Display after (47) — E_{P̂n}[|L − 1|^p]^{1/p} ≤ n^{−1/p} √(ρ C_f)
ProvedGenEmpLik.Consistency.display_after_47empirical-likelihoodf-divergencep2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1
Let , (the exponent of (47)), , and let be a constant as in Lemma 13: for every in the ball . Then for every in that ball, with likelihood ratio ,
where . Since , the right side tends to as , which makes the reweighting term of (47) vanish.
Formalization Note The page writes with undefined ("where is as in the lemma"); Lemma 13's bound is , so . Lemma 13 enters as the hypothesis on ; the paper's exponent is called .
Preamble
import Mathlib import Definitions.Def_PhiDivRobust_Counterpart_uncertaintySet
Formal statement
namespace GenEmpLik.Consistency
/-- The display after (47) (arXiv:1610.03425v3, App. E.1, proof of Theorem 7, p. 46). Let
`n ≥ 1`, `r ≥ 2` (the paper's `p = max{2, 1 + 1/ε}`), and let `C_f` be a constant as in
Lemma 13: `‖n w − 𝟙‖₂ ≤ √(ρ C_f)` for every `w` in the ball `{w ≥ 0 : ∑ wᵢ = 1,
∑ᵢ f(n wᵢ) ≤ ρ}`. Then for every `w` in the ball, with `L(ξᵢ) = n wᵢ`,
`E_{P̂n}[|L − 1|^r]^{1/r} = n^{−1/r} ‖n w − 𝟙‖_r ≤ n^{−1/r} ‖n w − 𝟙‖₂ ≤ n^{−1/r} √(ρ C_f)`.
(The page writes `√(ρ/γ_f)` with `γ_f` undefined; it is `√(ρ C_f)`.) -/
theorem display_after_47 (f : ℝ → EReal) (ρ : ℝ) (n : ℕ) (hn : 0 < n) (r : ℝ) (hr : 2 ≤ r)
(Cf : ℝ)
(hCf : ∀ w ∈ PhiDivRobust.Counterpart.probUncertaintySet f
(fun _ : Fin n => (1 : ℝ) / n) (ρ / n),
Real.sqrt (∑ i, ((n : ℝ) * w i - 1) ^ 2) ≤ Real.sqrt (ρ * Cf))
(w : Fin n → ℝ) (hw : w ∈ PhiDivRobust.Counterpart.probUncertaintySet f
(fun _ : Fin n => (1 : ℝ) / n) (ρ / n)) :
((1 / (n : ℝ)) * ∑ i, |(n : ℝ) * w i - 1| ^ r) ^ (1 / r) =
(n : ℝ) ^ (-(1 / r)) * (∑ i, |(n : ℝ) * w i - 1| ^ r) ^ (1 / r) ∧
(n : ℝ) ^ (-(1 / r)) * (∑ i, |(n : ℝ) * w i - 1| ^ r) ^ (1 / r) ≤
(n : ℝ) ^ (-(1 / r)) * Real.sqrt (∑ i, ((n : ℝ) * w i - 1) ^ 2) ∧
(n : ℝ) ^ (-(1 / r)) * Real.sqrt (∑ i, ((n : ℝ) * w i - 1) ^ 2) ≤
(n : ℝ) ^ (-(1 / r)) * Real.sqrt (ρ * Cf) := by sorry
end GenEmpLik.Consistency
Source
Duchi, Glynn & Namkoong, Statistics of Robust Optimization: A Generalized Empirical Likelihood Approach, arXiv:1610.03425v3, p. 46, App. E.1, proof of Theorem 7, display after (47)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.