Optimization under rare events: scaling laws for linear chance-constrained programs

TL;DR

This paper analyzes how the optimal solutions of linear chance-constrained programs scale as the violation probability becomes very small, revealing the emergence of convex programs and assessing the performance of popular approximation methods.

Contribution

It introduces a scaling law framework for chance-constrained programs under rare events and compares the effectiveness of CVaR and sample approximations in different tail regimes.

Findings

Optimal value scales predictably as violation probability decreases.

CVaR and sample approximations are optimal under light-tailed distributions.

These methods are sub-optimal in heavy-tailed scenarios.

Abstract

We consider a class of chance-constrained programs in which profit needs to be maximized while enforcing that a given adverse event remains rare. Using techniques from large deviations and extreme value theory, we show how the optimal value scales as the prescribed bound on the violation probability becomes small and how convex programs emerge in the limit. We use our results to analyze the performance of existing popular approaches in the rare-event regime. We show that the popular CVaR and sample approximations have optimality properties under light-tailed assumptions on the randomness, while they behave sub-optimal in a heavy-tailed setting. Our results are derived using large deviations theory, extreme value theory, process techniques, and random set theory.