Spatial medians, depth functions and multivariate Jensen's inequality

Abstract

For any given partial order in a $d$-dimensional euclidean space, under mild regularity assumptions, we show that the intersection of closed (generalized) intervals containing more than 1/2 of the probability mass, is a non-empty compact interval. This property is shared with common intervals on real line, where the intersection is the median set of the underlying probability distribution. So obtained multivariate medians with respect to a partial order, can be observed as special cases of centers of distribution in the sense of type D depth functions introduced by Y. Zuo and R. Serfling, {\em Ann. Stat.}, {\bf 28} (2000), 461-482. We show that the halfspace depth function can be realized via compact convex sets, or, for example, closed balls, in place of halfspaces, and discuss structural properties of halfspace and related depth functions and their centers. Among other things, we prove that, in general, the maximal guaranteed depth is $\frac{1}{d+1}$. As an application of these results, we provide a Jensen's type inequality for functions of several variables, with medians in place of expectations, which is an extension of the previous work by M. Merkle, {\em Stat. Prob. Letters}, {\bf 71} (2005), 277--281.}