Topological Combinatorics * * * * * *

Transcription

1 Topological Combinatorics * * * * * * Anders Björner Dept. of Mathematics Kungl. Tekniska Högskolan, Stockholm * * * * * * MacPherson 60 - Fest Princeton, Oct. 8, 2004

2 Influence of R. MacPherson on topological combinatorics: Intersection homology Convex polytopes (via toric varieties), toric g-vector Bruhat order (via Schubert varieties) Subspace arrangements Goresky-MacPherson formula Application to complexity Oriented matroids CD (combinatorial differential) manifolds MacPhersonians (discrete Grassmannians) And more... 1

3 Two topics for this talk: Goresky-MacPherson formula, with an application to complexity Bruhat order with an application of intersection cohomology 2

4 Connections Topology Combinatorics Simplest case: Space Triangulation Example: The real projective plane RP 2 c f a d b e c {abd, acf, adf, ace, abf, aef, bcd, bcf, cde, def} b a 3

5 Topic 1: Goresky-MacPherson formula for subspace arrangements A def = collection of affine subspaces of R d an arrangement M A def = R d \ A its complement L A def = family of nonempty intersections of members of A, ordered by reverse containment its intersection semi-lattice. 4

6 THM ( Goresky-MacPherson formula ): H i (M A ) = x L A, x>ˆ0 ********* Proof: Stratified Morse Theory (1988) Other proofs by several authors ********* Here (ˆ0, x) is the simplicial complex of such that H codim(x) 2 i ( (ˆ0, x)) {z 1, z 2,..., z k } ˆ0 < z 1 < z 2 < z k < x called the order complex of the open interval (ˆ0, x) in L A. 5

7 f g b c d h ˆ0 e A small poset Values of its Möbius function µ(ˆ0, x) µ(ˆ0, x) def = ˆ0 y<x µ(ˆ0, y) µ(ˆ0, x) = Euler char( (ˆ0, x)) 1 6

8 Special cases of G-M formula: Hyperplane arr ts over R Zaslavsky s formula for number of connected components of M A Hyperplane arr ts over C Brieskorn-Orlik-Solomon formula for cohomology groups of M A 7

9 Application of G-M formula to complexity of algorithms Given: a string of real numbers x 1, x 2,..., x n Sought: Efficient algorithms to decide some property of the sequence or to restructure it using only pairwise comparisons. The question: How many such comparisons must be made in the worst case when using the best algorithm? This number, c(n), is called the complexity of the problem. Note: c(n) n 2 is immediate 8

10 Well-known examples: 1. Sorting. To rearrange the n numbers increasingly x i1 x i2 x in requires Θ(nlog n) comparisons. 2. Median. To find j such that x j is in the middle requires Θ(n) comparisons, where 2n Θ(n) 3n. 3. Distinctness. To decide whether all entries x i are distinct (i.e., if x i x j when i j) requires Θ(nlog n) comparisons. 9

11 A generalization of the distinctness problem (the k = 2 case). The k-equal problem: for k 2, decide whether some k entries are equal, that is, can we find i 1 < i 2 < < i k such that x i1 = x i2 = = x ik? For example, are there nine equal entries in the following list of numbers?

12 Question repeated: are there nine equal entries in the following list of numbers? Answer: Yes, there are nine copies of the number 4. Are there ten equal entries? Answer: No. 11

13 THM (Bj-Lovász-Yao 92) The complexity of the k-equal problem is Θ(nlog 2n k ). More precisely, where C 1 nlog 2n k c k(n) C 2 nlog 2n k, C 2 C Upper bound: Sorting algorithms Lower bound: Topological method (involving G-M formula) 12

14 Before sketch of lower bound argument, need more tools Examples of interesting subspace arr ts in codimension k 1: A n,k def = {x i1 = = x ik 1 i 1 < < i k n} D n,k def = {ε 1 x i1 = = ε k x ik 1 i 1 < < i k n, ε i {±1}} B n,k def = D n,k {x j1 = = x jk 1 = 0 1 j 1 < < j k 1 n} Note: for k = 2 get Coxeter reflection arrangements 13

15 Computing cohomology of complement of A reduces (via G-M formula) to computing homology of order complex of L A. How compute homology of poset L A? combinatorial method that works surprisingly often: lexicographic shellability 14

16 P a poset with ˆ0 and ˆ1 E(P) = {(x, y) P P x y} its covering relation Def: An EL-labeling of P is a map λ : E(P) Z, such that for every interval [x, y]: 1. there is a unique maximal chain m [x,y] whose associated label λ(m [x,y] ) = (a 1,..., a p ) is increasing a 1 < a 2 < < a p, 2. if m is any other maximal chain in [x, y] then λ(m ) > λ(m [x,y] ) in the lexicographic order on strings with elements from Z. The poset P is said to be lexicographically shellable (or for short: EL-shellable) if it admits an EL-labeling. 15

17 EL-shellability, when applicable, reduces homology computations for posets to a combinatorial labeling game. Call a maximal chain ˆ0 = x 0 x 1 x k = x, falling if λ(x 0 x 1 ) λ(x 1 x 2 )... λ(x k 1 x k ). THM (Bj-Wachs 96) EL-shellable (ˆ0, x) has the homotopy type of a wedge of spheres, for x > ˆ0. Furthermore, for any fixed EL-labeling: H i ( (ˆ0, x); Z) = Z # falling chains of length (i + 2) a basis for i-dimensional (co)homology is induced by the falling chains of length i

18 with lexico- Combining Goresky-MacPherson formula for M A graphic shellability of L A we get: THM For arrangement A, suppose L A is EL-shellable. Then H i (M A ) is torsion-free, and the Betti number β i (M A ) is equal to the number of falling chains ˆ0 = x 0 x 1 x g such that codim(x g ) g = i. 17

19 THM EL-shellability works for (1) hyperplane arr ts (over any field) (2) A n,k and B n,k, (3) some other cases... Conjecture: Works for D n,k. Incidentally, THM (Khovanov 96) Complements of A n,3 and B n,3 are K(π,1) spaces. Conjecture: Complement of D n,3 is a K(π,1) space. 18

20 Example: EL-shellability-based computation for A n,k (Following 4 slides are based on joint work with M. Wachs 95 and V. Welker 95.) 19

21 Let Π n,k be family of all partitions of {1,2,..., n} that have no parts of sizes 2, 3,..., k 1. Order them by refinement. Fact: The intersection lattice of A n,k is (isomorphic to) Π n,k Π 4,2 Π 4,3 20

22 Labeling of L An,k = Πn,k with elements from the totally ordered set 1 < 2 < < n < 1 < 2 < < n Covering Label New k-block B created from singletons max(b) Non-singleton block B merged with singleton {a} a Two non-singleton blocks B 1 and B 2 merged max(b 1 B 2 ) For instance, the following maximal chain in Π 8,3 (only nonsingleton blocks are shown) ˆ ˆ1 receives the label (8, 4, 7, 8, 6). This is an EL-labeling of Π n,k. 21

23 Hence, * n,k = (ˆ0,ˆ1) has homotopy type of a wedge of spheres * the Betti numbers β d n,k = rank H d ( n,k ; Z) satisfy and β d n,k 0 iff d = n 3 t(k 2) for some 1 t n k β n 3 t(k 2) n,k = (t 1)! = j 1 + +j t =n j i k 0=i 0 i t =n tk t 1 j=0 ( n 1 j 1 1, j 2,..., j t, ( n jk i j 1) (j + 1) i j+1 i j k 1 ) t i=1 ( j i 1 k 1 ). 22

24 Also, * the Betti numbers of the complement M n,k of A n,k : β i (M n,k ) 0 i = t(k 2), for 0 t n k. 23

25 Back to algorithmic k-equal problem: geometric point of view (following Bj-Lovász 94) The k-equal problem is to determine whether a given point x = (x 1, x 2,..., x n ) R n lies in the union of all the subspaces x i1 = x i2 = = x ik. Equivalently, does it belong to M n,k, the complement of the k- equal arrangement A n,k. 24

26 Let β(m n,k ) = n i=0 rank H i (M n,k ). Fact 1. The complexity of the k-equal problem is at least log 3 β(m n,k ). Note: Recall computation of β(m n,k ) via GM+EL method, messy sums/products of binomial coeff s... 25

27 Let µ n,k be the Möbius function computed over the poset Π n,k. Fact 2. β(m n,k ) µ n,k. (Again based on G-M formula) We turn to generating functions and prove: exp n 1 x n µ n,k = 1 + x + x2 n! 2! + + xk 1 (k 1)!. This implies Fact 3. Let α 1, α 2,..., α k 1 be the complex roots of the polynomial 1 + x + x2 2! + + xk 1 (k 1)!. Then µ n,k = (n 1)! ( α n 1 + α n α n k 1 ). 26

28 Collecting the facts, we have: c k (n) log 3 β(m n,k ) log 3 µ n,k Now either estimate β(m n,k ) via expressions given by EL-shelling, or estimate µ n,k via Fact 3. This gives: c k (n) C nlog 2n k Q.E.D. 27

29 Remark : Goresky-MacPherson formula over finite fields l-adic étale cohomology H i (X; Q l ) versions of G-M: Bj-Ekedahl 97, Yan 00, Deligne-Goresky-MacPherson 00,... Briefly: Let A be a d-dim l subspace arr t in F n q, q = pr. Let α i,j be the eigenvalues of Frobenius acting on Hc(M i A ; Q l ), and P i (t) def = j(1 α i,j t). Then, P i (t) = d j=0 (1 q j t) β j i 2j 2, where β j def i = sum over x L A such that dim(x) = j of i-th Betti number of (ˆ0, x) (i.e., order complex homology). 28

30 Suggestion: Étale cohomology could be a secret tool for complexity theory. Observation: Boolean function f : {0,1} n {0,1} is simply a subset of affine n-space over GF(2). Program: 1. Find good description of some NP-complete f as a variety, 2. Compute the étale Betti numbers of f, 3. Show that big Betti numbers force big Boolean circuits, 4. Conclude that NP P. 29

31 Topic 2: Bruhat order Figures on following slides are mostly taken from the book Combinatorics of Coxeter Groups by Björner Brenti, Springer Help with figures by F. Incitti and F. Lutz is gratefully acknowledged. 30

32 The pair (W, S) is a Coxeter group (Coxeter system) if W is a group with presentation { Generators: S Relations: (ss ) m(s,s ) = e, (s, s ) S 2, where m : S S {1,2,..., } satisfies m(s, s ) = m(s, s); m(s, s ) = 1 s = s. In particular, and s 2 = e, for all s S, s s s s s... }{{} m(s,s ) = s s s s s... }{{} m(s,s ) classification of finite (affine, hyperbolic) Coxeter groups: type A n, B n,...etc. 31

33 Bruhat order: For u, w W: u w def for reduced expressionw = s 1 s 2... s q a reduced subexpressionu = s i1 s i2... s ik, 1 i 1 <... < i k q. abab = baba aba bab ab ba a b e Bruhat order of B 2 32

34 Bruhat order of B 3. 33

35 analogy Intervals [e, w] in Bruhat order Face lattices of convex polytopes Weyl group rational polytope Schubert variety toric variety Kazhdan-Lusztig polynomial g-polynomial Also: Both determine regular CW decompositions of a sphere Intersection cohomology lurks in the background Remark: For all polytopes: combinatorial intersection cohomology theory satisfying hard Lefschetz (recent work of K. Karu and others) Question:??? combinatorial intersection cohomology theory for all Coxeter groups ( virtual Schubert varieties )? 34

36 THM (Bj-Wachs 82, Bj 84) Let [u, w] be a Bruhat interval. Then regular CW decomposition Γ u,w of the (l(w) l(u) 2)- dimensional sphere with cells σ x, u < x < w, such that and dim(σ x ) = l(x) l(u) 1 σ x σ z x z. Proof idea: via lexicographic shellability of Bruhat order 35

37 Example: Lex. shelling of W = S 3, generators S = {a, b} aba = bab ab a e ba b Choosing aba as reduced expression for the top element the induced labels of the four maximal chains are λ(aba ba a ) = (1,2,3), λ(aba ba b ) = (1,3,2), λ(aba ab b ) = (3,1,2), λ(aba ab a ) = (3,2,1). 36

38 A B C a B α β γ δ α A γ β a b c b C δ c Regular CW interpretation of a Bruhat interval. 37

39 y 1 y 2 y 3 y k x 1 x 2 x 3 x k A k-crown. All Bruhat intervals of length 3 are k-crowns, k 2. Finite case only k = 2,3,4 possible. (And for type H also k = 5.) 38

40 THM (Dyer 91). For each m, there exist only finitely many isomorphism classes of length m intervals in finite Coxeter groups. THM (Hultman 03). There are 24 types of length 4 intervals in finite Weyl groups. Only 7 of them occur in the symmetric groups. All 24 show up in F 4. 39

41 A Bruhat interval of length 4 (rendered as a CW complex) 40

42 All length 4 intervals that appear in finite Weyl groups. 41

43 Shape of lower interval [e, w]: m def f w i f w = {f w 0, fw 1,..., fw m }, = l(w) def = number of elements x w of length i. Note: Analogy with f-vector of convex polytope 42

44 Known for f-vector of simplicial (d + 1)-dimensional polytope: (1) f i f j if i < j d i. In particular, f 0 f 1 f d/2 and f i f d i (2) f 3d/4 f (3d/4) 1 f d (3) The bounds d/2 and 3d/4 are best possible. Conjecture: (2) is true for all polytopes. 43

45 Does it make sense to ask such questions for f w -vectors of Bruhat intervals [u, w]? Perhaps... consider this: THM (Carrell-Peterson 94) A Shubert variety X w is rationally smooth fi w = fm i w for all i 44

46 THM (Bj-Ekedahl 04) For the f w -vector f w = {f 0, f 1,..., f m } of interval [e, w] in a (Kac-Moody) Weyl group: (1) f i f j if i < j m i. In particular, f 0 f 1 f m/2 and f i f m i (2) If finite then also f f +1 f m (*= to be explained) Conjecture: This is true for all Coxeter groups. 45

47 f w -vector of Bruhat interval [e, w] 46

48 Idea of proof of (1): For X = X w, H (X, Q l ) IH (X, Q l ) is an H (X, Q l )-module map for i j m i it commutes with multiplication by c 1 (L) j i commutative diagram H 2i (X, Q l ) IH 2i (X, Q l ) c 1 (L) j i c 1 (L) j i H 2j (X, Q l ) IH 2j (X, Q l ). The horisontal maps are injective and the right vertical map is an injection by hard Lefschetz. Hence the left vertical map is injective, giving f w i = dim Ql H 2i (X, Q l ) dim Ql H 2j (X, Q l ) = f w j. 47

49 For monotonicity at upper end (part (2)), all we can prove is For k > 1 N k such that for finite Weyl group and w W such that m = l(w) N k : f m k f m k+1 f m Question: Does there exist α < 1 such that for all w in all Coxeter groups? f αm f αm +1 f m Conjecture: Yes, and α = 3/4 will work 48

50 Let Invol(W) def = involutions of W with induced Bruhat order. Studied by Richardson-Springer 94, Incitti 03, Hultman 04. Has wonderful properties as poset, much as W itself: pure, regular CW spheres for the classical Weyl groups (via ELshellability), intervals= homology spheres in general,... Poset rank function: rk(w) = l(w)+al(w) 2, where al(w) is absolute length 49

51 Involutions in S 4 Invol(S 4 ) 50

52 Happy Birthday, Bob! 51