acceptodds
Under review as a conference paper at ICLR 2027

Learning Generative Dynamics with Soft Path-Law Constraints: A McKean-Vlasov FBSDE Approach

Abstract

We introduce a unified framework for learning stochastic dynamics from distributional observations and discretely observed trajectories. The key contribution is to combine soft law constraints, which regularize fitting to empirical distributions, with path-dependent supervision, which captures temporal structure that marginal distributions alone cannot identify. We formulate the problem as a discrete path-dependent McKean–Vlasov control problem, penalizing endpoint, intermediate marginal, and joint sampled-path laws within a single objective. Its optimality system takes the form of a McKean–Vlasov forward-backward stochastic differential equation (FBSDE), in which the marginal terms act continuously in time while the path-law term instead induces conditional adjoint jumps at the observation dates. We derive the reduced FBSDE for quadratic control cost and constant diffusion, establish local well-posedness for a Maximum Mean Discrepancy (MMD) path-law formulation, and develop a sample-based neural solver with causal representations. Experiments on synthetic transport, latent face transport, and human motion validate the framework. In particular, a two-state AMASS experiment under path-law supervision captures both low-to-high and high-to-low transitions, which marginal supervision alone cannot distinguish.

Primary area:
generative models
Keywords:
Mean-field-control, Generative modeling, McKean-Vlasov FBSDEs, Soft-law constraints

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

The paper proposes a formulation of diffusion as a control problem that has energy loss, the mismatch in the marginal distribution at every time loss, and the mismatch of the joint finitely many observation time path loss. This fits into the framework of MecKean-Vlasov forward-backward-SDE. The paper proposes a way to use that for training a neural network in order to fit a drift.

Critical strengths

It is quite interesting to see a new formulation of diffusion as a control problem.

Critical weaknesses

It is not clear whether the algorithm 1 will converge or not. My understanding is that the paper just shows that a stationary point (or approximate due to discretization) exists due to Theorem 1, but it is not clear whether such stationary point can be obtained. Also, algorithm 1 requires the compression of the past history with a fixed and exogenously given function. This seems to limit the expressivity of the numerical conditional expectation computation. I also fail to see the novelty of the algorithm given that it essentially is the same as attempting to solve the system of PDE which is known to the control community and fit the model with neural network. It could be that I am wrong on this and that the novelty of applying the known technique from control is sufficiently good.

General diffusion paper normally addresses what the algorithm would be if there is unlimited amount of data and expressivity of neural network. In this paper, it is unclear what the limit will be. Currently, the use of soft path-law constraint seems like an excuse so that we can use the forward-backward SDE instead of the paper claim about that soft constraint seems to genuinely good for generalization and prevent overfitting.

Questions for authors

Can the author explain more about the experiment in 4.1? In the current framework, the energy part of the loss cannot make the coefficient to be location or time specific, so how do we add the partitions into the calculation?

How do we normally get an observation of a continuous time process? Is this done by discretization? If so, would it be better to use ushc discretized process and the joint path instead of the integral of marginal loss?

5 comments
AuthorsOpenReview ↗

Rebuttal Answer

We thank the reviewer for his comment and address your comments below.

General comment. The optimal control is not obtained from a PDE system but from the probabilistic FBSDE representation: αt⋆\alpha_t^{\star} is the pointwise minimiser of the Hamiltonian a↦H(t,X,PX,Xt,PXt,Yt,Zt,a)a \mapsto H(t,\mathbf{X},\mathbb{P}_{\mathbf{X}},X_t,\mathbb{P}_{X_t},Y_t,Z_t,a), i.e. αt⋆=a^(t,X,PX,Xt,PXt,Yt,Zt)\alpha_t^{\star} = \hat{a}(t,\mathbf{X},\mathbb{P}_{\mathbf{X}},X_t,\mathbb{P}_{X_t},Y_t,Z_t). In our setting αt⋆=−Yt\alpha^{\star}_t = -Y_t, but the FBSDE must still be solved numerically, which is one of our contributions.

Convergence. A general convergence proof is beyond the scope of the paper, but convergence holds for benchmark problems (see our answer to Q3 of Reviewer Jv5k). Dimensionality reduction for conditional expectations is used only in the high-dimensional AMASS experiment, not in the detour one. Our method approximates the FBSDE directly, with neural networks for the backward process; this differs from direct optimisation of the control ("Direct SDE" in the paper).

Motivation for soft constraints. The penalty λ>0\lambda>0 trades off control energy against fidelity to the target, and the relaxation leads to a mean-field control problem whose analysis requires extending the stochastic maximum principle to our path-dependent setting with discrete observation times, a main mathematical contribution. The hard constraint is recovered as λ→∞\lambda\to\infty. With Jλ(α):=E[12∫0T∣αt∣2dt]+λD(P(Xs1α,…,XsKα)∣ρ)J_{\lambda}(\alpha) := \mathbb{E}[\frac{1}{2}\int_0^T|\alpha_t|^2\mathrm{d}t] + \lambda D(\mathbb{P}_{(X_{s_1}^{\alpha},\ldots,X_{s_K}^{\alpha})} \mid \rho), a minimiser αλ\alpha^{\lambda} and any feasible αˉ\bar{\alpha} (with D=0D=0) give

λD(P(Xs1αλ,…,XsKαλ)∣ρ)≤Jλ(αλ)≤Jλ(αˉ)=:C<∞,\displaystyle \lambda D(\mathbb{P}_{(X_{s_1}^{\alpha^{\lambda}},\ldots,X_{s_K}^{\alpha^{\lambda}})} \mid \rho) \leq J_{\lambda}(\alpha^{\lambda}) \leq J_{\lambda}(\bar{\alpha}) =: C < \infty,

so D≤C/λ→0D \leq C/\lambda \to 0. The FBSDE formulation is also interpretable: the control is characterised by the backward process, directly linked to the gradients of the distributional discrepancies. (Q2 of Jv5k)

New experiment: soft vs hard constraints with little data. We compared our marginal-only solver with a neural multi-marginal Schrödinger bridge, using 24 and 96 training observations per date, independent test observations, and the same shared network and diffusion coefficient. Schrödinger bridges impose the marginals as equality constraints (Léonard, 2014; multi-marginal: Di Marino & Gerolin, 2020); our comparator adapts the empirical construction of Pooladian & Niles-Weed (2025) to multiple dates. Our penalty weights are fixed across dataset sizes.

Obs./dateMethodTrain MMD2\mathrm{MMD}^2Test MMD2\mathrm{MMD}^2Change
24Ours (soft)0.034980.03422−2.2%
24Schrödinger bridge0.010430.01994+91.2%
96Ours (soft)0.035270.03414−3.2%
96Schrödinger bridge0.010500.01300+23.9%

Change = relative train-to-test change of MMD2\mathrm{MMD}^2 (one seed). The bridge fits more tightly (lower absolute test MMD2\mathrm{MMD}^2), but its discrepancy nearly doubles on unseen data with 24 observations (+24% with 96), whereas ours is stable at both sizes. This supports the regularising role of finite penalties: our fit transfers to unseen data with essentially no deterioration, even with little data.

Q1. We solve a single control problem over the whole horizon, so no partition is needed. With observation grid Πobs={s1,…,sK}⊆Πsim={t0,…,tN}\Pi_{\mathrm{obs}} = \lbrace s_1,\ldots,s_K \rbrace \subseteq \Pi_{\mathrm{sim}} = \lbrace t_0,\ldots,t_N \rbrace, the control αt=Yt\alpha_t = Y_t is càdlàg: at sℓ=tms_\ell = t_m, αtm=αtm−+ΔYsℓ\alpha_{t_m} = \alpha_{t_m^-} + \Delta Y_{s_\ell}, where the jump is induced by the discrepancy gradient (Eq. (2.8)). These jumps are thus already in the scheme, and ∫0T∣αt∣2dt≈∑m=0N−1∣αtm∣2Δtm\int_0^T |\alpha_t|^2\mathrm{d}t \approx \sum_{m=0}^{N-1} |\alpha_{t_m}|^2\Delta t_m on the simulation grid.

Q2. Observation times are fixed in advance on Πobs⊆Πsim\Pi_{\mathrm{obs}} \subseteq \Pi_{\mathrm{sim}}, while the process evolves continuously; continuous time is more tractable for the mean-field FBSDE, though a discrete-time version is possible. Our constraint involves the joint law P(Xs1,…,XsK)\mathbb{P}_{(X_{s_1},\ldots,X_{s_K})}, not the full path law PX⋅∧t\mathbb{P}_{X_{\cdot\wedge t}}, so the marginal flow (PXt)t≥0(\mathbb{P}_{X_t})_{t\geq 0} carries extra information between observations. A full-path formulation is possible but numerically harder; ours stays realistic since data are observed at discrete times. Since Πobs\Pi_{\mathrm{obs}} can be strictly smaller than Πsim\Pi_{\mathrm{sim}}, integrals are approximated on the whole simulation grid.

Thank you for your rebuttal. I may mistakenly think that the FBSDE can be changed into PDEs, but it seems like it cannot, and the solver in your paper just do numerical iteration (where in each iteration we more or less do numerical mean for Yt), right?

I am not so sure how solving FBSDE in a numerical and a very starightforward manner can be counted as a good contribution. Are you thinking more like some empirical proof that the method can work in this area?

Thank you for your remark. In fact, there is a dual representation between FBSDE and PDEs in our path-dependent formulation. Indeed, solving the optimal control problem would lead to a system of coupled PDEs where each PDE is written over each the subintervals [si,si+1)[s_i,s_{i+1}) and are linked through the jump terms in our paper. In this paper, we wanted to focus more on the FBSDE representation which we believe has a straightforward interpretation and relies on more probabilistic tools. Concerning the numerical solver, we are indeed performing an iterative approach. However, we want to point out that our algorithm is first sample-based in the sense that the law-functionals (law-gradient terms) need to be estimated from observed samples and trajectories generated by the current control and recomputed as the generated laws change. Moreover, it needs to tackle the jump terms, which is also a novelty of our path-dependent approach. As we mention, the solver has to handle this jumps by computing the conditional expectation with respect to the available trajectory history.

By the way, what do you mean by numerical mean of YtY_t ?

Reviewer_ctQ3OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose a method to learn stochastic dynamics, taking into account both the marginals and the paths, in an interesting way. Results are shown on a toy problem, on human motion and on optimal transport in the image domain.

Critical strengths
  • In my opinion, this research topic is valuable to the field, and somewhat under-explored.
  • The experiments provide a clear view of how this method can be useful, and take the time to carefully explain this, also with qualitative results.
Critical weaknesses
  • The experimental validation is somewhat lacking, especially for comparing to related works or methods.
Questions for authors
  • The related work section is very brief, for example, are you aware of these related works, in general working on learning SDE models?

    • Grigory Bartosh, Dmitry Vetrov, and Christian A Naesseth. SDE Matching: Scalable and simulation-free training of latent stochastic differential equations. In International Conference on Machine Learning. PMLR, 2025.
    • Kevin Course and Prasanth B. Nair. Amortized reparametrization: efficient and scalable variational inference for latent SDEs. In Advances in Neural Information Processing Systems, volume 36, 2023a
  • Can you compare your method experimentally with other (recent) works?

2 comments
AuthorsOpenReview ↗

Rebuttal Answer

Thank you for your positive comments on our paper. We address your comments below and all the responses will be added to the revised paper.

Q.1 :\textbf{Q.1 :} We were not aware of these works, as they do not directly consider stochastic control problems, which is the starting point of our approach. After carefully reviewing these papers, we believe that they propose different approaches and address different optimization problems, although they share the common goal of sequential data generation. We will add these references and briefly discuss their differences with our approach in the revised version.

Q.2 :\textbf{Q.2 :} We provide a comparison on the detour experiment. SDE Matching (Bartosh et al., 2025) achieves a route accuracy of 0.33 and a path MMD of 2.123. While Deep MKV achieves better generation quality (Table 1), SDE Matching is significantly more computationally efficient, highlighting a trade-off between generation quality and computational cost.

Q1. Your answer is contradictory. "although they share the common goal of sequential data generation". Your work does not exist in a vacuum. Furthermore, these other works can definitely be seen as stochastic control problems, there is a very strong relationship between optimal stochastic control and variational inference of stochastic dynamics, see e.g. https://arxiv.org/abs/2505.17150 or older papers by Kappen. And even if not, it is the application that defines the scope, not the specific method.

I retain my score

Reviewer_Jv5kOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper aims to incorporate soft constraints on exact marginals and some observed path data to solve for trajectories via a stochastic control approach with quadratic control cost. The result is a FBSDE of McKean-Vlasov type, with the majority of the theoretical basis borrowed from Hafsi et al. 2026. A neural parameterization of the optimal control is used and the resulting optimality system is solved by forward simulation, estimation of adjoint targets, and fitting of the control to them. Finally, three experiments show that the method can successfully infer appropriate paths in a synthetic setting, in a human motion dataset, and lastly on a high-dimensional image-transport experiment.

Critical strengths
  1. The paper introduces a novel approach to incorporating both marginals and sequence/path observations, via an optimal control approach.
  2. The experiments demonstrate that the approach works in principle and is able to leverage such information to come up with more faithful trajectories.
Critical weaknesses
  1. The main theoretical basis of the work seems to be contained in Hafsi et al. 2026, so perhaps the main contribution lies in the realization of the Deep MKV solver of Section 3.
  2. The experiments are simply a proof-of-concept and are perhaps not the most compelling examples. The computational costs appear relatively high, and comparison is only done to a direct neural SDE method that is not fully described (is it Issa et al. 2023 or Kidger et al.?)
  3. Robustness to parameter variation and convergence are not explored or considered at all.
  4. The writing is rather opaque and symbol-heavy. It would be useful to have a notation table in the appendix and to try to describe the high-level intuition of the control setting. This would improve the reach of the article and elucidate the advantages and disadvantages of the approach, in my opinion.
Questions for authors

Merely to address the weaknesses above.

1 comment
AuthorsOpenReview ↗

Rebuttal answer

Thank you for your positive comments on our paper. We answer your comments below.

Q1:\textbf{Q1}:

Our work indeed builds indeed upon the recent theoretical results of Hafsi et al. (2026). However, we would like to emphasize that our main contribution is to propose a new generative approach for sequential data based on McKean–Vlasov control with discrete path dependence.

The idea is to start from a reference stochastic model (e.g., a standard diffusion model) and regularize its dynamics through a soft distributional constraint, where the parameter λ\lambda allows us to balance regularization and fitting the target data distribution.

We then design a scalable numerical algorithm, Deep MKV, to solve this problem. Therefore, our contribution lies both in the formulation of a new control-based approach for sequential data generation and in its numerical implementation.

Q2:\textbf{Q2}:

While our experiments are mainly proof-of-concept, we believe they provide interesting and original illustrations of our approach. In particular, the AMASS experiment required the construction of a dedicated dataset from real human motion sequences, allowing us to illustrate the importance of path dependence in capturing temporal dynamics beyond marginal distributions.

By Direct Neural SDE, we refer to the direct minimization of JJ, where the control is parametrized by a neural network. The mentioned works use similar parametrizations but consider different objectives.

We also emphasize the interpretability of Deep MKV, where the control α\alpha is computed from the adjoint process YY, which is related to the gradients of the discrepancies. In particular, for KL-based discrepancies, the terminal condition and the drift of YY involve score functions, providing a connection with score-based diffusion models.

Regarding computational costs, we thank the reviewer for this remark. The reported training times are approximately 6–8 minutes for the synthetic experiments, 1 hour 54 minutes for AFHQ, and 3 hours for AMASS, using a single NVIDIA A100 GPU.

We would like to emphasize that these are offline training costs. Once the controller is trained, generating new trajectories only requires forward simulations, without recomputing the adjoint processes or retraining the estimators.

We will clarify these computational costs and the baseline configuration in the revised version.

Q.3:\textbf{Q.3}:

To investigate numerical convergence, we conducted an additional experiment on an exactly solvable path-dependent mean-field linear–quadratic problem, allowing us to evaluate the MP residual, control error, and objective gap.

Starting from zero initialization, under a 90-second budget, the MP residual decreases from 0.4540.454 to 0.01920.0192 for λ=0.2\lambda=0.2, and from 11.36211.362 to 0.25180.2518 for λ=5\lambda=5, with corresponding objective gaps of 1.79×10−41.79\times10^{-4} and 1.67×10−21.67\times10^{-2}.

These results provide numerical evidence of convergence toward the MP solution under different coupling strengths. We will include the convergence curves and evaluation protocol in the revised paper.

Q.4:\textbf{Q.4}:

In the revised version, we will add a notation table in the appendix and provide more intuition on the underlying control problem. We will also discuss more clearly the advantages and limitations of our approach to make the paper easier to follow.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.