The effect of non-optimal bases on the convergence of Krylov subspace methods (Q2501255)

The authors investigate the effect of non-orthogonal bases for Krylov subspace methods for the solution of the linear sytem \(A x = b\). While these typically require less storage and computational effort, the convergence may be delayed. Identities for the residual gap are proved and give an indication when the convergence delay is small. The results confirm that in exact arithmetics not orthogonality but rather linear independence is the requirement for good convergence. Also, not the symmetry of the matrix \(A\) matters, but other properties like its spectral properties. In particular, some numerical experiments show that the eigenvector matrix should not be too ill-conditioned for the non-optimal (i.e. non-orthogonal) Krylov methods to be competitive.