Characterization of the asymptotic behavior of $U$-statistics on row-column exchangeable matrices

Tâm Le Minh

Published 2024-01-15Version 1

We consider $U$-statistics on row-column exchangeable matrices. We derive a decomposition for them, based on orthogonal projections on probability spaces generated by sets of Aldous-Hoover-Kallenberg variables. The specificity of these sets is that they are indexed by bipartite graphs, which allows for the use of concepts from graph theory to describe this decomposition. The decomposition is used to investigate the asymptotic behavior of $U$-statistics of row-column exchangeable matrices, including in degenerate cases. In particular, it depends only on a few terms of the decomposition, corresponding to the non-zero elements that are indexed by the smallest graphs, named principal support graphs, after an analogous concept suggested by Janson and Nowicki (1991). Hence, we show that the asymptotic behavior of a $U$-statistic and its degeneracy are characterized by the properties of their principal support graphs. Indeed, their number of nodes gives the convergence rate of a $U$-statistic to its limit distribution. Specifically, the latter is degenerate if and only if this number is strictly greater than 1. Also, when the principal support graphs are connected, we find that the limit distribution is Gaussian, even in degenerate cases.