Topology trivialization and large deviations for the minimum in the simplest random optimization

TL;DR

This paper investigates the landscape and large deviations of the minimum in a simple random optimization problem, revealing two regimes of topology trivialization and deriving the distribution of the global minimum energy.

Contribution

It identifies two large-N regimes with distinct critical point counts and derives the probability distribution of the global minimum energy using the replica method.

Findings

Two regimes of topology trivialization with different critical point counts

Distribution of the global minimum energy related to Tracy-Widom statistics

Large deviation rate function for the minimum energy derived using replica method

Abstract

Finding the global minimum of a cost function given by the sum of a quadratic and a linear form in N real variables over (N-1)- dimensional sphere is one of the simplest, yet paradigmatic problems in Optimization Theory known as the "trust region subproblem" or "constraint least square problem". When both terms in the cost function are random this amounts to studying the ground state energy of the simplest spherical spin glass in a random magnetic field. We first identify and study two distinct large-N scaling regimes in which the linear term (magnetic field) leads to a gradual topology trivialization, i.e. reduction in the total number N_{tot} of critical (stationary) points in the cost function landscape. In the first regime N_{tot} remains of the order $N$ and the cost function (energy) has generically two almost degenerate minima with the Tracy-Widom (TW) statistics. In the second regime the number of critical points is of the order of unity with a finite probability for a single minimum. In that case the mean total number of extrema (minima and maxima) of the cost function is given by the Laplace transform of the TW density, and the distribution of the global minimum energy is expected to take a universal scaling form generalizing the TW law. Though the full form of that distribution is not yet known to us, one of its far tails can be inferred from the large deviation theory for the global minimum. In the rest of the paper we show how to use the replica method to obtain the probability density of the minimum energy in the large-deviation approximation by finding both the rate function and the leading pre-exponential factor.