acceptodds
Under review as a conference paper at ICLR 2027

Radon Spectral Neural Operator: Learning PDE Operators in the Radon-Fourier Domain

Abstract

Standard Fourier neural operators use frequency-wise multipliers in their linear spectral branches. We introduce the Radon Spectral Neural Operator (RSNO), a hybrid neural operator that combines multiscale spatial processing with structured global interactions in Radon-Fourier coordinates. The spatial branch aggregates fine- and coarse-resolution convolutional features, while the spectral branch applies orientation-dependent channel filtering and cross-orientation mixing within radial frequency shells. Their outputs are fused in each residual block, making multiscale aggregation and global spectral processing the two central architectural components. We prove an exact shellwise representation of the continuous branch, establish Sobolev stability, and show that different angular actions range from standard Fourier multipliers to nonlocal Fourier-integral structures. Across four PDE benchmarks spanning Darcy flow, Navier-Stokes forecasting, Helmholtz waves, and Burgers dynamics, RSNO delivers consistently strong performance against a broad set of neural-operator baselines. Ablations show that removing the Radon spectral branch increases mean error across all evaluated benchmark settings, whereas removing cross-orientation coupling produces substantially smaller changes. Together, these results support shellwise Radon–Fourier processing as a principled and analyzable global component within a multiscale neural-operator architecture.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
PDE, Neural Operator, Radon Transform, Multiscale representations

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

This study improves neural operators by using multiscale spatial branch and Radon–Fourier spectral branch in residual block of , which is called RSNO. The spectral branch uses orientation-dependent channel filtering and cross-orientation mixing in the same residual shell. In particular, it is realized by polar sampling and compensated adjoint gridding without a sinogram.

Critical strengths

(1) This paper addresses multiscale structure, which is an important topic in learning PDEs, and extending FNO with orientation-aware interactions is an attractive direction.

(2) The paper provides many proofs, which show the effort to give the proposed method a theoretical foundation.

Critical weaknesses

To be honest, the current manuscript is very hard to read. The Abstract is sometimes inconsistent with the real contributions or explanations in the main body. The definitions in earlier sections are overridden in later sections. Although I spent a fair amount of time to read it, I am not very confident that I understood it correctly.

(1) angular mixing The contribution of the angular mixing in the spectral branch is unclear. Comparing Full RSNO (e.g., 0.4426 ± 0.1870 for Darcy) and No angular mixing (0.4455 ± 0.2033) in Table 5, the improvement by the angular mixing is limited and almost negligible, given the standard deviation in Table 4. So, Abstract and Introduction are somewhat overclaiming the contribution. The improvement by the spectral branch seem significant. No visualizations of angular map are provided. They might help understanding its contribution.

(2) Baseline implementations It is unclear why No Radon spectral branch in Table 5 is much better than multiscale baselines (e.g., U-NO, MgNO) in Table 1 on Navier-Stokes. These architectures are all multiscale. So, the difference might come from some difference in training procedures, and comparisons might have been unfair. Detailed experimental settings and/or computational budgets are expected to be disclosed. I found that reported results of baselines (e.g., GNOT, RNO) tend to be worse than those provided in the paper of RNO. It is also unclear why the training epochs for FNO, U-NO, F-FNO are truncated, as mentioned in Table 10.

(3) Inconsistency between Theorems/Contributions and Implementations Abstract says "show that different angular actions range from standard Fourier multipliers to nonlocal Fourier-integral structures" and "principled and analyzable global component". However, in the real implementation, Appendix B says "only indices 1, . . . , Nr − 1 are passed to the mixer; the spectral-branch output at DC and beyond the cutoff is zero." Then, Theorems A.x say nothing. I found other explanation not to hold for the actual implementations, such as "the linear branch does not move information between different radial frequencies."

(4) Details of implementations. cross-orientation mixing requires the spatial origin of the orientation, but it is not mentioned clearly. "When a Radon spectral branch is applied, a latent channel is extended by zero outside Ω, yielding a function on Rd". Then, the periodicity of Navier-Stokes is not guaranteed, whicle the spatial branch did. In contrast, for Burgers equation, "Circular convolutional padding is applied along both axes as an architectural convention," but the time axis should not be periodic.

Questions for authors

See above.

9 comments
AuthorsOpenReview ↗

Response to Reviewer PoQw (part I)

Thank you for the careful reading and constructive comments on the presentation. We recognize that the original manuscript did not sufficiently distinguish the central architectural idea from its mathematical analysis and numerical realization. We have clarified these distinctions in the revised manuscript.

The central idea of RSNO is straightforward. The Fourier Neural Operator (FNO) provides a familiar starting point for learning PDE solution operators. In a standard FNO spectral layer, each retained Fourier mode has a learned matrix that mixes channels at that frequency. The layer can therefore change the amplitude and phase of an existing mode, but it cannot directly transfer its coefficient to a different frequency. Some operators require precisely this transfer: a rotation, for example, moves a Fourier coefficient from one direction to another while preserving its frequency magnitude. Learning an independent coupling between every pair of retained frequencies would make the parameter count grow quadratically with the number of modes. RSNO instead writes a frequency in polar form, ξ=rθ\xi=r\theta, and allows coefficients at different orientations θ\theta to interact at a fixed radius rr. The Fourier-slice theorem gives this construction its Radon interpretation: the Fourier transform of a Radon projection at orientation θ\theta and detector frequency rr gives the coefficient v^(rθ)\widehat v(r\theta). This is the central design idea behind RSNO. Our main representation theorem makes it precise for the continuous linear branch: at each radius, the output is obtained by applying a learned angular and channel map to the input coefficients on that same shell.

For a two-dimensional field, the Radon transform records its integrals along lines. The unit normal θ\theta specifies the direction perpendicular to a line, and the signed offset ss specifies its position. The Fourier-slice theorem gives Fs[Rv](θ,r)=v^(rθ)\mathcal F_s[\mathcal R v](\theta,r)=\widehat v(r\theta). At fixed rr, the coefficients therefore describe different directions at the same spatial frequency magnitude. An orientation-dependent multiplier changes the amplitude or phase of the coefficient already present in a direction. Cross-orientation mixing additionally allows the output in that direction to depend on coefficients in other directions on the same shell.

The rotation example in Appendix A (after THeorem A.5) illustrates the connection to wavefront transport. Under the map v(x)↦v(QTx)v(x)\mapsto v(Q^T x), a wave at kk becomes a wave at QkQk. For a nonsmooth field, the same map carries a wavefront at position and covector (x0,ξ0)(x_0,\xi_0) to (Qx0,Qξ0)(Qx_0,Q\xi_0). The angular action specifies the transport; the Radon–Fourier coordinates make it possible to express it. The theorem concerns specified continuous angular actions under high-frequency assumptions. Our trained models have a finite frequency cutoff. The rotation experiment below uses smooth inputs. Its measured quantity is directional transfer at resolved frequencies.

We have also clarified the scope of the theoretical results. The exact shellwise representation, Sobolev estimates, and wavefront characterizations concern the continuous linear spectral branch under their stated assumptions. They are not claimed as exact properties of the complete finite-dimensional nonlinear network. Sections 3.5 and Appendix A.6 now make this distinction explicit.

Finally, the revised experiments distinguish the contribution of the Radon spectral branch from the additional angular mixing. We acknowledge that angular mixing yields relatively small improvements on the main PDE benchmarks, while the new rotation experiment demonstrates its effectiveness for a task requiring directional frequency transfer.

AuthorsOpenReview ↗

Response to Reviewer PoQw (part II)

A1. On Darcy, the seed-0 test error was 0.4455% without angular mixing and 0.4426% with it. We agree that this is a small increment. We also tested a task that explicitly requires a directional transfer. Each example is a random sum of 24 Fourier modes; the target rotates each wavevector by 90° while preserving its magnitude and coefficient. Across paired seeds 0, 1, and 2, turning mixing on reduced mean test error from 4.4845% to 0.6382%. The paired reduction was 3.8463 ± 0.0623 percentage points (mean ± sample SD across seeds). The enabled model has 161,090 parameters versus 119,618. As a separate check, we zeroed only the off-diagonal angular weights in the three trained models: validation error rose from 0.60–0.66% to 100.20–100.69%. This establishes that the fitted models used those connections for the rotation. The individual seed results are given in our response to Reviewer uZaQ.

Figure 7 in the revised appendix shows the learned angular mixing weights for Darcy flow and the synthetic 90° rotation task on the same color scale. On the rotation task, a pronounced band connects input and output orientations separated by 90°, as prescribed by the target map. The Darcy panel shows no comparable band. Each cell shows the root-mean-square magnitude of the learned complex update Δi,o\Delta_{i,o} from input angle ii to output angle oo. On the rotation task a band follows o≈i+90∘o\approx i+90^\circ, the Darcy weights are more dispersed.

A2. On NS, the methods use the same 1,000 training trajectories and 200 test trajectories and are scored on the same ten-frame forecast with their final checkpoint. The spatial-only RSNO in Table 5 retains dilated local convolutions and learned coarse-grid branches. U-NO uses a multilevel operator encoder–decoder, and MgNO uses a multigrid V-cycle. “Multiscale” describes a feature shared by all three, not an identical backbone. The archived NS runs also show substantial differences in optimization budget.

NS model (seed 0)Trainable parametersBatch sizeEpochsOptimizer updatesTraining time (min)Test relative L² (%)
RSNO-H10, spectral branch disabled373,105850062,500192.37.3771
U-NO30,478,0331650031,50068.413.7822
MgNO2,922,5085050010,00098.612.5206

The RSNO ablation used teacher forcing, a growing training horizon, and a relative L² loss over that horizon. U-NO and MgNO used prediction feedback over ten steps and summed per-step relative errors. The evaluation is shared, and the measured gap combines architecture and training-protocol effects. To faithfully reproduce the results reported in the original papers, we used each baseline’s original training protocol without modification. For all methods, the anonymous repository provides the implementations, detailed experimental settings, and computational budgets.

The original GNOT paper (Section 4.1 and Table 1) identifies Darcy and time-dependent Navier–Stokes as FNO benchmarks, but it does not specify the spatial resolution used for its Darcy result or the viscosity used for its Navier–Stokes result. The errors reported for the other baselines in its Table 1 are also insufficient to determine these settings. To evaluate GNOT on the same samples as the other methods in our manuscript, we retrained it on our datasets, using an 85 × 85 grid for Darcy and a viscosity of ν=10−5\nu=10^{-5} for Navier–Stokes. Differences in the experimental settings may contribute to this gap, although the published information does not establish its cause.

The RNO paper (Section 4 and Table 1) describes a common training protocol for the competing methods, but it does not document a separate GNOT run for either benchmark. Several baseline entries in its Table 1 closely match those in the Transolver paper. In particular, both papers report exactly 1.05% on Darcy and 13.8% on Navier–Stokes for GNOT. Transolver explains in Appendix B.3 that it used previously published results for benchmarks already evaluated in the GNOT paper. The agreement across the three papers therefore does not establish that the reported GNOT values were obtained on the same dataset configurations and sample splits used in our manuscript. For a discussion of the discrepancy between our RNO reproduction and the published RNO results, please see our response to Reviewer eByq.

AuthorsOpenReview ↗

Response to Reviewer PoQw (part III)

On Burgers, the stopping criterion for FNO, U-NO, and F-FNO was a fixed budget of 50,000 optimizer updates per model. We chose this budget to give the three one-step models the same number of updates. Each of the 9,000 training trajectories provides 40 adjacent-frame pairs, so one epoch contains 360,000 pairs and, at batch size 64, requires 5,625 updates. All three runs therefore completed eight full epochs and 5,000 updates in the ninth, which is what the 9* entries in Table 10 denote. The stopping rule was set before training and was independent of test error. RSNO predicts the full trajectory from the initial condition and completed 141,000 updates over 500 epochs.

A3. In the linear spectral layer of an FNO, the learned matrix mixes channels independently at each Fourier frequency: \widehat{v}{\mathrm{out}}(\xi)=W(\xi)\widehat{v}{\mathrm{in}}(\xi). Its continuous analogue is a matrix-valued Fourier multiplier. Proposition A.4 recovers precisely this case when angular mixing is disabled. RSNO extends this parameterization by writing ξ=rθ\xi=r\theta and allowing the output at (r,θ)(r,\theta) to depend on inputs at (r,θ′)(r,\theta') for other directions θ′\theta'. For a smooth angular kernel satisfying the symbol assumptions of Theorem A.6, this interaction can give rise to a nonlocal Fourier-integral structure. The theorem identifies the corresponding canonical relation: away from the spatial origin, it can connect distinct points xx and yy with ∣x∣=∣y∣|x|=|y| and radial covectors of the same signed magnitude.

Appendix B describes how this construction is discretized. At r=0r=0, every polar angle denotes the same Fourier coefficient, so there is no distinct orientation to mix. The discrete reconstruction ramp also vanishes there. We therefore omit DC from the angular mixer, avoiding an ill-defined angular parameterization at the origin. The implementation retains radial indices 1,…,Nr−11,\ldots,N_r-1 and sets the polar output to zero beyond the cutoff. Omitting DC is consistent with the treatment of the frequency origin in Theorem A.6, which assumes that the angular kernel vanishes in a neighborhood of r=0r=0.

The statement that “the linear branch does not move information between different radial frequencies” refers to the whole-space polar representation. For every r>0r>0, the output on the shell of radius rr depends only on input coefficients on that same shell. The discrete mixer preserves this property exactly at the level of polar indices: an output at index kk uses only inputs at index kk, even when their angles differ.

Reviewer_eByqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the Radon Spectral Neural Operator (RSNO), a hybrid neural operator for learning solution mappings of partial differential equations. The proposed architecture combines a multiscale convolutional spatial branch with a Radon–Fourier spectral branch, which performs orientation-dependent channel mixing and cross-orientation interactions within fixed radial frequency shells. The authors also provide a continuous operator formulation, establish shellwise representations and Sobolev stability results under certain assumptions, and evaluate their method on four PDE benchmarks: Darcy flow, Navier–Stokes, Helmholtz, and Burgers.

The paper reports competitive performance against several existing neural operator baselines and includes architectural ablation studies. However, I have significant concerns regarding the reliability of the baseline reproductions, particularly RNO, the fairness of the experimental comparisons, and whether the proposed spectral mechanism provides a sufficiently meaningful and well-validated contribution. These issues substantially undermine my confidence in the paper's main empirical conclusions.

Critical strengths

The paper attempts to provide a mathematically structured formulation of spectral interactions in Radon–Fourier coordinates, supported by continuous operator analysis and experiments across several PDE benchmarks. The connection between the proposed architecture and shellwise spectral representations is potentially interesting.

However, the theoretical analysis does not convincingly establish the practical advantages of the proposed architecture, and the empirical evidence is weakened by substantial concerns about baseline reliability and experimental fairness. I therefore do not find these strengths sufficient to support acceptance.

Critical weaknesses

1. Serious concerns about the correctness and reproducibility of the RNO baseline, which substantially undermine the credibility of the experimental comparisons.

My primary concern is the reliability of the RNO baseline reproduction. RNO is one of the most directly relevant prior methods because it also incorporates Radon-domain representations for PDE operator learning. Therefore, a correct and faithful comparison with RNO is particularly important for evaluating the contribution of RSNO.

However, the RNO results reported in Table 1 are highly questionable. Specifically, the authors report an RNO-code relative L2 error of 0.7813% on Darcy and 19.9753% on Navier–Stokes, whereas the corresponding RNO-paper results cited in the same table are 0.510% and 8.94%, respectively. The reproduced Navier–Stokes error is more than twice the value reported in the original RNO paper. This represents a substantial discrepancy that cannot be dismissed without a careful technical explanation.

More importantly, I have independently reproduced RNO and obtained results largely consistent with those reported in the original RNO paper. My reproduction results are therefore substantially better than the RNO-code results presented in this submission. This independent observation further reinforces my concern that the RNO baseline in this paper may not have been reproduced or evaluated correctly.

The authors state that they use the released implementation of RNO under a shared evaluation protocol. However, using a common evaluation protocol does not automatically guarantee a fair comparison, particularly when it requires modifications to the original model's preprocessing, normalization, training strategy, or inference procedure. A reliable reproduction should first establish that the original method can achieve performance consistent with its published results, followed by a careful examination of any changes introduced by the new experimental protocol.

Unfortunately, the paper does not sufficiently explain why its RNO reproduction performs so much worse than the published reference. It also does not provide convincing diagnostic evidence demonstrating that the degradation is an unavoidable consequence of a legitimate protocol difference rather than an implementation, configuration, or training issue.

This problem is particularly serious for the Navier–Stokes benchmark, where the reproduced RNO error of 19.9753% makes the proposed RSNO models appear dramatically better. If RNO is substantially underperforming because of reproduction issues, the resulting performance comparison is potentially misleading and cannot provide reliable evidence of RSNO's superiority.

Given the direct methodological relevance of RNO to the proposed approach, I consider this discrepancy a fundamental weakness rather than a minor experimental detail. The authors must provide a convincing explanation and a verified reproduction of RNO before their comparative performance claims can be considered credible.

2. Insufficient evidence for the necessity and effectiveness of the proposed spectral mechanism, together with potentially confounded experimental comparisons.

My second major concern is that the experiments do not convincingly demonstrate that the proposed Radon–Fourier cross-orientation interactions are responsible for the reported improvements.

The ablation results in Tables 3 and 5 suggest that the contribution of global angular mixing is extremely limited. For example, in Table 3, the full RSNO-H10 achieves a relative L2 error of 6.3473% on Navier–Stokes, while removing global angular mixing results in 6.3811%. The absolute difference is only 0.0338 percentage points. Similarly, in Table 5, the Darcy result changes only from 0.4426% to 0.4455% when angular mixing is removed.

These improvements are marginal, and the paper does not provide sufficient repeated-seed evidence or statistical analysis to establish their robustness. In particular, the reported per-sample standard deviations should not be confused with uncertainty across independent training runs.

In contrast, the multiscale spatial branch appears to be considerably more influential. For example, reducing the spatial scales from {1, 2, 4} to {1} increases the Navier–Stokes error from 6.3473% to 11.2242%. This suggests that much of the observed performance improvement may come from multiscale convolutional processing rather than the distinctive cross-orientation interactions that motivate the proposed architecture.

Although removing the entire Radon spectral branch causes a larger degradation, this ablation also removes a substantial number of trainable parameters. As the authors acknowledge in Section 4.4, these comparisons are not parameter-matched causal estimates. Consequently, the experiments do not adequately distinguish the benefits of the proposed Radon–Fourier representation from those of introducing an additional high-capacity spectral branch.

Furthermore, there are concerns about the fairness of the temporal prediction comparisons. Appendix D.2 indicates that RSNO-H1 is trained using all 19 adjacent-frame pairs from each Navier–Stokes training trajectory, whereas RSNO-H10 uses a different training formulation with a horizon curriculum. Despite these differences, the paper presents both models alongside baselines in the same performance comparison without adequately isolating the effects of temporal formulation, training-example construction, optimization budget, and architecture.

The theoretical contribution also has a limited demonstrated connection to the empirical results. The continuous analysis establishes shellwise properties and stability bounds for a class of linear operators under specific assumptions. However, the actual implementation employs finite angular sampling, interpolation, density compensation, spatial cropping, nonlinear operations, and multiscale convolutional components. The paper does not establish that the theoretical properties translate into the claimed empirical improvements in the implemented architecture.

There is also a concrete inconsistency in the numerical implementation description. Appendix B.1 states that square-grid inputs are zero-padded from H × H to D × D with D ≥ 2H. However, Appendix C.2 states that the 85 × 85 Darcy inputs are padded to 121 × 121, which does not satisfy the stated condition. This discrepancy raises further concerns about the precision and reproducibility of the implementation details.

Overall, I do not find the current evidence sufficient to establish a compelling architectural contribution. The reported performance advantages are difficult to attribute to the proposed spectral mechanism, and the substantial RNO reproduction discrepancy further undermines the credibility of the empirical evaluation.

In summary, the paper has fundamental weaknesses in its experimental reliability and in the validation of its central contribution. The concerns about RNO reproduction are especially serious, as they directly affect the credibility of one of the most relevant comparisons. I therefore do not believe that the current submission meets the standards required for acceptance at ICLR.

Questions for authors

Q1. RNO reproduction discrepancy. Why does the reproduced RNO achieve 19.9753% relative L2 error on Navier–Stokes, while the published RNO result cited in your own Table 1 is 8.94%? What specific implementation, configuration, or evaluation differences explain this substantial degradation?

Q2. RNO reproduction fidelity. Can the authors reproduce the original RNO results using the officially released implementation and its original experimental configuration before introducing any modifications for the shared evaluation protocol? If so, please report these results and clearly document the changes responsible for the performance degradation.

Q3. Reliability of the baseline comparisons. I have independently reproduced RNO and obtained results largely consistent with the original publication, which are substantially better than the RNO-code results reported here. Can the authors provide convincing evidence that the RNO implementation used in this submission is correct and that its unusually poor performance is not caused by implementation or training issues?

Q4. Temporal prediction fairness. Would RSNO-H1 retain its reported advantage if the competing methods, particularly RNO, were trained using an equivalent single-state autoregressive formulation, comparable training samples, and controlled optimization budgets?

Q5. Contribution of angular mixing. Given the marginal improvement from cross-orientation mixing in Tables 3 and 5, what evidence demonstrates that this component is necessary? Can the authors provide parameter-matched and repeated-seed comparisons against simpler Fourier multiplier branches and other spectral alternatives?

Q6. Contribution of the Radon spectral branch. Since removing the entire Radon branch also removes substantial model capacity, can the authors provide capacity-matched alternatives to demonstrate that the observed improvements arise specifically from the Radon–Fourier representation rather than from additional trainable parameters?

Q7. Numerical implementation consistency. Appendix B.1 specifies D ≥ 2H for square-grid zero-padding, while Appendix C.2 uses D = 121 for H = 85. Which description reflects the implementation used in the actual experiments, and how does this inconsistency affect the numerical reconstruction and reproducibility of the reported results?

7 comments
AuthorsOpenReview ↗

Response to Reviewer eByq (part I)

Response to Reviewer eByq

Thank you for their comments on the RNO baseline. RNO and RSNO share a Radon-based component but use fundamentally different mechanisms to model global interactions. RNO builds on Transolver’s physics-attention (PA) mechanism. In the released Darcy configuration, approximately 88.9% of its trainable parameters belong to the attention modules, while the four Radon blocks together account for less than 0.31%. RNO’s strong performance appears to stem primarily from its Transolver backbone, with the Radon transform playing a comparatively minor role. RSNO follows the Fourier neural operator (FNO) approach, parameterizing its spectral operator in the Radon–Fourier domain and combining it with multiscale spatial convolutions. RSNO is therefore not a variant or extension of RNO. The similarity in their names therefore overstates the architectural resemblance between the two methods. The table reports RNO’s parameter counts for the two Darcy configurations.

Darcy model parameter counts8 PA layers32 PA layers
PA attention modules only2,504,00010,016,000
Full PA backbone, including attention, feedforward networks, normalization, and the final output head2,772,67311,089,537
Four Radon blocks8,4488,448
Full model2,815,73111,132,595
Radon blocks as a share of total parameters0.300%0.076%

Nevertheless, we report both the originally published RNO results and our independent reproductions to ensure transparency in the experimental comparisons.

  1. Our RNO results were obtained from documented experiments, not from deliberately weakened configurations.

We have revisited the original RNO paper, its supplementary materials, and the public GitHub repository. We found differences in the reported architecture and released configurations. As of November 8, 2026, the official RNO repository provides a Darcy experiment driver but no Navier–Stokes (NS) training or evaluation script. Even for Darcy, these sources specify different models. RNO Appendix G, Table 8 lists 8 Physics-Attention layers and 4 Radon blocks, and the supplementary model (NeurIPS supplementary code) applies PA before the Radon blocks. The GitHub Darcy launcher (Darcy launch script) specifies 32 PA layers, and its model applies 4 Radon blocks first, fuses their output with a spatial branch, and then applies PA. We therefore used the depths documented in RNO Appendix G, Table 8 for the main comparison and also tested the deeper Darcy setting. In Table 1 of the original manuscript, RNO-code reports our runs with 8 PA layers on both tasks, using 4 Radon blocks for Darcy and 2 for NS. The gray RNO-paper row gives the original authors’ published results for reference. A reliable reproduction should follow the configuration specified in the published paper and the publicly released code implementation exactly, without modifying either.

The Darcy result reported as RNO-code in the original manuscript (0.7813% relative L² error) was obtained with a reimplementation. It followed the GitHub model’s ordering: the Radon and local features were fused before the physics-attention backbone. The main configuration differences between the original Darcy reproduction and the official GitHub launch script concerned batch size, positional encoding (--unified_pos), and gradient clipping (--max_grad_norm). Since the paper does not report these settings for Darcy, the original reproduction followed the defaults in exp_darcy.py: a batch size of 8, raw spatial coordinates (--unified_pos 0), and no gradient clipping. The official launch script instead specifies a batch size of 4, --unified_pos 1, and --max_grad_norm 0.1.

AuthorsOpenReview ↗

Response to Reviewer eByq (part II)

To accommodate the available GPU, the RNO evaluation used Python 3.10 with the CUDA 12.8 builds of PyTorch 2.7.1 and torchvision 0.22.1, along with NumPy 1.22.4, SciPy 1.7.3, einops 0.6.1, timm 0.9.2, Matplotlib, and tqdm. The following Darcy runs used the released RNO implementation with the two fixes described above. Each command is followed by its final test relative L² error.

  1. 32 PA layers with the official launch settings

    bash scripts/RNO_Darcy.sh
    

    Final test relative L² error: 0.5928%. Parameters: 11165875

  2. 8 PA layers as specified in the paper, with script defaults for the slice count, positional encoding, and gradient clipping

    python exp_darcy.py \
        --data_path ./data \
        --n-hidden 128 \
        --n-heads 8 \
        --n-layers 8 \
        --batch-size 4 \
        --epochs 500 \
        --lr 0.001
    

    Final test relative L² error: 0.6492%. Parameters: 2815731

After testing several configurations, we obtained a Darcy test error close to the published result by combining the 8 PA layers specified in the RNO paper with the remaining settings from the official launch script, although neither source specifies this exact combination.

  1. 8 PA layers with the official launch settings

    python exp_darcy.py \
        --gpu 0 \
        --model RNO_2D \
        --n-hidden 128 \
        --n-heads 8 \
        --n-layers 8 \
        --lr 0.001 \
        --max_grad_norm 0.1 \
        --batch-size 4 \
        --slice_num 64 \
        --unified_pos 1 \
        --ref 8 \
        --eval 0 \
        --downsample 5 \
        --save_name darcy_RNO_layer_8
    

    Final test relative L² error: 0.5314%. Parameters: 2835955

Across the configurations tested, the Darcy result reported in the RNO paper has been difficult to reproduce. Although one configuration yielded an error close to the published value, it remains unclear whether the original authors used that exact combination of settings. A strict reproduction should follow either the official launch script (the first configuration) or the settings reported in the paper, with the script defaults used for unspecified parameters (the second configuration); agreement in test error alone does not establish the third, undocumented combination as a reliable reproduction of the original experiment. We ultimately selected Configuration 1 as the Darcy reproduction result reported in the revised manuscript, while keeping the Navier–Stokes results unchanged. Since the reviewer has reproduced results close to the published values for Darcy as well as for Navier–Stokes, could you share their reproduction code, launch commands, and parameter settings through an anonymous repository? The Navier–Stokes scripts would be particularly valuable given their absence from the official release.

  1. Our main conclusions remain valid when using the original RNO results.

RNO and RSNO use fundamentally different mechanisms to model global interactions. Given this architectural distinction and the discrepancy between the reproduced and published RNO results, Table 1 of our manuscript includes the original authors’ results in a separate RNO-paper row so that the comparison can also be assessed against their reported accuracy. Using those published values as the reference, RSNO still achieves the lowest mean relative L² error among the methods in Table 1 on both Darcy and Navier–Stokes. On Darcy, RSNO achieves 0.4426% compared with the published RNO result of 0.510%. On Navier–Stokes, RSNO-H10 and RSNO-H1 achieve 6.3473% and 3.7623%, respectively, compared with RNO’s published 8.94%. The parameter and training-time comparisons come from our runs on the same GPU. On Darcy, RSNO used 317,757 parameters and trained in 57.9 minutes, compared with 2,815,731 parameters and 92.4 minutes for RNO. On Navier–Stokes, RSNO-H10 used 3.19M parameters and trained in 337.1 minutes, compared with 11.21M parameters and 933.0 minutes for RNO. The anonymous repository provides the RSNO source code, model definitions, and trained checkpoints for all reported experiments. We invite the reviewer to load the checkpoints and evaluate them directly on the specified test splits.

AuthorsOpenReview ↗

Response to Reviewer eByq (part III)

  1. We have also strengthened the architectural evidence independently of RNO.

A new parameter-matched Cartesian Fourier control compares the two spectral representations within the same multiscale backbone. On Darcy, both variants have exactly 303,933 trainable parameters, while the polar variant achieves 0.4455% error compared with 0.5455% for the Cartesian variant. Please see the response to Reviewer uZaQ for details.

We additionally introduce a controlled rotation experiment demonstrating the representational role of cross-orientation mixing. These experiments directly address whether the proposed spectral architecture provides benefits beyond additional model capacity. Please see the response to Reviewer uZaQ for details.

We hope these clarifications and additional experiments address the central concerns regarding the reliability of our evaluation and the independent contribution of the proposed spectral representation.

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

The paper introduces RSNO, a hybrid neural operator combining multiscale convolutional processing with a Radon–Fourier spectral branch. The latter performs orientation-dependent channel filtering and learned cross-orientation mixing at fixed radial frequencies. The authors characterize the continuous linear branch through an exact shellwise representation, Sobolev bounds, and microlocal results for specified angular actions. Experiments on Darcy flow, Navier–Stokes, Helmholtz, and Burgers show a compact and accurate Darcy configuration, strong Navier–Stokes results, and competitive performance on the other two tasks.

Critical strengths
  1. The exact shellwise characterization and conditional Sobolev bounds provide a useful account of the linear branch. The relationship between pointwise angular filtering and Fourier multipliers is explicitly acknowledged, alongside the more general angular actions.

  2. On Darcy, RSNO achieves 0.4426% relative L2 error with 317,757 parameters, compared with HPM’s 0.4553% and 2,864,321 parameters. Both NS formulations outperform the evaluated baselines. The H10 ten-seed mean is 6.4861%, compared with 7.3088% for the strongest reported baseline, providing positive evidence independently of the larger H1 improvement. Helmholtz and Burgers results are competitive, although not the best .

  3. The paper includes branch and architectural ablations, clearly defined error statistics, training/checkpoint details, parameter counts, and recorded training times. Distinguishing rerun RNO results from quoted results and explicitly acknowledging unmatched ablation capacities are appreciated.

Critical weaknesses
  1. The source of the hybrid architecture's gains could be isolated more clearly. Tables 3 and 5 provide useful component controls, including a no-angular variant that is a Fourier multiplier in the continuum but retains polar sampling and compensated gridding. Removing the entire spectral branch also changes capacity substantially: NS-H10 drops from 3,190,137 to 373,105 parameters. A Cartesian Fourier branch in the same multiscale backbone, with comparable capacity and training, would better distinguish the specific polar parameterization from generic global spectral processing. This is an attribution limitation; the existing results still support the effectiveness of the integrated architecture. The smaller, single-seed angular gains on Darcy and NS would also benefit from paired variability estimates.
Questions for authors
  1. Beyond the reported no-angular polar Fourier-multiplier control, is there an existing comparison with a Cartesian Fourier branch in the same multiscale backbone under comparable capacity and training? Which result best isolates the benefit of the polar parameterization itself?
2 comments
AuthorsOpenReview ↗

Response to Reviewer uZaQ

A1. Thank you for suggesting a Cartesian Fourier control in the same multiscale backbone and repeated runs to assess the variability of the angular-mixing effect. The Darcy comparison now includes training seeds 0, 1, and 2 for all three spectral-branch configurations. At exactly 303,933 trainable parameters, the polar branch without angular mixing achieved a mean test relative L² error of 0.4405% across seeds, compared with 0.5244% for the Cartesian branch. The polar branch achieved lower error in each of the three runs.

We kept the four-block, width-20 backbone, including the local and coarse-grid branches, lifting, and projection. The Cartesian replacement applies a 2D FFT, a learned complex channel matrix at each retained frequency, and an inverse FFT. We used Cartesian mode counts (6, 6) to match the parameter count of the polar model without angular mixing. Both models used the same 1,000/24/200 train/validation/test split, relative H¹ training loss, and 500-epoch AdamW/OneCycle schedule. In each run, validation relative L² selected the checkpoint.

Darcy spectral branchTrainable parametersSeed 0 (%)Seed 1 (%)Seed 2 (%)Mean ± SD across seeds (%)
Cartesian Fourier303,9330.54550.52920.49860.5244 ± 0.0238
Polar, without cross-orientation mixing303,9330.44550.45500.42090.4405 ± 0.0176
Polar, with cross-orientation mixing (full RSNO)317,7570.44260.44410.42930.4387 ± 0.0081

Each seed-specific entry is the mean relative L² error over the same 200 test fields. The final column reports the mean and sample standard deviation across the three training seeds.

For the parameter-matched comparison, the Cartesian-minus-polar error differences were 0.1000, 0.0743, and 0.0777 percentage points at seeds 0, 1, and 2, respectively. Their mean and sample standard deviation were 0.0840 ± 0.0140 percentage points.

When does angular mixing help?

On Darcy, the input is a permeability field and the target is its pressure solution. This map imposes no prescribed rotation of input directions. Adding angular mixing reduced error by only 0.0029 percentage points at seed 0. To test its role more directly, we constructed a periodic problem with the target y(x)=u(QTx)y(x)=u(Q^T x), where QQ is a 90° rotation. Each input is a random sum of 24 Fourier modes. For example, the target transfers the coefficient at wavevector (3,4)(3,4) to (−4,3)(-4,3), preserving its amplitude, phase, and radial frequency. An angle-wise Fourier multiplier can reweight an existing coefficient, but only a cross-angle map can make this transfer within the linear spectral branch.

To isolate cross-orientation transfer, we constructed a synthetic dataset on a periodic 32×3232\times32 grid. Each input field is a sum of sine and cosine waves at 24 fixed integer wavevectors kjk_j in the first quadrant, with 3≤∥kj∥2≤73\leq\lVert k_j\rVert_2\leq7. For each sample, we independently draw the 48 sine and cosine coefficients from a standard Gaussian distribution and scale their sum by 1/241/\sqrt{24}. The target is an exact 90° rotation, v(x)=u(Q⊤x)v(x)=u(Q^\top x): each coefficient at kj=(kx,ky)k_j=(k_x,k_y) is assigned to Qkj=(−ky,kx)Qk_j=(-k_y,k_x), preserving its amplitude, phase, and radial frequency. We generate the targets analytically, so the labels introduce no rotation-interpolation error. The fixed split contains 512 training, 128 validation, and 256 test fields, with independently drawn coefficients in each split. For each of three seeds (0, 1, and 2), we train RSNO with and without angular mixing using identical initial values for shared parameters and the same minibatch order. Each model is trained for 500 epochs with AdamW, a batch size of 4, and a OneCycle schedule with a peak learning rate of 5×10−45\times10^{-4}. Training minimizes the mean per-sample relative L2L^2 error in physical units. The checkpoint with the lowest validation error is evaluated on the test set.

90° rotation results by training seed. Within each pair, the two models used the same data, initial values for their shared parameters, and minibatch order. The test columns report mean relative L² error over 256 fields. The last two columns use the same 128 validation fields and the same trained mixing model, before and after zeroing only its off-diagonal angular weights at inference.

SeedTest error, mixing off (%)Test error, mixing on (%)Validation error (%)Validation error, off-diagonal weights zeroed (%)
04.45780.66700.6603100.2717
14.47930.64470.6418100.2028
24.51640.60280.6047100.6941

Across seeds, mean test error fell from 4.4845% to 0.6382%. Enabling mixing increased the parameter count from 119,618 to 161,090. The validation intervention shows that each fitted model relied on its cross-angle connections to perform the prescribed rotation.

Thank you again for your careful review. As the discussion period draws to a close, we would appreciate any feedback on whether our responses have addressed your main concerns. Please let us know if any concerns remain unresolved or if further clarification would be helpful. Thank you for your time and consideration.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response to Reviewer eByq (part I)

We thank the reviewer for the constructive follow-up and for emphasizing methodological novelty, scientific understanding, and the objective presentation of experimental evidence. We agree that the contribution of RSNO should be assessed primarily through its proposed spectral interaction mechanism rather than marginal improvements in benchmark rankings.

We appreciate the suggestion to evaluate RSNO on more complex and realistic physical systems. Our original manuscript already includes the one-dimensional Burgers dataset from PDEBench, but we agree that this alone does not establish generalization to more challenging physical settings.

Additional evaluation on RealPDEBench. Building on our preliminary experiments, we have extended the evaluation to RealPDEBench (Hu et al., ICLR 2026), which provides paired numerical simulations and real measurements for five physical systems: cylinder flow, controlled cylinder flow, fluid–structure interaction, foil flow, and combustion.

We evaluated RSNO under all three benchmark training regimes: simulation-only training, real-data training, and simulation pretraining followed by real-data fine-tuning, covering 15 scenario–training-regime configurations.

Across all five scenarios, we use the same RSNO backbone hyperparameters and optimization settings: width 48, eight residual blocks, 16 angles, four radial locations, and 10,000 optimizer updates per training stage. No scenario-specific hyperparameter tuning was performed.

The competing-method results are taken directly from the published RealPDEBench study rather than obtained through our own reruns. Accordingly, these are comparisons with published reference results, whose training configurations differ from ours.

Dataset version and comparability. The published RealPDEBench baselines were evaluated using v2.0.0, whereas our experiments use the official v2.0.1 release (September 18, 2026; revision 0ae4426). We verified that the numerical training data and the train/validation/test splits are identical across the two versions. The v2.0.1 update affects 12 real-data Arrow shards across Cylinder (1), Controlled Cylinder (2), and Foil (9), but the precise changes within these files could not be determined. Both versions use the same normalization procedure and evaluation metrics; v2.0.1 additionally provides precomputed normalization statistics, which can be recomputed from the unchanged numerical training data using the official implementation. Thus, the data splits, normalization methodology, and evaluation protocol are consistent, although the exact real-data values in the updated shards may differ. We therefore present the published baseline results as reference comparisons rather than strictly version-matched evaluations, and do not attribute the observed performance differences solely to model architecture.

Results on measured physical systems. The following table summarizes single-round prediction errors (relative L², %) on real test data under real-data training and simulation-pretrained fine-tuning. The reference column reports the lowest published error for each regime.

Physical systemBest published reference (Real / Fine-tuning)RSNO (Real / Fine-tuning)
Cylinder7.01 / 7.285.93 / 5.92
Controlled cylinder5.55 / 5.435.38 / 5.34
Fluid–structure interaction5.83 / 5.795.74 / 5.71
Foil1.59 / 1.450.84 / 0.84
Combustion53.31 / 53.1854.13 / 53.79

Under both real-data training and fine-tuning, RSNO achieves lower reported relative L² errors than the published reference methods on all four flow scenarios. The numerical differences are particularly pronounced on the foil dataset, but their magnitude should be interpreted subject to the dataset-version and training-setting limitations above.

Across all five scenarios and three training regimes, RSNO achieves lower reported relative L² errors and RMSE values than the published FNO baseline in 14 of 15 comparisons. The exception is simulation-only training on combustion. RSNO also does not achieve the lowest error on combustion under the other two regimes.

These results provide additional evidence that the complete RSNO architecture is applicable to more realistic measured physical dynamics. However, they do not isolate the contribution of angular mixing, establish superiority under matched dataset versions, normalization statistics, and training budgets, or demonstrate long-horizon forecasting performance. The parameter-matched Cartesian/polar comparison and the rotation experiment address complementary questions about the spectral representation and its interaction mechanism.

We appreciate the reviewer's suggestion, which has motivated us to broaden the empirical evaluation while maintaining a clear distinction between architectural evidence, practical performance, and the remaining limitations.

AuthorsOpenReview ↗

Response to Reviewer eByq (part II)

Cylinder

MethodParameters (M)Simulation: Rel. L² (%)Simulation: RMSEReal data: Rel. L² (%)Real data: RMSEFine-tuning: Rel. L² (%)Fine-tuning: RMSE
U-Net23.021.650.07587.010.07007.280.0632
DeepONet3.635.920.086315.280.071315.030.0661
FNO50.425.750.07398.550.05857.800.0545
MWT2.922.400.07338.320.05408.500.0584
Galerkin84.429.410.08768.980.11969.080.0995
CNO8.018.490.07298.970.042410.130.0403
Transolver4.332.240.112118.870.109317.230.0965
DPOT-S-FT30.814.740.05139.830.05867.660.0440
DPOT-L-FT649.814.460.04868.120.03907.330.0394
WDNO104.320.530.069916.660.06899.690.0513
RSNO (ours)2.62715.010.02825.930.01175.920.0117

Controlled cylinder

MethodParameters (M)Simulation: Rel. L² (%)Simulation: RMSEReal data: Rel. L² (%)Real data: RMSEFine-tuning: Rel. L² (%)Fine-tuning: RMSE
U-Net23.013.600.01955.550.00805.430.0079
DeepONet3.642.560.054023.990.030922.840.0291
FNO50.420.070.02857.230.00977.020.0094
MWT2.915.940.02207.470.01027.460.0102
Galerkin50.818.160.02617.670.01037.480.0101
CNO8.012.090.01675.830.00815.740.0080
Transolver4.318.940.029412.000.017111.760.0168
DPOT-S-FT30.817.220.02335.980.00846.150.0085
DPOT-L-FT649.817.840.02486.030.00846.110.0085
WDNO359.818.290.02409.270.01157.520.0101
RSNO (ours)2.62411.990.01755.380.00785.340.0077

Fluid–structure interaction (FSI)

MethodParameters (M)Simulation: Rel. L² (%)Simulation: RMSEReal data: Rel. L² (%)Real data: RMSEFine-tuning: Rel. L² (%)Fine-tuning: RMSE
U-Net23.015.890.02235.830.00855.790.0084
DeepONet3.446.060.063725.020.035023.680.0332
FNO268.530.950.04268.920.01298.810.0127
MWT2.924.870.03399.100.01289.120.0128
Galerkin67.622.410.03079.160.01278.980.0124
CNO8.017.240.02417.410.01056.790.0096
Transolver4.321.190.030515.670.023614.800.0223
DPOT-S-FT41.318.860.02607.460.01057.010.0099
DPOT-L-FT673.518.980.02626.870.00996.680.0097
WDNO91.727.000.03698.170.01178.240.0116
RSNO (ours)2.62715.880.02245.740.00825.710.0082

Foil

MethodParameters (M)Simulation: Rel. L² (%)Simulation: RMSEReal data: Rel. L² (%)Real data: RMSEFine-tuning: Rel. L² (%)Fine-tuning: RMSE
U-Net23.07.450.02721.590.01001.450.0094
DeepONet3.65.650.03393.630.02223.750.0226
FNO50.45.400.02742.280.01302.060.0120
MWT2.94.920.02642.270.01332.100.0125
Galerkin50.85.120.02772.450.01422.370.0138
CNO8.04.380.02272.530.01362.060.0114
Transolver4.34.650.03453.700.02202.370.0138
DPOT-S-FT41.33.970.02211.660.01061.740.0109
DPOT-L-FT673.54.020.02291.590.01051.610.0109
WDNO358.44.900.02424.440.01621.810.0106
RSNO (ours)2.6275.120.01090.840.00190.840.0019

Combustion

MethodParameters (M)Simulation: Rel. L² (%)Simulation: RMSEReal data: Rel. L² (%)Real data: RMSEFine-tuning: Rel. L² (%)Fine-tuning: RMSE
U-Net23.372.900.035854.870.021654.030.0213
DeepONet3.585.650.040357.510.022957.230.0227
FNO67.176.060.036356.640.022656.800.0225
MWT2.975.360.036755.600.022155.490.0220
Galerkin67.691.430.040060.830.024764.210.0258
CNO8.082.740.040160.300.024858.770.0238
Transolver4.390.560.044278.250.037578.360.0375
DPOT-S-FT41.578.420.039353.490.020953.780.0211
DPOT-L-FT674.277.500.037853.310.020853.180.0206
WDNO122.7110.160.043986.050.038074.410.0314
RSNO (ours)2.66578.140.037754.130.021553.790.0213
AuthorsOpenReview ↗

Response to Reviewer eByq (part I)

We thank the reviewer for identifying the central question concerning the methodological novelty of RSNO. We agree that neither the Fourier-slice theorem, parameter efficiency, nor the combination of existing architectural components constitutes sufficient novelty by itself. Our contribution is a structured, learnable interaction between different frequency directions at a fixed frequency magnitude, together with its continuous characterization and numerical realization.

1. What is new beyond a Fourier reparameterization?

A standard FNO linear spectral branch applies a learned channel matrix independently at each Fourier frequency: Bdiagv^(ξ)=W(ξ)v^(ξ).\widehat{\mathcal B_{\mathrm{diag}}v}(\xi)=W(\xi)\widehat v(\xi). Simply rewriting ξ=rθ\xi=r\theta does not change this operation. RSNO introduces an additional angular action: Bv^(rθ)=(Ar[W(⋅,r)v^(r ⋅)])(θ),r>0,\widehat{\mathcal Bv}(r\theta)=\left(\mathcal A_r\left[W(\cdot,r)\widehat v(r\,\cdot)\right]\right)(\theta),\qquad r>0, where Ar\mathcal A_r can couple Fourier coefficients at distinct orientations on the same radial frequency shell.

The essential difference is not a change of coordinates, but a change in the dependencies between spectral coefficients. The linear RSNO spectral branch permits non-diagonal interactions within frequency shells while excluding arbitrary cross-radius coupling. It therefore provides a structured intermediate choice between frequency-wise Fourier multipliers and unrestricted frequency coupling.

This restriction defines a specific inductive bias, with parameter efficiency arising as a consequence rather than constituting the novelty itself. The distinction concerns the linear spectral branches; complete nonlinear FNO networks can also generate cross-frequency interactions through nonlinear composition.

2. Why the Radon–Fourier formulation?

The Fourier-slice theorem gives Fs[Rv](θ,r)=v^(rθ){\mathcal F_s}[{\mathcal Rv}](\theta,r)={\widehat v}(r\theta). We agree that explicitly computing the Radon transform is not indispensable. Our implementation uses FFT-based polar sampling and compensated gridding without constructing a sinogram, and equivalent structured interactions could in principle be implemented directly in Cartesian Fourier coordinates.

The value of the Radon–Fourier formulation is its geometric organization of spectral interactions: frequency magnitude determines the shell, while orientation identifies the directions to be coupled. This organization also provides a continuous analysis–interaction–synthesis framework through which we establish the exact shellwise representation, Sobolev stability, and conditional wavefront characterizations for specified angular actions.

We do not claim that Radon coordinates contain information unavailable in Fourier coordinates or that the continuous results hold exactly for the full finite-dimensional nonlinear network.

AuthorsOpenReview ↗

Response to Reviewer eByq (part II)

3. Relation to existing hybrid neural operators

We agree that combining local, multiscale, and global representations is not itself new. The distinguishing feature of RSNO is the explicit shell-constrained, non-diagonal spectral interaction, together with its mathematical analysis and FFT-based realization.

The multiscale spatial branch complements this restricted spectral operation through local and coarse-resolution processing. Thus, the methodological contribution is not simply a two-branch architecture, but a specifically structured global spectral interaction embedded in a multiscale neural operator.

4. When is angular mixing useful? What do the experiments show?

We agree that angular mixing need not benefit every PDE operator equally. Its usefulness depends on the spectral structure of the target mapping.

The new 90° rotation benchmark directly tests a case where directional transfer is required. The target v(x)=u(Q⊤x)v(x)=u(Q^\top x) transfers Fourier coefficients from wavevector kk to QkQk, preserving their magnitude. A frequency-diagonal linear multiplier cannot directly perform this transfer between distinct frequencies, whereas the angular mixer explicitly permits it.

Across three paired training seeds, angular mixing reduces the mean test relative L2L^2 error from 4.4845% to 0.6382%, an approximately 85.8% reduction (Table 7). Moreover, zeroing the learned off-diagonal angular weights after training increases validation error from approximately 0.6% to 100%, showing that the fitted models rely on these connections. The learned angular matrices also exhibit the expected 90° coupling pattern (Figure 7), providing a direct connection between the learned interaction and the geometry of the target operator.

By contrast, on Darcy, enabling angular mixing reduces the three-seed mean error only from 0.4405% to 0.4387%. These findings indicate that cross-orientation mixing is an operator-dependent mechanism: it may provide limited additional benefit when directional transfer is not essential, but can be particularly effective when the target operator requires such interactions.

The rotation experiment establishes this targeted representational capability and its empirical effectiveness. It does not establish superiority over every possible structured Cartesian Fourier operator or imply that angular mixing is necessary for general PDE learning.

Separately, the parameter-matched Darcy comparison evaluates the discrete spectral representation itself. Across three seeds, the polar branch without angular mixing achieves 0.4405% error versus 0.5244% for the Cartesian Fourier branch, with exactly 303,933 parameters in each model (Table 6). These two experiments therefore address complementary questions: the practical effect of the discrete polar representation and the additional expressive capability of cross-orientation mixing.

In summary, RSNO contributes a geometrically organized and analyzable family of shellwise spectral interactions, not a new mathematical transform or a claim of universal superiority. We appreciate the reviewer's emphasis on clarifying both the scientific motivation and the conditions under which this inductive bias is useful.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.