Langevin iteration with finite precision: explicit weak-error budget
ProvedCritWinSpin.langevin_weak_budgetfinite-precisionlangevinprobabilitysampling
设 带欧氏范数, 为连续势函数, 是密度正比于 的概率测度, 为连续漂移。假定 以常数 强单调、Lipschitz 常数至多为 ,且 。记 ,其中 为一维标准高斯。
令 作用于有界可测测试函数,保持可测性和一致界,并在一致范数中不扩张。对于任意有界 -Lipschitz 测试函数,假设 与高斯 Euler 步 的单步期望差至多为 。
取 ,要求 、、,并要求这些矩确有可积性。若 、、、,且
则每个有界、可测且 的 Lipschitz 测试函数满足
这是 Theorem 4.4 有限精度采样器所用的通用 Langevin 弱误差估计。它把势函数的几何条件、两个矩界与实际数值算子的单步误差合并成显式的有限步精度保证。
Preamble
import Mathlib open MeasureTheory ProbabilityTheory open scoped ENNReal RealInnerProductSpace
Formal statement
namespace CritWinSpin
theorem langevin_weak_budget (d : ℕ) (U : EuclideanSpace ℝ (Fin d) → ℝ)
(g : EuclideanSpace ℝ (Fin d) → EuclideanSpace ℝ (Fin d))
(μ : Measure (EuclideanSpace ℝ (Fin d)))
(hμ : μ = ((∫⁻ x : EuclideanSpace ℝ (Fin d),
ENNReal.ofReal (Real.exp (-U x)) ∂volume)⁻¹) •
(volume : Measure (EuclideanSpace ℝ (Fin d))).withDensity
(fun x => ENNReal.ofReal (Real.exp (-U x))))
(a s e R : ℝ)
(hU : Continuous U) (hg : Continuous g)
(hiU : Integrable (fun x => Real.exp (-U x)))
(hmono : ∀ x y, a * ‖x - y‖ ^ 2 ≤ inner ℝ (x - y) (g x - g y))
(hTaylor : ∀ x y, |U y - U x - inner ℝ (y - x) (g x)| ≤ ‖y - x‖ ^ 2)
(hLip : ∀ x y, ‖g x - g y‖ ≤ 4 * ‖x - y‖)
(hiG : Integrable (fun x => ‖g x‖ ^ 6) μ)
(hiNorm : Integrable (fun x => ‖x‖) μ)
(ha : 0 < a) (hs : 0 < s) (hs1 : s ≤ 1) (he0 : 0 ≤ e)
(hsmall : a * s ^ 2 ≤ 1) (hstep : 16 * s ^ 2 ≤ a)
(hR : 2 ≤ R) (haR : 1 ≤ a * R) (hsR : s ≤ 1 / (2048 * R ^ 4))
(he : e ≤ s ^ 2 / (16 * R ^ 3))
(hM : (∫ x, ‖x‖ ∂μ) ≤ R)
(hH : 2 + (∫ x, ‖g x‖ ^ 6 ∂μ) +
(∫ t : ℝ, |t| ^ 6 ∂gaussianReal 0 1) * (d : ℝ) ^ 3 ≤ R)
(Q : (EuclideanSpace ℝ (Fin d) → ℝ) → EuclideanSpace ℝ (Fin d) → ℝ)
(hQ : ∀ (f : EuclideanSpace ℝ (Fin d) → ℝ) (B : ℝ),
Measurable f → (∀ x, |f x| ≤ B) → Measurable (Q f) ∧ ∀ x, |Q f x| ≤ B)
(hQC : ∀ (f k : EuclideanSpace ℝ (Fin d) → ℝ) (B D : ℝ),
Measurable f → Measurable k → (∀ x, |f x| ≤ B) → (∀ x, |k x| ≤ B) →
(∀ x, |f x - k x| ≤ D) → ∀ x, |Q f x - Q k x| ≤ D)
(happrox : ∀ (f : EuclideanSpace ℝ (Fin d) → ℝ) (B L : ℝ), 0 ≤ L →
Measurable f → (∀ x, |f x| ≤ B) → (∀ x y, |f x - f y| ≤ L * dist x y) →
∀ x, |Q f x - ∫ z, f (x - s ^ 2 • g x + (Real.sqrt 2 * s) • z)
∂stdGaussian (EuclideanSpace ℝ (Fin d))| ≤ L * e)
(f : EuclideanSpace ℝ (Fin d) → ℝ) (B L : ℝ)
(hL : 0 ≤ L) (hLR : L ≤ R) (hf : Measurable f)
(hfB : ∀ x, |f x| ≤ B) (hfL : ∀ x y, |f x - f y| ≤ L * dist x y)
(n : ℕ) (hn : 16 * R ^ 4 / s ^ 2 ≤ (n : ℝ)) :
|(Q^[n] f) 0 - ∫ x, f x ∂μ| ≤ 1 / (2 * R) := by sorry
end CritWinSpin
Source
New auxiliary finite-dimensional weak-approximation theorem supporting Theorem 4.4, section4.2, pp16-17, of The Critical Window for Sampling and Learning-to-Sample in Spin Systems. Official manuscript https://github.com/sigact-stoc/stoc2027/releases/download/papers-2/943.pdf (SHA256 f77584dab87d9e6d0936a083a2e3d26476a91a9d73dc303674514a3257c0de95). The explicit R-budget statement and Metropolis-flux proof are new formal support results, not a separately numbered lemma in that manuscript.