acceptodds
Desk rejected at ICLR 2027

Posterior Means Underestimate Diffusion: Auditable Identification of Position-Only Stochastic Dynamics

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Learning stochastic dynamics from noisy positions requires estimating forces and diffusion without observing velocity. We show that the standard smooth-then-regress approach systematically underestimates diffusion in our benchmarks because posterior-mean reconstruction discards uncertainty in velocity increments. We introduce OMWG-C, an endpoint-corrected weak-moment estimator that restores this missing variance from smoother cross-time covariances. Drift and diffusion are unknown functions represented in supplied candidate dictionaries; the second-order kinematic structure is assumed. In a frozen synthetic confirmation, OMWG-C lowers diffusion error in every unit, with a substantial median reduction relative to the posterior-mean baseline, while development results show little change in one-step prediction. For data without ground truth, we introduce a pre-set cross-sampling-rate audit that refits the dynamics after downsampling. Endpoint-corrected fits fail this audit in all cell-migration cohorts, despite better prediction in most cohorts, whereas ocean-drifter point estimates pass despite worse prediction. Passing tests rate stability, not physical correctness. These results establish predictive fit, parameter recovery, and sampling-rate stability as distinct dimensions of position-only equation learning.

Primary area:
learning on time series and dynamical systems
Keywords:
stochastic differential equations, system identification, partially observed dynamical systems, diffusion estimation, time series analysis

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_rVmZOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies the problem of learning stochastic dynamics from noisy position-only trajectories, where velocities are unobserved. The authors argue that conventional two-stage approaches, which first reconstruct latent velocities and then estimate drift and diffusion, systematically underestimate diffusion because posterior-mean reconstruction discards uncertainty.

To address this issue, they propose Observation-Marginalized Weak Generator with endpoint correction (OMWG-C), which incorporates posterior velocity-increment variance obtained from a Gaussian smoother into weak-moment diffusion estimation. The method is evaluated on synthetic stochastic systems, where it improves diffusion recovery relative to a posterior-mean baseline, although an established Stochastic Force Inference (SFI) method generally achieves better joint parameter recovery. The authors also introduce a cross-sampling-rate audit to evaluate the stability of estimated diffusion coefficients when ground truth is unavailable. Experiments on cell-migration and ocean-drifter trajectories demonstrate that predictive performance and sampling-rate stability can differ substantially.

Critical strengths

None

Critical weaknesses

1. Insufficient scientific clarity and motivation. The manuscript is difficult to follow, particularly in the first three sections. The introduction lacks sufficient context and references to relevant literature, introduces mathematical quantities without adequate explanation, and reports numerical results without clearly explaining their significance. The theoretical development in Section 3 also lacks sufficient motivation, making it difficult to understand why the proposed equations are necessary and what they contribute. These issues substantially undermine the accessibility and scientific clarity of the paper.

2. Limited methodological novelty and insufficient evidence of advantages over established methods. The central theoretical contribution relies largely on standard conditional variance identities, and the paper does not sufficiently explain how the proposed correction advances existing approaches to stochastic system identification. Although OMWG-C improves upon several baselines, the established SFI parametric method achieves lower joint recovery error in 51 out of 72 matched synthetic experiments. Furthermore, the ablation study shows that a simpler single-lag correction already accounts for much of the improvement achieved by the full method. These results raise concerns about whether the additional methodological complexity is justified and whether the paper offers a sufficiently substantial contribution for ICLR.

Overall assessment. In my view, the manuscript requires major revision, particularly in its scientific presentation, motivation, and positioning relative to existing methods. The current version does not communicate a sufficiently clear and compelling methodological contribution, while the experimental evidence does not convincingly establish its advantages over strong existing approaches. Given these substantial concerns, I recommend rejection in its current form.

Questions for authors

See critical weaknesses.

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

This paper proposes OMWG-C, a confirmed endpoint-corrected variant of the Observation-Marginalized Weak Generator. It explores applicability of the proposed method on synthetic and real-world experiments. The main reason for low score is clarity and difficulty to follow the work as well as understand the gap it closes in the current literature.

Critical strengths
  • The paper demonstrates proposed method lowers the error on 72 units considered in synthetic experiments
  • The paper shows it provides lower CPU time in comparison to other approaches
Critical weaknesses
  • Understanding the problem scope: I think introduction would benefit from better framing the problem and what gap OMWG-C fills that other methods do not. It is currently unclear I.e what properties a desired estimator should satisfy and how the identified strategy positions in comparison to those. It is unclear how significant lowering diffusion error in 72 confirmation units is, and what TwoStage is (it sounds like this is a competitive approach and should therefore be cited?). The second paragraph also feels like it belongs more in the Experiments section as a deep-dive into experimental validation
  • Clarity and presentation: The paper is missing a background section which I believe would be helpful to define terminology used in the work and also explaining principle behind Observation-Marginalized Weak generator, in section 2 notation is not well defined (it is unclear what β\beta or γ\gamma are among other variables for example). I have similar comments for section 3 (it is unclear what D is, or B or R since they are not explicitly defined). I would strongly encourage to add small descriptors before writing identities and propositions as it is quite difficult to follow at the moment. In section 4 baselines used should be cited, ideally results should be split into a diagram per experiment, I.e one for cells, one for drifters.
Questions for authors
  • In the introduction you mention latent velocity however I don’t think this latent space is defined. Could you clarify?
  • I understand posterior should represent states that are not observed Y_i|Y_k, could you explicitly define it though since it seems to be the central object of consideration in the paper? I think adding it into the introduction after defining Y_k would be helpful
  • What are the 72 development units? How is this data constructed?
  • Could you explain what pre-set re-identification test means in contribution 3?
  • Could you explain what neural controls you are referring to in contribution 3? Is latent neural SDE part of your method? I think this should be highlighted earlier on if so. What is meant by “passing the audit”?
  • In figure 4, could you explain the pre-set band? Is it the same for both datasets? On cell datasets I assume three groups come from sd01, sd02 and sd03, where do two groups come from on the drifters data?
Reviewer_HiV8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

his paper investigates how to learn stochastic dynamics from noisy position observations. The authors propose a simple extension to Kalman smoothing by adding a posterior variance correction when estimating diffusion, accounting for uncertainty lost when using only the smoothed mean trajectory.

Critical strengths

Figure 1 illustrates the problem of underestimating diffusion, and the ablation experiments show that adding the variance correction improves diffusion estimation.

Critical weaknesses

The writing needs improvement. For example, in the introduction, what does “72 confirmation units” mean? What does “in the development comparison” mean? Where do Equations 3 and 13 come from? Detailed explanations are missing. The writing makes it difficult to judge the correctness of the formulation.

Given the paper’s confusing presentation, the proposed method’s novelty seems limited. Retaining posterior second moments is already standard in EM, and the proposed endpoint correction applies this principle. I am not convinced that this constitutes a substantial methodological advance.

Empirically, the proposed method is not a strong alternative to baselines such as Kalman maximum likelihood.

Questions for authors

Posterior second moments are already used in EM. Could the authors explain precisely what is new in the proposed endpoint correction?

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

Related papers

Discussion (0)

Sign in to comment.