acceptodds
Under review as a conference paper at ICLR 2027

Recovering Governing Dynamics from Fragmented Observations via Exact Distributed Spline Merging

Abstract

Scientific observations are often fragmented across space, time, and institutions. If such fragments can be combined into a continuous, differentiable field, its derivatives yield governing physical parameters through regression. This paper makes two contributions in this setting. First, the established additive structure of fixed-basis ridge-regression statistics is applied to tensor-product spline fields: each data holder computes a local Gram matrix and moment vector, and the merged solution is mathematically identical to centralized fitting, with no raw data shared and no iterative synchronization. This property is specific to the fixed-feature squared-error setting; the present derivation does not establish an analogous guarantee for general jointly trained multilayer networks. Second, a complete pipeline connects distributed observations to physical parameter inference through field reconstruction, derivative extraction, and linear regression. The pipeline is validated on four PDEs: diffusion, wave, heat-with-source, and the nonlinear viscous Burgers equation, recovering governing parameters to sub-percent accuracy in the linear cases and 5% for Burgers. In all cases, distributed merging introduces zero degradation relative to centralized fitting. Synthetic experiments validate parameter recovery; application to 41 years of NOAA sea-surface temperature data validates field reconstruction and aggregation equivalence on real spatiotemporal observations. Source code to reproduce all experiments is available at https://anonymous.4open.science/r/splinemerge-F887.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
distributed learning, model merging, spline fields, sufficient statistics, physical parameter recovery, ridge regression, B-splines, PDE inference

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_jgWvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a distributed framework for reconstructing continuous spatiotemporal fields from fragmented observations and recovering parameters of known PDEs. Using a fixed tensor-product B-spline basis, each data holder shares only local Gram matrices and moment vectors, whose aggregation exactly reproduces centralized ridge-regression fitting without iterative communication. The reconstructed field is then differentiated and used to estimate PDE parameters via least squares. Experiments on four synthetic PDEs and 41 years of NOAA sea-surface temperature data demonstrate near-exact agreement with centralized reconstruction and accurate parameter recovery.

Critical strengths
  1. The paper tackles a practically important problem of learning from fragmented scientific observations in distributed settings.

2.The overall framework is simple, interpretable, and mathematically transparent, with a clear separation between distributed field reconstruction and downstream PDE parameter estimation.

Critical weaknesses
  1. The core distributed merging method has limited methodological novelty, as it directly applies the classical additive sufficient-statistics property of fixed-feature ridge regression to tensor-product spline reconstruction.

  2. The experiments mainly compare the proposed Gram merge against centralized spline fitting, which is sufficient to verify the claimed exact equivalence between distributed and centralized ridge-regression solutions. However, this comparison does not establish whether the proposed approach is practically preferable to alternative distributed reconstruction methods.

Questions for authors

see weakness.

Reviewer_uw56OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper fits a spatiotemporal field with a fixed tensor-product B-spline basis, where each data holder shares only a local Gram matrix and moment vector. Summing these gives coefficients identical to centralized ridge regression, with no raw data exchanged and no iteration. The authors then differentiate the merged field by finite differences and use least squares to recover one physical parameter from a known PDE (diffusion, wave, heat-with-source, Burgers). They report sub-percent errors for the linear cases and 5% for Burgers, plus a field-reconstruction demo on 41 years of NOAA sea-surface temperature data.

Critical strengths
  1. The paper states clearly what the exactness guarantee covers: fixed features with a quadratic loss, not trained multilayer networks.

  2. The paper also honestly admits that field-approximation and derivative error remain, that privacy is not analyzed, and that dense summaries can exceed the raw data size.

  3. The derivation and path from basis to normal equations to sufficient statistics to additive merging looks sound. The remark on rank-deficient local Gram matrices and the extension to any shared quadratic penalty are correct and useful.

  4. The paper distinguishes aggregation error (zero), reconstruction error, and parameter-estimation error.

  5. The experiments seem transparent and reproducible. The paper uses float64, fixed seeds, a multi-seed noise study with signed-bias analysis, and a grid-convergence check. It reports unflattering results too (Burgers 5%, noise degradation, MNIST extrapolation failure) and provides anonymous code.

Critical weaknesses
  1. Limited novelty. The core result is the additivity of XᵀX and Xᵀy, which the paper calls classical and credits to Alsulaimawi (2026) for the one-shot federated formulation. The added pieces are the spline basis, a standard derivative-and-regression pipeline, and a NOAA application. Polar & Poluektov (2025) on merging KANs is arguably the closest prior work, yet it gets only a clause saying it addresses "different settings."

  2. Weak real-data and nonlinear evidence. NOAA is an already-interpolated product, no physical parameter is inferred from it, and the 0.64 °C RMSEs have no baseline (I suspect a climatology-plus-trend model would be competitive). The held-out splits (random points, every 6th month) likely leave leakage through autocorrelation. For Burgers, the only nonlinear case, the 5% error is not decomposed; the authors say representation and derivative errors "have not been isolated." The synthetic solutions are low-frequency periodic modes, which favor a smooth spline surrogate.

  3. No method baselines for physics recovery. PINNs and weak SINDy are cited and not run, and federated SciML comparison is deferred to future work. The only control is finite differences on exact fields, so the 0.09–5% errors are hard to interpret. The title's "recovering governing dynamics" also overclaims, since the method estimates one coefficient of an assumed PDE.

  4. The communication and privacy motivation is not supported by the experiments. With the dense Gram matrices used, summaries are far larger than the raw data in the synthetic settings (P=1,331 gives ~1.8M entries, or ~0.9M with symmetry, versus 10K observations) and about the same order as the data for NOAA (~2.2–4.4M versus 4.41M values). The paper acknowledges this and leaves sparsity to future work. It also says privacy is not analyzed, though the abstract highlights "no raw data is shared."

Questions for authors
  1. Beyond choosing a spline basis, what does this add to Alsulaimawi (2026) and to Polar & Poluektov (2025) on merging KANs trained on disjoint data? Is there a case where the spline structure gives something the generic ridge formulation does not?

  2. In your experiments the dense Gram matrices are as large as or larger than the raw data. Can you show a setting where the summaries are smaller (for example, by exploiting sparsity), or where they provide a privacy benefit?

  3. How do the recovered parameters compare with PINNs, weak SINDy, or a direct nonlinear least-squares fit of the PDE parameter?

  4. Can you decompose the 5% Burgers error into spline-fit, finite-difference, and regression contributions?

  5. For NOAA Can you add a baseline (such as climatology plus trend) and a held-out test that is not heavily autocorrelated, such as a withheld region or a later time period?

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

The paper studies physical parameter estimation from distributed observations of a shared spatiotemporal field. Each data holder computes sufficient statistics for a fixed tensor-product spline representation, which are aggregated to reproduce the centralized ridge-regression solution. The reconstructed field is then used to estimate parameters of a known PDE through derivative-based regression. Experiments cover four synthetic PDEs, MNIST, and NOAA sea-surface temperature data.

Critical strengths

None

Critical weaknesses

Limited methodological novelty. The approach combines established spline fitting, additive least-squares statistics, and parameter regression for known PDE structures. The paper acknowledges that the aggregation principle is classical, but does not clearly establish a substantive methodological contribution beyond this combination.

Insufficient comparative evaluation. The main comparison is against the mathematically equivalent centralized estimator. This verifies implementation consistency but provides limited evidence of an advantage over alternatives. The parameter-recovery and noise experiments provide useful feasibility evidence, yet comparisons with relevant distributed reconstruction or parameter-estimation methods are missing.

Unclear practical motivation. The paper does not sufficiently explain when transmitting sufficient statistics is preferable to collecting raw observations. Dense summaries exceed raw-data size in the synthetic experiments, and no formal privacy guarantee is provided. The practical constraints that justify this protocol need clarification.

Weak organization and exposition. Figure 1 is crowded and difficult to read. Important reconstruction visualizations appear late in the document, with small labels and limited quantitative guidance, making the experimental findings harder to assess. The manuscript’s organization and allocation of detail are not well aligned with the focused presentation expected of an ICLR paper.

Questions for authors
  1. What is technically new beyond combining established spline regression, sufficient-statistic aggregation, and PDE parameter estimation?

  2. When is this protocol preferable to sharing raw data or using existing distributed methods, and can matched comparisons demonstrate that advantage?

Reviewer_7kK6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper uses Gram matrices and moment vectors from fixed spline basis functions to let multiple data holders merge their results in one shot, strictly reproducing centralized ridge regression. It then estimates parameters for known PDEs from the derivatives of the reconstructed field. Experiments on four synthetic PDEs and NOAA sea temperature data back up the merging and reconstruction results. However, the precision really only applies to the merging step; parameter accuracy is still affected by the basis functions, derivative errors, and the equation form. Also, storage and communication costs for the statistics can get pretty high in high-dimensional settings.

Critical strengths

1.Merging summary statistics once faithfully reproduces centralized ridge regression.

2.No need to transmit raw data or repeatedly sync parameters.

3.Clearly distinguishes "merge consistency" from "physically accurate parameters."

4.Validated on four synthetic PDEs, with noise impact studied.

5.Tested on NOAA sea surface temperature data for both reconstruction and held-out prediction.

Critical weaknesses

1.Exact merging is limited to classical fixed-feature quadratic regression.

2.Tensor-product basis functions and dense statistics scale poorly with dimension.

3.Parameter estimation assumes the equation form is known and is affected by reconstruction/derivative errors.

4.Physical parameters weren't actually recovered on real data.

5.Privacy risks from sharing summary statistics aren't analyzed.

Questions for authors

1.Where exactly does the ~5% parameter error in the Burgers equation come from?

2.How do the number of basis functions, node positions, and regularization strength affect results?

3.How does it perform when different data holders have varying noise levels and sampling densities?

4.How much accuracy improvement would analytical spline derivatives bring?

5.After implementing sparse statistics, how much communication is actually saved?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.