Theorem 3.6, p. 8 — asymptotic consistency: P^∞-a.s. Ĵ_N → J⋆, and accumulation points of x̂_N are optimal for (1)
ProvedWassDDRO.Consistency.theorem_3_6Let be a finite-dimensional real normed space with its Borel -algebra, , and a probability distribution on supported on . Suppose Assumption 3.3 holds: for some , ; let be constants for which the concentration inequality (7) holds, and the radius (8). Let , , satisfy
Let , and with measurable for every . Let be i.i.d. with law (joint law ), and for each let be an optimizer, and the optimal value, of the distributionally robust program (5) with ambiguity set built from the first samples. Let be the optimal value of (1).
- If is upper semicontinuous in for every , and there is with
then -almost surely for all sufficiently large and as . 2. If the assumptions of (1) hold, is closed, and is lower semicontinuous in for every , then -almost surely every accumulation point of is an optimal solution of (1): and for all .
With radii shrinking at a rate dictated by the summable confidence levels (for example ), the Wasserstein distributionally robust certificates and decisions are therefore consistent for the true stochastic program.
Formalization Note The paper writes ""; its proof establishes for all large and , which is what is stated (convergence from above, not monotonicity). The constants are any constants satisfying (7), which Theorem 3.4 supplies under Assumption 3.3 for ; (7) and (8) are printed only for . The loss is real-valued (the paper allows extended values, but the growth bound makes finite on ). The optimizers are required to exist only for samples in , which holds -almost surely. The worst-case expectation counts distributions under which is integrable; under the growth bound every distribution in the ball integrates , and Assumption 3.3 makes the Bochner integrals genuine. "Accumulation point" is a cluster point of the sequence along ; values of are irrelevant.
import Mathlib import Definitions.Def_WassDDRO_Consistency_Setting open MeasureTheory Filter Topology
namespace WassDDRO.Consistency
/-- Theorem 3.6 (Asymptotic consistency), p. 8. Under Assumption 3.3, with constants `c₁, c₂`
for which (7) holds (`m ≠ 2`), `β_N ∈ (0, 1)` summable with `ε_N(β_N) → 0`, and `x̂_N` an
optimizer of (5) with ambiguity set `B_{ε_N(β_N)}(P̂_N)`:
(i) if `h(x, ·)` is upper semicontinuous on `Ξ` and `|h(x, ξ)| ≤ L(1 + ‖ξ‖)` on `X × Ξ`, then
`P^∞`-a.s. `Ĵ_N ≥ J⋆` eventually and `Ĵ_N → J⋆`;
(ii) if moreover `X` is closed and `h(·, ξ)` is lower semicontinuous on `X` for every `ξ ∈ Ξ`,
then `P^∞`-a.s. every accumulation point of `(x̂_N)` is an optimal solution of (1). -/
theorem theorem_3_6 {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)
(β : ℕ → ℝ) (hβ : ∀ N, 0 < β N ∧ β N < 1) (hsum : Summable β)
(hlim : Tendsto (fun N => radius c₁ c₂ a (Module.finrank ℝ E) N (β N)) atTop (𝓝 0))
{n : ℕ} (X : Set (EuclideanSpace ℝ (Fin n))) (h : EuclideanSpace ℝ (Fin n) → E → ℝ)
(hmeas : ∀ x, Measurable (h x))
(xhat : (N : ℕ) → (Fin N → E) → EuclideanSpace ℝ (Fin n))
(hopt : ∀ (N : ℕ) (ξhat : Fin N → E), 1 ≤ N → (∀ i, ξhat i ∈ Ξ) →
xhat N ξhat ∈ X ∧ ∀ x ∈ X,
WassersteinDRO.Duality.worstCaseRisk (radius c₁ c₂ a (Module.finrank ℝ E) N (β N)) 1 Ξ
(WassersteinDRO.Duality.empiricalDistribution ξhat) (h (xhat N ξhat)) ≤
WassersteinDRO.Duality.worstCaseRisk (radius c₁ c₂ a (Module.finrank ℝ E) N (β N)) 1 Ξ
(WassersteinDRO.Duality.empiricalDistribution ξhat) (h x)) :
((∀ x ∈ X, UpperSemicontinuousOn (h x) Ξ) → ∀ L : ℝ, 0 ≤ L →
(∀ x ∈ X, ∀ ξ ∈ Ξ, |h x ξ| ≤ L * (1 + ‖ξ‖)) →
∀ᵐ ω ∂(P_inf P),
(∀ᶠ N in atTop, trueValue X h P ≤
droValue X h (radius c₁ c₂ a (Module.finrank ℝ E) N (β N)) Ξ (fun i : Fin N => ω i)) ∧
Tendsto (fun N => droValue X h (radius c₁ c₂ a (Module.finrank ℝ E) N (β N)) Ξ
(fun i : Fin N => ω i)) atTop (𝓝 (trueValue X h P))) ∧
((∀ x ∈ X, UpperSemicontinuousOn (h x) Ξ) → ∀ L : ℝ, 0 ≤ L →
(∀ x ∈ X, ∀ ξ ∈ Ξ, |h x ξ| ≤ L * (1 + ‖ξ‖)) →
IsClosed X → (∀ ξ ∈ Ξ, LowerSemicontinuousOn (fun x => h x ξ) X) →
∀ᵐ ω ∂(P_inf P), ∀ xstar : EuclideanSpace ℝ (Fin n),
MapClusterPt xstar atTop (fun N => xhat N (fun i : Fin N => ω i)) →
xstar ∈ X ∧ ∀ x ∈ X, ∫ ξ, h xstar ξ ∂P ≤ ∫ ξ, h x ξ ∂P) := by sorry
end WassDDRO.Consistency
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.