Time-homogeneity: shifting the trajectory is one step of
ProvedMarkovChainCLT.markovChainKernel_map_shiftLet be a Markov transition kernel on and let denote the law of the trajectory of the time-homogeneous Markov chain started at . Write for the shift on path space, . Then
i.e. shifting the trajectory by one time step is the same as taking one step of first and then running the chain. Equivalently, as an identity of kernels , the pushforward of the trajectory kernel along equals its composition with .
This is the formal content of time-homogeneity of the chain. It is intuitively obvious and is used silently throughout the Markov chain literature, but it is not available for free in a formalization built on the Ionescu-Tulcea theorem: Kernel.traj is constructed for a general, possibly time-inhomogeneous family , and its API (traj_comp_partialTraj, map_traj_succ_self, traj_map_frestrictLe) never refers to the special structure of the constant family. Nothing in it transports the fact that the mission's one-step kernels satisfy for every . Proving the identity means comparing all finite-dimensional marginals, via an eq_traj-style argument together with the partialTraj recursion.
Why this is worth isolating. It is the single missing ingredient beneath several other targets of this mission. Most directly, shift-invariance of the chain started from an invariant distribution — i.e. strict stationarity of the coordinate process — follows from it in a few lines: pushing the identity through gives
and an induction on the shift amount, using , gives invariance under every shift. It is also the engine of the conditional-expectation argument that bounds the strong mixing coefficients of the chain by its total-variation convergence rate, and of the transfer of a central limit theorem from the stationary start to an arbitrary initial distribution.
import Definitions.Def_MarkovChainPathMeasure open MeasureTheory ProbabilityTheory Filter open scoped ENNReal NNReal Topology ProbabilityTheory
theorem MarkovChainCLT.markovChainKernel_map_shift {X : Type*} [MeasurableSpace X]
(P : Kernel X X) [IsMarkovKernel P] :
(BanditAlgorithm.markovChainKernel P).map (fun ω : ℕ → X => fun n => ω (n + 1))
= (BanditAlgorithm.markovChainKernel P) ∘ₖ P := by sorry