Using Krylov approximations to the matrix exponential operator in Davidson's method (Q1807760)

The paper deals with the problem of computing the rightmost eigenvalues of a large sparse nonsymmetric square matrix \(A\). The approach here has as starting point that such eigenvalues are transformed into the dominant eigenvalues of the exponential \(e^{\tau A}\). These dominant eigenvalues are often well separated and this accounts for better convergence of the associated eigenvectors in a Krylov space of that exponential. Accordingly the paper concentrates on the efficient construction of the Krylov space spanned by the exponential when applied to any one vector and with the computation of eigenpairs from the Galerkin projection of \(A\) on that space. In the paper a distinction is drawn between a Davidson and an Arnoldi approach in the sense that the former updates eigenvectors locally while the latter does it globally, because it does not use the information provided by a single eigenpair. An important conclusion in the paper is that the Davidson approach seems to be more efficient than the Arnoldi approach. The paper presents a convergence theory for eigenpairs for both Davidson and Arnoldi type methods in the present context. The numerical examples presented by the authors are abundant and provide the theory with thorough and ample illustration. Comparison with the Jacobi-Davidson method shows that the Davidson method yields better starting eigenvectors in order to obtain convergence to the rightmost eigenvalue. On the other hand, since Jacobi-Davidson is guaranteed to eventually converge, this suggests that a combination of both approaches could result in a both efficient and reliable eigensolver.