Convergence analysis of the block Householder diagonalization algorithm (Q2568633)

The authors develop algebraic and convergence properties of the left and right block reflectors used in block diagonalization algorithms. For this purpose, block Householder transformations are used. More precisely, the convergence of the algorithm is shown under a weak condition. In addition, it is proved that if there exists a gap between the singular values, i.e., if \(\alpha=\sigma_{\max}(D_{22})/\sigma_{\min}(D_{11})<1\), where \(D_{11},\,D_{22}\) are the limits of the diagonal blocks, then the off-diagonal blocks converge to zero at the rate of a geometric progression and that asymptotically, the rate of convergence depends essentially on \(\alpha^2.\) \textit{R. Mathias} and \textit{G. W. Stewart} [Linear Algebra Appl. 182, 91--100 (1993; Zbl 0769.65012)] analyzed the convergence behavior of a block-diagonalization process with general unitary transformations. Both approaches show the importance of a sufficient gap in the singular values of the matrix. As presented in this paper, this condition is required for the speed of convergence, not for the convergence.