Theorem 3.1 (Generalized Duality), p. 5 — the Kernel DRO primal (2) equals the dual (4)
ProvedKernelDRO.Duality.theorem_3_1Let be a compact metric space with its Borel -algebra and the set of Borel probability measures on . Let be a real Hilbert space with a continuous feature map , so that acts by , and let be closed and convex with nonempty ambiguity set . Let be any set of decisions and a loss such that every is upper semicontinuous and proper (not identically ). Then
- for every ,
where is the support function of ; 2. consequently the Kernel DRO problem (2) and its dual (4) have the same value:
The worst-case expected loss over a set of distributions described through their kernel mean embeddings is thus computed by a single minimization over an RKHS function that majorizes the loss; no Lipschitz constant or RKHS norm of is needed.
Formalization Note is the published extended-real expectation erealExpectation; all values are extended reals. The paper's Slater clause names no topology on measures and is used in the proof only to guarantee a primal solution (p. 15); it is read as , which it implies, so the statement is at least as strong as the paper's. Continuity of (equivalently of the kernel) is not in Assumption 3.1 but is used by Appendix D. Every printed "min" is read as an infimum: dual attainment is not proved and can fail. A compact is an instance of the compact metric space.
import Mathlib import Definitions.Def_WassersteinDRO_Duality_erealExpectation import Definitions.Def_KernelDRO_Duality_Setting open MeasureTheory
namespace KernelDRO.Duality
open WassersteinDRO.Duality
/-- **Theorem 3.1** (Generalized Duality) (Zhu, Jitkrittum, Diehl & Schölkopf, Kernel Distributionally
Robust Optimization, arXiv:2006.06981v3, Theorem 3.1, p. 5; proved on p. 16 from Proposition A.1 by
(15)). Under Assumption 3.1, the Kernel DRO problem
`(P) := min_θ sup_{P,µ} {∫ l(θ, ξ) dP(ξ) : ∫ φ dP = µ, P ∈ 𝒫, µ ∈ 𝒞}` (2) is equivalent to
`(D) := min_{θ, f₀∈ℝ, f∈ℋ} f₀ + δ*_𝒞(f)` subject to `l(θ, ξ) ≤ f₀ + f(ξ), ∀ ξ ∈ 𝒳` (4), where
`δ*_𝒞(f) := sup_{µ∈𝒞} ⟨f, µ⟩_ℋ`; i.e. `(P) = (D)`, and strong duality holds for the inner moment
problem for every `θ` pointwise.
**Formalization Note.** The decision variable `θ` ranges over an arbitrary type `Θ` ("The theorem
holds regardless of the dependency of l on θ", p. 5) and the loss is `l : Θ → 𝒳 → [−∞, ∞)`. Standing
hypotheses (Assumption 3.1, p. 3, and §2, p. 2): `𝒳` is a compact metric space with its Borel
σ-algebra (a compact `𝒳 ⊂ ℝᵈ` is an instance); the RKHS is in feature-map form, a real Hilbert space
`ℋ` with a continuous feature map `φ` (continuity of the kernel is not written in Assumption 3.1 but
is used by §D.2 through a universal, hence continuous, kernel); `𝒞 ⊆ ℋ` is closed and convex; every
`l(θ, ·)` is upper semicontinuous, never `+∞`, and proper (not identically `−∞`, p. 2). The Slater
clause "ri(𝒦_𝒞) ≠ ∅" names no topology on measures and is used in the proof only as "the primal
problem has a non-empty solution set" (p. 15); it is read as `𝒦_𝒞 ≠ ∅`, which it implies under any
reading, so this statement is at least as strong as the paper's. Every "min" (the outer `min_θ` of
(2) and (4) and the inner `min_{f₀, f}`) is read as an infimum: dual attainment is not proved in the
paper and can fail. `∫ l dP` is the published `erealExpectation` and all values are extended
reals. The first conjunct is the pointwise strong duality, the second the literal `(P) = (D)`. -/
theorem theorem_3_1 {X : Type*} [MetricSpace X] [CompactSpace X] [MeasurableSpace X]
[BorelSpace X] {H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
{Θ : Type*} (φ : X → H) (hφ : Continuous φ) (C : Set H) (hC_closed : IsClosed C)
(hC_convex : Convex ℝ C) (l : Θ → X → EReal) (hl_usc : ∀ θ, UpperSemicontinuous (l θ))
(hl_top : ∀ θ ξ, l θ ξ ≠ ⊤) (hl_proper : ∀ θ, ∃ ξ, l θ ξ ≠ ⊥)
(hK : (ambiguitySet φ C).Nonempty) :
(∀ θ, primalValue φ C (l θ) = dualValue φ C (l θ)) ∧
(⨅ θ, primalValue φ C (l θ)) = ⨅ θ, dualValue φ C (l θ) := by sorry
end KernelDRO.Duality
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.