A fixed boundary term scaled by vanishes in probability
ProvedMarkovChainCLT.tendstoInMeasure_inv_sqrt_coord_subLet be a Harris ergodic chain run from and let . Then
Every coordinate of the stationary chain has law , so and have the same fixed law for all ; in particular is a constant. Markov's inequality then gives, for ,
No independence, mixing or ergodicity beyond stationarity of the marginals is used — only that the law of does not drift with , so the boundary term is bounded in probability while the normalisation grows.
This is the step that discards the telescoping remainder in the martingale-approximation proof of the Markov chain CLT. After solving the Poisson equation one has ; dividing by , this lemma applied to shows the difference between the normalised sum and the normalised martingale tends to in probability, so Slutsky's theorem transfers the martingale central limit theorem to the sample average.
import Definitions.Def_MarkovErgodicity import Definitions.Def_MarkovChainPathMeasure open MeasureTheory ProbabilityTheory Filter open scoped ENNReal NNReal Topology ProbabilityTheory
theorem MarkovChainCLT.tendstoInMeasure_inv_sqrt_coord_sub {X : Type*} [MeasurableSpace X]
(P : Kernel X X) [IsMarkovKernel P] (π : Measure X) [IsProbabilityMeasure π]
(hP : HarrisErgodic P π) (h : X → ℝ) (hhm : Measurable h) (hh : Integrable h π) :
TendstoInMeasure (chainMeasure P π)
(fun (n : ℕ) (ω : ℕ → X) => (Real.sqrt n)⁻¹ * (h (ω 0) - h (ω n))) atTop 0 := by sorry