Researcher profile

L. Pastur

L. Pastur contributes to research discovery and scholarly infrastructure.

ResearcherAffiliation not importedOpen to collaborate

math-ph math.MP math.PR quant-ph cond-mat.dis-nn Machine Learning math.SP

Trust snapshot

Quick read

Trust 21 - Emerging

7works

0followers

7topics

4close collaborators

Actions

Decide how to stay connected

Follow researcher0

Inspect adjacent work, topics, institutions and collaborators without jumping out to a separate graph page.

Building this graph slice

BZPEER is loading the nearby papers, people, topics and institutions for this page.

preprint2022arXiv

On Random Matrices Arising in Deep Neural Networks: General I.I.D. Case

We study the distribution of singular values of product of random matrices pertinent to the analysis of deep neural networks. The matrices resemble the product of the sample covariance matrices, however, an important difference is that the population covariance matrices assumed to be non-random or random but independent of the random data matrix in statistics and random matrix theory are now certain functions of random data matrices (synaptic weight matrices in the deep neural network terminology). The problem has been treated in recent work [25, 13] by using the techniques of free probability theory. Since, however, free probability theory deals with population covariance matrices which are independent of the data matrices, its applicability has to be justified. The justification has been given in [22] for Gaussian data matrices with independent entries, a standard analytical model of free probability, by using a version of the techniques of random matrix theory. In this paper we use another, more streamlined, version of the techniques of random matrix theory to generalize the results of [22] to the case where the entries of the synaptic weight matrices are just independent identically distributed random variables with zero mean and finite fourth moment. This, in particular, extends the property of the so-called macroscopic universality on the considered random matrices.

preprint2016arXiv

Large Block Properties of the Entanglement Entropy of Disordered Fermions

We consider a macroscopic disordered system of free $d$-dimensional lattice fermions whose one-body Hamiltonian is a Schrödinger operator $H$ with ergodic potential. We assume that the Fermi energy lies in the exponentially localized part of the spectrum of $H$. We prove that if $S_Λ$ is the entanglement entropy of a lattice cube $Λ$ of side length $L$ of the system, then for any $d \ge 1$ the expectation $\mathbf{ E}\{L^{-(d-1)}S_Λ\}$ has a finite limit as $L \to \infty$ and we identify the limit. Next, we prove that for $d=1$ the entanglement entropy admits a well defined asymptotic form for all typical realizations (with probability 1) as $ L \to \infty$. According to numerical results of [33] the limit is not selfaveraging even for an i.i.d. potential. On the other hand, we show that for $d \ge 2$ and an i.i.d. random potential the variance of $L^{-(d-1)}S_Λ$ decays polynomially as $L \to \infty$, i.e., the entanglement entropy is selfaveraging.

preprint2014arXiv

On the Area Law for Disordered Free Fermions

We study theoretically and numerically the entanglement entropy of the $d$-dimensional free fermions whose one body Hamiltonian is the Anderson model. Using basic facts of the exponential Anderson localization, we show first that the disorder averaged entanglement entropy $\langle S_Λ\rangle$ of the $d$ dimension cube $Λ$ of side length $l$ admits the area law scaling $\langle S_Λ\rangle \sim l^{(d-1)}, \ l \gg 1$ even in the gapless case, thereby manifesting the area law in the mean for our model. For $d=1$ and $l\gg 1$ we obtain then asymptotic bounds for the entanglement entropy of typical realizations of disorder and use them to show that the entanglement entropy is not selfaveraging, i.e., has non vanishing random fluctuations even if $l \gg 1$.

preprint2013arXiv

The Central Limit Theorem for Linear Eigenvalue Statistics of the Sum of Independent Matrices of Rank One

We consider $n\times n$ random matrices $M_{n}=\sum_{α=1}^{m}{τ_{α}}\mathbf{y}_{α}\otimes \mathbf{y}_{α}$, where $τ_{α}\in \mathbb{R}$, $\{\mathbf{y}_{α}\}_{α=1}^{m}$ are i.i.d. isotropic random vectors of $\mathbb{R}^n$, whose components are not necessarily independent. It was shown in arXiv:0710.1346 that if $m,n\rightarrow \infty$, $m/n\rightarrow c\in \lbrack 0,\infty )$, the Normalized Counting Measures of $\{τ_{α}\}_{α=1}^{m}$ converge weakly and $\{\mathbf{y}_α\}_{α=1}^m$ are \textit{good} (see corresponding definition), then the Normalized Counting Measures of eigenvalues of $M_{n}$ converge weakly in probability to a non-random limit found in \cite{Ma-Pa:67}. In this paper we indicate a subclass of good vectors, which we call \textit{very good} and for which the linear eigenvalue statistics of the corresponding matrices converge in distribution to the Gaussian law, i.e., the Central Limit Theorem is valid. An important example of good vectors, studied in arXiv:0710.1346 are the vectors with log-concave distribution. We discuss the conditions for them, guaranteeing the validity of the Central Limit Theorem for linear eigenvalue statistics of corresponding matrices.

preprint2011arXiv

Non-Gaussian Limiting Laws for the Entries of Regular Functions of the Wigner Matrices

This paper is a continuation of our paper "Fluctuations of Matrix Elements of Regular Functions of Gaussian Random Matrices", J. Stat. Phys. (134), 147--159 (2009), in which we proved the Central Limit Theorem for the matrix elements of differential functions of the real symmetric random Gaussian matrices (GOE). Here we consider the real symmetric random Wigner matrices having independent (modulo symmetry conditions) but not necessarily Gaussian entries. We show that in this case the matrix elements of sufficiently smooth functions of these random matrices have in general another limiting law which coincides essentially with the probability law of matrix entries.

preprint2002arXiv

On the Spectral and Propagation Properties of the Surface Maryland Model

We study the discrete Schrödinger operator $H$ in $\ZZ^d$ with the surface potential of the form $V(x)=g δ(x_1) \tan π(α\cdot x_2+ ω)$, where for $x \in \ZZ^d$ we write $x=(x_1,x_2), \quad x_1 \in \ZZ^{d_1}, x_2 \in \mathbb{Z}^{d_2}, α\in \R^{d_2}, ω\in [0,1)$. We first consider the case where the components of the vector $α$ are rationally independent, i.e. the case of the quasi periodic potential. We prove that the spectrum of $H$ on the interval $[-d,d]$ (coinciding with the spectrum of the discrete Laplacian) is absolutely continuous. Then we show that generalized eigenfunctions corresponding to this interval have the form of volume (bulk) waves, which are oscillating and non decreasing (or slow decreasing) in all variables. They are the sum of the incident plane wave and of an infinite number of reflected or transmitted plane waves scattered by the "plane" $\ZZ^{d_2}$. These eigenfunctions are orthogonal, complete and verify a natural analogue of the Lippmann-Schwinger equation. We also discuss the case of rational vectors $α$ for $d_1=d_2=1$, i.e. a periodic surface potential. In this case we show that the spectrum is absolutely continuous and besides volume (Bloch) waves there are also surface waves, whose amplitude decays exponentially as $|x_1| \to \infty$. The part of the spectrum corresponding to the surface states consists of a finite number of bands. For large $q$ the bands outside of $[-d,d]$ are exponentially small in $q$, and converge in a natural sense to the pure point spectrum, that was found in [KP] in the case of the Diophantine $α$'s.

preprint2000arXiv

On the Law of Addition of Random Matrices

Normalized eigenvalue counting measure of the sum of two Hermitian (or real symmetric) matrices $A_{n}$ and $B_{n}$ rotated independently with respect to each other by the random unitary (or orthogonal) Haar distributed matrix $U_{n}$ (i.e. $A_{n}+U_{n}^{\ast}B_{n}U_{n}$) is studied in the limit of large matrix order $n$. Convergence in probability to a limiting nonrandom measure is established. A functional equation for the Stieltjes transform of the limiting measure in terms of limiting eigenvalue measures of $A_{n}$ and $B_{n}$ is obtained and studied. Keywords: random matrices, eigenvalue distribution