Lemma 1, p. 16 — Γ* ≠ ∅ and ℓ_rob < ∞ above κ√δβᵀA(x)⁻¹β, ℓ_rob = ∞ below; 𝕌₁ ⊆ 𝕌 ⊆ 𝕌₂
ProvedOTDRO.WorstCase.effective_domaindistributionally-robust-optimizationdualityeffective-domainp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Assume Assumptions 1 and 2, let and let be convex. Write for the growth exponent of , for the set of maximizers of , and . Then for every , and :
- if , then and ;
- if , then and .
Consequently, with ,
The lemma identifies the effective domain of the dual objective up to its boundary ; it is the reason a dual optimizer always satisfies and the region where the worst case of Theorem 6(c) is analysed.
Formalization Note The page prints the threshold as "", a typo for . The inclusions are stated pointwise: for and , implies , and implies . is a supremum of real numbers over , so it is never , and "finite" is written . The same statement is drafted in the sibling mission on convexity of the dual objective.
Preamble
import Mathlib import Definitions.Def_OTDRO_Dual_Setting open MeasureTheory
Formal statement
namespace OTDRO.WorstCase
/-- **Lemma 1** (arXiv:1810.02403v3, p. 16; proof p. 47). (a) If `λ > κ√δ βᵀA(x)⁻¹β` then
`Γ*(β, λ; x)` is nonempty and `ℓ_rob(β, λ; x)` is finite; (b) if `λ < κ√δ βᵀA(x)⁻¹β` then
`Γ*(β, λ; x) = ∅` and `ℓ_rob(β, λ; x) = ∞`; consequently `𝕌₁ ⊆ 𝕌 ⊆ 𝕌₂`, where
`𝕌 = {(β, λ) ∈ B × ℝ₊ : f_δ(β, λ) < ∞}`, `𝕌₁` (resp. `𝕌₂`) is `λ > λ_thr(β)` (resp. `≥`). -/
theorem effective_domain {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 ∈ B, ∀ lam : ℝ, 0 ≤ lam → ∀ x : EuclideanSpace ℝ (Fin d),
(OTDRO.Dual.growthKappa ℓ * Real.sqrt δ * OTDRO.Dual.quadInv A β x < lam →
(OTDRO.Dual.maximizers ℓ A δ β lam x).Nonempty ∧ OTDRO.Dual.ellRob ℓ A δ β lam x ≠ ⊤) ∧
(lam < OTDRO.Dual.growthKappa ℓ * Real.sqrt δ * OTDRO.Dual.quadInv A β x →
OTDRO.Dual.maximizers ℓ A δ β lam x = ∅ ∧ OTDRO.Dual.ellRob ℓ A δ β lam x = ⊤)) ∧
(∀ β ∈ B, ∀ lam : ℝ, 0 ≤ lam → OTDRO.Dual.lamThr P0 ℓ A δ β < lam → OTDRO.Dual.fDelta P0 ℓ A δ β lam < ⊤) ∧
(∀ β ∈ B, ∀ lam : ℝ, 0 ≤ lam → OTDRO.Dual.fDelta P0 ℓ A δ β lam < ⊤ → OTDRO.Dual.lamThr P0 ℓ A δ β ≤ lam) := by sorry
end OTDRO.WorstCase
Source
arXiv:1810.02403v3, Lemma 1, p. 16; (11), p. 15; proof p. 47
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.