acceptodds
Under review as a conference paper at ICLR 2027

Haar Paraproduct Neural Operator

Abstract

Physical-field prediction involves interactions between regional means and internal variations. Bounding individual coupling coefficients, however, need not control their accumulated effect across a hierarchy. We introduce the Haar Paraproduct Neural Operator (HPNO), which learns these interactions through two input-conditioned bounded functions, called symbols. Their Haar coefficients define a mean-to-detail paraproduct and a detail-to-mean adjoint paraproduct on a weighted binary hierarchy, while a bounded Haar multiplier reweights existing components. Regional orthogonality converts pointwise symbol bounds into joint control of squared transfer coefficients within every subtree. This yields a weighted bound for each fixed-input kernel action, independent of hierarchy size and depth. We also characterize both directional responses across arbitrary tree cuts and express their exact norms through regional symbol variation. HPNO achieves the lowest reported relative errors among the compared methods on all six PDE benchmarks spanning regular grids, structured meshes, and point clouds, with competitive results on two unstructured computational fluid dynamics benchmarks. Ablations on Darcy and Elasticity support bidirectional coupling and bounded-symbol construction. Paired Darcy tests further show lower regional response errors than HNO and LinearNO under permeability rearrangements that preserve regional input means.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operator, Scientific Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_sDXGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes HPNO, a neural operator whose kernel lives on a weighted binary Haar tree. Each kernel action combines a bounded Haar multiplier with a mean-to-detail paraproduct and a detail-to-mean adjoint paraproduct. The coefficients of the two paraproducts are the Haar coefficients of two input-conditioned, tanh-bounded symbols. The authors show that the cross-cut transfer norms equal the regional standard deviation of the symbol. Using Carleson embedding, they also show that each fixed-input kernel action is bounded by a constant that does not depend on tree size or depth. Experiments cover six standard PDE benchmarks and two CFD datasets, plus ablations on Darcy and Elasticity and a paired Darcy test where permeability is rearranged without changing regional means.

Critical strengths

I like the core idea. Generating all coefficients in each direction from a single bounded symbol gives joint subtree control almost for free. It also makes the mean-to-detail and detail-to-mean responses easy to interpret, and the two-pass evaluation keeps the cost linear. Fig. 4b makes the point well: when the coefficients are predicted directly, they break the subtree bound. The results are also consistently strong across grids, meshes and point clouds. The ablation in Table 3, run inside the same pipeline, does support using both directions, untied symbols and bounded-symbol generation.

Critical weaknesses

My main concern is that I can't tell how much of the improvement comes from the kernel itself. Appendix E.2 says the baseline numbers are taken from earlier papers. HPNO, however, has extra task-specific machinery: a derivative loss on Darcy, local convolution, full-resolution refinement and a coarse-mean path. Table 3 makes this concrete. The multiplier-only model, which has no paraproducts at all, already gets 0.00345 on Darcy, better than every baseline in Table 1 (HNO is 0.0045). The paraproducts then take it from 0.00345 to 0.00311. So most of the gap to prior work seems to come from the rest of the predictor. The paired Darcy result has the same problem: a 34–40% lower response error could simply reflect HPNO's lower overall error.

My second concern is that the theory is largely classical and not clearly linked to accuracy. Theorem 1 is essentially the standard finite-tree Carleson embedding (Treil, 2013) applied to bounded symbols. It bounds one linear action at a fixed input. It says nothing about the stability or generalization of the full model, and the paper doesn't explain why this bound should lead to lower error. Also, on the CFD tasks the tree is built over slice slots, so the physical mean-versus-detail motivation doesn't really apply there.

Minor issues

Several citations failed to compile and show up as question marks: FNO, RNO and GEO-FNO in Tables 1 and 2, "(?Li et al., 2023b)" in Section 5 and Appendix E.2, and "benchmark interfaces of ?" in Appendix E.3.

In Table 1, WNO's Plasticity entry is just a dash with no explanation. In Definition 1, the difference between λ₀ (the root mean) and λ_S ("the root detail when present") is easy to misread and should be spelled out.

Naming is inconsistent: the dataset is "Shape-Net Car" in Table 2 but "ShapeNet Car" in the text, and the same variant is "Normalized-direct" in Fig. 4 but c̄ in Table 3.

Section 5.2 presents Normalized-direct staying under the bound as a finding, but Appendix F.2 shows the bound is enforced by construction, so it isn't evidence of anything. In Section 2, "Concurrently" is used for work from 2019 to 2023, which isn't concurrent. Finally, the main text doesn't report parameter counts, runtime or memory against the baselines.

Questions for authors

Could you put the HNO and LinearNO kernels inside your own predictor, with the same auxiliary paths, loss and training budget, and compare on Darcy, Elasticity and Plasticity? That would show what the kernel itself contributes, and it could change my score. In the paired Darcy test, how does the multiplier-only ablation do? What do the response errors look like after normalizing by each model's overall error? Do you have evidence that the bounded construction helps beyond the small gains in Table 3, for example with deeper trees, larger N or a change in resolution?

5 comments
AuthorsOpenReview ↗

Response to Reviewer sDXG (1/4): Isolating Kernel Contribution and Multiplier-Only Model (W1)

Thank you for the detailed review and for recognizing the bounded-symbol construction, two-pass evaluation, and architectural ablations. We have revised the manuscript to directly test the kernel's contribution and clarify the theory and response diagnostics.

We organize our response into four posts:

  1. Post 1/4 (W1): Controlled kernel replacements and the multiplier-only model.
  2. Post 2/4 (W1 & Questions): Paired Darcy responses and normalization by overall field error.
  3. Post 3/4 (W2 & Question): Theoretical contribution, larger hierarchies, and CFD support interpretation.
  4. Post 4/4 (Minor Issues): Computational costs and the remaining notation and presentation issues.

All references refer to the revised manuscript. The new kernel-replacement experiments are in Table 3; the original architectural ablations are now Table 4. W1 and W2 below denote your two main concerns.


[W1] Isolating the kernel contribution

  • "My main concern is that I can't tell how much of the improvement comes from the kernel itself."
  • "Could you put the HNO and LinearNO kernels inside your own predictor, with the same auxiliary paths, loss and training budget, and compare on Darcy, Elasticity and Plasticity? That would show what the kernel itself contributes, and it could change my score."

1. The requested comparisons under a shared predictor architecture. We performed these comparisons on all three requested tasks and extended them to Transolver++ and all eight benchmarks. Within each host predictor, we hold the feature interface, representation/reconstruction maps, residual and auxiliary paths, loss, data split, training budget, and final-epoch evaluation rule fixed. Every variant is retrained.

Table 3(a) gives the following relative ℓ2\ell_2 errors within the HPNO predictor:

KernelDarcyElasticityPlasticity
HNO0.00400.00360.0008
LinearNO0.00430.00430.0009
Transolver++0.00420.00450.0009
HPNO0.00310.00310.0004

We also performed the reverse replacements. Table 3(b) reports native-kernel error →\to HPNO-kernel error within each baseline predictor:

Host predictorDarcyElasticityPlasticity
HNO0.0045 →\to 0.00360.0037 →\to 0.00330.0009 →\to 0.0006
LinearNO0.0050 →\to 0.00390.0050 →\to 0.00370.0011 →\to 0.0007
Transolver++0.0049 →\to 0.00380.0052 →\to 0.00380.0011 →\to 0.0007

The improvements in both directions show that the surrounding task-specific pipeline alone does not explain the gains. Table 3 also reports the remaining tasks, including CFD metrics where the effect depends on the host representation and output quantity.

2. What the multiplier-only result establishes. The multiplier-only ablation retains a learned, input-conditioned Haar kernel: M\mathcal{M} reweights the root mean and every Haar detail. Removing the two paraproducts therefore does not remove the kernel or its multiscale representation. The difference between multiplier-only HPNO and literature-reported HNO cannot, by itself, be assigned to the surrounding predictor.

The new controls separate these effects. With the HPNO predictor architecture held fixed, alternative kernels obtain Darcy errors of 0.0040 to 0.0043 (Table 3(a)), the Haar multiplier alone obtains 0.00345, and adding both directional transfers gives 0.00311 (Table 4). On Elasticity, alternative kernels obtain 0.0036 to 0.0045, while adding both transfers to M\mathcal{M} lowers error from 0.00364 to 0.00314. Table 3 tests the kernel choice within each host architecture, while Table 4 tests directional coupling within the Haar kernel.

Revisions in the manuscript: Section 5.1 and Table 3 (reciprocal kernel replacements); Appendix E.2, "Controlled kernel comparisons" (interfaces and training controls); Section 5.2, Table 4, and Appendix F.2 (multiplier-only interpretation and ablations).

AuthorsOpenReview ↗

Response to Reviewer sDXG (2/4): Paired Darcy Responses and Field-Error Normalization (W1 & Questions)

[W1 & Questions] Paired Darcy responses and field-error normalization

  • "In the paired Darcy test, how does the multiplier-only ablation do?"
  • "What do the response errors look like after normalizing by each model's overall error?"

1. Multiplier-only responses under the same paired-input protocol. We added multiplier-only HPNO to Figure 5 and report its results at every rearrangement level in Table F.1. Permeability is rearranged within fixed regions while preserving their weighted means; the domain, forcing, and boundary conditions remain unchanged. Frozen predictors estimate the solution change Δu^=u^+−u^−\Delta\hat{u}=\hat{u}^{+}-\hat{u}^{-} against Δu=u+−u−\Delta u=u^{+}-u^{-}, without response-specific training.

The response score is the weighted RMS error of regional mean changes. Full HPNO has the lowest aggregate response error at every tested rearrangement level (Figure 5(b), Table F.1). At r=1.0r=1.0, response RMS is 3.0556 for full HPNO, 3.5034 for M\mathcal{M}-only HPNO, 4.6046 for HNO, and 5.0841 for LinearNO, in units of 10−510^{-5}. Across the 50 matched pairs at this level, HPNO has lower error in 31 cases (62.0%) against M\mathcal{M}-only, 39 (78.0%) against HNO, and 45 (90.0%) against LinearNO (Figure 5(c)).

2. Normalization by each model's overall field error. We report the requested normalization in Appendix F.3 and Table F.2: Q=Eresp,pool/EfieldQ = E_{\mathrm{resp,pool}} / E_{\mathrm{field}}. Here, EfieldE_{\mathrm{field}} is the pooled field RMS over both members of the same input pairs, and Eresp,poolE_{\mathrm{resp,pool}} is the pooled regional response RMS. Both use the same decoded physical fields and node weights. The results pool 50 base cases and ten rearrangement levels; both RMS columns below are in units of 10−510^{-5}, and QQ is dimensionless.

ModelField RMSRegional response RMSQQ
HNO5.39773.49720.6479
LinearNO6.10633.70960.6075
HPNO (M\mathcal{M}-only)4.89772.75850.5632
HPNO3.87022.35030.6073

Full HPNO has the lowest pooled field and regional response errors. Its normalized ratio is lower than HNO's, close to LinearNO's, and higher than multiplier-only HPNO's. Adding the directional terms reduces field RMS from 4.8977 to 3.8702 and response RMS from 2.7585 to 2.3503; QQ increases from 0.5632 to 0.6073 because field error decreases proportionally more.

We have revised the interpretation accordingly. The normalized results do not establish a uniformly smaller response-to-field error ratio. The paired experiment demonstrates improved absolute prediction of rearrangement-induced changes within the same predictor; the reciprocal replacements in Table 3 provide separate controlled evidence for the kernel's contribution beyond the surrounding pipeline.

Revisions in the manuscript: Section 5.2 and Figure 5 (multiplier-only and case-wise response comparisons); Appendix F.3 and Table F.1 (paired protocol and per-level results); Equations (F.15)-(F.16) and Table F.2 (field-error normalization and interpretation).

AuthorsOpenReview ↗

Response to Reviewer sDXG (3/4): Theory, Hierarchy Scaling, and CFD Support (W2 & Question)

[W2 & Question] Theoretical contribution and larger hierarchies

  • "Theorem 1 is essentially the standard finite-tree Carleson embedding (Treil, 2013) applied to bounded symbols. It bounds one linear action at a fixed input. It says nothing about the stability or generalization of the full model, and the paper doesn't explain why this bound should lead to lower error."
  • "Do you have evidence that the bounded construction helps beyond the small gains in Table 3, for example with deeper trees, larger N or a change in resolution?"

1. Theoretical contribution and scope. We credit the embedding estimate to classical Carleson theory. Our contribution is an input-conditioned coefficient construction: each bounded symbol generates all transfer coefficients in one direction, with explicit regional responses and joint control across nested regions. Proposition 1 and Equations (9)-(11) express both directional responses and their exact norms through regional symbol variation.

Theorem 1 bounds each fixed-input kernel action; it does not alone establish full-predictor stability or generalization. Appendix C.3 separates feature variation from input-induced kernel variation on a fixed weighted tree and bounds the latter through symbol and multiplier changes. Full-predictor stability additionally depends on the coefficient generators, representation maps, and nonlinear updates.

2. Connection to predictive accuracy. Direct bounds coefficients individually without enforcing joint subtree control. Normalized-direct enforces that control through a global rescaling factor for each direction, determined by the largest normalized subtree energy. HPNO instead generates Haar contrasts of pointwise-bounded symbols, satisfying the same subtree constraint without global rescaling (Appendix F.2).

In Table 4, Direct, Normalized-direct, and HPNO obtain Darcy errors of 0.00338, 0.00327, and 0.00311, and Elasticity errors of 0.00402, 0.00348, and 0.00314, respectively. These comparisons support bounded-symbol parameterization as an inductive bias beyond enforcing the bound alone. Figure 4(b) and Section 5.2 now explicitly state that HPNO and Normalized-direct satisfy the unit bound by construction.

3. Evidence beyond the default setting.

Resolution and hierarchy scaling (Figure 4(c)). Each variant is trained separately at each resolution. HPNO has lower error at every resolution, and the relative excess errors over HPNO increase across the four settings:

Darcy resolution(N,D)(N,D)HPNO error (×10−3\times10^{-3})Direct excessNorm-direct excess
43×4343\times43(225,8)(225,8)7.86+1.4%+2.7%
85×8585\times85(841,10)(841,10)3.11+8.7%+5.1%
141×141141\times141(2209,12)(2209,12)2.10+13.1%+9.7%
211×211211\times211(5041,13)(5041,13)1.91+20.8%+18.6%

Resolution, support size, and depth change together, so this is evidence across these combined settings rather than an isolated causal effect of depth. Table E.6 reports the corresponding inference latency and memory.

Support changes at fixed input resolution (Table E.5). With Darcy inputs fixed at 85×8585\times85, varying patch size changes Haar support while preserving the physical input resolution:

Haar support NNHPNODirectNormalized-direct
2890.003350.003430.00346
8410.003110.003380.00327
1,8490.003220.003410.00335

HPNO has lower error than both direct variants at each support size, including the variant that already satisfies the subtree bound.


[W2] Interpretation of CFD supports

  • "Also, on the CFD tasks the tree is built over slice slots, so the physical mean-versus-detail motivation doesn't really apply there."

For CFD, the hierarchy acts on learned slice features. Its means and details describe averages and contrasts among latent slots and need not correspond to physical neighborhoods or length scales. The weighted Haar identities and subtree bounds remain valid; slicing and deslicing connect these feature interactions to physical predictions. We now state this distinction explicitly. The paired Darcy experiment separately evaluates physical regional responses of the complete predictor.

Revisions in the manuscript: Sections 4.1, 4.2, and 4.4; Proposition 1 and Theorem 1; Appendices C.1-C.3 (theory and scope); Section 5.2, Figure 4(b,c), and Table 4 (diagnostics and ablations); Appendix E.4, Table E.5 (support accuracy), and Table E.6 (resolution-dependent costs); Appendices E.3 and F.2 (representations and variant definitions).

Reviewer_N8dsOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces the Haar Paraproduct Neural Operator (HPNO), a novel architecture that learns input-conditioned kernel actions using Haar wavelets. By coupling an exact Haar multiplier with directional mean-detail transfers (paraproducts), the authors establish theoretical weighted L2L^2 bounds for the kernel action that are independent of hierarchy size and depth. The authors demonstrate HPNO's effectiveness across six PDE benchmarks and two CFD tasks, supporting their theoretical claims with empirical ablations.

Critical strengths
  • Solid Theoretical Foundation: The paper provides a rigorous mathematical formulation connecting directional response strength to regional symbol variation, and establishes a strict weighted L2L^2 bound independent of tree depth.
  • Novel Architecture: The integration of bounded symbols with Haar paraproducts to enforce subtree Carleson conditions is a fresh and mathematically grounded approach to multiscale neural operators.
  • Comprehensive Benchmarking: The model is evaluated across a diverse set of physical tasks (grids, meshes, point clouds), showing consistent improvements over strong baselines (e.g., Transolver, HNO, LinearNO).
Critical weaknesses

1. Insufficient Readability of Figures

  • Figure 2 is visually dense, featuring small fonts and a high density of mathematical formulas. The caption lacks sufficient explanation of key symbols and intermediate quantities, making it difficult to follow the data flow.
  • Figure 3 lacks color bars and error range annotations, which severely hinders the reliable visual comparison of signed prediction errors across different models.  2. Limited Discussion on Graph Aggregation and Multiscale Methods
  • The paper lacks a thorough discussion on how the binary-tree Haar Paraproduct differs fundamentally from kNN local aggregation and graph hierarchy methods. The benefits and limitations of the proposed tree structure over graph-based approaches regarding irregular sampling and local information modeling are not fully articulated.
  • While the paper provides theoretical analysis, the mechanistic comparison with existing spatial multiscale neural operators (e.g., MNO[1], AMG[2]-based multi-graph neural operators) remains insufficient to justify HPNO's unique advantages.  3. Unresolved Citation Placeholders
  • The manuscript contains unresolved citation placeholders. For instance, citations for FNO, RNO, and GEO-FNO in Tables 1–2 still display as "(?)". Similar unresolved references appear in the experimental setup and the appendix. This is a significant formatting error that severely affects literature traceability and requires immediate correction.  4. Unclear Provenance and Fairness of Baseline Comparisons
  • Although the appendix acknowledges that not all baselines were retrained, the paper does not itemize whether each reported result is from the original paper or reproduced by the authors. Without clear verification of data splits, input information, and training budgets, the fairness of the comparisons remains ambiguous.
  • The analysis of key multiscale baselines (e.g., MWT, WNO, SAOT) is superficial, focusing primarily on overall rankings without dissecting the underlying reasons for performance differences.  5. Incomplete Reporting for Reproducibility and Practical Cost
  • The paper fails to report essential training configurations (e.g., learning rate, training epochs, batch size, network scale/parameter counts, and amplitude bounds).
  • While asymptotic complexity is provided in the appendix, there is no evaluation of practical computational efficiency. Metrics such as parameter counts, training/inference time, GPU memory usage, and the overhead of tree building are entirely missing, making it impossible to evaluate the trade-off between accuracy gains and computational costs.
Questions for authors

1. Regarding Figures:

  • Revise Figure 2 to simplify mathematical details and clearly highlight the inputs, outputs, and data flow of each module.
  • Update Figure 3 to include color bars and explicit error range annotations for reliable comparison.

2. Regarding Comparisons with Multiscale/Graph Methods:

  • Explicitly discuss the advantages of the binary-tree Haar Paraproduct over graph-based multiscale methods (e.g., MNO [1] or multi-graph neural operators [2]).
    • [1] MNO: Multiscale Neural Operator for 3D Computational Fluid Dynamics
    • [2] Harnessing scale and physics: A multi-graph neural operator framework for PDEs on arbitrary geometries

3. Regarding Configurations and Computational Efficiency:

  • Provide a detailed table of the hyperparameters used in the experiments.
  • Report the actual computational costs (e.g., parameter count, training/inference time, and memory usage).

Reply to author rebuttal

Thank you for the detailed response and revisions. The reciprocal kernel swaps address my concerns about the source of the performance gains, and the revised figures and cost analysis make the results easier to assess. I maintain my rating of 3 (Weak Accept) and raise my confidence from 2 to 3.

5 comments
AuthorsOpenReview ↗

Response to Reviewer N8ds (1/4): Overview, Figures, and Citations (W1, W3, Q1)

Thank you for your careful review and for recognizing HPNO's theoretical foundation, architectural contribution, and benchmark coverage. Your feedback helped us improve the presentation, clarify the comparisons, and document the practical costs.

We have revised the manuscript and organized our response into four posts:

  1. Post 1/4 (W1, Q1, W3): Redrawn Figure 2, numerical color scales in Figures 3 and F.1, and citation fixes.
  2. Post 2/4 (W2, Q2): Comparison with graph-based multiscale operators (MGNO, MNO, AMG) and partition locality.
  3. Post 3/4 (W4): Baseline provenance, reciprocal kernel replacements across all eight benchmarks, and multiscale analysis.
  4. Post 4/4 (W5, Q3): Hyperparameters, training/inference costs, tree-construction overhead, and scaling.

All references refer to the revised manuscript. The new reciprocal kernel-replacement experiments are in Table 3; the original architectural ablations are now Table 4.


[W1 & Q1] Readability of Figures 2 and 3

  • "Revise Figure 2 to simplify mathematical details and clearly highlight the inputs, outputs, and data flow of each module."
  • "Update Figure 3 to include color bars and explicit error range annotations for reliable comparison."
  • Figure 2 (Redrawn): Separates the overall predictor pipeline (panel a) from the computation inside one Haar paraproduct kernel (panel b). Solid arrows trace incoming features ff and dashed arrows indicate coefficients generated from conditioning features v0v_0. The multiplier M\mathcal{M} and the two directional transfers (Πα\Pi_\alpha, Πβ∗\Pi_\beta^*) appear as color-coded parallel branches accumulating into Kf\mathcal{K}f, with the key quantities and data flow explained in the caption.
  • Figures 3 and F.1 (Color bars and shared ranges): Numerical color bars are provided for ground-truth fields and signed error maps (u^−u\hat{u} - u). Within each case and physical quantity, signed errors share an identical symmetric scale across HPNO, HNO, and LinearNO, while ground-truth fields use separate scales.

Revisions in the manuscript: Figure 2 (diagram and caption); Figure 3 and Appendix F.4, Figure F.1 (color bars, shared symmetric error scales, and captions).


[W3] Unresolved Citation Placeholders

  • "The manuscript contains unresolved citation placeholders."

We corrected all unresolved citations and cross-references across the main text, tables, and appendix, and verified the compiled PDF:

  • Distinguished FNO (Li et al., 2021) from Geo-FNO (Li et al., 2023a) and standardized Geo-FNO naming across Tables 1-2.
  • Clarified in the Table 1 caption that "-" denotes an entry not reported in the comparison source.
  • Documented the primary and secondary numerical sources for baseline results in Appendix E.2.

Revisions in the manuscript: Tables 1-2 (citations and caption); Section 2, Section 5, References, and Appendix E.

AuthorsOpenReview ↗

Response to Reviewer N8ds (2/4): Comparison with Graph-Based Multiscale Methods (W2, Q2)

[W2 & Q2] Graph-Based Multiscale Methods and Partition Locality

  • "The benefits and limitations of the proposed tree structure over graph-based approaches regarding irregular sampling and local information modeling are not fully articulated."
  • "Explicitly discuss the advantages of the binary-tree Haar Paraproduct over graph-based multiscale methods (e.g., MNO [1] or multi-graph neural operators [2])."

We expanded Section 2 and added a dedicated comparison in Appendix E.2 covering MGNO (Li et al., 2020), MNO (Wang et al., 2026), and AMG (Li et al., 2025). Whereas MNO combines global low-rank attention with local kkNN graph attention and AMG couples local/global graphs through virtual physics nodes, HPNO makes regional mean-detail responses and their coefficient control explicit:

1. Structural Advantages on Irregular Weighted Supports

  1. Weighted multiscale representation: Haar detail functions maintain zero weighted mean and unit norm even when child-region masses differ. The basis remains orthonormal under the weighted inner product on irregular supports without requiring equal-mass partitions.
  2. Input-conditioned directional coupling: Separate bounded symbols generate mean-to-detail and detail-to-mean coefficients. Their joint subtree constraints control accumulation across nested regions and yield a weighted L2(μ)L^2(\mu) bound for each fixed-input kernel action, independent of hierarchy size and depth (Theorem 1).
  3. Exact two-pass evaluation: Shared upward and downward passes evaluate the action without assembling a dense kernel matrix. While linear complexity is also achievable in graph-based designs, HPNO's distinction lies in its explicit regional responses and joint coefficient control.

2. Limitation and Empirical Sensitivity

  • Partition-dependent locality: On physical supports, nearby nodes separated by a binary split interact through common ancestors, whereas a local graph can connect them directly. Thus, the partition determines which nodes are grouped into local contrasts. In Figure 6(b), Darcy relative ℓ2\ell_2 errors remain within 0.00311 to 0.00330 across median, spatial-midpoint, and random-axis median split rules before and after joint 90∘90^\circ field rotations.
  • Latent supports: On slice-based supports, tree regions group feature slots and need not correspond to physical neighborhoods. Regional means and details therefore represent averages and contrasts among learned features.

Revisions in the manuscript: Section 2 (expanded discussion and citations); Section 4.1 ("Weighted hierarchy"); Section 5.3 and Figure 6(b) (partition and rotation study); Appendix E.2 ("Comparison with graph-based multiscale operators").

AuthorsOpenReview ↗

Response to Reviewer N8ds (3/4): Baseline Provenance, Fairness, and Multiscale Analysis (W4)

[W4] Baseline Provenance, Fairness, and Multiscale Analysis

  • "Although the appendix acknowledges that not all baselines were retrained, the paper does not itemize whether each reported result is from the original paper or reproduced by the authors."
  • "The analysis of key multiscale baselines (e.g., MWT, WNO, SAOT) is superficial, focusing primarily on overall rankings without dissecting the underlying reasons for performance differences."

1. Result Provenance (Appendix E.2)

All baseline errors in Tables 1-2 are literature-reported results. We prioritize original-paper numbers and itemize secondary benchmark sources, including reproductions by those authors, as follows:

  • Six PDE benchmarks: Wu et al. (2024) for MWT, Geo-FNO, and GNOT; Luo et al. (2025) for Transolver-series comparisons; Zhou et al. (2026) for WNO; and Pei et al. (2026) for RNO.
  • CFD benchmarks: Wu et al. (2024) for Geo-FNO, GNOT, and Transolver; and Hu et al. (2026) for LNO and Transolver++.
  • HPNO setup: Our evaluations use the released benchmark test splits. Input definitions, data splits, and training budgets are documented in Tables E.1 and E.3.

2. Controlled Comparison Fairness (Section 5.1, Table 3)

We added reciprocal kernel replacements across all eight benchmarks with HNO, LinearNO, and Transolver++: evaluating their kernels inside the HPNO predictor (Table 3(a)) and the HPNO kernel inside each baseline predictor (Table 3(b)). Within each host, the feature interface, representation/reconstruction maps, residual and auxiliary paths, loss, data split, training budget, and final-epoch evaluation rule are held fixed. Each variant is retrained. Representative Darcy results are shown below:

Host predictorNative / baseline kernelWith HPNO kernel
HPNO predictor (Table 3(a))0.0040 (HNO) / 0.0043 (LinearNO) / 0.0042 (Transolver++)0.0031
HNO predictor (Table 3(b))0.00450.0036
LinearNO predictor (Table 3(b))0.00500.0039
Transolver++ predictor (Table 3(b))0.00490.0038

These controls show that the surrounding task-specific pipeline alone does not explain HPNO's gains, while the CFD results in Table 3 also show where performance depends jointly on the kernel and host representation.

3. Mechanistic Analysis (Section 5.2, Table 4, Appendix E.2)

MWT couples scaling and wavelet coefficients, WNO applies learned wavelet-domain kernels within nonlinear layers, and SAOT combines input-dependent wavelet attention with Fourier processing. HPNO generates separate mean-to-detail and detail-to-mean coefficients from input-conditioned bounded symbols with joint subtree control. In the profiled Darcy setup, it also uses fewer parameters, lower latency, and less peak allocated memory than MWT, WNO, and SAOT (Table E.4(a)). Fixed-predictor ablations (Table 4) isolate two effects:

  • Directional coupling across retained coordinates: M\mathcal{M}-only retains the root mean and every Haar detail; adding Πα+Πβ∗\Pi_\alpha + \Pi_\beta^* lowers error from 0.00345 to 0.00311 on Darcy and 0.00364 to 0.00314 on Elasticity.
  • Bounded-symbol parameterization vs. norm control alone: Direct bounds individual coefficients but does not impose a joint subtree constraint. Subtree normalization (Norm-direct) lowers Darcy error from 0.00338 to 0.00327, while bounded-symbol HPNO reaches 0.00311 under the same subtree bound. The corresponding Elasticity errors are 0.00402, 0.00348, and 0.00314, respectively.

Revisions in the manuscript: Section 5.1 and Table 3 (reciprocal kernel swaps); Section 5.2, Table 4, and Appendix F.2 (ablations and labels); Appendix E.2 (baseline sources, swap protocol, and multiscale analysis).

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

The paper proposes the Haar Paraproduct Neural Operator. Its kernel is built on a weighted binary Haar tree over the kernel's points. It has the lowest reported errors on six PDE benchmarks.

Critical strengths
  • Efficient and works on any geometry since no dense kernel is formed.
Critical weaknesses
  • The gains can't be attributed to the kernel alone. HPNO adds task-specific extras that the baselines lack, such as local convolution and a gradient loss on Darcy, full-resolution refinement on Airfoil, pressure correction on ShapeNet Car.
Questions for authors
  • How does HPNO perform when all baselines are retrained under the same pipeline (patching, losses, extra paths)? Or when the HPNO kernel is swapped into the HNO backbone and vice versa, to isolate the kernel's contribution?
  • Is there any runtime or memory comparison? What are wall-clock time, peak memory and param count against HNO, LinearNO and Transolver at increasing N?
  • How do results change with the split rule (median vs. dyadic vs. random), with rotating the domain, or with a learned or adaptive tree?
  • Does HPNO support zero-shot super-resolution / discretization transfer, given that the tree changes with resolution?
5 comments
AuthorsOpenReview ↗

Response to Reviewer siaJ (1/4): Isolating the Kernel Contribution (W1, Q1)

Thank you for your review and for recognizing the geometric flexibility and efficiency of the kernel construction. Your questions motivated controlled kernel comparisons across all eight benchmarks, together with evaluations of computational cost, partition sensitivity, and discretization transfer.

We organize our response into four posts:

  1. Post 1/4 (W1, Q1): Reciprocal kernel replacements, focusing on the task-specific processing highlighted in your review.
  2. Post 2/4 (Q2): Parameters, runtime, peak memory, and scaling with physical input and Haar support size.
  3. Post 3/4 (Q3): Partition rules, rotation sensitivity, and adaptive hierarchies.
  4. Post 4/4 (Q4): Frozen-checkpoint discretization transfer and the scope of resolution-dependent hierarchy changes.

All references refer to the revised manuscript. Table 3 contains the new reciprocal kernel-replacement experiments; the architectural ablations are now Table 4.


[W1 & Q1] Isolating the kernel contribution

  • "The gains can't be attributed to the kernel alone. HPNO adds task-specific extras that the baselines lack, such as local convolution and a gradient loss on Darcy, full-resolution refinement on Airfoil, pressure correction on ShapeNet Car."
  • "How does HPNO perform when all baselines are retrained under the same pipeline (patching, losses, extra paths)? Or when the HPNO kernel is swapped into the HNO backbone and vice versa, to isolate the kernel's contribution?"

1. Reciprocal replacements with the surrounding architecture fixed. We implemented your proposed reciprocal replacements with HNO and extended them to LinearNO and Transolver++ across all eight benchmarks:

  • Table 3(a): These three baseline kernels are retrained inside the HPNO predictor, sharing its task-specific processing, including Darcy local convolution and gradient loss, Airfoil full-resolution refinement, and ShapeNet Car pressure correction.
  • Table 3(b): The HPNO kernel is retrained inside each baseline predictor, preserving that host's native representation, losses, and auxiliary paths.

Within each host, token order, feature dimensions, representation/reconstruction maps, residual processing, auxiliary paths, loss, data split, training budget, and final-epoch evaluation rule remain fixed. Every variant is retrained.

2. Results on the tasks highlighted in your review. The following rows summarize Table 3(a). Metrics follow Tables 1-2; lower is better.

Kernel in HPNO predictorDarcyAirfoilShapeNet Car volumeShapeNet Car surface
HNO0.00400.00460.02100.0760
LinearNO0.00430.00470.01990.0756
Transolver++0.00420.00470.02180.0781
HPNO0.00310.00430.02050.0753

Table 3(b) reports native-kernel error →\to HPNO-kernel error within each baseline predictor:

Host predictorDarcyAirfoilShapeNet Car volumeShapeNet Car surface
HNO0.0045 →\to 0.00360.0048 →\to 0.00450.0214 →\to 0.02080.0764 →\to 0.0757
LinearNO0.0050 →\to 0.00390.0049 →\to 0.00450.0194 →\to 0.01960.0754 →\to 0.0752
Transolver++0.0049 →\to 0.00380.0049 →\to 0.00450.0226 →\to 0.02150.0800 →\to 0.0776

The full eight-benchmark results are in Table 3. These reciprocal controls show that HPNO's task-specific pipeline alone does not explain the gains. The ShapeNet Car volume results also show where performance depends on the kernel and host representation, including the LinearNO exception above.

Tables 1-2 contain literature-reported baseline results; Table 3 contains the controlled retrained comparisons. Their provenance and protocols are documented separately in Appendix E.2.

Revisions in the manuscript: Section 5.1 and Table 3 (reciprocal replacements across all eight benchmarks); Appendix E.2, "Controlled kernel comparisons" (interfaces and training controls); Tables E.2-E.3 (task-specific processing and configurations).

AuthorsOpenReview ↗

Response to Reviewer siaJ (2/4): Runtime, Memory, Parameter Count, and Scaling (Q2)

[Q2] Parameters, runtime, memory, and scaling

  • "Is there any runtime or memory comparison? What are the wall-clock time, peak memory and param count against HNO, LinearNO and Transolver at increasing N?"

1. Complete-predictor costs under a common profiling setup. We added the accuracy-cost comparison in Figure 6(c), numerical costs for all 12 models in Table E.4(a), and the profiling protocol in Appendix E.4. For Darcy at 85×8585\times85, batch size four, on an NVIDIA RTX PRO 6000, the comparison includes both the original Transolver and Transolver++:

ModelReported Darcy errorParameters (M)Inference (ms/batch)Peak allocated memory (GB)
HNO0.00450.81591.28800.0755
Transolver0.00582.82699.97320.3073
LinearNO0.00501.766112.77060.3370
Transolver++0.00492.836213.85950.4255
HPNO0.00311.18902.08310.1001

Baseline errors are literature-reported, and HPNO's error is from our evaluation; costs are profiled separately. In this Darcy setup, HPNO has fewer parameters, lower latency, and lower peak memory than the profiled Transolver, LinearNO, and Transolver++ configurations. HNO remains smaller, faster, and less memory-intensive, so we present an accuracy-cost trade-off.

Table E.3(c) also reports estimated single-GPU training times: 0.95 h for Darcy and 0.73 to 6.48 h across all eight tasks.

2. Scaling with physical input node count (nXn_{\mathcal{X}}). Table E.4(b) compares HNO, LinearNO, Transolver++, and HPNO on AirfRANS at batch size one. HPNO's physical input count varies from 2k to 128k while its latent Haar support remains fixed at N=64N=64, D=6D=6. Each cell reports inference time in ms / peak allocated memory in GB:

ModelParameters (M)2k input nodes32k input nodes128k input nodes
HNO3.89392.7624 / 0.071115.5976 / 0.399266.2054 / 1.1797
LinearNO3.35882.6363 / 0.067227.7645 / 0.3655122.2528 / 1.3087
Transolver++6.00562.3237 / 0.078016.5402 / 0.377277.5265 / 1.3235
HPNO7.61087.2138 / 0.093220.3035 / 0.416871.1705 / 1.1968

At 128k nodes, HPNO has lower latency and peak memory than LinearNO and Transolver++, while HNO remains faster and uses slightly less memory.

3. Scaling with Haar support size (NN). Table E.6 profiles HPNO at batch size one as Darcy resolution, Haar support, and tree depth increase together:

Darcy resolutionNNDDInference (ms)Peak allocated memory (GB)
43×4343\times4322581.74350.0424
85×8585\times85841102.06470.0536
141×141141\times1412,209122.35320.0734
211×211211\times2115,041132.55080.1339

These measurements distinguish growth in the physical input from growth in the Haar hierarchy.

Profiling scope. Measurements use FP32 with TF32 disabled, 30 warm-ups, and 100 synchronized forward passes. Timings use cached hierarchies and include input-conditioned coefficient generation. Tree construction is measured separately: 2.3838 ms for Darcy with N=841N=841 at batch size one. AirfRANS timings measure one subset forward pass and exclude repeated-subset aggregation for full-mesh accuracy evaluation.

Revisions in the manuscript: Section 5.3 and Figure 6(c); Appendix E.3 and Table E.3 (configurations and per-task costs); Appendix E.4 and Tables E.4(a,b) and E.6 (profiling, tree construction, and scaling).

AuthorsOpenReview ↗

Response to Reviewer siaJ (3/4): Split Rules, Domain Rotation, and Adaptive Trees (Q3)

[Q3] Partition rules, rotation, and learned hierarchies

  • "How do results change with the split rule (median vs. dyadic vs. random), with rotating the domain, or with a learned or adaptive tree?"

1. Partition choice and rotation. We added a Darcy comparison in Figure 6(b) using three partition rules: the default median, spatial-midpoint, and random-axis median splits. The spatial-midpoint variant provides the geometric-midpoint comparison. Reported relative ℓ2\ell_2 errors before and after joint 90∘90^\circ field rotation are:

Partition ruleOriginal orientationJoint 90∘90^\circ rotation
Median0.003110.00312
Spatial midpoint0.003170.00316
Random-axis median0.003270.00330

Median and spatial-midpoint partitions give similar errors, while random-axis median gives slightly higher errors. Across the six tested settings, errors range from 0.00311 to 0.00330, indicating limited sensitivity to these partition choices and the tested rotation.

2. Training and evaluation protocol. We reuse the median-partition checkpoint and train the spatial-midpoint and random-axis median variants separately under the same training budget. Each final-epoch checkpoint is then frozen and evaluated on both the original test fields and the jointly 90∘90^\circ-rotated permeability and target pressure fields on the fixed grid. Thus, each partition has its own trained model, and the orientation comparison uses no rotation-specific retraining.

3. Learned or adaptive trees. We have not evaluated learned or task-adaptive tree topologies. The weighted Haar construction and Theorem 1 apply to any fixed finite binary partition tree with positive support weights and nonempty children, including unequal child masses. Median splitting is therefore not required by the kernel formulation. Section 6 identifies adaptive hierarchies, alongside amortized multi-physics foundation models, as future work.

Revisions in the manuscript: Section 5.3 and Figure 6(b) (partition and rotation results); Appendix E.4, "Discretization and rotation protocols" (training and evaluation details); Section 4.1 and Theorem 1 (tree assumptions); Section 6 (adaptive hierarchies).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Additional Experiments

We thank all three reviewers (N8ds, sDXG, and siaJ) for their careful reviews and constructive suggestions. We appreciate your positive assessment of HPNO's theoretical foundation and architecture (N8ds), bounded-symbol construction and ablations (sDXG), and ability to handle diverse geometries without assembling dense kernels (siaJ).

Following your feedback, we conducted extensive controlled experiments and revised the manuscript to isolate the kernel's contribution and clarify its theoretical and practical scope.

Guide to the revised PDF: Within the 10-page main-text limit, the revision incorporates a new subsection (Section 5.3), a new appendix section (Appendix E.4), six new/expanded tables (Tables 3, E.3–E.6, and F.2), and six new, updated, or redrawn figures (Figures 2–6 and F.1). All references follow the revised numbering (Table 3 reports the reciprocal kernel replacements; Table 4 reports the architectural ablations), and substantive revisions are highlighted in \textcolor[RGB]{0, 115, 235}{\textbf{blue}}.

  1. Reciprocal Kernel Replacements (N8ds, sDXG, siaJ): We retrained HNO, LinearNO, and Transolver++ inside the HPNO predictor (Table 3(a)) and the HPNO kernel inside each baseline's native predictor (Table 3(b)) across all eight benchmarks under fixed host architectures, auxiliary paths, losses, and budgets. Improvements in both directions confirm the kernel's contribution beyond surrounding pipeline components.

    \textcolor[RGB]{0, 115, 235}{\Rightarrow} Revised PDF: New Table 3(a,b) · Section 5.1 (expanded) · Appendix E.2 (new swap protocol)

  2. Paired Darcy Controls and Error Normalization (sDXG): We added the multiplier-only (M\mathcal{M}-only) baseline, matched-pair comparisons, and overall field-error normalization (Q=Eresp,pool/EfieldQ = E_{\mathrm{resp,pool}} / E_{\mathrm{field}}), separating absolute regional-response gains from field-error-normalized ratios.

    \textcolor[RGB]{0, 115, 235}{\Rightarrow} Revised PDF: New Table F.2 · Figure 5(b,c) and Table F.1 (updated with M\mathcal{M}-only) · Section 5.2 · Appendix F.3 (new Eqs. (F.15)–(F.16))

  3. Theoretical Scope and Multiscale Comparisons (N8ds, sDXG): We state the fixed-input scope of Theorem 1 explicitly, separate the bounded-symbol parameterization from the classical Carleson embedding estimate, and compare local symbol generation against global subtree rescaling (Norm-direct). We also expanded mechanistic comparisons with graph-based (MGNO, MNO, AMG) and wavelet (MWT, WNO, SAOT) operators, and clarified latent slice supports on CFD tasks.

    \textcolor[RGB]{0, 115, 235}{\Rightarrow} Revised PDF: Sections 2, 4.1, and 4.4 (expanded) · Table 4 (clarified labels) · Appendices C.3, E.2, and F.2 (expanded analyses)

  4. Scaling, Partition Sensitivity, and Discretization Transfer (sDXG, siaJ): Multi-resolution Darcy experiments (43×4343\times 43 to 211×211211\times 211, depths D=8D=8 to 1313) and fixed-grid support studies show consistent gains over Direct and Norm-direct. We also evaluated partition rules (median, spatial-midpoint, and random-axis median) under joint 90∘90^\circ rotations on Darcy and frozen-checkpoint sampling-density transfer (2k, 32k, 128k nodes) on AirfRANS, distinguishing fixed-latent-tree transfer from rebuilt spatial hierarchies.

    \textcolor[RGB]{0, 115, 235}{\Rightarrow} Revised PDF: New Section 5.3 · New Figures 4(c) and 6(a,b) · New Table E.5 · New Appendix E.4 · Section 5.2

  5. Computational Costs, Reproducibility, and Presentation (N8ds, sDXG, siaJ): We report 12-model Darcy inference profiling, AirfRANS and Darcy scaling costs, tree-construction overhead, estimated training hours, full hyperparameters, and itemized literature sources for baseline results. We also redrew Figure 2, added color bars and shared symmetric error ranges to Figures 3 and F.1, and resolved all citation and notation issues.

    \textcolor[RGB]{0, 115, 235}{\Rightarrow} Revised PDF: New Figure 6(c) · New Tables E.3, E.4, and E.6 · Redrawn Figure 2 · Figures 3 and F.1 (color bars & shared error scales) · Tables 1–2 · Appendices E.2–E.4 and F.4

Detailed point-by-point responses are posted directly under each review. Your constructive feedback has substantially strengthened our manuscript, and we hope the new controlled experiments and clarifications thoroughly address your concerns. We remain actively available throughout the discussion period and warmly welcome any further questions or follow-up discussions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.