The MCS estimator of location and scatter

摘要

In this note we introduce a new robust estimator of multivariate location and scatter. Like MVE and MCD it searches for an h-subset which minimizes a criterion. The difference is that the new criterion attempts to measure the cohesion of the h-subset. The optimal h-subset is called the Most Cohesive Subset (MCS). The MCS estimator uses projections and is affine equivariant. We construct a fast algorithm for the MCS estimator, and simulate its bias under various outlier configurations.