acceptodds
Under review as a conference paper at ICLR 2027

Local Basis Neural Fields: Scalable Scientific Signal Fitting with Explicit Order of Accuracy for Values and Derivatives

20 views

Abstract

Neural field representations are increasingly used to represent, visualize, and analyze large-scale scientific signals, such as astronomical simulations and turbulent flows. These applications demand scalable representations that are accurate not only in reconstructed values, but also under differentiation. However, purely MLP-based networks scale poorly to large signals, while hybrid methods that pair MLPs with grid-based features (e.g., hash grids) introduce significant artifacts under differentiation. To bridge this gap, we introduce Local Basis Neural Fields (LBNFs), a framework that represents a signal using a set of overlapping local polynomial functions placed adaptively across the domain, paired with a lightweight MLP that decodes these local features into signal values. The polynomial degree and the placement density together give LBNFs explicit, independent control over modeling complexity and its allocation around complex signal structures. Because LBNFs build on classical partition-of-unity (PoU) approximation theory, they inherit its convergence guarantees: for smooth targets, this yields convergence in value and convergence for derivatives of order , where is the spacing between polynomials and their degree. Empirically, LBNFs realize these -approximation rates on analytic benchmarks while achieving significantly lower absolute error than classical approximants. On large-scale scientific data, LBNFs scale comparably to state-of-the-art neural field methods (e.g., Instant-NGP and NeuRBF), matching or improving their reconstruction error while reducing first- and second-order gradient errors by an order of magnitude.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
implicit neural representations, hp-approximation, scientific computing, scientific visualization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_LbauOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper identifies the challenge of Neural Fields to model large-scale signals such as astronomical simulations and fluid flows, using pure MLP networks, as they tend towards global representation rather than focusing on local features. It is pointed out that scientific signal representation should be evaluated using signal derivatives, as they represent the dynamics of the signal, rather than just measuring only the reconstruction accuracy. The introduced Local Basis Neural Fields (LBNFs) follow a hybrid method using polynomials combined with lightweight MLPs, where the input signal coordinates are mapped to polynomial features. These features are given as the input to the lightweight MLP to output the signal value for a given coordinate. The proposed LBNF increases the controllability of the representation by selecting the approximation level of the polynomials. Further, training anchor points are either fixed as a grid or varied during training to learn complex signal components more accurately. The results show that LBNF competes with the SOTA algorithms in representing large signals, achieving comparatively high performance.

Critical strengths
  1. Introducing polynomials combined with lightweight MLPs as a hybrid approach increases the controllability of the representation. When using a pure MLP based network, the approximation level is mainly determined through the number of parameters it possess, while it is controlled by changing the depth and the width of the MLP in general. From the proposed LBNF, the architectural changes in MLPs for approximation are replaced with polynomials.

  2. The representation results for signal reconstruction are at an acceptable level. The quantitative results in Table 01 show a significant increase in the LBNF’s PSNR improvement. The qualitative improvement in Figure 1 shows that LBNF tries to correctly learn the signal structure with fewer incorrect predictions during inference.

  3. The additional ablation studies further experimentally provide the model performance under different scenarios and parameter variations, as shown in Table 2 and the appendix.

Critical weaknesses
  1. Although the overall contribution is explained in the paper, it is hard to identify the different contributions to distinguish the novelty in the methodology. Specifically explain whether the Adaptive Anchor placement is a novel contribution or it is an existing method. If it exists, need to explain the novel way that has been used in LBNF.

  2. Although the authors point out that it is essential to compare the derivatives of scientific signals, according to Table 01, the best performance in some derivative errors can be seen in SIREN, but not for LBNF. Need a detailed explanation as a justification for the result.

Questions for authors
  1. Distinguish the contributions clearly. Is it only the LBNF, or are there multiple contributions? Clarify the novelty of the adaptive anchor points used in the methodology.

  2. Provide justifications for why SIREN has better derivative preservation compared to LBNF, as in Table 01.

2 comments
AuthorsOpenReview ↗

Response to Lbau

We thank the reviewer for their positive assessment and constructive feedback. Below, we clarify our contributions, including the role of adaptive anchor placement, and explain the SIREN derivative results.

CW1/Q1: Contributions and novelty of adaptive anchor placement.

LBNF makes multiple contributions, including a new anchor-placement mechanism. Its components (local polynomials, Gaussian weighting, a small neural decoder) are individually established. Our contribution lies in how they are combined and placed, what can be proven about the result, and the accuracy this achieves on large scientific signals.

  1. Adaptive anchor relocation. While adaptive basis placement is a well explored idea in classical numerical methods, our contribution is the relocation mechanism itself. LBNFs start from random anchors and relocate them during training through residual-driven split-and-merge combined with gradient updates (Sec. 3.2). Not only is this parameter neutral, but it also removes the need for signal-dependent initialization. Many other methods either fix their bases (Instant-NGP, ∂∞\partial^\infty-Grid) or rely on a special initialization (NeuRBF, efunc). See Figures 3 and 20 for more details.

  2. Approximation guarantees. The approximation rates themselves are classical. However, to the best of our knowledge we are the first to show that a hybrid neural field can attain these rates (Sec. 3.3). Because the rates are explicit in the polynomial order pp and anchor spacing hh, this makes pp and hh interpretable accuracy controls, and we show that trained LBNFs meet the predicted rates empirically (Figure 7, Appendix C.2). We have additionally clarified in Sec. 3.3 how these results explicitly relate to classical theory.

CW2/Q2: SIREN derivative results in Table 1.

SIREN and LBNF have comparable derivative accuracy in Table 1, with SIREN modestly ahead on some maximum-error and second-derivative metrics. This is not surprising, since SIREN is a single global network with smooth sinusoidal activations, so its derivatives are smooth everywhere. However, this global parameterization comes at a substantial cost in value accuracy and training memory (0.32 GB for LBNF versus 3.00 GB for SIREN). The LBNF results in Table 1 are also truncated. Each query point blends only its K=32K=32 nearest anchors rather than all anchors, which intentionally keeps memory low on the larger scientific signals (Appendix D.2). Unlike SIREN, LBNFs expose truncation as an explicit control over this tradeoff, since blending more neighbors (larger KK) reduces derivative error at higher memory cost (Table 2, Appendix C.4). This control matters for large scientific signals. All representations need more parameters for larger signals, but in a global network, such as SIREN, every query passes through the full network, so per-query cost grows with capacity. LBNFs instead add capacity through more anchors while each query blends only KK of them, leaving users free to choose how much derivative accuracy (inter-neighborhood smoothness) to buy with additional compute.

Thank you for the clarifications. I agree with the explanations about the derivative concern. However, I am not fully convinced that the level of novelty could increase my score. Hence, I'll keep my acceptance rate the same.

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

The authors propose an n-dimensional field representation based on polynomial kernels that are mixed through Gaussian weighting. This representation is infinitely smooth and supports accurate derivative computation, which is useful for scientific applications where previous grid-based methods give zero or undefined derivatives. The polynomial representation also admits theoretical error bounds. The results show improved signal reconstruction on scientific datasets and significantly better computation of derivative-based properties of the signal.

Critical strengths

1- the paper provides an infinitely smooth alternative representation for signal fitting 2- the proposed representation achieves improved reconstruction quality on scientific data while allowing derivative-based property computation without failure.

Critical weaknesses
  1. I think the choice of polynomials as the anchor functions is not experimentally supported. From what I understand, the polynomials define a local function in space that provides an anchor feature given a query point. This is similar to the setup of KiloNeRF [1], where instead of polynomials, tiny MLPs are used, which can have more expressive power than the hand-designed polynomial forms used here. MLPs with smooth activations would also be infinitely smooth, so I'm not sure how the decision to use polynomials is justified in the paper. I would like to see a comparison with tiny MLPs instead of polynomials, combined with the proposed Gaussian-weighted mixture.

  2. The second reason provided in the paper for using polynomials, besides smooth derivatives, is the theoretical error bounds that polynomials provide. The proof seems to assume that the MLP decoder behaves like a linear function, but there is no guarantee that the trained MLP will behave this way. Therefore, I'm not sure to what extent these bounds apply to the actual trained method. Also, a similar approximation argument could potentially be made for the KiloNeRF-style setup.

[1] Reiser et al., KiloNeRF: Speeding Up Neural Radiance Fields with Thousands of Tiny MLPs, ICCV 2021.

Questions for authors

My main concerns and requests that need answer and additional evidence were posed in the Critical weaknesses section.

2 comments
AuthorsOpenReview ↗

Response to QVXZ

We thank the reviewer for their positive assessment and constructive feedback. In response, we have added an ablation comparing polynomial and MLP local bases (Appendix C.7) and clarified what the approximation guarantees cover with respect to the trained decoder.

CW1: Choice of polynomial local bases.

The LBNF framework is not restricted to polynomial bases, and we agree the choice should be justified. We chose polynomials for two reasons, and have added the requested comparison against smooth-MLP local bases.

  1. Analytic structure gives inherited rates. Degree-pp polynomial anchors contain the classical local Taylor approximant exactly, and we show that LBNFs inherit classical value and derivative rates in pp and hh with no intermediate approximation step (Sec. 3.3). With MLP anchors, one standard approach to obtaining rates is to show that each local MLP approximates a classical basis with known approximation rates (e.g., polynomials) in both value and derivatives (Yarotsky, 2017; Opschoor et al., 2020; De Ryck et al., 2021). This adds a network-size-dependent error and loses an explicit and more interpretable order control such as pp.

  2. Polynomial anchors perform better empirically. In the new Appendix C.7, we compare cubic polynomial anchors with sinusoidal and tanh⁡\tanh MLP anchors under the same Gaussian-weighted aggregation and decoder. Even with slightly fewer parameters, polynomial LBNFs achieve the highest PSNR and the lowest gradient and Hessian errors.

Local basisParamsPSNR ↑\uparrowRel. L2↓L^2 \downarrowRel. L∇2↓L^2_{\nabla} \downarrowRel. L∇22↓L^2_{\nabla^2} \downarrow
Polynomial (p=3p=3)86.7k90.031.3e-40.00100.0080
sin-MLP91.8k86.082.0e-40.00140.0096
tanh⁡\tanh-MLP91.8k83.402.7e-40.00190.0145

CW2: Applicability of the bounds to the trained model.

This is a fair point. Our results are existence guarantees and they show that the LBNF hypothesis class can attain the classical rates, so adding a neural decoder does not cost approximation order in theory. The near-linear decoder is a construction used in the proof, not an assumption or requirement on the trained decoder, and we do not claim that gradient-based training is guaranteed to find these solutions. We clarified this scope in Sec. 5 and Appendix C.2.

We do, however, test whether trained LBNFs attain the predicted rates. By refining a uniform anchor grid on Sinusoid for p∈{0,…,4}p \in \lbrace 0,\dots,4 \rbrace we show that trained LBNFs are capable of empirically meeting the predicted value rate O(hp+1)O(h^{p+1}) with and without truncation (Figure 7), and the gradient rate O(hp)O(h^{p}) (Figure 9, Appendix C.2). These experiments use a hybrid training procedure that combines gradient descent with Gauss–Newton. Reaching these rates with first-order optimizers such as Adam alone is harder, reflecting a limitation of the optimizer rather than of the inherent capabilities of the representation. Hybrid schemes combining gradient descent with least-squares solves are similarly used to realize convergence rates in other neural partition-of-unity models (Lee et al., 2021).

For the KiloNeRF-style setup, see CW1 (point 1).

Thanks for the experiments and clarifications provided. They partially address both of my concerns. I see that the new results show gain with the polynomial bases, but I still think MLP-based functions offer more flexibility and potentially less tuning as universal function estimators, and I am unsure how practically significant the reported PSNR improvements are given the already very high PSNR values (I might be wrong here, as I work with images and mostly under 30db PSNRs). Regarding the theoretical bounds, I appreciate the point that these are existence guarantees, although this limits how directly they explain the behavior of the trained model. Therefore, I'm happy to maintain my original score of weak acceptance.

Reviewer_cnrjOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a scalable neural field representation for scientific signals. The new representation is a hybrid (classical & neural) local basis neural field (LBNF), that represents a signal using a set of overlapping local polynomial functions placed adaptively across the domain, paired with a lightweight MLP that decodes these local features into signal values. This representation allows accurate derivate estimate, while benefitting from the scalability of local features. Moreover, the work proposes explicit relationship between representation complexity and approximation order. The method is validated on large scientific dataset (including Rayleigh–Taylor instability, Asteroid impact and Supernova) and convergence behavior on smaller analytic benchmarks.

Critical strengths

Presentation

Generally, the paper is well-written, well-structured and easy to follow. The figures, tables for all the experimental results are quite neat as well.

Soundness

The work bridges classical polynomial approximation theory with scalable neural representations. The method is technically sound, and experiments are well-designed to test the claim of fitting large signals and the convergence claims are well-supported with careful evaluation against analytic reference solutions. I could not verify the all the math like proofs/lemmas but generally appear okay.

Significance

The paper addresses an important problem of scalability and accurate higher order derivatives needed for scientific signals. The derivative is not required for the representing visual signals, thus has been generally ignored in most of the prior literature. Thus, the problem is timely.

Originality

I found the connection to classical approximation theory and the convergence analysis of systematically varying approximation order and anchor density on analytic benchmarks, quite original and insightful.

Critical weaknesses

Originality/Significance/Soundness

  • The proposed paper doesn’t properly position itself in in the context of some of the prior literature. For example, [Biao Zhang et al., arXiv 2025] has proposed the same polynomial function basis, the offsetting trick, and the Gaussian/RBF interpolation. So, I have to question the originality of this work, in that context. What is new?
  • Another closest work is [Kairanda et al. 2026]. It is a hybrid grid neural field that support derivatives (similar to the proposed method with Gaussian interpolation) and solves PDEs. It is only mentioned in L120 in passing, but incorrectly described to say its doesn’t support derivatives. This method should be compared to.
  • Overall, proposed method is compared to the grid-based methods (K-Planes, FHash, Instant-NGP, NeuRBF) which are not designed for accurate derivatives (so they trivially fail), but there has been works (He et al. 2026, Kairanda et al. 2026, Chetan et al., CVPR 2025) which support higher order derivatives (thus a fair comparisons), those are precisely the most relevant works that has neither been cited or compared to in the current paper.

Presentation

  • In Sec 3.3, please include appropriate citations for classical approximation theory. Moreover, it is a straightforward implementation of the classical theory? Or did you have to adapt it? If so, how? What part is precisely your contribution? It is unclear where the proof is coming from. The content looks heavily AI generated, where I cannot asses what new things you are bringing to the theory!
  • There are too many equations and math, it’s hard to verify the correctness or even understand the paper without proper visualization of the method architecture/basis functions! The paper only contains result figures. See INGP, Din-Grid, K-plan, Zhang et al., arXiv May 2025 paper for figure inspirations.

References

[He et al. 2026] Hermite-NGP: Gradient-Augmented Hash Encoding for Learning PDEs, ICML 2026

[Kairanda et al. 2026] ∂∞-Grid: A Neural Differential Equation Solver with Differentiable Feature Grids, ICLR 2026

[Zhang et al., arXiv May 2025] efunc: An Efficient Function Representation without Neural Networks, https://arxiv.org/pdf/2505.21319

[Chetan et al., CVPR 2025] Accurate Differential Operators for Hybrid Neural Fields

Questions for authors

Please see the weaknesses.

2 comments
AuthorsOpenReview ↗

Response to cnrj (part 1/2)

We thank the reviewer for recognizing the importance of scalable representations with accurate derivatives and for their positive assessment of our experimental design and approximation analysis. We appreciate the detailed concerns and have revised the paper to address them, including new baseline comparisons, additional ablations, a clearer statement of the theoretical contributions, and a methods figure.

CW1: Novelty relative to efunc (Zhang, 2025) and other works mentioned.

We agree that polynomial bases, coordinate offsets, and Gaussian/RBF interpolation are established techniques from classical numerical methods and do not claim these components individually as novel. Without a nonlinear MLP decoder, they form a classical partition-of-unity (PoU) approximant. efunc applies a related decoder-free representation to 3D SDFs for geometry, and its authors list extensions to other data types as future work. LBNFs target a broader class of scientific signals, including signals with complex multiscale structure, high-frequency oscillations, and sharp interfaces, which we address with three additions:

  1. Neural decoder. Local polynomials alone can be insufficient for approximating the function characteristics typical of scientific data. LBNF jointly learns polynomial features, their locations and a nonlinear neural decoder. In Section 4.2 and Appendix C.5 (Table 4, Figure 13), we directly show how a nonlinear decoder improves reconstruction PSNR and substantially reduces derivative error compared to a classical PoU approximant. On scientific signals efunc's accuracy is also considerably lower than most neural-based comparisons and well below LBNF's (Tables 1, 7, 8).

  2. Approximation rates. Our work connects the representation to classical approximation orders, with explicit O(hp+1)\mathcal{O}(h^{p+1}) value and O(hp+1−ν)\mathcal{O}(h^{p+1-\nu}) derivative guarantees through the decoder (Section 3.3), confirmed empirically (Section 4.2). To our knowledge, none of the related representation works (efunc, Hermite-NGP, ∂∞\partial^\infty-Grid, Chetan et al.) establish approximation rates for their learned representations. efunc appeals only to Taylor's theorem.

  3. Anchor relocation. Adaptive basis placement has a long history in numerical methods. Our contribution is the relocation mechanism itself, which allocates parameters to signal complexity during training, without any signal-dependent initialization. LBNFs start from random anchors and relocate them in a parameter-neutral way through residual-driven split-and-merge combined with gradient updates (Section 3.2). The anchors naturally allocate themselves around complex signal structure. We show the effectiveness of this mechanism (Figure 3, Figure 20). Prior methods use fixed grids (Instant-NGP, K-Planes, Hermite-NGP, ∂∞\partial^\infty-Grid) or rely on signal-dependent initialization. NeuRBF initializes with weighted kk-means and keeps locations fixed during training. efunc does optimize its offsets during training, but initializes them from surface samples and reports this initialization as critical for accuracy.

AuthorsOpenReview ↗

Response to cnrj (part 2/2)

CW2: Relationship to ∂∞\partial^\infty-Grid (Kairanda et al., 2026).

We thank the reviewer for flagging this, as the phrasing in our related work could be misconstrued, and have clarified the wording to reflect ∂∞\partial^\infty-Grid's differentiability. We now compare directly against ∂∞\partial^\infty-Grid on all scientific benchmarks (Tables 1, 7, 8 and Figures 16–19 in the paper). While ∂∞\partial^\infty-Grid maintains derivative accuracy comparable to (and sometimes better than) LBNF, its value accuracy is much lower and its computational cost higher. On Argon bubble, ∂∞\partial^\infty-Grid reaches 36.72 dB versus 78.33 dB for LBNF, with ∂xf\partial_x f MAE of 0.97 versus 0.77 and and roughly 20×20\times higher training memory (6.68GB versus 0.32GB) (Table 1, Figure 16).

CW3: Comparisons with derivative-aware methods.

Our original grid-based baselines already included cubic B-spline variants of K-Planes, Instant-NGP, and FHash with smooth tanh⁡\tanh decoders. These configurations are C2C^2 and support meaningful derivative evaluation rather than trivially failing. We also agree that the cited derivative-aware methods are important comparisons and thank the reviewer for pointing this out. We have expanded the related work (Appendix E) and the experiments accordingly, and summarize these additional comparisons below:

  1. Derivative-aware baselines added. ∂∞\partial^\infty-Grid, Hermite-NGP, and efunc are compared on all scientific signals, with configuration sweeps at matched parameter budgets (Appendix D.2, Tables 1, 7, 8, Figures 16–19). We also apply the post-hoc operator of Chetan et al. to Instant-NGP (Polyfit-NGP) in the same comparisons (Figures 16–19). Results from two of the scientific signal fitting experiments are in the table below:
DatasetMethodPSNR ↑\uparrowNRMSE∇↓\mathrm{NRMSE}_{\nabla} \downarrowNRMSE∇2↓\mathrm{NRMSE}_{\nabla^2} \downarrowTrain mem. (GB) ↓\downarrow
Rayleigh–TaylorLBNF (ours)66.220.020.060.32
D-∞\infty42.430.020.063.68
Hermite-NGP53.040.060.290.46
efunc42.530.020.110.03
Argon bubbleLBNF (ours)78.330.010.100.32
D-∞\infty36.720.010.106.68
Hermite-NGP71.620.173.780.46
efunc40.900.030.150.03

Hermite-NGP and ∂∞\partial^\infty-Grid were originally trained primarily on derivative-based losses. For consistency, we train all methods on values alone.

  1. Post-hoc smoothing (Chetan et al.). We added a new section applying the operator to LBNF, NeuRBF, and Instant-NGP on two analytic benchmarks. While the operator improves less accurate derivatives, it oversmooths interfaces and creases (Appendix C.6, Figure 14). On scientific data, Polyfit-NGP improves Instant-NGP but generally remains behind LBNF (e.g., ∂xf\partial_x f MAE 1.37 versus 0.97 on the Rayleigh–Taylor dataset, Figure 19). The oversmoothing is also visible qualitatively for both the analytic and scientific datasets. See Appendix C.6 for a more thorough discussion.

CW4: Citations and contribution in Sec. 3.3.

We revised Sec. 3.3 to separate classical results from our own contribution and have included additional relevant citations.

  1. What is classical, and where the proof comes from. The approximation rate itself is classical: a partition of unity blending local degree-pp Taylor polynomials attains O(hp+1)O(h^{p+1}) value error and O(hp+1−ν)O(h^{p+1-\nu}) error in the ν\nu-th derivative (Babuška and Suri, 1994). Lee et al. (2021) give value-error rates for learned partitions with least-squares polynomial fits. Both are cited in Sec. 3.3, and our proofs (Appendix B) follow this Taylor/partition-of-unity argument.

  2. What we adapted, and how. The classical argument does not apply directly, because LBNFs pass local polynomial features through a nonlinear neural decoder. We adapt the argument in three places. First, we show the decoder can pass these features through without losing value or derivative accuracy (Lemma 1, Appendix B.1–B.2). Second, in the untruncated regime our Gaussian weights overlap everywhere rather than vanishing outside a small patch, so we bound how quickly they decay in place of the classical compact-support assumption (Appendix B.3). Third, when LBNFs are truncated to evaluate only the KK nearest anchors, we show the value rate survives under a neighborhood-radius assumption (Corollary 2).

  3. Our contribution. To our knowledge, no existing theory for hybrid neural fields characterizes how value and derivative error scale with the resolution of the representation. We establish these rates for LBNFs (Theorem 1, Corollaries 1–2), making pp and hh interpretable accuracy controls, and Figure 7 confirms the predicted value rates empirically for p∈{0,…,4}p \in \lbrace 0,\dots,4 \rbrace, with and without truncation.

CW5: Method visualization. Figure 2 in the paper now illustrates the full LBNF pipeline.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.