Formulation and applications of complex network theory.

Abstract

In recent years, there has been a great surge of interest among physicists in modeling social, technological, or biological systems as networks. Analyses of large-scale networks such as the Internet have led to discoveries of many unexpected network properties, including power-law degree distributions. These discoveries have prompted physicists to devise novel ways to model networks, both computational and theoretical. In this dissertation, we present several network models and their applications. First, we study the theory of Exponential Random Graphs. We derive it from the principle of maximum entropy, thereby showing that it is the equivalent of the Gibbs ensemble for networks. Using tools of statistical physics, we solve well-known and new examples that include power-law networks and the two-star model. Our solutions confirm the existence of a first-order phase transition for the latter whose exact behavior has not been presented previously. We also study degree correlations and clustering in networks. Degrees of adjacent vertices are positively correlated in social networks, whereas they are negatively correlated in other types of networks. We demonstrate that a negative degree correlation is a more natural state of a network, and therefore that social networks are an exception. We argue that variations in the number of vertices in social groups cause positive degree correlations, and analyze a model that incorporates such a mechanism. The model indeed shows a high level of degree correlation and clustering that is similar in value to those of real networks. Finally, we develop algorithms for ranking vertices in networks that represent pairwise comparisons. The first algorithm is based on the familiar concept of indirect wins and losses. The second algorithm is based on the concept of retrodictive accuracy, which is maximized by positioning as many winners above the losers as possible. We compare the rankings of American college football teams generated by our algorithms with those officially used in college football's postseason proceedings, highlighting their simliarities and differences. These studies will show that the methods of statistical physics are well-suited for studying networks. As networks of interest become larger and more complex, they will become more relevant and necessary.