Lattice paths and orthogonal polynomials: a case study Michael Anshelevich April 13, 2007 P LAN . • Formulation of the problem: free Meixner polynomials and their free cumulants. • General orthogonal polynomials. • Motzkin paths and moments. • Modified Motzkin model for the free Meixner polynomials. • Multivariate case: operators. • Operators on Fock spaces. 1 F ORMULATION OF THE PROBLEM . µt = probability measure on R, mean zero, variance t: m1(µt) = µt(x) = m2(µt) = µt(x2) = Z R Z R x dµt(x) = 0, x2 dµt(x) = t. 2 F ORMULATION OF THE PROBLEM . µt = probability measure on R, mean zero, variance t: m1(µt) = µt(x) = m2(µt) = µt(x2) = Z R Z R x dµt(x) = 0, x2 dµt(x) = t. {Pn} = {P0, P1, P2, . . .} = MOPS (Monic Orthogonal Polynomial System) for µt if Pn (x) = xn + . . . and hPn, Pk i = for n 6= k. Z R Pn(x)Pk (x) dµt(x) = 0 3 Definition. {Pn} are free Meixner polynomials, and µt is a free Meixner distribution, if ∞ X 1 Pn (x)z = F (z) 1 − xG(z) n=0 n for some F (z), G(z). 4 Definition. {Pn} are free Meixner polynomials, and µt is a free Meixner distribution, if ∞ X 1 Pn (x)z = F (z) 1 − xG(z) n=0 n for some F (z), G(z). Example. Un = Chebyshev polynomials of the 2nd kind. Orthogonal with respect to the semicircular distribution 1 dµt(x) = 4t − x2 dx. 2πt q 0.6 0.5 0.4 0.3 0.2 0.1 –1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1 x ∞ X n=0 Un(x)z n = 1 1 1 = . 1 − xz + tz 2 1 + tz 2 1 − x tz 2 2 1+tz 5 Answer. {Pn } satisfy recursion relations xP0 = P1, xP1 = P2 + bP1 + t, xPn = Pn+1 + bPn + (t + c)Pn−1 for n ≥ 2. 6 Answer. {Pn } satisfy recursion relations xP0 = P1, xP1 = P2 + bP1 + t, xPn = Pn+1 + bPn + (t + c)Pn−1 for n ≥ 2. Complete description: q 4(t + c) − (x − b)2 1 dµt(x) = dx + zero, one, or two atoms. 2πt 1 + (b/t)x + (c/t2)x2 7 Particular cases: The semicircular (free Gaussian) distributions 1 4t − x2 dx. 2πt q The Marchenko-Pastur (free Poisson) distributions 1 2π q 4t − (x − b)2 t + bx dx + possibly one atom. 8 Other appearances: Szegö (1922), Bernstein (1930), Boas & Buck (1956), Carlin & McGregor (1957), Geronimus (1961), Allaway (1972), Askey & Ismail (1983), Cohen & Trenholme (1984), Kato (1986), Freeman (1998), Saitoh & Yoshida (2001), M.A. (2003), Kubo, Kuo & Namli (2006), Belinschi & Nica (2007), . . . 9 F REE PROBABILITY CONNECTION : FREE CUMULANTS. mn(µt) = µt(xn) = E.g. m1 = mean. Z R xn dµt(x) = moments of µt. m2 − m2 1 = variance. 10 F REE PROBABILITY CONNECTION : FREE CUMULANTS. mn(µt) = µt(xn) = E.g. m1 = mean. Z R xn dµt(x) = moments of µt. m2 − m2 1 = variance. More generally, define another collection of numbers r1, r2, r3, . . . 11 X rn = mn − Y r|B|. π non-crossing partition of n B class of π π6=1̂ r1 = m1, r2 = m2 − r1r1, r3 = m3 − r2r1 = m3 − − r1r2 − − r2r1 − − r1r1r1 − . 12 X rn = mn − Y r|B|. π non-crossing partition of n B class of π π6=1̂ r1 = m1, r2 = m2 − r1r1, r3 = m3 − r2r1 = m3 − − r1r2 − − r2r1 − − r1r1r1 − . {rn} = free cumulants of µt. 13 Question. Find explicit (combinatorial) formulas for the moments and free cumulants of the free Meixner distributions. 14 Favard’s (Stone’s) Theorem. Let {Pn} be a monic polynomial system. {Pn} are orthogonal with respect to some measure µ if and only if they satisfy a three-term recursion relation xPn = Pn+1 + βnPn + γn Pn−1, for some {βn, γn }. 15 Favard’s (Stone’s) Theorem. Let {Pn} be a monic polynomial system. {Pn} are orthogonal with respect to some measure µ if and only if they satisfy a three-term recursion relation xPn = Pn+1 + βnPn + γn Pn−1, for some {βn, γn }. Example. Free Meixner: β0 = 0, βn = b, γ1 = t, γn = t + c. 16 Favard’s (Stone’s) Theorem. Let {Pn} be a monic polynomial system. {Pn} are orthogonal with respect to some measure µ if and only if they satisfy a three-term recursion relation xPn = Pn+1 + βnPn + γn Pn−1, for some {βn, γn }. Example. Free Meixner: β0 = 0, βn = b, γ1 = t, γn = t + c. Comment. Orthogonal ⇒ Recursion: nice exercise. Recursion ⇒ Orthogonal with respect to a positive functional on polynomials: harder. Recursion ⇒ Orthogonal with respect to a measure: hard operator theory. 17 Use recursion to calculate moments: P0 = 1. x = xP0 = P1 + β0P0, m1 = µ(x) = hx, P0i = β0. 18 Use recursion to calculate moments: P0 = 1. x = xP0 = P1 + β0P0, m1 = µ(x) = hx, P0i = β0. x2 = x · x = x(P1 + β0P0) = P2 + β1P1 + γ1P0 + β0P1 + β0β0P0, m2 = x2, P0 D E = γ1 + β0β0. 19 Use recursion to calculate moments: P0 = 1. x = xP0 = P1 + β0P0, m1 = µ(x) = hx, P0i = β0. x2 = x · x = x(P1 + β0P0) = P2 + β1P1 + γ1P0 + β0P1 + β0β0P0, m2 = x2, P0 D E = γ1 + β0β0. x3 = x · x2 = x(P2 + (β1 + β0)P1 + (γ1 + β0β0)P0) = P3 + β2P2 + γ2P1 + (β1 + β0)P2 + β1(β1 + β0)P1 + γ1(β1 + β0)P0 + (γ1 + β0β0)P1 + β0(γ1 + β0β0)P0, m3 = γ1β1 + γ1β0 + β0γ1 + β0β0β0. 20 mn = coefficient of P0 in xn 21 mn = coefficient of P0 in xn xPn = Pn+1 + βnPn + γn Pn−1. Motzkin paths: right-to-left. step up, weight 1. step horizontally, on level n weight βn. Step down, on level n weight γn . 22 Always stay above the x-axis. xn = all paths of length n. mn = paths ending on level zero. 23 x= + β0 m1 = β0. 24 x= + β0 m1 = β0. β1 γ1 x2 = + + β0 + β0 β0 + m2 = γ1 + β0β0 25 β2 γ2 β1 x3 = + β1 + β1 + β1 γ1 + γ1 + γ1 + + β0 β1 β0 + + β0 β0 γ1 β0 + β0 β0 β0 + β0 + 26 Free Meixner case: β0 = 0, βn = b, γ1 = t, γn = t + c. No horizontal steps on level zero. m1 = 0. t m2 = = t. b t = tb. m3 = t+c b b t t m4 = + t t + . 27 Free cumulants? Motzkin paths ↔ non-crossing partitions. 28 Free cumulants? Motzkin paths ↔ non-crossing partitions. 29 Free cumulants? Motzkin paths ↔ non-crossing partitions. 30 Step up = beginning of class, weight 1. Step down = end of class, weight t or t + c. Step horizontally = middle of class. Non-crossing partitions with no singletons. 31 Example. Semicircular: b = c = 0. No horizontal steps = Dyck paths. Steps down all weight t. m2n+1 = 0, m2n = t2n/2 X π non-crossing pair partition of 2n n = t · # {non-crossing pair partitions of 2n} = tn · n’th Catalan number. Free cumulants: r2 = t, rn = 0, mn = X Y r|B|. π∈NC (n) B∈π 32 Example. Free Poisson / Marchenko-Pastur: c = 0, take b = 1. Steps down all weight t, horizontal steps weight 1. t|π| X mn = π non-crossing partition of n For t = 1, mn = # {non-crossing partitions of n} = n’th Catalan number. For general t: free cumulants rn = t. mn = X Y r|B|. π∈NC (n) B∈π 33 Free cumulants for general b, c, t? Separate t, c: two types of steps down. b t+c t+c t+c b t+c t b t+c t+c t+c t 34 Free cumulants for general b, c, t? Separate t, c: two types of steps down. b b c t t c c c c b t t 35 Free cumulants for general b, c, t? Separate t, c: two types of steps down. b b c t t c c c c b t t 36 Motzkin path with colored descents → non-crossing partition. b b c t t c c c c b t t 37 Motzkin path with colored descents → non-crossing partition. b b c t t c c c c b t t 38 On each class of this partition → further non-crossing partition, one outer class. b b c t t c c c c b t t 39 On each class of this partition → further Motzkin path, One t-descent.. c t b t b b c t c c c t 40 One t-descent. X mn = Y X . non-crossing partitions classes of the partition Motzkin paths with one t-descent So rn = X Motzkin paths of length n with one t-descent =t X Motzkin paths of length n − 2 with only c-descents 41 Comment. rn(µt) = t · rn(µ1). Comment. Motzkin paths, horizontal weight b, descent weight c, no restrictions. xPn = Pn+1 + bPn + cPn−1 (x − b)Pn = Pn+1 + cPn−1 Moments of shifted semicircular distribution, mean b, variance c. 42 M ULTIVARIATE CASE . N ON - COMMUTATIVE POLYNOMIALS. Rhxi = Rhx1, x2, . . . , xdi Involution (x1 x2 x1 x1 x3)∗ = x3 x1 x1 x2 x1. 43 N ON - COMMUTATIVE POLYNOMIALS. Rhxi = Rhx1, x2, . . . , xdi Involution (x1 x2 x1 x1 x3)∗ = x3 x1 x1 x2 x1. ϕ a state on Rhxi: linear functional, ϕ [1] = 1, ∗ ϕ P = ϕ [P ] , ∗ ϕ P P ≥ 0. 44 Multivariate monic orthogonal polynomials Pi(x) = xi + . . . , Pij (x) = xixj + . . . n {P~u(x)} = 1, Pi (x), Pij (x), Pijk (x), . . . o 45 Multivariate monic orthogonal polynomials Pi(x) = xi + . . . , Pij (x) = xixj + . . . n {P~u(x)} = 1, Pi (x), Pij (x), Pijk (x), . . . o Gram-Schmidt: can make orthogonal hP~u, P~v i = ϕ P~u∗P~v = 0 unless ~ u = ~v . 46 Theorem. A family {P~u } is a MOPS for some state ϕ if and only if xi = Pi + Bi,∅,∅, xiPu = P(i,u) + xiP~u = P(i,~u) + d X Bi,w,uPw + δi,uCu, w=1 X Bi,w,~ ~ uPw ~ + δi,u(1)C~ u P(u(2),u(3),...,u(k)), |w|=|~ ~ u| with C~u ≥ 0 and, denoting ~sj = (s(j), . . . , s(k)), Bi,~s,~u k Y j=1 C~sj = Bi,~u,~s k Y j=1 C~uj . Note: can replace the whole ~ u by w, ~ but can remove only the first term of ~ u. 47 Moments: ϕ [xi] = Bi,∅,∅, h i ϕ xixj = δij Ci + Bi,∅,∅Bi,∅,∅, h i ϕ xixj xk = CiBj,i,k + Bi,∅,∅δjk Cj + δij CiBk,∅,∅ + Bi,∅,∅Bj,∅,∅Bk,∅,∅. Motzkin paths, colored, complicated. Another approach: use operators. 48 B ACK TO ONE - DIMENSIONAL CASE . xPn = Pn+1 + βnPn + γn Pn−1, Vector space P of polynomials, orthogonal basis {Pn}. Operator X of multiplication by x: β0 1 0 X= 0 0 ... P0 γ1 β1 1 0 0 ... P1 0 γ2 β2 1 0 ... P2 0 0 γ3 β3 1 ... P3 0 0 0 γ4 β4 ... P4 ... ... ... ... ... ... P5 P0 P1 P2 P3 P4 P5 mn = µ(xn) = hP0, X n P0i . 49 More abstractly, in a Hilbert space F (C) with orthogonal basis Ω = “vacuum vector”. n o ⊗2 ⊗3 Ω, e, e , e , . . . ... β0 γ1 0 0 0 ... 1 β1 γ2 0 0 ... 0 1 β2 γ3 0 1 β3 γ4 . . . X= 0 0 0 0 0 1 β4 . . . ... ... ... ... ... ... Ω e e⊗2 e⊗3 e⊗4 e⊗5 Ω e e⊗2 e⊗3 e⊗4 e⊗5 mn = hΩ, X nΩi . 50 In the free Meixner case, ... 0 t 0 0 0 ... 1 b t+c 0 0 ... 0 1 b t+c 0 1 b t + c ... X= 0 0 ... 0 0 0 1 b ... ... ... ... ... ... Ω e e⊗2 e⊗3 e⊗4 e⊗5 Ω e e⊗2 e⊗3 e⊗4 e⊗5 X = sum of four operators: a+ (e⊗k ) = e⊗(k+1), B(e⊗k ) = be⊗k , B(Ω) = 0 a−(e⊗k ) = e⊗(k−1), a−C(e⊗k ) = ce⊗(k−1), a−C(e) = 0. 51 X = a+ + B + ta− + a−C, Moreover, D E + − − n n mn = hΩ, X Ωi = Ω, (a + B + ta + a C) Ω . E + − n rn+2 = t e, (a + B + a C) e . D 52 Definition. A state ϕ is a free Meixner state if its monic orthogonal polynomials have a generating function X P~u(x)z~u = 1 + u ~ X i Pi(x)zi + X Pij (x)zizj + . . . i,j = F (z) 1 − X i −1 xiGi(z) for some F (z), Gi(z). Theorem. h i D E ϕ xu(1)xu(2) . . . xu(n) = Ω, Xu(1) Xu(2) . . . Xu(n)Ω . 53 Here − − Xk = a+ + B + ta + a k k k kC operators on a Fock space FC (Cd ). B1, . . . , Bd symmetric d × d matrices. C diagonal d2 × d2 matrix, ! ! Bi 0 Bi 0 C=C . 0 I 0 I 54 Moreover, the free cumulant functional h i D E Rϕ xixu(1)xu(2) . . . xu(n)xj = t ei , Su(1)Su(2) . . . Su(n)ej , where − Sk = a+ + B + a k k k C. 55