Theses and Dissertations

Date of Award

8-1-2026

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Applied Mathematics

First Advisor

Josef Sifuentes

Second Advisor

Brandt Kronholm

Third Advisor

Erwin Suazo

Abstract

Often in solving large linear systems of equations, direct methods like Gaussian elimination are infeasible due to how computationally expensive this can be. Often, one uses Krylov subspace-based iterative methods like Generalized Minimal Residuals (GMRES) or Full Orthogonalization Method (FOM) to solve such systems, especially when the coefficient matrix is sparse or can be multiplied by a vector faster than n squared complexity. One situation where this occurs is in solving large systems of differential equations utilizing implicit time stepping methods. In this situation, one must solve a sequence of related linear systems. In this paper, we propose preconditioning strategies that glean information from previous runs of GMRES to build an effective preconditioner for subsequent solves in order to lessen the number of iterations required per step. We explore implementations of this preconditioning method both in serial and in parallel. We demonstrate that these approaches are successful in accelerating convergence for GMRES iterative methods. For linear systems that are nearly singular, but well conditioned on a subspace, we expand upon known deflation techniques in FOM to GMRES. We also demonstrate that deflation preconditioning approaches are also effective. We show both approaches are successful at solving the linear system on the subspace for which the matrix is well conditioned.

Comments

Copyright 2026 Marina Nacianceno. All Rights Reserved. https://proquest.com/docview/3409151117

Share

COinS