Discriminant analysis under elliptical populations (Q1336493)

This paper consists of two parts. In Part I, the Fisher's linear discriminant function (LDF) under elliptical populations with common covariance matrix is investigated. A minimax classification rule between two elliptical populations and an expression for the conditional misclassification probabilities (MP) of the LDF are given. In order to derive asymptotic expansion formulas for the MP, an asymptotic expansion of the joint distribution of the sample mean and the sample covariance matrix under an elliptical population is derived. The distributions of the conditional MP of Fisher's linear discrimination are considered. The asymptotic expansions of the expected probabilities of each kind of misclassification are obtained. An estimator of the MP which is unbiased up to the order \((n_ 1+ n_ 2)^{-3/2}\) is given. In Part II, the use of \(M\)-estimators is considered in order to get a robust classification rule. After a general setup of the estimation problem in discriminant analysis, some definitions and properties related with influence functions are considered. Then a measure of sensitivity and a measure of efficiency of the estimator based on the loss function are defined. The optimal robust \(M\)-estimators are obtained, and equivariant \(M\)-estimators are considered. Finally, the above results are applied to the elliptical model.