The stabilized \(V\)-cycle method (Q2564246)

The multilevel methods are often used for constructing preconditioners to the iterative solution methods for solving the systems of linear equations arising after the discretization of partial differential equations. One type of such methods is based on the recursive use of a nested sequence of finite element meshes and associated matrix blocks, using either standard basic functions or hierarchical basis functions. Another type of multilevel iteration methods has been based on a sequence of node sets defined from the matrix graph, associated with the given matrix. Both types of methods have been called \textit{Algebraic Multilevel Iteration} (AMLI) methods and the author proposes in this paper to rename the finite element version to FEMLI and to keep the name AMLI for the matrix graph version. It is shown that for the FEMLI method the condition number gets essentially stabilized when the short version of the method (only few coarse levels) is used. The stabilization is stronger for spatially three-dimensional than for two-dimensional problems. This holds for quite general classes of problems and for both \(h\)- and \(p\)-versions of finite element methods. In this paper the \(h\)-version using piecewise linear basis functions for a triangulation of a polygonal domain is considered but the proposed technique can be extended straightforwardly to three-dimensional (polytope) problems. Each iteration step of the FEMLI method can be implemented efficiently on massively parallel computers. It is shown that by balancing the size of the coarsest mesh to the finest and the number of processors used to these mesh sizes properly one can achieve a method of \(V\)-cycle FEMLI form which has both an essentially constant computational complexity per mesh point and an asymptotically optimal \textit{speedup} or \textit{efficiency}. These results have been shown to hold for elliptic problems with anisotropy and arbitrary jumps of the coefficients in the differential equations between the elements of the coarsest mesh, which is quite a fine mesh. Since many more general problems can be presented as a sequence of elliptic solvers, the FEMLI method is applicable also in such a context.