ∈ F), not all zero, such that the corresponding linear combination of the
rows of I ∗ (F, s) is zero (in Fp ). We shall use the phrase “F is s∗ -dependent in characteristic
p with coefficients γ(E)” to describe this circumstance (including the condition that not all
the γ(E) are zero).
We rephrase the definition of linear dependence in a more combinatorial form, avoiding
matrix language.
Let F be a set system over the universe X of n points. For a subset A ⊆ X, let
def
X
δ(A) = γ(E). (32)
A⊆E∈F

Observation 7.24. A family F is s∗ -dependent in characteristic p with coefficients γ(E) ∈

Fp , if and only if for each B ⊆ X of cardinality |B| ≤ s,

δ(B) = 0. (33)

For s-dependence in characteristic p with coefficients γ(E) ∈ Fp it is necessary and sufficient

that (33) holds for each B ⊆ X of cardinality |B| = s.

By definition, δ(A) is the sum of the coefficients γ(E) for those edges E which contain A.
It will be useful to consider, more generally, the sum over those edges which split A in a
prescribed way. For A0 ⊆ A ⊆ X, let
def
X
δ(A, A0 ) = γ(E). (34)
E∈F , E∩A=A0

——————————————————————— 161

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 7. A CLASS OF HIGHER INCIDENCE MATRICES
The Inclusion-Exclusion Formula (Exercise 7.4.1) provides the following expression for
δ(A, A0 ).
Proposition 7.25. For any A0 ⊆ A ⊆ X,
X
δ(A, A0 ) = (−1)|B−A0 | δ(B). (35)
A0 ⊆B⊆A

From this formula we derive another simple criterion of s∗ -independence.

Theorem 7.26. For s ≥ 0, if |F| < 2s+1 then F is s∗ -independent in any characteristic.
In other words, s∗ -dependent families cannot have fewer than 2s+1 edges. What we actually
obtain is explicit structural information on s∗ -dependent families.
Theorem 7.27 (Frankl–Pach, 1983). If F is s∗ -dependent in some characteristic, then
there exists a subset A ⊆ X of cardinality |A| = s + 1 such that every subset B of A occurs
as B = A ∩ E for some E ∈ F.
The set {A ∩ E : E ∈ F} is called the trace of F in A. For the set A in the theorem, the
trace of F in A is all of 2A .
Proof. Let γ(E) be the coefficients satisfying (33). Replacing F by a subfamily if necessary,
we may assume γ(E) 6= 0 for every E ∈ F.
Let A ⊆ X be a set of minimum cardinality such that δ(A) 6= 0. (Clearly |A| ≤
maxE∈F |E|.) We have |A| ≥ s + 1 by (33). Therefore it will be sufficient to show that for
every subset A0 ⊆ A there exists E ∈ F such that A ∩ E = A0 .
Let us fix A0 ⊆ A and use Proposition 7.25 to calculate δ(A, A0 ). By the minimal
choice of A we see that all terms on the right hand side of (35) vanish except for the term
corresponding to B = A. We infer that
X
γ(E) = δ(A, A0 ) = (−1)|A−A0 | δ(A) 6= 0.
E∈F , E∩A=A0

So the sum on the left hand side cannot be empty.

The Shatter Function Theorem (Theorem 7.22) is now an immediate corollary.
Corollary 7.28. If F is a family of m > si=0 ni subsets of an n-set X then there exists
P

a subset A ⊆ X of cardinality |A| = s + 1, shattered by F.

Proof. It is trivial that a family of this size must be s∗ -dependent (Proposition 7.4).
This immediately implies the second statement in the Shatter Function Theorem. The
first statement then follows by replacing F with its trace in A (observing that the VC
dimension of the trace TA (F) is not greater than the VC dimension of F.)

Exercises
♦Ex. 7.4.1. Prove the Inclusion–Exclusion formula, as stated in Proposition 7.25.
Ex. 7.4.2. Prove: the VC dimesnion of a finite set-system F is at most log2 |F|.

162 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 7.4. VC DIMENSION
Ex. 7.4.3. Prove that the VC dimension of the set of halfspaces in Rn is n + 1.
Hint. To prove the upper bound, apply Radon’s Lemma (Lemma 3.21).

Ex. 7.4.4. Let x ∈ Rn . Consider the set of those half-spaces in Rn which do not contain x. Prove
that the VC dimension of this family of half-spaces is n.
Hint. To prove the upper bound, apply Radon’s Lemma (Lemma 3.21) to a set of n + 2 points
which contains x.

Ex. 7.4.5 (E. F. Harding, 1967). Let S be a set of n points in Rd . A planar partition of S is a
partition S = S1 ∪ S2 such that S1 and S2 are separated by a hyperplane (i. e., the convex hulls of
S1 and S2 are disjoint). (a) Prove that the number of planar partitions of S is at most
d  
X n−1
. (36)
d
i=0

(b) Show that this bound is tight.

Hint. (a) Call one of the points of S the root. Apply Ex. 7.4.4 (x = the root), and use the
Fundamental Theorem. (b) Use the moment curve.

Ex. 7.4.6 (L. Schläfli, d

P 1852).
Prove: the maximum number of regions into which R is partitioned
d n
by n hyperplanes is i=1 i .
Hint. Use induction.
♦Ex. 7.4.7.* The V C dimension of balls in Rn is n + 1, i.e. the same as that of halfspaces.

General comment: When looking for the V C dimension of a family of convex sets, one can
always assume that the set A establishing the dimension (the largest shattered set) has all
its points as vertices on its convex hull conv(A). Otherwise, x ∈ conv(A \ {x}) would hold
for some x ∈ A. Thus no convex set C could satisfy C ∩ A = A \ {x}.
♦Ex. 7.4.8.* The V C dimension of convex k-gons is 2k + 1.

* * *

In computational learning theory we consider a family F ⊆ 2X referred to as the concept

class. Members of F are the concepts, X is the sample space (usually Rn ). The teacher
selects a target concept F0 ∈ F which we wish to learn under some specific learning scenario.
The scenario to be considered in the following sequence of exercises is called “learning
from counterexamples.” We are allowed to specify a subset Y ⊆ X as our hypothesis. If
Y = F0 , the learning has been completed; otherwise the teacher reveals some element of the
symmetric difference Y 4F0 (a counterexample to our hypothesis). We consider the worst
case complexity: at most how many queries are required to learn an arbitrary F0 ∈ F?
♦Ex. 7.4.9. Prove that the worst case complexity of learning a concept class F is greater than
the VC dimension of F.
♦Ex. 7.4.10 (Polynomial learnability). Usually the sample space X is infinite but our horizon

——————————————————————— 163

c László Babai and Péter Frankl. 1988, 1992, 2020.
 CHAPTER 7. A CLASS OF HIGHER INCIDENCE MATRICES
is limited to just a finite (but variable and perhaps growing) subset Y ⊂ X. Let us say that we
learned a concept F0 ∈ F over Y when we arrived at a hypothesis to which no counterexample
exists in Y . Prove: if the VC dimension of F is finite then the number of queries required to learn
a concept over Y is bounded by |Y |c for some constant c (independent of Y ).

Ex. 7.4.11* (Maass, Turán, 1989). Let X = {0, 1}n ⊂ Rn and let F consist of those subsets F
of X sepatated from X \ F by a hyperplane. Prove that the number of queries required to learn a
concept from this class is Θ(n2 ) in the worst case.

* * *

♦Ex. 7.4.12. Prove that the bound in Theorem 7.26 is tight for every p.
Hint. Let F be the family consisting of all subsets of an (s + 1)-set. Prove that this family is
s∗ -dependent with coefficients ±1.
The following three exercises are from Deza–Frankl–Singhi (1983). They lead to yet another
nonuniform modular version of the RW Theorem.
♦Ex. 7.4.13. Suppose F is s∗ -dependent in characteristic p with coefficients γ(E) ∈ Fp . Assume
that p is either zero or greater than s. Let f (x) be a polynomial of degree at most s over Fp . Then
X
γ(E)f (|E ∩ A|) = 0 for all A ⊆ X. (37)
E∈F

♦Ex. 7.4.14. Suppose the family F is s∗ -dependent in characteristic p with coefficients γ(E) ∈ Fp .
Then for each E0 ∈ F with γ(E0 ) 6= 0, at least one of the following holds:

(i) |E0 ∩ E| ≡ |E| (mod p) for some E 6= E0 , E ∈ F.

(ii) |E0 ∩ E| assumes at least s + 1 distinct values mod p for E 6= E0 , E ∈ F.

We admit the possibility p = 0 here. Naturally, a ≡ b mod 0 means a = b.

♦Ex. 7.4.15. Let L be a set of s integers and p a prime number or zero. Assume F is a family of
subsets of a set of n elements such that

(a) |E| ∈
/ L (mod p) for E ∈ F;

(b) |E ∩ F | ∈ L (mod p) for E, F ∈ F, E 6= F .

Then F is s∗ -independent in characteristic p. Consequently,

     
n n n
|F| ≤ + + ··· + .
s s−1 0

♦Ex. 7.4.16. Deduce the Nonuniform RW Theorem (Theorem 7.8) from Exercise 7.4.14.
♦Ex. 7.4.17. Deduce the Frankl–Wilson Theorem (Theorem 7.15) from Exercise 7.4.15.

164 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 7.5. SHADOWS OF S-INDEPENDENT FAMILIES
7.5 Shadows of s-independent families
In this section we derive further consequences of s-independence. Let F be a system of m
subsets of the n-element universe X.
For 0 ≤ i ≤ n, we define the i-shadow (or shadow of rank i) of the family F, denoted
i
∂ F, to consist of those i-subsets of X contained in at least one member of F :
   
i X
∂F = A∈ : ∃E ∈ F, A ⊆ E . (38)
i

Note that an i-subset A belongs to ∂ i F precisely if the column corresponding to A in I(F, i)

is not all zero.
As the all-zero columns do not contribute to the rank of a matrix, we observe that

rk I(F, s) ≤ |∂ s F| (39)

and
rk I ∗ (F, s) ≤ |∂ s F| + · · · + |∂ 0 F|. (40)
Since, obviously, |∂ i F| ≤ ni , a slight improvement of various versions of the RW Theo-

rem is immediate.
Corollary 7.29. Under the conditions of the Frankl–Wilson Theorem (Theorem 7.15), we
have
|F| ≤ |∂ s F|. (41)
Under the conditions of the Nonuniform RW Theorem (Theorem 7.8), we have
s
X
|F| ≤ |∂ i F|. (42)
i=0

For an s-independent family, by inequality (39) we have

|∂ s F|
≥ 1. (43)
|F|

It seems natural to ask, how small other shadows of an s-independent family can be. We
have an answer for uniform families.
Theorem 7.30 (Frankl–Füredi, 1984). Let F be an s-independent, k-uniform family
(s ≤ k). Then
k+s


|∂ t F| t
≥ k+s
for s ≤ t ≤ k. (44)
|F| s

Note that inequality (44) is best possible, as the complete k-uniform family on n = k + s
points demonstrates.
For the proof, we introduce some notation.

——————————————————————— 165

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 7. A CLASS OF HIGHER INCIDENCE MATRICES
Let us fix a point u of the k-uniform family F. Define F(u) = {E ∈ F : u ∈ E}),
F(e
u) = {E \ {u} : u ∈ E ∈ F}. Let us divide the columns of I(F(u), i) into two blocks:
I(F(u), i) = I 0 (F, u, i)|I 1 (F, u, i) ,
 

where the columns of I 0 (F, u, i) correspond to those i-subsets of X not containing u. Observe
that
I 0 (F, u, i − 1) = I 1 (F, u, i). (45)
Moreover, if we list the members of F(e u) in the order of the corresponding members of F(u),
then
I 0 (F, u, i) = I(F(e
u), i). (46)
Lemma 7.31. If F is k-uniform and s-independent then for each point u, I 0 (F, u, s) has
full row rank.
Proof. From equations (45), (46), and (23) we obtain
I 1 (F, u, s) = I 0 (F, u, s − 1) = I(F(e u), s) (47)
1
= k−(s−1)
I(F(e u), s)In−1 (s, s − 1) (48)
s−(s−1)
1
= I 0 (F, u, s)In−1 (s, s − 1) (49)
k−s+1
Clearly, F(u) is s-independent (because F is), so the (column) rank of I(F(u)) is |F(u)|.
By comparing the far ends of equations (47)-(49) we notice that the column space of
I 1 (F, u, s) is contained in the column space of I 0 (F, u, s) so the rank of the latter is the
same, i. e., it equals the number of rows of I 0 (F, u, s).
Proof of Theorem 7.30. Note that inequality (44) trivially holds for s = 0, also (by (43)) for
t = s, and trivially for t = k. Suppose 1 ≤ s < t < k and apply induction on k. Observe
that s-independence of F implies (s − 1)-independence of F(e u) by Lemma 7.31 and equation
(46). We can therefore apply the induction hypothesis to F(e u).
k+s−1

k+s


t−1 t−1 t t
|∂ F(e u)| ≥ |F(e u)| k+s−1
= |F(u)| k+s
. (50)
s
k s
P
Using the trivial identities u∈X |F(u)| = k|F| and
X
| ∂ t−1 F(e
u)| = t|∂ t F|,
u∈X

we infer from (50) that

k+s


t 1 X t−1 1X t t
|∂ F| = |∂ F(e u)| ≥ |F(u)| k+s


t u∈X t u∈X k s
k+s k+s



t 1X t
= k+s
|F(u)| = |F| k+s
.
s
k u∈X s

166 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 7.5. SHADOWS OF S-INDEPENDENT FAMILIES
Exercises
♦Ex. 7.5.1 (G. O. H. Katona 1964). Let F be a k-uniform family and assume |E ∩ F | ≥ r for
every pair E, F of distinct members of F. Prove that for every t, if k − r ≤ t ≤ k then
2k−r


|∂ t F| t
≥ 2k−r
. (51)
|F| k

* * *

The following sequence of five exercises introduces the left compression method and culmi-
nates in the Kruskal–Katona Theorem (Exercise 7.5.6). F will be a family of subsets of [n]
throughout.
Ex. 7.5.2 (The shift operator). Let 1 ≤ i < j ≤ n. The left shift operator Sij : 2[n] → 2[n] is
defined as follows. The family F 0 = Sij (F) is obtained by replacing every F ∈ F by F 0 = F \{j}∪{i}
if j ∈ F , i ∈ / F , and F 0 ∈
/ F. Otherwise we keep F . — Prove that the left shift operators
semicommute with the shadow operators in the following sense:

∂ t Sij F ⊆ Sij ∂ t F.

Ex. 7.5.3. We say that F is left compressed if Sij F = F for every i, j (1 ≤ i < j ≤ n). For
u ∈ [n], set F(u) = {F ∈ F : u ∈/ F }. Using the notation from the proof of Theorem 7.30, prove:
if F is left compressed and k-uniform then

1) ⊇ ∂ k−1 (F(1)).
F(e

Ex. 7.5.4. Prove: if F is k-uniform and left compressed then

∂ k−1 (F) = ∂ k−1 (F(1)).

Ex. 7.5.5. Prove: if F is k-uniform and left compressed then

|∂ k−1 (F)| = |F(1)| + |∂ k−2 (F(e
1))|.

Ex. 7.5.6* (J. B. Kruskal, 1963; G. O.

H. Katona, 1967). Let α ∈ R, α > k. If F is a k-uniform
family and |F| ≥ αk then |∂ t F| ≥ αt for every t ≤ k.

Hint. It suffices to prove the result for t = k − 1. Proceed by induction on n. The starting cases
k = 1 and n = k are trivial. We may assume F is left compressed. (Keep performing left shifts
until F becomes left compressed. By Exercise 7.5.2, this will not increase the size of the t-shadow.)
Claim. |F(1)| ≥ α−1


k−1 . (For otherwise
   
α α−1
F(1) ≥ − |F(1)| > .
k k

——————————————————————— 167

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 7. A CLASS OF HIGHER INCIDENCE MATRICES
But then, by Exercise 7.5.3,
|F(1)| = |F(e1)| ≥ |∂ k−1 (F(1))|;
and the right hand side is greater than α−1


k−1 by the inductive hypothesis.)
Finally, by Exercise 7.5.5,
|∂ k−1 F| = |F(1)| + |∂ k−2 F(e
1)|.
The first term on the right hand side is ≥ α−1 α−1



k−1 by the Claim above; the
second term is ≥ k−2
α
by the inductive hypothesis. The sum of these two lower estimates is k−1 .

168 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
Chapter 8

Applications of inclusion matrices

The applications of intersection theorems to Ramsey Theory and geometry, discussed in

the last two sections of Chapter 5, originally arose as applications of the inclusion matrix
technique since the polynomial space proofs of the Ray-Chaudhuri–Wilson theorem and its
variants were not available at the time. In this chapter we give three more applications of the
inclusion matrix technique; two of them to graph theory, the third one to partially ordered
sets.

8.1 The edge-reconstruction problem

Let G = (X, E) and K = (Y, F ) be graphs with the same number of vertices: |X| = |Y | = n.
Let c(G, K) denote the number of copies of K in G, i. e., the number of subgraphs of G,
isomorphic to K. Let m(G, K) denote the number of bijections Y → X which map K to
a subgraph of G. Different bijections may map K to the same subgraph of G; the number
of ways this can happen is the number of isomorphisms of K with its copy in G. This is
the same as the number of automorphisms of K. (An automorphism of K is by definition a
K → K isomorphism.) To sum up, we have

m(G, K) = aut(K)c(G, K), (1)

where aut(K) denotes the number of automorphisms of K.

We say that the graphs G1 and G2 have the same deck of k-edge-deleted subgraphs if for
every graph K with m − k edges,

c(G1 , K) = c(G2 , K). (2)

A graph H is a reconstruction of G from its k-edge-deleted subgraphs if G and H have the

same deck of k-edge-deleted subgraphs. We say that G is reconstructible from its k-edge-
deleted subgraphs if all reconstructions of G from its k-edge-deleted subgraphs are isomorphic
to G. The following conjecture was stated by Harary (1964).

c László Babai and Péter Frankl. 1988, 1992, 2020.

169
CHAPTER 8. APPLICATIONS OF INCLUSION MATRICES
Edge-Reconstruction Conjecture. Every graph with at least four edges is reconstructible
from its 1-edge-deleted subgraphs.
This is actually a weaker version of a much older problem, the Reconstruction Conjecture,
formulated by P. J. Kelly and S. M. Ulam in 1942, stating that every graph with at least
three vertices is reconstructible from its vertex-deleted subgraphs. An extensive survey on
the reconstruction problems was written by Bondy and Hemminger (1972). An exposition of
some of the key ideas can be found in Nash-Williams (1978). Chapter 14 of Lovász (1979c)
contains substantial material on this and several quite different (algebraic) reconstruction
questions.
While the Reconstruction Conjecture still remains essentially intractable, the Edge-
reconstruction Problem has been confirmed for a surprising general class of graphs. In 1972,
Lovász came out
with an inclusion-exclusion argument proving the conjecture for graphs
with m > 12 n2 edges, or, in other words, for graphs with edge density > 1/2.
(The edge
n
density is the ratio of the number of edges to the maximum possible number, 2 .) A further
significant step was subsequently taken by V. Müller who succeeded in drastically reducing
the density assumption.
Theorem 8.1 (V. Müller (1977)). Let G be a graph with n vertices and m edges. If
2m−k > n! then G is reconstructible from its k-edge-deleted subgraphs.
This indeed is a substantial relaxation of the density constraint: the inequality

m − k ≥ n(log2 n − 1)

is sufficient for the condition of the theorem to hold.

The original proof and several later versions were based on Inclusion–Exclusion. We
present a proof, due to Godsil, Krasikov, and Roditty (1987), which shows that the result
follows from Theorem 7.26, one of the simplest combinatorial conditions guaranteeing t∗ -
independence of a set system.
We remark that Müller stated the theorem for k = 1 only. The more general result
also follows by his method (see also Nash-Williams (1978)), but the first paper to state this
explicitly seems to be Godsil et al. (1987).

Proof. Suppose we have two nonisomorphic graphs, G1 and G2 , with the same deck of k-
edge-deleted subgraphs. Both graphs have n
vertices and m edges.
Let V be a set of n elements and X = V2 . Consider the m-uniform families Ei (i = 1, 2)
on the set X where a subset E ⊆ X belongs to Ei if the graph (V, E) is isomorphic to Gi . We
observe that the two families have no common member. Since there are n! bijections from
the vertex set of Gi to V , the number of members of Ei is at most n!.
Claim 8.2. The family F = E1 ∪ E2 is (m − k)-dependent.
Since F is uniform, it follows from the Claim that F is (m − k)∗ -dependent as well. In view
of Theorem 7.26, we then infer that

|F| ≥ 2m−k+1 . (3)

170 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 8.2. CHROMATIC CRITICAL GRAPHS
m−k+1
Consequently, 2n! > 2 , proving the Theorem.
In order to justify the Claim, we have to assign a coefficient to each member of F to
produce a nontrivial linear relation of the rows of the inclusion matrix I(F, m − k).
There will be two different coefficients only. Let αi denote the number of automorphisms
of Gi .
For E ∈ Ei , we define the coefficient γ(E) to be (−1)i αi . Note that apart from the sign,
this is the number of isomorphisms Gi → (V, E).
Now let S ⊂ X. For i = 1, 2 let
X
δi (S) = γ(E). (4)
S⊂E∈Ei

According to Observation 7.24, in order to verify that the γ(E) are indeed the coefficients
of a linear relation between the rows of I(F, m − k), we have to check that

δ1 (S) + δ2 (S) = 0 (5)

for each S ⊂ X, |S| = m − k.

It is clear that (−1)i δi (S) is the number of those bijections of V to the vertex set of Gi
which map S into the edge set of Gi . By assumption, this number is the same for i = 1 and
2.

Exercises
We call a graph vertex-reconstructible if it is reconstructible from its 1-vertex-deleted sub-
graphs. Edge-reconstructible graphs are defined analogously.
Ex.
 8.1.1
 . Prove: almost every graph is vertex-reconstructible. (“Almost every” means all but
n n
o 2( 2 ) of the 2( 2 ) graphs on vertex set [n].)

Ex. 8.1.2. Prove: the deck of 1-edge-deleted subgraphs determines the deck of k-edge-deleted
subgraphs.

Ex. 8.1.3. Prove: if a graph is vertex-reconstructible then it is edge-reconstructible.

Ex. 8.1.4. Disconnected graphs are vertex-reconstructible.

Ex. 8.1.5* (P. J. Kelly, 1957). Trees are vertex-reconstructible.

Ex. 8.1.6. Infinite forests are not vertex-reconstructible.

Ex. 8.1.7* (L. Pyber, 1987). Hamiltonian graphs are edge-reconstructible.

Ex. 8.1.8** (W. T. Tutte, 1979). If two graphs have the same deck of 1-vertex-deleted subgraphs
then their adjacency matrices have the same characteristic polynomial.

——————————————————————— 171

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 8. APPLICATIONS OF INCLUSION MATRICES
8.2 Chromatic critical graphs
“Finite basis theorems” play a key role in algebraic theories. The general idea is to show for
some infinite set of objects that it has a finite subset from which all elements in the set can
be generated by some elementary operations.
The archetype of such results is Hilbert’s Basis Theorem asserting that every ideal in the
ring of polynomials in several variables over a field is finitely generated (Hungeford (1974),
p. 391).
In combinatorics, Kuratowski’s characterization of nonplanar graphs can be regarded as
being of this kind: every nonplanar graph arises from the two basic ones, K5 and K3,3 , by
simple operations (adding new vertices and edges, splitting an edge by a new vertex). A far-
reaching generalization of this result has recently been obtained by Robertson and Seymour
who proved that for any set of graphs closed under forming subgraphs and contracting edges,
the class of graphs not in this set has a finite basis in Kuratowski’s sense. Although the size
of the basis in the Robertson–Seymour Theorem is quite unmanageable, the theoretical
significance of the existence of a finite basis has, like in Hilbert’s case, enormous.
In this section we shall present a very simple “finite basis theorem” in the chromatic
theory of graphs and then go on to giving an essentially sharp estimate on the size of the
basis.
The classes to be shown to have a “finite basis” in some sense will be defined in terms of
two parameters, the chromatic number and the covering number.
Let us recall the definitions. G = (V, E) will denote a graph with n vertices throughout.
The chromatic number χ(G) of G is the smallest integer k such that there exists a partition
V = V1 ∪ · · · ∪ Vk of the vertex set V into independent sets Vi , i = 1, . . . , k. The Vi are the
color classes.
A cover of a graph G is a set T of vertices such that at least one end of each edge of
G belongs to T . In other words, T is a cover if V \ T is an independent set. The covering
number τ (G) is defined as the minimum number of vertices that cover G. By the preceding
remark,
τ (G) = n − α(G), (6)

where α(G) is the maximum size of independent sets in G.

Let T (k, τ ) be the class of graphs with chromatic number k and covering number ≤ τ .
We shall show that for every k and τ , this class has a finite “basis” of minimal members.
G is said to be critically k-chromatic if χ(G) = k but the deletion of an arbitrary vertex
reduces the chromatic number. G is chromatic critical if it is critically k-chromatic for
k = χ(G).
Clearly, every k-chromatic graph contains a critically k-chromatic induced subgraph. In
each class T (k, τ ), the minimal members (those containing no other member as an induced
subgraphs) are chromatic critical. Our “finite basis theorem” (Proposition 8.4) asserts that
these minimal members form a finite set.
This is easy to verify for k ≤ 3. The only critically 2-chromatic graph is K2 , a single
edge. The critically 3-chromatic graphs are precisely the cycles of odd lengths.

172 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 8.2. CHROMATIC CRITICAL GRAPHS
No such simple characterization of critically k-chromatic graphs exists for k ≥ 4. Nev-
ertheless, the finiteness result is easy to prove. We first derive a lemma on the structure of
critically k-chromatic graphs.
Lemma 8.3. In a critically k-chromatic graph, the neighborhood of a vertex cannot be a
subset of the neighborhood of another one.
Proof. Assume all neighbors of u are also neighbors of v. Delete u. Then, by assumption,
the remaining graph can be colored by k − 1 colors. Now give u the color of v. This is a
legal coloring, in contradiction with the assumption that the chromatic number of the graph
is k.
Proposition 8.4. If G is critically k-chromatic and has a cover of size τ then
 
τ
n≤τ+ ≤ 2τ . (7)
bτ /2c

Proof. Let T be a cover, |T | ≤ τ . Then the neighborhood of each vertex in V \ T is entirely

in T , and these sets form a Sperner family

according to the lemma. Therefore by Sperner’s
Theorem (Section 4.4), |V \ T | ≤ bτ τ/2c .

L. Lovász succeeded in proving a a much better upper bound. The proof employs a
rank argument for mod 2 inclusion matrices. We should also mention that the proof was
inspired by a well-known combinatorial proof of the Brouwer Fixed Point Theorem (see, e.g.,
Pontriagin (1952)).
Theorem 8.5 (Lovász, 1973). If G is a critically k-chromatic graph with n vertices and
covering number τ then  
τ −1
n≤τ+ . (8)
k−2
Proof. Let A be an arbitrary independent set of G, and T = V \ A the corresponding cover.
We shall prove the theorem by showing that
 
|T | − 1
|A| ≤ . (9)
k−2

Delete an arbitrary vertex y ∈ A from G and fix a (k − 1)-coloring of the rest, Y1 ∪ · · · ∪

Yk−1 = V \ {y}. Since G is not (k − 1)-chromatic, y must have a neighbor in each class Yi ,
i = 1, . . . , k − 1. Let xi ∈ Yi be such a neighbor: {y, xi } ∈ E. Set F (y) = {x1 , . . . , xk−1 }.
Note that F (y) ⊂ T because A is an independent set.
T


We shall prove that the (k−2)-inclusion matrix of the family F = { F (y) : y ∈ A } ⊂ k−1
has full row rank mod 2. In other words, we claim:
The family F is (k − 2)-independent mod 2.
This immediately implies that the number of rows of the
(k − 2)-inclusion matrix of |F|
|T |
is not greater than the number of columns, i. e., |A| ≤ k−2 , which is only slightly weaker
than (9).

——————————————————————— 173

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 8. APPLICATIONS OF INCLUSION MATRICES
In order to derive (9) from our claim, we have to recall that the mod 2 rank
of the
m−1
(k − 2)-inclusion matrix of the complete (k − 1)-uniform family on m points is k−2 (Prop-
osition 7.17).
Now
let M be the (k − 2)-inclusion matrix of F. The columns of M are indexed by
T
G ∈ k−2 , the rows by the vertices in A, and the (y, G)-entry is 1 or 0 according to whether
or not G ⊂ F (y).
Consider a nontrivial linear combination over F2 of the rows. Since the coefficients are
0 or 1, this corresponds to assigning coefficient 1 to the members of a nonempty subset A0
of A. In order to verify that this linear combination is non-zero we have to prove that some
(k − 2)-set is contained in an odd number of sets F (y), y ∈ A0 .
Let us fix y0 ∈ A0 . Consider the corresponding coloring Y1 ∪ · · · ∪ Yk−1 = V \ {y0 }. We
shall compute the parity of the number of pairs (G, y) with G ⊂ F (y), y ∈ A0 , |G| = k − 2,
|G ∩ Yi | = 1 for i = 1, . . . , k − 2.
Precisely one such set G0 will correspond to y0 , namely, G0 = F (y0 ) − Yk−1 . We claim
that for any y 6= y0 , y ∈ A, there are 0 or 2 corresponding sets G. Indeed, if such G exists
then y ∈ Yk−1 and the single vertex u in F (y) \ G belongs to some Yi , 1 ≤ i ≤ k − 2.
(Observe that u 6= y0 because A is an independent set in G). Now, G0 = (G \ Yi ) ∪ {u} is
clearly the only other (k − 2)-set appropriate for y.
Thus the total

number of pairs (G, y) with the properties required is odd. Consequently
T
some G ∈ k−2 is contained in an odd number of F (y), y ∈ A0 , proving our claim.

Exercises
Ex. 8.2.1. (a) Prove: if G is critically k-chromatic than
τ (G) > (k/e)n1/(k−2) .
(b)* (Lovász, 1973) For some function c(k) of k, construct critically k-chromatic graphs G satisfying
τ (G) < c(k)n1/(k−2) .

Hint. (a) Invert inequality (8).

(b)

8.3 Partially ordered sets, unimodal sequences, and

the Sperner property
In this section we shall see Sperner’s Theorem, and much more, to be consequences of
Gottlieb’s Theorem.
Let us call two k-subsets A, B ⊆ Fdp equivalent if ϕ(A) = B for some linear automorphism
ϕ : Fnp → Fnp . Let fp (d, k) denote the number of resulting equivalence classes. What can we
say about the integer sequence
fp (d, 0), fp (d, 1), . . . , fp (d, n),

174 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 8.3. UNIMODALITY, SPERNER PROPERTY
d
where n = p ?
We can determine the first few terms precisely. We may also observe that

fp (d, k) = fp (d, n − k).

After some experimentation we may conclude that this sequence will probably grow for a
while and then decrease; if this is the case, the peak must occur in the middle.
This is an instance of a surprisingly general result. A finite sequence of real numbers is
called unimodal if it increases till its maximum and then decreases. (Equalities—plateaus—
are permitted.)
Theorem 8.6. Let G be a group of permutations of [n]. Call two sets A, B ⊆ [n] G-
equivalent if ϕ(A) = B for some ϕ ∈ G. Let fG (k) denote the number of G-equivalence
classes of k-subsets. Then the sequence

fG (0), fG (1), . . . , fG (n)

is unimodal.
Of course, the peak again occurs in the middle, since obviously

fG (k) = fG (n − k).

If this result was not general enough, here is an even more general one, at the same time
giving the clue to the proof.
A partially ordered set or poset is a set P together with a relation ≤ that has the usual
properties: reflexive, transitive, and the relations a ≤ b and b ≤ a imply a = b. Two elements
a, b ∈ P are comparable if a ≤ b or b ≤ a holds.
An isomorphism of two posets is a bijection between them that preserves the ordering
relation. An automorphism of the poset (P, ≤) is an isomorphism of (P, ≤) to itself.
Let (P, ≤) be a finite poset. A chain of length k is a k-subset of P , linearly ordered by the
relation ≤. The rank rk(x) of x ∈ P is the maximum length of chains consisting of elements
< x (i. e., ≤ x and 6= x). The height of (P, ≤) is 1+ the maximum rank of its elements.
A level consists of the elements of equal rank. A set of pairwise incomparable elements is
called an antichain. Obviously, every level is an antichain. The poset (P, ≤) is said to have
the Sperner property if no antichain has more elements than the largest level.
The automorphisms of (P, ≤) form a group of permutations of the set P . Clearly, au-
tomorphisms preserve levels. Therefore any subgroup G of the group of automorphisms of
(P, ≤) divides each level into G-equivalence classes. These classes are called the orbits of G.
The set of G-orbits is denoted by P/G. This set inherits a partial order from P , defined as
follows: for U, V ∈ P/G, we say U ≤ V if ∃u ∈ U , v ∈ V such that u ≤ v. If this is the case,
then such v clearly exists for every u ∈ U and conversely. We call P/G the factor of P by
G.
Let X and Y be two subsets of P . The incidence matrix IP (X, Y ) is an |X| × |Y | matrix;
the entry in position (x, y) is 1 or 0 depending on whether or not x ≤ y (x ∈ X, y ∈ Y ).

——————————————————————— 175

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 8. APPLICATIONS OF INCLUSION MATRICES
We are ready to state the main result of this section.
Theorem 8.7. Let G be a subgroup of the group of automorphisms of the poset (P, ≤).
Let X and Y be two distinct levels of P . Let r and s be the number of G-orbits in X and
Y , resp. Assume that the incidence matrix IP (X, Y ) has full rank. Then r ≤ s if and only if
|X| ≤ |Y |. Moreover, if this is the case, then there exists a one-to-one map ψ : X/G → Y /G
such that for every U ∈ X/G, we have U ≤ ψ(U ) in the ordering of P/G.
Theorem 8.6 is now an immediate consequence. Indeed, let P = 2[n] , with set inclusion
being the partial order.
G induces a group of automorphisms of (P, ≤). Take two neigh-
[n] [n]


boring levels: X = k−1 and Y = k ; assume k ≤ (n + 1)/2. By Gottlieb’s Theorem
n
≤ nk , it fol-



(Theorem 7.16), the incidence matrix IP (X, Y ) has full rank. Since k−1
lows that fG (k − 1) ≤ fG (k). For k ≥ (n − 1)/2, the inequality will go the other way by
symmetry.
The proof of Theorem 8.7 will follow from a matrix partitioning lemma, valid over an arbi-
trary field F.
We call a matrix A ∈ Fk×n column-regular if its columns have equal sum. We call A
row-regular if AT is column-regular; and biregular if both conditions hold. P
r
Ps Let r, s, ki , nj (1 ≤ i ≤ r, 1 ≤ j ≤ s) be positive integers such

that i=1 ki = k, and
i=1 ni = n. Then the matrix A can be partitioned as A = Aij into a r × s hypermatrix
(matrix with matrix entries), where Aij ∈ Fki ×nj . We call this partition column-regular if
each matrix Aij is column-regular.
Let A be partitioned as above. Let βij denote the sum of the entries of Aij . We call the
matrix B = (βij ) ∈ Fr×s the factor of A by the given partition.
Lemma 8.8. Let A = (Aij ) be a column-regular partition of the matrix A, with factor B.
If A has full row rank then B has full row rank, too.
Proof. Consider the intermediate partition into r ×n submatrices of dimensions ki ×1. Let C
be the corresponding factor matrix. It is clear that each row of C is a sum of the correspond-
ing rows of A. So any nontrivial linear relation between the rows of C would expand to a
similar relation between the rows of A. But the rows of A are linearly independent, hence so
are the rows of C. So far we did not need the column-regularity assumption. We use it now,
with the consequence, that C has at most r different columns: each column corresponding
to the same block in the partition is equal. Throwing all but one of them away in each block
of columns will not alter the rank, and will result in an r × s matrix D which still has full
row rank. Now multiplying the j th column of D by nj , we obtain B, which therefore has the
same rank as D, i. e., full row rank.
Proof of Theorem 8.7. Let X1 , . . . , Xr and Y1 , . . . , Ys be the G-orbits on X and on Y , resp.
Partition the incidence matrix IP (X, Y ) correspondingly. It is clear that this is a biregular
partition. Assume |X| ≤ |Y |. Then, by assumption, IP (X, Y ) has full row rank. Therefore
the r × s factor matrix B has full row rank, too. In particular, we have r ≤ s.
In order to obtain the required one-to-one map Qψ : X → Y , select a nonsingular r × r
minor in B, and find a nonzero expansion term ri=1 biji of det(B). The correspondence
ψ : i 7→ ji is clearly right.

176 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 8.3. UNIMODALITY, SPERNER PROPERTY
In the case |X| ≥ |Y |, we apply the same argument to the transpose of the incidence
matrix.
Theorem 8.7 has further important consequences. Let fi (P ) denote the size of the ith level
of P . A poset is unimodal if the sequence f0 (P ), f1 (P ), . . . is unimodal. The following is a
far-reaching generalization of Sperner’s Theorem (Section 4.4).
Corollary 8.9. Let (P, ≤) be a unimodal poset and G a group of automorphisms of P . If
the incidence matrices between each pair of neighboring levels of P have full rank, then P/G
is unimodal and has the Sperner property.
To see that this result includes Sperner’s Theorem, we take, as before, (P, ≤) := (2[n] , ⊆).
We ignore G (let G be the group consisting of the identity only). Gottlieb’s Theorem
guarantees that the rank condition is met, and the Sperner property for this poset, stated
by Sperner’s Theorem, follows.
Proof. Let L0 , L1 , . . . , Lh be the levels of P and M0 , M1 , . . . , Mh the levels of Q := P/G. Let
Lk have maximal size among the levels of P . Unimodality of Q with peak at Mk immediately
follows by an application of Theorem 8.7 to all pairs of neighboring levels of P .
Theorem 8.7 also guarantees the existence of a family of functions ψi , i = 0, . . . , h − 1,
such that for i ≤ k − 1, ψi is a one-to-one map Mi → Mi+1 , and for i ≥ k it is a one-to-
one map Mi+1 → Mi , such that x and ψi (x) are always comparable. Let us combine these
functions to a single function ψ : Q → Q defined by ψ(x) = x for x ∈ Mk and ψ(x) = ψi (x)
otherwise, for the unique i such that x is in the domain of ψi . S For each u ∈ Mk , define
C(u) = {x ∈ Q : ψ h (x) = u}. It is clear that C(u) is a chain and u∈Mk C(u) = Q.
Summarizing, we managed to divide Q into |Mk | chains. Clearly, any antichain A can
have at most one element from each of these chains, hence |A| ≤ |Mk |. This proves the
Sperner property of Q.

Exercises
Ex. 8.3.1. Let i(n, m) denote the number of isomorphism-types of graphs with n vertices and m
n


edges. Prove that for every n, the sequence i(n, 0), i(n, 1), . . . , i(n, 2 ) is unimodal.

——————————————————————— 177

c László Babai and Péter Frankl. 1988, 1992, 2020.
178 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
Chapter 9

Partially ordered sets

: CHAPTER TO BE WRITTEN; ALL TEXT BELOW TENTA-

TIVE AND INCOMPLETE
In this chapter we show that much of the material we have learned so far can be adapted
context of astonishing generality. After all the down-to-Earth set-system theory of the
previous chapters, the unprepared reader should brace herself1 for a “culture shock” in the
high-flight universe of partially ordered sets. In this world, familiar binomial coefficients are
replaced not even by their q-analogues, the Gaussian binomial coefficients, but remote and
sterile “Whitney numbers.” Yet, the new form of the old results seems perfectly streamlined,
the essence and unexpectedly broad scope of some of the old concepts become transparent.

9.1 Geometric semilattices

9.1.1 Matroids
Matroids, coordinatization, graphic matroids

9.1.2 Geometric lattices

9.1.3 RW-type theorems for semilattices
q-analogues of the RW Theorem
Equicardinal geometric semilattices
(Alon–Babai–Suzuki, 1991)

1
In the authors’ native Hungarian, pronouns have no gender, “himself” and “herself” is the same word.
Unfortunately this piece of grammatical wisdom of a small language community has so far had zero impact
on the debates of Indo-Europeans, in spite of the immense headaches continually facing “politically correct”
writers from Doctor Spock to the authors of bylaws of professional associations.

179
CHAPTER 9. PARTIALLY ORDERED SETS
9.2 Incidence matrices of full rank
State Kantor’s Theorem: q-analog of Gottlieb’s
Greatly amplifies the scope of the methods of Section 8.3.

*********** tentative text ************

Definition 9.1. Let P be a ranked poset, and let 0 ≤ k ≤ `. Define the wk × w` matrix
IP (k, `) = (ζab ) to be the incidence matrix of rank k elements versus rank ` elements in L,
given by ζab = 1 if a ≤ b, and 0 otherwise, where rk(a) = k and rk(b) = `.

Definition 9.2. A ranked semi-lattice is rank-regular if for each i, j, k, `, there exists a

number sijk` such that for every a and b, if rk(a) = rk(b) = k, and rk(a ∧ b) = i then we have

#{ c | rk(c) = `, c ≥ a, rk(b ∧ c) = j } = sijk` .

We note that a certain degree of symmetry implies rank-regularity.

Proposition 9.3. Assume the ranked semi-lattice L admits a group G of automorphisms
with the following transitive property: for all k and i, G is transitive on the set of pairs
{(a, b) | rk(a) = rk(b) = k, rk(a ∧ b) = i}. Then L is rank-regular. In particular, the Boolean
lattice and the lattice of subspaces of a vector space over a finite field are rank-regular.
The following general result includes both Gottlieb’s Theorem (Theorem 7.16) and Kan-
tor’s (Theorem 9.2) (see Cor. 9.5). It was formulated by B. Guiduli; the proof generalizes
the approach of Graver and Jurkat (1973) to this much more general setting.
Theorem 9.4 (Guiduli, 1992). Let L be a rank-regular semi-lattice and fix 0 ≤ k ≤ `.
Let sij = sijk` and assume that sii > 0 for all i = 0, . . . , k. Then the matrix IL (k, `) has full
row rank.
Proof. Let M = IL (k, `). We construct a w` × wk matrix N , such that M N = Iwk , thus
showing that M has full row rank. Let N = (nab ) where rk(a) = `, and rk(b) = k. We need
to show that there exists a matrix N satisfying:

(M N )ab = δa,b .

We are looking for a solution satisfying the additional constraint that nab depends only on
rk(a ∧ b): nab = trk(a∧b) . Let i = rk(a ∧ b). We have:

X k
X X `
X
(M N )ab = mac ncb = tj mac = sij tj .
c j=0 c j=0
rk(c) = ` rk(c) = `
rk(c ∧ b) = j

180 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 9.3. THE MÖBIUS FUNCTION
We need to show the existence of a solution to the following system of k + 1 linear equations
in the k + 1 unknowns t0 , . . . , tk :
k
X
sij ti = δi,k .
i=0

If i > j then sij = 0, and if i = j, then sij > 0, by hypothesis. This shows that S = (sij )
is an upper triangular matrix, with nonzero entries on the diagonal, thus a solution exists,
showing that M has full row rank.

Corollary 9.5. (1) (Gottlieb’s Theorem.) Let L be the Boolean lattice for a set of size
n. If k + ` ≤ n, then IL (k, `) has full row rank. (2) (Kantor’s Theorem.) Let L be the
lattice of subspaces of a vector space over a finite field. If k + ` ≤ n, then IL (k, `) has full
row rank.
Proof. In both cases, L is rank-regular (by Prop. 9.3) and siik` > 0 for i = 0, . . . , k.

Exercises
     
n n n
Ex. 9.2.1. Prove that the sequence of Gaussian binomial coefficients , , . . .,
0 q
1 q
n q
is unimodal.

Ex. 9.2.2. Prove that the lattice of subspaces of a finite linear space has the Sperner property.
Ex. 9.2.3. Define action of GL(n, q) on Fqn×n by conjugation. Prove that the lattice of orbits of
subspaces has the Sperner property.

9.3 The Möbius function

Much of the groundwork on the Möbius function is due to Gian Carlo Rota (1964). The first
two sections of this chapter follow his work.

9.3.1 Möbius inversion

Let (P, ≤) be a finite poset. Define the matrix Z = (ζ(x, y)) by
(
1 if x ≤ y
ζ(x, y) = (1)
0 otherwise.

Let f : P → C be any function. Then we may define the function g : P → C of partial

sums of f : X
g(y) = f (x) (2)
x≤y

——————————————————————— 181

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 9. PARTIALLY ORDERED SETS
In some cases we may want to invert (2) and recover f from g. To see how to do this, observe
that (2) may be written in matrix form as

g = f Z. (3)

The following is a consequence of the exercises at the end of this section:

Proposition 9.6. Z has an inverse M = (µ(x, y)), with the property that µ(x, y) depends
only on the structure of the interval [x, y]. In particular, µ(x, y) = 0 unless x ≤ y.
Therefore, X
f (y) = g(x)µ(x, y). (4)
x≤y

The function µ is the Möbius function, and (4) is the Möbius inversion formula.

Exercises
Ex. 9.3.1. Show that Z is invertible.
Hint. With an appropriate ordering of the basis, Z is upper triangular with 1’s on the diagonal.

Ex. 9.3.2. Fix P

x, y ∈ P. For k ≥ 0 let pk denote the number of chains x = a0 < a1 < . . . < ak = y.
Then µ(x, y) = k≥0 (−1)k pk . Deduce Proposition 9.6.
Hint. Let N = Z − I. N is nilpotent, so Z −1 = (I + N )−1 = I − N + N 2 − . . .. What is the x, y
entry of N k ?

Ex. 9.3.3. Show that µ(x, x) = 1, and for x 6= y,

X
µ(x, y) = − µ(x, a). (5)
x≤a<y

Hint. Look at the x, y entry in the matrix equation M Z = I.

Ex. 9.3.4. Let (P, ≤P ) and (Q, ≤Q ) be posets. Define the direct product poset P × Q, ≤P×Q
by (x1 , x2 ) ≤P×Q (y1 , y2 ) iff x1 ≤P y1 and x2 ≤Q y2 . Show that µP×Q ((x1 , x2 ), (y1 , y2 )) =
µP (x1 , y1 )µQ (x2 , y2 ).
Hint. ZP×Q = ZP ⊗ ZQ .

Ex. 9.3.5. Let (P, ≤) be a poset. Define the opposite poset (P op , ≤op ) by reversing the order:
x ≤op y iff y ≤ x. Show that µop (x, y) = µ(y, x).
Hint. Z op = Z T .

9.3.2 The Möbius function in geometric lattices

We have seen that the Möbius function µ(x, y) vanishes unless x ≤ y. The converse, however,
is not true. It is possible that x ≤ y but µ(x, y) = 0 (consider µ(0, 1) in a chain of length 2).

182 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 9.3. THE MÖBIUS FUNCTION
In this section we will show that the converse does indeed hold if the poset is a geometric
lattice.
Theorem 9.7 (Rota 1964). Let L be a geometric lattice. Then for any x, y ∈ L with
x ≤ y, µ(x, y) is nonzero and has sign (−1)rk(x)−rk(y) .
Lemma 9.8. Let P, Q be posets with 0. Let p : P → Q satisfy
1. For all x1 , x2 ∈ P with x1 ≤ x2 , p(x1 ) ≤ p(x2 ).
2. For any y ∈ Q there is an x ∈ P such that the inverse image of [0, y] is [0, x].
3. The inverse image of 0 contains at least two points.
Then for any y ∈ Q, X
µ(0, x) = 0
x
p(x)=y

Proof. Note that {x : p(x) ≤ y} = [0, r] for some r 6= 0 ∈ P. By equation (5) we have:
X
µ(0, x) = 0 (6)
x
p(x)≤y

We proceed by induction on y. If y = 0 then

X X
µ(0, x) = µ(0, x) = 0. (7)
x x
p(x)=0 p(x)≤0

Now suppose that the statement is true for z < y. Then

X X X X
0= µ(0, x) = µ(0, x) = µ(0, x) (8)
x z≤y x x
p(x)≤y p(x)=z p(x)=y

as desired.
Corollary 9.9. Let L be a lattice. For any a, b ∈ L with a 6= 0,
X
µ(0, x) = 0
x
x∨a=b

Proof. Apply Lemma 9.8 with P = L, Q = [a, 1], and p(x) = x ∨ a.

Proof of Theorem 9.7. By considering the interval [x, 1] we may assume that x = 0. We
proceed by induction on y. If y = 0, then µ(0, y) = 1. If y is an atom, then µ(0, y) = −1.
Now suppose that the theorem is true for z < y. Let a < y be an atom. By Corollary 9.9
we have: X
µ(0, y) = − µ(0, z) (9)
z<y
z∨a=y

Let z < y be such that z ∨ a = y. Then z ∧ a = 0, and by the submodular inequality,

rk(z) ≥ rk(y) − 1. Therefore in all of the terms on the right hand side of (9), z is a dual
atom and µ(0, z) has sign (−1)rk(y)−1 .

——————————————————————— 183

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 9. PARTIALLY ORDERED SETS
Exercises
Ex. 9.3.6 (Wilf, 1968). Let L be a ∧-semilattice. Suppose the functions f and gQ
satisfy g = f Z.
Let A be the matrix (a(x, y)), where a(x, y) = g(x ∧ y). Show that det(A) = x∈L f (x), and
rk(A) = |{x ∈ L : f (x) 6= 0}|.
Hint. Let D = diag(f (x)). What is Z T DZ?

Ex. 9.3.7. Let L be a geometric lattice, and let A be the matrix (a(x, y)) where
(
1 if x ∧ y = 0
a(x, y) =
0 otherwise.

Show that A is nonsingular.

Hint. Use the previous exercise.

Ex. 9.3.8. Let L be a lattice. For a, b in L such that b ≤ a < 1 show:

X
µ(x, 1) = 0
x
x∧a=b

Hint. Mimic the proof of Corollary 9.9.

Ex. 9.3.9.* Use the previous exercise to calculate µ(0, 1) in the lattice of partitions of an n-element
set.

Ex. 9.3.10* (Graham-Pollak 1971). Let T be a tree with vertex set V = {x1 , x2 , . . . , xn }.
Let A be the matrix (a(x, y))x,y∈V where a(x, y) is the distance from x to y in T . Show that
det(A) = −(n − 1)(−2)n−2 .
Hint: Define a poset on V so that a(x, y) = f1 (x) + f2 (y) + f3 (x ∧ y). What is M T AM ? See also
(Lovasz 1979c), exercise 3.33.

9.3.3 Whitney number inequalities

In this section we show, roughly, that in a geometric lattice there are at least as many
elements of high rank as there are of low rank. Let L be a geometric lattice of rank r, and
for 0 ≤ k ≤ r let Wk denote the number of elements of L of rank k. The Wk are the Whitney
numbers of the second kind.
Theorem 9.10 (Dowling-Wilson (1975)). For any 1 ≤ k < r the Whitney numbers
satisfy
W1 + W2 + . . . + Wk ≤ Wr−k + . . . + Wr−2 + Wr−1 . (10)
Proof. Let D be the diagonal matrix diag(µ(x, 1)), and let A = ZDZ T . Then the entries of
A are given by (
1 if x ∨ y = 1
a(x, y) = (11)
0 otherwise.

184 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 9.3. THE MÖBIUS FUNCTION
Let P be the diagonal matrix
(
1 if rk(x) ≤ k
p(x, x) =
0 otherwise.

Pick any y ∈ L with rk(y) < r − k. By the submodular inequality if a(x, y) = 1 then
p(x, x) = 0 since x ∨ y = 1 implies rk(x) > k. Therefore, the y column of P A is zero, and

rk(P A) ≤ Wr−k + . . . + Wr−1 + Wr (12)

since the right hand side is the number of y for which the y column of P A could be nonzero.
On the other hand, A is nonsingular by Theorem 9.7, so

rk(P A) = rk(P ) = W0 + W1 + . . . Wk (13)

The Theorem follows by combining (12) and (13) and observing that W0 = Wr = 1.
Ex. 9.3.11. Show that if L is modular, then (10) holds with equality.
Hint. For y of rank ≥ r − k, what is the y column of P A?

Ex. 9.3.12.** Show conversely that if for some 1 ≤ k ≤ r − 2 equation (10) holds with equality
then L is modular.
Hint. See (Dowling–Wilson, 1975).

Ex. 9.3.13. Let L be a geometric lattice. Show that there exists a permutation π of L such that
for all x ∈ L we have x ∨ π(x) = 1.
Hint. Let A be as above, and use the fact that det(A) 6= 0.

Ex. 9.3.14. Deduce the Dowling-Wilson Theorem from the previous exercise.
Hint. Use the submodular inequality.

9.3.4 The VC dimension revisited: shattered elements in a poset

In this section we give a far reaching generalization of the Shatter Function Theorem, the
fundamental theorem of the theory the Vapnik-Chervonenkis dimension. (Theorem 7.22).
Let L be a meet semi-lattice with smallest element 0.
Definition 9.11. The trace of a subset F ⊆ L in x ∈ L is the set Tx (F) = {x ∧ y : y ∈ F}.
We say that x is shattered by F if for every z ≤ x, there exists a y ∈ F such that z = x ∧ y.
In other words, x is shattered if the trace of F in x is the entire interval [0, x].

Definition 9.12. The shatter function of F is the function σF : L → Z defined by

σF (x) = |Tx (F)| . (14)

——————————————————————— 185

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 9. PARTIALLY ORDERED SETS
The following result applies among others to all geometric semilattices. The special case
of the Boolean lattice includes the original Shatter Function Theorem, as more readily seen
from the Corollary below.
Theorem 9.13 (Generalized Shatter Function Theorem). Let L be a meet semi-
lattice with non-vanishing Möbius function, and let H ⊆ L. If F ⊆ L, and |F| > |H|, then
there exists an element x ∈ L \ H shattered by F.
Proof. Let I(H, F) = (ζ(x, y)) be the F vs. H incidence matrix, given by
(
1 x≤y
ζ(x, y) := (x ∈ H, y ∈ F). (15)
0 otherwise

The columns of I(H, F) must be linearly dependent because |F| > |H|, so let γ(x) be
coefficients (not all zero) such that
X
γ(x)ζ(x, y) = 0, for all y ∈ H. (16)
x∈F

For x ∈ L, define X
f (x) := γ(z). (17)
z∈F
x≤z

With this notation, (16) says that

f (x) = 0, whenever x ∈ H. (18)
On the other hand, f is not identically zero. Indeed, let x ∈ F be maximal such that
γ(x) 6= 0. Then the sum (17) has a single nonzero term, f (x) = γ(x) 6= 0.
For x, y ∈ L, set X
η(x, y) := γ(z). (19)
z∈F
z∧y=x

If η(x, y) 6= 0 then necessarily z ∧ y = x for some z ∈ F (otherwise the sum (19) would be
empty). So if η(x, y) 6= 0 for all x, x ≤ y, then y is shattered by F.
Since f is not identically zero, we may take a minimal u such that f (u) 6= 0. From
eqn. (18) we see that u 6∈ H. Our claim is that u is shattered by F . This now follows from
the claim below.
Claim. η(x, u) 6= 0 for all x, x ≤ u.
Proof of the Claim.
X X X
η(x, u) = γ(z) = γ(z) µ(x, y)
z∈F z∈F x≤y≤z∧u
z∧u=x
X X X
= µ(x, y) γ(z) = µ(x, y)f (y)
x≤y≤u z∈F x≤y≤u
y≤z
= µ(x, u)f (u) 6= 0.

186 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 9.3. THE MÖBIUS FUNCTION
The last equality follows by the minimality of u.
Definition 9.14. Let L be a ranked meet-semilattice. The VC-dimension of F ⊆ L is the
maximum rank such that some element of rank r is shattered by F.

Corollary 9.15. Let L be a ranked meet-semilattice with non-vanishing Möbius function

and Whitney numbers w0 , w1 , . . ., and let F ⊆ L. If the VC dimension of F is s then
|F| ≤ w0 + w1 + . . . ws , and more generally for every x ∈ L,

σF (x) ≤ w0 (x) + w1 (x) + . . . ws (x), (20)

where wi (x) is the ith Whitney number of the interval [0, x].

——————————————————————— 187

c László Babai and Péter Frankl. 1988, 1992, 2020.
CHAPTER 9. PARTIALLY ORDERED SETS

188 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
Chapter 10

Applications to the Theory of

Computing

TO BE WRITTEN

10.1 Communication complexity theory

10.2 Overview
– Representations of graphs – orthogonal, projective–affine: Pudlak–Rodl
– Matrix methods for lower bounds: Razborov
– Razborov–Smolensky lower bounds in ckt complexity ??
– derandomization

189
190 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
Answers to the exercises

A.1 Chapter 1
1.1.1. Consider equation (4) with rational λi . Assume, for a contradiction, that not all
the λi are zero. Multiply with the common denominator to obtain integer coefficients; then
divide by the g.c.d. of these coefficients. So we may assume (4) holds with integer coefficients
whose g.c.d. is 1. Now take the inner product of each side by v1 . By equation 1 it follows that
λ1 is even. A similar argument shows that all the λi are even, contradicting the assumption
that their g.c.d. is 1.
1.1.2. We have to prove that the rank of the incidence matrix is m. The rank of the
intersection matrix A is not greater than the rank of M (in fact now they are equal), so it
suffices to prove that A has rank m. This will follow if we prove that the determinant of A
is not zero. The value of this determinant is an integer. We claim it is odd. In order to see
this, let us have a look at A modulo 2. By conditions (a) and (b), the diagonal entries of A
are odd, all the other entries are even, therefore A is congruent modulo 2 to I, the identity
matrix, and det A ≡ det I = 1 mod 2.
1.1.3. Let a devoted new membership collector set up residence in the town and immediately
join all existing clubs. In addition, let him form a new one-member club. Apply the Oddtown
Theorem to the resulting situation.
1.1.4. (b) The determinant is odd if m is even. (c) If m is odd, the determinant is even,
therefore rk2 (Jm − Im ) ≤ m − 1. On the other hand, Jm − Im has Jm−1 − Im−1 as a minor,
and by part (b), this minor has full rank.
1.1.5. To obtain this many clubs, form n−1 clubs of two members each such that the mayor
belongs to each club. For odd n, add one club of size n − 1 that includes everyone but the
mayor.
In order to prove that it is impossible to form more clubs under the given rules, we
observe that Exercise 1.1.3 says just this if n is odd. We may therefore assume that n is
even. We have to work a little harder to shave one off the easy upper bound m ≤ n. Assume
for a contradiction that m = n.
The intersection matrix A = M M T will now have even numbers in the diagonal and odd
numbers elsewhere. By part (b) of the preceding exercise, A has full rank m over F2 (because

191
ANSWERS TO THE EXERCISES
now m = n is even); therefore rk2 M = m, i. e., the rows of M are linearly independent over
F2 .
Pthe incidence vectors belong to the (n − 1)-
On the other hand, since all clubs are even,
dimensional subspace defined by the equation ni=1 xi = 0 over F2 . Consequently, m ≤ n−1,
contradicting our assumption that m = n.
1.1.6. Show that det A is odd by looking at A modulo 2. Use Ex. 1.1.4.
1.1.11. Distribute all citizens into n − 2t disjoint clubs.
1.1.12. Construct 3 clubs with a total of 7 members such that each club has an even number
of members, their pairwise intersections are also even, but the intersection of all the three is
odd. Show that such a system cannot be part of the “married couples” solution described in
the main text. Use Exercise 1.1.10 to extend this system of 3 clubs to an extremal solution.
1.1.13. Add one point to the projective plane over F5 , the field of 5 elements. Add this
point to each line to obtain the 31 small clubs. The plane itself will be the large club.
1.1.14. Let n = 2k. From a k × k (0, 1)-matrix A ∈ Fk×k
2 , construct an n × n matrix B as
shown here:  
A + Ik A
B= ∈ Fn×n
2 .
A A + Ik
If A is symmetric (A = AT ), it is easy to see that BB T = In . This means B is an Oddtown
k+1
incidence matrix. The number of symmetric k ×k matrices A is 2( 2 ) = 2n(n+2)/8 . This is an
overestimate of the number of club systems obtained, but not by too much. Permutations of
the rows correspond to renaming clubs; permutations of the columns mean renaming citizens
(isomorphism). Thus dividing by a factor of (n!)2 , we obtain a lower bound.
2n

In order
n2
to obtain an upper bound, we note that there are altogether no more than
n
< 2 /n! ways to form n clubs in a town of n citizens.
1.1.15. Form an m × n matrix whose rows are the vi . Use the determinant characterization
of the rank. The result will not depend on what extension field of the rationals has been
chosen.
1.1.17. Consider the n × n matrix with aii = ai,i+1 = 1 for i = 1, . . . , n − 1; zero elsewhere
in the first n − 1 rows, and indeterminates u1 , . . . , un in the last row:
 
1 1

 1 1 


 1 1 


 · 


 · 


 · 

 1 1 
u1 u2 u3 · · · un−1 un
The determinant of this matrix is un −un−1 +− · · ·+(−1)n−1 u1 . (Prove!) Now substitute
zeros and ones for the ui to obtain the desired value.

192 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 A.1. CHAPTER 1
1.1.18. Let m = n and make the determinant divisible by 10 but not by 3. (Use the
preceding exercise.)
1.1.25. First construct 4 clubs in a town of 6 citizens such that each citizen is a member
of precisely two clubs and no two citizens belong to the same pair of clubs. The clubs
constructed will have 3 members each, and each pair will share 1 member. Add 3 new
citizens to the town, and let them join each of the four clubs. Now we have four clubs in a
town of 9 citizens, each club has 6 members, and each pair of clubs shares 4 members. The
figure shows the incidence matrix of this system of clubs.
 
1 1 1 0 0 0 1 1 1
 1 0 0 1 1 0 1 1 1 
 
 0 1 0 1 0 1 1 1 1 
0 0 1 0 1 1 1 1 1

Figure A.1: Four Mod-4-town clubs which are dependent mod 2.

Nine is the smallest population and four the smallest number of clubs that can realize
the conditions required in the Exercise. (Prove!)
1.1.27. Divide the set of clubs into c(s) classes according to which maximal prime power
divisor of s does not divide the size of the club.
1.2.3. Consider the incidence vectors of
all 2-subsets of a set of n + 1 elements. They form√
a 2-distance set of cardinality m = n+1 2
= n(n + 1)/2 in Rn+1 . (The two distances are 2
and 2. (Prove!) Which oneP occurs more frequently?) Actually, this set is on the hyperplane
defined by the equation xi = 2 and therefore it can be viewed as a subset of Rn .
1.2.6. Assume for a contradiction that det A = 0. Then, for some nonzero x = (x1 , . . . , xm ) ∈
Rm , we have AxT = 0. Let |xi | = max{|x1 | , . . . , |xm |}. Now,
Xm 
0 =  aij xj  ≥ |aii xi |
 
j=1
X X
− |aij xj | ≥ |xi | (aii − |aij |) > 0,
j6=i j6=i

a contradiction.
1.2.7. Assume all distances are between δ1 −ε and δ1 +ε or between δ2 −ε and δ2 +ε. Consider
the same functions F (x, y) and fi defined in the proof of Theorem 1.3. Prove that the matrix
(F (ai , aj ))m 2 2
i,j=1 is nonsingular. The diagonal entries of this matrix are approximately δ1 δ2 ; the
off-diagonal entries are bounded in absolute value by approximately 2δ1 (δ22 − δ12 )ε (assuming
δ1 < δ2 ). Therefore, for sufficiently small ε, the matrix will be diagonally dominated and the
result of the preceding exercise applies. (It suffices to assume ε ≤ δ1 /(3n2 ).)
1.2.8. The radius of the sphere clearly does not matter. For the lower bound consider
(1, −1)-vectors in Rn+1 with exactly two negative entries. Proceed like in the solution of

——————————————————————— 193

c László Babai and Péter Frankl. 1988, 1992, 2020.
ANSWERS TO THE EXERCISES
Exercise 1.2.3, noting that the intersection of a sphere and a hyperplane is a lower dimensional
sphere.
For the upper bound, proceed like in the main text, restricting the domain of the functions
fi to the unit sphere. The functions will remain linearly independent (as members of the
space of SS n−1 → R functions), but we shall find that they reside in a lower dimensional
space than before. Indeed, of the generators listed under P (10), the first two types are now
covered by the third and the last types, resp., since now nk=1 x2i = 1. Also, we may drop
n−1
x2n from the list since x2n = 1 − k=1 x2i . The new list is thus
P

xi xj (1 ≤ i < j ≤ n), x2i (1 ≤ i ≤ n − 1), xi (1 ≤ i ≤ n), 1. (1)

The tally is n(n − 1)/2 + (n − 1) + n + 1 = n(n + 3)/2.

Note. There are three values n ≥ 2 where the upper bound ms (n) ≤ n(n + 3)/2 for spherical
two-distance sets is known to be tight. These are the pentagon (n = 2, m = 5), the so-called
6-dimensional Gosset polytope (n = 6, m = 27) (see Exercise 1.2.10), and a set of m = 275
points in dimension n = 22.
1.2.10. (a1) The count is 2 82 = 56. (a2) All points given lie in the hyperplane 8i=1 xi = 0.

P
√ √
(a3)√The distance of each point from the origin is 24. (a4) Only 3 distances occur: 4 2,
8, 4 6. Note that the largest distance only occurs between antipodal pairs (v, −v). (a5)
Only the first two of these distances occur between the points with six positive coordinates.
(b) Take the hyperplane x1 + x2 = 2. This reduces the dimension to six and restricts the
first two coordinates to (1, 1), (3, −1), or (−1, 3). The number of corresponding
√ points is
15+12+12=27. No antipodal pairs occur in this set, so the distances are 4 2 and 8.
1.2.12. Let A ⊂ Rn be a 2-distance set. Take a very large sphere in Rn+1 , tangent near
the points of A to the hyperplane Rn containing A. Project A to the sphere from its center.
The projection will be a spherical approximate 2-distance set in Rn+1 and therefore have at
most (n + 1)(n + 4)/2 points.
1.2.13. Represent each set Ai by its (1, −1)-incidence vector vi (replace zeros by −1’s in √the
usual definition of incidence vectors). All these vectors belong now to the sphere of radius n
in Rn . The distance kvi − vj k is determined by the size of the symmetric difference Ai ⊕ Aj .
(How?) Therefore the bound of Exercise 1.2.8 for spherical two-distance sets applies.
1.2.14. Combine the solutions of the preceding exercise and Exercise 1.2.8 (spherical 2-
distance bound). Restrict the domain of the functions fi to the set Ω = {1, −1}n (the set of
(1, −1)-vectors in Rn ). The functions fi : Ω → R remain linearly independent. On the other
hand, they are now generated by an even smaller number of functions. Taking the equalities
x2i = 1 into account, all that remains from the list (10) is

xi xj (1 ≤ i < j ≤ n), xi (1 ≤ i ≤ n), 1.

The tally is n(n − 1)/2 + n + 1 = 1 + n(n + 1)/2.

1.2.18. No. Let A = {(x1 , . . . , x4 ) : x21 + x22 = 1, x3 = x4 = 0} and B = {(x1 , . . . , x4 ) : x1 =
x2 = 0, x23 + x24 = 1}.

194 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 A.1. CHAPTER 1
1.2.21. The lower bound is a straightforward generalization of Exercise 1.2.3. For the upper
bound, let {δ1 , . . . , δs } be the distances permitted. Following the main text, we define the
polynomial F (x, y) (x, y ∈ Rn ) by
s
Y
F (x, y) := (kx − yk2 − δk2 ). (2)
k=1

We observe that the zero-nonzero alternative in (8) still holds. We set fi (x) := F (x, ai ) and
conclude, as before, that the fi are linearly independent over R.
What remains is to Pestimate from above the dimension of the subspace containing all
n 2
the fi . Let us set z = i=1 xi . Let us expand the norm-square expressions in each factor in
(12) and collect the squares into a single term z. Let us multiply the constant terms in each
factor by another new variable t. This way (12) becomes a product of s homogeneous linear
forms
in n + 2 variables. The number of degree s monomials formed from n + 2 variables is
n+s+1
s
(see Ex. 2.1.2), the desired upper bound.
1.2.22. Let S = {a1 , . . . , am }, and let C be the convex hull of S. We may assume that
a1 = 0. Let Ci = C + ai be the translation of C by the vector ai . (So C1 = C.) Then the Ci
are pairwise internally disjoint (have no common interior points). (Prove!) Let now D = 2C
(enlarge C by a factor of 2 by a homothety with center a1 = 0). Show that D contains all
the Ci . A comparison of the volumes shows that m ≤ 2n . (Where did we use that S was
nondegenerate?) For full details, see Boltyanski–Gohberg (1985), Section 18.)
1.2.23. Let N denote the unit ball of the given norm, so N = −N is a bounded closed
convex set, centrally symmetric about the origin. Let S ⊂ Rn be a set with all pairwise
distances equal to 1 in this norm. We may assume S is nondegenerate (aff(S) = Rn ). We
shall prove that S is a Klee-set, thereby reducing the problem to the preceding exercise.
Let a, b ∈ S. We have to construct two parallel support planes of S through a and b,
resp. We may assume a + b = 0. Let us consider N 0 = (1/2)N (reduced copy of N ). Clearly,
both a and b are on the boundary of N . Let U be a support hyperplane of N 0 through a,
and let V = −U be its reflection in the origin. So V is a support plane of N 0 through b,
parallel to U .
We claim that all of S is between the hyperplanes U and V , proving that they are support
hyperplanes to S. To this end, let us blow up N 0 by a homothety with center a and coefficient
2. The set M obtained is clearly the unit ball with center b, hence S ⊂ M . On the other
hand, it is clear that M lies on one side of U since U was transformed into itself by this
homothety.
1.2.25. We have to prove that the rank of B is n + 1. This follows from the fact that the
rows of B span Rn+1 , a fact that is easily seen to be equivalent to the condition that the
affine hull of the rows of A is Rn .
1.2.26. If det(B T B) = 0 then B T Bx = 0 for some nonzero Pm column vector x. Let Bx =
T T T T 2
y = (α1 , . . . , αm ) . Consequently, 0 = x B Bx = y y = i=1 αi . We conclude that y = 0,
which means the columns of B are linearly dependent, contrary assumption.
1.2.27. Let ai = (αi1 , . . . , αin ) (i = 1, . . . , m). Set αi0 = 1; let B be the m × (n + 1) matrix
B = (αij )1≤i≤m,0≤j≤n . We may assume the affine hull of the ai is Rn (otherwise all the points

——————————————————————— 195

c László Babai and Péter Frankl. 1988, 1992, 2020.
ANSWERS TO THE EXERCISES
belong to a lower dimensional space). Now a combination of the preceding two exercises
imply that the (n + 1) × (n + 1) matrix B T B is nonsingular.
Assume u = (λ1 , . . . , λm ) ∈ Rm and w = (µ0 , . . . , µn ) ∈ Rn+1 satisfy
m
X n
X
λi f i + µ0 + µj xj = 0. (3)
i=1 j=1

We have to prove that u = 0 and w = 0. In order to eliminate the variables µj , take, for
every i, j (1 ≤ i < j ≤ m), the second partial derivative ∂ij = ∂ 2 /∂xi ∂xj of (3). The result,
divided by 8, is:
m
X m
X m
X m
X
xi xj λ k − xi λk αkj − xj λk αki + λk αki αkj = 0. (4)
k=1 k=1 k=1 k=1

This being an identity, the coefficient of each term must vanish. The fact that the coefficients
of the first two terms in (4) vanish for every j can be condensed into the matrix equation

uB = 0. (5)

Let us now employ the trick used in the main text: substitute aj for (x1 , . . . , xn ) in (3).
Almost every λi disappears, and we obtain the equation
n
X
2
(δ1 δ2 ) λj + µ0 + µt αjt = 0. (6)
t=1

The fact that this equation holds for every j can be condensed into the matrix equation

(δ1 δ2 )2 u + wB T = 0. (7)

Multiplying by B on the right we obtain, in view of (5):

wB T B = 0. (8)

This is a nonsingular system of n + 1 homogeneous linear equations in the n + 1 variables

µ0 , . . . , µn . Hence w = 0. From this and (7) we infer that u = 0 as well. The proof is
complete.
1.3.1. Solution 1. 1. First assume the weights are integers. Since the condition does not
change if we subtract the same number from each weight, we may assume one of the weights
is zero. If all of them are zero, we are done. Otherwise we may assume that the g.c.d. of
the weights is 1; in particular at least one of them is odd. We shall show this is impossible.
Let us now consider the conditions mod 2 and work over F2 . Let the weights (mod 2)
be w1 , . . . , w13 . The condition implies that the sum of any twelve of them is 0. Since one
of the weights, say the last one, is w13 = 0, it follows that the sum of any 11 of the weights
w1 , . . . , w12 is also 0. This is a system of 12 homogeneous equations in the 12 unknowns.

196 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 A.1. CHAPTER 1
The matrix of the system is J12 − I12 , nonsingular by Ex. 1.1.4. This means the system has
no nontrivial solution: w1 = . . . = w13 = 0.
Translating back to integers, this means that all weights must be even. This contradicts
our g.c.d. assumption.
2. To reduce the problem with rational weights to integral, multiply each weight by their
least common denominator.
3. Assume now that the weights belong to some linear space over Q; let V be the span of
the weights. So V has finite dimension d ≤ 13. Select a basis in V and represent the weights
by their coordinates as d-tuples of rationals. Now the condition implies that for every i
(1 ≤ i ≤ d), the ith coordinates of the weights form a set of admissible weights themselves,
and are therefore equal. This being true for each i, all vectors must be equal.
4. The general case follows, observing that R is a linear space over Q.
Solution 2. Let w1 , . . . , w13 ∈ R be the weights. As before, we may assume one of the weights
is zero, say w13 = 0. Now we have 12 homogeneous linear equations for the remaining 12
weights. The matrix of this system is of the kind described in Ex. 1.1.6. Therefore this
matrix is nonsingular and the system has no nontrivial solution: w1 = . . . = w12 = 0.
1.3.2.
P Let xu = µ({u}) denote the (unknown) measure of u ∈ X. Then by additivity,
u∈E u = µ0 (E) for every E ∈ F. This system of linear equations in the xu (u ∈ X) must
x
have a solution for any choice of the right hand side. For this to happen, it is necessary and
sufficient for the left hand sides of the equations to be linearly independent (cf. Ex. 2.2.1).
1.3.3. The “hard part” is solved by Figure A.2. Here we assume that the ratio of the sides
AB and BC is of the rectangle is less than two; AP QR is a square. The condition on the
ratios ensures that the segment BR intersects the segments DZ and P Z. The two shaded
right triangles are congruent, and so are the larger, overlapping right triangles with their right
angles at Q and C, resp. Three pieces thus suffice to demonstrate the equidecomposability of
ABCD and AP QR (the pentagonal region AP Y XD plus one of each kind of right triangle
described). For more details, see Boltyanski (1978), pp. 49–56.

Figure A.2: Squaring a rectangle by dissection.

1.3.4. 1. First turn Hill’s tetrahedron H into a right triangular prism based on the same
equilateral right triangle as the bsae of H. (So it will be 1/3 as tall as H.) 4 pieces suffice (see
Boltyanski (1978), p. 99). Then, using the method of Bolyai and Gerwien (Ex. 1.3.3), use
cuts perpendicular to the base of the prism to turn it into a right prism with a rectangular

——————————————————————— 197

c László Babai and Péter Frankl. 1988, 1992, 2020.
ANSWERS TO THE EXERCISES
base such that one of the sides of the base be of the final length (third root of the volume
of H, i. e., (1/6)1/3 ). Finally, turn the side perpendicular to this edge into a square by the
same method.
2. To prove that G is not equidissectible with the cube, use the method of √ the proof
of Dehn’s Theorem. Note that the dihedral angles of G are π/2 and arccos(1/ 3). Use
Exercise 1.3.12 to construct the requisite function f for a Dehn invariant that will tell G and
the cube apart.
1.3.5. Use the center of the inscribed circle to form three deltoids. (Deltoids are quadrilat-
erals with a symmetry axis.)

Figure A.3: A good cut for the pastry chef.

1.3.6. First dissect P into tetrahedra. Consider a tetrahedron T and the center O of its
inscribed sphere. With each edge e of T , associate the polyhedron spanned by the following
five points: O, its projections to the faces adjacent to e, and the two ends of e. Show that the
six polyhedra obtained combine to T (without overlap), and each of them is symmetrical.
(The plane of symmetry passes through e and O.) (For more details, see Boltyanski (1978),
pp. 95-96.
1.3.8. Let ` be a line which cuts the plane inPtwo half, say the “upper” and the “lower” half.
For a polygon P , let us define Φ(P ) to be ±|ei |, where the summation extends over all
edges of P parallel to `; and the sign is positive if and only if P touches ei from the “upper”
side.
Φ is clearly invariant under translations; and it is easy to see that it satisfies eqn. (16).
We conclude that if P1 and P2 are translation-equidissectible then Φ(P1 ) = Φ(P2 ).
Furthermore, if P is centrally symmetric, then Φ(P ) vanishes for any choice of `. On the
other hand, if we choose ` parallel to one of the sides of the triangle, then the Φ value of
that triangle will not be zero.
1.3.10. To prove that cos(`α) is of the form stated in the hint, proceed by induction on `.
For ` = 0, we have cos 0 = 1/30 . For ` = 1, we have cos α = 1/3 by definition. For ` ≥ 2,
use the identity cos(β + γ) = 2 cos β cos γ − cos(β − γ). Set β = (` − 1)α, γ = α. By the
induction hypothesis applied to the two preceding terms, we obtain
2s1 1 s2 2s1 − 9s2
cos(`α) = 2 cos((` − 1)α) cos α − cos((` − 2)α) = · − = . (9)
3`−1 3 3`−2 3`
1.3.14. From the last equation in the hint, we obtain that
Z π
f (x) sin x = F (π) + F (0). (10)
0

198 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 A.1. CHAPTER 1
Now both F (π) and F (0) are integers. But for 0 < x < π,
π n an
0 < f (x) sin x < , (11)
n!
so that the integral (10) is positive, but smaller than π n+1 an /n!. The right hand side is
positive but tends to zero as n → ∞.
1.3.16. Let ak = a0 q k be a Fibonacci-type geometric progression. Then either a0 = 0 or
q k = q k−1 + q k−2 for all k ≥ 2. In particular, for k = 2 we have q 2 = q + 1. On the other
hand, this single
√ equation clearly implies all the others. The solution is ak = a0 φki where
φ1,2 = (1 ± 5)/2. Note that φ1 is the golden ratio.
1.3.26. As before, find the geometric progressions satisfying the given recurrence. They
correspond to the roots of the polynomial x3 − 2x2 + x − 2 = (x2 + 1)(x − 2). Therefore the
geometric progressions
√ ik , (−i)k , and 2k form a basis of the space of sequences of this type,
where i = −1. Infer the formula

ak = (−1 + 2i)ik − (1 + 2i)(−i)k + 2k+1 . (12)

Note that we needed complex numbers to express the general term of this integer sequence
0, 0, 10, 22, 36, 52, . . ..
1.4.1. (c) We may assume n0 = n. Let M be the m × n incidence matrix of the system of
clubs (cf. Section 1.1). Assume for a contradiction that m > n. This implies that the columns
of M are linearly dependent. (Remember,
Pn that the columns correspond to the citizens.) Let
u1 , . . . , un be the columns, and i=1 λi ui = 0 a nontrivial linear relation between them. This
means a number λi is assigned to citizen i such that the numbers in each club add up to
zero. Trade red hats for positive numbers, blue hats for negative ones. Let Y = {i : λi = 0}.
These are the citizens left without a hat so far. Note that at this moment, any club having
members outside Y have members with hats of each color. By our assumption n = n0 ,
everybody is a member of some club, so the number of clubs without hats is ≤ m − 1. By
the assumption that the system was critical, there exists a hat-assignment to the set Y such
that all clubs composed entirely from Y receive each color.
P P
1.4.2. Set uk = i∈Xk xi and vk = i∈Yk xi . Now the fact of decomposition is expressed by
the equation
X Xm
xi xj = uk vk .
i<j k=1
1 1
P 2 P 2
(uk + vk )2 −
P
Write the left hand side as 2
(( x i ) − ( x i )), the right hand side as 4
(
(uk − vk )2 ). Rearrange the equation such that x2i stands alone on the left hand side. It
P P
follows from Sylvester’s Law that if this sum is represented as a linear combination of squares
of linear forms with real coefficients then the number of positive terms in this representation
is at least n.
1.4.3. The adjacency matrix of the complete graph is Jn − In . Let Ck be the adjacency
matrix of the bipartite graph Bk . The combinatorial condition is now expressed by the

——————————————————————— 199

c László Babai and Péter Frankl. 1988, 1992, 2020.
ANSWERS TO THE EXERCISES
congruence m
P
k=1 Ck ≡ Jn − In mod 2. The rank over F2 of Jn − In is ≥ n − 1; the rank of
each Ci is 2.
1.4.5. Let us first split the set of n vertices into two nonempty parts and take the complete
bipartite graph with these parts. Removing the edges of this graph results in two disjoint
complete graphs. Procede recursively. Let F (n) denote the number of nonisomorphic de-
compositions obtained in this fashion. Prove that

F (1) = F (2) = F (3) = 1, F (4) = 2, F (5) = 3, F (6) = 6,

F (7) = 11, (8) = 23, F (9) = 46, F (10) = 98.

P
Prove that F (n) ≥ 1≤i<n/2 F (i)F (n − i). Use the first four terms of the right hand side for
an inductive proof of the inequality F (n) > 2n−4 .
1.4.6. We do it for n = 5; the pattern should be clear. Assign the vertices of K5 to the
following faces of the 4-cube (to be squashed): (1, ∗, ∗, ∗), (0, 1, ∗, ∗), (0, 0, 1, ∗), (0, 0, 0, 1),
(0, 0, 0, 0).
1.4.7. Let x, y ∈ S m . If d(x, y) = 1, there is exactly one position, say the j’th, where
{xj , yj } = {0, 1}. Let us say in this case that j is responsible for this distance.
Let now Xk denote the set of those vertices i of Kn which receive 0 in the kth entry of
their address, and Yk those which receive 1. Clearly Xk ∩ Yk = ∅. Let Hk be the complete
bipartite graph
with color classes (Xk , Yk ). We claim that the edge sets of H1 , . . . , Hm
[n]
partition 2 , the edge set of Kn . Indeed, the edge joining i and j belongs to Hk precisely
if the kth coordinate is responsible for the distance of i and j.
On the other hand, from a partition of the edge set of Kn into m complete bipartite
graphs, we obtain an addressing into the squashed m-cube by associating a position in the
strings with each bipartite constituent, and reversing the process described.

A.2 Chapter 2
2.1.2. Let x1 , . . . , xn denote the indeterminates. The monic monomials of degree k form a
basis of the space of homogeneous polynomials of degree k. A monic monomial has the form
xt11 xt22 · · · xtnn , where
t1 + t2 + · · · + tn = k, (ti ∈ Z, ti ≥ 0). (13)
The number of such monomials is the number of solutions (t1 , . . . , tn ) of (13). Let ui =
i + t1 + · · · + ti . Observing that un = n + k, it is easy to check that the correspondence
established by these equations between the solutions (t1 , . . . , tn ) of (7) and the (n − 1)-tuples
(u1 , . . . , un−1 ) satisfying the inequalities

1 ≤ u1 < u2 < · · · < un ≤ n + k − 1

n+k−1 n+k−1

is simply selecting an (n − 1)-
is one-to-one. But what choosing (u1 , . . . , un−1 )
amounts to
subset of {1, 2, . . . , n + k − 1}. There are n−1 = k
ways to do this.

200 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 A.2. CHAPTER 2
2.1.3. The monic multilinear monomials form a basis of this space. Such a monomial is just
the product of an arbitrary subset of the indeterminates.
2.1.41. The result is straightforward for k ≤ 1. Assume k ≥ 2.
We just have to prove (by Ex 2.1.33) that (Lk : Lk−1 ) = 2. Suppose the contrary. Then
√ √
for some α, β ∈ Lk−2 we have α + β pk−1 = pk .
√
If β = 0 then pk = α ∈ Lk−2 contradicting the induction hypothesis.
√
If α = 0 then pk pk−1 = βpk−1 ∈ Lk−2 again contradicting the induction hypothesis.
(Here we applied the induction hypothesis to the k − 1 pairwise relatively prime square-free
numbers p1 , . . . , pk−2 , pk−1 pk . Note that the inductive proof would not go through if the pi
were restricted to be primes; the generalization suggested in the Hint was crucial.)
Suppose finally that αβ 6= 0. Then
√
α2 + 2αβ pk−1 + pk−1 = pk
√
hence pk−1 ∈ Lk−2 , a contradiction again.
2.2.1. The condition means every b ∈ Fk must belong to the column space of A.
2.2.4. (a) Observe that the column space of the matrix A + B is a subspace of the sum of
the column spaces of A and B.
(b) Observe that the column space of AB is a subspace of the column space of A.
2.2.5. (F32 )× has seven elements, corresponding to seven 1-dimensional subspaces, the points
of the Fano plane. It also has seven lines consisting of 3 points each; they correspond to the
lines in the figure, including the circle. Specific coordinates are assigned to the points in the
figure below. (Find which line corresponds to each of the equations x2 = 0, x1 + x2 = 0.
What is the equation of the line represented by the circle?)

Figure A.4: The Fano plane, coordinatized

2.2.6. The three vertices of the trianle in the figure and the center of the circle are four
points in general position. Following the Hint, we can assume that these four vertices are
coordinatized as in Figure A.4. Now the equations of all lines except the “circle” can be
calculated; e. g., the equation of the line connecting (1, 0, 0) and (0, 0, 1) is x2 = 0; the
equation of the line connecting (0, 1, 0) and (1, 1, 1) is x1 − x3 = 0. Thus the intersection of
these two lines is the point (1, 0, 1). Similarly we recover the coordinates of all the remaining
points in Figure A.4. Finally, we have to check whether or not the three edge-midpoints are
collinear. This depends on whether or not their determinant is zero. But
 
 0 1 1 
 
 1 0 1 =1+1
 
 1 1 0 

——————————————————————— 201

c László Babai and Péter Frankl. 1988, 1992, 2020.
ANSWERS TO THE EXERCISES
which is zero if and only if the characteristic of F is 2.
2.3.12. The case n ≤ 6 is easy. Assume n ≥ 7. As in the previous solution, divide the
set S of incidence vectors into even and odd part: S0 and S1 . Let mi = |Si |. Now, S1
is linearly independent mod 2 (Oddtown Theorem), and S0 is linearly independent mod 3
(Mod-3-town). Therefore m0 , m1 ≤ n. Let us now view S as a subset of Fn2 , and let Ui be the
span of Si . Let ni = dim(Ui ). Again, U0 ⊥ U1 , and therefore U0 ∩ U1 = 0. (Prove!) Hence, as
in the previous solution, 2n0 + n1 ≤ n. Now m1 = n1 (since S1 is linearly independent), and
m0 ≤ 2n0 (trivially). We conclude that m1 = n1 ≤ n − 2n0 ≤ n − 2 log2 m0 , and therefore
m ≤ n + m0 − 2 log2 m0 ≤ 2n − 2 log2 n (since n ≥ 7).

A.3 Chapter 3
3.1.1. To obtain an (n+1)×n matrix whose rows are in general position, extend the identity
matrix In by an all-ones row.
3.1.2. The extra point is (0, . . . , 0, 1). (Verify that this point can indeed be added.)
3.1.5. Let S ⊂ Fnq be a set in general position. Take n − 2 vectors from S. They span
a subspace U of codimension 2. This subspace is contained in exactly q + 1 subspaces
of codimension 1. Each of those codimension 1 subspaces contain at most one additional
element of S.
3.1.8. g(3, 4) ≤ 3 + 4 − 1 = 6 by Exercise 3.1.4. The previous exercise implicitly suggests
a way to construct an appropriate 6 × 3 matrix A over F4 . The figure shows AT , using the
notation of Exercise 2.1.12 for the elements of F4 .
 
1 0 0 1 1 1
0 1 0 1 ω 1 + ω
0 0 1 1 1+ω ω

Figure A.5: Six vectors in general position in F34

3.1.10. By Exercise 3.1.1, g(n, q) ≥ n + 1. Assume now that A is an (n + 2) × n matrix over

Fq whose rows are in general position. As before, we may assume the top part of A is In . We
may also assume the (n + 1)st row of A is all ones. Now all elements of the last row must be
different, and none of them zero. (Verify!) This implies q − 1 ≥ n, contrary to assumption.
3.1.11. For a nonempty affine subspace S of W , let (S)0 denote the linear space of which it
is a translate. Clearly, ϕ is injective on S if and only if it is injective on (S)0 . Let us select ϕ
to be in general position with respect to the subspaces (Si )0 , where Si = aff{Ui , Vi }. If both
Ui and Vi are linear subspaces then Si is their sum and so dim(Si ) ≤ dim(Ui ) + dim(Vi )≤ t.
Otherwise, by Corollary 2.23, dim(Si ) ≤ 1 + dim(Ui ) + dim(Vi ) ≤ t.
3.2.1. Let a0 , . . . , an be an affine basis of Rn . The convex hull of the ai is a full-dimensional
simplex. For i = 0, . . . , n, let Ai = {aj : 1 ≤ j ≤ n, j 6= i} and Ci = conv(Ai ). The Ci are
the facets of the simplex. Now ai belongs to all the Cj except for Ci , so every n of the Cj

202 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 A.4. CHAPTER 4
intersect. On the other hand, there is no point shared by all the Ci . In fact, no point is
common even to the n+1 hyperplanes aff(Ai ). Indeed, assume x is such a point. Then x is an
affine combination of the ai . At least one of the ai , say a0 , appears with nonzero coefficient
in this combination. But since x ∈ aff(A0 ), we infer that x is also an affine combination of
just a1 , . . . , an . By subtracting this representation of x from the preceding one we obtain a
nontrivial linear relation among the ai with zero-sum coefficients. This contradicts the affine
independence of the ai .
3.2.7. For a contradiction assume SS d−1 = F1 ∪ . . . ∪ Fd where each Fi is closed and has
diameter less than 2. This means, no Fi contains a pair of antipodal points. For x ∈ SS d−1 ,
let fi (x) denote the distance between x and Fi . (This means the distance to the nearest
point of Fi .) Set f (x) = (f1 (x), . . . , fd−1 (x)); this is a continuous map SS d−1 → Rd−1 . By
Borsuk’s “sphere squashing” theorem, f (x) = f (y) for some antipodal pair x, y ∈ SS d−1 .
Now let x ∈ Fi , y ∈ Fj . Since i 6= j, at least one of them is different from d, say i 6= d. This
means fi (x) = 0 but fi (y) 6= 0, contradicting the statement f (x) = f (y) since i ≤ d − 1.

A.4 Chapter 4
4.1.5. This is again a sweet one. Teachers, please make note.
As in the main text, we first dispose of the case when one of the sets has size λ. Hence-
forth, |Ci | = λ + γi , where γi ≥ 1.
Let vi ∈ Rn denote the incidence vector of the set Ci . The conditions of the Theorem
can be rephrased in terms of the (standard) inner product:
(
λ + γi if j = i ;
vi · vj = (14)
λ if j 6= i .

Assume now that a linear relation n

X
αi v i = 0 (15)
i=1
Pn
exists between the vi . Multiply each side of (15) by vj . Setting β = i=1 αi , we obtain

λβ + αj γj = 0 (j = 1, . . . , m). (16)

In other words,
λ
αj = − β (j = 1, . . . , m). (17)
γj
If β = 0, it follows that all the αj are zero. Otherwise, let us add up all the m equations
(17). We obtain !
m m
X X 1
β= αj = −λ β, (18)
j=1
γ
j=1 j

a contradiction, because the signs of the two sides differ.

——————————————————————— 203

c László Babai and Péter Frankl. 1988, 1992, 2020.
ANSWERS TO THE EXERCISES
Note
P that when
 moving all terms to the left side in equation (bose7eq), the expression
m 1
1+λ j=1 γj occurs, the same as in the expansion of det A in Exercise 4.1.4. (No surprise–
why?)
4.2.1. The first inequality is immediate:
 
 n s n n n nn−1 n−s+1 n
= ··· ≤ ··· = .
s ss s s s−1 1 s

For the second inequality observe that for every α > 0,

n    
nα n
X n j n s
e > (1 + α) = α > α.
j=0
j s

Subtituting α = s/n we obtain   

s n s s
e > ,
s n
as desired.
4.2.10. Take the disjoint union of t − 1 copies of Kt−1 . Make the edges red, connect the
nonadjacent pairs by blue.
4.2.16. Let X = {x1 , . . . , x2t−1 } and |Y | = t4t /2. Assume the edges of the complete
bipartite graph with vertex classes X and Y are colored red and blue. Set B0 = Y . Let
Bi be the larger of the two sets: the set of red neighbors of xi in Bi−1 , and the set of blue
neighbors of xi in Bi−1 , for i = 1, . . . , 2t − 1. Show that |B2t−1 | ≥ t and that there is a
set A ⊆ X, |A| ≥ t, such that the bipartite subgraph with vertex classes A and B2t−1 is
monochromatic.
4.2.17. We use the language of probabilities; this is equivalent to a counting argument
analagous to the solution of Ex. 4.2.9. Consider a random 2-coloring of the edges of Kn,n .
2
(Selecting any of the 2n colorings with equal probability.) The probability that a given
2
bipartite subgraph of size (t, t) is monochromatic is 21−t . Hence the probability that there

2 2
is a monochromatic bipartite subgraph of size (t, t) is less than nk 21−t . Verify that for
n ≤ 2t/2 , this quantity is less than 1. Conclude that BP2 (t) > 2t/2 .
4.3.2. Observe that |F1 ∩ F2 | ≤ 1, and this remains true if we replace F2 by its image F2σ
under any permutation σ of the universe [n]. In particular, the average size (over all σ) of
the intersection of F1 and F2σ is ≤ 1. Compare this with the result of Ex. 4.3.1.
4.3.11. Consider the 3-dimensional subspaces in the projective spaces over Fp .

A.5 Chapter 5
5.1.6. Take the set of the convex hulls of each set of k consecutive vertices of a regular
2k + 1-gon.

204 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 A.5. CHAPTER 5
5.1.8. Let |X| = r + s; let the Ai be all the r-subsets of X; and Bi = X \ Ai .
5.1.9. Solution 1 (Bollobás,1965). Induction on n, the size of the universe.
Solution 2 (for part (a) only) (G. O. H. Katona, 1972). Use a variant of Katona’s “Cyclic
Permutation Method ” introduced in Section 4.4.
5.1.13. We want to reproduce precisely the same situation as guaranteed by the conditions of
the Theorem in a space T of dimension r + 1. To this end, we apply a linear map ϕ : W → T
and hope that ϕ maps all the Bi and Ui injectively, thereby preserving the cardinalities of
the Bi and the dimensions of the Ui , and in addition it preserves the disjointness of each pair
(Bi , Ui ). This is a large number of conditions, each itself easily satisfied. The idea of “general
position” discussed in Section 3.1.3 is that then, all these conditions can simultaneously be
satisfied, assuming the field is large enough. Specifically for the conditions just listed, their
simultaneous feasibility is stated as Corollary 3.15 in Section 3.1.3.
5.3.6. Let V = {1, . . . , n}. With every pair {a, b} ∈ / E (a, b ∈ V ) associate a coordinate.
Make this coordinate of f (v) different for all vertices v except for a and b. Add one more
coordinate to make f one-to-one if G is either complete or the complete graph minus an
edge.
(n)
5.4.1. Observe that for 1 ≤ k ≤ s ≤ n, k−1 = k/(n − k + 1) ≤ s/(n − s + 1). Setting
(nk)
α = s/(n − s + 1), it follows that for s ≤ n/2 we have
∞
s  
!
n
 

X n n X
i s
≤ · α = . (19)
k=0
k s i=0
1−α

Substituting the value of α we obtain the desired inequality.

5.4.2. The monic multilinear monomials of degree k are in one-to-one correspondence with
the k-subsets of the set of variables; therefore their number is nk . Adding up these binomial

coefficients for k ≤ s to obtain an upper bound on m.

5.5.2. Let ABC be an equilateral triangle with unit side length. Let D be the reflection
of A in the line BC. If we could color the unit-distance graph with three colors, A and D
would certainly get the same color. Now rotate the parallelogram ABCD about A so that
the new position D0 of D be at unit distance from D. The color of A would be shared by D0
as well, so D and D0 , at unit distance, would have the same color. (See Figure A.6.)

Figure A.6: Three colors don’t suffice.

5.5.4. First take an (infinite) set H ⊂ Rn , maximal with respect to the property that the
distance between any two of its points is ≥ 1/2. (Such sets clearly exist, you don’t have to

——————————————————————— 205

c László Babai and Péter Frankl. 1988, 1992, 2020.
ANSWERS TO THE EXERCISES
construct them explicitly.) Let H(r) denote the union of the open balls of radius r about
each point in H. Observe that the H(1/2) = Rn .
Let now G be the (infinite) graph with vertex set H, and two points adjacent if their
distance is ≤ 2. Prove that the degree of each vertex in this graph is < 9n . (Use the
observation that the open balls of radius 1/4 about the points of H are disjoint, and at most
9n such balls fit inside a ball of radius (9/4)n .)
Use this information to color G by ≤ 9n colors. Finally, for each u ∈ Rn , find a point
h(u) ∈ H within distance < 1/2 from u, and assign u the color of h(u). Verify that points
at unit distance receive different colors.
5.5.5. The Prime Number Theorem asserts that for every  > 0 and sufficiently large x,
the number of primes not greater than x is between the bounds (1 ± )x/ ln x. Let p be the
largest prime such that n > 4p − 1. By the Prime Number Theorem, for every ε > 0 and
every sufficiently large x there exists a prime number p between (1 − ε)x and x. Applying
this to x = n/4, we find a prime number p such that (1 − ε)n < 4p < n, therefore

c(n) > c(4p − 1) > 1.13974p−1 > 1.1397(1−ε)n > 1.139n

when ε is chosen to be small enough.

5.5.7. Using the estimate from Ex. 5.4.4, and setting α = (p − 1)/n and β = (2p − 1)/n, we
have to maximize the expression
αα (1 − α)1−α
(20)
β β (1 − β)1−β .
Noting that β ≈ 2α, it is a calculus exercise
√ to see, that, setting β = 2α, the maximum
2− 2
in (20) will be reached when α = 4 = 0.1464 . . . . The maximum value of (20) is
1.207 . . . . By the Prime Number Theorem (see the solution of Ex. 5.5.5), we can select a
prime p = (1 + √ o(1))αn. With this choice, we infer that for large n, the chromatic number
of the distance- 2p graph of the unit cube is > 1.2n , and thereby c(n) > 1.2n .
5.5.8. The square of a distances is the sum of the squares of two rationals. Assume (p/q)2 +
(r/s)2 = 3 where p, q, r, s are integers and qs 6= 0. Multiplying by the common denominator
we are led to an equation of the form x2 + y 2 = 3z 2 where x, y, z are integers, z 6= 0. We may
assume g.c.d.(x, y, z) = 1. Since x2 + y 2 is divisible by 3, both x and y must be divisible by
3. (Verify!) But then x2 + y 2 is divisible by 9, therefore z 2 must be divisible by 3. This is a
contradiction with our g.c.d. assumption.
5.5.9. Following the idea of the previous solution, reduce the problem to solving the equation
x2 + y 2 + z 2 = 7w2 in integers, where w 6= 0. Again, we may assume g.c.d.(x, y, z, w) = 1.
First we claim that w must be odd. Indeed otherwise x2 + y 2 + z 2 would be divisible by 4,
forcing each of x, y, z to be even. (To see this, check that if x is odd then x2 ≡ 1 (mod 8).)
But now 7w2 ≡ 7 (mod 8), while each of x2 , y 2 , z 2 is ≡ 0, 1 or 4 (mod 8), a contradiction.
5.5.13. Let r2 = α2 + β 2 where α, β ∈ Q. The matrix
 
α β
(21)
−β α

206 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 A.5. CHAPTER 5
accomplishes the desired transformation.
5.5.14.  
α β γ δ
 β −α δ −γ 

 γ −δ −α
 (22)
β 
δ γ −β −α

5.5.15. By Ex. 5.5.12, r2 can be written as r2 = α2 + β 2 + γ 2 + δ 2 (α, β, γ, δ ∈ Q). Now the

matrix of the preceding exercise accomplishes the desired transformation.
5.5.16. To obtain the desired transformation in dimension 4k, create a 4k × 4k matrix
consisting of 4 × 4 diagonal blocks from Exercise 5.5.14, and zeros elsewhere.
5.5.17. The properties of A listed in the hint can be summarized in the equation AT A = r2 In .
Consequently, (det A)2 = det(r2 In ) = r2n . Since rn−1 ∈ Q (n is odd) but r 6∈ Q, we infer
that det A is irrational, a contradiction, proving that such A does not exist.
5.5.18.√ First, let n = 4` +  where 0 ≤  ≤ 3, ` ∈ Z. As in the solution of Ex. 5.5.7, set
α = 2−4 2 = 0.1464 . . .. Select
√ a prime number p closest to α · 4`. By Ex. 5.5.6, the chromatic
number of the distance- 2p graph on Ω(4k, 2p − 1) is greater than 1.2074k . By Ex. 5.5.16,
an isometric copy T of √12p Ω(4k, 2p − 1) resides in Q4k ⊆ Qn . The unit-distance graph on T
√
is isomorphic to the distance- 2p graph on Ω(4k, 2p − 1); therefore the chromatic number
of the unit-distance graph on Qn is greater than 1.2074k > 1.2n (for large n).
5.8.1. Let U ≤ Rn be a subspace of codimension 2 in general position (in this case, it
should not intersect D2 ). Each v ∈ Rn can be uniquely written as v = ψ(v) + π(v) where
ψ(v) ∈ U and π(v) ⊥ U . Let τ be a rotation of the 2-dimensional plane U ⊥ such that
π(D1 ) ∩ τ (π(D2 )) = ∅. Set ϕ(v) = ψ(v) + τ (π(v)). Obviously, D1 ∩ ϕ(D2 ) = ∅, since even the
projections of these sets (via π) are disjoint. Verify that ϕ is an orthogonal transformation.
(It is the rotation of the same angle as τ , about the (n − 2)-dimensional “axis” U .)
5.8.2. Let a, b, c be integers such that a2 +b2 = c2 (Pythagorean triples), and g.c.d.(a, b, c) =
1. There are infinitely many such triples: take a = u2 − v 2 , b = 2uv, c = u2 + v 2 , where
u, v are relatively prime integers. Set α = a/c, β = b/c, and take the matrix (21) from the
solution of Ex. 5.5.13.
5.8.4. Follow the lines of the proof in the main text, with the following modifications.
Observe that αi /r is the square root of a rational number. Define Si ⊂ Qn as an isometric
copy of (αi /r)S inside Qn . Such a copy exists, since by Ex. 5.5.16, Qn and (αi /r)Qn are
isometric. Finally, use Ex. 5.8.3 in place of Lemma 5.28.
5.9.2. Since qr = qr q−1 r−1 r−1




, we observe that q always divides r q−1 . Therefore, if p does
r−1


not divide q−1 , then q must divide r. Conversely, assume that q divides r. In this case,
for 1
≤ Q
i ≤ q − 1, the integers r − i and q − i contain the same power of p and therefore
r−1 q−1 r−i
q−1
= i=1 q−i is not divisible by p.

5.9.3. Take αi disjoint copies of [n]

i
for i = 0, . . . , r; this makes a total of f (n) points which
will constitute the universe of G. Call these ri=0 αi sets blocks. Within each block [n]
P

i
, every

——————————————————————— 207

c László Babai and Péter Frankl. 1988, 1992, 2020.
ANSWERS TO THE EXERCISES
subset B ⊆ [n] induces the corresponding subset Bi . Let f (B) denote the union over all

blocks of the sets induced by B. Clearly, |f (B)| = f (|B|). Let, finally, G = {f (A) : A ∈ F}.
It is clear that G is f (k)-uniform. Moreover, since f (A1 ∩ A2 ) = f (A1 ) ∩ f (A2 ) (verify!),
we see that G is f (L)-intersecting.
n−3


5.9.4. Let F = {A ⊂ [n] : |A| = 11; 1, 2, 3 ∈ A}. This family is L-intersecting and has 8
members. The resulting family G is f (11) = 55-uniform, f (L)-intersecting, on n2 points,

and has      4
n−3 16 n
> −
8 8! 2
members ( > 0 arbitrarily small, n sufficiently large).
On the other hand, f (L) ≡ {0, 3, 4} (mod 6); and 55 ≡ 1 ∈ / f (L) (mod 6). We obtained
three residue classes for the intersections, but the family size is of fourth order.

n−2

Let F = {A ⊂ [n] : |A| = q − 1; 1, 2 ∈ A}. This family
5.9.5.
2
is L-intersecting and has
q−3
members. The resulting family G is f (q − 1) = (q − 1) -uniform, f (L)-intersecting, on
n points, and has n−2
2


q−3
members. For fixed q and n → ∞, this number is

Ω(nq−3 ).

On the other hand, f (L) (mod q) consists of 0 and those residue classes mod q which are
not divisible by p and are quadratic residues mod p. The total number of such classes is
1 + (q − p)/2, and the corresponding RW–type bound would be

n2
 
q−p+2
q−p <n ,
1+
2
by a factor of np−5 smaller than nq−3 .
5.10.1. Consider the following polynomials fi ∈ Fp [x1 , . . . , xn ].
Y
fi (x) = (x · vi − `).
`∈Li

Verify that eqn. (38) of the proof of the Nonuniform RW Theorem (Theorem 5.34) holds for
these polynomials, and finish along the lines of that proof.
To deduce Theorem 5.34 from this result, let |A1 | ≤ . . . ≤ |Am |. Set Li = {` ∈ L : ` <
|Ai |}, and select a prime p > n.
5.10.2. The upper part of equation (38) will fail if |Ai | ∈ L. Formula (37) prevents this
from happening, but now the lower part of (38) would be violated, if for some j < i we had
Ai ⊂ Aj . This possibility is eliminated by the ordering of the Ai .
5.12.3. If W 6⊆ Y then the sum is empty. Assume now that W ⊆ Y . We may assume
W = ∅ (otherwise delete the elements of W both from Y and from each T ). Now if Y = ∅,
the sum is 1. Otherwise let y0 ∈ Y , and let us pair up the subsets of Y by adding/deleting
y0 . It is clear that the contributions of the two sets in each pair cancel out.

208 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 A.6. CHAPTER 6
5.12.4. Let β be the zeta transform of α and δ the Moebius transform of β. We have to
prove that δ = α. We do this by substituting and switching the order of summations: We
have
X
δ(W ) = (−1)|W | (−1)|T | β(T )
T ⊆W
X X
|W |
= (−1) (−1)|T | α(U )
T ⊆W U ⊆T
X X
= α(U ) (−1)|W |+|T | .
U ⊆X U ⊆T ⊆W

Note that for T ⊆ W we have (−1)|W |+|T | = (−1)|W \T | . Therefore by Ex. 5.12.3, the only
term remaining in the outer sum corresponds to U = W ; hence the sum reduces to a single
term equal to α(W ), as desired.
5.12.5.
X
(−1)|T | β(T ) =
W ⊆T ⊆Y
X X
(−1)|T | α(U ) =
W ⊆T ⊆Y U ⊆T
X X
α(U ) (−1)|T | .
U ⊆Y U ∪W ⊆T ⊆Y

By Ex. 5.12.3, the nonvanishing terms of the outer sum correspond to U ∪ W = Y , i. e.,
Y \ W ⊆ U ⊆ Y ; the term corresponding to such a U is equal to (−1)|Y | .

A.6 Chapter 6
6.1.1. Fix the variables in all positions other than the ith and the j th . Set g(wi , wj ) =
f (. . . , wi , . . . , wj , . . .). We have to prove that for x, y ∈ W , g(x, y) = −g(y, x). From (4) we
take that g(x, x) = g(y, y) = g(x + y, x + y) = 0. On the other hand, by (3),
g(x + y, x + y) = g(x, x) + g(x, y) + g(y, x) + g(y, y),
hence 0 = g(x, y) + g(y, x), as claimed.
6.1.2. Using the notation of the preceding solution, we now have g(x, y) + g(y, x) = 0 for
x, y ∈ W . Setting y = x we obtain 2g(x, x) = 0 which implies g(x, x) = 0 unless char F = 2.
6.1.3. Condition (7) now says g(x, y) = g(y, x). In order to construct such a k-linear
function, assume W = Fn , and let αi denote the first coordinate of wi ∈ W (i = 1, . . . , k).
Set, for example, f (w1 , . . . , wk ) = α1 · . . . · αk .
6.2.1. Let Gn be the complement of the perfect matching with 2n vertices. Show that
adim(Gn ) = 2. On the other hand, infer form Lovász’s theorem (Theorem 6.14) that
pdim(Gn ) = Ω(log n).

——————————————————————— 209

c László Babai and Péter Frankl. 1988, 1992, 2020.
ANSWERS TO THE EXERCISES
A.7 Chapter 7
7.4.1. See Exercise 5.12.3 and the proof of equation (17) in Section 5.12.
7.4.7. Suppose that we have a set X of n+2 points in Rn . By Radon’s theorem (Theorem 6)
there exists a partition X = A0 ∪A˙ 1 such that conv(A0 ) ∩ conv(A1 ) 6= ∅. Suppose there are
balls Q0 and Q1 , with centers a0 and a1 , respectively, such that Ai = X ∩ Qi (i = 0, 1). The
point ai is closer to every point of Ai than to any point of A1−i . Therefore,
(a1 − a0 )(x1 − x0 ) > 0 for all x1 ∈ A1 , x0 ∈ A0 . (23)
(We have taken inner
P products.) By P Radon’s theorem there exist coefficients α(x) ≥
0, β(y) ≥ 0, with x∈A0 α(x) = 1 = y∈A1 β(y) satisfying
X X
α(x)x = β(y)y (24)
x∈A0 y∈A1

(this is a common point of conv(A0 ) and conv(A1 ).) From eqn. (24) it follows that
X
α(x)β(y)(y − x) = 0.
x∈A0 , y∈A1

Taking the inner product of each side with a1 − a0 we obtain a contradiction with (23).
7.4.8. Taking the 2k + 1 vertices of a regular 2k + 1-gon shows V C-dim ≥ 2k + 1. Suppose
|X| = 2k + 2, X ⊂ R2 . We have to show that for some A ⊂ X there is no convex k-gon K
with X ∩ K = A. By the preceding variant we may assume that X = {x2 , x3 . . . . , x2k+2 } is
the vertex set of convex 2k +2-gon. Set A = {x1 , x3 , . . . , x2k+1 }. Suppose for contradiction
that X ∩K = A for some convex k-gon K. Take K minimal so that every side contains a point
of A. Let H1 , . . . , Hk be the halfplanes determined by the edges of K and not containing K.
By the pigeonhole principle ∃i such that Hi contains at least two of the s + 1 points in X \ A.
Let x2a and x2b be those vertices. Then the angle x2a xu x2b is concave for xu ∈ A ∩ ∂Hi , a
contradiction.
7.4.9. Let A = {a1 , . . . , as } be a largest set shattered by F. Let the teacher follow the
following strategy: whatever the for the first s queries, he responds by stating that ai is a
counterexample to the ith query (i ≤ s). These responses are consistent with F since A is
shattered.
7.4.10. Guessing one concept for each member of thePtrace of
F in Y is certainly sufficient.
s
Hence the number of queries needed is bound by i=0 s
|Y |
< |Y |s where s is the VC
dimension of F.
7.4.12. Let the coefficient of the set E ∈ F be γ(E) = (−1)|E| .
7.4.13. As in equation (9), we write f (x) as a unique linear
combination of the polynomials
x x


i
, 0 ≤ i ≤ s. (Note that in order for the polynomial i to make sense over Fp , we need
that p does not divide s!.)
X  x
f (x) = αi , αi ∈ Fp .
0≤i≤s
i

210 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 A.8. CHAPTER 8
x


It is therefore sufficient to prove (37) for f (x) = i
. In that case the left hand side is
  X X
X |E ∩ A| X
γ(E) = γ(E) = δ(B).
E∈F
i E∈F B∈( E∩A A
i )
B∈( i )

Each term on the right hand side vanishes by Observation 7.24.

7.4.14. Let l1 , l2 , . . . , lt be representatives of the distinct residue classes mod p of the integers
|E ∩ E0 | (E 6= E0 , E ∈ F).
Clearly t ≤ p; if t = p then (i) holds. If t ≥ s + 1 then Qt (ii) holds. So we may assume that
t ≤ min{p − 1, t} and apply Exercise 7.4.13 to f (x) = j=1 (x − lj ). For this polynomial we
have f (|E ∩ E0 |) ≡ 0 (mod p) for all E 6= E0 , E ∈ F. This reduces (37) (with A = E0 ) to
t
Y
γ(E0 ) (|E0 | − lj ) ≡ 0 (mod p).
j=1

Consequently |E0 | ≡ lj (mod p) for some j, proving that (i) holds.

7.4.15. Condition (b) rules out option (ii) of the preceding exercise, and condition (a) rules
out option (i). Hence, F cannot be s-dependent mod p.
The next two exercises show that the two main versions of the RW Theorem discussed
previously are immediate consequences of Exercise 7.4.14.
7.4.16. Assume for contradiction that F is s∗ -dependent. Set p = 0 and choose E0 to
have maximum cardinality among the edges with nonzero coefficient. Then alternative (i)
of Exercise 7.4.14 is ruled out. Alternative (ii) contradicts the assumption of Theorem 7.8.
7.4.17. F is uniform now, so, in order to prove F s-independent it suffices to prove its
s∗ -independence. Exercise 7.4.15 asserts even more: F is s∗ -independent in characteristic p.
7.5.1. Observe that there are at most k − r intersection sizes. Therefore F is (k − r)-
independent by the RW Theorem. Now apply Theorem 7.30 with s := k − r.

A.8 Chapter 8
A.9 Chapter 9

——————————————————————— 211

c László Babai and Péter Frankl. 1988, 1992, 2020.
212 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
Bibliography

H. L. Abbott (1972), A note on Ramsey’s theorem, Canad. Math. Bull. 15 (1972) 9–10.
[87, 90]
Martin Aigner (1979), Combinatorial Theory, Springer 1979.
P. S. Aleksandrov (1969), Problemy Gilberta (Hilbert’s Problems), Nauka, Moscow,
1969 (Russian). Translated to German: “Hilbertsche Probleme,” Leipzig, Akad. Verlag, 1971
[16]
Noga Alon (1986a), Eigenvalues and expanders (1986), Combinatorica 6 (1986), 83–96.
Noga Alon (1986b), Covering graphs by the minimum number of equivalence relations,
Combinatorica 6 (1986), 201–206.
Noga Alon (1986c), Eigenvalues, geometric expanders, sorting in rounds, and Ramsey
theory, Combinatorica 6 (1986), 207–219.
Noga Alon, László Babai, Hiroshi Suzuki (1991), Multilinear polynomials and
Frankl–Ray-Chaudhuri–Wilson type intersection theorems, J. Combin. Th. A 58 (1991),
165–180. [110, 127, 129 129, 130, 133 ]
Noga Alon, László Babai, Alon Itai (1986), A fast and simple randomized parallel
algorithm for the maximum independent set problem, J. Algorithms 7 (1986), 567–583.
Noga Alon, Pt́er Frankl, László Lovász (1986), The chromatic number of Kneser
hypergraphs, Trans. Amer. Math. Soc. 298 (1986), 359–370.
Noga Alon, Péter Frankl, Vojtěch Rödl (1985), Geometrical realization of set
systems and probabilistic communication complexity, Proc. 26th IEEE Symp. on Found.
Comp. Sci., Portland OR, pp. 277–280.
Noga Alon, Gil Kalai, A simple proof of the upper bound theorem, European J. Comb.
6 (1985), 211–214. [76, 102, 136]
Noga Alon, Vitali D. Milman, λ1 , isoperimetric inequalities for graphs and supercon-
centrators, J. Combin. Th. B 38 (1985), 73–88.
Brian Alspach, Moshe Rosenfeld (1977), On embedding triangle-free graphs in unit
spheres, Discrete Math. 19 (1977), 103–111. [106]

c László Babai and Péter Frankl. 1988, 1992, 2020.

213
BIBLIOGRAPHY
Ian Anderson (1974), A First Course in Combinatorial Mathematics, Clarendon Press,
Oxford 1974.
Ian Anderson (1987), Combinatorics of finite sets, Oxford University Press, 1987.
László Babai (1987), On the nonuniform Fisher inequality, Discrete Math. 66 (1987),
303–307. [84]
László Babai (1988), A short proof of the non-uniform Ray-Chaudhuri–Wilson inequality,
Combinatorica 8 (1988), 133–135. [126]
László Babai, Péter Frankl (1980), On set intersections, J. Combin. Th. A 28 (1980),
103–105. [7]
László Babai, Vera T. Sós (1985), Sidon sets in groups and induced subgraphs of
Cayley graphs, European J. Comb. 6 (1985), 101–114.
Eiichi Bannai, Etsuko Bannai (1981), An upper bound for the cardinality of an s-
distance subset in real Euclidean space, Combinatorica 1 (1981), 99–102. [9, 11]
Eiichi Bannai, Etsuko Bannai, Dennis Stanton (1983), An upper bound for the
cardinality of an s-distance subset in real Euclidean space II, Combinatorica 3 (1983),
147–152. [11]
Eiichi Bannai, R. M. Damerell (1979), Tight spherical designs, I, J. Math. Soc. Japan
31 (1979), 199–207.
Eiichi Bannai, Tatsuro Ito (1984), Algebraic Combinatorics 1, Benjamin/Cummings,
Menlo Park, 1984. [iii]
Imre Bárány (1978), A short proof of Kneser’s Conjecture, J. Combin. Th. A 25 (1978),
325–326. [79]
J. G. Basterfield, Leroy Milton Kelly (1968), A characterization of sets of n points
which determine n hyperplanes, Proc. Cambr. Phil. Soc. 64 (1968), 585–588.
József Beck (1981), Van der Waerden and Ramsey type games, Combinatorica 1 (1981),
103–116.
József Beck, Tibor Fiala (1981), Integer making theorems, Discrete Appl. Math. 3
(1981), 1–8.
Claude Berge (1973), Graphs and Hypergraphs, North–Holland, Amsterdam 1973.
Elwyn R. Berlekamp (1969), On subsets with intersections of even cardinality, Canad.
Math. Bull. 12 (1969), 363–366. [3, 5, 56]
Abram Samoilovitch Besicovitch (1940), On the linear independence of fractional
powers of integers, J. London Math. Soc. 15 (1940), 3–6. [43]
Norman L. Biggs (1974), Algebraic Graph Theory, Cambridge University Press, Cam-
bridge 1974. [iii]

214 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 BIBLIOGRAPHY
Norman L. Biggs, M. Hoare (1983), The sextet construction for cubic graphs, Combi-
natorica 3 (1983), 153–165.
Anders Björner (1979), Fixed point in partially ordered sets, Adv. in Math. 31 (1979),
263–287. [104]
Anders Björner (1981), Homotopy type of posets and lattice complementation, J. Com-
bin. Th. A 30 (1981), 187–188. [104]
Anders Björner (1981), Fixed points and complements in finite lattices, J. Combin. Th.
A 30 (1981), 335–338. [104]
Anders Björner, Péter Frankl, Richard M. Stanley (1987), The number of faces
of balanced Cohen-Macaulay complexes and a generalized Macaulay theorem, Combinatorica
7 (1987), 23–34.
Aart Blokhuis (1981), A new upper bound for the cardinality of 2-distance sets in
Euclidean space, Eindhoven Univ. Technology, Mem. 1981–04. [9, 128]
Aart Blokhuis (1982), An upper bound for the cardinality of s-distance sets in E d and
H d , Eindhoven Univ. Technology, Mem. 1982–68. [11]
Aart Blokhuis (1984), Few distance sets, C.W.I. Tracts No.7, Mathematisch Centrum,
Amsterdam 1984. [11, 128]
Aart Blokhuis (1990), Solution of an extremal problem for sets using resultants of poly-
nomials, Combinatorica 10 (1990), 393–396. [104]
Anselm Blumer, Andrzej Ehrenfeucht, David Haussler, Manfred K. War-
muth (1989), Learnability and the Vapnik–Chervonenkis dimension, J. of ACM 36 (1989),
929–965. [160]
Béla Bollobás (1965), On generalized graphs, Acta Math. Acad. Sci. Hungar. 16 (1965),
447–452. [96, 99, 100, 101, 205]
Béla Bollobás (1978), Extremal Graph Theory, Academic Press, London 1978.
Béla Bollobás (1979), Graph Theory, An Introductory Course, Springer 1979. [xiv]
Béla Bollobás (1986), Combinatorics, Cambridge University Press 1986. [95]
Béla Bollobás, Pierre Duchet (1979), Helly families of maximal size, J. Combin. Th.
A 26 (1979), 197–200.
Vladimir G. Boltyansky (1978), Hilbert’s Third Problem, Winston 1978. [14, 16,
18, 197, 197]
Vladimir G. Boltyansky, Israel Ts. Gohberg (1985), Results and Problems in
Combinatorial Geometry, Cambridge University Press 1985. [195, 117]
J. Adrian Bondy (1972), Induced subsets, J. Combin. Th. B 12 (1972), 201–202. [83]
J. Adrian Bondy, Robert L. Hemminger (1977), Graph reconstruction – A survey,
J. Graph Theory 1 (1977), 227–268. [170]

——————————————————————— 215

c László Babai and Péter Frankl. 1988, 1992, 2020.
BIBLIOGRAPHY
J. Adrian Bondy, U. S. Rama Murty (1976), Graph Theory with Applications,
Macmillan, London, 1976. [xiv]
Karol Borsuk (1933), Drei Sätze über die n-dimensionale euklidische Sphäre, Fund.
Math. 20 (1933), 177–190. [77, 80, 116]
Raj Chandra Bose (1949), A note on Fisher’s inequality for balanced incomplete block
designs, Ann. Math. Stat. 20 (1949), 619–620. [81]
Raj Chandra Bose (1963), Strongly regular graphs, partial geometries, and partially
balanced designs, Pacific J. Math. 13 (1963), 389–419.
Raj Chandra Bose (1984), Introduction to Combinatorial Theory, Wiley, 1984.
N. Bourbaki (1974) Algebra I, Addison–Wesley, 1974.
Andries E. Brouwer, Arjeh M. Cohen, Arnold Neumaier (1989), Distance-Regular
Graphs, Springer, 1989. [iii]
Felix E. Browder, ed. (1976), Mathematical developments arising from Hilbert prob-
lems, (Proc. Symp. Pure Math., Vol. XXVIII., Northern Illinois Univ., De Kalb, IL, 1974),
AMS, Providence, R.I., 1976. [16]
W. G. Bridges (1977), A characterization of type-1 λ-designs, J. Combin. Th. A 22
(1977), 361–367. [85]
R. C. Buck (1943), Partition of space, Amer. Math. Monthly 48 (1943), 541–544.
Constantin Carathéodory (1907), Über den Variabilitätsbereich der Koeffizienten von
Potenzreihen, die gegebene Werte nicht annehmen, Math. Annalen 64 (1907), 95–115. [62,
74]
Benny Chor, Oded Goldreich, Johan Håstad, Joel Friedman, Steven Rudich,
Roman Smolensky (1985), The bit extraction problem or t-resilient functions, in: Proc.
26th IEEE Symp. on Foundations of Computer Science, Portland, OR 1985.
Václav Chvátal (1983), Linear programming, W.H. Freeman, N.Y. 1983. [71]
Paul J. Cohen (1966), Set Theory and the Continuum Hypothesis, Benjamin, New York
1966. [14]
Harold S. M. Coxeter (1927), The pure Archimedean polytopes in six and seven di-
mensions, Proc. Cambr. Phil. Soc. 24 (1927), 1–9. [10]
Harold S. M. Coxeter (1968), Twelve Geometric Essays, Southern Illinois University
Press, Carbondale 1968. [10]
Harold S. M. Coxeter (1973), Regular polytopes, Dover, N.Y., 1973. [10]
Peter Crawley, Robert P. Dilworth (1973), Algebraic Theory of Lattices, Prentice-
Hall, 1973.
Dragoš M. Cvetković, Michael Doob, Horst Sachs (1980), Spectra of graphs,
Academic Press, New York 1980.

216 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 BIBLIOGRAPHY
Ludwig Danzer, Brako Grünbaum (1962), Über zwei Probleme bezüglich konvexer
Körper von P. Erdős und von V. L.Klee, Math. Zeitschrift 79 (1962), 95–99. [12]
Martin Davis, Yury V. Matiyasevich, Julia Robinson (1976), Hilbert’s tenth prob-
lem. Diophantine equations: positive aspects of negative solution, Proc. Symp. Pure Math.
28 (1976), 223–378. [14]
Nicolaas Govert de Bruijn, Paul Erdős (1948), On a combinatorial problem, Proc.
Konink. Nederl. Akad. Wetensch. Ser. A. 51 (1948), 1277–1279 (=Indagationes Math. 10
(1948), 421–423.) [82]
Nicolaas Govert de Bruijn, Paul Erdős (1951), A colour problem for infinite graphs
and a problem in the theory of relations, Proc. Konink. Nederl. Akad. Wetensch. Ser. A 54
(1951), 371–373. [78]
Max Dehn (1900), Über raumgleiche Polyeder, Nachr. Akad. Wiss. Göttingen Math.-Phys.
Kl. 55 (1900), 345–354. [14]
Paul Delsarte (1973), An algebraic approach to the association schemes of coding theory,
Philips Res. Rept. Suppl. 10 (1973).
Paul Delsarte, Jean-Marie Goethals, Johan Jacob Seidel (1975), Bounds for
systems of lines and Jacobi polynomials, Philips Res. Repts. 30 (1975), 91–105.
Paul Delsarte, Jean-Marie Goethals, Johan Jacob Seidel (1977), Spherical
codes and designs, Geometriae Dedicata 6 (1977), 363–388. [10, 10, 11]
Peter Dembowski (1968), Finite Geometries, Springer 1968.
Michel Deza (1973), Une propriété extrémale des plans projectifs finis dans une classe
de codes equidistants, Discrete Math. 6 (1973), 343–352. [90]
Michel Deza, Paul Erdős, Péter Frankl (1978), Intersection properties of systems
of finite sets, Proc. London Math. Soc. 36 (1978), 369-384.
Michel Deza, Paul Erdős, Navin M. Singhi (1978), Combinatorial problems on
subsets and their intersections, Advances in Mathematics, Suppl. Studies 1 (1978), 259–265.
Michel Deza, Péter Frankl (1981), Every large set of equidistant (0, +1, −1)-vectors
forms a sunflower, Combinatorica 1 (1981), 225–231.
Michel Deza, Péter Frankl (1982), On the vector space of 0-configurations, Combi-
natorica 2 (1982), 341–345.
Michel Deza, Péter Frankl, Navin M. Singhi (1983), On functions of strength t,
Combinatorica 3 (1983), 331–339. [109, 164]
Jean Dieudonné et al. (1978), Abrégé d’histoire des mathématiques 1700–1900, Her-
rmann, Paris 1978.
Jean Dieudonné (1979), The tragedy of Grassmann, Linear and Multilinear Alg. 8 (1979),
1–14. [135, 136]

——————————————————————— 217

c László Babai and Péter Frankl. 1988, 1992, 2020.
BIBLIOGRAPHY
Thomas A. Dowling, Richard M. Wilson (1975), Whitney number inequalities for
geometric lattices, Proc. A. M. S. 47 (1975), 504–512. [184]
Andres Dress, László Lovász (1987), On some combinatorial properties of algebraic
matroids, Combinatorica 7 (1987), 39–48.
B. Dushnik, E. W. Miller (1941), Partially ordered sets, Amer. J. Math. 63 (1941),
600–610. [104]
Jack Edmonds (1967), Systems of distinct representatives and linear algebra, J. Res. Nat.
Bur. Standards 71B (1967), 241–245. [69]
Georgy P. Egorychev (1981), The solution of van der Waerden’s problem for perma-
nents, Advances in Math. 42 (1981), 299–305.
Mohamed M. El-Zahar, Norbert W. Sauer (1985), The chromatic number of the
product of two 4-chromatic graphs is 4, Combinatorica 5 (1985), 121–126. [107]
Paul Erdős (1947), Some remarks on the theory of graphs, Bull. Amer. Math. Soc. 53
(1947), 292–294. [87]
Paul Erdős (1950), Some remarks on set theory, Proc. Amer. Math. Soc. 1 (1950), 127–
141.
Paul Erdős (1959), Graph theory and probability, Can. J. Math. 11 (1959), 34–38.
Paul Erdős (1976), Problems and results in graph theory and combinatorial analysis, in:
Proc. Fifth British Comb. Conf., Aberdeen 1975, Congr. Num. 15, Utilitas Math., Winnipeg
1976, pp.169–192.
Paul Erdős (1976), Problems and results in combinatorial analysis, in: Colloq. Intern.
sulle Teorie Combin., Roma 1973, Acad. Naz. Lincei, Roma 1976, Vol. 2, pp. 3–17.
Paul Erdős (1973), The Art of Counting, (Joel H. Spencer, ed.) M.I.T. Press, 1973. [86]
Paul Erdős, András Hajnal (1961) On a property of families of sets, Acta Math. Acad.
Sci. Hung. 12 (1961), 87–123. [25]
Paul Erdős, András Hajnal (1966) On chromatic number of graphs and set systems,
Acta Math. Acad. Sci. Hung. 17 (1966), 61–99. [58, 78]
Paul Erdős, András Hajnal, J. W. Moon (1964), A problem in graph theory, Amer.
Math. Monthly 71 (1964), 1107–1110. [99, 100]
Paul Erdős, Frank Harary, William T. Tutte (1965), On the dimension of a graph,
Mathematika 12 (1965), 118–122. [106]
Paul Erdős, Chao Ko, Richard Rado (1961), Intersection theorems for systems of
finite sets, Quart. J. Math. Oxford 12 (1961), 313–320. [96]
Paul Erdős, Péter Frankl, Vojtěch Rödl (1986), The asymptotic number of graphs
not containing a fixed subgraph and a problem for hypergraphs having no exponent, Graphs
and Combinatorics 2 (1986), 113–121.

218 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 BIBLIOGRAPHY
Paul Erdős László Lovász (1975), Problems and results on 3-chromatic hypergraphs
and some related questions, in: Infinite and Finite Sets, Colloq. Math. Soc. J. Bolyai 10,
Keszthely, Hungary 1973 (A. Hajnal, R. Rado, V. T. Sós, eds.), North-Holland 1975, pp.
609–628.
Paul Erdős, P. Montgomery, Bruce L. Rothschild, Joel H. Spencer, Ernst
G. Straus, (1973), Euclidean Ramsey theorems, I, J. Combin. Th. A 14 (1973), 341–363;
Coll. Math. Soc. J. Bolyai 10, Infinite and Finite Sets, Keszthely, Hungary, North–Holland
1973, pp. 529–557 and 559–583.
Paul Erdős, P. Montgomery, Bruce L. Rothschild, Joel H. Spencer, Ernst
G. Straus (1975), Euclidean Ramsey theorems II–III, in: Infinite and Finite Sets, Colloq.
Math. Soc. J. Bolyai 10, Keszthely, Hungary 1973, North–Holland 1975, pp. 529–557 and
559–583.
Paul Erdős, Richard Rado (1960), Intersection theorems for systems of sets, J. London
Math. Soc. 35 (1960), 85–90. [97]
Paul Erdős, Miklós Simonovits (1980), On the chromatic number of geometric graphs,
Ars Combinatoria 9 (1980), 229–246. [106]
Paul Erdős, Joel H. Spencer (1974), Probabilistic Methods in Combinatorics, Akadémiai
Kiadó, Budapest 1974. [25, 58]
Paul Erdős, György Szekeres (1935), A combinatorial problem in geometry, Compo-
sitio Math. 2 (1935), 464–470. [86, 89]
H. Eves (1963), A survey of geometry, Allyn & Bacon, Boston, 1963. [13]
Dmitry I. Falikman (1981), A proof of the van der Waerden conjecture on the perma-
nent of a doubly stochastic matrix, Math. Zametki 29 (1981), 931–938 (Russian). English
translation: Math. Notes of the Academy of Sci. of the USSR 29, 1981, 475–479.
Miroslav Fiedler (1973), Algebraic connectivity of graphs, Czechoslovak Math. J. 23
(1973), 298–305.
Ronald Aylmer Fisher (1935), The design of experiments, Oliver and Boyd, Edinburgh
1935.
Ronald Aylmer Fisher (1940), An examination of the different possible solutions of a
problem in incomplete blocks, Annals of Eugenics (London) 10 (1940), 52–75. [81]
D. E. Flath (1989), Introduction to Number Theory, Wiley, 1989. [115]
Péter Frankl (1977), A constructive lower bound for Ramsey numbers, Ars Combina-
toria 3 (1977), 297–302. [120]
Péter Frankl (1982), An extremal problem for two families of sets, European J. Comb.
3 (1982), 125–127.
Péter Frankl (1983a), On the trace of finite sets, J. Combin. Th. A 34 (1983), 41–45.

——————————————————————— 219

c László Babai and Péter Frankl. 1988, 1992, 2020.
BIBLIOGRAPHY
Péter Frankl (1983b), An extremal set theoretic characterization of some Steiner sys-
tems, Combinatorica 3 (1983), 193–199.
Péter Frankl (1984), Families of finite sets with three intersections, Combinatorica 4
(1984), 141–148.
Péter Frankl (1986a), Orthogonal vectors in the n-dimensional cube and codes with
missing distances, Combinatorica 6 (1986), 279–285.
Péter Frankl (1986b), All rationals occur as exponents, J. Combin. Th. A 42 (1986),
200–206.
Péter Frankl (1986c), Intersection theorems and geometric applications, Proc. Internat.
Congress Math., Berkeley 1986, AMS 1988, 1419–1430.
Péter Frankl, Intersection theorems for sets and constructive Ramsey bounds, in The
Mathematics of Ramsey theory, (Nešetřil–Rödl ed.) Springer Verlag, 1990.
Péter Frankl (1989), Helly-type theorems for varieties, European J. Comb. 10 (1989),
253–245.
Péter Frankl (1990), Intersection theorems and mod p rank of inclusion matrices, J.
Combin. Th. A 54 (1990), 85–94.
Péter Frankl, The convex hull of antichains in posets, Combinatorica 12 (1992), 493–
496.
Péter Frankl, H. Enomoto, Noboru Ito, K. Nomura (1987), Codes with given
distances, Graphs and Combinatorics 3 (1987), 25–38.
Péter Frankl, Zoltán Füredi (1983), A new generalization of the Erdős–Ko–Rado
theorem, Combinatorica 3 (1983), 341–349. [165]
Péter Frankl, Zoltán Füredi (1984), On hypergraphs without two edges intersecting
in a given number of vertices, J. Combin. Th. A 36 (1984), 230–236.
Péter Frankl, Zoltán Füredi (1986), Finite projective spaces and intersecting hyper-
graphs, Combinatorica 6 (1986), 335–354.
Péter Frankl, Zoltán Füredi (1988), Solution of the Littlewood–Offord problem in
high dimensions, Annals of Math. 128 (1988), 259–270.
Péter Frankl, Zoltán Füredi, Gil Kalai (1989), Shadows of colored complexes,
Math. Scandinavica, 63 (1989), 169–178.
Péter Frankl, Zoltán Füredi, János Pach (1987), Bounding one-way differences,
Graphs and Combinatorics 3 (1987), 341–347.
Péter Frankl, Ronald L. Graham (1987), The Radon transform on Abelian groups,
J. Combin. Th. A 44 (1987), 168–171.
Péter Frankl, H. Maehara (1986), Embedding the n-cube in lower dimensions, Eu-
ropean J. Comb. 7 (1986), 221–225.

220 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 BIBLIOGRAPHY
Péter Frankl, H. Maehara (1988a), On the contact dimension of a graph, Discr.
Comput. Geom. 3 (1988), 89–96. [106]
Péter Frankl, H. Maehara (1988b), The Johnson–Lindenstrauss lemma and the
sphericity of some graphs, J. Combin. Th. B 44 (1988), 355–362.
Péter Frankl, Andrew M. Odlyzko (1983), On subsets with cardinalities divisible
by a fixed integer, European J. Comb. 4 (1983), 215–220.
Péter Frankl, János Pach (1983), On the number of sets in a null t-design, European
J. Comb. 4 (1983), 21–23. [162]
Péter Frankl, János Pach (1984), On disjointly representable sets, Combinatorica 4
(1984), 39–45.
Péter Frankl, Vojtěch Rödl (1984), Hypergraphs do not jump, Combinatorica 4
(1984), 149–159.
Péter Frankl, Vojtěch Rödl (1986), All triangles are Ramsey, Transactions AMS
297 (1986), 777–779.
Péter Frankl, Vojtěch Rödl (1987), Forbidden intersections, Transactions AMS 300
(1987), 259–286. [111]
Péter Frankl, Vojtěch Rödl, Richard M. Wilson (1988), The number of sub-
matrices of a given type in a Hadamard matrix and related results, J. Combin. Th. B 44
(1988), 317–328.
Péter Frankl, Ivo G. Rosenberg (1981), A finite set intersection theorem, European
J. Comb. 2 (1981), 127–129.
Péter Frankl, Navin M. Singhi (1983), Linear dependencies among subsets of a finite
set, European J. Comb. 4 (1983), 313–318.
Péter Frankl, Richard M. Wilson (1981), Intersection theorems with geometric
consequences, Combinatorica 1 (1981), 357–368. [113, 120, 125, 130, 153, 155, 156]
Péter Frankl, Richard M. Wilson (1986), The Erdős–Ko–Rado theorem for vector
spaces, J. Combin. Th. A 43 (1986), 228–236.
Ferdinand Georg Frobenius (1912), Über Matrizen aus nicht negativen Elementen,
Sitzungsber. Königl. Preuss. Akad. Wiss. 26 (1912), 456–477.
Ferdinand Georg Frobenius (1917), Über zerlegbare Determinanten, Sitzungsber.
Königl. Preuss. Akad. Wiss. XVIII (1917), 274–277.
Zoltán Füredi (1981), Maximum degree and fractional matchings in uniform hyper-
graphs, Combinatorica 1 (1981), 155–162.
Zoltán Füredi (1984), Geometrical solution of an intersection problem for two hyper-
graphs, European J. Comb. 5 (1984), 133–136. [142, 145]
Zoltán Füredi, Jerrold R. Griggs (1986), Families of finite sets with minimum shad-
ows, Combinatorica 6 (1986), 355–363.

——————————————————————— 221

c László Babai and Péter Frankl. 1988, 1992, 2020.
BIBLIOGRAPHY
Zoltán Füredi, Jerrold R. Griggs, R. Holzman, Daniel J. Kleitman (1990),
Representations of families of triples over GF (2), J. Combin. Th. A 53 (1990), 306–315.
David Gale (1956), Neighboring vertices on a convex polyhedron, in: Linear inequalities
and related systems (ed. H.W. Kuhn, A.W. Tucker), Princeton University Press, Princeton,
1956, pp. 255–263. [62, 74, 76, 76]
David Gale (1963), Neighborly and cyclic polytopes, Proc. Symp. Pure Math. 7 (Con-
vexity), 225–232.
David Gale (1964), On the number of faces of a convex polytope, Can. J. Math. 16 (1964),
12–17.
Felix Ruvimovich Gantmacher (1960), The Theory of Matrices I, II, Chelsea, New
York 1960.
Richard J. Gardner, Stanley Wagon (1989), At long last, the circle has been squared,
Notices of the AMS Dec (1989), 1338-1343. [17]
Anthony Vito Geramita, Jennifer Seberry (1979), Orthogonal designs, Marcel
Dekker, New York 1979.
Kurt Gödel (1931), Über formal unentscheidbare Sätze der Principia Mathematica und
verwandter Systeme, I. Monatsh. Math. Phys., 38 (1931), 173–198. [14]
Kurt Gödel (1940), The Consistency of the Continuum Hypothesis, Ann. Math. Studies,
Vol. 3, Princeton University Press [14]
Chris D. Godsil (1987), Polynomial spaces, Res. Rept. CORR 87-17, Fac. Math., Uni-
versity of Waterloo, 1987
Chris D. Godsil (1989), Some Algebraic Combinatorics, Dept. of Combinatorics and
Optimization, University of Waterloo, 1989. [iii]
Chris D. Godsil, I. Krasikov, Y. Roditty (1987), Reconstructing graphs from their
k-edge-deleted subgraphs, J. Combin. Th. B 43 (1987), 360–363. [170]
Jean-Marie Goethals, Johan Jacob Seidel (1970), Strongly regular graphs derived
from combinatorial designs, Canad. J. Math. 22 (1970), 597-614.
Martin C. Golumbic (1980), Algorithmic Graph Theory and Perfect Graphs, Academic
Press, New York, 1980.
Daniel Gorenstein (1985), The Enormous Theorem, Scientific American 253, Dec. 1985,
104–114. [67]
T. Gosset (1900), On the regular and semiregular figures in space of n dimensions, Mes-
senger of Mathematics 29 (1900), 43–48. [10]
Daniel H. Gottlieb (1966), A certain class of incidence matrices, Proc. A. M. S. 17
(1966), 1233–1237. [157, 181]
Ronald L. Graham (1980), On partitions of En , J. Combin. Th. A 28 (1980), 89-97.

222 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 BIBLIOGRAPHY
Ronald L. Graham (1988), Isometric embeddings of graphs, in: Selected Topics in Graph
Theory 3 (L. W. Beineke, R. J. Wilson, eds.), Academic Press, 1988, pp. 133–150. [24]
Ronald L. Graham, S. Y. R. Li, Wen-Ching Winnie Li (1980), On the structure of
t-designs, SIAM J. on Alg. Discr. Methods 1 (1980), 8–14.
Ronald L. Graham, Lászlo Lovász (1978), Distance matrix polynomials of trees,
Advances in Math. 29 (1978), 60–88.
Ronald L. Graham, Henry O. Pollak (1971), On the Addressing Problem for Loop
Switching, Bell Sys. Tech. J. 50 (1971), 2495–2519. [184]
Ronald L. Graham, Henry O. Pollak (1972), On embedding graphs in squashed
cubes, in: Graph Theory and Appl., Proc. Conf. Western Mich. Univ., Kalamazoo, Springer
Lecture Notes in Math. 303 (1972), 99-110. [22, 23, 25, 106]
Ronald L. Graham, Bruce L. Rothschild, Joel H. Spencer (1980), Ramsey The-
ory, Wiley 1980. [85, 89]
Jack E. Graver (1975), Boolean designs and self-dual matroids, Lin. Alg. Appl. 10
(1975), 111–128. [5, 56]
Jack E. Graver, Wolfgang B. Jurkat (1973), The module structure of integral
designs, J. Combin. Th. A 15 (1973), 75–90. [180]
Curtis Greene, Daniel J. Kleitman (1976), On the structure of Sperner k-families,
J. Combin. Th. A 20 (1976), 41-68.
D. Greenwell, László Lovász (1974), Applications of product coloring, Acta Math.
Acad. Sci. Hungar. 25 (1974), 335–340.
Jerrold R. Griggs, J. W. Walker (1989), Anticlusters and intersecting families of
subsets, J. Combin. Th. A 51 (1989), 90–103.
Branko Grünbaum (1967), Convex Polytopes, Wiley–Interscience 1967 [71, 74]
Hugo Hadwiger (1944), Überdeckungssätze für den Euklidischen Raum, Portugaliae
Math. 4 (1944), 140–144. [112]
Hugo Hadwiger, P. Glur (1951), Zerlegungslleichheit ebener Polygone, Elem. Math. 6
(1951), 97–106. [18]
Willem Haemers (1979), On some problems of Lovász concerning the Shannon capacity
of a graph, IEEE Trans. Inform. Theory 25 (1979), 231-232. [147]
Willem Haemers (1981), An upper bound for the Shannon capacity of a graph, in:
Algebraic Methods in Graph Theory, Coll. Math. Soc. J. Bolyai, 25, 1981, pp. 267–272.
[147]
Roland Häggkvist, Pavol Hell, D. J. Miller, Vı́ctor Neumann-Lara (1988),
On multiplicative graphs and the product conjecture, Combinatorica 8 (1988), 63–74.
András Hajnal (1985), The chromatic number of the product of two ℵ1 -chromatic graphs
can be countable, Combinatorica 5 (1985), 137–139. [107]

——————————————————————— 223

c László Babai and Péter Frankl. 1988, 1992, 2020.
BIBLIOGRAPHY
Marshall Hall, Jr. (1967), Combinatorial Theory , Blaisdell, 1967.
R. R. Hall (1971), On pseudo-polynomials, Mathematika 18 (1971), 71–77. ((Sect.5.3))
[159]
Georg Hamel (1905), Eine Basis aller Zahlen und die unstetigen Lösungen der Funktion-
algleichung: f (x + y) = f (x) + f (y), Math. Ann. 60 (1905), 459–462. [21]
Frank Harary (1964), On the reconstruction of a graph from a collection of subgraphs, in:
Theory of Graphs and its Applications (M. Fieder, ed.) Czechoslovak Academy of Sciences,
Prague, 1964, pp. 47–52; reprinted by Academic Press, New York. [169]
Frank Harary (1969), Graph Theory, Addison–Wesley, 1969. [104]
Godfrey Harold Hardy, Edward Maitland Wright (1979), An Introduction to
the Theory of Numbers, Oxford Science Publ., 1979. [115]
E. F. Harding (1967) The number of partitions of a set of N points in k dimensions
induced by hyperplanes, Proc. Edinburgh Math. Soc. 15 (1967), 285–290. [163]
M. J. M. Hill (1895), Determination of the volume of certain species of tetrahedra without
employment of the method of limits, Proc. London Math. Soc. 27 (1895), 39–53. [18]
R. Hill, R. W. Irving (1982), On group partitions associated with lower bounds for
symmetric Ramsey numbers, European J. Comb. 3 (1982), 35–50.
Alan J. Hoffman, R. R. Singleton (1960), On Moore graphs with diameters 2 and 3,
IBM J. Res. Devel. 4 (1960), 497–504. [27]
Alan J. Hoffman (1975), Eigenvalues of graphs, in: Studies in Graph Theory, Part II
(D.R. Fulkerson, ed.), Math. Assoc. Amer. 1975, pp. 225–245.
József Kürschák, György Hajós, Gyula Neukomm, János Surányi (1963), Hun-
garian Problem Book, Random House N.Y. 1963.
Thomas William Hungerford (1974), Algebra, Springer 1974. [172]
Wilfried Imrich (1984), Explicit construction of regular graphs without small cycles,
Combinatorica 4 (1984), 53–59.
John R. Isbell (1959), An inequality for incidence matrices, Proc. Amer. Math. Soc. 10
(1959), 216–218. [82, 83]
François Jaeger, C. Payan (1971), Nombre maximal d’arétes d’un hypergraphe critique
de rang h, Comptes Rendus Acad. Sci. Paris 273 (1971), 221–223. [100]
S. Jimbo, A. Maruoka (1987), Expanders obtained from affine transformations, Combi-
natorica 7 (1987), 343–355.
Anatole Joffe (1974), On a set of almost deterministic k-independent variables, Ann.
Probability 2 (1974), 161–162.
Ferenc Juhász (1982), The asymptotic behaviour of Lovász’ ϑ function for random
graphs, Combinatorica 2 (1982), 153–155.

224 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 BIBLIOGRAPHY
Jeff Kahn, Gil Kalai (1992), A counterexample to Borsuk’s conjecture, manusript,
1992. [i, 117]
Gil Kalai (1984), Characterization of f -vectors of families of convex sets in Rd , Part I:
Necessity of Eckhoff’s condition, Israel J. Math. 48 (1984) 175–195. [104, 135]
Gil Kalai (1986), Characterization of f -vectors of families of convex sets in Rd , Part II:
Sufficiency of Eckhoff’s condition, J. Combin. Th. A 41 (1986) 167–188. [135]
Ravi Kannan, Achim Bachem (1979), Polynomial time algorithms for computing the
Smith and Hermite normal forms of an integer matrix, SIAM J. Comput. 8 (1979), 499–507.
William M. Kantor (1972), On incidence matrices of finite projective and affine spaces,
Math. Z. 124 (1972), 315–318. [180]
Irving Kaplansky (1977), Hilbert’s Problems, U. of Chicago, Lecture Notes in Math.
(1977). [16]
Narendra Karmarkar (1984), A new polynomial-time algorithm for linear program-
ming, Combinatorica 4 (1984), 373–395.
Gyula O. H. Katona (1964), Intersection theorems for systems of finite sets, Acta Math.
Acad. Sci. Hung. 15 (1964), 329–337. [167]
Gyula O. H. Katona (1967), A theorem of finite sets, in: Theory of Graphs, (Proc. Conf.
Tihany, Hungary, 1966), Akadémiai Kiadó, Budapest 1967, 187–207. [167]
Gyula O. H. Katona (1972), A simple proof of the Erdős–Ko–Rado theorem, J. Combin.
Th. B 13 (1972), 183–184. [97, 205]
Gyula O. H. Katona (1974), Solution of a problem of Ehrenfeucht and Mycielski, J.
Combin. Th. A 17 (1974), 265–266. [100]
Paul J. Kelly (1957), A congruence theorem for trees, Pacific J. Math. 7 (1957), 961–968.
[171]
Daniel J. Kleitman, K. J. Winston (1981), Forests and score vectors, Combinatorica
1 (1981), 49–54.
Daniel J. Kleitman, J. Shearer, Dean G. Sturtevant (1981), Intersections of
k-element sets, Combinatorica 1 (1981), 381–384.
Martin Kneser (1955), Aufgabe 300, Jber. Deutsch. Math. Verein. 58 (1955). [78]
Dénes König (1915), Line systems and determinants, Mat. Természettud. Értesı́tő 33
(1915), 221–229. (Hungarian)
Tom H. Koornwinder (1976), A note on the absolute bound for systems of lines, Proc.
Konink. Nederl. Akad. Wet. Ser A 79 (1976),152–153. [9]
Igor Křı́ž (1984), A class of dimension-skipping graphs, Combinatorica 4 (1984), 317–319.
[108]

——————————————————————— 225

c László Babai and Péter Frankl. 1988, 1992, 2020.
BIBLIOGRAPHY
Igor Křı́ž (1989), A hypergraph-free construction of highly chromatic graphs without
short cycles, Combinatorica 9 (1989), 227–230. [58]
Joseph Bernard Kruskal, Jr. (1963), The number of simplices in a complex, in: Math-
ematical Optimization Techniques, Univ. California Press, Berkeley 1963, pp. 251–278. [167]
Miklós Laczkovich (1990), Equidecomposability and discrepancy; a solution of Tarski’s
circle-squaring problem, J. reine angew. Math. 404 (1990), 77–117. [17]
√
David G. Larman (1978), A triangle free graph which cannot be 3-imbedded in any
Euclidean unit sphere, J. Combin. Th. A 24 (1978), 162–169. [106]
David G. Larman, Claude Ambrose Rogers (1972), The realization of distances
within sets in Euclidean space, Mathematika 19 (1972), 1–24. [10, 113, 121]
David G. Larman, Claude Ambrose Rogers, Johan Jacob Seidel (1977), On
two-distance sets in Euclidean space, Bull. London Math. Soc. 9 (1977), 261-267. [9, 11]
Bernt Lindström (1985), A Desarguesian theorem for algebraic combinatorial geome-
tries, Combinatorica 5 (1985), 237–239.
Bernt Lindström (1988), A generalization of the Ingleton–Main lemma and a class of
non-algebraic matroids, Combinatorica 8 (1988), 87–90.
Nathan Linial, László Lovász, Avi Wigderson (1988), Rubber bands, convex em-
beddings and graph connectivity, Combinatorica 8 (1988), 91–102.
Nathan Linial, Bruce L. Rothschild (1981), Incidence matrices of subsets – a rank
formula, SIAM J. Alg. Discr. Methods 4 (1981), 333–340. [157]
Jacobus H. van Lint, Alexander Schrijver (1981), Construction of strongly regular
graphs, two-weight codes, and partial geometries by finite fields, Combinatorica 1 (1981),
63–73.
Nick Littlestone (1987), Learning qickly when irrelevant attributes abound: a new
linear–threshold algorithm, Machine Learning 2 (1987), 285–318. [160]
László Lovász (1968), On chromatic number of finite set-systems Acta Math. Acad.
Scient. Hung. 19 (1968), 59–67. [58]
László Lovász (1972), A note on the line reconstruction problem, J. Combin. Th. B 13
(1972), 309–310. [170]
László Lovász (1973), Independent sets in critical chromatic graphs, Studia Sci. Math.
Hungar. 8 (1973), 165–168. [173, 174]
László Lovász (1977a), A homology theory for spanning trees of a graph, Acta Math.
Acad. Sci. Hung. 30 (1977), 241–251.
László Lovász (1977b), Flats in matroids and geometric graphs, in: Combinatorial sur-
veys, Proc. 6th British Comb. Conf., Egham 1977 (P.J. Cameron, ed.), Acad. Press, London
1977, pp. 45–86. [100, 102, 135, 143]

226 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 BIBLIOGRAPHY
László Lovász (1978a), Kneser’s conjecture, chromatic number, and homotopy, J. Com-
bin. Th. 25 (1978), 319–324. [78]
László Lovász (1978b), Some finite basis theorems in graph theory, in: Combinatorics,
Vol. II (A. Hajnal, V.T. Sós, eds.), Colloq. Math. Soc. J. Bolyai 18, North-Holland 1978, pp.
717–729. [100]
László Lovász (1979a), Topological and algebraic methods in graph theory, in: Graph
Theory and Related Topics, Proc. Conf. in Honor of W.T. Tutte, Waterloo 1977 (J.A. Bondy,
U.S.R. Murty, eds.), Academic Press, New York 1979, pp. 1–15.
László Lovász (1979b), On the Shannon capacity of a graph, IEEE Trans. Inform. Theory
25 (1979), 1-7. [148]
László Lovász (1979c), Combinatorial Problems and Exercises, Akadémiai Kiadó, Bu-
dapest and North-Holland, Amsterdam, 1979. [iii, xiv, 71, 83, 90, 95, 170, 184]
László Lovász (1979d), On determinants, matchings, and random algorithms, in: Fun-
damentals of Computation Theory, FCT’79 (L. Budach, ed.), Berlin 1979, Akademie–Verlag,
pp. 565–574.
László Lovász, Jaroslav Nešetřil, Alev s Pultr (1980), On a product dimension
of graphs, J. Combin. Th. B 29 (1980), 47-67. [105, 107]
D. Lubell (1966), A short proof of Sperner’s lemma, J. Combin. Th. 1 (1966), 299. [96]
Alex Lubotzky, Ralph S. Phillips, Peter Sarnak (1988), Ramanujan graphs, Com-
binatorica 8 (1988), 261–277. [58]
Wolfgang Maass, György Turán (1989), On the complexity of learning from coun-
terexamples, Proc. 30th IEEE Symp. on Found. Comp. Sci., Research Triangle Park, NC,
pp. 262–267. [164]
Wolfgang Maass, György Turán (1990), On the complexity of learning from coun-
terexamples and memberhsip queries, Proc. 31st IEEE Symp. on Found. Comp. Sci., St.
Louis MI, pp. 203–210. [160]
Florence Jessie Macwilliams, Neil J. A. Sloane (1978), The Theory of Error-
correcting Codes, North-Holland, Amsterdam 1978.
H. Maehara (1984), Space graphs and sphericity, Discrete Appl. Math. 7 (1984), 55–64.
[106]
K. N. Majumdar (1953), On some theorems in combinatorics relating to incomplete block
designs, Ann. Math. Stat. 24 (1953), 377–389. [82, 83]
Marvin Marcus, Henryk Minc (1964), A survey of Matrix Theory and Matrix Inequal-
ities, Allyn & Bacon, Boston, 1964.
Grigory A. Margulis (1982), Explicit constructions of graphs without short cycles and
low density codes, Combinatorica 2 (1982), 71–78.

——————————————————————— 227

c László Babai and Péter Frankl. 1988, 1992, 2020.
BIBLIOGRAPHY
Grigory A. Margulis (1988), Explicit group theoretic construction of combinatorial
schemes and their application for the construction of expanders and concentrators, J. Prob-
lems of Information Transmission (1988) (Russian) [58]
Yury V. Matiyasevich, (1970), Recursively enumerable sets are Diophantine, Dokl.
Akad. Nauk SSSR 191 (1970), 279–282. (Russian) [14]
William S. Massey (1977), Algebraic Topology, An Introduction, Springer-Verlag, New
York, 1977.
Peter McMullen (1970), The maximum number of faces of a convex polytope, Mathe-
matika 17 (1970), 179–184. [75]
Peter McMullen, Geoffrey Colin Shephard (1971), Convex Polytopes and the
Upper Bound Conjecture, London Math. Soc. Lect. Note 3, Cambridge Univ. Press 1971.
Vladimı́r Müller (1977), The edge reconstruction hypothesis is true for graphs with
more than n log2 n edges, J. Combin. Th. B 22 (1977), 281–283. [170, 170]
Ketan Mulmuley (1987), A fast parallel algorithm to compute the rank of a matrix over
an arbitrary field, Combinatorica 7 (1987), 101–104.
Ketan Mulmuley, Umesh V. Vazirani, Vijay V. Vazirani (1987), Matching is as
easy as matrix inversion, Combinatorica 7 (1987), 105–113.
Zsigmond Nagy (1972), A certain constructive estimate of the Ramsey number (Hungar-
ian), Matematikai Lapok 23 (1972), 301–302 [87]
C. St. J. A. Nash-Williams (1978), The reconstruction problem, in: Selected Topics in
Graph Theory (L.W. Beineke, R.J. Wilson, eds.), Academic Press 1978. [170]
Arnold Neumaier (1980), t 21 -designs, J. Combin. Th. A 28 (1980), 226–248.
Arnold Neumaier (1980), Distances, graphs and designs, European J. Comb. 1 (1980),
163–174.
Jaroslav Nešetřil, Aleš Pultr (1977), A Dushnik–Miller type dimension of graphs
and its complexity, in: Fundamentals of Computation Theory, Proc. Conf. Poznań–Kórnik,
1977, Springer Lecture Notes in Comp. Sci. 56, Springer 1977, pp. 482–493. [104]
Ivan Niven (1947), A simple proof that π is irrational, Bull. Amer. Math. Soc. 53 (1947),
509. [19]
Andrew M. Odlyzko (1981), On the ranks of some (0,1)-matrices with constant row
sums, J. Austral. Math. Soc. 31 (1981), 193–201.
G. W. Peck (1979), Maximum antichains of rectangular arrays, J. Combin. Th. A 2 7
(1979), 397-400.
Micha Perles, Saharon Shelah (1972), see: S. Shelah (1972) [160]
J. R. Pierce (1972), Network for block switching of data, Bell Systems Tech. J. 51 (1972),
1133–1145. [23]

228 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 BIBLIOGRAPHY
J. E. Pin (1981), On two combinatorial problems arising from automata theory, in: Combi-
natorial Mathematics (Marseille-Luminy, 1981) North-Hollandm, 1983, pp. 535–548. [102]
Lev Semyonovich Pontriagin (1952), Foundations of Combinatorial Topology, Gray-
lock Press, Rochester NY, 1952. [105, 173]
R. A. Proctor (1982), Solution of two difficult problems with linear algebra, Amer. Math.
Monthly 89 (1982), 721–734.
R. A. Proctor, Michael E. Saks, Dean G. Sturtevant (1980), Product partial
orders with the Sperner property, Discrete Math. 30 (1980), 173–180.
Pavel Pudlák, Vojtěch Rödl (1992), A combinatorial approach to complexity, Com-
binatorica 12 (1992), 221–226. [146, 189]
László Pyber (1987), Hamiltonian graphs are edge-reconstructible. [171]
Johann Radon (1921), Mengen konvexer Körper, die einen gemeinsamen Punkt erhalten,
Math. Ann. 83 (1921), 113–115. [72, 163]
Frank Plumpton Ramsey (1930), On a problem of formal logic, Proc. London Math.
Soc. 30 (1930), 264-286. [86]
Dijen K. Ray-Chaudhuri, Richard M. Wilson (1975), On t-designs, Osaka J. Math.
12 (1975), 737-744. [91, 92, 108, 128, 155]
Alexander A. Razborov (1987), Lower bounds for the size of circuits of bounded depth
with basis AND, XOR, Matem. Zametki 41 (1987), 598–607 (Russian). (English translation
in: Mathem. Notes of the Acad. Sci. USSR 41:4, 333–338. [189]
Alexander A. Razborov (1990), Applications of matrix methods to the theory of lower
bounds in computational complexity, Combinatorica 10 (1990), 81–93. [189]
Jan Reiterman, Vojtěch Rödl, Edita Šiňajová (1988), On the scalar product di-
mension of graphs, Discrete Math. 74 (1989), 291–319. [106]
Neil Robertson, Paul D. Seymour (1985), Graph minors – a survey, in: Surveys in
Combinatorics (I. Anderson, ed.), Cambridge Univ. Press, 1985, pp. 153–171. [172]
Vojtěch Rödl (1984), On combinatorial properties of spheres in Euclidean spaces, Com-
binatorica 4 (1984), 345–349. [106]
Gian-Carlo Rota (1964), On the Foundations of Combinatorial Theory I. Theory of
Möbius Functions, Zeitschr. für Wahrscheinlichkeitstheorie und Verwandte Gebiete 2 (1964),
340–368. [183]
Imre Z. Ruzsa (1972), On congruence-preserving functions (Hungarian, with English sum-
mary), Matematikai Lapok 22 (1971), 125–134 (1972); MR 48# 2044. [159]
Herbert J. Ryser (1950), A note on a combinatorial problem, Proc. A.M.S. 1 (1950),
422–424. [85]
Herbert J. Ryser (1963), Combinatorial Mathematics, Carus Math. Monograph 14,
Math. Assoc. of America, 1963.

——————————————————————— 229

c László Babai and Péter Frankl. 1988, 1992, 2020.
BIBLIOGRAPHY
Herbert J. Ryser (1968), An extention of a theorem of de Bruijn and Erdős on combi-
natorial designs, J. Algebra 10 (1968), 246–261. [84, 85]
Norbert Sauer (1972), On the density of families of sets, J. Combin. Th. A 13 (1972),
145–147. [160]
Ludwig Schläfli (1852), Theorie der vielfachen Kontinuität, in: Gesamm. Math. Abh.
von L. Schläfli, Basel 1950. See Vol. I, p. 209. [163]
Wolfgang M. Schmidt (1976), Equations over Finite Fields: An Elementary Approach,
Springer Lect. Notes in Math. 536, 1976.
Oded Schramm (1988), Illuminating sets of constant width, Mathematika 35 (1988),
180–189. [117]
Alexander Schrijver, (1978), Vertex-critical subgraphs of Kneser graphs, Nieuw Arch.
Wiskunde 26 (1978), 454–461. [80]
Alexander Schrijver, ed. (1979), Packing and Covering in Combinatorics, Mathemat-
ical Centre Tracts 108, Amsterdam 1979.
Alexander Schrijver (1981), Association schemes and the Shannon capacity: Eberlein-
polynomials and the Erdős–Ko–Rado theorem, in: Algebraic Methods in Graph Theory (L.
Lovász, V.T. Sós, eds.) Vol. I, North-Holland 1981, pp. 671–689. [147]
Issai Schur (1916), Über die Kongruenz xm + y m ≡ z m (modp) Jber. Deutsch. Math.-
Verein. 25 114–117. [89]
Jacob T. Schwartz (1980), Fast probabilistic algorithms for verification of polynomial
identities, J. Assoc. Computing Machinery 27 (1980), 701–717. [63, 66, 68]
Ákos Seress (1989), Some characterizations of type-1 λ-designs J. Combin. Th. A 52
(1989), 288–300. [84, 85]
Paul Seymour (1974), On the two-coloring of hypergraphs, Quarterly J. Math Oxford
(2nd Ser.) 25 (1974), 303–313. [25]
L. Shader (1976), All right triangles are Ramsey in E2 !, J. Combin. Th. A 20 (1976),
385-389.
Claude E. Shannon (1956), The zero-error capacity of a noisy channel, IRE Trans.
Inform. Theory 3 (1956), 3–15. [148]
Saharon Shelah (1972), A combinatorial problem; stability and order for models and
theories in infinitary languages, Pacific. J. Math. 41 (1972), 247–261. [160]
Sharadchandra Shankar Shrikhande, Navin M. Singhi (1976), On the λ-design
conjecture, Utilitas Math. 9 (1976), 301–313.
Roman Smolenski (1987), Algebraic methods in the theory of lower bounds for boolean
circuit complexity, in: Proc. 19th ACM Symp. on Theory of Computing, 1987, pp. 77–82.
[189]

230 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 BIBLIOGRAPHY
Robert M. Solovay, Volker Strassen (1977), A fast Monte-Carlo test for primality,
SIAM J. Comput. 6 (1977), 84–85.
Vera T. Sós (1976), Remarks on the connection of graph theory, finite geometry and
block designs, in: Colloq. Intern. sulle Teorie Combin., Roma 1973, Acad. Naz. Lincei, Roma
1976, Vol. 2, pp. 3–17.
Edwin H. Spanier (1977), Algebraic Topology, Springer-Verlag, New York, 1977. [105]
Joel H. Spencer (1975), Ramsey’s theorem – a new lower bound, J. Combin. Th. A 18
(1975), 108–115.
Joel H. Spencer (1977), Asymptotic lower bounds for Ramsey functions, Discrete Math.
20 (1977), 69–76. [87]
Joel H. Spencer (1986), Balancing vectors in the max norm, Combinatorica 6 (1986),
55–65.
Emanuel Sperner (1928), Ein Satz über Untermengen einer endlichen Menge, Math. Z.
27 (1928), 544–548. [96]
Richard P. Stanley (1975), The upper bound conjecture and Cohen–Macauley rings,
Studies in Apll. Math. LIV (1975), 135–142. [76]
Richard P. Stanley (1980), Weyl groups, the hard Lefschetz theorem, and the Sperner
property, SIAM J. Alg. Discr. Methods 1 (1980), 168–184. [104]
Richard P. Stanley (1983), Combinatorics and Commutative Algebra, Birkhäuser,
Boston, 1983. [104]
Richard P. Stanley (1984), Quotients to Peck posets, Order 1 (1984), 29–34.
Volker Strassen (1969), Gaussian elimination is not optimal, Numerische Mathematik
13 (1969), 354–356.
Volker Strassen (1973), Vermeidung von Divisionen, J. Reine Angew Math. 264 (1973),
184–202. [69]
László A. Székely (1984), Measurable chromatic number of geometric graphs and sets
without some distances in Euclidean space, Combinatorica 4 (1984), 213–218.
László A. Székely, Nicholas C. Wormald (1989), Bounds on the measurable chro-
matic number of Rn , Discrete Math. 75 (1989), 343–372. [10]
Endre Szemerédi (1975), On sets of integers containing no k elements in arithmetic
progression, Acta Arithm. 27 (1975), 199-245.
Tamás Tarján (1975), Complexity of lattice-configurations, Studia Sci. Math. Hungar.
10 (1975), 203–211.
Pál Turán (1941), Egy gráfelméleti szélsőértékfeladatról (On an extremal problem in
graph theory), Math. Phys. Lapok 48 (1941), 436–452 (Hungarian) [82]

——————————————————————— 231

c László Babai and Péter Frankl. 1988, 1992, 2020.
BIBLIOGRAPHY
William T. Tutte (1947), The factorization of linear graphs, J. London Math. Soc. 22
(1947), 107–111.
William T. Tutte (1979), All the king’s horses, in: Graph Theory and Related Topics
(J. A. Bondy, U. S. A. Murty, eds.), Academic Press, 1979, pp. 15–33. [171]
Zsolt Tuza (1984), Helly-type hypergraphs and Sperner families, European J. Comb. 5
(1984), 185–187.
Leslie G. Valiant (1979), The complexity of computing the permanent, Theor. Comp.
Sci. 8 (1979), 189–201.
Leslie G. Valiant (1984), A theory of learnable, Commun. ACM 27 (1984), 1134–1142.
V. N. Vapnik, A. Ya. Chervonenkis (1971), On the uniform convergence of relative
frequencies of events to their probabilities, Theory of Probability and Appl. XVI (1971),
264–280. [160]
B. L. van der Waerden (1926), Aufgabe 45, Jber. Deutsch. Math. Verein. 35 (1926),
117.
B. L. van der Waerden (1927), Beweis einer Baudetschen Vermutung, Nieuw Arch.
Wisk. 15 (1927), 212–216. [85]
Stanley Wagon (1985), The Banach–Tarski Paradox, Cambridge University Press, New
York, 1985. [17]
A. Weiss (1984), Girth of bipartite sextet graphs, Combinatorica 4 (1984), 241–245.
B. Weiss (1989), A combinatorial proof of the Borsuk–Ulam antipodal point theorem Israel
J. of Math. 66 (1989), 364–368. [77]
J. A. Dominic Welsh (1976), Matroid Theory, Acad. Press, London, 1976.
J. A. Dominic Welsh (1983), Randomised algorithms, Discrete Appl. Math. 5 (1983),
133–145.
Herbert S. Wilf (1968), Hadamard determinants, Möbius functions, and the chromatic num-
ber of a graph. Bull. Amer. Math. Soc. 74 (1968), 960–964. [184]
Richard M. Wilson (1973), The necessary conditions for t-designs are sufficient for
something, Utilitas Math. 4 (1973), 207–215.
Richard M. Wilson (1983), Inequalities for t-designs, J. Combin. Th. A 34 (1983),
313–324.
Richard M. Wilson (1984), The exact bound in the Erdős–Ko–Rado theorem, Combi-
natorica 4 (1984), 247–257. [147]
Peter M. Winkler (1983), Proof of the squashed cube conjecture, Combinatorica 3
(1983), 135–139. [23, 23, 25]
Peter M. Winkler (1987), The metric structure of graphs: theory and applications in:
Surveys in Combinatorics 1987 (C. Whitehead, ed.) London Math. Soc. Lect. Note 123,
Cambridge Univ. Press 1987, pp. 197–221. [24]

232 ———————————————————————

c László Babai and Péter Frankl. 1988, 1992, 2020.
 BIBLIOGRAPHY
P. de Witte (1967), Some new properties of semi-tactical λ-spaces, Bull. Soc. Math. Belg.
19 (1967), 13–24.
D. R. Woodall (1970), Square λ-linked designs, Proc. London Math. Soc. 20 (1970),
669–687. [84]
Thomas Zaslavsky (1975), Facing up to arrangements: for partitions of space by hyper-
planes, Memoirs of the American Mathematical Society, Number 154, Volume 1, 1–99.

——————————————————————— 233

c László Babai and Péter Frankl. 1988, 1992, 2020.