Efficient Iterative Methods for Saddle Point Problems

Abstract

This thesis investigates efficient iterative methods for a type of saddle-point problem, namely the generalized Stokes problem, which arises frequently in the simulation of time-dependent Navier-Stokes equations for incompressible fluid flow. These systems are indefinite due to a set of linear constraints on the velocity, causing difficulty for most preconditioners and iterative methods. A multilevel algorithm is proposed for the solution of such systems, which uses a novel technique for the construction of a basis for the space satisfying the constraints. The proposed algorithm achieves faster convergence on account of implicit preconditioning of the linear system, and can be implemented efficiently on parallel processors. Along with a scalable parallel implementation described in the thesis, the multilevel algorithm yields a competitive parallel preconditioned iterative method for the solution of these problems.