Theorem 6(c), p. 14 — if λ*(β) > λ_thr(β), the law of X + √δ G A(X)⁻¹β, G = ZG₋ + (1 − Z)G₊, attains the worst case with E[c(X, X*)] = δ
ProvedOTDRO.WorstCase.worst_case_randomized_mongeAssume Assumptions 1 and 2, and convex; fix with and a dual optimizer . Suppose . Then there are measurable functions such that, for -almost every , and belong to and
the moments and are finite; lies in ; and, with a Bernoulli variable independent of , and
the joint law of satisfies and attains the worst case:
The theorem describes an adversarial distribution explicitly: each data point is moved along the single direction by an amount read off the one-dimensional problem (7), randomizing between two extreme maximizers only to spend the transport budget exactly.
Formalization Note The page defines and (with for ). Its proof (p. 40) instead takes measurable selections with and , and only that choice makes follow from the first-order conditions. With the printed choice the statement can fail: for , , , , with large and uniform on two suitable points, is a dual optimizer with at one point and at the other, so the printed and no mixture has cost . The statement follows the proof. "The law of attains the supremum in (6)" is encoded with the published primal problem of Blanchet–Murthy: is a coupling with first marginal and cost at most (primalFeasible) whose value equals the supremum primalValue over all such couplings; the cost is the lower Lebesgue integral of . The random is encoded through the mixture law (mixtureCoupling). The convexity of is the paper's standing assumption (p. 1).
import Mathlib import Definitions.Def_ModelRiskOT_Duality_primalFeasible import Definitions.Def_ModelRiskOT_Duality_primalValue import Definitions.Def_OTDRO_Dual_Setting import Definitions.Def_OTDRO_WorstCase_Transport open MeasureTheory
namespace OTDRO.WorstCase
/-- **Theorem 6(c)** (arXiv:1810.02403v3, p. 14; proof §5.4, pp. 39–40). Under Assumptions 1–2,
for `β ≠ 0` and a dual optimizer `λ*(β) > λ_thr(β)`, there are measurable selections `G₋, G₊` of
`Γ*(β, λ*(β); ·)` such that, with `c̄ = E[G₊² βᵀA⁻¹β]`, `c̲ = E[G₋² βᵀA⁻¹β]` and
`q = (c̄ − 1)/(c̄ − c̲) ∈ [0, 1]`, the joint law of `(X, X*)`,
`X* = X + √δ (Z G₋ + (1 − Z) G₊) A(X)⁻¹β` with `Z ∼ Bernoulli(q)` independent of `X`, is a
feasible coupling that attains the worst-case value and has transport cost exactly `δ`.
`G₋, G₊` are the elements of `Γ*` of least and greatest square, as in the proof (p. 40). -/
theorem worst_case_randomized_monge {d : ℕ} (P0 : Measure (EuclideanSpace ℝ (Fin d)))
[IsProbabilityMeasure P0] (A : EuclideanSpace ℝ (Fin d) → Matrix (Fin d) (Fin d) ℝ)
(ρmin ρmax : ℝ) (hA : OTDRO.Dual.Assumption1 P0 A ρmin ρmax) (ℓ : ℝ → ℝ) (h2 : OTDRO.Dual.Assumption2 P0 ℓ)
(δ : ℝ) (hδ : 0 < δ) (B : Set (EuclideanSpace ℝ (Fin d))) (hB : Convex ℝ B)
(β : EuclideanSpace ℝ (Fin d)) (hβB : β ∈ B) (hβ : β ≠ 0)
(lamStar : ℝ) (hlam0 : 0 ≤ lamStar)
(hmin : IsMinOn (fun l => OTDRO.Dual.fDelta P0 ℓ A δ β l) (Set.Ici 0) lamStar)
(hthr : OTDRO.Dual.lamThr P0 ℓ A δ β < lamStar) :
∃ Gminus Gplus : EuclideanSpace ℝ (Fin d) → ℝ,
AEMeasurable Gminus P0 ∧ AEMeasurable Gplus P0 ∧
(∀ᵐ x ∂P0, Gminus x ∈ OTDRO.Dual.maximizers ℓ A δ β lamStar x ∧
Gplus x ∈ OTDRO.Dual.maximizers ℓ A δ β lamStar x ∧
∀ g ∈ OTDRO.Dual.maximizers ℓ A δ β lamStar x, Gminus x ^ 2 ≤ g ^ 2 ∧ g ^ 2 ≤ Gplus x ^ 2) ∧
Integrable (fun x => Gminus x ^ 2 * OTDRO.Dual.quadInv A β x) P0 ∧
Integrable (fun x => Gplus x ^ 2 * OTDRO.Dual.quadInv A β x) P0 ∧
0 ≤ bernoulliWeight (sqMoment P0 A β Gplus) (sqMoment P0 A β Gminus) ∧
bernoulliWeight (sqMoment P0 A β Gplus) (sqMoment P0 A β Gminus) ≤ 1 ∧
mixtureCoupling P0 A δ β Gminus Gplus ∈
ModelRiskOT.Duality.primalFeasible (OTDRO.Dual.mahalCost A) P0 δ ∧
ModelRiskOT.Duality.primalObj (fun x => ℓ (inner ℝ β x))
(mixtureCoupling P0 A δ β Gminus Gplus) =
ModelRiskOT.Duality.primalValue (OTDRO.Dual.mahalCost A) (fun x => ℓ (inner ℝ β x)) P0 δ ∧
∫⁻ p, ENNReal.ofReal (OTDRO.Dual.mahalCost A p.1 p.2) ∂(mixtureCoupling P0 A δ β Gminus Gplus) =
ENNReal.ofReal δ := by sorry
end OTDRO.WorstCase
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.