On spectrum of sample covariance matrices from large tensor vectors

Abstract

In this paper, we investigate the limiting empirical spectral distribution (LSD) of sums of independent rank-one $k$-fold tensor products of $n$-dimensional vectors as $k,n \to \infty$. Assuming that the base vectors are complex random variables with unit modular, we show that the LSD is the Mar\v{c}enko-Pastur law. Comparing with the existing results, our limiting setting allows $k$ to grow much faster than $n$. Consequently, we obtain the necessary and sufficient conditions for Mar\v{c}enko-Pastur law to serve as the LSD of our matrix model. Our approach is based on the moment method.