Abstract:This thesis is concerned with two classical topics in matrix computations : The QR algorithm for solving nonsymmetric eigenvalue problems and the computation of matrix exponentials for two types of structured matrices. We focus on the performance in the former topic and on accuracy in the latter one. For computing all eigenvalues of a non-Hermitian matrix, the QR algorithm which iteratively computes a Schur decomposition of the matrix is the method of choice. We present a new parallel implementation of the multishift QR algorithm targeting distributed memory architectures. Starting from recent developments of the parallel multishift QR algorithm, we propose a number of algorithmic and implementation improvements. Guidelines concerning several important tunable algorithmic parameters are also provided. Numerous computational experiments confirm that our new implementation significantly outperforms previous parallel implementations of the QR algorithm. The computation of the exponential of a square matrix is also an important task in matrix computations. For a general dense matrix, the scaling and squaring method coupled with Pade approximation is the most popular approach. However, for an essentially nonnegative matrix (a real square matrix with nonnegative off-diagonal entries), truncated Taylor series rather than Pade approximation is preferred to achieve componentwise accuracy in the matrix exponential. We propose a method which efficiently computes all entries of the exponential of an essentially nonnegative matrix to high relative accuracy. Truncation and rounding error bounds, as well as numerical experiments demonstrate the efficiency and accuracy of our method. When the matrix is banded, the entries of its matrix exponential decay exponentially away from the main diagonal. We analyze the decay property for the exponentials of several classes of doubly-infinite skew-Hermitian matrices. Then finite section methods based on the decay property are established. We also propose a repeated doubling strategy which works well even when a priori error estimates are pessimistic or not easy to compute. Finally, numerical experiments are presented to illustrate the effectiveness of the finite section method.

A Parallelizable Method for Computing the Singular Subspaces of a General Matrix

A Cubically Convergent Parallelizable Method for the Hermitian Eigenvalue Problem

Fast parallelizable methods for the Hermitian eigenvalue problem

Fast Parallelizable Methods for Computing Invariant Subspaces of Hermitian Matrices

A Quadratically Convergent QR-like Method Without Shifts for the Hermitian Eigenvalue Problem

Convergence of a Ulm-like Method for Square Inverse Singular Value Problems with Multiple and Zero Singular Values.

An Extended Ulm-like Method for Inverse Singular Value Problems with Multiple And/or Zero Singular Values

A Parallel Iterative Method For Solving Symmetrical Tridiagonal Extreme Eigenpair Problems

Dense and Structured Matrix Computations —the Parallel QR Algorithm and Matrix Exponentials

Extended Newton-type Method for Inverse Singular Value Problems with Multiple And/or Zero Singular Values

A parallel iterative method for solving symmetric tridiagonal extreme eigenpair problems

Towards a more robust algorithm for computing the restricted singular value decomposition

The Singular Value Decomposition, Applications and Beyond.

Newton-type Methods for Inverse Singular Value Problems with Multiple Singular Values

A Contour Integral-Based Algorithm for Computing Generalized Singular Values

An Augmented Matrix-Based CJ-FEAST SVDsolver for Computing a Partial Singular Value Decomposition with the Singular Values in a Given Interval

On Choices of Formulations of Computing the Generalized Singular Value Decomposition of a Large Matrix Pair.

Limited Memory Block Krylov Subspace Optimization for Computing Dominant Singular Value Decompositions.

An Ulm-like Cayley Transform Method for Inverse Eigenvalue Problems with Multiple Eigenvalues

The Convergence of Krylov Subspace Methods for Large Unsymmetric Linear Systems

A Hybrid Algorithm for Computing a Partial Singular Value Decomposition Satisfying a Given Threshold