acceptodds
Under review as a conference paper at ICLR 2027

WHAT CAN A REPRESENTATION IDENTIFY? SPECTRAL LIMITS OF LEARNING PDE DYNAMICS FROM SPATIAL DATA

Abstract

Inferring dynamical laws from spatial observations is fundamentally limited by the information preserved in the representation supplied to the inference method. We study this limitation for PDE identification from one or a few spatial observations through a representation-dependent view of identifiability. For linear translationinvariant PDEs, we show that power spectral density (PSD) captures only the real part of the Fourier symbol, while phase-sensitive temporal information is required to recover its imaginary part. Consequently, mechanisms such as advection and odd-order dispersion are invisible to magnitude-only spectral representations. We characterize when PDE parameters can be recovered from spectral data and show how identifiability depends jointly on the governing dynamics, observation protocol, and representation used for inference. When elapsed time is unknown, only parameter ratios, parameter combinations, or spectral shape information may remain identifiable, while a single spatial snapshot generally confounds the initial condition with subsequent dynamics. Numerical experiments distinguish structural non-identifiability from practical estimation error and representation-induced information loss. These results motivate a representation-first approach to learning dynamics from spatial data: before selecting an inference architecture, one should determine which dynamical distinctions the chosen representation makes learnable.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation learning, identifiability, partial differential equations, spectral methods, power spectral density, system identification

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_idZcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

this work builds a representation-dependent theory of PDE identifiability, showing that PSD magnitude reveals only the real fourier symbol while phase information is required for advection and others. The work validates the results over 22 pde families.

Critical strengths

The paper has some theoretical analysis. It separates practical estimation difficulty, structural non-identifiability, and information destroyed by the representation. It has a nice demonstration that good spectral fit can coexist with incorrect inferred parameter ratios

Critical weaknesses

The central theory assumes linear, translation-invariant/constant-coefficient dynamics and much of the numerical study uses controlled periodic simulations. With that said, there is a practical usability gap of the identifiability theory, and the noisy, heterogeneous, and nonlinear biological imaging.

Questions for authors
  • could authors include a real-image experiment?
  • which core results can be extended to slowly varying coefficients or weakly nonlinear dynamics?
  • another question: Why is the identifiability of PDE a topic or interest for ICLR audience?
Reviewer_JP9yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies representation-dependent identifiability of linear, constant-coefficient PDEs from spatial observations. It organizes observation protocols according to the available temporal information and knowledge of the initial state. The analysis separates the real Fourier symbol, accessible through PSD evolution, from the imaginary component, accessible through phase-sensitive observations. For a single frame at unknown time, the paper proposes a shape-and-level estimator using an initial-spectrum model and growth calibration from the image mean. Experiments across 22 named PDE families examine spectral-class selection, parameter-ratio recovery, and failures caused by mismatched initial-state or generative assumptions.

Critical strengths
  1. The paper provides a useful framework for interpreting spectral inference under different observation protocols. The distinction between absolute coefficients, scale-invariant combinations, and quantities requiring an initial-spectrum model or growth calibration is practically meaningful. In particular, the single-frame shape-and-level analysis makes explicit the additional assumptions needed to interpret a fitted spectral quantity as a mechanistic parameter. This is the paper’s strongest potential contribution.

  2. The model-mismatch experiments address an important failure mode of scientific inference. Section 4.5 demonstrates that an excellent spectral fit can coexist with substantially incorrect parameter ratios. Appendix G.1 also evaluates the shoulder estimator under its matching persistent-source model, while Appendix H examines the breakdown under nonlinear growth. These controls give concrete support to the warning that goodness of fit alone does not establish valid parameter recovery.

Critical weaknesses
  1. The contribution beyond directly relevant spectral-identification work is not sufficiently established. He, Zhao, and Zhong [1] already separate even- and odd-order coefficients using Fourier amplitude and phase, respectively, and establish two-snapshot identifiability under spectral-excitation and sufficiently small time-gap conditions (Section 3.1, Eqs. (3.5)–(3.6), Theorem 3.2). This is directly relevant to contributions 1 and 4 of the present manuscript: the PSD ratio is the squared magnitude of the Fourier ratio, and the cross-spectrum has the same phase as that ratio. The paper does not cite or distinguish itself from this work. The single-frame calibration may provide a meaningful additional contribution, but the manuscript needs to establish that increment explicitly rather than present established two-frame spectral identification alongside it as new. This omission materially weakens the novelty argument.

  2. The claimed phase recovery is not justified by the stated alias-control procedure, and its main illustration is inconsistent with the equations. Algorithm 4 computes ϕ=arg⁡C\phi=\arg C and then selects a “phase-unambiguous” band using ∣ϕ∣<π|\phi|<\pi. Since the principal argument already lies in (−π,π](-\pi,\pi], this test cannot detect wrapped observations. Appendix F correctly gives the required condition ∣ω(k)Δt∣<π|\omega(k)\Delta t|<\pi in Eq. (39), but the algorithm does not explain how that condition is ensured without knowing the true coefficient. In addition, Fig. 1(f) shows a positive phase-versus-k3k^3 slope for a positive β3\beta_3, whereas Eq. (35) and Algorithm 4 require the opposite sign under the stated convention. These issues appear fixable, but they concern a central claimed recovery mechanism. The assumptions, actual implementation, and reported figure need to be reconciled before this part of the evidence can be considered reliable.

[1] Yuchen He, Hongkai Zhao, and Yimin Zhong. How Much Can One Learn a Partial Differential Equation from Its Solution? Foundations of Computational Mathematics, 24, 1595–1641, 2024. https://doi.org/10.1007/s10208-023-09620-z

Questions for authors
  1. Clarification of novelty: Which central results extend beyond He et al. [1], and what is the precise new identifiability result supplied by the single-frame shape-and-level calibration? Please distinguish the established two-frame results from the additional contribution.

  2. Clarification and verification of phase recovery: How is the phase-unambiguous band selected without access to the true ω(k)\omega(k)? Please specify the assumed parameter/displacement bounds or the implemented alias-handling procedure. Can the authors also reconcile Fig. 1(f) with Eq. (35) and Algorithm 4, and confirm whether the discrepancy is confined to the plotted sign or affects the reported recovery?

  3. Clarification of the single-frame observation model: Equations (5)–(6) concern ensemble expectations, whereas Eqs. (7)–(8) use an observed image mean. Is Proposition 2 intended as a population-identifiability statement or an exact statement for a finite realized image? In particular, does known n0n_0 mean the point-process intensity or the realized initial cell density? This distinction affects the interpretation of the main single-frame contribution.

Reviewer_U1e5OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The broad question of the paper is: which PDEs can be recovered once the data is compressed into PSD magnitude?

While the exposition of the paper is acceptable, the remaining parts are unintelligible. The paper seems to be heavily AI generated and almost impossible to read. I believe this paper does not fit the publication standards of ICLR and should be rejected. I also believe that the author's AI statement is missleading: AI usage seems to go well beyond "assist[ing] with manuscript editing and with reviewing and revising research code."

Critical strengths

None.

Critical weaknesses

This paper is heavily AI written and unfit for this conference.

Questions for authors

I believe that this paper is entirely AI generated and is of poor quality. I urge the authors not to submit papers like this.

Reviewer_ES4XOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies identifiability in the context of partial differential equations (PDEs), discussing what influence different observation protocols, underlying dynamics, and architectural assumptions have on which properties of the PDE can be recovered (or not). The paper shows with numerical experiments what remains unidentifiable vs. which mismatches are caused by approximation errors.

Critical strengths
  • I like the idea of the paper and its connection between realistic measurement protocols and the limits of how constrained a model inferred from these measurements can be.
  • I especially like the (very important) point about goodness of fit not implying good identifiability (i.e. the shoulder template result), which is clearly shown and argued.
Critical weaknesses
  • The paper is not really integrated into the wider literature. It lacks a related work section, and there are only 7 citations in total, some of which (e.g. Brunton's SINDy or Raissi) are somewhat general and not specific to the problem at hand. It seems only Pasetto&Daher are actually directly relevant. For instance, the main vocabulary of the paper derives from the identifiability literature, which is completely ommited, and the field has clearly already studied some of these questions (e.g.that equal-time spectra reveal only the real part of the dispersion relation is standard knowledge). As a result, the novelty of the core contributions is difficult to assess.
  • Tumor biopsy seems central to the paper's framing (first sentence, §4.5, discussion), yet no real or microscopy-realistic data is analyzed, and the discussion's tumor paragraph discusses requirements without arguing in which conditions they would be satisfied.
  • The entire framework assumes the underlying dynamics are linear with constant coefficients. Real systems, including the tumor growth that motivates the paper, are not necessarily like this, yet the paper tests only a single non-linear setting (logistic growth) where parameter errors become large while the goodness of fit stays essentially perfect. Doesn't this suggest the linear assumption is doing a lot of work behind the scenes? How limiting is the linear assumption for the motivating application? I think it would significantly strengthen the paper if the taxonomy of the paper results would be more clearly extended to the nonlinear case/this case would be more thoroughly discussed.
  • The repository link is not actually provided, but still contains the placeholder.
  • The space is not used effectively (only 7/9 pages), which could have been used for instance for an illustrative figure of the overall framework and approach.
Questions for authors

Some small comments:

  • The figure reference in Fig.2 is unresolved

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.