Theorem 5, p. 15 — a geometric Lyapunov function bounds E_π Φ by φ(t₀)K/(1 − γ)
ProvedGJNSteadyState.Interchange.theorem_5Let be the time- transition kernel of a Markov process on and let be a stationary distribution, so that . Let be measurable and a geometric Lyapunov function with geometric drift size and exception parameter , and suppose is finite. Then
The bound holds for every stationary distribution, unique or not.
It is the basic tool of the paper: applied to it yields the exponential tail bound of Theorem 6.
Formalization Note The Markov process enters only through its time- transition kernel and the invariance of a stationary law, which is all the statement uses. The hypothesis is the page's remark "the bound is only meaningful when is finite"; it is needed in arithmetic, where would otherwise turn into a false claim. enters as ; for the geometric drift holds everywhere and both sides are . The expectation is a lower Lebesgue integral and may a priori be infinite.
import Mathlib import Definitions.Def_GJNSteadyState_Interchange_Lyapunov open MeasureTheory ProbabilityTheory
namespace GJNSteadyState.Interchange
/-- Theorem 5 (p. 15): if `π` is a stationary distribution of a Markov process whose time-`t₀`
transition kernel is `κ`, and `Φ` is a geometric Lyapunov function with parameters `γ, t₀, K`,
then `E_π Φ(Ξ(0)) ≤ φ(t₀) K / (1 − γ)`. -/
theorem theorem_5 {X : Type*} [MeasurableSpace X] (κ : Kernel X X) [IsMarkovKernel κ]
(π : Measure X) [IsProbabilityMeasure π] (hπ : κ.Invariant π)
(Φ : X → NNReal) (hΦ : Measurable Φ) (γ K : ℝ)
(hgeo : IsGeomLyapunov (kernelExp κ) Φ γ K)
(hphi : phiSup (kernelExp κ) Φ < ⊤) :
∫⁻ x, (Φ x : ENNReal) ∂π ≤
phiSup (kernelExp κ) Φ * ENNReal.ofReal K / ENNReal.ofReal (1 - γ) := by sorry
end GJNSteadyState.Interchange
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.