On the perturbation bound in unitarily invariant norms for subunitary polar factors (Q929490)

Let the singular-value decomposition of an \(m\)-by-\(n\) rank-\(r\) complex matrix \(A\) be \[ A= U\begin{pmatrix} \Sigma_1 & 0\\ 0 & 0\end{pmatrix} V^*, \] where \(U\) and \(V\) are \(m\)-by-\(m\) and \(n\)-by-\(n\) unitary matrices, respectively, and \(\Sigma_1= \text{diag}(\sigma_1,\dots, \sigma_r)\) with \(\sigma_1\geq\cdots\geq\sigma_r> 0\), the positive singular values of \(A\). Let \(U= [U_1, U_2]\) and \(V= [V_1, V_2]\), where \(U_1\) and \(V_1\) are \(m\)-by-\(r\) and \(n\)-by-\(r\) matrices, respectively, and let \(Q=U_1 V^*_1\) and \(H= V_1\Sigma_1 V^*_1\). Then \(A= QH\) is called the generalized polar decomposition of \(\widetilde A\). Let \(\widetilde A=\widetilde Q\widetilde H\) be the generalized polar decomposition of another \(m\)-by-\(n\) rank-\(r\) matrix \(A\). The main result of this paper gives a new upper bound for \(\widetilde Q- Q\) in any unitarily invariant norm \(\|.\|\). More precisely, it is shown that \[ \|\widetilde Q- Q\|\leq\begin{cases} 3\|\widetilde A-A\|/(\sigma_r+ \widetilde\sigma_r)\quad &\text{if }r\leq n\leq m,\\ 2\|\widetilde A- A\|/\max\{\sigma_n, \widetilde\sigma_n\}\quad &\text{if }r= n< m.\end{cases} \] For a \(Q\)-norm \(\|.\|_Q\) (meaning that there exists another unitarily invariant norm \(\|.\|'\) such that \(\| Y\|_Q=\| Y*Y\|^{\prime 1/2}\) for all matrices \(Y\)), the perturbation bounds can be improved to \[ \|\widetilde Q- Q\|_Q\leq \begin{cases} (1+ \sqrt{3})\|\widetilde A- A\|_Q/(\sigma_r+ \widetilde\sigma_r)\quad &\text{if }r\leq n\leq m,\\ (1+ \sqrt{2})\|\widetilde A- A\|_Q/(\sigma_n+ \widetilde\sigma_n)\quad &\text{if }r= n< m.\end{cases} \] It is shown that these bounds are sharper than the previous ones obtained by the author and others. Whether \(\|\widetilde Q- Q\|\leq 2\|\widetilde A- A\|/(\sigma_r+ \widetilde\sigma_r)\) holds for every unitarily invariant norm \(\|.\|\) and every rank \(r\) is still open although many special cases of it are known to be true.