acceptodds
Under review as a conference paper at ICLR 2027

Dynamic Neural Optimal Transport for Subset Alignment

Abstract

We study optimal transport between distributions when only a subset of the target population is relevant to the source. The goal is to map the source distribution to a reweighted target distribution whose density ratio with respect to the full target lies in . This constraint enables the transported distribution to select an appropriate target subpopulation while limiting how strongly any region of the target can be emphasized. We propose a dynamic neural optimal transport framework for this setting using two neural networks. The first parameterizes a time-dependent potential whose spatial gradient defines the velocity field and whose terminal value enforces the density-ratio constraint. The second parameterizes an interpolation between the source distribution and the learned terminal distribution. Unlike prior neural dynamic optimal transport methods, which typically assume a fixed terminal distribution, our formulation learns the terminal distribution within the support of the target. We demonstrate the method on positive-unlabeled learning, unpaired domain translation, and trajectory inference, where meaningful alignments exist between source and target distributions, but direct distribution matching is inappropriate because the target contains additional or irrelevant components.

Primary area:
generative models
Keywords:
Dynamic Optimal Transport, Wasserstein Distance, PU-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 3 reviews, above 27% of the others.
Reviewer_7Eb5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a dynamic neural optimal transport framework that maps a source distribution to a reweighted target distribution. This framework has connections to unbalanced and dynamic optimal transport. In the proposed semi-relaxed method, the authors formulate a Lagrangian that is optimized over two neural network variables. The algorithm then becomes a min-max procedure (gradient descent-ascent) over these two variables. Experiments demonstrate some advantages in practice.

Critical strengths

The formulation yields a novel min-max formulation that is different from existing unbalanced and partial OT methods. The formulation captures both dynamic and static formulations, which appear to be solvable effectively in practice.

Experiments demonstrate increased accuracy in domain translation, unlabeled learning, and image translation datasets.

Critical weaknesses

Not enough time is spent really digging into why, theoretically, this formulation is better than or motivated over existing unbalanced OT and partial OT frameworks, which also relax the marginal constraints. Furthermore, these also have dynamic formulations. While empirically there appear to be differences, this work would be stronger if these differences were based on foundational theoretical results stating the advantages of this work.

To optimize the Lagrangian, the authors formulate a min-max problem with neural network variables. This is reminiscent of GANs and other critic-based methods in ML. One concern is that these types of problems are frequently unstable and hard to optimize in practice. However, the authors do not explore this aspect of the problem. I am curious how hard it is to train these variables and what guarantees can be given that the problem can be solved effectively in general.

Questions for authors

If the authors could comment on the critical weaknesses above, that would be helpful. In other words, what theoretically differentiates the new approach? Examples and simple experiments could help to illustrate this.

Second, is the min-max procedure more stable than other variants that have appeared in the ML literature (e.g., GANs), or does this formulation suffer from the same issues? Is there a reformulation we could use without the min-max?

4 comments
AuthorsOpenReview ↗

Reply to Reviewer 7Eb5

We thank the reviewer for recognizing the formulation and its empirical benefits, and for raising two important questions.

Answer to Question 1

"...What theoretically differentiates the new approach?" We discuss the positioning of our work in Section 1, under “Positioning and contributions,” and in Section 2. The central motivation is to learn transport dynamics when the relevant terminal distribution is unknown. Our framework preserves the full source marginal and conserves mass along the path, while allowing the terminal distribution to vary under the constraint ρ1≤cν\rho_{1}\leq c\nu. This constraint provides an explicit bound on local reweighting. For any measurable target region A\mathcal{A}, ρ1(A)≤cν(A).\rho_1(\mathcal{A})\leq c\nu(\mathcal{\mathcal{A}}). For example, if A\mathcal{A} contains (10%) of the target mass and c=2c=2, it can receive at most (20%) of the transported mass, regardless of how inexpensive it is to reach. Moreover, any region containing all transported mass must have target mass at least 1c\frac{1}{c}. These properties give cc a direct interpretation and control concentration on a small part of the target.

Dynamic unbalanced OT also relaxes marginal matching, including through mass growth and destruction. Our formulation combines conservation of the entire source mass with a hard terminal density cap. A finite marginal-divergence penalty instead balances marginal discrepancy against transport cost. The hard cap is therefore motivated by settings where preserving the source population and controlling target reweighting are explicit requirements.

The endpoint constraint is closely related to partial and incomplete transport. Our contribution develops a dynamic neural formulation and solver for this free terminal marginal, including the corresponding boundary term and a learned probability path. The method jointly learns terminal reweighting and transport dynamics from samples without requiring ODE integration during each training step. It also provides intermediate distributions and a velocity field for trajectory generation and reverse integration. These capabilities motivate the dynamic formulation in applications such as population trajectory inference.

Answer to Question 2 (Stability and Alternatives)

Since our method uses neural networks in a saddle-point optimization problem, it inherits the associated optimization challenges. Training can depend on initialization, learning rates, and the relative frequency of potential and interpolant updates. The underlying population problem nevertheless has a convex formulation in density and momentum, with a convex kinetic action, a linear continuity constraint, and a convex terminal density constraint. This provides a principled variational basis, although it does not imply convergence of the nonconvex neural updates.

Our experiments, including repeated runs on 20 PU-learning datasets, demonstrate practical applicability and report variability in task performance. These results support the effectiveness of the training procedure in the evaluated settings, while a general convergence guarantee or stability advantage over GAN-style optimization is not established.

The min–max structure arises from our variational solution method. The underlying problem also admits a constrained primal minimization. For finite empirical distributions, one can solve a linear program with fixed source weights and capped target weights, avoiding alternating neural updates.

Another possibility is to construct minibatch couplings and train a flow from them. UOT-FM [1] follows this general strategy using unbalanced transport couplings. Adapting this strategy to our setting would require capped couplings that preserve the source marginal. Each coupling would then optimize over the target points available within its batch. Consequently, solving these local problems does not by itself guarantee recovery of the globally optimal capped transport.

Our approach directly optimizes the sample-based variational objective using shared potential and interpolant networks, avoiding a discrete transport solution within every batch. This is the computational motivation for the proposed saddle-point formulation. Both approaches involve approximation, but they differ in whether transport is learned through the variational objective or through separately computed minibatch couplings. We would further like to point out sections 1 and 2 from paper discussion positioning our work with respect to literature and the appendix C also discuss related work from both static and dynamic optimal transport.

We hope these clarifications would motivate the reviewer to reconsider their assessment of our work.

[1] Eyring, Luca, et al. “Unbalancedness in Neural Monge Maps Improves Unpaired Domain Translation.” International Conference on Learning Representations, 2024.

AuthorsOpenReview ↗

Follow-up to our response to Reviewer 7Eb5

Dear Reviewer 7Eb5, Thank you again for your feedback. With the discussion period ending on November 18, we would appreciate your assessment of our clarifications regarding the theoretical distinction from related transport methods, optimization stability, and alternatives to min–max training. We remain available to address any outstanding questions.

Thank you for taking the time to better explain what sets your method apart, and clearly writing this into text would be helpful. I am inclined to raise my score.

AuthorsOpenReview ↗

Reply to official comment by 7Eb5

We thank the reviewer 7Eb5 for thoughtful review and reply to our rebuttal. We would like to inform the reviewer 7Eb5 that the updated manuscript has been uploaded and all changes to the initial manuscript (both main body and appendix) are colored purple. Due to page length, we had to make the exposition in the rebuttal more compact, but we made the following changes to the text to better explain our method:

  1. Lines 119-125 (Section 2, Related Work) specifically discuss a mini-batch alternative for subset alignment and its difference from our approach and cite the related reference.
  2. For limitations regarding stability associated with mini-max training we updated the text between lines 515-519 (Section 5, Limitations) and connected it to the updated Related Work via a footnote.

Additionally, we would like to mention that we made minor changes to Sections 1 and 2 to better position our work in relation to directly related work. While not updated, detailed literature surveys for both static and dynamic neural optimal transport are given in Appendix C.

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

They propose an optimal transport approach that achieves alignment between the source and target distribution where only a subset of the target distribution is relevant to the source. To achieve it, they propose to constrain the matched target samples to lie in some range of the region. Building on top of this idea, they introduce a dynamic neural optimal transport framework using two networks. In my understanding, one network is learned to choose which target samples are relevant to the source, i.e., defining the true target distribution. The second one is learned to move the source samples to the target. Empirical results indicate that the proposed approach can be effective in pu-learning, domain translation, and trajectory inference.

Critical strengths
  1. The optimal transport to the subset of the target data is practical and interesting. They propose a method to tackle the task, and the method formulation overall sounds reasonable.
  2. They test their approach on diverse settings, trying to show how generalizable their approach is. They are not performing best in all settings, but overall, the performance is good compared to the baselines.
Critical weaknesses
  1. My critical concern is about how to choose the subset of the target distribution. I guess their proposed approach is measuring the distance between the source and target using some metric, either in an implicit or explicit way. However, how we define the similarity will dramatically change the results of the experiments. For example, in the case of the digit experiment, if the similarity metric favors class-related features, we can correctly choose the target subsets from 0-9 classes. However, it is also possible that the metric favors the style of the digits. Then, the metric probably choose target digits that have similar writing style to the source examples.

  2. Another point is how to define the range, c. In my understanding, the value defines how much of the target data to choose as the real matching target. If we know the range beforehand, we can remove some proportion of the target data from training data, and learn full optimal transport from the source and reduced target?

  3. In summary, my concerns of two above indicate that the metric function to measure source and target distance, and the proportion of the target data to choose can matter the most. However, they do not seem to have the analysis for that.

Questions for authors

Please address the question 1 and 2, which will address my concerns.

4 comments
AuthorsOpenReview ↗

Reply to Reviewer DNSY

We thank the reviewer for the comments on our work. Below are our answers to the questions regarding the choice of metric and the parameter cc.

Answer to Question 1

This is an important practical concern. Optimal transport requires a transport cost through which source and target data points can be meaningfully compared. In our experiments, we use the quadratic cost associated with the Euclidean metric in the chosen data representation. Therefore, both the representation and the metric influence the selected terminal distribution.

We agree that a metric useful for label alignment might not be equally useful for style alignment, and vice versa. In the digit example, if the representation space primarily reflects handwriting style, the transport can favor transporting near target samples with similar styles. Our formulation identifies a low-cost terminal distribution subject to the density-ratio constraint. Its usefulness for semantic alignment therefore depends on whether the chosen metric reflects the relevant class structure. Other task-relevant representations, or transport costs other than Euclidean, can also be considered, with corresponding adaptations to the formulation when needed. One example is that one could use metric learning or learn a new encoding to create a latent representation space that emphasizes important aspects, coupled with a decoding to map back to the input space.

The reviewer’s observation therefore identifies an important modeling assumption underlying all optimal transport based approaches. Our image experiments demonstrate empirical alignment under the chosen representations using Euclidean distance, rather than a guarantee of semantic alignment if it is not well represented.

Answer to Question 2

Depending on the problem, cc can be specified using available prior information. For example, in our PU-learning experiments, we use c=1π+c=\frac{1}{\pi_+}, where π+\pi_+ is the known positive-class proportion in the unlabeled population. When this proportion is unknown, mixture proportion estimation methods such as [1] can provide an estimate under their identification assumptions. Incorporating an estimated prior is a possible extension that we have not evaluated in the reported experiments and is left for future work.

We would also like to clarify that cc bounds the amount of reweighting rather than fixing the exact proportion of points selected. For any measurable target region A\mathcal{A}, the constraint gives ρ1(A)≤cν(A).\rho_1(\mathcal{A})\leq c\nu(\mathcal{A}). If all transported mass lies in A\mathcal{A}, then

\qquad\text{and hence}\qquad \nu(A)\geq\frac1c.$$ Thus, $\frac{1}{c}$ is a lower bound on the target mass needed to support the terminal distribution. The solution can retain a larger region and assign different weights within it. At $c=1$, normalization forces $\rho_{1} = \nu$, while larger values permit stronger reweighting and exclusion. If the relevant subset were already known, one could filter the target and perform balanced optimal transport to its conditional distribution, provided that distribution satisfies the cap. However, knowing the proportion alone does not identify which points should be retained. Their selection still depends on the transport cost and the geometry of the source and target distributions. Moreover, filtering fixes the relative weights within the retained population, whereas our formulation jointly optimizes those weights and the transport. As already discussed in Section 5 (Limitations), our framework could also be extended toward automatic selection of $c$. Specifically, we discuss using the resulting transport cost to guide its selection, motivated by previous work on partial optimal transport in the discrete setting (Phatak et al., 2023). We also propose conditioning the networks on $c$, through scalar embeddings analogous to the time embeddings, and jointly training over a prescribed range of values. This would allow a shared model to represent solutions for different values of $c$ and facilitate their comparison. These are directions toward automatic selection, while the current experiments either use known prior information or evaluate (c\in{1,2,4,8}). We hope these clarifications would motivate the reviewer to reconsider their assessment of our work. [1] Garg, Saurabh, et al. “Mixture Proportion Estimation and PU Learning: A Modern Approach.” Advances in Neural Information Processing Systems 34 (2021): 8532–8544.

Thanks for the response. The response clarified my concerns about the first question.

I have another questions about c. In many tables, the authors are showing results on different c for the proposed approach. In some cases, the baseline is also sweeped on the hyper-parameters, in some cases, there is no sweep. I would like to ask how we should interpret the difference.

Also, what is the intuition from Table 5 and Fig. 5?

AuthorsOpenReview ↗

Answer Regarding “c” to Reviewer DNSY

Table 5 and figure 5 correspond to each other. Other baselines except static and UOTM has been reported from Gazdieva et al (2024). Whereas static baseline is static version of our own framework. This is the reason we varied the c parameter for that too. UOTM framework is very close our static framework, we just used our own static neural networks with slight modifications to implement UOTM and we used best performing tau parameter from UOTM paper to run it. So essentially we did parameter sweeping for static version of our own framework.

The key takeaway from these experiments was the fact that as c is increased, subset selection selection choses target points which on average are close to source, which results in decreased target accuracy but increase in source accuracy. This observation is consistent with other methods too. Methods with better performance in terms of target accuracy do bad in terms of source accuracy and vice versa. The reasoning for c is the fact that as c is increased velocity fields lands points near and near the source points, thereby reducing target accuracy but source accuracy remains higher.

AuthorsOpenReview ↗

Follow-up to our response to Reviewer DNSY

Dear Reviewer DNSY, Thank you for engaging in the discussion and acknowledging our clarification regarding the metric. As the discussion period ends on November 18, we would appreciate any further comments on our response regarding the choice of cc, the baseline settings, and the interpretation of Table 5 and Figure 5.

Reviewer_2YzWOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a neural solver for dynamic optimal transport with a relaxed terminal marginal constrained by (\rho_1\le c\nu). It jointly learns a potential and a distributional path, with applications to image translation, positive-unlabeled learning, and cellular dynamics. The formulation is useful, but the contribution appears primarily algorithmic. An incorrect subset-identification claim, inconsistent numerical results, and insufficiently controlled comparisons currently prevent a convincing assessment of the method.

Critical strengths
  1. The problem is well motivated: matching an entire target distribution can be inappropriate when it contains irrelevant components. The density-ratio bound controls reweighting in an interpretable way.
  2. The method trains on samples without integrating the ODE during each step. Experiments cover several applications, include static comparisons and different values of (c), and report variability across 20 runs for PU learning.
Critical weaknesses
  1. The central claim that the potential’s sign identifies the selected subset is false as stated. Equations (9), (53), and (55) require (\phi_1>0) on the rejected region. The optimality conditions only imply (\phi_1\ge0) there; zero potential is compatible with both selected and rejected points. For example, with (\mu=U[0,1]), (\nu=U[0,2]), and (c=2), the identity map and (\phi_t\equiv0) are optimal, but thresholding at (\phi_1\le0) selects the entire target. This affects PU learning: under SCAR, (\mu=P_+), (\nu=\pi P_++(1-\pi)P_-), and (c=1/\pi) imply (\mu\le c\nu). The identity map and zero potential are again optimal. Additional assumptions or a different identification rule are needed.
  2. The reported results contain substantial inconsistencies, and the static baseline’s pseudocode has an incorrect update sign. For (c=4), the confusion matrices in Figure 13 yield approximately 72.95% and 73.66% accuracy for one-step and ODE generation, whereas Table 1 reports 93.57% and 92.68%. Several borrowed PU results also differ from the published source: for example, NTC-MI on Spambase is reported as 0.887 rather than 0.837. Teisseyre et al., 2025. Moreover, Algorithm 2 performs gradient ascent on the negative of the potential-dependent objective in (27). This may be a pseudocode error, but the actual implementation must be verified before relying on comparisons against the static method.
  3. The novelty is narrower than the positioning suggests, and important comparisons are missing. The constraint (T_#\mu\le c\nu) already defines incomplete transport in Gazdieva et al., 2023, while the potential/path parameterization and saddle-point training closely follow Wasserstein Lagrangian Flows. The specific contribution is the dynamic solver with a free terminal marginal and the corresponding density-cap boundary term. UOT-FM already learns dynamic transport between reweighted distributions and is a particularly relevant missing baseline. It does not impose the same hard density cap, but comparison is necessary to establish the benefit of that distinction. For PU learning, comparisons with partial OT and nnPU are also missing.
  4. The experiments do not establish that the method reliably solves the stated problem or improves over alternatives. There are no quantitative checks of density-cap violations, transport-cost optimality, or agreement between the learned interpolant and ODE-generated distributions. Image-classification accuracy can reward preserving the input rather than successful translation, making an identity baseline essential. PU comparisons confound the method with model capacity and prior information: the proposed models use deep MLPs and a known class prior, whereas the borrowed UCI baselines use logistic regression, and PUSB estimates its prior. For melanoma, the best proposed MMD occurs at (c=1), providing no evidence for the benefit of subset selection. Furthermore, filtering (\nu) by a selected set produces (\nu(\cdot\mid A)), which need not equal the learned reweighted measure (r\nu). Consequently, the reverse-transport evaluation requires additional justification.
Questions for authors
  1. What assumptions resolve the zero-potential ambiguity and justify subset identification, particularly in the SCAR setting?
  2. Can the tables and confusion matrices be reconciled, and can the actual update sign used for the static baseline be verified?
  3. Does the method retain an advantage over UOT-FM, partial OT, and nnPU under comparable architectures, prior information, and computational budgets?
  4. Can you report constraint violations and transport optimality, and justify why the filtered melanoma population represents the learned terminal measure?
3 comments
AuthorsOpenReview ↗

Reply to Reviewer 2YzW part 1

We are grateful to the reviewer for a thorough and comprehensive review of our work. It truly helped us analyze our work better and bring rigor in our statements. A revised manuscript has been updated, changes made are colored purple, and in the discussion below we will mention the changes made in accordance with weakness pointed out and questions asked by reviewer.

Weakness 1 and Question 1

We thank the reviewer for finding an ambiguity in our statement regarding the sign of potential function. We have updated the discussion with equation 9 and appendix B. In appendix B, it is shown that under the condition that there is no overlap between support of source and rejected set, one can find an optimal potential, where sign-based subset selection is valid up to ν\nu-null set. But we would like to emphasize that although such optimal potentials exist, our algorithm does not guarantee finding any such solutions, therefore our sign-based selections are essentially heuristics which happen to work in many scenarios.

Weakness 2 and 3 and Questions 2 and 3:

We are grateful to the reviewer for careful reading and finding important presentation related errors and typos in paper. Here is brief summary of changes made as suggested by reviewer:

  1. To make Table 1 and Figure 13 compatible, we updated the actual confusion matrices corresponding to results reported in Table 1. We also updated the confusion matrices for the EMNIST-trained classifier evaluated on EMNIST and MNIST. Table 1 has been updated with the result for identity baseline too.
  2. We also reverified the Table 2 results from Teisseyre et al. (2025) and fixed the previous typos.
  3. Algorithm 2 has been fixed to clearly make it compatible with corresponding objective functions.
  4. For PU learning, in addition to baselines from Teisseyre et al. (2025), as suggested by the reviewer, we did additional comparisons with nnPU, UOT, UOT-FM and POT. Table 2 in the updated manuscript summarizes the comparison of average balanced accuracies. We made sure that neural network based methods (nnPU and UOT-FM) have comparable architectures to that of our method, architectural and algorithmic details for all baselines have been updated in Appendix E.5. We observed that our algorithm maintains a comparable or advantageous position over neural network-based methods throughout, especially in the cases where positive and unlabeled classes are geometrically separable.
  5. Regarding Novelty Concerns In some aspects our framework lies between Wasserstein Lagrangian Flows and incomplete transport (Gazdieva et al., 2023) and in other aspects it extends beyond both. We agree that our framework is similar to Wasserstein Lagrangian Flows because of the saddle point nature of problem and to incomplete transport due to the transport to target subset but we would like to emphasize that to the best of our knowledge, our framework is first to extend the subset alignment to dynamic settings. We would like to further point out that our framework can be extended to subsetting across multiple chronologically ordered marginals sampled from some process, with the same potential function, whereas static setting of incomplete transport would require multiple parametric models for each subset alignment.
AuthorsOpenReview ↗

Reply to Reviewer 2YzW part 2

Weakness 4 and Question 4

Density cap violations and Optimality We updated Section 4.3 to contain a discussion on density cap violations and optimality of the transport solution. Detailed analysis of density cap violations and trajectory curvature for PU learning with dynamic subset alignment is given in Appendix H of the revision. Table 3 in the main body gives a summary of density cap violations and straightness analysis, whereas Table 9 in appendix H gives the average metrics for 20 runs for each dataset. In order to evaluate “agreement between the learned interpolant and ODE-generated distributions”, we consider the Wasserstein distance and MMD between the empirical distribution of the actual positive subset of the unlabeled compared to either the output of the trajectory or the interpolant at t=1t=1. In the vast majority of cases outputs of trajectories generated by flows are closer to actual positives in terms of Wasserstein distance as compared to interpolant at t=1t=1 (see Table 9). This gap highlights clear room for improving the interpolant parameterization QtQ_t.

Filtered melanoma population and the terminal measure We thank the reviewer for raising this distinction. Let B=x:φt1(x)≤0\mathcal{B}=\\{x:\varphi_{t_1}(x)\leq0\\} denote the retained region and ν\nu the observed 8-hour population. Filtering gives ν(⋅∣B)\nu(\cdot\mid \mathcal{B}), which equals the learned terminal measure ρt1\rho_{t_1} only when

dρt1dν(x)=1B(x)ν(B)ν-almost everywhere.\frac{d\rho_{t_1}}{d\nu}(x) = \frac{\mathbf{1}_\mathcal{B}(x)}{\nu(\mathcal{B})} \qquad \nu\text{-almost everywhere}. Thus, identifying the selected region is insufficient unless its relative weights also agree. At optimality, complementary slackness justifies excluding the strictly positive-potential region, but does not establish this equality, particularly on the zero level set.

Our melanoma experiment therefore evaluates reverse transport from a sign-filtered population. The reported MMD and W2W_2 assess agreement with the observed 48-hour population after reverse transport; they do not independently establish that the filtered 8-hour population equals the learned terminal measure. The sign rule is motivated by the optimality conditions, while equality of these distributions requires the additional condition above.

Regarding minimum MMD at c=1 We agree that minimum MMD is achieved at c=1c=1; nonetheless, c=2c=2 achieves the best W2W_2 across all methods. This pattern matches the pattern for Wasserstein Lagrangian flows, where minimum MMD is achieved for no growth but the W2W_2 is lower when growth is allowed. This could be due to the fact that empirical estimators for MMD are less sensitive to outlier points compared to W2W_2, and W2W_2 is the cost that is minimized. Therefore, for our purposes, where our aim is to use data geometry and Wasserstein distance-based approach for subsetting and measuring reweighting, W2W_2 is the consistent evaluation metric.

AuthorsOpenReview ↗

Follow-up to our response to Reviewer 2YzW

Dear reviewer 2YzW, Thank you again for your detailed feedback. With the discussion period ending on November 18, we would appreciate your assessment of our responses and the revised manuscript, particularly the corrected sign conditions, additional baseline comparisons, and transport diagnostics. We remain available to clarify any outstanding concerns.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Message to reviewers and AC part 1

Dear Reviewers and Area Chair,

We thank the reviewers for their thoughtful comments and discussion, which helped us improve the theoretical statements, experimental evaluation, and presentation of our work. The updated manuscript has been uploaded, and all changes to the initial submission, both in the main body and appendix, are colored purple\textcolor{purple}{\text{purple}}. Below, we summarize the changes made in response to the reviewers' concerns.

Sign conditions and subset identification\textcolor{blue}{\text{Sign conditions and subset identification}}

Following Reviewer 2YzW\texttt{2YzW}'s helpful observation, we updated the discussion accompanying Equation 9 and expanded Appendix B to clarify the sign conditions obtained from complementary slackness and the possible ambiguity on the zero level set. Under population SCAR and an additional support condition, we showed that one can construct an optimal potential for which thresholding identifies the selected region up to a ν\nu-null set. We further showed that this potential is optimal in both the endpoint and dynamic formulations. This provides a theoretical basis for thresholding under the stated conditions. We also clarified the distinction between this construction and the potentials obtained during neural network training. Accordingly, sign-based and sorting-based selection are described as empirical rules motivated by the optimality conditions.

Corrections and additional comparisons\textcolor{blue}{\text{Corrections and additional comparisons}}

As suggested by Reviewer 2YzW\texttt{2YzW}, we updated the confusion matrices corresponding to the results in Table 1 and added the identity baseline. We also reverified the borrowed PU-learning results against the published source and corrected the corresponding entries. The static implementation used the correct update; the error was confined to the pseudocode, which has been corrected in Algorithm 2.

For PU learning, we added comparisons with nnPU, partial OT, UOT, and UOT-FM. Table 2 summarizes the results across 20 datasets, while Appendix E.5 gives the architectural and training details for the added baselines. Under the reported experimental settings, dynamic subset alignment with sorting achieves the best average rank of 2.950 among the evaluated methods.

Density-cap violations and transport diagnostics\textcolor{blue}{\text{Density-cap violations and transport diagnostics}}

We updated Section 4.3 and added Appendix H to evaluate estimated density-cap violations, trajectory straightness, and distances to the held-out positive-class distribution. Table 3 gives a summary, whereas Table 9 reports the mean and standard deviation over 20 seeds separately for each dataset. The estimated fraction of transported points exceeding the cap is at most 0.1 on 13 of 20 datasets, and normalized trajectory curvature is at most 0.1 on 16 datasets. The discussion also identifies datasets with larger estimated violations.

Additionally, we evaluate both the flow output and interpolant endpoint against the held-out positive-class reference using W2 and MMD. These measurements provide information about different aspects of the learned transport. We explicitly clarify that trajectory straightness alone does not establish global transport optimality.

AuthorsOpenReview ↗

Message to reviewers and AC part 2

Choice of metric and parameter c\textcolor{blue}{\text{Choice of metric and parameter }c}

In response to Reviewer DNSY\texttt{DNSY}, we clarified how the transport cost and the chosen data representation influence the selected terminal distribution. We further explained that cc controls the permitted reweighting rather than fixing the exact proportion of selected points. Knowledge of this bound does not identify which points should be retained; their selection also depends on the geometry of the source and target distributions.

Our follow-up response explains the baseline parameter settings and the trade-off between source preservation and target alignment observed in Table 5 and Figure 5. As discussed in Section 5, automatic selection of c is a possible extension of the framework.

Melanoma evaluation\textcolor{blue}{\text{Melanoma evaluation}}

Following Reviewer 2YzW\texttt{2YzW}'s comment, we clarified in Section 4.5 and Appendix B the distinction between the sign-filtered population and the learned terminal measure. Identifying the retained region is sufficient to recover the terminal measure only when the relative weights within that region also agree. Therefore, the experiment is described as reverse transport from a sign-filtered population. The reported distances evaluate agreement with the observed 48-hour population after reverse transport.

Positioning and optimization stability\textcolor{blue}{\text{Positioning and optimization stability}}

Following Reviewer 7Eb5\texttt{7Eb5}'s feedback, we made changes to Sections 1 and 2 to better position our work in relation to incomplete transport, Wasserstein Lagrangian Flows, and UOT-FM. Our framework jointly learns transport dynamics and a free terminal marginal under a hard density cap while preserving the source mass.

Lines 119–125 in Section 2 discuss an alternative based on capped minibatch couplings and explain its difference from our approach. For the limitations regarding stability associated with neural min–max training, we updated the discussion in lines 515–519 of Section 5. Appendix C additionally provides a detailed discussion of related work on static and dynamic neural optimal transport.

We thank the reviewers and Area Chair for their time and consideration. We hope that the clarifications, additional comparisons, and transport diagnostics help address the concerns raised during the review. We remain available to clarify any outstanding questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.