- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies the convergence of exponential-integrator samplers for probability flow ODEs. Under weak log-concavity and -Lipschitz score assumptions, the authors derive error bounds with polynomial dependence on the step size . The theoretical results also apply to diffusion models with general time-dependent drift and diffusion coefficients.
Critical strengths
- The paper relaxes the strong log-concavity assumption to weak log-concavity and establishes error bounds for PF-ODEs with general noise coefficients.
Critical weaknesses
-
It is unclear how the weak log-concavity assumption compares with the bounded-support or finite-second-moment assumptions used in previous work. In practical applications, the bounded-support assumption is often reasonable, since data such as images and videos are typically normalized to a bounded domain. However, weak log-concavity and -Lipschitz score assumptions may not hold for such data distributions. Moreover, the paper provides limited discussion of how the weak log-concavity parameters and affect the error bounds in Theorem 5 for specific distributions. Although Proposition 7 establishes the necessity of exponential dependence for the initialization error , it remains unclear how these parameters influence the discretization and score-matching errors.
-
The extension from constant to time-dependent drift and diffusion coefficients seems incremental. It seems to be a reparameterization of the noise schedule through a rescaling of time. The authors should clarify the novelty of this extension.
-
The claim that PF-ODEs are more efficient than SDEs is not sufficiently justified by comparing their upper bounds on sampling complexity alone.
Questions for authors
-
Do the weak log-concavity and -Lipschitz score assumptions hold for practical data distributions, such as images and videos? If so, how large can the regularity parameters and be, and how do they affect the error bounds?
-
Does a similar exponential dependence on the weak log-concavity parameters and , as established in Proposition 7 for the initialization error, also arise in the discretization and score-matching errors? How tight are the corresponding bounds in Theorem 5?
3 comments
Thank you for the further clarification and the additional theoretical results.
As I understand it, Proposition 7 establishes a worst-case lower bound using a specifically constructed two-component Gaussian mixture and a constant score perturbation . It remains unclear how this result should be interpreted under the assumptions of the paper. Could the authors further clarify the practical implications of the upper and lower bounds, and whether the predicted exponential error amplification can be observed in numerical experiments?
We thank the reviewer for this suggestion. We have added numerical simulations in Appendix D.3 to illustrate the exponential error amplification established in Proposition 7. We compare the Gaussian mixture used in the lower-bound construction with a Gaussian-smoothed uniform distribution, which is log-concave. The results show pronounced amplification of all three error components for the Gaussian mixture, consistent with the predicted exponential dependence, while the log-concave distribution exhibits substantially milder growth. These findings support the necessity of exponential dependence for uniform error bounds under weak log-concavity, while illustrating that such amplification need not occur for distributions satisfying stronger assumptions.
Thank you for your continued engagement with our response and for the opportunity to further clarify these points.
Exponential dependence. We agree that the exponential dependence is a crucial point. In fact, in the revised manuscript, we have established a new lower bound (Proposition 7) showing that exponential dependence on M0/α0 is in general unavoidable, which complements our preceeding results in this direction on initialization error (Proposition 7 in our original manuscript). We believe these new results substantially address the concern and would be happy to hear the reviewer's thoughts on them. We provide a more detailed summary of our changes in the general comment above.
Regarding early stopping, it is not required under our assumptions. Indeed, unlike settings involving bounded-support or singular targets, where the score may become irregular as the smoothing noise vanishes, our regularity assumptions allow us to control the score down to time zero (Lemma 26). Our guarantees therefore concern the original target p0, without introducing an early-stopping bias.
General noise schedules. Treating time-dependent coefficients allows us to quantify how the noise schedule affects convergence. In particular, it determines the transition time from weak to strong log-concavity, the contraction properties of the error-propagation factors γk,h (Proposition 18), and the accumulation of discretization errors. Moreover, our analysis of time regularity (Lemma 26 in the revised version) shows that a time-uniform Lipschitz constant for the score cannot generally be taken for granted: additional conditions on f and g are needed. We believe these insights provide a substantive motivation for considering general noise schedules.
PF-ODE efficiency comparison. We agree that it is important not to overclaim results, and we do not intend to do so. Our contribution provides tables guiding the choice of hyperparameters based on upper bounds. A natural question is how these compare with analogous results available in the literature for SDE samplers, particularly those of Gao et al. (2025, Table 2). We believe this comparison is interesting, especially since our analysis extends the proof techniques of Gao and Zhu (2025) for probability flow ODEs. Of course, a comparison based solely on upper bounds is not conclusive, and we do not claim otherwise. Nevertheless, the existing Wasserstein bounds guarantee convergence of ODE samplers under less restrictive step-size conditions, whereas the same conditions are not sufficient to establish convergence through the corresponding SDE bounds. This does not imply that SDE samplers cannot achieve comparable efficiency, but rather highlights an interesting difference between the currently available theoretical guarantees.