acceptodds
Under review as a conference paper at ICLR 2027

Light Entropic Optimal Transport on Riemannian Manifolds

Abstract

Entropic Optimal Transport (EOT) has become a practical framework for learning stochastic couplings between complex distributions, with applications in generative modeling and domain adaptation. However, most EOT solvers are designed for Euclidean spaces, while manifold extensions remain limited and often rely on costly iterative methods, simulated dynamics, or generic neural models that do not fully exploit the underlying geometry. We introduce ManifoldLightOT, a light approach for learning kernel-induced EOT couplings directly on common manifolds. Using the kernel form of the EOT solution, we construct geometry-specific Gibbs kernels together with compatible potential parameterizations for spheres, tori, , and . These choices yield closed-form normalization and directly sampleable conditional distributions. Our formulation naturally extends to products of manifolds, making it applicable to more complex geometries. The parameters of the potentials are optimized directly from samples using Monte Carlo estimates of the learning objective. Through synthetic and real-world experiments, we show that ManifoldLightOT often outperforms existing manifold OT methods while retaining direct sampling.

Primary area:
generative models
Keywords:
Entropic Optimal Transport, Manifolds, Deep Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_GC6uOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes an extension of an existing continuous entropy-regularized optimal transport solver to certain set of Riemannian manifolds by choosing a parametrization of the corresponding Gibbs kernel in such a way that the loss introduced by [1] remains tractable. Specifically the authors consider parametrizations for the following group manifolds (Proposition 1 and following): Sphere, Torus, SO(3) and SE(3) and show that their parametrization implies tractable loss also for arbitrary products of these manifolds (Proposition 2). Further the authors provide numerical experiments on synthetic as well as two real world datasets.

Critical strengths

The paper notes that for some particular choices of Gibbs kernel and manifolds the loss proposed in [1] remains tractable.

Critical weaknesses

The reviewer leans towards that the paper in its current form should be rejected, because of the following core issues: (1) Limitation in choice and lack of motivation in OT cost functions that can be used with their method, (2) lack of adequate experimental metrics that measure the quality of the proposed solution. In the following we describe them in more detail:

  1. Main problem: In usual entropy-regularized OT the kernel KϵK_\epsilon stems from a single cost function c=c(x,y)c = c(x,y) via Kϵ=exp⁡(−c/ϵ)K_\epsilon = \exp(-c/\epsilon). In this case the parameter ϵ\epsilon has a clear interpretation as a regularizer of the problem. However, the method proposed here takes as its design object of choice the Gibbs kernel KϵK_\epsilon itself, which then induces the cost cc. In particular, for the proposed kernels of the authors only the kernel for the sphere and torus admit such a representation with a fixed cost cc (independent of ϵ\epsilon). For the case of sphere the induced cost is c(x,y)=−x⊤yc(x,y) = -x^\top y and for the torus we have c(x,y)=−∑j=1dcos⁡(xj−yj)c(x,y) = -\sum_{j=1}^d \cos(x^j - y^j). However, it is not clear why these cost functions would be interesting for these manifolds in the first place. For instance, the cost on the sphere is up to constants equivalent to c(x,y)=12∣∣x−y∣∣2c(x,y) = \frac{1}{2}||x-y||^2 , the ambient space cost, which seems to defeat the purpose of using Riemannian OT at all. Furthermore, for the manifolds SO(3)SO(3) and SE(3)SE(3) the induced cost even depends on ϵ\epsilon, which makes the interpretation of ϵ\epsilon as a regularization parameter to some fixed OT-problem inapplicable. In this case it is not clear what problem the authors are solving.
  2. Lack of metrics: While claiming to address the entropy-regularized OT problem, the authors seem only to evaluate their method by using a metric on the transported marginal of their computed transport plan and ν\nu (see e.g. synthetic experiments). The reviewer found no reported expected transport costs, EOT objective values relative to e.g. a reference optimum or errors of their computed plan against a known optimal coupling. In this sense their reported metrics would be optimal already for the trivial coupling q=μ⊗νq = \mu \otimes \nu, which defeats the purpose of solving a EOT problem.

[1] Alexander Korotin, Nikita Gushchin, and Evgeny Burnaev. Light Schrodinger bridge. In The Twelfth International Conference on Learning Representations, 2024.

Questions for authors

The authors should spend more effort in reasoning why these choices of costs matter in applications and what problem they are solving: generative modeling (transporting μ\mu to ν\nu) or EOT (transporting μ\mu to ν\nu while minimizing the objective) and add corresponding metrics and adequate comparisons with other methods solving the problem with the same cost and entropy parameters.

5 comments
AuthorsOpenReview ↗

Answer [1/1]

We thank the reviewer for the comments and observations about our work, below we provide detailed answers for the raised concerns.

[W1] About geometric interpretations and interpretation of ε\varepsilon.

We thank the reviewer for the observations, below we provide detailed answers for them. For clarity, we split the first Weakness into two.

[W1.1] Geometric interpretation of the induced costs.

We would like to clarify that the proposed kernel-induced costs are closely related to geodesic distances on the corresponding manifolds, although they are not generally identical to them.

For the sphere, c(x,y)=−⟨x,y⟩=−cos⁡dM(x,y)c(x,y)=-\langle x,y\rangle=-\cos d_{\mathcal M}(x,y) is an increasing function of the geodesic distance and satisfies c(x,y)=−1+12dM(x,y)2+O(dM(x,y)4)c(x,y)=-1+\frac{1}{2}d_{\mathcal M}(x,y)^2+\mathcal O(d_{\mathcal M}(x,y)^4). Similarly, on the torus, c(x,y)=−∑j=1dcos⁡(xj−yj)c(x,y)=-\sum_{j=1}^{d}\cos(x^j-y^j) is periodic and, away from the cut locus, satisfies c(x,y)=−d+12dM(x,y)2+O(dM(x,y)4)c(x,y)=-d+\frac{1}{2}d_{\mathcal M}(x,y)^2+\mathcal O(d_{\mathcal M}(x,y)^4).

For SO(3)\mathrm{SO}(3), ∣⟨qx,qy⟩∣=cos⁡(dSO(3)(x,y)/2)|\langle q_x,q_y\rangle|=\cos(d_{\mathrm{SO}(3)}(x,y)/2). Hence, the antipodally symmetric kernel-induced cost is increasing with the rotational geodesic distance and is locally a positive multiple of dSO(3)(x,y)2d_{\mathrm{SO}(3)}(x,y)^2. On SE(3)\mathrm{SE}(3), the cost adds a squared Euclidean translation term to this rotational cost and is therefore locally equivalent to a weighted sum of squared translational and rotational geodesic distances.

Thus, these costs respect the local geometry of their respective manifolds while enabling closed-form normalization and direct conditional sampling, which are the main computational advantages of our construction. The updated manuscript details these equivalences in Appendix A.1 at the end of each manifold-specific parameterization.

[W1.1] Dependence on ε\varepsilon.

Regarding the dependence on ε\varepsilon, we clarify that each proposed kernel KεK_\varepsilon defines a well-posed EOT problem with induced cost cε(x,y)=−εlog⁡Kε(x,y)c_\varepsilon(x,y)=-\varepsilon\log K_\varepsilon(x,y), under the usual assumptions. For spheres and tori, these costs are independent of ε\varepsilon, so it acts as the standard entropy regularization parameter. For SO(3)\mathrm{SO}(3), the induced cost depends on ε\varepsilon, but this dependence is controlled: cε(x,y)=−εlog⁡cosh⁡(λR⟨x,y⟩/ε)c_\varepsilon(x,y)=-\varepsilon\log\cosh(\lambda_R\langle x,y\rangle/\varepsilon) converges uniformly to the fixed cost c0(x,y)=−λR∣⟨x,y⟩∣c_0(x,y)=-\lambda_R|\langle x,y\rangle|, with 0≤cε(x,y)−c0(x,y)≤εlog⁡20\leq c_\varepsilon(x,y)-c_0(x,y)\leq\varepsilon\log 2. For SE(3)\mathrm{SE}(3), our construction uses σε2=εst2\sigma_\varepsilon^2=\varepsilon s_t^2, making the translational cost ∥xt−yt∥2/(2st2)\|x^t-y^t\|^2/(2s_t^2) independent of ε\varepsilon, while the rotational component satisfies the same bound as on SO(3)\mathrm{SO}(3). Consequently, although ε\varepsilon affects both the cost and entropy regularization in the rotational constructions, the induced costs converge to fixed, geometrically meaningful costs as ε→0\varepsilon\to0. Moreover, the corresponding EOT objective values differ from those of the fixed-limit-cost formulations by at most εlog⁡2\varepsilon\log 2, providing a precise connection to conventional fixed-cost EOT.

[W2] Lack of metrics

In order to address the lack of metrics, we have introduced a new benchmark construction in Section 3.4, together with conditional and unconditional tangent-space Bures-Wasserstein unexplained variance percentages (cTBW-UVP and TBW-UVP), alongside MMD2\mathrm{MMD}^2. These metrics allow us to evaluate the learned conditional distributions against known optimal EOT conditionals, rather than relying exclusively on marginal metrics. Further details about the experimental construction and results are provided in our response to Reviewer K4Eg, [W1]. We also include these experiments in the updated manuscript.

On the other hand, it is important to clarify that evaluating continuous EOT couplings on manifolds remains challenging. Existing evaluation protocols are not always suitable for our method. For instance, RNOT uses metrics tailored to deterministic transport maps, which are not directly applicable to our stochastic couplings. Meanwhile, ERNOT relies on discrete reference solutions with small support sizes. Although this approach can be reasonable in low dimensions, such limited supports may poorly approximate continuous couplings in higher dimensions, making these metrics less reliable. Therefore, our contribution extends beyond the proposed solvers by introducing a new benchmark setup for evaluating continuous Riemannian EOT solvers that can accommodate general cost functions.

Concluding remarks. Kindly let us know if the clarifications provided address your concerns about our work. We're happy to discuss any remaining questions during the discussion phase. If you find our responses satisfactory, we would appreciate it if you could consider raising your score.

Thank you for your reply and clarifying the geometric properties and the dependence of on ε\varepsilon for the considered costs.

My main concern is why these particular costs define useful transport problems for the applications considered. It seems that they were chosen mainly because of their computational tractability and convenience in this work and its not clear why they are relevant (note that some geometric justification for these costs does not imply they are relevant).

As mentioned before, the application metrics (e.g. in Continental drift and Crystallographic texture) only assess the matching of the (marginal) target distribution (so do not assess the cost relevance). While the new synthetic benchmark in Section 3.4 does measure (some) quality of the conditional distribution (and thus that of the coupling), it validates recovery under the selected costs themselves (so it again can't serve as a justification for the choice of these costs).

I am happy to increase my score if the authors provide a concrete modeling justification for these costs or test whether their coupling preserves or predicts a relationship that matters in the concrete application.

We thank the reviewer for the follow-up response and the specific suggestions. To address this concern, we constructed an alternative Continental Drift experiment. In this setup, we retain the same source distribution at 150 million years ago (150 Ma), but obtain the target distribution by reconstructing the source locations to the present day (0 Ma) using the plate tectonic model of Müller et al. (2019). This construction provides paired geographic locations at 150 Ma and 0 Ma, allowing us to evaluate the learned transport coupling against the reconstructed correspondences. Importantly, the methods are trained using only unpaired source and target samples, while the correspondences are used only for evaluation on unseen testing data. We trained ManifoldLightOT, ERNOT with geodesic and kernel-infered cost.

The computed metrics are the following:

  • Spherical MMD²: Evaluates marginal distribution matching.
  • Expected Geodesic Error (°): Measures correspondence accuracy by averaging the geodesic distances between samples from the learned conditional distributions and their corresponding reconstructed destinations.
  • Energy Score: Evaluates conditional prediction accuracy and dispersion as E[∥Y−y∗(x)∥]−12E[∥Y−Y′∥]\mathbb{E}[\|Y-y^*(x)\|]-\frac12\mathbb{E}[\|Y-Y'\|], averaged over source points, where Y,Y′Y,Y' are independent samples from q^(⋅∣x)\widehat q(\cdot|x).

For each source point, we generate 128 samples from the learned conditional distribution to estimate the expected geodesic error and energy score. Using multiple conditional samples allows us to evaluate the stochastic coupling rather than relying on a single prediction, providing more reliable estimates of correspondence accuracy and predictive dispersion. Results can be found in the table below:

MethodMMD² (×10⁻²) ↓Expected Geodesic Error (°) ↓Energy Score ↓
ManifoldLightOT0.070 ± 0.00514.854 ± 0.0320.162 ± 0.001
ERNOT (geodesic)0.174 ± 0.04721.585 ± 0.2350.166 ± 0.002
ERNOT (kernel-induced)0.271 ± 0.01328.999 ± 0.0940.190 ± 0.002
Identity6.564 ± 0.00030.446 ± 0.0000.521 ± 0.000

These results demonstrate that ManifoldLightOT and its kernel-induced costs can learn meaningful couplings. A similar result is expected for applications on different manifolds, as each of our kernel-induced costs is closely related to the corresponding geodesic distance (as we explained in our previous answer and in the updated manuscript). We hope our answer addresses the reviewer's concerns in this regard.

Müller et al. (2019), A Global Plate Model Including Lithospheric Deformation Along Major Rifts and Orogens Since the Triassic, published in Tectonics.

Reviewer_HW91OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies entropic optimal transport (EOT) on manifolds. EOT is a practical way to learn stochastic couplings between distributions, with uses in generative modeling and domain adaptation. However, most EOT solvers are designed for Euclidean space. Manifold versions are still limited. A line of research in this field often rely on costly iterative methods when sampling the particle of dynamics, or neural network-based methods that merely respect the geometry. The authors propose a new method called aa "ManifoldLightOT", a light-weight method that learns kernel-induced EOT couplings directly on common manifolds. The method is started from the kernel form of the EOT solution. For the well-known model manifolds such as spheres, tori, SO(3), and SE(3), they build a Gibbs kernel for each geometry with a matching potential parameterization and estabilish GMM like coupled densities. Based on explicit and analytic computation of geometric obejcts, the normalization is in closed form and the conditional distributions can be sampled directly. The method also extends to products of manifolds. The potential parameters are learned directly from samples, using Monte Carlo estimates of the objective.

Critical strengths

None

Critical weaknesses

The contribution of this paper is clear but very marginal. Riemannian generative modeling has already been widely explored in the literature. In particular, in connecting it with OT, the limitation of this paper is that its core contribution has no originality in aspects such as curvature effects, modeling on principal bundles, or deriving principles with general coordinates and geometric tools rather than results specified to particular manifolds.

Also, surprisingly, despite the Riemannian setting, the paper lacks definitions and descriptions of the basic mathematical objects that it should have, as well as general propositions that lead the discussion at a high level. For example, the propositions in the paper are not really theoretical findings but computations worked out on model manifolds, and personally, compared with prior work that mainly uses geometric methods in current ML, I think they have no theoretical value. In addition, the whole paper relies on the ambient Rn\mathbb{R}^n coordinates given by the embedding of embedded manifolds, and there is no description of local objects, no explicit statement of the metric, and no general statement of the Gaussian measure that follows from it at all. Instead of enumerating detailed mathematical derivations for each system in the main text, I would like the authors to first state the principles with formal mathematical statements and move all of those derivations to the appendix.

As my final judgment, this work depends strongly on prior work. I gave a weak accept, but in its current state, I would not mind if the paper were rejected.

Questions for authors
  1. Could the authors write the formal definition of the Gaussian measure, and the GMM based on it, in the local coordinates of the manifolds, and rewrite Proposition 1 of the paper based on them?

  2. There is no theoretical analysis at all, and depending on where the target and the local “empirical measure” are located in the space, curvature effects may cause an OT cost close to infinity in some cases. Is it possible to first develop formal normal coordinates (or any type of coordinates if necessary) -> the Riemannian metric in these coordinates -> the definition of sectional curvature, and then use these mathematical objects to analyze the behavior of the loss function?

  3. Beyond the spaces where the closed form proposed in the model is available, is it possible to extend the method to a wider variety of manifolds by expressing an approximate form of the Gaussian measure with spectral modes?

3 comments
AuthorsOpenReview ↗

Answer [1/2]

We thank the reviewer for the comments and observations about our work, below we provide detailed answers for the raised concerns.

[W1] Contribution of the paper.

We respectfully disagree with the reviewer's assessment that our contribution is marginal. While the suggested directions are relevant to Riemannian generative modeling, our work addresses a different and equally relevant problem. Its main contribution is a tractable framework for EOT on different manifolds that enables closed-form normalization and direct conditional sampling. We believe the novelty of the paper should be assessed in relation to its stated objectives and contributions, rather than according to specific theoretical directions that fall outside its scope.

[W2] About the Riemmanian setting.

[W2.1] Description of mathematical setting.

We partially agree with the reviewer's observation. We agree that the manuscript could benefit from a more explicit theoretical justification of the learning objective in Equations 4 and 5 for the Riemannian setting. To address this, we have introduced Theorem 1 in the updated manuscript, which formally establishes the equivalence between minimizing the forward KL divergence and the proposed learning objective on Riemannian manifolds.

In particular, since both couplings are defined with respect to the same Riemannian volume measure, their common terms μ(x)Kε(x,y)\mu(x)K_\varepsilon(x,y) cancel in the KL density ratio, yielding q⋆(x,y)/qθ(x,y)=rθ(x)v⋆(y)/(r⋆(x)vθ(y))q^{\star}(x,y)/q_\theta(x,y)=r_\theta(x)v^{\star}(y)/(r^{\star}(x)v_\theta(y)). Using the marginal constraints, we directly obtain KL(q⋆∥qθ)=L(θ)−L(v⋆)\mathrm{KL}(q^{\star}\|q_\theta)=\mathcal L(\theta)-\mathcal L(v^{\star}). We provide the complete proof in the appendix, showing that the LightSB learning objective extends to Riemannian manifolds without requiring a particular choice of local coordinates.

On the other hand, we also clarify that the current manuscript already introduces the main mathematical objects in the Notation paragraph and Background section. In particular, the Notation paragraph defines the Riemannian volume measure mm, probability densities, and marginal constraints, while the Background section (\S 2) formulates the EOT problem directly on Riemannian manifolds. The subsequent theoretical developments are also formulated on manifolds, including the general principles established in Propositions 1 and 2.

[W2.2] Theoretical value of Propositions.

We respectfully disagree with the reviewer's assessment that our propositions have no theoretical value. Both propositions establish general principles that enable the subsequent construction of tractable EOT parameterizations, rather than aiming to provide new geometric results. These principles are not restricted to the manifolds considered in our paper and can be applied to other manifolds whenever the required conditions are satisfied. More importantly, these constructions lead to continuous EOT solvers that admit closed-form normalization and direct conditional sampling, avoiding additional inner optimization and enabling out-of-sample transport on different manifolds. We believe these computational properties constitute a meaningful contribution, even though they do not directly rely on curvature analysis.

[W2.3] Regarding Gaussian measure.

We would like to clarify that the general EOT formulation and theoretical results are stated directly on Riemannian manifolds. However, some of our manifold-specific parameterizations use convenient ambient representations, such as unit vectors for spheres and quaternions for SO(3)\text{SO}(3). The chosen representations facilitate the construction of tractable kernels and potentials, while the underlying EOT formulation and normalization are defined with respect to the Riemannian volume measure. Since the proposed parameterizations and theoretical results are valid under these conventions, we believe that reformulating them in local coordinates is not strictly necessary. To make this distinction clearer, we have added a clarification at the beginning of Subsection 3.2.

AuthorsOpenReview ↗

Answer [2/2]

[Q1] Gaussian measure.

To clarify how the proposed reformulation would look, we provide a brief formulation below.

A Gaussian-type density on M\mathcal M, with center p∈Mp\in\mathcal M and scale σ>0\sigma>0, can be defined with respect to the Riemannian volume measure mm as

NM(y∣p,σ)=1Z(p,σ)exp⁡(−dM2(p,y)2σ2),\mathcal N_{\mathcal M}(y|p,\sigma)=\frac{1}{Z(p,\sigma)}\exp\left(-\frac{d_{\mathcal M}^2(p,y)}{2\sigma^2}\right),

where Z(p,σ)=∫Mexp⁡(−dM2(p,y)/(2σ2)) dm(y)Z(p,\sigma)=\int_{\mathcal M}\exp(-d_{\mathcal M}^2(p,y)/(2\sigma^2))\,dm(y) is the normalizing constant. A Gaussian mixture potential can then be defined as vθ(y)=∑k=1KαkNM(y∣pk,σk)v_\theta(y)=\sum_{k=1}^{K}\alpha_k\mathcal N_{\mathcal M}(y|p_k,\sigma_k). Proposition 1 applies to this family whenever each product Kε(x,y)NM(y∣pk,σk)K_\varepsilon(x,y)\mathcal N_{\mathcal M}(y|p_k,\sigma_k) admits closed-form normalization and direct conditional sampling.

However, adopting this formulation throughout the paper would obscure our main contribution as our framework is not restricted to Gaussian-type distributions but applies to compatible kernels and potential families that admit closed-form normalization and direct conditional sampling. These properties are not generally guaranteed by Gaussian mixtures on manifolds. Reformulating the paper around such distributions and local coordinates would therefore shift its focus toward a different geometric approach rather than emphasizing the tractable EOT constructions developed in our work.

[Q2]Theoretical Analysis.

We would like to clarify that the proposed kernel-induced costs are uniformly bounded on spheres, tori, and SO(3)\mathrm{SO}(3), since their defining inner products and trigonometric terms are bounded. For SE(3)\mathrm{SE}(3), the rotational cost is bounded, while the expected translational cost remains finite under finite second moments. These properties also extend to Cartesian products. To address the reviewer's concern, we have included a brief analysis of these bounds in Appendix B, with a corresponding reference in the main text.

[Q3] Extension to more manifolds

Yes, that is possible, Proposition 1 is not restricted to the manifolds considered in our paper. It provides a general construction of qθ(y∣x)q_\theta(y|x) with closed-form normalization and direct conditional sampling whenever a suitable kernel and potential family are available. Thus, our framework can extend to other manifolds satisfying these conditions. However, finding compatible kernels and potentials that preserve these properties is nontrivial.

On the other hand, spectral approximations of Gaussian-type distributions could provide a way to model distributions on more general Riemannian manifolds, but they do not necessarily satisfy the tractability conditions required by our framework. For example, replacing the Gaussian-type density introduced in [Q1] with a spectral approximation N~M(y∣p,σ)\widetilde{\mathcal N}_{\mathcal M}(y|p,\sigma) yields

rθ(x)≈∑k=1Kαk∫MKε(x,y)N~M(y∣pk,σk) dm(y).r_\theta(x)\approx\sum_{k=1}^{K}\alpha_k \int_{\mathcal M}K_\varepsilon(x,y) \widetilde{\mathcal N}_{\mathcal M}(y|p_k,\sigma_k)\,dm(y).

Some of our existing parameterizations can also be expressed using Gaussian-type constructions, while preserving their closed-form properties. However, these are not generally equivalent to Gaussian densities defined through squared geodesic distances, as introduced in [Q1]. The tractability of our constructions follows from the specific compatibility between the kernels and potential families, rather than their representation in Gaussian notation. We consider this approach appropriate for our work, as our primary objective is to obtain closed-form normalization and direct conditional sampling. While approximation-based extensions are an interesting research direction, they would shift the focus of our work toward a different class of parameterizations.

Concluding remarks. Kindly let us know if the clarifications provided address your concerns about our work. We're happy to discuss any remaining questions during the discussion phase. If you find our responses satisfactory, we would appreciate it if you could consider raising your score.

Thank you for the careful rebuttal. Since all the concerns are resolved, I will maintain my initial score.

Reviewer_K4EgOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper extends LightSB to manifolds such as spheres, tori, SO(3), SE(3), and their product spaces. For each manifold, the authors pick a Gibbs kernel that is compatible with a specific distribution family: the von Mises-Fisher (vMF) distribution for spheres, the von Mises distribution for tori, a cosh-based form (symmetric with respect to quaternions ±q) for SO(3), and a Gaussian distribution for translational components. The potential function is parameterized as a weighted sum of compatible potential components, which ensures a closed-form expression for the normalization function and allows for direct sampling from the conditional distributions. Training follows the original LightSB objective function and requires only unpaired samples. Experiments compare the method with RNOT, ERNOT, and NeuralOT on synthetic data using MMD, and include two applications: continental drift and the crystallographic texture of a titanium alloy.

Critical strengths

1.The method is elegant and conceptually clean. The authors adopt the design principle of choosing compatible kernel–potential pairs whose products admit closed-form normalization and direct conditional sampling. Constructions for various manifolds are derived from this principle, so no neural networks, inner-loop optimization, or trajectory simulation are needed. I particularly like the cosh-based form on SO(3), which neatly handles the fact that quaternions q and −q represent the same rotation.

2.Training and sampling are simple and should be cheap. The only parameters are the mixture components; training relies on mini-batch Monte Carlo estimation, and sampling is performed directly.

Critical weaknesses
  1. The evaluation does not directly establish the accuracy of the learned conditional distributions. Squared MMD measures agreement between the push-forward and target marginals, but the paper provides no guarantee that small marginal errors imply accurate conditionals, particularly at ϵ=0.01\epsilon = 0.01. Direct comparisons against ground-truth conditionals, as in Gushchin et al. (2023b) and LightSB, would provide stronger evidence. The closed-form families used here appear well suited to constructing such benchmarks.

  2. The baseline comparisons do not use a common transport objective. RNOT and ERNOT use squared geodesic costs, whereas the proposed method uses a kernel-induced cost, and the regularization strengths also differ between ERNOT and the proposed method. The results therefore primarily compare target-marginal fit rather than accuracy on the same EOT problem. Comparisons with matched costs and regularization, using an adapted ERNOT or a discrete Sinkhorn reference, would be more informative.

  3. Several baseline results raise concerns about convergence or configuration. NeuralOT's reported MMD2\mathrm{MMD}^2 remains around 61.3 to 61.6 on tori with D≥20D \geq 20, in units of 10−210^{-2}, and Figure 2 shows collapse to a single point in texture transfer. On S100S^{100}, its standard deviation also exceeds its mean. These results warrant further tuning and convergence checks before drawing conclusions about relative performance.

  4. The real-world experiments provide limited evidence of practical effectiveness. Continental drift is illustrated by a single visualization without quantitative metrics, and the RNOT and ERNOT results are missing despite their hyperparameters being listed. For texture transfer, variability across runs is not reported, and the handling of hexagonal crystal symmetry in the transport cost and ODF-based metrics remains unclear.

Questions for authors

1.Why did you choose these particular synthetic distributions and applications? Work on manifold generative models and OT commonly uses more standard datasets, such as geoscientific data on spheres (e.g., earthquake or volcanic activity locations), dihedral angles of proteins and RNA on tori, or the benchmark tasks used in the ERNOT paper. Could results be provided for some of these datasets?

2.Could you report runtime for the proposed method and the baselines?

3 comments
AuthorsOpenReview ↗

Answer [1/3]

We thank the reviewer for the overall review. Below we provide detailed answers to the observations and questions.

[W1] Benchmark Construction.

We thank the reviewer for this suggestion. We have added Section 3.4, introducing a benchmark for direct conditional evaluation, together with new experimental results comparing ManifoldLightOT and the baselines. The construction follows Proposition 1, which establishes closed-form normalization and direct conditional sampling for suitable kernels KεK_\varepsilon and potentials vθv_\theta. By fixing the potential parameters θ\theta and a source distribution μ\mu, we construct a target distribution ν\nu for which qθ(y∣x)q_\theta(y|x) is the a known optimal EOT conditional q⋆(y∣x)q^{\star}(y|x). We can then generate optimal pairs (x,y)(x,y) by sampling x∼μx\sim\mu and y∼q∗(⋅∣x)y\sim q^*(\cdot|x), enabling direct evaluation of conditional accuracy.

For each manifold-aware potential vθ,εv_{\theta,\varepsilon}, we select parameters yielding visually plausible target distributions (Appendix C). We consider three values of ε∈{0.1,1,10}\varepsilon\in \lbrace 0.1, 1, 10 \rbrace. We evaluate ManifoldLightOT and ERNOT using unconditional and conditional tangent-space Bures-Wasserstein unexplained variance percentages (cTBW-UVP and TBW-UVP), alongside MMD2\mathrm{MMD}^2. The underlying squared tangent-space Bures--Wasserstein distance is defined as

TBW22((μ1,Σ1),(μ2,Σ2))=∥μ1−μ2∥2+Tr⁡(Σ1+Σ2−2(Σ11/2Σ2Σ11/2)1/2)\mathrm{TBW}_2^2((\mu_1,\Sigma_1),(\mu_2,\Sigma_2))=\|\mu_1-\mu_2\|^2+\operatorname{Tr}\left(\Sigma_1+\Sigma_2-2(\Sigma_1^{1/2}\Sigma_2\Sigma_1^{1/2})^{1/2}\right)

where μi\mu_i and Σi\Sigma_i denote the means and covariances computed in the corresponding tangent-space representation. The manifold-specific computations and UVP normalization are detailed in Appendix D.

For conditional evaluation, we consider 1,000 source samples and generate 500 samples per source point from both the known optimal conditional q∗(⋅∣x)q^*(\cdot|x) and the learned conditional q^(⋅∣x)\widehat q(\cdot|x).

For ERNOT, we evaluate using our kernel-induced costs. We exclude RNOT and NeuralOT because both learn deterministic transport maps, making conditional distribution metrics less meaningful. Additionally, adapting RNOT to our kernel-induced costs is not straightforward. Results of the benchmark experiments are summarized in the table below and a full version appears in Table 3 (main text for ε=0.1\varepsilon=0.1) and Tables 7 and 8 (Appendix F) of the updated text.

ε\varepsilonMethodS2\mathbb{S}^{2}S64\mathbb{S}^{64}T2\mathbb{T}^{2}T64\mathbb{T}^{64}SO(3)\mathrm{SO}(3)SE(3)\mathrm{SE}(3)
0.1ManifoldLightOT0.196±0.004\mathbf{0.196 \pm 0.004}4.985±0.013\mathbf{4.985 \pm 0.013}0.464±0.052\mathbf{0.464 \pm 0.052}0.831±0.308\mathbf{0.831 \pm 0.308}6.956±0.0216.956 \pm 0.0219.115±0.052\mathbf{9.115 \pm 0.052}
0.1ERNOT (kernel induced)0.919±0.0590.919 \pm 0.05925.442±0.34325.442 \pm 0.3430.774±0.0900.774 \pm 0.0900.877±0.1530.877 \pm 0.1536.583±0.200\mathbf{6.583 \pm 0.200}9.960±0.4449.960 \pm 0.444
1ManifoldLightOT0.623±0.018\mathbf{0.623 \pm 0.018}4.627±0.022\mathbf{4.627 \pm 0.022}0.645±0.136\mathbf{0.645 \pm 0.136}1.032±0.147\mathbf{1.032 \pm 0.147}0.643±0.024\mathbf{0.643 \pm 0.024}0.788±0.020\mathbf{0.788 \pm 0.020}
1ERNOT (kernel induced)2.156±0.2322.156 \pm 0.23261.408±0.01061.408 \pm 0.0104.318±0.4414.318 \pm 0.44112.236±0.10212.236 \pm 0.1025.758±1.1105.758 \pm 1.1107.984±0.5387.984 \pm 0.538
10ManifoldLightOT0.598±0.035\mathbf{0.598 \pm 0.035}4.769±0.027\mathbf{4.769 \pm 0.027}0.643±0.138\mathbf{0.643 \pm 0.138}0.933±0.024\mathbf{0.933 \pm 0.024}0.581±0.022\mathbf{0.581 \pm 0.022}0.735±0.022\mathbf{0.735 \pm 0.022}
10ERNOT (kernel induced)11.010±0.13411.010 \pm 0.13462.437±0.01562.437 \pm 0.015102.030±0.769102.030 \pm 0.769198.029±0.321198.029 \pm 0.32128.652±0.08628.652 \pm 0.086112.085±0.641112.085 \pm 0.641

From these results, we observe that ManifoldLightOT consistently outperforms ERNOT. However, this is expected, as ManifoldLightOT is based on the same principles underlying the construction of the optimal pairs in this setup. This inductive bias provides a natural advantage, suggesting that ManifoldLightOT should be considered as an oracle-like method in this specific experimental setting.

AuthorsOpenReview ↗

Answer [2/3]

[W2] Baseline Comparison.

We thank the reviewer for the comment and suggestion. Below, we include an updated version of Table 2 for the synthetic experiments that includes an adapted ERNOT that uses our kernel-induced cost functions. To make this adaptation, we replace the squared geodesic cost in the ERNOT objective and conditional sampling procedure with the corresponding kernel-induced cost cε(x,y)=−εlog⁡Kε(x,y)c_{\varepsilon}(x,y)=-\varepsilon\log K_{\varepsilon}(x,y). We also use the same value of ε\varepsilon as ManifoldLightOT. The remaining ERNOT architecture and training procedure are unchanged. This allows ERNOT and ManifoldLightOT to optimize the same EOT problem, while we retain the original ERNOT results with geodesic costs for completeness.

MethodS60\mathbb{S}^{60}S100\mathbb{S}^{100}T60\mathbb{T}^{60}T100\mathbb{T}^{100}SO(3)\mathrm{SO}(3)SE(3)\mathrm{SE}(3)
ManifoldLightOT (ours)0.07 ± 0.000.45 ± 0.050.08 ± 0.000.08 ± 0.000.03 ± 0.010.18 ± 0.01
RNOT0.18 ± 0.041.47 ± 0.212.15 ± 0.2023.15 ± 0.570.61 ± 0.270.23 ± 0.04
ERNOT (geodesic)0.23 ± 0.010.33 ± 0.130.27 ± 0.030.27 ± 0.020.47 ± 0.095.54 ± 0.72
ERNOT (kernel induced)0.23 ± 0.020.33 ± 0.100.27 ± 0.030.27 ± 0.030.06 ± 0.040.50 ± 0.09
NeuralOT0.27 ± 0.064.38 ± 8.2461.59 ± 0.0661.61 ± 0.071.27 ± 2.580.27 ± 0.05

As can be seen, ERNOT performs similarly with both types of costs on spheres and tori. Nevertheless, using the kernel-induced costs improves its performance on SO(3)\mathrm{SO}(3) and SE(3)\mathrm{SE}(3). The updated manuscript includes this table, as well as an updated version of Table 4 from Appendix A.5, which we omit here to keep the rebuttal concise. The updated manuscript also includes results for ERNOT with kernel-induced cost for the rest of experiments.

[W3] Fine tunning.

We thank the reviewer for raising this concern. We performed additional tuning but could not find parameters that consistently improved NeuralOT’s results. Addressing the observed instability may require more expressive architectures, improved parameter initialization, or stronger manifold-specific inductive biases. Such study is a good future direction of research, but falls outside the current scope of the paper.

[W4] Real-world experiments.

We thank the reviewer for these observations. Below, we address each concern separately.

Continental Drift. We first note that the original RNOT paper primarily presents qualitative results for this experiment and does not report MMD-based comparisons. To provide a direct quantitative comparison between the methods, we compute the same MMD2\mathrm{MMD}^2 metric used in our synthetic sphere experiments. The results are reported below.

MethodSpherical MMD ↓ (×10−2\times 10^{-2})
ManifoldLightOT0.0693
RNOT0.2524
ERNOT (geodesic)0.5265
ERNOT (kernel-induced)0.8400

As shown in the table, ManifoldLightOT also outperforms the other methods on this task, further demonstrating the suitability of our kernel-induced cost functions beyond synthetic experiments. The updated manuscript includes these results in Table 4 (main text).

Crystal experiments. We thank the reviewer for raising this point. We clarify that we do not perform a region reduction during preprocessing. The preprocessing only normalizes the orientation quaternions and resolves the double-cover ambiguity q≡−qq\equiv -q. Mapping every orientation to a fixed representative of the hexagonal fundamental region is not required for our distribution-matching task and can introduce artificial discontinuities at the boundaries of that region. Moreover, all methods are trained using the same quaternion representation and their original SO(3)\text{SO}(3) transport costs.

Nevertheless, we agree that the physical evaluation of the predicted texture should account for hexagonal crystal symmetry. In the updated experiments, the SO(3)\text{SO}(3) heat kernel used by the MMD and ODF-based metrics is averaged over the 12 proper symmetry operations of the HCP crystal structure. Therefore, crystallographically equivalent orientations produce identical evaluation results. The MM-index is also computed using the corresponding symmetry-aware disorientation angles. We emphasize that this modification concerns the evaluation metrics; the transport costs used during training are unchanged. We have clarified this distinction in the updated manuscript and also reported variability across five independent runs.

AuthorsOpenReview ↗

Answer [3/3]

[Q1] About synthetic experiments.

We thank the reviewer for this suggestion. We selected these experiments because they follow established benchmarks in the literature and cover several of the settings suggested by the reviewer. For instance, the synthetic experiments on SO(3)\mathrm{SO}(3) and SE(3)\mathrm{SE}(3) are directly adopted from the ERNOT paper (Subsection 5.1 and Appendix C.2), while the sphere and torus experiments follow RNOT (Subsection 6.2), as they allow us to evaluate higher-dimensional settings. The standard real-world Continental Drift setup is also adopted from RNOT (Subsection 6.1), providing a realistic geoscientific application on the sphere. Finally, we consider crystallographic texture transfer on SO(3)\mathrm{SO}(3) as an additional real-world task. Regarding the protein-ligand task from ERNOT (Subsection 5.3), we initially considered including this experiment. However, we found that the paper does not provide sufficient details to reproduce its preprocessing procedure, and the corresponding code is not publicly available. Given the complexity of this application, we decided not to include it.

Overall, the updated experimental section provides a broad comparison with existing methods across established synthetic and real-world benchmarks, complemented by our new benchmark construction and crystallographic texture transfer experiment.

[Q2] Runtimes.

Regarding computational time, we report approximate wall-clock training times until convergence for a single seed on an NVIDIA A100 GPU. Across the considered experiments, ERNOT typically converges within 2-5 minutes, RNOT within 5-15 minutes, and ManifoldLightOT within 8-20 minutes, depending on the manifold and dimensionality. Although ManifoldLightOT generally requires longer training, it provides closed-form normalization, direct conditional sampling and better performance in almost all tasks. These estimates are indicative rather than a controlled runtime comparisons due to the limited time during rebuttal.

Concluding remarks. Kindly let us know if the clarifications provided address your concerns about our work. We're happy to discuss any remaining questions during the discussion phase. If you find our responses satisfactory, we would appreciate it if you could consider raising your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.