Spatial median and directional data

Abstract

We introduce the normalized spatial median as an estimator of location for rotationally symmetric distributions on the hypersphere. We investigate some of its asymptotic properties and use them to obtain confidence regions for the modal direction of a distribution on the hypersphere. These results are then applied to the von Mises-Fisher distribution and to a contamination model. It is seen that the normalized spatial median can perform more efficiently than the normalized mean in presence of outliers.