acceptodds
Under review as a conference paper at ICLR 2027

An approximate operator-based learning method for the numerical solution of stochastic differential equations

Abstract

Stochastic differential equation (SDE in short) solvers find numerous ap- plications across various fields. However, in practical simulations, we usu- ally resort to using Itˆo-Taylor series-based methods like the Euler-Maruyama method. These methods often suffer from the limitation of fixed time scales and recalculations for different Brownian motions, which lead to computa- tional inefficiency, especially in generative and sampling models. To address these issues, we propose a novel approach: learning a mapping between the solution of SDE and corresponding Brownian motion. This mapping exhibits versatility across different scales and requires minimal paths for training. Specifically, we employ the DeepONet method to learn a nonlinear mapping. And we also assess the efficiency of this method through simulations con- ducted at varying time scales. Additionally, we evaluate its generalization performance to verify its good versatility in different scenarios

Primary area:
learning on time series and dynamical systems
Keywords:
stochastic differential equations, operator-based learning, DeepONet

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_JGPnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a recurrent DeepONet to learn the mapping from Brownian noise paths to numerical SDE solutions. A GRU/Transformer causally encodes the noise history, a branch network fuses the initial condition, a trunk network processes time queries, and their inner product yields the predicted state. The method is extended to McKean–Vlasov particle systems, coupled vector dynamics, and SPDE field prediction, and uses a Doob h-transform to construct a generative sampling SDE for single-pass sample generation.

Critical strengths
  1. This paper formulates a recurrent DeepONet for learning numerical SDE solution paths from Brownian histories and extend it to direct SPDE field prediction.

  2. A six-architecture benchmark and repeated comparisons isolate the effects of noise representation, training data, and optimization budget.

  3. This paper measures the cost of generating labels, training, and repeated prediction. It also compresses the trained output network and bound the resulting change in predictions on the training grid.

Critical weaknesses
  1. Section 4.2's "learn SDE from data and generate samples" is essentially a generative modeling task, yet the paper makes no experimental comparison with diffusion models, consistency models, or flow matching. Related work cites these directions, but the experimental baselines are only MLP, CNN, FNO, and randomized signatures, which are operator-learning-domain baselines rather than generative-model baselines familiar to the deep learning community. What is the advantage of this approach over directly training a conditional diffusion model?
  2. The paper achieves good results on low-dimensional scalar tasks with only 20 training paths. How does the training data requirement scale to higher-dimensional, more complex systems (e.g., 100+ dimensions)? Where is the scalability limit of the method?
  3. It would be helpful to include an overlay plot of predicted vs. true trajectories in the main text, to help readers outside the SDE field intuitively assess prediction quality.
Questions for authors

1.What is the essential difference between your method and diffusion models? Diffusion models also learn a noise-to-data SDE, and your recurrent DeepONet also learns a noise-to-solution mapping. Why not use a U-Net score network directly? What advantage does the branch-trunk DeepONet design have over standard diffusion architectures?

2.The SPDE field prediction is essentially a spatiotemporal prediction problem. How does your method compare to video prediction approaches? What advantages does your causal encoder + temporal trunk conditioning have over spatiotemporal attention mechanisms?

3.The model makes causal step-by-step predictions. Does the accumulated error diverge as the number of prediction steps grows? How stable is long-horizon prediction?

4.Currently, each SDE requires training a separate model. Can the model generalize across SDE parameters? Can it adapt to new SDE parameters with a small amount of fine-tuning, similar to transfer learning?

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

The paper studies supervised learning of numerical SDE solution maps from initial conditions and Brownian observations. A causal recurrent encoder summarizes the observed noise history, and a DeepONet branch–trunk output predicts the state at each query time. The approach is extended to SPDE fields and a data-defined sampling SDE. Theoretical results characterize conditional approximation error on a fixed grid and bound prediction changes after output-network compression. Experiments include scalar and vector SDEs, interacting particle systems, SPDEs, and controlled comparisons of noise representations, encoders, and output architectures.

Critical strengths
  1. The architectural comparisons provide credible evidence that the branch–trunk output is useful. The independent study in Table 13 matches parameter counts, data, and encoder initialization, includes a FiLM control, and uses validation-based model selection. Branch–trunk achieves lower mean error in five of six conditions against MLP and all six against FiLM. Its approximately 33% improvement over MLP on the 8D increment-input task is meaningful. The reported reversal on periodic 2D also appropriately limits claims of universal superiority.

  2. The paper carefully distinguishes several sources of approximation error. The conditional-mean interpretation correctly handles information unavailable to an individual-particle predictor, and Appendix I includes independent-system testing. The shared-noise refinements and exact-time heat reference distinguish fitting numerical labels from approximating the underlying solution. The fitted-network compression bound is also a useful, explicitly scoped guarantee. These checks make the empirical conclusions more credible.

Critical weaknesses
  1. The advance over existing stochastic operator learning remains insufficiently established. Learning a Brownian-to-solution map is already central to SDEONet, while Expanding the Chaos also reconstructs full stochastic trajectories from noise. The paper acknowledges these predecessors; its distinguishing contribution is therefore the learned causal recurrent representation and its branch–trunk output. The matched-head experiments support this architectural choice, but do not establish an advantage of the complete method over a directly relevant stochastic operator learner. Theorem 1 provides limited additional differentiation: it establishes finite-grid approximation of conditional expectations with a fixed encoder and unrestricted capacity, rather than a benefit specific to the proposed representation. I therefore find a useful architectural study, but insufficient evidence of a substantial methodological advance.

  2. The practical solver benefits motivating the paper are not demonstrated. In Table 14, network inference is slower than Euler–Maruyama for both the linear and double-well tasks: 0.0195 versus 0.0135 ms/path and 0.0139 versus 0.0080 ms/path, respectively. The favorable mean-field timing requires approximately one million paths to amortize the reported offline cost, and the paper explicitly notes that this is not a matched-accuracy comparison. Likewise, §4.1 states that the available experiment does not establish step-size transfer, despite the abstract claiming versatility across scales. These limitations do not invalidate fixed-grid regression, but leave the central practical justification unresolved. A convincing accuracy–cost advantage or controlled resolution-transfer result would matter more than additional demonstrations that the network can fit numerical trajectories.

Questions for authors
  1. Comparative evidence: Can you provide a targeted comparison with one directly relevant stochastic operator method on a shared task, controlling the observed noise, numerical labels, training/validation label budgets, and evaluation reference? For example, an established noise-conditioned SPDE operator would be relevant to Appendix L. The important question is whether the complete recurrent approach improves the accuracy–data–compute trade-off beyond the demonstrated output-head benefit.

  2. Computational evidence: Is there a demonstrated regime in which the method improves accuracy at comparable cost, or reduces total cost at matched accuracy? A comparison on a nonlinear vector SDE or SPDE should include label generation, training, recurrent inference, and the number of predictions needed for amortization. For mean-field dynamics, the timed reference and accuracy target should be consistent. This would directly affect my assessment of practical significance.

  3. Resolution claim and supporting experiment: Does the claimed versatility across scales mean transfer of a fixed trained model to a new time grid? If so, can you provide a controlled experiment specifying the physical horizon, noise coupling, normalization, and resampling procedure? If that capability is not established, the abstract should be narrowed, and the contribution assessed as fixed-grid surrogate learning.

Reviewer_mm1rOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies supervised approximation of the causal map from an initial state and a Brownian-motion history to the numerical solution of a fixed stochastic differential equation. The proposed model uses a recurrent encoder for the observed noise prefix and a DeepONet-style branch-trunk output network queried at time, with an analogous construction for stochastic PDE fields. The theoretical results characterize the best achievable squared error from a fixed finite-dimensional causal representation on a finite query grid, give a validation-selection bound for a finite set of trained candidates, and derive a perturbation bound for post-training compression of the output network. The experiments cover scalar SDEs, finite-particle McKean-Vlasov systems, coupled 2D and 8D systems, stochastic heat and nonlinear reaction equations, and a synthetic Doob-transform sampling task. The reported results show that the branch-trunk output often improves over matched MLP or FiLM outputs and can predict held-out paths accurately for several fixed equations and grids.

Critical strengths
  1. Careful causal formulation and evaluation discipline. The model respects the filtration by using only the Brownian prefix available at each query time and imposes the initial condition explicitly. The paper also distinguishes prediction error against numerical labels from discretization error, uses shared-noise refinements and exact-time checks where available, and discusses irreducible conditional uncertainty for individual particles in interacting systems. These details make the fixed-grid empirical results more credible than a comparison only against coarse numerical labels.

  2. Broad and transparent empirical investigation. The paper evaluates scalar, vector, interacting-particle, SPDE, and sampling examples; uses repeated experiments and independent test paths in the main controlled studies; and separates encoder, noise-representation, output-network, training-budget, and compression effects. The matched branch-trunk/MLP/FiLM comparisons provide useful evidence that the branch-trunk output can be beneficial under the studied training budgets.

Critical weaknesses
  1. The main generality claims exceed what the method and theory establish. Each equation is trained separately on a fixed observation grid, and the main approximation theorem fixes the encoder and considers only finitely many query nodes, with unrestricted width and latent dimension. It therefore largely reduces to L2L^2 approximation of conditional expectations at those nodes and gives no guarantee for learning the representation, changing the time resolution, or generalizing across equation parameters. Consistent with this limitation, the paper explicitly states that its cross-step-size experiment is incompletely specified and does not establish step-size transfer. This is in tension with the abstract's claims of versatility across scales and training with minimal paths.

  2. The practical and comparative case is not yet convincing. A "single forward evaluation" of the recurrent model still consumes the full Brownian path through KK sequential recurrent updates, so terminal sampling remains O(K)O(K) in sequence length even though it avoids explicit Euler updates. The measured model is slower than Euler-Maruyama on the linear and double-well tasks; the mean-field break-even point is approximately one million paths and is not evaluated at matched accuracy. Moreover, although the paper discusses SDEONet, MFNO, NeuralChaos, Neural SPDEs, and Wiener-chaos-based operators, the experiments compare mainly with generic MLP/CNN/FNO/signature baselines and internal architectural variants. Without matched comparisons to the closest specialized methods, it is difficult to assess novelty, accuracy, sample efficiency, or computational benefit.

Questions for authors
  1. Does a single trained model generalize to unseen time resolutions, horizons, or equation parameters? If so, please provide a fully specified experiment, including the training resolution, test resolutions, horizon, Brownian coupling/resampling procedure, input normalization, and sensor handling. If not, would the authors revise the abstract and introduction to describe the method explicitly as a fixed-equation, fixed-grid surrogate?

  2. Can the authors provide matched comparisons with the closest specialized baselines cited in the paper, such as SDEONet, MFNO, NeuralChaos, and an appropriate Neural-SPDE or Wiener-chaos operator on overlapping tasks? Comparisons should control the number of independent training trajectories, parameter/compute budget, numerical-label accuracy, and model-selection protocol.

  3. What is the end-to-end inference complexity and wall-clock scaling as the number of time steps KK increases? In particular, please include full Brownian-path construction and recurrent sequence encoding, and compare with the same numerical target at matched accuracy. This would clarify whether "single forward evaluation" provides a practical advantage beyond packaging KK recurrent updates into one model call.

  4. What is the intended nontrivial implication of Theorem 1 for the implemented learner? Since the encoder is fixed and the trunk can interpolate a finite query set with latent dimension scaling with the number of nodes, the result does not appear to control representation learning, finite-capacity approximation, optimization, or resolution transfer. A clearer statement of what new guarantee is obtained over standard finite-dimensional universal approximation would materially affect my assessment of the theoretical contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.