acceptodds
Under review as a conference paper at ICLR 2027

Sharp Wasserstein-1 Stability for Diffusions under Radial Drift Bounds

Abstract

We study how drift and initialization errors propagate in diffusions that may expand locally but contract at large distances. Under a radial constraint on the reference drift, we derive a finite-time Wasserstein-1 bound for diffusions with the same additive isotropic noise coefficient bounded away from zero. The bound separates initialization error from expected drift error along the perturbed process. Each term is optimal when the other error is absent. Reflection coupling reduces the analysis to a one-dimensional diffusion absorbed at zero. For fixed radial parameters allowing expansion, the amplification coefficient has a polynomial peak as the noise lower bound decreases. Its integral over all time grows exponentially; accumulating a fixed fraction takes an exponentially long time. A reverse Gaussian mixture shows why a sharp uniform coefficient can still overestimate a specified error: the optimal coefficient grows quadratically with mode separation, while adding a fixed positive constant to the drift gives bounded error per unit amplitude.

Primary area:
learning theory
Keywords:
Score-based diffusion models, Wasserstein stability, Drift perturbations

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YQNcOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper establishes bounds on the Wasserstein-1 distance between the laws of two diffusion processes differing in their drifts and initializations. The results assume a dissapitivity-type condition (Eq. 2) under which bounds on the Wasserstein-1 distance is established.

Critical strengths

I do not see any critical strengths in the paper. The problem statement and motivation is unclear.

Critical weaknesses

This paper is unfortunately unreadable. It is unclear why this manuscript considers diffusion processes driven by what is called "radial drift bounds". It is unclear what implications the results have, or what technical difficulty is resolved. The related work section lists references without describing how they relate to the paper. Is the relation in the underlying setting? Result? What are the differences/similarities?

Constructive comments

The paper needs significant revision with a clear scope. In my best efforts to guide the authors, I state some comments as a starting point.

  1. The law μ^t\hat{\mu}_t is defined as "the law of an approximation with drift ...". I suppose the intent is to define it as the law of the solution to the SDE with btb_t replaced by b^t\hat b_t.

  2. The sentences "Our objective is to ... which we use to establish sharpness" are unclear. "sharpness" does not have a reference; the last sentence reads as using "the linear transport cost" to establish sharpness, but it's unclear what "sharpness" is being referred to. I suppose the authors are trying to say that they establish sharp bounds on the Wasserstein-1 distance. Moreover, the title states Wasserstein-1 distance, but the introduction defines it in three different ways "expectation errors over 1-Lipschitz observables", "their largest errors", "the linear transport cost".

  3. The introduction needs to define the problem and a motivation for the problem. Equation (2) can be satisfied by a choice of bb, but why should one consider such diffusion processes? What is a "reference" diffusion"? These should be clearly stated.

  4. Regarding the math, the paper needs to justify many reasons for considering the assumptions of this paper. For which problems does bb satisfy Assumption 2.1, and how is this relevant? Even if bb satisfies Assumption 2.1, why should b^\hat b satisfy Assumption 2.1? Can't the results of this be established under sample path distances, not just distances between laws? The choice of assumptions and metric need to be carefully thought out.

Questions for authors

Please see the weaknesses section.

Reviewer_JtFjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies finite-time Wasserstein-1 stability of diffusions with shared nondegenerate additive isotropic noise, under radial drift bounds permitting local expansion and contraction at large distances. Using reflection coupling and a one-dimensional absorbed comparison diffusion, it derives a bound separating initialization error from expected drift error along the perturbed process. Both terms are proved separately optimal over the prescribed class, with an extension incorporating a one-sided Lipschitz bound. The paper also characterizes polynomial peak amplification and exponentially large time-integrated amplification in the small-noise regime. A reverse Gaussian-mixture example shows that a sharp uniform coefficient can substantially overestimate the response to a spatially constant perturbation. Validation consists of mathematical proofs, smooth constructions approaching the optimal bounds, and analytical examples; no experiments with trained diffusion models are presented.

Critical strengths
  1. The paper establishes sharp finite-time Wasserstein-1 stability bounds under radial drift constraints that permit local expansion. The separate optimality results for initialization and drift errors are supported by smooth constructions that preserve the stated constraints. This provides a precise benchmark for the strongest uniform guarantees available under these assumptions.

  2. The Gaussian-mixture example clearly distinguishes worst-case amplification from sensitivity to a specific perturbation. The optimal uniform coefficient grows quadratically with mode separation, whereas the error per unit constant forcing remains bounded. This illustrates why the spatial location of drift errors matters and why even sharp uniform bounds can poorly describe particular perturbations.

Critical weaknesses

The significance of the sharp bounds for learned diffusion models remains insufficiently established. The optimality results concern expected L1 drift errors, whereas standard score-matching objectives involve L2 errors under the forward marginals. The paper provides a connection, but this requires radial control of the learned drift, and optimality for the training-relevant L2 metric is not established. The bounded-denoiser example identifies an admissible model class, but the paper does not demonstrate that the resulting coefficients provide informative guarantees at realistic dimensions and noise levels. Consequently, while the worst-case characterization is mathematically meaningful, its practical or predictive advantage over existing score-error analyses remains unclear. This limits the demonstrated significance of the contribution for diffusion modeling.

Questions for authors

Clarification of the implications for score learning: For the bounded-denoiser application in Section 2.2, can you identify a concrete parameter regime where the resulting forward-marginal L2 score-error bound is informative and quantitatively improves on existing guarantees under comparable assumptions? Please clarify how this comparison depends on the data radius, minimum smoothing variance, and radial constraint on the learned drift. A comparison derived from the current results would help assess whether the sharp L1 coefficients yield a meaningful advance for learned diffusion models; additional experiments or a new L2 optimality theorem are not required.

Reviewer_n7PYOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper gives sharp finite-time Wasserstein-1 stability bounds for diffusions whose reference drift satisfies a radial contraction condition with bounded noncontractive allowance. An absorbed one-dimensional comparison diffusion defines a concave modulus Φt\Phi_t for initial-distribution error, whose boundary derivative κ(t)\kappa(t) controls the accumulation of drift error measured under the perturbed marginals. The paper proves the separate optimality of these two terms within the stated class and derives the time scales of amplification. Furthermore, this paper provides an explicit reverse Gaussian-mixture example in which a class-optimal bound substantially overestimates a particular constant perturbation.

Critical strengths
  1. The paper identifies the optimal finite-time Euclidean stability envelope under a meaningful drift condition. The absorbed comparison mean provides a nonlinear initial-error modulus, which yields the drift-error coefficient. Establishing sharpness using admissible smooth processes, including a fixed spatial perturbation profile, gives the result more substance than an upper bound obtained from standard coupling alone.

  2. The paper explains two distinct sources of pessimism in stability estimates. With χ=A2/(mσ2)\chi=A^2/(m\sigma^2), the maximal amplification is 2χ+O(χ)2\chi+O(\sqrt{\chi}), whereas its integral over dimensionless time q=mtq=mt is asymptotic to 2πχeχ2\sqrt{\pi\chi}e^\chi and the exponential scale comes from diffusion time. Separately, the reverse-mixture example has an optimal fixed-process coefficient growing quadratically with mode separation, while the response to a fixed positive constant drift error tends to log⁡2\log 2 after normalization. This shows concretely how the spatial placement of error matters even after class-level looseness has been removed.

Critical weaknesses
  1. As the discretization error is also a important part when using diffusion models in applicaiton. It would be better to show the influence of discretization error in the analysis. As this work studied the W1W1 distance, differenc from KL or TV, it can not directly use the Girsanov theorem.
Questions for authors

Please see the weakness part.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.