Lemma 4.3, pp.15–16 — Hessian bounds, second moments and , Lipschitz
ProvedCritWinSpin.lemma_4_3p2o-gran-per-chapterp2o-plan-paperp2o-v1spin-systems
There are depending only on such that for all , with , , and :
- for a.e. , and ;
- with and ,
- for all , .
These are the inputs to the Langevin analysis of the sampler.
Differences from the source:
- encoding: ∇²W(z) at a.e. z → the second Fréchet derivative, with its existence at a.e. z a conclusion, as a quadratic form; reason: the Hessian as a form; W is C¹ with a piecewise-C² profile
- encoding: expectations and TV of densities → Bochner integrals of normalised densities with Integrable conclusions, TV = ½∫|f − g|; reason: rules out junk values of the integral
- reading: constants depending only on ρ → ∃ M₀, C_K before N, L, s, α, h; reason: §3, p.10, and the lemma's wording
Preamble
import Mathlib import Definitions.Def_IsingLTL_BeliefProp_TVDist import Definitions.Def_CritWinSpin_Wpot open MeasureTheory IsingLTL.BeliefProp set_option autoImplicit false
Formal statement
namespace CritWinSpin
/-- **Lemma 4.3** (Liu–Zhang–Ren, STOC 2027 submission, manuscript sha256 f77584da…, pp.15–16). If
`LLᵀ ⪯ s I_N`, `τ < s ≤ 2τ + 1`, `α ∈ (0, 1]`, there are `M₀, C_K > 0` depending only on `ρ` such that for
all `h`: (i) for a.e. `z`, `0 ⪯ ∇²W ⪯ I_N` and `α I_N ⪯ ∇²U_α ⪯ (1 + α) I_N`; (ii) the second moments under
`q ∝ e^{−W}` and `ν_α ∝ e^{−U_α}` are at most `M₀N` and `TV(ν_α, q) ≤ αM₀N/2`, eq. (23);
(iii) `TV(K_z, K_w) ≤ C_K √N ‖z − w‖`. -/
theorem lemma_4_3 (S : SpinLaw) (hS : IsGHS S) :
∃ M₀ Cₖ : ℝ, 0 < M₀ ∧ 0 < Cₖ ∧
∀ (N : ℕ) (L : Matrix (Fin N) (Fin N) ℝ) (s α : ℝ) (h : Fin N → ℝ),
tau S < s → s ≤ 2 * tau S + 1 →
(s • (1 : Matrix (Fin N) (Fin N) ℝ) - L * L.transpose).PosSemidef →
0 < α → α ≤ 1 →
(∀ᵐ z : Fin N → ℝ,
DifferentiableAt ℝ (fderiv ℝ (Wpot S s L h)) z ∧
DifferentiableAt ℝ (fderiv ℝ (Ualpha S s α L h)) z ∧
∀ v : Fin N → ℝ,
0 ≤ fderiv ℝ (fderiv ℝ (Wpot S s L h)) z v v ∧
fderiv ℝ (fderiv ℝ (Wpot S s L h)) z v v ≤ normSq v ∧
α * normSq v ≤ fderiv ℝ (fderiv ℝ (Ualpha S s α L h)) z v v ∧
fderiv ℝ (fderiv ℝ (Ualpha S s α L h)) z v v ≤ (1 + α) * normSq v) ∧
(Integrable (fun z : Fin N → ℝ => Real.exp (-Wpot S s L h z)) ∧
Integrable (fun z : Fin N → ℝ => normSq z * Real.exp (-Wpot S s L h z)) ∧
Integrable (fun z : Fin N → ℝ => Real.exp (-Ualpha S s α L h z)) ∧
Integrable (fun z : Fin N → ℝ => normSq z * Real.exp (-Ualpha S s α L h z)) ∧
max (∫ z : Fin N → ℝ, normSq z * normDens (Wpot S s L h) z)
(∫ z : Fin N → ℝ, normSq z * normDens (Ualpha S s α L h) z) ≤ M₀ * (N : ℝ) ∧
(1 / 2) * (∫ z : Fin N → ℝ,
|normDens (Ualpha S s α L h) z - normDens (Wpot S s L h) z|) ≤
α * M₀ * (N : ℝ) / 2) ∧
(∀ z w : Fin N → ℝ,
tvDist (kernelK S s L h z) (kernelK S s L h w) ≤
Cₖ * Real.sqrt (N : ℝ) * Real.sqrt (normSq (z - w))) := by sorry
end CritWinSpin
Source
Liu, Zhang, Ren, The Critical Window for Sampling and Learning-to-Sample in Spin Systems, STOC 2027 submission (manuscript sha256 f77584da…), Lemma 4.3, pp.15–16, eq. (23)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.