Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Remark 4, (11) for f=1Af = 1_Af=1A​ — ε\varepsilonε-optimal plans nearly satisfy complementary slackness

Proved
ModelRiskOT.WorstProb.remark_4_eq_11

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

dualityoptimal-transportp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1worst-case-probability

Let SSS be a Polish space, ccc a cost satisfying (A1), μ\muμ a probability measure on SSS, δ>0\delta > 0δ>0, A⊆SA \subseteq SA⊆S nonempty and closed, and I=sup⁡{π(S×A):π∈Φμ,δ}I = \sup\{\pi(S \times A) : \pi \in \Phi_{\mu,\delta}\}I=sup{π(S×A):π∈Φμ,δ​}. Let λ∗≥0\lambda^* \ge 0λ∗≥0 be such that (λ∗,φλ∗)(\lambda^*, \varphi_{\lambda^*})(λ∗,φλ∗​) is dual optimal, where φλ∗(x)=(1−λ∗c(x,A))+\varphi_{\lambda^*}(x) = (1 - \lambda^* c(x,A))^+φλ∗​(x)=(1−λ∗c(x,A))+; that is,

I=λ∗δ+Eμ[(1−λ∗c(X,A))+].I = \lambda^*\delta + E_\mu\big[(1 - \lambda^* c(X,A))^+\big].I=λ∗δ+Eμ​[(1−λ∗c(X,A))+].

Let ε>0\varepsilon > 0ε>0 and let πε∈Φμ,δ\pi_\varepsilon \in \Phi_{\mu,\delta}πε​∈Φμ,δ​ satisfy πε(S×A)≥I−ε\pi_\varepsilon(S \times A) \ge I - \varepsilonπε​(S×A)≥I−ε. Then

∫(φλ∗(x)−(1A(y)−λ∗c(x,y))) dπε(x,y)≤ε,\int \Big(\varphi_{\lambda^*}(x) - \big(1_A(y) - \lambda^* c(x,y)\big)\Big)\,d\pi_\varepsilon(x,y) \le \varepsilon,∫(φλ∗​(x)−(1A​(y)−λ∗c(x,y)))dπε​(x,y)≤ε,

and, if λ∗>0\lambda^* > 0λ∗>0,

(δ−ελ∗)+≤∫c dπε≤δ.\Big(\delta - \frac{\varepsilon}{\lambda^*}\Big)^+ \le \int c\,d\pi_\varepsilon \le \delta .(δ−λ∗ε​)+≤∫cdπε​≤δ.

The two bounds say that an ε\varepsilonε-optimal transport plan almost moves mass only to maximizers of 1A(y)−λ∗c(x,y)1_A(y) - \lambda^* c(x,y)1A​(y)−λ∗c(x,y) and almost exhausts the budget; they are the quantitative form of the complementary slackness conditions, and Lemma 4 and Lemma 2 are built on them.

Formalization Note The page prints the first integrand as φλ∗(x)−f(y)−λ∗c(x,y)\varphi_{\lambda^*}(x) - f(y) - \lambda^* c(x,y)φλ∗​(x)−f(y)−λ∗c(x,y); the statement uses the integrand of (10), φλ∗(x)−(f(y)−λ∗c(x,y))\varphi_{\lambda^*}(x) - (f(y) - \lambda^* c(x,y))φλ∗​(x)−(f(y)−λ∗c(x,y)), which is nonnegative. The second bound is stated only for λ∗>0\lambda^* > 0λ∗>0, where ε/λ∗\varepsilon/\lambda^*ε/λ∗ is defined (the only case the paper uses). Dual optimality is stated as the equality above, which is I=J(λ∗,φλ∗)I = J(\lambda^*, \varphi_{\lambda^*})I=J(λ∗,φλ∗​) after the identity sup⁡y{1A(y)−λc(x,y)}=(1−λc(x,A))+\sup_y\{1_A(y) - \lambda c(x,y)\} = (1 - \lambda c(x,A))^+supy​{1A​(y)−λc(x,y)}=(1−λc(x,A))+. The condition πε(S×A)≥I−ε\pi_\varepsilon(S\times A) \ge I - \varepsilonπε​(S×A)≥I−ε is written I≤πε(S×A)+εI \le \pi_\varepsilon(S \times A) + \varepsilonI≤πε​(S×A)+ε in [0,∞][0,\infty][0,∞].

Preamble
import Mathlib
import Definitions.Def_ModelRiskOT_WorstProb_WorstCaseProb

open MeasureTheory
open scoped ENNReal
Formal statement
namespace ModelRiskOT.WorstProb

/-- Remark 4, (11), p. 8, for `f = 1_A`. Let `(λ*, φ_{λ*})` be dual optimal, where by §2.4
`φ_{λ*}(x) = (1 − λ* c(x, A))⁺`; i.e. `I = λ*δ + E_μ[(1 − λ* c(X, A))⁺]` with
`I = sup {π(S × A) : π ∈ Φ_{μ,δ}}`. Then for `ε > 0` and every `π_ε ∈ Φ_{μ,δ}` with
`π_ε(S × A) ≥ I − ε`:
`∫ (φ_{λ*}(x) − (1_A(y) − λ* c(x, y))) dπ_ε ≤ ε`, and, when `λ* > 0`,
`(δ − ε/λ*)⁺ ≤ ∫ c dπ_ε ≤ δ`.
The integrand is the corrected one of (10) (the page prints `φ_{λ*}(x) − f(y) − λ* c(x, y)`). -/
theorem remark_4_eq_11 {S : Type*} [TopologicalSpace S] [PolishSpace S] [MeasurableSpace S]
    [BorelSpace S] (c : S → S → ℝ) (hc : ModelRiskOT.Duality.AssumptionA1 c) (μ : Measure S) [IsProbabilityMeasure μ]
    (δ : ℝ) (hδ : 0 < δ) (A : Set S) (hA : IsClosed A) (hAne : A.Nonempty)
    (lamStar : ℝ) (hlam : 0 ≤ lamStar)
    (hdual : worstProbCoupling c μ δ A = obj13 c μ δ A lamStar)
    (ε : ℝ) (hε : 0 < ε) (π : Measure (S × S)) (hπ : π ∈ Phi c μ δ)
    (hπε : worstProbCoupling c μ δ A ≤ π (Set.univ ×ˢ A) + ENNReal.ofReal ε) :
    ∫⁻ p, ENNReal.ofReal (max (1 - lamStar * costToSet c A p.1) 0 -
        (A.indicator (fun _ => (1 : ℝ)) p.2 - lamStar * c p.1 p.2)) ∂π ≤ ENNReal.ofReal ε ∧
    (0 < lamStar →
      ENNReal.ofReal (δ - ε / lamStar) ≤ ∫⁻ p, ENNReal.ofReal (c p.1 p.2) ∂π ∧
      ∫⁻ p, ENNReal.ofReal (c p.1 p.2) ∂π ≤ ENNReal.ofReal δ) := by sorry

end ModelRiskOT.WorstProb
Source
Blanchet & Murthy, Quantifying Distributional Model Risk via Optimal Transport, arXiv:1604.01446v2, p. 8, Remark 4, Eq. (11), for f = 1_A
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