acceptodds
Withdrawn from ICLR 2027

Shrinkage-Weighted Bellman Regression for Neural PDE Solvers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Random-walk PDE solvers fit neural networks to sampled Bellman targets whose noise varies across the domain. Weighting by return spread can control noisy updates, but the spread estimates are themselves uncertain and downweighting can weaken learning near steep solution features. We propose bounded shrinkage weighting for Bellman regression. The method accounts for sampling uncertainty in the estimated spread and shrinks weights toward uniformity when spatial differences are poorly resolved. Clipping and a uniform weight floor limit concentration and preserve a minimum contribution from every point. Our analysis distinguishes minimax noise-amplitude weighting from minimum-variance weighting, bounds the bias caused by dependence between weights and targets, and characterizes how reweighting changes a local least-squares fit. In a baseline study covering seven Burgers settings with 40 paired trials per setting, plain inverse-spread weighting lowers mean sampled hard-region error by approximately 42–71% relative to uniform regression, while individual runs can be less accurate. These results identify the practical challenge addressed by the safeguards: controlling noisy updates without losing useful corrections near sharp layers. The bounded-shrinkage comparisons and component ablations remain pending.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural PDE solvers, Bellman regression, derivative-free loss methods, heteroscedastic noise, shrinkage weighting, stochastic PDE methods

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_hfwvOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a bounded-shrinkage weighting method for stochastic Bellman regression in the DFLM framework for neural PDE solvers, accounting for location-dependent target spread and uncertainty in its estimation. It combines a gate that adjusts weighting strength based on this uncertainty, clipping that limits extreme weights, and a uniform floor that maintains a minimum contribution from each point by mixing with uniform weights, aiming to suppress target noise while preserving the learning signal in regions where the solution varies rapidly.

Critical strengths

S1. Section 3 analyzes the relationship between return variability and the spatial gradient of the current approximate solution, explaining how noise-suppressing weights can also weaken the learning signal needed in regions where the solution changes rapidly. By distinguishing noise magnitude, bias arising from weight–target dependence, and changes in the regression solution induced by reweighting, it clarifies why noise reduction does not necessarily improve solution accuracy and motivates the design of the proposed method.

Critical weaknesses

W1. [Writing and Organization] The presentation would benefit from a clearer explanation of how the regression objective in Eq. (4) follows from the stochastic formulation used in prior DFLM work [1, 2]. Although Eqs. (6) and (7) are briefly explained, their connection to Eq. (4) and their role in motivating bounded-shrinkage weighting should be developed more systematically. Furthermore, terms such as PPO clipping and group normalization (line 95) are introduced without sufficient explanation of their roles in the estimator comparisons. The authors should also clarify which aspects of the theoretical analysis build on [2] and what additional insights or guarantees the current results provide for the proposed weighting scheme.

W2. [Theory] The analysis in Section 3 quantifies noise and local fitting tradeoffs and motivates the design of bounded-shrinkage weighting. However, it remains unclear how these results translate into improved PDE approximation accuracy relative to uniformly weighted regression. The authors should clarify the extent to which the current theoretical results support such an improvement and under what conditions.

W3. [Experiments] The performance comparisons and component ablations for the proposed bounded-shrinkage method remain incomplete, leaving the central contribution insufficiently validated. The completed results are limited to comparisons between inverse-spread weighting and uniformly weighted regression, and are therefore insufficient to establish improvements in accuracy, stability, or computational efficiency for the proposed method.

[1] Han, J., Nica, M., & Stinchcombe, A. R. (2020). A derivative-free method for solving elliptic partial differential equations with deep neural networks. Journal of Computational Physics, 419, 109672. https://doi.org/10.1016/j.jcp.2020.109672

[2] Han, J., & Lee, Y. (2025). An analysis of the derivative-free loss method for solving PDEs. arXiv preprint arXiv:2309.16829v2.

Questions for authors

[Q] I understand that the Bellman targets are constructed using the Feynman–Kac formulation underlying [1, 2]. Could the authors clarify the relationship between the underlying SDE and the proposed weighting scheme? Beyond using the empirical spread of the sampled returns, does the specific bounded-shrinkage rule admit a theoretical justification or interpretation in terms of properties of the underlying stochastic process?

[1] Han, J., Nica, M., & Stinchcombe, A. R. (2020). A derivative-free method for solving elliptic partial differential equations with deep neural networks. Journal of Computational Physics, 419, 109672. https://doi.org/10.1016/j.jcp.2020.109672

[2] Han, J., & Lee, Y. (2025). An analysis of the derivative-free loss method for solving PDEs. arXiv preprint arXiv:2309.16829v2.

Reviewer_WjAWOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies the weighting of the regression loss in the Derivative-Free Loss Method for solving PDEs by training neural networks. The Monte Carlo estimation of Bellman targets introduces noise into the regression gradients, with the noise level varying across the domain. A way to control this is to weight the loss using the estimated return spread, for example through inverse-spread or inverse-variance weighting.

The authors identify two potential problems with this approach. First, regions with high return spread often coincide with sharp solution features, so downweighting them suppresses not only gradient noise but also useful learning signal. Second, the spread itself is estimated from finitely many walkers, meaning that apparent spatial differences in spread may partly reflect sampling uncertainty.

To address these issues, the authors propose a heuristic weighting scheme comprising three components:

  1. An adaptive shrinkage gate that interpolates between uniform and inverse-spread weighting, depending on how reliably spatial differences in spread can be estimated.
  2. Clipping of relative weights to limit their concentration.
  3. A uniform weight floor to prevent excessively suppressing the learning signal at high-spread locations.

The design is motivated by theoretical results concerning gradient variance, bias arising from dependence between weights and targets, and the effect of weighting on finite-capacity regression.

Critical strengths
  1. The paper studies a real problem with practical implications.
  2. The proposed method, while being heuristic, is well motivated in the theoretical results under certain assumptions.
Critical weaknesses
  1. The authors describe the empirical evaluation of the method as “pending”. This is a substantial omission, as empirical validation is essential to assessing the proposed method. The evaluation should have been completed before submission, rather than leaving this crucial part of the work to be assessed at a later stage.
Questions for authors

None

Reviewer_Q55POpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes "bounded shrinkage weighting for Bellman regression."

Critical strengths

None.

Critical weaknesses

This paper 1) is incredibly poorly written and 2) does not contain empirical results of the proposed method.

The Abstract states "the bounded-shrinkage comparisons and component ablations remain pending"; the Introduction does not set the stage well--it took me a while to figure out the context and how the proposed methodology is different from existing techniques; Section 6 has text that was meant to be commented out, unless the authors wanted readers to know that results are "outstanding" (please confirm); and there are no empirical results.

Questions for authors

None.

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

Related papers

Discussion (0)

Sign in to comment.