Restrictions on implicit filtering techniques for orthogonal projection methods (Q1301271)

The application of a class of orthogonal projection methods for solving iteratively large eigenvalue problems is considered. It is shown that an implicit filtering procedure can only be applied to rank conservative methods or methods that have numerically a low residual rank. The larger the rank of the residual matrix is, the more inaccurate the implicit filtering will be or the more basis vectors must be sacrificed. It is shown that the set of rank conservative eigenvalue solvers is restricted to those methods that use an Arnoldi type scheme or that perform rational Krylov like computations. Other methods and inexact variants of rank conservative methods cannot be filtered implicitly in combination with an implicit restart. The study shows that the true implicit filtering can be only applied in a very limited set of cases. If a better filter must be applied, then the cost (in terms of ``lost'' subspace vectors), will be larger.