Computable Convergence Bounds for GMRES (Q4943666)

The author derives new computable convergence bounds for the generalized minimal residual (GMRES) method. Usually such GMRES convergence bounds are derived from the so-called ``ideal GMRES'' minimization problem of the spectral norm \(\|p_n(A)\|\) of the matrix polynomial \(p_n(A)\) over the set of all polynomials of the degree at most \(n\) with value 1 at the origin, where \(A\) denotes the system matrix of the system \(Ax= b\) that is to be solved by the GMRES. It is clear that such bounds will usually overestimate the true convergence behaviour considerably, because they are independent of the initial residual \(r_0= b- Ax_0\). In contrast to this approach, the bounds proposed in this paper are computed during the iteration, i.e. they are of a posteriori type, and they depend on the initial guess. The numerical examples presented in paper show that this approach works well.