Central limit theorem for ϵ \epsilon -independent products and higher-order tensors
Guillaume Cébron and Patrick Oliveira Santos and Pierre Youssef · 2025
Read Central limit theorem for ϵ \epsilon -independent products and higher-order tensors by Guillaume Cébron and Patrick Oliveira Santos and Pierre Youssef (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-2504-10059.txt.
Introduction
Three mathematicians prove a central limit theorem for products of nearly independent random variables, in this 2025 paper. The classic theorem says that sums of many independent random quantities settle into the familiar bell curve. Modern probability stretches that idea to stranger dependence patterns, where variables are independent only in a loose, approximate sense described by a small parameter.
The authors handle families of products of such variables, using graphon limits, mathematical objects that capture how dependence evolves across a large system, to pin down the limiting distribution. Their framework covers a wide class of dependence structures. As a special case it yields a limit theorem for tensor products of free random variables, objects from the non-commutative probability used in quantum theory and random matrices. The proof is shorter and more unified than earlier routes to similar results.
Catalog record
| CREATOR | Guillaume Cébron and Patrick Oliveira Santos and Pierre Youssef |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-10059
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Central limit theorem for ϵ \epsilon -independent products and higher-order tensors Guillaume Cébron and Patrick Oliveira Santos and Pierre Youssef Abstract . We establish a central limit theorem (CLT) for families of products of ϵ \epsilon -independent random variables. Building on the approach in [ 4 ] , we utilize graphon limits to encode the evolution of independence and characterize the limiting distribution. Our framework subsumes a wide class of dependency structures and includes, as a special case, a CLT for higher-order tensor products of free random variables. Our results generalize those in [ 4 ] , extend the findings of [ 8 ] , which were limited to order-2 tensors, and recover as a special case the recent tensor-free CLT of [ 10 ] , which was obtained through the development of a tensor analogue of free probability. In contrast, our approach is more direct and provides a unified and concise derivation of a more general CLT via graphon convergence. 1. Introduction The Free Central Limit Theorem (CLT) serves as a cornerstone of free probability theory [ 13 ] , [ 11 , Lecture 8] . It parallels the classical CLT, replacing classical independence with free independence and the Gaussian limit with the semi-circle law. Recently, the authors [ 4 ] unified these two by establishing a central limit theorem for \epsilon -independent variables. Let (\mathcal{A},\tau) be a noncommutative probability space. Based on the graph product of groups [ 6 ] , the notion of \epsilon -independence was introduced in [ 9 ] and elaborated in several subsequent works [ 1 , 5 , 12 ] (see also [ 2 , 3 ] for matrix models). Given a loopless graph g_{n}=([n],E) over n vertices, we say that subalgebras \mathcal{A}_{1},\ldots,\mathcal{A}_{n}\subset\mathcal{A} are g_{n} -independent if (1) For every (i,j)\in E , the subalgebras \mathcal{A}_{i},\mathcal{A}_{j} are classical independent (in particular, they commute); (2) For every k\geq 1 , i\in[n]^{k} such that for all j_{1}<j_{2} with i_{j_{1}}=i_{j_{2}} , there exists j_{1}<j_{3}<j_{2} with (i_{j_{1}},i_{j_{3}})\notin E , then for any centered random variables a_{j}\in\mathcal{A}_{i_{j}} , j\in[k] , we have \displaystyle\tau(a_{1}\cdots a_{k})=0. Note that when the graph g_{n} is complete (resp. edgeless), the above corresponds to classical independence (resp. freeness). The authors in [ 4 ] used the theory of graphons to track the evolution of the graphs of independence. Recall that a graphon consists of a measurable bounded symmetric function w:[0,1]^{2}\to[0,1] . Given a graphon w and graphs f=(V(f),E(f)),g=(V(g),E(g)) , we define \displaystyle\rho(f,w):=\int_{[0,1]^{V(f)}}\prod_{(u,v)\in E}w(x_{u},x_{v})\,\text{d}x; \displaystyle\rho(f,g)=\frac{1}{|V(g)|^{|V(f)|}}\sum_{\phi:V(f)\to V(g)}\prod_{(u,v)\in E(f)}\mathbf{1}_{(\phi(u),\phi(v))\in E(g)}. A sequence of simple graphs g_{n} in n nodes converges to a graphon w if, for every simple finite graph f=(V(f),E(f)) , we have \displaystyle\rho(f,g_{n})\to\rho(f,w). We say that a sequence of (self-adjoint) variables (a_{n})_{n\in\mathbb{N}} in \mathcal{A} converges in distribution to a\in\mathcal{A} if, for every p\in\mathbb{N} , \tau(a_{n}^{p})\to\tau(a^{p}) as n\to\infty . More generally, we define the convergence in * -moments of a sequence of not necessarily self-adjoint variables (b_{n})_{n\in\mathbb{N}} towards some variable z if for any k\geq 1 , t\in\{1,*\}^{k} , we have \displaystyle\tau(b_{n}^{t})\to\tau(z^{t}), where \displaystyle b_{n}^{t}:=b_{n}^{t_{1}}\cdots b_{n}^{t_{k}}. The authors in [ 4 ] established the following CLT for \epsilon -independent variables encompassing the classical and free central limit theorems. The combinatorial proof of those limit theorems latter deals with pair partitions. We denote by P_{2}(p) the set of pair partitions of [p] and for a pair partition \pi\in P_{2}(p) , we define its intersection graph f_{\pi} whose vertex set consists of the blocks of \pi , and there exists an edge between two blocks if they cross. Theorem 1.1 ( [ 4 ] ) . Let g_{n} be a graph over [n] , and a\in\mathcal{A} be a self-adjoint random variable with mean \lambda and variance \sigma^{2} . Assume that g_{n}\to w in the graphon sense. Let a_{1},\ldots,a_{n}\in\mathcal{A} be g_{n} -independent and identically distributed random variables with common law a . Let \displaystyle S_{n}=\frac{1}{\sigma\sqrt{n}}\sum_{k\in[n]}(a_{k}-\lambda). Then, S_{n} converges in distribution to a measure \mu_{w} that depends only on w , whose odd moments vanish and even moments are equal to \displaystyle\int x^{2p}\,\text{d}\mu_{w}=\sum_{\pi\in P_{2}(2p)}\rho(f_{\pi},w). The class of measures \mu_{w} is quite vast, encompassing well-known distributions like q -Gaussians as well as many exotic laws (see [ 4 ] for details). In another direction motivated by the study of the limiting behavior of random quantum channels, [ 8 ] establishes a central limit theorem for tensor product random variables of the form c_{k}=a_{k}\otimes a_{k} , where (a_{k})_{k\in\mathbb{N}} is a free family of variables. Theorem 1.2 ( [ 8 ] ) . Let a\in\mathcal{A} be a self-adjoint random variable with mean \lambda and variance \sigma^{2} , and denote \delta^{2}:=\sigma^{2}(\sigma^{2}+2\lambda^{2}),\quad\alpha:=\frac{2\lambda^{2}}{\sigma^{2}+2\lambda^{2}}\in[0,1]. Given (a_{k})_{k\in\mathbb{N}} a sequence of free copies of a , the distribution \mu_{S_{n}} of the normalized sum \displaystyle S_{n}:=\frac{1}{\delta\sqrt{n}}\sum_{k\in[n]}(a_{k}\otimes a_{k}-\lambda^{2}) converges in distribution as n\to\infty to \sqrt{\alpha}\left(\frac{1}{\sqrt{2}}s_{1}+\frac{1}{\sqrt{2}}s_{2}\right)+\sqrt{1-\alpha}\,s_{3}, where s_{1},s_{2},s_{3} are semi-circle random variables, s_{1},s_{2} are classical independent, both free independent from s_{3} . The limiting law appearing in the above theorem is partially reminiscent of the laws uncovered in Theorem 1.1 . For instance, the law of \frac{1}{\sqrt{2}}s_{1}+\frac{1}{\sqrt{2}}s_{2} coincides with that of \mu_{w} when the graphon w is given by \mathbf{1}_{[0,1/2)\times[1/2,1]}+\mathbf{1}_{[1/2,1]\times[0,1/2)} (see [ 4 , Example 7.4] ). In view of this, it is natural to seek a connection between both results. Moreover, the analysis of higher-order tensors naturally generalizes the results for order-two tensors, models more intricate structures, and introduces a rich interplay between free and classical independence. Notably, tensors of random variables correspond to products of independent variables. Indeed, tensors of the form a_{k}^{(1)}\otimes\ldots\otimes a_{k}^{(L)} coincide with the product of L independent variables b_{j}=1\otimes\ldots\otimes a_{k}^{(j)}\otimes\ldots\otimes 1 , 1\leq j\leq L . As such, a CLT for fixed-order tensors can be interpreted as a study of the free sum of products of independent variables. This realization provides a new perspective, allowing us to model these products and sums through mixtures of free and classical independence. To formalize this idea, we employ the concept of \epsilon -independence mentioned above, enabling us to consider a general setting beyond the fixed-order tensors, considering mixtures of free and independent variables, as well as mixtures of their sums and products. In [ 4 ] , we demonstrated the utility of graphon theory in tracking the evolution of these mixtures and establishing the corresponding limit theorems. Here, we also make use of these techniques to analyze the sum of mixtures of products modeled by \epsilon -independence. More precisely, let g_{L} be a graph over [L] , and (\mathcal{A}_{l})_{l\in[L]}\subset\mathcal{A} be g_{L} -independent subalgebras. For every 1\leq l\leq L and every n\geq 1 , let g^{\prime}_{n} be a graph over [n] and consider variables (a_{k}^{(l)})_{k\in[n]}\subset\mathcal{A}_{l} which are g^{\prime}_{n} -independent. We aim to capture the limiting behavior, as n\to\infty , of \frac{1}{\sqrt{n}}\sum_{k=1}^{n}\big(\prod_{l=1}^{L}a_{k}^{(l)}-\lambda^{L}\big). Theorem 1.1 corresponds to the case L=1 , while Theorem 1.2 corresponds to L=2 with g_{2} being the complete graph and g_{n}^{\prime} being the edgeless graph. Our main theorem generalizes both results to cover any depth L and any graphs of independence g_{L} and g_{n}^{\prime} . Theorem 1.3 . Let g_{L} be a graph over [L] , g^{\prime}_{n} be a graph over [n] , and a\in\mathcal{A} be a self-adjoint random variable with mean \lambda and variance \sigma^{2} . Assume that g^{\prime}_{n}\to w in the graphon sense. Let (\mathcal{A}_{l})_{l\in[L]} be g_{L} -independent subalgebras and (a_{k}^{(l)})_{k\in[n]}\subset\mathcal{A}_{l} be g^{\prime}_{n} -independent and identically