Time-homogeneity: the shift commutes with one-step extension of a trajectory
ProvedMarkovChainCLT.partialTraj_map_shift_succLet be a Markov kernel on and write for the shift on initial segments, the map given by , which drops the first coordinate of a partial trajectory. Then for every and every ,
In words: extending a trajectory by one step and then dropping its first coordinate is the same as dropping the first coordinate and then extending. The shift commutes with one-step extension.
Why it is the crux. This is the exact formal content of time-homogeneity, and it is where the two structural facts about the constant one-step family are actually consumed:
- Markov property. The one-step extension of uses the kernel , while the one-step extension of uses . These are the same kernel because — the shift preserves the last coordinate, which is the only thing the transition law depends on.
- Time-homogeneity. The kernel used at time and the kernel used at time are both . For a general, possibly time-inhomogeneous family the statement is simply false: one would be comparing with .
Once the two transition kernels are identified, what remains is a purely combinatorial identity between the two gluing maps: appending a new coordinate at position and then shifting, versus shifting and then appending at position . Both produce the function on equal to for and to the new coordinate at .
Where it leads. Iterating this identity along the partialTraj recursion partialTraj(0, n{+}1) = partialTraj(n, n{+}1) \circ partialTraj(0, n) propagates the shift through all finite-dimensional marginals of the trajectory law. Combined with the fact that a path measure is determined by those marginals, it yields the global statement that shifting a trajectory by one time step is the same as taking one step of and then running the chain — the identity underlying stationarity of a chain started from an invariant measure, the Markov property on path space, and every mixing-coefficient estimate that rests on it.
Proof. Rewrite both sides with the explicit pushforward description of a one-step extension, so each becomes a pushforward of a single measure along an explicit gluing map. Composing the pushforwards reduces the claim to the pointwise identity between the two maps, which is settled by splitting on whether the index satisfies : in the first case both sides read off , in the second both return the newly drawn coordinate (using that for a constant family the singleton-block identification is the constant map).
import Definitions.Def_MarkovChainPathMeasure open Filter Finset Function MeasurableEquiv MeasurableSpace MeasureTheory Preorder ProbabilityTheory open scoped ENNReal NNReal Topology
theorem MarkovChainCLT.partialTraj_map_shift_succ {S : Type*} [MeasurableSpace S]
(P : Kernel S S) [IsMarkovKernel P] (n : ℕ) (u : Π _j : Finset.Iic (n + 1), S) :
(Kernel.partialTraj (X := fun _ : ℕ => S) (BanditAlgorithm.markovChainStep P)
(n + 1) (n + 1 + 1) u).map
(fun v i => v ⟨i.1 + 1, Finset.mem_Iic.2 (Nat.succ_le_succ (Finset.mem_Iic.mp i.2))⟩)
= Kernel.partialTraj (X := fun _ : ℕ => S) (BanditAlgorithm.markovChainStep P) n (n + 1)
(fun i => u ⟨i.1 + 1, Finset.mem_Iic.2 (Nat.succ_le_succ (Finset.mem_Iic.mp i.2))⟩) := by sorry