Probability densities from distances and discrimination

Given a population and a random vector X, by using distances between observations of X, we prove that it is, in general, possible to construct probability densities for X. This distance-based approach can present problems, from a multidimensional scaling point of view, for some monotonic density functions, where the construction must be made on the basis of symmetric functions instead of distances. A measure of divergence between the true density and this construction is given. The procedure aims to offer alternative methods for performing discriminant analysis.