Inner iterations in the shift-invert residual Arnoldi method and the Jacobi-Davidson method (Q477157): Difference between revisions

The authors consider the computation of the eigenvalue closest to a given target \(\sigma \in \mathbb{C}\) and of the associated eigenvector of a sparse matrix eigenvalue problem \(Ax = \lambda x\) with \(A \in \mathbb{C}^{n \times n} \). At first, the shift-invert residual Arnoldi (SIRA) and the Jacobi-Davidson (JD) algorithms are described and compared. One of the differences are the systems of linear equations to be solved in each iteration step. It is shown that the SIRA and JD methods behave very similar when the systems of linear equations are solved with the same accuracy. Furthermore, it is shown that the number of outer iterations of the SIRA methods with exact solution of the linear systems and inexact solution of these systems is nearly the same, when in the inexact method the systems are solved with low or modest accuracy. Practical stopping criteria for the inner iterations in the SIRA and JD methods are derived. The presented theory is confirmed by numerical experiments. These experiments show that the inexact SIRA and JD methods are similarly effective and behave like the exact SIRA algorithm.