Stationary point complexity via minimal supersymmetry breaking

Abstract

The statistics of stationary points are a powerful way to understand mean-field random landscapes, and the Kac--Rice formula is a general way to compute them. A longstanding technical barrier to these calculations is the presence of the absolute value of the determinant of the Hessian. Neglecting the absolute value produces an elegant 2-index supersymmetric representation of the problem, but is often incorrect. We develop an expanded 4-index supersymmetric representation of the complexity problem which incorporates the absolute value naturally via spontaneous supersymmetry breaking along a particular superspace direction. Positing that no additional symmetry breaking occurs implies the reduction to five order parameters corresponding to elements of a superspace operator algebra generated by the spontaneously SUSY-breaking operator. We relate the order parameters to the geometry and spectra of stationary points, showing that the SUSY-breaking order parameter corresponds to the spectral density of the Hessian at zero eigenvalue. We give examples of this formalism applied to calculate the annealed complexity of several models, including the perceptron and the Sherrington--Kirkpatrick model. The framework is naturally extended to quenched complexity, where each order parameter corresponds to a replica matrix.