Limit theorems and the Chernoff product formula Michael Anshelevich (joint work with John Williams) Texas A&M University October 8, 2012 Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Compositional square root. Let A = F : C+ → C+ analytic, F (z) ∼ z as z → ∞ . Note A closed under composition. Question. Given F ∈ A, does there exist g ∈ A such that F = g ◦ g? Answer. In general, no. (But, if it exists, it is unique!) Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Infinite divisibility. Definition. If F ∈ A and for any n ∈ N, there exists gn ∈ A such that F = gn ◦ gn ◦ . . . ◦ gn , | {z } n times say F is infinitely divisible. May write gn = F ◦(1/n) . Theorem. (Belinschi 2005) If F ∈ A is infinitely divisible, then F ◦t is defined for all t ≥ 0: there exist {Ft : t ≥ 0} which form a semigroup under composition: Ft ◦ Fs = Ft+s , Michael Anshelevich (joint work with John Williams) F1 = F . Limit theorems and the Chernoff product formula Limit theorems. Question. Given an infinitely divisible F ∈ A and an increasing sequence (kn ). Under what conditions on {gn : n ∈ N} do gn ◦ gn ◦ . . . ◦ gn → F {z } | kn uniformly on compact sets? Will answer, but the answer in terms of measures not functions. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Limit theorems in probability. {µn : n ∈ N} = probability measures on R. Under what conditions do µ n ∗ µn ∗ · · · ∗ µn → ν | {z } kn weakly? Example. Central Limit Theorem: if Z supp(µn ) ↓ {0} , x dµn (x) = 0, R Z n x2 dµn (x) → 1, R then µ∗n n → normal distribution. Example. Poisson Limit Theorem: if npn → 1, then ((1 − pn )δ0 + pn δ1 )∗n → Poisson distribution. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Infinitely divisible measures. Such limit measures are necessarily infinitely divisible: ν = ν1 , {νt : t ≥ 0} , νt ∗ νs = νt+s . Lévy-Hinchin formula. Infinitely divisible measures are precisely those with Fourier transforms of the form Z x2 + 1 γ,σ iθx dσ(x) (Fν∗ )(θ) = exp iγθ + (e − 1 − iθx) x2 R for γ ∈ R and σ a finite Borel measure on R. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Limit theorems. Theorem. (Hinchin, Gnedenko) The sequence µn ∗ µn ∗ · · · ∗ µn → ν∗γ,σ {z } | kn weakly if and only if kn x2 dµn (x) → dσ(x) x2 + 1 weakly, and Z kn R x2 x dµn (x) → γ. +1 Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Other convolutions. µ = probability measure on R. Its Cauchy transform Z 1 Gµ (z) = dµ(x), z ∈ C+ , R z−x and Fµ (z) = 1 . Gµ (z) Define three new convolutions , ], B. Binary operations on measures. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Other convolutions. Free convolution µ1 µ2 : Fµ−1 (z) = Fµ−1 (z) + Fµ−1 (z) − z, 1 2 1 µ2 z in a domain. Boolean convolution µ1 ] µ2 : Fµ1 ]µ2 (z) = Fµ1 (z) + Fµ2 (z) − z, z ∈ C+ . Monotonic convolution µ1 B µ2 : Fµ1 Bµ2 (z) = Fµ1 (Fµ2 (z)), z ∈ C+ . Theorems (Voiculescu, Maassen, Bercovici; Speicher, Woroudi; Muraki, Franz): each operation preserves positivity. All come from appropriate notions of independence. Free convolution most complicated. But: monotonic convolution non-commutative. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Infinitely divisible distributions. ν -infinitely divisible = ν belongs to a -convolution semigroup. Same for ] and B. Theorem. (Bercovici, Voiculescu 1992, 1993; Speicher, Woroudi 1997) -infinitely divisible measures are precisely those ν with Z 1 + xz −1 Fν γ,σ (z) − z = γ + dσ(x), z ∈ C+ , R z−x ]-infinitely divisible measures are precisely those ν with Z 1 + xz dσ(x), z ∈ C+ , z − Fν]γ,σ (z) = γ + z − x R for γ ∈ R and σ a finite Borel measure on R. γ,σ Correspondences ν∗γ,σ ↔ ν ↔ ν]γ,σ = Bercovici-Pata bijections. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Limit theorems. Theorem.(Bercovici, Pata 1999). The following are equivalent. 1 The sequence µn ∗ µn ∗ · · · ∗ µn → ν∗γ,σ weakly. | {z } kn 2 γ,σ The sequence µn µn · · · µn → ν weakly. | {z } kn 3 The sequence µn ] µn ] · · · ] µn → ν]γ,σ weakly. | {z } kn 4 kn x2 x2 + 1 Z dµn (x) → σ weakly, and kn R x dµn (x) → γ. x2 + 1 Note: monotone convolution missing. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Proof based on kn log Fµn (θ) = log Fµn∗kn (θ) → log Fν∗γ,σ , −1 kn (Fµ−1 (z) − z) = (F −1 kn (z) − z) → Fν γ,σ (z), n µn kn (Fµn (z) − z) = (Fµ]k n (z) − z) → Fν γ,σ (z) − z. ] n But for monotonic convolution, want Fµ◦kn n (z) = FµnBkn (z) → FνBγ,σ (z). Much more complicated. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Chernoff product formula. Chernoff product formula, discrete version. Let {kn : n ∈ N} be an increasing sequence of positive integers, and {Vn : n ∈ N} a family of contractions on a Banach space X . Suppose A is an unbounded operator which generates a strongly continuous semigroup of contractions {T (t) : t ≥ 0}, D is a core for A, and for each f ∈ D, lim kn (Vn − I)f → Af . n→∞ Then for each f ∈ X , lim Vnkn f = T (1)f . n→∞ Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Classical convolution. This approach used for classical convolution and limit theorems by Feller (1971), Goldstein (1976). But note: classical convolution operators all commute. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Monotone convolution semigroups. z ∈ C+ . Thus Recall Fµ1 Bµ2 (z) = Fµ1 (Fµ2 (z)), B − convolution semigroups ↔ semigroups of analytic transformation of C+ under composition. Results of Berkson and Porta (1978) imply: each monotonic convolution semigroup {νt : t ≥ 0}, νt B νs = νt+s satisfies ∂Fνt = Φ(Fνt ). ∂t Moreover, by Nevanlinna-Pick theory Z 1 + xz −Φ(z) = γ + dσ(x) R z−x for some γ, σ. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Monotone convolution semigroups. Thus extend the Bercovici-Pata bijection: for γ, σ as before, denote Φγ,σ (z) = Fν]γ,σ (z) − z, let {νt : t ≥ 0} be the monotone convolution semigroup it generates, and define ν.γ,σ = ν1 . Limit theorems? Vn = composition operator, Vn f = f ◦ Fµn . Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Monotone limit theorem. Theorem. (A, Williams) The sequence µn ] µn ] · · · ] µn → ν]γ,σ | {z } kn weakly if and only if γ,σ µn B µn B · · · B µn → νB {z } | kn weakly. Known results: Muraki 2001 (CLT and Poisson limit for compactly supported), Wang 2012 (full CLT, convergence to stable laws). Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Proof (sketch) Step I. Assume n µ]k → ν]γ,σ n where {µn : n ∈ N} , ν]γ,σ all have uniformly bounded variance. Let C+ 1 = {z ∈ C : =z ≥ 1}, D = Span ({Gρ : ρ ∈ M2 }), and X its completion with respect to the norm kf k = sup |f (z)|. z∈C+ 1 Then each composition operator by Fµ is a contraction on X leaves D invariant. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Proof in the finite variance case. The operator Ah = Φγ,σ h0 generates the semigroup of composition operators corresponding to γ, σ has D as a core for h ∈ D, kn (h ◦ Fµn − h) − Φγ,σ h0 → 0 So the result follows from the Chernoff product formula. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Proof of the general result. Step II. n µ]k → ν]γ,σ n ⇒ n µ.k → ν.γ,σ n in general. Cutoff the measures, and approximate (carefully). Step III. The converse: n µ.k → ν.γ,σ n ⇒ n µ]k → ν]γ,σ . n In general, the converse to the Chernoff theorem is false. True in our case. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Key lemma. Let µn B µn · · · B µn → ν weakly. Then | {z } kn 1 The family of measures n o B µn j : n ∈ N, j = 1, . . . , kn is tight. 2 For every > 0 there exist α, β > 0 such that kn =(Fµn (z) − z) ≤ 2=(Fν (z) − z) ≤ |z| for large n ∈ N and z ∈ Γα,β . Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Compositional limit theorem. Recall A = {F : C+ → C+ analytic, F (z) ∼ z as z → ∞}. Theorem. (A, Williams) For {gn : n ∈ N} , F ∈ A, F infinitely divisible, and an increasing sequence (kn ), gn ◦ gn ◦ . . . ◦ gn → F {z } | kn uniformly on compact sets if and only if kn (gn (z) − z) → Φ, where F = F1 , ∂Ft ∂t = Φ(Ft ). Equivalent conditions on µn but not on gn . Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Composition in the unit disk. Similar but different results for the unit disk. Straightforward equivalence is false: ηµn (z) z kn → ην∪x z 6⇔ ηµ◦knn (z) → ην (z). But true under additional assumptions. Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Non-identically distributed arrays. Let {µnk : n ∈ N, 1 ≤ k ≤ kn } be an infinitesimal array, that is, for every ε > 0 lim max µnk ({x ∈ R : |x| ≥ ε}) = 0. n→∞ 1≤k≤kn Denote Z ank = x dµnk (x), dµ◦nk (x) = dµnk (x + ank ). (−1,1) Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula Non-identically distributed arrays. Theorem.(Gnedenko; Chistyakov, Götze; Bercovici, Wang; Wang). The following are equivalent. 1 The sequence µn1 ∗ µn2 ∗ · · · ∗ µnkn → ν∗γ,σ weakly. 2 γ,σ The sequence µn1 µn2 · · · µnkn → ν weakly. 3 The sequence µn1 ] µn2 ] · · · ] µnkn → ν]γ,σ weakly. 4 kn X k=1 x2 dµ◦ (x) → σ x2 + 1 nk weakly, and Z kn X ank + k=1 R Michael Anshelevich (joint work with John Williams) x ◦ dµ (x) → γ. x2 + 1 nk Limit theorems and the Chernoff product formula Generalization of Chernoff formula. Corresponding Chernoff-type result: if cnk uniformly small and kn X cnk → a, k=1 then kn Y (1 + cnk ) → ea . k=1 True for numbers, for commuting bounded operators. For matrices? Michael Anshelevich (joint work with John Williams) Limit theorems and the Chernoff product formula