Poisson-Corrector Complexity Bounds for Moreau--Yosida Unadjusted Langevin Sampling
cs.LG
Quick verdict
MYULA invariant bias O(h)+Õ(h^{3/4}) via Poisson corrector, giving Õ(ε^{-4/3}) complexity; rigorous and novel, but key proofs were absent from the excerpt.
Summary
This is a pure theory paper on the Moreau–Yosida unadjusted Langevin algorithm (MYULA) for sampling from π ∝ e^{−f−g}, with f strongly convex and C² with Lipschitz gradient, and g convex and G-Lipschitz (possibly nonsmooth). Under the step-size restriction h(L_f+λ^{−1}) ≤ c, the authors prove the invariant-measure bias bound √m W₂(π_λ, π̂_{λ,h}) ≤ C₁h + C₂h^{3/4}ℓ², where ℓ depends on λ^{−1} only logarithmically — the headline feature. Combining this with the Moreau bias √m W₂(π_λ,π) ≤ G²λ/4 and choosing λ = ε/G² yields a sufficient Õ(ε^{−4/3}) iteration complexity for √m W₂(μ_N,π) ≤ ε; the authors contrast this with the Õ(ε^{−2}) obtained by specializing Pedrotti–Whalley's smooth-target bound. The proof welds a discrete Poisson corrector, active-trace averaging from the authors' companion preprint, and a new shared-noise estimate E tr A_g ≤ CG√(d/h) for the exact–Euler coupling; appendices develop spectral calculus and an Lp multiplier theorem with unusual care. Limitations: the h ≲ λ step-size restriction remains, sharpness of the h^{3/4} rate and of the d-dependences is not discussed, the table of contents shows no numerical experiments, and the argument builds on two unpublished 2026 preprints. Complexity counts exact proximal iterations only.
Review complete
Assessment at a glance
3
risk flags identified
● high 0
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Under Assumptions 4.1–4.2 and h(L_f+λ^{-1}) ≤ c, the MYULA invariant law satisfies √m W₂(π_λ, π̂_{λ,h}) ≤ C₁h + C₂h^{3/4}ℓ²_{h,λ} with coefficients depending on λ^{-1} only logarithmically (Theorem 5.1).
- With λ = ε/G² and the step size (5.6), a sufficient Õ(ε^{-4/3}) iteration complexity holds for √m W₂(μ_N,π) ≤ ε (Theorem 5.2), improving the Õ(ε^{-2}) obtained by specializing Pedrotti–Whalley.
- The enabler is a new shared-noise bound E tr A_g^k ≤ CG√(d/h) for the exact–Euler two-point curvature, letting averaged Moreau curvature replace the worst-case λ^{-1}.
Strengths
Quantified improvement over a concrete baseline
Section 2 explicitly computes that specializing Pedrotti–Whalley's global-smoothness bound with λ ≍ ε/G² yields Õ(ε^{-2}); the averaged-curvature refinement delivers Õ(ε^{-4/3}) (Theorem 5.2, eq. 5.10), so the gain is stated against a computed baseline rather than asserted.
Explicit, checkable constants
R and K are defined in (5.1) and ℓ_{h,λ} in (5.2); Theorem 5.1's bound C[R(√m+√L_f)+L_f√d]h + C[G√R+K√G d^{1/4}]ℓ²h^{3/4} and the complexity display (5.9) expose full parameter dependence instead of hiding it in O(·).
Unusually careful functional analysis
Genuinely new technical estimate
Honest oracle accounting
Weaknesses
Step-size restriction persists
No sharpness discussion
Dependence on non-refereed preprints
No numerical component
Redundant standing assumption
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper is a pure-theory analysis of MYULA for targets π(dx) ∝ e^{−f(x)−g(x)}dx under Assumptions 4.1–4.2: f ∈ C², m-strongly convex with L_f-Lipschitz gradient; g convex and globally G-Lipschitz (nonsmooth penalties such as ℓ1 are the motivation). Two results are stated. Theorem 5.1 (fixed smoothing): under h(L_f+λ^{−1}) ≤ c, √m W₂(π_λ, π̂_{λ,h}) ≤ C[R(√m+√L_f)+L_f√d]h + C[G√R+K√G d^{1/4}]ℓ²_{h,λ}h^{3/4}, with R, K from (5.1) and ℓ_{h,λ} = 1+log(e+√h G√d/λ) from (5.2) — the coefficients depend on λ^{−1} only logarithmically. Theorem 5.2: with λ_ε = ε/G², the Moreau bias √m W₂(π_λ,π) ≤ G²λ/4 (cited from Xin–Zhang [13, Prop 5.9]) and Wasserstein contraction give the complexity display (5.9), hence N(ε) = Õ(ε^{−4/3}) (5.10) for fixed parameters and initialization. Section 2 anchors the contribution: specializing Pedrotti–Whalley's linear-in-h smooth-target bound with worst-case curvature λ^{−1} yields Õ(ε^{−2}); the Õ(ε^{−4/3}) comes from averaging the Moreau curvature.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 3 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What iteration complexity does this paper prove for MYULA?
A sufficient Õ(ε^{-4/3}) iterations — logarithmic factors suppressed, with model parameters and initialization fixed — to reach √m W₂(μ_N, π) ≤ ε for the original composite target, using smoothing λ = ε/G² and the step size h_ε defined in eq. (5.6) of the paper.
What is new compared with Pedrotti and Whalley's analysis?
Specializing the Pedrotti–Whalley corrector bound to a Moreau-smoothed target, using the worst-case curvature λ^{-1}, gives Õ(ε^{-2}) complexity. This paper instead averages the Moreau curvature — via active traces and a new shared-noise bound — obtaining an O(h) + Õ(h^{3/4}) invariant-measure bias and hence Õ(ε^{-4/3}).
Does the paper include experiments or code?
No. The table of contents lists only theory sections (1–9) and two appendices (A: weak Hessians and integration by parts; B: operator facts and an Lp multiplier theorem); no numerical experiments or code are mentioned.
What assumptions does the main theorem need?
f ∈ C²(R^d) that is m-strongly convex with L_f-Lipschitz gradient; g convex and globally G-Lipschitz (nonsmooth penalties like ℓ1 are allowed); and a step size satisfying h(L_f + λ^{-1}) ≤ c for a small universal constant c.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with three marked gaps: parts of Section 2 (the Õ(ε^{-2}) calculation), Sections 6.3–6.5 (the one-step matrix inequality and the theorem proofs), and parts of 7.5/8.1 (Proposition 8.1's statement and proof) were skipped; judgments about those parts rest on the visible statements of Props 6.2–6.3 and the surrounding estimates.
- External results the proof imports (Carbonaro–Dragičević [3], Chewi [4], Bakry–Gentil–Ledoux [1]) could not be checked offline; their use is assessed only as presented in the excerpt.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.