acceptodds
Under review as a conference paper at ICLR 2027

Non-asymptotic error bounds for probability flow ODEs under weak log-concavity

Abstract

Score-based generative modeling, implemented through probability flow ODEs, has shown impressive results in numerous practical settings. However, most convergence guarantees rely on restrictive regularity assumptions on the target distribution—such as strong log-concavity or bounded support—or focus on the specific case of constant drift and diffusion coefficients. This work establishes non-asymptotic convergence bounds in 2-Wasserstein distance for probability flow ODEs when the target is only weakly log-concave, allowing multimodal distributions such as Gaussian mixtures and time-dependent noise schedules. Our bounds account for initialization, discretization, and score approximation errors. We show that, for fixed regularity parameters of the target distribution, weakening strong to weak log-concavity leaves the asymptotic dependence on the time horizon, step size, and score matching error. Yet, the relaxation incurs a potentially large non-asymptotic price: the bounds contain an exponential dependence on the deviation from log-concavity. We complement the upper bound with a lower bound construction, showing that this exponential dependence is unavoidable under weak log-concavity alone.

Primary area:
generative models
Keywords:
Score-based generative modeling, Diffusion models, ODE-based samplers, Convergence analysis

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_E6FnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the W2W_2 convergence of exponential-integrator samplers for probability flow ODEs. Under weak log-concavity and L0L_0-Lipschitz score assumptions, the authors derive error bounds with polynomial dependence on the step size hh. 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 W2W_2 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 L0L_0-Lipschitz score assumptions may not hold for such data distributions. Moreover, the paper provides limited discussion of how the weak log-concavity parameters α0\alpha_0 and M0M_0 affect the error bounds in Theorem 5 for specific distributions. Although Proposition 7 establishes the necessity of exponential dependence for the initialization error E0E_0, 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 W2W_2 upper bounds on sampling complexity alone.

Questions for authors
  • Do the weak log-concavity and L0L_0-Lipschitz score assumptions hold for practical data distributions, such as images and videos? If so, how large can the regularity parameters α0\alpha_0 and M0M_0 be, and how do they affect the error bounds?

  • Does a similar exponential dependence on the weak log-concavity parameters α0\alpha_0 and M0M_0, 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 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/α0M_0/\alpha_0 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 p0p_0, 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\gamma_{k,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 ff and gg 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.

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 sθ=s+εs_\theta=s+\varepsilon. 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.

Reviewer_PUXXOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper derives non-asymptotic W2W_2 error bounds for probability-flow ODE samplers under weak log-concavity, allowing multimodal targets and time-dependent noise schedules. The bounds account for initialization, discretization, and score-estimation errors. Its analysis uses the eventual emergence of strong log-concavity under forward noising. It also claims that exponential dependence on the degree of non-log-concavity is unavoidable, supported by a lower-bound construction.

Critical strengths

The lower bound is the main strength of the paper. This results provides that exponential dependence on the weak log-concavity parameters is unavoidable for the probability-flow initialization error, identifying an intrinsic limitation rather than merely a weakness of the upper-bound analysis.

Critical weaknesses

Limited conceptual novelty. The weak log-concavity framework, its propagation under forward noising, and the transition between contraction regimes closely follow Gentiloni-Silveri & Ocello, 2025. The submission largely reuses this existing conceptual structure.

The significance of the ODE extension is insufficiently established. Probability-flow ODEs constitute a distinct setting, but the manuscript does not clearly identify which substantial technical obstacles require new ideas beyond adapting the existing SDE analysis. The change of setting alone does not establish a sufficiently significant contribution.

The noise-schedule extension appears incremental. The linear forward process admits a representation through Gaussian smoothing, deterministic scaling, and time change. The authors should clarify what their extension requires beyond applying existing regularization arguments through this representation.

The novelty rests heavily on the lower bound. This appears to be the clearest additional contribution, but it concerns initialization error for the exact probability flow. It does not establish sharpness of the complete bound incorporating discretization and score-estimation errors, which limits how broadly its significance can be interpreted.

Questions for authors

See weaknesses part.

1 comment

As mentioned in our earlier response, we have now uploaded a revised version of the manuscript with several new results. In particular, we have extended the lower bound result (Proposition 7) to cover all three error sources, showing that exponential dependence on M0/α0M_0/\alpha_0 is unavoidable. This addresses your concern on the scope of the lower bound. Moreover, we have replaced Assumption 2 with a condition on the third derivative of log⁡p0\log p_0, addressing a limitation of existing probability flow ODE error bounds by replacing the temporal regularity condition along the forward process with a direct condition on the target distribution that is easier to interpret and verify. Notably, the new assumption does not change the error bounds.

We provide a more detailed summary of our changes in the general comment above. We would be happy to hear your thoughts on them and hope that they address your concerns regarding the scope of the lower bound and the ODE setting. The remaining concerns were already discussed in our previous response.

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

In this paper, the authors studied non-asymptotic bounds in 2-Wasserstein distance for the probability flow ODE. In particular, they studied the case in which the target distribution is weakly log-concave and not strongly log-concave, as often assumed. They discover that, in this more general regime, many parameters (like time horizon, step size and score matching error) maintain the same asymptotic price, while in the non-asymptotic regime an exponential dependence on the deviation from the log-concavity appears. To finish, they also show that the exponential dependence is unavoidable in this setting via a lower bound.

Critical strengths

The paper fills a gap in the literature: strong log-concavity and general schedules were not assumed at the same time. Moreover, the time analysis is explicit, with closed-form expressions of the time dependence in the constants, and explicit regime shift with time. Finally, the presence of a lower bounds answer to the most natural question, which is if the exponential dependence is avoidable or not.

Critical weaknesses

It may have been nice to have some experiments to check the results

Questions for authors
  1. Regarding the assumptions, is the second still holding for example in Example 1 GMM? Moreover, assumption 3 is formulated on the law of the algorithm’s iterates, which depends on sθs_\theta. Can it be transferred to a score-matching error measured under ptp_t, the quantity that training actually controls?

  2. In Appendix C you decide to ignore constants in the exponents. For eTβ2(T)he^{T\beta^2(T)h} shifts the admissible threshold on hh by a multiplicative factor. What is the actual constant, and does it change the entries of Table 1 and Table 2?

2 comments

As mentioned in our earlier response, we have now uploaded a revised version of the manuscript with several new results. In particular, we have replaced Assumption 2 with a condition on the third derivative of log⁡p0\log p_0, addressing a limitation of existing probability flow ODE error bounds by replacing the temporal regularity condition along the forward process with a direct condition on the target distribution that is easier to interpret and verify. We have also included a discussion of this assumption for a GMM (Example 4 in Appendix F). Notably, the new assumption does not change the error bounds.

We provide a more detailed summary of our changes in the general comment above. We would be happy to hear your thoughts on them and hope that they address your concerns regarding Assumption 2. The remaining concerns were already discussed in our previous response.

We thank the reviewer again for their positive assessment and helpful suggestion. Following their recommendation, we have added a numerical experiment illustrating the exponential error amplification predicted by our theoretical results, together with a comparison against a log-concave distribution.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised version

We thank the reviewers for their constructive feedback. We have uploaded a revised manuscript addressing the two main concerns:

1. Lower bound for all three error sources. The revised Proposition 7 establishes lower bounds for initialization, discretization, and propagated score-matching error. These show that an exponential dependence on M0/α0M_0/\alpha_0, which measures the deviation from log-concavity, is unavoidable for each error source. The original manuscript established this only for the initialization error. All three lower bounds use the same target distribution: a one-dimensional, unbalanced two-component Gaussian mixture. We have also strengthened Proposition 6. While the original statement compared the asymptotic behavior of our error bound with that of Gao & Zhu (2025), the revised statement provides an explicit non-asymptotic comparison. This quantifies the additional dependence on M0/α0M_0/\alpha_0 arising from our weaker assumption and, together with Proposition 7, shows why an exponential dependence cannot generally be avoided.

2. Replacing the temporal Lipschitz assumption. We have replaced Assumption 2 with a condition on the third derivative of log⁡p0\log p_0. The new condition concerns the target distribution directly, making it easier to interpret and verify than a temporal regularity condition along the forward process. Notably, the new assumption does not change the error bounds. Temporal Lipschitz assumptions also appear in prior probability flow ODE analyses, including Gao & Zhu (2025) and Gao, Nguyen, and Zhu (2025), whereas reverse-SDE analyses can avoid such assumptions. Our revision therefore addresses a limitation of existing probability flow ODE error bounds by replacing this assumption with a condition on the target density.

We have highlighted corresponding changes in the main text of the manuscript in blue. The proof of the new Proposition 6 and 7 can be found in Appendix D, and all results related to the replaced temporal Lipschitz assumption are gathered in Appendix F. We have prioritized these two substantive changes in the uploaded revision. We will also incorporate the remaining minor revisions described in our responses to the reviewers into the final manuscript.

We thank the reviewers again for their constructive feedback. We believe that these revisions substantially strengthen the paper and further underscore the novelty and significance of its contributions.

We included a new numerical experiment in the appendix. The simulations support the predicted exponential error amplification, while illustrating milder behavior for a log-concave distribution. We thank the reviewers for their valuable feedback, which helped us improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.