Monotonicity and existence of the critical radius (Lemma 13.6), for non-degenerate star-shaped classes
ProvedHighDimStat.NonparametricLS.thm13_6_critical_radius_exists_v2Lemma 13.6 (p. 425). Let be fixed design points, i.i.d. standard Gaussian noise, and let be a star-shaped function class whose local Gaussian complexity (Eq. (13.16)) has a measurable defining supremum for every . Then the map is non-increasing on . Consequently, if is non-degenerate on the design (some has ), then for any constant the inequality
has a smallest positive solution.
This lemma is what makes sense of the phrase "let be any positive solution of the critical inequality" in Theorems 13.5 and 13.13: it guarantees that valid radii exist at all, and identifies the smallest one, , as the natural choice.
Formalization Note. The retired version (thm13_6_critical_radius_exists) admitted
degenerate classes — , , , or any class all of whose
members vanish at the design points — for which , every solves
the inequality and no smallest positive solution exists (accepted disproof with
, ). The printed sentence is literally false for such (it
tacitly assumes a nontrivial class), so the correction is the hypothesis
hH0 : ∃ h ∈ H, empiricalNorm x h ≠ 0, which excludes exactly the classes with
(for Gaussian noise, for small as soon as
some has , by star-shapedness) and in particular forces . The
hypothesis hmeas (integrability of the defining supremum for every ) makes explicit
the field's standing measurability convention; the supremum is automatically bounded by
(Cauchy–Schwarz), so it is a pure measurability condition, without which
Mathlib's Bochner integral would return the junk value at some radii and break the
monotonicity claim. is localGaussianComplexity from the mission's Core;
the existence claim is stated as IsLeast of the solution set, matching "smallest positive
solution" literally.
import Mathlib import Definitions.Def_HighDimStat_NonparametricLS_Core
namespace HighDimStat.NonparametricLS
open MeasureTheory
/-- **Lemma 13.6** (p. 425, PDF 445). For any star-shaped function class `H`, the map
`δ ↦ Gₙ(δ; H)/δ` is non-increasing on `(0, ∞)`. Consequently, for any constant `c > 0`, the
inequality `Gₙ(δ; H)/δ ≤ cδ` has a smallest positive solution — provided `H` is
non-degenerate on the design, i.e. some `h ∈ H` has `‖h‖ₙ ≠ 0`.
Corrections: the retired version admitted degenerate classes (`H = ∅`, `H = {0}`, `n = 0`, or
any `H` all of whose members vanish at the design points), for which `Gₙ ≡ 0`, every `δ > 0`
solves the inequality and no smallest positive solution exists; the printed sentence is false
for such `H` (it tacitly assumes a nontrivial class), so the non-degeneracy hypothesis `hH0`
is added as the correction. The integrability hypothesis `hmeas` makes explicit the field's
standing measurability convention for the supremum defining `Gₙ(δ; H)` (the supremum is
automatically bounded by `‖w‖ₙ δ`, so this is purely a measurability condition); without it
Mathlib's Bochner integral returns the junk value `0` for a non-measurable integrand. -/
theorem thm13_6_critical_radius_exists_v2 {X Ω : Type*} [MeasurableSpace Ω] {n : ℕ}
(x : Fin n → X) (w : Fin n → Ω → ℝ) (P : Measure Ω) [IsProbabilityMeasure P]
(hw : IsIIDStdGaussian P w)
(H : Set (X → ℝ)) (hH : IsStarShaped H)
(hH0 : ∃ h ∈ H, empiricalNorm x h ≠ 0)
(hmeas : ∀ δ : ℝ, 0 < δ →
Integrable (fun ω => ⨆ h : {h : X → ℝ // h ∈ H ∧ empiricalNorm x h ≤ δ},
|(∑ i, w i ω * h.1 (x i)) / n|) P) :
(∀ δ t : ℝ, 0 < δ → δ ≤ t →
localGaussianComplexity x w P H t / t ≤ localGaussianComplexity x w P H δ / δ) ∧
∀ c : ℝ, 0 < c →
∃ δ, IsLeast {δ : ℝ | 0 < δ ∧ localGaussianComplexity x w P H δ / δ ≤ c * δ} δ := by sorry
end HighDimStat.NonparametricLS