acceptodds
Under review as a conference paper at ICLR 2027

Dynamic Generalized Gromov-Wasserstein Optimal Transport

Abstract

Gromov–Wasserstein optimal transport (GW-OT) extends classical optimal transport by introducing structure-aware transport cost. This is particularly relevant for spatial transcriptomics, where dynamical reconstruction should preserve tissue structure in addition to matching expression patterns. While static formulations have been widely used for such structure-aware alignment, a general dynamic formulation for reconstructing continuous trajectories is still missing. We introduce **T**ravelling **P**air **D**ynamical **A**lignment and **T**rajectory **E**stimation (TP-DATE), a theoretical and computational framework to generalize GW-OT dynamically in a simulation-free manner. We formulate a broad class of static and dynamic Quadratic-form OT (QOT) through path actions and prove the static dynamic equivalence. We further develop travelling-pair flow matching, which allows interacting conditional paths and marginalizes their interactions into a single vector field. On synthetic and real spatial transcriptomics data, TP-DATE better preserves spatial structure and improves continuous 3D dynamics reconstruction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Gromov-Wasserstein optimal transport, dynamic optimal transport, flow matching

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_BM4KOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces TP-DATE, which incorporates pairwise structure into continuous transport paths and uses analytical paths for flow matching training. I find this a promising and useful approach. The rotation experiments clearly illustrate the motivation, while the theoretical development and scalable implementation give the method substance. My overall impression is positive, and I would welcome clarification on the scope of the theoretical guarantees and the interpretation of the empirical results.

Critical strengths

1.The rotation experiments make the motivation easy to understand and clearly demonstrate why intermediate paths matter for preserving structure.

2.The analytical paths directly provide position and velocity targets, making them a practical contribution. Comparing different paths under the same coupling also helps isolate the contribution of path design.

Critical weaknesses

1.The connection between the theory and the final implementation remains somewhat unclear. The experiments learn a single-particle velocity field, which the paper acknowledges cannot generally recover the full two-particle dynamics. I would appreciate clarification on which structural and action guarantees carry over to the generated trajectories.

2.The experiments leave the broader applicability uncertain. The synthetic tasks focus on rigid rotation, and each real dataset mainly tests one held-out snapshot. Additional nonrigid transformations and held-out positions would help establish when the structural prior is beneficial and how far the trajectory and 3D reconstruction claims extend.

3.One aspect of the path ablation deserves further explanation. Under the reported weights, the conditional paths for Mouse Brain and ARTISTA appear close to straight lines, yet some expression metrics improve noticeably. Could the authors clarify how the weights enter the implementation and how the resulting supervision differences explain these gains?

4.The efficiency results are encouraging, but the comparison protocol needs clarification. Please specify which stages are included in the timing and report the training budgets for all methods. This would help readers assess the practical computational cost.

Questions for authors

1.Which pairwise structural guarantees remain valid for the single-particle implementation, and how does it relate to the BB optimum?

2.Do the gains remain consistent across additional held-out positions and nonrigid transformations?

3.How different are the position and velocity targets from those of straight paths, and can these differences explain the ablation gains?

4.Does the timing include graph construction and coupling optimization, and are the training budgets comparable across methods?

1 comment

Thank you for the rebuttal and substantial revisions. I appreciate the authors' efforts in addressing my concerns through additional experiments, clearer theoretical explanations, and more detailed implementation analyses. These additions strengthen the empirical evidence and help clarify the scope and practical significance of the proposed framework. While some aspects could still benefit from further discussion, I consider my main concerns to have been satisfactorily addressed. Overall, I remain positive about the theoretical and methodological contributions of this work and will maintain my original rating of 3.

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

This paper introduces a dynamic quadratic optimal transport (QOT) formulation. The method uses travelling pairs and minimizes an objective that penalizes the deformation between pairs. They solve dynamic QOT using flow matching by training a neural network to predict a single-particle velocity field obtained by marginalizing the pair interactions.

On two synthetic rotation datasets, the method improves reconstruction accuracy while maintaining low pair distortion, as evaluated using spatial MSE, pair distortion, and Wasserstein distance. The authors also demonstrate competitive performance on the Mouse Brain and ARTISTA datasets and evaluate continuous 3D reconstruction using a tumor spatial transcriptomics dataset.

Critical strengths
  • The paper extends dynamic OT to QOT, which is, in my opinion, an important contribution. The authors establish equivalence between the static and travelling-pair dynamic formulations and provide extensive theoretical analysis of their method. Moreover, they introduce a general Lagrangian framework covering a broad class of pairwise interactions.

  • The experiments span both synthetic and real datasets, showcasing the possible applications of this method for spatial transcriptomics and 3D reconstruction.

Critical weaknesses
  • The measure of deformation is only computed at one intermediate point on the trajectory for both synthetic and real datasets. This gives a limited understanding of the ability of the method to preserve the structure of the dataset along the trajectory. Providing multiple values of the deformation along the trajectory would strengthen the claim.
  • On real datasets, the reported metrics show the reconstruction accuracy of spatial and expression but do not directly quantify the spatial structure distortion. Given the method’s focus on structure preservation, reporting FGW or GW would be a more direct evaluation of the claim.
Questions for authors
  • Do the runtimes in Figures 5 and 7 cover the full pipeline (optimizing the quadratic objective+training) or just flow matching training? If it is just training, can you report the time for quadratic objective optimization?
Reviewer_Lj11OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies a generalization of Gromov-Wasserstein optimal transport, called Quadratic-form Optimal Transport (QOT). Typical formulations of QOT are static as they are generalization of static Gromov-Wasserstein distance. The main contribution of this work is to develop a dynamic formulation of QOT. This is then paired with a flow-matching framework to learn the dynamics and solve the dynamic QOT problem in practice. Applications of the method in biology demonstrate some increased accuracy in trajectory reconstruction.

Critical strengths

The developed framework studies an important generalization of OT theory by showing how one can translate the dynamic OT framework into a dynamic QOT. This allows for a more flexible framework for modeling and comparing flows over probability distributions on different metric spaces. It is interesting to see the extension of path action Lagrangian formulations to the Gromov Wasserstein setting.

The method outperforms existing methods on biological data. It appears to be practically implementable, and extensions of it could offer more insight into real data.

Critical weaknesses

The paper could use some rewriting for clarity. While easier to follow than some, it still feels like large portions of this paper are AI-generated and not structured in a way that is easy for a human to read and understand. This includes patterns like frequent lists, jargon/terms that are not properly defined or motivated when introduced, and not enough expository text that allows for alternative explanations, along with signposting and roadmaps that make it easier for a reader to navigate.

The QOT problem is nonconvex, and the new formulation inherits it. The authors do not take any time or effort to address the optimality of their solutions to this problem here, or even why we might expect certain benign characteristics of the landscape. Experiments along these lines could be sufficient.

Questions for authors

It would be helpful to call out exactly what is meant by TP-DATE. Currently, it is a bit hidden and unclear. Does it refer to (10), (11), or does it refer to the flow-matching form (21)-(23)?

If the authors could comment on the nonconvexity and why we expect to be able to effectively optimize the objective here, that would be useful.

Reviewer_fR5sOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TP-DATE defines dynamic quadratic-form optimal transport through the cost of moving pairs of points. It derives closed-form paths for an interaction that penalizes changes in pairwise distance, then trains a single velocity network by averaging over these paths. The paper connects this formulation to standard OT and GW-OT. Experiments test spatial transcriptomics interpolation and reconstruction of tissue sections at unseen depths.

Critical strengths

The pair-based cost provides a concrete way to penalize changes in spatial structure throughout transport. The closed-form paths in unrestricted Euclidean space make this a useful approach to training without simulating trajectories during optimization.

The experiments include synthetic and real data, repeated runs, and an ablation that keeps the coupling fixed while changing the paths (Section C.1.3). This ablation helps show that the path construction contributes to the reported improvements.

Critical weaknesses
  1. The closed-form solution can give a cost lower than any path allowed by the problem. Section 4 allows a general space X⊂Rd\mathcal{X} \subset \mathbb{R}^d, with paths staying in X\mathcal{X}. But separating the midpoint cc and relative position qq does not remove the constraint that both c+q/2c + q/2 and c−q/2c - q/2 must stay in X\mathcal{X}. Eqs. (45)-(48) ignore this constraint. For a concrete example, take X=[−0.5,0.5]×[−0.05,0.05]\mathcal{X} = [-0.5, 0.5] \times [-0.05, 0.05], λ=34\lambda = \frac{3}{4}, both endpoint midpoints equal to zero, q0=(1,0)q_0 = (1, 0), and q1=(−cos⁡(0.1),sin⁡(0.1))q_1 = (-\cos(0.1), \sin(0.1)). All four endpoints lie in X\mathcal{X}. Eq. (45) gives a cost of about 1.901.90. Yet any valid path must reverse the horizontal component of qq. When that component reaches zero, the narrow rectangle forces ∥q∥≤0.1\lVert q \rVert \leq 0.1. The radius must therefore fall from 11 to at most 0.10.1 and return to 11. The radial energy alone is at least (0.9+0.9)2=3.24(0.9 + 0.9)^2 = 3.24. Thus, 1.901.90 cannot be the minimum cost of a valid path. These formulas determine both the coupling cost and training velocities. They solve a different problem unless X\mathcal{X} is unrestricted or the paths are proved to stay inside it.

  2. The GW limit in Theorem 4.4 is false for some probability densities allowed by the paper. The theorem claims that λ−1QOTS,λ(μ0,μ1)→GW-OT(μ0,μ1)\lambda^{-1}\mathrm{QOT}_{S,\lambda}(\mu_0,\mu_1) \to \mathrm{GW\text{-}OT}(\mu_0,\mu_1) as λ→∞\lambda \to \infty, but no finite second-moment assumption is stated. Consider a density f(s,u)=2/s3f(s,u) = 2/s^3 for s≥1s \geq 1 and 0≤u≤10 \leq u \leq 1, and let the target be its reflection through the origin. Reflection preserves every pairwise distance, so GW-OT(μ0,μ1)=0\mathrm{GW\text{-}OT}(\mu_0,\mu_1) = 0. However, the source has an infinite second moment, and every target point has a negative first coordinate. Every coupling therefore has infinite expected squared displacement. The kinetic term makes QOTS,λ\mathrm{QOT}_{S,\lambda} infinite for every finite λ\lambda, so λ−1QOTS,λ\lambda^{-1}\mathrm{QOT}_{S,\lambda} cannot approach zero. Eq. (104) only proves convergence for fixed endpoints; it does not justify passing the limit through the integral and minimization. This directly affects the claimed link to GW-OT. Finite second moments could repair this issue, and finite datasets satisfy that condition, but the theorem and proof need to state and use it.

Questions for authors
  1. Is X intended to be all of Rd\mathbb{R}^d? If bounded domains are included, how do Eqs. (45)-(48) enforce that both particles stay in X?
  2. What moment assumptions are required for Theorem 4.4, and how does the proof pass from the fixed-endpoint limit in Eq. (104) to the optimized transport cost?
  3. Under what conditions does the single-particle neural ODE reproduce the probability path in Theorem 5.2? How does this apply to the empirical distributions used by the implementation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.