Lemma D.1, pp. 22–23 — the conic constraint is equivalent to on
ProvedKernelDRO.Duality.lemma_D_1dualitymoment-problemp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let be a compact metric space with its Borel -algebra, a continuous feature map into a real Hilbert space, and upper semicontinuous. For and , the following are equivalent:
- for every Borel probability measure on (the conic constraint , with the dual cone of the probability simplex);
- for all , the semi-infinite constraint (26).
This turns the dual cone constraint of the Lagrangian derivation into the semi-infinite constraint of the dual problem (12).
Formalization Note The dual cone of the conic hull is the same as that of , since for ; the statement quantifies over probability measures. The integral of the extended-real function is erealExpectation.
Preamble
import Mathlib import Definitions.Def_WassersteinDRO_Duality_erealExpectation import Definitions.Def_KernelDRO_Duality_Setting open MeasureTheory
Formal statement
namespace KernelDRO.Duality
open WassersteinDRO.Duality
/-- **Lemma D.1** (Zhu, Jitkrittum, Diehl & Schölkopf, Kernel Distributionally Robust Optimization,
arXiv:2006.06981v3, Lemma D.1, pp. 22–23). The conic constraint `l − f − f₀ ∈ −𝒦*`, where `𝒦*` is
the dual cone of the probability simplex `𝒫`, i.e. `∫ (l − f₀ − f) dP ≤ 0` for every `P`, is
equivalent to the semi-infinite constraint (26): `l(ξ) ≤ f₀ + f(ξ)` for all `ξ ∈ 𝒳`.
**Formalization Note.** The dual cone is that of the conic hull `co(𝒫)` (p. 13: `𝒦* := {h : ∫ h dm ≥ 0,
∀ m ∈ 𝒦}`); since `∫ h d(cP) = c ∫ h dP` for `c ≥ 0`, quantifying over probability measures is the
same constraint, which is the normalization the proof of "⇐" makes. The integral `∫ (l − f₀ − f) dP`
of the extended-real function is `erealExpectation`; `θ` is suppressed (p. 13). Standing
hypotheses: `𝒳` compact metric with its Borel σ-algebra, `φ` continuous, `l` upper
semicontinuous with values in `[−∞, ∞)`. -/
theorem lemma_D_1 {X : Type*} [MetricSpace X] [CompactSpace X] [MeasurableSpace X]
[BorelSpace X] {H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
(φ : X → H) (hφ : Continuous φ) (l : X → EReal) (hl_usc : UpperSemicontinuous l)
(hl_top : ∀ ξ, l ξ ≠ ⊤) (f₀ : ℝ) (f : H) :
(∀ P : ProbabilityMeasure X,
erealExpectation (P : Measure X)
(fun ξ => l ξ - ((f₀ + rkhsEval φ f ξ : ℝ) : EReal)) ≤ 0) ↔
IsDualFeasible φ l f₀ f := by sorry
end KernelDRO.Duality
Source
Zhu, Jitkrittum, Diehl & Schölkopf, Kernel Distributionally Robust Optimization, arXiv:2006.06981v3, Lemma D.1, (26), pp. 22–23; dual cone defined in App. A, Notation, p. 13
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.