Composite orthogonal projection methods for large matrix eigenproblems (Q1585577)

Within the Ritz or Galerkin variational framework, an efficient partial numerical diagonalization of a Hermitian operator A can be performed within a suitable truncated \(m\)-dimensional basis. In the parallel non-Hermitian analysis a key difficulty emerges in connection with the possible mathematical failure of the variational background of the whole scheme. An explanation can be provided when we compare the implementations of the recipe in two different bases: The projection of the eigenvalue problem (generically, on a subspace of the whole Hilbert space) may happen to be badly ill-conditioned. The key idea of the author is to employ his own results on the general Arnoldi algorithm [cf., e.g., his joint publication with \textit{L. Elsner} in J. Comput. Math. 18, No.~3, 265-276 (2000; reviewed below) for an updated reference], emphasizing the fact that a bad convergence of the eigenvectors themselves is usually the very core of computational difficulties [i.a., \textit{J. H. Wilkinson}, Acta Univ. Carolinae, Math. Phys. 15(1974), No.~1/2, 173-177 (1975; Zbl 0315.65024) recommends the inverse iterations strongly]. Here, an alternative ``refined'' recipe is analysed, re-calculating the eigenvectors via a minimization technique. In such a setting the paper offers a number of estimates (a priori error bounds) and conditions of convergence.