Smooth uniform approximation of Euclidean-Lipschitz functions on preserving the Lipschitz constant
ProvedGaussianMatrix.lipschitz_smooth_approxconvolutionlipschitzmollificationsmooth-approximation
Let be a finite index set, , and let satisfy
Then for every there is a function satisfying the same Lipschitz bound,
and for every .
One may take , the convolution of with a smooth probability density supported in a small ball. This lemma lets one prove functional inequalities for Lipschitz functions from their versions for smooth functions with bounded gradient, without Rademacher's theorem; here it transfers the Gaussian log-Sobolev inequality to the entropy bound for with Lipschitz.
Formalization Note. Smoothness is expressed as ∀ n : ℕ, ContDiff ℝ n g. No sign condition on is imposed: for nonempty the hypothesis forces , and for everything is constant.
Preamble
import Definitions.Def_GaussianMatrix_basic open MeasureTheory ProbabilityTheory open scoped Matrix
Formal statement
namespace GaussianMatrix
theorem lipschitz_smooth_approx {ι : Type*} [Fintype ι] (f : (ι → ℝ) → ℝ) (L : ℝ)
(hLip : ∀ x y, |f x - f y| ≤ L * Real.sqrt (∑ i, (x i - y i) ^ 2)) (ε : ℝ) (hε : 0 < ε) :
∃ g : (ι → ℝ) → ℝ, (∀ n : ℕ, ContDiff ℝ n g) ∧
(∀ x y, |g x - g y| ≤ L * Real.sqrt (∑ i, (x i - y i) ^ 2)) ∧
∀ x, |g x - f x| ≤ ε := by
sorry
end GaussianMatrix
Source
Standard fact (mollification): if is -Lipschitz for the Euclidean distance and is a function with supported in a ball of radius , then is , -Lipschitz, and (up to the comparison between the ball's norm and the Euclidean norm); see e.g. L. C. Evans, R. F. Gariepy, Measure Theory and Fine Properties of Functions (CRC Press, 1992), §4.2.1 (mollifiers), cited from memory.