Composition-product commutes with relabelling the first coordinate
ProvedProbabilityTheory.compProd_map_leftRelabelling the first coordinate commutes with forming a composition-product.
Let be an s-finite measure on , let be measurable, and let be an s-finite kernel from to . Then
where is the push-forward and is the kernel pulled back along .
Both sides describe the same experiment: draw the first coordinate, relabel it by , and then draw the second coordinate from the kernel evaluated at the relabelled point. The only choice is whether the relabelling happens before or after the second coordinate is drawn, and since the kernel is evaluated at the relabelled point either way, the two agree.
The identity is what one needs to transport a composition-product between two indexings of the same underlying space. A typical use: an interaction history of length can be indexed by Fin n or by an initial segment Iic (n-1) of the naturals, and results about the infinite-horizon trajectory measure are naturally stated in the second indexing while the finite-horizon model uses the first.
import Mathlib.Probability.Kernel.Composition.MeasureCompProd open MeasureTheory ProbabilityTheory Set open scoped ENNReal
theorem ProbabilityTheory.compProd_map_left {α α' β : Type*}
{mα : MeasurableSpace α} {mα' : MeasurableSpace α'} {mβ : MeasurableSpace β}
(μ : Measure α) [SFinite μ]
{ι : α → α'} (hι : Measurable ι) (η : Kernel α' β) [IsSFiniteKernel η] :
(μ.map ι) ⊗ₘ η = (μ ⊗ₘ (η.comap ι hι)).map (Prod.map ι id) := by
sorry