Proof of Theorem 4.2, p. 13 — the case ε = 0: (10) is the sample average and (12b) tends to it as λ → ∞
ProvedWassDDRO.Reduction.worstCase_eq_sampleAverage_of_zerodistributionally-robust-optimizationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1sample-average-approximationwasserstein
Let , , and samples be as in Theorem 4.2 (, each measurable), and suppose Assumption 4.1 holds. Then:
- with radius , the worst-case expectation (10) equals the sample average,
- the objective of (12b) at converges to the sample average as :
- consequently (12a) holds with equality also at : (10) with equals the optimal value of (12c) with .
Together with the case this shows that (10) equals (12c) for all .
Formalization Note The sums are taken in EReal. Under Assumption 4.1 no takes the value , so each , and for large enough each inner supremum is ; the sums therefore never meet where it matters, and a sample with makes the average , as in the paper. The limit is in the order topology of .
Preamble
import Mathlib import Definitions.Def_WassDDRO_Reduction_Setting open MeasureTheory Filter Topology
Formal statement
namespace WassDDRO.Reduction
/-- Proof of Theorem 4.2, p. 13 (the case `ε = 0`): under Assumption 4.1, (10) with
`ε = 0` is the sample average `(1/N) Σ ℓ(ξ̂ᵢ)`, the objective of (12b) at `ε = 0` converges
to that sample average as `λ → ∞`, and (12a) holds with equality at `ε = 0`. -/
theorem worstCase_eq_sampleAverage_of_zero {E : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [FiniteDimensional ℝ E]
[MeasurableSpace E] [BorelSpace E]
{K N : ℕ} (hK : 0 < K) (hN : 0 < N) (Ξ : Set E) (ℓ : Fin K → E → EReal)
(hmeas : ∀ k, Measurable (ℓ k)) (ξhat : Fin N → E) (hξ : ∀ i, ξhat i ∈ Ξ)
(hA : Assumption41 Ξ ℓ) :
worstCaseExpectation 0 Ξ ξhat (maxLoss ℓ)
= ((1 / (N : ℝ) : ℝ) : EReal) * (∑ i, maxLoss ℓ (ξhat i)) ∧
Tendsto (fun lam : ℝ => ((1 / (N : ℝ) : ℝ) : EReal) *
∑ i, (⨆ ξ ∈ Ξ, maxLoss ℓ ξ - ((lam * ‖ξ - ξhat i‖ : ℝ) : EReal)))
atTop (𝓝 (((1 / (N : ℝ) : ℝ) : EReal) * (∑ i, maxLoss ℓ (ξhat i)))) ∧
worstCaseExpectation 0 Ξ ξhat (maxLoss ℓ) = program12cValue 0 Ξ ξhat (maxLoss ℓ) := by sorry
end WassDDRO.Reduction
Source
Mohajerin Esfahani & Kuhn, arXiv:1505.05116v3, proof of Theorem 4.2, p. 13 (the case ε = 0)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.