Squared event form of Pinsker’s inequality
ProvedBanditAlgorithm.pinsker_squared_event_differencebandit-algorithmsinformation-theoryprobability
For probability measures and with finite relative entropy and every measurable event , the squared difference of the event probabilities satisfies . This is the eventwise squared form of Pinsker’s inequality; the sign convention is immaterial because the difference is squared.
Preamble
import Mathlib.InformationTheory.KullbackLeibler.Basic open MeasureTheory InformationTheory Real open scoped ENNReal
Formal statement
theorem BanditAlgorithm.pinsker_squared_event_difference {Ω : Type} {mΩ : MeasurableSpace Ω} (P Q : Measure Ω) [IsProbabilityMeasure P] [IsProbabilityMeasure Q] {A : Set Ω} (hA : MeasurableSet A) (hD : klDiv P Q ≠ ∞) : 2 * (Q.real A - P.real A) ^ 2 ≤ (klDiv P Q).toReal := by sorrySource
T. Lattimore and C. Szepesvári, Bandit Algorithms (Cambridge University Press, 2020), Note 5, Eq. (14.12), printed p. 193 / PDF p. 202.