Increasing and Decreasing Subsequences

Transcription

1 Increasing and Decreasing Subsequences Richard P. Stanley M.I.T.

2

3

4

5

6 Permutations

7 Permutations First lecture: increasing and decreasing subsequences

8 Permutations First lecture: increasing and decreasing subsequences Second lecture: alternating permutations

9 Permutations First lecture: increasing and decreasing subsequences Second lecture: alternating permutations Third lecture: reduced decompositions

10 Definitions (increasing subsequence) (decreasing subsequence) is(w) = longest i.s. = 4 ds(w) = longest d.s. = 3

11 Application: airplane boarding Naive model: passengers board in order w = a 1 a 2 a n for seats 1, 2,..., n. Each passenger takes one time unit to be seated after arriving at his seat.

12 Boarding process

13 Boarding process

14 Boarding process

15 Boarding process

16 Results Easy: Total waiting time = is(w). Bachmat, et al.: more sophisticated model.

17 Results Easy: Total waiting time = is(w). Bachmat, et al.: more sophisticated model. Two conclusions: Usual system (back-to-front) not much better than random. Better: first board window seats, then center, then aisle.

18 Partitions partition λ n: λ = (λ 1, λ 2,... ) λ 1 λ 2 0 λi = n

19 Young diagrams (Young) diagram of λ = (4, 4, 3, 1):

20 Conjugate partitions λ = (4, 3, 3, 2), the conjugate partition to λ = (4, 4, 3, 2) λ λ

21 Standard Young tableau standard Young tableau (SYT) of shape λ n, e.g., λ = (4, 4, 3, 1): < <

22 f λ f λ = # of SYT of shape λ E.g., f (3,2) = 5:

23 f λ f λ = # of SYT of shape λ E.g., f (3,2) = 5: simple formula for f λ (Frame-Robinson-Thrall hook-length formula)

24 f λ f λ = # of SYT of shape λ E.g., f (3,2) = 5: simple formula for f λ (Frame-Robinson-Thrall hook-length formula) Note. f λ = dim(irrep. of S n ), where S n is the symmetric group of all permutations of 1, 2..., n.

25 RSK algorithm RSK algorithm: a bijection w rsk (P, Q), where w S n and P, Q are SYT of the same shape λ n. Write λ = sh(w), the shape of w.

26 RSK algorithm RSK algorithm: a bijection w rsk (P, Q), where w S n and P, Q are SYT of the same shape λ n. Write λ = sh(w), the shape of w. R = Gilbert de Beauregard Robinson S = Craige Schensted (= Ea Ea) K = Donald Ervin Knuth

27 Example of RSK w = 4132: φ

28 Example of RSK w = 4132: (P, Q) = ( , ) φ

29 Schensted s theorem Theorem. Let w rsk (P, Q), where sh(p ) = sh(q) = λ. Then is(w) = longest row length = λ 1 ds(w) = longest column length = λ 1.

30 Schensted s theorem Theorem. Let w rsk (P, Q), where sh(p ) = sh(q) = λ. Then is(w) = longest row length = λ 1 ds(w) = longest column length = λ 1. ( ) 1 2 Example rsk 1 3 3, is(w) = 2, ds(w) = 3.

31 Erdős-Szekeres theorem Corollary (Erdős-Szekeres, Seidenberg). Let w S pq+1. Then either is(w) > p or ds(w) > q.

32 Erdős-Szekeres theorem Corollary (Erdős-Szekeres, Seidenberg). Let w S pq+1. Then either is(w) > p or ds(w) > q. Proof. Let λ = sh(w). If is(w) p and ds(w) q then λ 1 p and λ 1 q, so λ i pq.

33 An extremal case Corollary. Say p q. Then #{w S pq : is(w) = p, ds(w) = q} = ( ) 2 f (pq )

34 An extremal case Corollary. Say p q. Then #{w S pq : is(w) = p, ds(w) = q} = ( ) 2 f (pq ) By hook-length formula, this is ( ) 2 (pq)! p p (p + 1) p q p (q + 1) p 1 (p + q 1) 1

35 Romik s theorem Romik: let w S n 2, is(w) = ds(w) = n. Let P w be the permutation matrix of w with corners (±1, ±1). Then (informally) as n almost surely the 1 s in P w will become dense in the region bounded by the curve (x 2 y 2 ) 2 + 2(x 2 + y 2 ) = 3, and will remain isolated outside this region.

36 An example w = 9, 11, 6, 14, 2, 10, 1, 5, 13, 3, 16, 8, 15, 4, 12, 17

37 (x 2 y 2 ) 2 + 2(x 2 + y 2 ) = 3 1 y x 0.5 1

38 Area enclosed by curve α = (1 t2 )(1 (t/3) 2 ) dt (t/3) 2 1 t 2 dt = 4( )

39 Expectation of is(w) E(n) = expectation of is(w), w S n = 1 n! = 1 n! is(w) w S n ( λ ) 1 f λ 2 λ n

40 Expectation of is(w) E(n) = expectation of is(w), w S n = 1 n! = 1 n! w S n is(w) λ n λ 1 ( f λ ) 2 Ulam: what is distribution of is(w)? rate of growth of E(n)?

41 Work of Hammersley Hammersley (1972): c = lim n n 1/2 E(n), and π 2 c e.

42 Work of Hammersley Hammersley (1972): c = lim n n 1/2 E(n), and Conjectured c = 2. π 2 c e.

43 c = 2 Logan-Shepp, Vershik-Kerov (1977): c = 2

44 c = 2 Logan-Shepp, Vershik-Kerov (1977): c = 2 Idea of proof. E(n) = 1 n! ( λ ) 1 f λ 2 λ n 1 n! max λ n λ 1 ( f λ ) 2. Find limiting shape of λ n maximizing λ as n using hook-length formula.

45 A big partition

46 The limiting curve

47 Equation of limiting curve x = y + 2 cos θ y = 2 (sin θ θ cos θ) π 0 θ π

48 is(w) 2 u k (n) := #{w S n : is n (w) k}.

49 is(w) 2 u k (n) := #{w S n : is n (w) k}. J. M. Hammersley (1972): a Catalan number. u 2 (n) = C n = 1 n + 1 ( 2n n ),

50 is(w) 2 u k (n) := #{w S n : is n (w) k}. J. M. Hammersley (1972): a Catalan number. u 2 (n) = C n = 1 n + 1 ( 2n For >170 combinatorial interpretations of C n, see n ), www-math.mit.edu/ rstan/ec

51 Gessel s theorem I. Gessel (1990): n 0 u k (n) x2n n! 2 = det [ I i j (2x) ] k i,j=1, where I m (2x) = j 0 x m+2j j!(m + j)!, a hyperbolic Bessel function of the first kind of order m.

52 The case k = 2 Example. n 0 u 2 (n) x2n n! 2 = I 0 (2x) 2 I 1 (2x) 2 = n 0 C n x 2n n! 2.

53 Painlevé II equation Baik-Deift-Johansson: Define u(x) by d 2 dx 2u(x) = 2u(x)3 + xu(x) ( ), with certain initial conditions.

54 Painlevé II equation Baik-Deift-Johansson: Define u(x) by d 2 dx 2u(x) = 2u(x)3 + xu(x) ( ), with certain initial conditions. ( ) is the Painlevé II equation (roughly, the branch points and essential singularities are independent of the initial conditions).

55 Paul Painlevé 1863: born in Paris.

56 Paul Painlevé 1863: born in Paris. 1890: Grand Prix des Sciences Mathématiques

57 Paul Painlevé 1863: born in Paris. 1890: Grand Prix des Sciences Mathématiques 1908: first passenger of Wilbur Wright; set flight duration record of one hour, 10 minutes.

58 Paul Painlevé 1863: born in Paris. 1890: Grand Prix des Sciences Mathématiques 1908: first passenger of Wilbur Wright; set flight duration record of one hour, 10 minutes. 1917, 1925: Prime Minister of France.

59 Paul Painlevé 1863: born in Paris. 1890: Grand Prix des Sciences Mathématiques 1908: first passenger of Wilbur Wright; set flight duration record of one hour, 10 minutes. 1917, 1925: Prime Minister of France. 1933: died in Paris.

60 The Tracy-Widom distribution F (t) = exp ( t ) (x t)u(x) 2 dx where u(x) is the Painlevé II function.

61 The Baik-Deift-Johansson theorem Let χ be a random variable with distribution F, and let χ n be the random variable on S n : χ n (w) = is n(w) 2 n n 1/6.

62 The Baik-Deift-Johansson theorem Let χ be a random variable with distribution F, and let χ n be the random variable on S n : Theorem. As n, χ n (w) = is n(w) 2 n n 1/6. χ n χ in distribution, i.e., lim Prob(χ n t) = F (t). n

63 Expectation redux Recall E(n) 2 n.

64 Expectation redux Recall E(n) 2 n. Corollary to BDJ theorem. E(n) = 2 ( ) n + t df (t) n 1/6 + o(n 1/6 ) = 2 n ( )n 1/6 + o(n 1/6 )

65 Proof of BDJ theorem Gessel s theorem reduces the problem to just analysis, viz., the Riemann-Hilbert problem in the theory of integrable systems, and the method of steepest descent to analyze the asymptotic behavior of integrable systems.

66 Origin of Tracy-Widom distribution Where did the Tracy-Widom distribution F (t) come from? F (t) = exp ( t ) (x t)u(x) 2 dx d 2 dx 2u(x) = 2u(x)3 + xu(x)

67 Gaussian Unitary Ensemble (GUE) Analogue of normal distribution for n n hermitian matrices M = (M ij ):

68 Gaussian Unitary Ensemble (GUE) Analogue of normal distribution for n n hermitian matrices M = (M ij ): dm = i dm ii Zn 1 e tr(m 2) dm, i<j d(r M ij )d(i M ij ), where Z n is a normalization constant.

69 Tracy-Widom theorem Tracy-Widom (1994): let α 1 denote the largest eigenvalue of M. Then lim n Prob (( α 1 ) 2n) 2n 1/6 t = F (t).

70 Random topologies Is the connection between is(w) and GUE a coincidence?

71 Random topologies Is the connection between is(w) and GUE a coincidence? Okounkov provides a connection, via the theory of random topologies on surfaces. Very briefly, a surface can be described in two ways: Gluing polygons along their edges, connected to random matrices via quantum gravity. Ramified covering of a sphere, which can be formulated in terms of permutations.

72 Two variations 1. Matchings 2. Pattern avoidance

73 Matching collaborators Joint with: Bill Chen Eva Deng Rosena Du Catherine Yan

74 Complete matchings (complete) matching:

75 Complete matchings (complete) matching: total number of matchings on [2n] := {1, 2,..., 2n} is (2n 1)!! := (2n 1).

76 Crossings and nestings 3 crossing 3 nesting

77 Crossing and nesting number M = matching cr(m) = max{k : k-crossing} ne(m) = max{k : k-nesting}

78 Crossing and nesting number M = matching cr(m) = max{k : k-crossing} ne(m) = max{k : k-nesting} Theorem. The number of matchings on [2n] with no crossings (or with no nestings) is C n := 1 ( ) 2n. n + 1 n

79 Main result on matchings Theorem. Let f n (i, j) = # matchings M on [2n] with cr(m) = i and ne(m) = j. Then f n (i, j) = f n (j, i).

80 Main result on matchings Theorem. Let f n (i, j) = # matchings M on [2n] with cr(m) = i and ne(m) = j. Then f n (i, j) = f n (j, i). Corollary. # matchings M on [2n] with cr(m) = k equals # matchings M on [2n] with ne(m) = k.

81 Oscillating tableaux φ shape (3, 1), length 8

82 Oscillating tableaux φ shape (3, 1), length 8 M Φ (M) φ φ φ φ

83 Proof sketch Φ is a bijection from matchings on 1, 2,..., 2n to oscillating tableaux of length 2n, shape. Corollary. Number of oscillating tableaux of length 2n, shape, is (2n 1)!!.

84 Proof sketch Φ is a bijection from matchings on 1, 2,..., 2n to oscillating tableaux of length 2n, shape. Corollary. Number of oscillating tableaux of length 2n, shape, is (2n 1)!!. (related to Brauer algebra of dimension (2n 1)!!).

85 Schensted for matchings Schensted s theorem for matchings. Let Φ(M) = ( = λ 0, λ 1,..., λ 2n = ). Then cr(m) = max{(λ i ) 1 : 0 i n} ne(m) = max{λ i 1 : 0 i n}.

86 Schensted for matchings Schensted s theorem for matchings. Let Φ(M) = ( = λ 0, λ 1,..., λ 2n = ). Then cr(m) = max{(λ i ) 1 : 0 i n} ne(m) = max{λ i 1 : 0 i n}. Proof. Reduce to ordinary RSK.

87 An example M Φ (M) φ φ φ φ

88 An example M Φ (M) φ φ φ φ cr(m) = 2, ne(m) = 2

89 M Now let cr(m) = i, ne(m) = j, and Φ(M) = ( = λ 0, λ 1,..., λ 2n = ). Define M by Φ(M ) = ( = (λ 0 ), (λ 1 ),..., (λ 2n ) = ). By Schensted s theorem for matchings, cr(m ) = j, ne(m ) = i.

90 Conclusion of proof Thus M M is an involution on matchings of [2n] interchanging cr and ne. Theorem. Let f n (i, j) = # matchings M on [2n] with cr(m) = i and ne(m) = j. Then f n (i, j) = f n (j, i).

91 Simple description? Open: simple description of M M, the analogue of a 1 a 2 a n a n a 2 a 1, which interchanges is and ds.

92 g k (n) g k (n) = number of matching M on [2n] with cro(m) k (matching analogue of u k (n))

93 Grabiner-Magyar theorem Theorem. Define H k (x) = n g k (n) x2n (2n)!. Then Hk (x) = det [ I i j (2x) I i+j (2x) ] k i,j=1 where as before. I m (2x) = j 0 x m+2j j!(m + j)!

94 Gessel s theorem redux Compare: I. Gessel (1990): n 0 u k (n) x2n n! 2 = det [ I i j (2x) ] k i,j=1.

95 Noncrossing example Example. k = 1 (noncrossing matchings): H 1 (x) = I 0 (2x) I 2 (2x) = j 0 C j x 2j (2j)!.

96 Baik-Rains theorem Baik-Rains (implicitly): lim Prob n ( cr n (M) 2n (2n) 1/6 t 2 ) = F 1 (t), where F 1 (t) = F (t) exp ( 1 2 t ) u(s)ds, where F (t) is the Tracy-Widom distribution and u(t) the Painlevé II function.

97 Bounding cr(m) and ne(n) g j,k (n) := #{matchings M on [2n], cr(m) j, ne(m) k}

98 Bounding cr(m) and ne(n) g j,k (n) := #{matchings M on [2n], cr(m) j, ne(m) k} g j,k (n) = #{( = λ 0, λ 1,..., λ 2n = ) : λ i+1 = λ i ±, λ i j k rectangle}, a walk on a graph L(j, k).

99 L(2, 3) φ L(2,3)

100 Transfer matrix generating function A = adjacency matrix of H(j, k) = adjacency matrix of H(j, k) { }. A 0

101 Transfer matrix generating function A = adjacency matrix of H(j, k) = adjacency matrix of H(j, k) { }. A 0 Transfer-matrix method g j,k (n)x 2n = det(i xa 0) det(i xa). n 0

102 Zeros of det(i xa) Theorem (Grabiner, implicitly) Every zero of det(i xa) has the form 2(cos(πr 1 /m) + + cos(πr j /m)), where each r i Z and m = j + k + 1.

103 Zeros of det(i xa) Theorem (Grabiner, implicitly) Every zero of det(i xa) has the form 2(cos(πr 1 /m) + + cos(πr j /m)), where each r i Z and m = j + k + 1. Corollary. Every irreducible factor of det(i xa) over Q has degree dividing 1 φ(2(j + k + 1)), 2 where φ is the Euler phi-function.

104 An example Example. j = 2, k = 5, 1 2φ(16) = 4: det(i xa) = (1 2x 2 )(1 4x 2 + 2x 4 ) (1 8x 2 + 8x 4 )(1 8x 2 + 8x 3 2x 4 ) (1 8x 2 8x 3 2x 4 )

105 Another example j = k = 3, 1 2φ(14) = 3: det(i xa) = (1 x)(1 + x)(1 + x 9x 2 x 3 ) (1 x 9x 2 + x 3 )(1 x 2x 2 + x 3 ) 2 (1 + x 2x 2 x 3 ) 2

106 An open problem rank(a) =? Or even: when is A invertible?

107 An open problem rank(a) =? Or even: when is A invertible? Eigenvalues are known (and A is symmetric)

108 An open problem rank(a) =? Or even: when is A invertible? Eigenvalues are known (and A is symmetric) Cannot tell from the trigonometric expression for the eigenvalues when they are 0.

109 Pattern avoidance v = b 1 b k S k w = a 1 a n S n w avoids v if no subsequence a i1 a ik of w is in the same relative order as v.

110 Pattern avoidance v = b 1 b k S k w = a 1 a n S n w avoids v if no subsequence a i1 a ik of w is in the same relative order as v does not avoid 3142.

111 Pattern avoidance v = b 1 b k S k w = a 1 a n S n w avoids v if no subsequence a i1 a ik of w is in the same relative order as v does not avoid w has no increasing (decreasing) subsequence of length k w avoids 12 k (k 21).

112 The case k = 3 Let v S k. Define S n (v) = {w S n : w avoids v} s n (v) = #S n (v).

113 The case k = 3 Let v S k. Define S n (v) = {w S n : w avoids v} s n (v) = #S n (v). Hammersley-Knuth-Rotem: s n (123) = s n (321) = C n.

114 The case k = 3 Let v S k. Define S n (v) = {w S n : w avoids v} s n (v) = #S n (v). Hammersley-Knuth-Rotem: Knuth: s n (123) = s n (321) = C n. s n (132) = s n (213) = s n (231) = s n (312) = C n.

115 Generating trees Chung-Graham-Hoggatt-Kleiman, West: define u v if u is a subsequence of v

116 123 and 132-avoiding trees black: 123 avoiding magenta: 132 avoiding

117 Structure of the tree k children 2, 3,..., k+1 children

118 Wilf equivalence Define u v if s n (u) = s n (v) for all n.

119 Wilf equivalence Define u v if s n (u) = s n (v) for all n. One equivalence class for k = 3.

120 Wilf equivalence Define u v if s n (u) = s n (v) for all n. One equivalence class for k = 3. Three equivalence classes for k = 4.

121 The three classes for k = 4 Gessel: s n (1234) = 1 (n + 1) 2 (n + 2) n j=0 ( 2j j )( )( ) n + 1 n + 2 j + 1 j + 2

122 The three classes for k = 4 Gessel: s n (1234) = 1 (n + 1) 2 (n + 2) n j=0 ( 2j j )( )( ) n + 1 n + 2 j + 1 j + 2 Bóna: s n (1342)x n = n 0 32x x 8x 2 (1 8x) 3/2,

123 The three classes for k = 4 Gessel: s n (1234) = 1 (n + 1) 2 (n + 2) n j=0 ( 2j j )( )( ) n + 1 n + 2 j + 1 j + 2 Bóna: s n (1342)x n = n 0 32x x 8x 2 (1 8x) 3/2, Open: s n (1324)

124 Typical application Ryan, Lakshmibai-Sandhya, Haiman: Let w S n. The Schubert variety Ω w in the complete flag variety GL(n, C) is smooth if and only if w avoids 4231 and 3412.

125 Typical application Ryan, Lakshmibai-Sandhya, Haiman: Let w S n. The Schubert variety Ω w in the complete flag variety GL(n, C) is smooth if and only if w avoids 4231 and n 0 where f(n)x n = 1 x x2 1 x C(x) = n 0 ( 1 ), 2x 1+x (1 x)c(x) 1 C n x n = 1 1 4x 2x.

126