Convergence of permuted products of exponentials

Michael Anshelevich, Anh Nguyen · 2025

Read Convergence of permuted products of exponentials by Michael Anshelevich, Anh Nguyen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-02681.txt.

Introduction

Michael Anshelevich and Anh Nguyen proved in 2025 a result about multiplying many exponentials of matrices in random order. They consider a triangular array whose row averages settle toward a fixed limit. The entries are shuffled by a random permutation, and their exponentials are multiplied over any stretch of the array.

When the array contains few distinct elements relative to its size, the shuffled product almost surely converges to a single clean exponential of the limit. For ordinary finite matrices, the restriction is not needed at all. The proof combines a general estimate for abstract algebras with a concentration inequality for random matrices. The result clarifies when disorder in multiplication order washes out in the limit.

Catalog record

CREATORMichael Anshelevich, Anh Nguyen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-02681

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Convergence of permuted products of exponentials Michael Anshelevich, Anh Nguyen Abstract . Let { A i , n } \left\{A_{i,n}\right\} be a triangular array of elements in a Banach algebra, whose norms do not grow too fast, and whose row averages converge to A A . Let σ ∈ S ⁡ ( n ) \sigma\in S(n) be a permutation drawn uniformly at random. If the array only contains o ⁡ ( n / log ⁡ n ) o(n/\log n) distinct elements, then almost surely, for all 0 < s < t < 1 0<s<t<1 , the permuted product of their exponentials ∏ i = [ s ​ n ] [ t ​ n ] e A σ ⁡ ( i ) , n / n \prod_{i=[sn]}^{[tn]}e^{A_{\sigma(i),n}/n} converges in norm to e ( t − s ) ​ A e^{(t-s)A} . For an array of finite-dimensional matrices, convergence holds without this restriction. The proof of the latter result consists of an estimate valid in a general Banach algebra, and an application of a matrix concentration inequality. 2010 Mathematics Subject Classification Primary 15A16 1. Introduction Let \mathcal{M} be a Banach algebra, A\in\mathcal{M} , and \left\{A_{i,n}:1\leq i\leq n,n\in\mathbb{N}\right\}\subseteq\mathcal{M} a triangular array whose row averages \frac{1}{n}\sum_{i=1}^{n}A_{i,n}\rightarrow A. It is easy to come up with examples, even of two-by-two matrices, for which \prod_{i=1}^{n}e^{A_{i,n}/n}\not\rightarrow e^{A}. In this note we show that, while possible, such behavior is quite unusual. Under two sets of assumptions, for most rearrangements of the array \left\{A_{i,n}:1\leq i\leq n\right\} , not only do the corresponding products converge to e^{A} , but in fact we have simultaneous approximation of e^{tA} for all t\in[0,1] . More precisely, for each n , pick a permutation \sigma\in S(n) uniformly at random. Then under the assumptions listed below, permuted products converge uniformly in probability (1) \mathnormal{P}\left(\left\|{\prod_{i=1}^{[tn]}e^{A_{\sigma(i),n}/n}-e^{tA}}\right\|<\varepsilon\text{ for all }0\leq t\leq 1\right)\rightarrow 1\qquad\forall\varepsilon>0 or even uniformly almost surely (2) \mathnormal{P}\left(\left\|{\prod_{i=1}^{[tn]}e^{A_{\sigma(i),n}/n}-e^{tA}}\right\|\rightarrow 0\text{ for all }0\leq t\leq 1\right)=1. The first type of assumption is when the n elements in each row of the array fall into only o(n/logn) distinct types. Theorem 1 . In a Banach algebra \mathcal{M} , let A\in\mathcal{M} , and \left\{A_{i,n}:1\leq i\leq n,n\in\mathbb{N}\right\}\subseteq\mathcal{M} be a uniformly bounded triangular array which satisfies \frac{1}{n}\sum_{i=1}^{n}A_{i,n}\rightarrow A . Let c_{n}=o(n/\log n) be a sequence of positive integers. Suppose that for each n , there is a family \left\{B_{j,n}:1\leq j\leq c_{n},n\in\mathbb{N}\right\} and an assignment j(i) such that \frac{1}{n}\sum_{i=1}^{n}\left\|{A_{i,n}-B_{j(i),n}}\right\|=o(1). Then permuted products converge almost surely in the sense of equation ( 2 ). If the Banach algebra \mathcal{M} has dimension d<\infty , the unit ball in \mathcal{M} has an \varepsilon -net with \sim\varepsilon^{-d} elements. It follows that for matrix algebras, the assumptions of Theorem 1 hold for any uniformly bounded array, and the corresponding permuted products converge uniformly almost surely. One can allow the bound on the array to grow with n , at the cost of lower growth rate for c_{n} . But in the finite-dimensional case, a better growth rate is achieved by a different approach, Theorem 2 . Let A and \left\{A_{i,n}\right\} be as above in the algebra of d\times d complex matrices (with the operator norm). Denote L^{\infty}_{n}=\max_{1\leq i\leq n}\left\|{A_{i,n}}\right\|,\quad L^{1}_{n}=\frac{1}{n}\sum_{i=1}^{n}\left\|{A_{i,n}}\right\|. If for some \delta>0 , (3) (L^{1}_{n})^{4+\delta}e^{3L^{1}_{n}}L^{\infty}_{n}=o(n), then we have uniform convergence ( 1 ) of permuted products in probability. If in addition, (4) (L^{1}_{n})^{3}e^{3L^{1}_{n}}L^{\infty}_{n}=o(n/\log n). the uniform convergence ( 2 ) is almost surely. We next list several examples where the assumptions (in the matrix case) take a more concrete form. Example 3 . If L^{\infty}_{n}=o(n) and (L^{1}_{n}) are uniformly bounded, we have convergence of the permuted products in probability. Recall that these are precisely the standard assumptions on a triangular array of scalars with \sum_{i=1}^{n}\alpha_{i,n}/n\rightarrow\alpha to satisfy \prod_{i=1}^{n}\left(1+\frac{\alpha_{i,n}}{n}\right)\rightarrow e^{\alpha}. If L^{\infty}_{n}=o(n/\log n) , the convergence is a.s. The following is a more general example. Example 4 . Take k_{n} matrices of the same norm L^{\infty}_{n}=o(n) , with the remaining n-k_{n} matrices uniformly bounded in n . Then for k_{n}=\frac{n}{3L^{\infty}_{n}}(\log\frac{n}{L^{\infty}_{n}}-(4+2\delta)\log\log\frac{n}{L^{\infty}_{n}}), we have L^{1}_{n}\approx\frac{k_{n}}{n}L^{\infty}_{n}\approx\frac{1}{3}\log\frac{n}{L^{\infty}_{n}} unbounded, but (L^{1})^{4+\delta}e^{3L_{1}}L^{\infty}_{n}=o(n) , so the permuted products still converge in probability. If we take k_{n}=\frac{n}{3L^{\infty}_{n}}(\log\frac{n}{L^{\infty}_{n}}-\log\log n-(3+\delta)\log\log\frac{n}{L^{\infty}_{n}}), then (L^{1})^{3}e^{3L_{1}}L^{\infty}_{n}=o(\frac{n}{\log n}) , and so the permuted products converge almost surely. In particular, we have such convergence for large L^{\infty}_{n} , L^{\infty}_{n}=\frac{n}{\log n(\log\log n)^{3+2\delta}},\quad k_{n}\approx\frac{1}{3}\log n(\log\log n)^{3+2\delta}. In the opposite direction, we may take all matrices to be uniformly bounded by (unbounded in n ) L^{\infty}_{n}=\frac{1}{3}(\log n-(5+\delta)\log\log n). For a final example of an intermediate regime, for 0<t\leq 1 and 0<\alpha<1 or \alpha=1 , \beta\leq 0 , L^{\infty}_{n}=\frac{1}{3t}n^{1-\alpha}(\log n)^{1-\beta},\quad k_{n}\approx\alpha tn^{\alpha}(\log n)^{\beta}. To prove Theorem 1 , we first show the result when o(n/\log n) distinct elements occur in the array with the same frequency, and then upgrade it to arbitrary frequencies. The proof of Theorem 2 naturally splits into two parts. In Proposition 8 , we translate the conditions on multiplicative convergence into conditions on the additive convergence. This part of the result still holds in an arbitrary Banach algebra. Then we prove additive convergence in Lemma 13 by using a concentration inequality. This part is proven only for finite dimensional matrices. Thus we do not know if the full result holds in a general Banach algebra. Remark 5 (Literature review) . There is a number of articles treating topics related to this work. The following list is not meant to be complete. The techniques are typically more sophisticated than ours. In [ EH18 ] , the authors consider a sequence (A_{n}) of matrices such that \frac{1}{n}\sum_{i=1}^{n}A_{i}\rightarrow A and L_{n}^{1} are uniformly bounded. Then for any t\in\mathbb{C} , deterministically \prod_{i=1}^{n}(1+\frac{t}{n}A_{i})\rightarrow e^{tA}. Note that (although the authors do not state this) their results appear to hold in a general Banach algebra. There has been substantial follow-up work based on this article. In [ HW20 ] , the authors assume the matrices to be random, with mean A , uniformly bounded. They obtain a bound on \left\|{\prod_{i=0}^{n}(1+\frac{1}{n}A_{i})-e^{A}}\right\| which holds with almost full probability. These results are further improved in [ KMS20 ] and [ HNWTW22 ] . This last article implies that, in our context, with probability at least 1-\delta , for large n , \left\|{\prod_{i=1}^{n}e^{A_{\sigma(i),n}/n}-e^{A}}\right\|\leq\frac{L^{\infty}_{n}e^{\left\|{A}\right\|}}{\sqrt{n}}\ \sqrt{2e^{2}\log(d/\delta)}. See also [ CHKT21 ] . All in all, the key aim of these papers is to find the fastest rate of convergence, while our (of course, closely related) aim is to find the fastest rate of growth for the matrices under which convergence still holds. We also emphasize that the convergence occurs uniformly to the whole path e^{tA} , t\in[0,1] , as well as the interplay between L^{\infty}_{n} and L^{1}_{n} . Finally, in earlier work [ AP23 ] of the first author with Pritchett, we considered arrays consisting of multiple copies of finitely many elements in a Banach algebra. These results are extended in Theorem 1 . For other results in a Banach algebra see [ Len71 ] and [ Pop25 ] . Remark 6 . Concentration inequalities hold in some, but not all, Banach spaces, see for example [ Pin94 ] and [ CHL24 ] . So our results can be extended from matrices to these settings. Concentration inequalities have also been extended to the context of (finite) von Neumann algebras [ JZ15 ] , with probability replaced by the trace of the spectral projection. We however only use

Read the complete text