In Distributionally Robust Optimisation (DRO), the optimiser does not trust the probability distribution you hand it. Classical stochastic programming assumes you know the distribution of uncertain parameters exactly, then minimises expected cost. Classical robust optimisation assumes you only know a set of possible parameter values, then minimises worst-case cost. DRO sits in between: it assumes you have an estimated distribution, probably from data, but that the true distribution is only known to lie within an ambiguity set centred on the estimate. It then minimises the worst-case expected cost over every distribution in that set.
The consequence is a decision whose performance is guaranteed not just for your empirical distribution, but for every distribution close enough to it. The radius of "close enough" is a single tunable parameter, and moving it smoothly interpolates between stochastic programming (zero radius) and fully classical robust optimisation (infinite radius). That one-knob property is why DRO has displaced both extremes in several applied domains over the past decade.
The setup. A retailer has two years of daily sales data for a fast-fashion item and needs to decide how much to stock for the next six weeks. Holding inventory costs money; stock-outs cost more. The classical stochastic-programming answer is to fit a distribution to the two years of history and solve the newsvendor critical-ratio expression against it.
The problem. The two years do not cover last winter's supply-chain shock, the upcoming social-media trend cycle, or the competitor's pricing campaign that will launch next month. The empirical distribution is almost certainly wrong — the only honest question is how wrong.
The DRO response. The planner draws a ball around the empirical distribution in Wasserstein distance (an optimal-transport metric that measures how much probability mass must be moved to turn one distribution into another) and says: whatever the true demand distribution is, it lives somewhere inside this ball. The order quantity is then chosen to minimise the worst-case expected cost over every distribution in the ball. The output is a safety stock larger than the stochastic-programming answer and smaller than the fully robust answer, and the size of the difference is controlled by the ball radius.
The shape of the ambiguity set is where the applied craft lives. Two families dominate the literature.
Moment-based sets. The set contains every distribution whose first one or two moments (mean and covariance) match those of the empirical distribution, possibly within tolerances. These sets are analytically convenient — many DRO problems with moment-based ambiguity sets reduce to tractable semidefinite programs — but they are often too loose, in the sense that the worst-case distribution can be pathological (a two-point distribution placing all mass on extremes).
Wasserstein balls. The set contains every distribution within a fixed Wasserstein distance of the empirical distribution. These sets are tighter than moment-based sets because they preserve distributional shape, and they come with a clean data-driven interpretation: the radius shrinks as more data arrives. Wasserstein DRO is the dominant choice in recent supply-chain, energy, and machine-learning applications.
Other constructions — phi-divergence balls (Kullback-Leibler, chi-squared), kernel-based mean embeddings, factor-model ambiguity sets — trade off between tractability, interpretability, and how tightly they wrap the data. The choice of ambiguity-set family is the single largest modelling decision in DRO and the one most often left implicit in software.
Classical robust optimisation specifies an uncertainty set over parameter values (for example, "demand is between 100 and 200 units") and hedges against the worst parameter realisation inside that set. The decision must be feasible for every point in the uncertainty set.
Distributionally Robust Optimisation specifies an ambiguity set over probability distributions and hedges against the worst expected value across that set. The decision must minimise worst-case expected cost across every distribution in the set, not worst-case cost across every scenario.
The distinction matters because the DRO optimal decision is typically less conservative than the classical-robust optimal decision: it exploits the fact that all distributions inside the ball share structure (for example, the same empirical mean, or a common support), even when you cannot say which one is correct. A classical-robust model faced with "demand might be up to 200" plans for demand of 200. A DRO model faced with "demand averages around 150 but the distribution is uncertain" plans for an expectation hedged against the worst expectation in the ball, which lands somewhere meaningfully below 200.
A common misread is that DRO is equivalent to inflating the cost function or tightening a constraint by some margin. It is not. The ambiguity-set construction is the whole model: changing from a moment-based set to a Wasserstein ball, or changing the radius, changes the optimal decision in ways that do not correspond to any single safety-factor adjustment.
The test for whether an applied model is genuinely DRO is simple: ask what happens as more data arrives. A DRO model shrinks the ambiguity set (the radius is a function of sample size), so the decision converges to the stochastic-programming answer as the empirical distribution becomes trustworthy. A safety-margin model does not have this property; the margin is usually a constant chosen by policy, not by statistics.
This matters for product design. When a vendor claims a planning tool is “robust”, the diagnostic question is whether the conservatism parameter has a data-driven interpretation (radius shrinks with n) or whether it is a static slider the user tunes by feel. Only the first is DRO in any meaningful sense.
The first question to ask when a paper or product claims “robust optimisation”: is the uncertainty over outcomes (classical robust), over distributions (DRO), or simply over a known distribution (stochastic)? The answer determines the conservatism profile, the solver structure, and the tuning surface the user actually interacts with.
DRO looks intractable on its face: the inner problem is an optimisation over distributions, which is infinite-dimensional. The breakthrough of the past fifteen years is that for well-chosen ambiguity sets, the inner problem admits an exact dual reformulation as a finite convex program.
Moment-based DRO reformulates as a semidefinite program when the cost function is linear or quadratic in the uncertain parameters — tractable but scale-limited in practice.
Wasserstein DRO reformulates, under mild conditions on the cost function, as a regularised version of the empirical-risk problem. For certain classes of loss functions the Wasserstein-DRO problem reduces to a standard stochastic programme plus a regularisation term whose coefficient equals the ball radius. This is the result that made Wasserstein DRO deployable: a practitioner can reuse existing stochastic-programming solvers and add a single regularisation term, and the problem retains convexity.
For non-convex inner problems — the case in most industrial integer models — DRO is solved by scenario decomposition or cutting-plane methods, with the scenario generator reshaped to explore worst-case distributional perturbations rather than worst-case point realisations.
Distributionally Robust Optimisation is a hedge against being wrong about the probability distribution you assumed, not just against bad outcomes under the distribution you trust.