Entropy dissipation along the Ornstein–Uhlenbeck semigroup:
ProvedGaussianMatrix.ou_entropy_hasDerivAtentropyfisher-informationgaussianornstein-uhlenbecksemigroup
Let be the standard Gaussian measure on . For and a function , let
be the Ornstein–Uhlenbeck (Mehler) semigroup. Let and , and let satisfy and . Then for every the function is differentiable at , and
where is the derivative of .
This is the de Bruijn-type identity behind the Bakry–Émery proof of the Gaussian log-Sobolev inequality: the entropy decreases with rate equal to the Fisher information. The proof has three ingredients:
- differentiation under the integral sign;
- the generator identity , valid for even though only is assumed, because Gaussian integration by parts in the noise variable gives ;
- Gaussian integration by parts in .
Formalization Note. is written out explicitly. The restriction is needed, since is not differentiable at .
Preamble
import Definitions.Def_GaussianMatrix_basic open MeasureTheory ProbabilityTheory open scoped Matrix
Formal statement
namespace GaussianMatrix
theorem ou_entropy_hasDerivAt (f : ℝ → ℝ) (hf : ContDiff ℝ 1 f) (δ C : ℝ) (hδ : 0 < δ)
(hlow : ∀ x, δ ≤ f x) (hup : ∀ x, f x ≤ C) (hdf : ∀ x, |deriv f x| ≤ C) (t : ℝ)
(ht : 0 < t) :
HasDerivAt
(fun s => ∫ x,
(∫ y, f (Real.exp (-s) * x + Real.sqrt (1 - Real.exp (-(2 * s))) * y) ∂(gaussianReal 0 1))
* Real.log (∫ y, f (Real.exp (-s) * x + Real.sqrt (1 - Real.exp (-(2 * s))) * y)
∂(gaussianReal 0 1)) ∂(gaussianReal 0 1))
(-∫ x,
deriv (fun z => ∫ y, f (Real.exp (-t) * z + Real.sqrt (1 - Real.exp (-(2 * t))) * y)
∂(gaussianReal 0 1)) x ^ 2
/ ∫ y, f (Real.exp (-t) * x + Real.sqrt (1 - Real.exp (-(2 * t))) * y) ∂(gaussianReal 0 1)
∂(gaussianReal 0 1)) t := by
sorry
end GaussianMatrix
Source
D. Bakry, I. Gentil, M. Ledoux, Analysis and Geometry of Markov Diffusion Operators (Springer, 2014), §2.7.1 (Ornstein–Uhlenbeck semigroup, Mehler formula) and §5.7 (entropy dissipation ); M. Ledoux, The Concentration of Measure Phenomenon (AMS, 2001), §5.1 (proof of Theorem 5.1). Section numbers are cited from memory.