distributionrobust

Distribution-robust, often written as distributionally robust, refers to approaches that guard against uncertainty in the data-generating distribution by optimizing across a set of plausible distributions, called an ambiguity set. Unlike standard empirical risk minimization, which minimizes loss under the empirical distribution, distribution-robust methods aim for solutions that perform well under the worst-case distribution within the ambiguity set.

Formulations usually take a minimax form: min_x sup_{P in A} E_P[L(x, Z)]. Ambiguity sets are defined by statistical distances or constraints, such as Wasserstein balls around the empirical distribution, phi-divergences (KL, chi-square), or moment and shape constraints. Many choices yield tractable reformulations via duality, often producing convex programs or semidefinite programs.

Applications span machine learning (classification and regression under distribution shift), finance (portfolio optimization under model risk), and operations research (inventory, scheduling, and supply chain planning under uncertain demand). It also appears in energy systems and health care decision making.

Relation to other ideas: It extends robust optimization to distributional uncertainty and connects to robust statistics and risk measures like CVaR. It contrasts with Bayesian approaches that encode uncertainty with priors.

Challenges include selecting an appropriate ambiguity set, balancing conservatism with empirical performance, and computational complexity for large problems. The field has grown since the early 2010s, with foundational work by Delage and Ye and continued advances across theory and applications.