Lemma 6 (case 1: π΅ bounded) β minima and minimizers of Δ_N(Β·|x) converge to v*(x) and π΅*(x)
ProvedPrescAnalytics.KNN.lemma_6_bounded_caseFix a covariate value and one sample path of data: weights that are nonnegative and sum to for all large , and observations . Let be closed, bounded and nonempty. Suppose is -integrable for every , is equicontinuous in (Assumption 4), , and for every . Then , for all large , and every sequence (for all large ) satisfies
This is the deterministic optimization step of the paper's argument: once the sample objective converges to the true conditional objective at , the optimal values and the optimal decisions converge as well.
Formalization Note The paper states Lemma 6 under Assumption 5 in both cases, with the extra hypothesis weakly. As printed it is false in case 2 ( unbounded): with , , , for , and , every hypothesis holds but while . The Lean therefore states case 1 only; the weak-convergence hypothesis is used only in case 2 and is dropped. is read as for each , and as at a minimizer; the attainment of both minima is part of the conclusion. Integrability at , , and the normalization of the weights are read as in Lemma 5. Metric.infDist is the distance to , which is nonempty by the first conclusion.
import Mathlib import Definitions.Def_PrescAnalytics_KNN_Basic open MeasureTheory ProbabilityTheory Filter Topology
namespace PrescAnalytics.KNN
/-- **Lemma 6, case 1 of Assumption 5 (`π΅` bounded)** (arXiv:1402.5481v4, pp. 45β47). Fix `x` and one
sample path: weights `w N i` that are eventually nonnegative and sum to 1, observations `ys i β π΄`.
Let `π΅` be closed, bounded and nonempty, `c(z; Β·)` integrable under `ΞΌ_{Y|x}` for every `z β π΅`,
`c` equicontinuous in `z` on `π΄` with `ΞΌ_{Y|x}(π΄) = 1`, and `Δ_N(z|x) β C(z|x)` for every `z β π΅`.
Then `π΅*(x) β β
`, the argmin of `Δ_N(Β·|x)` over `π΅` is eventually nonempty, and every sequence
`z_N` of eventual minimizers of `Δ_N(Β·|x)` over `π΅` satisfies, for every `zβ β π΅*(x)`,
`min_{z β π΅} Δ_N(z|x) = Δ_N(z_N|x) β v*(x) = C(zβ|x)` and `C(z_N|x) β v*(x)`, and
`inf_{z β π΅*(x)} βz - z_Nβ β 0`. The printed hypothesis `ΞΌΜ_{Y|x,N} β ΞΌ_{Y|x}` weakly is used only in
case 2 and is dropped. -/
theorem lemma_6_bounded_case
{dx dy dz : β}
(ΞΌ : Measure (EuclideanSpace β (Fin dx) Γ EuclideanSpace β (Fin dy))) [IsProbabilityMeasure ΞΌ]
(c : EuclideanSpace β (Fin dz) β EuclideanSpace β (Fin dy) β β)
(Z : Set (EuclideanSpace β (Fin dz))) (Ys : Set (EuclideanSpace β (Fin dy)))
(hZ_closed : IsClosed Z) (hZ_bdd : Bornology.IsBounded Z) (hZ_ne : Z.Nonempty)
(x : EuclideanSpace β (Fin dx))
(w : β β β β β) (ys : β β EuclideanSpace β (Fin dy))
(hys : β i, ys i β Ys) (hYx : βα΅ y β(ΞΌ.condKernel x), y β Ys)
(hw : βαΆ N in atTop, (β i < N, 0 β€ w N i) β§ β i β Finset.range N, w N i = 1)
(h3x : β z β Z, Integrable (c z) (ΞΌ.condKernel x))
(h4 : Assumption4 c Z Ys)
(hconv : β z β Z, Tendsto (fun N => weightedCost c w ys N z) atTop (π (condCost ΞΌ c z x))) :
(optSet ΞΌ c Z x).Nonempty β§
(βαΆ N in atTop, (weightedArgmin c w ys Z N).Nonempty) β§
β zN : β β EuclideanSpace β (Fin dz), (βαΆ N in atTop, zN N β weightedArgmin c w ys Z N) β
(β zstar β optSet ΞΌ c Z x,
Tendsto (fun N => weightedCost c w ys N (zN N)) atTop (π (condCost ΞΌ c zstar x)) β§
Tendsto (fun N => condCost ΞΌ c (zN N) x) atTop (π (condCost ΞΌ c zstar x))) β§
Tendsto (fun N => Metric.infDist (zN N) (optSet ΞΌ c Z x)) atTop (π 0) := by sorry
end PrescAnalytics.KNN
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.