acceptodds
Under review as a conference paper at ICLR 2027

Discrete Action Matching: Learning Stochastic Dynamics from Samples via State Graphs

15 views

Abstract

Learning population dynamics from unpaired temporal marginals is an ill-posed inverse problem that requires structural assumptions on the underlying dynamics. We introduce _Discrete Action Matching_ (DAM), a finite-state counterpart of Action Matching based on discrete Wasserstein geometry on graphs. For a prescribed marginal path and transport geometry, we derive an action-minimization objective for its canonical minimum-kinetic-energy current. Our key observation is that the density dependence of the discrete action reduces to neighboring density ratios. Along an empirical interpolation of the snapshots, DAM first estimates these ratios and then learns an action potential. The learned fields also define a graph-supported Markov sampler. Experiments on controlled synthetic dynamics and real mouse gastrulation data evaluate marginal reconstruction and interpolation. Additional experiments approximate numerical surface-transport paths from paired samples.

Primary area:
generative models
Keywords:
discrete action matching, learning discrete population dynamics, discrete dynamic optimal transport

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_knx3OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors consider the problem of learning the population dynamics on a graph when only unpaired temporal marginals are provided. Their approach prescribes an undirected, symmetrically weighted state graph with discrete Wasserstein geometry and introduces the continuity equation on that graph. This equation requires a notion akin to a probability mass function (PMF) or density defined on edges rather than vertices, which the authors resolve by introducing mobility θ(pi,pj)\theta(p_i,p_j), e.g. the mean of the probability mass of the endpoints. Although the continuity equation allows multiple solutions, one can ask, analogous to Action Matching (Neklyudov et al., 2023), for the solution with the minimal kinetic energy, which is unique. The main contribution of the paper is the method that finds this solution, which the authors refer to as Discrete Action Matching (DAM).

The crucial difference with Action Matching in the continuous space is the need to estimate the mobility. This can be done by using the 1-homogeneity of the mobility, which allows theta to be re-written in terms of the ratio of PMFs. The quantities required for the DAM approach can be obtained by sampling these local density ratios. Building on this idea, the paper splits training into two stages: first it fits a scalar field to represent the unnormalized log-density, and then it freezes it and proceeds with fitting the action potential according to the Action Matching method. The fields learned define a graph-supported Markov sampler. Experiments compare against results from free-energy gradient flow (FEGF) (Rancati et al., 2026) on eleven synthetic graph families and mouse gastrulation leave-one-stage-out interpolation, as well as two surface transport mesh problems.

Critical strengths
  • Population dynamics is an active area of research in single cell biology and considering the problem on a graph is a valuable contribution to the field
  • All underlying theory and methods are carefully reasoned and defended. This is especially notable when certain necessary conditions seem not to be met. The paper is honest about limitations and what it does not establish.
Critical weaknesses
  • The paper presents two variants of DAM, one approach using histograms and another using learned ratios. The latter is achieved by training unnormalized PMFs, and this is the variant that the paper seems to present as the main contribution. On the other hand, the histogram approach is emphasized less and not mentioned in the abstract, despite outperforming the learned-ratios variant for the most part. The paper could be greatly strengthened by providing more motivation and support for the learned-ratios approach, including an experiment that demonstrates a way in which the learned-ratios approach is superior.
  • In Figure 2, the new method is stated to outperform the baseline of free-energy gradient flow (FEGF) (Rancati et al., 2026). Such figures throughout the paper would be improved if a noise floor was provided, e.g. a Hellinger distance of 0.09. In particular, in Figure 2 DAM appears to be especially superior to FEGF for two graph families, Euclid-MST and Delaunay, and outperform FEGF for the other graph families only slightly. Some discussion of Euclid-MST and Delaunay would help delineate the cases in which this method best demonstrates its usefulness.
  • Currently, the contribution of the paper is mostly theoretical. However, more practical experiments on large-scale datasets and in high dimensions would significantly strengthen the submission.
Questions for authors
  • What are the reasons for the drastic performance difference on Euclid-MST and Delaunay? Is it related to their sparse connectedness, and would that also explain the difference in performance seen with the sphere and hand transport problems?
  • In Appendix A.3 "Relation to the Maas–Rancati parameterization," it is noted that the conductances of DAM are not necessarily equivalent to those of Mass-Rancati as in FEGF. In particular, it is not equivalent for the logarithmic mean and nonuniform pi. Does this affect the comparison between DAM and FEGF, as in Figure 2, where logarithmic mean is displayed? If so, how is it affected?
2 comments
AuthorsOpenReview ↗

Official Comment by Authors. Part 1

Dear Reviewer knx3,

Thank you for recognizing the value of our contribution to learning population dynamics and for your positive assessment of the theory and clarity of the manuscript. We also appreciate your comments on Figure 1. Below, we address each of your concerns in turn.


1. "The paper presents two variants of DAM, one approach using histograms and another using learned ratios. [...] the histogram approach is emphasized less and not mentioned in the abstract, despite outperforming the learned-ratios variant for the most part. [...]"

Thank you for this constructive suggestion. We revised the abstract to describe both ways of estimating the density ratios and now introduce Histogram DAM (h-DAM) and Log-ratio DAM (lr-DAM) explicitly in the contribution statement. We also use these names consistently in the experimental figures and tables.

Both variants learn the action potential from samples. Their difference lies in how they obtain the neighboring density ratios required by the discrete action. h-DAM computes these ratios from interpolated empirical histograms. It is a natural choice when the state space can be enumerated and the snapshot counts are sufficiently informative. lr-DAM learns the ratios from sampled states and their neighbors, then holds the ratio model fixed while learning the action. This gives a sample-based counterpart to continuous Action Matching [2] without requiring a full marginal histogram.

We agree that h-DAM is an important result in its own right. For example, on mouse gastrulation with harmonic mobility, h-DAM achieves lower mean held-out Hellinger error than lr-DAM (0.203 versus 0.220, Table 3). The revised presentation gives both estimators equal prominence and identifies the settings in which each performs better.


2. "[...] In particular, in Figure 2 DAM appears to be especially superior to FEGF for two graph families, Euclid-MST and Delaunay, and outperform FEGF for the other graph families only slightly. Some discussion of Euclid-MST and Delaunay would help delineate the cases in which this method best demonstrates its usefulness. [...] What are the reasons for the drastic performance difference on Euclid-MST and Delaunay? [...]"

Euclid-MST and Delaunay had the smallest spectral gaps among the eleven graph families in Figure 2. This matters because FEGF [1] estimates a velocity target vv from the difference between consecutive sampled marginals. It obtains vv by first solving for a potential through a weighted graph Laplacian. Along a mode with eigenvalue λk\lambda_k, this solve amplifies sampling error in the corresponding potential coefficient by 1/λk1/\lambda_k. DAM learns the potential directly from samples and avoids this intermediate velocity estimate.

We tested this explanation by weakening the connection between two graph blocks while keeping the sampled marginals fixed. As the graph spectral gap decreased from 0.2 to 0.002, FEGF’s mean TV error increased by 6.15 percentage points when mass moved across the weak connection and by 5.04 points when it moved within the blocks. This supports poor conditioning as one reason why the FEGF–DAM gap is especially large on Euclid-MST and Delaunay.

AuthorsOpenReview ↗

Official Comment by Authors. Part 2


3. "[...] The paper could be greatly strengthened by providing more motivation and support for the learned-ratios approach, including an experiment that demonstrates a way in which the learned-ratios approach is superior. [...] Currently, the contribution of the paper is mostly theoretical. However, more practical experiments on large-scale datasets and in high dimensions would significantly strengthen the submission."

The clearest advantage of learned ratios in our experiments occurs when each snapshot contains relatively few observations. In Appendix D.1, h-DAM and lr-DAM use the same prescribed interpolation, action architecture, and selection criterion. The table gives mean Hellinger distance to held-out simulator marginals for cubic smoothstep interpolation (Eq. 17), using the results in Figure 6.

Samples per snapshotDirect interpolationh-DAM (ours)lr-DAM (ours)
1280.19970.20420.1660
1,0000.06880.06360.0641

At 128 samples per snapshot, learning the ratios reduces mean error from 0.2042 to 0.1660 relative to the histogram estimator. At 1,000 samples, the two estimators are nearly tied. This comparison identifies a setting in which the additional learning step improves prediction and one in which the simpler estimator works equally well.

We also added a downstream generation task on the discretized Swiss roll in Appendix D.3. With 512 training pairs and 32 sampling steps, lr-DAM achieves mean TV 0.2885, compared with 0.2952 for h-DAM and 0.3017 for SEDD [3] (Table 5). The advantage depends on the metric: h-DAM places less mass outside the observed clean reference support than lr-DAM (5.50% versus 6.21%).

The computational motivation for lr-DAM is that its training objective can be evaluated using sampled states and local ratios without constructing a joint histogram. The Swiss-roll state space still permits histogram evaluation, and the revised experiments do not establish an empirical advantage in high dimensions. They demonstrate a benefit under sparse sampling and comparable or modestly better generation accuracy on the tested state space.


4. "In Appendix A.3 "Relation to the Maas–Rancati parameterization," it is noted that the conductances of DAM are not necessarily equivalent to those of Mass-Rancati as in FEGF. In particular, it is not equivalent for the logarithmic mean and nonuniform pi. Does this affect the comparison between DAM and FEGF, as in Figure 2, where logarithmic mean is displayed? If so, how is it affected?"

Yes. In Figure 2, both methods use the logarithmic mean, but they apply it to different quantities. Let pip_i be the probability at vertex ii, let πi\pi_i be its stationary probability under the reversible kernel KK, and write ωij=πiKij\omega_{ij}=\pi_iK_{ij}. With θlog⁡\theta_{\log} denoting the logarithmic mean, the conductances used by the two methods are

=\omega_{ij}\theta_{\log}(p_i,p_j), \qquad c_{ij}^{\mathrm{FEGF}}(p) =\omega_{ij}\theta_{\log}\left(\frac{p_i}{\pi_i},\frac{p_j}{\pi_j}\right).$$ When $\pi$ is uniform, homogeneity of the logarithmic mean makes the two conductances proportional by a single constant. When $\pi$ is nonuniform, the correction generally depends on the edge and the current marginal $p$. As a result, an identical potential difference across an edge can produce a different probability current under the two geometries. Conversely, fitting the same marginal evolution can require different potential differences. Figure 2 therefore compares marginal reconstruction under each method’s own geometry. Its errors quantify how well the complete methods reconstruct the same held-out distributions. They do not determine how much of the gap is due to the conductance choice or whether that choice favors either method on a particular graph. We now state this distinction explicitly in Appendix D.1 and give the conductance formulas in Appendix A.3. --- **Conclusion** Thank you again for prompting us to give h-DAM greater prominence, identify where learning the ratios improves prediction, and clarify how the choice of conductances affects the baseline comparison. We hope the revised manuscript and additional experiments address your concerns, and we would appreciate your reconsideration of the paper in light of these results. **References** [1] Rancati, Dario, et al. “Learning Discrete Diffusion on Graphs via Free-Energy Gradient Flows.” *Proceedings of the 43rd International Conference on Machine Learning*, PMLR, 2026. [2] Neklyudov, Kirill, et al. “Action Matching: Learning Stochastic Dynamics from Samples.” *Proceedings of the 40th International Conference on Machine Learning*, PMLR, 2023. [3] Lou, Aaron, et al. “Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution.” *Proceedings of the 41st International Conference on Machine Learning*, PMLR, 2024.
Reviewer_DSkGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper focuses on learning stochastic population dynamics from observations collected at different time points without tracking individual transitions. This problem is common in single-cell biology, where data show how cell populations change over time but not how individual cells move between states. And since different processes can produce the same observed distributions, recovering the underlying dynamics requires additional assumptions.

The authors propose here Discrete Action Matching (DAM), an extension of Action Matching to discrete state spaces that they represent as graphs. The method uses discrete Wasserstein geometry to describe how probability mass moves between the connected states. The method aims to find the probability currents that minimise the kinetic energy, while simultaneously matching the observed changes in population distributions. The learning objective recovers these currents without requiring individual transition data or explicit estimates of how the distributions change over time.

The method has 2 learning stages. First, a neural network estimates density ratios between neighboring states. These estimates are then used to learn a time-dependent action potential, and together they define a continuous-time Markov chain, that reproduces the observed distribution changes.

The authors test the method on synthetic graph dynamics, mouse gastrulation data, and transport problems on surface meshes. They demonstrate that the method performs better than a graph-based free-energy gradient-flow approach, but simple histogram-based estimates sometimes match or outperform the learned density ratios.

Critical strengths
  • The first figure is very impressive

  • The authors use local density ratios instead of full probability distributions which makes the optimisation more practical.

  • The authors show how the learned potential and density ratios can define a valid Markov process that reproduces the prescribed marginal distributions

  • They evaluate the method across different graph structures, mobility functions, and datasets

Critical weaknesses
  • The method relies on a predefined interpolation between observations thereby, the learned dynamics may reflect the chosen interpolation rather than the true population behavior. This sensitivity is not thoroughly explored.
  • The experiments mainly measure how well DAM reconstructs marginal distributions, without showing whether it captures the actual transition rates or probability currents.
  • Simple histogram-based estimates perform similarly or sometimes better, raising questions about whether the additional learning step is necessary.
  • The authors do not examine how density-ratio estimation errors affect the learned dynamics.
  • The main baseline uses different modelling assumptions, making it difficult to determine whether DAM’s improvements comes from
Questions for authors
  • How much do the learned dynamics change when different interpolation methods are used between observations?
  • Can DAM recover the actual transition rates or probability currents on synthetic datasets where the ground truth is known?
  • Since histogram-based estimation performs similarly or better, can the authors specify when does learning density ratios offer a clear advantage?
  • How do errors in density-ratio estimation affect the learned potential and Markov dynamics, especially for rare states or sparsely connected graphs?
3 comments
AuthorsOpenReview ↗

Official Comment by Authors. Part 1

Dear Reviewer DSkG,

Thank you for your careful review and for highlighting the practical value of training directly from samples. We appreciate your positive assessment of our evaluation across graph structures, mobility functions, and datasets, as well as your comments on Figure 1. Below, we address each of your concerns in turn.


1. "The method relies on a predefined interpolation between observations thereby, the learned dynamics may reflect the chosen interpolation rather than the true population behavior. This sensitivity is not thoroughly explored. [...] How much do the learned dynamics change when different interpolation methods are used between observations?"

We agree that the choice of interpolation affects the learned dynamics. Given a graph and mobility, DAM selects the minimum-energy current for the prescribed marginal path. Different paths through the same snapshots can therefore give different currents.

Appendix D.1, Figure 6 compares linear, cubic smoothstep (Eq. 17), log-linear, and Fisher–Rao interpolation. For each path, we evaluate both the direct interpolant and the fitted DAM rollouts against simulator marginals at 25%, 50%, and 75% of each observation gap. With 128 samples per snapshot, lr-DAM improves on its direct interpolant in 11 of the 12 path–time comparisons. With 1,000 samples, both DAM variants improve on direct smoothstep interpolation on average, while direct evaluation gives lower mean error for the other three paths.

The physical experiment in Appendix D.4 shows why interpolation matters even when two paths agree at the evaluation time. Linear and cubic smoothstep interpolation give exactly the same midpoint marginal, but the midpoint slope of smoothstep is 1.5 times that of linear interpolation. On the chain, this gives a midpoint current 1.5 times as large on every edge. At the larger sample budget, their relative current errors are 0.222 for linear and 0.590 for smoothstep (Table 7). Thus, agreement on the marginal at a held-out time can conceal a substantial difference in the inferred motion.


2. "The experiments mainly measure how well DAM reconstructs marginal distributions, without showing whether it captures the actual transition rates or probability currents. [...] Can DAM recover the actual transition rates or probability currents on synthetic datasets where the ground truth is known?"

We added an experiment with known physical currents in Appendix D.4. We simulate the discrete Schrödinger dynamics of Chow, Li, and Zhou [2] on a chain and train lr-DAM using only independently sampled state occupancies. The simulator’s phase and signed edge currents are withheld from training. Starting from the first smoothed histogram, we propagate the learned model without resetting it and compare its current with the simulator current across all 31 edges at all 16 midpoints between observations.

Current estimator4,096 samples per snapshot16,384 samples per snapshot
lr-DAM (ours)0.2210.109
Validation-selected smoothing and inversion0.2970.239
Direct linear interpolation0.4800.222
Direct cubic smoothstep interpolation0.8650.590

Entries are relative errors in the signed current (Table 7). lr-DAM achieved the lowest error at both sample budgets. Figure 11 also shows that the learned rollout tracks the reversal of flow through the central edge.

Regarding individual transition rates, the signed current on an edge is Jij=piqij−pjqjiJ_{ij}=p_iq_{ij}-p_jq_{ji}. Increasing the probability flow in both directions by the same amount preserves this current. Consequently, even exact recovery of the current would leave the two directed rates underdetermined. DAM constructs one compatible Markov process. The experiment demonstrates recovery of net probability flow from finite snapshots on this chain.

AuthorsOpenReview ↗

Official Comment by Authors. Part 2


3. "Simple histogram-based estimates perform similarly or sometimes better, raising questions about whether the additional learning step is necessary. [...] Since histogram-based estimation performs similarly or better, can the authors specify when does learning density ratios offer a clear advantage?"

We agree that learning the ratios is optional when empirical histograms are readily available and sufficiently accurate. Both h-DAM and lr-DAM learn the action potential. They differ in how they obtain the density ratios used to train it. h-DAM reads the ratios from interpolated histograms. lr-DAM learns them from sampled states and their neighbors, so it can evaluate a local ratio without constructing a full joint histogram.

The synthetic interpolation ablation gives a setting where this distinction affects accuracy. With 128 samples per snapshot and the smoothstep path, mean held-out Hellinger error is 0.1660 for lr-DAM, 0.2042 for h-DAM, and 0.1997 for direct interpolation. With 1,000 samples, the two DAM variants are nearly tied, at 0.0641 for lr-DAM and 0.0636 for h-DAM (Appendix D.1, Figure 6). The clearest empirical advantage of learning the ratios in this comparison occurs when the snapshots are sparsely sampled.

The advantage also depends on the evaluation metric. On Swiss-roll generation with 512 training pairs, lr-DAM gives slightly lower mean TV than h-DAM (0.2885 versus 0.2952), while h-DAM places less mass outside the observed clean reference support (5.50% versus 6.21%, Table 5). These results support using h-DAM when histograms are practical. The additional reason to use lr-DAM is that its ratio estimator can be trained locally from samples.


4. "The authors do not examine how density-ratio estimation errors affect the learned dynamics. [...] How do errors in density-ratio estimation affect the learned potential and Markov dynamics, especially for rare states or sparsely connected graphs?"

Appendix B.3 describes two ways in which a ratio error enters the dynamics. During action training, the fitted ratios determine the effective weight of each edge in the kinetic energy. An error changes that weight, so fitting the potential with the ratios held fixed can change the learned potential differences. During rollout, the fitted ratios also enter the forward and reverse jump rates directly. With exact ratios, the two orientations use the same mobility-weighted conductance. With approximate ratios, these weights can differ. Under the smooth rate split, this imbalance can produce a nonzero net current even on an edge where the potential difference is zero.

The Swiss-roll ablation in Appendix D.3 separates these effects empirically by varying the ratio estimator and the way the action is obtained. With a learned action, replacing empirical ratios by learned ratios changes mean TV from 0.5172 to 0.5093. With a pointwise action solve, the corresponding change is from 0.5147 to 0.5095 (Table 6). These comparisons measure sensitivity to the choice of ratio estimator. They do not directly measure error relative to the true ratio.

For rare states, an empirical endpoint count can vanish, making its unsmoothed log-ratio undefined. Uniform reference mixing gives every state positive mass and makes the ratio finite, at the cost of changing the target near low-mass states. The graph-family experiments include sparse topologies, but their aggregate marginal errors do not isolate the effect of a ratio error at a rare state or on a particular edge. The manuscript therefore explains how such errors propagate, while their magnitude in those specific cases remains open.

AuthorsOpenReview ↗

Official Comment by Authors. Part 3

5. "The main baseline uses different modelling assumptions, making it difficult to determine whether DAM’s improvements comes from. […]"

We agree that FEGF and DAM make different assumptions about how particles move between observed marginals. However, both methods make the same assumption about the available data: at each observation time, they receive unpaired samples from the population. In Figure 2, both methods use the same graph and sampled snapshots, start from the same empirical initial distribution, and propagate to the remaining observation times without reinitialization. We evaluate their predictions against the same independent test histograms. We therefore consider this a fair comparison of marginal reconstruction from unpaired snapshots.

The modeling difference concerns how each method explains the evolution between snapshots. On the synthetic benchmark, FEGF [1] assumes that an autonomous free energy drives the population dynamics. DAM prescribes an interpolation between the observed marginals and learns the minimum-energy current compatible with that path under the chosen graph geometry. Since individual particle trajectories are unobserved, both methods need an assumption about the motion. Figure 2 evaluates how accurately the resulting models reconstruct the observed population changes. We agree that the marginal errors alone cannot establish which assumption better describes the underlying particle motion.

We clarified this distinction in Appendix D.1. To evaluate motion directly, we added a separate experiment in Appendix D.4, where the simulator provides ground-truth edge currents for evaluation (Table 7 and Figure 11).


Conclusion

Thank you again for prompting us to examine the role of interpolation, compare learned and empirical ratios more directly, and evaluate a physical quantity beyond the marginals. We hope the revised experiments clarify both what DAM can recover from unpaired snapshots and where its conclusions depend on the chosen path and ratio estimator. We would appreciate your reconsideration of the paper in light of these results.

References

[1] Dario Rancati, Jan Maas, and Francesco Locatello. Learning Discrete Diffusion on Graphs via Free-Energy Gradient Flows. ICML, 2026.

[2] Shui-Nee Chow, Wuchen Li, and Haomin Zhou. A Discrete Schrödinger Equation via Optimal Transport on Graphs. Journal of Functional Analysis, 2019.

Reviewer_B5tROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies population dynamics when only snapshots are available, rather than individual trajectories. Given a graph of possible transitions and an interpolation between snapshots, the proposed approach learns a minimum-energy flow. The method, Discrete Action Matching, is a finite-state counterpart of Action Matching grounded in discrete Wasserstein geometry on graphs. Its central idea is to use relative frequencies of neighbouring states rather than the full distribution over states. This allows the authors to simulate trajectories consistent with the observed population changes, without recovering the actual trajectories of individuals.

Critical strengths

Theory appears carefully-developed and sound. The main body of the work reads rigorously, the presentation of the theoretical contribution seems complete, and formal proofs are provided in the Appendix. It states caveats, such as unidentifiability of microscopic dynamics, and clearly lays out its assumptions, such as smoothness and positivity of the mobility. This aspect represents a reliable reference for future research.

Critical weaknesses
  • Experimental scope is insufficient to show usefulness of DAM. If I understand correctly, an optimal model reproduces the probability path prescribed in Eq. 17. However, in the small state spaces tested, this equation may be used to compute the marginal probability at any time tt exactly. As the experiments evaluate only marginal distributions, they do not demonstrate a regime where a learning method is necessary. Neither do they demonstrate that a learning method is useful, since the exactly computable marginal distributions are not used as baselines. Furthermore, if learning the potential has value beyond reproducing the marginal distributions, it is not demonstrated, since the metrics used evaluate only the marginal distributions. Altogether, this work does not evaluate a setting where learning is necessary and does not provide evidence that in the settings tested learning is useful. In Section 5.2, it is stated that “Comparable accuracy to histogram DAM motivates learned log-ratios as a route to larger state spaces.”, but this claim is untested.

  • Experimental design and baseline performance require justification. The FEGF experimental protocol is substantially changed without clearly-articulated reasons. For instance, in the graphs experiment N=60N=60 with 10001000 samples per snapshot are used, instead of N=30N=30 with 50005000 samples per snapshot used by Rancati et al. (Table 1, p. 7), and only five marginals are retained, whereas Rancati et al. use a regular time grid (their time-grid ablation spans 2020 to 1000010000 steps). Moreover, in the mouse gastrulation experiment the number of cells decreased from approximately 116000116000 to 89,26789,267 and the number of developmental timepoints increased from 8 to 9. Lastly, if I am not mistaken, in the mouse gastrulation experiment, the FEGF is implemented as a time-indexed scalar potential table, whereas Rancati et al. use a neural network with a time input and a learned entropy coefficient. This makes the reliability and significance of the experimental section and the FEGF baseline unclear. For example, Rancati et al. report 0.06 mean Hellinger distance in the mouse gastrulation experiment, whereas in this work it varies between around 0.2 and 0.5 across the omitted stages. Whether this difference is due to the differences in experimental setup, in the FEGF implementation, or their combination is unclear. Overall, without further justification of the modifications to the experimental setup and the baseline implementation, the significance and fairness of the results is unclear.

Questions for authors

Clarifications

  • Could you please list the differences in experimental setting between your work and the FEGF paper, justifying why each choice was made and their effects on fairness of the comparison?
  • Have you managed to reproduce FEGF results in their original setup?
  • What is the intended use of the learned potential beyond reproducing marginals? Could you demonstrate it?

Experiments

  • How does DAM compare against the exact marginal distribution computed using Eq. 17?
  • Could you please demonstrate the scalability of DAM to a setting where computing the marginal distributions exactly is impractical or disadvantageous?
  • Could you please evaluate the neural network implementation of FEGF with time input and a learned entropy coefficient, as in Rancati et al., on the mouse gastrulation experiment?
3 comments
AuthorsOpenReview ↗

Official Comment by Authors. Part 1

Dear Reviewer B5tR,

Thank you for your thoughtful and constructive feedback. We appreciate your positive assessment of the theoretical rigor, clarity of presentation, and careful discussion of the assumptions underlying our method. Below, we address your concerns point by point.


1. "Experimental scope is insufficient to show usefulness of DAM. If I understand correctly, an optimal model reproduces the probability path prescribed in Eq. 17. [...] since the exactly computable marginal distributions are not used as baselines. [...] How does DAM compare against the exact marginal distribution computed using Eq. 17?"

Thank you for suggesting this baseline. We agree that optimal DAM reproduces the prescribed marginal path when initialized from its starting marginal. With finitely many samples, the empirical interpolant can differ from the underlying population between snapshots. We therefore compare the direct interpolant and the fitted DAM rollout against the same held-out simulator marginals.

Appendix D.1, Figure 6 compares four paths at 25%, 50%, and 75% of each observation gap. The table reports how often lr-DAM improves on its direct interpolant and gives mean Hellinger distances for cubic smoothstep (Eq. 17).

Samples per snapshotlr-DAM wins across four pathsDirect Eq. 17h-DAM (ours)lr-DAM (ours)
12811 of 120.19970.20420.1660
1,0004 of 120.06880.06360.0641

At 128 samples, lr-DAM reduced the mean error of Eq. 17 by 16.9% and improved on direct interpolation everywhere except the linear midpoint. At 1,000 samples, both DAM variants improved on direct smoothstep interpolation on average. For linear, log-linear, and Fisher–Rao interpolation, direct evaluation gave lower mean error. The benefit for marginal prediction therefore depends on the sample budget and the prescribed path.


2. "Experimental design and baseline performance require justification. The FEGF experimental protocol is substantially changed without clearly-articulated reasons. For instance, in the graphs experiment N=60N=60 with 10001000 samples per snapshot are used, instead of N=30N=30 with 50005000 samples per snapshot used by Rancati et al. (Table 1, p. 7), and only five marginals are retained, whereas Rancati et al. use a regular time grid (their time-grid ablation spans 20 to 10000 steps)."

Why we changed the protocol. We chose 60 states, 1,000 samples per snapshot, and five equally spaced observation times to test reconstruction when observations are sparse in both state and time. The larger graph and smaller sample budget give fewer observations per state, while the sparse time grid leaves most of the evolution unobserved. FEGF and both DAM variants receive the same five sampled marginals. Each model starts from the first training histogram and propagates through the remaining observation times without resets. We score its predictions against independent test histograms. Figure 5 varies graph size, sample budget, and entropy strength separately.

Rancati et al.’s Table 1. We also reconstructed the dense setting cited in your review, with 30 states and 5,000 samples at each of 101 equally spaced times. The public release does not identify the original graph instances, so we generated fresh instances from the same graph families. We used the released FEGF and Direct-FP objectives. To give all methods access to the same observations, we constructed FEGF’s velocity targets from adjacent empirical densities using their actual time separation. All methods started from the same empirical initial law and were scored against the simulator marginals.

MethodMean Hellinger distance
FEGF0.1411
Direct-FP0.0342
lr-DAM (ours)0.0369
h-DAM (ours)0.0385
Sampling reference0.0266

Direct-FP achieved the lowest mean error in the dense setting, followed by lr-DAM and h-DAM. Appendix D.1, Table 2 reports these results, and Figure 4 gives the comparison by graph family. The sampling reference measures the error obtained by drawing the same number of particles directly from the simulator law.

AuthorsOpenReview ↗

Official Comment by Authors. Part 2


3. "Moreover, in the mouse gastrulation experiment the number of cells decreased from approximately 116000 to 89267 and the number of developmental timepoints increased from 8 to 9. [...] Could you please evaluate the neural network implementation of FEGF with time input and a learned entropy coefficient, as in Rancati et al., on the mouse gastrulation experiment?"

Rancati et al. [1] describe approximately 116,000 cells at eight timepoints. We followed their released pipeline, which uses a processed scVelo atlas of 89,267 cells at nine measured stages (E6.5–E8.5). These stages define eight adjacent intervals, and the released evaluator scores seven destination stages.

We evaluated the requested time-input neural FEGF variant with a learned entropy coefficient alongside the released implementation. The table reports mean Hellinger distance. Both DAM variants used linear interpolation under both protocols.

MethodAll-stage reconstructionLeave-one-stage-out interpolation
FEGF, released potential table with fixed near-zero entropy0.04470.3201
FEGF, time-input neural variant with learned entropy0.05390.2871
h-DAM (ours)0.06560.1849
lr-DAM (ours)0.06810.1966

Both all-stage FEGF scores are close to the 0.06 reported by Rancati et al. [1]. Leave-one-stage-out interpolation is a harder task than all-stage reconstruction. The released all-stage evaluator uses each destination stage’s observed cell-type proportions to construct the transitions. In our protocol, we withhold the entire stage before training and checkpoint selection, so each method must infer its proportions from the surrounding stages. We chose this evaluation because predicting a held-out timepoint is common in the single-cell dynamics literature, including TrajectoryNet [2], PRESCIENT [3], and simulation-free Schrödinger bridges [4].

Mean error increased more than fivefold for both FEGF implementations under leave-one-stage-out evaluation. The requested neural variant improved FEGF’s held-out error from 0.3201 to 0.2871. Both DAM variants achieved lower held-out error.


4. "[...] Furthermore, if learning the potential has value beyond reproducing the marginal distributions, it is not demonstrated, since the metrics used evaluate only the marginal distributions. Altogether, this work does not evaluate a setting where learning is necessary and does not provide evidence that in the settings tested learning is useful. [...] What is the intended use of the learned potential beyond reproducing marginals? Could you demonstrate it?"

The learned potential determines jump rates. These rates let us estimate how much probability crosses an edge, in which direction, and when the flow reverses. To evaluate this use directly, we added an experiment on the discrete Schrödinger dynamics of Chow, Li, and Zhou [5] in Appendix D.4. We train lr-DAM on independently sampled state occupancies along a chain. The physical phase and signed edge currents are withheld from training.

This experiment distinguishes marginal accuracy from current accuracy. Linear and cubic smoothstep interpolation give exactly the same marginal halfway between adjacent snapshots. Their derivatives differ, however, so they predict different currents at that same time. Table 7 compares the current from a free lr-DAM rollout with the simulator current across all 31 edges and all 16 interval midpoints.

Current estimator4,096 samples per snapshot16,384 samples per snapshot
lr-DAM (ours)0.2210.109
Validation-selected smoothing and inversion0.2970.239
Direct linear interpolation0.4800.222
Direct cubic smoothstep interpolation0.8650.590

Entries are relative errors in the signed current. lr-DAM achieved the lowest error at both sample budgets. Figure 11 also shows that its rollout tracks the reversal of flow through the central edge. On a chain, the exact time derivative of the marginal determines the current analytically. The value of learning in this experiment is that, given only finite unpaired snapshots, the fitted potential estimates this physical observable more accurately than the direct estimators we evaluated.

AuthorsOpenReview ↗

Official Comment by Authors. Part 3


5. "[...] In Section 5.2, it is stated that “Comparable accuracy to histogram DAM motivates learned log-ratios as a route to larger state spaces.”, but this claim is untested. [...] Could you please demonstrate the scalability of DAM to a setting where computing the marginal distributions exactly is impractical or disadvantageous?"

We agree that comparable accuracy on mouse gastrulation alone does not demonstrate scalability. Our motivation for lr-DAM concerns the information needed during training. h-DAM constructs an empirical histogram and reads the masses at both endpoints of an edge. lr-DAM learns the corresponding density ratio from sampled states and their neighbors. It can therefore train the action potential without constructing a full joint histogram.

Appendix D.3 tests whether replacing histogram ratios with learned ratios preserves generation accuracy on the discretized Swiss roll. With 512 training pairs and 32 sampling steps, lr-DAM achieved a mean TV of 0.2885, compared with 0.2952 for h-DAM (Table 5). Table 6 separately examines the effect of learning the ratios and the action potential.

These results show that learned local ratios can replace histogram ratios while retaining comparable generation accuracy on this task. A full histogram is still feasible for the discretized Swiss roll. The revised experiments therefore support the use of learned ratios, while their performance in a regime where a full histogram is left for future work.


Conclusion

Thank you again for suggesting the direct interpolation baseline and for asking how the learned potential can be evaluated beyond marginal reconstruction. We hope the revised experiments clarify when direct interpolation is sufficient and how the learned potential can be used to estimate edge currents from unpaired snapshots. We would appreciate your reconsideration of the paper in light of these results.

References

[1] Dario Rancati, Jan Maas, and Francesco Locatello. Learning Discrete Diffusion on Graphs via Free-Energy Gradient Flows. ICML, 2026.

[2] Tong, Alexander, et al. “TrajectoryNet: A Dynamic Optimal Transport Network for Modeling Cellular Dynamics.” Proceedings of the 37th International Conference on Machine Learning, vol. 119, PMLR, 2020, pp. 9526–9536.

[3] Yeo, Grace Hui Ting, et al. “Generative Modeling of Single-Cell Time Series with PRESCIENT Enables Prediction of Cell Trajectories with Interventions.” Nature Communications, vol. 12, 2021, article 3222.

[4] Tong, Alexander Y., et al. “Simulation-Free Schrödinger Bridges via Score and Flow Matching.” Proceedings of the 27th International Conference on Artificial Intelligence and Statistics, vol. 238, PMLR, 2024, pp. 1279–1287.

[5] Shui-Nee Chow, Wuchen Li, and Haomin Zhou. A Discrete Schrödinger Equation via Optimal Transport on Graphs. Journal of Functional Analysis, 2019.

Comments on the paper

1 comment

Dear Reviewers,

We sincerely thank all reviewers for their thoughtful feedback. We appreciate the recognition of the theoretical rigor, clarity of presentation, and careful discussion of the assumptions underlying DAM (B5tR, knx3). We also thank DSkG for highlighting the value of training directly from samples, the evaluation across graph structures and mobility functions, and the illustration in Figure 1.

Our main contribution is methodological. Building on discrete Wasserstein geometry [2] and the study of population dynamics on graphs by Rancati et al. [1], we extend Action Matching [3] to finite state spaces. The central observation is that the density dependence of the discrete action can be expressed through ratios of probability masses at neighboring vertices. This allows us to learn an action potential from unpaired marginal samples and use it to simulate a graph-supported Markov process.

In response to your suggestions, we revised the main text and reorganized Appendix D by additional experiments. Changes are highlighted in orange in the updated manuscript.

  • (knx3, DSkG, B5tR) We revised the abstract and contribution statement to give both DAM variants explicit treatment. Histogram DAM (h-DAM) computes neighboring density ratios from empirical histograms and is a natural choice when the states are well sampled. Log-ratio DAM (lr-DAM) learns these ratios from sampled states and their neighbors. Once fitted, lr-DAM can generate a trajectory from a given starting state using its learned fields, without tracking an evolving histogram or particle ensemble.

  • (B5tR, DSkG) We added direct interpolation baselines to distinguish the choice of marginal path from the dynamics learned along it. Appendix D.1 compares linear, cubic smoothstep (Eq. 17), log-linear, and Fisher–Rao interpolation against h-DAM and lr-DAM at held-out times between snapshots (Figure 6). Appendix D.2 gives the corresponding comparison on mouse gastrulation: direct evaluation of the selected path and its pointwise minimum-action dynamics yield nearly identical held-out errors (Table 4).

  • (B5tR, knx3, DSkG) We reconstructed the dense synthetic benchmark from Table 1 of Rancati et al. [1] and clarified why our main experiment uses fewer samples and observation times. Appendix D.1 reports aggregate results (Table 2) and a breakdown by graph family (Figure 4), including Euclid-MST and Delaunay. Both DAM variants improve on FEGF on these two families, while Direct-FP achieves the lowest error on both comparable with DAM.

  • (knx3, DSkG) We clarified the role of the graph geometry in the FEGF comparison. We express the action in terms of vertex probabilities because these are the quantities observed through samples. Applying the mobility directly to these probabilities is a separate modeling choice. As detailed in Appendix A.3 and made explicit in Appendix D.1, FEGF applies the logarithmic mean to densities relative to the stationary distribution. Consequently, the conductances in Figure 2 can differ on irregular graphs. The figure compares the complete methods under their respective geometries.

  • (knx3, DSkG, B5tR) We added a Swiss-roll generation experiment to evaluate the learned ratios in a downstream task. Appendix D.3 compares h-DAM and lr-DAM with FEGF and SEDD at two training sample budgets (Table 5 and Figure 9). We also vary the ratio estimator and the action solver separately (Table 6 and Figure 10). Learned ratios give slightly lower generation TV at both budgets, while histogram ratios place less mass outside the observed reference support.

  • (B5tR, DSkG) We added a physical experiment that evaluates signed edge currents directly. In Appendix D.4, lr-DAM observes only unpaired state occupancies from discrete Schrödinger dynamics on a chain. We reserve the simulator currents for evaluation. The learned rollout achieves lower current error than direct interpolation and smoothing followed by inversion at both sample budgets (Table 7), and it captures the reversal of flow through the central edge (Figure 11). This experiment demonstrates a use of the learned potential beyond reproducing marginal distributions.

Thank you again for prompting us to test where learning improves on direct interpolation, where histograms suffice, and what the learned potential can reveal about motion between snapshots. Our detailed responses to the individual comments follow.


References

[1] Rancati, Dario, et al. “Learning Discrete Diffusion on Graphs via Free-Energy Gradient Flows.” Proceedings of the 43rd International Conference on Machine Learning, PMLR, 2026.

[2] Maas, Jan. “Gradient Flows of the Entropy for Finite Markov Chains.” Journal of Functional Analysis, vol. 261, no. 8, 2011, pp. 2250–2292.

[3] Neklyudov, Kirill, et al. “Action Matching: Learning Stochastic Dynamics from Samples.” Proceedings of the 40th International Conference on Machine Learning, PMLR, 2023.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.