Samplinglib
Lean gate passed 2026-08-19T06:04:36.257124+00:00 · 77184245109a
12.1 · Book p. 283 · PDF p. 295

Introduction

Introduces forward noising and reverse-time SDEs, identifying the marginal score in the reverse drift.

Open this section in the canonical August 9 source ↗
Section 12.1

The reverse-time drift uses the score of the forward marginal, and approximation errors become drift errors.

A forward diffusion transports data toward a tractable reference law. Under suitable marginal regularity, the reverse SDE contains the time-dependent score \(\nabla\log\pi_t\); replacing that score and discretizing time create distinct error terms.

Why is this valid?

Choose a density representative, establish spatial and temporal regularity, state the time-reversal theorem, and only then analyze learned-score and discretization errors.

Source assumptions

  • regular time marginals
  • available score approximation

Formal assumptions

  • score representative
  • time-reversal regularity
  • law under which score error is controlled
View Lean formalization
partial · faithful paraphrase

ASTIS exposes Fisher-information roots but does not yet claim a reverse-time SDE theorem.

AutoSamplingTheory.lsiKlFiSqrtDensityFisherChainIntegralHandoffScalar

compiled Samplinglib leaf · AutoSamplingTheory/Probability.lean:861
theorem lsiKlFiSqrtDensityFisherChainIntegralHandoffScalar
    {α ι : Type*} [MeasurableSpace α] [Fintype ι] (mu : Measure α)
    {r : α → Real} {dirichlet fisher : Real} {dr dSqrt dLog : ι → α → Real}
    (hdirichlet : dirichlet = ∫ x, (∑ i, dSqrt i x ^ 2) ∂mu)
    (hfisher : fisher = ∫ x, (r x * ∑ i, dLog i x ^ 2) ∂mu)
    (hr : ∀ᵐ x ∂mu, 0 < r x)
    (hdSqrt : ∀ᵐ x ∂mu,
      ∀ i, dSqrt i x = (1 / (2 * Real.sqrt (r x))) * dr i x)
    (hdLog : ∀ᵐ x ∂mu, ∀ i, dLog i x = dr i x / r x) :
    dirichlet = (1 / 4) * fisher := by
  rw [hdirichlet, hfisher]
  rw [← integral_const_mul]
  exact lsiKlFiSqrtDensityFisherChainIntegralFiniteSum mu hr hdSqrt hdLog

/-- Scalar rearrangement behind the one-sided use of the cited DV formula.

This is not a proof of Donsker--Varadhan.  It starts after a cited or
eventually formalized entropy-duality theorem has supplied the variational
upper bound for an admissible test.
-/
Imports
  • Mathlib.Data.Real.Basic
  • Mathlib.Algebra.Order.Archimedean.Real.Basic
  • Mathlib.Analysis.Real.Sqrt
  • Mathlib.Analysis.Calculus.ParametricIntegral
  • Mathlib.Analysis.SpecialFunctions.Log.Basic
  • Mathlib.InformationTheory.KullbackLeibler.Basic
  • Mathlib.MeasureTheory.Measure.LogLikelihoodRatio
  • Mathlib.MeasureTheory.Measure.Tilted
  • Mathlib.Probability.Kernel.Condexp
  • Mathlib.Probability.Moments.IntegrableExpMul
  • AutoSamplingTheory.Core
Local dependencies

Downstream consumers

  • diffusion generative models