(15), proof of Theorem 4, p. 18 — a strictly positive ℙ₂ with I(ℙ′₀,ℙ₂) < r and c(x,ℙ₀) < c(x,ℙ₂) + ϵ
ProvedDROOptimal.Predictor.eq_15distributionally-robust-optimizationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1relative-entropy
Let be compact, , continuous in for each , , and the relative entropy. Let , , and let satisfy and . Then for every there is with
In the proof of Theorem 4, is a maximizer in (10); the model is almost as costly as , strictly inside the relative entropy ball, and strictly positive, so that the large deviation lower bound (7b) applies to it.
Formalization Note The paper also records . That clause is dropped: it is never used, and it can fail (for the relative entropy vanishes identically). The decision is fixed; may depend on it.
Preamble
import Mathlib import Definitions.Def_DROOptimal_Predictor_Setting
Formal statement
namespace DROOptimal.Predictor
/-- Display (15), proof of Theorem 4, pp. 17–18: let r > 0, x ∈ X, ℙ′₀ ∈ 𝒫 and let ℙ₀ ∈ 𝒫 satisfy
I(ℙ′₀, ℙ₀) ≤ r and ĉ_r(x, ℙ′₀) = c(x, ℙ₀) (as in (14)). Then for every ϵ > 0 there is ℙ₂ ∈ 𝒫 with ℙ₂ > 0,
I(ℙ′₀, ℙ₂) < r and c(x, ℙ₀) < c(x, ℙ₂) + ϵ. -/
theorem eq_15 {n d : ℕ} (X : Set (EuclideanSpace ℝ (Fin n))) (hX : IsCompact X)
(γ : ↥X → Fin d → ℝ) (hγ : ∀ i, Continuous (fun x : ↥X => γ x i))
(r : ℝ) (hr : 0 < r) (x : ↥X) (ℙ'₀ ℙ₀ : Δ d) (h14 : relEntropy ℙ'₀ ℙ₀ ≤ (r : EReal))
(hOpt : drPredictor γ r x ℙ'₀ = cost γ x ℙ₀)
(ε : ℝ) (hε : 0 < ε) :
∃ ℙ₂ : Δ d, (∀ i, 0 < (ℙ₂ : Fin d → ℝ) i) ∧ relEntropy ℙ'₀ ℙ₂ < (r : EReal) ∧
cost γ x ℙ₀ < cost γ x ℙ₂ + ε := by sorry
end DROOptimal.Predictor
Source
Van Parys, Mohajerin Esfahani & Kuhn, From Data to Decisions: Distributionally Robust Optimization is Optimal, arXiv:1704.04118v3, p. 18, display (15) in the proof of Theorem 4 (with (14), p. 17)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.