Skip to content

DCT Lexicon · Governing Dynamics

Wasserstein Distributionally Robust Optimization

When the probability law itself is uncertain, optimise against the worst law within a Wasserstein distance of the data. The ball contains the distributions an honest analyst cannot rule out.

DCT Governing Dynamics

In depth

Formal statement

With empirical law P^N\hat P_N and radius ε\varepsilon, the ambiguity set is a Wasserstein ball, and the robust problem minimises the worst expected loss over it.

Bε(P^N)={Q:Wp(Q,P^N)≤ε},Wp(Q,P)p=inf⁡γ∈Π(Q,P)∫∥ξ−ζ∥p dγ\mathcal{B}_\varepsilon(\hat P_N) = \big\{ Q : W_p(Q, \hat P_N) \le \varepsilon \big\},\qquad W_p(Q,P)^p = \inf_{\gamma \in \Pi(Q,P)} \int \|\xi - \zeta\|^p\, d\gammasup⁡Q∈BεEQ[ℓ(x,ξ)]  =  inf⁡λ≥0{λεp+1N∑i=1Nsup⁡ξ[ℓ(x,ξ)−λ∥ξ−ξ^i∥p]}\sup_{Q \in \mathcal{B}_\varepsilon} \mathbb{E}_Q[\ell(x,\xi)] \;=\; \inf_{\lambda \ge 0} \Big\{ \lambda \varepsilon^p + \frac1N \sum_{i=1}^N \sup_{\xi} \big[ \ell(x,\xi) - \lambda \|\xi - \hat\xi_i\|^p \big] \Big\}

DCT Governing Dynamics: the element's drawer (/governing-dynamics)

Strong duality and finite-sample guarantees

The worst-case expectation over a Wasserstein ball equals the finite-dimensional dual above (Mohajerin Esfahani–Kuhn; Gao–Kleywegt). For light-tailed laws, a radius shrinking at the rate of the empirical measure's concentration makes the optimal value an upper bound on the out-of-sample cost with prescribed confidence.

DCT Governing Dynamics: the element's drawer (/governing-dynamics)

In DCT

The records of funding closures and deviations are short. Designing against the records' family, not one fitted law, prices ambiguity explicitly: the difference between the robust and the nominal programme is the price of not knowing the law.

DCT Governing Dynamics: the element's drawer (/governing-dynamics)

Limits

The radius is a modelling choice with no free lunch: too small and the guarantee fails, too large and the design is paralysed. Rates suffer the curse of dimensionality in the dimension of ξ\xi.

DCT Governing Dynamics: the element's drawer (/governing-dynamics)

Go next

  • Chance Constraints and Certificates: Distributionally robust chance constraints certify across the ball (Pathway 10).
  • Robust Viability: Robust viability keeps the kernel under the worst law.
  • Dynamic Risk Measures: Coherent risk measures are suprema over ambiguity sets.

DCT Governing Dynamics: the element's drawer (/governing-dynamics)

References

  • Mohajerin Esfahani & Kuhn (2018)
  • Gao & Kleywegt (2023)
  • Villani (2009), Optimal Transport: Old and New

DCT Governing Dynamics: the element's drawer (/governing-dynamics)

In the book: Vol. II, Ch. 11: Distributionally Robust Enterprise Optimization

Related: Chance Constraints and Certificates · Robust Viability · Dynamic Risk Measures

GD.17