Some remarks on allocatory and separatory linear discrimination

A random vector originates from one of r known normal populations having a common covariance matrix. We wish to reduce the dimension of the vector by means of a linear map from the original space down to a space of lower dimension while keeping the populations as separate as possible. The commonly used linear maps which are optimal for a class of measures of separation may be very poor in terms of a different criterion: the probability of correct classification calculated with no prior information about the population of origin.