On the relation between S-estimators and M-estimators of multivariate location and covariance (Q915298)

M-estimators of multivariate location and covariance based on \(\underset{\tilde{}} x_ 1,...,\underset{\tilde{}} x_ n\in {\mathbb{R}}^ p\) are considered as defined by \textit{P. J. Huber} [Robust statistics. (1981; Zbl 0536.62025)]. For the same problem, the S- estimator based on \(\rho:\;{\mathbb{R}}\to [0,\infty)\) and satisfying certain conditions, is defined as the solution \({\underset{\tilde{}} \theta}_ n=(\underset{\tilde{}} t_ n,\underset{\tilde{}} C_ n)\) to the problem of minimizing det(\(\underset{\tilde{}} C)\) subject to \[ n^{-1}\sum^{n}_{i=1}\rho [\{(\underset{\tilde{}} x_ i- \underset{\tilde{}} t)^ T\underset{\tilde{}} C^{- 1}(\underset{\tilde{}} x_ i-\underset{\tilde{}} t)\}^{1/2}]=b_ 0, \] for a certain constant \(b_ 0\). S-estimators are shown to satisfy the first-order conditions of M-estimators, as is the case in the estimation in multiple regression where S-estimators were originally proposed [see \textit{P. Rousseeuw} and \textit{V. Yohai}, Robust and nonlinear time series analysis, Proc. Workshop, Heidelberg/Ger. 1983, Lect. Notes Stat. 26, 256-272 (1984; Zbl 0567.62027)]. It is shown that the influence function of S-functionals exists and is the same as that of the corresponding M-functionals, and also the S- estimators have a limiting normal distribution which is similar to the limiting normal distribution of M-estimators. Finally the author concludes that the S-estimator is able to achieve the asymptotic variances attained by the M-estimator, but in addition it has a breakdown point that becomes considerably higher when the dimension p increases.