Some insights into depth estimators for location and scatter in the multivariate setting

Abstract: The concept of statistical depth has received considerable attention as a way to extend the notions of the median and quantiles to other statistical models. These procedures aim to formalize the idea of identifying deeply embedded fits to a model that are less influenced by contamination. Since contamination introduces bias in estimators, it is well known in the location model that the median minimizes the worst-case performance, in terms of maximum bias, among all equivariant estimators. In the multivariate case, Tukey's median was a groundbreaking concept for location estimation, and its counterpart for scatter matrices has recently attracted considerable interest. The breakdown point and the maximum asymptotic bias are key concepts used to summarize an estimator's behavior under contamination. For the location and scale model, we consider two closely related depth formulations, whose deepest estimators display significantly different behavior in terms of breakdown point. In the multivariate setting, we analyze recently introduced concentration inequalities that provide a unified framework for studying both the statistical convergence rate and robustness of Tukey's median and depth-based scatter matrices. We observe that slight variations in these inequalities allow us to visualize the maximum bias behavior of the deepest estimators. Since the maximum bias for depth-based scatter matrices had not previously been derived, we explicitly calculate both the breakdown point and the maximum bias curve for the deepest scatter matrices.