The coordinate process of a chain started from an invariant measure is strictly stationary
ProvedMarkovChainCLT.isStrictlyStationary_chainMeasureLet be a Markov kernel with invariant probability measure . Then the coordinate process on path space, under the law of the chain started from , is strictly stationary: for every , the shifted sequence has the same law, as a random element of , as .
Why this is the form actually needed. Strict stationarity — equality of the laws of the whole shifted sequence, not just equality of finite-dimensional marginals one at a time — is the standing hypothesis of the classical central limit theorems for dependent sequences: Ibragimov and Linnik's theorem for strongly mixing sequences, Ibragimov's for -mixing, Billingsley's for -mixing, and Doukhan–Massart–Rio. Each of these is stated for a strictly stationary sequence, and the way they are applied to a Markov chain is: start the chain from , observe that the coordinate process is then strictly stationary, and feed it in. This theorem is that observation.
It is also what makes the mixing coefficients of the chain well defined as functions of the lag alone: the definitions of , and take a supremum over the split point , and it is stationarity that makes the quantity being supremised independent of .
Proof. By induction on from the one-step shift-invariance . For there is nothing to prove. For the step, the -shift factors as the -shift composed with , so pushing forward along it is pushing forward along first — which leaves unchanged — and then along the -shift, which is the induction hypothesis.
import Definitions.Def_MarkovChainPathMeasure import Definitions.Def_MixingCoefficients import Mathlib.Probability.Kernel.Invariance open MeasureTheory ProbabilityTheory Filter open scoped ENNReal NNReal Topology ProbabilityTheory open MarkovChainCLT
theorem MarkovChainCLT.isStrictlyStationary_chainMeasure {X : Type*} [MeasurableSpace X]
(P : Kernel X X) [IsMarkovKernel P] (π : Measure X) [IsProbabilityMeasure π]
(hinv : Kernel.Invariant P π) :
IsStrictlyStationary (chainMeasure P π) (fun i (ω : ℕ → X) => ω i) := by sorry