Computation and analysis of spectra of large undirected networks

---

Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Uygulamalı Matematik Enstitüsü, Türkiye

Tezin Onay Tarihi: 2010

Öğrenci: ÖZGE ERDEM

Danışman: BÜLENT KARASÖZEN

Özet:

Many interacting complex systems in biology, in physics, in technology and social systems, can be represented in a form of large networks. These large networks are mathematically represented by graphs. A graph is represented usually by the adjacency or the Laplacian matrix. Important features of the underlying structure and dynamics of them can be extracted from the analysis of the spectrum of the graphs. Spectral analysis of the so called normalized Laplacian of large networks became popular in the recent years. The Laplacian matrices of the empirical networks are in form of unstructured large sparse matrices. The aim of this thesis is the comparison of different eigenvalue solvers for large sparse symmetric matrices which arise from the graph theoretical epresentation of undirected networks. The spectrum of the normalized Laplacian is in the interval [0 2] and the multiplicity of the eigenvalue 1 plays a particularly important role for the network analysis. Moreover, the spectral analysis of protein-protein interaction networks has revealed that these networks have a different distribution type than other model networks such as scale free networks. In this respect, the eigenvalue solvers implementing the well-known implicitly restarted Arnoldi method, Lanczos method, Krylov-Schur and Jacobi Davidson methods are investigated. They exist as MATLAB routines and are included in some freely available packages. The performances of different eigenvalue solvers PEIG, AHBEIGS, IRBLEIGS, EIGIFP, LANEIG, JDQR, JDCG in MATLAB and the library SLEPc in C++ were tested for matrices of size between 100-13000 and are compared in terms of accuracy and computing time. The accuracy of the eigenvalue solvers are validated for the Paley graphs with known eigenvalues and are compared for large empirical networks using the residual plots and spectral density plots are computed.