acceptodds
Under review as a conference paper at ICLR 2027

Beyond Points: Chebyshev Surrogates for Region Queries on Neural Fields

Abstract

Neural fields provide compact and expressive representations of signals and geometry, mapping coordinates in a continuous input domain to output values. However, many applications require or benefit from reasoning over continuous regions of that domain rather than evaluating the network at individual coordinates. Existing approaches either compute a linear surrogate over each region and reduce it to a pair of output bounds, as in range analysis, or require dedicated architectures and training to answer region queries, such as integrals. Instead, we approximate a trained neural field over a region with a closed-form surrogate, obtained by propagating a Chebyshev polynomial through the network in a single forward pass. The output of each activation is refitted with a discrete cosine transform, which prevents the degree of the surrogate from growing across layers and exposes it as a knob, trading computation against approximation accuracy. Since it describes how the field varies within the region, it enables a broad range of queries directly on the field, without further network evaluations. This lets us locate the sub-regions where the field takes a given value, instead of only accepting, rejecting, or subdividing the region. Moreover, the coefficients of the surrogate give the integral of the field over the region in closed form. We apply our framework to mesh extraction on neural signed distance functions, and to anti-aliased, adaptive-resolution reconstruction of neural images, improving both speed and accuracy over sampling-based and range-analysis methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Implicit Neural Representations, Geometric Queries, Signed Distance Functions

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_i44pOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for querying finite regions of trained neural fields using Chebyshev polynomial surrogates. Affine layers propagate the polynomial coefficients exactly, while DCT-based refitting after each activation controls polynomial degree. The resulting surrogate supports range estimation, approximate root finding, and integration without retraining the network. Experiments demonstrate applications to mesh extraction and anti-aliased reconstruction of neural images.

Critical strengths

A coherent and useful framework for region queries. Retaining a local polynomial representation enables several operations within a common framework, including locating level-set crossings and integrating the represented field. The method applies to existing networks without architectural modifications or retraining. The DCT-based refitting provides a practical way to prevent polynomial degree from growing across layers.

Promising improvements on relevant applications. Table 1 reports substantial acceleration of Marching Cubes while preserving the reported Hausdorff distance, and improvements in both runtime and geometric accuracy for Dual Contouring. The image discretization results provide an additional application beyond geometry. Experiments with sine, Gaussian, and GELU activations also support applicability across multiple smooth activation functions.

Critical weaknesses

The reliability of the empirical error control needs stronger validation. Degree selection relies on fitted relationships analyzed primarily on one network and shape. It is unclear whether these fits transfer across trained networks or require additional calibration, and whether calibration costs are included in the reported timings. Local truncation errors can accumulate and be amplified through subsequent layers. Since approximate ranges are used to discard regions, underestimated errors could cause missed surface components. The limitations section appropriately acknowledges the absence of certified bounds, but empirical measurements of missed intersections and sensitivity to tolerance would strengthen confidence in the approach.

The comparative evaluation and implementation details remain incomplete. Section 5.2 compares against fitting a polynomial directly to network samples and reports lower maximum errors. However, the sampling budgets and computational costs are insufficiently specified, and application-level runtime–accuracy comparisons with this alternative are missing. The benchmark also contains only two images, and per-shape results, hardware, numerical precision, and memory consumption are not reported. Additionally, page 6 refers to a set C and tolerance τ without clearly defining the corresponding selection rule.

Questions for authors

Are the degree-selection fits recalibrated for each trained network? Please clarify the calibration procedure, safety margin, and associated cost.

How often does surrogate-based filtering incorrectly discard regions containing the surface, particularly for thin structures or highly oscillatory fields?

Under matched computational budgets, how does layer-wise propagation compare with direct sampling and polynomial fitting in the final applications?

1 comment

We thank the reviewer for the constructive feedback. We have expanded the evaluation and clarified the calibration procedure, filtering reliability, and computational comparisons.

Q1. Degree selection and calibration. We clarify that surrogate-construction degrees are selected separately for each trained network and configuration, using calibration regions and a conservative empirical error threshold (Sec. 5.2). For mesh extraction, the coarse-stage schedule is shared across shapes, while the fine-stage degrees are selected per network (Appendix C). In the initial submission, we manually selected one schedule based on the results we obtained from the Armadillo network. Now we report the measured selection costs between these approaches and evaluate transfer between networks (Appendix D, Table 4).

Q2. Incorrect filtering decisions. We now test the filter against sign-changing cells and independently sampled surface points across multiple neural SDFs (Appendix D, Table 5). Our selected configuration produces no incorrect exclusions among these references, while lower-degree configurations can fail. We also analyze the effect of retaining approximation remainders (Appendix B, Fig. 8) and clarify that our results do not constitute a certification.

Q3. Matched computational budgets. We now compare propagation with direct Chebyshev fitting at equal construction time and equal approximation error (Sec. 5.2, Fig. 5). Application-level comparisons are reported for mesh extraction at matched runtime (Sec. 5.3, Table 1; Appendix D, Table 3) and image discretization across runtime–accuracy budgets (Fig. 7, Appendix E).

We have additionally expanded the neural-image evaluation, added qualitative comparisons, clarified the filtering rule and implementation details, and documented the limitations of our empirical error control.

Reviewer_CMJpOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a method for evaluating continuous regions of an implicit neural representation by passing functions as input rather than points. Specifically, the paper proposes to pass a Chebyshev polynomial through the network; this can be done by applying the network's affine transformation to the polynomial and then approximating the non-linear function via discrete sampling and then re-fitting of the Chebyshev polynomials. The method offers practical improvements over other techniques for finding zero level sets of neural representations used to represent SDFs, and it is also demonstrated for representing images. Overall it's a very nice technical idea and I quite enjoyed reading the paper.

Critical strengths

The paper is very well written and I was able to follow the method quite easily. I've worked in the area for a while and hadn't considered this approach, so I think it's quite clever!

The rest of the paper is also in good shape: the positioning compared to other techniques seems clear, the method appears novel and clearly conveyed, and the evaluation, while not completely rock-solid (see my notes below) shows some benefits to the method.

I think the paper would be of interest, particularly to folks in computational geometry and graphics.

Critical weaknesses

I have a few nitpicks about the paper. For such an interesting technical approach, I was a little disappointed that only two applications were shown in the end.

Some of the formatting could be improved, too. There's space at the end of the paper, but Fig. 5 in particular was so small that it was very difficult to read the axis labels, and there's no qualitative comparison to the baseline sampling method (are the improvements actually due to anti-aliasing? It would be nice to see this visually).

Likewise there are no qualitative results for the adaptive mesh extraction technique. The method also seems to perform with the same accuracy as the Spelunking baseline, albeit with some computational speedups. It wasn't clear to me why the accuracy should be the same when the introduction seems to imply that the Spelunking method is less accurate at locating the zero level set.

I was also a little bit confused about the paragraph on L402 ("how...compare to fitting a polynomial on network samples?). This paragraph seems to suggest somewhat modest improvements to pointwise error compared to this naive baseline, and I wasn't sure if this comparison was performed fairly, under equal-compute conditions.

Questions for authors
  • Can it work for deeper networks?

  • Was the analysis of L402 ("how does...compare to fitting a polynomial...") conducted with equal compute for each method? For example, if you (1) run the proposed method and then (2) query the network at an increased number of sample points such that the total compute matches (1), and then you fit a polynomial to the output of the network, how does the approximation error generally compare?

  • Section 5.3 Adaptive Mesh Extraction mentions using "the same two-stage strategy", but I don't think this strategy has been introduced yet?

  • The font in Fig. 5 is a little small.

  • How exactly is the super-sampling conducted in Fig. 5 for the sampling baseline? How does the PSNR/time tradeoff change if fewer samples are used for the sampling baseline?

  • In Table 1 the analysis shows the speedups under equal Hausdorff distance, but what if we constrain the time for the proposed method and baseline to be the same — does the proposed method improve the Hausdorff distance?

  • Is this applicable at all to neural radiance fields? How does it fare when applied to high dimensional signals such as a 5D radiance field?

2 comments

We thank the reviewer for the encouraging assessment and constructive suggestions. We have revised the manuscript to address the questions and improve the experimental evaluation.

Q1. Deeper networks. Our formulation is not restricted to a particular network depth, although our experiments cover three- and four-layer networks (Sec. 5.1). Required polynomial degrees increase in later layers (Sec. 5.2, Fig. 3), which limits scalability (Appendix A).

Q2. Direct fitting at equal computational cost. We now compare layer-wise propagation with direct Chebyshev fitting at matched wall-clock budgets (Sec. 5.2, Fig. 5). Across ten independently trained neural SDFs representing different shapes, our method achieves lower approximation error at equal construction time and faster construction at equal error. We also clarify the degree-selection procedure and calibration costs.

Q3. Two-stage mesh extraction. Section 5.3 now explicitly describes coarse filtering followed by higher-degree localization of zero crossings along grid edges. Appendix C details degree selection, while Appendix D evaluates filtering reliability.

Q4. Figure readability and qualitative results. We have redesigned the image comparisons and added qualitative results (Fig. 7), with additional runtime–accuracy plots in Appendix E (Fig. 10). Figure 6 provides a qualitative comparison for mesh extraction.

Q5. Supersampling. Uniform supersampling averages network evaluations at the midpoints of an (n\times n) grid within each pixel. We now vary the sampling density and report the resulting PSNR–runtime trade-offs (Fig. 7, Appendix E).

Q6. Hausdorff distance at equal runtime. We now explicitly compare mesh-extraction methods at matched runtime (Sec. 5.3, Table 1, Appendix D). Our revised extractor also locates zero crossings using surrogate roots, improving reconstruction accuracy. We reran the extraction experiments to evaluate this extension, which explains the updated results.

Q7. Radiance fields. Our formulation extends to higher-dimensional regions (Sec. 4), but its computational cost grows rapidly with input dimensionality. We acknowledge this limitation in Appendix A and leave practical evaluation on five-dimensional radiance fields for future work.

Reviewer_CMJpOpenReview ↗

response

Thanks for the response and for the additional experiments and clarifications. I think this has helped to strengthen the paper and still believe it would be of interest to the community. Overall, I'd maintain my positive rating.

Reviewer_rqi1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studied region queries on neural SDFs. Previous studies use linear functions to bound the region queries, and this paper proposes to improve its tightness and quality using Chebyshev polynomials. The method propogates a Chebyshev polynomial through the neural SDF (basically a MLP); special care is taken when propagating through the sin layers (e.g., using DCT). The polynomial can then give estimated bounds for a set of inputs (region queries) of this MLP; this is useful for checking whether the field crosses a value, or detect whether a marching cube cell is empty, and also improves rendering quality.

Critical strengths

The application of Chebyshev polynomials for range analysis is novel and interesting idea. Previous analysis were mostly based on linear bounds or intervals. The effectiveness of higher order polynomials on neural fields have never been demonstrated in prior works.

Refitting after each activation with DCT to avoid degree explosion is a good idea. It addresses the common problem of polynomial based methods where the order can explode after propagation.

The method does not require retraining and the authors shows good speedups compared to existing methods.

Critical weaknesses

Using polynomials for verifying neural networks and neural network controlled systems is a well-known idea in the field of neural network verification. The difference is application. The mathematical techniques used in this paper is thus not new.

Unlike existing verification methods, this paper drops the remainder term after propagation, so the bounds are not sound. Although the authors acknowledge this limitations, this is a clear distinction to existing work such as Spelunking the Deep. The discarded remainder term can cause soundness issue, such as a cell is incorrectly discarded; but there is no study that looks at this potential failure.

The empirical results on the new application is relatively weak and can be enhanced. Only one mesh table and one image figure with two network sizes are shown. If we consider this paper as an application paper, we hope to see more diverse settings and more tests on its practicability and scalability.

The method is limited to smooth activations and small MLPs. The cost for activation refit is actually high, and it is unclear how this method can scale to higher dimensions. That said, if the paper can demonstrate the scalability honestly and set up a baseline for future work, this paper is acceptable. Many verification based methods do not scale when they were first proposed.

Questions for authors

Can you report the effect of dropping the remainder term? How much does the bound change? Will it lead to significant unsound issue? Especially, we need test cases for shapes with thin features and high-frequency details to see how this method works and whether dropping the remainders is actually safe.

For comparison to spelunking the deep, I suggest using a version without dropping the remainder, since that’s the fair comparison using certified bounds.

The core claim of the paper is better accuracy/cost, so it is important to how the quality varies with different costs (e.g., plot error against wall-clock)

Testing more networks, especially with positional encodings, will be helpful.

2 comments

We thank the reviewer for the positive assessment and constructive feedback. We have revised the manuscript to address the concerns about approximation reliability, computational cost, and experimental scope.

Q1–Q2. Approximation remainders and filtering reliability. We now directly compare affine range analysis, polynomial propagation with retained remainders, and our approximate filter (Appendix B, Fig. 8). Conservative bounds become increasingly loose on larger regions, while discarding the remainder substantially improves filtering. We also evaluate incorrect exclusions on sign-changing cells and independently sampled surface points, including challenging shapes and configurations that produce false rejections (Appendix D, Table 5). These experiments support the practical reliability of our selected settings but do not establish certified bounds.

Q3. Accuracy versus computation time. We now compare layer-wise propagation with direct Chebyshev fitting at matched computational budgets (Sec. 5.2, Fig. 5). Across independently trained SIREN SDFs representing ten different shapes, propagation is 4.2× faster at equal approximation error and achieves 6.1× lower error at equal construction time. We also clarify the degree-selection procedure, calibration costs, and their exclusion from the reported runtimes (Sec. 5.2, Appendix C).

Q4. Additional networks and scalability. We extend the evaluation to Gaussian and GELU activations (Sec. 5.4, Table 2) and discuss how polynomial degree, network depth, and input dimensionality affect computational cost (Sec. 5.2, Appendix A). Novelty and experimental scope. We have clarified the distinction from polynomial methods in neural network verification. Our contribution is to retain the propagated polynomial as a reusable local representation for root finding, integration, and other region queries, rather than using it solely for certification (Secs. 1–2). We have also expanded the mesh-extraction and neural-image evaluations, including matched-runtime comparisons and qualitative results (Sec. 5.3, Figs. 6–7, Appendices D–E).

Thank you for the response. These added results are helpful. I am happy to keep my positive rating of this paper.

Comments on the paper

2 comments

We thank all reviewers for their careful reading and constructive feedback. We have revised the manuscript to address their concerns. All changes introduced in this revision are highlighted in blue.

  • Approximation reliability. We investigate the effect of retaining or discarding approximation remainders and evaluate the reliability of our surface filter on sign-changing cells and independently sampled surface points. We also examine configurations that lead to incorrect filtering decisions and clarify that our approximate range estimates are not certified (Appendices B and D, Fig. 8).
  • Mesh extraction and revised results. We now use the surrogate both to filter cells and to locate zero crossings along grid edges, improving reconstruction accuracy while retaining substantial speedups over uniform extraction. We have rerun the mesh-extraction experiments accordingly. We also provide per-shape evaluations and comparisons at matched runtime (Sec. 5.3, Table 1, Fig. 6, Appendix D).
  • Runtime–accuracy comparisons. Following the reviewers' suggestions, we now compare layer-wise propagation against direct Chebyshev fitting at matched computational budgets. We demonstrate improvements in both construction time and approximation accuracy, and clarify the degree-selection procedure and its calibration costs (Sec. 5.2, Fig. 5, Appendix C).
  • Additional experiments and presentation. We expand the neural-image experiments to more fields and resolutions, including qualitative results and runtime-accuracy comparisons against direct fitting and uniform supersampling. We also evaluate additional activation functions, clarify the extraction procedure, and improve the figures and experimental descriptions (Secs. 5.3–5.4, Fig. 7, Table 2, Appendix E).

Dear Reviewers,

Thank you for your careful reviews. It is encouraging to see that all of you lean toward acceptance. The authors have now responded to each of your comments, and it would be very helpful to hear your thoughts on their rebuttal. In particular, a short note on whether the remaining concerns behind your weak-accept ratings have been addressed would help us finalize a well-grounded decision. If anything is still unclear, please don't hesitate to ask the authors in the discussion thread.

Best regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.