Subspace Expansion in the Shift-Invert Residual Arnoldi Method and the Jacobi--Davidson Method: Theory and Algorithms

Zhongxiao Jia,Cen Li

2011-01-01

Abstract:We give a quantitative analysis of the Shift-Invert Residual Arnoldi (SIRA) method and the Jacobi–Davidson (JD) method for computing a simple eigenvalue nearest to a target σ and/or the associated eigenvector. In SIRA and JD, subspace expansion vectors at each step are obtained by solving certain (different) inner linear systems, respectively. We show that (i) SIRA and the JD method with the fixed target σ are mathematically equivalent when the inner linear systems are solved exactly and (ii) the inexact SIRA is asymptotically equivalent to the JD method when the inner linear systems in them are solved with the same accuracy. Remarkably, we prove that the inexact SIRA and JD methods mimic the exact SIRA well provided that the inner linear systems are iteratively solved with a fixed low or modest accuracy. It is opposed to the inexact Shift-Invert Arnoldi (SIA) method, where the inner linear system involved must be solved with very high accuracy whenever the approximate eigenpair is of poor accuracy and is only solved with decreasing accuracy after the approximate eigenpair starts converging. We also show that SIRA and JD expand subspaces in a computationally optimal way. We propose restarted SIRA and JD algorithms and design practical stopping criteria for inner solvers. Numerical experiments confirm our theory and the considerable superiority of the (nonrestarted and restarted) inexact SIRA and JD to the inexact SIA, and demonstrate that the inexact SIRA and JD are similarly effective and mimic the exact SIRA very well.