scientific article; zbMATH DE number 19010

From MaRDI portal
Publication:3976436





[For the entire collection see Zbl 0731.00019.] Let \({\mathcal M}_ p({\mathcal A})\) be the set of all probability measures on some sample space \((\Omega,{\mathcal A})\) and consider a parametric model \[ {\mathcal P}:=\{P_ \theta\mid\theta\in\mathbb{R}^ d\}\subset{\mathcal M}_ p({\mathcal A}). \] Decompose the parameter \(\Omega\in\mathbb{R}^ d\) into a \(d_ 1\)-dimensional parameter \(\theta_ 1\) of interest and a \(d_ 2\)- dimensional nuisance component \(\theta_ 2\), i.e. \(\theta=(\theta_ 1,\theta_ 2)\), \(\theta_ 1\in R^{d_ 1}\), \(\theta_ 2\in R^{d_ 2}\). Then tests for multisided, multiparameter hypotheses about \(\theta_ 1\) using i.i.d. observations \(x_ 1,\dots,x_ n\in\mathbb{R}\) are desired. Let \(T=T(\psi)\) be a functional that is Fisher consistent and differentiable at the model with influence curve \(\psi\). The multisided testing problem on \(T\), \[ H:T'\hbox{Cov}(\psi)^{-1}T\leq b^ 2\hbox{ vs. } K:T'\hbox{Cov}(\psi)^{-1}T\geq c^ 2, \] which extends the multiparameter hypotheses to nonparametric measures is studied using a locally asymptotic robust approach. Under the assumption of null infinitesimal neighbourhoods the author derives a locally asymptotic maximin (LAM) upper bound for the power of the test at level \(\alpha\). The selection of the functional \(T(\psi)\) to be tested leads to nontrivial optimality problems. In the one-dimensional, one-sided case, \[ H:T\leq b\hbox { vs. } K:T\geq c, \] one may want to test \(\epsilon\)- contamination or total variation neighbourhoods using functionals. Then maximization of the corresponding LAM bound yields an asymptotic version of Huber-Strassen's maximin test. In the multidimensional, multisided case, the same idea leads to the minimization of the maximum eigenvalue of the covariance standardized by the Fisher information subject to a bound on the self-standardized sensitivity.











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 Q3976436)