Inexact inverse subspace iteration for generalized eigenvalue problems (Q630516)

To solve the generalized eigenvalue problem \(Ax=\lambda Bx\), some inverse subspace iteration (outer iteration iterating based on \(A^{-1}B\)) is used. In each iteration step, one has to solve a linear system, which is again solved by an iterative procedure (inner iteration) to within some threshold accuracy. This paper deals with the influence of the inexact inner iteration solves on the convergence of the outer iteration. Several eigenvalues (e.g. the \(p\) smallest ones) and their eigenvectors are approximated simultaneously. It is shown that the outer iteration converges linearly with a rate given by \(\max\{\gamma,\rho\}\) where \(\epsilon_k=a\gamma^k\), \(\gamma<1\) is the threshold used in step \(k\) of the inner iteration in a stopping criterion and \(\rho=|\lambda_p|/|\lambda_{p+1}|<1\) is the ratio of the bordering eigenvalues measuring the gap between the \(p\) smallest eigenvalues and the others. It is however generally independent of the particular iteration or the preconditioner used in the inner loop. The number of inner iterations does not blow up as \(k\) increases and depends only modestly on \(\gamma\). Also the basis vectors of the subspaces converge so that an approximation of a step can be used as an initial condition for the next one.