|
Introduction
The interest in parallel solvers for partial differential equations has
risen in recent years.
Thus the domain decomposition method (DD) became increasingly popular
since it contains a natural parallel structure.
As it is usual with this method, we utilize the resulting block system
of linear equations
(cf. section 3 for a more detailed description). The Additive Schwarz Method has been employed to construct preconditioners of the type for the (parallel) conjugate gradient method (cf. [12,13], and section 3). Two of the three components of this DD preconditioner, namely the (modified) Schur complement preconditioner , and the local Dirichlet problem preconditioner , have been studied intensively by the DD community (cf. the proceedings of the international Symposia on ``DD methods for partial differential equations'' since 1987 [7,3,4,8,,1,19], and also [2,5,6]). Haase/Langer/Meyer [12,13] proved that the quality of the DD preconditioner is dominantly influenced by the third component, the basis transformation . This basis transformation determines the perturbation of the Schur complement which, in turn, influences the spectral radius and finally the relative condition number of the preconditioner, (provided that and are chosen appropriately). Several ideas to choose the basis transformation which is defined implicitly by some iteration method are summarized in section 3. Our interest is focused on the multigrid method for which it is known that
In this paper we investigate the use of one full multigrid cycle for the definition of the basis transformation. Special interest is paid firstly as to whether the increasing condition number can be overcome as , and secondly to the computational expense. Sections 2 is devoted to the model problems whereas section 3 deals with the DD preconditioner and the basis transformation . The theoretical analysis of the full multigrid basis transformation is performed in section 4. Unfortunately the analysis turned out to be so technical that only the full two-grid operator on a simple model problem could be treated. In section 5 the numerical experiments with the full multigrid operator are given. Section 6 summarizes the results obtained.
|
|