Il problema della riduzione del numero di parametri nell'analisi discriminante

Abstract

In this paper we discuss a statistical procedure for reducing the number of parameters to be estimated in the quadratic discriminant function. This procedure is based on the specification of relationships between covariance matrices of the populations. In the problem of discrimination between two populations, we have considered two particular relationships: the first specifies a common principal components model and the second equal correlation matrices. After a brief review of the methods, we compare the apparent error rate of the discriminant functions associated with the proposed models with those of the classical linear and quadratic functions for a set of simulated data.