Proposition 3, p. 609 — the bounds of Proposition 2 for the squared weights
ProvedStoneRegression.Criterion.proposition_3nonparametric-regressionp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1weight-functions
Let be i.i.d. -valued with law and let be a sequence of Borel weights (of any sign) satisfying conditions (1)–(3) of Theorem 1. Suppose that there are sequences and of nonnegative constants such that
Let be a nonnegative Borel function on with . Then
and
It is used in the case of Theorem 1, where the second moment of the noise part equals with .
Formalization Note. As in Proposition 2, all expectations and limits inferior and superior are in . The paper does not assume nonnegative weights here, and neither does the statement. Weights are assumed jointly Borel (standing assumption).
Preamble
import Mathlib import Definitions.Def_StoneRegression_Criterion_Setting open MeasureTheory ProbabilityTheory Filter Topology open scoped ENNReal NNReal BoundedContinuousFunction
Formal statement
namespace StoneRegression.Criterion
/-- Proposition 3 (Stone 1977, p. 609). Weights (of any sign) satisfying (1)–(3), constants `Mₙ, Nₙ ≥ 0`
with `P(Mₙ ≤ ∑ᵢ W²_{ni}(X) ≤ Nₙ) → 1`, and a nonnegative Borel `f` with `Ef(X) < ∞`. Then
`liminf E ∑ᵢ W²_{ni}(X) f(Xᵢ) ≥ (liminf Mₙ) Ef(X)` and `limsup E ∑ᵢ W²_{ni}(X) f(Xᵢ) ≤ (limsup Nₙ) Ef(X)`,
all in `ℝ≥0∞`. -/
theorem proposition_3 {d : ℕ} (μ : Measure (EuclideanSpace ℝ (Fin d))) [IsProbabilityMeasure μ]
(W : WeightSeq d) (hW : MeasurableWeights W)
(h1 : ∃ C : ℝ≥0, 1 ≤ C ∧ Cond1 μ W C) (h2 : ∃ D : ℝ, 1 ≤ D ∧ Cond2 μ W D) (h3 : Cond3 μ W)
(M N : ℕ → ℝ≥0)
(hMN : Tendsto (fun n => xLaw μ {ω | (M n : ℝ) ≤ ∑ i : Fin n, wAt W n ω i ^ 2 ∧
∑ i : Fin n, wAt W n ω i ^ 2 ≤ (N n : ℝ)}) atTop (𝓝 1))
(f : EuclideanSpace ℝ (Fin d) → ℝ≥0) (hf : Measurable f) (hfi : ∫⁻ x, f x ∂μ < ∞) :
liminf (fun n => (M n : ℝ≥0∞)) atTop * ∫⁻ x, f x ∂μ ≤
liminf (fun n => ∫⁻ ω, ∑ i : Fin n,
ENNReal.ofReal (wAt W n ω i ^ 2) * (f (ω (i.val + 1)) : ℝ≥0∞) ∂(xLaw μ)) atTop ∧
limsup (fun n => ∫⁻ ω, ∑ i : Fin n,
ENNReal.ofReal (wAt W n ω i ^ 2) * (f (ω (i.val + 1)) : ℝ≥0∞) ∂(xLaw μ)) atTop ≤
limsup (fun n => (N n : ℝ≥0∞)) atTop * ∫⁻ x, f x ∂μ := by sorry
end StoneRegression.Criterion
Source
Stone (1977), Ann. Statist. 5, Proposition 3, p. 609
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.