Measures, orthogonal polynomials, and continued fractions Michael Anshelevich November 7, 2008 M EASURES AND ORTHOGONAL POLYNOMIALS . µ a positive measure on R. A linear functional µ[P ] = R P (x) dµ(x), µ : R[x] → R. Positive: µ[P (x)2] ≥ 0. Inner product hP, Qi = µ[P Q]. Gram-Schmidt n 1, x, x2, x3, . . . o ⇒ ⇒ monic orthogonal polynomials {P0 = 1, P1, P2, P3, . . .}. 1 Theorem. (Favard, Stone, etc.) For some βi ∈ R, γi ≥ 0 xPn = Pn+1 + βnPn + γnPn−1. 2nd order recursion relation. Two independent solutions {Pn, Qn}. Initial conditions P−1 = 0, P0 = 1, Q0 = 0, P1 = x − β0, Q1 = 1, Exercise. Qn(x) = (I ⊗ µ) " Q2 = x − β1. # Pn(x) − Pn(y) . x−y 2 ( µ↔ (β0, β1, β2, . . .) ) equivalent. (γ1, γ2, γ3, . . .) More explicit relation? ( If know (β0, β1, β2, . . .) (γ1, γ2, γ3, . . .) ) , how to recover µ? Without going through {Pn}. Cauchy transform Z Gµ(z)= · 1 1 dµ(x) = µ z−x z−x ¸ µ[1] µ[x] µ[x2] µ[x3] = + 2 + + + ... 3 4 z z z z 3 µ[1] µ[x] µ[x2] µ[x3] Gµ(z) = + 2 + + + . . .. z z z3 z4 Theorem. Also 1 Gµ(z) = γ1 z − β0 − z − β1 − γ2 γ3 z − β2 − z − ... Same coefficients as in the recursion. Note: µ ≥ 0 ⇔ all γ ≥ 0, no determinants. Proof II. Flajolet (1980): lattice paths. 4 Proof I. Look at 1 G(z) = γ1 z − β0 − γ2 z − β1 − γ3 ... z − β2 − z − βn−1 − γnH polynomial . G= polynomial Claim. Qn − γnQn−1H G= , Pn − γnPn−1H where {P, Q} with recursion {βi, γi}. 5 Proof. By induction. G= 1 Q − γ1Q0H0 . = 1 z − β0 − γ1H0 P1 − γ1P0H0 Assume Qn − γnQn−1H G= Pn − γnPn−1H and say H= 1 . z − βn − γn+1K Then G= 1 Qn − γnQn−1 z−βn−γ n+1 K 1 Pn − γnPn−1 z−βn−γ n+1 K zQn − βnQn − γn+1QnK−γnQn−1 = zPn − βnPn − γn+1PnK−γnPn−1 Qn+1 − γn+1QnK = . Pn+1 − γn+1PnK 6 F INITE CONTINUED FRACTIONS . Let Gn = G cut off at level n. 1 Gn(z) = γ1 z − β0 − γ2 z − β1 − γ3 ... z − β2 − z − βn−1 − 0 H = 0. Qn − γnQn−1H Qn Gn = = . Pn − γnPn−1H Pn 7 Qn Gn = . Pn Pn monic, n real roots n Y Pn(x) = (x − xi). i=1 Z Gn(z)= 1 Qn(z) dµn(x) = Q (z − xi) z−x = (partial fractions) = X ai . z − xi Gn = Cauchy transform of µn = xi = roots of Pn, X ai = aiδxi , Qn(xi) . 0 Pn(xi) 8 1 Gn(z) = γ1 z − β0 − γ2 z − β1 − γ3 ... z − β2 − z − βn−1 − 0 Gn approximate G. Gn = Gµn , µn = n X i=1 aiδxi . Do µn approximate µ? Yes. In fact, µn[P (x)] = µ[P (x)] for deg P ≤ 2n − 1. 9 G AUSSIAN QUADRATURE . Want to evaluate Z n X f (x) dµ(x) ≈ aif (xi). i=1 “Riemann sums”. R Want P (x) dµ(x) = Pn i=1 aiP (xi ) for P of low degree. How to choose ai, xi? Answer: take xi = roots of Pn. Choose ai so that Z xk dµ(x) = X aixki , k = 0, 1, . . . , n − 1 (n equations, n unknowns). Our ai = Qn(xi) work. 0 Pn(xi) 10 Proof. Lagrange interpolation: for any P with deg P < n, P (x) = Note " # n X P (xi)Pn(x) . 0 i=1 Pn (xi )(x − xi ) " # Pn(x) Pn(x) − Pn(xi) µ =µ = Qn(xi) x − xi x − xi so µ[P (x)] = n X n X P (xi) Qn(xi) Q (x ) = δxi [P ] n i 0 0 i=1 Pn (xi ) i=1 Pn (xi) µ[P (x)] = n X i=1 aiδxi [P (x)] µ[P (x)] = µn[P (x)] for deg P < n. In fact, same xi, ai work for k = n, n + 1, . . . , 2n − 1. 11 For Pk , n ≤ k ≤ 2n − 1, Pk (x) = A(x)Pn(x) + B(x), deg A, B ≤ n − 1. µ[Pk ] = h1, Pk i = 0. To show: µn[Pk ] = 0. µ[APn] = hA, Pni = 0 ⇒ µ[B] = 0 ⇒ deg A < n µn[B] = 0. Finally, µn[APn] = X aiA(xi)Pn(xi) = 0, so µn[Pk ] = 0. So µn → µ, Gn → G, and therefore Gµ = G. 12 If know {βi, γi}, can find µ? Usually hard: 1 Gµ(z) = γ1 z − β0 − z − β1 − γ2 γ3 z − β2 − z − ... an infinite expression. Class of explicit examples. Semicircle law: q 1 4t − x21[−2√t,2√t] dx. 2πt 2.0 1.5 y 1.0 0.5 0.0 −3 −2 −1 0 1 2 3 x 13 Semicircle law: q 1 4t − x21[−2√t,2√t] dx. 2πt 2.0 1.5 y 1.0 0.5 0.0 −3 −2 −1 0 1 3 2 x Marchenko-Pastur distributions: q 1 · 2π 4t − (x − 1 − t)2 x 1[1+t−2√t,1+t+2√t](x) dx + max(1 − t, 0)δ0. 1.0 1.0 0.75 0.75 y 0.5 y 0.5 0.25 0.25 0.0 0.0 0 1 2 x 3 4 0 1 2 3 4 x 14 Semicircular, Marchenko-Pastur orthogonal polynomials satisfy ∞ X Pn (x)un n=0 A(u) = . 1 − B(u)x In general: free Meixner distributions q 4(t + c) − (x − b)2 1 √ √ · 1 dx [b−2 t+c,b+2 t+c] 2 2 2πt 1 + (b/t)x + (c/t )x +0, 1, 2 atoms. ◦ AC support an interval √ polynomial ◦ polynomial ◦ limited atoms ◦ outside of the AC support ◦ Not SC part 15 P ERIODIC CONTINUED FRACTIONS . βi+n = βi, γi+n = γi. H = G in 1 G(z) = γ1 z − β0 − γ2 z − β1 − γ3 ... z − β2 − z − βn−1 − γnH G= Qn − γnQn−1G . Pn − γnPn−1G γnPn−1G2 − (γnQn−1 + Pn)G + Qn = 0. Quadratic equation! D = (γnQn−1 + Pn)2 − 4γnQnPn−1. 16 √ (γnQn−1 + Pn) − D G= 2γnPn−1 Z = 1 dµ(x). z−x Stieltjes inversion formula: dµ(x) = − q dµ(x) > 0 if 1 lim Im G(x + iy). π y→0 D(x) ∈ iR, i.e. D < 0. D degree 2n, n intervals for D ≤ 0. No SC. atoms: roots of Pn−1, at most (n − 1). Recall D = (γnQn−1 + Pn)2 − 4γnQnPn−1. So if Pn−1(a) = 0, then D(a) ≥ 0. Atoms outside of the AC support. 17 Eventually constant continued fractions (n = 1) βi = β, γi = γ for i ≥ N . p ⇔ polynomial on one interval. polynomial If β = 0, γ = 1 (Peherstorfer?) q 4 − x2 Bernstein-Szegő class on [−2, 2]. polynomial Weyl’s Theorem. If βi → 0, γi → 1, then σess(µ) = [−2, 2]. Denisov-Rakhmanov Theorem. If σess(µ) = AC support of µ = [−2, 2], then βi → 0, γi → 1. 18 √ polynomial Eventually periodic ⇒ polynomial on n intervals. √ polynomial polynomial on n intervals 6⇒ eventually periodic. Weyl: if {βi, γi} asymptotically periodic, same essential spectrum as for actually periodic. Converse false. Last, Simon: if {βi, γi} approaches the isospectral torus of a periodic sequence, same essential spectrum. Damanik, Killip, Simon: converse true. 19 Q UESTIONS . ◦ Connection between random matrices and (eventually) periodic continued fractions (Pastur). ◦ Multivariate (non-commutative) orthogonal polynomials, states, continued fractions. All exist. Continued fractions matricial. ◦ Formulas for states with periodic continued fractions. Free Meixner states known; “constant” after step 2. If not formulas for states, description of their operator algebras. ◦ Connection between multi-matrix models and states with multivariate (eventually) periodic continued fractions. 20