acceptodds
Under review as a conference paper at ICLR 2027

Trajectory inference via Acceleration Matching

Abstract

Trajectory inference is a fundamental problem in many scientific domains: given a collection of unpaired snapshots of observations at discrete time points, the goal is to generate smooth trajectories that best resemble and interpolate the data. Existing numerical approaches to this problem have various computational burdens, such as explicitly computing spline constructions or simulating trajectories during training. In order to overcome these limitations, we propose a new algorithm called Acceleration Matching (AM). Our approach consists of lifting the original interpolation problem to phase space and then regressing onto an explicit conditional acceleration field that induces random trajectories with smooth positional paths that, under suitable assumptions, agree with the prescribed marginals. Importantly, our resulting training algorithm only requires positional data, avoids trajectory simulation during training, and uses direct Gaussian conditioning to sample training states and targets. We provide ample numerical evidence suggesting that AM is competitive with or superior to existing algorithms on several benchmark problems from the existing literature.

Primary area:
generative models
Keywords:
trajectory inference, simulation-free, acceleration matching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ZpR1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors consider a new approach to trajectory inference utilizing AM, drawing inspiration from kinetic Langevin dynamics, but using the flow-matching formulation. The authors provide theory for flow matching with multiple marginals in this setting, which informs the training of a neural network to fit the acceleration field. Experiments are provided on low-dimensional data, along with real-world single-cell biology datasets.

Critical strengths
  1. The approach seems novel, and seems to be potentially useful if the population dynamics follow some kind of velocity field.
  2. The authors provide some theoretical grounding for their method, which may potentially be useful for other applications.
Critical weaknesses
  1. Often, the applications of trajectory inference in single-cell biology usually assume the model follows an overdamped SDE (see Eq. (1.1) in Lavenant et al. (2024)). This may mean the method is not as practically relevant for single-cell biology applications.
  2. In Lavenant et al. (2024), the goal is to recover the underlying distribution on path-space, and indeed their estimators, and some followup works consider consistent estimators in this sense (in some suitable limit). It is not clear to me whether AM can do this.
Questions for authors
  1. Concerning weakness 1, can the authors test the method on datasets where there is some kind of acceleration field involved in the model where data is drawn from? I would expect that AM would have better accuracy in this setting, compared to some of the baselines.
  2. Is it possible to show consistency of AM as an estimator (on path-space) under some kind of acceleration field model, as does Lavenant et al. (2024)? Additionally, as a case study, can the authors also test the experiment from Figure 4, [GCG25]? It would be nice to see whether AM can recover the ``correct’’ path-space distribution in this setting.

The first question is most relevant for my assessment in the rebuttal as it concerns the applicability of the method. The second question would help strengthen the paper.

[GCG25] Gu, Chien, Greenewald. Partially observed trajectory inference using optimal transport and a dynamics prior. ICLR 2025.

2 comments

We thank the reviewer for the careful reading and constructive feedback. Below we address the comments individually.

Often, the applications of trajectory inference in single-cell biology usually assume the model follows an overdamped SDE (see Eq. (1.1) in Lavenant et al. (2024)). This may mean the method is not as practically relevant for single-cell biology applications.

We agree with the reviewer that a large body of the single-cell literature follows an overdamped SDE system. However, we believe our work develops algorithmic insights and explores the possibility that perhaps learning underdamped dynamics is more suitable (and scalable) for learning from physical systems, and that this was not possible with previous approaches.

Concerning weakness 1, can the authors test the method on datasets where there is some kind of acceleration field involved in the model where data is drawn from? I would expect that AM would have better accuracy in this setting, compared to some of the baselines.

We wholly agree with the reviewer, and we included a new experiment in the paper in Section 4.1. Specifically, we consider data generated by a nonlinear damped oscillator where the equations governing the dynamics are given by dXt=Vt dt ,dVt=(−∇U(Xt)−Vt) dt+0.2 dBt .{\rm d}X_t = V_t {\, \rm d}t\,, \quad {\rm d}V_t = (-\nabla U(X_t)-V_t)\,{\rm d}t+0.2\,{\rm d}B_t\,. We simulate these dynamics for T=10T=10 seconds and use 41 equispaced times for training data and evaluate on 20 other held-out times, and report the resulting W1W1 hold-out error; see Appendix C.3 for more details. Table 1 now includes the following column

MethodHeld-out W1
AM (indep.)0.316 ± 0.091
OT-CFM0.864 ± 0.165
MMFM0.635 ± 0.196
MMFM (prior)2.074 ± 1.261

which shows that AM has the lowest mean held-out error among the methods we consider in our paper, and Figure 3 plots sample trajectories for the methods we consider.

In Lavenant et al. (2024), the goal is to recover the underlying distribution on path-space, and indeed their estimators, and some followup works consider consistent estimators in this sense (in some suitable limit). It is not clear to me whether AM can do this.

Is it possible to show consistency of AM as an estimator (on path-space) under some kind of acceleration field model, as does Lavenant et al. (2024)?

We thank the reviewer for opening the discussion on this relevant theoretical aspect. Lavenant et al. (2024) consider the statistical problem of recovering an underlying path-space distribution from observations of its temporal marginals. Their consistency results rely on assumptions on the data-generating process, notably that the dynamics belong to a suitable class of potential-driven diffusion processes. In contrast, our objective is not to recover an assumed underlying stochastic process. Rather, we propose a model-free approach: a generative methodology for constructing stochastic dynamics that match a prescribed collection of JJ marginal distributions, without assuming that these marginals were generated by a particular class of SDEs. We therefore consider path-space statistical consistency to be an interesting but distinct and orthogonal research direction. Our approach is similar in spirit to Chen et al. (2023) and Theodoropoulos et al. (2025), which also employ underdamped reference dynamics to construct generative models satisfying multiple marginal constraints, with applications to single-cell trajectory inference. Nevertheless, regarding the possibility of establishing a consistency result for AM, we believe that such a result could be obtained under suitable well-specified model assumptions, for instance, by assuming that the observations are generated by dynamics belonging to the class of acceleration-field models considered in our framework. However, establishing such a result would require additional identifiability assumptions, as well as a careful analysis of the statistical and approximation errors associated with the learned acceleration field. We believe that these questions are beyond the scope of the present work and leave them for future work.

Thanks for the responses; I will increase my score to a 3.

Reviewer_to9LOpenReview ↗
Rating
4(+2)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Acceleration matching (AM) is a new multi-marginal trajectory inference method that employs a regression loss to learn an acceleration field -- this is different from flow matching where the vector field is learned via regression. As formulated by the authors, AM thus "lifts" FM to phase space. The paper is thorough in its derivation of the method and applies the method to low- and high-dimensional datasets with varying results (mostly good performance on the low-dimensional data). The way of validation of the method is standard in the field.

Critical strengths

Solid method exposition: The proposed method, AM, is fleshed out, and extensively explained in the main text and, especially, considering all the theoretical results, in Appendix A. The closed-form generation of velocity vectors given the positional knots vector is a nice result which makes the acceleration field network efficient to train.

Tackles an important question: I also agree with the narrative of the story in the sense that the field needs more methods that avoid point-wise interpolation. Theoretically grounded methods that yield smooth trajectories are nice contributions to the field of multi-marginal trajectory inference that will be appreciated. Furthermore, the method cleverly does inference in another space, which is interesting.

Critical weaknesses

Critical weakness 1.

While the authors have put commendable effort into the derivation of the AM method and its theoretical characterization, I am worried that the current state of the field of multi-marginal FM is not sufficiently described in the article. My main concerns are with regards to 1) the portrayal of piece-wise FM, and 2) existing FM methods that do not interpolate piece-wise.

1) Starting from Fig. 1: I am not aware of any interesting implementation of a multi-marginal FM method where this would be the resulting vector field. For instance, in OT-CFM paper by Tong et al. (2024), the piece-wise nature arises from linear interpolation between a sequence of points. The vector field is then regressed onto the piece-wise linear interpolant. From what I have experienced, the resulting vector field will not pass through points point-wise and should be smoother than what is depicted in Figure 1. Could the authors expand on how these trajectories were generated? On the same point, exact definitions of the piece-wise FM methods considered should be included in the paper. Neither OT-CFM nor OT-MFM are both formulated in the bi-marginal setting, so it is necessary to be clear about how these methods were implemented here.

2) The ALI-CFM paper (Kviman et al. (2026)) is an important missing piece in the narrative of this article: this method already avoids point-wise interpolations by learning interpolants that fit the marginal distributions instead, with the same motivation of obtaining smoother trajectories in multi-marginal problems.

These two points together with the incorrect use of the term "discontinuity" --- the vector field is not discontinuous; the interpolants may be non-smooth but they are not discontinuous either --- puts the submission farther away from an acceptance level. However, these points are actionable and if properly addressed I can consider raising my score.

Related work

Through a quick search I could find at least two works that consider OT or FM in velocity + position space

Kinetic Optimal Transport by Brigati et al. (2025) might interest you as they do OT in the velocity + position space.

OAT-FM (Yue et al. (2025)) is an (unpublished) paper deveop an FM algo that also works in velocity + position space.

Critical weakness 2.

The exposition of the method is very extensive, however I had difficulties understanding the inferred law. What are the formulations of the SDEs and ODEs one gets after running AM? In Eq. 5 I see the expression for the SDE with the target acceleration field drift, and the ODE for the positional knots, but what is the expression for the SDE with a_t^\theta?

On the same note, I couldn't understand if the inferred distribution of the generated trajectories are guaranteed to match the prescribed marginal distributions?

This weakness section is meant to point out that, although the derivations of the method are extensive, I found parts of the method description difficult to grasp, tending to abstract. Addressing the two points above will surely help.

To be clear, I think the submission is very interesting and a good contribution, however the two concerns above need addressing before I feel confident voting for acceptance.

Minor:

Please explain asterisk, in the caption of Table 2.

I have not previously seen the definition of a function as it is written in line 108: (x,v)↦F(x,v)∈R(x,v) \mapsto F(x,v)\in \mathbb{R}, it seems incorrect but I may be wrong. Compare e.g. with the definition of ata_t in 073.

In Algorithm 2 the conditional sampler has a hat in the requirements, but qϕq_\phi in the algo doesn't have the hat.

References

Kviman et al. "Multi-marginal flow matching with adversarially learnt interpolants." ICLR 2026. Brigati et al. "Kinetic Optimal Transport (OTIKIN) – Part 1: Second-Order Discrepancies Between Probability Measures", https://arxiv.org/pdf/2502.15665 Yue et al. "OAT-FM: Optimal Acceleration Transport for Improved Flow Matching", https://arxiv.org/pdf/2509.24936

Questions for authors

I am not sure I follow the argument in lines 248-251. I agree that AM samples the velocity given a sequence of "positional knots", but what do is meant by disjoint intervals? A spline is fitted to control points / a sequence of "positional knots" which then results in a smooth, continuous interpolation (at least in theory). Could you please explain how this paragraph is intended to separate AM from other methods?

The field has access to many FM methods with various interesting insights and techniques to improve performance -- how transferable are FM techniques to AM? Please excuse the absence of examples of FM techniques.

There are two questions in the Weaknesses cell, too.

5 comments

We first thank the reviewer for the careful reading of our paper and for providing us with constructive comments, and for recognizing the theoretical benefits of our estimator. Below we address your comments individually

Critical Weakness 1

Starting from Fig. 1: I am not aware of any interesting implementation of a multi-marginal FM method where this would be the resulting vector field…

Thank you for raising this point; we hope to clarify our intent with our response.

To create Figure 1(a), we simply learned a single FM vθ(t,xt)v_\theta(t, x_t) that is trained on the 4 Gaussian marginals using linear interpolation targets on consecutive intervals. In Figure 1(b), we used the same data but learned a single acceleration field aθ(t,xt,vt)a_\theta(t, x_t, v_t) using our approach.

In the case of (any) FM method, the piecewise-linear training interpolants are indeed continuous, but the learned velocity fields are not provably continuous in time and can change sharply at the knots. This is unlike our method, where the second-order nature of our system allows for continuously differentiable sample paths for an arbitrary (piecewise) acceleration field. The generation of non-smooth trajectories for FM is not uncommon in the multimarginal setting; see for instance Figure 2 of Petrović, Katarina, et al. “Curly flow matching for learning non-gradient field dynamics.”

Our implementation of OT-CFM uses linear interpolants between consecutive retained marginals, whereas MMFM uses cubic splines; this is outlined in Section 4.3 and follows standard practice from other papers. Regarding OT-MFM, we did not implement it ourselves; the values in the table were taken directly from Kapuśniak et al. (2024).

The ALI-CFM paper (Kviman et al. (2026)) is an important missing piece in the narrative of this article. [...]

Through a quick search I could find at least two works that consider OT or FM in velocity + position space [...]

Thank you for bringing these papers to our attention, as they are quite important to the overall literature. We are quite aware of the paper by Brigati et al. and in fact we thought we cited it but turns out we did not; we now mention it in Remark 3.1. In the same section, we have placed OAT-FM, as a related work that consider phase space information to improve flow matching. We also added a citation and discussion of ALI-CFM (Kviman et al., 2026) to Section 4.3, which we acknowledge as an important contribution in this field.

Critical weakness 2

The exposition of the method is very extensive. [...] What are the formulations of the SDEs and ODEs one gets after running AM?

Section 3.3 now explicitly provides the SDE associated with our method. This SDE is second-order of the form (3) and given by dXt=Vt dt,dVt=aθ(t,Xt,Vt) dt+ε dBt,\mathrm{d}X_t = V_t\,\mathrm{d}t, \qquad \mathrm{d}V_t = a_\theta(t,X_t,V_t)\,\mathrm{d}t + \sqrt{\varepsilon}\,\mathrm{d}B_t,

where aθa_{\theta} is the learned acceleration field. By construction, then, the trajectories (Xt)t≥0(X_t)_{t \geq 0} are continuously differentiable in time, almost surely, which is not the case if the trajectories were merely given by an arbitrary vector field, e.g., piecewise time-continuous.

On the same note, I couldn't understand if the inferred distribution of the generated trajectories are guaranteed to match the prescribed marginal distributions?

We clarify that the law of the generated trajectories is guaranteed to match the prescribed marginal distributions in the case of no discretization error and exact acceleration learning. In fact, Theorem 1 proves that the diffusion process obtained as a solution of the presented SDE with the Markovian acceleration has marginal laws that match the interpolant’s one. And then, by construction the marginals of the interpolant correspond to the observed distributions.

We report the answers to the remaining questions in the next comment.

We now address the remaining questions.

Minor

Please explain asterisk, in the caption of Table 2.

Thank you for catching this; the asterisk means that the provided results are taken from the original paper. We have modified the caption of the table also in the main paper. We have now explicitly mentioned that the data comes from Kapuśniak et al. (2024, Tables 4 and 6).

I have not previously seen the definition of a function as it is written in line 108

This notation is standard in mathematics; see for example Chapter 0, §2 in [1]

In Algorithm 2 the conditional sampler has a hat in the requirements, but qϕq_\phi in the algo doesn't have the hat.

Thank you for catching this; we have fixed this typo in the paper; Algorithm 2 now consistently uses q^ϕ\widehat{q}_\phi.

Questions

I am not sure I follow the argument in lines 248-251. I agree that AM samples the velocity given a sequence of "positional knots", but what do is meant by disjoint intervals?

As you mentioned, the velocities are sampled jointly, conditional on all observed positions, rather than independently on each interval. However, conditional on the resulting collection of position–velocity pairs, the interpolation values can then be sampled independently across intervals from the corresponding reference kinetic bridges.

The global dependence here comes from jointly conditioning all velocity knots on all position knots before sampling an interval. Other multi-marginal methods can also use global information, for example through spline or learned interpolants.

The field has access to many FM methods with various interesting insights and techniques to improve performance -- how transferable are FM techniques to AM?

We agree with the author that the literature has now important extensions and variants of FM. Due to the simplicity of our method (namely, an understanding of Gaussian conditioning and simulation-free training), we believe there are many viable extensions from FM that can be brought to AM, both theoretical and empirical, which we highlighted in our conclusion. For instance, we demonstrated that using OT-couplings as in Tong et al. and Pooladian et al. remains quite feasible, and we believe other extensions, such as Generator Flow Matching by Holderrieth et al., 2024, are also possible, among others. We leave it for future work.

[1] Serge Lang, Undergraduate Analysis, 2nd edition, Springer, 1997.

Dear authors, thank you for your thorough response!

I greatly appreciate the example from Curly flows -- I was admittedly incorrect here. I am however less surprised to see the kinks in the trajectories induced by the sequence of distributions in Fig. 2 in Curly flows, as the second marginal distribution separates distribution 1 and 3 much more sharply than in your Fig. 1.

So just to double-check: are the trajectories plotted in Fig. 1 actually produced via simulation (I couldn't find attached code)? If so, then I am happy to drop this discussion and raise my score!

Thank you for updating Sec. 3.3!

Reviewer_4YCQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose acceleration matching, an extension of the idea of flow matching to second-order stochastic dynamics, to address the issue that previous approaches for trajectory inference sometimes produce non smooth trajectories By modelling both position and velocity, AM produces continuously differentiable trajectories and avoids trajectory simulations during training.

The method connects observations across time using a kinetic Brownian process, and while observations contain only positions of trajectories, the authors derive a distribution of hidden velocities analytically through Gaussian conditioning, and construct an acceleration field that describes how velocity changes between observations. They employ a neural network to learn this field using analytically sampled training examples, and during inference, generate initial velocities with a conditional normalising flow, and integrate numerically the learned dynamics.

The authors test the proposed method on ocean current data, Lotka–Volterra dynamics, and single-cell datasets, and quantify the performance using Wasserstein distance between predicted and observed distributions. The proposed method is shown to performs competitively and achieves lower interpolation errors on several benchmark settings, but it does not consistently outperform existing methods. The momentum multi-marginal Schrödinger bridge performs better on both low-dimensional datasets, while standard flow-matching methods achieve better results on the single-cellq data. The experiments also show that performance depends strongly on how observations are coupled across time.

Critical strengths
  • Nice extension of flow matching to second-order stochastic dynamics that provides differentiable position trajectories

  • The training procedure is simulation free

  • Unlike methods that independently interpolate consecutive distributions, the proposed method samples velocity knots conditioned on all observed positions

Critical weaknesses
  • the authors do not provide enough evidence for the claimed computational efficiency

  • The inferred trajectories depend heavily on the diffusion coefficient and initial velocity distribution of the kinetic Brownian process. While these choices can affect the learned dynamics and accuracy, the authors do not thoroughly examine their impact, despite acknowledging the method’s sensitivity to hyperparameters.

  • Several baseline results are taken directly from previously published papers, while others are obtained using the authors’ own implementations

Questions for authors
  • How sensitive is the method to the choice of diffusion coefficient and initial velocity distribution, and how do these affect the learned trajectories?

  • Since the Markovian projection preserves marginal distributions but does not preserve temporal correlations, how can we be sure that AM captures the true underlying dynamics rather than simply connecting distributions?

  • How well do the theoretical guarantees hold when the Lipschitz assumption is not fully satisfied or when neural-network approximation and numerical errors are introduced?

  • Considering the extra steps involved in Gaussian conditioning and initial-velocity sampling, how does AM compare with existing methods in terms of training time, memory usage, and scalability?

4 comments

We thank the reviewer for the careful reading and constructive feedback. Below we address the comments individually.

the authors do not provide enough evidence for the claimed computational efficiency

Considering the extra steps involved in Gaussian conditioning and initial-velocity sampling, how does AM compare with existing methods in terms of training time, memory usage, and scalability?

Thank you for highlighting this important point. Over the course of running the experiments, we learned that our initial velocity sampler can be substantially simplified; see Section 3.3 and Appendix B.1. In summary, the initial velocity can be written as V0=U+ζV_0=U+\zeta, where U=∑j=1JαjXtjU=\sum_{j=1}^J\alpha_jX_{t_j} and ζ∼N(α0x0,σpost2I)\zeta\sim\mathcal N(\alpha_0x_0,\sigma_{\rm post}^2I) is independent of UU and αj\alpha_j are deterministic coefficients that arise from Gaussian conditioning. Therefore, we only need to learn an unconditional sampler for UU (which is significantly easier than learning a conditional sampler), and then add the Gaussian shift when sampling V0V_0 given X0=x0X_0=x_0. For pairwise OT coupling, we still use the conditional NF as before.

Table 15 in Appendix D.2 (displayed below) reports measured training times on one NVIDIA GeForce GTX 1080, where we compare AM to OT-CFM and MMFM. Since AM with independent coupling also avoids storing and computing dense pairwise OT plans, which require O(Jn2)O(Jn^2) entries for nn samples per marginal, and the initial velocity sampler is much easier than before, we significantly outperform the other methods.

DatasetMethodOT couplingInitializer fittingMain networkTotal
GoMAM (indep.)—0.488 ± 0.1238.174 ± 0.0668.662
GoMOT-CFM0.012—4.073 ± 0.0144.085
GoMMMFM0.012—34.037 ± 0.07234.049
LVAM (indep.)—0.827 ± 0.3262.630 ± 0.0533.457
LVOT-CFM0.005—1.438 ± 0.0161.443
LVMMFM0.004—8.374 ± 0.0548.377
CITE5AM (indep.)—6.000 ± 0.42533.923 ± 0.53339.924
CITE5OT-CFM67.873—118.339 ± 0.722186.212
CITE5MMFM67.873—759.351 ± 4.594827.224
CITE50AM (indep.)—52.749 ± 3.28146.583 ± 0.21499.333
CITE50OT-CFM83.296—142.667 ± 14.347225.963
CITE50MMFM83.296—882.507 ± 43.834965.803
EB5AM (indep.)—4.896 ± 0.51637.790 ± 0.82342.686
EB5OT-CFM15.035—225.618 ± 0.658240.652
EB5MMFM15.035—1155.470 ± 2.2891170.505
EB100AM (indep.)—53.186 ± 2.40766.736 ± 1.675119.923
EB100OT-CFM9.529—128.503 ± 1.308138.032
EB100MMFM9.529—822.615 ± 2.358832.144

The inferred trajectories depend heavily on the diffusion coefficient and initial velocity distribution of the kinetic Brownian process. While these choices can affect the learned dynamics and accuracy, the authors do not thoroughly examine their impact, despite acknowledging the method’s sensitivity to hyperparameters.

How sensitive is the method to the choice of diffusion coefficient and initial velocity distribution, and how do these affect the learned trajectories?

We agree that our choices require a closer examination. Following your suggestion, we performed a hyperparameter sweep across several orders of magnitude for the noise amplitude ε\sqrt{\varepsilon} and the prior velocity variance σv2\sigma_v^2 for two of our datasets (GoM and LV). Figure 11 reports the held-out W2W_2 over the sweep; we see that GoM is more sensitive to noise amplitude, whereas LV shows a stronger interaction between the parameters and. We additionally plot GoM trajectories (Figure 12) under these choices which ultimately do not vary too significantly.

Several baseline results are taken directly from previously published papers, while others are obtained using the authors’ own implementations

Thank you for pointing this out. We now distinguish our OT-CFM and MMFM implementations from quoted published results more clearly. We also clarified in the table captions where the reported values come from; for example the caption of Table 2 identifies Kapuśniak et al. (2024, Tables 4 and 6) as their source. We report the published 3MSBM values because we were unfortunately not able to reproduce them using their released code.

We report the answers to the remaining questions in the next comment.

We now address the remaining questions.

Since the Markovian projection preserves marginal distributions but does not preserve temporal correlations, how can we be sure that AM captures the true underlying dynamics rather than simply connecting distributions?

The reviewer is correct—there is no guarantee that the dynamics produced by AM are the correct ones for a given dataset. However, up to our knowledge, in the absence of further structural assumptions, the same is true for any method due to basic identifiability issues in the data. Our goal here was to provide an algorithm that is scalable in a certain sense, and does not hinge on OT-type solvers as algorithmic primitives that can hinder the runtime on large-scale problems.

How well do the theoretical guarantees hold when the Lipschitz assumption is not fully satisfied or when neural-network approximation and numerical errors are introduced?

We agree that these are both interesting questions. We would like first to emphasize that we view our core contribution as methodological (already with quite strong theoretical guarantees compared to prior work) and purposefully did not focus on adding more theoretical contributions, as they would overcomplicate the paper given the stringent page limit. Nevertheless, we believe that theoretical analysis is of great importance and interest, and would be an exciting direction to pursue for future work.

We believe the Lipschitz assumption is a technical requirement that could be relaxed by imposing suitable conditions directly on the Fokker–Planck equations associated with the Markovian projection (we note that this is often done implicitly for methodology papers such as Flow Matching). For example, we conjecture that the Lipschitz assumption can instead be replaced by a moment condition on the marginal distributions (in particular, when all marginals are Gaussian, all issues are expected to disappear). Establishing such results would, however, require significant additional theoretical developments, which we believe could constitute a conference paper in their own right. Regarding approximation guarantees: we believe that it is possible to establish consistency of our method to recover the prescribed JJ marginal distributions as the number of samples nn tends to infinity, at least in a suitably well-specified setting. As before, such a result would be of significant interest, but would require an analysis that is beyond the scope of our present submission, and we therefore believe leaving this question is an exciting direction for future work.

I would like to thank the authors for their detailed responses and the additional experiments and clarification. The updated manuscript and responses have addressed some of my initial concerns, although some of them remain partially covered.

I detail my concerns on a point mby point manner below:

  • W1/Q4. (Computational efficiency, memory, and scalability) partially ok. Table 15 provides useful runtime comparisons, including the cost of fitting the initialiser. However, the results cover only independent-coupling AM, while OT-CFM remains faster on the GoM and LV. It would be helpful to discuss runtime together with accuracy and report timings for OT-coupled AM.

  • W2/Q1. (Sensitivity to the reference process and initial velocity) partially ok. The new parameter experiments help to demonstrate the sensitivity of the framework to noise amplitude and prior variance. The sensitivity to errors in the learned initial-velocity sampler is not yet specified, as this is not covered by the reference-parameter tests.

  • W3. (Baseline comparisons) ok. The revised captions now distinguish published results from the authors’ own implementations, which resolves the attribution issue. However, I would prefer if the authors could more clearly distinguish the results obtained from published literature literature from their controlled comparisons more clearly, i.e use different tables

  • Q2. (Recovering dynamics versus interpolating distributions) ok, please update the manuscript accordingly!

  • Q3. (Theoretical assumptions and approximation errors) unresolved, but fine.

I think the additional experiments address several of my concerns. However, as mentioned above, the distinction between marginal interpolation and dynamical recovery, as well as the gap between the theoretical guarantees and their practical implementation, still needs to be made clearer in the updated manuscript.

I will think within the next days whether I will update the score or not.

Thank you for pushing us to have the best possible version of our paper; we hope the following modifications will be satisfactory as we believe they address all points you mentioned.

W1/Q4. (Computational efficiency, memory, and scalability) partially ok. Table 15 provides useful runtime comparisons, including the cost of fitting the initialiser. However, the results cover only independent-coupling AM, while OT-CFM remains faster on the GoM and LV. It would be helpful to discuss runtime together with accuracy and report timings for OT-coupled AM.

We have since updated Table 18 (used to be Table 15) to include the training times for OT-AM. The related section is updated referring also to the accuracy of the methods and underlying that AM can achieve competitive results while drastically reducing the training time that can be. (Note that the table uses 3k iterations for the initial velocity sampler as opposed to the original 20k iterations — we observed that AM works just as well in this case; we include both values in the appendix for completeness as yet a further ablation, see Appendix B.1.4).

W2/Q1. (Sensitivity to the reference process and initial velocity) partially ok. The new parameter experiments help to demonstrate the sensitivity of the framework to noise amplitude and prior variance. The sensitivity to errors in the learned initial-velocity sampler is not yet specified, as this is not covered by the reference-parameter tests.

Thank you for emphasizing this point. Appendix B.1 now contains two additional studies for the initial velocity sampler.

We first compare the trained sampler against direct sampling (i.e. sampling velocities through Gaussian conditioning using the positional samples at all the training marginals; see Appendix B.1.2). We observed that we are able to perform comparatively to using the direct sampler (sometimes slightly better, sometimes slightly worse).

We also assess the impact of the “quality” of the learned sampler on hold-out error. For EB5 and CITE5, we show how the number of training samples for GMM (and consequently runtime) impacts hold-out error; see Table 4 (old Table 18). In particular, we see that dramatically increasing the number of samples from n=500n=500 to greater than 20,000 does not lead to further improvements.

Finally, in appendix B.1.4, we show that changing the number of iterations for training the initial velocity sampler for OT-AM barely changes the outcome, and so our default number of iterations for this procedure is now significantly less than before, resulting in improved runtimes for this method.

W3. (Baseline comparisons) ok. The revised captions now distinguish published results from the authors’ own implementations, which resolves the attribution issue. However, I would prefer if the authors could more clearly distinguish the results obtained from published literature literature from their controlled comparisons more clearly, i.e use different tables

We agree that this is an important point. In order to make this more explicit without ruining the flow of the paper and still obeying space constraints, we have now mentioned that results from published works are highlighted in addition to the asterisk, and are also mentioned in table captions, as is common practice.

Q2. (Recovering dynamics versus interpolating distributions) ok, please update the manuscript accordingly!

We are now clarifying this after Theorem 1. Specifically, we now write (in purple):

"In particular, we do not claim that \texttt{AM} learns the true underlying distribution of the data, but that the ideal acceleration field will match the marginals that we have access to at training time; we leave open the possibility to explore the theoretical assumptions, consistency of our method as the number of samples grows, and incorporation of approximation error."

Comments on the paper

1 comment

We begin by thanking the reviewers for taking the time to carefully read our manuscript and provide constructive feedback, and the AC for handling the paper. We believe the comments have greatly helped us to clarify certain points and ultimately strengthen the paper in several directions.

We further emphasize that Reviewer 4YCQ describes AM as a “nice extension of flow matching to second-order stochastic dynamics”, Reviewer to9L highlights our “solid method exposition”, and that Reviewer ZpR1 observed that our approach “ seems novel” and that we gave “theoretical grounding for [our] method”.

In this general response, we address comments raised by some of the reviewers that are worth highlighting for everyone. We also included a point-by-point response to the individual reviewers below. These papers are addressed in the new rebuttal version, where we made the edits in purple.

A new initial velocity sampler

Our newly written Section 3.3 and Appendix B.1 now describe a simpler initial velocity sampler when using the independent coupling. In particular, we realized that the initial velocity sampler in this case can be written into a Gaussian component and an unconditional sampler, which is much easier to learn.

Runtime comparisons

Per Reviewer 4YCQ, we added Table 15 in Appendix D.2 that reports training times comparing our method to the others. In all cases, we verified that AM (with independent coupling) has the lowest total time on the four larger datasets even when we include the additional step of learning an initial velocity sampler.

Sensitivity to hyperparameters

Following another comment of 4YCQ, Appendix D.3 now includes a hyperparameter sweep across the reference process parameters. We compute held-out W2W_2 values for the GoM and LV datasets across the sweep and also plot the trajectories for GoM. We notice that the GoM dataset is more sensitive to noise amplitude whereas for LV, a wider range of parameters perform well. In particular, the “optimal” parameters are certainly dataset dependent, which is expected.

Additional experiment with known second-order dynamics

Per Reviewer ZpR1, we added a new dataset, trajectories given by a nonlinear damped oscillator. This dataset falls in the “well-specified” regime and we observe that using only unpaired positions, independent AM achieves a lower held-out W1W_1 error of 0.316 ± 0.091 across five seeds, compared with 0.864 ± 0.165 for OT-CFM, 0.635 ± 0.196 for MMFM. The newly added Figure 4 shows generated trajectories, and more details appear in the appendix.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.