A preliminary test in discriminant analysis (Q1914700)

One of the aims of discriminant analysis is the allocation of unknown entities to populations that are known a priori. A preliminary matter for consideration before an outright or probabilistic allocation is made for an unclassified entity \(X\) is to test the assumption that \(X\) belongs to one of the predefined groups \(\pi_i (i = 1, 2, \dots, n)\). One way of approaching this question is to test if the smallest distance between \(X\) and \(\pi_i\) is null or not. Most of the results were obtained in the case of linear discriminant analysis where the law of the population is assumed to follow a normal and having a common variance-covariance matrix \(\Sigma\). The assumption of normality can be unrealistic. For example, in taxonomy or in medicine, discrete and continuous measurements are taken. The aim of this article is to propose a preliminary test to the general parametric case. In Section 2 we present a generalization of the Mahalanobis distance; then we study its distributional properties. In Section 4 we derive the distribution of the minimum of the divergence between \(X\) and \(\pi_i\). The implications of these results are discussed. An exposition of the proposed methodology as applied to the homoscedastic normal cases is presented in Section 5.