Proposition 2, (11), p. 15 — ĉ_r(x,ℙ′) = min_{α≥γ̄(x)} α − e^{−r} Π_i (α−γ(x,i))^{ℙ′(i)}, with a bracketed minimizer
ProvedDROOptimal.Predictor.proposition_2distributionally-robust-optimizationdualityp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1relative-entropy
Let be compact, , continuous in for each , , and let be the distributionally robust predictor (10). If and is the worst-case cost, then for every and ,
and this minimum is attained at some with
The dual representation reduces the evaluation of to a one-dimensional convex problem and is the basis of the continuity of (Proposition 3).
Formalization Note "min" is stated as: is a value of the objective at some and is below all such values. Powers are real powers with , so outcomes with contribute a factor . is written as , which is the maximum because is finite and nonempty whenever is.
Preamble
import Mathlib import Definitions.Def_DROOptimal_Predictor_Setting
Formal statement
namespace DROOptimal.Predictor
/-- Proposition 2 (Dual representation of ĉ_r), (11), p. 15: if r > 0 and
γ̄(x) = max_{i ∈ Ξ} γ(x, i), then
ĉ_r(x, ℙ′) = min_{α ≥ γ̄(x)} α − e^{−r} Π_{i ∈ Ξ} (α − γ(x, i))^{ℙ′(i)},
and problem (11) has a minimizer α⋆ with γ̄(x) ≤ α⋆ ≤ (γ̄(x) − e^{−r} c(x, ℙ′)) / (1 − e^{−r}).
Powers are `Real.rpow` (0^0 = 1). -/
theorem proposition_2 {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' : Δ d) :
IsLeast {v : ℝ | ∃ α : ℝ, (⨆ i, γ x i) ≤ α ∧
v = α - Real.exp (-r) * ∏ i, (α - γ x i) ^ ((ℙ' : Fin d → ℝ) i)}
(drPredictor γ r x ℙ') ∧
∃ αstar : ℝ, (⨆ i, γ x i) ≤ αstar ∧
αstar ≤ ((⨆ i, γ x i) - Real.exp (-r) * cost γ x ℙ') / (1 - Real.exp (-r)) ∧
drPredictor γ r x ℙ' =
αstar - Real.exp (-r) * ∏ i, (αstar - γ x i) ^ ((ℙ' : Fin d → ℝ) i) := by sorry
end DROOptimal.Predictor
Source
Van Parys, Mohajerin Esfahani & Kuhn, From Data to Decisions: Distributionally Robust Optimization is Optimal, arXiv:1704.04118v3, p. 15, Proposition 2, (11); proof in Appendix A, p. 28
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.