Entropy of Exchangeable Random Graphs
Anda Skeja,Sofia C. Olhede
2024-03-15
Abstract:In this paper, we propose a complexity measure for exchangeable graphs by considering the graph-generating mechanism. Exchangeability for graphs implies distributional invariance under node permutations, making it a suitable default model for a wide range of graph data. For this well-studied class of graphs, we quantify complexity using graphon entropy. Graphon entropy is a graph property, meaning it is invariant under graph isomorphisms. Therefore, we focus on estimating the entropy of the generating mechanism for a graph realization, rather than selecting a specific graph feature. This approach allows us to consider the global properties of a graph, capturing its important graph-theoretic and topological characteristics, such as sparsity, symmetry, and connectedness. We introduce a consistent graphon entropy estimator that achieves the nonparametric rate for any arbitrary exchangeable graph with a smooth graphon representative. Additionally, we develop tailored entropy estimators for situations where more information about the underlying graphon is available, specifically for widely studied random graph models such as Erdős-Rényi, Configuration Model and Stochastic Block Model. We determine their large-sample properties by providing a Central Limit Theorem for the first two, and a convergence rate for the third model. We also conduct a simulation study to illustrate our theoretical findings and demonstrate the connection between graphon entropy and graph structure. Finally, we investigate the role of our entropy estimator as a complexity measure for characterizing real-world graphs.
Information Theory,Combinatorics,Statistics Theory