acceptodds
Under review as a conference paper at ICLR 2027

Wrong Rays: In-Distribution Accuracy Does Not Certify Transport in Learned PDE Propagators

Abstract

Neural PDE models are commonly evaluated by prediction error on held-out trajectories, but downstream physical tasks may depend on dynamical responses that are weakly constrained by this metric. We ask when in-distribution prediction accuracy certifies the transport of a localized disturbance. We find that low prediction error can coexist with systematically incorrect packet trajectories, or wrong rays, along spectral directions weakly weighted by the standard distribution, from an exact non-identifiability construction to trained models and released PDEBench data. We develop a task-conditioned characterization of this gap: the input distribution determines which operator directions are strongly constrained, whereas transport can be sensitive to different directions. In clean transport settings, this mismatch appears as phase-slope error that accumulates with rollout horizon; on released linear advection, phase-only correction accounts for \(81.7%\) of the improvement achieved by full-complex correction. The error is not removed by the tested representation-consistency, activation-anti-aliasing, or longer-training interventions. An information-presence test further shows that explicit spectral estimation recovers the corrective response from the same broadband observations that the tested end-to-end fine-tuning procedure fails to exploit reliably. These results distinguish predictive accuracy from certification of task-relevant dynamics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, operator learning, scientific machine learning, partial differential equations, spectral analysis, distribution shift, transport dynamics, model reliability

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

This paper points out that in-distribution accuracy does not support the accurate transport behavior. Also, it tries to demonstrate and explain this gap. In addition to that, the authors shows some experimental evidence to clarify their claim.

Critical strengths

(S1) The paper addresses an important problem such as the gap between the prediction accuracy and physical fidelity. Also, the starting point of the discussion is understandable and good.

(S2) As authors mentioned, this topic might be connected to the OOD behavior of the learning-based simulator. In this sense, the direction of the research question makes sense.

Critical weaknesses

(W1) The most important point is that the paper is not well-organized at all. Therefore, for me, it is difficult to catch the motivation, contribution, and importance of this work. Also, the visualization by each figure is not effectively used. In addition, some part of figures are broke, for example, the numbers in Figure 1 (c) are overlapped by plots.

(W2) As I mentioned in the strengths, the starting point of the discussion is understandable. However, the importance or naturally of the direction of discussion is not clearly written. Also, the merits or soundness of the analysis for their motivation is not well validated.

(W3) The experimental and analytical evidence is limited. I cannot catch the generality of the discussion, which is related to (W1).

(W4) The discussion is limited to advection, which is somehow minor.

Overall, author should comprehensively reconsider the organization and presentation of concept of this work to make it easy to understand for broader readers.

Questions for authors

(Q1) Please explain again the concept and main contribution of this work in language that even non-experts can understand.

(Q2) How general is the method? For example, do you think that this type of discussion can be applied for non advection physics, and how?

(Q3) The figures are created by AI? Personally, it is OK, but it is better to arrange each component in tidy way for clarify the meaning of the figures.

Reviewer_fU96OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper wonders whether low in-distribution prediction error in learned PDE propagators guarantees accurate transport of localized disturbances. Models with low standard prediction error are shown to exhibit substantial wave-packet displacement errors in weakly represented spectral directions. The authors propose a task-conditioned sensitivity framework, derive bounds relating spectral coverage and phase errors to transport fidelity, and evaluate the phenomenon using synthetic systems and PDEBench datasets. They also investigate spectral corrections, architectural effects, and alternative explanations.

Critical strengths

The phase/amplitude intervention experiments provide useful evidence about the mechanisms underlying packet displacement. The authors also include controls and report negative results that help delimit their conclusions.

The paper is also easy to follow, although I found the writing style rather austere (which is a personal taste and do not impact my rating).

Critical weaknesses
  • The conceptual and theoretical novelty is quite limited:

The main theoretical results are largely direct applications of established mathematical principles: covariance-weighted error norms, dual-norm sensitivity, Cauchy-Schwarz inequalities, Fourier identities, and classical group-velocity analysis. Related work, including Trefethen (1982), Dong and Cheng (2026), and Gao et al. (2026), already establishes important components of the proposed interpretation. The task-conditioned formulation is coherent but does not constitute a sufficiently substantial theoretical advance.

  • The empirical significance is too limited (for now) which limits the generalizability of the claims

The strongest results concern relatively simple linear transport settings. The released-data experiments rely primarily on selected PDEBench problems, while the Navier–Stokes evaluation uses only a 32 × 32 representation and achieves moderate prediction accuracy. More challenging, high-resolution benchmarks, such as The Well, are absent. Moreover, Appendix E.2 reports substantial full-field prediction errors on the out-of-distribution packets themselves, making the observed transport failures less surprising.

Although the paper includes extensive ablations and controls, I do not think the empirical contribution is sufficiently significant to compensate for the limited theoretical novelty or to establish a broadly useful certification framework.

Questions for authors
  1. What theoretical result is fundamentally new beyond existing dual-norm sensitivity and classical numerical dispersion analysis? Or is the main novelty primarily in applying these principles to learned PDE propagators? If so, what new theoretical insight does this application provide? Perhaps this is partly a presentation issue, but despite the formal development (that seems correct), I found the underlying mathematical ideas rather classical and I do not manage to identify what constitutes the main theoretical advance in this work.

  2. Can the proposed certificate provide quantitatively informative bounds on actual transport errors, rather than merely identifying poorly constrained spectral directions?

I am open to revise my score if convinced during the rebuttal.

Reviewer_p1ZdOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors consider neural approximates of discretized PDE propagation operators. They investigate the question whether the standard prediction risk can certify the transport behavior of downstream tasks. Their case in point is the obsevable defining the position of a localized packet. In that context, the authors experimentally demonstrate a transport-certification gap; they characterize that gap through task-conditioned sensitivity and isolate it from the effects of representation, optimization, and information availability. The authors use advection equation and Schroedinger dynamics as their chief examples. The authors’ main theoretical result is Corollary 1, where they provide an upper bound on phase shift along a band in terms of both the prediction risk as well as the alignment of the input distribution with the band. This demonstrates that a considerable phase shift is possible despite low prediction risk. Corollary 1 is established under strict Fourier-diagonalizability assumption for both (discretized) propagator and its approximator. As such, it is not directly applicable to neural propagator approximators.

Critical strengths

Demonstrating that a considerable phase shift is possible despite low prediction risk

Critical weaknesses

Linear, translation-equivariant discretized propagators are a limiting assumption that the authors’ analysis depends on in order to use diagonalizing Fourier modes. In constant-velocity advection the discrete propagator is linear and can be made translation-equivariant with the right choice of a discretization scheme. When the velocity field varies spatially, the propagator remains linear, but loses translation equivariance.

I looked in the main body of the manuscript and the appendix for a formula for c(k), but couldn’t find one. The estimated complex correction c(k) is critical to any assertion regarding error reduction with targeted spectral correction (e.g., L198-9). Without such a formula, error reduction assertions are impossible to evaluate. In particular, I am left wondering if there hasn’t been some information transfer from the true target to G_\theta via the correction c(k).

The authors state that an “agreement on the observed subspace does not determine off-support transport” while “packet transport can depend strongly on weakly weighted directions”. This I find unsurprising: if \mu provides no meaningful information regarding T then surely learning T will fail for downstream tasks. An extreme case would be a situation where \mu is localized along directions where T is constant or almost so. But I don’t see this phenomenon as a failure of NRMSE. As such, I would have found a direct certification of alignment between \mu and T more illuminating: the phenomenon in itself is task agnostic.

Overall, I found the manuscript disorganized and, as a results, hard to asses. Basic mathematical constituents were missing or incomplete, and I was left having to guess them. It is a shame, because presenting the basic math ingredients of the problem in an explicit, precise fashion, moving from the continuous to the discrete setting and using a running example such as the one-dimensional advection with constant velocity, should have been perfectly suitable for a reader with a grad education in math / CS / physics / etc. As a specific example of the manuscript’s opacity, on Section 3.2, I could not infer how c(k) is meant to be estimated or perhaps learned? I did not understand what \hat{u}\theta^{\text{base}}(k) was, so I guessed it to be the k-th coefficient of the N-discrete Fourier of G\theta. Given that these two are fundamental to the analysis, they deserve a clear exposition. IMO, getting the manuscript to the point where it is ready for publication would require substantial intervention beyond the scope of an ICLR review iteration.

Local issues: L117; equation (2) overflow L053, L128, L241: “weights” is not a verb. Do you mean “prediction risk weighs operator error”? L151: What is a: a \in \mathbb{R}{>0}? a \in \mathbb{C}? Because you’re working sometimes over \mathbb{R} and sometimes over \mathbb{C}, it’s important to make these distinctions clear. L156: \phi\theta(k) —> \phi_\theta(k;a) L250: You need to add the assumption that \mu has finite second moment L261: B_F is the gradient of F at T — the T dependence should be made clear. Clearly state that you are using the Riesz representative. L264: the sup is over all perturbations A such that … L269: equation (9) \dagger/2 (and later in L286-7) —> explain notation L282: be consistent in the use of arg or Arg, also, what is \bar{g_j} L284: what is q??

Questions for authors
  • Is there any way to push your analysis beyond linear, translation-equivariant? On the face of it, the phase-slope analysis in (5) loses its interpretation as a transport velocity, and the corrections in (6) are no longer justified as reconstructions of the true dynamic.

  • c(k) = \tilde{g}(k)/\hat{g}_\theta(k)??

  • I couldn’t make out whether targeted spectral correction is offered as a practical correction method or a diagnostics tool.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.