Anasazi: BlockKrylovSchur reports convergence when it shouldn't
Created by: amklinv
@hkthorn @trilinos/anasazi
The BlockKrylovSchur eigensolver reports convergence when its residual estimate drops below the tolerance. For ill-conditioned problems, this may be a bad estimate. For instance, I was computing the smallest eigenpairs of a graph Laplacian with a tolerance of 1e-8, and the actual residual achieved was roughly 1e-3. Perhaps we should do something like Belos::Gmres, where they test the true residual after the estimate becomes small?