Lemma D.1 — duality for conditional value-at-risk
ProvedLatentMixDRO.lemma_D_1covariate-shiftdistributionally-robust-optimizationp2o-gran-per-chapterp2o-plan-paperp2o-v1statistical-learning-theory
Let be a probability space, an integrable random variable and . Then
This identifies the supremum over bounded likelihood ratios with the Rockafellar–Uryasev form of conditional value-at-risk, and is the source of the dual objective (7).
Differences from the source:
- encoding: of two real quantities →
IsLUBandIsGLBof one common value ; reason: Lean convention; the suprema of real sets are otherwise junk-valued on unbounded sets
Preamble
import Mathlib open MeasureTheory
Formal statement
namespace LatentMixDRO
/-- Lemma D.1: duality for conditional value-at-risk. For an integrable `W` on a probability space,
`sup {E[𝓛 W] | 𝓛 : Ω → [0, 1/α₀] measurable, E[𝓛] = 1} = inf_{η ∈ ℝ} {(1/α₀) E[(W - η)₊] + η}`.
Duchi, Hashimoto & Namkoong, Distributionally Robust Losses for Latent Covariate Mixtures,
arXiv:2007.13982v2, Lemma D.1, p.33.
**Formalization Note.** The supremum and the infimum are stated as `IsLUB` and `IsGLB` of one common
value `v`, so no junk value of a real `sSup`/`sInf` enters. -/
theorem lemma_D_1 {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ]
(W : Ω → ℝ) (hW : Integrable W μ) (α₀ : ℝ) (hα₀ : 0 < α₀) (hα₀' : α₀ < 1 / 2) :
∃ v : ℝ,
IsLUB {s | ∃ 𝓛 : Ω → ℝ, Measurable 𝓛 ∧ (∀ a, 0 ≤ 𝓛 a ∧ 𝓛 a ≤ 1 / α₀) ∧
(∫ a, 𝓛 a ∂μ) = 1 ∧ s = ∫ a, 𝓛 a * W a ∂μ} v ∧
IsGLB {s | ∃ η : ℝ, s = (1 / α₀) * (∫ a, max (W a - η) 0 ∂μ) + η} v := by sorry
end LatentMixDRO
Source
Duchi, Hashimoto & Namkoong, Distributionally Robust Losses for Latent Covariate Mixtures, arXiv:2007.13982v2, Lemma D.1, p.33
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.