Multivariate M-extremiles

Abstract

Extremiles represent an important class of distributional functionals defined through weighted squared loss functions, providing a valuable alternative to quantiles and expectiles. To date, their development has remained confined to the univariate setting. Here we introduce a robust multivariate extension of extremiles for random vectors by combining a directional weighting mechanism with Huber's multidimensional loss function. The proposed definition yields a family of directional extremiles indexed by a tuning parameter controlling the degree of robustness to outliers, with the quadratic-loss geometric extremile arising as a limiting case. We prove their main theoretical properties and introduce the notion of extremile sets. From an inferential perspective, we develop an efficient two-step procedure based on iteratively reweighted least squares, and derive asymptotic properties using the theory of Z-estimators. The finite-sample behavior of the proposed methodology is investigated through simulation experiments under Gaussian, contaminated, and heavy-tailed settings. An application to daily returns of three major cryptocurrencies illustrates its practical utility on financial data.