Theorem 3.9 (Consistency), pp. 21–22 — measurable optimal estimates converge μ-a.s. to the unique minimizer in D
ProvedDupacovaWets.Consistency.consistency_of_measurable_estimatesThe standing data are those of §3: is a Polish space with its Borel -field, a probability measure on it, a probability space with an increasing sequence of sub--fields , random probability measures on , and an integrand with domain ; , and the measures satisfy Assumptions 3.4 and 3.5. Write and , with the convention of p. 10.
Suppose Assumptions 3.4 and 3.5 hold and there is a compact set such that for every the set is nonempty -almost surely, and
Then there exist -measurable selections of such that
and also -almost surely.
This is the strong consistency statement of the paper: estimators that use only the information available at stage and solve the estimated problem converge almost surely to the true solution, and the estimated optimal values converge to the true optimal value. It extends Wald's and Huber's consistency theorems to constrained problems with extended-real-valued, merely lower semicontinuous criteria.
Formalization Note Lean writes as EuclideanSpace ℝ (Fin n), and the sample index is shifted: Lean's Pν k ζ and 𝔽 k are the paper's and , while is kept as the separate argument P. The paper's selections are maps , measurable for the trace of on a full-measure set ; here they are total maps , -measurable on all of , with for all for -almost every . The two forms are equivalent: a function into that is measurable for a trace -field extends measurably to the whole space, and has full measure. The limit of the infima is in . The paper also says that is the support of ; this hypothesis is not used in the proofs of §3 and is omitted, which makes the statement stronger.
import Mathlib import Definitions.Def_DupacovaWets_Consistency_expect import Definitions.Def_DupacovaWets_Consistency_ExtendedFunctions import Definitions.Def_DupacovaWets_Consistency_Assumptions open MeasureTheory Filter Topology
namespace DupacovaWets.Consistency
/-- Dupačová–Wets (WP-86-41), pp. 21–22, Theorem 3.9 (Consistency), "In particular": under
Assumptions 3.4 and 3.5, if `D ⊆ ℝⁿ` is compact, `(argmin E^ν f) ∩ D` is nonempty `μ`-a.s. for
every `ν`, and `{x*} = argmin Ef ∩ D`, then there are `F^ν`-measurable selections `x^ν` of
`argmin E^ν f` with `x^ν(ζ) → x*` for `μ`-almost all `ζ`, and `inf E^ν f → inf Ef` `μ`-a.s.
(`Pν k`, `𝔽 k`, `x k` are the paper's `P^{k+1}`, `F^{k+1}`, `x^{k+1}`.) -/
theorem consistency_of_measurable_estimates {n : ℕ} {Ξ : Type*} [TopologicalSpace Ξ] [PolishSpace Ξ] [MeasurableSpace Ξ] [BorelSpace Ξ]
(P : Measure Ξ) [IsProbabilityMeasure P]
{Z : Type*} [mZ : MeasurableSpace Z] (μ : Measure Z) [IsProbabilityMeasure μ]
(𝔽 : Filtration ℕ mZ) (Pν : ℕ → Z → Measure Ξ)
(f : EuclideanSpace ℝ (Fin n) → Ξ → EReal) (S : Set (EuclideanSpace ℝ (Fin n)))
(h34 : Assumption3_4 f S) (h35 : Assumption3_5 f S P μ 𝔽 Pν)
(D : Set (EuclideanSpace ℝ (Fin n))) (hD : IsCompact D)
(hDne : ∀ k, ∀ᵐ ζ ∂μ, (argminSet (expectFn (Pν k ζ) f) ∩ D).Nonempty)
(xstar : EuclideanSpace ℝ (Fin n)) (hx : argminSet (expectFn P f) ∩ D = {xstar}) :
(∃ x : ℕ → Z → EuclideanSpace ℝ (Fin n), (∀ k, Measurable[𝔽 k] (x k)) ∧
(∀ᵐ ζ ∂μ, ∀ k, x k ζ ∈ argminSet (expectFn (Pν k ζ) f)) ∧
(∀ᵐ ζ ∂μ, Tendsto (fun k => x k ζ) atTop (𝓝 xstar))) ∧
∀ᵐ ζ ∂μ, Tendsto (fun k => ⨅ y, expectFn (Pν k ζ) f y) atTop (𝓝 (⨅ y, expectFn P f y)) := by sorry
end DupacovaWets.Consistency
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.