acceptodds
Under review as a conference paper at ICLR 2027

Refinement Does Not Amplify Discretization Mismatch in Fourier Neural Operators

Abstract

Neural operators provide a powerful framework for learning mappings between function spaces, including solution operators for partial differential equations (PDEs). Fourier neural operators (FNOs) are particularly attractive because their parameters are independent of the spatial resolution, enabling a model trained on one discretization to be evaluated on another. However, resolution-independent parameters do not imply discretization-invariant predictions: the outputs of an FNO can differ when the same input function is represented on different discretizations. This discretization mismatch is important for assessing the reliability of FNOs across discretizations, yet a precise theoretical characterization of how the output discrepancy depends on the two discretization sizes remains lacking. In this work, we derive an explicit upper bound for the pointwise discretization mismatch error (DME) of FNOs. Our bound characterizes the dependence of the DME on the size of the base discretization and the relative refinement between the two discretizations. In particular, for any fixed base discretization, the bound remains uniformly controlled as the finer discretization becomes arbitrarily fine. Thus, refining the finer discretization does not amplify the mismatch beyond a bound determined by the base discretization. Moreover, the bound converges to zero as the base discretization is refined. These results provide a quantitative characterization of how discretization refinement affects pointwise mismatch in FNOs. Numerical experiments on representative FNO settings support our theoretical results and demonstrate the refinement-uniform behavior predicted by the bound.

Primary area:
learning theory
Keywords:
Fourier Neural Operators, Discretization Mismatch, Discretization Error, Neural Operators, Operator Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_fdz7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the output discrepancy of a fixed FNO evaluated on two nested grids representing the same input. For a one-dimensional, single-Fourier-layer model with linear lifting/projection, H^s input with s > 3/2, Lipschitz activation, and bounded parameters, it gives a common-grid pointwise bound proportional to (1 - N/M)/N. This bound is uniform in the finer resolution M and vanishes as the base resolution N increases. An appendix transfers it to squared residuals evaluated on the common coarse grid. Experiments use two synthetic inputs, 100 random parameter choices, and one to four layers to illustrate base-grid convergence and apparent saturation with increasing refinement.

The intended estimate is recoverable by a short quadrature argument. However, the central coefficient proof is missing, the stated real-valuedness condition is incorrect for channel matrices, and the smooth single-layer experiment should have zero mismatch in exact arithmetic. Theoretical novelty relative to existing discretization analysis is also limited.

Critical strengths
  1. An explicit and interpretable consistency estimate. The formula separates base resolution from refinement ratio and makes the dependence on activation, input regularity, mode cutoff, and parameter bounds visible. This can help distinguish implementation mismatch from target-operator approximation error, provided the narrow single-layer scope is maintained.

  2. A clean comparison at shared physical points. Holding the input function and parameters fixed avoids mixing discretization effects with retraining or interpolation choices. The paper also acknowledges that its discontinuous input falls outside the theoretical regularity regime. These choices give a reasonable starting point for a focused analysis.

Critical weaknesses
  1. Limited theoretical novelty and scope. Lanthaler et al. (2024), Remark 4 and Theorem 5, already imply refinement-uniform common-grid and reconstructed L2 discrepancies under smooth-activation assumptions, by a triangle inequality through the continuous model. The submission allows merely Lipschitz activations, but analyzes only one layer; its explicit ratio factor follows from elementary paired quadrature for any integer refinement. Under its stated Sobolev assumptions, standard aliasing estimates also yield faster base-grid decay for this one-layer model. The work does not analyze repeated nonlinear aliasing in deep FNOs or establish why the new estimate is consequential beyond existing theory.

  2. Incomplete proof and unresolved empirical validation. Appendix B.3 does not contain the proof of the key Lemma 3.4. Moreover, the smooth test input has Fourier support only at +/-1 and +/-2, fully resolved on all tested grids. For the defined T = 1 architecture, all retained coefficients and common-point local terms therefore agree exactly, so the mismatch is zero in exact arithmetic for every parameter choice. The nonzero curves in Figures 1a and 2a require reconciliation with sampling, precision, and implementation before they can support the claimed behavior. The discontinuous-input and multilayer tests are outside the theorem. Together these issues leave no convincing nontrivial numerical validation within the stated setting.

Questions for authors
  1. Can you provide the intended proof of Lemma 3.4, replace the duplicated Appendix B.3, and state all assumptions used? Please also clarify whether the matrix symmetry condition was intended to mean entrywise conjugation rather than the explicitly defined conjugate transpose.

  2. How do you explain nonzero T = 1 mismatch for the band-limited smooth input? Please provide the exact sampling/restriction operations, full implemented architecture, activation, parameter initialization, dtype, and code revision. Do float64 and exact-Fourier-coefficient controls give zero up to roundoff? A smooth input with a nonzero unresolved spectral tail would provide a nontrivial test of the theorem.

  3. What substantive theoretical conclusion is new beyond existing continuous-versus-discrete convergence results and standard quadrature/aliasing estimates? Please distinguish the weaker activation assumption and explicit ratio factor from the broader refinement-uniform conclusion, and justify the suggestion that whole-domain L2 comparison causes refinement-dependent growth.

  4. Is Theorem C.2 intended only for fine outputs restricted to the coarse grid? Its loss uses an unnormalized sum over common points and is not the ordinary fine-grid training objective. Please state precisely what training-related conclusion, if any, follows from it.

Reviewer_8f9xOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the consistency of the outputs of a fixed-parameter Fourier Neural Operator (FNO) across different discretization resolutions. For a one-dimensional, single-layer FNO, the authors introduce the pointwise discretization mismatch error (DME) to quantify the discrepancy between FNO outputs evaluated on two different discretizations of the same input function, and prove an upper bound of the form DME(x)≤CN(1−NM)\mathrm{DME}(x) \leq \frac{C}{N}\left(1-\frac{N}{M}\right). Thus, for a fixed resolution NN, the bound remains uniform with respect to the finer-grid resolution MM, and it converges to zero as NN increases. The authors conduct numerical validation using randomly initialized FNOs on two one-dimensional functions and also investigate the behavior of multilayer models.

Critical strengths
  1. The paper addresses a meaningful question: in neural operator learning, resolution-independent parameters do not imply discretization-invariant predictions.
  2. The upper bound explicitly distinguishes the effects of the base resolution and the refinement level, clearly illustrating the central point that refining the finer grid does not amplify the output mismatch at shared grid points without bound.
  3. The authors explicitly state that the theory applies to single-layer models and acknowledge that discontinuous inputs do not satisfy its regularity assumptions. These scope limitations help readers understand the conditions under which the results apply.
Critical weaknesses
  1. The central proof is incomplete in the current manuscript. On page 7, the proof of Lemma 3.4 is deferred to Appendix B.3. However, B.3 is identical to B.2, which proves upper bounds on the parameter norms, leaving the key proof missing. Furthermore, the conclusion of Lemma 3.4 uses the parameter bound DD from A3, whereas its statement lists only A1.
Questions for authors
  1. Please provide the complete proof of Lemma 3.4. Appendices B.2 and B.3 are currently identical.
  2. Please explain why a1a_1 produces a nonzero DME in the single-layer model. What gives rise to the nonzero curves in Figures 1 and 2? Please specify the sampling coordinates, grid alignment procedure, numerical precision, and any additional operations used.
Reviewer_ss7bOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how the discretization mismatch error from the previous work of Gao et al. (2025) behaves when a pointwise error measure is used, which the authors call the pointwise discretization mismatch error (DME). Instead of considering the error in a function-space norm as in Gao et al., the authors compare the FNO outputs at common grid points of two different discretizations. In Theorem 3.5, they derive an upper bound of order O((1−2−r)/N)O((1-2^{-r})/N), where NN is the coarse grid size and M=2rNM=2^rN is the finer grid size. The bound shows that the pointwise error remains uniformly bounded as the finer discretization is refined. To support this theoretical result, the authors conduct numerical experiments using two synthetic input functions, one smooth and one discontinuous. The outputs are compared on two nested grids, with the coarser grid points also contained in the finer grid. For the experiments, they use FNOs with one to four layers and randomly generated network parameters. The theoretical bound is derived only for a single-layer FNO.

Critical strengths

Strength 1:

The authors provide an explicit upper bound for the pointwise discretization mismatch error. The bound shows how the error depends on the coarse grid resolution and the refinement level, and that it remains bounded as the finer grid is refined.

Strength 2:

The numerical experiments explore how the discretization mismatch changes with different grid resolutions. The authors use two synthetic input functions and 100 randomly generated sets of FNO parameters to examine whether the observed behavior is consistent across different settings.

Critical weaknesses

Weakness 1:

The paper builds on the discretization mismatch problem studied by Gao et al. (2025), but considers a pointwise error instead of a function-space norm. While the authors derive an upper bound that remains bounded as the finer resolution increases, it is not clear what additional theoretical insight this provides compared to previous works. The authors mention in Section 1.3.3. that the bound by Gao et al. (2025) contains a term BMB_M that "may grow with the finer discretization size MM", but this is a structural argument, which is now shown in this paper. Since the two papers consider different error measures, the bounds are not directly comparable. The authors should clarify what can be concluded from the pointwise bound that was not already covered by existing results.

Weakness 2:

The theoretical result is derived only for a single-layer FNO, while the numerical experiments also consider networks with up to four layers. However, the paper does not explain how the theoretical findings relate to these multilayer experiments. It remains unclear what conclusions can be drawn from the multilayer experiments and how the theoretical error bound relates to the numerical results for these networks.

Questions for authors
  1. Theorem 3.5 establishes the pointwise discretization mismatch bound only for a single-layer FNO, while the experiments also consider networks with up to four layers. Could the authors explain why they include multilayer FNOs in the experiments and how these results support their theoretical findings?

  2. The experiments show similar discretization mismatch behavior for both smooth and discontinuous input functions. Could the authors discuss whether the regularity assumption in Theorem 3.5 could be relaxed?

  3. Could the authors compare the pointwise discretization mismatch with the function-space mismatch considered by Gao et al. (2025) in their numerical experiments, to clarify what additional insight the proposed error measure provides?

  4. Could the authors plot the bound against the measured errors to visualize its tightness and the discrepancy?


Review edited in the rebuttal phast:

The extension to multilayer FNOs in Theorem 3.6 in the revised version of the paper addresses my main concern regarding the limited scope of the original theoretical results. I appreciate that the authors have provided a detailed proof accounting for the propagation of discretization mismatch through successive Fourier layers.

I consider this a meaningful improvement that strengthens the theoretical contribution of the paper. As my main concern has been addressed, I improve my rating to Weak Accept.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.