Date of Award

August 2017

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Mathematics

Advisor(s)

Yuesheng Xu

Second Advisor

Uday Banerjee

Keywords

Banded matrices, Ill-conditioning, Large-scale linear systems, Parallel computing, Sparse orthogonal factorization

Subject Categories

Physical Sciences and Mathematics

Abstract

Sequential and parallel algorithms based on the LU factorization or the QR factorization have been intensely studied and widely used in the problems of computation with large-scale ill-conditioned banded matrices. Great concerns on existing methods include ill-conditioning, sparsity of factor matrices, computational complexity, and scalability. In this dissertation, we study a sparse orthogonal factorization of a banded matrix motivated by parallel computing. Specifically, we develop a process to factorize a banded matrix as a product of a sparse orthogonal matrix and a sparse matrix which can be transformed to an upper triangular matrix by column permutations. We prove that the proposed process requires low complexity, and it is numerically stable, maintaining similar stability results as the modified Gram-Schmidt process. On this basis, we develop a parallel algorithm for the factorization in a distributed computing environment. Through an analysis of its performance, we show that the communication costs reach the theoretical least upper bounds, while its parallel complexity or speedup approaches the optimal bound. For an ill-conditioned banded system, we construct a sequential solver that breaks it down into small-scale underdetermined systems, which are solved by the proposed factorization with high accuracy. We also implement a parallel solver with strategies to treat the memory issue appearing in extra large-scale linear systems of size over one billion. Numerical experiments confirm the theoretical results derived in this thesis, and demonstrate the superior accuracy and scalability of the proposed solvers for ill-conditioned linear systems, comparing to the most commonly used direct solvers.

Access

Open Access

Download

Share

COinS