Modified block-approximate factorization strategies (Q758105): Difference between revisions

Incomplete block-matrix factorizations have been recognized as useful preconditioning techniques for the iterative solution of large sparse linear systems of algebraic equations, particularly efficient when combined with a conjugate gradient like method or when used for the construction of smoothers in multigrid methods or of domain decomposition preconditioners. Among these factorizations, ``modified'' versions for which the original matrix and its approximate factorization have equal row sum, are known to be powerful for reducing the spectral condition number. Unfortunately, performances of modified methods are sometimes spoiled by round off errors in the presence of strongly isolated largest eigenvalues. Two new variants are proposed which not only reduce the largest eigenvalues but also (implicitly) cluster the interior eigenvalues and ensure a \({\mathcal O}(h^{-1})\) behaviour for the associated spectral condition number, when applied to discrete multidimensional self-adjoint (elliptic) partial differential equations with (average) mesh size parameter h. Both rigorous theoretical support and numerical evidence are given, displaying the superiority of the new versions over the standard (``unmodified'' and ``modified'') versions, even in the case of quasi- singular problems.