Theorem 3.6, p. 722 — Ent_π(μH_t) ≤ Ent_π(μ)e^{−2αt}, and ≤ Ent_π(μ)e^{−4αt} when (K, π) is reversible
ProvedLogSobolevMC.Entropy.theorem_3_6Let be a Markov kernel on a finite set with invariant probability , for all , and let be its log-Sobolev constant,
where and . Let and, for a probability measure on , let . Write for the relative entropy. Then for every probability measure :
- if is reversible ( for all ), then
The log-Sobolev constant is thus an exponential rate of convergence to stationarity in relative entropy, for every initial distribution. Combined with Pinsker's inequality it gives total variation bounds such as (3.3).
Formalization Note The chain is not assumed irreducible: the statement holds for every stochastic with a positive invariant probability (for a reducible chain and the claim is that entropy does not increase). is a real infimum, which is on a one-point space; no positivity of is assumed.
import Mathlib import Definitions.Def_mm_basic import Definitions.Def_mm_continuous import Definitions.Def_LogSobolevMC_Entropy_Setting
namespace LogSobolevMC.Entropy
theorem theorem_3_6 {V : Type*} [Fintype V] [DecidableEq V]
(K : Matrix V V ℝ) (hK : MarkovMixing.IsStochastic K)
(π : V → ℝ) (hπ : MarkovMixing.IsStationary K π) (hπpos : ∀ x, 0 < π x)
(μ : V → ℝ) (hμ : MarkovMixing.IsDist μ) :
(∀ t : ℝ, 0 < t →
relEnt π (measHeat K t μ) ≤ relEnt π μ * Real.exp (-2 * LogSobolevMC.ChiSquare.logSobolev K π * t)) ∧
(MarkovMixing.DetailedBalance K π → ∀ t : ℝ, 0 < t →
relEnt π (measHeat K t μ) ≤
relEnt π μ * Real.exp (-4 * LogSobolevMC.ChiSquare.logSobolev K π * t)) := by sorry
end LogSobolevMC.Entropy
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.