Nested Lanczos: Implicitly restarting an unsymmetric Lanczos algorithm (Q1272642)

The paper presents a generalization of the concept of implicitly restarting and iterative algorithm to the Lanczos method for eigenvalue problems. It is shown that a full run of the Lanczos algorithm on the small, projected eigenvalue problem generates a biorthogonal factorization for the Lanczos bases. This factorization can be computed by a Lanczos algorithm with special starting vectors or by the two sided Gram-Schmidt algorithm. The resulting implicitly restarted Lanczos method is called ``nested Lanczos''. This method can be used with an implicit filter and exact shifts in order to remove eigenvalues from the approximation. Also, it is shown that implicitly restarted Lanczos may be used as an alternative for lookahead, in case of breakdown.