Lemma 13 — divergence-ball weights stay near uniform
ProvedGenEmpLik.Coverage.lemma_13empirical-processf-divergencep2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1
Let satisfy Assumption A and fix . There are constants such that every feasible weight vector at every positive sample size satisfies
and at some positive sample size a feasible vector satisfies . Thus the supremum over all sample sizes and feasible weights lies between these two bounds.
Formalization Note The proof on p. 35 chooses truncation points using , so the constants may depend on both and . Its factor of is absorbed into the existential constants.
Preamble
import Mathlib import Definitions.Def_GenEmpLik_Coverage_AssumptionA import Definitions.Def_GenEmpLik_Coverage_divergenceBall
Formal statement
namespace GenEmpLik.Coverage
/-- Lemma 13, p. 34, with constants allowed to depend on f and ρ, as in its proof. -/
theorem lemma_13 (f : ℝ → EReal) (hA : AssumptionA f)
(ρ : ℝ) (hρ : 0 < ρ) :
∃ cf Cf : ℝ, 0 < cf ∧ cf ≤ Cf ∧
(∀ n : ℕ, 0 < n → ∀ p ∈ divergenceBall f ρ n,
Real.sqrt (∑ i : Fin n, ((n : ℝ) * p i - 1) ^ 2) ≤ Real.sqrt (ρ * Cf)) ∧
(∃ n : ℕ, 0 < n ∧ ∃ p ∈ divergenceBall f ρ n,
Real.sqrt (ρ * cf) ≤ Real.sqrt (∑ i : Fin n, ((n : ℝ) * p i - 1) ^ 2)) := by sorry
end GenEmpLik.Coverage
Source
Duchi, Glynn & Namkoong, Statistics of Robust Optimization: A Generalized Empirical Likelihood Approach, arXiv:1610.03425v3, p. 34, Lemma 13; proof p. 35
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.