Title: A new non-degenerate test for model selection based on maximum-mean-discrepancy
Authors: Florian Brueck - Technical University Munich (Germany) [presenting]
Jean-David Fermanian - Ensae-Crest (France)
Aleksey Min - Technical University of Munich (Germany)
Abstract: The purpose is to investigate the influence of parameter estimation on the asymptotic distribution of the two-sample test based on the Maximum-Mean-Discrepancy. To circumvent the problem of determining quantiles of an infinite sum of Chi-squared distributed random variables, we propose a new two-sample test based on Maximum-Mean-Discrepancy, which solely requires to determine quantiles of the standard normal distribution, while approximately keeping the power of the original two-sample test. Moreover, we deduce a new test for model comparison based on Maximum-Mean-Discrepancy, which only requires determining quantiles of the standard normal distribution.