Universality of covariance matrices (Q2454401): Difference between revisions

Let \(X= (x_{i,j})\) be an \(M\times N\) data matrix with independent centered real-valued entries such that \(\mathbb{E}x_{i,j}=0\) and \( \mathbb{E}x^{2}_{i,j}=M^{-1},\). Further let assume that the entries have a sub-exponential decay or sufficiently high number of moments. The covariance matrix corresponding to \(X\) is \(H= X^{\dag} X.\) Let \(d_{N}=N/M\), with \(\lim_{N\rightarrow\infty} d_{N}\neq 0, \infty\). The general theory of the asymptotic distribution of the eigenvalues of such covariance matrices is not exhaustively analyzed up to some cases under specific assumptions on the distributions of their entries. For instance, in the Gaussian case, the covariance matrix, the so-called Wishart matrix, is a closed form expression for the joint distribution of the eigenvalues. On the other hand, the limiting distribution of the smallest eigenvalue of squared matrices with standard Gaussian entries was computed by \textit{A. Edelman} [Linear Algebra Appl. 159, 55--80 (1991; Zbl 0738.15010)]. Using the classical Wigner approach in physical models, the universality of covariance matrices obtained from a data matrix \(X\) satisfying the conditions as above is studied. The universality of the extremal singular values for rectangular matrices \(X\) or, equivalently, the eigenvalues of the covariance matrix \(H\) at both edges of the spectrum is deduced. When \(\lim_{N\rightarrow\infty} d_{N}=1\), the analysis is focused only in the largest eigenvalue. The approach to deal with universality is based on the consideration of a novel version of the Green function comparison theorem. A strong local Marcenko-Pastur law constitutes the main tool for the proof of rigidity of eigenvalues and universality. In the mean time, the authors show that the Stieltjes transform of the eigenvalue distribution of \(H\) is given by the Marcenko-Pastur law uniformly up to the edges of the spectrum with an error \(O( N\eta)\), depending on \(N\) and the imaginary part, \(\eta\), of the spectral parameter in the Stieltjes transform. For covariance matrices whose entries have a symmetric probability density (including the Wishart matrices) it is well known that the largest and smallest \(k\) eigenvalues, after an appropriate normalization (centering and rescaling), converge in distribution to the Tracy-Widom law (see [\textit{O. N. Feldheim} and \textit{S. Sodin}, Geom. Funct. Anal. 20, No. 1, 88--123 (2010; Zbl 1198.60011)]). On the other hand, the bulk universality of eigenvalues for covariance matrices in almost optimal scale is deduced using the above strong Marcenko-Pastur law as well as some previous results by \textit{L. Erdős} et al. [Ann. Inst. Henri Poincaré, Probab. Stat. 48, No. 1, 1--46 (2012; Zbl 1285.82029)]. Notice that for Gaussian divisible ensembles the covariance matrix is embedded into a stochastic flow of matrices in such a way that the eigenvalues evolve according to the Dyson Brownian motion. Such an approach is followed in the paper under review as well as, again, the Green function comparison method. \textit{T. Tao} and \textit{V. Vu} [Ann. Probab. 40, No. 3, 1285--1315 (2012; Zbl 1247.15036)] deduced the bulk universality under the requirement that the first four moments of the entries of the two covariance matrices are equal. As a remarkable conclusion of this nice contribution, from an asymptotic point of view, you get that the distribution of the local statistics of eigenvalues of the covariance matrices \(H\) associated with \(X\) are identical to those of the Wishart matrix. In particular, the edge universality under the assumption that the first two moments of the matrix entries are equal to the corresponding ones of the standard Gaussian is pointed out.