LIN] KAR ALG] EBRA MI ETHODS IN COMBINATORICS with Applications to Geometry and Compute Science LASZLO BABAI PETER FRANKL Department of Computer Science The University of Chicago Preliminary Version 2 (September 1992)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 years. The combinatorial nature of the determinant expansion (and the related difficulty in teaching it) may hint for 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 Cayley in the last century on counting spanning trees in a network. It is much less known, however, that quite apart from the theory of determinants, the elements of the theory of linear spaces have striking ap- plications to the theory of families of finite sets. With a mere knowledge of the concept of linear independence, completely 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 at 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 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 recent disproof of Borsuk’s 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 cen- trally 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 by a trifle: it produced an infinite family where the minimum number of pieces grew as an exponential function of /7i. Yet the proof took only a page to describe, with reference to a combinatorial result: which occupies a central place in this book. 1 Babai-Frankl: Linear Algebra Methods in Combinatorics. © Lésalé Babai and Péter Frankl. September 1992. 1 Sections 5.4 and 5.6 together give a complete and self-contained proof of this sur-Rather than presenting as many results as possible, we have concen- trated on developing techniques and showing different methods to yield different proofs and a variety of generalizations to a small set of focal re- sults, 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 “flints”; and full solutions are given in an appendix (“Answers to exer- cises”) to those exercises marked with a diamond (%). Asteriscs indicate the degree of difficulty. Most results are motivated by applications. This book is geometry all over. Applications to the theory of computing ae prominent in several sec- tions (computational learning theory (Section 7.4), communication com- plexity 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 connec- tion 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; and in the few successful attempts on record, methods of algebra and num- ber theory have been the winners. In this volume, explicit Ramsey graph constructions (Sections 4.2, 5.7) serve as simple illustrations of the phe- nomenon. 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 all this, naturally, the prime application area of the meth- ods presented remains combinatorics, especially the theory of extremal set systems. We have made an effort to motivate each combinatorial applica- tion 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 exposure 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 mod p where p is a prime number; and the term “field of characteristic zero” to mean the domain Q of rational 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 to minimize the overlap with existing texts and mono- graphs, we have omitted major areas that would fit the title of the book The most painful omission is that of eigenvalue methods. However, excel- lent 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 Lovasz (1979), Biggs (1974), Godsil (1989), Brouwer, Co- prising result.iii hen, Neumaier (1980), Bannai, Ito (1984)). The list of major relevant areas omitted or hardly touched upon in the present volume includes eigenvalue techniques in the study of highly regular structures (strong- ly regular graphs, association schemes, designs, finite geometries); algebraic theory of error correcting codes; — eigenvalue techniques in the study of properties of graphs such as connectivity and expansion, 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 tech- niques in various models 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 researchers working in discrete mathematics, discrete geome- try, the theory of computing, applications 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, leading to research level open problems. Some other recipes for classroom use: ¢ Use Chapters 1-5 embedded in an introductory course on combina- torics. 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. Use Chapter 6 to introduce and motivate wedge products. Select material from Chapters 3, 5, and 6 to show delightful applica- tions 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 introductory 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 Bardny of Kneser’s Conjecture on the chromatic number of certain graphs. A puzzling 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 combina- torics. A combination with one of the following is especially recommended: 1. Anderson, Combinatorics of finite sets, Oxford University Press, 1987; : B. Bollobés, Combinatorics, Cambridge University Press 1986; or in an advanced seminar with Chapters 8, 9, and 13 of L. Lovdsz, 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 Eétvés 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 Friedi, Robert Freud, Albert Goodman, Barry Guiduli, Péter Hajnal, Zsolt Hats4gi, Penny Haxell, Anna Lubiw, Akos Seress, Mario Szegedy, Laszlé Székely, Norihide Tokushige, and Maté Wierd] for their generous support. In addition to the institutions mentioned above, special thanks are due to the Institute fir Okonometrie und Operations Research, Universitat 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 improving communication between the coauthors. Thanks are due to Richard Carnes of the University of Chicago for making his TpXnical 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 teach- ers and mentors at Hétvés 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 Pal Erdés, everybody’s “Pali bacsi” [paw-lee but-chee] who took interest in us at our epsilon age; Pal Turan, Vera T. Sés, Andrés Hajnal, Gyula O. A. Katona, Lajos Pésa, Laszlé Lovasz. We feel fortunate to have grown up in the mathematical environment they helped create. For the sake of later improved versions of this volume, we ask the readers to send comments to L. Babai Department of Computer Science University of Chicago 1100 E 58th St Chicago, IL 60637.Contents Preface , i Notation and terminology 3 1 Warm-up 7 1.1 Counting clubs in Oddtown ............0.000., 7 Exercises 22... 2... eee eee Lee. 10 1.2 Point sets in R" with only two distances... ........~ 13 Exercises 2.02... ..0.., i 4 1.3 Two solutions to a jigsaw puzzle? . . . 7 Exercises . 2.0... 21 14 Addressing into the squashed cube... . 25 Exercises ....... Lees . 28 1.5 Beauty israre............. . . 2% 2 Basic linear algebra and combinatorics 33 2.1 A guide to basic abstract algebra . . . 33 2.1.1 Fundamental structures 33 2.1.2 Polynomials . . .. 35 2.1.3 Limear spaces»... 2... 36 2.1.4 Criteria of linear independence ........ 38 Exercises 40 2.2 Affine subspaces, linear equations, rank... .... 2... 44 2.2.1 Inequalities for subspaces 44 2.2.2 Linear maps. . 45 2.2.3 Matrices, rank a 45 2.2.4 Systems of linear equations. Affine subspaces . 47 2.2.5 Projective spaces... .......... er) 2.2.6 Extending the field... .. Lae 50 Exercises . 52 2.3 Orthogonality . . 52 23.1 Inner product spaces 52 2.3.2. Eventown revisited . . . 54 Exercises . 55 2.4 Graphs and set systems ee) 24.1 Notation, terminology ............ . 2 56 2.4.2 Chromatic number and short cycles 57 2.4.3. Block designs 58 3 “General position” arguments 59 3.1 Configurations in general position. ........ 59 3.1.1 Points in general position. The moment « 59 3.1.2. Subspace in general position w.r.t. a family of subspaces 60 viiviii 3.1.3 Linear maps in general position . 3. Exercises 3.2 Convexity . 3.2.1 Terminology Helly’s ‘Theorem A polytope with Exercises... .. 3.2.6 Linear and statistical independence many faces Distributing points on the sphere . . Borsuk’s and Kneser’s graphs . . . . 4 Set systems with restricted intersections 4.1 When all intersections are equal size . . . Exercises 2... . 4.2 Ramsey theory - a constructive lower bound Exercises 4.3 Restricted intersections Exercises... .... 4.4 Extremal set theory: the classics 5 Spaces of polynomials 5.1 Helly-type theorems for Exercises 2... 5.2 Resultants... 2... finite sets 5.3 The Prague dimension of graphs Exercises... ..- 5.4 Sets with few intersection sizes mod p Exercises CONTENTS ‘4 Checking identities: a Monte Carlo algorithm . 5.5 Geometric application: unit distance is hard to miss Exercises 5.6 Reducing the diameter proved . . Exercises of bodies: Borsuk’s conjecture dis- 5.7 Constructive Ramsey graphs via intersection theorems |. - Exercises... 2... 5.8 Geometric application: any distance is hard to miss Exercises 2... 0. 5.9 Prime power moduli Exercises .. 2... 5.10 The nonuniform RW Theorem Exercises 2.2... 5.11 The Ray-Chaudhuri ~ Wilson Theorem . Exercises Exercises... 2... 6 Tensor product methods 6.1 Wedge products — a concrete introduction .1 The Laplace expansion of determinants 6.1.2 Alternating k-linear functions... . 6.1 6.1.3 Exterior powers Exercises 6.2. Bollobas-type theorems 5.12 A modular Ray-' Cl haudhuri - of EF" Wilson Theorem. ....... 62 64 66 68 “68 69 70 72 73 75 76 77 120 123 125 126 126 126 128 130 132CONTENTS ix Exercises... Lees rn 135 6.3 Symmetric products Leas . 135 6.4 The Shannon capacity of agraph................ 135 7 A class of higher incidence matrices: the inclusion matrix 137 7.1 The inclusion matrix; s-independent families ....... . 137 Exercises 26... 0.000000. 139 7.2. Extended inclnsion matrices. ‘The Nonuniform RW Theorem revisited . 139 Exercises 11 7.3. Inclusion matrices of uniform families. 142 Exercises... .......000. 146 7.4 Linear dependencies among the rows of inclusion matrices and the Vapnik-Chervonenkis dimension... .. . . 146 Exercises»... 0.0.0... beeen eee ee 149 7.5 Shadows of s-independent families... |... vee TBI Exercises 6... eee ne Cy 8 Applications of inclusion matrices 155 8.1 The edge-reconstruction problem ............. 155 Exercises Lae 157 8.2 Chromatic critical sraphs 157 Exercises . 159 8.3. Partially ordered sets, unimodal sequences, and the Sperner property... 2.0.0... 160 Exercises 162 9 Partially ordered sets 163 9.1 Geometric semilattices .... . . Lee eee 163 91.1 Matroids 2.2.2... 2000000008. . 163 9.1.2 Geometric lattices .. 2... « Lee ... 168 9.1.3 RW-type theorems for semilattices ...... 0... 163 9.2. Incidence matrices of full rank... .. . en 163 Exercises. . eee ee eee 165 9.3 The Mébius function... ..........0- 0. 165 9.3.1 Mébius inversion... [Link]... 000005 165 9.3.2 ‘The Mébius function in geometric lattices... .. 166 9.3.3 Whitney number inequalities... . . 167 9.3.4 The VC dimension revisited: shattered elements in a poset... 6... eee eee ee 168 10 Applications to the Theory of Computing 171 10.1 Communication complexity theory ....... 0... 171 10.2 Overview oo. . 0... eee ee Lees .. 171 Answers to the exercises 173 AA Chapter... 00. eee .. 18 AQ Chapterr2........ eee eee eee 181 A3 Chapter3.............00.. 183 A4 Chapter 4 . . 184 A.5 Chapter 5 185 A.6 Chapter 6 190 A.7 Chapter 7 190 A8 Chapter 8 192 A.9 Chapter 9 192Index 193 Bibliography 199List of Figures A6 Clubs in Oddtown 2.2.00... 00000 Lee 8 Squaring a regular triangle: Dudeney’s 4-piece jigsaw puzzle 25 Three minimal decompositions of Ky... ... . 2 26 Squashing the cube. ............. ne) The complete graph K'4 as a squashed 3-cube. . . . .. 28 Petersen’s graph. (Which one is it?) ........ .... 80 The multiplication table of Fa... ......0000. .. 4i (a) The Fano plane. (b) The anti-Fano configuration. .. 52 Four Mod-4-town clubs which are dependent mod 2... .. 175 Squaring a rectangle by dissec A good cut for the pastry chef. The Fano plane, coordinatized Six vectors in general position in F3 Three colors don’t suffice. 183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}. 2* is the set of all subsets of the set X. If |X| =n then |2*| = 2”. ({) denotes the set of all k-subsets of the set X (k > 0) If |X| =n then /=@) (Z,) denotes the set of all subsets of X of size < k. If |X| = n then (2) = rho (9). We call the set {0,1}" := {(41,.--,€n) : € € {0,1}} the n-cube. It has 2” points. The incidence vector of the set A C [n] is (a1,..., an) € {0,1}", where aafl ified; ‘Vo itig a. Strings S$” denotes the set of ordered n-tuples from the set S. In this context we sometimes refer to $ as the alphabet; the members of S" are strings of length n over S. The Hamming distance of two strings x = (21,...,2n) € S" and y = (y1,-+.,9n) € S" is defined as the number of places i where x; # y;. 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” is the Hamming-distance of z and (0,...,0) € S". If S is a subset of an abelian group then the Hamming distance of 2, y € S” is the weight of ¢ — 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 A; 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. Let L be aset of integers. A set system F is L-intersecting if |ENF|€ L for any two distinct E, F € F. (For a modular version of this concept, see Def. 5.14.)4 NOTATION AND TERMINOLOGY Graphs A graph G = (V,B) is a pair consisting of a set V of vertices and a 2-uniform set system EZ 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 (8). If all the (3) 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, £) is another graph G on the same vertex set, having complementary edge-set: G = (V,B) where B = (4) \ L. The adjacency matrix of G = ((n], E) is ann x n (0, 1)-matrix A = (aij) where aij = 1 if vertices i and j are adjacent; ai; = 0 otherwise. Note that A= AT? and the diagonal elements of A are aij = 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 Ks. ‘A cycle of length k in a graph is a sequence of k > 3 distinct vertices (a1,...,a,) such that each a; is adjacent to aj41 (i = 1,...,k—1), and ax 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 Sec- tion 2.4.1, For the elements of graph theory we refer the reader to Bol- lobas (1979) or Bondy-Murty (1976). For more advanced problem-solvers, Lovasz (1979c) is an invaluable source. Fields, rings, matrices Z ~ set of integers @Q ~ set of rational numbers B ~ set of real numbers C ~ set of complex numbers F, ~ field of g elements (same as GF(q); q is a prime power). If ¢ = p is a prime, then F, can be viewed as the set {0,1,...,p—1} with addition and multiplication performed modulo p. F - any field F(z] — ting polynomials in one indeterminate over F E[21,...,2n] ~ ring of polynomials in n indeterminates over F DEX" — set of k x n matrices over the domain D D" = D"*! — 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 Tn ~ nx n identity matrix Jn — nx n all-ones matrix (every entry is 1) le] = S32 oF — Euclidean norm of a vector x = (a1,...,@m) € R” |Je|oo = max"; a; — maximum norm Multivariate polynomialsmonomial ~ a product of variables with a scalar coefficient, e. g., 3032423 degree of monomial - sum of the exponents polynomial of degree # — maximum degree of the monomials appearing in its full expansion. ‘The degree of the zero polynomial is —co homogeneous polynomial of degree k ~ a sum of monomials of degree & monic monomial ~ monomial with coefficient 1, e. g., e3x423 multilinear polynomial — the degree in each variable is < 1,e. g., 3¢124— Sa2tar7 More algebra terminology can be found in Sections 2.1 and 2.2 Comparison of orders of magnitude! of two sequences {an} and {by} of real numbers: an = O(bn) (“dig-oh of bn”) means |an| < Cbn for some constant C’ and all sufficiently large n; an = bn) means by = O(an), i.¢., dn > ¢[bn| for some constant ¢ > 0 and all sufficiently large n; Gn = O(bn) means an = O(bn) and ay = (bn) simultaneously, i.e., 0 7, there exist nonisomorphic extremal solutions to the Eventown problem. (An eztremal solution is a system of maximum possible number of clubs, i.e., m = 2'"/7J_ 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. OEx.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 Hint. Use projective planes (Sec. 2.4.3). OEx. 1.1.14 (M. Szegedy, 1988). Show that if n is even then there exist at least an(4+2)/8 /(n!)? nonisomorphic extremal solutions to the Oddtown problem. (This is a very large number: for large 1, it is greater than 2”’/°. (Prove!)) Prove that there are no more than 2"” /n! extremal solutions. The next sequence of problems investigates the connections between linear independence over various fields. OEx.1.1.15. Let v1,...,%m be vectors with n rational entries. Prove that the v; are linearly independent over Q if and only if they are linearly independent over R. Ex. 1.1.16. Let v1,....um 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. OEx.1.1.17. Given an integer d, exhibit a square (0, 1)-matrix with determi- nant d. Ex. 1.1.18. Construct a set of (0,1)-vectors, linearly independent over Q and F but linearly dependent over Fz and Fs. Ex. 1.1.19. Prove: if a set of (0,1)-vectors is linearly independent over @, then they are linarly independent over Fp for every sufficiently large prime p