arXiv:math/0010135v1 [math.CO] 13 Oct 2000 n hi plctos ahmtclSine eerhInstitut Research Sciences Mathematical 2001. Press, : University Applications” Their and rnesUiest,Wlhm as044 S.Emi:vanmoerbe acknow gratefully Foundation is Science grant E-mail: Foundation National Francqui USA. a a and of FNRS 02454, a support Nato, Mass The Waltham, @math.brandeis.edu. University, Brandeis nerbeLtie:Rno arcsadRandom and Matrices Random Lattices: Integrable ∗ † xaddvrino etrsa SI ekly pig19.T a To 1998. Spring Berkeley, MSRI, at lectures of version Expanded eateto ahmtc,Uiested ovi,14 Louvain 1348 Louvain, de Universit´e Mathematics, of Department ieeta qain o h rgnlprobabilities. original the for equ equations integrable the differential determina with integ Fredholm combined for which or constraints, also integrals rasoro matrix but These lattices, Toeplitz int equation. and KdV matrix Pfaff these Toda, that the tau-functio A like showing natural are to spaces. time-parameters, devoted adding symmetric s priately is over symmetric lectures integrals to these as of space finally tangent and the groups over over integrals determinan Fredholm as T (or naturally integrals permutations. matrix random suggest and infinite) problems and (finite matrices dom hs etrspeetasre frcn eeomnsi th in developments recent of survey a present lectures These irevnMoerbeke van Pierre Permutations ac 5 2000 25, March Abstract 1 ∗ † sfritgal lattices, integrable for ns ledged. ulctos#0 Cambridge #40, Publications e tosla o(partial) to lead ations al D’,lk the like PDE’s, rable s,wihaievery arise which ts), rn M-84570 a DMS-98-4-50790, # grant pa n”admMatrices ”Random in ppear gas pnapro- upon egrals, t lostsyVi- satisfy also nts ae,a integrals as paces, motn part important n eeprobabilistic hese l-ev,Blimand Belgium -la-Neuve, ego.c.cb and [email protected] rao ran- of area e 0, p.2 § Contents
0 introduction 3
1 Matrix integrals, random matrices and permutations 5 1.1 Tangent space to symmetric spaces and associated random matrix en- sembles ...... 6 1.2 Infinite Hermitian matrix ensembles ...... 12 1.3 Integralsoverclassicalgroups ...... 15 1.4 Permutationsandintegralsovergroups ...... 17
2 Integrals, vertex operators and Virasoro relations 26 2.1 Beta-integrals ...... 26 2.1.1 Virasoro constraints for Beta-integrals ...... 26 2.1.2 Proof: β-integrals as fixed points of vertex operators ...... 29 2.1.3 Examples ...... 32 2.2 Doublematrixintegrals...... 34 2.3 Integralsovertheunitcircle ...... 36
3 Integrable systems and associated matrix integrals 38 3.1 Toda lattice and Hermitian matrix integrals ...... 38 3.1.1 Toda lattice, factorization of symmetric matrices and orthogonal polynomials ...... 38 3.1.2 SketchofProof ...... 41 3.2 Pfaff Lattice and symmetric/symplectic matrix integrals ...... 47 3.2.1 Pfaff lattice, factorization of skew-symmetric matrices and skew- orthogonalpolynomials...... 47 3.2.2 SketchofProof...... 51 3.3 2d-Toda lattice and coupled Hermitian matrix integrals ...... 56 3.3.1 2d-Toda lattice, factorization of moment matrices and bi-orthogonal polynomials ...... 56 3.3.2 Sketchofproof ...... 58 3.4 The Toeplitz Lattice and Unitary matrix integrals ...... 59 3.4.1 Toeplitz lattice, factorization of moment matrices and bi-orthogonal polynomials ...... 59 3.4.2 SketchofProof ...... 63
4 Ensembles of finite random matrices 64 4.1 PDE’s defined by the probabilities in Hermitian, symmetric and sym- plecticrandomensembles...... 64 4.1.1 Gaussian Hermitian, symmetric and symplectic ensembles . . . 66 4.1.2 Laguerre Hermitian, symmetric and symplectic ensembles . . . . 66 4.1.3 Jacobi Hermitian, symmetric and symplectic ensembles . . . . . 67
2 0, p.3 §
4.2 ODE’s, when E hasoneboundarypoint ...... 68 4.3 ProofofTheorems4.1,4.2and4.3 ...... 71 4.3.1 GaussianandLaguerreensembles ...... 71 4.3.2 Jacobiensemble...... 72 4.3.3 Inserting partials into the integrable equation ...... 73
5 Ensembles of infinite random matrices: Fredholm determinants, as τ-functions of the KdV equation 75
6 Coupled random Hermitian ensembles 84
7 Random permutations 85
8 Random involutions 89
9 Appendix: Chazy classes 91
0 introduction
The purpose of these lectures is to give a survey of recent interactions between statistical questions and integrable theory. Two types of questions will be tackled here: (i) Consider a random ensemble of matrices, with certain symmetry conditions to guarantee the reality of the spectrum and subjected to a given statistics. What is the probability that all its eigenvalues belong to a given subset E ? What happens, when the size of the matrices gets very large ? The probabilities here are functions of the boundary points ci of E. (ii) What is the statistics of the length of the largest increasing sequence in a random permutation, assuming each permutation is equally probable ? Here, one considers generating functions (over the size of the permutations) for the probability distributions, depending on the variable x. The main emphasis of these lectures is to show that integrable theory serves as a useful tool for finding equations satisfied by these functions of x, and conversely the probabilities point the way to new integrable systems. These questions are all related to integrals over spaces of matrices. Such spaces can be classical Lie groups or algebras, symmetric spaces or their tangent spaces. In infinite- dimensional situations, the ” -fold” integrals get replaced by Fredholm determinants. ∞ During the last decade, astonishing discoveries have been made in a variety of di- rections. A first striking feature is that these probabilities are all related to Painlev´e equations or interesting generalizations. In this way, new and unusual distributions have entered the statistical world. Another feature is that each of these problems is related to some integrable hier- archy. Indeed, by inserting an infinite set of time variables t1, t2, t3, ... in the integrals
3 0, p.4 §
∞ t yi or Fredholm determinants - e.g., by introducing appropriate exponentials e 1 i in P the integral - this probability, as a function of t1, t2, t3, ..., satisfies an integrable hier- archy. Korteweg-de Vries, KP, Toda lattice equations are only a few examples of such integrable equations. Typically integrable systems can be viewed as isospectral deformations of differential or difference operators . Perhaps, one of the most startling discoveries of integrable theory is that can beL expressed in terms of a single“τ-function” τ(t , t , ...) (or vector L 1 2 of τ-functions), which satisfy an infinite set of non-linear equations, encapsulated in a single “bilinear identity”. The ti account for the commuting flows of this integrable hierarchy. In this way, many interesting classical functions live under the same hat: characters of representations, Θ-functions of algebraic geometry, hypergeometric func- tions, certain integrals over classical Lie algebras or groups, Fredholm determinants, arising in statistical mechanics, in scattering and random matrix theory! They are all special instances of “τ-functions”. The point is that the probabilities or generating functions above, as functions of t1, t2, ... (after some minor renormalization) are precisely such τ-functions for the cor- responding integrable hierarchy and thus automatically satisfy a large set of equations. These probabilities are very special τ-functions: they happen to be a solution of yet another hierarchy of (linear) equations in the variables ti and the boundary points ci, J(2) J(2) namely k τ(t; c) = 0, where the k form -roughly speaking- a Virasoro-like algebra:
J(2), J(2) =(k ℓ) J(2) + ... k ℓ − k+ℓ h i The point is that each integrable hierarchy has a natural “vertex operator”, which auto- matically leads to a natural Virasoro algebra. Then, eliminating the partial derivatives in t from the two hierarchy of equations, the integrable and the Virasoro hierarchies, and finally setting t = 0, lead to PDE’s or ODE’s satisfied by the probabilities. In the table below, we give an overview of the different problems, discussed in this lecture, the relevant integrals in the second column and the different hierarchies satisfied by the integrals. To fix notation, , , refer to the Hermitian, symmetric and Hℓ Sℓ Tℓ symplectic ensembles, populated respectively by ℓ ℓ Hermitian matrices , symmetric × matrices and self-dual Hermitian matrices, with quaternionic entries. ℓ(E), ℓ(E), (E) are the corresponding set of matrices, with all spectral points belongingH S to E. Tℓ U(ℓ) and O(ℓ) are the unitary and orthogonal groups respectively. In the table below, i Vt(z) := V0(z)+ tiz , where V0(z) stands for the unperturbed problem; in the last integral V˜t(z) is a more complicated function of t1, t2, ... and z, to be specified later. P
4 1, p.5 §
Probability problem underlying t-perturbed corresponding integral, τ-function of integrable −→ hierarchies
∞ i P (M (E)) eTr( V (M)+ 1 tiM )dM Toda lattice n n(E) − ∈H H P R KP hierarchy
Tr( V (M)+ ∞ t M i) P (M (E)) e 1 i dM Pfaff lattice n n(E) − ∈ S S P R Pfaff-KP hierarchy
Tr( V (M)+ ∞ t M i) P (M (E)) e 1 i dM Pfaff lattice n n(E) − ∈T T P R Pfaff-KP hierarchy
P ((M1, M2) 2 (E) dM1dM2 2d-Toda lattice ∈ Hn Tr(Vt(M1) Vs(M2) cM1M2) KP-hierarchy (E ) (E )) eR− − − Hn 1 ×Hn 2
P (M (E)) det (I Kt(y, z)IEc (z)) KdV equation ∈H∞ − ( Fredholm determinant)
∞ i ¯ i Tr 1 (tiM siM ) longest increasing sequence U(ℓ) e − dM Toeplitz lattice P in random permutations R 2d-Toda lattice
Tr(xM+V˜t(M)) longest increasing sequence O(ℓ) e dM Toda lattice in random involutions R KP-hierarchy
Acknowledgment: These lectures represent joint work especially with (but also in- spired by) Mark Adler, Taka Shiota and Emil Horozov. Thanks also for many informa- tive discussions with Jinho Baik, Pavel Bleher, Edward Frenkel, Alberto Gr¨unbaum, Alexander Its, espacially Craig Tracy and Harold Widom, and with other participants in the semester at MSRI. I wish to thank Pavel Bleher, David Eisenbud and Alexander Its for organizing a truly stimulating and enjoyable semester at MSRI.
1 Matrix integrals, random matrices and permuta- tions
5 1, p.6 §
1.1 Tangent space to symmetric spaces and associated random matrix ensembles Random matrices provided a model for excitation spectra of heavy nuclei at high ex- citations (Wigner [76], Dyson [27] and Mehta [49]), based on the nuclear experimental data by Porter and Rosenzweig [56]; they observed that the occurrence of two levels, close to each other, is a rare event (level repulsion), showing that the spacing is not Poissonian, as one might expect from a naive point of view. Random matrix ideas play an increasingly prominent role in mathematics: not only have they come up in the spacings of the zeroes of the Riemann zeta function, but their relevance has been observed in the chaotic Sinai billiard and, more generally, in chaotic geodesic flows. Chaos seems to lead to the “spectral rigidity”, typical of the spectral distributions of random matrices, whereas the spectrum of an integrable system is random (Poisson)! (e.g., see Odlyzko [53] and Sarnak [60]). All these problems have led to three very natural random matrix ensembles: Hermi- tian, symmetric and symplectic ensembles. The purpose of this section is to show that these three examples appear very naturally as tangent spaces to symmetric spaces. A symmetric space G/K is given by a semi-simple Lie group G and a Lie group involution σ : G G such that → K = x G, σ(x)= x . { ∈ } Then the following identification holds: 1 G/K = gσ(g)− with g G , ∼ { ∈ } and the involution σ induces a map of the Lie algebra σ : g g such that (σ )2 =1, ∗ −→ ∗ with k = a g such that σ (a)= a g = k p with { ∈ ∗ } ⊕ p = a g such that σ (a)= a { ∈ ∗ − } and [k, k] k, [k, p] p, [p, p] k. ⊂ ⊂ ⊂ 1 Then K acts on p by conjugation: kpk− p for all k K and p is the tangent space to G/K at the identity. The action of K on⊂p induces a∈ root space decomposition, with a being a maximal abelian subalgebra in p:
p = a + pα, with mα = dim pα. α ∆ X∈ Then, according to Helgason [35], the volume element on p is given by
dV = α(z)mα dz ...dz , 1 n α ∆+ Y∈ 6 1, p.7 §
where ∆+ is the set of positive roots; see [35, 36, 61, 62]. This will subsequently be worked out for the three so-called An-symmetric spaces. See also Sarnak’s MSRI-lecture [59] in these proceedings, who deals with more general symmetric spaces. I like to thank Chuu-Lian Terng for very helpful conversations on these matters.
Examples :
(i) Hermitian ensemble
1 1 Consider the non-compact symmetric space SL(n, C)/SU(n) with σ(g)=¯g⊤− . Then
SL(n, C)/SU(n) = gg¯⊤ g SL(n, C) { | ∈ } = positive definite matrices with det = 1 { } with
1 K = g SL(n, C) σ(g)= g = g SL(n, C) g− =g ¯⊤ = SU(n). { ∈ | } { ∈ | }
Then σ (a) = a¯⊤ and the tangent space to G/K is then given by the space p = n ∗ of Hermitian matrices− H
sl(n, C)= k p = su(n) , i.e., a = a + a , a su(n), a . ⊕ ⊕ Hn 1 2 1 ∈ 2 ∈ Hn If M , then the M , M and M (1 i < j n) are free variables, so that ∈ Hn ii ℜ ij ℑ ij ≤ ≤ Haar measure on M takes on the following form: ∈ Hn n dM := dM (d M d M ). (1.1.1) ii ℜ ij ℑ ij 1 1 i A maximal abelian subalgebra a p = n is given by real diagonal matrices z = diag(z ,...,z ). Each M p = can⊂ be writtenH as 1 n ∈ Hn A A A M = e z e− , e K = SU(n), ∈ with2 A = (a (e e )+ ib (e + e )) k = su(n), a =0. (1.1.2) kℓ kℓ − ℓk kℓ kℓ ℓk ∈ ℓℓ 1 k ℓ n ≤X≤ ≤ Notice that e e and i(e + e ) k = su(n) and that kℓ − ℓk kℓ ℓk ∈ [e e , z]=(z z )(e + e ) p = kℓ − ℓk ℓ − k kℓ ℓk ∈ Hn 1The corresponding compact symmetric space is given by (SU(n) SU(n))/SU(n). 2e is the n n matrix with all zeroes, except for 1 at the (k,ℓ)th× entry. kℓ × 7 1, p.8 § [i(e + e ), z]=(z z )i(e e ) p = . (1.1.3) kℓ ℓk ℓ − k kℓ − ℓk ∈ Hn 3 Incidentally, this implies that ekℓ + eℓk and i(ekℓ + eℓk) are two-dimensional eigenspaces of (ad z)2 with eigenvalue (z z )2. From (1.1.2) and (1.1.3) it follows that ℓ − k [A, z]=(z z ) (a (e + e )+ ib (e e )) p = (1.1.4) ℓ − k kℓ kℓ ℓk kℓ kℓ − ℓk ∈ Hn 1 k<ℓ n ≤X≤ and thus, for small A, we have4 A A dM = d(e z e− ) = d(z + [A, z]+ ... ) n = dz d((z z )a )d((z z )b ), using (1.1.4) and (1.1.1) i ℓ − k kℓ ℓ − k kℓ 1 1 k<ℓ n ≤ ≤ Yn Y 2 = dzi∆n(z) dakℓdbkℓ. (1.1.5) 1 1 k<ℓ n Y ≤Y≤ Therefore ∆2(z) is also the Jacobian determinant of the map M (z, U), such that 1 → M = UzU − , and thus dM admits the decomposition in polar coordinates: ∈ Hn dM = ∆2 (z)dz ...dz dU, U SU(n). (1.1.6) n 1 n ∈ In random matrix theory, is endowed with the following probability, Hn trV (M) V (z) P (M dM)= c e− dM, ρ(dz)= e− dz, (1.1.7) ∈ n where dM is Haar measure (1.1.6) on and c is the normalizing factor. Since dM Hn n as in (1.1.6) contains dU and since the probability measure (1.1.7) only depends on the trace of V (M), dU completely integrates out. Given E R, define ⊂ (E) := M with all spectral points E R . (1.1.8) Hn { ∈ Hn ∈ ⊂ } ⊂ Hn Then 2 n Tr V (M) En ∆ (z) 1 ρ(dzk) − P (M n(E)) = cne dM = 2 n . (1.1.9) ∈ H n(E) Rn ∆ (z) 1 ρ(dzk) ZH R Q As the reader can find out from the excellent bookR by MehtaQ [49], it is well known that, if the probability P (M dM) satisfies the following two requirements: (i) in- ∈ 1 variance under conjugation by unitary transformations M UMU − , (ii) the random variables M , M , M , 1 i < j n are independent,7→ then V (z) is quadratic ii ℜ ij ℑ ij ≤ ≤ (Gaussian ensemble). 3ad x(y) := [x, y]. 4∆ (z)= (z z ) is the Vandermonde determinant. n i − j ≤ ≤ 1 i (ii) Symmetric ensemble 5 1 Here we consider the non-compact symmetric space SL(n, R)/SO(n) with σ(g)= g⊤− . Then SL(n, R)/SO(n) = gg⊤ g SL(n, R) { | ∈ } = positive definite matrices with det = 1 { } with 1 K = g SL(n, R) σ(g)= g = g SL(n, R) g⊤ = g− = SO(n). { ∈ | } { ∈ | } Then σ (a)= a⊤ and the tangent space to G/K is then given by the space p = n of ∗ − S symmetric matrices, appearing in the decomposition of sl(n, R), sl(n, R)= k p = so(n) , i.e., a = a + a , a so(n), a ⊕ ⊕ Sn 1 2 1 ∈ 2 ∈ Sn with Haar measure dM = dM on . ij Sn 1 i j n A maximal abelian subalgebra≤Y≤ ≤ a p = is given by real traceless diagonal matrices ⊂ Sn z = diag(z ,...,z ). Each M p = conjugates to a diagonal matrix z 1 n ∈ Sn A A A M = e z e− , e K = SO(n), A so(n). ∈ ∈ A calculation, analogous to example (i)(1.1.5) leads to dM = ∆ (z) dz ...dz dU, U SO(n). n 1 n ∈ Random matrix theory deals with the following probability on : Sn trV (M) V (z) P (M dM)= c e− dM, ρ(dz)= e− dz, (1.1.10) ∈ n with normalizing factor cn. Setting as in (1.1.8): n(E) n is the subset of matrices with spectrum E. Then S ⊂ S ∈ n Tr V (M) En ∆(z) 1 ρ(dzk) P (M n(E)) = cne− dM = | | n . (1.1.11) ∈ S n(E) Rn ∆(z) 1 ρ(dzk) ZS R | | Q As in the Hermitian case, P (M dM) is Gaussian,R if P (M QdM) satisfies ∈ ∈ 1 (i) invariance under conjugation by orthogonal conjugation M OMO− , (ii) Mii, Mij (i < j) are independent random variables. → 5The compact version is given by SU(n)/SO(n). 9 1, p.10 § (iii) Symplectic ensemble 6 1 1 Consider the non-compact symmetric space SU ∗(2n)/USp(n) with σ(g)= Jg⊤− J − , where J is the 2n 2n matrix: × 0 1 1 0 − 0 1 1 0 2 J := with J = I, (1.1.12) − − 0 1 1 0 − . .. and 1 G = SU ∗(2n)= g SL(2n, C) g = JgJ¯ − , { ∈ | } K = g SU ∗(2n) σ(g)= g := Sp(n, C) U(2n) { ∈ | } ∩ 1 = g SL(2n, C) g⊤Jg = J g SL(2n, C) g− =g ¯⊤ { ∈ }∩{ ∈ } 1 1 = g SL(2n, C) g− =g ¯⊤and g = JgJ¯ − { ∈ } =: USp(n). 1 Then, σ (a)= Ja⊤J − and ∗ − k = a su∗(2n) σ (a)= a = sp(n, C) u(2n) { ∈ ∗ } ∩ 2n 2n 1 = a C × a⊤ = a,¯ a = JaJ¯ − { ∈ − } p = a su∗(2n) σ (a)= a = su∗(2n) iu(2n) { ∈ ∗ − } ∩ 2n 2n 1 = a C × a⊤ =¯a, a = JaJ¯ − { ∈ } (0) (1) Mkℓ Mkℓ ¯ C2 2 = M =(Mkℓ)1 k,ℓ n, Mkℓ = with Mℓk = Mkℓ⊤ × ≤ ≤ (1) (0) ∈ M¯ M¯ − kℓ kℓ = self-dual n n Hermitean matrices, with quaternionic entries ∼ { × } =: . T2n The condition on the 2 2 matrices Mkℓ implies that Mkk = MkI, with Mk R and the 2 2 identity I. Notice× USp(n) acts naturally by conjugation on the tangent∈ space × 6The corresponding compact symmetric space is SU(2n)/Sp(n). 10 1, p.11 § p to G/K. Haar measure on is given by T2n n (0) ¯ (0) (1) ¯ (1) dM = dMk dMkℓ dMkℓ dMkℓ dMkℓ , (1.1.13) 1 1 k<ℓ n Y ≤Y≤ since these M are the only free variables in the matrix M . A maximal abelian ij ∈ T2n subalgebra in p is given by real diagonal matrices of the form z = diag(z1, z1, z2, z2, ..., zn, zn). Each M p = can be written as ∈ T2n A A A M = e ze− , e K = USp(n), (1.1.14) ∈ with (a , b ,c ,d R) kℓ kℓ kℓ kℓ ∈ A = a (e(0) e(0))+ b (e(1) + e(1))+ c (e(2) e(2))+ d (e(3) + e(3)) k kℓ kℓ − ℓk kℓ kℓ ℓk kℓ kℓ − ℓk kℓ kℓ ℓk ∈ 1 k ℓ n (1.1.15) ≤X≤ ≤ in terms of the four 2 2 matrices 7 × 1 0 i 0 0 1 0 i e(0) = , e(1) = , e(2) = , e(3) = . 0 1 0 i 1 0 i 0 − − Since e(0) e(0), z = (z z )(e(0) + e(0)) p kℓ − ℓk ℓ − k kℓ ℓk ∈ (1) (1) (1) (1) ekℓ + eℓk , z = (zℓ zk)(ekℓ eℓk ) p − − ∈ (2) (2) (2) (2) ekℓ eℓk , z = (zℓ zk)(ekℓ + eℓk ) p − − ∈ (3) (3) (3) (3) ekℓ + eℓk , z = (zℓ zk)(ekℓ eℓk ) p, (1.1.16) − − ∈ [A, z] p has the following form: it has 2 2 zero blocks along the diagonal and from (1.1.16)∈ and (1.1.15), × akℓ + ibkℓ ckℓ + idkℓ ((k,ℓ)th block in [A, z]) = (zℓ zk) , (k<ℓ). (1.1.17) − ckℓ + idkℓ akℓ ibkℓ − − Therefore, using (1.1.17), Haar measure dM on equals T2n A A dM = d(e ze− ) = d(I + A + ... )z(I A + ... ) − = d(z + [A, z]+ ... ) = dz d((z z )(a + ib ))d((z z )(a ib )) k ℓ − k kℓ kℓ ℓ − k kℓ − kℓ 1 k n 1 k<ℓ n ≤Y≤ ≤Y≤ d((z z )(c + id ))d((z z )( c + id )) ℓ − k kℓ kℓ ℓ − k − kℓ kℓ = ∆4(z)dz dz 4da db dc dd . 1 ··· n kℓ kℓ kℓ kℓ 1 k<ℓ n ≤Y≤ 7e(i) in (1.1.15) refers to putting the 2 2 matrix e(i) at place (k,ℓ). kℓ × 11 1, p.12 § As before, define 2n(E) 2n as the subset of matrices with spectrum E and define the probability:T ⊂ T ∈ 4 n Tr V (M) En ∆ (z) 1 ρ(dzk) − P (M 2n(E)) = cne dM = 4 n . (1.1.18) ∈ T 2n(E) Rn ∆ (z) 1 ρ(dzk) ZT R Q R Q Remark: Notice 2n is called the symplectic ensemble, although the matrices in p = 2n are not at all symplectic;T but rather the matrices in k are. T 1.2 Infinite Hermitian matrix ensembles Consider now he limit of the probability ∆2(z) n ρ(dz ) P (M (E)) = En 1 k , when n . (1.2.1) ∈ Hn ∆2(z) n ρ(dz ) ր∞ RRn Q1 k Dyson [27] (see also Mehta [49]) usedR the followingQ trick, to circumvent the problem of dealing with -fold integrals. Using the orthogonality of the monic orthogonal ∞ R 2 2 polynomials pk = pk(z) for the weight ρ(dz) on , and the L -norms hk = R pk(z)ρ(dz) of the p ’s, one finds, using (det A)2 = det(AA ), k ⊤ R n 2 ∆ (z) ρ(dzi) Rn 1 Z Y n = det(pi 1(zj))1 i,j n det(pk 1(zℓ))1 k,ℓ n ρ(dzk) Rn − ≤ ≤ − ≤ ≤ Z k=1 n Y π+π′ = ( 1) pπ(k) 1(zk)pπ′(k) 1(zk)ρ(dzk) − − ′ − R π,π σn k=1 X∈ Y Z n 1 n 1 − − 2 = n! pk(z)ρ(dz)= n! hk. (1.2.2) R 0 0 Y Z Y For the integral over an arbitrary subset E R, one stops at the second equality, since ⊂ the pn’s are not necessarily orthogonal over E. This leads to the probability (1.2.1), P (M (E))) ∈ Hn 1 n = n det pj 1(zk)pj 1(zℓ) ρ(dzi) − − n! 1 hi 1 En ! − Z 1 j n 1 k,ℓ n 1 ≤X≤ ≤ ≤ Y 1 Q n = det(Kn(zk, zℓ))1 k,ℓ n ρ(dzi), (1.2.3) n! n ≤ ≤ E 1 Z Y 12 1, p.13 § in terms of the kernel n pj 1(y) pj 1(z) Kn(y, z) := − − . (1.2.4) j=1 hj 1 hj 1 X − − p p The orthonormality relations of the pk(y)/√hk lead to the reproducing property for the kernel Kn(y, z): Kn(y, z)Kn(z,u)ρ(dz)= Kn(y,u), Kn(z, z)ρ(dz)= n. (1.2.5) R R Z Z n k n k Upon replacing E by dz R − in (1.2.3), upon integrating out all the remaining 1 i × variables zk+1, ..., zn and using the reproducing property (1.2.5), one finds the n-point correlation function Q P (one eigenvalue in each [zi, zi + dzi], i =1, ..., k) k = cn det (Kn(zi, zj))1 i,j k ρ(dzi). (1.2.6) ≤ ≤ 1 Y Finally, by Poincar´e’s formula for the probability P E , the probability that no ∪ i spectral point of M belongs to E is given by a Fredholm determinant P (M (Ec)) = det(I λKE) ∈ Hn − n k ∞ k E = 1+ ( λ) det Kn (zi, zj) ρ(dzi), − 1 i,j k z1 ... zk ≤ ≤ 1 Xk=1 Z ≤ ≤ Y E for the kernel Kn (y, z)= Kn(y, z)IE(z). Wigner’s semi-circle law: For this ensemble (defined by a large class of ρ’s, in partic- •ular for the Gaussian ensemble) and for very large n, the density of eigenvalues tends to Wigner’s semi-circle distribution on the interval [ √2n, √2n]: − 1 = √2n z2dz, z √2n π − | | ≤ density of eigenvalues =0, z > √2n. | | Bulk scaling limit: From the formula above, it follows that the average number of eigenvalues• per unit length near z = 0 (“the bulk”) is given by √2n/π and thus the average distance between two consecutive eigenvalues is given by π/√2n. Upon using this rescaling, one shows ([43, 48, 52, 55, 39]) π πx πy sin π(x y) lim Kn , = − (Sine kernel) n √2n √2n √2n π(x y) ր∞ − 13 1, p.14 § and ( 1)k ∂ k P (exactly k eigenvalues [0, a]) = − det(I λKI[0,a]) ∈ k! ∂λ − λ=1 with πa f(x; λ) det(I λKI ) = exp dx, (1.2.7) − [0,a] x Z0 where f(x, λ) is a solution to the following differential equation, due to the pioneering work of Jimbo, Miwa, Mori, Sato [39], (′ = ∂/∂x) 2 2 λ (xf ′′) = 4(xf ′ f)( f ′ xf ′ + f) , with f(x; λ) = x for x 0. − − − ∼ −π ≃ (Painlev´eV) (1.2.8) Edge scaling limit: Near the edge √2n of the Wigner semi-circle, the scaling is √2n1/6 • and thus the scaling is more subtle: (see [21, 30, 51, 49, 64]) u y = √2n + , (1.2.9) √2n1/6 and so for the kernel Kn as in (1.2.4), with the pn’s being Hermite polynomials, 1 u v lim Kn √2n + , √2n + = K(u, v), n √2n1/6 √2n1/6 √2n1/6 ր∞ where ∞ ∞ iux x3/3 K(u, v)= A(x + u)A(x + v)dx, A(u)= e − dx. 0 Z Z−∞ Relating y and u by (1.2.9), the statistics of the largest eigenvalue for very large n is governed by the function, 2 λmax P (λmax y) = P 2n 3 1 u , for n , ≤ √2n − ≤ ր∞ ∞ 2 = det(I KI( ,u]) = exp (α u)g (α)dα , − −∞ − − Zu with g(x) a solution of 3 g′′ = xg +2g 3 2 2 (Painlev´eII) (1.2.10) e− 3 x ( g(x) = 1/4 for x . ∼ − 2√πx ր∞ The latter is essentially the asymptotics of the Airy function. In section 5, I shall derive, via Virasoro constraints, not only this result, due to Tracy-Widom [64], but 14 1, p.15 § also a PDE for the probability that the eigenvalues belong to several intervals, due to Adler-Shiota-van Moerbeke [11, 12]. Hard edge scaling limit: Consider the ensemble of n n random matrices for the La- • × ν/2 z/2 guerre probability distribution, thus corresponding to (1.1.9) with ρ(dz)= z e− dz. One shows the density of eigenvalues near z = 0 is given by 4n for very large n. At this edge, one computes for the kernel (1.2.4) with Laguerre polynomials pn [52, 30]: 1 (ν) u v (ν) lim Kn , = K (u, v), (1.2.11) n 4n 4n 4n ր∞ (ν) where K (u, v) is the Bessel kernel, with Bessel functions Jν, 1 1 K(ν)(u, v) = xJ (xu)J (xv)dx 2 ν ν Z0 J (u)√uJ ′ (v) J (√v)√vJ ′ (√u) = ν ν − ν ν . (1.2.12) 2(u v) − Then x f(u) P ( no eigenvalues [0, x]) = exp du , ∈ − u Z0 with f satisfying 2 2 2 (xf ′′) 4 xf ′ f f ′ + (x ν )f ′ f f ′ =0. (Painlev´eV) (1.2.13) − − − − This result due to Tracy-Widom [65] and a more general statement, due to [11, 12] will be shown using Virasoro constraints in section 5. 1.3 Integrals over classical groups The integration on a compact semi-simple simply connected Lie group G is given by the formula (see Helgason [36]) 1 α(iH) 1 H f(M)dM = 2 sin dt f(utu− )du, t = e , (1.3.1) G W T 2 U Z | | Z α ∆ Z Y∈ where A G is a maximal subgroup, with g and a being the Lie algebras of G and A. Let du and⊂ dt be Haar measures on G, A respectively such that dt = du = 1; ZA ZU ∆ denotes the set of roots of g with respect to a; W is the order of the Weyl group of G. | | 15 1, p.16 § ∞ t Tr M i Integration formula (1.3.1) will be applied to integrals of f = e 1 i over the groups SO(2n), SO(2n+1) and Sp(n). Their Lie algebras (over C) areP given respectively by dn, bn, cn, with sets of roots: (e.g., see [20]) ∆ = εe , 1 i k, (e + e ), (e e ), 1 i < j n , n {± i ≤ ≤ ± i j ± i − j ≤ ≤ } with ε = 0 for dn = so(2n) ε = 1 for bn = so(2n + 1) ε = 2 for cn = sp(n). Setting H = iθ, we have, in view of formula (1.3.1), α(iH) 2 sin dt 2 α ∆ Y∈ 2 θ θ θ +θ n j − k j k cn 1 j n n ∞ t Tr M k ∞ t n P(eikθj +e−ikθj ) 2 ∞ t cos kθ 2 t T (z ) e 1 k = e 1 k j=1 = e k=1 k j = e k k j , (1.3.2) P P P j=1 P j=1 P Y Y 16 1, p.17 § where Tn(z) are the Tchebychev polynomials, defined by Tn(cos θ) := cos nθ; in partic- ular T1(z)= z. iθ iθ For M SO(2n + 1), the eigenvalues are given by 1, e j and e− j , 1 j n, ∈ ≤ ≤ which is responsible for the extra-exponential e ti appearing in (1.3.2). Before listing various integrals, define the JacobiP weight ρ (z)dz := (1 z)α(1 + z)βdz, (1.3.3) αβ − and the formal sum ∞ g(z) := 2 tiTi(z). 1 X The arguments above lead to the following integrals, originally due to H. Weyl [75], and in its present form, due to Johansson [40]; besides the integrals over SO(k) = O+(k), the integrals over O (k) and U(n) will also be of interest in the theory of random − permutations: n ∞ i 1 titrM 2 g(zk) e dM = ∆n(z) e ρ( 1 , 1 )(zk)dzk n 2 2 O(2n)+ P [ 1,1] − − Z Z − k=1 Y n ∞ i ∞ 1 titrM 1 ti 2 g(zk) e dM = e ∆n(z) e ρ( 1 , 1 )(zk)dzk n 2 2 O(2n+1)+ P P [ 1,1] − Z Z − k=1 n Y ∞ i titrM 2 g(zk) e 1 dM = ∆n(z) e ρ 1 1 (zk)dzk. ( 2 , 2 ) Sp(n) P [ 1,1]n Z Z − k=1 Y n 1 ∞ i ∞ − 1 titrM 1 2t2i 2 g(zk) e dM = e ∆n 1(z) e ρ( 1 , 1 )(zk)dzk n−1 − 2 2 O(2n)− P P [ 1,1] Z Z − k=1 nY ∞ i ∞ i 1 titrM 1 ( 1) ti 2 g(zk) e dM = e − ∆n(z) e ρ( 1 , 1 )(zk)dzk n 2 2 O(2n+1)− P P [ 1,1] − Z Z − Yk=1 n ∞ i i ∞ i −i tr(tiM siM¯ ) 1 2 (tiz siz ) dzk e 1 − dM = ∆n(z) e 1 k− k U(n) P n! (S1)n | | P 2πizk Z Z Yk=1 (1.3.4) 1.4 Permutations and integrals over groups 0 Let Sn be the group of permutations π and S2n the subset of fixed-point free involutions 0 0 2 0 π (i.e., (π ) = I and π (k) = k for 1 k 2n ). Put the uniform distribution on Sn 0 6 ≤ ≤ and S2n ; i.e., all permutations or involutions have equal probability: 2nn! P (π )=1/n! and P (π0 )= ; (1.4.1) n 2n (2n)! 17 1, p.18 § 0 0 πn refers to a permutation in Sn and π2n to an involution in S2n. An increasing subsequence of π S or S0 is a sequence 1 j < ... < j n, such ∈ n n ≤ 1 k ≤ that π(j1) < ... < π(jk). Define L(πn) = length of the longest increasing subsequence of πn . (1.4.2) Example: for π = (3, 1, 4, 2, 6, 7, 5), we have L(π7) = 4. Around 1960 and based on Monte-Carlo methods, Ulam [70] conjectured that E(L ) lim n = c exists. n →∞ √n An argument of Erd¨os & Szekeres [28], dating back from 1935 showed that E(Ln) 1 ≥ 2 √n 1, and thus c 1/2. In ’72, Hammersley [33] showed rigorously that the limit exists.− Logan and Shepp≥ [46] showed the limit c 2, and finally Vershik and Kerov ≥ [74] that c = 2. In 1990, I. Gessel [31] showed that the following generating function is the determinant of a Toeplitz matrix: ∞ n 2π t 2√t cos θ i(k m)θ P (L ℓ) = det e e − dθ . (1.4.3) n! n ≤ n=0 0 0 k,m ℓ 1 X Z ≤ ≤ − The next major contribution was due to Johansson [41] and Baik-Deift-Johansson [17], who prove that for arbitrary x R, we have a ”law of large numbers” and a ”central limit theorem”, where F (x) is the∈ statistics (1.2.10), L L 2√n lim n =1, and P n − x F (x), for n . n 2√n n1/6 ≤ −→ −→ ∞ →∞ A next set of ideas is due to Diaconis & Shashahani [26], Rains [57, 58], Baik & Rains [18]. For a nice state-of-the-art account, see Aldous & Diaconis [14]. An illustration is contained in the following proposition; the first statement is essentially Gessel’s and the next statement is due to [26, 58, 18]. Proposition 1.1 The following holds ∞ tn P (L(π ) ℓ) = e√t Tr(M+M¯ )dM (1.4.4) n! n ≤ n=0 U(ℓ) X Z dθk = eiθj eiθk 2 e2√t cos θk . ℓ | − | 2π [0,2π] 1 j ∞ (t2/2)n P (L(π0 ) ℓ)= et Tr M dM. (1.4.5) n! 2n ≤ n=0 O(ℓ) X Z 18 1, p.19 § The proof of this statement will be sketched later. The connection with integrable systems goes via the following chain of ideas: Combinatorics ↓ Robinson-Schensted-Knuth correspondence ↓ Theory of symmetric polynomials ↓ Integrals over classical groups ↓ Integrable systems All the arrows, but the last one, will be explained in this section; the last arrow will be discussed in sections 7 and 8. We briefly sketch a few of the basic well known facts going into these arguments. They can be found in MacDonald [47], Knuth [45], Aldous- Diaconis [14]. Useful facts on symmetric functions, applicable to integrable theory, can be found in the appendix to [1]. Let me mention a few of these facts: A Young diagram λ is a finite sequence of non-increasing, non-negative integers • λ λ ... λ 0; also called a partition of n = λ := λ + ... + λ , with 1 ≥ 2 ≥ ≥ ℓ ≥ | | 1 ℓ λ being the weight. It can be represented by a diagram, having λ boxes in the | | 1 first row, λ2 boxes in the second row, etc..., all aligned to the left. A dual Young ˆ ˆ ˆ diagram λ = (λ1 λ2 ...) is the diagram obtained by flipping the diagram λ about its diagonal.≥ ≥ A Young tableau of shape λ is an array of positive integers a (at place (i, j) in • ij the Young diagram) placed in the Young diagram λ, which are non-decreasing from left to right and strictly increasing from top to bottom. A standard Young tableau of shape λ is an array of integers 1, ..., n placed in the • Young diagram, which are strictly increasing from left to right and from top to bottom. The number of Young tableaux of a given shape λ =(λ ... λ ) is 1 ≥ ≥ m 19 1, p.20 § 8 given by a number of formulae (for the Schur polynomial sλ, see below) f λ = # standard tableaux of shape λ { } = coefficient of x1x2 ...xn in sλ(x) λ ! 1 = | | = λ ! det (1.4.6) hλ | | (λ i + j)! all i,j ij i − m 1 = Qλ ! (h h ) , with h := λ i + m, m := λˆ . | | i − j h ! i i − 1 1 i sλ(x1, x2, ...) := xaij . (1.4.7) a tableaux of λ ij { ij } X Y The linear space Λn of symmetric polynomials in x1, ..., xn with rational coefficients • comes equipped with the inner product n 1 2 dzk f,g = f(z1, ..., zn)g(¯z1, ..., z¯n) zk zℓ h i n! n | − | 2πizk (S1) 1 k<ℓ n 1 Z ≤Y≤ Y = f(M)g(M¯ )dM. (1.4.8) ZU(n) An orthonormal basis of the space Λ is given by the Schur polynomials s (x , ..., x ), • n λ 1 n in which the numbers aij are restricted to 1, ..., n. Therefore, each symmetric func- tion admits a “Fourier series” f(x , ..., x )= f,s s (x , ..., x ), with s ,s ′ = δ ′ . (1.4.9) 1 n h λi λ 1 n h λ λ i λλ λ Xwith λˆ1 n ≤ In particular, one proves (see (1.4.6) for the definition of f λ) k λ (x1 + ... + xn) = f sλ, (1.4.10) |λ|=k ˆ λX1≤n If λ =(λ ... λ > 0), with9 λˆ = ℓ > n, then obviously s = 0. 1 ≥ ≥ ℓ 1 λ 8 λ ˆ hij := λi + λj i j + 1 is the hook length of the i, jth box in the Young diagram; i.e., the length of the hook formed− by− drawing a horizontal line emanating from the center of the box to the right and a vertical line emanating from the center of the box to the bottom of the diagram. 9 Remember, from the definition of the dual Young diagram, that λˆ1 = the length of the first column of λ 20 1, p.21 § Robinson-Schensted-Knuth correspondence: There is a 1-1 correspondence • (P, Q), two standard Young S tableaux from 1,... ,n, where n −→ P and Q have the same shape Given a permutation i1, ..., in, the correspondence constructs two standard Young tableaux P, Q having the same shape λ. This construction is inductive. Namely, having obtained two equally shaped Young diagrams Pk, Qk from i1, ..., ik, with the numbers (i1, ..., ik) in the boxes of Pk and the numbers (1, ..., k) in the boxes of Qk, one creates a new diagram Qk+1, by putting the next number ik+1 in the first row of P , according to the following rule: (i) if i all numbers appearing in the first row of P , then one creates a new k+1 ≥ k box with ik+1 in that box to the right of the first column, (ii) if not, place ik+1 in the box (of the first row) with the smallest number higher than ik+1. That number then gets pushed down to the second row of Pk according to the rule (i) or (ii), as if the first row had been removed. The diagram Q is a bookkeeping device; namely, add a box (with the number k + 1 in it) to Qk exactly at the place, where the new box has been added to Pk. This produces a new diagram Qk+1 of same shape as Pk+1. The inverse of this map is constructed essentially by reversing the steps above. Example: π = (5, 1, 4, 3, 2) S , ∈ 5 5 1 14 13 12 55 4 3 5 4 5 1 1 13 13 13 22 2 2 4 4 5 1 2 1 3 3 2 Hence π (P (π), Q(π)) = , −→ 4 4 5 5 and so L5(π) = 2 = #columns ofPor Q. The Robinson-Schensted-Knuth correspondence has the following properties 1 π (P, Q), then π− (Q, P ) • 7→ 7→ 21 1, p.22 § length (longest increasing subsequence of π) = # (columns in P ) • length (longest decreasing subsequence of π) = # (rows in P ) • π2 = I, then π (P,P ) • 7→ π2 = I, with k fixed points, then P has exactly k columns of odd length. • (1.4.11) From representation theory (see Weyl [75] and especially Rains [57]), one proves: Lemma 1.2 The following perpendicularity relations hold: (i) s (M)s (M¯ )dM = s ,s = δ λ µ h λ µi λµ ZU(n) (ii) s (M)dM = 1 for λ =(λ ... λ 0), k n, λ even λ 1 ≥ ≥ k ≥ ≤ i ZO(n) = 0 otherwise. (iii) s (M)dM = 1 for λˆ even, λˆ 2n , λ i 1 ≤ ZSp(n) = 0 otherwise. (1.4.12) Proof of Proposition 1.1 : On the one hand, (x + ... + x )k, (x + ... + x )k h 1 n 1 n i = f λf µ s ,s h λ µi |λ|=|µ|=k ˆ λ1X,µˆ1≤n = (f λ)2 |λ|=k ˆ λX1≤n = (f λ)2 |λ|=k λX1≤n = # (P, Q), standard Young tableaux, each of arbitrary shape λ { with λ = k, λ n | | 1 ≤ } = # π S such that L(π ) n . (1.4.13) { k ∈ k k ≤ } On the other hand, notice that, upon setting θ = θ′ + θ for 2 j n, the expres- j j 1 ≤ ≤ sion eiθj eiθk 2 is independent of θ . Then, setting z = eiθk , one computes: | − | 1 k 1 j 22 1, p.23 § (x + ... + x )k, (x + ... + x )ℓ h 1 n 1 n i 1 k ℓ iθj iθk 2 = (z1 + ... + zn) (¯z1 + ... +¯zn) e e dθ1 ...dθn n! n | − | [0,2π] 1 j (tr(M + M¯ ))kdM ZU(n) k j k j = ((trM) (trM) − )dM j 0 j k U(n) ≤X≤ Z 0, if k is odd (because then j = k j for all 0 j k) 6 − ≤ ≤ = k (1.4.15) trM kdM, if k is even. k/2 U(n) | | R Combining the three identities (1.4.13), (1.4.14) and (1.4.15) leads to 1 2k − # π S such that L(π ) n = (Tr(M + M¯ ))2kdM. (1.4.16) { k ∈ k k ≤ } k ZU(n) Finally 23 1, p.24 § ∞ tn P (L(π ) ℓ) n! n ≤ n=0 X n ∞ t # πn Sn L(πn) ℓ = { ∈ ≤ } n! n! n=0 X 1 ∞ tn 2n − = (tr(M + M¯ ))2ndM (n!)2 n n=0 U(ℓ) X Z ∞ (√t)2n = (tr(M + M¯ ))2ndM (2n)! n=0 U(ℓ) X Z = e√t Tr(M+M¯ )dM ZU(ℓ) 1 √t(z +z−1+...+z +z−1) iθ iθ 2 dθk = e 1 1 ℓ ℓ e j e k , ℓ! ℓ | − | 2π [0,2π] 1 j 1 ℓ dz √t(zk+¯zk) k = ∆ℓ(z)∆ℓ(¯z) e ℓ! (S1)ℓ 2πizk Z Yk=1 1 ℓ dz k 1 m 1 √t(zk+¯zk) k = det zσ(−m)z¯σ(−m) e ℓ! (S1)ℓ 1 k,m ℓ 2πizk σ Sℓ ≤ ≤ k=1 Z X∈ Y 1 dz k 1 m 1 √t(zk+¯zk) k = det zk− z¯k − e ℓ! 1 2πizk σ S S 1 k,m ℓ ∈ ℓ Z ≤ ≤ X 2π 2√t cos θ i(k m)θ = det e e − dθ , 0 1 k,m ℓ Z ≤ ≤ confirming Gessel’s result (1.4.3). The proof of the second relation (1.4.5) of Proposition 1.1 is based on the following computation: 24 1, p.25 § k λ (Tr M) dM = f sλ(M)dM, using (1.4.10), O(n) |λ|=k O(n) Z ˆ Z λX1≤n = f λ , using Lemma 1.2, λ =k | | λˆ1 n λi Peven≤ = f λ , using duality, λ =k |λ1| n ≤ λˆi Peven (P,P ),P standard Young tableau of shape λ = # with λ = k, λ n, λˆ even | | 1 ≤ i = # π0 S0, no fixed points and L(π0) n . k ∈ k k ≤ (1.4.17) In the last equality, we have used property (1.4.11): an involution has no fixed points iff all columns of P have even length. Since all columns λˆi have even length, it follows that λ = k is even and then only is (Tr M)kdM > 0; otherwise = 0 . Finally, one | | O(n) computes, R ∞ (t2/2)k P L(π0 ) n, π0 S0 k! 2k ≤ 2k ∈ 2k k=0 X ∞ t2k 2kk! = # π0 S0 , L(π0 ) n , using (1.4.1), 2kk! (2k)! { 2k ∈ 2k 2k ≤ } Xk=0 ∞ tk = # π0 S0, L(π0) n k! { k ∈ k k ≤ } Xk=0 ∞ tk = (Tr M)kdM , using (1.4.17), k! O(n) Xk=0 Z = et Tr M dM, ZO(n) ending the proof of Proposition 1.1. 25 2, p.26 § 2 Integrals, vertex operators and Virasoro relations In section 1, we discussed random matrix problems over different finite and infinite ma- trix ensembles, generating functions for the statistics of the length of longest increasing sequences in random permutations and involutions. One can also consider two Hermi- tian random matrix ensembles, coupled together. All those problems lead to matrix integrals or Fredholm determinants, which we list here: (β =2, 1, 4) n T rV (M) β e− dM = cn ∆n(z) ρ(zk)dzk • n | | n(E), n(E) or n(E) E 1 ZH S T Z Y 1 2 2 Tr(M +M 2cM1M2) dM1dM2 e− 2 1 2 − 2 • (E1 E2) Z ZHn × ex Tr M dM • ZO(n) e√x Tr(M+M¯ )dM • ZU(n) det I λK(y, z)I (z) with K(y, z) as in (1.2.4). • − E (2.0.1) The point is that each of these quantities admit a natural deformation, by inserting time variables t1, t2, ... and possibly a second set s1,s2, ..., in a seemingly ad hoc way. Each of these integrals or Fredholm determinant is then a fixed point for a natural vertex operator, which generates a Virasoro-like algebra. These new integrals in t1, t2, ... are all annihilated by the precise subalgebra of the Virasoro generators, which annihilates τ0. This will be the topic of this section. 2.1 Beta-integrals 2.1.1 Virasoro constraints for Beta-integrals V (z) Consider weights of the form ρ(z)dz := e− dz on an interval F = [A, B] R, with ⊆ rational logarithmic derivative and subjected to the following boundary conditions: i ρ′ g 0∞ biz k = V ′ = = , lim f(z)ρ(z)z = 0 for all k 0, (2.1.1) − ρ f ∞ a zi z A,B ≥ P0 i → and a disjoint union of intervals,P 2r E = [c2i 1,c2i] F R. (2.1.2) − ⊂ ⊂ 1 [ 26 2, p.27 § These data define an algebra of differential operators 2r ∂ = ck+1f(c ) . (2.1.3) Bk i i ∂c 1 i X Take the first type of integrals in the list (2.0.1) for general β > 0, thus general- izing the integrals, appearing in the probabilities (1.1.9),(1.1.11) and (1.1.18). Con- sider t-deformations of such integrals, for general (fixed) β > 0: (t := (t1, t2, ...), c =(c1,c2, ..., c2r) and z =(z1, ..., zn)) n ∞ i β tiz In(t, c; β) := ∆n(z) e 1 k ρ(zk)dzk for n> 0. (2.1.4) En | | P Z kY=1 The main statement of this section is Theorem 2.1, whose proof will be outlined in the next subsection. The central charge (2.1.9) has already appeared in the work of Awata et al. [16]. Theorem 2.1 (Adler-van Moerbeke [3, 6]) The multiple integrals n ∞ i β tiz In(t, c; β) := ∆n(z) e 1 k ρ(zk)dzk , for n> 0 (2.1.5) En | | P Z kY=1 and n ∞ i 4 4/β tiz In(t, c; ) := ∆n(z) e 1 k ρ(zk)dzk , for n> 0, (2.1.6) β En | | P Z kY=1 with I =1, satisfy respectively the following Virasoro constraints10 for all k 1: 0 ≥ − + a βJ(2) (t, n) b βJ(1) (t, n) I (t, c; β)=0, −Bk i k+i,n − i k+i+1,n n i 0 ! X≥ βt 2n βb βt 2n 4 + a βJ(2) , + i βJ(1) , I (t, c; )=0, −Bk i k+i,n − 2 − β 2 k+i+1,n − 2 − β n β i 0 ! X≥ (2.1.7) in terms of the coefficients ai, bi of the rational function ( log ρ)′ and the end points ci Z − βJ(2) βJ(1) of the subset E, as in (2.1.1) to (2.1.2). For all n , the k,n(t, n) and k,n(t, n) form a Virasoro and a Heisenberg algebra respectively,∈ interacting as follows k3 k βJ(2) βJ(2) βJ(2) k,n, ℓ,n = (k ℓ) k+ℓ,n + c − δk, ℓ − 12 − h i βJ(2) βJ(1) βJ(1) k,n, ℓ,n = ℓ k+ℓ,n + c′k(k + 1)δk, ℓ. − − h i k βJ(1) βJ(1) k,n, ℓ,n = δk, ℓ, (2.1.8) β − 10 When E equalsh the whole rangei F , then the ’s are absent in the formulae (2.1.7). Bk 27 2, p.28 § with central charge 1/2 1/2 2 β β − 1 1 c =1 6 and c′ = . (2.1.9) − 2 − 2 β − 2 ! βJ(2) Remark 1: The k,n’s are defined as follows: β β βJ(2) = : βJ(1) βJ(1) :+ 1 (k + 1) βJ(1) kJ(0) . (2.1.10) k,n 2 i,n j,n − 2 k,n − k,n i+Xj=k Componentwise, we have βJ(1) β (1) (0) βJ(0) (0) k,n(t, n)= Jk + nJk and k,n = nJk = nδ0k and βJ(2) k,n(t, n) β β β = βJ (2) + nβ +(k + 1)(1 ) βJ (1) + n (n 1) +1 J (0), 2 k − 2 k − 2 k where β (1) ∂ 1 Jk = + ( k)t k (2.1.11) ∂tk β − − 2 β (2) ∂ 2 ∂ 1 Jk = + iti + 2 itijtj. ∂ti∂tj β ∂tj β i+j=k i+j=k i j=k X −X −X− βJ(2) We put n explicitly in ℓ,n(t, n) to indicate the nth component contains n explicitly, besides t. Note that for β = 2, (2.1.10) becomes particularly simple: βJ(2) 2J(1) 2J(1) k,n = : i,n j,n : . β=2 i+j=k X Remark 2: The Heisenberg and Virasoro generators satisfy the following duality prop- erties: 4 (2) (2) βt 2n β J t, n = βJ , , n Z ℓ,n ℓ,n − 2 − β ∈ 4 (1) β β (1) βt 2n β J t, n = J , , n> 0. (2.1.12) ℓ,n − 2 ℓ,n − 2 − β βJ(2) In (2.1.7), ℓ,n ( βt/2, 2n/β) means that the variable n, which appears in the nth component, gets− replaced− by 2n/β and t by βt/2. − Remark 3: Theorem 2.1 states that the integrals (2.1.5) and (2.1.6) satisfy two sets of differential equations (2.1.7) respectively. Of course, the second integral also satisfies the first set of equations, with β replaced by 4/β. 28 2, p.29 § 2.1.2 Proof: β-integrals as fixed points of vertex operators Theorem 2.1 can be established by using the invariance of the integral under the trans- k+1 formation zi zi +εf(zi)zi of the integration variables. However, the most transpar- ent way to prove7→ Theorem 2.1 is via vector vertex operators, for which the β-integrals are fixed points. This is a technique which has been used by us, already in [2]. In- deed, define the (vector) vertex operator, for t = (t , t , ...) C∞, u C, and setting 1 2 ∈ ∈ χ(z) := (1,z,z2, ...). −i ∞ i ∞ u ∂ 1 tiu β 1 i ∂t β Xβ(t, u)=Λ− e 1 e− i χ( u ), (2.1.13) P P | | 11 which acts on vectors f(t)=(f0(t), f1(t), ...) of functions, as follows ∞ i n 1 X 1 tiu β − 1 β(t, u)f(t) n = e u fn 1(t β[u− ]). P | | − − For the sake of these arguments, it is convenient to introduce the following vector βJ(i) βJ(i) Virasoro generators: k (t) := ( k,n(t, n))n Z. ∈ k ∂ k+1 Proposition 2.2 The multiplication operator z and the differential operators ∂z z with z C∗, acting on the vertex operator X (t, z), have realizations as commutators, ∈ β βJ(1) βJ(2) in terms of the Heisenberg and Virasoro generators k (t, n) and k (t, n) : kX βJ(1) X z β(t, z) = k (t), β(t, z) ∂ h i zk+1X (t, z) = βJ(2)(t), X (t, z) . (2.1.14) ∂z β k β h i Corollary 2.3 Given a weight ρ(z)dz on R satisfying (2.1.1), we have ∂ zk+1f(z)X (t, z)ρ(z)= a βJ(2) (t) b βJ(1) (t) , X (t, z)ρ(z) . (2.1.15) ∂z β i k+i − i k+i+1 β " i 0 # X≥ Proof: Using (2.1.13) in the last line, compute ∂ zk+1f(z)X (t, z)ρ(z) ∂z β ρ (z) ∂ = ′ f(z) zk+1X (t, z)ρ(z)+ ρ(z) zk+1f(z)X (t, z) ρ(z) β ∂z β ∞ ∂ ∞ = b zk+i+1X (t, z) ρ(z)+ ρ(z) a zk+i+1X (t, z) − i β ∂z i β 0 ! 0 ! X X ∞ ∞ = b βJ(1) , X (t, z)ρ(z) + a βJ(2) , X (t, z)ρ(z) − i k+i+1 β i k+i β " 0 # " 0 # X X (2.1.16) 2 3 11For α C, define [α] := (α, α , α , ...) C∞. The operator Λ is the shift matrix, with zeroes ∈ 2 3 ∈ everywhere, except for 1’s just above the diagonal, i.e., (Λv)n = vn+1. 29 2, p.30 § establishing (2.1.14). Giving the weight ρE(u)du = ρ(u)IE(u)du, with ρ and E as before, define the integrated vector vertex operator Yβ(t, ρE) := duρ(u)Xβ(t, u). (2.1.17) ZE and the vector operator := Dk Bk − Vk 2r ∂ := ck+1f(c ) a βJ(2) (t) b βJ(1) (t) , (2.1.18) i i ∂c − i k+i − i k+i+1 1 i i 0 X X≥ consisting of a c-dependent boundary part and a (t, n)-dependent Virasoro part . Bk Vk Proposition 2.4 The following commutation relation holds: [ , Y (t, ρ )] = 0 (2.1.19) Dk β E Proof: Integrating both sides of (2.1.14) over E, one computes: ∂ 2r dz zk+1f(z)X (t, z)ρ(z) = ( 1)ick+1f(c )X (t, c )ρ(c ) ∂z β − i i β i i ZE 1 X 2r ∂ = ck+1f(c ) X (t, z)ρ(z)dz i i ∂c β 1 i E X Z = [ , Y (t, ρ )] , (2.1.20) Bk β E while on the other hand dz a βJ(2) b βJ(1) , X (t, z)ρ(z) i k+i − i k+i+1 β E " i 0 # Z X≥ βJ(2) βJ(1) X = ai k+i bi k+i+1 , dzρE (z) β(t, z) − R " i 0 # X≥ Z = [ , Y (t, ρ )] . (2.1.21) Vk β E Subtracting both expressions (2.1.19) and (2.1.20) yields 0 = [ , Y (t, ρ )] = [ , Y (t, ρ )] , Bk −Vk β E Dk β E concluding the proof of proposition 2.4. 30 2, p.31 § Proposition 2.5 The column vector, n ∞ i β tiz I(t) := ∆n(z) e 1 k ρ(zk)dzk En | | P Z k=1 !n 0 Y ≥ is a fixed point for the vertex operator Yβ(t, ρE): (see definition (2.1.17)) (Y (t, ρ )I) = I , n 1. (2.1.22) β E n n ≥ Proof: We have n ∞ i β tiz In(t) = ∆n(z) e 1 k IE(zk)ρ(zk)dzk Rn | | P Z kY=1 ∞ i tiu β(n 1) = duρE(u)e 1 u − R | | Z P n 1 n 1 − β − ∞ i zk β tiz 1 ∆n 1(z) e 1 k ρE(zk)dzk Rn−1 − u | − | P Z kY=1 Yk=1 ∞ i tiu β(n 1) = duρE(u)e 1 u − R | | Z P −i n 1 ∞ u ∂ − ∞ i β 1 i ∂t β tiz e− i ∆n 1(z) e 1 k ρE(zk)dzk P Rn−1 | − | Z k=1 P Y−i ∞ i ∞ u ∂ β(n 1) tiu β 1 i ∂t = duρE(u) u − e 1 e− i τn 1(t) R | | P − Z P = Yβ(t, ρE)I(t) . (2.1.23) n Proof of Theorem 2.1: From proposition 2.4 it follows that for n 1, ≥ 0 = [ , (Y (t, ρ ))n] I Dk β E = Y (t, ρ )nI Y (t, ρ )n I. (2.1.24) Dk β E − β E Dk −i i u ∂ tiu β i ∂t Taking the nth component for n 1 and k 1, and setting Xβ(t, u)= e e− i , ≥ ≥ − P and using (2.1.21) P 0 = ( I Y (t, ρ )n I) Dk − β E Dk n β n 1 = ( I) duρ (u)X (t; u)( u ) − ... duρ (u)X (t; u)( I) Dk n − E β | | E β Dk 0 Z Z = ( I) . Dk n 31 2, p.32 § β (2) β (1) (0) Indeed ( kI)0 = 0 for k 1, since τ0 = 1 and k involves Jk , Jk and Jk for k 1: D ≥ − D ≥ − β (2) Jk is pure differentiation for k 1; β (1) ≥ − Jk is pure differentiation, except for k = 1; but β (1) − J 1 appears with coefficient nβ, which vanishes for n = 0; (0)− J appears with coefficient n((n 1) β + 1), vanishing for n = 0. k − 2 2.1.3 Examples Example 1 : Gaussian β-integrals The weight and the ai and bi, as in (2.1.1), are given by V (z) z2 ρ(z)= e− = e− , V ′ = g/f =2z, a0 =1, b0 =0, b1 =2, and all other ai, bi =0. From (2.1.5), the integrals n 2 ∞ i β z + tiz In = ∆n(z) e− k i=1 k dzk (2.1.25) En | | P Z kY=1 satisfy the Virasoro constraints, for k 1, ≥ − 2r ∂ I = ck+2 I = βJ(2) a βJ(1) + βJ(1) I . (2.1.26) −Bk n − i ∂c n − k+1,n − k+1,n k+2,n n 1 i X Introducing the following notation i +1 i +1 σ =(n )β + i +1 b =(n )β + i +1, (2.1.27) i − 2 − 0 − 2 the first three constraints have the following form, upon setting Fn = log In, ∂ ∂ ∂ ∂ n −1F = 2 iti F nt1 , 0F = 2 iti F σ1 −B ∂t − ∂t − − −B ∂t − ∂t − 2 1 ≥ i 1 2 ≥ i Xi 2 Xi 1 ∂ ∂ ∂ F = 2 σ it F. (2.1.28) −B1 ∂t − 1 ∂t − i ∂t 3 1 ≥ i+1 Xi 1 For later use, take the linear combinations 1 1 1 σ1 1 = 1, 2 = 0, 3 = 1 + 1 , (2.1.29) D −2B− D −2B D −2 B 2 B− ∂F such that each i contains the pure term ∂F/∂ti, i.e., iF = + ... D D ∂ti 32 2, p.33 § Example 2 (Laguerre β-integrals) Here, the weight and the ai and bi, as in (2.1.1), are given by V a z g z a e− = z e− , V ′ = = − , f z a =0, a =1, b = a, b =1, and all other a , b =0. 0 1 0 − 1 i i Thus from (2.1.4), the integrals n ∞ i β a zk+ i=1 tizk In = ∆n(z) zk e− dzk (2.1.30) En | | P Z kY=1 satisfy the Virasoro constraints, for k 1, ≥ − 2r ∂ I = ck+2 I = βJ(2) a βJ(1) + βJ(1) I . (2.1.31) −Bk n − i ∂c n − k+1,n − k+1,n k+2,n n 1 i X Introducing the following notation i +1 i +1 σ = =(n )β + i +1 b =(n )β + i +1+ a, i − 2 − 0 − 2 the first three have the form, upon setting F = Fn = log In, ∂ ∂ n − F = it F (σ + a) −B 1 ∂t − i ∂t − 2 1 1 ≥ i Xi 1 ∂ ∂ ∂ F = σ it F −B0 ∂t − 1 ∂t − i ∂t 2 1 ≥ i+1 Xi 1 ∂ ∂ ∂ β ∂2 β ∂F 2 F = σ it F . −B1 ∂t − 2 ∂t − i ∂t − 2 ∂t2 − 2 ∂t 3 2 ≥ i+2 1 1 Xi 1 ∂F Again, replace the operators i by linear combinations i, such that iF = + ..., B D D ∂ti 1 = 1, 2 = 0 σ1 1, 3 = 1 σ2 0 σ1σ2 1. (2.1.32) D −B− D −B − B− D −B − B − B− Example 3 (Jacobi β-integral) This case is particularly important, because it covers not only the first integral, but also the second integral in the list (2.0.1), used in the problem of random permutations. The weight and the ai and bi, as in (2.1.1), are given by V a b g a b +(a + b)z ρ(z) := e− = (1 z) (1 + z) ,V ′ = = − − f 1 z2 − 33 2, p.34 § a =1, a =0, a = 1, b = a b, b = a + b, and all other a , b =0. 0 1 2 − 0 − 1 i i The integrals n ∞ i β a b tiz In = ∆n(z) (1 zk) (1 + zk) e i=1 k dzk (2.1.33) En | | − P Z kY=1 satisfy the Virasoro constraints (k 1): ≥ − 2r ∂ I = ck+1(1 c2) I − Bk n − i − i ∂c n 1 i X = βJ(2) βJ(2) + b βJ(1) + b βJ(1) I . (2.1.34) k+2,n − k,n 0 k+1,n 1 k+2,n n Introducing σ =(n i+1 )β + i +1+ b , the first four have the form (k = 1, 0, 1, 2) i − 2 1 − ∂ ∂ ∂ −1F = σ1 + iti iti F + n(b0 t1) −B ∂t ∂t − ∂t − − 1 ≥ i+1 ≥ i 1 Xi 1 Xi 2 ∂ ∂ ∂ ∂ β ∂2 β ∂F 2 n F = σ + b + it ( )+ F + (σ b ) −B0 2 ∂t 0 ∂t i ∂t − ∂t 2 ∂t2 2 ∂t − 2 1 − 1 2 1 ≥ i+2 i 1 1 Xi 1 ∂ ∂ ∂ ∂ ∂ ∂2 F = σ + b (σ b ) + it ( )+ β F −B1 3 ∂t 0 ∂t − 1 − 1 ∂t i ∂t − ∂t ∂t ∂t 3 2 1 ≥ i+3 i+1 1 2 Xi 1 ∂F ∂F +β ∂t1 ∂t2 ∂ ∂ ∂ ∂ ∂ F = σ + b (σ b ) + it ( ) −B2 4 ∂t 0 ∂t − 2 − 1 ∂t i ∂t − ∂t 4 3 2 ≥ i+4 i+2 Xi 1 β ∂2 ∂2 ∂2 β ∂F ∂F ∂F ∂F + ( +2 ) F + ( )2 ( )2 +2 . 2 ∂t2 − ∂t2 ∂t ∂t 2 ∂t − ∂t ∂t ∂t 2 1 1 3 2 1 1 3 (2.1.35) 2.2 Double matrix integrals Consider now weights of the form i j ρ(x, y)= e i,j≥1 rij x y ρ(x)˜ρ(y), (2.2.1) P defined on a product of intervals F F R2, with rational logarithmic derivative 1 × 2 ⊂ i ˜ i ρ′ g i 0 bix ρ˜′ g˜ i 0 biy = = ≥ and = = ≥ , i ˜ i − ρ f i 0 aix − ρ˜ f i 0 a˜iy P ≥ P ≥ P 34 P 2, p.35 § satisfying lim f(x)ρ(x)xk = lim f˜(y)˜ρ(y)yk = 0 for all k 0. (2.2.2) x ∂F y ∂F˜ ≥ → → Consider subsets of the form r s 2 E = E1 E2 := i=1[c2i 1,c2i] i=1[˜c2i 1, c˜2i] F1 F2 R . (2.2.3) × ∪ − × ∪ − ⊂ × ⊂ A natural deformation of the second integral in the list (2.0.1) is given by the following integrals: n ∞ i i (tix siy ) In(t,s,r; E)= ∆n(x)∆n(y) e i=1 k− k ρ(xk,yk)dxkdyk (2.2.4) En P Z Z Yk=1 J(i) J˜(i) In the theorem below, k,n and k,n are vectors of operators, whose components are given by the operators (2.1.10) for β = 1; i.e., J(i) βJ(i) J˜(i) βJ(i) k,n(t)= k,n(t) , k,n(s) := k,n(t) ; β=1 β=1, t s 7→− thus, from (2.1.10) and (2.1.11), one finds: 1 J(2) (t)= J (2)(t)+(2n + k + 1)J (1)(t)+ n(n + 1)J (0) , (2.2.5) k,n 2 k k k satisfying the Heisenberg and Virasoro relations (2.1.8), with central charge c = 2 and − c′ =1/2. The ai, a˜i, bi, ˜bi,ci, c˜i,rij given by (2.2.1), (2.2.2) and (2.2.3) define differential oper- ators: 2r ∂ ∂ k+1 J(2) J(1) k,n : = ci f(ci) ai( k+i,n + mrmℓ ) bi k+i+1,n D ∂ci − ∂rm+k+i,ℓ − 1 i 0 m,ℓ 1 X X≥ X≥ 2r ∂ ∂ ˜ k+1 ˜ J˜(2) ˜ J˜(1) k,n : = c˜i f(˜ci) a˜i( k+i,n + ℓrmℓ ) bi k+i+1,n . D ∂c˜i − ∂rm,ℓ+k+i − 1 i 0 m,ℓ 1 X X≥ X≥ (2.2.6) Theorem 2.6 (Adler-van Moerbeke [3, 4]) Given ρ(x, y) as in (2.2.1), the integrals n ∞ i i (tix siy ) In(t,s,r; E) := ∆n(x)∆n(y) e i=1 k− k ρ(xk,yk)dxkdyk (2.2.7) En P ZZ Yk=1 satisfy two families of Virasoro equations for k 1: ≥ − I (t,s,r; E)=0 and ˜ I (t,s,r; E)=0. (2.2.8) Dk,n n Dk,n n 35 2, p.36 § The proof of this statement is very similar to the one for β-integrals. Namely, define the vector vertex operator, −i −i ∞ i i ∞ u ∂ v ∂ 1 (tiu siv ) 1 ( i ∂t i ∂s ) X12(t, s; u, v)=Λ− e 1 − e− i − i χ(uv), (2.2.9) P P J(i) which, as a consequence of Proposition 2.2, interacts with the operators k (t) = J(i) k,n(t, n) , as follows: n Z ∈ kX J(1) X u 12(t, s; u, v) = k (t), 12(t, s; u, v) ∂ h i uk+1X (t, s; u, v) = J(2)(t), X (t, s; u, v) . (2.2.10) ∂u 12 k 12 h i k ∂ k+1 k A similar statement can be made, upon replacing the operators u and ∂u u by v ∂ k+1 J˜(i) and ∂v v , and upon using the k (s)’s. Finally, one checks that the integral vertex operator Y(t, s; ρE) := dxdyρ(x, y)X12(t, s; x, y) (2.2.11) Z ZE commutes with the two vectors of differential operators k =( k,n)n Z, as in (2.2.6): D D ∈ , Y(t, s; ρ ) = ˜ , Y(t, s; ρ ) =0, Dk E Dk E h i h i and that the vector I =(I0 =1,I1, ...) of integrals (2.2.7) is a fixed point for Y(t, s; ρE), Y(t, s; ρE)I(t,s,r; E)= I(t,s,r; E). Then, as before, the proof of Theorem 2.6 hinges ultimately on the fact that Dk,0 annihilates I0 = 1. 2.3 Integrals over the unit circle We now deal with the fourth type of integral in the list (2.0.1), which we deform, this time, by inserting two sequences of times t1, t2, ... and s1,s2, .... The following theorem holds: Theorem 2.7 (Adler-van Moerbeke [7]) The multiple integrals over the unit circle S1, n ∞ i −i 2 (tiz siz ) dzk In(t, s)= ∆n(z) e 1 k− k , n> 0, (2.3.1) (S1)n | | P 2πizk Z Yk=1 36 2, p.37 § with I0 =1, satisfy an SL(2,Z)-algebra of Virasoro constraints: k = 1, θ =0 − θ I (t, s)=0, for k =0, θ arbitrary only, (2.3.2) Dk,n n k =1, θ =1 where the operators θ := θ (t, s, n), k Z, n 0 are given by Dk,n Dk,n ∈ ≥ θ J(2) J(2) J(1) J(1) k,n := k,n(t, n) k,n( s, n) k θ k,n(t, n)+(1 θ) k,n( s, n) , (2.3.3) D − − − − − − − J(i) βJ(i) with k,n(t, n) := k,n(t, n) , as in (2.1.11). β=1 The explicit expressions are ∂ ∂ ∂ − I = (i + 1)t (i 1)s − + n t + I =0 D 1 n i+1 ∂t − − i 1 ∂s 1 ∂s n ≥ i ≥ i 1 Xi 1 Xi 2 ∂ ∂ I = it is I =0 (2.3.4) D0 n i ∂t − i ∂s n ≥ i i Xi 1 ∂ ∂ ∂ I = (i + 1)s + (i 1)t − + n s + I =0. D1 n − i+1 ∂s − i 1 ∂t 1 ∂t n ≥ i ≥ i 1 Xi 1 Xi 2 Here the key vertex operator is a reduction of X12(t, s; u, v), defined in the pre- Z θ vious section (formula (2.2.9)). For all k , the vector of operators k(t, s) = θ (t, s, n) form a realization of the first∈ order differential operators Dd uk+1, us- Dk,n n Z du ing the vertex∈ operator X (t, s; u,u 1), namely 12 − d X (t, s; u,u 1) X (t, s; u,u 1) uk+1 12 − = θ(t, s), 12 − . (2.3.5) du u Dk u Indeed, d k 1 u u X (t, s; u,u− ) du 12 ∂ k+1 ∂ 1 k k k = u v − kθu k(1 θ)v− X (t, s; u, v) ∂u − ∂v − − − 12 v= u − J(2) J(2) J(1) J(1) X = k (t) k( s) k θ k (t)+(1 θ) k ( s) , 12(t, s; u, u) − − − − − − − θ 1 = h (t, s), X (t, s; u,u− . i Dk 12 The θ := θ(t, s) satisfy Virasoro relations with central charge = 0, Dk Dk θ, θ =(k ℓ) θ , (2.3.6) Dk Dℓ − Dk+ℓ 37 3, p.38 § and, from (2.3.5) the following commutation relation holds: θ du 1 k(t, s), Y(t, s) =0, with Y(t, s) := X12(t, s; u,u− ). (2.3.7) D 1 2πiu ZS The point is that the column vector I(t, s)=(I0,I1, ...) of integrals (2.3.1), is a fixed point for Y(t, s): (Y(t, s)I) = I , n 1, (2.3.8) n n ≥ which is shown in a way, similar to Proposition 2.5. Proof of Theorem 2.7: Here again the proof is similar to the one of Theorem 2.1. Taking the nth component and the nth power of Y(t, s), with n 1, and noticing from θ ≥ the explicit formulae (2.3.4) that k(t, s)I 0 = 0, we have, by means of a calculation similar to the proof of Theorem 2.1,D 0 = [ θ, Y(t, s)n]I Dk n = θY(t, s)nI Y(t, s)n θI Dk − Dk n = θI Y(t, s)n θI = θI . Dk − Dk n Dk n 3 Integrable systems and associated matrix inte- grals 3.1 Toda lattice and Hermitian matrix integrals 3.1.1 Toda lattice, factorization of symmetric matrices and orthogonal poly- nomials V (z) Given a weight ρ(z)= e− defined as in (2.1.1), the inner-product over E R, ⊆ i f,g = f(z)g(z)ρ (z)dz, with ρ (z) := e tiz ρ(z), (3.1.1) h it t t ZE P leads to a moment matrix i j mn(t)=(µij(t))0 i,j ∂µij ∂m k commuting = µi+k,j, and thus ∞ =Λ m . (3.1.3) ∂t ∂t ∞ vector fields k k 12 H¨ankel means µij depends on i + j only 38 3, p.39 § Another important ingredient is the factorization of m into a lower- times an upper- ∞ triangular matrix13 1 1 m (t)= S(t)− S(t)⊤− , S(t) = lower triangular with ∞ non-zero diagonal elements. The main ideas of the following theorem can be found in [2, 5]. Remember c = (c1, ..., c2r) denotes the boundary points of the set E. dM refers to properly normalized Haar measure on . Hn Theorem 3.1 The determinants of the moment matrices n 1 2 τn(t, c) := det mn(t, c) = ∆n(z) ρt(zk)dzk n! En Z kY=1 tr( V (M)+ ∞ t M i) = e − 1 i dM, (3.1.4) n(E) P ZH satisfy (i) Virasoro constraints (2.1.7) for β =2, 2r ∂ ck+1f(c ) + a J(2) b J(1) τ (t, c)=0. (3.1.5) − i i ∂c i k+i,n − i k+i+1,n n 1 i i 0 ! X X≥ (ii) The KP-hierarchy14(k =0, 1, 2,... ) ∂ 1 ∂ 1 ∂ 1 ∂2 s , , ,... τ τ =0, k+4 ∂t 2 ∂t 3 ∂t − 2 ∂t ∂t n ◦ n 1 2 3 1 k+3 of which the first equation reads: ∂ 4 ∂ 2 ∂2 ∂2 2 +3 4 log τ +6 log τ =0. (3.1.6) ∂t ∂t − ∂t ∂t n ∂t2 n 1 2 1 3 ! 1 (iii) The standard Toda lattice, i.e., the symmetric tridiagonal matrix 1/2 ∂ τ1 τ0τ2 log 2 0 ∂t1 τ0 τ1 1/2 1/2 τ0τ2 ∂ τ2 τ1τ3 1 τ 2 ∂t log τ τ 2 L(t) := S(t)ΛS(t)− = 1 1 1 2 (3.1.7) 1/2 τ1τ3 ∂ τ3 0 τ 2 ∂t log τ 2 1 2 . .. 13 This factorization is possible for those t’s for which τn(t) := det mn(t) = 0 for all n> 0. 14 6 Given a polynomial p(t1,t2, ...), define the customary Hirota symbol p(∂t)f g := ∞◦ i ∂ ∂ i p( , , ...)f(t + y)g(t y) . The s ’s are the elementary Schur polynomials e 1 t z := ∂y1 ∂y2 ℓ − y=0 P s (t)zi and s (∂˜) := s ( ∂ , 1 ∂ ,... ). i≥0 i ℓ ℓ ∂t 1 2 ∂t2 P 39 3, p.40 § satisfies the commuting equations15 ∂L 1 k = (L )s, L . (3.1.8) ∂t 2 k (iv) Eigenvectors of L: The tridiagonal matrix L admits two independent eigenvectors: p(t; z)=(pn(t; z))n 0 satisfying (L(t)p(t; z))n = zpn(t; z), n 0, where • ≥ ≥ pn(t; z) are nth degree polynomials in z, depending on t C∞, orthonormal with respect to the t-dependent inner product16 (3.1.1) ∈ p (t; z),p (t; z) = δ ; h k ℓ it kℓ they are eigenvectors of L, i.e., L(t)p(t; z)= zp(t; z), and enjoy the following 2 representations: (χ(z)=(1,z,z , ...)⊤) pn(t; z) := (S(t)χ(z))n 1 m (t) z 1 n = det . τn(t)τn+1(t) . n µn,0 ... µn,n 1 z p − 1 n 1/2 τn(t [z− ]) τn+1(t) = z hn− − , hn := (3.1.9) τn(t) τn(t) pn(t;u) q(t, z)=(qn(t; z))n 0, with qn(t; z) := z Rn ρt(u)du, satisfying (L(t)q(t; z))n = • ≥ z u zq (t; z), n 1; q (t; z) enjoys the following− representations: n ≥ n R pn(t; u) 1 1 ⊤− − qn(t; z) := z ρt(u)du = S (t)χ(z ) n Rn z u Z − 1 = S(t)m (t)χ(z− ) ∞ n 1 n 1/2 τn+1(t + [z− ]) = z− hn− . τn(t) (3.1.10) In the case β = 2, the Virasoro generators (2.1.11) take on a particularly elegant form, namely for n 0, ≥ J(2) J(1) J(1) (2) (1) 2 k,n(t) = : i,n(t) j,n(t): = Jk (t)+2nJk (t)+ n δ0k i+Xj=k J(1) (1) k,n(t) = Jk (t)+ nδ0k, 15 ()s means: take the skew-symmetric part of (); for more details, see subsection 3.1.2. 16The explicit dependence on the boundary points c will be omitted in this point (iv). 40 3, p.41 § with17 2 (1) ∂ 1 (2) ∂ ∂ 1 Jk = + ( k)t k , Jk = + iti + itijtj. (3.1.11) ∂tk 2 − − ∂ti∂tj ∂tj 4 i+j=k i+j=k i j=k X −X −X− Statement (i) is already contained in section 2, whereas statement (ii) will be established in subsection 3.1.2, using elementary methods. 3.1.2 Sketch of Proof Orthogonal polynomials and τ-function representation: The representation (3.1.4) of the determinants of moment matrices as integrals follows immediately from the fact that the square of a Vandermonde determinant can be represented as a sum of determinants 2 ℓ+k 2 ∆ (u1,...,un)= det uσ(k)− . 1 k,ℓ n σ Sn ≤ ≤ X∈ Indeed, n!τn(t) = n! det mn(t) ℓ+k 2 = det zσ(k)− ρt(zσ(k))dzσ(k) E 1 k,ℓ n σ Sn X∈ Z ≤ ≤ ℓ+k 2 = det zσ(k)− ρt(zσ(k))dzσ(k) En 1 k,ℓ n σ Sn ≤ ≤ ∈ Z X n 2 = ∆n(z) ρt(zk)dzk, En Z kY=1 whereas the representation (3.1.4) in terms of integrals over Hermitian matrices follows from section 1.1. The Borel factorization of m is responsible for the orthonormality of the polyno- ∞ mials pn(t; z)=(S(t)χ(z))n; indeed, pk(t; z),pℓ(t; z) 0 k,ℓ< = Sχ(z)(Sχ(z))⊤ρt(z)dz = Sm S⊤ = I. h i ≤ ∞ ∞ ZE Note that Sχ(z)(Sχ(z))⊤ should be viewed as a semi-infinite matrix obtained by multi- plying a semi-infinite column and row. The determinantal representation (3.1.9) follows k at once from noticing that pn(t; z), z = 0 for 0 k n 1, because taking that inner-product produces twoh identical columnsi in the≤ matrix≤ − thus obtained. From the 1 n 1/2 same representation (3.1.9), one has pn(t; z)= hn− z + ... , where hn := (τn+1/τn(t)) . 17 (1) The expression Jk = 0 for k = 0. 41 3, p.42 § The “Sato” representation (3.1.9) of pn(t; z) in terms of the determinant τn(t) of the moment matrix can be shown by first proving the Heine representation of the orthogonal polynomials, which goes as follows: hnpn(t; z) 1 1 mn(t) z = det . τn . n µn,0 ... µn,n 1 z − 0 1 n 1 u1 u2 un− 1 1 2 ··· n u1 u2 un z n 1 . ··· . . = det . . . ρt(ui)dui τ n n E n 1 n 2n 2 n 1 1 Z u1− u2 un − z − Y n n+1 ··· 2n 1 n u u u − z 1 2 ··· n 0 0 0 u1 u2 un 1 1 1 ··· 1 u1 u2 un z n 1 . ··· . . 0 1 n 1 = det . . . u1u2 un− ρt(ui)dui τ n ··· n E n 1 n 1 n 1 n 1 1 Z u1− u2− un− z − ··· Y un un un zn 1 2 ··· n 0 0 0 uσ(1) uσ(2) uσ(n) 1 1 1 ··· 1 uσ(1) uσ(2) uσ(n) z 1 . ··· . . = det . . . τn En n 1 n 1 n 1 n 1 Z u − u − u − z − σ(1) σ(2) ··· σ(n) un un un zn σ(1) σ(2) ··· σ(n) n 0 1 n 1 u u u − ρ (u )du , σ(1) σ(2) ··· σ(n) t σ(i) σ(i) 1 Y for any permutation σ Sn 0 0 0 ∈ u1 u2 un 1 1 1 ··· 1 u1 u2 un z 1 . ··· . . = det . . . τn En n 1 n 1 n 1 n 1 Z u1− u2− un− z − ··· un un un zn 1 2 ··· n n σ 0 1 n 1 ( 1) u u u − ρ (u )du − σ(1) σ(2) ··· σ(n) t i i 1 n Y 1 2 = ∆n(u) (z uk)ρt(uk)duk, upon summing over all σ. n!τn En − Z Yk=1 42 3, p.43 § Therefore, using again the representation of ∆2(z) as a sum of determinants, Heine’s formula leads to hnpn(t, z) n n z ℓ+k 2 uσ(k) = det uσ(k)− 1 ρt(uσ(k))duσ(k), n!τn En 1 k,ℓ n − z σ Sn ≤ ≤ k=1 Z ∈ n X Y z ℓ+k 2 1 ℓ+k 1 = det uσ(k)− uσ(k)− ρt(uσ(k))duσ(k) n!τn En − z 1 k,ℓ n σ Sn Z X∈ ≤ ≤ zn 1 = det µ µ τ ij − z i,j+1 n 0 i,j n 1 n ≤ ≤ − z 1 = det µij(t [z− ]) 0 i,j n 1 τn − ≤ ≤ − 1 n τn(t [z− ]) = z − , (3.1.12) τn(t) invoking the fact that −i ∞ z i u 1 i+j 1 ti i u i+j µij(t [z− ]) = u e − ρ(u)du = u 1 ρ(u)du − P − z Z Z 1 = µ µ . i+j − z i+j+1 Formula (3.1.10) follows from computing on the one hand S(t)m χ(z) using the ∞ explicit moments µij, together with (3.1.12), and on the other hand the equivalent 1 1 n k expression S⊤− (t)χ(z− ). Indeed, using (S(t)χ(z))n = pn(t; z)= 0 pnk(t)z , j j P (Sm ) z− = z− pnℓ(t)µℓj ∞ nj j 0 j 0 ℓ 0 X≥ X≥ X≥ j ℓ+j = z− pnℓ(t) u ρt(u)du j 0 ℓ 0 E X≥ X≥ Z u j = p (t)uℓ ρ (u)du nℓ z t E ℓ 0 j 0 Z X≥ X≥ p (t, u)ρ (u) = z n t du. z u ZE − Mimicking computation (3.1.12), one shows 1 1 j τn+1(t + [z− ]) n h S⊤− (t) z− = z− , n nj τ (t) j 0 n X≥ 43 3, p.44 § 1 from which (3.1.10) follows, upon using Sm = S⊤− . Details of this and subsequent ∞ derivation can be found in [5, 6]. 2 The vectors p and q are eigenvectors of L. Indeed, remembering χ(z)=(1,z,z , ...)⊤, and the shift (Λv)n = vn+1, we have 1 1 Λχ(z)= zχ(z) and Λ⊤χ(z− )= zχ(z− ) ze , with e = (1, 0, 0, ...)⊤. − 1 1 1 1 Therefore, p(z)= Sχ(z) and q(z)= S⊤− χ(z− ) are eigenvectors, in the sense 1 Lp = SΛS− Sχ(z)= zSχ(z)= zp 1 1 1 1 1 1 1 L⊤q = S⊤− Λ⊤S⊤S⊤− χ(z− )= zS⊤− χ(z− ) zS⊤− e = zq zS⊤− e . − 1 − 1 Then, using L = L⊤, one is lead to ((L zI)p) =0, for n 0 and ((L zI)q) =0, for n 1. − n ≥ − n ≥ Toda lattice and Lie algebra splitting: The Lie algebra splitting of semi-infinite matrices and the corresponding projections (used in (3.1.8)), denoted by ( )s and ( )b are defined as follows: s = skew-symmetric matrices gl( )= s b { } . ∞ ⊕ b = lower-triangular matrices { } Conjugating the shift matrix Λ by S(t) yields a matrix 1 L(t) = S(t)ΛS(t)− 1 1 = SΛS− S⊤− S⊤ = SΛm S⊤, using (3.1.3), ∞ = Sm Λ⊤S⊤, using Λm = m Λ⊤, ∞ ∞ ∞ 1 1 = S(S− S⊤− )Λ⊤S⊤, using again (3.1.3), 1 = (SΛS− )⊤ = L(t)⊤, which is symmetric and thus tridiagonal. Moreover, from (3.1.3) one computes k ∂m 0 = S Λ m ∞ S⊤ ∞ − ∂t k k 1 ∂ 1 1 = SΛ S− S (S− S⊤− )S⊤ − ∂tk k ∂S 1 1 ∂S⊤ = L + S− + S⊤− . ∂tk ∂tk Upon taking the () and ()0 parts of this equation (A means the lower-triangular part − − of the matrix A, including the diagonal and A0 the diagonal part) leads to k ∂S 1 1 ∂S⊤ ∂S 1 1 k (L ) + S− + S⊤− = 0 and S− = (L )0. − ∂t ∂t ∂t −2 k k 0 k 0 44 3, p.45 § Upon observing that for any symmetric matrix a c a 0 a c a c = =2 , c b b 2c b c b − c b 0 − it follows that the matrices L(t), S(t) and the vector p(t; z)=(pn(t; z))n 0 = S(t)χ(z) ≥ satisfy the (commuting) differential equations and the eigenvalue problem ∂S 1 k = (L )bS, L(t)p(t; z)= zp(t; z), (3.1.13) ∂tk −2 and thus ∂L 1 k ∂p 1 k = (L )b, L , = (L )bp. ∂t − 2 ∂t −2 k k (Standard Toda lattice) The bilinear identity: The functions τ (t) satisfy the following identity, for n n ≥ m +1, t, t′ C, where one integrates along a small circle about , ∈ ∞ ′ i 1 1 (ti t )z n m 1 τn(t [z− ])τm+1(t′ + [z− ])e − i z − − dz =0. (3.1.14) z= − P I ∞ An elementary proof can be given by expressing the left hand side of (3.1.14), in terms of pn(t; z) and pm(t, z), using (3.1.9) and (3.1.10). One uses below the following identity (see [2]) g(u) f(z)g(z)dz = f, du , (3.1.15) R R z u Z Z − ∞ 18 involving the residue pairing . So, modulo terms depending on t and t′ only, the left hand side of (3.1.14) equals ∞ ′ i ∞ ′ i n (ti t )z n m 1 m+1 pm(t′; u) t u dz z− pn(t; z)e 1 − i z − − z e 1 i ρ(u)du z= P R z u P I ∞ Z − ′ i ′ i (ti t )z t z = pn(t; z)e − i pm(t′; z)e i ρ(z)dz, using (3.1.15), R Z P P i tiz = pn(t; z)pm(t′; z)e ρ(z)dz =0, when m n 1. (3.1.16) R P ≤ − 18 Z i + −j−1 The residue pairing about z = between f = i≥0 aiz and g = j∈Z bj z is defined as: ∞ ∈ H ∈ H Pdz P f,g ∞ = f(z)g(z) = a b . h i 2πi i i z=∞ ≥ I Xi 0 45 3, p.46 § The KP-hierarchy: Setting n = m + 1, shifting t t y, t′ t + y, evaluating the 7→ − 7→ residue and Taylor expanding in yk and using the Schur polynomials sn, leads to (see footnote 14 for the definition of p(∂ )f g.) t ◦ 1 ∞ 2y zi 1 1 0 = dz e− 1 i τ (t y [z− ])τ (t + y + [z− ]) 2πi n − − n I P 1 ∞ ∞ ∞ ∂ i j 1 yk ∂t = dz z si( 2y) z− sj(∂˜) e k τn τn 2πi − P ◦ 0 ! 0 ! I X X ∞ ∂ ∞ 1 yk ∂t = e k si( 2y)si+1(∂˜)τn τn P − ◦ 0 X ∞ ∂ ∂ ∞ = 1+ y + O(y2) + s (∂˜)( 2y + O(y2)) τ τ j ∂t ∂t i+1 − i n ◦ n 1 j ! 1 1 ! X X ∂ ∞ ∂ ∂ = + y 2s (∂˜) τ τ + O(y2), ∂t k ∂t ∂t − k+1 n ◦ n 1 1 k 1 ! X thus yielding (3.1.6), taking into acount the fact that (∂/∂t )τ τ = 0 and the coefficient 1 ◦ of yk is trivial for k =1, 2. The Riemann-Hilbert problem: Observe that, as a function of z, the integral (3.1.10) has a jump accross the real axis 1 pn(t; u) 1 pn(t; u) lim ρt(u)du = pn(t, z)ρt(z)+ lim ρt(u)du, ′ ′ 2πi z →z R z′ u 2πi z →z R z′ u Iz′<0 Z − Iz′>0 Z − and thus we have: (see [29, 19, 5]) Corollary 3.2 The matrix −1 −1 τn(t [z ]) n τn+1(t+[z ]) n 1 − z z− − τn(t) τn(t) Yn(z)= −1 −1 τn−1(t [z ]) n 1 τn(t+[z ]) n − z − z− τn(t) τn(t) satisfies the Riemann–Hilbert problem19: 1. Yn(z) holomorphic on the C+ and C − 1 2πiρt(z) 2. Yn (z)= Yn+(z) (Jump condition) − 0 1 n z− 0 1 3. Y ′(z) =1+ O(z− ), when z . n 0 zn →∞ 19 C+ and C− denote the Siegel upper and lower half plane. 46 3, p.47 § 3.2 Pfaff Lattice and symmetric/symplectic matrix integrals 3.2.1 Pfaff lattice, factorization of skew-symmetric matrices and skew- orthogonal polynomials Consider an inner-product, with a skew-symmetric weightρ ˜(y, z), i i ti(y +z ) f,g t = f(y)g(z)e ρ˜(y, z)dy dz, withρ ˜(z,y)= ρ˜(y, z). (3.2.1) h i R2 − Z Z P Since f,g = g, f , the moment matrix, depending on t =(t , t ,... ), h it −h it 1 2 i j mn(t)=(µij(t))0 i,j n 1 =( y , z t)0 i,j n 1 ≤ ≤ − h i ≤ ≤ − is skew-symmetric. It is clear from formula (3.2.1) that the semi-infinite matrix m ∞ evolves in t according to the commuting vector fields: ∂µij ∂m k k = µi+k,j + µi,j+k, i.e., ∞ =Λ m + m Λ⊤ . (3.2.2) ∂tk ∂tk ∞ ∞ Since m is skew-symmetric, m does not admit a Borel factorization in the stan- ∞ ∞ dard sense, but m admits a unique factorization, with an inserted semi-infinite, skew- ∞ symmetric matrix J, with J 2 = I, of the form (1.1.12): (see [2]) − 1 1 m (t)= Q− (t)J Q⊤− (t), ∞ where . .. 0 0 Q 0 2n,2n 0 Q Q(t)= 2n,2n K. (3.2.3) ∈ Q2n+2,2n+2 0 ∗ 0 Q2n+2,2n+2 . .. K is the group of lower-triangular invertible matrices of the form above, with Lie algebra k. Consider the Lie algebra splitting, given by k = lower-triangular matrices of the form (3.2.3) gl( )= k n { } ∞ ⊕ n = sp( )= a such that Ja⊤J = a , ∞ { } 47 3, p.48 § (3.2.4) with unique decomposition20 a = (a)k +(a)n 1 = (a J(a+)⊤J)+ (a0 J(a0)⊤J) − − 2 − 1 + (a + J(a )⊤J)+ (a + J(a )⊤J) . (3.2.5) + + 2 0 0 Consider as a special skew-symmetric weight (3.2.1): (see [13]) α V˜ (y) ρ˜(y, z)=2D δ(y z)˜ρ(y)˜ρ(z) , with α = 1, ρ˜(y)= e− , (3.2.6) − ∓ together with the associated inner-product21 of type (3.2.1): i i ti(y +z ) α f,g t = f(y)g(z)e 2D δ(y z)˜ρ(y)˜ρ(z)dy dz (3.2.7) h i R2 − Z Z P ∞ t (yi+zi) f(y)g(z)e 1 i ε(y z)˜ρ(y)˜ρ(z)dy dz, for α = 1 R2 P − − = ZZ ∞ 2t yi 2 f,g (y)e 1 i ρ˜(y) dy, for α = +1 R{ } P Z in terms of the Wronskian f,g := ∂f g f ∂g . The moments with regard to these { } ∂y − ∂y inner-products (with that precise definition of time t!) satisfy the differential equations ∂µij/∂tk = µi+k,j + µi,j+k, as in (3.2.2). It is well known that the determinant of an odd skew-symmetric matrix equals 0, whereas the determinant of an even skew-symmetric matrix is the square of a polynomial in the entries, the Pfaffian 22. Define now the “Pfaffian τ-functions”, defined with regard 20 a± refers to projection onto strictly upper (strictly lower) triangular matrices, with all 2 2 × diagonal blocks equal zero. a0 refers to projection onto the “diagonal”, consisting of 2 2 blocks. 21ε(x) = sign x, having the property ε′ =2δ(x). × 22 with a sign specified below. So det(m2n−1(t)) = 0 and 1/2 (det m2n(t)) = pf(m2n(t)) n 1 −1 = (dx dx ... dx − ) µ (t)dx dx . n! 0 ∧ 1 ∧ ∧ 2n 1 ij i ∧ j ≤ ≤ − 0 i 48 3, p.49 § to the inner-products (3.2.7) : ∞ i i pf ykzℓε(y z)e 1 ti(y +z )ρ˜(y)˜ρ(z)dydz ,α = 1 R2 − P 0 k,ℓ 2n 1 − τ2n(t) := ZZ ≤ ≤ − ∞ i k ℓ 1 2tiy 2 pf y ,y e ρ˜ (y)dy , α = +1 R{ } P 0 k,ℓ 2n 1 Z ≤ ≤ − (3.2.8) Setting ρ˜(z)= ρ(z)I (z) for α = 1 E − ρ˜(z)= ρ1/2(z)I (z), t t/2 for α = +1, E → in the identities (3.2.8) leads to the identities (3.2.9) between integrals and Pfaffians, spelled out in Theorem 3.3 below. Remember c =(c1, ..., c2r) stands for the boundary points of the disjoint union E R. (E) and (E) denotes the set of symmet- ⊂ Sn Tn ric/symplectic ensemble with spectrum in E. Theorem 3.3 (Adler-Horozov-van Moerbeke [9], Adler-van Moerbeke [7]) The follow- ing integrals In(t, c) are Pfaffians: n ∞ i β tiz In = ∆n(z) e 1 k ρ(zk)dzk En | | P Z kY=1 β =1 Tr( V (X)+ ∞ t Xi) = e − 1 i dX, (n even) n(E) P ZS k ℓ ∞ t (yi+zi) = n!pf y z ε(y z)e 1 i ρ(y)ρ(z)dydz E2 − P 0 k,ℓ n 1 ZZ ≤ ≤ − = n!τn(t, c) = (3.2.9) β =4 ∞ i Tr( V (X)+ tiX ) = e − 1 dX, (n arbitrary) 2n(E) P ZT k ℓ ∞ t yi = n!pf y ,y e 1 i ρ(y)dy { } E P 0 k,ℓ 2n 1 Z ≤ ≤ − = n!τ2n(t/2,c). The In and τn’s satisfy (i) The Virasoro constraints23 (2.1.7) for β =1, 4, 2r ∂ ck+1f(c ) + a βJ(2) b βJ(1) I (t, c)=0. (3.2.10) − i i ∂c i k+i,n − i k+i+1,n n 1 i i 0 ! X X≥ ′ i 23 ρ biz i here the ai’s and bi’s are defined in the usual way, in terms of ρ(z); namely, ρ = aiz − P P 49 3, p.50 § (ii) The Pfaff-KP hierarchy: (see footnote 14 for notation) 2 ˜ 1 ∂ ˜ sk+4(∂) τn τn = sk(∂) τn+2 τn 2 (3.2.11) − 2 ∂t ∂t ◦ ◦ − 1 k+3 n even, k =0, 1, 2, ... . of which the first equation reads (n even) 4 2 2 2 2 ∂ ∂ ∂ ∂ τn 2τn+2 +3 4 log τ +6 log τ = 12 − . ∂t ∂t − ∂t ∂t n ∂t2 n τ 2 1 2 1 3 ! 1 n (iii) The Pfaff Lattice: The time-dependent matrix L(t), zero above the first superdiag- onal, obtained by dressing up Λ, 1 ∗ (h /h )1/2 O 2 0 ∗ 1 ∗ 1 (h /h )1/2 L(t)= Q(t)ΛQ(t)− = 4 2 (3.2.12) ∗ ∗ . .. ∗ satisfies the Hamiltonian commuting equations ∂L i = [ (L )k, L]. (Pfaff lattice) (3.2.13) ∂ti − (iv) Skew-orthogonal polynomials: The vector of time-dependent polynomials q(t; z) := (qn(t; z))n 0 = Q(t)χ(z) in z satisfy the eigenvalue problem ≥ L(t)q(t, z)= zq(t, z) (3.2.14) 50 3, p.51 § and enjoy the following representations (with h = τ2n+2(t) ) 2n τ2n(t) 1 1/2 m2n+1(t) z h2−n . q2n(t; z) = pf . τ2n(t) 2n z 1 z . . . z2n 0 − − − 1 2n 1/2 τ2n(t [z− ]) 2n 1/2 = z h2−n − = z h2−n + ... , τ2n(t) 1 µ0,2n+1 m2n(t) z µ1,2n+1 1/2 . . h− . . q (t; z) = 2n pf . . . 2n+1 2n 1 τ2n(t) z − µ2n 1,2n+1 2n 1 2−n+1 1 ... z − 0 z − − 2n+1 − µ2n+1,0 ... µ2n+1,2n 1 z 0 − 2n 1/2 1 ∂ 1 2n+1 1/2 = z h− z + τ (t [z− ]) = z h− + .... 2n τ (t) ∂t 2n − 2n 2n 1 (3.2.15) They are skew-orthogonal polynomials in z; i.e., q (t; z), q (t; z) = J . h i j it ij The hierarchy (3.2.11) already appears in the work of Kac-van de Leur [42] in the context of, what they call the DKP-hierarchy, and interesting further work has been done by van de Leur [71]. 3.2.2 Sketch of Proof Skew-orthogonal polynomials and the Pfaff Lattice: The equalities (3.2.9) be- tween the Pfaffians and the matrix integrals are based on two identities [49], the first one due to de Bruyn, 1 n dyi det Fi(y1) Gi(y1) ... Fi(yn) Gi(yn) n! Rn 0 i 2n 1 1 ≤ ≤ − Z Y 1/2 = det (Gi(y)Fj(y) Fi(y)Gj(y))dy R − 0 i,j 2n 1 Z ≤ ≤ − and (Mehta [50]) 4 i i i i i i ∆n(x) = det x1 (x1)′ x2 (x2)′ ... xn (xn)′ 0 i 2n 1 . ≤ ≤ − 51 3, p.52 § i tiz On the one hand, (see Mehta [49]), setting in the calculation below ρt(z)= ρ(z)e IE(z) P and x i i Fi(x) := y ρt(y)dy and Gi(x) := Fi′(x)= x ρt(x), Z−∞ i tiz one computes: (ρt(z) := ρ(z)e ) P 1 2n ∆2n(z) ρt(zi)dzi (2n)! R2n | | Z i=1 Y 2n = det zi ρ (z ) dz , j+1 t j+1 0 i,j 2n 1 i