Title: Phase transitions in optimal unsupervised learning

Abstract: We determine the optimal performance of learning the orientation of the symmetry axis of a set of P = alpha N points that are uniformly distributed in all the d irections but one on the N-dimensional sphere. The components along the symmetry breaking direction, of unitary vector B, are sampled from a mixture of two gaus sians of variable separation and width. The typical optimal performance is measu red through the overlap Ropt=B.J* where J* is the optimal guess of the symmetry breaking direction. Within this general scenario, the learning curves Ropt(alpha) may present first order transitions if the clusters are narrow enough. Close t o these transitions, high performance states can be obtained through the minimiz ation of the corresponding optimal potential, although these solutions are metas table, and therefore not learnable, within the usual bayesian scenario.