Proof of Lemma 12 (App. D.1.1), p. 47 — R_k(Z) = z₀ + (z₁ − z₀) sup{q ∈ [0,1] : (1−p)^{1−k}(1−q)^k + p^{1−k}q^k ≤ c_k^k}
ProvedUniformDRO.LowerBound.primal_two_pointcressie-readdistributionally-robust-optimizationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let , and . Let , , and let with probability and with probability . Then
The supremum runs over the mass that an alternative distribution places on ; the constraint is the Cressie–Read divergence condition for two-point laws, rewritten with . This primal formula is how the proof of Lemma 12 obtains the lower bound of the third claim.
Formalization Note. The page uses the formula for , where the negative powers and are finite; the statement keeps that range (Lean's Real.rpow at base would give a meaningless value). The set contains and lies in , so the real supremum is genuine.
Preamble
import Mathlib import Definitions.Def_UniformDRO_LowerBound_Setting
Formal statement
namespace UniformDRO.LowerBound
theorem primal_two_point (k ρ : ℝ) (hk : 1 < k) (hρ : 0 < ρ) (z₀ z₁ p : ℝ) (hz : z₀ ≤ z₁)
(hp : p ∈ Set.Ioo (0 : ℝ) 1) :
UniformDRO.Concentration.robustRisk k ρ (twoPointLaw z₀ z₁ p) id =
z₀ + (z₁ - z₀) * sSup {q ∈ Set.Icc (0 : ℝ) 1 |
(1 - p) ^ (1 - k) * (1 - q) ^ k + p ^ (1 - k) * q ^ k ≤ UniformDRO.Concentration.ck k ρ ^ k} := by sorry
end UniformDRO.LowerBound
Source
Duchi & Namkoong, arXiv:1810.08750v6, proof of Lemma 12 (App. D.1.1), p. 47, first display
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.