scientific article; zbMATH DE number 16766

From MaRDI portal
Publication:3973451





[For the entire collection see Zbl 0725.00013.] In previous papers the author has studied the following minimax problem: Let \(\phi_ 0\) be a simple hypothesis and let \(\phi_ N\) be a sequence of composite alternatives approaching \(\phi_ 0\) as \(N\to \infty\). Then at what rate of convergence of \(\phi_ N\) to \(\phi_ 0\) in this metric or another is it still possible to distinguish between \(\phi_ 0\) and \(\phi_ N\) with probabilities of errors \(\alpha_{Nn}\) and \(\beta_{Nn}\) tending to 0 when the sample size \(n\to \infty\)? In the present paper this approach to minimax testing is carried further in several directions: the null hypothesis \(\phi_ 0\) is nonparametric (like the hypothesis of independence); as test statistics \(U\)-statistics are considered. The kernels \(K_ N\) of these \(U\)-statistics incorporate randomness (in a certain sense they are ``estimated by a subsample). The conditions which guarantee \(\alpha_{Nn}+\beta_{Nn}\to 0\) are given essentially in terms of ratios of different norms of the kernels \(K_ N\).











This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3973451)