Interaction-correlated random matrices

Abstract

We introduce a family of random matrices where correlations between matrix elements are induced via interaction-derived Boltzmann factors. Varying these yields access to different ensembles. We find a universal scaling behavior of the finite-size statistics characterized by a heavy-tailed eigenvalue distribution whose extremes are governed by the Fréchet extreme value distribution for the case corresponding to a ferromagnetic Ising transition. The introduction of a finite density of nonlocal interactions restores standard random-matrix behavior. Suitably rescaled average extremes, playing a physical role as an order parameter, can thus discriminate aspects of the interaction structure; they also yield further nonuniversal information. In particular, the link between maximum eigenvalues and order parameters offers a potential route to resolving long-standing problems in statistical physics, such as deriving the exact magnetization scaling function in the two-dimensional Ising model at criticality.

4 figures

References in corpus (7)

• Machine learning and deep learning
• Topological and Dynamical Complexity of Random Neural Networks
• Random matrix analysis of complex networks
• Condensation and Extreme Value Statistics
• Breakdown of random-matrix universality in persistent Lotka--Volterra communities
• Exact finite-size scaling for the random-matrix representation of bond percolation on square lattice
• Universal scaling and criticality of extremes in random matrix theory

Cited by in corpus (2)

• Phase diagram and eigenvalue dynamics of stochastic gradient descent in multilayer neural networks
• q-Gaussian Crossover in Overlap Spectra towards 3D Edwards-Anderson Criticality