Thesis

Advanced polynomial matrix factorization techniques

Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2024
Thesis identifier
  • T16925
Person Identifier (Local)
  • 202171539
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • The advancements in polynomial matrix algebra have significantly bolstered the formulationof broadband array problems, enabling more effective solutions. This is particularlyevident through the utilization of polynomial matrix factorization techniquessuch as the polynomial eigenvalue decomposition (PEVD), the polynomial singularvalue decomposition (PSVD), and the polynomial QR decomposition (PQRD), whichhave notably revolutionized the approach to solving these complex problems. Therefore,these polynomial matrix factorization methods have received significant attentionover the last decades. Still, the polynomial order of the decompositions, complexityscaling with spatial and temporal dimensions of the matrix, and issues regarding parallelizabilityremain pertinent. Therefore, this thesis address these issues in these threementioned polynomial factorization methods.This thesis demonstrates that in most practical situations, eigen- and singular valueswill be spectrally majorised. We exploit this property in the algorithms proposed inthis thesis. The first and foremost algorithm is for applications such as data compactionor dominant component extraction, where the power method is extended to the para-Hermitian polynomial matrices for the extraction of a dominant eigenvector. Thisapproach prevents computing an entire PEVD of a para-Hermitian matrix. Later, thisextension is combined with a deflation approach to compute the PEVD of a low rankpolynomial matrix. Perturbation bounds are computed which show that these boundsincrease with repeated deflations. Ensemble tests reveal its superior performance overstate-of-the-art PEVD algorithms. To accommodate non-para-Hermitian polynomialmatrices, the polynomial power method is extended to general polynomial matricesfor the extraction of dominant left and right singular vectors and the corresponding dominant singular value.To reduce iterations in the polynomial power method to just one, the provided polynomialmatrix is decomposed into a sum of rank-one matrices using a computationallyinexpensive approach in the discrete Fourier transform (DFT) domain. Subsequently,each rank-one term is subjected to a single iteration of the polynomial power method,resulting in the corresponding eigenpair or the left- and right singular vectors and theassociated singular value. For the PQRD, rank one terms are obtained by computingQRDs in the DFT bins. To obtain any column of the paraunitary matrix and thecorresponding row of the upper-right triangular matrix, any column of the respectiverank one matrix is normalized to unit norm on the unit circle. This rank one decompositionbased method is termed as unified-I algorithm which is highly parallelizableand outperforms all state-of-the-art algorithms in accuracy of the decomposition andexecution time.Lastly, this thesis demonstrates that a further reduction in the order of any ofthe above decompositions can be achieved via assessing the auto- and cross-correlationterms of eigen- or singular vectors. A spectral factorisation of these terms can then leadto the compact polynomial order factors. This spectral factorization is obtained viapolynomial root finding method to avoid the positive definite restriction of a Laurentpolynomial on the unit circle. Ensemble results show its superior performance overstate-of-the-art algorithms including the rank decomposition method proposed in thisthesis.
Advisor / supervisor
  • Weiss, Stephan, 1968-
Resource Type
Note
  • Previously held under moratorium from 29th April 2024 until 22nd May 2025.
DOI

Relations

Items