acceptodds
Under review as a conference paper at ICLR 2027

Continuous-Time Trajectory Generation from Discrete Observations with Stochasticity

20 views

Abstract

Physical systems evolve continuously in time, yet their states are typically observed only at discrete times. Generating trajectories consistent with their probability densities from such observations therefore requires capturing the continuous-time evolution rather than only learning transition mappings between consecutive observations. We propose , a framework that directly estimates the probability velocity field induced by the stochastic differential equation (SDE) which governs this continuous-time distributional evolution. PhiBE-Flow learns from discrete observations using a model-free approach requiring neither known SDE coefficients nor score estimation. We establish convergence guarantees for the method, accounting for both time-discretization and finite-sample errors. We evaluate PhiBE-Flow on systems of increasing complexity, from controlled stochastic numerical systems to Navier–Stokes dynamics and real-world videos. Our results show that PhiBE-Flow accurately recovers probability flows of stochastic dynamics, preserves multiscale physical statistics, and improves video generation performance over representative baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Generative modeling, Continuous-time generation, Stochastic dynamical systems, Probability flows, Video generation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_dvKoOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes PhiBE-Flow, a framework for learning continuous-time probability velocity fields from discrete observations of stochastic dynamical systems. By combining PhiBE approximations with the Fokker–Planck equation, the authors derive a variational objective that avoids explicit estimation of SDE coefficients and density scores. The paper provides theoretical convergence guarantees for time-discretization and finite-sample errors. Experiments on stochastic dynamical systems, Navier–Stokes equations, and video generation demonstrate the effectiveness of the proposed method, with improved performance over representative baselines on high-dimensional generation tasks.

Critical strengths
  1. The paper presents a well-motivated learning formulation that combines the Fokker–Planck equation, PhiBE approximations, and integration by parts to directly learn probability velocity fields without requiring known SDE coefficients or explicit score estimation.

  2. The experiments cover a diverse range of stochastic systems and demonstrate promising performance, particularly on Navier–Stokes dynamics and video generation.

Critical weaknesses
  1. The paper lacks ablation studies to validate the contributions of its key components. In particular, the role of the Jacobian term in the proposed objective is not isolated, making it difficult to assess its importance in capturing stochastic diffusion effects.

  2. The experimental comparisons are limited to representative generative baselines. Comparisons with methods that explicitly estimate SDE drift and diffusion coefficients or learn probability flows through score estimation would provide stronger evidence for the advantages of the proposed direct learning approach.

Questions for authors
  1. Could the authors provide an ablation study removing the Jacobian term and comparing against direct drift regression to clarify its contribution to learning the probability velocity field?

  2. Could the authors include comparisons with approaches that first estimate SDE coefficients or density scores and then construct the corresponding probability flow, under matched experimental settings?

3 comments

We thank the reviewer for the careful reading and constructive suggestions. We address the two questions below.

1. Ablation of the Jacobian term

We added a controlled ablation on both reversible and non-reversible OU examples, comparing the full PhiBE-Flow objective with ordinary increment regression using the same five-layer, width-50 MLP. The latter removes the Jacobian term and minimizes

Ldrift(θ)=1Nx(Nt−1)∑k=1Nx∑i=0Nt−2∥vθ(xti(k),ti)−xti+1(k)−xti(k)Δt∥22.\mathcal{L}_{\mathrm{drift}}(\theta)=\frac{1}{N_{x}(N_{t}-1)}\sum_{k=1}^{N_{x}}\sum_{i=0}^{N_{t}-2}\left\lVert v_{\theta}(x_{t_{i}}^{(k)},t_{i})-\frac{x_{t_{i+1}}^{(k)}-x_{t_{i}}^{(k)}}{\Delta t}\right\rVert_{2}^{2}.

For the reversible/non-reversible cases, the full objective reduces:

  • Velocity RMSE: from 0.9796/1.0697 to 0.3917/0.4088.
  • Probability-flow path RMSE: from 0.1937/0.2787 to 0.1090/0.1294.
  • Endpoint covariance error: from 0.2244/0.3524 to 0.0715/0.1081.

Affine-model ablations, obtained by setting C=0C=0, show consistent improvements. These results support the importance of the Jacobian term for recovering probability velocities and marginal distributions.

The Fig. 9 in our revised manuscript compares analytical probability-flow trajectories with drift-only MLP and affine rollouts. The drift-only models fail to preserve the correct ensemble spread, illustrating the effect of omitting the Jacobian correction.

These reduced-scale experiments use T=1, 2,048 trajectories, 201 observation times, an 80/20 trajectory-level split, and one seed. Velocity and path errors are evaluated against the analytical OU probability velocity and probability-flow trajectories, respectively.

In the revised manuscript, we report the MLP ablation in Section 5.1 and the experimental details in Appendix B.2.

2. Comparison with separate SDE-coefficient and score estimation

We compare PhiBE-Flow with a two-stage approach that estimates the drift and diffusion coefficients, learns the marginal score using denoising score matching (DSM), and then constructs the probability velocity.

To test this comparison in higher dimension, we considered 12-dimensional OU systems with rotational-coupling strengths 0, 2, and 5, corresponding to reversible, non-reversible, and stronger-rotation cases. For each setting, we used 512 trajectories (409 for training), 51 observation times on [0,1][0,1], and 12 paired data seeds with fixed SDE coefficients.

We compared the time-indexed affine PhiBE field with a two-stage estimator using pooled affine drift regression, residual-based diffusion estimation, and a time-indexed affine score trained with the DSM loss (Song et al., 2021), with σdsm=0.05\sigma_{\mathrm{dsm}}=0.05. Both methods use the same training trajectories.

As shown in the Fig. 10 in our revised manuscrips, the time-averaged Gaussian W2/dW_{2}/\sqrt{d} errors, averaged over seeds, were:

SettingPhiBESDE+DSM
Reversible OU0.06660.1156
Non-reversible OU0.06660.1146
Stronger rotation0.06860.1147

PhiBE had lower error in all 12 paired seeds in each setting. These results provide evidence that, in the tested high-dimensional affine OU settings, jointly learning the probability velocity can reduce errors accumulated by this separate coefficient-and-DSM-score pipeline.

This comparison also highlights a limitation of identifying continuous-time SDE coefficients from discrete observations. Consider two linear SDEs with drifts

b0(x)=−x,b1(x)=(−I+2πhJ)x,J=(0−110),b_{0}(x)=-x,\qquad b_{1}(x)=\left(-I+\frac{2\pi}{h}J\right)x,\qquad J=\begin{pmatrix}0&-1\\\\1&0\end{pmatrix},

and the same isotropic diffusion loading σI\sigma I. At observation interval hh, both induce exactly the same transition kernel,

Ph(⋅∣x)=N!(e−hx,,σ22(1−e−2h)I),P_{h}(\cdot\mid x)=\mathcal{N}\\!\left(e^{-h}x,\\,\frac{\sigma^2}{2}(1-e^{-2h})I\right),

despite having different intermediate-time dynamics. Thus, even perfect knowledge of the sampled transitions does not uniquely identify the continuous-time SDE.

PhiBE-Flow does not require SDE coefficient identification as an intermediate step: it directly estimates a finite-step approximation to the probability velocity through its variational objective. For the non-rotating OU system with a centered Gaussian marginal, minimizing the full objective over linear velocity fields gives

vh=αhv⋆,αh=1−e−2h2h=1−h+O(h2),v_{h}=\alpha_{h}v^\star,\qquad \alpha_{h}=\frac{1-e^{-2h}}{2h}=1-h+O(h^2),

making the finite-step approximation bias explicit without selecting an SDE-generator branch.

In the revised manuscript, we add these evaluations and their experimental settings to Appendix B.8, with a brief reference at the end of the OU experiment subsection.

References

Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. In International Conference on Learning Representations, 2021. Paper.

Thank you for the helpful response. My main concerns have been addressed, and I am raising my score to 3.

We sincerely thank you for your constructive feedback and careful consideration.

Reviewer_tWt5OpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces PhiBE-Flow, a framework for learning continuous-time probability flows from discrete observations of stochastic dynamical systems. Starting from the Fokker–Planck equation, it derives a score-free variational objective and replaces unknown drift and diffusion terms with short-time increment estimates. The resulting deterministic dynamics aim to reproduce the time-marginal distributions of the underlying SDE. The paper establishes population approximation bounds and finite-sample guarantees for fixed linear-basis models, and evaluates the method on OU processes, stochastic double pendulums, Navier–Stokes dynamics, and KTH videos.

Critical strengths
  • An appealing objective for learning probability velocity directly. The integration-by-parts formulation removes the need for explicit score estimation, while local increment statistics make the objective accessible without separately fitting complete drift and diffusion models. This provides a useful connection between stochastic dynamics and deterministic probability flows.

  • Validation across analytical and applied settings. The OU experiments permit comparison against an analytical probability velocity, providing a meaningful check of the formulation. The fluid and video experiments also show encouraging improvements over MCVD and PFI, including better preservation of multiscale fluid statistics.

Critical weaknesses

Major weaknesses

  • Marginal distribution matching does not establish stochastic trajectory or conditional forecasting accuracy. The formulation learns a deterministic ODE that matches the SDE’s time-marginal distributions. For a fixed initial state, however, this ODE produces a single future trajectory, whereas the SDE generally induces a nondegenerate distribution of futures. Matching marginals therefore does not imply matching transition distributions, temporal correlations, or the joint distribution over trajectories.

  • The experiments do not sufficiently isolate the benefit of the proposed stochastic correction or continuous-time formulation. The comparisons against MCVD and PFI are useful, but do not establish whether the gains arise from the covariance-weighted Jacobian term, the continuous-time parameterization, architecture choices, or optimization settings. A particularly important baseline is the same velocity network trained by ordinary increment regression, obtained by removing the Jacobian term. Comparisons with neural ODE and neural or latent SDE approaches would also more directly test the claimed advantages. Furthermore, continuous-time generation is a central motivation, yet the reported experiments primarily evaluate trajectories at sampled times. Evaluation at withheld intermediate times, across observation intervals and rollout step sizes, would provide stronger evidence for this contribution. The computational cost of the Jacobian-dependent objective also needs characterization, including how it is evaluated in high-dimensional experiments and how training time and memory compare with the baselines.

Minor weaknesses

  • Presentation issues At L1118, “such OU process with Gaussian initial” should read “such an OU process with a Gaussian initial distribution.” Theorem 3 uses WW for the Brownian motion, whereas the corresponding SDE uses BtB_t. Please make the notation consistent.
Questions for authors

see weaknesses.

5 comments

For major weakness 1: Marginal and Pathwise Accuracy

We thank the reviewer for pointing out this important distinction. We agree that matching the SDE's time-marginal distributions does not, by itself, establish matching transition distributions, temporal correlations, or the joint distribution over trajectories.

PhiBE-Flow learns a deterministic probability-flow ODE whose population evolution is designed to reproduce the SDE's time marginals. Consequently, for a fixed initial state (or conditioning context), the ODE produces a deterministic rollout and does not represent a non-degenerate conditional distribution over future paths. Our method therefore targets the physical-time evolution of the SDE's marginals; we do not claim to recover its full stochastic path law or to provide calibrated conditional uncertainty.

The deterministic formulation is useful for applications such as video generation, where a coherent and reproducible rollout may be desired. Once initialized, the ODE generates a continuous-time sequence without requiring stochastic increments during integration. This is a practical motivation for the deterministic formulation, rather than a claim that stochastic variability is undesirable or that the ODE captures all possible futures. The experiments evaluate generation quality and physical or video statistics, and should be interpreted within this scope.

In the revised manuscript, we clarify this scope in the Introduction by describing PhiBE-Flow as a deterministic probability flow representing the physical-time evolution of the underlying SDE's marginal distributions. Specifically, we revise the statement beginning at line 074.

For major weakness 2: Isolating the Method's Empirical Benefits

We thank the reviewer for this constructive and detailed comment. We understand it as raising four related questions:

  • (a) A controlled ablation isolating the contribution of the Jacobian term.
  • (b) Comparisons with neural ODE and neural or latent SDE approaches.
  • (c) Evaluation at withheld intermediate times and across different observation intervals and rollout step sizes.
  • (d) The computational cost of the Jacobian-dependent objective.

We address these points separately below.

(a) Controlled ablation of the Jacobian term

We added a controlled ablation on both reversible and non-reversible OU examples, comparing the full PhiBE-Flow objective with ordinary increment regression using the same five-layer, width-50 MLP. The latter removes the Jacobian correction and minimizes

Ldrift(θ)=1Nx(Nt−1)∑k=1Nx∑i=0Nt−2∥vθ(xti(k),ti)−xti+1(k)−xti(k)Δt∥22.\mathcal{L}_{\mathrm{drift}}(\theta)=\frac{1}{N_{x}(N_{t}-1)}\sum_{k=1}^{N_{x}}\sum_{i=0}^{N_{t}-2}\left\lVert v_{\theta}(x_{t_{i}}^{(k)},t_{i})-\frac{x_{t_{i+1}}^{(k)}-x_{t_{i}}^{(k)}}{\Delta t}\right\rVert_{2}^{2}.

For the reversible/non-reversible cases, the full objective reduces:

  • Velocity RMSE: from 0.9796/1.0697 to 0.3917/0.4088.
  • Probability-flow path RMSE: from 0.1937/0.2787 to 0.1090/0.1294.
  • Endpoint covariance error: from 0.2244/0.3524 to 0.0715/0.1081.

Affine-model ablations, obtained by setting C=0C=0, show consistent improvements. These results support the importance of the Jacobian correction for recovering probability velocities and marginal distributions.

The Fig. 9 in our revised manuscript compares analytical probability-flow trajectories with drift-only MLP and affine rollouts. The drift-only models fail to preserve the correct ensemble spread, illustrating the effect of omitting the Jacobian correction.

These reduced-scale experiments use T=1, 2,048 trajectories, 201 observation times, an 80/20 trajectory-level split, and one seed. Velocity and path errors are evaluated against the analytical OU probability velocity and probability-flow trajectories, respectively.

In the revised manuscript, we report the MLP ablation in Section 5.1 and the experimental details in Appendix B.2.

(b) Relation to neural ODE and neural or latent SDE approaches

We compare PhiBE-Flow with a two-stage approach that estimates the drift and diffusion coefficients, learns the marginal score using denoising score matching (DSM), and then constructs the probability velocity.

To test this comparison in higher dimension, we considered 12-dimensional OU systems with rotational-coupling strengths 0, 2, and 5, corresponding to reversible, non-reversible, and stronger-rotation cases. For each setting, we used 512 trajectories (409 for training), 51 observation times on [0,1][0,1], and 12 paired data seeds with fixed SDE coefficients.

We compared the time-indexed affine PhiBE field with a two-stage estimator using pooled affine drift regression, residual-based diffusion estimation, and a time-indexed affine score trained with the DSM loss (Song et al., 2021), with σdsm=0.05\sigma_{\mathrm{dsm}}=0.05. Both methods use the same training trajectories.

As shown in the Fig. 10 in our revised manuscrips, the time-averaged Gaussian W2/dW_{2}/\sqrt{d} errors, averaged over seeds, were:

SettingPhiBESDE+DSM
Reversible OU0.06660.1156
Non-reversible OU0.06660.1146
Stronger rotation0.06860.1147

PhiBE had lower error in all 12 paired seeds in each setting. These results provide evidence that, in the tested high-dimensional affine OU settings, jointly learning the probability velocity can reduce errors accumulated by this separate coefficient-and-DSM-score pipeline.

This comparison also highlights a limitation of identifying continuous-time SDE coefficients from discrete observations. Consider two linear SDEs with drifts

b0(x)=−x,b1(x)=(−I+2πhJ)x,J=(0−110),b_{0}(x)=-x,\qquad b_{1}(x)=\left(-I+\frac{2\pi}{h}J\right)x,\qquad J=\begin{pmatrix}0&-1\\\\1&0\end{pmatrix},

and the same isotropic diffusion loading σI\sigma I. At observation interval hh, both induce exactly the same transition kernel,

Ph(⋅∣x)=N!(e−hx,,σ22(1−e−2h)I),P_{h}(\cdot\mid x)=\mathcal{N}\\!\left(e^{-h}x,\\,\frac{\sigma^2}{2}(1-e^{-2h})I\right),

despite having different intermediate-time dynamics. Thus, even perfect knowledge of the sampled transitions does not uniquely identify the continuous-time SDE.

PhiBE-Flow does not require SDE coefficient identification as an intermediate step: it directly estimates a finite-step approximation to the probability velocity through its variational objective. For the non-rotating OU system with a centered Gaussian marginal, minimizing the full objective over linear velocity fields gives

vh=αhv⋆,αh=1−e−2h2h=1−h+O(h2),v_{h}=\alpha_{h}v^\star,\qquad \alpha_{h}=\frac{1-e^{-2h}}{2h}=1-h+O(h^2),

making the finite-step approximation bias explicit without selecting an SDE-generator branch.

In the revised manuscript, we add these evaluations and their experimental settings to Appendix B.8, with a brief reference at the end of the OU experiment subsection.

(c) Withheld intermediate times, rollout step sizes, and observation intervals

We conducted supplementary experiments on both reversible and non-reversible OU processes.

Withheld intermediate times. We trained using an observation interval of Δt=0.005\Delta t=0.005 and evaluated at 40 physical times absent from the training grid. Using 2,048 independent initial samples, we compared the generated distributions with the analytical OU marginals. With Euler step size h=1/1280h=1/1280, the time-averaged sliced W2W_{2} errors are 0.0678 and 0.0766 for the reversible and non-reversible processes, respectively.

Rollout step sizes. Keeping the trained models fixed, we varied the Euler step size over

h∈1/160,1/320,1/640,1/1280.h\in\\{1/160,1/320,1/640,1/1280\\}.

The average sliced W2W_{2} error varies by less than 2.5% relative to its minimum across these steps in each example. RK4 reference rollouts with h=1/640h=1/640 give errors of 0.0678 and 0.0771, respectively. These results indicate limited sensitivity to the tested rollout step sizes at Δt=0.005\Delta t=0.005.

Observation intervals. We additionally conducted a separate study with

Δt∈0.0125,0.025,0.05,\Delta t\in\\{0.0125,0.025,0.05\\},

using a fixed training-update budget and evaluating all settings at the same withheld times. The mean sliced W2W_{2} errors are:

Observation intervalReversible OUNon-reversible OU
0.01250.18950.2724
0.0251.56152.3110
0.054.25884.7555

In the revised manuscript, we add these evaluations and their experimental settings to Appendix B.3, with a brief reference at the end of the OU experiment subsection.

(d) Computational cost of the Jacobian-dependent objective

We profile the KTH model implementations at batch size 32. Training measurements include loss computation, backpropagation, and optimizer updates.

MethodTraining time (s/batch)Inference time (s/batch)Training peak memory (GiB)
PhiBE-Flow0.572.294.66
PhiBE-Flow (without Jacobian)0.252.292.66
MCVD1.48165.379.45
PFI0.32245.242.69

The controlled comparison shows that the correction increases training-update time by 2.23× and peak allocated memory by 1.75× at this batch size. These measurements demonstrate the feasibility of the exact VJP implementation for the KTH latent model while making its additional training cost explicit.

In the revised manuscript, we add implementation details and a computational-cost comparison to Section 5.3. We also clarify that the correction is used only during training and adds no computation at inference when the architecture and solver settings are unchanged. The profiling protocol and sampling settings are provided in Appendix B.7.

Presentation issues

We thank the reviewer for pointing out these issues. We have unified the Brownian-motion notation in Theorem 3 with the corresponding SDE and proof, using BB consistently. We have also corrected the wording to “such an OU process with a Gaussian initial distribution.”

References

  1. Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. In International Conference on Learning Representations, 2021. Paper.

  2. Yifan Chen, Mark Goldstein, Mengjian Hua, Michael Samuel Albergo, Nicholas Matthew Boffi, and Eric Vanden-Eijnden. Probabilistic Forecasting with Stochastic Interpolants and Föllmer Processes. In Proceedings of the 41st International Conference on Machine Learning, 2024. Paper.

  3. Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. MCVD: Masked Conditional Video Diffusion for Prediction, Generation, and Interpolation. In Advances in Neural Information Processing Systems, 35, 2022. Paper.

  4. Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a local SVM approach. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. DOI.

Reviewer_xhutOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors attempt to learn a velocity model v(x,t)=xt˙v(x,t) = \dot{x_t} from paired transitions (xt,xt+δt)(x_t, x_{t+\delta t}). This is in contrast to, say, flow matching or stochastic interpolants which construct hand-crafted probability paths that are empirical useful for learning (usually some form of denoising). For this, the authors directly regress onto xt+δt−xtx_{t+\delta t} - x_t with an added Jacobian-vector product loss term. Authors show the validity of the method on small scale, toy-ish experiments.

Critical strengths
  • The method introduces a new jvp (to me at least) which makes the loss diverge from regular regression, and improves results.
  • Results are strong compared to baselines evaluated (caveat here, see weaknesses)
Critical weaknesses
  • I don't fully understand the motivation for the introduction of the jvp, can the authors elaborate in the text maybe via an example?
  • The experiment section is very limited to low dimensional data. I guess this is due to needing to compute a jvp. Have the authors considered any heuristics to avoid computing the jvp? I imagine this must be a more fundamental problem that appears in many circles.
  • Paper mentions several methods that also learn in this regime of pairwise data, but doesn't include them in the experiments. This makes it hard to understand the gain from the new velocity objective.
Questions for authors
  • I understand the evals are meant to compare against current methods. However, the method only requires pair-wise samples instead of trajectories (which is good). Have the authors considered use-cases where only this type of information is available, and not trajectories?
  • See weaknesses, thanks.
4 comments

We thank the reviewer for the careful reading and constructive feedback. We organize our response around the following issues.

1. Why the Jacobian correction matters

The Jacobian correction accounts for stochastic diffusion in the evolution of the density. Ordinary increment regression approximates the SDE drift, which generally differs from the probability velocity. For example, consider a stationary OU process:

dXt=−γXtdt+2DdBt,X0∼N(0,D/γ),γ,D>0.dX_t=-\gamma X_t dt+\sqrt{2D} dB_t,\qquad X_0\sim\mathcal{N}(0,D/\gamma),\qquad \gamma,D>0.

Its drift is −γx-\gamma x, but its probability velocity is

v⋆(x)=−γx−D∂xlog⁡ρ(x)=0.v^\star(x)=-\gamma x-D \partial_x\log\rho(x)=0.

Drift and diffusion balance to preserve the stationary density; integrating the drift alone as an ODE would instead contract it.

Moreover, we added a controlled ablation on both reversible and non-reversible OU examples, comparing the full PhiBE-Flow objective with ordinary increment regression using the same five-layer, width-50 MLP. The latter removes the Jacobian correction and minimizes

Ldrift(θ)=1Nx(Nt−1)∑k=1Nx∑i=0Nt−2∥vθ(xti(k),ti)−xti+1(k)−xti(k)Δt∥22.\mathcal{L}_{\mathrm{drift}}(\theta)=\frac{1}{N_{x}(N_{t}-1)}\sum_{k=1}^{N_{x}}\sum_{i=0}^{N_{t}-2}\left\lVert v_{\theta}(x_{t_{i}}^{(k)},t_{i})-\frac{x_{t_{i+1}}^{(k)}-x_{t_{i}}^{(k)}}{\Delta t}\right\rVert_{2}^{2}.

For the reversible/non-reversible cases, the full objective reduces:

  • Velocity RMSE: from 0.9796/1.0697 to 0.3917/0.4088.
  • Probability-flow path RMSE: from 0.1937/0.2787 to 0.1090/0.1294.
  • Endpoint covariance error: from 0.2244/0.3524 to 0.0715/0.1081.

Affine-model ablations, obtained by setting C=0C=0, show consistent improvements. These results support the importance of the Jacobian correction for recovering probability velocities and marginal distributions.

The Fig. 9 in our revised manuscript compares analytical probability-flow trajectories with drift-only MLP and affine rollouts. The drift-only models fail to preserve the correct ensemble spread, illustrating the effect of omitting the Jacobian correction.

These reduced-scale experiments use T=1T=1, 2,048 trajectories, 201 observation times, an 80/20 trajectory-level split, and one seed. Velocity and path errors are evaluated against the analytical OU probability velocity and probability-flow trajectories, respectively.

In the revised manuscript, we report the MLP ablation in Section 5.1 and the experimental details in Appendix B.2.

2. Structural simplification, dimensionality, and computational cost

(a) Structural simplification

Our formulation admits a structural simplification in the constant-isotropic diffusion setting. As established in Appendix A.4, when Σ=σ2I\Sigma=\sigma^2I, the covariance-weighted Jacobian term reduces to

Σ:Jxvθ=σ2∇x⋅vθ.\Sigma:J_xv_\theta=\sigma^2\nabla_x\cdot v_\theta.

This simplifies the treatment of the diffusion covariance to a scalar coefficient multiplying the divergence.

More generally, the empirical correction can be evaluated exactly using a single Jacobian–vector product (JVP) or vector–Jacobian product (VJP), without constructing the full Jacobian. Specifically, for a fixed observed increment δ\delta,

δ⊤JxvθδΔt=δ⊤JVP⁡x(vθ;δ)Δt=VJP⁡x(vθ;δ)⊤δΔt.\frac{\delta^\top J_{x}v_{\theta}\delta}{\Delta t}=\frac{\delta^\top\operatorname{JVP}_{x}(v_{\theta};\delta)}{\Delta t}=\frac{\operatorname{VJP}_{x}(v_{\theta};\delta)^\top\delta}{\Delta t}.

Our implementation uses the single-VJP formulation followed by an inner product. Thus, avoiding explicit Jacobian construction does not require an additional heuristic approximation. Training does, however, differentiate through this contraction, and we quantify that overhead below.

In the revised manuscript, we added a separate Appendix B.7 explaining the JVP/VJP implementation and its differentiation requirements.

(b) Dimensionality and computational measurements

Our application experiments extend beyond low-dimensional synthetic systems. In particular, the KTH velocity model (Schuldt et al., 2004) operates on a 4×8×84\times8\times8 latent tensor, corresponding to a 256-dimensional state.

We profile the KTH model implementations at batch size 32. Training measurements include loss computation, backpropagation, and optimizer updates.

MethodTraining time (s/batch)Inference time (s/batch)Training peak memory (GiB)
PhiBE-Flow0.572.294.66
PhiBE-Flow (without Jacobian)0.252.292.66
MCVD1.48165.379.45
PFI0.32245.242.69

The controlled comparison shows that the correction increases training-update time by 2.23× and peak allocated memory by 1.75× at this batch size. These measurements demonstrate the feasibility of the exact VJP implementation for the KTH latent model while making its additional training cost explicit.

In the revised manuscript, we add a computational-cost discussion at the end of the Video Generation subsection, together with Table 3 on page 10.

3. Comparison with stochastic-interpolant-based methods

We would like to clarify that our experiments already include PFI (Chen et al., 2024), a stochastic-interpolant-based forecasting method that learns from observed transitions, on both the Navier–Stokes and KTH benchmarks.

However, we agree that these comparisons alone do not isolate the effect of the proposed objective from architectural and optimization differences. We therefore added the controlled Jacobian ablation described above.

Specifically, we compare the full PhiBE-Flow objective with ordinary increment regression obtained by removing the Jacobian correction. The full objective achieves lower velocity RMSE, probability-flow trajectory RMSE, and endpoint covariance error in both examples, supporting the contribution of the correction to learning the probability velocity and recovering marginal distributions.

4. Applications with paired observations rather than complete trajectories

We thank the reviewer for highlighting this potential advantage. Our method requires only paired observations at known time intervals, rather than complete trajectories. This data setting arises naturally in experimental fluid mechanics and single-cell biology, where an individual particle or cell can be observed at two distinct times even when continuous tracking is unavailable.

For example, double-frame particle tracking velocimetry matches particles between two images acquired at a known time separation, without tracking them across successive image pairs (Baek and Lee, 1996; Fuchs et al., 2017).

Similarly, Live-seq enables transcriptomic measurements of the same living cell before and after a biological transition, as demonstrated for macrophage responses to LPS stimulation and adipogenic differentiation (Chen et al., 2022).

These settings provide potential applications of our framework, subject to its modeling and time-discretization assumptions.

References

  1. S. J. Baek and S. J. Lee. A new two-frame particle tracking algorithm using match probability. Experiments in Fluids, 22(1):23–32, 1996.

  2. Wanze Chen, Orane Guillaume-Gentil, Pernille Yde Rainer, Christoph G. Gäbelein, Wouter Saelens, Vincent Gardeux, Amanda Klaeger, Riccardo Dainese, Magda Zachara, Tomaso Zambelli, et al. Live-seq enables temporal transcriptomic recording of single cells. Nature, 608(7924):733–740, 2022.

  3. Yifan Chen, Mark Goldstein, Mengjian Hua, Michael Samuel Albergo, Nicholas Matthew Boffi, and Eric Vanden-Eijnden. Probabilistic Forecasting with Stochastic Interpolants and Föllmer Processes. In Proceedings of the 41st International Conference on Machine Learning, 2024. Paper.

  4. Thomas Fuchs, Rainer Hain, and Christian J. Kähler. Non-iterative double-frame 2D/3D particle tracking velocimetry. Experiments in Fluids, 58(9):119, 2017.

  5. Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a local SVM approach. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. DOI.

Thank you for clarifying and answering my questions. I agree the method is interesting for this pairwise setting. I will increase my score.

We sincerely thank you for your time, constructive feedback and careful consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision 1: Summary of Changes

We thank all reviewers for their careful reading and constructive feedback. We have revised the manuscript as summarized below. Our individual responses provide further details.

  • Scope clarification: Clarified that PhiBE-Flow models the evolution of time-marginal distributions, rather than recovering the full stochastic path distribution (Introduction).
  • Additional experiments: Added controlled Jacobian ablations on reversible and non-reversible OU processes, and a 12-dimensional comparison with a separate SDE-coefficient and DSM-score estimation pipeline (Section 5.1; Appendices B.2 and B.8).
  • Continuous-time evaluation: Added evaluations at withheld intermediate times and across observation intervals and rollout step sizes (Appendix B.3).
  • Implementation and computational cost: Added the exact VJP implementation, training-time and memory measurements, and inference-time comparisons (Section 5.3; Appendix B.7).
  • Identifiability and presentation: Discussed the limitations of continuous-time SDE identification from discrete observations (Appendix B.8).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.