The coupling method in extreme value theory (Q2040094): Difference between revisions

This is an interesting paper showing how the coupling method can be used in univariate extreme value theory. First, the authors introduce the required background on coupling theory and Wasserstein spaces. Second, an important tool establishing equality of the Wasserstein distance between the distributions of the empirical measures of two iid samples of the same size and the Wasserstein distance of the underlying distributions generating the samples is proved. The proof uses elementary coupling arguments together with the more subtle Kantorovitch duality. Third, this tool is applied to EVT, mostly in the PoT framework. For the sake of clarity the heavy-tailed case is considered first, and a sharp estimate between the empirical measure of observations above high-threshold and the empirical measure associated with iid Pareto observations is given. Statistical consequences for the Hill estimator and Weissman extreme quantile estimates are presented as corollaries. Finally, extensions covering the asymptotic regime where bias occurs, the generalization to all domain of attractions, and some results in the BM framework are considered. All proofs are gathered in Section 4, and for the sake of brevity, some proofs are deferred to the supplementary material.