Ordered overlaps in disordered mean-field models (Q5954656)

This paper studies large deviation properties in general mean-field disordered spin systems, such as Hopfield models or some generalizations. \(N\) spins are considered and the disorder is described by Bernoulli random variables called patterns, which enter in the definition of the mean-field type Hamiltonian. The authors focus here on the case of a sublinear number of patterns \(M= o(N)\), which goes to infinity when the number of spins \(N\) goes to infinity. They provide under the original Gibbs measure a full LDP for some ordered overlap, quantities introduced to describe the interaction of the spins and the patterns. The rate function is studied in details and a LDP is also proved for the distribution of the \(L^2\)-norm of the overlap, which appears to be a crucial order parameter in a large class of examples. As an interpretation of their results, they retrieve the phenomenon of condensation of a single pattern in the case of the Hopfield model, and describe a new and very different phenomenon, called Gaussian noise condensation, in more general cases. This helps to understand the non replica-symmetry breaking in this sublinear case, and might be a step to study physicists' prediction at the mathematical level in the more general regime \(M=O(N)\).