Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Proof of Theorem 6(c), p. 39 — on 𝕌₁, ∂±f_δ/∂λ = √δ(1 − E[βᵀA(X)⁻¹β inf/sup_{g∈Γ*} g²])

Proved
OTDRO.WorstCase.dual_lambda_derivatives

by mikedeng1 · Oct 7, 2026 · Mathlib 0df444a (Lean v4.33.1)

directional-derivativedistributionally-robust-optimizationdualityp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1

Assume Assumptions 1 and 2, δ>0\delta>0δ>0, B⊆RdB\subseteq\mathbb R^dB⊆Rd convex, and fix β∈B\beta\in Bβ∈B with β≠0\beta\neq0β=0. Then fδ(β,λ)f_\delta(\beta,\lambda)fδ​(β,λ) is finite for every λ>λthr(β)\lambda>\lambda_{thr}(\beta)λ>λthr​(β), and for every such λ\lambdaλ: for P0P_0P0​-almost every xxx the set Γ∗(β,λ;x)\Gamma^*(\beta,\lambda;x)Γ∗(β,λ;x) is nonempty and {g2:g∈Γ∗(β,λ;x)}\{g^2:g\in\Gamma^*(\beta,\lambda;x)\}{g2:g∈Γ∗(β,λ;x)} is bounded, the functions x↦βTA(x)−1βinf⁡g∈Γ∗g2x\mapsto\beta^{\mathsf T}A(x)^{-1}\beta\inf_{g\in\Gamma^*}g^2x↦βTA(x)−1βinfg∈Γ∗​g2 and x↦βTA(x)−1βsup⁡g∈Γ∗g2x\mapsto\beta^{\mathsf T}A(x)^{-1}\beta\sup_{g\in\Gamma^*}g^2x↦βTA(x)−1βsupg∈Γ∗​g2 are P0P_0P0​-integrable, and the right and left derivatives of λ↦fδ(β,λ)\lambda\mapsto f_\delta(\beta,\lambda)λ↦fδ​(β,λ) are

∂+fδ∂λ(β,λ)=δ(1−EP0[βTA(X)−1βinf⁡g∈Γ∗(β,λ;X)g2]),∂−fδ∂λ(β,λ)=δ(1−EP0[βTA(X)−1βsup⁡g∈Γ∗(β,λ;X)g2]).\frac{\partial_+f_\delta}{\partial\lambda}(\beta,\lambda)=\sqrt\delta\Bigl(1-E_{P_0}\Bigl[\beta^{\mathsf T}A(X)^{-1}\beta\inf_{g\in\Gamma^*(\beta,\lambda;X)}g^2\Bigr]\Bigr),\qquad \frac{\partial_-f_\delta}{\partial\lambda}(\beta,\lambda)=\sqrt\delta\Bigl(1-E_{P_0}\Bigl[\beta^{\mathsf T}A(X)^{-1}\beta\sup_{g\in\Gamma^*(\beta,\lambda;X)}g^2\Bigr]\Bigr).∂λ∂+​fδ​​(β,λ)=δ​(1−EP0​​[βTA(X)−1βg∈Γ∗(β,λ;X)inf​g2]),∂λ∂−​fδ​​(β,λ)=δ​(1−EP0​​[βTA(X)−1βg∈Γ∗(β,λ;X)sup​g2]).

At a dual optimizer λ∗(β)>λthr(β)\lambda_*(\beta)>\lambda_{thr}(\beta)λ∗​(β)>λthr​(β) the first-order conditions ∂+fδ≥0≥∂−fδ\partial_+f_\delta\ge0\ge\partial_-f_\delta∂+​fδ​≥0≥∂−​fδ​ turn these formulas into c‾≤1≤cˉ\underline c\le1\le\bar cc​≤1≤cˉ, which is what lets the randomized worst case of Theorem 6(c) spend exactly the budget δ\deltaδ.

Formalization Note This is an unnumbered display in the proof of Theorem 6(c), which the page derives from Proposition 2 and a general result on directional derivatives of integral functionals ([2, Proposition 2.1]); it is stated here under Assumptions 1–2 only, without Proposition 2's max-of-C1C^1C1 form, which is what the proof of Theorem 6 uses. β≠0\beta\neq0β=0 is the hypothesis of Theorem 6, in whose proof the display appears; at β=0\beta=0β=0, Γ∗=R\Gamma^*=\mathbb RΓ∗=R and sup⁡g2=∞\sup g^2=\inftysupg2=∞. Finiteness of fδf_\deltafδ​ (neither +∞+\infty+∞ nor −∞-\infty−∞) on {λ>λthr(β)}\{\lambda>\lambda_{thr}(\beta)\}{λ>λthr​(β)} is a conjunct, so the derivatives of λ↦fδ(β,λ)\lambda\mapsto f_\delta(\beta,\lambda)λ↦fδ​(β,λ) are taken of its real values. Infimum and supremum over Γ∗\Gamma^*Γ∗ are real sInf/sSup of {g2}\{g^2\}{g2}; nonemptiness and boundedness are conjuncts, and integrability is a conjunct so that the expectations are genuine.

Preamble
import Mathlib
import Definitions.Def_OTDRO_Dual_Setting

open MeasureTheory
Formal statement
namespace OTDRO.WorstCase

/-- **One-sided `λ`-derivatives of `f_δ` on `𝕌₁`** (arXiv:1810.02403v3, §5.4, proof of
Theorem 6(c), p. 39; unnumbered). Under Assumptions 1–2, for `β ∈ B`, `β ≠ 0`: `f_δ(β, λ)` is
finite for `λ > λ_thr(β)`, and at every such `λ`
`∂₊f_δ/∂λ(β, λ) = √δ(1 − E[βᵀA(X)⁻¹β inf_{g∈Γ*(β,λ;X)} g²])`,
`∂₋f_δ/∂λ(β, λ) = √δ(1 − E[βᵀA(X)⁻¹β sup_{g∈Γ*(β,λ;X)} g²])`. -/
theorem dual_lambda_derivatives {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) :
    (∀ l : ℝ, OTDRO.Dual.lamThr P0 ℓ A δ β < l → OTDRO.Dual.fDelta P0 ℓ A δ β l ≠ ⊤ ∧ OTDRO.Dual.fDelta P0 ℓ A δ β l ≠ ⊥) ∧
    ∀ lam : ℝ, OTDRO.Dual.lamThr P0 ℓ A δ β < lam →
      (∀ᵐ x ∂P0, (OTDRO.Dual.maximizers ℓ A δ β lam x).Nonempty ∧
        BddAbove ((fun g : ℝ => g ^ 2) '' OTDRO.Dual.maximizers ℓ A δ β lam x)) ∧
      Integrable (fun x => OTDRO.Dual.quadInv A β x *
        sInf ((fun g : ℝ => g ^ 2) '' OTDRO.Dual.maximizers ℓ A δ β lam x)) P0 ∧
      Integrable (fun x => OTDRO.Dual.quadInv A β x *
        sSup ((fun g : ℝ => g ^ 2) '' OTDRO.Dual.maximizers ℓ A δ β lam x)) P0 ∧
      HasDerivWithinAt (fun l => (OTDRO.Dual.fDelta P0 ℓ A δ β l).toReal)
        (Real.sqrt δ * (1 - ∫ x, OTDRO.Dual.quadInv A β x *
          sInf ((fun g : ℝ => g ^ 2) '' OTDRO.Dual.maximizers ℓ A δ β lam x) ∂P0))
        (Set.Ici lam) lam ∧
      HasDerivWithinAt (fun l => (OTDRO.Dual.fDelta P0 ℓ A δ β l).toReal)
        (Real.sqrt δ * (1 - ∫ x, OTDRO.Dual.quadInv A β x *
          sSup ((fun g : ℝ => g ^ 2) '' OTDRO.Dual.maximizers ℓ A δ β lam x) ∂P0))
        (Set.Iic lam) lam := by sorry

end OTDRO.WorstCase
Source
arXiv:1810.02403v3, §5.4, proof of Theorem 6(c), p. 39
Human review
  • Endorsed by Shuze Chen · Oct 8, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 8, 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