Discarding the first steps restarts the chain from
ProvedMarkovChainCLT.markovChainKernel_map_shift_iterLet be a Markov kernel on , let be the law of the trajectory started at , and let be the -fold shift on path space, . Then for every ,
i.e. as an identity of kernels, the pushforward of the trajectory kernel along the -shift is its composition with the -step transition kernel .
What it says. Discarding the first observations of a Markov chain started at leaves a Markov chain started from the -step distribution . This is the "restart" or regeneration principle in its simplest, unconditional form, and it is the workhorse behind every quantitative mixing estimate: bounding the dependence between the past and the future always proceeds by re-expressing the future as a chain restarted from and then invoking a rate of convergence .
Proof. Induction on , with the one-step identity as the base mechanism. The key bookkeeping is the factorisation , which holds definitionally in this indexing (unlike the equally true , since addition on recurses on its second argument). Pushing forward along first turns the trajectory kernel into its composition with ; the induction hypothesis then handles , and reassociating leaves . A short separate induction shows , which converts this into as defined by the recursion .
That commutation step is not a triviality to be skipped: with defined as , the composite produced by the shift argument is , and identifying the two is exactly the statement that the -step kernel can be built by taking the extra step either at the beginning or at the end.
import Definitions.Def_MarkovChainPathMeasure import Definitions.Def_MarkovIterKernel open MeasureTheory ProbabilityTheory Filter open scoped ENNReal NNReal Topology ProbabilityTheory open MarkovChainCLT
theorem MarkovChainCLT.markovChainKernel_map_shift_iter {X : Type*} [MeasurableSpace X]
(P : Kernel X X) [IsMarkovKernel P] (k : ℕ) :
(BanditAlgorithm.markovChainKernel P).map (fun ω : ℕ → X => fun n => ω (n + k))
= (BanditAlgorithm.markovChainKernel P) ∘ₖ (iterKernel P k) := by sorry