Network representation using graph root distributions (Q2039784)

The author introduces a new parameterization of exchangeable random graphs (i.e., random graphs which are probabilistically invariant under a permutation of the vertices) satisfying some mild conditions on a related spectral decomposition. This parameterization is in terms of a family of probability distributions (called \textit{graph root distributions}) on a separable Kreĭn space. Issues of identifiability are discussed. It is shown that closeness of two graph root distributions in a certain Wasserstein distance implies closeness of the corresponding graphons in cut distance. Statistical questions of estimation using graph root distributions are considered for both dense and sparse random graphs. The paper concludes with numerical examples to illustrate its results.