One step of the trajectory kernel is an explicit pushforward of
ProvedMarkovChainCLT.partialTraj_succ_self_apply_eq_mapLet be a Markov transition kernel on and let be the associated time-homogeneous one-step family used by the Ionescu–Tulcea construction, . For every and every partial trajectory ,
i.e. extending a trajectory by one time step is the pushforward of along the map that appends as the new last coordinate, glueing the block with the singleton block .
Why this is worth isolating. In Mathlib the one-step extension partialTraj κ m (m+1) is defined as
a composite of a product kernel, a kernel pushforward, and a gluing map between dependent function types indexed by Finset.Iic and Finset.Ioc. That form is well suited to the abstract development, but it is unusable for hands-on computation: every concrete manipulation has to unfold the product kernel, discharge the measurability of the gluing map, and fight the dependent indexing.
This lemma performs that unfolding once and delivers the elementary description that any concrete argument actually wants: a single pushforward of along an explicit map. Two things become immediate from it:
- the one-step law depends on only through its last coordinate — this is precisely the Markov property, which is invisible in the definitional form;
- the transition kernel is the same for every — time-homogeneity, which the general
Kernel.trajAPI deliberately does not know, since it is built for possibly time-inhomogeneous families.
Both facts are exactly what is needed to prove that shifting a trajectory by one time step is the same as taking one step of and then running the chain, and hence to establish stationarity of the chain started from an invariant measure, the Markov property on path space, and the mixing-coefficient estimates that depend on it.
Proof. Unfold partialTraj_succ_self, evaluate the product kernel at (the first factor is Kernel.id, so it contributes the Dirac mass ), and use Measure.dirac_prod to rewrite as the pushforward of along . Composing the three pushforwards with Measure.map_map and observing by definition gives the claim definitionally.
import Definitions.Def_MarkovChainPathMeasure open Filter Finset Function MeasurableEquiv MeasurableSpace MeasureTheory Preorder ProbabilityTheory open scoped ENNReal NNReal Topology
theorem MarkovChainCLT.partialTraj_succ_self_apply_eq_map {S : Type*} [MeasurableSpace S]
(P : Kernel S S) [IsMarkovKernel P] (m : ℕ) (v : Π _i : Finset.Iic m, S) :
Kernel.partialTraj (X := fun _ : ℕ => S) (BanditAlgorithm.markovChainStep P) m (m + 1) v
= (P (v ⟨m, Finset.mem_Iic.2 le_rfl⟩)).map
(fun w => IicProdIoc (X := fun _ : ℕ => S) m (m + 1)
(v, MeasurableEquiv.piSingleton (X := fun _ : ℕ => S) m w)) := by sorry