What are the differences and connections between various Robust Optimizations?

Robust optimization represents a family of methodologies designed to make decisions perform well under uncertainty, and its various forms are distinguished primarily by their underlying philosophy for modeling that uncertainty and their resulting computational tractability. The most fundamental division lies between "worst-case" approaches, typified by classical robust optimization rooted in the work of Soyster and later Ben-Tal and Nemirovski, and "probabilistic" or "distributionally robust" approaches. The classical worst-case paradigm constructs a single, deterministic uncertainty set—such as a box, ellipsoid, or polyhedron—and optimizes for the best performance guaranteed against any realization within that set. This connection to convex optimization and duality theory makes it tractable for many problems but is often criticized for excessive conservatism, as protecting against all possible, even extreme, scenarios can lead to overly costly or performance-poor solutions. In contrast, distributionally robust optimization (DRO) occupies a middle ground, acknowledging that some distributional information is available without requiring full knowledge of the exact probability law. It optimizes against the worst-case probability distribution from an ambiguity set—a family of distributions defined by known moments, support, or divergence measures from a reference distribution. This creates a direct connection to, and generalization of, both classical robust optimization (by using an ambiguity set of all distributions on a fixed support) and traditional stochastic programming (by using a singleton ambiguity set containing only the nominal distribution).

The differences between these paradigms manifest in their data requirements, computational complexity, and the nature of the solutions they yield. Classical robust optimization requires only an uncertainty set, making it data-light but potentially overly conservative. Stochastic programming requires a known precise distribution, which is often unavailable, and its solutions can be fragile to misspecification. DRO explicitly trades off these two, using available data to construct an ambiguity set that reflects statistical confidence, thereby seeking less conservative solutions than classical robust optimization while offering stronger guarantees than stochastic programming if the true distribution deviates from the nominal estimate. The specific geometry of the uncertainty or ambiguity set is the primary lever controlling this trade-off. For instance, a box uncertainty set in classical robust optimization leads to linear programming formulations but is the most conservative, while an ellipsoidal set, connected to conic optimization, can reduce conservatism by excluding extreme corner cases. In DRO, an ambiguity set based on Wasserstein distance or moment constraints connects to tractable reformulations, often via duality, that can be solved with convex or semi-definite programming, but the complexity increases with the sophistication of the set.

The connections between these approaches are deep and often formal. Technically, they are linked through a hierarchy of conservatism: a classical robust formulation can frequently be recovered as a special case of a DRO model where the ambiguity set contains all distributions supported on the uncertainty set. Conversely, a DRO model can be viewed as a stochastic program with a worst-case distribution. Furthermore, from a data-driven perspective, many modern robust optimization techniques are connected through their statistical interpretation. The parameters defining an uncertainty set in classical robust optimization (like the size of a budget-of-uncertainty set) or an ambiguity set in DRO (like the radius of a Wasserstein ball) can be calibrated from data to provide probabilistic guarantees on constraint satisfaction, known as chance-constraint approximations. This bridges the methodologies to statistical learning theory. Ultimately, the choice between them is not merely technical but strategic, dictated by the available information quality, the operational cost of constraint violation, and the computational budget for solving the resulting optimization problem. The field continues to evolve by creating new hybrids, such as incorporating adjustable or recoverable robustness into these frameworks to add decision-stage adaptability, further blurring the traditional categorical lines while expanding the toolkit for managing uncertainty.