(A.3), p. 747 — a minimizer u of (A.2) solves 2u log u − 2u log‖u‖₂ − (1/α)(I − K)u = 0
ProvedLogSobolevMC.TwoPoint.eq_A_3Let be a Markov kernel on a finite set , reversible with respect to its invariant probability (that is, ), and suppose its log-Sobolev constant is positive. Let be a positive function with that attains the infimum (A.2), . Then at every point of
This Euler–Lagrange equation is what makes a minimizer for take at most two values, reducing Theorem A.1 to the two-point Theorem A.2.
Formalization Note The page says "any minimizer u of (A.2)" in a context where minimizers are nonnegative. Three hypotheses are added and disclosed: reversibility (for a nonreversible kernel the variational equation involves the symmetric part rather than ; the page applies (A.3) only to the reversible chain ); positivity of (the logarithm at a zero of would take Lean's default value); and (so that is meaningful).
import Mathlib import Definitions.Def_mm_basic import Definitions.Def_mm_lower import Definitions.Def_LogSobolevMC_TwoPoint_Setting
namespace LogSobolevMC.TwoPoint
open MarkovMixing
/-- (A.3), p. 747: a positive minimizer `u` of (A.2) for a reversible chain satisfies
`2u log u − 2u log ‖u‖₂ − (1/α)(I − K)u = 0`. -/
theorem eq_A_3 {V : Type*} [Fintype V] [DecidableEq V]
(K : Matrix V V ℝ) (hK : IsStochastic K) (π : V → ℝ) (hπ : IsStationary K π)
(hπpos : ∀ x, 0 < π x) (hrev : DetailedBalance K π)
(u : V → ℝ) (hu : ∀ x, 0 < u x) (hne : LogSobolevMC.ChiSquare.entL π u ≠ 0)
(hmin : LogSobolevMC.ChiSquare.dirichlet K π u u = LogSobolevMC.ChiSquare.logSobolev K π * LogSobolevMC.ChiSquare.entL π u)
(hα : 0 < LogSobolevMC.ChiSquare.logSobolev K π) :
∀ x, 2 * u x * Real.log (u x) - 2 * u x * Real.log (LogSobolevMC.ChiSquare.lpNorm π 2 u)
- (1 / LogSobolevMC.ChiSquare.logSobolev K π) * (u x - K.mulVec u x) = 0 := by sorry
end LogSobolevMC.TwoPoint
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.