acceptodds
Under review as a conference paper at ICLR 2027

Sharp Convergence and Sampling Trade-offs for Riemannian Diffusion under Nonnegative Ricci Curvature

Abstract

Diffusion models have emerged as state-of-the-art generative models, with recent extensions from Euclidean spaces to Riemannian manifolds. However, existing convergence guarantees for Riemannian diffusion models typically require score evaluations, with potentially unfavorable dependence on the dimension. In this work, we develop a general framework that separates score discretization from Brownian-motion simulation and allows multiple geodesic random-walk steps per score evaluation. Under nonnegative Ricci curvature assumption and an exact Brownian-motion simulation oracle, we show that score evaluations suffice to achieve an KL divergence from the target distribution, matching the existing convergence rate of Euclidean diffusion models. We further show that geodesic random-walk steps suffice to approximate the required drifted Brownian motion to total variation error. Combining these results yields a sampling scheme with score evaluations and geodesic random-walk steps, motivating multiple random-walk steps between consecutive score evaluations. Our results provide a sharper characterization of the convergence and sampling complexity of Riemannian diffusion models.

Primary area:
generative models
Keywords:
Riemannian diffusion models, Diffusion models, Generative models, Sampling algorithms

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_PQBGOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studied sampling guarantee of Riemannian diffusion models. Existing works only proved polynomial guarantee, but didn't explicitly work out the specific order. This paper tried to gave specific order.

Critical strengths

Personally I think this is a very interesting problem. In the sampling literature, to the best of my knwledge, analyzing the manifold Brownian motion simulation error (e.g., geodesic random walk) is a difficult, and the only existing result is a Wasserstein analysis under synchronous coupling. Recent paper Xu et al. (2026) proved polynomial guarantee, but the specific order is still open in KL or TV.

Possibly a question that might be out of the scope of the submitted paper could be, to trace the order of Wasserstein error bound of geodesic random walk (As I recall it might be like d3d^3 but I am not sure) in Cheng et al. (2022) and compare with the error bound in this paper. Just a comparison on manifold Brownian motion simulation itself.

Critical weaknesses

The main text is well written and easy to follow, but I feel some parts of the appendix difficult to follow. I think a revision on Appendix is needed.

Questions for authors

Some questions:

I feel Appendix D contain some typos and is difficult to follow.

Appendix D.4.1 ANALYSING THE DENSITY RATIO: for example, around line 1500, I don't see how the inequality was obtained. The rest of this subsection is also difficult to follow for me.

line 1481-1482: I didn't find where the author discuss the required range of h′h'. I think h′h' should be upper bounded by something depending on dd, and this is important.

As this is not the first convergence proof on Riemannian diffusion model, I think it's better to avoid using OO and oo notation through the proof, and only use it when it comes to the conclusion/final order. Moreover, I noticed that a universal constant C was used throughout the proof. I think it's better to define constants separately and track them (I feel this might make the proof easier to follow).

Line 448, "for sufficiently small ϵ>0\epsilon > 0" I think this should be stated explicit (e.g., require ϵ<d−5\epsilon < d^{-5} or so). I am not saying such condition is not ok. I understand manifold Brownian motion is complicated and difficult to analyze. But as I see from the proof (if I am wrong please let me know) the small epsilon played a role, in some sense absorbed some extra dimension dependency. What I mean is, back to the proof of Xu et al. (2026), they imposed some upper bound (say, t≤t0t \le t_0) on time interval of Brownian motion simulation. And in the submitted paper I think the author also imposed a similar condition. Now if this upper bound t0t_0 on the time variable which control the maximal allowed Brownian motion simulation time, has a very bad dependency on some variables, e.g. t0≲d−10t_0 \lesssim d^{-10} or even t0≲exp⁡(−d)t_0 \lesssim \exp(-d), the required small ϵ\epsilon would need to carry such order, to make hh small (so that the t0t_0 condition is satisfied).

possible typo: line 1228, clipped or truncated?

Summary: interesting paper, but the proof need more polishing, and track the constants more carefully.

7 comments

Thank you for your time and effort in reviewing our paper! We are grateful for your constructive suggestions, which have significantly guided our improvements. Please find our responses to your comments below.

Appendix D.4.1 ANALYSING THE DENSITY RATIO: for example, around line 1500, I don't see how the inequality was obtained. The rest of this subsection is also difficult to follow for me.

A: Thank you for your suggestion. We have expanded the derivation of the inequality and provided additional explanations to make the proof easier to follow. In particular, for the inequality you are asking, we now explicitly write the derivation as ∣P^(t,x,y)Q(t,x,y)−1∣≤∣P^(t,x,y)q^t(x,y)u0(x,y)−1∣+∣P^(t,x,y)q^t(x,y)u0(x,y)−P^(t,x,y)qt(x,y)u0(x,y)j(x)−1∣+∣P^(t,x,y)qt(x,y)u0(x,y)j(x)−1−P^(t,x,y)Q(t,x,y)∣\left|\frac{\hat{P}(t,x,y)}{Q(t,x,y)}-1\right|\le\left|\frac{\hat{P}(t,x,y)}{\hat{q}_t(x,y)u_0(x,y)}-1\right|+\left|\frac{\hat{P}(t,x,y)}{\hat{q}_t(x,y)u_0(x,y)}-\frac{\hat{P}(t,x,y)}{q_t(x,y)u_0(x,y)j(x)^{-1}}\right|+\left|\frac{\hat{P}(t,x,y)}{q_t(x,y)u_0(x,y)j(x)^{-1}}-\frac{\hat{P}(t,x,y)}{Q(t,x,y)}\right| ≤∣P^(t,x,y)q^t(x,y)u0(x,y)−1∣+P^(t,x,y)q^t(x,y)u0(x,y)⋅∣q^t(x,y)qt(x,y)j(x)−1−1∣+P^(t,x,y)qt(x,y)u0(x,y)j(x)−1⋅∣qt(x,y)u0(x,y)j(x)−1Q(t,x,y)−1∣.\le \left|\frac{\hat{P}(t,x,y)}{\hat{q}_t(x,y)u_0(x,y)}-1\right|+\frac{\hat{P}(t,x,y)}{\hat{q}_t(x,y)u_0(x,y)}\cdot\left|\frac{\hat{q}_t(x,y)}{q_t(x,y)j(x)^{-1}}-1\right|+\frac{\hat{P}(t,x,y)}{q_t(x,y)u_0(x,y)j(x)^{-1}}\cdot\left|\frac{q_t(x,y)u_0(x,y)j(x)^{-1}}{Q(t,x,y)}-1\right|. Here, the fractions P^(t,x,y)q^t(x,y)u0(x,y)\frac{\hat{P}(t,x,y)}{\hat{q}_t(x,y)u_0(x,y)} and P^(t,x,y)qt(x,y)u0(x,y)j(x)−1\frac{\hat{P}(t,x,y)}{q_t(x,y)u_0(x,y)j(x)^{-1}} are close to 1 (because all the absolute values on the right hand side are small), and are therefore upper bounded by constants.

We hope these revisions address your concerns.

line 1481-1482: I didn't find where the author discuss the required range of h′h'. I think should be upper bounded by something depending on dd, and this is important.

A: Thank you for pointing this out. In the revised version, we have explicitly summarized the required conditions on the step sizes in Theorem 5.3, which is h′≤cmin⁡{d−9T−1log⁡−1(d),d−4δ2,d−3δ3}h'\le c\min\{d^{-9}T^{-1}\log^{-1}(d),d^{-4}\delta^2,d^{-3}\delta^3\}. In particular, this leads to ϵ≤min⁡{d−5/2log⁡1/2d,δ1/2d−1/2}\epsilon\le\min\{d^{-5/2}\log^{1/2}d,\delta^{1/2} d^{-1/2}\} in Corollary 5.4 to make sure the complexity bound is valid.

As this is not the first convergence proof on Riemannian diffusion model, I think it's better to avoid using oo and OO notation through the proof, and only use it when it comes to the conclusion/final order. Moreover, I noticed that a universal constant C was used throughout the proof. I think it's better to define constants separately and track them (I feel this might make the proof easier to follow).

A: Thank you for your suggestion. In the revised version, we use separately numbered constants (e.g., C1,C2C_1,C_2) within each section to make the derivations easier to follow. At the end of each section, we summarize the corresponding results using a universal constant CC, making the main conclusions of each separate section clearer. Moreover, we now avoid using oo and OO notations during the proof, particularly for the derivation of bounding u1,∇u1u_1,\nabla u_1 and Δu1\Delta u_1. In the revised version, oo and OO are only used for a brief view of the bounds. We hope these revisions address your concerns.

Line 448, "for sufficiently small ϵ>0\epsilon>0" I think this should be stated explicit.

A: Thank you for your suggestion. In the revised version, we have explicitly stated the conditions on ϵ\epsilon required to ensure that the corresponding step size h′h' is sufficiently small. In our revised Corollary 5.4, the condition for the (simplified) score-evaluation complexity is ϵ≤d−1T−1\epsilon\le d^{-1}T^{-1}, and the condition for GRW complexity is ϵ≤min⁡{d−5/2log⁡1/2d,δ1/2d−1/2}\epsilon\le\min\{d^{-5/2}\log^{1/2}d,\delta^{1/2} d^{-1/2}\}.

possible typo: line 1228, clipped or truncated?

A: It should be "truncated". Thank you for pointing out this typo, and it is now revised.


We hope our response addresses your concerns. If so, we wonder if you could kindly consider raising your score? We will also be happy to answer any further questions you may have. Thank you very much!

Reviewer_PQBGOpenReview ↗

Response

Thank the author for the response. I think there might be some other typos, and I strongly encourage the author to check them. I think some typos, after correction, would also affect other parts of the proof so a more careful check is needed. I'm willing to raise my score if these can be addressed.

Line 1522, the equation, if I expect the integral of P^loc\hat{P}_{loc} to be 1, should there be an indicator function?

Line 1557, the equation, should it be 1−P(resample)1-P(resample) or P(resample)P(resample)?

The equation around line 1610, if I take t=0t=0, I get 0=10=1.

Line 1603, I think the author plug in u0u_0 but the sign might be a typo? I mean maybe it should be something like exp⁡(...+∫01<>ds)\exp(... + \int_0^1 <> ds).

Line 1708, the expression of h′h', the d9d^9 seems to be a typo. (also the log⁡d\log d term?)

Thank you very much for your suggestions. We have corrected the typos you pointed out and also carefully checked the remaining parts of the proof, making further corrections to typos in Appendices B and D. We hope these revisions address your concerns.

Reviewer_dJcEOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper improves the sampling complexity upper bound of score-based diffusion models on Riemannian compact manifolds. In particular the authors decompose the target distribution error in KL-divergence in an error due the discretization of the reverse SDE drift and an error due to the simulation of Brownian motion by a one-step Gaussian. For the former error they show (under non-negativity of Ricci curvature) a linear dependence on the dimension d, thereby improving a previous polynomial bound (Theorem 4.3). Furthermore, the authors suggest replacing the simulation of Brownian motion by one-step Gaussians via multi-step simulation of another SDE with a modified drift term and use it in the reverse SDE for more accurate sampling (Algorithm 1). They also provide an error bound on this new simulation which is of order O(d4)O(d^4) in dimension (Theorem 5.3).

In all the reviewer has found the paper's contributions understandable and the exposition easy to read. However, the reviewer thinks that the considered problem is not sufficiently motivated, both from theory and practice.

Critical strengths
  • Idea that the sampling error can be decomposed in two parts: discretization of the reverse SDE drift and simulation of Brownian motion and that both parts can be improved upon independently, e.g. by using a more accurate Brownian motion oracle which is cheaper than neural network score evaluation.
  • New analysis technique (see Appendix A) that allows to show the linear dependence of the first error via two new inequalities on the expected value of the squared norm on the score
  • Matching asymptotic complexity on the dimension as for the Euclidean case
Critical weaknesses
  • The main weakness is motivation why polynomial or linear dependence on dimension is important here. While polynomial input-dependence is an important class in algorithmic complexity theory, it is not clear what special mathematical role do polynomials or linear functions get in the sampling discretization error problem class, in particular the w.r.t. the space/manifold dimension. Moreover, the quantity log⁡(T/δ)\log(T/\delta) is not motivated or explained (i.e. why should the bound get worse as δ→0\delta \to 0). Unfortunately the authors do not motivate this question strongly enough, either from applications or theory (as neither do many previous papers on this topic).
  • No lower bounds, particular examples or numerical experiments are provided to show that the bound derived is not vacuous and can actually be observed in practice. Thus it is unclear how far from being optimal the derived bound on the first error is (e.g. is dimension-free possible?).
  • No numerical experiments or concrete (analytical) examples are provided to show that the proposed improved Brownian motion oracle actually improves the sampling. Such an example (even a toy one) would greatly improve the motivation and exposition for this paper.
  • The paper assumes that the score can already be approximated within some fixed accuracy (w.r.t. the L2L^2-norm), but the main difficulty is exactly this approximation. It could be the case that this approximation takes data, training time, or model size in quantities that are superpolynomial in dimension (although the reviewer is unfamiliar of such results in general. In particular this would depend clearly on the distribution), so that the whole process of diffusion sampling is superpolynomial.
  • The paper does not consider the intrinsic dimension of the distribution itself (e.g. what if p0p_0 is supported on a much lower dimensional submanifold?). In this case the reviewer suspects that the bound can be possibly improved.
Questions for authors

See weaknesses, in particular about motivation and lower bounds/optimality

4 comments

Thank you for your time and effort in reviewing our paper! We are grateful for your constructive suggestions, which have significantly guided our improvements. Please find our responses to your comments below.

The main weakness is motivation why polynomial or linear dependence on dimension is important here. While polynomial input-dependence is an important class in algorithmic complexity theory, it is not clear what special mathematical role do polynomials or linear functions get in the sampling discretization error problem class, in particular the w.r.t. the space/manifold dimension. Moreover, the log⁡(T/δ)\log(T/\delta) quantity is not motivated or explained (i.e. why should the bound get worse as δ→0\delta\to0). Unfortunately the authors do not motivate this question strongly enough, either from applications or theory (as neither do many previous papers on this topic).

A: The dependence on the dimension dd is important because the underlying data space can be high-dimensional in practical applications. In particular, understanding how the sampling complexity scales with dd is crucial for assessing the computational efficiency and scalability of diffusion models. This issue has also received considerable attention in the literature for Euclidean diffusion models (e.g., Benton et al., 2024; Li and Yan, 2025).

In this work, our goal is to establish sharper dimension dependence for Riemannian diffusion sampling. Specifically, we improve the existing complexity bounds and obtain a sharper dependence on dd for the number of score evaluations under nonnegative Ricci curvature.

Regarding the parameter δ\delta, it represents the early-stopping time in the reverse diffusion process. Early stopping is commonly adopted in diffusion models because the score function ∇log⁡pt(⋅)\nabla\log p_t(\cdot) can become unbounded as t→0t\to 0, making the final stage of reverse diffusion particularly challenging to analyze and discretize. Moreover, logarithmic dependence on the inverse early-stopping time (i.e., log⁡(1/δ)\log(1/\delta)) frequently appears in existing convergence analyses of diffusion models.

No lower bounds, particular examples or numerical experiments are provided to show that the bound derived is not vacuous and can actually be observed in practice. Thus it is unclear how far from being optimal the derived bound on the first error is (e.g. is dimension-free possible?).

A: The primary goal of this work is to improve the theoretical sampling complexity of Riemannian diffusion models. For sufficiently small target error ϵ>0\epsilon>0, we establish a O~(ϵ−2d)\tilde{O}(\epsilon^{-2}d) complexity bound for the number of score evaluations (Theorem 4.3), together with an O~(ϵ−2(Td4+δ−2d2))\tilde{O}(\epsilon^{-2}(Td^4+\delta^{-2}d^2)) omplexity bound for the number of GRW steps (Theorem 5.3).

Regarding optimality of overall error, although it remains unclear whether its dependence on step size hh can be further improved (e.g., by improving the algorithm), a linear dependence on dd in the KL divergence is known to be generally unavoidable. This can be seen by considering a product product data distribution pdata⊗np_{\text{data}}^{\otimes n} on the product manifold Mn\mathcal{M}^n. Thus, the nearly linear dependence on dd established in Theorem 4.3 is consistent with this fundamental dimensional scaling. For the GRW complexity, to the best of our knowledge, no corresponding lower bound is currently available. In particular, developing more efficient algorithms for simulating Riemannian Brownian motion remains an important open problem. We therefore view the GRW complexity bound as a theoretical upper bound rather than claiming its optimality. Still, we believe that this would be an interesting question, and we will leave it to future work.

No numerical experiments or concrete (analytical) examples are provided to show that the proposed improved Brownian motion oracle actually improves the sampling. Such an example (even a toy one) would greatly improve the motivation and exposition for this paper.

A: We appreciate this suggestion. We would like to first clarify that the main contribution of this work is theoretical: we develop a framework for analyzing Riemannian diffusion sampling and establish improved complexity bounds for score evaluations and Brownian motion simulations. Our results concern the generation-time complexity of the sampling process, rather than the practical implementation or training of specific neural network architectures. Indeed, numerical experiments are not provided in many existing works, particularly those focusing on theoretical analyses (e.g., Benton et al., 2024; Xu et al., 2026; Chen et al., 2026).

A numerical evaluation would require additional algorithmic and experimental considerations, including the choice of network architectures and training procedures, and practical implementations of the Brownian motion simulation oracle. These aspects are beyond the scope of the present theoretical study. Nevertheless, we agree that numerical investigations would be valuable for understanding the practical implications of our theoretical results and leave this direction for future work.

The paper assumes that the score can already be approximated within some fixed accuracy (w.r.t. the L2L_2-norm), but the main difficulty is exactly this approximation. It could be the case that this approximation takes data, training time, or model size in quantities that are superpolynomial in dimension (although the reviewer is unfamiliar of such results in general. In particular this would depend clearly on the distribution), so that the whole process of diffusion sampling is superpolynomial.

A: This is a standard assumption in the theoretical analysis of diffusion models, where the output of the score estimator (i.e., the neural network) is assumed to be L2L_2-close to the ground-truth score (see, e.g., Benton et al. 2024; Xu et al., 2026; Chen et al. 2026). In particular, in Theorem 4.3 and many existing convergence analyses of diffusion models, the score estimator is often treated as a general function, without imposing any specific structure on it. Finally, the training time is independent with our analysis, as our work focuses on the generation-time complexity.

The paper does not consider the intrinsic dimension of the distribution itself (e.g. what if p0p_0 is supported on a much lower dimensional submanifold?). In this case the reviewer suspects that the bound can be possibly improved.

A: Thank you for the suggestion! We agree that exploiting the intrinsic dimension of the data distribution could potentially lead to improved complexity bounds. The primary goal of our work is to establish improved sampling complexity guarantees for Riemannian diffusion models under general assumptions on the data distribution, without requiring additional structural assumptions on its support.

In particular, our analysis allows the initial data distribution dd to be arbitrary on the ambient manifold M\mathcal{M}, and does not require it to be supported on the entire manifold. Developing a refined analysis that explicitly exploits the potentially lower intrinsic dimension of the data distribution is an interesting direction for future research.


We hope our response addresses your concerns. If so, we wonder if you could kindly consider raising your score? We will also be happy to answer any further questions you may have. Thank you very much!

References

Joe Benton, Valentin De Bortoli, Arnaud Doucet, and George Deligiannidis. Nearly d-linear convergence bounds for diffusion models via stochastic localization. In International Conference on Learning Representations, 2024.

Gen Li and Yuling Yan. O(d/t)O(d/t) convergence theory for diffusion probabilistic models under minimal assumptions. In The Thirteenth International Conference on Learning Representations, 2025.

Xingyu Xu, Ziyi Zhang, Yorie Nakahira, Guannan Qu, and Yuejie Chi. Polynomial convergence of Riemannian diffusion models. In The Fourteenth International Conference on Learning Representations, 2026.

Fan Chen, Sinho Chewi, Constantinos Daskalakis, and Alexander Rakhlin. High-accuracy sampling for diffusion models and log-concave distributions, 2026.

Dear Reviewer,

We greatly appreciate the effort and expertise you have contributed to reviewing our paper. Since the author reviewer discussion period is ending soon, we hope that our responses have addressed your concerns. If that is the case, we wonder if you could kindly consider raising your score rating? Should you have any additional questions, we are more than happy to provide further assistance. Thank you very much for the support!

Reviewer_dJcEOpenReview ↗

Change of Rating

The authors have cleared up the reviewer's concern on purpose and relevance of their work. I decide to increase the rating to 3 (Weak acceptance).

Reviewer_yEB3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies sampling guarantees for diffusion models on compact Riemannian manifolds. The main idea is to decouple score evaluations from simulating the drifted Brownian motion, so several geodesic random-walk steps can be taken between two score evaluations. Under nonnegative Ricci curvature, the authors get a better bound on the number of score evaluations and an explicit bound on the number of random-walk steps. I think this is a good question to ask, since score evaluations are the expensive part. I have two main concerns. First, the nearly linear dependence on dd only holds once the target error is small enough. Second, I don’t think the derivative estimates behind the random-walk guarantee are right as written.

Critical strengths
  • Counting score evaluations and simulation steps separately makes it clear where the computational cost actually goes.
  • Using Ric≥0\mathrm{Ric} \ge 0 to control the integrated squared Hessian of log⁡pt\log p_t is a nice idea, and it does sharpen earlier estimates.
  • The analysis permits arbitrary initial data distributions and uses an expected squared score-error assumption, without requiring a smooth, strictly positive data density.
Critical weaknesses

mathematical Concerns In Appendix D.2, page 26, lines 1374–1377, substituting the score bound drops powers of the drift magnitude. This can be checked locally on a flat circle, where the extended drift is a constant bb. The parametrix coefficients then satisfy $\tau(x,y)=e^{b(x-y)}, \ u_1(x,y)=-\frac{b^2}{2}e^{b(x-y)}, $ and hence $ |\Delta_x u_1(y,y)|=\frac{b^4}{2}. $ Assumption 5.2 permits bb of order 1/δk1/\delta_k, where δk=T−tk\delta_k=T-t_k. The left-hand side can therefore grow as δk−4\delta_k^{-4}, whereas line 1377 gives only O(δk−3)O(\delta_k^{-3}) for fixed dimension and geometry. The subsequent bound on Bu1Bu_1 in line 1403 has a similar problem. These estimates enter the remainder analysis for Theorem 5.3. These bounds need correction.

The main dimension dependence is not a uniform finite-accuracy guarantee and only holds at high accuracy. Writing L=log⁡(T/δ)L=\log(T/\delta), the full score-evaluation bound is $ N=O\left(\frac{dL^2}{\epsilon^2} +\frac{d^2TL}{\epsilon} +\frac{d^{7/3}TL}{\epsilon^{2/3}}\right). $ Reducing this to the first term requires ϵ\epsilon to be sufficiently small relative to dimension and the time horizon. The paper acknowledges this, but does not quantify the regime. Similarly, the random-walk proof uses dimension-dependent localization conditions and discards higher-order terms without making the resulting step-size threshold explicit. The random-walk step-size cap h′≲d−9T−1h′ \lesssim d^{-9}T^{-1} forces at least about d9Td^9 T random-walk steps at any accuracy. The abstract and Table 1 omit these conditions, so the comparison with Benton et al., whose rate holds for every ϵ\epsilon, is overstated.

Limited Novelty The comparison with Euclidean sampling theory should include Chen et al. (2026), who obtain polylogarithmic dependence on inverse accuracy using approximate score evaluations. Intrinsic dimension-dependent rates are covered by Chen et al, for which differentiable compact manifolds are a special case. Also, the analysis builds on Xu et al.’s parametrix approach and Benton et al.’s KL argument. The authors should clarify the scope of this comparison and explain which geometric obstacles prevent a similar rejection-based approach.

References

Chen, Fan, Sinho Chewi, Constantinos Daskalakis, and Alexander Rakhlin. "High-accuracy sampling for diffusion models and log-concave distributions." arXiv preprint arXiv:2602.01338 (2026).

Questions for authors
  1. Can you fix the derivative and remainder estimates in Appendix D and show whether Theorem 5.3 retains its stated dependence on dimension and early stopping?
  2. What explicit conditions on ϵ\epsilon, dd, TT, and δ\delta justify the simplified complexities, including all localization restrictions?
  3. What sampling guarantee and complexity can be stated when δ\delta must decrease below order d−2d^{-2}?
6 comments

Thank you for your time and effort in reviewing our paper! We are grateful for your constructive suggestions, which have significantly guided our improvements. Please find our responses to your comments below.

Q1: Can you correct the derivative and remainder estimates in Appendix D and show whether Theorem 5.3 retains its stated dependence on dimension and early stopping? (Also Weakness 1.)

A: Thank you for pointing this out. We have corrected the derivative and remainder estimates in Appendix D, and the revised bounds are summarized in Lines 1388–1403. The error bound in Theorem 5.3 remains unchanged, except that we have relaxed the assumption on DD from D=O(1)D=O(1) to D=O(d)D=O(\sqrt{d}) to accommodate its potential dependence on the dimension dd (see Line 340). Here, this relaxation introduces a slightly larger dependence on δ\delta.

Q2: What explicit conditions on ϵ\epsilon, dd, TT, and δ\delta justify the simplified complexities, including all localization restrictions? (Also weakness 2.)

A: We have revised the discussion of our results to explicitly state the conditions on ϵ\epsilon, as presented in Corollary 5.4.

For the number of score evaluations, the condition ϵ<d−1T−1\epsilon < d^{-1}T^{-1} is required only to simplify the resulting complexity bound. For the number of GRW steps, we require ϵ≤min⁡{d−5/2log⁡1/2d,δ1/2d−1/2}\epsilon\le\min\{d^{-5/2}\log^{1/2}d,\delta^{1/2} d^{-1/2}\} to obtain the complexity N′=O(ϵ−2[d4(T+log⁡(1/δ))+δ−2d2]log⁡(T/δ)),N'=O\left(\epsilon^{-2}\left[d^4(T+\log(1/\delta))+\delta^{-2}d^2\right]\log(T/\delta)\right), where the condition comes from the restriction on step sizes in the analysis.

Q3: What sampling guarantee and complexity can be stated when δ\delta must decrease below order d−2d^{-2}?

A: We have revised our results to explicitly retain the dependence on the early-stopping parameter δ\delta. In particular, Theorem 5.3 and Corollary 5.4 now provide a complexity bound of N′=O(ϵ−2[d4(T+log⁡(1/δ))+δ−2d2]log⁡(T/δ))N'=O\left(\epsilon^{-2}\left[d^4(T+\log(1/\delta))+\delta^{-2}d^2\right]\log(T/\delta)\right) on the number of GRW steps.

Weakness 3: (Novelty.) The comparison with Euclidean sampling theory should include Chen et al. (2026), who obtain polylogarithmic dependence on inverse accuracy using approximate score evaluations. Intrinsic dimension-dependent rates are covered by Chen et al, for which differentiable compact manifolds are a special case. The authors should clarify the scope of this comparison and explain which geometric obstacles prevent a similar rejection-based approach.

A: Thank you for your suggestion. We have incorporated Chen et al. (2026) into the revised literature review. We would also like to clarify the distinctions between their work and ours.

First, their method addresses sampling in the Euclidean setting. In particular, even when the underlying data distribution is supported on a known manifold, the generated samples are not necessarily guaranteed to lie exactly on that manifold. In contrast, the Riemannian diffusion models considered in our work perform the entire generation process intrinsically on the manifold. Second, Chen et al. (2026) propose an algorithmic improvement to achieve improved sampling complexity. In contrast, our work focuses primarily on improving the theoretical understanding of Riemannian diffusion model, particularly by establishing sharper bounds on the number of score evaluations under nonnegative Ricci curvature. The resulting trade-off between score evaluations and BM simulation steps emerges naturally from our analysis.

Regarding the geometric obstacles to extending the rejection-based approach of Chen et al. (2026), a fundamental difficulty is that the transition kernel of Riemannian Brownian motion generally does not admit a closed-form expression.

To elaborate, in the Euclidean setting, their idea is to sample from the from the reverse conditional distribution ρk(x∣x′)∝xpk(x)exp⁡(−∣x−αk−1x′∣22ηk).\rho_k(x|x')\propto_xp_k(x)\exp\left(-\frac{|x-\alpha_k^{-1}x'|^2}{2\eta_k}\right). By applying a first-order expansion to log⁡pk(x)\log p_k(x), they find it convenient to sample from q(x)∝exp⁡(−⟨x−x+,∇log⁡pk(x+)⟩−∣x−αk−1x′∣22ηk),q(x)\propto\exp\left(-\langle x-x_+,\nabla\log p_k(x_+)\rangle-\frac{|x-\alpha_k^{-1}x'|^2}{2\eta_k}\right), so that a rejection sampling is performed on these samples. Note that this is a Gaussian distribution with a nonzero mean and can therefore be sampled from efficiently.

However, in the Riemannian setting, the corresponding reverse conditional distribution takes the form ρk(x∣x′)∝xpk(x)P(hk,x′,x)\rho_k(x|x')\propto_xp_k(x)P(h_k,x',x) where PP denotes the transition kernel of Riemannian Brownian motion. Applying an analogous first-order expansion to and proceed similarly would lead to the distribution q(x)∝exp⁡(−⟨log⁡x+(x),∇log⁡pk(x+)⟩)⋅P(hk,x′,x).q(x)\propto\exp\left(-\langle\log _{x_+}(x),\nabla\log p_k(x_+)\rangle\right)\cdot P(h_k,x',x).Unlike the Euclidean case, this distribution does not have an immediately tractable form, and its efficient sampling is not straightforward. To the best of our knowledge, overcoming this difficulty remains an open challenge.


We hope our response addresses your concerns. If so, we wonder if you could kindly consider raising your score? We will also be happy to answer any further questions you may have. Thank you very much!

Dear Reviewer,

We greatly appreciate the effort and expertise you have contributed to reviewing our paper. Since the author reviewer discussion period is ending soon, we hope that our responses have addressed your concerns. If that is the case, we wonder if you could kindly consider raising your score rating? Should you have any additional questions, we are more than happy to provide further assistance. Thank you very much for the support!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.