Sub-gamma Chernoff bound gives a Bennett–Bernstein tail
Provedsubgamma_chernoff_tailconcentrationprobability
Sub-gamma Chernoff / Bennett–Bernstein tail: if a centered variable has, for every , a sub-gamma MGF with variance proxy , then for , . The reusable concentration brick turning a sub-gamma MGF bound (entropy-method / modified-LSI output) into the closed-form Bennett–Bernstein deviation tail.
Preamble
import Mathlib.Probability.Moments.Basic import Mathlib.Tactic open Real MeasureTheory ProbabilityTheory open scoped ENNReal
Formal statement
theorem subgamma_chernoff_tail
{Ω : Type*} {mΩ : MeasurableSpace Ω} (μ : MeasureTheory.Measure Ω) [MeasureTheory.IsProbabilityMeasure μ]
(Y : Ω → ℝ) (v x : ℝ) (hv : 0 < v) (hx : 0 < x)
(hint : ∀ t : ℝ, MeasureTheory.Integrable (fun ω => Real.exp (t * Y ω)) μ)
(hsg : ∀ t : ℝ, 0 ≤ t → t < 3 → ProbabilityTheory.mgf Y μ t ≤ Real.exp ((Real.exp t - 1 - t) * v)) :
μ.real {ω | x ≤ Y ω} ≤ Real.exp (- x ^ 2 / (2 * (v + x / 3))) := by sorrySource
Boucheron–Lugosi–Massart, Concentration Inequalities (OUP 2013) §2.4 Thm 2.10; Klein–Rio 2005 Ann.Probab.33 Thm 1.1(c) (arXiv:math/0506594); Bousquet 2002 C.R.Acad.Sci.334.