Median balls: An extension of the interquantile intervals to multivariate distributions (Q1375106): Difference between revisions

For a probability distribution \(P\) on a Banach space \((X,\;\|.\|)\), interquantile intervals are defined by \([F^{-1} (\alpha), F^{-1} (1-\alpha)]\), and are used as central regions for studies of location, dispersion, skewness, or kurtosis. A median ball of \(P\) with nonnegative radius \(\lambda\) is any closed ball \(B(c^*,\lambda)\) with center \(c^*\), such that there exists a mapping \(\varphi^*\) in the set \(\Phi (c^*, \lambda)\) of measurable mappings from \(X\) to \(X\) for which \(\varphi^* (X) \subset B (c^*, \lambda)\), satisfying \(f(\varphi^*) =\inf_{c\in X} \inf_{\varphi\in \Phi(c, \lambda)}\). When \(\lambda =0\) they give the spatial median, and for a real distribution they are the interquantile intervals. Some properties are studied, and in particular it is shown that they possess robustness and equivariance properties similar to those of the spatial median.