Bounds on extremal functions of forbidden patterns by Jesse Geneson WW Submitted to the Department of Mathematics in partial fulfillment of the requirements for the degree of U z __J C WI:; Doctor of Philosophy U <0 at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY June 2015 @ Massachusetts Institute of Technology 2015. All rights reserved. Signature redacted A uthor ....................... Department of Mathematics May 10, 2015 Signature redacted Certified by................. Peter Shor Professor Thesis Supervisor Signature redacted Accepted by .................. Michel Goemans Chairman, Department Committee on Graduate Theses U)2 C) (I cc C0 I? Bounds on extremal functions of forbidden patterns by Jesse Geneson Submitted to the Department of Mathematics on May 10, 2015, in partial fulfillment of the requirements for the degree of Doctor of Philosophy Abstract Extremal functions of forbidden sequences and 0 - 1 matrices have applications to many problems in discrete geometry and enumerative combinatorics. We present a new computational method for deriving upper bounds on extremal functions of forbidden sequences. Then we use this method to prove tight bounds on the extremal functions of sequences of the form (12 ... 1)' for 1 > 2 and t > 1, abc(acb)t for t > 0, and avav'a, such that a is a letter, v is a nonempty sequence excluding a with no repeated letters and v' is obtained from v by only moving the first letter of v to another place in v. We also prove the existence of infinitely many forbidden 0 - 1 matrices P with non-linear extremal functions for which every strict submatrix of P has a linear extremal function. Then we show that for every d-dimensional permutation matrix P with k ones, the maximum number of ones in a d-dimensional matrix of sidelength n that avoids P is 20(k) dThesis Supervisor: Peter Shor Title: Professor 3 4 Acknowledgments Thanks to Peter Shor for advising my thesis research and for improving the bounds on S2 (m) in Lemma 11. Thanks also to Jacob Fox for research advice and for help proving lower bounds on extremal functions of multidimensional permutation matrices. I also thank Henry Cohn for comments to improve the clarity of the 0 - 1 matrix proofs and for ideas about formation width. Also I thank Peter Shor, Jacob Fox, and Henry Cohn for being on my thesis committee. In addition, I thank Joe Gallian for introducing me to extremal functions of 0 - 1 matrices. Thanks to Peter Tian for collaborating on the multidimensional 0 - 1 matrix results and on the Python code for computing formation width. Thanks also to Lilly Shen for collaborating on results about extremal functions of 2-dimensional 0 - 1 matrices, as well as Rohil Prasad and Jonathan Tidor for collaborating on results about formation width. Moreover, I thank Tanya Khovanova for advice on several research projects and on my thesis defense, and for collaborating on research about visibility graphs. Thanks to PRIMES and RSI for giving me the opportunity to collaborate on projects and co-author papers as a research mentor. Thanks also to the NSF and MIT for financial support during graduate school. I also thank Mary Thornton, David Geneson and Arianna Geneson for their support and advice. In addition, I thank Katherine Bian for support, advice, and help on my thesis defense. 5 6 Contents 1 2 3 9 Introduction 1.1 1.2 Matrix extremal functions . . . . . . . . . . . . . . . . . . . . . . . . Sequence extremal functions . . . . . . . . . . . . . . . . . . . . . . . 9 10 1.3 Order of results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 The 2.1 2.2 2.3 maximum number of distinct letters Upper Bounds . . . . . . . . . . . . . . . Lower bounds . . . . . . . . . . . . . . . Bounds on extremal functions of 0 - 1 Schinzel sequences . . . . . . . . . . . . in ababa-free sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . matrices using Davenport. . . . . . . . . . . . . . . . 13 14 16 The 3.1 3.2 3.3 3.4 . . . . 23 25 27 29 31 Computing I . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 formation width of sequences An extension of the Erd6s-Szekeres theorem . . . . Algorithm for computing fw . . . . . . . . . . . . . Using binary formations to compute fw . . . . . . . Bounding the formation width of binary formations 3.4.1 20 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 34 36 39 4 Linear extremal functions of forbidden 0 - 1 matrices 4.1 Facts about exs(n, S) and exsk(m, S) . . . . . . . . . . . . . . . . . . 4.2 Bar s-visibility hypergraphs and 0 - 1 matrices . . . . . . . . . . . . 4.3 Infinitely many minimal non-linear 0 - 1 matrices . . . . . . . . . . . 41 42 45 47 5 Extremal functions of forbidden multidimensional matrices 5.1 Upper and lower bounds for d-dimensional permutation matrices 5.2 Sharp bounds on m(n, R . k, d) . . . . . . . . . . . . . . . . . . 5.3 Upper bounds for d-dimensional double permutation matrices . 5.4 Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 50 . . . 56 . . . . . . 57 61 3.5 3.6 3.7 3.4.2 Computing r . . . . . . . . . Further bounds on extremal functions Further bounds on fw . . . . . . . . . Open Problems . . . . . . . . . . . . 7 . . . . . using fw . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 Chapter 1 Introduction Problems in the area of pattern containment and avoidance focus on determining whether specific substructures are present in a given structure [1, 5, 15, 27, 29, 34, 36, 43, 441. For example, does the sequence a have a subsequence that is isomorphic to v? Or, analogously, does the 0 - 1 matrix A have a submatrix that can be turned into B by possibly changing some ones to zeroes? We focus mainly on bounding extremal functions of sequences and 0 - 1 matrices that avoid forbidden patterns. For example, what is the maximum number of ones in an n x n 0 - 1 matrix that does not contain a forbidden pattern? Or, what is the minimum length k so that every sequence of length k with n distinct letters and no adjacent same letters contains an alternation of length 5? 1.1 Matrix extremal functions An early motivation for bounding matrix extremal functions was to use them for solving problems in computational and discrete geometry [4, 18, 371. Mitchell wrote an algorithm to find a shortest rectilinear path that avoids obstacles in the plane [371 and proved that the complexity of this algorithm is bounded from above in terms of a specific matrix extremal function, which was bounded by Bienstock and Gybri [4]. Firedi [181 used matrix extremal functions to derive an upper bound on Erd6s and Moser's [101 problem of maximizing the number of unit distances in a convex n-gon. Recent interest in the extremal theory of matrices has been spurred by the resolution of the Stanley-Wilf conjecture using the linearity of the extremal functions of forbidden permutation matrices [33, 36]. The 0 - 1 matrix A contains a 0 - 1 matrix M if some submatrix of A can be transformed into M by changing some ones to zeroes. If A does not contain M, then A avoids M. Let ex(n, M) be the maximum number of ones in an n x n 0 - 1 matrix that avoids M, and let exk(M, M) be the maximum number of columns in a 0 - 1 matrix with m rows that avoids M and has at least k ones in every column. Furedi and Hajnal asked for a characterization of all 0 - 1 matrices P such that ex(n, P) = O(n) [19]. A corresponding problem for the column extremal function is to characterize all 0 - 1 matrices P with j rows such that for every fixed k > j, 9 exk(m, P) = 0(m). Another related problem is to characterize all 0 - 1 matrices P for which eXk(m, P) = OF) ). A method for bounding ex(n, M) by using bounds on the maximum number of edges in bar visibility graphs was introduced in [171. By using a similar method with bar visibility hypergraphs, we obtain linear bounds on the extremal functions of other forbidden 0 - 1 matrices. Call N a minimal nonlinear 0 - 1 matrix if ex(n, N) = w(n) and ex(n, N') = O(n) for every N' properly contained in N. Call the k x 2k matrix P a double permutation matrix if P can be obtained from a k x k permutation matrix by replacing every column with two copies of itself. We prove that ex(n, P) = 2O(k)n for every k x 2k double permutation matrix P. Using this result, we show the existence of infinitely many minimal nonlinear 0 - 1 matrices. We also study the extremal functions of multidimensional 0 - 1 matrices. A ddimensional ni x ... x nd matrix is denoted by A = (a .... ,id), where 1 < i1 < ne for e = 1, 2,. .. , d. We may view a d-dimensional 0 - 1 matrix A (ail ..,id) as a d-dimensional rectangular box of lattice points with coordinates (i1 , ... , id). An f-cross section of matrix A is a maximal set of entries ail . ,id with i fixed. A d-dimensional k x ... x k 0 - 1 matrix is a permutation matrix if each of its i-cross sections contains a single one for every e = 1,... , d. A d-dimensional 0 - 1 matrix A avoids another d-dimensional 0 - 1 matrix P if no submatrix of A can be transformed into P by changing some ones to zeroes. The maximum number of ones in a d-dimensional n x ... x n matrix that avoids P is denoted by f(n, P, d). We exhibit a family of k x ... x k permutation matrices P for which f(3,d)4 has a lower bound of 2 Q(kl/d) for n > 2Lkl/dJ/20 Furthermore we improve the upper bound on A"P") from 2 0(klogk) to 2 0(k) for all k x - x k permutation matrices P, and we show for every fixed d > 2 that the new upper bound is also true for d-dimensional double permutation matrices of dimensions 2k x k x ... x k. 1.2 Sequence extremal functions A sequence s contains a sequence tt if some subsequence of s can be changed into u by a one-to-one renaming of its letters. If s does not contain u, then s avoids u. The sequence s is called r-sparse if any r consecutive letters in s are pairwise different. A Davenport-Schinzel sequence of order s is a 2-sparse sequence that avoids alternations of length s + 2. Upper bounds on the lengths of Davenport-Schinzel sequences provide bounds on the complexity of lower envelopes of solution sets to linear homogeneous differential equations of limited order [8] and on the complexity of faces in arrangements of arcs with a limited number of crossings [1. A generalized Davenport-Schinzel sequence is an r-sparse sequence that avoids a fixed forbidden sequence with r distinct letters. Fox et al. [161 and Suk et al. [451 used bounds on the lengths of generalized Davenport-Schinzel sequences to prove that k-quasiplanar graphs on n vertices with no pair of edges intersecting in more than t 10 points have at most (n log n)2a("-)' edges, where a(n) denotes the inverse Ackermann function and c is a constant that depends only on k and t. Our main contribution for sequence extremal functions is a new computational method for proving tight upper bounds using upper bounds that are already known. We also construct families of sequences avoiding alternations in order to prove lower bounds on extremal functions of forbidden alternations. Let As,k(m) be the maximum number of distinct letters in any sequence which can be partitioned into m contiguous blocks of pairwise distinct letters, has at least k occurrences of every letter, and avoids alternations of length s. Nivasch [381 proved that A 5 ,2 d+l(m) = O(mad(m)) for all fixed d > 2. We show that A,+, 8 (m) = ("F1) for all s > 2, A 5 ,6 (M) = 0(mlog log m), and A 5 ,2d+ 2 (M) = O(mad(M)) for all fixed d > 3. An (r, s)-formation is a concatenation of s permutations of r letters. If u is a sequence with r distinct letters, then let Ex(u, n) be the maximum length of any rsparse sequence with n distinct letters which avoids u. We introduce a computational method for deriving tight upper bounds on Ex(u, n): For every sequence u define fw(u), the formation width of u, to be the minimum s for which there exists r such that there is a subsequence isomorphic to u in every (r, s)-formation. We use fw(u) to prove upper bounds on Ex(u, n) for sequences a such that a contains an alternation with the same formation width as u. = 2t - 1 We generalize the bounds on Ex((ab)t , n) by showing that fw((12. .)t) and Ex((12... l)', n) = n2(t2)!a(n)-2O(a(n)t 3 ) for every I > 2 and t > 3, such that a(n) denotes the inverse Ackermann function. Upper bounds on Ex((12 ... 1)', n) were used to bound the maximum number of edges in k-quasiplanar graphs on n vertices with no pair of edges intersecting in more than 0(1) points. If u is any sequence of the form avav'a such that a is a letter, v is a nonempty sequence excluding a with no repeated letters and v' is obtained from v by only moving the first letter of v to another place in v, then we show that fw(u) = 4 and Ex(u, n) = ®(na(n)). Furthermore we prove that fw(abc(acb)t ) = 2t + 1 and Ex(abc(acb)t , n) = n 2 1.3 I)(n) O((f)) for every t > 2. Order of results The next two chapters are about sequences. In Chapter 2, we prove bounds on As,k(m), as well as corollaries related to extremal functions of interval chains and 0 - 1 matrices. In Chapter 3, we use formation width to prove tight bounds on extremal functions of forbidden sequences. The final two chapters are about 0-1 matrices. In Chapter 4, we show that double permutation matrices and classes of 0 - 1 matrices corresponding to bar visibility hypergraphs have linear extremal functions. In Chapter 5, we generalize Fox's results on permutation matrices, as well as our results on double permutation matrices, to d dimensions. The results in these chapters also appear in the papers [20, 21, 22, 23, 24]. 11 12 Chapter 2 The maximum number of distinct letters in ababa-free sequences The sequence s is called r-sparse if any r consecutive letters in s are pairwise different. Let D.(n) be the maximum length of any 2-sparse sequence with n distinct letters which avoids alternations of length s. Nivasch [38] and Klazar [32] proved that lim D5 (n) 2, such that a(n) denotes the inverse Ackermann function. Agarwal, Sharir, Shor [21 and Nivasch [38] proved the bounds D8 (n) = n2t!(n)O(o(n)t-1) for Pettie [42] derived sharp bounds on D (n) for all odd s. even s > 6 with t == ". To define the Ackermann hierarchy let A1(n) = 2n and for k > 2, Ak(0) = 1 and Ak(n) = Akl(Ak(n - 1)) for n > 1. To define the inverse functions let ak(x) = min {n Ak(n) > 4} for all k > 1. We define the Ackermann function A(n) to be A,,(3) as in [38]. The inverse Ackermann function a(n) is defined to be min {x : A(x) > n}. Collections of contiguous distinct letters in a sequence are called blocks. Nivasch's bounds on D8 (n) were derived using an extremal function which maximizes number of distinct letters instead of length. Let A.,k(m) be the maximum number of distinct letters in any sequence on m blocks avoiding alternations of length s such that every letter occurs at least k times. Clearly A,(m) = 0 if m < k and A 8 ,k(M) = o0 if k < s - 1 and k < m-. Nivasch proved that A5,2d+1(mn) = O(Mad(m)) for each fixed d > 2 (but noted that the bounds on A 5,k(m) were not tight for even k). Sundar [46] derived similar bounds in terms of m on functions related to the Deque conjecture. In [381, similar bounds were also derived for a different sequence extremal function. Let an (r, s)-formation be a concatenation of s permutations of r distinct letters. For example abcddcbaadbcis a (4, 3)-formation. Define Fr,s(n) to be the maximum length of any r-sparse sequence with n distinct letters which avoids all (r, s)-formations. Klazar [30] proved that F,2 (n) = O(n) and Fr,3 (n) = O(n) for every r > 0. Nivasch proved that F,,4 (n) = 0(na(n)) for r > 2. Agarwal, Sharir, Shor [21 and Nivasch [381 showed that F,s(n) n2 * )'1) for all r > 2 and odd s > 5 with t = Y-2. Let F,,,,k(m) be the maximum number of distinct letters in any sequence on m blocks avoiding every (r, s)-formation such that every letter occurs at least k times. - 13 Clearly F,,,,k(M) = 0 if m < k and F,,,,k(M) = oo if k < s and k K m. Every (r, s)formation contains an alternation of length s + 1 for every r > 2, so AS+1,k(m) Fr,,k(m) for every r > 2. Nivasch proved for r > 2 that F,,4,2d+1(M) = O(mad(m)) for each fixed d > 2. The recursive inequalities for the upper bounds on F,,4,2d+1 (in) in [38] also imply that F,,4, 6(M) = O(m log log 7n) and F,4,2d+2(mn) = O(Mad(m)). Similar bounds were also derived on an extremal function related to interval chains. A k-chain on [1, m] is a sequence of k consecutive, disjoint, nonempty intervals of the form [ao, a,][ai + 1, a2 ... [ak_1 + 1, ak] for integers 1 < ao K a1 < ... < ak < rM. An s-tuple is a set of s distinct integers. An s-tuple stabs an interval chain if each element of the s-tuple is in a different interval of the chain. Let (,,k(m) denote the minimum size of a collection of s-tuples such that every k-chain on [1, m] is stabbed by an s-tuple in the collection. Clearly ((,k(m) = 0 if m < k and (sk(m) is undefined if k < s and k K m. Alon et al. [3] showed that 3,s(m) = (" ) for s > 1, ( 3 ,4 (m) O(mlogm), (m log logm), and (3,k(M) = (maLkj (M)) for k > 6. (3,5(m) Let rr,s,k(mi) denote the maximum size of a collection X of not necessarily distinct k-chains on [1, m] so that there do not exist r elements of X all stabbed by the same s-tuple. Clearly r,,,k(m) = 0 if m < k and 7r,,,k(m) = oo if k < s and k n. In Section 2.1 we show that r,,,,,k(M) =F,sl+,k+l(m-i+ 1) for all r > 1 and 1 K s K k K m. Since (.,k(i) rq2,,k(m) for all 1 K s K k K m, then AS+ 2 ,k+l(m+1) K (M) for all 1 K s K k K m. This implies the bounds A,+1 , 8(in) < (" 7il) for all s > 2, A 5 ,6 (m) = O(m log log m), and A5, 2 d+ 2 (M)= O(mad(in)) for d > 3. In Section 2.2 we construct alternation-avoiding sequences to prove lower bounds on As,k(m). We prove that A,+1,8 (m) = (" ) for all s > 2. Furthermore we show that A 5 ,6 (M) = Q(mloglogim) and A 5 ,2d+ 2 (m) = Q(jMad(M)) for d > 3. Thus the bounds on A 5 ,a(m) have a multiplicative gap of O(d) for all d. 2.1 Upper Bounds We show that F,,s+1,k+l(m + 1) = rlr,s,(M) for all r 1 and 1 K s K k K M using maps like those between matrices and sequences in [7] and [40]. Lemma 1. F,,sl+,k+l(M + 1) < Tlr,,k(m) for all r > 1 and 1 K s K k K m. Proof. Let P be a sequence with Fr,s +l,k+1(m + 1) distinct letters and m + 1 blocks 1,.. . , m + 1 such that no subsequence is a concatenation of s + 1 permutations of r different letters and every letter in P occurs at least k + 1 times. Construct a collection of k-chains on [1, m] by converting each letter in P to a k-chain: if the first k + 1 occurrences of letter a are in blocks ao, ... , ak, then let a* be the k-chain with ith interval [ai_ 1 , a, - 1]. Suppose for contradiction that there exist r distinct letters q1,. . ., q, in P such that q*, ... , q* are stabbed by the same s-tuple 1 < Ji < ... < j 5 K <m. Let jo = 0 and js+1 m + 1. Then for each 1 K i K s + 1, q,, occurs in some block b,,i such 14 that b (r, s + c,i[j1i- + 1, ji] for every 1 < n K r. Hence the letters qi,.. . , q, make an l 1)-formation in P, a contradiction. Corollary 2. AS+ 2 ,k+1(m + 1) K Q(m) for all 1 K s < k K m. The bounds on (,,k(m) in [3] imply the next corollary. Corollary 3. A,+i,8 (m) < (1 1 ) for s > 2, A 5 ,5 (m) for k > 7. O(m log log m), and A 5 ,k(m) = O(mak21J(m)) 'jj O(mlogrm), A 5 ,6 (m) = To prove that F,,,+,k+1(M + 1) ms,k(m) for all r > 1 and 1 K s K k K m, we convert collections of k-chains into sequences with a letter corresponding to each k-chain. Lemma 4. Fr,s+1,k+1(m + 1) r,s,k(m) for all r > 1 and 1 K s < k K m. Proof. Let X be a maximal collection of k-chains on [1, m] so that there do not exist r elements of X all stabbed by the same s-tuple. To change X into a sequence P create a letter a for every k-chain a* in X, and put a in every block i such that either a* has an interval with least element i or a* has an interval with greatest element i - 1. Order the letters in blocks starting with the first block and moving to the last. Let Ai be the letters in block i which also occur in some block j < i and let Bi be the letters which have first occurrence in block i. All of the letters in Ai occur before all of the letters in Bi. If a and b are in Aj, then a appears before b in block i if the last occurrence of a before block i is after the last occurrence of b before block i. The letters in Bi may appear in any order. P is a sequence on m + 1 blocks in which each letter occurs k + 1 times. Suppose for contradiction that there exist r letters q, ..., q,. which form an (r, s + 1)-formation in P. List all (r, s + 1)-formations on the letters qi, .. . , q, in P lexicographically, so that formation f appears before formation g if there exists some i > 1 such that the first i - 1 elements of f and g are the same, but the ith element of f appears before the ith element of g in P. Let fo be the first (r, s + 1)-formation on the list and let 7ir (respectively pi) be the number of the block which contains the last (respectively first) element of the i < s + 1. Suppose for contradiction that for some ith permutation in fo for 1 1 < i K s, 7i = Pi+1. Let a be the last letter of the ith permutation and let b be the first letter of the (i + 1)st permutation. Then a occurs before b in block ri and the b in 7i is not the first occurrence of b in P, so the a in wi is not the first occurrence of a in P. Otherwise a would appear after b in 7ri. Since the a and b in wi are not the first occurrences of a and b in P, then the last occurrence of a before 7i must be after the last occurrence of b before irr. Let fi be the subsequence obtained by deleting the a in ir from fo and inserting the last occurrence of a before 7i. Then fi is an (r, s + 1)-formation and fi occurs before fo on the list. This contradicts the definition of fo, so for every 1 K i K s, 7ri < Pi+i. j < r and 1 K i K s + 1, the letter qj appears in some block between For every 1 w) stabs 71,..., pi and Tr inclusive. Since 7i < Pj+j for every 1 < i K s, the s-tuple ( 15 each of the interval chains q*,.. , q,*, a contradiction. Hence P contains no (r, s + 1)formation. The idea for the next lemma is similar to a proof about doubled formation free matrices in [7]. Lemma 5. F,,s,s (m) < (r - 1)("- IV) for every r > 1 and I < s < m. Proof. Let P be a sequence with m blocks such that no subsequence of P is a concatenation of s permutations of r distinct letters and every letter of P occurs at least s times. An occurrence of letter a in P is called even if there are an odd number of occurrences of a to the left of it. Otherwise the occurrence of a is called odd. Suppose for contradiction that P has at least 1 + (r - 1) (7ni 1) distinct letters. + The number of distinct tuples (i, .... iLJ) for which a letter could have even occurrences in blocks i,.. . , 2 j is equal to the number of positive integer solutions to the equation (1 I x 1 ) + ... + (1 + xLj) + xI+LJ = + 1 if s is even and (1 x1 ) + ... + (1 + X +H)x1+[IJ = M if s is odd. Then by the pigeonhole principle there are at least r distinct letters q 1 , ... , q, with even occurrences in the same [LJ blocks. Then P contains a concatenation of s permutations of the letters q 1 ,... q, a contradiction. l The last lemma is an alternate proof that A,+1, 8(m) < ("r[l) since A8 +,(rm) < Fr,,,s(mfl) for all s, i > 1 and r > 2. 2.2 Lower bounds In the last section we showed that A,+1,8 (n) ( for s > 2. The next lemma provides a matching lower bound. (" 71) for all s > 2 and m > s + 1. Proof. For every s > 1 and m > s + 1 we build a sequence X,(m) with (j"- ) Lemma 6. As+ 1 ,(n) > distinct letters. First consider the case of even s > 2. The sequence X,(rn) is the concatenation of i - 1 fans, so that each fan is a palindrome consisting of two blocks of equal length. First assign letters to each fan without ordering them. Create a letter for every -tuple of non-adjacent fans, and put each letter in every fan in its 1-tuple. Then order the letters in each fan starting with the first fan and moving to the last. Let Ai be the letters in fan i which occur in some fan j < i and let B be the letters which have first occurrence in fan i. In the first block of fan i all of the letters in Ai occur before all of the letters in Bi. If a and b are in Ai, then a occurs before b in the first block of fan i if the last occurrence of a before fan i is after the last occurrence of b before fan i. If a and b are in Bi, then a occurs before b in the first block of fan i if the first fan which contains a without b is before the first fan which contains b without a. 16 Consider for any distinct letters x and y the maximum alternation contained in the subsequence of X,(m) restricted to x and y. Start building the alternation with a fan that contains only x. Any other fans which contain x without y or y without x add at most 1 to the alternation length. Any fans which contain both x and y add 2 to the alternation length. If x and y occur together in i fans, then the length of their alternation is at most (s - i) + (I - i) + 2i = s. Every pair of adjacent fans have no letters in common, so every pair of adjacent blocks in different fans can be joined as one block when the rn-1 fans are concatenated to form Xs(rn). Thus X,(m) has m blocks and ("2_) letters, and each letter occurs s times. For odd s > 3, construct X,(m) by adding a block r after Xs-1(m - 1) containing all of the letters in X,-1(m - 1) such that a occurs before b in r if the last occurrence of a in X,_(m - 1) is after the last occurrence of b in X,_1(m - 1). Then X,(m) contains no alternation of length s + 1 since X,_1(m - 1) contains no alternation of length s. Moreover X,(m) has m blocks and (" q--+ r 1 ) letters, and each letter occurs 2 s times. l The next lemma shows how to extend the lower bounds on As+,,,(m) to F,,,,,(m). Lemma 7. F,s,k(m) (r - 1)F2 ,,,k(M) for all r > 1 and 1 < s < k < m. - Proof. Let P be a sequence with F 2 ,,,k(m) distinct letters and m blocks such that no subsequence is a concatenation of s permutations of two distinct letters and every letter occurs at least k times. P' is the sequence obtained from P by creating r - 1 new letters a 1 , .. , a,-, for each letter a and replacing every occurrence of a with the sequence a, ... a. Suppose for contradiction that P' contains an (r, s)-formation on letters q 1 ,... , q,. Then there exist indices i, j, k, 1 and distinct letters a, b such that qi = a3 and qk= bl. P' contains a (2, s)-formation on the letters qi and qk, so P contains a (2, s) formation on the letters a and b, a contradiction. Then P' is a sequence with (r 1)F2,s,k(m) distinct letters and m blocks such that no subsequence is a concatenation of s permutations of r distinct letters and every letter occurs at least k times. El Corollary 8. Fr,,,, (m) = (r - 1)(" ) for all r > 1 and m > s > 1. The proof of the next lemma is much like the proof that A 5 ,2 d+1(m) - Q(Mad(n)) for d > 2 in [381. Lemma 9. A 5 ,6 (m) = Q(m log log m) and A5,2d+ 2 (m) = Q(!mad(m)) for d > 3. For all d, m > 1, we inductively construct sequences Gd(M) in which each letter appears 2d + 2 times and no two distinct letters make an alternation of length 5. This proof uses a different definition of fan: fans will be the concatenation of two palindromes with no letters in common. Each palindrome consists of two blocks of equal length. The sequences G1(m) are the concatenation of m + 1 fans. In each fan the second block of the first palindrome and the first block of the second palindrome make one 17 block together since they are adjacent and have no letters in common. The first palindrome in the first fan and the second palindrome in the last fan are empty. There is a letter for every pair of fans and the letter is in both of those fans. The letters with last appearance in fan i are in the first palindrome of fan i. They appear in fan i's first palindrome's first block in reverse order of the fans in which they first appear. The letters with first appearance in fan i are in the second palindrome of fan i. They appear in fan i's second palindrome's first block in order of the fans in which they last appear. By construction G1 (m) contains no alternation of length 5. For all d > 1 the sequence Gd(1) consists of 2d + 2 copies of the letter 1. The first and last copies of 1 are both special blocks, and there are empty regular blocks before the first 1 and after the last 1. For d, m > 1 the blocks in Gd(m) containing only first and last occurrences of letters are called special blocks. Let Sd(m) be the number of special blocks in Gd(m). Every letter has its first and last occurrence in a special block, and each special block in Gd(m) has m letters. Blocks that are not special are called regular. No regular block in Gd(m) has special blocks on both sides, but every special block has regular blocks on both sides. The sequence Gd(m) for d, m > 2 is constructed inductively from Gd(m - 1) and Gd_1(Sd(m - 1)). Let f = Sd(m - 1) and g = Sd_1(f). Make g copies X 1 ,..., X of Gd(m - 1) and one copy Y of Gdl(f), so that no copies of Gd(m - 1) have any letters in common with Y or each other. Let Ai be the ith special block of Y. If the 1" element of Ai is the first occurrence of the letter a, then insert aa right after the Ith special block of Xi. If the 1 t" element of Ai is the last occurrence of a, then insert aa right before the lPh special block of Xi. Replace A in Y by the modified Xi for every i. The resulting sequence is Gd(m). Lemma 10. For all d and m, Gd(m) avoids ababa. Proof. Given that the alternations in Gi(m) have length at most 4 for all m > 1, then F-1 the rest of the proof is the same as the proof in [381 that Zd(m) avoids ababa. Let Ld(m) be the length of Gd(m). Observe that Ld(M) = (d + 1)mSd(m) since each letter in Gd(m) occurs 2d + 2 times, twice in special blocks, and each special block has m letters. Define Nd(m) as the number of distinct letters in Gd(m) and MAd(m) as the number of blocks in Gd(m). Also let Xd(m) = Ad() ". We bound Xd(m) Id(mf) Sd (M) and Vlj(m) = Ld and Vd,(m) as in [381. Lemma 11. For all m, d > 1, Xd(mn) < 2d + 2 and V1(m) > M. Proof. By construction Si(m) = m + 1 for m > 1, Sd(1) = 2 for d> 2, and Sd(7n) Sd(m -1)Sdl_(Sd(m-1)) for d, m > 2. Furthermore M1 (m) = 3m-+3, MAd(1) = 2d+4 ford> 2, and Aid(m) = A(m-)Sdl(Sd(m--1))+MAd(S(m-1))-Sdl_(Sd(m- 1)) for d, m > 2. Thus S 2 (m) = S 2 (m - 1)(S 2 (m - 1) + 1) 2(S 2 (m - 1))2 and S 2 (1) = 2. Since S'(m) = 22-1 satisfies the recurrence S'(1) = 2 and S'(m + 1) = 2(S'(m)) 2 , then 18 22m1 < 3 x S 2 (m) < 2 For d > 2, Sd(2) = 2Sd_1(2) and S 1 (2) = 3. So Sd(2) = 2 d-1 For d > 2, MAd(2) = (2d + 3)(3 x (6d + 3 ) 2d-1 2 d-2) d+2 for all d > 2, and Xd(2) = 2d+ 1 for all Then X,1(m) = 3 for all m, Xd(1) d > 2. For d, m > 2, Xd(m) = Xd(m + Mdl_(2) and M 1(2) = 9. Hence MAd(2) = - 1) + Xd-1(Sd(m-1))--1 Sd(M-1) We prove by induction on d that Xd(m) 2d + 2 for all m, d > 1. Observe that . the inequality holds for Xi(m), Xd(1), and Xd(2) for all m, d. Xdj(m - 1) + s2Fix d and suppose Xd1(m) < 2d for all m. Then Xd(m) 1 -~=2 Sd(n)- 1 -2 Sd(n)) = 2d + 1 + (2d - 1) Hence Xd(m) < Xd(2) + (2d - 1) 2SK(2)-l = ' 2 Sd(n))< Since Sd(m) > 2Sd(m - 1) for all d, m > 2, then 1 < 1 foral so Ld(m) > (r) m1 2d+2. Hence Vd(m) -Xd(m) all d> , x2d-2 - 2d1 _ -(d+1)MSd(rn) - The following analysis demonstrates the lower bounds on (m) for each A5,2d+2 2, let mi = M2 (i) and ni = N2 (i). Then mi = X 2 (i)S2 (i) < 6S2 (i) < = V(i)N 2 (i) > i 2 () L 2 (i) 6(221-1) < 22"+2 for i > 1. Then i = Q(log log mi), so ni 6 12 Q(mi log log mi). d > 2. If d = -6 We use interpolation to extend the bound from mi to m. Let i and t satisfy mi < m < mi+ 1 and t = L[J. Concatenate t copies of G 2 (i) with no letters in common for a total of at least [- Jni = Q(m log log m) letters. Hence A 5 ,6 (m) = Q(m log log mn). We prove that S3 (m) < A 3 (2m) following the method of [381. Since S 3 (n) S3 (m - 1)S 2 (S 3 (m - 1)) < S2 (S 3 (n - 1))2 < 22s3(m-'+-2, then let F(in) = 22m+1_2 and G(m) = 22m. Then 2F(n) = 22m+_1 < 222, = G(2m) for every m > 0. Thus + S 3 (m) < F(m-l)(S 3 (1)) < 2F(M-n)(S 3 (1)) 5 G(m- 1)(2S 3 (1)) = A 3 (2mi). Let mi = A13 (i) and ni = N3 (i). Therefore mi = X 3 (i)S 3 (i) K 8S 3 (i) < A 3 (2i =_ 3()M i >3iiMV (i . 3 = ,> 2) for i > 1. So i = Q(a3 (Mi)) and ni = N 3 (i) = Q(mina3 (mi)). Then A 5 ,s(M) Q(ma 3 (m)) by interpolation. Ad(m + 2) for m > 1 by induction on For each d > 4 we prove that Sd(m) d. Since S 4 (in) = S4 (m - 1)S3 (S 4 (m - 1)) < S 3 (S4 (n - 1))2, then let F(m) = S 3 (im) 2 . Since 4F(m) < A 3 (4m) and A 4 (3) > 4S4 (1), then S 4 (m) < F(rM- (S4 (1)) < 4F(m- 1 ) (S4 (4)) < 4(fl-l)(4S4(1)) < A 4 (m + 2). Fix d > 4 and suppose that Sd-_(m) < A-1 (m + 2). Define F(m) = Sd-(m) 2 Since 4F(m) < Ad_1(4m) and Ad(3) > 4Sd(1), then Sd(m) < F(m-1 ) (Sd(1)) < + 4F(m-1 ) (Sd(1)) < A_" 1"(4Sd(1)) < Ad(m + 2). Fix d > 4. Let mi = Md(i) and ni = Nd(i). Then mi = Xd(i)Sd(i) < (2d < (2d + 2)Ad(i + 2) < Ad(i + 3). Then i > ad(Mi) - 3, so ni =di iMA(i) = Q(lMiad(Mi)). By interpolation A,2d+2(m) = Q(1 Mad(M)) for - 2)Sd(i) Vd(i)Md(i) > d > 4. Corollary 12. For r > 2, F,4 , 6(m) = r.,3,(m) 7Ir,3,2d+1(n) = Q(jm!ed(M)) for d > 3. 19 = Q(miloglognm) and Fr,4,2d+2(m)= 2.3 Bounds on extremal functions of 0 - 1 matrices using Davenport-Schinzel sequences Let IF, (m, n) be the maximum length of a sequence with ?n blocks and n letters which avoids alternations of length s + 2. Nivasch showed that TV(m, n) < k(HFI'm) + n) for all k. We modify the matrix-sequence transformations in [401 to show bounds on exk(m, P) for alternating patterns P. Let P, be the 0 - 1 matrix with s rows 0, ... , s - 1 and 2 columns 0, 1 such that the number in each entry is the sum of its row and column mod 2. Lemma 13. T, (m, n) = ex(n,n, P+1) for for s, k, m > 1. 'm, n, s > 1 and H(m) = exk(MPs+1) Proof. The sequence to matrix transformation in [40] starts with a sequence Q with n letters on m blocks which avoids alternations of length s + 2, and results with a 0 - 1 matrix A with n columns and m rows which avoids the matrix P,+. The letters of Q are named 0,...,n -I by first occurrence and the entry in column i and row j of A is a one if and only if the letter i occurs in block j of Q. Suppose for contradiction that A contains P,+1. Then there is a submatrix of A with 2 columns co < ci and s + 1 rows ro, ... , r8 such that the entry in column ci and row rj is one if i + j is odd. The one entries in this submatrix correspond to an alternation c 1co... of length s + 1 in Q. However, the first occurrence of co is before the first occurrence of c1 in Q, so Q contains an alternation of length s + 2, a contradiction. This implies both T, (m, n) < ex(m, n, P,+1 ) for m, n, s > 1 and Hs(M) exk(M, P+l) for s, k, m > 1. ) Pettie used a matrix to sequence transformation to show that ex(m, n, P,+ 1 I(m, n) + n. 0 - 1 matrix A with m rows 0,..., m - I and n columns 0,. .,rn - 1 is converted to a sequence Q with m blocks 0,..., m - 1 and n letters 0,.. .,i - 1. Letter i occurs in block j of Q if and only if the entry of A in column i and row j is a one. Let C be the letters in block j of Q which occur in no block before j and let Dj be the letters in block j of Q which occur in a block before j. All letters in Di occur before all letters in Cj in block j, and letters in Dj occur in block j in reverse order of their last appearance before block j. In [401 the letters in C were ordered arbitrarily, but here letter x in Ci occurs before letter y in Cj if and only if x < y. Suppose that Q contains an alternation of length s + 2 on letters x and y such that x < y. List all alternations of length s+2 on the letters x and y in Q lexicographically, so that alternation f appears before alternation g if there exists some i > 1 such that the first i - 1 elements of f and g are the same, but the i''h element of f appears before the ith element of g in Q. Let fo be the first alternation on the list and let 7i be the number of the block which contains the ith element of fo for 1 < i < s + 2. Suppose for the sake of contradiction that for some 2 < i K s + 1, i = wi+ 1 . Let a be the ith element of fo and let b be the (i + 1)'t element of fo. Then a occurs before b in block ri and the b in 7ir is not the first occurrence of b in Q, so the a in 7i is not the first occurrence of a in Q. Otherwise a would appear after 20 . b in 1ri. Since the a and b in 7i are not the first occurrences of a and b in Q, then the last occurrence of a before 7i must be after the last occurrence of b before 7ri. Let fi be the subsequence obtained by deleting the letter a in 7i from fo and inserting the last occurrence of a before 7ri. Then fi is an alternation of length s +2 and fi occurs before fo on the list. This contradicts the definition of fo, so for every 2 < i < s + 1, 7i < 7rj+ 1 . We now consider two cases. If the first element of fo is x, then the submatrix of A with 2 columns x, y and s + 1 rows 7r2, -- -,7r,+2 contains P+1 since x < y. If the first element of fo is y, then suppose for contradiction that 7r, = r2 . The occurrences of x and y in block 7r, are both first occurrences of x and y in Q. Otherwise, fo would not be the first alternation on the list. Since x and y are in Cr, then x appears before y in block 7r1 because x < y, a contradiction. Then the submatrix of A with 2 columns ,r+ 7r,... contains Ps+1 x, y and s + 1 rows Therefore ''(m, n) > ex(m, n, P+1) for m, n, s > 1 and I >-(m) exk(M, P+l) for s, k, m > 1. for all s > 2, ex6(m, P4 ) = (m loglog m), and (Mrad(m)) for all fixed d > 3. Corollary 14. ex,(m,Ps) =("j.) ex2d+2(m, P4 ) = 6 21 22 Chapter 3 The formation width of sequences If u is a sequence with r distinct letters, then let Ex(u, n) be the maximum length of any r-sparse sequence with n. distinct letters that avoids u. If a and b are single letters, then Ex(a, n) = 0, Ex(ab, n) = 1, Ex(aba, n) = n and Ex(abab, n) = 2n - 1. Nivasch [381 and Klazar [32] determined that Ex(ababa, n) ~ 2na(n). Agarwal, Sharir, and Shor [21 proved the lower bound and Nivasch [381 proved the upper bound to show ') for 2O(a(l)t that if u is an alternation of length 2t + 4, then Ex(u, n) = n9nCdn) t > 1. If u is a sequence with r distinct letters and c > r, then let Exc(u, n) be the maximum length of any c-sparse sequence with n distinct letters which avoids u. Klazar [30] showed that Exc(u, n) = E(Exd(u, n)) for all fixed c, d > r. Lemma 15. [30] For every sequence u with r distinct letters, Exd(u, n) Ex,(u, n) (1 + Exc(u, d - 1))Exd(u, n) for all n > 1 and d > c > r. An (r, s)-formation is a concatenation of s permutations of r distinct letters. For example abcddcbaadbc is a (4, 3)-formation. Definition 16. F,s (n) is the maximum length of any r-sparse sequence with n distinct letters that avoids every (r, s)-formation. Klazar [30] proved that Fr2 (n) = O(n) and F,3 (n) = O(n) for every r. Nivasch [38] proved that F,4 (n) = O(na(n)) for r > 2. Agarwal, Sharir, and Shor [2] proved the lower bound and Nivasch [38] proved the upper bound to show that F,s()= n2Aa()+!(a(n)-) for all r > 2 and odd s > 5 with t = Nivasch [381 proved that Ex(u, n) < Fr,s-.r+i(n) for any sequence u with r distinct letters and length s by showing that every (r, s - r + 1) formation contains u. Definition 17. The formation width of u, denoted by fw(u), is the minimum value of s such that there exists an r for which every (r, s)-formation contains u. The formation length of u, denoted by fl(u), is the minimum value of r such that every (r, fw(u))-formation contains u. By Nivasch's proof, fw(u) < s - r + 1 for every sequence u with r distinct letters and length s. The next two facts follow from the definition of fw. 23 Lemma 18. If u contains v, then fw(v) fw(u). Lemma 19. If u begins with the letter a, then fw(au) = fw(u) + 1. Lemma 15 implies that fw(u) and fl(u) can be used to obtain upper bounds on Ex(u, n). - Lemma 20. For any sequence u with r distinct letters and fixed integer c with c > r, Exc(u, it) = 0 (Ffl() (n)) In this paper we use fw(n) primarily in order to prove tight upper bounds on Ex (u, n) for several classes of sequences u such that a contains an alternation with the same formation width as u. We also bound and evaluate fw for various other families of sequences in order to develop a classification of all sequences in terms of their formation widths. If at is an alternation of length t for t > 2, then fw(at) < t - 1 since every (r, t - 1) formation contains at for r > 2. Any (r, t - 2)-formation in which order of letters reverses in adjacent permutations avoids at, so fw(at) = t - 1. Pettie [411 used the fact that every (4,4)-formation contains abcacbc to prove the upper bound Ex(abcacbc, n) = O(na(n)). Since any (r, 3) formation with order reversing in adjacent permutations would avoid abcacbc, then fw(abcacbc) = 4. Similarly fw(abcadcbd) = 4. Definition 21. An (r, s)-formation f is called binary if there exists a permutation p on r letters such that every permutation in f is either the same as p or the reverse of p. Most of the proofs in this paper depend on the fact that if u is a sequence with r distinct letters, then every binary (r, s)-formation contains u if and only if s > fw(u). We use the following notation to describe binary formations more concisely. Definition 22. I is the increasing sequence 1 ... c on c letters and D, is the decreasing sequence c . . 1 on c letters. Given a permutation 7r C Sc, the sequences I, and D, are 7r(1) ... 7r(c) and 7(c) ... w(1) respectively. We focus especially on two classes of binary formations in order to derive bounds on fw(u). The sequence up(l, t) is I, repeated t times, and alt(l, t) is a concatenation of t permutations, starting with I and alternating between I1 and D1 . For example, up(3, 3) = 123123123 and alt(3,3) = 123321123. Definition 23. If u is a sequence with c distinct letters, then l(u) is the smallest k such that up(c, k) contains a, and r(a) is the smallest k such that alt(c, k) contains a. Then fw(u) > l(a) and fw(u) > r(a). We evaluate both 1(u) and r(u) for every binary formation U. In Section 3.1 we prove that -y(r, s) = (r-1)2 -1+1 is the minimum value for which every (-y(r, s), s)-formation contains a binary (r, s)-formation. It follows that if u has 24 r distinct letters, then fw(u) is the minimum s for which every binary (r, s)-formation contains u. In Section 3.3 we prove that fw(u) = t - 1 for every sequence u with two distinct letters and length t. We also determine every sequence u for which fw(u) < 3. In addition, we show that fw(up(c,t)) = 2t - 1 for all c > 2 and t > 1. This t3 implies that Ex(up(l, t), n) = n2 -- (n) t 2s(Q(n) ) for all 1 > 2 and t > 3 and that fw(u) < 21(u) - 1 for every sequence u. In Section 3.4.1 we compute l(u) and use the result to bound fw(u) up to a factor of 2 for every binary formation u. In particular we prove the following bounds on fw(u). Theorem 24. Fix c > 2 and let u = I," D 2 I 12 ... Cn, where L is I if n is odd and D if n is even, and ej > 0 for all i. Define A = >1 e 2i- 1 and B = E M = max(A,B) and let m = min(A,B). Then (c- 1)m + M + [ n e 2i. Let fw(u) 2(c - 1)m + 2M + 2[nJ - 1. In Section 3.4.2 we compute r(u) for every binary formation u. Specifically we prove that if c > 2, then r(Ig1Dc2Ig3 .... Cen) = 2Z" e -nwhere L is I if n is odd and D if n is even. In Section 3.5 we use fw(u) to derive tight bounds on Ex(u, n) for other sequences u besides up(l, t). Let u be any sequence of the form avav'a such that a is a letter, v is a nonempty sequence excluding a with no repeated letters and v' is obtained from v by only moving the first letter of v to another place in v. We show that fw(u) = 4, implying that Ex(u, n) = E(na(n)). We also prove that Ex(abc(acb)t , n) = for all t > 2. n 2 (Tt (n)(a(f)2 In Section 3.6 we compute fw for various classes of binary formations. In particular we show for c > 2 and k > 1 that fw(IcDcIc) = c + 3, fw(IcDc) = c + 2k - 1, fw(IcDcIcDc) = 2c+3, fw(alt(c, 2k)) > k(c+2)-1, and fw(alt(c, 2k+1)) > k(c+2)+1. In Section 3.7 we discuss some unresolved questions. 3.1 An extension of the Erd6s-Szekeres theorem The following upper bound is obtained by iterating the Erd6s-Szekeres theorem as in [30]. Lemma 25. Every ((r - 1)2'- + 1, s)-formation contains a binary (r, s)-formation. Proof. We prove by induction on s that every ((r - 1)2'-' + 1, s)-formation contains a binary (r, s)-formation. Clearly this is true for s = 1. For the inductive hypothesis fix s and suppose for every r > 1 that each ((r - 1)281 + 1, s)-formation contains a binary (r, s)-formation. Consider any ((r - 1)2" +1, s+1)-formation F. Without loss of generality suppose that the first permutation of F is I(r_1)2-. By inductive hypothesis the first s permutations of F contain a binary ((r-1)2 +1, s)-formation f. By the Erd6s-Szekeres theorem, every sequence of (x - 1)2 + 1 distinct integers contains an increasing or 25 decreasing subsequence of length x. Therefore the last permutation of F contains an increasing or decreasing subsequence of length r on the letters of f. Thus F contains l a binary (r, s + 1)-formation. Corollary 26. If u has r distinct letters, then every binary (r, s)-formation contains u if and only if s fw (u). Proof. If for some s every binary (r, s)-formation contains t, then there exists a function y(r, s) such that every (?(r, s), s)-formation contains i. Thus fw(u) < s. If some binary (r, s - 1)-formation f avoids i, then for every z > r the binary (z, s - 1)-formations which contain f will avoid u. Hence fw(u) > s - 1. l Corollary 27. If i is a nonempty sequence and v is obtained from u by inserting a single occurrence of a letter which has no occurrence in i, then fw(u) =fw(v). Proof. If 't has r distinct letters, then every binary (2r + 1,fw(u))-formation F with first permutation I2-+1 has a copy of u using only the even numbers 2,...,2r. Since there is at least one odd number between every pair of even numbers in F, then the copy of t in F can be extended to a copy of v using an odd number. El 1. Proof. Since every binary (r, fw(u))-formation contains u, then every ((r -1)2mu)-1 1,fw(u))-formation contains u. + Corollary 28. If a has r distinct letters, then fl(u) < (r - 1)2ft (u+ l The next theorem shows that the upper bound in Lemma 25 is tight. Theorem 29. For every r, s > 1 there exists a ((r every binary (r, s)-formation. - 1)2'1, s)-formation that avoids Proof. We construct the desired formation Fa(r, s) one permutation at a time. Define an a-block in Fa(r, s) to be a block of numbers in a permutation from positions (k-1)(r-1)a+1 to k(r-1) for some k. For k < s-I define a k-swap on a permutation of length (r - 1)2-1 as follows: For every even i, 1 < i < 2k, a k-swap reverses the placement of the (i - 1)2'-k-'-blocks in each i2s-k-'-block. For example if (r, s) (3, 3), then a 1-swap on 1234567890ABCDEF produces CDEF90AB56781234. - Let permutation 1 of Fa(r, s) be the identity permutation on the letters 1, . . . , (r 1)2-'. To form permutation k + 1 of Fa(r, s), perform a k-swap on permutation k. The next lemma about Fa(r, s) will imply that Fa(r, s) avoids every binary (r, s)formation. Lemma 30. Consider any set B of distinct numbers occurring in each of the first k permutations of Fa(r, s) with the same or reverse order in adjacent permutations. Let i(k) = ej2k-j~1 where ej = 1 if the elements in B reverse order from permutation j to permutation j + 1 and ej = 0 otherwise. Then in permutation k the elements of B are contained in different i(k)2sk-blocks, but the same (i(k) + 1) 2 s-kblock. 26 Proof. We induct on k. When k = 1, i(k) = 0. The entire permutation is a 2~1-block and 0-blocks are individual elements, so the lemma is true when k = 1. For the inductive hypothesis, suppose that in permutation k the elements of B are contained in a single (i(k) + 1)2,-k-block but different i(k)2s-k-blocks. Consider any set B of distinct numbers occurring in each of the first k + 1 permutations of Fa(r, s) with the same or reverse order in adjacent permutations. Now consider the k-swap that sends permutation k of Fa(r, s) to permutation k + 1. The parts of the swap that reverse the placement of the (j - 1)2s~k-'-blocks in each j2s-k---block for even j > 2i(k) + 4 do not affect the order of the elements of B since the elements of B are contained in a single (2i(k) + 2)2s-k-l-block. The parts of the swap that reverse the placement of the (j - 1)2s-k-'-blocks in each j2'-k-'-block for even j < 2i(k) also do not affect the order of the elements of B since the elements of B are contained in different (2i(k))2s-k-'-blocks. Thus the only part of the swap which is relevant to the order of the elements in B is the reversal of the placement of the (2i(k) + 1)2s-k-l-blocks inside each (2i(k) + 2)2s-k-l-block. If the order of elements in B reverses from permutation k to permutation k + 1, then i(k + 1) = 2i(k) + 1. All the elements of B must be contained in different (2i(k)+1)2s-k-l-blocks, or else the k-swap would not reverse their order. By inductive hypothesis the elements of B are contained in the same (i(k + 1) + 1)2s-k-l-block. If the order of elements in B is the same in permutation k and permutation k + 1, then i(k + 1) = 2i(k). The elements of B must be contained in the same (2i(k)+1)2s-k-l-block, or else the k-swap would not preserve their order. By inductive l hypothesis the elements of B are contained in different i(k + 1)2s-k-l-blocks. Given any set B of distinct numbers contained in every permutation of Fa(r,s) whose order either stays the same or reverses between adjacent permutations, there is some i such that the elements of B are in different i-blocks, but the same (i + 1)-block of permutation s. Since there are r - 1 i-blocks in each (i + 1)-block, then r - 1 is E] the maximum possible number of elements in B. 3.2 Algorithm for computing fw The following algorithm for computing fw(u) is an implementation in Python of the method for computing formation width in Corollary 26. Specifically if u is a nonempty sequence with r distinct letters, then the algorithm increments s starting from 1 until it finds that every binary (r, s)-formation contains it. If some binary (r, s)-formation f contains u, then for every s' > s the algorithm does not check for containment of u in any binary (r, s')-formations f' for which f' restricted to its first s permutations is equal to f. from itertools import permutations #determines whether one sequence is a subsequence of another: def issubseq(seq, subseq): 27 if len(subseq) == 0: return True else: if len(seq) == 0: return False elif seq[-1] == subseq[-1]: return issubseq(seq[:-11,subseq[:-1]) elif seq[-1] != subseq[-1]: return issubseq(seq[:-1],subseq) #determines the formation width of u: def fw(u): 1=len(set(u)) v = list(u) rsformset = set() rsformsetl = set() q = tuple(range(l)) qi = q[::-1] rsformset.add(q) rsforml=q while len(rsformset)!=0: for rsforms in rsformset: done=False for perms in permutations(range(l)): for i in range(len(u)): v[i] = perms[u[i]] if issubseq(rsforms, v): done=True break if not done: rsformsetl.add(rsforms+q) rsformsetl.add(rsforms+ql) rsforml=rsforms+q rsformset.clear() for rsform in rsformsetl: rsformset.add(rsform) rsformsetl.clear() return len(rsformi)//1 #u must be nonempty tuple with letters 0,1,2,..., e.g.: print fw((0,1,2,3,4,5,0,2,3,1,4,5,0)) 28 3.3 Using binary formations to compute fw If u has one distinct letter, then fw(u) is the length of it. If u has two distinct letters, then fw(u) also depends only on the length of u. Lemma 31. If u has two distinct letters and length t, then fw(u) = t - 1. Proof. By Lemma 19 it suffices to prove this lemma for sequences with different first and second letters. The upper bound follows since every (2, t - 1)-formation contains u. For the lower bound it suffices to construct a (2, t - 1) formation f(u) which only contains copies of u for which the last letter of the copy of u is the last letter of f(u). Therefore the (2, t - 2)-formation in the first t - 2 permutations of f(u) avoids u, so fw(u) > t - 2 by Corollary 26. Assume without loss of generality that u starts with xy. Construct f(u) by ignoring the leading x and replacing every x in u by ba and every y by ab. Let it' denote the sequence obtained by deleting the last letter of it. We prove by induction on the length of i that f(u) contains only copies of u for which the last letter of the copy of u is the last letter of f(u). The first case to consider is i = xy. Since f(xy) = ab, then f(xy) contains exactly one copy of the sequence xy and the last letter of the copy of xy is the last letter of f(xy). Suppose by inductive hypothesis that f(u') contains only copies of u' for which the last letter of the copy of u' is the last letter of f(u'). If the last two letters of u are the same, then the first letter of the last permutation of f(u) is different from the last letter of f(u'), so the last letter of f(u) will be the last letter of any copy of u in f((u). If the last two letters of u are different, then the first letter of the last permutation of f(u) is the same as the last letter of f(,'), so the last letter of f(u) will be the last letter of any copy of El it in f(it). If u has at least three distinct letters, then fw(u) cannot be determined solely from the length of u and the number of distinct letters in u. For example fw(abcabc) = 3 and fw(abccba) = 4. The next lemma identifies all sequences i for which fw(u) = 3. As a result of Corollary 27, deleting any letters which occur just once in u will not change the value of fw(u) unless a has no other letters. We call a sequence reduced if every distinct letter in the sequence occurs at least twice. By Lemma 31, fw(u) = 1 if and only if it is nonempty and each distinct letter in it occurs once, and fw(u) = 2 if and only if one letter in u occurs twice and every other distinct letter occurs once. Lemma 32. If u is reduced and fw(u) = 3, then either there exists some 1 > 2 for which u is isomorphic to up(l, 2) or u is isomorphic to one of the sequences aaa, aabb, abba, abcacb, abcbac, abccab, or abcdbadc. Proof. Since u is reduced, then every distinct letter in u occurs at least twice. If any letter in u occurs three times, then it is the only letter in it and u is isomorphic to aaa, or else fw(a) > 4 by Lemma 31. If u is not isomorphic to aaa, then every distinct letter in it occurs twice. 29 Suppose u is not isomorphic to up(l, 2) for any I > 2. Then there exist two distinct letters x and y in u for which the subsequence consisting of occurrences of x and y is isomorphic to aabb or abba. If x and y are the only distinct letters in u, then u is isomorphic to aabb or abba. If u has three distinct letters, then t is isomorphic to a sequence obtained by adding two occurrences of c anywhere in aabb or abba, so we consider 30 cases. If u had the form xxv or vxx for some letter x and sequence v of length 4 with two distinct letters not equal to x, then fw(v) = 3 by Lemma 31, so fw(u) = 4 by Lemma 19. This eliminates the cases aabbcc, aabcbc, aacbbc, aabccb, aacbcb, aaccbb, acacbb, caacbb, accabb, cacabb, ccaabb, abbacc, and ccabba. The binary (3,3)-formation xyzxyzxyz avoids caabbc, abbcca, accbba, cabbac, acbbca, and abccba. The binary (3,3)-formation xyzzyxxyz avoids acabcb, abcbca, and acbeba. The binary (3, 3)-formation xyzzyxzyx avoids acabbc and cacbba. So its reverse avoids caabcb and abbcac. Thus each of these sequences have formation width at least 4 by Corollary 26. If u is one of the remaining sequences abcbac, acbbac, cabbca, or cabcba, then fw(u) = 3. Thus every reduced sequence ? with three distinct letters for which fw(u) = 3 is a (3, 2)-formation. Note that acbbac and cabbca are isomorphic to abccab, and cabcba is isomorphic to abcacb. If a has four distinct letters, then u is isomorphic to a sequence obtained by adding two occurrences of d to the sequence abcabc, abcacb, abcbac, or abccab. If u was not a (4, 2)-formation, then u would contain a reduced sequence v with three distinct letters which was not a (3, 2)-formation, so fw(7) > 4. We consider all (4, 2)-formations with first permutation abcd. The binary (4,3)formation xyzwxyzwxyzw avoids abcdadcb, abcdbdea, abcdcbad, abcdcbda, abcddacb, abcddbac, abcddbca, abcddcab, and abcddcba. The formation xyzwxyzwwzyx avoids abcdbacd, abcdcabd, and abcdcadb. The formation xyzw'wzyxxyzw avoids abcdacbd, abcdacdb, abcdbcad, abcdbcda, abcdcdab, and abcdcdba. The formation xyzwwzyxwzyx avoids abcdabdc, abcdadbc, abcdbdac, and abcddabc. Thus each of these (4, 2)formations have formation width at least 4 by Corollary 26. If u is abcdbadc, then fw(u) = 3. If a had five distinct letters, then a must be a (5, 2)-formation or else fw(u) > 4. If a was any (5, 2)-formation with first permutation abcde, then every (4, 2)-formation in u would be isomorphic to abcdbadc or up(4, 2). It is impossible for a (5, 2)-formation to have both a subsequence isomorphic to abcdbadc and another subsequence isomorphic to up(4, 2), so every (4,2)-formation in u would be isomorphic to abcdbadc or else a would be isomorphic to up(5, 2). In particular u must have both abcdbadc and acdecaed as subsequences, a contradiction. If a had r distinct letters for some r > 5 and a was not isomorphic to up(r, 2), then u would contain a subsequence of length 10 with five distinct letters that was not isomorphic to up(5, 2), so fw(u) > 3. l The last lemma can also be verified by using the Python formation width algorithm. The next lemma provides an upper bound on fa (u) for every binary formation u. It is tight if u = up(1, t) for any 1 > 2 and t > 1. 30 Lemma 33. Let u = Ie1D 23... that ej > 0 for each i and E', L-, where L is I if n is odd and D if n is even so ej = k. Then fw(u) < c(k - em) + 2e, - 1 for all m. Proof. Let k1 = ej and k2 = E ml e.. In any binary (c, c(k - em) + 2em - I)formation f, there is a copy of up(c, em) in permutations cki + 1 through ck 1 +2em -I of f by the pigeonhole principle. This copy of up(c, em) can be extended to make a copy of u in f by using one letter from each of the remaining cki + ck2 permutations of f. Thus fw(u) c(k - em) + 2em - 1 by Corollary 26. El Theorem 34. fw(up(l, t)) = 2t - 1 for every 1 > 2 and t > 1. Proof. For the lower bound fw(up(l, t)) > fw((ab)t) = 2t - 1 since up(1, t) contains (ab)t. The upper bound fw(up(l, t)) < 2t - 1 follows from Lemma 33. l Therefore fw(u) = 2t - 1 for every sequence a such that a contains (ab)t and there exists I > 2 for which up(l, t) contains u. As a corollary this implies the upper bounds in the next result, which gives nearly tight asymptotic bounds on Ex(up(l, t), n). The lower bounds in the next corollary follow from the lower bounds on Ex((ab)t , n) in [2] by Lemma 15. Corollary 35. Ex(up(1, t), n) = n2( a(n) + (a(n)-) for all 1 > 2 and t > 3. As a result, the constant c improves in the (n log n)2a(n)c upper bound from [451 on the maximum number of edges in k-quasiplanar graphs on n vertices with no pair of edges intersecting in more than 0(1) points, since their proof used the bounds Ex(up(l, t), n) nl21t-3(10l)oa(n)" from [30]. 3.4 Bounding the formation width of binary formations In this section we compute the exact values of l(u) and r(u) for all binary formations u. This yields upper and lower bounds on fw(u) which differ by at most a factor of two for each binary formation a. 3.4.1 Computing 1 If 7r E S, and u is a sequence on the letters 1, ... ,c, then let 4,(u) = k if u is a subsequence of I, but u is not a subsequence of Ig~1. It follows that 1(u) = min,asc{4,(u)}. Lemma 36. If 1,(Ic) = a and 1,(Dc) = b, then a + b c +1. Proof. Represent the permutation K by the set of points (i, r(i)). Connect points (i, 7r(i)) and (j,7r(j)) if i <j and r(j) = ir(i) + 1. This partitions the points into a connected sections. In a different representation connect points (i, 7r(i)) and (j, 7r(j)) if i <j and 7r(j) = 7r(i) - 1. This partitions the points into b connected sections. 31 We count the total number of endpoints of connected sections of points in both representations in two ways so that each connected section of points is considered to have two endpoints, even when the section consists of a single point. Since every connected section has two endpoints, then there are a total of 2(a + b) endpoints. Alternatively every point (i, ir(i)) contributes two endpoints, unless r(i) = 1 or ir(i) = c, in which case (i, 7r(i)) contributes three endpoints. Thus there are a total of 2c + 2 endpoints, so a + b = c+ 1. El Corollary 37. l(ID,) = c + 1 for every c > 1. Corollary 38. fw(IcDc) = c + 1 for every c > 1. Proof. The upper bound is trivial. The lower bound follows since I, avoids ID,. E If u and v are sequences on the letters 1, . .. , c, then 4,(u) + 4,(v) - 1 < 4r(Uv) 41 (u) + 4,(v). Say that u and v 7r-overlap if 4,(uv) = 1,(u) + 1,(v) - 1. Then u and v 7r-overlap if and only if the last letter of u and the first letter of v 7r-overlap. For each 7r E Sc, the sequences I, and D, do not 7r-overlap since the last letter of Ic is the first letter of Dc, and D, and I, do not r-overlap since the last letter of D, is the first letter of Ic. Furthermore if c > 2, then exactly one of the two sequences I, or D, T-overlaps itself, depending on the order in which the first and last letters of I, occur in I,. Moreover for any sequence u, if 1,(u) = 1 then u does not 7r-overlap itself. The next theorem implies Theorem 24 since l(u) fw(u) < 21(u) - 1. Theorem 39. Fix c > 2 and let u = I,"D 2 13 ... L, where L is I if n is odd and D if n is even, and ej > 0 for all i. Define A =Li> e 2 i 1 and B = >1 e 2i. Let Al = max(A, B) and let rn min(A, B). Then 1(u) = (c - 1)m + M + Lj. Proof. Fix an arbitrary 7r c S, and let l(Ic) = a and 4,(D,) = b. We show 1,(u) > L1J by considering two cases depending on whether I, or D, 7r-overlaps (c- 1)m++A' itself. Case 1: Ic ir-overlaps itself. In this case /,(Iel) = (a - 1)ej + 1 and i,(D'i) bei. Since I, and D, do not 7r-overlap and D, and Ic do not 7r-overlap, then 41(u) (a - 1)A + bB+ [+~. Lemma 36 implies that (a - 1) + b = c, while b > 0 and a > 1 since Ic r-overlaps itself. Then - ,(u) > (c- 1)n + M + [. Case 2: D, r-overlaps itself. In this case 41(IL,) = aej and 1,(De) = (b -1)e+1, so 1,(u) = aA+(b- 1)B Moreover a + (b - 1) = c, a > 0, and b > 1. Then 1,(u) > (c - 1)m + M + [l. [J. Thus in either case 1,(u) > (c - 1)m + M + LL J. If A > B, then this value is attained by letting 7r be the identity permutation. If B > A, then this value is attained by letting 7r(1) = 1 and ir(i) = c+ 2 - i for 2 < i < c. l 32 3.4.2 Computing r For every binary formation u we compute r(u), and we identify when r(u) > l(u). Theorem 40. If c > 2 and ei > 0 for all i, then r(If1D1I23 where L is I if n is odd and D if n is even. .. . L-) = 2 t ei - n, Proof. First we show that r(I) = 2x - 1 for every x > 0. The upper bound is trivial. For the lower bound we also show that alt(c, 2x - 1) has the subsequence Ix only if r(c) = c. We proceed by induction on x. Clearly r(Ic) = 1. In addition, , is a subsequence of Ic only if 7 is the identity permutation, so ir(c) = c. For the inductive hypothesis assume that r(Ix) = 2x - 1 and that alt(c, 2x - 1) has the subsequence Ix only if 7r(c) = c. We claim that r(I+1 ) = 2x + 1 and that alt(c, 2x + 1) has the subsequence I ?-+ only if ir(c) = c. 4 Let 7r be an arbitrary permutation. We will first show that Ix+1 is not a subsequence of alt(c, 2x). Suppose for contradiction that Ix+1 is a subsequence of alt(c, 2x). Then Ix is a subsequence of alt(c, 2x - 1), so 7r(c) = c. Then the last letter in Ix+1 must be the first letter of the last permutation of alt(c, 2x), a contradiction. Thus r(I+ 1 ) = 2x + 1. We still must show that alt(c, 2x + 1) has the subsequence I+l only if 'r(c) = c. Suppose 7r(c) = i for some 1 < i < c, and assume for contradiction that Ix+l is a subsequence of alt(c, 2x +1). Since Ix is not a subsequence of alt(c, 2x -1), then the second to last i in Ix+1 must occur in the second to last permutation of alt(c, 2x + 1) and the last i in I+1 must occur in the last permutation of alt(c, 2x + 1). Since i < c, then there are at most c - 2 distinct letters between the occurrences of i in the last two permutations of alt(c, 2x + 1), a contradiction. Thus alt(c, 2x + 1) has the subsequence Ix+l only if wr(c) = c. This completes the induction. By symmetry we find that r(Dx) = 2x- 1 for every x > 0. We now prove the claim that r(Ig1D 2I 3 ... L-) = 2Z ei - n. The upper bound is trivial since the copy of Ie, D12I ... L'- can be selected greedily from left to right in alt(c, 2 EZ ei - n). For the lower bound, suppose for some k and permutation 7r that alt(c, k) has the subsequence I,, D143 ... e with n sections of the form I4 or Dx. No section I4 or Dx can occur in fewer than 2x - 1 adjacent permutations of alt(c, k). Furthermore no different sections have letters occurring in the same permutation. Thus alt(c, k) contains at least 2ZL 1 ei - n permutations, so k > 2 E= ei - n. Corollary 41. Fix c > 2 and let u = I D ... fe, where L is I if n is odd and D if n is even, and ei > 0 for all i. Define A = E> e 2i- 1 and B = Zi> e 2 i. Let M = max(A, B) and let m = min(A, B). Then r(u) > l(u) if and only if M > (c - 3)m + n+ ["].Proof. This follows from setting 2 E= ei - n > (c - 1)m + M + [nJ since En m + M. ei = l 33 3.5 Further bounds on extremal functions using fw The lemmas in this section use Corollary 26 to identify sequences u with fw(u) > 3 for which fw(u) provides tight upper bounds on Ex(u, n), starting with an infinite set of sequences which contain ababa. Lemma 42. If u is any sequence of the form avav'a such that a is a letter, v is a nonempty sequence excluding a with no repeated letters and v' is obtained from v by only moving the first letter of v to another place in v, then fw(u) = 4. , Proof. Since t contains an alternation of length 5, then fw(u) > 4. Suppose a has r distinct letters for r > 2. In order to prove that fw(u) < 4, it suffices by Corollary 26 to show that u is contained in every binary (r, 4)-formation. First note that binary (r, 4)-formations isomorphic to I, or IrD, contain a copy of i which uses every letter in the first permutation. Furthermore if the position in V' of the occurrence of the first letter of v is right after the occurrence in v' of the ith letter of v, then I DrI, has a subsequence a' isomorphic to u such that the /th letter of a' is given by (r - i + j - 2 mod r) + 1 for each 1 < j - r. In particular the subsequence u' includes the last i + 1 letters in the first permutation of I,?DI,, all of the letters except r - i + 1 in the second permutation, the single letter r - i + 1 in the third permutation, and the first r - i letters in the last permutation. Thus every binary (r, 4)-formation isomorphic to I DrI contains a copy of u. Since every other binary (r, 4)-formation has a subsequence isomorphic to IrD then it suffices to observe that I,D' contains a copy of a that uses every letter in the third permutation. D Corollary 43. If u is any sequence of the form avav'a such that a is a letter, v is a nonempty sequence excluding a with no repeated letters and v' is obtained from v by only moving the first letter of v to another place imuv, then Ex(U, T) = (na(n)). Proof. The upper bound follows from the last lemma and Lemma 20, while the lower bound follows by Lemma 15 since a contains ababa. El The next corollary is obtained by reversing the sequences considered in the last lemma. Corollary 44. If u is any sequence of the form avav'a such that a is a letter, v is a nonempty sequence excluding a with no repeated letters and v' is obtained from v by moving a single letter in v to the end of v, then fw(u) = 4 and Ex(u, n) = e(na(n)). The next lemma implies that if v and v' are nonempty permutations of the same distinct letters excluding a, then fw(avav'a) = 4 if and only if v' is obtained from v by only moving the first letter of v to another place in v or by only moving a single letter in v to the end of v. Lemma 45. Let u be any sequence of the form avav'a such that a is a letter, v is a nonempty sequence excluding a with no repeated letters, and v' is a permutation of v which cannot be obtained from v by only moving the first letter of v to another place in v or by only moving a single letter in v to the end of v. Then fw(u) > 4. 34 Proof. If x is one of the sequences abcdadbca,abcdadcba,abcdeabdcea, or abcdeacbeda, then fw(x) > 4 . This can be verified using the formation width algorithm. Suppose u is a sequence of the form OvOv'0 for which fw(u) = 4, v is the sequence 12.. r, and v' is the permutation 7r72... Ir, of 12. .. r. Since u avoids abcdadbca and abcdadcba, then 7ri < i + 1 for each 1 Ki K r. Consider two cases. In the first case, 7r 1 = 1. If 7r i for each 1 K i K r then fw(u) = 4 since fw(up(r + 1,2)) = 3. Otherwise let m be minimal for which 7rm = m + 1. Then 7r j for each j < m. Since u avoids abcdeabdcea, then r, = m. Moreover 7r= j + 1 for m K j < r since 7r i + 1 for each 1 Ki <r. Thus v' can be obtained from v by only moving a single letter in v to the end of v. In the second case, 71 = 2. Let m be the index for which lrm = 1. Then 7r = J+1 for 1 < j < m since ri < i + 1 for each 1 < i K r. Since u avoids abcdeacbeda, then 7r1j=j for each j > m. Thus v' can be obtained from v by only moving the first letter of v to another place in v. For t < 4 the next lemma exhibits sequences with three distinct letters and t occurrences of each letter which contain (ab)' and have formation width 2t - 1. Lemma 46. If t is 2, 3, or 4 and z is any sequence of the form ax1 ax 2 ... axt such that a is a letter and xi is a sequence equal to either bc or cb for each 1 < i < t, then fw(z) = 2t - 1. Proof. The lower bound follows since z contains (ab)t. By Corollary 26, the upper bound is verified by checking that every binary (3, 2t - 1)-formation contains z. This can be done with the formation width algorithm. l ) Corollary 47. If t is 3 or 4 and z is any sequence of the form ax1 ax 2 ... axt such that a is a letter and xi is a sequence equal to either bc or cb for each 1 K i K t, then (n Q Ex(z, n) = n2 (t 2 )! Proof. The upper bounds follow from the last lemma and Lemma 20. The lower bounds follow from the lower bounds on Ex((ab)t , n) in [21 by Lemma 15. El There are sequences z of the form axiax 2 ax 3ax 4 ax5 such that a is a letter and xi is a sequence equal to either bc or cb for each 1 K i < 5 for which fw(z) > 9. For example fw(abcacbacbabcacb) = 10. The following lemma presents another infinite class of forbidden sequences with three distinct letters for which formation width yields tight bounds on extremal functions. Lemma 48. fw(abc(acb)t ) = 2t + 1 for t > 0. Proof. The proof is trivial for t = 0, so suppose that t > 0. Since abc(acb)' contains an alternation of length 2t + 2, then fw(abc(acb)t ) > 2t + 1. In order to prove that fw(abc(acb)t ) < 2t+ 1, it suffices by Corollary 26 to show that every binary (3, 2t+ 1)formation contains abc(acb)'. 35 . Consider any binary (3, 2t+1)-formation f with permutations xyz and zyx. Without loss of generality suppose that the last 2t - 1 permutations of f have the subsequence (xyz)'. Then f has the subsequence xzy(xyz) t unless the first six letters of f are zyxxyz. If the first six letters of f are zyxxyz, then f has the subsequence zyx(zXy) t Corollary 49. Ex(abc(acb)t ,n) - n2 rO(n)-1 (a(n)") for t > 2. Proof. The upper bounds follow from the last lemma and Lemma 20. The lower bounds follow from the lower bounds on Ex((ab)t , n) in [2] by Lemma 15. E] 3.6 Further bounds on fw For c > 2 the bounds on 1(u) imply that (c + 1)k K fw(alt(c, 2k)) < 2(c + 1)k - 1 and (c + 1)k + 1 < fw(alt(c, 2k + 1)) < 2(c + 1)k + 1 for every k. In this section we derive improved bounds on fw(alt(c, 2k)) and fw(alt(c, 2k + 1)) using Corollary 26. First we compute fw(alt(c, 3)) for all c > 2. In [411, Pettie proved for all c > 1 that Ex(alt(c, 3), n) = 0(n). Theorem 50. If c > 2, then fw(IcDcIc) = c + 3. Proof. First we prove for every permutation 7r c S, that IDI, is not a subsequence of the binary (c, c + 2)-formation 1,D2. Assume for contradiction that ID2 has the subsequence IDI, for some permutation r E Sc. Since l(IcDc) = c+ 1 by Corollary 37, then the last letter of D, must be in the first Dc in IcD . However, the first letter of I, is the same as the last letter of D, so the first letter of the last I, in IDI, must be in the last D, in IcD2. Then I, = D,, so the last letter of D, is c. This would imply that Ic has the subsequence I,D,. Since the last letter of I-' is c, then ID, would be a subsequence of Ic, a contradiction. Thus IcD2 does not have 1,DI, as a subsequence for any permutation 7 c Sc. Thus fw(IcDcIc) > c + 2 by Corollary 26. It remains to show that every binary (c, c + 3)-formation f has a subsequence I,D,I, for some permutation 7r E Sc. Without loss of generality suppose the first permutation of f is Ic. If f is I,+3, then f has IcDcIe as a subsequence. If f has an alternation of Ic and Dc terms of length at least 3, then also f must have IcDcIc as a subsequence. Otherwise f has the form IcfD' with a + b = c + 3, a > 0 and b > 0. If a K 2, then f has IcDcIe as a subsequence. If b K 2, then f has DcIcDc as a subsequence. Otherwise f has the subsequence IDI., such that I, is the sequence Ib2C... (b - 1) consisting of the integers from 1 to b - 2 followed by the integers in reverse from c to b - 1. In other words I, is obtained by reversing the last a - 1 letters of Ic. Thus fw (IcDcIc) < c + 3 by Corollary 26. El The next two lemmas are used for the lower bounds in the remaining theorems. Lemma 51. If c > 2 and w E Se, then 1,D, is a subsequence of IcDe if and only if 7r(1) < 7r(2). 36 , , Proof. Let 7r E Sc and suppose I,D, is a subsequence of IDc. Then the last letter of IrD7, namely 7r(1), occurs in the last D, of IDc since l(IcDc) = c + 1 by Corollary 37. If ir(1) is not the only letter of I,D, occurring in the last Dc, then 7r(2)ir(1) is a subsequence of D,. This is possible only if i(1) < 7(2). If the final D, contains no letters in ID, besides 7(1), then the last 7r(2) in I, D occurs in some I,. If ir(1) > 7(2), then the last w(1) in ID 1, can be replaced with the 7Tr(1) in the same permutation as the last 7r(2) in ID,. This would imply that I, D is a subsequence of Ic, which is impossible since l(IcDc) = c + 1. Thus ir(1) < 7r(2). For the other direction suppose that wr(1) < 7r(2). Then I,D, is a subsequence of I+1 with exactly one letter in the last permutation of Ic+1. Thus I,D, is a El subsequence of ItDc. Define the reverse permutation 7r, E S, so that 7r,(i) = c+ 1 - i for 1 < i < c. Corollary 52. If c > 2 and w E S,, then I,D, is a subsequence of DcIf if and only if 7r(2) <wr(1). Proof. By reflection, ID , is a subsequence of DcIf- if and only if ID, is a subsequence of DcIc. Moreover I,D, is a subsequence of D'Ic if and only if 27,(ID,) is a subsequence of IDc. By Lemma 51, 7r,(ID,) is a subsequence of IgDc if and only if 7r,(7r(1)) < r(ir(2)). Since 7r,(7(1)) < 7r,(ir(2)) if and only if 7(2) < 7(1), then ID, is a subsequence of DcIc if and only if r(2) < 7r(1). 0 Using these facts we determine fw(IcDc) and fw(alt(c, 4)). Pettie in [411 showed bounds of e(na(n)) on the maximum lengths of sequences with n distinct letters avoiding both ababab and alt(c, 4) for some c. This improved an upper bound by Ezra, Aronov, and Sharir in [12] on the complexity of the union of n 6-fat triangles. Theorem 53. If c > 2 and k > 1, then fw(ILkDc) = c + 2k - 1. Proof. The upper bound follows since fw(IDc) < fw(I,) + c. For the lower bound let Tk be the (c, c + 2k - 2)-formation obtained by concatenating alt(c, 2k - 2) and I,. We show that Tk avoids 4kDc by induction on k. This is clearly true for k = 1 since i(IcDc) = c + 1 by Corollary 37. For the inductive hypothesis assume that Tk avoids IcDc. Suppose for contradiction that Tk+1 has the subsequence Ik+1D, for some permutation 7r E Sc. The proof of Theorem 40 showed that r(I) =2k - 1 and I is a subsequence of alt(c, 2k - 1) only if ir(c) = c, so the last I,D, of I+ 1D, must be a subsequence of the rightmost DcIc in Tk+1. Then ir(1) > 7r(2). Since T avoids IcDe, then the first letter 7(1) of the second I, in I+1D, must occur in the initial IcDc of Tk 1. Thus 7r(1)7r(2)7r(1) must be a subsequence of IDce. This contradicts 7(1) > 7r(2), so Tk+1 avoids Ik+lDc. Thus fw(Ik-Dc) > c+2k-2 for every c > 2 and k > 1 by Corollary 26. El Theorem 54. If c > 2, then fw(IcDcIcDc) = 2c + 3. 37 Proof. Since c + fw(IcDcIc) > fw(IcDcIcDc), then 2c + 3 > fw(IcDcIcDc). As for the lower bound, the (c, 2c + 2)-formation F = IcD'Ic avoids I,D,ID, for all permutations 7r G Sc. To see this assume for contradiction that F contains IDID, for some permutation i c Sc. Since Ic does not contain I,D, by Corollary 37, then the first 1,D, is in the first IDc of F and the second I.D, is in the last DcI; of F. This is a contradiction by Lemma 51 and Corollary 52. Thus fw(IcDcIcDc) > 2c + 2 by Corollary 26. El We extend the technique used in the last proof to get an improved lower bound on fw(alt(c, k)) for all c > 2 and k > 5. Theorem 55. If c > 2 and k > 1, then fw(alt(c, 2k)) > k(c+2)-1 and fw(alt(c, 2k+ 1)) > k(c+ 2) + 1. Proof. Define TI = , T- = T 2 .D, and T2kTkI for k> 1. We prove that avoids alt(c, k) by induction on k. This implies that fw(alt(c, 2k)) > k(c+ 2) - 2 and fw(alt(c, 2k + 1)) > k(c + 2) by Corollary 26. Theorems 50 and 54 proved that Tk_1 T2 avoids alt(c, 3) and T3 avoids alt(c, 4). For the inductive hypothesis there are two cases. First assume that T-1 avoids alt(c, j) for all j < 2k-1, but suppose for contradiction that T2k-1 has the subsequence (ID, )k for some permutation i E Sc. Let G be the leftmost (I,D,)k-1 in the subsequence ( 1,D,)k. Since the leftmost T2k-3 in T2k-1 avoids alt(c, 2k - 2), then the last letter of G must occur somewhere in the rightmost D'Ic in T2k-1. Moreover the letter directly after G in (I,D,)k is the same as the last letter of G, so these two letters cannot occur in the same permutation. Thus the last I,D, in (ITD,)k must be a subsequence of the last DcI; in T2k-1. Then 7r(2) < r(1) by Corollary 52. Let H be the rightmost (ID,)k-1 in the subsequence (I,D,)k. Since the rightmost T2k-3 in T2k-1 avoids alt(c, 2k - 2), then the first letter of H must occur somewhere in the leftmost ID 2 in T2k-1. Moreover the letter directly before H in (I,D,)k is the same as the first letter of H, so these two letters cannot occur in the same permutation. Thus the first 1,D,, in (I,D,)k must be a subsequence of the first ID, in T2k-1. Then 7r(1) < 7(2) by Lemma 51, a contradiction. For the second case of the inductive hypothesis, assume that T_ 1 avoids alt(c, j) for all j < 2k, but suppose for contradiction that T2k has the subsequence (ID, )kjI2 for some permutation 7r E Sc. Let G be the leftmost (I,D,)k in the subsequence (I,D,)kI,. Since the leftmost T2k-1 in T2k avoids G, then the last letter of G must occur in the last D in T2k. The last letter of G is equal to the first letter of the last permutation of (I,D, )kI, so the last I, of (I,D,)kI, must be a subsequence of the final D, in T2k. Therefore 1, = D,, so the last letter of D, is c. This implies that (I.D )k is a subsequence of T2 k-iC, so (ID,) would be a subsequence of T2k-1, a contradiction. Thus (I-D,)kI is not a subsequence of T2k for any permutation 7 E Sc. El 38 3.7 Open Problems Many questions about formation width are left unresolved by the results in this chapter. We found several classes of sequences u for which u contained an alternation with the same formation width as u, which implied tight bounds on Ex(u, n). One problem is to find all sequences u for which u contains an alternation with the same formation width as u. We showed that fw(abc(acb)') = 2t + 1 for t > 0. This implied for t > 2 that Ex(abc(acb)t , n) = n2 t.a(n)t (o(n)t2. We conjecture the following result, which would imply nearly tight bounds on Ex(abc(acb)t abc, n). Conjecture 56. fw(abc(acb)t abc) = 2t + 3 for t > 0. We identified the set of all sequences u for which fw(u) K 3. These are all the sequences for which the value of fw(u) implies linear bounds on Ex(u, n). A next step would be to identify all sequences u for which fw(u) < 4, since these are all of the sequences for which the value of fw(u) implies O('ntar()) upper bounds on Ex(u, n). We also determined the values of 1(u) and r(u) for every binary formation u. Since both of these functions provide lower bounds on fw(u), it would be useful to compute the values of 1(u) and r(u) for every sequence u. On a related note, the values of 1(u) implied bounds on fw(u) within a factor of 2 for every binary formation u. What is the exact value of fw(u) for every binary formation u? We also obtained bounds on fw(alt(c, k)) for every k > 1. In particular we determined the exact values for k < 4. What is the exact value of fw(alt(c, k)) for each k > 5? In addition we proved that fl(u) < (r - 1) -1+ 1 for all sequences u with r distinct letters. What is the exact value of fl(u) for every sequence u? 39 40 Chapter 4 Linear extremal functions of forbidden 0 - 1 matrices For a collection S of 0 - 1 matrices, let the weight extremal function exs(n, S) denote the maximum number of ones in an n x n 0 - 1 matrix which avoids every matrix in S, and let the column extremal function exsk(m, S) denote the maximum number of columns in a 0 - 1 matrix with m rows which avoids every matrix in S and has at least k ones in every column. The column extremal function of matrices is the analogue of a sequence extremal function, the maximum number of letters in a sequence with m contiguous blocks of distinct letters that avoids a fixed subsequence (or collection of subsequences) and has at least k occurrences of each letter. This sequence extremal function was bounded for alternations and collections of sequences called (r, s)-formations in [38]. The column extremal functions for collections of matrices called doubled (r, s)-formations were bounded in [7]. Both the weight and the column extremal functions of permutation matrices were used in [15] to improve upper bounds on the Stanley-Wilf limits of forbidden permutations. Since an n x n matrix has n2 entries, then exs(n, S) < n2 for every collection S of 0 - 1 matrices. If S contains only one matrix M, then we write exs(n, S) = ex(n, M) and exsk(M, S) = exk(m, M). If M has at least two ones, then ex(n, M) ;> n since a matrix with ones only in a single column or a matrix with ones only in a single row avoids M. Fulek [17] found bounds on the extremal functions of the patterns L1 and L 2 (Figure 4-1) using visibility representations constructed by treating rows of a given n x n matrix as vertices and projections of ones on lower rows as edges. He bounded the number of ones in the matrix by limiting the multiplicity of edges in the visibility representation based on the forbidden 0 - 1 matrices. 0 1 1 1 0 1 0 0 0 1 0 0 1 0 0 0 1 1 0 1 0 0 1 0 1 0 0) Figure 4-1: L1 and L 2 41 In Section 4.1 we prove some general facts about exs(n, S) and exsk(m, S). In Section 4.2 we define bar visibility hypergraphs, bound their number of edges, and use these bounds to show that ex(n, L 3) = 0(n) and exk(m, L 3 ) = O(2) (Figure 4-2). We also bound the extremal functions of forbidden collections of 0 - 1 matrices corresponding to bar visibility hypergraphs. 0 1 1 1 0 0 0 0 0 0 0 0 0 1 0 0 0 1 0 Figure 4-2: L 3 4.1 Facts about exs(n, S) and exsk(m, S) If k > m, then exsk(m, S) = 0 for any collection S. If K < in and every matrix in L has at least k rows with ones, then exsj(M, S) = oc for j < k since a matrix with n rows, any number of columns, and ones in every entry in the first j rows avoids every matrix in S. If some matrix in S has k rows and c columns, then exsj(m, S) < (c - 1) (m) for all j > k. This is because a matrix A with (c - 1) (') + 1 columns, in which every column contains at least j ones, must contain c columns with ones in the same k rows by the pigeonhole principle, so A contains the matrix in S with k rows and c columns. Let ex(m, n, P) (resp. exs(m, n, S)) be the maximum weight of an in x n matrix which avoids the 0 - 1 matrix P (resp. collection S), so that ex(n, P) = ex(n, n, P). For every column of P, draw a segment connecting the topmost and bottommost ones. Call P range-overlapping if, for every pair of columns of P, there exists a horizontal line passing through the corresponding segments of both columns. U U * U * * U U Figure 4-3: The pattern on the left is range-overlapping. The pattern on the right is not range-overlapping because its final two columns have disjoint ranges. Theorem 57. For any range-overlappingP, ex(m, n, P) < k(exk(m, P) 42 + n). Proof. Let A be a 0 - 1 matrix with m rows, n columns, and ex(m, n, P) ones which avoids P. Consider column c in A. Starting from the top of c, divide its ones into clusters of size k, deleting the at most k - 1 remaining ones in the column. Move each cluster g after the first cluster horizontally to a new column immediately to the right of the column containing the cluster which was above g in the original column. Delete any columns with no ones. Call the newly formed matrix A', and suppose that A' has n' columns. Suppose for contradiction that A' contains P, and consider two cases. If the copy of P in A' contained at least two columns that originated from the same column c, then P could not be range-overlapping because the construction separated columns of A into clusters with non-range-overlapping row indices to obtain A'. If the copy of P in A' only contained columns derived from different original columns, then the original matrix A also contained P, a contradiction. Therefore, A' cannot contain P. In our construction, we deleted a maximum of k(n' + n) < n(k - 1) ones. Each column of A' contains k ones, so ex(m, n, P) l k(exk(m, P) + n). Lemma 58. If S is a collection of 0 - 1 matrices, c is a constant, and g is a function ) for k > c. such that exs(m, n, S) < g(m)+cn for all m and n, then exsk(m, S) < Proof. Fix k > c. Any matrix with m rows, exsk(m, S) columns, and k ones per column which avoids S has at most exs(m, exsk(m, S), S) < g(m) + exsk(m, S)c ones. Therefore exsk(m, S)k < g(m) + exsk(m, S)c, so exsk(m, S) < g() for all El M. . Corollary 59. If S is a collection of 0 - 1 matrices such that exs(n, S) = O(n) and every matrix in S contains the 2 x 2 identity matrix or two ones in the same row, then exsk(m, S) = ( Proof. Since exs(m, n, S) < max {exs(m, S), exs(n, S)}, the upper bound follows from Lemma 58. For the lower bound, let K be the matrix with m rows and [j] columns obtained from an [J x [IJ identity matrix by replacing every row r in the identity matrix with the k x [-] submatrix consisting of k adjacent copies of r and then adding rows full of zeroes at the bottom until K has m rows. Let K' be obtained - 1 since K' avoids from K by reflecting over a vertical line. Then exsk(m, S) ;> El S. Marcus and Tardos [36] showed that ex(n, P) = O(n) for any permutation matrix P, solving the Stanley-Wilf conjecture. Call a 0 - 1 matrix light if it has no pair of ones in the same column. The Marcus Tardos theorem was generalized in [201 to show that ex(n, Q) O(n) for any light 0 - 1 matrix Q such that for every two columns co and ci in Q with ones in the same row r, all columns between co and ci also contain ones in row r. Corollary 59 can be applied to the matrices Q to show that exk(m, Q) = 6(j), as well as to the matrices L 1 and L 2 in [17]. Let Pr,c denote the r x c matrix filled with ones. We find the exact value of exk(m, Pk,c) using a simple counting argument, and then bound exk(m, P,c) for all k >r. 43 (c - 1) () . Lemma 60. exk(m, Pk,) Proof. There exist (7) possible configurations By the pigeonhole principle, a matrix with (c configuration of k ones that occurs in at least c (c - 1) (7). We can obtain the same lower bound possible configuration c - 1 times; this matrix of k ones in columns of height m. - 1) (7) + 1 columns must have a columns. Therefore, exk(m, P,,) < by constructing a matrix with every avoids Ph,c. Hence, exk(m, Pk,c) l (c - 1)(7). Theorem 61. For all fixed k > r, eXk(m, P,,2) (mr). , . . Proof. First observe that eXr(M, P,2) = 0(mr), according to the last lemma. Since exk(m, P) is decreasing in k for every P, exk(M, P,2) = Q(mr) for k > r. For any 0 - 1 matrix M, let Gm be the graph obtained from M by letting every column of M be a vertex and adding an edge between two vertices if and only if their corresponding columns have ones in exactly r - 1 common rows. Note that r - 1 is the maximum number of rows a pair of columns may share without containing P,,2 We proceed by induction on k to prove that exk(n, Pr,2) = Q The first case is k = r. The last lemma implies that ex,(m, P,,2 ) ("). Observe that if M is a 0 - 1 matrix with m rows and r ones in each column which avoids Pr,2 then the maximum degree of any vertex of Gm, A(Gm), is at most r(m - r). Fix k > r. Let M be the matrix obtained in the last inductive step avoiding Pr,2 with m' < xm rows, (') columns, and k - 1 ones in each column, such that A(GM) < ym for some constants x and y that depend only on r and k. Using a greedy algorithm, color the vertices in Gm using A(GA) + 1 colors so that no two vertices with a common edge have the same color. Let C : VGm -+ {1,... , A(Gm) + 1} be the coloring function. Construct a new matrix A' from M by adding A(Gm) + 1 new rows to the bottom of Al with a one in row m' + r of column b if and only if G(b) = r. Since A(GA) < yin, the resulting matrix M' has at most xm + ym + 1 rows. Fix a column b of M'. A column c is a neighbor of b in Gm, but not in Gm only if there are exactly r - 2 rows in M which have ones in both b and c, and b and c were assigned the same number in the coloring of Gm. Since b has k - 1 rows with ones in Ak, there are (-1) ways to choose r - 2 rows that contain ones in column b of Al. For every set of r - 2 rows in which column b has ones in M, there are at most xm-(k-1) possible new neighbors c of b in M' with ones in those rows, since every pair o new neighbors of b in Gm' with ones in those r - 2 rows get the same color in Gm and thus have no common rows in M containing ones besides the r - 2 rows each neighbor shares with b. Then A(Gm,) < "--(k-1) k-1) + ym, completing the induction. l Thus for each k > r, there exists a constant ark such that exk(ar,km, P,2) > (M) To generalize the bounds for all integral values of m, let n satisfy ar,kn < m ar,k (n+ 1) and observe that exk(ar,kn, Pr,2 ) > (n) > ( all fixed k > r. ~k). Hence exk(M, P,2 ) = (ir) for Corollary 62. For all fixed k > r and c > 2, exk(m, Pc) 44 O(inr). Proof. For c > 2, exk(rm, Pr,c) > exk(M, P,,2) since P,, contains P,,2. Therfore exk(m, P,,,) = Q(mr) for all k > r. Since ex,(m, Pr,c) (c-i)(), then eXk(rM, P,c) = O(m') for all k > r. This shows that exk(m, P,c) = O(m') for all fixed k > r and c > 2. l Corollary 63. If P is a 0-1 matrix with r rows which contains P,2 , then exk(m, P) 0(n'r) for all fixed k > r. 4.2 Bar s-visibility hypergraphs and 0 - 1 matrices A bar s-visibility hypergraph's vertices are finite, disjoint, horizontal bars in the plane and its edges are subsets of vertices of size s + 2 for which a vertical segment exists which intersects only the vertices in the edge. Using the fact that every bar visibility graph (i.e. bar 0-visibility hypergraph) is a planar graph, Fulek converted 0 - 1 matrices into bar 0-visibility hypergraphs to prove linear bounds on the extremal functions of patterns L1 and L 2 [17]. For r, s > 0, define T,,, to be the collection of matrices M which have r + s + 2 rows and r + 2s +2 columns such that Al restricted to the first s +1 columns and rows 2,. . . , s + 2 is a (s 1) x (s + 1) permutation matrix, M restricted to the last s + 1 columns and rows 2, ... , s + 2 is a (s + 1) x (s + 1) permutation matrix, M restricted to the middle r columns and the last r rows is an r x r permutation matrix, and M has ones in the middle r columns in row 1. For example T1,0 contains a single 3 x 3 matrix with four ones in a diamond formation. One of the patterns in T4,1 appears in Figure 4-4, with black squares representing ones and white squares representing zeroes. We generalize the technique from [171 to show for all r, s > 0 that exs(n, Tr,,s) 0(n) and exsk(m, T,,,) = 0(1). The first step is to prove a linear bound on the number of edges in a bar s-visibility hypergraph with n vertices. This proof is similar to the proof of the maximum number of edges in bar s-visibility graphs with n vertices in [9]. We assume that all bar endpoints have distinct coordinates since this does not decrease the maximum number of edges. Lemma 64. All bar s-visibility hypergraphs with n vertices have at most (2s + 3)n edges. Proof. Scan any representation of the given bar s-visibility hypergraph in the plane from left to right, making a list of distinct edges. Add an edge to the list whenever the scan for the first time shows part of the representation in which a vertical segment can be drawn which intersects each of the vertices in the edge. Then edges will only be added to the list whenever the scan passes the left or right end of some bar. For each bar B, the maximum possible number of edges added to the list when the scan passes the left end of B is s + 2 since there are at most s + 2 vertical segments representing different edges which pass through the left end of B and through s + 1 other bars. The maximum possible number of edges added to the list when the scan passes the right endpoint of B is s + 1, since at the right endpoint of B there are at 45 Figure 4-4: A matrix in T4,1 most s + 1 vertical segments representing different edges which pass through s + 2 bars other than B, at least one of which is below B and at least one of which is above B. Then there are at most (2s + 3)n edges on the list since the representation contains 5 n bars. In the next lemma, we change 0 - 1 matrices avoiding T,,, into bar s-visibility hypergraphs, and then show that the resulting hypergraphs have edge multiplicity at most r - 1. Theorem 65. For all r > 0 and s > 0, exs(n, Tr,s) = O(n). Proof. Let M be an n x n matrix which avoids T,,,. Define M' to be the matrix obtained from M by deleting the first s + 1 and last s + 1 ones in every row, and the last r ones in every column. Construct a representation of a bar s-visibility hypergraph H from M' by drawing a bar in each row that contains a one in M' with left end at the first one of M' in the row and right end at the last one of M' in the row. For each one in M' with at least s + 1 ones below it in M', draw a vertical line segment representing an edge starting from the one and extending through s bars until reaching the (s + 1)" bar below the one. Suppose for contradiction that H contains some edge e with multiplicity at least r. Let u, ... , u+ 2 , in order from top to bottom, be the rows of M' which contain the vertices in the edge e, and let c 1,... , c, be columns of M' which contain r vertical segments representing the copies of e. Let v 1 ,... , v, be distinct rows of M such that vi contains one of the bottommost r ones of ci for each i = 1, . . . , r; let di, ... , d,+ 1 be distinct columns of M such that di contains one of the s+1 leftmost ones of ui for each i = 2, ... , s+2; and let e, ... , e.+1 be distinct columns of M such that ej contains one of the s+1 rightmost ones of ui for each i = 2,. . . , s+2. Then the submatrix of M consisting of rows u1 ,...,us+2 , vI,...,v,. and columns ci,...,c,,d, .. ,d s +1,e, ... , e,+ contains a matrix in T,,,, a contradiction. 46 Then every edge of H has multiplicity less than r, so the number of ones in M is El at most (3s + 3 + r)n + (r - 1)(2s + 3)(n - r). Observe that every element of T,,, contains the pattern L 3 for r which implies the following corollary. 3 and s > 1, Corollary 66. ex(n, L 3 ) = 0(n). The next two corollaries follow from Corollary 59. Corollary 67. exsk(m, T,,,) = 0(). Corollary 68. exk(m, L 3) - 0(M). 4.3 Infinitely many minimal non-linear 0 -1 matrices In this section we prove that any k x 2k double permutation matrix has a 20(k)n extremal function. Let the pattern P be any k x 2k double permutation matrix with k > 1. Fix an integer n that is divisible by k. Choose any n x n matrix A with ex(n, P) ones that avoids P. We now partition A into (,)2 submatrices, each with k rows and k columns. Define block Sij, a square submatrix of A, to be the intersection of rows k(i - 1) + 1 through ki of A with columns k(j - 1) + 1 through kj of A. We construct an 1 x n 0 - 1 matrix Q. Define the ones of Q row by row from top to bottom, in each row from left to right such that qij = 1 if: (1) Sij is the leftmost block with a one in its row of blocks or (2) Sij has a one and some row of A contained in the blocks between Sij and Sij inclusive (where j' is the greatest number less than j such that q, = 1) has at least two ones. Q avoids P, or else A would contain P. We also define chunks of blocks, Cij, which are sets of blocks in the same blockrow R,. A chunk Cij consists of a block Sij with qij = 1 on the left and all blocks, including empty ones, between it and Sij, the next block to the right of Sij with qij = 1. The chunk Cij does not include Sti. If Sij is the rightmost block in its block-row with qij = 1, then its chunk contains it and all blocks in its block-row to the right of it. Hence the chunks partition the non-empty blocks of A. Note that if a chunk has two ones in the same row, then both ones must be in the same block. Note also that if block Sij has two ones in the same row, then qij = 1 and qij' = 1 for the next non-empty Sij, because the row with two ones is contained between Sij and Sij inclusive. Hence if a chunk has two ones in the same row, then there is a single non-empty block in the chunk. Thus all chunks fall into two categories: (1) chunks with a single non-empty block or (2) chunks with at least two non-empty blocks and no rows with more than a single one. Call chunk Cij wide if it has at least two ones in a single row. Call Cij tall if it has ones in every row. Chunks that are neither wide nor tall have at most k - 1 ones total. 47 Lemma 69. The number of wide chunks in A is at most ex(!', P'), where P' is the k x k permutation matrix obtained from P by removing duplicate columns. Proof. If A had more than ex(!, P') wide chunks, then the entries in those wide chunks would form a copy of P. This would be a contradiction. l Lemma 70. In each row of chunks Ri = USj for 1 < j chunks is less than 2k. 1g, the number of tall Proof. If the number of tall chunks in Ri was at least 2k, then Ri would contain a copy of P. This would be a contradiction. El + Lemma 71. Suppose that P is the k x 2k double permutation matrix correspondingto the k x k permutation matrix P. For n divisible by k, ex(n, P) < 2nk+ex(!-, P')k2 (k - 1)ex(n, P). Proof. We account for all of the ones in A. First, there are at most ex(L, P') wide chunks in A, so there are at most ex T, P')k2 ones in wide chunks in A. There are at most 2k2 = 2n tall chunks in A, so there are at most 2nk ones in chunks that are tall but not wide. The remaining chunks are neither tall nor wide, so they each have at most k - 1 ones. There are a total of at most ex(, P) chunks since Q avoids P, so the number of ones in A is at most 2nk + ex(i, P')k2 + (k - 1)ex(11, P). l Theorem 72. If P is a k x 2k double permutation matrix, then ex(n, P) = 20 ()n. ) + Proof. Let P' be the k x k permutation matrix obtained by removing duplicate columns from P. Fox proved that ex(n, P') = 2O(k)n [15]. Let c > 1 be a constant such that ex(n, P') < 2ckn for all n. By the last lemma, ex(n, P) < 2nk 2cknk 2 + (k - 1)ex(L, P) < 22 ckn + (k - 1)ex(-, P) for every n divisible by k. For n < k, it is clear that ex(n, P) 5 2k22 ckn. FLx m, suppose that ex(n, P) < 2k2 2 ckn for all n < m, and consider the case n = m. Let N be the largest multiple of k less than or equal to n. Then ex(n, P) < ex(N, P)+2kn - 22 ckN +k -e Ji 2kn < ( 2 2ck + 2(k - 1) 2 2ck + 2k)n < 2k2 2 ckn. l Keszegh defined an infinite sequence of 0 - 1 matrices Hk for k > 0 and proved ex(n, Hk) = Q(n log n) for all k > 0 [28]. Hk is the 3k + 4 x 3k + 4 0 - 1 matrix with all entries zero except h4 , 1 = h1,2 = hl,3 = h3k+3,3k+4 =h3k+2,3k+4 = 1 and h3i+4,3i+1= h 3i- 1 ,3 +3 =3i,3i+2 = 1 for each 1 K i K k. Corollary 73. There exist infinitely many pairwise distinct minimal nonlinear 0 - 1 matrices. Proof. This result follows because for every k there are at least k +1 ones that can be deleted from HA to produce a matrix with a linear extremal function. In particular, for each 0 K i K k, the matrix obtained by deleting the one in Hk at h3 i+4 ,3i+1 has a linear extremal function. This follows from Theorem 72 and the following fact proven by Keszegh [28]: if P is any 0 - 1 matrix with a one in its bottom right corner, Q is any 0 - 1 matrix with a one in its top left corner, and R is obtained by joining P and Q so that they overlap only in the bottom right corner of P and the top left corner of Q, then ex(n, R) < ex(n, P) + ex(n, Q). 48 l Chapter 5 Extremal functions of forbidden multidimensional matrices A d-dimensional ni x ... x nd matrix is denoted by A = (a.....,id), where 1 < il < ne for f = 1, 2,... , d. An -cross section of matrix A is a maximal set of entries ail,. ,id with il fixed. An f-row of matrix A is a maximal set of entries ail . ,,d with ij fixed for every j $ 1. A d-dimensional k x ... x k 0 - 1 matrix is a permutation matrix if each of its i-cross sections contains a single one for every f = 1, ... , d. The Kronecker product of two d-dimensional 0 - 1 matrices M and N, denoted by M D N, is the d-dimensional matrix obtained by replacing each 1 in M with a copy of N and each 0 in Al with a 0-matrix the same size as N. As with two dimensions, we say that a d-dimensional 0 - 1 matrix A contains another d-dimensional 0 - 1 matrix P if A has a submatrix that can be transformed into P by changing any number of ones to zeroes. Otherwise, A is said to avoid P. Let f(n, P, d) be the maximum number of ones in a d-dimensional n x ... x n 0 - 1 matrix that avoids a given d-dimensional 0 - 1 matrix P. It is easy to obtain trivial lower and upper bounds on f(n, P, d). Proposition 74. If P is a d-dimensional0- 1 matrix that contains at least two ones, then nd-i < f (n, P, d) < nd. Proof. For every P, we can choose a d-dimensional n x ... x n zero-one matrix A, with ones in a single 1-cross section for some 1 and zeroes elsewhere, such that A avoids P. This implies the lower bound, the upper bound is the number of entries in a d-dimensional n x - - - x n matrix. L The upper bound in Proposition 74 is a factor of n higher than the lower bound. Most work on improving this bound has been for the case d = 2. Pach and Tardos proved that f(n, P, 2) is super-additive in n [39]. By Fekete's Lemma on superadditive sequences [13], the sequence {"f(nP,2) } is convergent. The limit is known as the Fiiredi-Hajnal limit [6, 15]. Cibulka [6] showed that this limit is at least 2(k - 1) when P is a k x k permutation matrix and that the limit is exactly 2(k - 1) when P is the k x k identity matrix. 49 Marcus and Tardos [36] showed that the Ffiredi-Hajnal limit has an upper bound of 20(k log k) for every k x k permutation matrix P, and Fox [15] improved this upper bound to 2 0(k). Klazar and Marcus [341 bounded the extremal function when the d-dimensional matrix P is a permutation matrix of size k x - x k, generalizing the d = 2 result [36] by proving that f(n, P, d) = O(nd-1). In particular, they showed that f(",fPd) = 2 0(klogk), which generalizes Marcus and Tardos' upper bound on the ndI Fiiredi-Hajnal limit [36]. We observe that f(n, P, d) is super-homogeneous in higher dimensions for certain matrices P, i.e., f(sn, P, d) > Ksd-lf(n, P, d) for some positive constant K. This super-homogeneity is key to our proof for n> 2 LkI/dJ /20 that f jP has a lower bound of 2(k1/d) for a family of k x ... x k permutation matrices, generalizing Fox's result [14] from d= 2 to d> 2. For each fixed d > 2 we also show an upper bound of 2 0(k) on f_ 7 Od) for every k x ...x k permutation matrix P. This improves the previous upper bound of 2 0(k log k) for d > 2 [341, and it also generalizes Fox's bound 20(k) on the Fiiredi-Hajnal limit when d = 2 [15]. Moreover for each fixed d > 2 we prove that this upper bound 2 0() is also true for all d-dimensional double permutation matrices P of dimensions 2k x k x .-.x k. 5.1 Upper and lower bounds for d-dimensional permutation matrices When d > 2, it is still an open problem to prove the convergence of the sequence { f(nP,d) Theorem 75. if P is a d-dimensionalkx -- : x k permutation matrix, then "aPI = 20(k) The proof uses the notions of cross section contraction and interval minor containment [15]. Contracting several consecutive i-cross sections of a d-dimensional matrix means replacing these i-cross sections by a single i-cross section and placing a one in an entry of the new cross section if and only if at least one of the corresponding entries in the original I-cross sections is a one. We say that A contains B as an interval minor if we can use repeated cross section contraction to transform A into a matrix which contains B. A avoids B as an interval minor if A does not contain B as an interval minor. The containment in previous sections is at least as strong as interval minor containment. Indeed, A contains B implies that A contains B as an interval minor. However, since a permutation matrix has only one 1-entry in every cross section, containment of a permutation matrix P is equivalent to containment of P as an interval minor. Analogous to f(n, P, d), we define m(n, P, d) to be the maximum number of ones in a d-dimensional n x - - - x n zero-one matrix that avoids P as an interval minor. 50 Let Rkl..'kd be the d-dimensional matrix of dimensions k, x ... equal to 1. We observe that f(n, P, d) m(n, R'..-'k, d) x kd with every entry (5.1) for every k x ... x k permutation matrix P. This follows from the fact that containment of Rk,.-k as an interval minor implies containment of P. Lemma 76. m(tn, R k..,k d) < sdm(n, R'.k---, d) + dntdfk .... k(n, t, s, d), (5.2) where fA,,...,kd(nt,s,d) denotes the maximum number of 1-rows that have at least s ones in a d-dimensional t x n x --- x n matrix that avoids Rkl..-kd as an interval minor. Proof. Let A be a d-dimensional In x ... x In matrix that avoids R . as an interval minor and has m(tn, Rk k,.., d) ones. The S-submatrices of A are constructed by dividing A into nd disjoint submatrices of size t x ... x t and labeling these submatrices as S(ii, ... , is). The submatrix S(ii, ... , id) is defined to contain the entries a 1,...,d of A such that t(ie - 1) + I j< < tit for each?= 1,..., d. The contraction matrix of A is defined to be the d-dimensional n x ... x n matrix 1 if S(ii,i2 ,...,id) is a nonzero matrix and C = (ci,4 2. d) such that ci, 2 . d is a zero matrix. Cil2,...,id = 0 if S(ii, i2 ,... , i) We do casework based on whether an S-submatrix of A has s nonzero 1-cross sections for some 1. We first count the number of ones from the S-submatrices that do not have s nonzero i-cross sections for any 1. The contraction matrix C has at most m(n, Rk,--) d) ones, otherwise C would contain Rk---,k as an interval minor, a contradiction. Hence, A has at most m(n, Rk-.k, d) such S-submatrices, each of which contains at most (s - 1)d < sd ones. Thus there are at most sdm(n, R k..-'k d) ones from the Ssubmatrices of this type. We next count the number of ones from the S-submatrices that have s nonzero i-cross sections for some l. Without loss of generality let 1 = 1. Let A' be the matrix obtained from A by changing all the ones from the S-submatrices of A that do not have s nonzero i-cross sections to zeroes. Divide A' into n blocks, each of which is a t x tn x ... x In submatrix of A'. The Ith block of A' is defined to contain the entries 1,..., n. For each block, of A' such that t(i - 1) + 1 < j< K ti for each i a' contract every t consecutive j-cross sections uniformly for every j # 1 to obtain a t x n x - x n matrix, which has at most fk,....k(n, t, s, d) 1-rows that contain at least s ones. In each block there are at most fk,...,k(n, t, s,d) nonzero S-submatrices. Ranging over all n blocks as 1 ranges from 1 to d, there are at most dntdfk,..k.,(n, t, s, d) ones from the S-submatrices of this type. l Summing both cases proves the lemma. Lemma 77. fAk..,kd(n, 2t, 2s, d) < 2 fk 1 1...,kd(n, t, 2s, d) + 2 fk. 51 -,2,...,kd 1 (n, t, s, d). Proof. Let A be a d-dimensional 2t x n x - x n matrix that avoids Rk, ,kd as an interval minor and has fki,...,kd(n, 2t, 2s, d) 1-rows, each with at least 2s ones. The first type of 1-rows have all their ones among their first t or last t entries. There are clearly at most 2 fk 1 ,. kd(n, t, 2s, d) such 1-rows in A. The other type of 1-rows must have at least one 1 among both the first t and the last t entries. Since each 1-row has at least 2s ones, there are at least s ones among either the first or last t entries. Without loss of generality, we consider those 1-rows in which the first t entries contain at least s ones. Let A' be the matrix obtained from A by changing all ones to zeroes in all other 1-rows, and then contracting the last t 1-cross sections. Hence, the last entry in each nonzero 1-row of A' is a one. The first t 1-cross sections of A' must avoid Rkl-l,k2,...,kd as an interval minor, otherwise A' contains Rk,..-,kd as an interval minor and so does A, a contradiction. Thus, there are at most 2 fk 1 -1,k 2 ,...,kd(n, t, s, d) 1-rows in which both the first t and last t entries include at least one 1. Summing both cases proves the lemma. l Lemma 78. If s, t are powers of 2 and 2'-1 < s K t, then fe,k,...,k(n, t, s, d) < 2f 1t 2 m,(n, Rk,.d 1). Proof. We induct on f. For e = 1, we prove that fl,k,...,k(n, t, s, d) m(n, Rk,. d-1). Suppose for contradiction that f1,,...,k(n, t, s, d) > m(n, Rk, -. "', d - 1). Then there is a t x n x ... x n matrix A which avoids Rl',...,k as an interval minor and has more than m(n, R k,.'k, d - 1) 1-rows with at least s ones. Let B be the 1 x n x ... x n matrix obtained from A by contracting all the 1-cross sections. Then B, which can be viewed as a (d - 1)-dimensional matrix, has over it(n, R ak d -1) ones and thus contains the (d - 1)-dimensional matrix R ,.k as an interval minor. Consequently, A contains the d-dimensional R,-.-,k as an interval minor, a contradiction. Therefore, f1,k,..,k(n, t, s, d) < m(n, R-.', d- 1) < 21 -1t2 d - 1), m(n, R.', which proves the base case. We assume for all s and t that are powers of 2 satisfying 2e-2 < s < t that 2t 2 fe_1,,.(n, t, s, d) < s r(n, m Rk'..'k d - 1) (5.3) for some f > 2. We need to show that fe,k,...,k(n, t, s, d) < t m(n, Rk'.-.k' d- 1) 2 (5.4) for all s and t that are powers of 2 satisfying 2" < s < t. The case t = s > 2 -1 is trivial. If fe,k,...,k(n, t, s, d) < 2 m(n, Rk.,k, d - 1) for some t > s > 2 e-1 that is a 52 power of 2, then we prove the same inequality for 2t. By Lemma 77, ff,k,...,k(n, 2t, s, d) < < 2fe,k,...,k (n, t, s, d) + 2fe_1,k,...,k (n, t, s/2, d) 2e-1t 2 2e-2t22 m(n, Rk'--.k + 2 , + s/2 s = " S , d - 1) m(n, Rk'--'k d- 1) where we use both inductive assumptions in the second inequality. Proof of Theorem 75. We first bound the right hand side of inequality (5.1). claim for each fixed d > 2 that n- 20(k) El We (5.5) Fox proved the base case d = 2 [15]. Assuming that (5.5) is true for d - 1, we combine Lemmas 76 and 78 to get m(tn, Rk-.--k, d) < sdm(n, Rk.--k, d) + dtd 2k-1t2 2O(k)nd-1. S Choosing t = and s = 2 dk 2 k-1 yields m(2dkn, Rk,-.,k d ) < 2(k- 1 )dm(n, Rk.', In particular, if n is a positive integer power of d) + d 2 kd(d+2) 2 (k)nd-1. 2 dk, iterating this inequality yields Rk.-k, d) + d2 kd(d+2) 2 0(k)( 2 dk)(L-1)(d-1) < m((2dk)L kd(d+2) 2 0(k) (1 + 2d(dk2 +1) ( 2 dk)(L-1)(d-1) 2 2(k-1)dm(( 2 dk)L-2, Rk,.-k, d) + d 2 dk)(L-1)(d-1) +1) + 2 2tdkk+1) +--) d 2 (1 2 L(k-1)dm(1, Rk'.k, d)+d2kd(d+2)20(k) k,...,k - 2 d) < 2 (k-1)d ((2dk)L-1, 0(k)( 2 dk)(L-1)(d-1)* Suppose that ( 2 kd)L-1 < n < ( 2 kd) L. Then m(n, Rk.k, d) < m(( 2 dk)L k,...,k d) 20(k)nd- 1 . This completes the induction on d, and hence (5.5) l is proved. 2 0(k)( 2 dk)(L-1)(d-1) Next we extend Fox's probabilistic lower bound for the d = 2 case [15]. A key part of our approach is the observation that there exist d-dimensional permutation matrices P for which f(n, P, d) is super-homogeneous. Theorem 79. There exists a family of d-dimensionalk x ... x k permutation matrices P such that f (nPd) 2 "(k1/d) for n > 2 [kI/dj/20 In two dimensions, the extremal function was shown to be super-additive [39], i.e., f(m + n, P, 2) > f(m, P, 2) + f (n, P, 2). This implied the convergence of the sequence { f ,2)} for matrices P whose extremal functions are O(n). The limit is the well-known Firedi-Hajnal limit. The super-additivity of f(n, P, 2) also implies its super-homogeneity: f(sn, P, 2) > sf(n, P, 2) for every positive integer s. Define a corner entry of a ki x . x kd matrix 53 P =(...,id) to be an entry pi2,..., j located at a corner of P, i.e., e1 = 1 or i, = k, for every 1 < -r d. Lemma 80. If P is a d-dimensional matrix with a corner 1-entry, then f (sn, P, d) > S d_1 ,f~ P, d). - Proof. Without loss of generality, suppose that pi,...,1 = 1 is a corner 1-entry in P. Let M be a s x ... x s matrix with ones at the coordinates (ii,... , i ) where I1 + +'d i= s +d - 1 and zeroes everywhere else, so that Al has (d-2) >2 ones. Let N be an n x - x n matrix that avoids P and has f(n, P, d) ones. It suffices to prove that M 0 N avoids P. Suppose for contradiction that M 0 N contains P. Pick an arbitrary 1-entry p* in P other than Pi,...,1. Suppose that pi,..1 and p* are represented by el and e 2 in M 0 N, respectively. In particular, suppose that el and e 2 are in the S-submatrices S(i 1 ,. ..,i) and S(ji,.. . ,jd), respectively. Note that i1 -I- -+ .= + jd. Since each coordinate of p* is at least as great as each coordinate of Pi,...,1 in P, then each coordinate of e 2 must also be at least as great as each coordinate of ei in AM DN, and hence i4 j, for =1, 2, . . . , d. It then follows from i 1+ - + id= jI + .+ jd that ir = jr for r = 1, 2,. . . , d, i.e., the two entries ei and e 2 must be in the same S-submatrix in M 0 N. Since p* can be any arbitrary 1-entry other than pi,...,1 in P, a single S-submatrix contains all of P. However, this is a contradiction since each nonzero S-submatrix in M 0 N is a copy of N, which avoids P. Thus M 0 N avoids P. El Lemma 81. If P is a d-dimensional matrix with a corner 1-entry, then for any positive integers m and n > m, f (n, P, d) -1 f(m, P, d) 1 2d-1(d-1)! rd-1 Proof. Write n as n = sm + r, with 0 < r < m. f(sm,P,d) > P) f(nPd) >3_ f(m,) > i f(n,Pd) h Lem 1 (smn+r),d1 - (d-1)! (sm~r)d - (d-1)! ((s+1),n)d-1 - Then 2dl(d-1)! mn-, f(n_,Pd) - f(sm+r,Pd) where used in the second inequality. > ma 80 is El The following lemma gives a lower bound on the right hand side of the inequality in Lemma 81 for a particular m. The proof is based on Fox's methods for the d = 2 case in [14]. ( Nd-1/ 2 ) Lemma 82. There exists an N x ...x N matrix A, with N = 2 "(C), that has ones and avoids R-.. as an interval minor. + Proof. We prove the lemma for E that are multiples of 20, the result can then be easily extended to all f. Let r = -, q = 2 , and N = 2'. Define a dyadic interval to be a set of consecutive integers of the form {(s - 1) 2 ' 1, ... , s2t} for nonnegative integers s and t, and a dyadic hyper-rectangle to be the Cartesian product of d dyadic intervals. We consider only dyadic intervals that are subsets of {1,..., N}. Note that there are exactly E_ 2 = 2N - 1 such dyadic 54 intervals and (2N - 1)d such dyadic hyper-rectangles. There are (r + 1)" dyadic hyper-rectangles containing each entry ai1 . ,,d since each ij is contained in exactly r + 1 dyadic intervals. Let R be a random collection of the dyadic hyper-rectangles, each included independently with probability q. Define A to be the N x ... x N d-dimensional 0 - 1 matrix such that aj, . . . 1 if and only if (ii, ... , is) is not contained in any dyadic hyper-rectangle of R. The expected number of ones in A is (1 - q)(r+1)d Nd = (e-q(-r+1)"d)- where we use (1 - q)/ = [(1 - q)1/q]q(r+1)d N d 2 -(2r/ Nd) = O(Nd-1/ 2 ), E(e- 1 ) in the second equality. Denote by X and Y the events that A contains and avoids Re. as an interval minor, respectively. We bound the probability P(X). If B is a set of dyadic intervals, let X(B) be the number of dyadic intervals on {1,... , N} that contain at least one interval in B as a subset. Then define h(x) to be the number of sets B containing f dyadic intervals such that X(B) = x. If A contains Re.-- as an interval minor, then there are intervals of consecutive integers, denoted by Wi, .. ., W i ,j, partitioning the set {1, 2,... , N} in the ith dimension such that each submatrix W1,x j, x ... x Wd,j, of A contains a one. Let Ij be the unique smallest-length dyadic interval which contains W, 3 . If there are xi dyadic intervals which contain at least one of the dyadic intervals in Bi = {Ii, .. . , Iq}, then and there are h(xi) possible sets B. Thus there are h(xi) ... h(Xd) choices xi > for B 1 ,..., Bd and x 1 ... Xd dyadic hyper-rectangles which contain at least one of the d dyadic hyper-rectangles of the form I1,j, x ... X Id,id. Since these X1 ... Xd dyadic hyper-rectangles contain 1-entries of A, none of them is in R. Hence, e P(X) < Z (1 - q)x1.".Xdh(xi) ... h(Xd). (5.6) To bound h(x), we must bound the number of sets U = {u1,..., ue} of dyadic intervals such that X(U) = x. Let ,v1,... , ve be the integers such that vi = X({u1) andvi = X({u, . . ., uj}) - X({ui,..., 1uj}), for i= 2, . .. , e. Since vi + + vf = x, there are at most (2jey) possible values for vi, . . . , ve. Given v 1 , . . ., ve, there are 2j1-1 choices for ui and at most 2'i+1X({ui, ... , u2 }) choices for uj+1 . Since the order of the elements in U does not matter, then h (x) < hx 1f!< )2v'- i - 2i+1X({u1, . . . , ui}) (x+f-l)2xxe e_ Substituting this into (5.6) yields d q)(x ..-. )9P P (X ) d X1x,- di 55 (xi+f-12x i The summand r( , . I, (1 - ) d xi+ -1 _1 ( q) ) 2Xmi i=1 is symmetric in its variables. The ratio r(xi + 1, c 2 , . .. 2(1 - q) 4e -qX2 .-.X(1 + I 1 , xd)/r(Xi, . .. (I +) 1 )< 2...Xd X1 + 1 X1 )yj < eqi- -- 2 -2od-I/2& 4e <12, 1/e for q > 0 and (1 + 1/1x)x1 < e for where we use (1 - q) / ,X) is equal to 1 > 1. We also note that ,)) rq... e-fd (2 2e-12ee)d K 2 20d- le/2( 2 3f-le (_-) where we use Stirling's inequality and r(i,.-- [(Y- fjd (1- q)d < )d < N-10d, 5 22-. We now use the symmetry of , Xd) to obtain 00 P(X) < r(i,...,xd) d () < < 2 N- i=O X1,...,Xd>t Now we estimate the conditional expectation E( IY), where E( |Y)P(Y) = E( ) - E(>X)P(X) > O(Nd- 1 / 2 ) - Nd2dN-10d JAI. Note that E (Nd-1/ 2 ) So E( JY) = E(Nd-1/ ). Thus, there exists an N x x N d-dimensional 0 - 1 matrix A that avoids Rl-,' as an interval minor and has e(Nd-1/ 2 ) ones. El 2 Proof of Theorem 79. Let e = [k1/d], and let P be any d-dimensional permutation matrix of size k x ... x k that contains Rtee as an interval minor and has at least one corner 1-entry. P contains R.., as an interval minor, so f(N, P, d) > m(N, R....-'', d) for N = 2 /20, which by Lemma 81 and Lemma 82 implies for n > N that f(n, P, d) nd-1 1 2d-1(d - - 1)! m(N, R'.,, d) - 2 (k/d) Nd-1 El 5.2 Sharp bounds on m(n, Rk,.,k) d) We proved that rn(nR.kld) 20(k) In this section, we prove matching lower bounds for n> 2(-1)/20 on "("d I'd) - Lemma 83. For all s, n> 1, m(sn, R'.k, d) > (d- Fm(n, Rk~.k-, d). 56 ... x s matrix with ones at the coordinates (i, ... 1 and zeroes everywhere else, so that M has ( , id) where d- d-2) ones. Let N be an n x - - - x n matrix that avoids Rhi-l,---k-i as an interval minor d) ones. It suffices to prove that M 0 N avoids Rh'--.k as an and has m(n, Rhi.1, interval minor. Suppose for contradiction that M 0 N contains R k as an interval minor. Let . such that r* and r 1 . have no coordinates trary 1-entry r* in R-..k, other than r,. and r* are represented by ei and e 2 in MON, respecin common. Suppose that Ti. tively. In particular, suppose that el and e 2 are in the S-submatrices S(ii, ... , id) and +jd. Since each coordinate - id = Ji+ . .. , j), respectively. Note that iI + S0, then of r* is greater than each coordinate of r 1 . in an interval minor copy of Rh., each coordinate of e 2 must also be greater than each coordinate of el in M 0 N, and hence i, < j for T = 1, 2, ... ,d. It then follows fromZ1 + -- + id = ji + + id for T =1, 2,... , d, i.e., the two entries ei and e 2 must be in the same that iT = S-submatrix in M 0 N. Since r* can be any arbitrary 1-entry in Rh'..hk with no coordinates in common with ri., a single S-submatrix contains R-li---k-1 as an interval minor. However, this is a contradiction since each nonzero S-submatrix in M 0 N is a copy of N, which avoids Rh-1,..-'k-1 as an interval minor. Thus M 0 N avoids Rhk-.k as an interval minor. Lemma 84. For every t > 1 and n > t, "n'f.k"d) > __ Proof. Suppose that n sd-1 m(t,R k-l.-k-1,d) 1 (st+r)d(d-i)! > - = st + r, with 0 < r < t. Then " __ sd-1 n(t,Rk-l,...,k-1,d) ((s+1)t)d-1 (d-i)! > i - 2d-1(d-1)! Corollary 85. For n > 2(k-i)/20 mnR.d Proof. Let N - 2 (k-)/20. 2 27- '(d-i1)! - Then for n > N, m(tR-l..- k1, t- ) arbian pick and minimal, coordinate every with Rh.-...' of 1-entry corner the be Proof. Let M be a s x = s + d21 + - - +d > m(stR ',Rkkd) n m(t,Rk-1 ,-k-1,d) td-1 k1 (st+r)d ,d) > -I by Lemma 83. 2 mnRk.>d nI - 1 2d-(d-)! m(N,Rk-1--Nd-' ,-1d) El P(h) by Lemma 82 and Lemma 84. 2 Upper bounds for d-dimensional double permutation matrices 5.3 In this section, we bound the extremal functions of double permutation matrices. Suppose that P is a permutation matrix. A matrix P D Rhi .--. kd is called a j-tuple permutation matrix generated by P if one of k, ... , kd is equal to j and the rest are equal to 1. Let F(n, j, k, d) = maxm f(n, Al, d), such that the maximum ranges over all ddimensional j-tuple permutations matrices M that are generated by d-dimensional k x ... x k permutation matrices. 57 Theorem 86. For all fixed d, j > 2, F(n, j, k, d) = 2 We first reduce the problem to the case j n(k)rd. = 2 by applying the following lemma. Lemma 87. [28] F(n, 2, k, d) < F(n, j, k, d) < (j - 1)F(n, 2, k, d) for j > 2. By Lemma 87, it suffices to prove an upper bound on f(n, P, d), where P is any d-dimensional double permutation matrix of size 2k x k x ...x k. Suppose that A is an arbitrary d-dimensional kn x ... x kn matrix that avoids P. The S-submatrices of A are constructed by dividing A into nd disjoint submatrices of size k x ... x k and labeling these submatrices as S(ii,. . . , is). The submatrix S(ii, . .. , id) is defined to contain the entries aJl,.,d of A such that k(ie - 1) + 1 < je kIf for each e =1.. d. The contraction matrix of A is defined to be the d-dimensional n x ... x n matrix Co=.(ci2....id) such that cih..,id 1 if S(ii,z 2 ,... ,i) is a nonzero matrix and C.lh .... i = 0 if S(ii, i2, ... , id) is a zero matrix. We construct a d-dimensional 'n x - - - x n zero-one matrix Q=(q..d). Each entry q, . d is defined based on the S-submatrices of A. . d, = 0 if S(ii, ... , id) is a zero matrix. ,..,id = 1 if S(ii, i 2 , . ..., i) is a nonzero matrix and S(1, i2 , 1, i 2 , ... , id) are all zero matrices. 3. Let x be the largest integer less than i1 for which , id), . . . S(ii . . ., - 1. qji 1. Then define qxi.,'=d = 1 if the matrix consisting of the submatrices S(x, i 2 , . . . , Zi), ,... S(ii, i 2 , . . . , i) contains at least two 1-entries in the same 1-row, and qj, . otherwise. q., - Q avoids P. Proof. This follows from the definition of Q. El For each 1-entry qsii2,.., = 1 of Q, we define the chunk C*(ii, i2 , ... is a collection of consecutive S-submatrices. , id), which 1. If qa1,. . gd = 1 and it is the smallest integer greater than i 1 such that qi,. . '1 1, then the chunk C*(ii,i 2, .. . ,i) consists of the submatrices S(ii, i2 ,. ..,), 17,12,* .. , iid). 2. If qlh,...,d = *(', '2, ... 1 and if there is no i" > i1 such that qij , id) consists of the submatrices S(ii, i 2 ,. . .. . . . ,isz), . . . , =. S(r, i 2 , 1, then .. . ..,S(i4 - - Lemma 88. =0 We call a chunk wide if the chunk has at least two ones in the same 1-row. Every wide chunk has a single non-empty S-submatrix. We call a chunk j-tall, where j = 2, 3,.. ., d, if each of its j-cross sections contains a one. The (d - 1)-dimensional matrix A' = (m 1. . . . . . . . . id) is called the j-remainder of the d-dimensional matrix M = (mi,. , id) under the condition that M/= 1 if and only if there exists i1 such that , 1. 58 Lemma 89. The number of wide chunks of A is at most F(n, 1, k, d). Proof. If A had more than F(n, 1, k, d) wide chunks, then the entries in those wide F1 chunks would form a copy of P. This would be a contradiction. Lemma 90. For eachj 2,3,.. ., d and each m = 1, ... ,n, A has at most F(n, 1+ kd- 2 , k, d - 1) total j-tall chunks C*(ii,. . . , ij-1, m, i+ 1 ,. - ,). ) ) , , + Proof. Suppose for contradiction that A has r chunks C*,C*,..., C,*, where r > F(n, 1 + kd-2, k, d - 1), of the form C*(ii, ... , jj 1 , m, ij+1 , - . , id) that have ones in all of their j-cross sections. Let S1, S2,..., , be the initial S-submatrices of the , C, respectively. Let A' be the matrix formed by changing all chunks C*, C2*,ones in A that do not lie in S1,..., S, to zeroes. Let C be the contraction matrix of the j-remainder of A'. Then C is a (d - 1)-dimensional n x ... x n matrix and it has r ones, so it contains every (1 + kd- 2 )-tuple (d - 1)-dimensional permutation matrix of dimensions (1 + kd- 2 )k x k x ... x k. Since P is a d-dimensional double permutation matrix of size 2k x k x ... x k and j # 1, the j-remainder of P is a (d - 1)-dimensional double permutation matrix of size 2k x k x ... x k. Let P' be the (d - 1)-dimensional (1 + kd- 2 )-tuple permutation matrix of size (1 + kd- 2 )k x k x - -- x k such that P' and the j-remainder of P are generated from the same (d - 1)-dimensional permutation matrix. For each pair of ones in a 1-row of P with coordinates (x1, x 2 , ... , Xd) and (x1 1, x2 , . . , Xd), P' has 1 + kd- 2 ones with coordinates (sfi, x 2 , ... - -, (ij + k 2 ,1x2 Xj1, +1,.,d), Xj_1, Xj+1, . , Xe), (-1 + 1, x 2 ,.... , . . . , x- _1 , xj+1 , - - - , Xd) in a single 1-row. Since C contains P', each set of (1+ k 2 ones in P' is represented by ones with coordinates (t,(A), t 2 ,..., t._1, 2 tj +,..., td), where A = 1, 2,...,1 + kd- , in the same 1-row of C. Let S(t 1 (A), t 2 ,..,t_1,M, tj+1, . .. td), for each 1 < A < 1 + kd-2, be the corresponding S-submatrices of A. By the construction, these S-submatrices are the initial S-submatrices of a subset D of the chunks C1, . .. , C,*. Each of the (1 + kd- 2 chunks in D has ones in every j-cross section; in particular each chunk in D has a one with the same Jth coordinate (M - 1)k + xj. Among all of the chunks in D, there are at least 1 + kd- 2 ones with jth coordinate (m - 1)k + xj. Moreover there are kd-2 < k, there Kj 1-rows in each j-cross section of the chunks in D. Thus for each 1 < 1-row and have Jth are in the same in D that the chunks exists a pair of ones among coordinates equal to (m - 1)k + xj, by the pigeonhole principle. Hence, for each pair of ones in the same 1-row of P, there is a corresponding pair E of ones in the same 1-row of A. Thus A contains P, a contradiction. Next we prove a recursive inequality for F(n, 2, k, d) that will imply the desired upper bounds. Lemma 91. Let d, s, n be positive integers with d > 2. Then F(kn, 2, k, d) < (d - 1)nkd-lF(n, 1 + kd-2, k, d - 1) + kdF(n, 1, k, d) + (k - 1)d-lF(n, 2, k, d). Proof. We count the maximum possible number of 1-entries in A by considering three cases. 59 Case 1: chunk has two ones in the same 1-row Such a chunk has only one nonempty S-submatrix, so it has at most kd ones. By Lemma 89, there are at most F(n, 1, k, d) wide chunks, so wide chunks contain at most kdF(n, 1, k, d) nonzero entries. + Case 2: chunk is j-tall and has no ones in the same 1-row There are (d - 1) choices for j since j 2, 3,..., d. For each j, the integer m of Lemma 90 can be 1,.. ., n. A j-tall chunk with no ones in the same row has at most k-1 ones. For each pair of j and m, there are at most F(n, 1 + kd-2, k, d - 1) j-tall chunks by Lemma 90. In total, j-tall chunks contain at most (d - 1)nk d-F(n, 1 kd-2, k, d - 1) ones. Case 3: chunk is not j-tall and has no two ones in the same 1-row Such a chunk has at most (k - I)d-1 ones. By the definition of a chunk, the number of chunks is equal to the number of ones in matrix Q. by Lemma 88, Q has at most F(n, 2, k, d) ones. There are at most (k - 1)diF(n, 2, k, d) ones in chunks of this type. Summing the maximum possible number of ones for all cases proves the lemma. Proof of Theorem 86. We induct on d > 2. The base case of d = 2 is true by Theorem 72. We assume for some d > 2 that F(n,j,k,d 1) = 2O(k)rid 2 for each fixed j 1, and we prove that F(n, j, k, d) - 2 0(k)nd-1 for each fixed j > 1. First we simplify the inequality in Lemma 91. The inductive assumption im- plies that F(n, 1 + kd 2 , k, d - 1) = 20 (k)nr- 2 . We also proved that F(n, 1, k, d) 20(k)nd-1. Hence, there is a sufficiently large constant c such that the sum of the last two terms on the right hand side of the inequality in Lemma 91 is bounded by 2cknd-1. Therefore, F(kn, 2, k, d) < (k - 1)d F(n, 2, k, d) + 2 "knd-1 for all n. We use a strong induction on n to prove that F(n, 2, k, d) < k(2ck+dk)n d for all n. The base case of n < k is trivial. If F(n, 2, k, d) < k(2 ck + dk)nd-1 for all n < m, then we show that F("1, 2, k, d) < k(2k + dk)md-1. Let N be the maximum integer that is less than m and divisible by k. A ddimensional m x ... x m matrix has at most md - Nd < Md - (m - k)d < dkmdmore entries than an N x N x ... x N d-dimensional matrix. Thus F(m, 2, k, d) < F(N, 2, k, d) + dkn 1 . Since F(kn, 2, k, d) < (k 1)d-lF(n, 2, k, d) + 2cknd-1 for all n, then F(m, 2, k, d) < (k - 1)d- F (, 2,k,d +2 ck k (k (k - 1)d-1k( 2ck + dk) (N)k < < (k - 1)( 2ck + dk) Nd- + ( 2 k(2ck + dk)md-1 +2 k k - 1 + dkmd-1 + dk)md-1 , < + dkmd - where we used the strong inductive assumption in the second inequality. 60 L 5.4 Open Problems Since there has not been much research on d-dimensional 0--1 matrices outside of [34], bounds on f(n, P, d) are unknown for most d-dimensional 0 - 1 matrices P. Besides finding bounds on f(n, P, d) for specific d-dimensional 0 - 1 matrices P, there are a few more general questions about f(n, P, d) that are left unresolved by our work. Question 92. Does the sequence matrix P? {f ("i d} converge Question 93. Does the sequence {,,Pd) matrix P? for every d-dimensional 0 - 1 converge for every d-dimensional 0 - 1 Question 94. What is the maximum possible value of k x .. x k permutation matrices P? 61 f (?d) n for all d-dimensional 62 Bibliography [1] P.K. Agarwal and M. Sharir. Davenport-Schinzel sequences and their geometric applications. Cambridge University Press, Cambridge, 1995. [2] P.K. Agarwal, M. Sharir, and P. Shor. Sharp upper and lower bounds for the length of general Davenport-Schinzel sequences. Journal of CombinatorialThe- ory Series A, 52:228-274, 1989. 131 N. Alon, H. Kaplan, G. Nivasch, M. Sharir, and S. Smorodinsky. Weak epsilon- nets and interval chains. J. ACM, 55 (2008). [4] D. Bienstock and E. Gy6ri, An extremal problem on sparse 0-1 matrices, SIAM J. Discrete Math. 4 (1991), 17-27. [5] M. B6na, Combinatorics of Permutations, second edition, CRC Press-Chapman Hall, 2012. [6] J. Cibulka, On constants in the Ffiredi-Hajnal and the Stanley-Wilf conjecture, J. Combin. Theory Ser. A, 116 (2) (2009), 290-302. [71 J. Cibulka and J. Kyncl. Tight bounds on the maximum size of a set of permutations with bounded vc-dimension. Journal of Combinatorial Theory Series A, 119 (2012) 1461-1478. [8] H. Davenport and A. Schinzel. A combinatorial problem connected with differential equations. American Journal of Mathematics, 87:684-694, 1965. [91 A. M. Dean, W. Evans, E. Gethner, J. D. Laison, M. A. Safari, W. T. Trotter, Bar k-visibility graphs. J. Graph Algorithms and Applications, 11(1): 45-59 (2007). [10] P. Erd6s and L. Moser, Problem 11, Canadian Math. Bull. 2 (1959), 43. [111 P. Erdos and J. Spencer, Probabilistic Methods in Combinatorics, Academic Press, New York, 1974. [12] E. Ezra, B. Aronov, and M. Sharir. Improved bound for the union of fat triangles. Proceedings of the twenty-second ACM-SIAM Symposium on Discrete Algorithms, 1778-1785, 2011. 63 [13] M. Fekete, Uber die Verteilung der Wurzeln bei gewissen algebraischen Gleichungen mit. ganzzahligen Koeffizienten, Mathematische Zeitschrift 17 (1) (1923), 228-249. [14] J. Fox, Combinatorics of permutation, http://www.math.uni- frankfurt.de/dm2014/Fox.pdf, 2014 [15] J. Fox, Stanley-Wilf limits are typically exponential. CoRR abs/1310.8378 (2013) [16] J. Fox, J. Pach, and A. Suk. The number of edges in k-quasiplanar graphs. SIAM Journal of Discrete Mathematics, 27:550-561, 2013. [171 R. Fulek, Linear bound on extremal functions of some forbidden patterns in 0-1 matrices. Discrete Mathematics 309(6): 1736-1739 (2009). [18] Z. Furedi, The maximum number of unit distances in a convex n-gon, Journal of Combinatorial Theory Ser. A 55 (2), 316-320 (1990). [19] Z. Furedi, P. Hajnal, Davenport-Schinzel theory of matrices, Discrete Mathemat- ics, v.103 n.3, p. 2 3 3 - 2 5 1 (1992) [201 J.T. Geneson, Extremal functions of forbidden double permutation matrices, Journal of Combinatorial Theory Series A, v.116 Issue 7, 1235-1244 (2009) [211 J. Geneson, Improved bounds on maximum sets of letters in sequences with forbidden alternations. CoRR abs/1401.0063 (2014) [22] J. Geneson, R. Prasad, and J. Tidor. Bounding Sequence Extremal Functions with Formations. Electr. J. Comb. 21(3): P3.24 (2014) [23] J. Geneson and L. Shen, Linear bounds on matrix extremal functions using visibility hypergraphs. CoRR abs/1410.3147 (2014) [24] J. Geneson and P. Tian, Improved bounds on extremal functions of forbidden multidimensional 0 - 1 matrices. Preprint. (2014) [25] A. Hesterberg, preprint, 2009. Extremal functions of excluded block permutation matrices, [26] A. Hesterberg, Extremal functions of excluded tensor products of permutation matrices, Discrete Math., 312 (10) (2012), 1646-1649. [27] S. Heubach and T. Mansour, Combinatorics of Compositions and Words, CRC Press, 2009. [28] B. Keszegh, On linear forbidden submatrices, Journal of Combinatorial Theory Series A, v.116 n.1, p. 2 3 2 - 2 4 1 (2009) [29] S. Kitaev, Patterns in Permutations and Words, Springer-Verlag, 2011. 64 [30j M. Klazar. A general upper bound in the extremal theory of sequences. Commentationes Mathematicae Universitatis Carolinae, 33:737-746, 1992. [31J M. Klazar, Extremal problems for ordered (hyper) graphs: applications of Davenport-Schinzel sequences, European J. Combin. 25 (2004) 125-140. [321 M. Klazar. On the maximum lengths of Davenport-Schinzel sequences. Contemporary trends in discrete mathematics (DIMACS series in discrete mathematics and theoretical computer science), 49:169-178, 1999. [331 M. Klazar, The Fiiredi-Hajnal conjecture implies the Stanley-Wilf conjecture, in: D. Krob, A.A Mikhalev, A.V. Mikhalev (Eds), Formal Power Series and Algebraic Combinatorics, Springer, Berlin, 2000, 250-255. [34] M. Klazar, A. Marcus, Extensions of the linear bound in the Furedi-Hajnal con- jecture, Adv. in Appl. Math. 38, no. 2, 258-266 (2006) [351 T. K6vdri, V. T. S6s and P. Turin, On a problem of K. Zarankiewicz. Colloquium Math. 3 (1954), 50-57. [361 A. Marcus, G. Tardos, Excluded permutation matrices and the Stanley-Wilf conjecture, Journal of Combinatorial Theory Series A, v.107 n.1, p. 1 5 3 - 1 6 0 (2004) [371 J. Mitchell: Shortest rectilinear paths among obstacles, Department of Operations Research and Industrial Engineering Technical Report No. 739, Cornell University, Ithaca, New York (1987) [381 G. Nivasch. Improved bounds and new techniques for Davenport-Schinzel sequences and their generalizations. Journal of the A CM, 57: article 17, 2010. [391 J. Pach, G. Tardos, Forbidden paths and cycles in ordered graphs and matrices, Israel J. Math. 155, 309-334 (2006) [40J S. Pettie. Degrees of nonlinearity in forbidden 0-1 matrix problems. Discrete Mathematics 311 (2011) 2396-2410. [411 S. Pettie. On the structure and composition of forbidden sequences, with geometric applications. Proceedings of the twenty-seventh annual symposium on computational geometry, 370-379, 2011. [421 S. Pettie. Sharp bounds on Davenport-Schinzel sequences of every order. Symposium on Computational Geometry (2013): 319-328. [431 R. P. Stanley, Increasing and decreasing subsequences and their variants, in: Proceedings of the International Congress of Mathematicians, Plenary Lectures, vol. I, Madrid, Spain, 2006, 545-579. [441 E. Steingrimsson, Some open problems on permutation patterns, in: Surveys in combinatorics 2013, London Math. Soc. Lecture Note Ser. Cambridge University Press, 2013, 239-263. 65 [45] A. Suk and B. Walczak. New bounds on the maximum number of edges in kquasi-planar graphs. Twenty-first InternationalSymposium on Graph Drawing, 95-106, 2013. [46] R. Sundar. On the Deque conjecture for the splay algorithm. Combinatorica 12(1): 95-124 (1992). 66