Theorem 4, p. 17 — for r > 0 the relative-entropy DRO predictor ĉ_r is strongly optimal in (5)
ProvedDROOptimal.Predictor.theorem_4Let be compact, , and continuous in for each ; write for the expected cost under a model in the simplex , for the relative entropy, and for the empirical distribution of independent samples. For let
Then is strongly optimal in the vector optimization problem (5):
- is continuous on , and for all and ,
- every continuous with the same property satisfies for all and .
In words: among all data-driven predictors whose out-of-sample disappointment decays at exponential rate at least under every model, the relative-entropy distributionally robust predictor is the least conservative, uniformly in the decision and the observed frequencies.
Formalization Note The rate condition is the logarithm-free RateLE (equivalent to the displayed with ); the competitors range over all jointly continuous functions on and nothing else; feasibility of , continuity included, is part of the conclusion.
import Mathlib import Definitions.Def_DROOptimal_Predictor_Setting import Definitions.Def_DROOptimal_Predictor_Problem5
namespace DROOptimal.Predictor
/-- Theorem 4 (Optimality of ĉ_r), p. 17: under the standing assumptions of §2 (p. 5), if r > 0
then ĉ_r is strongly optimal in (5): it is a continuous predictor whose disappointment decays at rate
at least r under every model and decision, and ĉ_r(x, ℙ′) ≤ ĉ(x, ℙ′) for all x ∈ X, ℙ′ ∈ 𝒫 and every
predictor ĉ feasible in (5). -/
theorem theorem_4 {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) :
StronglyOptimal5 γ r (drPredictor γ r) := by sorry
end DROOptimal.Predictor
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.