Conditional relative entropy as an average, under a.e. absolute continuity
ProvedInformationTheory.klDiv_compProd_self_eq_lintegral_of_aeThe conditional relative entropy of two composition-products equals the average of the fibrewise divergences, assuming absolute continuity of the fibres only almost everywhere.
Let be a finite measure on and finite kernels from to with for -almost every . Then
Weakening the hypothesis from "for every " to "for -almost every " is what makes the identity usable in the situations it is meant for. A typical example: a learner whose action distribution never selects a particular arm on some event. Off that event the corresponding fibres need not be absolutely continuous at all — the fibrewise divergence may be genuinely infinite — yet the identity still holds, because those fibres are invisible to . Requiring absolute continuity everywhere would exclude exactly the cases where the conclusion is most useful.
Formalization Note Both sides are unchanged when the kernels are modified on a -null set, since the composition-product depends on the kernel only up to -almost-everywhere equality; this is what allows the everywhere-hypothesis in the companion statement to be relaxed.
import Mathlib.InformationTheory.KullbackLeibler.ChainRule import Mathlib.Probability.Kernel.RadonNikodym open MeasureTheory ProbabilityTheory InformationTheory open scoped ENNReal
theorem InformationTheory.klDiv_compProd_self_eq_lintegral_of_ae {α β : Type*}
{mα : MeasurableSpace α} {mβ : MeasurableSpace β}
[MeasurableSpace.CountableOrCountablyGenerated α β]
(μ : Measure α) [IsFiniteMeasure μ] (κ η : Kernel α β)
[IsFiniteKernel κ] [IsFiniteKernel η] (hac : ∀ᵐ a ∂μ, κ a ≪ η a) :
klDiv (μ ⊗ₘ κ) (μ ⊗ₘ η) = ∫⁻ a, klDiv (κ a) (η a) ∂μ := by
sorry