The symmetric eigenvalue problem: stochastic perturbation theory and some network applications.

Abstract

This thesis is concerned with stochastic perturbation theory of the symmetric eigen-value problem. In particular, we provide results about the probability of interchanges in the ordering of the eigenvalues and changes in the eigenvectors of symmetric matrices subject to stochastic perturbations. In this analysis we use a novel combination of traditional Numerical Linear Algebra, Perturbation Theory and Probability Theory. The motivation for this study arises from reliability of spectral clustering of networks, when network data is subject to noise. As far as we are aware, there is nothing comparable in the literature. Further, we make conjectures from which we derive an asymptotic relation between the distributions of the largest eigenvalue and the 2-norm of random symmetric ma- trices, whose entries above the main diagonal are independent, identically distributed random variables with probability density functions being symmetric with respect to zero, including matrices from the Gaussian Orthogonal Ensemble (GOE). As far as we know, some of these conjectures are not new (possibly only as conjectures) but we are not aware of any proofs. Also, we consider networks of coupled oscillators. In their analysis we use both, knowledge of dynamical systems and spectral properties of non-negative matrices. As a result, we present an algorithm, which uncovers the \master-slave" structure of the network. With its help, the analysis of the dynamics and the entrainment of the entire network can be reduced to considering only few of the oscillators, those whose dynamics determine the behaviour of the rest. This can be helpful in large networks exhibiting the \master-slave" structure. Finally, we consider similarities of spectral clustering with respect to di®erent matrices which can be associated with a given network. In particular, we compare clustering of products of Path graphs with respect to two di®erent matrices: the Laplacian and the Normalised Laplacian matrices of the graph. We make the comparison by constructing a Homotopy between two eigenvalue problems and, using some Linear Algebra techniques, we show that the two matrices give similar spectral clusterings when applied to products of Path graphs.

Date of Award	1 Oct 2008
Original language	English
Awarding Institution	University of Bath
Supervisor	Alastair Spence (Supervisor)