Theorem 2, p. 10 — f_δ : B × ℝ₊ → ℝ ∪ {∞} is proper and convex under Assumptions 1 and 2
ProvedOTDRO.Cvx.dual_proper_convexLet be a Borel probability measure on , let satisfy Assumption 1 and and satisfy Assumption 2, let and let the decision set be convex. Consider the dual objective
with values in . Then is proper and convex:
- for every ;
- for every there is with ;
- for and ,
By Theorem 1 of the paper, the distributionally robust objective equals , so the robust problem becomes a joint convex minimization of over ; this is the basis of the paper's stochastic gradient schemes.
Formalization Note is the EReal-valued integral . Convexity is the displayed inequality in EReal (with ), since ConvexOn does not apply to EReal-valued maps; together with item 1 the right-hand side never involves . "Proper" is item 1 together with item 2; item 2 is stated for each , which is stronger than "finite at some point of " whenever and avoids a nonemptiness hypothesis on (for the statement is vacuous). The convexity of is the paper's standing assumption (p. 1); is the paper's standing assumption on the budget.
import Mathlib import Definitions.Def_OTDRO_Dual_Setting
namespace OTDRO.Cvx
open MeasureTheory
/-- **Theorem 2**, p. 10: under Assumptions 1 and 2, the dual objective
`f_δ(β, λ) = E_{P₀}[ℓ_rob(β, λ; X)]`, a map `B × ℝ₊ → ℝ ∪ {∞}`, is proper and convex:
1. it never takes the value `−∞` on `B × ℝ₊`;
2. for every `β ∈ B` it is finite at some `λ ≥ 0`;
3. for `(βᵢ, λᵢ) ∈ B × ℝ₊` and `α ∈ [0, 1]`,
`f_δ(αβ₁ + (1 − α)β₂, αλ₁ + (1 − α)λ₂) ≤ α f_δ(β₁, λ₁) + (1 − α) f_δ(β₂, λ₂)`. -/
theorem dual_proper_convex {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) (δ : ℝ) (hδ : 0 < δ) (B : Set (EuclideanSpace ℝ (Fin d))) (hB : Convex ℝ B)
(h2 : OTDRO.Dual.Assumption2 P0 ℓ) :
(∀ β ∈ B, ∀ lam : ℝ, 0 ≤ 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 ≠ ⊤) ∧
(∀ β₁ ∈ B, ∀ β₂ ∈ B, ∀ lam₁ lam₂ : ℝ, 0 ≤ lam₁ → 0 ≤ lam₂ →
∀ α : ℝ, 0 ≤ α → α ≤ 1 →
OTDRO.Dual.fDelta P0 ℓ A δ (α • β₁ + (1 - α) • β₂) (α * lam₁ + (1 - α) * lam₂) ≤
(α : EReal) * OTDRO.Dual.fDelta P0 ℓ A δ β₁ lam₁ +
((1 - α : ℝ) : EReal) * OTDRO.Dual.fDelta P0 ℓ A δ β₂ lam₂) := by sorry
end OTDRO.Cvx
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.