Convergence of the block Lanczos method for eigenvalue clusters (Q495529)

The authors develop a new convergence theory for the block Lanczos method for solving large scale Hermitian eigenvalue problems. Unlike the approach of \textit{Y. Saad} [SIAM J. Numer. Anal. 17, 687--706 (1980; Zbl 0456.65016)], which bounded errors in each individual approximate eigenvector, their approach bounds approximation errors in the entire eigenspace associated with all eigenvalues in a tight cluster. This produces much tighter error bounds for (i) the canonical angles from the eigenspace to the generated Krylov space, (ii) the canonical angles between the eigenspace and its Ritz approximate subspace, and (iii) the total differences between the eigenvalues in the cluster and their corresponding Ritz values.