A general finite element preconditioning for the conjugate gradient method (Q582001)

Some preconditioning techniques like the (modified) incomplete Cholesky decomposition (MIC and IC) should not be applied to finite element discretizations of elliptic boundary value problems, when the FEM leads to a matrix which lacks the property M (f.i. quadratic triangular elements), because the existence of a preconditioning matrix cannot be proved then. The author introduces a method that constructs a preconditioning matrix for all finite element methods if only a mild condition for the node numbering is fulfilled: No global node should have a maximal number, that means, if all the neighbours of a node have a lower number, the node or one of its neighbours must be on the Dirichlet- type portion of the boundary. Such a numbering can be constructed using the reverse Cuthill-McKee algorithm starting with a node on the Dirichlet-type portion of the boundary. The main idea of the preconditioning method is to write the contribution of each finite element to the global matrix as a product of a lower triangular, a diagonal and an upper triangular matrix and assemble these matrices. The preconditioning matrix is defined as the product of the assembled global matrices. The idea can be generalized by assembling the sum of the factors of contributions of several finite elements to the global matrix. It is proved for a certain variant of the Cuthill-McKee algorithm, which fulfil a necessary and sufficient condition for the existence of the preconditioning matrix, that the smallest eigenvalue of this matrix is equal to 1. The application of the method to the (singular) Neumann problem is considered. The numerical experiments show, that the IC and MIC preconditioners give the fastest convergence of the conjugate gradient method in most cases, but for strange problems like a rhombus with vertex angle \(5\pi\) /6 the preconditioning matrix for the IC does not exist. The conclusion is that the MIC preconditioning for the conjugate gradient method is very effective even in the situation when the coefficient matrix is not an M-matrix. But the less efficient finite element preconditiong proposed by the author is robust and promising for higher order or non-symmetric problems.