Herbst's argument: integrating the entropy differential inequality
ProvedHerbst.herbst_differential_inequality_integrateHerbst's argument, calculus core (the entropy-method MGF-from-entropy integration step). Let be continuous on , with a right derivative on (i.e. has the within-derivative on at each ) satisfying . Then .
Role in the entropy method. This is exactly the integration step in Herbst's argument (Boucheron–Lugosi–Massart, Concentration Inequalities, OUP 2013, §5.2, proof of Theorem 5.3). A logarithmic Sobolev inequality applied to yields the differential inequality for the log-MGF ; dividing by identifies the left side as , so with (and by l'Hôpital) the LSI bound integrates to the slope bound , hence the Gaussian-shaped MGF bound . The same template (with a modified LSI giving a non-quadratic ) yields the Bennett/Bousquet inequality for suprema of empirical processes (BLM Ch.12). This isolated, source-faithful real-analysis lemma is the reusable Herbst integration brick; the Chern off MGFtail closure is already available in Mathlib (ProbabilityTheory.measure_ge_le_exp_mul_mgf), and the remaining irreducible step is the (modified) log-Sobolev inequality itself.
import Mathlib.Analysis.Calculus.Deriv.MeanValue import Mathlib.Analysis.Calculus.Deriv.Inv import Mathlib.Analysis.SpecialFunctions.Log.Deriv open Set Filter Topology
theorem Herbst.herbst_differential_inequality_integrate
(g g' : ℝ → ℝ) (m C lam : ℝ)
(hlam : 0 ≤ lam)
(hg_cont : ContinuousOn g (Set.Icc 0 lam))
(hg0 : g 0 = m)
(hg_deriv : ∀ x ∈ Set.Ico (0 : ℝ) lam, HasDerivWithinAt g (g' x) (Set.Ici x) x)
(hbound : ∀ x ∈ Set.Ico (0 : ℝ) lam, g' x ≤ C) :
g lam ≤ m + C * lam := by sorry