A Schatten-\(q\) low-rank matrix perturbation analysis via perturbation projection error bound (Q821008)

Let \(A\) be a rank-\(r\) real matrix. Consider \(\widehat{A}=\widehat{U}\widehat{\Sigma_1}\widehat{V}^T\), where \(\Sigma_1\) is a pseudo-diagonal matrix with decreasing singular values of \(A\). The authors search the estimation error in the matrix Schatten-\(q\) norm \(\|\widehat{A}-A\|_q\). One of the main results of the paper reads as follows. Suppose \(B=A+Z\). Let \(\widehat{A}=\widehat{U}\widehat{\Sigma_1}\widehat{V}^T\) be the best rank-\(r\) approximation of \(B\). Then \[\|\widehat{A}-A\|_q\leq\left\lbrace \begin{array}{cc} (2^q+1)^{\frac{1}{q}}\|Z_{\max (r)}\|_q, & 1\le q\le 2 \\ \sqrt{5}\|Z_{\max (r)}\|_q, & 2\le q<\infty \\ 2\|Z_{\max (r)}\|, & q=\infty \end{array} \right. ,\] where \(Z_{\max (r)}\) is the best rank-\(r\) approximation of \(Z\). The results in this paper may be useful in statistical analysis.