Zero-One Laws for Random Feasibility Problems

TL;DR

This paper introduces a general probabilistic model for combinatorial optimization problems with geometric structure, establishing strong concentration results for the feasibility margin under permutation symmetry, which unify and extend many existing results.

Contribution

It provides a unified framework for analyzing feasibility in combinatorial optimization problems using the margin, with new concentration results for a broad class of problems.

Findings

Strong concentration of the $oldsymbol{ ext{ell}^q}$-margin for $q o \infty$

Unified analysis of various optimization problems via the margin

Sharp threshold results for multiple combinatorial models

Abstract

We introduce a general random model of a combinatorial optimization problem with geometric structure that encapsulates both linear programming and integer linear programming. Let $Q$ be a bounded set called the feasible set, $E$ be an arbitrary set called the constraint set, and $A$ be a random linear transform. We define and study the $\ell^q$-margin, $M_q := d_q(AQ, E)$. The margin quantifies the feasibility of finding $y \in AQ$ satisfying the constraint $y \in E$. Our contribution is to establish strong concentration of the margin for any $q \in (2,\infty]$, assuming only that $E$ has permutation symmetry. The case of $q = \infty$ is of particular interest in applications -- specifically to combinatorial ``balancing'' problems -- and is markedly out of the reach of the classical isoperimetric and concentration-of-measure tools that suffice for $q \le 2$. Generality is a key feature of this result: we assume permutation symmetry of the constraint set and nothing else. This allows us to encode many optimization problems in terms of the margin, including random versions of: the closest vector problem, integer linear feasibility, perceptron-type problems, $\ell^q$-combinatorial discrepancy for $2 \le q \le \infty$, and matrix balancing. Concentration of the margin implies a host of new sharp threshold results in these models, and also greatly simplifies and extends some key known results.