acceptodds
Under review as a conference paper at ICLR 2027

Potential Matching Optimal Transport: Continuous Normalizing Flows for Exact -Wasserstein Dynamics

47 views

Abstract

We introduce Potential Matching Optimal Transport (PMOT), a potential-flow framework for general -cost optimal transport with . PMOT parameterizes the CNF velocity field with a scalar potential in the generalized Benamou–Brenier form for the chosen exponent . It trains the potential gradient with a self-induced matching loss along straight bridges determined by the model's own endpoints, while allowing flexible terminal distribution matching. Our main result establishes zero-loss exactness: under the stated regularity, exact terminal matching, and uniqueness assumptions, any zero-loss solution satisfies the generalized Benamou–Brenier optimality system and recovers the corresponding -optimal transport map and dynamics. On synthetic benchmarks, PMOT learns -specific maps that agree with the corresponding -matched OT references. It also remains competitive as a likelihood-based density model on high-dimensional tabular data, and an MMD-based color transformation experiment demonstrates flexible sample-based terminal matching.

Primary area:
generative models
Keywords:
optimal transport, p-Wasserstein geometry, potential matching, continuous normalizing flows, generative modeling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ve1mOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a continuous normalizing flow-based approach for dynamic Wasserstein-p dynamics. Training loss employed to train the flows consist of two weighted terms: (1) Potential matching loss, which is derived using Wasserstein-p extension of Benamou-Brenier formulation of dynamic optimal transport. (2) Terminal distribution loss, the term to enforce the matching for terminal distribution. The paper further applies the proposed methodology to toy transport geometry, density estimation and image color transformation problems.

Critical strengths

The paper studies the dynamic optimal transport, which is helpful in the better understanding of existing models and can be helpful in the development of new generative models.

Critical weaknesses
  1. The preliminaries section of the paper does not properly motivate the discussion or establish the background. None of the equations discussed in Section 2.1 (for example, the Kantorovich problem, McCann interpolation, and the Hamilton-Jacobi-Bellman dual corresponding to the continuity equation) were originally proposed by the authors, but there is no citation whatsoever. In line 92, the existence of the Monge map is discussed without even defining it. Sections 2.2 and 2.3 do not properly establish the background required to understand the work done in the paper and fail to cite any references.

  2. The main contributions of the paper are discussed in Section 3, but due to the poor development of ideas in Section 2, Section 3 also suffers. For example, Equation 1 is expressed as if it were a fact without any citations, and the discussion after Equation 2 contains redundant, non-standard notation, which requires careful reading for it to make sense. Also, the pushforward has been denoted by # in the main body and by ♯\sharp in the appendix.

  3. The discussion in Section 3 does not motivate the weighted objective; to fully understand the matching objective, one has to read lines 559–602 in the appendix. If one does not have sufficient background in dynamic optimal transport or generative models, it is almost impossible to follow Section 3.2. In Section 3.3, the discussion before and after Theorem 3.2 does not properly relate it to the subsequent development of the neural network parameterization framework.

  4. The proof of theorem in the appendix is difficult to follow. One primary reason for this is the fact that the equations are not numbered, and the arguments are stated without any discussion.

  5. Technically, the total potential matching optimal transport loss has three optimization variables, i.e., two weights and one flow variable. The paper fixes two scalars and parameterizes the potential using neural networks, which can result in suboptimality. The paper does not discuss how the scalar weights are chosen.

  6. The experiments are lacking in both breadth and depth. The majority of the experiments are done on toy point clouds and are not fully representative of the capabilities of neural optimal transport and related generative models.

  7. From section 3.2, it seems that optimal trajectories should be straight lines but trajectories in Figure 2(c) are not straight.

Questions for authors
  1. It seems that potential matching objective LPM\mathcal{L}_{\textrm{PM}} is directly related to OT conditional flow matching and rectified flow matching, authors should elaborate on that.

  2. In line 259, we see that one needs reference terminal density model to evaluate the negative log-likelihood, which was never mentioned before.

  3. As discussed in weakness 7 above, the trajectories of learned flow maps are not straight lines, which can be interpreted as suboptimality of learned flow maps, if in practice, transport plan is suboptimal after employing actual learnt flow maps, what is the advantage of potential matching loss as compared to Flow-matching or Stochastic interpolant type objectives, as all estimate certain Wasserstein distances.

6 comments
AuthorsOpenReview ↗

Response to Reviewer ve1m (1/2) : Weaknesses W1–W7

Thank you for taking the time to read our paper so carefully and for the thoughtful questions you raised about both the derivations and the experiments. In the revised manuscript, we mark changes addressing your comments in teal. Bracketed reviewer IDs indicate changes that address comments from more than one reviewer.

W1. Preliminaries and references.

A1. You are right that we omitted references for several standard results and used the term “Monge map” before defining it. We now cite Villani for static OT, Santambrogio for dynamic OT and displacement interpolation, and Chen et al. for CNFs. We also define couplings, Monge maps and pushforwards before explaining how the velocity along a displacement interpolation motivates our training target. (Sections 2.1–2.3, lines 91–107, 110–149 and 155–174.)

W2. Equation (1) and notation.

A2. We derive Eq. (1) by varying the action with respect to velocity and inverting the momentum relation. After Eq. (2), we define the trajectory, endpoint and time-tt law in turn. We use #\# for pushforwards throughout. (§§2.2 and 3.1, lines 119–136 and 186–206, Appendix A.1, lines 630–631.)

W3. The two losses and the neural model.

A3. A constant potential leaves particles in place. If the marginals differ, matching loss is zero but terminal loss is positive. Several maps can match the same target, so we also match momentum along each bridge. Under its assumptions, Theorem 3.2 gives the optimal map and dynamics at zero population loss. We train a neural potential on the empirical objective. (§§3.2–3.4, lines 210–225, 244–258 and 266–269.)

W4. Proof of Theorem 3.2.

A4. ODE uniqueness identifies the trajectory with its straight bridge in (A.1). The adjusted potential satisfies the Hamilton–Jacobi equation in (A.2). Fenchel’s inequality gives (A.3), showing that permuting destinations cannot lower total cost. (Appendix A.1, lines 721, 829 and 887.)

W5. Choosing the weights.

A5. One practical advantage was that PMOT needed no dataset-specific weight tuning. We used (λPM,λNLL)=(1,5)(\lambda_{\mathrm{PM}},\lambda_{\mathrm{NLL}})=(1,5) on 8-Gaussians, Pinwheel and three tabular datasets. On 8-Gaussians, PMOT’s map MSE was 0.01971 at weight 5 and 0.01924–0.01971 in a separate 60-fold NLL weight sweep. OT-Flow reached 0.04815 at 5 and comparable errors at 50 and 300 (0.02000 and 0.01969). On Pinwheel, OT-Flow at 100 approached PMOT at 5 in reverse W22W_2^2 (0.03430 vs. 0.03421). We fit the potential with fixed positive weights, preserving any common zero. (§3.2, lines 227–230, §4.2, Table 1, lines 424–431 and 450–466.)

W6. Breadth and depth of the experiments.

A6. We evaluated density estimation and generation on three high-dimensional tabular datasets (Table 2). In the reported runs, PMOT had lower test NLL than OT-Flow on MiniBooNE, POWER and HEPMASS. For generation, PMOT had lower MMD² on POWER and OT-Flow had lower MMD² on the other two. Figure 4 shows image color transfer using samples from the target color distribution. (§4.3, lines 470–477, 487–497 and 505–506.)

W7. Moons trajectories.

A7. Exact straightness holds at zero population loss. The fitted Moons model has a positive matching residual, so slight bending is expected. We chose (2,20)(2,20) by generation quality and used only eight RK4 steps per training solve. For all 900 points in Figure 3(c) (originally 2(c)), paths averaged 1.26% longer than their chords at both 32 and 64 evaluation steps. (§§3.3 and 4.1, lines 244–256, 414–418 and 445–447.)

AuthorsOpenReview ↗

Response to Reviewer ve1m (2/2): Questions Q1–Q3

Q1. How does PMOT relate to OT-CFM and rectified flow?

A1. At p=2p=2, a pair (x,y)(x,y) gives all three methods the same velocity target u=y−xu=y-x on the bridge x+tux+tu. The endpoint yy is chosen differently. OT-CFM obtains pairs by solving OT within each minibatch. Rectified flow fixes the coupling within each round and can use the fitted flow to generate the next round’s pairs. PMOT uses its own current endpoint y=Fθ(x)y=F_\theta(x). Gradients pass through yy in both x+tux+tu and the momentum target ∥u∥p−2u\|u\|^{p-2}u for general pp. (§§3.2, 3.4 and 5, lines 234–239, 273–289 and 514–522.)

Q2. Which terminal density enters the NLL?

A2. When the target log density is available, we evaluate it at Fθ(x)F_\theta(x) in the CNF likelihood. Training needs no target samples, even when sampling is difficult. With samples alone, the same flow can match endpoints by MMD or energy distance. We use a Gaussian density on tabular data and MMD on RGB samples. (§§2.3 and 4.3, lines 162–174 and 475–477.)

Q3. What does matching offer when paths bend?

A3. Finite training rarely attains zero loss, so fitted paths can bend. FM and stochastic interpolants fit a chosen interpolation. PMOT learns the pairing under the chosen pp-cost from unpaired data, without an OT solver. On two disks, PMOT and minibatch pp-OT-FM both sent every point to the correct disk after 1,200 updates. PMOT’s mean pointwise map MSE was about 1/16 of OT-FM’s at p=1.5p=1.5 and 1/283 at p=3p=3. On the square, its time-integrated MSE against the exact OT paths was 1.74×10−31.74\times10^{-3} and 3.19×10−43.19\times10^{-4}, respectively. (§4.1, lines 326–345 and 381–389.)

We hope the revised paper is easier to follow and that the additional experiments give you greater confidence in our results.

AuthorsOpenReview ↗

Follow-up on our response to Reviewer ve1m

Dear Reviewer ve1m,

We would appreciate your feedback on our point-by-point responses and revised manuscript. We have added experiments and theoretical clarifications addressing your comments, with references to the relevant sections and line numbers.

With the author–reviewer discussion closing on November 18, could you let us know whether our responses address your main concerns and which issues, if any, remain? Hearing from you before the deadline would give us time to provide further clarification or revisions.

Thank you for your time and consideration.

Reviewer_Fa17OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a new approach for calculating the solutions of the pp-Wasserstein optimal transport (OT) problem, i.e., recovering the associated transport maps and dynamics. The method belongs to the family of continuous normalizing flows (CNF) models whose optimization objective is inspired by the Benamou-Brenier optimality principles. The method is supported by the theoretical result showing that its objective allows for recovering the solutions of the pp-Wasserstein OT problem in the zero-loss setting and several experiments.

Critical strengths

The overall idea of finding solutions to the pp-cost OT problem using potential-based CNF is interesting.

Critical weaknesses

[W1] The method is supported theoretically by the so-called “zero-loss exactness” result (Theorem 3.2) which shows that the potential which delivers zero value to the method’s objective recovers true pp-optimal mapping and the corresponding dynamics. However, this zero-loss scenario almost never appears in practice. Thus, this theoretical result seems to be a rather weak justification of the developed approach. I would rather expect the result showing that the method’s loss upper bounds the error in the recovered mapping. Further theoretical investigation might also include the bounds on approximation and statistical errors, etc.

[W2] Another limitation of Theorem 3.2 is the reliance on the assumption of nondegeneracy of the support ptΦp_t^{\Phi}. Due to this limiting assumption, the established theory might not hold true in many experimental setups. For example, it seems that this assumption does not hold for the experiment with image color transformation (lines 409-411). I kindly suggest the authors comment on this point.

[W3 (a)] The provided experiments also do not sufficiently support the thesis that the method allows for recovering good approximations of optimal transport maps. The closest experimental setup is considered in section 4.1. Here the authors assess the results of their approach taking the results of discrete solvers as the ground truth. However, discrete OT plans endowed with barycentric projection strategy are known to produce mappings which might be quite far from actual OT ones. Thus, I kindly suggest the authors consider testing their approach in the experimental setup where the ground-truth mapping is analytically known, e.g., this should be true for the Gaussian case (with some pp).

[W3 (b)] This concern is also related to experiment in section 4.1. Here the assessment is performed in terms of the delivered cost and recovered maps; the results are summarized in their Table 1. However, it is hard to understand the quality of the approach without the comparison with the baselines, e.g., OT-Flow. I kindly suggest the authors to include the comparison with this approach in their Table 1.

[W4] Another weakness is related to the limited clarity of the paper, including non-explained abbreviations (BB in line 128, AC in line 167, HJB and RK4 in lines 267-268) and theoretical transitions without sufficient explanations, e.g., derivations of equations in lines 115, 119, 122, objective in line 182 and many others could be explained more deeply (maybe in Appendix if the space is limited). From my point of view, without these explanations the paper is hard to follow.

Overall, I am not convinced that the proposed approach indeed allows for obtaining good approximations of pp-optimal transport maps and dynamics. This is neither sufficiently supported from the theoretical side, nor from the practical one.

Questions for authors

[Q1] Am I right in understanding that the assumptions of your Theorem 3.2 do not hold for the experiment with image color transformation?

[Q2] Could you test your approach in the experimental setup where the ground-truth mapping is analytically known?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Fa17

Thank you for your careful review and for suggesting a test against an analytically known transport map. Edits specific to your review are blue. Bracketed IDs mark shared revisions.

W1. Can the loss control the error in the recovered map?

A1. Yes, under additional assumptions. Let FF be the learned map and T⋆T^\star the optimal map to the prescribed target μ1=N(0,I)\mu_1=\mathcal N(0,I). Appendix B bounds the population map error dF2=EX∼μ0∥F(X)−T⋆(X)∥2d_F^2=\mathbb E_{X\sim\mu_0}\|F(X)-T^\star(X)\|^2 even when both the momentum-matching loss ℓ\ell and the terminal mismatch η=KL(F#μ0∥μ1)\eta=\mathrm{KL}(F_\#\mu_0\Vert\mu_1) are positive. The small-loss bounds for dF2d_F^2 are

  • p=1.5p=1.5: O(ℓ2/9+η1/4)O(\ell^{2/9}+\eta^{1/4}),
  • p=2p=2: O(ℓ2/3+η1/2)O(\ell^{2/3}+\eta^{1/2}),
  • p=3p=3: O(ℓ1/3+η1/3)O(\ell^{1/3}+\eta^{1/3}).

These rates require time regularity of the bridge residual near t=1t=1, a differentiable convex terminal potential with the stated integrability, and uniform control of the relevant moments and growth constants. We have not verified these conditions for the fitted networks. Finite-sample and network approximation errors remain open. (Section 3.3, lines 254–258, and Appendix B, lines 1129–1158 and 1258–1275.)

W2 / Q1. Does Theorem 3.2 apply to image color transfer?

A2. You are right. The empirical RGB source is discrete, and its support stays finite under the logit transform and deterministic flow. Theorem 3.2 therefore gives no optimality guarantee here (lines 244–252).

Figure 4 illustrates the choice of terminal loss in practice. We use squared MMD to match the transformed source colors to samples from the target image (Section 4.3, lines 475–477, and Appendix C.6, lines 1435–1444). Energy distance also works with target samples. If sampling is difficult, PMOT can instead train by likelihood when the target log density can be evaluated at the flow endpoints (Section 2.3, lines 162–174).

W3(a) / Q2. How accurate is PMOT when the optimal map is known?

A3(a). We added two benchmarks with exact maps and trained on unpaired source and target samples. On the uniform square, we construct the target using an explicit nonlinear optimal map TpT_p. At p=1.5p=1.5 and p=3p=3, the mean pointwise map MSE was 2.82×10−32.82\times10^{-3} and 7.21×10−47.21\times10^{-4}, respectively. We also compared the learned trajectories with the exact paths (1−t)x+tTp(x)(1-t)x+tT_p(x). The corresponding time-integrated MSEs were 1.74×10−31.74\times10^{-3} and 3.19×10−43.19\times10^{-4} on 4,096 held-out points, averaged over five training seeds. (Section 4.1, Figure 1, lines 326–345, and Appendix C.7, construction and proof, lines 1514–1533.)

On the two-disk benchmark the marginals are fixed, and the exact map translates each disk. With 1,200 updates per method, the mean pointwise map MSE over three training seeds was

ppPMOTMB pp-OT-FMcc-RF (2 rounds)
1.50.003550.05680.657
30.0003460.09800.878

PMOT and pp-OT-FM both sent every test point to the correct disk. Within those disks, PMOT placed points closer on average to their exact destinations in all six paired runs. (Section 4.1, Figure 2, lines 373–389 and 409–412.)

W3(b). How does PMOT compare with OT-Flow on the original Table 1 tasks?

A3(b). PMOT used one likelihood weight of 5 across both datasets. At this shared weight, PMOT outperformed OT-Flow on all three metrics (mean map MSE against the Sinkhorn barycentric reference and mean forward and reverse W22W_2^2). Both methods used p=2p=2, 10,000 updates and three paired training seeds.

We also tested larger OT-Flow weights (50 and 300 on 8-Gaussians, 50 and 100 on Pinwheel). On 8-Gaussians, PMOT's map MSE at weight 5 was 0.01971, essentially matching OT-Flow's best value of 0.01969 at weight 300. On Pinwheel, the same PMOT setting gave lower map MSE (0.02508) and reverse W22W_2^2 (0.03421) than any tested OT-Flow setting. OT-Flow's best values for these metrics were 0.02818 at weight 5 and 0.03430 at weight 100, respectively. (Section 4.2, Table 1, lines 424–431 and 450–466.) The original PMOT cost ratios appear in Appendix C.2, Table 3 (lines 1313–1322).

W4. Derivations and notation.

A4. Section 2.2 now shows how varying the Benamou–Brenier action with respect to velocity gives the momentum relation and how variation in density gives the Hamilton–Jacobi–Bellman equation (lines 110–146). Section 3.2 derives the matching target from the velocity of the straight bridge to the current endpoint. It also explains why a separate terminal loss is needed. A constant potential leaves every point in place and gives zero matching loss even when the marginals differ (lines 210–225). We spell out BB and HJB in Section 2.2, write “absolutely continuous” in Section 3.1 (lines 192–197), and introduce RK4 as the fourth-order Runge–Kutta scheme in Section 3.4 (lines 316–318).

We hope the revised paper is easier to follow and that the new error bounds and experiments give you greater confidence in our results.

AuthorsOpenReview ↗

Follow-up on our response to Reviewer Fa17

Dear Reviewer Fa17,

We would appreciate your feedback on our point-by-point responses and revised manuscript. We have added experiments and theoretical clarifications addressing your comments, with references to the relevant sections and line numbers.

With the author–reviewer discussion closing on November 18, could you let us know whether our responses address your main concerns and which issues, if any, remain? Hearing from you before the deadline would give us time to provide further clarification or revisions.

Thank you for your time and consideration.

AuthorsOpenReview ↗

Follow-up on Our Response and Revised Manuscript

Dear Reviewer Fa17, we would appreciate your feedback on our responses and revised manuscript before the author discussion closes on November 18. Could you let us know whether your concerns have been addressed, or which specific issues remain? Thank you for your time.

AuthorsOpenReview ↗

Update on the Revised Manuscript

Dear Reviewer Fa17,

We would like to share a brief update on our revised manuscript.

Our earlier response addresses your main concerns through experiments with known optimal maps and trajectories (Section 4.1), comparisons with OT-Flow (Section 4.2), and map-error bounds at nonzero population loss under additional assumptions (Appendix B).

We have since extended the evaluation to unpaired image translation in 512-dimensional FFHQ representations (Section 4.3; Appendix C.7, Table 9 and Figures 8–9). We have also added an explicit limitations discussion covering the theoretical assumptions, computational cost, and scope of the empirical evaluation.

We hope these results and clarifications address your concerns about the method's transport accuracy and practical applicability. In light of these revisions, we would greatly appreciate your reconsideration of the rating and overall assessment of our work. If any concerns remain, we would welcome your feedback so that we can address them while the author discussion is still open.

Thank you very much for your time and consideration.

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

This paper proposes PMOT, a CNF framework that parameterizes the velocity field via a scalar potential in the generalized Benamou–Brenier form for general pp-cost optimal transport, and trains it with self-induced potential matching plus terminal distribution matching, without external OT plans or inner OT optimization. It proves zero-loss exact recovery under strong regularity and interior full-support assumptions, and reports toy, tabular, and color-transfer experiments. The theoretical contribution is substantive and the method is well motivated, but the empirical evidence and finite-loss analysis may be insufficient.

Critical strengths
  1. Principled method design. PMOT integrates a potential-velocity parameterization for general pp-power costs, self-induced potential-gradient matching between source samples and model-generated endpoints, and terminal distribution matching within a CNF. Training requires neither external OT plans nor inner OT optimization, offering a feasible approach for learning continuous transport under general pp-costs.

  2. Clear theoretical contribution. Theorem 3.2 establishes zero-loss exact recovery under realizability, Φ∈C2\Phi\in C^2, interior full support, and absolute continuity of μ0\mu_0.

  3. Initial empirical evidence. The pp-swap endpoint MSE in Figure 1 shows that each checkpoint is closer to the Sinkhorn reference computed with its matching pp. In the reported runs on three tabular datasets, PMOT achieves lower NLL than OT-Flow.

Critical weaknesses
  1. Missing discussion of c-rectified flow. Section 5 does not cite Liu (2022), which already uses a convex-cost-induced potential-velocity structure and a matching-type loss. I suggest discussing the similarities and differences.

  2. Strong assumptions and no finite-loss control. The zero-loss theorem relies on strong assumptions and does not quantify the mapping error or OT suboptimality for small but nonzero matching loss. Please clarify which experimental distributions satisfy these assumptions; empirical distributions generally do not have interior full support. Can the authors control the WpW_p error or OT optimality gap under additional assumptions?

  3. Smoothing vs. unsmoothed momentum for p≠2p\ne2. The stabilized velocity in Eq. (3) and the unsmoothed momentum in Eq. (4) no longer satisfy an exact duality relation when p≠2p\ne2, especially near g≈0g\approx0 or u≈0u\approx0. What is the effect on endpoints, potential-matching residuals, and OT optimality?

  4. Divergence implementation for general pp. For general pp, the CNF divergence involves the quadratic form g⊤Hgg^\top H g, unlike −ΔΦ-\Delta\Phi for p=2p=2. How is this implemented—explicit formula, autograd, or approximation—and how is it stabilized near g=0g=0?

Questions for authors

See Weaknesses

1 comment
AuthorsOpenReview ↗

Response to Reviewer YCrU

Thank you for reading our paper carefully and for pointing us to Liu’s work on cc-rectified flow. Revisions specific to your review appear in burgundy. Bracketed reviewer IDs mark shared revisions.

Q1. How does PMOT differ from cc-rectified flow?

A1. For c(u)=∥u∥p/pc(u)=\|u\|^p/p, both losses vanish on a given bridge when the predicted momentum equals mp(u)=∇c(u)m_p(u)=\nabla c(u) for its displacement uu. They differ in how the endpoint enters the gradient. In each cc-RF round, the coupling is fixed. For a sampled pair (x,y)(x,y), the bridge x+t(y−x)x+t(y-x) and displacement y−xy-x are held fixed when differentiating the loss. The fitted flow can generate new pairs for the next round. In PMOT, y=Fθ(x)y=F_\theta(x) is the endpoint of the very flow being updated. The gradient passes through Fθ(x)F_\theta(x) in both the bridge location and the target momentum mp(Fθ(x)−x)m_p(F_\theta(x)-x). A terminal loss on those same endpoints fits their distribution to the target (Sections 3.4 and 5, lines 273–289 and 516–521).

Figure 2 compares the methods on two disks with known optimal maps. At 1,200 updates per method (600 per cc-RF round), mean pointwise map MSE over three seeds was 0.00355 for PMOT and 0.657 for cc-RF at p=1.5p=1.5, and 0.000346 and 0.878, respectively, at p=3p=3 (Section 4.1, lines 372–387).

Q2. Which experiments satisfy full support, and what can we say at nonzero loss?

A2. Theorem 3.2 requires full support of the learned population law at each interior time (lines 244–252). The source distributions for 8-Gaussians, noisy Moons, 2-Spirals and Pinwheel have full support. The square, two-disk and Checkerboard distributions have bounded support. The empirical RGB distributions are atomic, and we have not established full support for the tabular populations. Without interior full support, zero loss can yield a suboptimal map even when the target itself has full support (Appendix A.2, lines 948–957).

Appendix B bounds the mean squared error dF2=E∥F(X)−T⋆(X)∥2d_F^2=\mathbb E\|F(X)-T^\star(X)\|^2 to the optimal map for the target μ1=N(0,I)\mu_1=\mathcal N(0,I), allowing the learned endpoint law to differ from that target. Write ℓ\ell for the population momentum-matching loss and η=KL(F#μ0∥μ1)\eta=\mathrm{KL}(F_\#\mu_0\Vert\mu_1) for the terminal loss. Under the stated time regularity of the bridge residual, convexity and integrability of the terminal potential, and the corresponding moment and growth conditions, the small-loss rates for dF2d_F^2 are

  • p=1.5p=1.5: O(ℓ2/9+η1/4)O(\ell^{2/9}+\eta^{1/4}),
  • p=2p=2: O(ℓ2/3+η1/2)O(\ell^{2/3}+\eta^{1/2}),
  • p=3p=3: O(ℓ1/3+η1/3)O(\ell^{1/3}+\eta^{1/3}).

For 1<p≤21<p\le2, the reverse triangle inequality gives ∣∥F(X)−X∥Lp−Wp(μ0,μ1)∣≤dF\big|\|F(X)-X\|_{L^p}-W_p(\mu_0,\mu_1)\big|\le d_F. These are population bounds. We have not verified the additional assumptions for the fitted networks (Section 3.3, lines 257–258, and Appendix B, lines 1129–1158 and 1258–1275).

Q3. How does smoothing affect training, the fitted endpoints, and OT optimality?

A3. Equation (3) stabilizes the velocity. Equation (4) gives the ideal momentum label (lines 266–284). In the original toy runs, computing this label also involved adding an offset to the squared displacement norm. We checked both effects at p=1.5p=1.5.

During training, reducing the momentum offset from 10−410^{-4} to 10−1210^{-12} with all other settings fixed changed mean empirical endpoint W1.5W_{1.5} on 8-Gaussians from 0.16833 to 0.16852 over three paired seeds. On Pinwheel, the smaller offset reduced endpoint W1.5W_{1.5} for all three seeds, though the improvement varied substantially across seeds.

After training, removing velocity smoothing from either fitted model shifted endpoints by less than 6.5×10−136.5\times10^{-13} RMS and changed the measured transport cost by less than 1.2×10−121.2\times10^{-12}. Using ideal momentum labels on the same bridges changed the mean squared matching residual by less than 5×10−75\times10^{-7}. These checks measure marginal fit and changes in the fitted map. The exact OT guarantee applies to the unsmoothed population objective.

Q4. How is the divergence computed for general pp, especially near g=0g=0?

A4. Let q=p/(p−1)q=p/(p-1), g=∇xΦθg=\nabla_x\Phi_\theta, H=∇x2ΦθH=\nabla_x^2\Phi_\theta and s=∥g∥2+ϵs=\|g\|^2+\epsilon. The divergence is

∇x⋅vθ=−s(q−2)/2tr⁡(H)−(q−2)s(q−4)/2g⊤Hg.\nabla_x\cdot v_\theta=-s^{(q-2)/2}\operatorname{tr}(H)-(q-2)s^{(q-4)/2}g^\top H g.

We compute the Hessian trace analytically. For the second term, automatic differentiation gives Hg=∇x(∥g∥2/2)Hg=\nabla_x(\|g\|^2/2), which we contract with gg. Since s≥ϵ>0s\ge\epsilon>0, both powers stay finite at g=0g=0. At p=2p=2, the formula reduces to −ΔΦθ-\Delta\Phi_\theta (Section 3.4, lines 304–308).

We have revised the exposition throughout the paper and hope the manuscript is now easier to follow.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.