p. 6 (and proof of Prop. A.1, p. 14) — weak duality
ProvedKernelDRO.Duality.weak_dualitydistributionally-robust-optimizationdualityp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let be a compact metric space with its Borel -algebra, continuous, any set, and upper semicontinuous. If (that is, ) and satisfies for all , then
Taking the supremum over and the infimum over gives weak duality for (11) and (12); the right-hand side is a computable bound on the worst-case risk.
Formalization Note is erealExpectation; is extended-real valued.
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
/-- **Weak duality** of (11) and (12) (Zhu, Jitkrittum, Diehl & Schölkopf, Kernel Distributionally
Robust Optimization, arXiv:2006.06981v3, §3.2, p. 6, "By the weak duality (P) ≤ (D) of (11) and
(12), we have ∫ l dP ≤ f₀ + δ*_𝒞(f)"; derived from the Lagrangian (13) in the proof of Proposition
A.1, p. 14). For every `P` in the ambiguity set `𝒦_𝒞` and every dual feasible `(f₀, f)`,
`∫ l dP ≤ f₀ + δ*_𝒞(f)`. Taking the supremum over `P` and the infimum over `(f₀, f)` gives
`(P) ≤ (D)`.
**Formalization Note.** `∫ l dP` is `erealExpectation` and `δ*_𝒞` is the `EReal` support function.
Standing hypotheses: `𝒳` compact metric with its Borel σ-algebra, `φ` continuous, `l` upper
semicontinuous with values in `[−∞, ∞)`. No hypothesis on `𝒞` is needed. -/
theorem weak_duality {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) (C : Set H) (l : X → EReal) (hl_usc : UpperSemicontinuous l)
(hl_top : ∀ ξ, l ξ ≠ ⊤) (P : ProbabilityMeasure X) (hP : P ∈ ambiguitySet φ C) (f₀ : ℝ)
(f : H) (hfeas : IsDualFeasible φ l f₀ f) :
erealExpectation (P : Measure X) l ≤ (f₀ : EReal) + supportFun C f := by sorry
end KernelDRO.Duality
Source
Zhu, Jitkrittum, Diehl & Schölkopf, Kernel Distributionally Robust Optimization, arXiv:2006.06981v3, §3.2, p. 6; Lagrangian (13) in the proof of Proposition A.1, p. 14
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.