acceptodds
Under review as a conference paper at ICLR 2027

Entropic Riemannian Neural Optimal Transport

Abstract

Many machine learning problems involve data supported on curved spaces such as spheres, rotation groups, hyperbolic spaces, and general Riemannian manifolds, where Euclidean geometry can distort distances, averages, and the resulting optimal transport (OT) problem. Existing manifold OT methods have pursued amortized out-of-sample maps, while entropic regularization has made discrete OT more scalable, but these advantages have remained largely disjoint. We propose Entropic Riemannian Neural Optimal Transport (Entropic RNOT), a unified framework that combines intrinsic entropic OT with amortized out-of-sample evaluation on Riemannian manifolds. Our method learns a single target-side Schr"odinger potential through a neural pullback parameterization, recovers the induced Gibbs coupling and uses the resulting conditional laws to construct intrinsic transport surrogates. These include barycentric projections on Cartan–Hadamard manifolds and heat-smoothed conditional surrogates on complete stochastically complete manifolds, the latter turning possibly atomic target laws into absolutely continuous ones. For fixed regularization \(\varepsilon>0\), we prove that the proposed hypothesis class recovers the entropic optimal coupling in strong probabilistic metrics. As consequences, barycentric surrogates converge in \(L^2\), while heat-smoothed surrogates are stable at fixed heat time and asymptotically unbiased as the heat time vanishes. The guarantees hold for compactly supported data on possibly noncompact manifolds. Empirically, our method matches or improves over Euclidean, tangent-space, and log-Euclidean baselines on benchmarks over \(\mathbb S^2\), \(SO(3)\), \(SPD(3)\), \(SE(3)\), and \(\mathbb H^2\), scales favorably relative to discrete manifold Sinkhorn, and in a protein–ligand docking application, refines poses on without retraining or per-instance optimization.

Primary area:
generative models
Keywords:
Optimal transport, Riemannian manifolds, entropic regularization, Schrödinger potentials, neural optimal transport, geometric machine learning.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_x1gtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Entropic Riemannian Neural Optimal Transport, which learns a target-side Schrödinger potential through a neural feature parameterization and reconstructs the corresponding Gibbs conditionals. These conditionals are used to obtain barycentric or smoothed transport surrogates. Under compact-support and universal-approximation assumptions, the paper establishes recovery of the entropic optimal coupling and derives convergence or stability results for the associated surrogates. Experiments cover several manifold geometries and an application to protein–ligand pose refinement.

Critical strengths
  • The combination of intrinsic entropic OT and a reusable neural potential is well motivated. The formulation provides a good way to construct conditional transport outputs for new source points.
  • The theoretical development distinguishes coupling recovery, barycentric projection, and smoothing of conditional laws. In particular, the KL identity gives a useful connection between potential approximation and recovery of the optimal coupling.
  • The experiments span several geometries, including compact and noncompact manifolds, and investigate both transport quality and computational efficiency. The pose-refinement application provides an additional practical motivation.
Critical weaknesses
  1. This is a suggestion about the quantitative strength of the result. Proposition G.7, Eq. (22), bounds the squared barycentric error by a quantity linear in the joint total variation distance. The pointwise estimate already established in its proof yields a stronger bound. Let EgE_g denote the squared barycentric error:

Eg=∫d(Tg(x),Tε(x))2,dμ(x).E_g = \int d(T_g(x),T_\varepsilon(x))^2\\,d\mu(x).

The pointwise barycenter contraction estimate gives

Eg≤D2∫∥πg,xε−πε,x⋆∥TV2,dμ(x).E_g \le D^2\int \lVert \pi_{g,x}^\varepsilon-\pi_{\varepsilon,x}^\star\rVert_{\mathrm{TV}}^2\\,d\mu(x).

Applying Pinsker's inequality to each conditional law and then using the KL chain rule yields

Eg≤D22KL⁡(πε⋆∥πgε).E_g \le \frac{D^2}{2}\operatorname{KL}(\pi_\varepsilon^\star\Vert\pi_g^\varepsilon).

Finally, the semidual-gap identity implies Eg≤D22ε(OT⁡ε−J(g))E_g \le \frac{D^2}{2\varepsilon}(\operatorname{OT}_{\varepsilon}-J(g))

  1. The theoretical guarantees seems to not cover the implemented minibatch training procedure. Theorem 4.1 establishes the existence of an approximating sequence in the hypothesis class. It does not show that the optimization procedure in Section 3 produces such a sequence. The authors replace a minibatch average by the target integral inside the logarithm, which introduces an additional issue beyond neural-network optimization.

For a fixed potential gg and source point xx, define

\qquad Y_j\sim\nu.$$ Jensen's inequality gives $$\mathbb{E}\left[-\varepsilon\log\left(\frac{1}{B}\sum_{j=1}^{B}U_j\right)\right] \ge -\varepsilon\log\mathbb{E}[U_1].$$ Consequently, the minibatch objective is generally a biased estimator of the population semidual. This bias depends on the potential, so it can change the optimization problem. As an illustrative limiting case, when $B=1$, the empirical objective reduces to $c(X,Y)$ and contains no learning signal for the potential. 3. The implemented smoothing kernel and extracted point predictions are not covered directly by Proposition 4.3. Proposition 4.3 concerns smoothing with the actual Riemannian heat kernel. Appendix C.1 instead implements the geodesic Gaussian kernel $$k_t(y,z)=\exp\left(-\frac{d(y,z)^2}{4t}\right).$$ The appendix acknowledges that this surrogate omits heat-kernel normalization factors, curvature corrections, and volume-distortion terms. On curved manifolds, it is generally not the heat kernel. The leading short-time logarithmic asymptotic alone does not provide an approximation-error guarantee at the bandwidths used in the experiments. Moreorver, the geodeisc Gaussian may fail to be positive defined in many curved manifolds, that would be another issues that may raise cocerns.
Questions for authors

How should the large tangent-space plan error on the log-Euclidean SPD benchmark be reconciled with the exact Euclidean representation of that geometry?

This is my main concern about the interpretation of the synthetic comparisons. Appendix C.2 defines the log-Euclidean distance as

dLE(X,Y)=∥log⁡X−log⁡Y∥F.d_{\mathrm{LE}}(X,Y)=\lVert\log X-\log Y\rVert_F.

Thus, the matrix logarithm is a global isometry to the Euclidean space of symmetric matrices. Entropic OT computed in matrix-logarithm coordinates should therefore reproduce the intrinsic LE reference plan, up to numerical error, when the supports, marginal weights, regularization, and solver settings are matched. More generally, a correctly metric-weighted tangent representation at any base point PP satisfies

= d_{\mathrm{LE}}(X,Y).$$ There is no curvature-induced single-chart approximation error in this geometry. Nevertheless, Table 3 reports a plan KL of $1.5911\pm0.0384$ for the tangent-space baseline on SPD(3) (LE), compared with $0.0280\pm0.0030$ for Entropic RNOT.Could the authors clarify the following? 1. Which Riemannian logarithm and tangent-space inner product were used for the LE row? 2. Were the evaluation supports, marginal weights, cost scaling, regularization, and Sinkhorn settings identical to those used for the LE reference? 3. Could the authors include an explicit Euclidean Sinkhorn baseline on the matrix logarithms, verify equality of its cost matrix with the intrinsic LE cost matrix, and report the resulting plan KL?
Reviewer_Xg1QOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper solves entropic optimal transport on Riemannian manifolds using neural optimal transport. They use amortization: for a fixed source and target distribution, they learn a transport plan that can then be used to transport unseen source points. They learn a neural network to predict one of the potentials, and recover the other using an approximation of the soft c-transform. They then recover the transport map using barycentric projection or heat-smoothed mode extraction. Their method is tested on a synthetic dataset on different Riemannian manifolds as well as a real protein-ligand docking dataset.

Critical strengths
  • The paper learns entropic optimal transport plans on manifolds, which is useful for distributions on the sphere, covariance matrices, and other geometrically structured data.
  • The authors do not limit the experiments to synthetic data and introduce a real application of their method to protein-ligand docking.
Critical weaknesses
  • There is insufficient positioning relative to existing neural entropic OT estimators and amortized solvers, which makes it hard to evaluate the novelty of the method. For example, positioning with respect to the following papers would help understand the contribution:

Wang, Tao, and Ziv Goldfeld. "Neural estimation of entropic optimal transport." 2024 IEEE International Symposium on Information Theory (ISIT). IEEE, 2024.

Brandon Amos, Giulia Luise, Samuel Cohen, Ievgen Redko. "Meta Optimal Transport", Proceedings of the 40th International Conference on Machine Learning, PMLR 202:791-813, 2023.

Jonathan Geuter, Gregor Kornhardt, Ingimar Tomasson, Vaios Laschos. "Universal neural optimal transport", Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:19196-19232, 2025.

Korotin, Alexander, Daniil Selikhanovych, and Evgeny Burnaev. "Neural optimal transport. », The Eleventh International Conference on Learning Representations, 2023

Moreover, among neural optimal transport methods, the performance is only evaluated against non-entropic RNOT. Evaluating against closely related neural OT methods such as the one from Wang and Goldfeld (2024), which also parametrize the entropic OT semidual with a neural network, would strengthen the contribution.

  • Table 1 evaluates transport accuracy on a fixed 200×200200\times200 source–target support, while Figure 2 compares runtime and memory usage across different support sizes. However, without corresponding accuracy measurements at these sizes, it is unclear whether the methods have converged to comparably accurate solutions. Reporting accuracy alongside the scaling results would strengthen the computational comparison.
Questions for authors
  • What is the sensitivity to the parameter epsilon? How do the training stability, accuracy, and runtime evolve as epsilon decreases?

  • A central claimed advantage of the method is its scalability relative to the dense Sinkhorn implementation used in the experiments. However, memory-efficient implementations, such as the online backend in GeomLoss, avoid storing the full cost matrix and have linear memory requirements. Could you compare against such an implementation?

Reviewer_S24vOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Entropic Riemannian Neural Optimal Transport (Entropic RNOT). The method first encodes manifold-valued samples as Euclidean features, then learns an entropic OT potential through a neural network and computes the conditional transport distribution for each source point, and finally extracts a deterministic transport map on the original manifold. Theoretically, the authors prove that couplings induced by neural potentials can approximate the optimal entropic coupling and establish convergence and stability guarantees for the associated transport surrogates. Experiments show that Entropic RNOT achieves higher transport accuracy than methods based on Euclidean embeddings and tangent-space approximations, while reducing GPU memory usage and improving computational efficiency compared with discrete Sinkhorn.

Critical strengths

The paper explores the application of entropic neural optimal transport to distributions supported on Riemannian manifolds. The method has a simple design and is easy to follow.

Critical weaknesses
  1. Limited methodological novelty. The manifold feature encoding used in this paper has already been employed in RNOT [1]. Neural parameterization of entropic OT semi-dual potentials, construction of Gibbs couplings from these potentials, and control of the coupling KL error through the population semi-dual objective gap have also been studied in prior work [2]. Thus, the main contribution is a Riemannian extension of existing neural entropic OT methods, with limited novelty in the core algorithmic mechanisms.

  2. Insufficient and unfair experimental comparisons. Most experiments rely on simplistic baselines or unfair comparison settings, making it difficult to adequately validate the effectiveness of the proposed method. Specifically, Tables 1 and 3 use weak baselines, Tables 2 and 4 lack comprehensive and fair baseline comparisons, and Tables 5 and 6 employ unfair experimental settings. These concerns are detailed in the Questions section.

References

[1] Riemannian Neural Optimal Transport. 2026

[2] Neural Estimation of Entropic Optimal Transport. 2024

Questions for authors
  1. What are the main advantages of the proposed method over RNOT? Given that RNOT is the most directly relevant baseline, the paper should provide a comprehensive comparison across the main experimental settings to establish these benefits.

  2. Why does Table 1 not include a Euclidean OT baseline operating in the authors’ landmark-based embedding space? Additionally, why are the Ambient Euclidean and Tangent-space baselines not implemented using ENOT [2]?

  3. In Table 2, Sinkhorn SE(3) only uses the source and target distributions from a single complex, whereas Entropic RNOT is trained on data pooled from multiple complexes. I suggest the authors consider the following experiments:

  • Train Entropic RNOT using the source and target distributions from a single complex, or run Sinkhorn SE(3) using pooled distributions from multiple complexes, to ensure a fair comparison.
  • Include more competitive baselines, such as RNOT and ENOT with a standard quaternion-based embedding.
  1. In Table 6, why is Entropic RNOT trained for 3,000 iterations while RNOT is trained for only 200? Could the authors also report the results of Entropic RNOT after 200 iterations to ensure a fairer comparison?
Reviewer_JKWpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Entropic Riemannian Neural Optimal Transport (Entropic RNOT), a framework that combines entropic optimal transport with neural Riemannian optimal transport to enable efficient, amortized transport on Riemannian manifolds. The method learns a target-side Schrödinger potential through neural parameterization and constructs a Gibbs coupling using the semi-dual formulation of entropic OT. It further develops intrinsic transport surrogates based on barycentric projections and heat-smoothed conditional distributions to accommodate different manifold geometries. The authors provide theoretical guarantees for the recovery of the entropic optimal coupling and the convergence and stability of the resulting transport surrogates. Experiments on five types of Riemannian manifolds demonstrate the effectiveness and computational efficiency of the proposed framework. The method is also evaluated on a real-world protein–ligand docking pose-refinement task, achieving improvements over the evaluated baselines.

Critical strengths
  1. The paper presents a clear research motivation and an interesting approach with a certain degree of novelty.

  2. The paper provides a solid theoretical analysis.

Critical weaknesses

1. The experimental comparisons and evaluation settings are insufficient to fully demonstrate the advantages of the proposed method.

Although RNOT, the most closely related method, is mentioned in the main text, the detailed quantitative comparisons are mainly presented in the appendix, with limited discussion of the performance differences between the two approaches. Furthermore, the main experimental comparisons primarily involve relatively traditional baselines, without sufficient comparisons with more recent advances in Neural OT. For example, Overcoming Spurious Solutions in Semi-Dual Neural Optimal Transport: A Smoothing Approach for Learning the Optimal Transport Plan is a relevant recent work that deserves further comparison or discussion where applicable. More comprehensive evaluations would help establish the effectiveness and distinctive advantages of the proposed method.

2. The discussion of related work is insufficient.

Optimal transport and neural approaches to OT have been extensively studied, with substantial developments in recent years. Although the proposed method has clear differences from existing approaches, the paper lacks a systematic discussion of related work and does not sufficiently explain the development and limitations of previous methods. A dedicated Related Work section would help clarify the connections and differences between the proposed framework and existing Neural OT, Entropic OT, and Riemannian OT methods, making the paper's research positioning and contributions easier to understand.

Questions for authors

1. Applicability of the proposed framework to broader real-world tasks.

The current experiments mainly evaluate the proposed framework on synthetic manifold benchmarks and a protein–ligand docking pose-refinement task. Could the authors discuss its applicability to a broader range of real-world applications, such as the geographic mapping task considered in RNOT and the image-to-image translation tasks studied in Overcoming Spurious Solutions in Semi-Dual Neural Optimal Transport: A Smoothing Approach for Learning the Optimal Transport Plan? Given the differences in data structures and geometric properties across these applications, can the proposed framework be directly applied, or would additional model designs or theoretical assumptions be required?

2. Satisfaction of the marginal constraints by the learned Gibbs coupling.

According to Eq. (9), the proposed framework constructs a Gibbs coupling by learning a target-side Schrödinger potential. However, for a neural potential that has not reached its optimum, the soft (c)-transform can ensure that the source marginal constraint is satisfied, while the target marginal may not exactly match (\nu).

Could the authors clarify how the deviation of the target marginal is measured and controlled in practice under finite-step neural network training? Furthermore, to what extent could such deviations affect the accuracy and stability of the subsequent barycentric projections or heat-smoothed transport surrogates?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.