Spin models on random graphs

Abstract

In the past decades complex networks and their behavior have attracted much attention. In the real world many of such networks can be found, for instance as social, information, technological and biological networks. An interesting property that many of them share is that they are scale free. Such networks have many nodes with a moderate amount of links, but also a significant amount of nodes with a very high number of links. The latter type of nodes are called hubs and play an important role in the behavior of the network. To model scale free networks, we use power-law random graphs. This means that their degree sequences obey a power law, i.e., the fraction of vertices that have k neighbors is proportional to k- for some > 1. Not only the structure of these networks is interesting, also the behavior of processes living on these networks is a fascinating subject. Processes one can think of are opinion formation, the spread of information and the spread of viruses. It is especially interesting if these processes undergo a so-called phase transition, i.e., a minor change in the circumstances suddenly results in completely different behavior. Hubs in scale free networks again have a large influence on processes living on them. The relation between the structure of the network and processes living on the network is the main topic of this thesis. We focus on spin models, i.e., Ising and Potts models. In physics, these are traditionally used as simple models to study magnetism. When studied on a random graph, the spins can, for example, be considered as opinions. In that case the ferromagnetic or antiferromagnetic interactions can be seen as the tendency of two connected persons in a social network to agree or disagree, respectively. In this thesis we study two models: the ferromagnetic Ising model on power-law random graphs and the antiferromagnetic Potts model on the Erd?os-Renyi random graph. For the first model we derive an explicit formula for the thermodynamic limit of the pressure, generalizing a result of Dembo and Montanari to random graphs with power-law exponent > 2, for which the variance of degrees is potentially infinite. We furthermore identify the thermodynamic limit of the magnetization, internal energy and susceptibility. For this same model, we also study the phase transition. We identify the critical temperature and compute the critical exponents of the magnetization and susceptibility. These exponents are universal in the sense that they only depend on the power-law exponent and not on any other detail of the degree distribution. The proofs rely on the locally tree-like structure of the random graph. This means that the local neighborhood of a randomly chosen vertex behaves like a branching process. Correlation inequalities are used to show that it suffices to study the behavior of the Ising model on these branching processes to obtain the results for the random graph. To compute the critical temperature and critical exponents we derive upper and lower bounds on the magnetization and susceptibility. These bounds are essentially Taylor approximations, but for power-law exponents 5 a more detailed analysis is necessary. We also study the case where the power-law exponent 2 (1, 2) for which the mean degree is infinite and the graph is no longer locally tree-like. We can, however, still say something about the magnetization of this model. For the antiferromagnetic Potts model we use an interpolation scheme to show that the thermodynamic limit exists. For this model the correlation inequalities do not hold, thus making it more difficult to study. We derive an extended variational principle and use to it give upper bounds on the pressure. Furthermore, we use a constrained secondmoment method to show that the high-temperature solution is correct for high enough temperature. We also show that this solution cannot be correct for low temperatures by showing that the entropy becomes negative if it were to be correct, thus identifying a phase transition.