acceptodds
Under review as a conference paper at ICLR 2027

Simulation-Free Schrödinger Bridges on Riemannian Manifolds

12 views

Abstract

Schrödinger bridges provide a principled framework for stochastic transport between distributions. Extending simulation-free Schrödinger bridge learning to Riemannian manifolds is challenging because neither Brownian transition kernels nor conditional bridge distributions are generally available in closed form. We introduce SSR-FM, a simulation-free framework for approximating Schrödinger bridges on Riemannian manifolds through flow and score matching. Using a small-noise heat-kernel expansion, SSR-FM constructs a geometry-aware approximation to the endpoint transport cost and obtains the corresponding entropic coupling by Sinkhorn. Conditional bridge marginals are then approximated by Gaussian fluctuations around minimizing geodesics, yielding tractable flow and score targets without simulating diffusion paths during training. Under regularity assumptions, we show that the reconstructed SDE reproduces the marginal distributions of the resulting approximate probability path. Across benchmarks with analytically specified and data-estimated geometries, SSR-FM preserves manifold structure while achieving competitive hidden-time distribution matching among the existing methods.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Schrödinger bridges, Riemannian manifolds, flow matching, score matching, entropic optimal transport

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_S8BqOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to learn Schrodinger bridges between distributions on Riemannian manifolds without simulating diffusion trajectories during training. The main difficulty is that Brownian transition densities and conditional bridge distributions generally do not have closed form expressions on Riemannian manifolds. To address this, the paper proposes SSR-FM, which uses an approximation of the Brownian transition density to construct a manifold transport cost. It then approximates conditional Brownian bridges using Gaussian distributions on minimizing geodesics, with covariance determined by the Hessian of the Brownian action. The authors theoretically justify these approximations when noise is small and show that the learned SDE recovers the marginals of the constructed probability path under ideal training and regularity assumptions. They evaluate SSR-FM on synthetic curved surfaces, geographic and molecular datasets on spheres and tori, and single cell datasets where the underlying data manifold is not known. SSR-FM generally achieves good results, although improvements over the strongest baselines are relatively small in several cases. The paper also provides ablations on the geometric covariance and endpoint coupling and shows substantially faster training than RDSB.

Critical strengths
  1. The paper proposes a framework for extending simulation-free Schrödinger bridge learning to Riemannian manifolds, where Brownian transition densities and conditional bridge distributions generally do not have closed-form expressions. The authors derive geometry aware transport costs and conditional Gaussian paths using approximations with small noise, and provide theoretical results that justify these constructions and show that the learned SDE recovers the constructed probability path under ideal conditions. These results provide theoretical motivation for the proposed method that is specific to Riemannian geometry.
  2. The method achieves competitive distribution matching while preserving manifold geometry. The authors also provide runtime comparisons showing that their method trains around 7 times faster than RDSB under similar update budgets.
Critical weaknesses
  1. The biggest weakness of the paper is that the improvements over existing methods are very marginal on most real-world datasets. Many of the absolute improvements are on the order of 10^-2 or 10^-3, while the W1 values themselves are on the order of 10^0 to 10^2, meaning that the relative improvements are often only a few tenths or hundredths of a percent. Also, the improvements are also often within standard deviation. For example, on NOAA, SSR-FM achieves 101.228±0.588101.228\pm0.588 compared to 101.273±0.103101.273\pm0.103 for RFM (0.04% improvement), and on IBTrACS, it achieves 119.722±0.425119.722\pm0.425 compared to 119.842±0.240119.842 \pm0.240 for SF²M (0.10% improvement). The improvements in the single-cell experiments are similarly marginal. These results do not provide sufficient evidence that the proposed method offers meaningful advantages over existing methods, especially considering the additional geometric computations required over Riemannian flow matching methods.
  2. The paper could also benefit from testing on more complicated manifolds than the torus and sphere, such as SO(n) or U(n).
Questions for authors

See weakness

5 comments
AuthorsOpenReview ↗

Response to reviewer S8Bq(1/2)

We sincerely thank the reviewer for the careful assessment and constructive suggestions. We have expanded the evaluation to examine distributional accuracy, geometric adherence, and computational cost, and have added experiments on SO(n)SO(n), U(n)U(n), and isometric embeddings. All additions and substantive revisions are highlighted in blue.

W1

Weakness 1. We appreciate the reviewer's concern about the practical gains. Among our baselines, only RDSB [1] targets the same Riemannian Schrödinger bridge problem, relying on iterative diffusion simulation during training, while SSR-FM addresses it by directly regressing the velocity and score fields. On the four real-world manifold tasks, SSR-FM (SDE) reduces W1W_1 by 19.6%–75.6% relative to RDSB; it also achieves lower W1W_1 on all six single-cell tasks, with a mean per-task relative reduction of 14.63% (Tables 3 and 4). Field training is also 7.08 times faster in the 200k-update EB100D benchmark (Appendix G.1). These results demonstrate substantial gains in both accuracy and training efficiency over an algorithm addressing the same class of Riemannian Schrödinger bridge problems.

Compared with SF²M, SSR-FM explicitly incorporates the manifold geometry into its reference diffusion and conditional paths. SF²M uses Euclidean Brownian bridge conditionals, which do not enforce manifold constraints and therefore do not guarantee that generated samples remain on the manifold. This lack of geometric control is one of the motivations for our work. In the original submission, we computed W1W_1 in the common ambient Euclidean space to ensure a fair and transparent comparison across methods. Similar ambient W1W_1 values may mask substantial off-manifold deviations in SF²M samples, so W1W_1 alone does not fully capture the geometric advantages of SSR-FM. For example, under ODE sampling, SSR-FM reduces the mean distance to the manifold from 8.273 to 0.020 km on IBTrACS and from 6.744 to 0.004 km on NOAA. Under SDE sampling, the corresponding errors on ALA2 and Chignolin decrease from 0.074 and 0.763 to 0.000 at the reported precision (Table 3). Across all reported experiments where distance to a prescribed manifold can be measured, SSR-FM substantially reduces this error compared with SF²M (Tables 1–3; Appendix C.4).

Compared with RFM, SSR-FM has lower W1W_1 on most evaluated tasks, while RFM's performance relative to other baselines is mixed across datasets (Tables 1–4). On the six single-cell tasks, SSR-FM achieves mean per-task relative W1W_1 reductions of 1.75% (ODE) and 2.72% (SDE) over RFM, with both samplers outperforming RFM on four of the six tasks (Table 4). RFM remains competitive on NOAA. Similar behavior was reported in the original SF²M study [2], where SF²M and OT-CFM perform comparably on some datasets. Introducing stochasticity may provide limited gains for data dominated by deterministic motion. When biological processes exhibit substantial variability, modeling stochastic fluctuations can be important. Our biological benchmarks include clearer improvements over RFM: SSR-FM (SDE) reduces W1W_1 by 11.2% on EB (10D) and 15.7% on Chignolin (Tables 3 and 4), supporting its value for modeling biological populations and molecular configurations.

Overall, we derive a simulation-free algorithm for approximating Schrödinger bridges on Riemannian manifolds from a common Riemannian Brownian reference, providing a capability missing from the prior methods compared in Appendix A.

AuthorsOpenReview ↗

Response to reviewer S8Bq(2/2)

W2

Weakness 2. Following the reviewer's suggestion, we have added experiments on SO(n)SO(n), U(n)U(n), and isometric embeddings. Table R9 reports the full comparison. These results are also presented in Section 6.2 of the revised manuscript (Figure 2 and Table 2), with experimental details in Appendix C.4 and additional trajectory visualizations in Appendix H.

Table R9. Hidden-time W1W_1 errors (↓\downarrow), averaged over t=0.25,0.5,0.75t=0.25,0.5,0.75.

Manifold familySO(n)SO(n)U(n)U(n)Isometric embeddings
ManifoldSO(5)SO(5)SO(10)SO(10)SO(15)SO(15)SO(32)SO(32)U(5)U(5)U(10)U(10)U(15)U(15)U(32)U(32)SO(14)SO(14)SO(14)SO(14)SO(14)SO(14)
Intrinsic dimension1045105496251002251024919191
Ambient dimension251002251024502004502048196100010000
OT-CFM0.5451.0861.4442.0360.8051.5862.0372.9021.2961.3051.365
SF²M0.5621.0751.4572.3940.8221.5962.0543.1001.3041.3491.392
Metric FM0.5371.1121.5052.0740.8541.6112.2943.1821.3271.3241.387
RFM0.374‾\underline{0.374}0.716‾\underline{0.716}1.122‾\underline{1.122}1.632‾\underline{1.632}0.571‾\underline{0.571}1.255‾\underline{1.255}1.658‾\underline{1.658}2.731‾\underline{2.731}1.001‾\underline{1.001}0.977‾\underline{0.977}0.877‾\underline{0.877}
RDSB0.6351.4102.0534.8791.4252.2253.1024.5781.7252.2322.849
SSR-FM0.1160.2330.2710.3360.1920.2680.2950.3530.2630.2640.322

SSR-FM achieves the lowest W1W_1 among all compared methods across all eleven settings in Table R9, with an 87.1% reduction relative to RFM on U(32)U(32). These results demonstrate a more pronounced advantage on complex, high-dimensional manifolds. The trajectory visualizations further show that SSR-FM more closely matches the reference process in its dominant transport branches and distributional spread (Figure 2 and Appendix H of the revised manuscript). We also evaluate isometric embeddings of SO(14)SO(14) in ambient dimensions 196, 1,000, and 10,000 while keeping its intrinsic dimension fixed at 91, with SSR-FM achieving the lowest W1W_1 in all three settings (Table R9; Table 2 in Section 6.2 of the revised manuscript).

Thank you for the constructive comments, which helped us present the evidence and limitations more clearly. We hope the revised paper better communicates the theoretical and algorithmic novelty of SSR-FM. We welcome additional questions or suggestions and will do our best to reply promptly after receiving them.

[1] Thornton, J., Hutchinson, M., Mathieu, E., De Bortoli, V., Teh, Y. W., & Doucet, A. (2022). Riemannian Diffusion Schrödinger Bridge. arXiv:2207.03024.

[2] Tong, A. Y., Malkin, N., Fatras, K., Atanackovic, L., Zhang, Y., Huguet, G., Wolf, G., & Bengio, Y. (2024). Simulation-Free Schrödinger Bridges via Score and Flow Matching. AISTATS, PMLR 238:1279–1287.

AuthorsOpenReview ↗

Brief follow-up to Reviewer S8Bq

Dear Reviewer S8Bq,

Before the discussion period ends on November 18, we would like to briefly follow up on your concerns about the practical advantages of our method. Our response discusses distributional accuracy, geometric adherence, and computational cost, alongside the newly added SO(n) and U(n) experiments. We would appreciate hearing whether this evidence addresses your concerns or whether further explanation would be useful. Thank you for your time and careful consideration.

Reviewer_NtLzOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a framework for computing Schrodinger bridges on Riemannian manifolds. The bridge is defined to be the optimal coupling with respect to KL divergence to a refernce measure: this is typically an ambient Brownian motion but, on a manifold, involves estimating the Riemannian heat kernel. The authors give an approximate cost using the Riemannian metric of the ambient coordinates, and show that this leads to approximate Schrodinger bridges without needing to simulate paths at training time. The method is evaluated with both closed-form and estimated Riemannian metrics, where it shows promising results. Overall, the paper is clearly motivated, nice to read, and well explained. The experiments are effective, but could be more extensive and would benefit from evaluation with recent developments in Riemannian metric learning.

The paper is an interesting algorithmic contribution. I would be willing to increase my score if my questions are addressed.

Critical strengths
  1. The problem is well motivated and explained. The writing is clear and effective. The mathematical explanations are direct and present the ideas straightforwardly. I really like this work!
Critical weaknesses

While the experiments are nice, they are limited in scope and would benefit from more extensive testing. I appreciate that could benchmarks are hard to come by for this kind of problem, but I think the work could be improved in two ways:

  1. The synthetic example is just a surface in 3d - it would be possible to explore more complicated geometries in different intrinsic and ambient dimensions. With the key real-world example being single-cell data in O(10^4) ambient dimensions, it would be useful to see synthetic results that model this range. For instance, I would like to know how performance of these methods varies under changes in both intrinsic and ambient dimension.
  2. The data-estimated geometry only uses the LAND metric, which is not theoretically guaranteed and is diagonal (so not a good model for general data geometry). These results would benefit significantly from using more recent developments in low-rank metric learning, e.g. [1]. One could also use the Jacobian of a score-based denoiser [2], although this is full-rank so less scalable.

[1] Bamberger, J., Gosztolai, A., Vandergheynst, P., Bronstein, M., & Jones, I. (2026). Riemannian Metric Matching for Scalable Geometric Modeling of Distributions. [2] Kharitenko, A., Shen, Z., De Santi, R., He, N., & Dorfler, F. (2026, April). Landing with the score: Riemannian optimization through denoising.

Questions for authors
  1. Single cell data typically have branching points so are not manifolds in the classical sense. While metrics like LAND and [1,2,3] do not require a manifold, the diffusion process underlying the Schrodinger bridge is only well understood in this setting. Do you expect this to pose a problem with branching or variable-dimensional data?
6 comments
AuthorsOpenReview ↗

Response to reviewer NtLz (1/2)

We are grateful to receive your recognition of our work and sincerely appreciate your constructive feedback. Your suggestions on more complex manifolds, alternative metric estimators, and branching data have helped us broaden the evaluation and clarify the scope of the method. We have revised the manuscript in response, with additions and substantive changes highlighted in blue.

Q1

Question 1. We understand your question as asking whether branching or locally varying data dimension is compatible with our assumption that the dynamics evolve on a Riemannian manifold. For example, consider a biological process with “Y” shaped dynamics. Along an individual branch, observations may vary mainly in one direction, whereas several directions coexist near the junction. This can make a single smooth one-dimensional description inadequate there and affect local dimension estimates.

In our formulation, the cell populations concentrate near branching structures within a smooth ambient state space. More specifically, in our experiments, we construct a single spatially varying LAND metric G(x)G(x) from reference data including all visible training snapshots (Appendix C.6). Its Gaussian weights and positive regularization define a smooth positive definite metric throughout RD\mathbb R^D [1]. In this sense, all observed times share a globally defined Riemannian manifold (RD,G)(\mathbb R^D,G). The reference generator is L=(ε/2)ΔG\mathcal L=(\varepsilon/2)\Delta_G. Branching and variations in the apparent local dimension of the observations therefore do not, by themselves, make this ambient reference diffusion ill-defined. We therefore model a dynamical system evolving on a Riemannian manifold. The effective degrees of freedom of the observed dynamics can vary across regions and over time, while the process as a whole remains defined on the same manifold.

We hope we have correctly understood your concern. Please feel free to let us know if you had a different setting in mind or have further questions; we would be happy to discuss further.

W1

Weakness 1. We thank the reviewer for this helpful suggestion. We have added high-dimensional SO(n)SO(n) and U(n)U(n) benchmarks and isometric embeddings, with results in Section 6.2 (Figure 2 and Table 2), experimental details in Appendix C.4, and additional trajectory visualizations in Appendix H.

Table R6. Hidden-time W1W_1 errors (↓\downarrow), averaged over t=0.25,0.5,0.75t=0.25,0.5,0.75.

Manifold familySO(n)SO(n)U(n)U(n)Isometric embeddings
ManifoldSO(5)SO(5)SO(10)SO(10)SO(15)SO(15)SO(32)SO(32)U(5)U(5)U(10)U(10)U(15)U(15)U(32)U(32)SO(14)SO(14)SO(14)SO(14)SO(14)SO(14)
Intrinsic dimension1045105496251002251024919191
Ambient dimension251002251024502004502048196100010000
OT-CFM0.5451.0861.4442.0360.8051.5862.0372.9021.2961.3051.365
SF²M0.5621.0751.4572.3940.8221.5962.0543.1001.3041.3491.392
Metric FM0.5371.1121.5052.0740.8541.6112.2943.1821.3271.3241.387
RFM0.374‾\underline{0.374}0.716‾\underline{0.716}1.122‾\underline{1.122}1.632‾\underline{1.632}0.571‾\underline{0.571}1.255‾\underline{1.255}1.658‾\underline{1.658}2.731‾\underline{2.731}1.001‾\underline{1.001}0.977‾\underline{0.977}0.877‾\underline{0.877}
RDSB0.6351.4102.0534.8791.4252.2253.1024.5781.7252.2322.849
SSR-FM0.1160.2330.2710.3360.1920.2680.2950.3530.2630.2640.322

On the native groups, increasing nn increases both intrinsic and ambient dimension. SSR-FM achieves the lowest W1W_1 in all eight settings, up to intrinsic dimension 1,024 on U(32)U(32), where its error is 0.353 versus 2.731 for RFM, an 87.1% reduction.

The isometric embeddings separately test ambient dimension: the same SO(14)SO(14) manifold, with intrinsic dimension fixed at 91 and unchanged intrinsic geometry, is represented in 196, 1,000, and 10,000 coordinates. SSR-FM remains the most accurate method in all three representations; at 10,000 dimensions, its W1W_1 is 0.322 versus 0.877 for RFM. These results show that its advantage persists both as group dimension grows and as the same intrinsic problem is embedded in a larger ambient space. The trajectory visualizations also show closer agreement with the reference transport patterns (Figure 2 and Appendix H).

AuthorsOpenReview ↗

Response to reviewer NtLz (2/2)

W2

Weakness 2. We sincerely thank the reviewer for these valuable suggestions, which motivated us to investigate whether richer geometry could further improve SSR-FM. We incorporated ordinary Riemannian metric matching [2] and a metric based on denoiser Jacobians [3] into SSR-FM, evaluating five representations across all three single-cell datasets: EMT, EB, and CITE-seq (Appendix F.1–F.2).

Table R7. Single-cell results for SSR-FM. Each entry lists W1W_1 / W2W_2 (↓\downarrow), using one matched training seed and the same SDE sampling settings.

Metric estimatorEMT (10D)EB (10D)EB (50D)EB (100D)CITE (50D)
LAND2.1138 / 2.16991.5121 / 1.58376.2787 / 6.361710.3258 / 10.39426.0415 / 6.1710
Metric matching2.0740‾\underline{2.0740} / 2.1364‾\underline{2.1364}1.5536 / 1.64676.2823‾\underline{6.2823} / 6.3763‾\underline{6.3763}10.4810 / 10.59746.0596‾\underline{6.0596} / 6.1965‾\underline{6.1965}
Denoiser Jacobian2.0665 / 2.12941.5518‾\underline{1.5518} / 1.6421‾\underline{1.6421}6.3692 / 6.465010.4109‾\underline{10.4109} / 10.4844‾\underline{10.4844}6.0607 / 6.1973

Both alternatives improve EMT, but LAND gives the lowest W1W_1 and W2W_2 on the EB and CITE-seq settings. We therefore did not observe a consistent advantage over LAND under this recipe. Although Metric matching predicts its cometric with a single forward pass, our path solver repeatedly evaluates and spatially differentiates the metric, making direct neural use expensive. We therefore cache cometric matrices at selected training locations and smoothly interpolate them with positive weights. The resulting approximation, together with finite-sample estimation error, may offset the benefit of richer geometry.

Following your guidance, we further explored this direction on MNIST, inspired by the within-class interpolation experiments in [2]. We test unpaired slant interpolation for digits 1, 5, and 8 in a shared 8D VAE space: lower and upper slant quartiles define the endpoint styles, and test images with intermediate slant provide the distributional reference. Here, Metric matching is evaluated directly in latent space (Appendix F.3; Figure F1).

Table R8. Mean MNIST errors across digits 1, 5, and 8 at t=0.5t=0.5 (↓\downarrow), using one training seed.

MethodLatent W1W_1Latent W2W_2Pixel W1W_1Pixel W2W_2
OT-CFM1.34201.42005.0782‾\underline{5.0782}5.2354‾\underline{5.2354}
SF²M1.36941.50835.13755.3850
Metric FM1.2915‾\underline{1.2915}1.35615.16235.3367
Metric matching + RFM2.01252.21686.10836.3883
Metric matching + SSR-FM1.28321.3754‾\underline{1.3754}5.01315.2195

Metric matching + SSR-FM achieves the lowest latent W1W_1 and pixel W1/W2W_1/W_2, while Metric FM achieves the lowest latent W2W_2. With the same Metric matching geometry, SSR-FM also improves on RFM in all four errors. These findings suggest that the benefit of a metric estimator may depend on the data-generating structure and the geometry of its representation, as well as on estimation and approximation quality.

We sincerely appreciate your thoughtful suggestions, which have helped us strengthen both the experiments and the presentation of our work. We would be glad to continue the discussion and address any remaining questions, and will respond immediately after hearing from you.

[1] Arvanitidis, G., Hansen, L. K., & Hauberg, S. (2016). A Locally Adaptive Normal Distribution. Advances in Neural Information Processing Systems, 29.

[2] Bamberger, J., Gosztolai, A., Vandergheynst, P., Bronstein, M., & Jones, I. (2026). Riemannian Metric Matching for Scalable Geometric Modeling of Distributions. arXiv:2606.14334.

[3] Kharitenko, A., Shen, Z., De Santi, R., He, N., & Dörfler, F. (2026). Landing with the Score: Riemannian Optimization through Denoising. ICLR.

AuthorsOpenReview ↗

Brief follow-up to Reviewer NtLz

Dear Reviewer NtLz,

Thank you again for your thoughtful suggestions, which have helped us strengthen our work. We have added experiments with the metric estimators you recommended and more complex manifolds, and clarified our formulation for branching data. With the discussion deadline approaching on November 18, we would warmly welcome your thoughts on these additions and explanations. Should any questions remain, we would be happy to discuss them further and respond promptly.

Reviewer_6SKsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a simulation-free framework for approximating Schrödinger bridges to Riemannian manifolds by constructing a geometry-aware approximation for transport coupling and Gaussian conditional paths centered around minimizing geodesics. The method is evaluated in both synthetic and real-world data compared with relevant baselines.

Critical strengths

The paper presents an extension of simulation-free Schrödinger bridges to Riemannian manifolds and provides tools for incorporating geometry into endpoint coupling and conditional stochastic interpolation. The results show evidence that incorporating the geometry shows improved interpolation performance in the majority of the cases.

Critical weaknesses
  1. The comparison with SF²M-Geo, which also has geometry-aware coupling, is missing.
  2. Given that on several real-world datasets the improvement in performance is small, this omission leads to the advantage of the proposed method over this closely related baseline not being established. 3)While the baseline SF²M shows results on high-dimensional data (ex: 1000), the evidence of how the proposed method scales to those high dimensions is not shown.
  3. It is not clear how the geometric preprocessing time and performance of the proposed framework scale beyond the evaluated dimensions.
Questions for authors

The paper discusses SF²M-Geo but doesn’t compare against. Could the authors explain the reason or provide results comparing the proposed method to it?

7 comments
AuthorsOpenReview ↗

Response to reviewer 6SKs (1/3)

We thank the reviewer for recognizing our extension of simulation-free Schrödinger bridge learning to Riemannian manifolds and for the constructive suggestions on the SF²M-Geo comparison and scalability. We have revised the manuscript accordingly, adding the requested baseline comparisons, benchmarking on manifolds of higher dimension, and a preprocessing-time analysis. All additions and substantive revisions are highlighted in blue in the revised manuscript.

W1&Q1

Weakness 1 & Question 1. We agree that SF²M-Geo is a relevant baseline and apologize for omitting it. Following several single cell trajectory inference studies, we used standard SF²M as an established and representative baseline and overlooked its geometric variant. SF²M-Geo [1] computes a geometry-aware coupling from a graph approximation of the manifold heat kernel using Geodesic Sinkhorn, while retaining Euclidean Brownian bridge conditionals. This geometric adaptation is not accompanied by a derivation of the coupling and conditional paths from a common Riemannian Brownian reference, and therefore does not solve the corresponding Riemannian Schrödinger bridge problem. SSR-FM provides the theoretical connection through small-noise approximations derived from that reference.

Nevertheless, to further demonstrate the advantages of our method, we have added experiments on the rolled saddle and helicoid, NOAA and IBTrACS, ALA2 and Chignolin, and six single-cell settings. Following the original SF²M paper [1], we implement a graph-based SF²M-Geo adaptation on our datasets using the official TorchCFM conditional matcher and GeoSinkhorn heat-kernel routines. Tables R1–R3 report the results, which are also included in Appendix C.7.

Table R1. Hidden-time W1W_1 and distance-to-manifold errors on curved surfaces (↓\downarrow), using the data and evaluation settings described in the manuscript.

MethodRolled saddle W1W_1Rolled saddle DMD_{\mathcal M}Helicoid W1W_1Helicoid DMD_{\mathcal M}
SF²M-Geo (ODE)0.6650.6231.9550.023
SF²M-Geo (SDE)0.6760.6381.9410.130
SSR-FM (ODE)0.115‾\underline{0.115}0.0000.083‾\underline{0.083}0.000
SSR-FM (SDE)0.1090.0000.0600.000

Table R2. Mean held-out W1W_1 and distance-to-manifold errors on geographic and molecular datasets (↓\downarrow), averaged over five training seeds (0–4) for both methods.

MethodNOAA W1W_1NOAA DMD_{\mathcal M}IBTrACS W1W_1IBTrACS DMD_{\mathcal M}ALA2 W1W_1ALA2 DMD_{\mathcal M}Chignolin W1W_1Chignolin DMD_{\mathcal M}
SF²M-Geo (ODE)124.02412.533130.58116.6830.5080.1052.0921.007
SF²M-Geo (SDE)124.19113.104134.98628.9270.4800.0982.084‾\underline{2.084}0.970
SSR-FM (ODE)101.2280.004119.7220.020‾\underline{0.020}0.314‾\underline{0.314}0.0002.1290.000
SSR-FM (SDE)101.284‾\underline{101.284}0.006‾\underline{0.006}120.159‾\underline{120.159}0.0110.3080.0001.9850.000

Table R3. W1W_1 errors for single-cell population interpolation (↓\downarrow), averaged over five training seeds (0–4) for both methods.

MethodEMT (10D)EB (10D)EB (20D)EB (50D)EB (100D)CITE (50D)
SF²M-Geo (ODE)2.12321.56623.11486.362010.42336.1272
SF²M-Geo (SDE)2.13371.57643.10636.3153‾\underline{6.3153}10.3495‾\underline{10.3495}6.1615
SSR-FM (ODE)2.09771.5174‾\underline{1.5174}3.0991‾\underline{3.0991}6.507910.53035.9968
SSR-FM (SDE)2.1000‾\underline{2.1000}1.50913.08346.279510.33016.0400‾\underline{6.0400}

With SF²M-Geo included as an additional baseline, the results continue to show an overall advantage for SSR-FM in distribution prediction and geometric adherence (Tables R1–R3). Although SF²M-Geo adapts the transport coupling to the data geometry, its Euclidean Brownian bridge conditionals do not constrain the stochastic process to remain on the manifold. The larger distance-to-manifold errors in Tables R1–R2 reflect this limitation and highlight the geometric advantage of SSR-FM.

W2

Weakness 2. This concern is closely related to Weakness 1. We hope that the additional experiments above address the omission in our original submission and more clearly demonstrate the advantages of SSR-FM.

AuthorsOpenReview ↗

Response to reviewer 6SKs (2/3)

W2 (3)

Weakness 2 (3). We thank the reviewer for this helpful suggestion. One motivation for this work is the low-dimensional manifold hypothesis: high-dimensional observations can concentrate near structures with much lower intrinsic dimension. For example, Levina and Bickel [2] estimated an intrinsic dimension of 3.1 for hand images with 245,760 pixel coordinates, while Facco et al. [3] reported an effective intrinsic dimension of 13.4 for 784-dimensional MNIST images of the digit “2”. These empirical findings motivated our initial focus on low-dimensional synthetic manifolds.

Nevertheless, we have sought to follow the reviewer's suggestion as closely as possible and further demonstrate the effectiveness and scalability of SSR-FM. We have therefore added experiments that vary intrinsic and ambient dimension. The SO(n)SO(n) and U(n)U(n) benchmarks reach intrinsic dimension 1,024, while isometric embeddings of the same SO(14)SO(14) manifold increase the ambient dimension from 196 to 10,000 with its intrinsic dimension fixed at 91. These results are now included in Section 6.2 (Figure 2 and Table 2), with experimental details in Appendix C.4 and additional visualizations in Appendix H.

Table R4. Hidden-time W1W_1 errors(↓\downarrow), averaged over t=0.25,0.5,0.75t=0.25,0.5,0.75.

Manifold familySO(n)SO(n)U(n)U(n)Isometric embeddings
ManifoldSO(5)SO(5)SO(10)SO(10)SO(15)SO(15)SO(32)SO(32)U(5)U(5)U(10)U(10)U(15)U(15)U(32)U(32)SO(14)SO(14)SO(14)SO(14)SO(14)SO(14)
Intrinsic dimension1045105496251002251024919191
Ambient dimension251002251024502004502048196100010000
OT-CFM0.5451.0861.4442.0360.8051.5862.0372.9021.2961.3051.365
SF²M0.5621.0751.4572.3940.8221.5962.0543.1001.3041.3491.392
Metric FM0.5371.1121.5052.0740.8541.6112.2943.1821.3271.3241.387
RFM0.374‾\underline{0.374}0.716‾\underline{0.716}1.122‾\underline{1.122}1.632‾\underline{1.632}0.571‾\underline{0.571}1.255‾\underline{1.255}1.658‾\underline{1.658}2.731‾\underline{2.731}1.001‾\underline{1.001}0.977‾\underline{0.977}0.877‾\underline{0.877}
RDSB0.6351.4102.0534.8791.4252.2253.1024.5781.7252.2322.849
SSR-FM0.1160.2330.2710.3360.1920.2680.2950.3530.2630.2640.322

SSR-FM achieves the lowest W1W_1 in all eleven settings, reducing error by 63.3%–87.1% relative to RFM, the strongest baseline. Its advantage is maintained as the ambient dimension of the same SO(14)SO(14) manifold increases from 196 to 10,000.

AuthorsOpenReview ↗

Response to reviewer 6SKs (3/3)

W3

Weakness 3. We thank the reviewer for this suggestion. We have added a preprocessing-time study in Appendix G.2 (Figure G2), covering SO(n)SO(n) and U(n)U(n) with intrinsic dimensions up to 1,024. We benchmark preprocessing using 4,096 observations per endpoint via Sinkhorn, compared with exact OT. Table R5 reports the total preparation time for the different methods.

Table R5. Numerical preparation time in seconds. Both SSR-FM and the Euclidean baseline use the Sinkhorn algorithm; only the input cost computation differs.

ManifoldIntrinsic dimensionAmbient dimensionExact OTEuclidean SinkhornSSR-FM
SO(3)SO(3)3936.5900.2030.419
SO(5)SO(5)102531.1600.2122.093
SO(10)SO(10)4510030.8671.4455.711
SO(15)SO(15)10522532.3945.3838.138
SO(32)SO(32)496102433.75810.43213.668
U(3)U(3)91832.3080.2162.562
U(5)U(5)255030.7820.8003.811
U(10)U(10)10020032.1647.5267.682
U(15)U(15)22545033.4018.40310.339
U(32)U(32)1024204835.36420.56619.982

At SO(32)SO(32), SSR-FM takes 13.668 seconds, compared with 10.432 seconds for Euclidean Sinkhorn. At U(32)U(32), the corresponding times are 19.982 and 20.566 seconds. The absolute gap in preprocessing time is less than four seconds at these largest tested dimensions.

Algorithmically, once the pairwise costs are available, both methods use the same Sinkhorn solver. For NN observations per endpoint, minibatch size bb, and II iterations per batch, the solver requires O((N/b)Ib2)=O(INb)O((N/b)Ib^2)=O(INb) work. For fixed bb and II, this stage scales linearly with sample count and has the same complexity as Euclidean Sinkhorn, while coordinate dimension affects the preceding cost construction. This shared computational structure is consistent with the limited absolute timing gaps observed above.

Overall, the moderate growth in preparation time as dimension increases supports the practical scalability of SSR-FM across the tested high-dimensional manifolds.

We appreciate your careful assessment and the suggestions that helped us strengthen the comparisons and scalability evaluation. We hope these additions clarify the theoretical and algorithmic contribution. Please feel free to raise any further questions; we welcome continued discussion and will reply as soon as we can.

[1] Tong, A. Y., Malkin, N., Fatras, K., Atanackovic, L., Zhang, Y., Huguet, G., Wolf, G., & Bengio, Y. (2024). Simulation-Free Schrödinger Bridges via Score and Flow Matching. AISTATS, PMLR 238:1279–1287.

[2] Levina, E., & Bickel, P. J. (2004). Maximum Likelihood Estimation of Intrinsic Dimension. Advances in Neural Information Processing Systems, 17.

[3] Facco, E., d’Errico, M., Rodriguez, A., & Laio, A. (2017). Estimating the intrinsic dimension of datasets by a minimal neighborhood information. Scientific Reports, 7, 12140.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their constructive feedback. We have revised the manuscript in two rounds, with changes from the first round highlighted in blue and those from the second round in purple.

In the first round, we added benchmarks on high-dimensional manifolds and isometric embeddings, expanded the baseline comparisons, and provided experiments with alternative metric estimators and analyses of computational cost.

The second round follows Reviewer NtLz’s further comments. We clarified the metric choices and compared Metric FM under LAND and analytic metrics on the datasets in Sections 6.2 and 6.3, updating the corresponding results and adding Appendix C.8. These comparisons support our original conclusions. Thank you again for helping us strengthen 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.