Kernel random matrices have attracted a lot of interest in recent years, from both practical and theoretical standpoints. Most of the theoretical work so far has focused on the case were the data is sampled from a low-dimensional structure. Very recently, the first results concerning kernel random matrices with high-dimensional input data were obtained, in a setting where the data was sampled from a genuinely high-dimensional structure---similar to standard assumptions in random matrix theory. In this paper, we consider the case where the data is of the type "information${}+{}$noise." In other words, each observation is the sum of two independent elements: one sampled from a "low-dimensional" structure, the signal part of the data, the other being high-dimensional noise, normalized to not overwhelm but still affect the signal. We consider two types of noise, spherical and elliptical. In the spherical setting, we show that the spectral properties of kernel random matrices can be understood from a new kernel matrix, computed only from the signal part of the data, but using (in general) a slightly different kernel. The Gaussian kernel has some special properties in this setting. The elliptical setting, which is important from a robustness standpoint, is less prone to easy interpretation.
First-order asymptotic expansions for spectral convergence of compact self-adjoint operators on general spectral subsets, with application to kernel Gram matrices
Universality of Kernel Random Matrices and Kernel Regression in the Quadratic Regime
Center for Machine Intelligence and Data Science Indian Institute of Technology Bombay(机器智能与数据科学中心 印度理工学院班加罗尔);International Computer Science Institute Department of Statistics University of California, Berkeley(国际计算机科学研究所 统计学系 加州大学伯克利分校);Department of Mathematics University of Southern California(数学系 华盛顿大学)