Invariant measures of kernels #
This file adds to mathlib's Kernel.Invariant the facts about invariant measures that a Markov
chain on a finite state space needs. An invariant measure is carried along a map that intertwines
two kernels, and it is invariant for every power of its kernel. One step of a Markov kernel leaves
the mass of a set unchanged exactly when the flow out of the set balances the flow into it.
On a finite state space with measurable singletons, every Markov kernel has an invariant probability measure. When the kernel is irreducible with respect to the counting measure, every state reaching every state, it has exactly one, and that measure charges every state; conversely a Markov kernel whose invariant probability measures all charge every state is irreducible.
Main statements #
ProbabilityTheory.Kernel.Invariant.map: ifη (f a) = (κ a).map ffor alla, an invariant measure ofκmaps alongfto an invariant measure ofη.ProbabilityTheory.Kernel.comp_apply_eq_iff_setLIntegral_compl:(κ ∘ₘ μ) s = μ sexactly when∫⁻ a in s, κ a sᶜ ∂μ = ∫⁻ a in sᶜ, κ a s ∂μ.ProbabilityTheory.Kernel.isIrreducible_count_iff: irreducibility with respect to the counting measure is reachability of every point from every point.ProbabilityTheory.Kernel.exists_isProbabilityMeasure_invariant: on a finite type, a Markov kernel has an invariant probability measure.ProbabilityTheory.Kernel.Invariant.eq_of_isIrreducible,ProbabilityTheory.Kernel.Invariant.measure_singleton_ne_zero: an irreducible kernel on a finite type has at most one invariant probability measure, and a nonzero invariant measure charges every point.ProbabilityTheory.Kernel.isIrreducible_count_of_measure_singleton_ne_zero: a Markov kernel on a finite type whose invariant probability measures all charge every point is irreducible.
Implementation notes #
The existence proof is linear algebra rather than compactness: the transition matrix P fixes the
constant vector, so P - 1 is singular and has a nonzero left null vector, whose absolute value
is a nonnegative fixed vector. Uniqueness is the minimum-ratio argument: subtracting from one
invariant law the largest multiple of another that stays nonnegative leaves an invariant measure
that vanishes at a point, hence everywhere. Neither needs the Perron–Frobenius theorem.
[UPSTREAM] candidate for Mathlib/Probability/Kernel/Invariance.lean.
An invariant measure of κ maps along f to an invariant measure of η when η moves the
image of a point as κ moves the point.
One step of a Markov kernel leaves the mass of a set unchanged exactly when the flow out of the set balances the flow into it.
The flow of an invariant measure out of a set balances the flow into it.
A path to b in n steps followed by a step from b to c reaches c in n + 1 steps.
On a space with measurable singletons, κ is irreducible with respect to the counting measure
when every point is reached from every point.
A nonzero invariant measure of an irreducible kernel charges every point.
On a finite type, invariance is the system of equations ∑ a, μ {a} κ a {b} = μ {b}.
A Markov kernel on a finite type has an invariant probability measure.
An irreducible Markov kernel on a finite type has at most one invariant probability measure.
A Markov kernel on a finite type is irreducible when each of its invariant probability measures charges every point.