Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 4.4 (Worst-case distributions), p. 14 — under Assumption 4.1, (10) equals the value of (13), and near-optimal points of (13) give distributions attaining (10) asymptotically

Proved
WassDDRO.Extremal.theorem_4_4

by mikedeng1 · Oct 6, 2026 · Mathlib 0df444a (Lean v4.33.1)

convex-reformulationdistributionally-robust-optimizationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1wassersteinworst-case-distribution

Let EEE be a finite-dimensional real normed space with its Borel σ\sigmaσ-algebra, let N,K≥1N, K \ge 1N,K≥1, let Ξ⊆E\Xi \subseteq EΞ⊆E contain the samples ξ^1,…,ξ^N\hat\xi_1,\dots,\hat\xi_Nξ^​1​,…,ξ^​N​, let ε≥0\varepsilon \ge 0ε≥0, and let ℓ1,…,ℓK:E→R‾\ell_1,\dots,\ell_K : E \to \overline{\mathbb R}ℓ1​,…,ℓK​:E→R be measurable with ℓ=max⁡kℓk\ell = \max_k \ell_kℓ=maxk​ℓk​. Suppose Assumption 4.1 holds: Ξ\XiΞ is convex and closed, each −ℓk-\ell_k−ℓk​ is proper, convex and lower semicontinuous, and no ℓk\ell_kℓk​ is identically −∞-\infty−∞ on Ξ\XiΞ. Then:

  1. The worst-case expectation (10) equals the optimal value of the finite convex program (13):
sup⁡Q∈Bε(P^N)EQ[ℓ(ξ)]  =  sup⁡αik,qik{1N∑i=1N∑k=1Kαikℓk(ξ^i−qikαik):1N∑i,k∥qik∥≤ε, ∑kαik=1, αik≥0, ξ^i−qikαik∈Ξ}.\sup_{\mathbb Q\in\mathbb B_\varepsilon(\widehat{\mathbb P}_N)}\mathbb E^{\mathbb Q}[\ell(\xi)] \;=\; \sup_{\alpha_{ik},q_{ik}}\Big\{\frac1N\sum_{i=1}^N\sum_{k=1}^K\alpha_{ik}\ell_k\Big(\hat\xi_i-\frac{q_{ik}}{\alpha_{ik}}\Big) : \frac1N\sum_{i,k}\|q_{ik}\|\le\varepsilon,\ \sum_k\alpha_{ik}=1,\ \alpha_{ik}\ge0,\ \hat\xi_i-\frac{q_{ik}}{\alpha_{ik}}\in\Xi\Big\}.Q∈Bε​(PN​)sup​EQ[ℓ(ξ)]=αik​,qik​sup​{N1​i=1∑N​k=1∑K​αik​ℓk​(ξ^​i​−αik​qik​​):N1​i,k∑​∥qik​∥≤ε, k∑​αik​=1, αik​≥0, ξ^​i​−αik​qik​​∈Ξ}.
  1. Let (αik(r),qik(r))r∈N(\alpha_{ik}(r), q_{ik}(r))_{r\in\mathbb N}(αik​(r),qik​(r))r∈N​ be feasible decisions of (13) whose objective values converge to the supremum of (13), and put ξik(r)=ξ^i−qik(r)/αik(r)\xi_{ik}(r) = \hat\xi_i - q_{ik}(r)/\alpha_{ik}(r)ξik​(r)=ξ^​i​−qik​(r)/αik​(r) and Qr=1N∑i,kαik(r)δξik(r)\mathbb Q_r = \frac1N\sum_{i,k}\alpha_{ik}(r)\delta_{\xi_{ik}(r)}Qr​=N1​∑i,k​αik​(r)δξik​(r)​. Then every Qr\mathbb Q_rQr​ lies in Bε(P^N)\mathbb B_\varepsilon(\widehat{\mathbb P}_N)Bε​(PN​), and
sup⁡Q∈Bε(P^N)EQ[ℓ(ξ)]=lim⁡r→∞EQr[ℓ(ξ)]=lim⁡r→∞1N∑i=1N∑k=1Kαik(r) ℓ(ξik(r)).\sup_{\mathbb Q\in\mathbb B_\varepsilon(\widehat{\mathbb P}_N)}\mathbb E^{\mathbb Q}[\ell(\xi)] = \lim_{r\to\infty}\mathbb E^{\mathbb Q_r}[\ell(\xi)] = \lim_{r\to\infty}\frac1N\sum_{i=1}^N\sum_{k=1}^K\alpha_{ik}(r)\,\ell(\xi_{ik}(r)).Q∈Bε​(PN​)sup​EQ[ℓ(ξ)]=r→∞lim​EQr​[ℓ(ξ)]=r→∞lim​N1​i=1∑N​k=1∑K​αik​(r)ℓ(ξik​(r)).

Fractions in (13) follow the paper's conventions: when αik=0\alpha_{ik} = 0αik​=0 the constraint forces qik=0q_{ik} = 0qik​=0 and the ikikik-th objective term is 000.

The theorem shows that the worst case over a Wasserstein ball around the empirical distribution can be approached by discrete distributions with at most NKNKNK atoms, computed from a finite convex program; this is what makes stress tests of a decision against its extremal distributions practical.

Formalization Note Values are in EReal, so the limits may be +∞+\infty+∞; no finiteness is assumed. The expectation is the extended expectation of p. 5. The printed second limit "lim⁡k→∞\lim_{k\to\infty}limk→∞​" is read as lim⁡r→∞\lim_{r\to\infty}limr→∞​ (the sequence index). The hypotheses N,K≥1N, K \ge 1N,K≥1, ξ^i∈Ξ\hat\xi_i \in \Xiξ^​i​∈Ξ and the measurability of each ℓk\ell_kℓk​ are the paper's standing assumptions (§2, p. 5 and p. 11).

Preamble
import Mathlib
import Definitions.Def_WassDDRO_Extremal_Setting

open MeasureTheory Filter Topology
Formal statement
namespace WassDDRO.Extremal

/-- Theorem 4.4 (Worst-case distributions), p. 14. Under Assumption 4.1 the worst-case
expectation (10) equals the optimal value of the finite convex program (13), for every ε ≥ 0;
and for every feasible sequence of (13) whose objective values converge to the supremum of (13),
the discrete distributions Q_r lie in the Wasserstein ball and attain (10) asymptotically. The
printed second limit index "k → ∞" is the sequence index r. -/
theorem theorem_4_4 {E : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [FiniteDimensional ℝ E]
    [MeasurableSpace E] [BorelSpace E]
    {K N : ℕ} (hK : 0 < K) (hN : 0 < N) (Ξ : Set E) (ℓ : Fin K → E → EReal)
    (hmeas : ∀ k, Measurable (ℓ k)) (ξhat : Fin N → E) (hξ : ∀ i, ξhat i ∈ Ξ)
    (ε : ℝ) (hε : 0 ≤ ε) (hA : Assumption41 Ξ ℓ) :
    worstCaseExpectation ε Ξ ξhat (maxLoss ℓ) = program13Value ε Ξ ξhat ℓ ∧
    ∀ (α : ℕ → Fin N → Fin K → ℝ) (q : ℕ → Fin N → Fin K → E),
      (∀ r, Feasible13 ε Ξ ξhat (α r) (q r)) →
      Tendsto (fun r => objective13 ξhat ℓ (α r) (q r)) atTop
        (𝓝 (program13Value ε Ξ ξhat ℓ)) →
      (∀ r, discreteQ ξhat (α r) (q r) ∈ WassersteinDRO.Duality.ambiguitySet ε 1 Ξ
          (WassersteinDRO.Duality.empiricalDistribution ξhat)) ∧
      Tendsto (fun r => DupacovaWets.Consistency.expect (discreteQ ξhat (α r) (q r)) (maxLoss ℓ)) atTop
        (𝓝 (worstCaseExpectation ε Ξ ξhat (maxLoss ℓ))) ∧
      Tendsto (fun r => ((1 / (N : ℝ) : ℝ) : EReal) *
          ∑ i, ∑ k, term13 (maxLoss ℓ) (ξhat i) (α r i k) (q r i k)) atTop
        (𝓝 (worstCaseExpectation ε Ξ ξhat (maxLoss ℓ))) := by sorry

end WassDDRO.Extremal
Source
Mohajerin Esfahani & Kuhn, arXiv:1505.05116v3, Theorem 4.4, p. 14; proof pp. 15–17
Human review
  • Endorsed by Shuze Chen · Oct 6, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 6, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me