Collaborative Research: AF: Small: Sampling and Optimization under Global Constraints

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

Sampling from complex, high dimensional probability distributions and optimizing high dimensional functions are two central computational tasks, with applications in machine learning, statistics, physics, and many other fields, both industrial and scientific. A widespread and very useful framework for studying both of these tasks is provided by probabilistic graphical models. This framework originates in statistical physics, and physical intuition has guided both the development of algorithms and the understanding of computational complexity. A common setting in both practical and scientific applications is the imposition of global constraints on a graphical model; examples including fixing the number of interacting particles in a system, fixing magnetization in a mathematical model of a magnet, or fixing the sizes of a partition of elements. A rigorous understanding of the performance of algorithms in this setting is still largely absent. The goal of this project is to develop new techniques in algorithms and complexity to understand how global constraints influence the tractability of these sampling and optimization problems. Guided by insights from statistical physics, this research project will develop new algorithms and study the limits of algorithmic approaches. The project also involves the training of graduate students in interdisciplinary research and a high-school outreach program. The main research goal of the project is to build a robust theory of sampling and optimization in the presence of global constraints in several algorithmic contexts: computational thresholds for approximate counting and sampling; Markov chain mixing times; and the structure of optimization problems on random graphs. Preliminary work shows that imposing a global constraint on a Markov random field can transform the behavior and complexity of the associated sampling and optimization problems. Specific problems the project focuses on include the determination of computational thresholds for the antiferromagnetic Ising model at fixed magnetization on bounded degree graphs; proving optimal mixing time bounds for conservative dynamics like the Kawasaki dynamics; and understanding the limiting values of optimization problems on sparse random graphs. The approach is guided by intuition from statistical physics, a field that provides a natural language and suite of techniques for analyzing graphical models and a wealth of examples of different phenomena that result from imposing global constraints. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.