Theorem 3.5, p. 8 — finite sample guarantee: P^N{E^P[h(x̂_N, ξ)] > Ĵ_N} ≤ β for the radius ε_N(β)
ProvedWassDDRO.Consistency.theorem_3_5Let be a finite-dimensional real normed space with its Borel -algebra, , and let be a probability distribution on supported on (that is, ). Suppose Assumption 3.3 holds: for some ,
and let be constants for which the concentration inequality (7) holds. Fix and , a feasible set and a loss with measurable for every and -integrable for every . Let be an optimizer of the distributionally robust program (5) with ambiguity set : for every sample , and minimizes over . Let be the optimal value of (5). Then the finite sample guarantee (2) holds:
The certificate is thus an upper confidence bound, at level , on the out-of-sample cost of the data-driven decision.
Formalization Note The probability is stated for the failure event, as an outer measure, because the success event need not be measurable ( is an arbitrary selection of optimizers). The constants enter as any constants satisfying (7) (Theorem 3.4 supplies them); is the scope of (7) and (8). The optimizer is required only for samples in , as in §2 (). The loss is real-valued and -integrability of on is assumed, a specialization of the paper's extended-valued loss: the worst-case expectation counts only distributions under which is integrable, so without this hypothesis itself would not be counted.
import Mathlib import Definitions.Def_WassDDRO_Consistency_Setting open MeasureTheory
namespace WassDDRO.Consistency
/-- Theorem 3.5 (Finite sample guarantee), p. 8. Under Assumption 3.3, with constants
`c₁, c₂` for which (7) holds (`m ≠ 2`), `β ∈ (0, 1)`, `N ≥ 1`, and `x̂_N` an optimizer of (5)
with ambiguity set `B_{ε_N(β)}(P̂_N)`, the guarantee (2) holds in failure-set form:
`P^N{E^P[h(x̂_N, ξ)] > Ĵ_N} ≤ β`. -/
theorem theorem_3_5 {E : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [FiniteDimensional ℝ E]
[MeasurableSpace E] [BorelSpace E]
(P : Measure E) [IsProbabilityMeasure P] (Ξ : Set E) (hP : P Ξᶜ = 0)
(a : ℝ) (ha : 1 < a) (hA : Integrable (fun ξ => Real.exp (‖ξ‖ ^ a)) P)
(hm : Module.finrank ℝ E ≠ 2)
(c₁ c₂ : ℝ) (hc₁ : 0 < c₁) (hc₂ : 0 < c₂) (h7 : Concentration7 P a c₁ c₂)
(N : ℕ) (hN : 1 ≤ N) (b : ℝ) (hb : 0 < b ∧ b < 1)
{n : ℕ} (X : Set (EuclideanSpace ℝ (Fin n))) (h : EuclideanSpace ℝ (Fin n) → E → ℝ)
(hmeas : ∀ x, Measurable (h x)) (hint : ∀ x ∈ X, Integrable (h x) P)
(xh : (Fin N → E) → EuclideanSpace ℝ (Fin n))
(hopt : ∀ ξhat : Fin N → E, (∀ i, ξhat i ∈ Ξ) →
xh ξhat ∈ X ∧ ∀ x ∈ X,
WassersteinDRO.Duality.worstCaseRisk (radius c₁ c₂ a (Module.finrank ℝ E) N b) 1 Ξ
(WassersteinDRO.Duality.empiricalDistribution ξhat) (h (xh ξhat)) ≤
WassersteinDRO.Duality.worstCaseRisk (radius c₁ c₂ a (Module.finrank ℝ E) N b) 1 Ξ
(WassersteinDRO.Duality.empiricalDistribution ξhat) (h x)) :
(Measure.pi fun _ : Fin N => P)
{ξhat | ¬ (((∫ ξ, h (xh ξhat) ξ ∂P : ℝ) : EReal) ≤
droValue X h (radius c₁ c₂ a (Module.finrank ℝ E) N b) Ξ ξhat)} ≤
ENNReal.ofReal b := by sorry
end WassDDRO.Consistency
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.