acceptodds
Under review as a conference paper at ICLR 2027

ConeNO: Enforcing Physical Information Flow in Neural Operators

Abstract

For finite-speed PDEs, the solution at a point depends only on data inside its backward cone. Standard neural operators need not respect this constraint: a global Fourier operator trained to low error changes its prediction when data outside the cone is edited while the reference stays fixed, and a mask applied after global mixing leaves the dependence in place. We introduce CaNO, a neural operator in which every read, including the inputs that determine its attention weights, stays inside a prescribed domain of dependence. We prove that this property is closed under composition and characterize what a Jacobian audit can and cannot certify. Under matched capacity and supervision on variable-coefficient transport and a finite-volume Burgers flow map, CaNO lowers error by 26–75% relative to convolutional and local-window Fourier operators that satisfy the same constraint. The constraint costs no in-distribution accuracy: CaNO matches an unconstrained global FNO on transport and reduces its error by 45–80% on Burgers, while having zero response to exterior interventions under which the error of global FNO grows nearly tenfold.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, hyperbolic PDEs, domain‑of‑dependence constraints, physics‑informed neural networks, numerical PDE

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_z936OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

In PDEs with finite propagation speed, the influence of a point is confined to a cone. The main contribution of this paper is ConeNO, a neural operator for fast PDE solving that improves performance by restricting attention to tokens inside this cone. The paper also proves mathematically that this restriction confines the influence to a cone-shaped region, and in particular that this still holds when the operation is applied repeatedly.

Overall, I think the problem setting and the approach are insightful. However, as described below, I am concerned that the computational cost may be a critical issue.

Critical strengths
  • The problem setting is clear, and the method fits it well. Solvers for PDEs with finite propagation speed have been studied in various ways in numerical analysis, but to my knowledge this idea, has not been introduced into neural operators before.
  • The performance is high, and the error is well controlled.
  • The method assumes that the propagation speed is known, but strictly speaking only an upper bound is needed. Also, if the purpose of a surrogate model is speed-up, it is reasonable to assume a setting where the equation and its theoretical propagation speed are known. So I do not consider this a real problem.
Critical weaknesses
  • The paper has two main technical contributions: the modified attention and the raw-value readout. I would like to see an ablation on these two components. Table 9 is a different kind of ablation.
  • The scale of the experiments seems small compared to prior work. The resolution is 64x64, and the rollout appears to go only up to T=2. This resolution is the smallest size used in the original FNO paper, and PDEBench contains datasets with higher resolution.
  • A comparison of computational cost is needed. Even though the attention range is restricted, the cost is still O(N^2) in the collocation points, so scalability is questionable. I expect the cost to be much larger than that of models built only from lightweight operations, such as FNO. ConeNO does not support irregular meshes, which is a major advantage of attention-based neural operators such as Transolver, so the computational cost issue becomes even more important. In the worst case, solving the original equation directly may be faster.
  • Some explanations are slightly inaccurate. The text right below Table 1 says "DeepONet is worse than the local models on every condition," but DeepONet is not worse for Coeff and SR2.

Once the concern about the computational cost is resolved, I am happy to raise the score.

Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

reply to Reviewer z936

We thank Reviewer z936. On a tau-step the softmax in Eq. (6) runs over J(p) only. With s_x = s_t = 4 and c_mask * tau = 1/6, the widest query on the N=256 prefix has 30 admissible keys. Rollout to T=2 is eighteen calls of that one-step cone; the counted maximum is 30 at calls 1, 6, 12, and 18. On one GPU, batch size 8, a step is 2.9 ms for ConeNO and 2.0 ms for global FNO on Burgers, 3.1 ms and 2.2 ms on transport, and 5.8 ms and 4.0 ms on the 64x64 wave map. Peak memory is 0.70 GB against 0.50 GB, 0.74 GB against 0.53 GB, and 1.10 GB against 0.80 GB. Training is 1.7 s/epoch against 1.2 on Burgers and 1.9 against 1.3 on transport. The step stays within a factor of 1.5 of global FNO, at a Burgers ID error of 1.86 against 5.04 and a transport bump error of 1.08 against 22.65.

The epoch column is train mode; the tau-step is eval mode.

TaskMethodTrain s/epochOne tau-step (ms)Rollout T=2 (ms)Peak mem. (GB)
1D transport, N=256Global FNO1.32.241.90.53
1D transport, N=256Conv. control1.62.750.20.64
1D transport, N=256ConeNO1.93.157.90.74
1D Burgers, N=256Global FNO1.22.037.40.50
1D Burgers, N=256Conv. control1.52.546.90.60
1D Burgers, N=256ConeNO1.72.953.40.70
2D wave, 64x64FNO-XY2.44.075.80.80
2D wave, 64x64Conv. control3.05.095.20.95
2D wave, 64x64ConeNO-2D3.55.8108.81.10

The rollout column is one trajectory of eighteen calls, including the prefix between calls. Isolated steps are 55.8 ms on transport, 52.2 ms on Burgers, and 104.4 ms on the wave map; the trajectories are 57.9, 53.4, and 108.8 ms. The difference is the prefix.

The references are a characteristic shift, a Godunov step, and a Fourier mode rotation. An exterior edit leaves each fixed. The cost table compares the learned maps with global FNO.

At a fixed cone radius and fixed strides, keys per query grow with N, from 30 at N=256 to 60 for a model trained from scratch at N=512, and the number of queries grows with N, so the number of admissible pairs is quadratic in the number of spatial nodes. Cone restriction reduces that count relative to mixing every pair. SR2 is separate: Appendix B.1 and Proposition 5.5 scale the residue stride with the grid, so each query keeps the physical key set of the coarse cone. The cone radius is set by tau, not by the horizon. SR2 evaluates the trained models at 512 nodes in 1D and at 128 x 128 on the wave map. Global FNO has no set J and mixes all 512 nodes. The convolutional control has no set J.

ModelParams (k)IDBumpCVSmax keys in J
Global FNO258.91.15+/-0.0719.74+/-4.280.0531+/-0.0071—
Conv. control264.31.95+/-0.162.28+/-0.470—
ConeNO264.91.18+/-0.071.06+/-0.21060

ID matches global FNO (1.18+/-0.07 and 1.15+/-0.07). On localized inputs ConeNO stays at 1.06+/-0.21 and global FNO moves to 19.74+/-4.28. CVS is 0 against 0.0531+/-0.0071.

On transport, DeepONet is above both locality-constrained baselines on ID (8.71 vs. 2.17 and 2.08), Bump (9.49 vs. 1.85 and 2.56), and T=2 (9.35 vs. 3.61 and 3.72). On Coeff it is 8.55, below local-window FNO at 12.27 and above the convolutional control at 7.98; on SR2 it is 8.71, below local-window FNO at 9.91 and above the convolutional control at 2.07. ConeNO is lowest among the locality-constrained models on Coeff and SR2, at 5.92 and 1.25, and on SR2 it lies within one seed SD of global FNO (1.25 against 1.24).

The two-component ablation is Appendix D.3.2, Tables 10 and 11 (five seeds, 80 epochs, matched widths). Every row has CVS 0.

Transport

VariantIDBumpCoeff.SR2T=2
Conv. control2.08+/-0.172.56+/-0.617.98+/-0.492.07+/-0.173.72+/-0.33
Latent attention1.66+/-0.091.58+/-0.186.85+/-0.241.70+/-0.072.90+/-0.12
Raw readout1.48+/-0.072.02+/-0.227.90+/-0.401.48+/-0.062.48+/-0.10
ConeNO1.26+/-0.081.08+/-0.265.92+/-0.421.25+/-0.072.14+/-0.13

Burgers

VariantIDBumpSR2T=2
Conv. control4.02+/-0.564.65+/-1.315.82+/-0.506.09+/-0.93
Latent attention2.35+/-0.141.70+/-0.164.20+/-0.163.30+/-0.15
Raw readout3.25+/-0.182.65+/-0.225.05+/-0.184.60+/-0.25
ConeNO1.86+/-0.181.15+/-0.133.64+/-0.242.74+/-0.20

On transport the readout carries ID, SR2, and T=2; latent attention carries Bump and Coeff. On Burgers each component alone improves every column over the control, and ConeNO improves further over either alone. On the wave map, SR2 is 5.27 against 12.64, T=2 is 4.18 against 6.95, and CVS is 0 against 0.318 (Table 5).

Thank you for the detailed response.

The ablation addressed my first comment for the 1D tasks, and the explanation of the DeepONet addressed the forth.

However, my main concern about the computational cost remains. While the reported time is not slow, it is because of the small sizes of the benchmark tasks. For practical sizes, the computational cost of O(N^2) is not negligible. The attention-based methods such as OFormer, GNOT and Transolver can handle irregular meshes, where interactions between nodes are unavoidable and the computational cost is justified. In contrast, ConeNO does not provide a clear advantage that justifies its computational cost. Moreover, ConeNO is restricted to periodic uniform grids.

I consider that the idea is sound and the advantages are established, and I appreciate the added ablation. My main concern is the computational cost against the benefit. I therefore keep my score. If the other reviewers endorse acceptance, I will not oppose it.

AuthorsOpenReview ↗

reply to Reviewer z936

We thank the reviewer for confirming the ablation and the DeepONet correction.

The quadratic count arises only when the key stride is held fixed in grid units. Under Proposition 5.5 the stride scales with the grid, so each query keeps the physical key set of the coarse cone. The keys per query stay constant and the admissible pairs grow linearly with the number of nodes M. SR2 is this evaluation: 512 nodes in 1D and 128 x 128 on the wave map, from the trained checkpoint, with tap offsets, residue strides, and raw-value read offsets multiplied by the refinement factor f (Appendix B.1), without retraining. The 60-key model in the previous response is a separate run, trained from scratch at M=512 with the stride fixed at 4. That is the quadratic setting reported there. The cone radius is set by the one-step interval tau, and the same mask construction is reused at every rollout step.

Global FNO mixes the grid by an FFT and a Fourier multiplier, so the key count and the node count are not a FLOP ratio. On one GPU, batch size 8, a step is 3.1 ms for ConeNO and 2.2 ms for global FNO on transport, 2.9 ms and 2.0 ms on Burgers, and 5.8 ms and 4.0 ms on the 64 x 64 wave map. ConeNO keeps the exterior response at 0. Table 1 bumps are 1.08+/-0.26 against 22.65+/-5.35 for global FNO, and Table 2 in-distribution error is 1.86+/-0.18 against 5.04+/-0.18 for global FNO. A local convolution also has exterior response 0 and is less accurate on the ablation.

The cone on irregular coordinates does not use the periodic index set of Eq. (16). A node is admissible when it lies inside Eq. (1), at distance at most c_mask times the elapsed time, and a delayed tap is kept only when its length satisfies Eq. (5). OFormer, GNOT, and Transolver do not enforce a receptive-field bound, so the comparison here is unrestricted full attention on the same nodes. We use one frozen set of 1024 nodes in the unit square. This run tests invariance on irregular coordinates. It does not subsample, and it is separate from the Proposition 5.5 configuration. The widest query reads 519 of the 1024 nodes, the ones that lie in one fixed physical disk. The reference at each node is the sample of the translated continuous field, not a spectral reconstruction of the nodal values, and an edit outside the cone leaves that sample unchanged. The convolutional control uses the same cone and has no attention keys, which is why its cell is blank.

The rows below are separate from Table 4. Relative L2 x 100, mean +/- SD over five seeds. Exterior change is the prediction change at the query nodes after perturbing inputs outside their cones. Reference change is the change in the exact solution.

ModelParams (k)IDBumpExterior changeReference changeNodes readOne step (ms)Peak mem. (GB)
Global attention140.62.87+/-0.2323.64+/-2.9113.28+/-1.86010244.60.91
Conv. control, coordinate cone139.83.54+/-0.2810.73+/-1.2400—1.80.42
ConeNO, coordinate cone140.42.16+/-0.167.38+/-0.67005193.30.61

ConeNO is lowest on ID and on bumps. Both cone-restricted models have zero exterior change. Global attention changes by 13.28+/-1.86 under the same edit, and the reference change is 0.

Reviewer_cA2dOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces ConeNO, a neural operator enforces prescribed domains of dependence for finite-speed hyperbolic dynamics. It combines delayed compact convolutions, cone-restricted latent attention, and a query-conditioned raw value readout, constraining both the aggregated information and the inputs used to generate adaptive weights. Main contributions are the proposed architecture which designed to preserve the dependency constraint under decomposition, and analysis which connect Jacobian audits to finite interventions.

Validations covers variable-coefficient transport, a Burgers finite volume flow map, 2D advection, and wave dynamics, using closely matched parameter counts and 5 seeds in the main comparisons. Evaluation includes prediction accuracy, distribution shifts, spatial refinement, autoregressive rollout, and exterior-dependency audits. The authors report consistently lower errors than the tested locality-constrained baselines and zero measured exterior sensitivity under the specified discrete masks. Performance relative to global FNO is task-dependent, with substantial gains on Burgers and competitive transport accuracy.

Critical strengths
  1. A clear and mathematically justified information-dependence guarantee. ConeNO constrains the complete computational pathway, including the inputs used to generate attention weights, gates, and normalization factors. The composition result establishes that these restrictions persist through network depth under explicit assumptions.

  2. Encouraging accuracy and robustness under meaningful controlled comparisons. The main experiments use 5 seeds and closely matched parameter counts, including baselines also designed for locality restriction. On Burgers, ConeNO achieves 1.86% ID error versus 2.85% for local-window FNO and 5.04% for global FNO. On transport, it retains comparable ID accuracy to global FNO while remaining invariant to exterior interventions that substantially increase the latter’s error.

Critical weaknesses

Limited validation of the motivating many-query applications. The introduction motivates ConeNO through inverse design and control, but the experiments evaluate forward prediction and dependency preservation without demonstrating downstream benefits. The Jacobian audits establish the absence of forbidden exterior sensitivities, but do not verify the accuracy of physically relevant sensitivities within the admissible domain. No parameter-recovery or design-optimization experiment establishes whether the proposed constraint improves these workflows. Furthermore, the absence of runtime and amortization measurements leaves the computational advantage for repeated queries unverified. This limits the demonstrated practical significance: the reported improvements have not been connected to the accuracy, reliability, or efficiency of the applications motivating the work.

Questions for authors
  1. Comparison with adaptive operators and attribution of gains. Can the authors establish whether the reported improvements persist against a representative attention-based operator restricted to the same admissible inputs under comparable training resources? The current convolutional control also differs from ConeNO in both latent attention and raw-value readout; separating these components would help determine which mechanism produces the gains. This evidence would materially affect whether the experiments establish the value of ConeNO’s specific architecture.

  2. Scope of the many-query application claims. Are inverse design and control presented solely as motivating applications, or does the paper claim demonstrated benefits for these workflows? If downstream benefits are claimed, what evidence supports accurate relevant sensitivities or a computational advantage after accounting for training and repeated inference? If these applications are only motivations, please clarify that scope; downstream demonstrations would then be suggestions for future work rather than required validation of the present contribution.

2 comments
AuthorsOpenReview ↗

reply to Reviewer cA2d

We thank Reviewer cA2d. Under one cone, accuracy follows how admissible values are mixed. The convolutional control replaces each latent-attention update by a pointwise MLP (264.3k). Latent cone attention restores that update (264.3k). The raw-value row adds the 573-parameter readout (264.9k). ConeNO uses both. Five seeds, 80 epochs, the widths of Tables 1 and 2. Table 9 uses latent attention only.

Transport (relative L2 x 100):

VariantParams (k)IDBumpCoeff.SR2T=2CVS
Conv. control (compact + MLP, no raw)264.32.08+/-0.172.56+/-0.617.98+/-0.492.07+/-0.173.72+/-0.330
Latent cone attention, no raw readout264.31.66+/-0.091.58+/-0.186.85+/-0.241.70+/-0.072.90+/-0.120
Raw-value readout, no latent attention264.91.48+/-0.072.02+/-0.227.90+/-0.401.48+/-0.062.48+/-0.100
ConeNO (both)264.91.26+/-0.081.08+/-0.265.92+/-0.421.25+/-0.072.14+/-0.130
Cone-masked MHA, no compact stack264.91.88+/-0.091.72+/-0.209.10+/-0.352.50+/-0.143.20+/-0.120

Burgers flow map:

VariantParams (k)IDBumpSR2T=2CVS
Conv. control (compact + MLP, no raw)264.34.02+/-0.564.65+/-1.315.82+/-0.506.09+/-0.930
Latent cone attention, no raw readout264.32.35+/-0.141.70+/-0.164.20+/-0.163.30+/-0.150
Raw-value readout, no latent attention264.93.25+/-0.182.65+/-0.225.05+/-0.184.60+/-0.250
ConeNO (both)264.91.86+/-0.181.15+/-0.133.64+/-0.242.74+/-0.200
Cone-masked MHA, no compact stack264.92.70+/-0.132.10+/-0.165.45+/-0.203.85+/-0.160

On transport the readout carries the larger reduction on ID (2.08 to 1.48), SR2 (2.07 to 1.48), and T=2 (3.72 to 2.48); latent attention carries Bump (2.56 to 1.58) and Coeff (7.98 to 6.85). On Burgers, each component alone improves every column over the control, and combining them improves further over either component alone. ConeNO is the lowest entry on every column. The last row is the attention operator on the same admissible inputs, at the same parameter count and training budget, without the delayed-compact stack. It leaves Coeff at 9.10 and SR2 at 2.50, where ConeNO is at 5.92 and 1.25. ConeNO is lower than this model on every column. Every cone-restricted row has CVS 0.

Inverse design and control are motivating applications in the introduction. The measurements that bear on them are the exterior audit, the interior derivatives, and the interior-field recovery. In Table 3, an exterior edit that leaves the reference fixed leaves ConeNO's query error at 0.0068, while global FNO on transport moves from 0.0067 to 0.0636. Section 5 calls CVS the Jacobian form of that invariance; Proposition 5.4 is the finite-edit bound.

On the trained Table 1 and Table 2 checkpoints, a bump of amplitude 1e-3 inside the cone gives a directional derivative. The cell is the relative error against the reference, x 100.

TaskModelInterior derivative error
TransportGlobal FNO6.47+/-0.82
TransportConv. control9.53+/-1.16
TransportConeNO5.76+/-0.68
BurgersGlobal FNO13.84+/-1.57
BurgersConv. control11.26+/-1.48
BurgersConeNO6.18+/-0.74

ConeNO is lowest on both tasks, at 5.76+/-0.68 on transport and 6.18+/-0.74 on Burgers.

Frozen Table 1 checkpoints recover the interior field so that the query matches the reference. Exterior nodes stay at the Table 3 donor, and a second donor repeats the recovery. Recovery error is the relative L2 of the optimized interior against the true interior, x 100.

ModelRecovery errorQuery errorChange under a new exterior
Global FNO21.86+/-3.741.48+/-0.2715.72+/-2.86
Conv. control13.94+/-2.152.18+/-0.360
ConeNO8.06+/-1.421.37+/-0.260

ConeNO recovers the interior at 8.06+/-1.42, against 21.86+/-3.74 for global FNO and 13.94+/-2.15 for the convolutional control. Its query error is 1.37+/-0.26, against 1.48+/-0.27 and 2.18+/-0.36. A new exterior donor leaves that interior unchanged for ConeNO and the convolutional control. Global FNO's recovered interior moves by 15.72+/-2.86.

A tau-step is 2.9 ms against 2.0 ms for global FNO on Burgers, 3.1 ms against 2.2 ms on transport, and 5.8 ms against 4.0 ms on the 64x64 wave map, with peak memory 0.70/0.50, 0.74/0.53, and 1.10/0.80 GB. Training is paid once: 1.7 s/epoch against 1.2 on Burgers and 1.9 against 1.3 on transport. On Burgers that step has ID error 1.86 against 5.04. The full cost table is in the response to Reviewer z936. On the wave map, SR2 is 5.27 against 12.64, T=2 is 4.18 against 6.95, and CVS is 0 against 0.318 (Table 5).

Thanks for the detailed ablation study. It provides evidences for the benefits of each components of the proposed model.

However, since the comparison with a representative attention-based operator restricted to the same admissible inputs under comparable training resources is still missing, it is difficult to fully decide the value of the proposed structure. The ablation studies are internal control while this proposed one is external.

Additionally, even though the authors clarify that the inverse design and control are only motivating applications, an example inverse problem with analysis of accuracy, robustness and computational costs will materially raise the value of this work.

Overall, I think this work has a very good starting point. But given the proposed experiments and analysis missing, I am not confident enough to accept it in current form.

After careful consideration, I decide to raise to 3, with a low confidence though.

P.S. A kind recommendation for drafting a response: compared to the number and results, the insights and conclusions behinds them are more important, so highlights of those with pointers to the reviewers' points would be more straightforward and effective.

Reviewer_UKTROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The work introduces an important inductive bias in the architecture of neural operators. They propose to restrict the flow of information from the part of the domain that remains outside the region of influence, or what they call the Domain of dependence. The proposed idea is intuitive, and the restriction is primarily obtained by restricting the attention span. Although limited, the author empirically showed the effectiveness of the proposed approach.

Critical strengths

The paper's primary strength is its inductive bias design. The structure restricts information flow from outside the domain of dependence and is intuitive and effective, as empirical experiments show.

Critical weaknesses
  • While such a structure is not unknown in this community, imposing such a structure is difficult in general PDEs. a. First, the J(P)J(P) (sample token inside the cone) is difficult to construct. This can be due to not knowing the velocity bound 'c ', which can also vary with time. For recurrent architectures, they need to be updated over time. At the same time, the neural operators, which are trained on multiple physical parameters (such as multiple viscosities), J(P)J(P) cannot be constructed effectively without knowing the physical parameters. So, even though the idea sounds very intuitive, it is prohibitive to apply to general setups. This is also evident from the simpler experimental setups used by the authors.

  • Proposing 5.3 connects the bias in the dataset with the spurious solution learned by the neural operator. Such situation often arises for high-dimensional data such as 3d PDEs or multi-variable PDEs (such as active matter). Experiments does not cover such cases.

Questions for authors
  1. Line 046: What is "Delayed compact transport?
  2. Should the 'Domain of dependence' be better understood as 'Minimal Domain of dependence'. Cause my understanding is that the whole domain can trivially be taken as the domain of dependence of any PDE. Thus, it contributes nothing significant.
  3. Line 162: the concept of time lag is not clear from here. It would be better to clarify it.
  4. How is J(p)J(p) constructed and maintained for all the points in the domain?
  5. What is 'r' in Eqn. 5?
  6. Did the authors generate their own data or use it from some standard, well-used benchmark like the " Well" dataset?
4 comments
AuthorsOpenReview ↗

reply to Reviewer UKTR

We thank Reviewer UKTR. Table 1 is spatial variable-coefficient transport. ConeNO and global FNO match in distribution (1.26+/-0.08 and 1.24+/-0.07). On localized inputs the global model moves to 22.65+/-5.35, while ConeNO stays at 1.08+/-0.26. On spatial modes outside training, global FNO is at 19.32+/-0.85 and ConeNO at 5.92+/-0.42. Proposition 5.3: on half-periodic data a Fourier multiplier has zero training error and, for small t, influence outside the cone. The Bump columns of Tables 1, 2, and 4 measure it. On 2D advection the global model reaches 26.41+/-2.83 on bumps and ConeNO-2D stays at 6.57+/-0.58, with CVS 0.411 against 0.

Separate from Table 1, scalar c(t) is drawn from [0.75, 1.25], c_mask = 1.5, and enters only as an input channel. On 128 fields the Table 3 edit leaves the reference at relative L2 x 100 of 0, including unseen time modes. J(p) in Eq. (16) is the same 30-key set for slow, fast, and unseen draws, and at rollout calls 1, 6, 12, and 18. Appendix C: nested cone support is closed under rollout.

ModelParams (k)IDBumpTimeCVS
Global FNO258.91.21+/-0.0818.42+/-4.158.06+/-0.940.0418+/-0.0054
Local-window FNO258.92.16+/-0.141.86+/-0.316.47+/-0.720
Conv. control264.32.04+/-0.162.38+/-0.475.12+/-0.580
ConeNO264.91.24+/-0.081.14+/-0.233.46+/-0.380

ID matches global FNO. On bumps ConeNO stays at 1.14 and global FNO moves to 18.42. On unseen time modes ConeNO is lowest, at 3.46. The cone-restricted models have CVS 0.

What the mask needs is a speed bound, not the speed field. Eq. (16) does not read c(x) or c(t). The bound is the maximum characteristic speed, the number a CFL condition uses. Experiments use 1.5 for speeds in [0.75, 1.25]; a larger bound only enlarges the index set. The realized speed is an input channel, and the index set comes from the query and the bound.

This covers the active-matter example, two-field run-and-tumble. tau_r is drawn from {0.5, 1, 2} as an input channel. The mask uses c = 1 only, c_mask = 1.5. The reference exterior response is 0 at each tau_r, with the same 30-key index set.

ModelParams (k)IDBumpCVS
Global FNO258.91.32+/-0.0916.84+/-3.270.0364+/-0.0047
Local-window FNO258.92.18+/-0.141.87+/-0.320
Conv. control264.32.07+/-0.162.43+/-0.410
ConeNO264.91.36+/-0.091.16+/-0.220

ID matches global FNO. ConeNO is lowest on bumps, at 1.16 against 16.84. CVS is 0 for the cone-restricted models.

Viscosity is not an input to J(p). One Burgers operator is trained with viscosity drawn from {0, 1e-4, 1e-3} as an input channel, under the Table 2 mask. The reference is the explicit finite-volume step with the viscous term inside that step. Its stencil lies inside the c_mask = 1.5 cone, and the zero exterior response is a property of this discrete map. The Table 3 edit measures reference response 0. ConeNO has 30 keys; global FNO reads all 256 nodes.

ModelParams (k)Test nuIDBumpCVSmax keys in J
Global FNO258.905.18+/-0.215.86+/-0.370.0096+/-0.0011—
Global FNO258.91e-45.11+/-0.195.74+/-0.330.0091+/-0.0010—
Global FNO258.91e-34.57+/-0.175.08+/-0.290.0114+/-0.0013—
ConeNO264.902.05+/-0.181.30+/-0.14030
ConeNO264.91e-42.00+/-0.181.28+/-0.14030
ConeNO264.91e-31.70+/-0.151.05+/-0.12030

ConeNO is lower than global FNO on ID and Bump at each viscosity, and CVS is 0. Grid-aligned advection at 16 cubed uses the same one-cell roll, with c_mask = 1.5.

ModelParams (k)IDBumpT=2CVS
FNO-XYZ158.22.68+/-0.2328.36+/-3.1426.08+/-1.940.438+/-0.027
Conv. control157.63.24+/-0.2710.47+/-1.288.14+/-0.910
ConeNO-3D158.22.17+/-0.157.18+/-0.636.05+/-0.480

ConeNO-3D is lowest on ID, Bump, and T=2. On bumps, 7.18 against 28.36, and CVS is 0 against 0.438.

Delayed compact transport (Section 4, Eq. 14) reads offsets |j| <= r only at lag d_l. In Eq. (5), r = 6 in 1D and r = 1 in 2D, with (r * d_l) / N <= c_mask * d_l * Delta t. J(p) in Eq. (16) is fixed before the forward pass from the query index, c_mask, and the strides. The whole domain satisfies Definition 3.1 and excludes no read. The operator uses the backward cone (1), of radius 1/6 at tau = 1/9, and a larger bound is a valid superset. The name is Domain of dependence. Fields are generated in Appendix D, not The Well. A mask after global mixing does not build this path (Proposition 5.2). Without the compact stack, transport Coeff is 9.10 and SR2 is 2.50, where ConeNO is 5.92 and 1.25 (Appendix D.3.2).

Reviewer_UKTROpenReview ↗

Response

Thanks for the response and the additional results on the 3D dataset.

The unanswered question is: how do we calculate the velocity bound from the dataset?

The issue I was pointing to is: For optimal design, i.e., if we want my model not to look at any spurious feature, meaning not looking into the region outside of the'minimal' domain of dependence, the index set needs to vary with time and position. This is because the velocity at a point changes with time and position. And calculating this can be time-consuming. As I understand it, the proposed design uses only a fixed bound, thereby avoiding this expensive computation, but the possibility of learning spurious features remains.

AuthorsOpenReview ↗

reply to Reviewer UKTR

We thank the reviewer. The bound is the maximum characteristic speed of the family, the speed that enters a CFL condition. For the transport family the equation is q_t + c(x) q_x = 0, and the data-generation protocol draws 0.75 <= c <= 1.25. The mask uses 1.5, a conservative number above that maximum. Equation (16) builds the index set from the query index, this bound, and the strides. It does not read c(x) or c(t).

The exterior edit checks this choice. Nodes outside the cone of radius 1.5 are perturbed, and the reference at the query is unchanged. That reference change is 0 on the reported tasks, including time modes of c(t) outside the training draw. Slow, fast, and unseen draws share the same 30-key index set.

The local speed varies with position and time, so the characteristic segment varies with the query. For this equation the value at a query depends on the initial datum at the foot of F and on c along the segment from that foot to the query. Every such segment lies inside Equation (1) with speed 1.5. The index set also contains the nodes between the segment and that cone. Equation (16) is this superset, written once from the bound, before the forward pass.

Proposition 5.3 is one construction. On half-periodic linear advection, a Fourier multiplier has zero training error, and for small t its influence lies outside the cone. The Bump columns are localized inputs under the fixed bound. Global FNO reaches 22.65+/-5.35 on Table 1, 18.42+/-4.15 on scalar c(t), and 28.36+/-3.14 on the 16^3 grid, where the speed has magnitude 1 and the mask is again 1.5. ConeNO stays at 1.08+/-0.26, 1.14+/-0.23, and 7.18+/-0.63. CVS on edits outside this prescribed cone is 0.

On the Table 1 family the characteristic is known from F. The characteristic set keeps, at every delayed tap and every attention key, the nodes in the one-cell tube around that segment. Equation (16) is not used for that model. The annulus edit changes nodes inside the cone of radius 1.5 and outside the tube, with the same donor construction as Table 3. Relative L2 x 100, mean +/- SD over five seeds. Mean keys are the average number of attention keys per query.

ModelParams (k)IDBumpAnnulus changeReference changeMean keys
Global FNO258.91.24+/-0.0722.65+/-5.3511.64+/-1.280—
ConeNO, fixed cone264.91.26+/-0.081.08+/-0.260.16+/-0.05013.5
ConeNO, characteristic set264.91.31+/-0.091.17+/-0.23000.80

The mean-key column counts attention keys. The fixed cone averages 13.5 per query, and its widest query remains the 30-key set above. The characteristic tube meets that stride lattice at 0.80 keys per query. ID is 1.26+/-0.08 for the fixed cone and 1.31+/-0.09 for the characteristic set. Bumps are 1.08+/-0.26 and 1.17+/-0.23. The annulus lies inside the cone of radius 1.5, so this column is separate from the exterior CVS of 0. The reference stays at 0. Global FNO changes by 11.64+/-1.28, and the fixed cone changes by 0.16+/-0.05. The characteristic set does not read the edited nodes, and its change is 0.

Reviewer_UKTROpenReview ↗

Response

Thanks for the response. Please explicitly mention these discussions in the paper. I have revised my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.