Lemma 3.7, p. 9 — any data-dependent Q̂_N ∈ B_{ε_N(β_N)}(P̂_N) converges to P in d_W, P^∞-almost surely
ProvedWassDDRO.Consistency.lemma_3_7Let be a finite-dimensional real normed space with its Borel -algebra, , and a probability distribution on supported on . Suppose Assumption 3.3 holds with exponent , and let be constants for which the concentration inequality (7) holds. Let , , satisfy
with the radius (8). Let , , be any distributions, possibly depending on the training data, with , where is the empirical distribution of the first samples of the i.i.d. sequence . Then
Every distribution in the shrinking Wasserstein balls thus converges to the true distribution; this is the step in the proof of Theorem 3.6 that turns worst-case expectations into true expectations in the limit.
Formalization Note The selection is any function of the whole sample sequence that lies in the ball whenever the samples lie in (as they do -almost surely); in the paper it may depend on the training data. "Almost surely" is stated as a -almost-everywhere statement. are any constants satisfying (7), which Theorem 3.4 provides for . The paper's parenthetical "and thus weakly" is a remark citing the literature and is not part of the statement.
import Mathlib import Definitions.Def_WassDDRO_Consistency_Setting open MeasureTheory Filter Topology
namespace WassDDRO.Consistency
/-- Lemma 3.7 (Convergence of distributions), p. 9. Under Assumption 3.3, with constants
`c₁, c₂` for which (7) holds (`m ≠ 2`), and `β_N ∈ (0, 1)` summable with `ε_N(β_N) → 0`,
any (data-dependent) selection `Q̂_N ∈ B_{ε_N(β_N)}(P̂_N)` satisfies
`P^∞{lim_N d_W(P, Q̂_N) = 0} = 1`. -/
theorem lemma_3_7 {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))
(Qhat : ℕ → (ℕ → E) → Measure E)
(hQ : ∀ N (ω : ℕ → E), 1 ≤ N → (∀ i, ω i ∈ Ξ) →
Qhat N ω ∈ WassersteinDRO.Duality.ambiguitySet
(radius c₁ c₂ a (Module.finrank ℝ E) N (β N)) 1 Ξ
(WassersteinDRO.Duality.empiricalDistribution (fun i : Fin N => ω i))) :
∀ᵐ ω ∂(P_inf P),
Tendsto (fun N => WassersteinDRO.Duality.wassersteinDistance 1 P (Qhat N ω)) atTop (𝓝 0) := by sorry
end WassDDRO.Consistency
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.