Spectral norm of products of random and deterministic matrices (Q718893): Difference between revisions

The spectral norm, i.e., the largest singular value, of an \(m \times n\) matrix with i.i.d. entries is a well-known topic both in asymptotic (see, e.g., [\textit{Y. Q. Yin, Z. D. Bai} and \textit{P. R. Krishnaiah}, Probab. Theory Relat. Fields 78, No.~4, 509--521 (1988; Zbl 0627.62022)]) and non-asymptotic (see [\textit{R. Latala}, Proc. Am. Math. Soc. 133, No.~5, 1273--1282 (2005; Zbl 1067.15022); \textit{Y. Seginer}, Comb. Probab. Comput. 9, No.~2, 149--166 (2000; Zbl 0969.15009)]) random matrix theory. The present paper replaces the assumption of independent matrix entries by the assumption that the matrix under consideration should be of the form \(W = BA\), where \(A\) is a random matrix with independent centered entries, and \(B\) is deterministic. The main result extends a known non-asymptotic bound from the independent case to this more general framework, which encompasses sample covariance matrices of random vectors of the form \(BX\), where \(X\) has i.i.d. centered entries. Specifically, the result is as follows. Let \(\varepsilon \in (0, 1)\) and let \(m, n, N\) be positive integers. Consider a random \(m \times n\) matrix \(W = BA\), where \(A\) is an \(N \times n\) random matrix the entries of which are independent random variables with mean zero and \((4 + \varepsilon)\)-th moment bounded by \(1\), and \(B\) is an \(m \times N\) non-random matrix with spectral norm \(\| B\| \leq 1\). Then \(\text{E}(\|W\|) \leq C(\varepsilon)(\sqrt{m} + \sqrt{n}),\) where \(C(\varepsilon)\) depends only on \(\varepsilon\). Note, in particular, that the bound is independent of the dimension \(N\).