Proposition 14, p. 616 — E|L(p|X)|ʳ and E|U(p|X)|ʳ are at most E|Y|ʳ/(p ∧ (1 − p))
ProvedStoneRegression.Quantile.proposition_14conditional-quantilesmoment-boundsnonparametric-regressionp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let and , let have law on and let be real with conditional law given . Then the lower and upper th conditional quantiles satisfy
where .
The bound shows that the conditional quantiles inherit the th moment of . It is the integrability input for the part of Theorem 3.
Formalization Note. The page states the proposition for ; the proof of Theorem 3 applies it for every , and the page's argument does not use . The Lean statement takes , the range the paper uses. Expectations are -valued lower integrals, so the inequality is meaningful (and true) also when .
Preamble
import Mathlib import Definitions.Def_StoneRegression_Criterion_Setting import Definitions.Def_StoneRegression_Quantile_Quantiles
Formal statement
namespace StoneRegression.Quantile
open MeasureTheory ProbabilityTheory Filter Topology
open scoped ENNReal NNReal
/-- Proposition 14, p. 616 (range of `r` corrected from the printed `r > 1` to `r ≥ 1`, the range the
proof of Theorem 3 uses): `E|L(p|X)|ʳ ≤ E|Y|ʳ / (p ∧ (1 − p))` and the same for `U(p|X)`, with all
expectations as `ℝ≥0∞`-valued lower integrals (both sides may be `∞`). -/
theorem proposition_14 {d : ℕ} (μ : Measure (EuclideanSpace ℝ (Fin d))) [IsProbabilityMeasure μ]
(κ : Kernel (EuclideanSpace ℝ (Fin d)) ℝ) [IsMarkovKernel κ]
(p : ℝ) (hp0 : 0 < p) (hp1 : p < 1) (r : ℝ) (hr : 1 ≤ r) :
∫⁻ x, ‖lowerQ κ p x‖ₑ ^ r ∂μ ≤
(∫⁻ q, ‖q.2‖ₑ ^ r ∂(μ ⊗ₘ κ)) / ENNReal.ofReal (min p (1 - p)) ∧
∫⁻ x, ‖upperQ κ p x‖ₑ ^ r ∂μ ≤
(∫⁻ q, ‖q.2‖ₑ ^ r ∂(μ ⊗ₘ κ)) / ENNReal.ofReal (min p (1 - p)) := by sorry
end StoneRegression.Quantile
Source
Stone (1977), Ann. Statist. 5, Proposition 14, p. 616 (§12); printed hypothesis r > 1 widened to r ≥ 1 as used in the proof of Theorem 3, p. 617
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.