Maximal entropy in the moment body

TL;DR

This paper introduces a novel, efficient method for moment body membership testing using a convex optimization approach inspired by quantum information theory, significantly improving scalability in large semidefinite programming problems.

Contribution

It proposes a new approach based on minimizing a smooth convex log-partition function, offering a geometric pre-conditioning technique and detailed complexity analysis for large-scale semidefinite problems.

Findings

Efficient handling of dense projections of size 1000 within seconds.

Cubic complexity dependence on matrix size, similar to eigenstructure computations.

Large-scale semidefinite programming bottleneck shifted to gradient storage and manipulation.

Abstract

A moment body is a linear projection of the spectraplex, the convex set of trace-one positive semidefinite matrices. Determining whether a given point lies within a given moment body is a problem with numerous applications in quantum state estimation or polynomial optimization. This moment body membership oracle can be addressed with semidefinite programming, for which several off-the-shelf interior-point solvers are available. In this paper, inspired by techniques from quantum information theory, we argue analytically and geometrically that a much more efficient approach consists of minimizing globally a smooth strictly convex log-partition function, dual to a maximum entropy problem. We analyze the curvature properties of this function and we describe a neat geometric pre-conditioning algorithm. A detailed complexity analysis reveals a cubic dependence on the matrix size, similar to a few eigenstructure computations. Basic numerical experiments illustrate that dense (i.e. non-sparse) projections of size 1000 of a dense semidefinite matrix of size 1000-by-1000 can be routinely handled in a few seconds on a standard laptop, thereby moving the main bottleneck in large-scale semidefinite programming almost entirely to efficient gradient storage and manipulation.