An isobar-surface approach to multidimensional outlier-proneness (Q1424672)

Notions of absolutely outlier resistant (AOR) and absolutely outlier prone (AOP) for multidimensional distributions are introduced. A univariate distribution \(F\) is AOR if \(X_{n,n}-X_{n,n-1}\to 0\) in probability (\(n\to\infty\)), where \(X_{j,n}\) are order statistics of i.i.d. samples from \(F\). A univariate distribution \(F\) is AOP if there are \(\varepsilon,\delta>0\) such that \(\Pr\{X_{n,n}-X_{n,n-1}> \varepsilon\}\geq\delta\). A possible characterization of multidimensional AOR is as follows: Let \(X\) be a r.v. in \(R^k\), \(R=\| X\| \), \(\Theta=X/\| X\| \), and \(F_\vartheta\) be the conditional distribution of \(R\) given \(\Theta=\vartheta\). The distribution \(F\) of \(X\) is AOR if \(F_\vartheta\) is AOR for all \(\vartheta\) and \(F\) is AOP if \(F_\vartheta\) is AOP for some \(\vartheta\). The authors consider different characterizations of AOR and AOP distributions in terms of distribution isobars. It is shown that a bivariate Gaussian distribution is AOR, and a bivariate Morgenstern distribution with density \[ f(x,y)= e^{-(x+y)}(1+\alpha(2e^{-x}-1)(2e^{-y}-1)), \quad -1\leq \alpha\leq 1, \] is AOP. An example of a distribution which is neither AOP nor AOR is presented.