Browse
- No file added yet -

FAST ALGORITHMS FOR MATRIX COMPUTATION AND APPLICATIONS

Version 2 2023-12-11, 14:28
Version 1 2023-12-10, 18:01
thesis
posted on 2023-12-11, 14:28 authored by

Matrix decompositions play a pivotal role in matrix computation and applications. While general dense matrix-vector multiplications and linear equation solvers are prohibitively expensive, matrix decompositions offer fast alternatives for matrices meeting specific properties. This dissertation delves into my contributions to two fast matrix multiplication algorithms and one fast linear equation solver algorithm tailored for certain matrices and applications, all based on efficient matrix decompositions. Fast dimensionality reduction methods in spectral clustering, based on efficient eigen-decompositions, are also explored.

The first matrix decomposition introduced is the "kernel-independent" interpolative decomposition butterfly factorization (IDBF), acting as a data-sparse approximation for matrices adhering to a complementary low-rank property. Constructible in $O(N\log N)$ operations for an $N \times N$ matrix via hierarchical interpolative decompositions (IDs), the IDBF results in a product of $O(\log N)$ sparse matrices, each with $O(N)$ non-zero entries. This factorization facilitates rapid matrix-vector multiplication in $O(N \log N)$ operations, making it a versatile framework applicable to various scenarios like special function transformation, Fourier integral operators, and high-frequency wave computation.

The second matrix decomposition accelerates matrix-vector multiplication for computing multi-dimensional Jacobi polynomial transforms. Leveraging the observation that solutions to Jacobi's differential equation can be represented through non-oscillatory phase and amplitude functions, the corresponding matrix is expressed as the Hadamard product of a numerically low-rank matrix and a multi-dimensional discrete Fourier transform (DFT) matrix. This approach utilizes $r^d$ fast Fourier transforms (FFTs), where $r = O(\log n / \log \log n)$ and $d$ is the dimension, resulting in an almost optimal algorithm for computing the multidimensional Jacobi polynomial transform.

An efficient numerical method is developed based on a matrix decomposition, Hierarchical Interpolative Factorization, for solving modified Poisson-Boltzmann (MPB) equations. Addressing the computational bottleneck of evaluating Green's function in the MPB solver, the proposed method achieves linear scaling by combining selected inversion and hierarchical interpolative factorization. This innovation significantly reduces the computational cost associated with solving MPB equations, particularly in the evaluation of Green's function.

Finally, eigen-decomposition methods, including the block Chebyshev-Davidson method and Orthogonalization-Free methods, are proposed for dimensionality reduction in spectral clustering. By leveraging well-known spectrum bounds of a Laplacian matrix, the Chebyshev-Davidson methods allow dimensionality reduction without the need for spectrum bounds estimation. And instead of the vanilla Chebyshev-Davidson method, it is better to use the block Chebyshev-Davidson method with an inner-outer restart technique to reduce total CPU time and a progressive polynomial filter to take advantage of suitable initial vectors when available, for example, in the streaming graph scenario. Theoretically, the Orthogonalization-Free method constructs a unitary isomorphic space to the eigenspace or a space weighting the eigenspace, solving optimization problems through Gradient Descent with Momentum Acceleration based on Conjugate Gradient and Line Search for optimal step sizes. Numerical results indicate that the eigenspace and the weighted eigenspace are equivalent in clustering performance, and scalable parallel versions of the block Chebyshev-Davidson method and OFM are developed to enhance efficiency in parallel computing.

Degree Type

• Doctor of Philosophy

• Mathematics

Campus location

• West Lafayette

Haizhao Yang

Xiangxiong Zhang

Yuan Gao

Jie Shen

Exports

RefWorks
BibTeX
Ref. manager
Endnote
DataCite
NLM
DC
figshare. credit for all your research.