A refined iterative algorithm based on the block Arnoldi process for large unsymmetric eigenproblems (Q1377506)

For large unsymmetric matrix eigenproblems, Arnoldi's method relies on Ritz values for Ritz vectors which in the block Arnoldi version often do not converge, even if the Ritz values do. This paper proposes instead to minimize the residual norm of the eigenvector equation over a small-dimensional Krylov subspace for a computed Ritz value and it does so by solving a small-dimensional singular value decomposition problem to obtain a `refined approximate eigenvector' in each iteration. Thus convergence is ensured and at a faster speed than block Arnoldi.