Multivariate Stieltjes continued fractions Michael Anshelevich June 17, 2008 µ = probability measure on R, with all moments finite. Identify with a functional µ on R[x]. Moments µ[x], µ[x2], µ[x3], . . . . Moment generating function M µ(z) = 1 + µ[x]z + µ[x2]z 2 + µ[x3]z 3 + . . . . 1 Has a (Stieltjes) continued fractions expansion M µ(z) = 1 + µ[x]z + µ[x2]z 2 + µ[x3]z 3 + . . . 1 . = 2 γ1z 1 − β0 z − γ2z 2 1 − β1 z − γ3z 2 1 − β2 z − 1 − ... Coefficients: Jacobi parameters. For {Pn(x)} monic orthogonal polynomials for µ, xPn(x) = Pn+1(x) + βnPn (x) + γnPn−1(x). 2 M ULTIVARIATE CASE ϕ = positive linear functional on Rhx1, x2, . . . , xdi, polynomials in non-commuting variables. Unital (= probability). A “state”. 3 M ULTIVARIATE CASE ϕ = positive linear functional on Rhx1, x2, . . . , xdi, polynomials in non-commuting variables. Unital (= probability). A “state”. Where do these appear? 4 M ULTIVARIATE CASE ϕ = positive linear functional on Rhx1, x2, . . . , xdi, polynomials in non-commuting variables. Unital (= probability). A “state”. Where do these appear? Let X1, X2, . . . , Xd be symmetric (better, self-adjoint) operators on a Hilbert space H, with common invariant dense domain containing a unit vector Ω. Then their joint distribution in the vector state Ω is the functional ϕ [P (x1, x2, . . . , xd)] = hΩ, P (X1 , X2, . . . , Xd)Ωi . 5 M ULTIVARIATE CASE ϕ = positive linear functional on Rhx1, x2, . . . , xdi, polynomials in non-commuting variables. Unital (= probability). A “state”. Where do these appear? Let X1, X2, . . . , Xd be symmetric (better, self-adjoint) operators on a Hilbert space H, with common invariant dense domain containing a unit vector Ω. Then their joint distribution in the vector state Ω is the functional ϕ [P (x1, x2, . . . , xd)] = hΩ, P (X1 , X2, . . . , Xd)Ωi . Moment generating function ϕ M (z1, z2, . . . , zd) = 1 + d X i=1 ϕ [xi] zi + d X i,j=1 h i ϕ xi xj z i z j + . . . . 6 Continued fraction expansion? M µ(z) = 1 + µ[x]z + µ[x2]z 2 + µ[x3]z 3 + . . . 1 = γ1z 2 1 − β0 z − γ2z 2 1 − β1 z − γ3z 2 1 − β2 z − 1 − ... can be done by induction, breaks down if some γ = 0. Multivariate: 1+ X aizi + X 1 P P P 1 − bizi − bij zizj − bijk zizj zk . . . 1 6= P P 1 − bizi − bij ziFij zj aij zizj + . . . = if e.g. bij = 0, bikj 6= 0. 7 P RODUCT- TYPE EXAMPLES Distributions with fixed marginals. For simplicity, restrict to the symmetric case: µ(1), µ(2), (i) (i) (i) (i) xiPn (xi) = Pn+1(xi) + γn Pn−1(xi), i = 1, 2. Commutative case: many measures µ on R2 with marginals µ(1), µ(2). Canonical one: product measure µ[x1k x2n] = µ[x1k ]µ[x2n], x1, x2 independent with respect to µ, orthogonal polynomials (1) Pi (2) (x1)Pj (x2) . 8 Non-commutative case: want ϕ on Rhx1, x2i with (1) n ϕ [xn = µ [x1 ], ] 1 (2) n ϕ [xn = µ [x2 ]. ] 2 Again many choices, more than one canonical choice. 9 Non-commutative case: want ϕ on Rhx1, x2i with (1) n ϕ [xn = µ [x1 ], ] 1 (2) n ϕ [xn = µ [x2 ]. ] 2 Again many choices, more than one canonical choice. Example. 1 1 + M (z1, z2) = 1− γ11z12 γ21z12 1− γ31z12 1− 1 − ... − γ12z22 γ22z22 1− γ32z22 1− 1 − ... 10 Corresponds to the rule h ϕ x1 u(1) x2 v(1) x1 u(2) x2 v(2) i h . . . = ϕ x1 u(1) i h ϕ x2 v(1) i h ϕ x1 u(2) i h ϕ x2 (recall x1, x2 do not commute). Boolean product. Not very natural since very degenerate. 11 v(2) i ... Example. Free product, free independence. Rule for u(1) v(1) u(2) v(2) x2 x1 x2 . . . ϕ x1 complicated, but appears in applications, e.g. random matrices. 12 1 1− γ11 z1|z1 γ12 z2 |z2 γ21 z1|z1 − 1− γ31 z1|z1 γ12 z2 |z2 γ21 z1|z1 γ22 z2|z2 1− − 1− − 1 − ... 1 − ... 1 − ... 1 − ... − γ12 z2 |z2 γ22 γ11 z1|z1 − 1− γ21 z1 |z1 γ12z2 |z2 γ11 z1| 1− − 1− 1 − ... 1 − ... 1−. 13 1 1− γ11z12 γ21z12 1− γ31z12 1− 1 − ... − γ12z22 γ22z22 1− γ32z22 1− 1 − ... 14 For a special but large class of ϕ, have a matricial continued fraction expansion. Namely, when ϕ has monic multivariate orthogonal polynomials: {P~u = x~u + . . .} , for~ u 6= ~v . ∗ ϕ P~u (x1, . . . , xd)P~v (x1, . . . , xd) = 0 Do not always have them. Existence equivalent to special (graded) Hilbert space H and operators Xi. 15 Theorem. Let ϕ be a state with a monic orthogonal polynomials. There exist matrices C (k) = diagonal non-negative dk × dk matrix, k = 1, 2, . . . and (k) Ti = dk × dk matrix, k = 0, 1, . . . , i = 1, 2, . . . , d such that 1 + M (z) = 1 . P P (1) | k1 Ek1 zk1 P j1 zj1 Ej1 C (0) 1 − i0 z i0 Ti − P (2)| P E z 0 z E C P j2 j2 j2 k2 k2 k2 (1) 1 − i1 z i1 Ti − 1 1 − ... 16 Here for matrices A, B ∈ Mdk ×dk ≃ Md×d ⊗ Md×d ⊗ . . . ⊗ Md×d, we use the notation D E EiA|Ej −1 = ei ⊗ I ⊗ . . . ⊗ I, AB (ej ⊗ I ⊗ . . . ⊗ I) ∈ Mdk−1×dk−1 . B Example. ∗ ∗ E1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ E2 = ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ . ∗ ∗ 17