54
35

Limiting spectral distribution of renormalized separable sample covariance matrices when p/n0p/n\to 0

Abstract

We are concerned with the behavior of the eigenvalues of renormalized sample covariance matrices of the form C_n=\sqrt{\frac{n}{p}}\left(\frac{1}{n}A_{p}^{1/2}X_{n}B_{n}X_{n}^{*}A_{p}^{1/2}-\frac{1}{n}\tr(B_{n})A_{p}\right) as p,np,n\to \infty and p/n0p/n\to 0, where XnX_{n} is a p×np\times n matrix with i.i.d. real or complex valued entries XijX_{ij} satisfying E(Xij)=0E(X_{ij})=0, EXij2=1E|X_{ij}|^2=1 and having finite fourth moment. Ap1/2A_{p}^{1/2} is a square-root of the nonnegative definite Hermitian matrix ApA_{p}, and BnB_{n} is an n×nn\times n nonnegative definite Hermitian matrix. We show that the empirical spectral distribution (ESD) of CnC_n converges a.s. to a nonrandom limiting distribution under some assumptions. The probability density function of the LSD of CnC_{n} is derived and it is shown that it depends on the LSD of ApA_{p} and the limiting value of n1\tr(Bn2)n^{-1}\tr(B_{n}^2). We propose a computational algorithm for evaluating this limiting density when the LSD of ApA_{p} is a mixture of point masses. In addition, when the entries of XnX_{n} are sub-Gaussian, we derive the limiting empirical distribution of {n/p(λj(Sn)n1\tr(Bn)λj(Ap))}j=1p\{\sqrt{n/p}(\lambda_j(S_n) - n^{-1}\tr(B_n) \lambda_j(A_{p}))\}_{j=1}^p where Sn:=n1Ap1/2XnBnXnAp1/2S_n := n^{-1} A_{p}^{1/2}X_{n}B_{n}X_{n}^{*}A_{p}^{1/2} is the sample covariance matrix and λj\lambda_j denotes the jj-th largest eigenvalue, when FAF^A is a finite mixture of point masses. These results are utilized to propose a test for the covariance structure of the data where the null hypothesis is that the joint covariance matrix is of the form ApBnA_{p} \otimes B_n for \otimes denoting the Kronecker product, as well as ApA_{p} and the first two spectral moments of BnB_n are specified. The performance of this test is illustrated through a simulation study.

View on arXiv
Comments on this paper