betaFeedback

WorksMathematics

Strong law of large numbers for random semigroups in norm topology

Abstract

This paper studies the strong law of large numbers for products of random uniformly continuous semigroups in Banach spaces. We propose a novel approach by treating the product of semigroups as a solution to an integral equation and establish a version of the SLLN in norm topology for arbitrary Banach spaces.

MSC 2020: 60B12, 60H25, 47D06, 60F15.

Introduction

The theory of compositions of random linear operators can nowadays be considered classical [1, 2, 3, 4]. Besides the purely theoretical interest, it has applications in random quantum dynamics, where either compositions of random linear operators in Hilbert spaces [5, 6] or compositions of random quantum channels [7, 8, 9, 10, 11] are considered.

The strong law of large numbers (SLLN) for random variables acting in a Banach space is well studied [12, 13]. Distributions of compositions of random linear operators were studied in [3, 4, 10, 14]. In particular, the law of large numbers for compositions of random operators in different operator topologies was studied in [6, 15, 16]. Most of the results of this type follow the classical approach of the Chernoff product formula [17, 18, 19], where the convergence of semigroups is typically described in strong operator topology (SOT) and occasionally in weak operator topology (WOT). For example, in [20], the SLLN in SOT was obtained for uniformly continuous random semigroups in uniformly smooth spaces; and in [21], a generalization was obtained for strongly continuous random semigroups. However, these and earlier papers only consider uniformly bounded random operators and usually rely on some geometric properties of the underlying Banach space (such as uniform smoothness), which limits the applicability of such results.

In this paper, we present another viewpoint on the product of semigroups by considering it as the solution of an integral equation. This allows us to derive the SLLN for random semigroups from the SLLN for their generators by employing the standard theorems of calculus for the Bochner integral. In particular, this approach produces the SLLN in norm topology for all independent, identically distributed (i.i.d.) random uniformly continuous semigroups in an arbitrary Banach space. This directly generalizes earlier results obtained in [20].

The paper is organized as follows. In § 2, the necessary notions and main results (Theorem 2.2 and Corollary 2.3) are formulated. In § 3, the proof of Theorem 2.2 is given. § 4 is devoted to counterexamples showing the tightness of the assumptions in Theorem 2.2. Finally, § 5 is devoted to some practical considerations and further ideas regarding our results.

Background and main results

Below are the conventions adopted in this paper.

  • We consider Banach spaces over the field R\mathbb{R} or C\mathbb{C}.

  • We work with a probability space (Ω,σΩ,P)(\Omega,\sigma_{\Omega},\mathrm{P}) for a σ\sigma-additive probability measure P\mathrm{P}.

  • We say that some property S⊆ΩS \subseteq\Omega holds almost surely if there is E∈σΩE \in\sigma_{\Omega} such that E⊆SE \subseteq S and P(E)=1\mathrm{P}(E)=1 (this clarification addresses the case of a non-complete measure).

  • N\mathbb{N} denotes the set of positive integers.

For a Banach space X\mathcal{X}, we say that x:Ω→Xx:\Omega\to\mathcal{X} is a random element if it is Bochner measurable. For a Bochner integrable random element xx, Ex\mathbb{E}x denotes the Bochner integral of xx. A random element A:Ω→L(X)A:\Omega\to\mathcal{L}(\mathcal{X}) is called a random operator with the corresponding random semigroup eAte^{At}. See [20], [22 Chapter 2] for some basic properties of random elements and random operators.

There is a classical result about the strong law of large numbers in Banach spaces:

Theorem 2.1 (Kolmogorov–Mourier SLLN [23], [12 Corollary 7.10]). Let X\mathcal{X} be a separable Banach space. Let {xn:Ω→X}n∈N\{x_n:\Omega\to\mathcal{X}\}_{n\in\mathbb{N}} be a sequence of i.i.d. Bochner integrable random elements. Then almost surely

∑i=1nxin→Ex1.\sum_{i=1}^{n}\frac{x_i}{n} \to\mathbb{E}x_1.

Actually, it is redundant to require separability of the Banach space in SLLN-type theorems, provided that the random elements are Bochner measurable (which some authors do not assume as part of the definition). This follows from the fact that a Bochner measurable random element is essentially separably valued, i.e., takes values in a fixed separable subset of X\mathcal{X} almost surely. Indeed, let En∈σΩE_n \in\sigma_{\Omega} and Sn⊆XS_n \subseteq\mathcal{X} be an event and a separable subset such that P(En)=1\mathrm{P}(E_n)=1 and xn(En)⊆Snx_n(E_n) \subseteq S_n. Then, almost surely, all the xnx_n take values in the closed linear span of ⋃n∈NSn\bigcup_{n\in\mathbb{N}}S_n, which is still separable.

By contrast, it is important that the random elements are identically distributed to guarantee the SLLN in arbitrary Banach spaces. A generalization of the SLLN to independent random elements with bounded second moments requires B-convexity of the Banach space [24].

We define the SLLN for random semigroups in three (generally distinct) operator topologies. Let X\mathcal{X} be a Banach space and let {An:Ω→L(X)}n∈N\{A_n:\Omega\to\mathcal{L}(\mathcal{X})\}_{n\in\mathbb{N}} be a sequence of Bochner integrable random operators.

  • The sequence of random semigroups {eAnt}n∈N\{e^{A_nt}\}_{n\in\mathbb{N}} satisfies the SLLN in norm topology if almost surely for every $T>0

lim⁡n→∞sup⁡t∈[0,T]∥eA1t/n⋯eAnt/n−eEA1t/n⋯eEAnt/n∥=0.\lim_{n\to\infty}\sup_{t\in[0,T]}\left\|e^{A_1t/n}\cdots e^{A_nt/n}-e^{\mathbb{E}A_1t/n}\cdots e^{\mathbb{E}A_nt/n}\right\|=0.
  • The sequence of random semigroups {eAnt}n∈N\{e^{A_nt}\}_{n\in\mathbb{N}} satisfies the SLLN in SOT if for every x∈Xx\in\mathcal{X} almost surely for every T>0T>0

lim⁡n→∞sup⁡t∈[0,T]∥(eA1t/n⋯eAnt/n−eEA1t/n⋯eEAnt/n)x∥=0.\lim_{n\to\infty}\sup_{t\in[0,T]}\left\|\left(e^{A_1t/n}\cdots e^{A_nt/n}-e^{\mathbb{E}A_1t/n}\cdots e^{\mathbb{E}A_nt/n}\right)x\right\|=0.
  • The sequence of random semigroups {eAnt}n∈N\{e^{A_nt}\}_{n\in\mathbb{N}} satisfies the SLLN in WOT if for every x∈Xx\in\mathcal{X} and f∈X∗f\in\mathcal{X}^* almost surely for every T>0T>0

lim⁡n→∞sup⁡t∈[0,T]∣f((eA1t/n⋯eAnt/n−eEA1t/n⋯eEAnt/n)x)∣=0.\lim_{n\to\infty}\sup_{t\in[0,T]}\left|f\left(\left(e^{A_1t/n}\cdots e^{A_nt/n}-e^{\mathbb{E}A_1t/n}\cdots e^{\mathbb{E}A_nt/n}\right)x\right)\right|=0.

The earlier papers usually reverse the order of quantifiers “for all T>0T>0” and “almost surely”. These formulations are equivalent, which follows from continuity of the probability and the fact that supremum over the segment [0,T][0,T] is monotone in TT.

Also, it natural to inquire whether the SLLN for random semigroups survives reversing the order of composition of semigroups. In the setting of Theorem 2.2, the reversed SLLN can indeed be established using the same proof strategy. However, in general, the two convergences are not equivalent, as shown in Example 4.6.

In [20], it was proved that the SLLN in WOT holds for random semigroups with i.i.d. uniformly bounded generators in all Banach spaces, while the SLLN in SOT was proved only in uniformly smooth spaces. It was claimed that by Schur’s property, the SLLN in SOT holds for ℓ1\ell_1, although ℓ1\ell_1 is not uniformly smooth. This argument does not work because the quantifiers “for all ff” and “almost surely” cannot be interchanged in general. However, we show that the conclusion is true for all Banach spaces, thus closing [20 Conjecture 2.5]. Moreover, we state the result in a more general way, using norm topology instead of SOT.

For a sequence of Bochner integrable random elements {xn:Ω→X}n∈N\{x_n:\Omega\to\mathcal{X}\}_{n\in\mathbb{N}} in a Banach space X\mathcal{X}, we define three properties for convenience:

  • Cesàro boundedness (CB for short), which means sup⁡n∈N1n∑i=1n∥xi∥<∞\sup_{n\in\mathbb{N}}\frac{1}{n}\sum_{i=1}^{n}\|x_i\|<\infty almost surely. This property also makes sense for sequences of deterministic elements regarded as constant random elements.

  • Cesàro boundedness of expectations (CBE): sup⁡n∈N1n∑i=1n∥Exi∥<∞\sup_{n\in\mathbb{N}}\frac{1}{n}\sum_{i=1}^{n}\|\mathbb{E}x_i\|<\infty.

  • strong law of large numbers (SLLN): 1n∑i=1n(xi−Exi)→0\frac{1}{n}\sum_{i=1}^{n}(x_i-\mathbb{E}x_i)\to0 in the norm of X\mathcal{X} almost surely.

Theorem 2.2 (SLLN in norm topology; proved in § 3 below). Let X\mathcal{X} be a Banach space. Let {An:Ω→L(X)}n∈N\{A_n:\Omega\to\mathcal{L}(\mathcal{X})\}_{n\in\mathbb{N}} be a sequence of Bochner integrable random operators that has properties CB, CBE and SLLN. Then, the sequence of random semigroups {eAnt}n∈N\{e^{A_nt}\}_{n\in\mathbb{N}} satisfies the SLLN in norm topology. If, in addition, 1n∑i=1nEAi\frac{1}{n}\sum_{i=1}^{n}\mathbb{E}A_i converges to A∈L(X)A\in\mathcal{L}(\mathcal{X}), then almost surely eA1t/n⋯eAnt/ne^{A_1t/n}\cdots e^{A_nt/n} converges to eAte^{At} uniformly on any segment [0,T][0,T].

Theorem 2.1 yields the following corollary for i.i.d. random operators:

Corollary 2.3 (SLLN for i.i.d. random semigroups). Let X\mathcal{X} be a Banach space. For any sequence {An:Ω→L(X)}n∈N\{A_n:\Omega\to\mathcal{L}(\mathcal{X})\}_{n\in\mathbb{N}} of i.i.d. Bochner integrable random operators, the sequence of random semigroups

{eAnt}n∈N\{e^{A_n t}\}_{n\in\mathbb{N}} satisfies the SLLN in norm topology. In particular, eA1t/n⋯eAnt/ne^{A_1t/n}\cdots e^{A_nt/n} converges almost surely to eEA1te^{\mathbb{E}A_1t} uniformly on any segment.

Proof. We need to check that the hypotheses of Theorem 2.2 are satisfied, that is, the sequence has the three main properties in this case.

Note that {∥An∥}n∈N\{\lVert A_n\rVert\}_{n\in\mathbb{N}} are i.i.d. integrable random variables. By the scalar SLLN, almost surely, the sequence 1n∑i=1n∥Ai∥\frac{1}{n}\sum_{i=1}^{n}\lVert A_i\rVert converges to E∥A1∥\mathbb{E}\lVert A_1\rVert; in particular, it is bounded.

The expectations are bounded since the moments are equal.

The SLLN for random operators follows from Theorem 2.1 (as it was said, separability is automatic).

Proof of Theorem 2.2

In our proof, we consider the Bochner integral over a segment [a,b][a,b] (with the Lebesgue measure), which we denote by ∫abx(t) dt\int_a^b x(t)\,dt. One of the basic properties of the Lebesgue integral generalizes to this case:

Proposition 3.1 (Fundamental theorem of calculus for the Bochner integral). Let X\mathcal{X} be a Banach space. Let x:[a,b]→Xx:[a,b]\to\mathcal{X} be continuous on [a,b][a,b] and differentiable on (a,b)(a,b) with x′x' Bochner integrable on [a,b][a,b]. Then,

∫abx′(t) dt=x(b)−x(a).\int_a^b x'(t)\,dt=x(b)-x(a).

Proof. It is sufficient to show that the equality holds in composition with an arbitrary bounded linear functional. Let f∈X∗f\in\mathcal{X}^*. Then, f∘xf\circ x is continuous on [a,b][a,b] and (f∘x)′(t)=f(x′(t))(f\circ x)'(t)=f(x'(t)). Since the Bochner integral commutes with bounded linear operators, we have

f(∫abx′(t) dt)=∫abf(x′(t)) dt=∫ab(f∘x)′(t) dt.f\left(\int_a^b x'(t)\,dt\right)=\int_a^b f(x'(t))\,dt=\int_a^b(f\circ x)'(t)\,dt.

By the Fundamental theorem of calculus for the Lebesgue integral,

∫ab(f∘x)′(t) dt=f(x(b))−f(x(a))=f(x(b)−x(a)).\int_a^b(f\circ x)'(t)\,dt=f(x(b))-f(x(a))=f(x(b)-x(a)).

Thus, for all f∈X∗f\in\mathcal{X}^*,

f(∫abx′(t) dt)=f(x(b)−x(a)).f\left(\int_a^b x'(t)\,dt\right)=f(x(b)-x(a)).

Corollary 3.2 (Integration by parts). Let X\mathcal{X} be a Banach space. Let U,V:[a,b]→L(X)U,V:[a,b]\to\mathcal{L}(\mathcal{X}) be continuous on [a,b][a,b] and differentiable on (a,b)(a,b) with U′,V′U',V' Bochner integrable on [a,b][a,b]. Then,

∫abU(t)V′(t) dt=U(b)V(b)−U(a)V(a)−∫abU′(t)V(t) dt.\int_a^b U(t)V'(t)\,dt=U(b)V(b)-U(a)V(a)-\int_a^b U'(t)V(t)\,dt.

These properties clearly extend to xx (or U,VU,V) being continuous and piecewise differentiable on [a,b][a,b].

We will also refer to a measure-theoretic version of a classical integral inequality:

Theorem 3.3 (Grönwall’s inequality [25 Appendix 5, Theorem 5.1]). Let μ\mu be a (non-negative) Borel measure on a finite interval [a,b][a,b], and let ff be a non-negative bounded Borel measurable function on [a,b][a,b] such that

f(t)≤c+∫[a,t)f(s) μ(ds)f(t) \le c + \int_{[a,t)} f(s)\,\mu(\mathrm{d}s)

for some c≥0c \ge0 and all t∈[a,b]t \in[a,b]. Then

f(t)≤ceμ([a,t))f(t) \le c e^{\mu([a,t))}

for all t∈[a,b]t \in[a,b].

For a general Borel measure, Grönwall’s inequality requires that the bounding integral is taken over the interval [a,t)[a,t) with the endpoint tt excluded. This detail is substantial if the measure has atoms; otherwise, integrating over [a,t][a,t] instead clearly makes no difference.

The core of the proof of Theorem 2.2 lies in the following deterministic lemma:

Lemma 3.4. Let X\mathcal{X} be a Banach space. Let {An∈L(X)}n∈N\{A_n \in\mathcal{L}(\mathcal{X})\}_{n\in\mathbb{N}} and {Bn∈L(X)}n∈N\{B_n \in\mathcal{L}(\mathcal{X})\}_{n\in\mathbb{N}} be two Cesàro bounded sequences such that 1n∑i=1n(Ai−Bi)→0\frac{1}{n}\sum_{i=1}^{n}(A_i-B_i) \to0. Then for every T>0T > 0

lim⁡n→∞sup⁡t∈[0,T]∥eA1t/n⋯eAnt/n−eB1t/n⋯eBnt/n∥=0.\lim_{n\to\infty}\sup_{t\in[0,T]}\left\|e^{A_1t/n}\cdots e^{A_nt/n}-e^{B_1t/n}\cdots e^{B_nt/n}\right\|=0.

Proof. Fix T>0T > 0. Let Ik:=(k−1n,kn)I_k := \left(\frac{k-1}{n},\frac{k}{n}\right) and I‾k:=[k−1n,kn]\overline{I}_k := \left[\frac{k-1}{n},\frac{k}{n}\right] for k=1,…,nk=1,\ldots,n. For t∈[0,T]t \in[0,T] and s∈I‾ks \in\overline{I}_k, define

PA,n(s,t)=eA1t/n⋯eAk−1t/neAkt(s−k−1n).P_{A,n}(s,t)=e^{A_1t/n}\cdots e^{A_{k-1}t/n}e^{A_kt(s-\frac{k-1}{n})}.

Note that

∂PA,n∂s(s,t)=tPA,n(s,t)Ak\frac{\partial P_{A,n}}{\partial s}(s,t)=tP_{A,n}(s,t)A_k

for s∈Iks \in I_k. Then, by the fundamental theorem of calculus for the Bochner integral, PA,nP_{A,n} satisfies the following integral equation for a variable s∈[0,1]s \in[0,1] and a parameter t∈[0,T]t \in[0,T]:

PA,n(s,t)=I+∫0stPA,n(τ,t)GA,n(τ) dτ,P_{A,n}(s,t)=I+\int_0^s tP_{A,n}(\tau,t)G_{A,n}(\tau)\,\mathrm{d}\tau,

where GA,n(s)=AkG_{A,n}(s)=A_k for s∈Iks \in I_k. Similarly, define GB,nG_{B,n} and PB,nP_{B,n}, which satisfies the analogous integral equation.

Put Dn(s,t):=PA,n(s,t)−PB,n(s,t)D_n(s,t):=P_{A,n}(s,t)-P_{B,n}(s,t). Note that PA,n(1,t)=eA1t/n⋯eAnt/nP_{A,n}(1,t)=e^{A_1t/n}\cdots e^{A_nt/n} and PB,n(1,t)=eB1t/n⋯eBnt/nP_{B,n}(1,t)=e^{B_1t/n}\cdots e^{B_nt/n}, so we need to show that ∥Dn(1,t)∥→0\|D_n(1,t)\|\to0 uniformly on [0,T][0,T]. Write Dn(s,t)D_n(s,t) using the integral representation of PA,nP_{A,n} and PB,nP_{B,n}:

Dn(s,t)=∫0st(PA,n(τ,t)GA,n(τ)−PB,n(τ,t)GB,n(τ)) dτ.D_n(s,t)=\int_0^s t\left(P_{A,n}(\tau,t)G_{A,n}(\tau)-P_{B,n}(\tau,t)G_{B,n}(\tau)\right)\,\mathrm{d}\tau.

Add and subtract tPA,n(τ,t)GB,n(τ)tP_{A,n}(\tau,t)G_{B,n}(\tau) inside the integral:

Dn(s,t)=∫0stDn(τ,t)GB,n(τ) dτ+∫0stPA,n(τ,t)(GA,n(τ)−GB,n(τ)) dτ⏟=:Rn(s,t).D_n(s,t)=\int_0^s tD_n(\tau,t)G_{B,n}(\tau)\,\mathrm{d}\tau+\underbrace{\int_0^s tP_{A,n}(\tau,t)(G_{A,n}(\tau)-G_{B,n}(\tau))\,\mathrm{d}\tau}_{=:R_n(s,t)}.

Taking norms, we get:

∥Dn(s,t)∥≤∥Rn(s,t)∥+∫0st∥Dn(τ,t)∥∥GB,n(τ)∥ dτ.\lVert D_n(s,t)\rVert\le\lVert R_n(s,t)\rVert+ \int_0^s t\lVert D_n(\tau,t)\rVert\lVert G_{B,n}(\tau)\rVert\,d\tau.

Bound ∥Rn(s,t)∥\lVert R_n(s,t)\rVert by its supremum over s∈[0,1]s \in[0,1] and apply Grönwall’s inequality with respect to the variable ss and the absolutely continuous measure μ(dτ)=t∥GB,n(τ)∥ dτ\mu(d\tau)=t\lVert G_{B,n}(\tau)\rVert\,d\tau:

∥Dn(1,t)∥≤(sup⁡s∈[0,1]∥Rn(s,t)∥)exp⁡(∫01t∥GB,n(τ)∥ dτ).\lVert D_n(1,t)\rVert\le\left(\sup_{s\in[0,1]}\lVert R_n(s,t)\rVert\right)\exp\left(\int_0^1t\lVert G_{B,n}(\tau)\rVert\,d\tau\right).

Since {An}\{A_n\} is CB, there is a constant MM such that

∫01∥GA,n(τ)∥ dτ=1n∑i=1n∥Ai∥≤M.\int_0^1\lVert G_{A,n}(\tau)\rVert\,d\tau=\frac{1}{n}\sum_{i=1}^n\lVert A_i\rVert\le M.

Then, for all s∈[0,1]s\in[0,1] and t∈[0,T]t\in[0,T] we have

∥PA,n(s,t)∥≤exp⁡(tn∑i=1n∥Ai∥)≤eTM=:C.\lVert P_{A,n}(s,t)\rVert\le\exp\left(\frac{t}{n}\sum_{i=1}^n\lVert A_i\rVert\right)\le e^{TM}=:C.

Similarly, since {Bn}\{B_n\} is CB, ∫01∥GB,n(τ)∥ dτ≤M\int_0^1\lVert G_{B,n}(\tau)\rVert\,d\tau\le M (for sufficiently large MM). Hence,

∥Dn(1,t)∥≤Csup⁡s∈[0,1]∥Rn(s,t)∥.\lVert D_n(1,t)\rVert\le C\sup_{s\in[0,1]}\lVert R_n(s,t)\rVert.

Define Δn(s):=∫0s(GA,n(τ)−GB,n(τ)) dτ\Delta_n(s):=\int_0^s(G_{A,n}(\tau)-G_{B,n}(\tau))\,d\tau. Rewrite Rn(s,t)R_n(s,t) using dΔn(τ)=(GA,n(τ)−GB,n(τ)) dτd\Delta_n(\tau)=(G_{A,n}(\tau)-G_{B,n}(\tau))\,d\tau and integrate by parts:

Rn(s,t)=t∫0sPA,n(τ,t) dΔn(τ)=tPA,n(s,t)Δn(s)−t∫0s∂PA,n∂τ(τ,t)Δn(τ) dτ.R_n(s,t)=t\int_0^sP_{A,n}(\tau,t)\,d\Delta_n(\tau)=tP_{A,n}(s,t)\Delta_n(s)-t\int_0^s\frac{\partial P_{A,n}}{\partial\tau}(\tau,t)\Delta_n(\tau)\,d\tau.

Substitute ∂PA,n∂τ(τ,t)=tPA,n(τ,t)GA,n(τ)\frac{\partial P_{A,n}}{\partial\tau}(\tau,t)=tP_{A,n}(\tau,t)G_{A,n}(\tau):

Rn(s,t)=tPA,n(s,t)Δn(s)−t2∫0sPA,n(τ,t)GA,n(τ)Δn(τ) dτ.R_n(s,t)=tP_{A,n}(s,t)\Delta_n(s)-t^2\int_0^sP_{A,n}(\tau,t)G_{A,n}(\tau)\Delta_n(\tau)\,d\tau.

Taking norms and substituting the uniform bounds ∥PA,n∥≤C\lVert P_{A,n}\rVert\le C and t≤Tt\le T,

∥Rn(s,t)∥≤TC∥Δn(s)∥+T2C∫0s∥GA,n(τ)∥∥Δn(τ)∥ dτ.\lVert R_n(s,t)\rVert\le TC\lVert\Delta_n(s)\rVert+T^2C\int_0^s\lVert G_{A,n}(\tau)\rVert\lVert\Delta_n(\tau)\rVert\,d\tau.

We can factor out the supremum of Δn\Delta_n:

∥Rn(s,t)∥≤(TC+T2CM)sup⁡τ∈[0,1]∥Δn(τ)∥.\lVert R_n(s,t)\rVert\le(TC+T^2CM)\sup_{\tau\in[0,1]}\lVert\Delta_n(\tau)\rVert.

Notice that for τ=k/n\tau=k/n, we have

Δn(kn)=1n∑i=1k(Ai−Bi).\Delta_n\left(\frac{k}{n}\right)=\frac{1}{n}\sum_{i=1}^k(A_i-B_i).

Now we use the last assumption that ∥Δn(1)∥→0\|\Delta_n(1)\| \to0. This implies max⁡k=1,…,n∥Δn(k/n)∥→0\max_{k=1,\ldots,n}\|\Delta_n(k/n)\| \to0. Also, Δn\Delta_n is linear on each interval I‾k\overline{I}_k, so for any τ∈I‾k\tau\in\overline{I}_k we can bound the norm of Δn(τ)\Delta_n(\tau) by the norm at the endpoints:

∥Δn(τ)∥≤max⁡(∥Δn(k−1n)∥,∥Δn(kn)∥).\|\Delta_n(\tau)\| \le\max\left(\left\|\Delta_n\left(\frac{k-1}{n}\right)\right\|,\left\|\Delta_n\left(\frac{k}{n}\right)\right\|\right).

Thus, we have the uniform bound:

sup⁡τ∈[0,1]∥Δn(τ)∥=max⁡k=1,…,n∥Δn(kn)∥→0.\sup_{\tau\in[0,1]}\|\Delta_n(\tau)\|=\max_{k=1,\ldots,n}\left\|\Delta_n\left(\frac{k}{n}\right)\right\|\to0.

Combining the obtained estimates, we arrive at the desired convergence:

sup⁡t∈[0,T]∥Dn(1,t)∥≤Csup⁡t∈[0,T]sup⁡s∈[0,1]∥Rn(s,t)∥≤(TC2+T2C2M)sup⁡τ∈[0,1]∥Δn(τ)∥→n→∞0.\sup_{t\in[0,T]}\|D_n(1,t)\|\le C\sup_{t\in[0,T]}\sup_{s\in[0,1]}\|R_n(s,t)\|\le(TC^2+T^2C^2M)\sup_{\tau\in[0,1]}\|\Delta_n(\tau)\|\xrightarrow{n\to\infty}0.

Proof of Theorem 2.2. By hypothesis, there are sets E1∈σΩE_1\in\sigma_\Omega where sup⁡n∈N∑i=1n∣Ai∣n<∞\sup_{n\in\mathbb{N}}\sum_{i=1}^n\frac{|A_i|}{n}<\infty and E2∈σΩE_2\in\sigma_\Omega where ∑i=1nAi−EAin→0\sum_{i=1}^n\frac{A_i-\mathbb{E}A_i}{n}\to0 that have probability 1. Then so does E:=E1∩E2E:=E_1\cap E_2. By Lemma 3.4, it holds pointwise on EE that for all T>0T>0

lim⁡n→∞sup⁡t∈[0,T]∥eA1t/n⋯eAnt/n−eEA1t/n⋯eEAnt/n∥=0.\lim_{n\to\infty}\sup_{t\in[0,T]}\left\|e^{A_1t/n}\cdots e^{A_nt/n}-e^{\mathbb{E}A_1t/n}\cdots e^{\mathbb{E}A_nt/n}\right\|=0.

This means, by definition, that the sequence of random semigroups {eAnt}n∈N\{e^{A_nt}\}_{n\in\mathbb{N}} satisfies the SLLN in norm topology. For the additional part of the theorem, apply Lemma 3.4 to sequences {EAn}n∈N\{\mathbb{E}A_n\}_{n\in\mathbb{N}} and {A}n∈N\{A\}_{n\in\mathbb{N}}, and use the triangle inequality.

Counterexamples

The CB and CBE hypotheses in 2.2 seem quite restrictive because they force both eA1t/n⋯eAnt/ne^{A_1t/n}\cdots e^{A_nt/n} and eEA1t/n⋯eEAnt/ne^{\mathbb{E}A_1t/n}\cdots e^{\mathbb{E}A_nt/n} to be bounded (almost surely). However, we show that none of the three conditions can be discarded.

Example 4.1 (SLLN hypothesis in theorem 2.2 is significant). Take a sequence of independent random variables {ξn}n∈N\{\xi_n\}_{n\in\mathbb{N}} such that P{ξn=n2}=n−2\mathbb{P}\{\xi_n=n^2\}=n^{-2} and ξn=0\xi_n=0 otherwise. Then, Eξn=1\mathbb{E}\xi_n=1, but the Borel–Cantelli lemma implies that ξn≠0\xi_n\ne0 only finitely often almost surely, so 1n∑i=1n∣ξi∣→0\frac{1}{n}\sum_{i=1}^n|\xi_i|\to0 a.s. and 1n∑i=1n(ξi−Eξi)→−1\frac{1}{n}\sum_{i=1}^n(\xi_i-\mathbb{E}\xi_i)\to-1 a.s. (Actually, any sequence with properties CB and CBE but not SLLN would suffice.)

Then, in any non-zero Banach space, the sequence of random semigroups generated by {An=ξnI}n∈N\{A_n=\xi_n I\}_{n\in\mathbb{N}} does not satisfy the SLLN even in WOT:

eA1t/n⋯eAnt/n−eEA1t/n⋯eEAnt/n=(etn∑i=1nξi−etn∑i=1nEξi)I→(1−et)I.e^{A_1t/n}\cdots e^{A_nt/n}-e^{\mathbb{E}A_1t/n}\cdots e^{\mathbb{E}A_nt/n} =\left(e^{\frac{t}{n}\sum_{i=1}^n\xi_i}-e^{\frac{t}{n}\sum_{i=1}^n\mathbb{E}\xi_i}\right)I\to(1-e^t)I.

The examples for the other two combinations share a common pattern, so let us introduce it beforehand.

Example 4.2. Let {ξn,ηn}n∈N\{\xi_n,\eta_n\}_{n\in\mathbb{N}} be two sequences of integrable random variables that satisfy the SLLN but

lim⁡‾n→∞1n2∣∑1≤i<j≤n(ξiηj−EξiEηj)∣>0\overline{\lim}_{n\to\infty}\frac{1}{n^2}\left|\sum_{1\leq i<j\leq n}(\xi_i\eta_j-\mathbb{E}\xi_i\mathbb{E}\eta_j)\right|>0

almost surely (such sequences appear in the examples below). Consider any Banach space X\mathcal{X} with two non-random operators U,V∈L(X)U,V\in\mathcal{L}(\mathcal{X}) such that U2=V2=VU=0U^2=V^2=VU=0 but UV≠0UV\ne0. For example, the following operators in R3\mathbb{R}^3 fit:

U∼(010000000),V∼(000001000).U\sim\begin{pmatrix}0&1&0\\0&0&0\\0&0&0\end{pmatrix},\qquad V\sim\begin{pmatrix}0&0&0\\0&0&1\\0&0&0\end{pmatrix}.

Consider the following sequence of random operators: A2k−1:=ηkVA_{2k-1}:=\eta_kV and A2k:=ξkUA_{2k}:=\xi_kU for k∈Nk\in\mathbb{N}. Due to the properties of UU and VV, for n=2kn=2k we have:

eA1t/n⋯eAnt/n−eEA1t/n⋯eEAnt/n=∏i=1k(I+tnηiV)(I+tnξiU)−∏i=1k(I+tnEηiV)(I+tnEξiU)e^{A_1t/n}\cdots e^{A_nt/n}-e^{\mathbb{E}A_1t/n}\cdots e^{\mathbb{E}A_nt/n} =\prod_{i=1}^{k}\left(I+\frac{t}{n}\eta_iV\right)\left(I+\frac{t}{n}\xi_iU\right)-\prod_{i=1}^{k}\left(I+\frac{t}{n}\mathbb{E}\eta_iV\right)\left(I+\frac{t}{n}\mathbb{E}\xi_iU\right)
=tn∑i=1k(ξi−Eξi)U+tn∑i=1k(ηi−Eηi)V+t2n2∑1≤i<j≤k(ξiηj−EξiEηj)UV.=\frac{t}{n}\sum_{i=1}^{k}(\xi_i-\mathbb{E}\xi_i)U+\frac{t}{n}\sum_{i=1}^{k}(\eta_i-\mathbb{E}\eta_i)V+\frac{t^2}{n^2}\sum_{1\leq i<j\leq k}(\xi_i\eta_j-\mathbb{E}\xi_i\mathbb{E}\eta_j)UV.

Because {ξn}\{\xi_n\} and {ηn}\{\eta_n\} satisfy the SLLN, the linear terms vanish almost surely. So, assuming t>0t>0,

lim⁡‾n→∞∥eA1t/n⋯eAnt/n−eEA1t/n⋯eEAnt/n∥≥lim⁡k→∞t24k2∣∑1≤i<j≤k(ξiηj−EξiEηj)∣∥UV∥>0a.s.\overline{\lim}_{n\to\infty}\left\|e^{A_1t/n}\cdots e^{A_nt/n}-e^{\mathbb{E}A_1t/n}\cdots e^{\mathbb{E}A_nt/n}\right\| \geq\lim_{k\to\infty}\frac{t^2}{4k^2}\left|\sum_{1\leq i<j\leq k}(\xi_i\eta_j-\mathbb{E}\xi_i\mathbb{E}\eta_j)\right|\|UV\|>0\quad\text{a.s.}

Thus, the SLLN fails for {eAnt}n∈N\{e^{A_nt}\}_{n\in\mathbb{N}}.

Example 4.3 (CBE hypothesis in theorem 2.2 is significant). Let {ξn,ηn}n∈N\{\xi_n,\eta_n\}_{n\in\mathbb{N}} be two sequences of jointly independent random variables such that for all nn, P{ξn=(−1)nn2}=n−3/2\mathbb{P}\{\xi_n=(-1)^n n^2\}=n^{-3/2}, ξn=0\xi_n=0 otherwise, and ηn\eta_n has the same distribution. Note that Eξn=(−1)nn\mathbb{E}\xi_n=(-1)^n\sqrt{n}. By the Borel–Cantelli lemma, ξn=ηn=0\xi_n=\eta_n=0 for all but finitely many nn almost surely (which implies boundedness). Denote this event by EE. On EE, we have

1n∑k=1n(ξk−Eξk)∼−1n∑k=1nEξk=−1n∑k=1n(−1)kk=O(nn)→0.\frac{1}{n}\sum_{k=1}^{n}(\xi_k-\mathbb{E}\xi_k)\sim-\frac{1}{n}\sum_{k=1}^{n}\mathbb{E}\xi_k=-\frac{1}{n}\sum_{k=1}^{n}(-1)^k\sqrt{k}=O\left(\frac{\sqrt{n}}{n}\right)\to0.

The same applies to ηn\eta_n. Also, observe that on EE,

1n2∑1≤i<j≤n(ξiηj−EξiEηj)∼−1n2∑1≤i<j≤nEξiEξj=−12(1n∑k=1nEξk)2+12n2∑k=1n(Eξk)2∼12n2∑k=1nk∼12n∑k=1nk∼14.\frac{1}{n^2}\sum_{1\leq i<j\leq n}(\xi_i\eta_j-\mathbb{E}\xi_i\mathbb{E}\eta_j)\sim-\frac{1}{n^2}\sum_{1\leq i<j\leq n}\mathbb{E}\xi_i\mathbb{E}\xi_j=-\frac{1}{2}\left(\frac{1}{n}\sum_{k=1}^{n}\mathbb{E}\xi_k\right)^2+\frac{1}{2n^2}\sum_{k=1}^{n}(\mathbb{E}\xi_k)^2\sim\frac{1}{2n^2}\sum_{k=1}^{n}k\sim\frac{1}{2n}\sum_{k=1}^{n}k\sim\frac{1}{4}.

Now, consider the random operators {An}n∈N\{A_n\}_{n\in\mathbb{N}} from Example 4.2. This sequence is bounded almost surely and satisfies the SLLN because {ξn}\{\xi_n\} and {ηn}\{\eta_n\} do. Additionally, the operators are independent. However, as the example shows, the sequence {eAnt}n∈N\{e^{A_nt}\}_{n\in\mathbb{N}} does not satisfy the SLLN.

Example 4.4 (CB hypothesis in theorem 2.2 is significant). Let {ξn,ηn}n∈N\{\xi_n,\eta_n\}_{n\in\mathbb{N}} be two sequences of jointly independent random variables such that for all nn, both ξn\xi_n and ηn\eta_n are distributed as follows:

{P{ξn=(−1)nn(1−n2)}=n−2,P{ξn=(−1)nn}=1−n−2.\begin{cases} \mathbb{P}\{\xi_n=(-1)^n\sqrt{n(1-n^2)}\}=n^{-2},\\ \mathbb{P}\{\xi_n=(-1)^n\sqrt{n}\}=1-n^{-2}. \end{cases}

Here, Eξn=Eηn=0\mathbb{E}\xi_n=\mathbb{E}\eta_n=0, but the Borel–Cantelli lemma ensures that ξn=ηn=(−1)nn\xi_n=\eta_n=(-1)^n\sqrt{n} for all but finitely many nn on a set E∈σΩE\in\sigma_\Omega of probability 1 (which is opposite to the previous example). Then, on EE, we similarly have 1n∑i=1nξi→0\frac{1}{n}\sum_{i=1}^{n}\xi_i\to0, 1n∑i=1nηi→0\frac{1}{n}\sum_{i=1}^{n}\eta_i\to0, and

1n2∑1≤i<j≤nξiηj∼1n2∑1≤i<j≤n(−1)i+jij=12(1n∑k=1n(−1)kk)2−12n2∑k=1nk∼−12n2∑k=1nk∼−14.\frac{1}{n^2}\sum_{1\le i<j\le n}\xi_i\eta_j\sim\frac{1}{n^2}\sum_{1\le i<j\le n}(-1)^{i+j}\sqrt{ij}=\frac{1}{2}\left(\frac{1}{n}\sum_{k=1}^{n}(-1)^k\sqrt{k}\right)^2-\frac{1}{2n^2}\sum_{k=1}^{n}k\sim-\frac{1}{2n^2}\sum_{k=1}^{n}k\sim-\frac{1}{4}.

Then, the sequence {An}n∈N\{A_n\}_{n\in\mathbb{N}} from Example 4.2 satisfies the SLLN and EAn=0\mathbb{E}A_n=0 for all nn, but the corresponding sequence of random semigroups does not satisfy the SLLN.

We also show that the direct analog of Theorem 2.2 for SOT or WOT does not hold in general, even for independent random operators.

Example 4.5. We construct a sequence of random operators {An}n∈N\{A_n\}_{n\in\mathbb{N}} in X=ℓ2(N0)\mathcal{X}=\ell_2(\mathbb{N}_0) (which can be replaced with any infinite-dimensional Banach space that admits a Schauder basis) with the following properties:

  • the random operators are independent;

  • CB;

  • CBE;

  • for all x∈Xx\in\mathcal{X}, the sequence {Anx}n∈N\{A_nx\}_{n\in\mathbb{N}} satisfies the SLLN;

  • the sequence of random semigroups {etAn}n∈N\{e^{tA_n}\}_{n\in\mathbb{N}} does not satisfy the SLLN in WOT.

For m∈Nm\in\mathbb{N}, define Km,Rm∈L(X)K_m,R_m\in\mathcal{L}(\mathcal{X}) as

Kmx:=xme0,Rmx:=x0em−xme0.K_mx:=x_me_0,\qquad R_mx:=x_0e_m-x_me_0.

Choose a sequence {nm}m∈N\{n_m\}_{m\in\mathbb{N}} such that

∑k=1m−1nknm→0,nm>2nm−1,\frac{\sum_{k=1}^{m-1}n_k}{n_m}\to0,\qquad n_m>2n_{m-1},

e.g., nm=2m2n_m=2^{m^2}. Let {εm}m∈N\{\varepsilon_m\}_{m\in\mathbb{N}} be a sequence of independent Rademacher variables, i.e., P{εm=1}=P{εm=−1}=1/2\mathbb{P}\{\varepsilon_m=1\}=\mathbb{P}\{\varepsilon_m=-1\}=1/2. Define

Anm:=nmεmKm,A2nm:=nmRm,A_{n_m}:=n_m\varepsilon_mK_m,\qquad A_{2n_m}:=n_mR_m,

and An:=0A_n:=0 for other nn. The random operators {An}n∈N\{A_n\}_{n\in\mathbb{N}} are independent since the only non-constant ones depend on distinct εm\varepsilon_m.

We have EA2nm=nmRm\mathbb{E}A_{2n_m}=n_mR_m and EAn=0\mathbb{E}A_n=0 for other nn, so the growth rate of nmn_m gives

sup⁡n∈N1n∑i=1n∥EAi∥=sup⁡m∈N12nm∑k=1m∥EA2nk∥=sup⁡m∈N12nm∑k=1mnk<∞.\sup_{n\in\mathbb{N}}\frac{1}{n}\sum_{i=1}^{n}\|\mathbb{E}A_i\|=\sup_{m\in\mathbb{N}}\frac{1}{2n_m}\sum_{k=1}^{m}\|\mathbb{E}A_{2n_k}\|=\sup_{m\in\mathbb{N}}\frac{1}{2n_m}\sum_{k=1}^{m}n_k<\infty.

Similarly,

sup⁡n∈N1n∑i=1n∥Ai∥≤sup⁡m∈N1nm∑k=1m(∥Ank∥+∥A2nk∥)=sup⁡m∈N2nm∑k=1mnk<∞.\sup_{n\in\mathbb{N}}\frac{1}{n}\sum_{i=1}^{n}\lVert A_i\rVert\le\sup_{m\in\mathbb{N}}\frac{1}{n_m}\sum_{k=1}^{m}\left(\lVert A_{n_k}\rVert+\lVert A_{2n_k}\rVert\right)=\sup_{m\in\mathbb{N}}\frac{2}{n_m}\sum_{k=1}^{m}n_k<\infty.

Also, for every xx and n∈[nm,nm+1)n\in[n_m,n_{m+1}),

1n∥∑i=1n(Ai−EAi)x∥=1n∥∑k=1mnkεkxke0∥≤1nm(∑k=1m−1nk∥x∥+nm∣xm∣)=1nm∑k=1m−1nk∥x∥+∣xm∣→n→∞0.\frac{1}{n}\left\lVert\sum_{i=1}^{n}(A_i-\mathbb{E}A_i)x\right\rVert=\frac{1}{n}\left\lVert\sum_{k=1}^{m}n_k\varepsilon_kx_ke_0\right\rVert\le\frac{1}{n_m}\left(\sum_{k=1}^{m-1}n_k\lVert x\rVert+n_m|x_m|\right)=\frac{1}{n_m}\sum_{k=1}^{m-1}n_k\lVert x\rVert+|x_m|\xrightarrow[n\to\infty]{}0.

So, the weakened assumptions are satisfied.

For m≥2m\ge2, denote

Pn(t):=eA1t/n…eAnt/n,Um(t):=P2nm−1(nm−1nmt),P_n(t):=e^{A_1t/n}\ldots e^{A_nt/n},\qquad U_m(t):=P_{2n_{m-1}}\left(\frac{n_{m-1}}{n_m}t\right),
Qn(t):=eEA1t/n…eEAnt/n,Vm(t):=Q2nm−1(nm−1nmt).Q_n(t):=e^{\mathbb{E}A_1t/n}\ldots e^{\mathbb{E}A_nt/n},\qquad V_m(t):=Q_{2n_{m-1}}\left(\frac{n_{m-1}}{n_m}t\right).

I.e., UmU_m and VmV_m are the products of the first 2nm−12n_{m-1} factors in P2nmP_{2n_m} and Q2nmQ_{2n_m} respectively. Then, for m≥2m\ge2, the following decomposition holds:

P2nm(t)−Q2nm(t)=Um(t)eεmKmt/2eRmt/2−Vm(t)eRmt/2.P_{2n_m}(t)-Q_{2n_m}(t)=U_m(t)e^{\varepsilon_mK_mt/2}e^{R_mt/2}-V_m(t)e^{R_mt/2}.

Since Km2=0K_m^2=0, the factor eεmKmt/2e^{\varepsilon_mK_mt/2} reduces to I+εmKmt/2I+\varepsilon_mK_mt/2, so we can group the terms as

P2nm(t)−Q2nm(t)=(Um(t)−Vm(t))eRmt/2+t2εmUm(t)KmeRmt/2.P_{2n_m}(t)-Q_{2n_m}(t)=(U_m(t)-V_m(t))e^{R_mt/2}+\frac{t}{2}\varepsilon_mU_m(t)K_me^{R_mt/2}.

From the mean value inequality in Banach spaces, we get

∥Um(t)−I∥≤t2nm(∑i=12nm−1∥Ai∥)∏i=12nm−1e∥Ai∥t/(2nm)=tnm(∑k=1m−1nk)e(t/nm)∑k=1m−1nk→m→∞0,\lVert U_m(t)-I\rVert\le\frac{t}{2n_m}\left(\sum_{i=1}^{2n_{m-1}}\lVert A_i\rVert\right)\prod_{i=1}^{2n_{m-1}}e^{\lVert A_i\rVert t/(2n_m)}=\frac{t}{n_m}\left(\sum_{k=1}^{m-1}n_k\right)e^{(t/n_m)\sum_{k=1}^{m-1}n_k}\xrightarrow[m\to\infty]{}0,

and

∥Vm(t)−I∥≤t2nm(∑i=12nm−1∥EAi∥)∏i=12nm−1e∥EAi∥t/(2nm)=t2nm(∑k=1m−1nk)et/(2nm)∑k=1m−1nk→m→∞0.\lVert V_m(t)-I\rVert\le\frac{t}{2n_m}\left(\sum_{i=1}^{2n_{m-1}}\lVert\mathbb{E}A_i\rVert\right)\prod_{i=1}^{2n_{m-1}}e^{\lVert\mathbb{E}A_i\rVert t/(2n_m)}=\frac{t}{2n_m}\left(\sum_{k=1}^{m-1}n_k\right)e^{t/(2n_m)\sum_{k=1}^{m-1}n_k}\xrightarrow[m\to\infty]{}0.

Now, it is easy to see that the SLLN in WOT fails on the vector e0e_0 and the coordinate functional f0f_0. Indeed, note that

eRmt/2e0=cos⁡(t/2)e0+sin⁡(t/2)em.e^{R_mt/2}e_0=\cos(t/2)e_0+\sin(t/2)e_m.

So, using the triangle inequality, the estimates above and the fact that ∥Km∥=∥Rm∥=1\lVert K_m\rVert=\lVert R_m\rVert=1, we can write

f0((P2nm(t)−Q2nm(t))e0)∼t2εmf0(KmeRmt/2e0)=t2εmsin⁡(t/2).f_0\left((P_{2n_m}(t)-Q_{2n_m}(t))e_0\right)\sim\frac{t}{2}\varepsilon_mf_0\left(K_me^{R_mt/2}e_0\right)=\frac{t}{2}\varepsilon_m\sin(t/2).

So, for t∉2πZt\notin2\pi\mathbb{Z}, the convergence in WOT fails.

Lastly, we show that the order of composition in the definition of the SLLN for random semigroups is significant in general.

Example 4.6. Let X=R2\mathcal{X} = \mathbb{R}^{2}. Take linear operators A:(x,y)↦(y,0)A:(x,y)\mapsto(y,0) and B:(x,y)↦(x,0)B:(x,y)\mapsto(x,0) on X\mathcal{X}. Set A1=εAA_{1}=\varepsilon A for ε\varepsilon a Rademacher random variable and An=(n−1)BA_{n}=(n-1)B for n≥2n\ge2. Then,

eA1t/n⋯eAnt/n−eEA1t/n⋯eEAnt/n=(eεAt/n−eEεAt/n)e∑k=2nB(k−1)t/n=tnεAeBt(n−1)/2=:Dn(t).e^{A_{1}t/n}\cdots e^{A_{n}t/n}-e^{\mathbb{E}A_{1}t/n}\cdots e^{\mathbb{E}A_{n}t/n}=\left(e^{\varepsilon At/n}-e^{\mathbb{E}\varepsilon At/n}\right)e^{\sum_{k=2}^{n}B(k-1)t/n}=\frac{t}{n}\varepsilon Ae^{Bt(n-1)/2}=:D_{n}(t).

We have Dn(t)(x,y)=tnεA(et(n−1)/2x,y)=tnε(y,0)D_{n}(t)(x,y)=\frac{t}{n}\varepsilon A(e^{t(n-1)/2}x,y)=\frac{t}{n}\varepsilon(y,0), so sup⁡t∈[0,T]∥Dn(t)∥→0\sup_{t\in[0,T]}\|D_{n}(t)\|\to0. However, the reversed order of composition gives

eAnt/n⋯eA1t/n−eEAnt/n⋯eEA1t/n=e∑k=2nB(k−1)t/n(eεAt/n−eEεAt/n)=tnεeBt(n−1)/2A=:En(t).e^{A_{n}t/n}\cdots e^{A_{1}t/n}-e^{\mathbb{E}A_{n}t/n}\cdots e^{\mathbb{E}A_{1}t/n}=e^{\sum_{k=2}^{n}B(k-1)t/n}\left(e^{\varepsilon At/n}-e^{\mathbb{E}\varepsilon At/n}\right)=\frac{t}{n}\varepsilon e^{Bt(n-1)/2}A=:E_{n}(t).

Here, En(t)(x,y)=tnεeBt(n−1)/2(y,0)=tnε(et(n−1)/2y,0)E_{n}(t)(x,y)=\frac{t}{n}\varepsilon e^{Bt(n-1)/2}(y,0)=\frac{t}{n}\varepsilon(e^{t(n-1)/2}y,0), so sup⁡t∈[0,T]∥En(t)∥→∞\sup_{t\in[0,T]}\|E_{n}(t)\|\to\infty for all T>0T>0. For a counterexample to the other implication, define A:(x,y)↦(0,x)A:(x,y)\mapsto(0,x) instead.

Remarks and further research

The proof of Theorem 2.2 actually gives the following estimate for some C=C(T,ω)C=C(T,\omega):

sup⁡t∈[0,T]∥eA1t/n⋯eAnt/n−eEA1t/n⋯eEAnt/n∥≤Cnmax⁡k=1,…,n∥∑i=1k(Ai−EAi)∥.\sup_{t\in[0,T]}\left\|e^{A_{1}t/n}\cdots e^{A_{n}t/n}-e^{\mathbb{E}A_{1}t/n}\cdots e^{\mathbb{E}A_{n}t/n}\right\|\le\frac{C}{n}\max_{k=1,\ldots,n}\left\|\sum_{i=1}^{k}(A_{i}-\mathbb{E}A_{i})\right\|.

Thus, the left-hand side automatically inherits any asymptotic upper bound for the right-hand side. For the latter, there is a classical result due to de Acosta [26] and independently to Azlarov and Volodin [27], which was initially proved for random variables by Marcinkiewicz and Zygmund [28]:

Theorem 5.1. Let X\mathcal{X} be a separable Banach space and let p∈[1,2)p\in[1,2). The following conditions are equivalent:

(1) X\mathcal{X} is of Rademacher type pp (see [12] for the definition);

(2) for any sequence of i.i.d. X\mathcal{X}-valued random elements {xn}n∈N\{x_{n}\}_{n\in\mathbb{N}} with Exn=0\mathbb{E}x_{n}=0 and E∥xn∥p<∞\mathbb{E}\|x_{n}\|^{p}<\infty,

1n1/p∑i=1nxi→0almost surely.\frac{1}{n^{1/p}}\sum_{i=1}^{n}x_{i}\to0\quad\text{almost surely}.

So, if {An}n∈N\{A_{n}\}_{n\in\mathbb{N}} is a sequence of i.i.d. Bochner integrable random operators such that E∥An∥p<∞\mathbb{E}\|A_{n}\|^{p}<\infty, and the closed linear span of the essential range of A1A_{1} is of Rademacher type pp, then almost surely

1n1/p∑i=1n(Ai−EAi)→0.\frac{1}{n^{1/p}}\sum_{i=1}^{n}(A_{i}-\mathbb{E}A_{i})\to0.

By the standard argument, this implies

1n1/pmax⁡k=1,…,n∥∑i=1k(Ai−EAi)∥→0,\frac{1}{n^{1/p}}\max_{k=1,\ldots,n}\left\|\sum_{i=1}^{k}(A_{i}-\mathbb{E}A_{i})\right\|\to0,

which gives the estimate

sup⁡t∈[0,T]∥eA1t/n⋯eAnt/n−eEA1t∥=o(n1p−1).\sup_{t\in[0,T]}\left\|e^{A_{1}t/n}\cdots e^{A_{n}t/n}-e^{\mathbb{E}A_{1}t}\right\|=o\left(n^{\frac{1}{p}-1}\right).

However, one should keep in mind that Rademacher type p>1p>1 is not automatic here unless the space is finite-dimensional. The essential range is necessarily separable due to Bochner measurability of AnA_{n}, but this does not imply non-trivial Rademacher type.

In practice, the assumptions of Theorem 2.2 might be infeasible to verify directly. To obtain special cases of the theorem, one imposes additional restrictions on either the range of the random operators or their measure-theoretic properties. The former includes independent uniformly bounded operators whose essential ranges are contained in a common BB-convex subspace or a compact subset of L(X)\mathcal{L}(\mathcal{X}). The latter includes Corollary 2.3 and stationary ergodic sequences of random operators (see [29, Section 3] for the corresponding ergodic theorem in Banach spaces). In general, there is a slightly stronger set of sufficient conditions:

sup⁡n∈N∑i=1nE∥Ai∥n<∞,\sup_{n\in\mathbb{N}} \frac{\sum_{i=1}^{n} \mathbb{E}\left\lVert A_i\right\rVert}{n}<\infty,
∑i=1n(∥Ai∥−E∥Ai∥)n→a.s.0,\frac{\sum_{i=1}^{n}\left(\left\lVert A_i\right\rVert-\mathbb{E}\left\lVert A_i\right\rVert\right)}{n}\xrightarrow{\text{a.s.}}0,
∑i=1n(Ai−EAi)n→a.s.0,\frac{\sum_{i=1}^{n}\left(A_i-\mathbb{E}A_i\right)}{n}\xrightarrow{\text{a.s.}}0,

where the second condition could be derived from one of the SLLN-type theorems for random variables. However, as already mentioned, the SLLN in Banach spaces may fail even if the random elements are independent and uniformly bounded but not identically distributed.

For non-identically distributed random semigroups, moving from norm topology to SOT or WOT could prove beneficial. Example 4.5 looks somewhat unnatural because it exploits very rare spikes of linear amplitude in the sequence. We believe that the SLLN in SOT and WOT can still be generalized under some stronger asymptotic assumptions. Moreover, it is reasonable to expect some connection between the SLLN for non-identically distributed random semigroups and the properties of the Banach space, as it exists in the case of the SLLN for random elements.

Acknowledgements

The author is grateful to S. V. Dzhenzher for valuable discussion and review throughout the creation of this paper.

References

  1. [1]M. L. Mehta. Random matrices. Amsterdam: Elsevier, 2004.
  2. [2]H. Furstenberg. “Non-commuting random products”. In: Trans. AMS 108.3 (1963), pp. 377–428.
  3. [3]V. N. Tutubalin. “A local limit theorem for products of random matrices”. In: Theory Probab. Appl. 22.2 (1978), pp. 203–214. DOI: 10.1137/1122028.DOI
  4. [4]A. V. Skorokhod. “Products of independent random operators”. In: Russ. Math. Surv. 38.(4) (1983), pp. 291–318. DOI: 10.1070/RM1983v038n04ABEH004213.DOI
  5. [5]J. Kempe. “Quantum random walks: an introductory overview”. In: Contemp. Phys. 44(4) (2003), pp. 307–327. DOI: 10.1080/00107151031000110776.DOI
  6. [6]Yu. N. Orlov, V. Zh. Sakbaev, and O. G. Smolyanov. “Feynman Formulas and the Law of Large Numbers for Random One-Parameter Semigroups”. In: Proc. Steklov Inst. Math. 306 (2019), pp. 196–211. DOI: 10.1134/S0081543819050171.DOI
  7. [7]Y. Aharonov, L. Davidovich, and N. Zagury. “Quantum random walks”. In: Phys. Rev. A (3) 48.2 (1993), pp. 1687–1690. DOI: 10.1103/PhysRevA.48.1687.DOI
  8. [8]A. S. Kholevo. Quantum probability and quantum statistics. Vol. 83. VINITI, Moscow: Probability theory -8, Itogi Nauki i Tekhniki. Series Sovrem. Probl. Mat. Fund. Napr., 1991, pp. 5–132.
  9. [9]J. Gough et al. “Markov Approximations of the Evolution of Quantum Systems”. In: Dokl. Math. 105.2 (2022), pp. 92–96. DOI: 10.1134/S1064562422020107.DOI
  10. [10]L. Pathirana and J. Schenker. “Law of large numbers and central limit theorem for ergodic quantum processes”. In: J. Math. Phys. 64.8 (Aug. 2023), p. 082201. DOI: 10.1063/5.0153483. eprint: https://pubs.aip.org/aip/jmp/article-pdf/doi/10.1063/5.0153483/18103512/082201_1_5.0153483.pdf.DOI
  11. [11]S. Dhamapurkar and O. Dahlsten. “Quantum walks as thermalisations, with application to fullerene graphs”. In: Physica A: Stat. Mech. Appl. 644 (2024), p. 129823. ISSN: 0378-4371. DOI: 10.1016/j.physa.2024.129823. URL: https://www.sciencedirect.com/science/article/pii/S0378437124003327.DOI
  12. [12]M. Ledoux and M. Talagrand. Probability in Banach Spaces: Isoperimetry and Processes. A Series of Modern Surveys in Mathematics Series. Springer, 1991. ISBN: 9783540520139. URL: https://books.google.ru/books?id=cykYDfvxRjSC.DOI
  13. [13]J. Hoffmann-Jørgensen and G. Pisier. “The Law of Large Numbers and the Central Limit Theorem in Banach Spaces”. In: Ann. Prob. 4.4 (1976), pp. 587–599. DOI: 10.1214/aop/1176996029.DOI
  14. [14]V. I. Oseledets. “Markov chains, skew products and ergodic theorems for general dynamic systems”. In: Theory Probab. Appl. 10.3 (1965), pp. 551–557. DOI: 10.1137/1110062.DOI
  15. [15]V. Zh. Sakbaev. “On the law of large numbers for compositions of independent random semigroups”. In: Russ. Math. 60.(10) (2016), pp. 72–76.DOI
  16. [16]V. Zh. Sakbaev. “Averaging of random flows of linear and nonlinear maps”. In: J. Phys. Conf. Ser. 990 (2018), p. 012012.DOI
  17. [17]P. R. Chernoff. “Note on product formulas for operator semigroups”. In: J. Funct. Anal. 2.2 (1968), pp. 238–242. ISSN: 0022-1236. DOI: 10.1016/0022-1236(68)90020-7.DOI
  18. [18]K.-J. Engel and R. Nagel. “Perturbation and Approximation of Semigroups”. In: One-Parameter Semigroups for Linear Evolution Equations. New York, NY: Springer New York, 2000, pp. 157–237. ISBN: 978-0-387-22642-2. DOI: 10.1007/0-387-22642-7_3.
  19. [19]O. E. Galkin and I. D. Remizov. “Upper and lower estimates for rate of convergence in the Chernoff product formula for semigroups of operators”. In: Isr. J. Math. 265 (2 Feb. 2025), pp. 929–943. DOI: 10.1007/s11856-024-2678-x.DOI
  20. [20]S. V. Dzhentzher and V. Zh. Sakbaev. “The Strong Law of Large Numbers for random semigroups on uniformly smooth Banach spaces”. In: Adv. Oper. Theory 11 (1 Nov. 2025), p. 4. DOI: 10.1007/s43036-025-00481-7.
  21. [21]S. V. Dzhentzher and V. Zh. Sakbaev. “The Strong Law of Large Numbers for random semigroups with unbounded generators on uniformly smooth Banach spaces”. In: Infin. Dimens. Anal. Quantum Probab. Relat. Top. (2026). DOI: 10.1142/S0219025726500049.
  22. [22]J. Diestel and J. J. Uhl. Vector Measures. Math. Surv. Monogr. AMS, 1977. ISBN: 9780821815151. URL: https://books.google.ru/books?id=fEKZAwAAQBAJ.
  23. [23]E. Mourier. “Éléments aléatoires dans un espace de Banach”. In: Ann. Inst. H. Poincaré 13.3 (1953), pp. 161–244. URL: http://www.numdam.org/item/AIHP_1953__13_3_161_0/.
  24. [24]A. Beck. “A convexity condition in Banach spaces and the strong law of large numbers”. In: Proc. Amer. Math. Soc. 13.2 (1962), pp. 329–334. DOI: 10.1090/S0002-9939-1962-0133857-9.DOI
  25. [25]S. N. Ethier and T. G. Kurtz. Markov Processes: Characterization and Convergence. Wiley Series in Probability and Mathematical Statistics. New York: John Wiley & Sons, 1986. DOI: http://dx.doi.org/10.1002/9780470316658.
  26. [26]A. de Acosta. “Inequalities for B-valued random vectors with applications to the strong law of large numbers”. In: Ann. Probab. 9.1 (1981), pp. 157–161. DOI: 10.1214/aop/1176994517.DOI
  27. [27]T. A. Azlarov and N. A. Volodin. “Laws of large numbers for identically distributed Banach-space valued random variables”. In: Theory Probab. Appl. 26.3 (1981). Russian original: Teor. Veroyatnost. i Primenen. 26 (1981), no. 3, 584–590, pp. 573–580.DOI
  28. [28]J. Marcinkiewicz and A. Zygmund. “Sur les fonctions indépendantes”. In: Fund. Math. 29 (1937), pp. 60–90. DOI: 10.4064/fm-29-1-60-90.DOI

Paper details

vlnc:W_nntmk8pp
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents