InformationTheory.Shannon.DeLaValleePoussin
The de la Vallée-Poussin criterion for uniform integrability #
A general-purpose, measure-agnostic de la Vallée-Poussin criterion: if there is a
superlinear control function G : ℝ≥0∞ → ℝ≥0∞ (meaning G t / t → ∞ as t → ∞)
whose composition with the norm has a uniform finite integral bound
∀ i, ∫⁻ x, G ‖f i x‖ₑ ∂μ ≤ C < ∞, then the family f is UnifIntegrable at
exponent 1.
This is the classical "forward" direction of the de la Vallée-Poussin theorem.
Mathlib does not have this lemma (loogle UnifIntegrable, ConvexOn = Found 0), but
the proof is short and reduces to MeasureTheory.unifIntegrable_of — crucially that
gateway does not require [IsFiniteMeasure μ], so the criterion applies to
infinite measures such as volume.
The intended downstream consumer is the EPI G2 Vitali witness chain for the approximate-identity
L¹ convergence of a density under vanishing-variance Gaussian convolution, where μ = volume and
f n = negMulLog (pX ∗ g_{u n}).
@audit:ok
InformationTheory.Shannon.Superlinear
sourceSuperlinear growth control, the genuine de la Vallée-Poussin hypothesis on
G : ℝ≥0∞ → ℝ≥0∞: for every slope K there is a finite threshold M beyond which
K * t ≤ G t. This is the non-degenerate ℝ≥0∞ reading of "G t / t → ∞" — note
that Filter.atTop on ℝ≥0∞ collapses to the singleton {∞} (since ∞ is the top),
so the literal Tendsto (G ·/·) atTop atTop would be vacuous; this threshold form is
the usable statement.
Equations
Instances For
Used by
InformationTheory.Shannon.unifIntegrable_of_superlinear_lintegral
sourceThe de la Vallée-Poussin criterion (forward direction).
If G : ℝ≥0∞ → ℝ≥0∞ is superlinear (Superlinear G, i.e. K * t ≤ G t eventually for
every slope K) and the family f satisfies a uniform bound
∀ i, ∫⁻ x, G ‖f i x‖ₑ ∂μ ≤ C with C ≠ ∞, then f is uniformly integrable at
exponent 1.
The measure μ is arbitrary — no [IsFiniteMeasure μ] is needed (the proof goes
through MeasureTheory.unifIntegrable_of, which is finite-measure-free), so this
applies to volume.