acceptodds
Under review as a conference paper at ICLR 2027

Learning Parametric Integrals from Integrands

Abstract

Parametric integration requires repeatedly evaluating families of integrals with varying integrands and integration domains. Learning a neural representation of such integrals can amortize this computation, but conventionally requires integral supervision, typically obtained by numerically integrating pointwise evaluations of the integrand. We introduce a neural representation that instead learns parametric integrals directly from pointwise integrand supervision. Our key idea is to decompose the integrand into a component that is constructed to integrate to zero and a second component that directly represents the integral. We realize the zero-integral component using the divergence of a neural field and handle parameterized integration domains through a change of variables. Once trained, our representation directly evaluates parametric integrals without numerical integration and remains differentiable with respect to its parameters. Across diverse applications such as filtering, rendering, and integral transforms, our approach achieves lower approximation error than supervision on Monte Carlo integral estimates under the same training-time budget, despite using fewer integrand evaluations overall.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
parametric integration, Monte-Carlo integration, neural fields, signal processing, differentiable optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_xAe1OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed a novel way to learn the integral with supervision without numerically computing it. The key idea is to use the divergence theorem to construct an integral that sums to 0 so that matching the integrand will requires the network to turn a learnable constant to the integral result.

The method is novel to the best of my knowledge. The read out of the integral is very efficient, since it will be a single parameter.

The validation is not complete. The author's claim that only two methods can perform all families of parametric integral is not correct. Li et. al. 2024 has demonstrated how to use antiderivative method to perform automatic integration in none-axis-aligned domain in their section 4.1. Moreover, even with this claim being correct, many application domains the paper validated on are axis-aligned, yet the authors failed to provide comparison with the antiderivative methods on those domains.

Critical strengths

The method is novel to the best of my knowledge. There are some unique advantage of the methods. For example, reading out the integrand could be very efficient.

Critical weaknesses

Validation is incomplete. Lack of critical baselines. The validation also fails to show the unique advantage of the method. Neither did the writing address that.

Questions for authors

L210-212 "Since the firstcomponent integrates to zero by construction, fitting fΘ to f directly learns c≈I, which can be read off after training without numerical integration." Is there any mathematical proof that minimizing this loss will lead to c that approximates I?

1 comment
AuthorsOpenReview ↗

Response to Reviewer xAe1

We thank the reviewer for recognizing the novelty of our method and the efficiency of reading out the integral directly, as well as for identifying the correction needed in our related-work discussion.

W1. Claim about Li et al. (2024). Li et al. (Section 4.1) indeed extend antiderivative-based integration to non-axis-aligned domains through a change of variables. We agree with the reviewer and have corrected our discussion in Section 2 and Table 1 accordingly. We now distinguish the methods by their derivative requirements. Antiderivative-based methods require successive differentiation along every integration dimension, resulting in a derivative order that grows with the dimension. Additionally, at inference, they require network evaluations at all 2d2^d corners of the domain for every query. In contrast, our method requires only first-order derivatives, irrespective of the integration dimension (Section 3.2, Table 1), and answers each query with a single network evaluation.

W2. Missing comparison with antiderivative methods. We have added the requested comparison in new experiments (Fig. 6, Appendix A.3). Under the same protocol as all other methods, our method achieves the lowest error on the Fourier transform and the highest PSNR on bilateral filtering and rendering at every budget. The antiderivative baseline is the strongest baseline at small budgets on bilateral filtering and rendering. On rendering, it attains higher SSIM and lower LPIPS up to 512 evaluations per iteration, and our method leads on all metrics from 1,024 on. The remaining applications are still training, and we will include them and move this comparison into Section 4 in the final version.

A second new experiment (Fig. 7, Appendix A.3) evaluates the neural control variate of Li et al., which on the unit hypercube is the antiderivative baseline with residual sampling at inference. Each residual sample costs an integrand evaluation and an additional network evaluation. Without residual samples, its MSE is 25 times higher than ours on the Fourier transform. At matched accuracy, it needs about 200 times as many evaluations per query as our method on bilateral filtering and about 8,000 times as many on the Fourier transform.

W3. "The validation also fails to show the unique advantage of the method. Neither did the writing address that." We agree that the efficient read-out the reviewer highlights is a unique advantage, in addition to learning the integral from integrand supervision alone. The revision now states this in Section 2 and Section 4.3 and validates it in Appendix A.3 and A.4. Our method answers every query with a single evaluation of the integral field. Antiderivative methods instead evaluate the network at all 2d2^d corners of the domain per query and require derivatives of order dd during training. Neural control variates do not return the integral directly. They estimate the residual by Monte Carlo sampling for every query, and each sample requires an integrand evaluation and an additional network evaluation. Three new results demonstrate our advantage. Fig. 6 shows that our method achieves the lowest error on the Fourier transform and the highest PSNR on bilateral filtering and rendering at every budget (W2). Fig. 7 shows that a neural control variate needs about 200 times as many evaluations per query to match our accuracy on bilateral filtering and about 8,000 times as many on the Fourier transform. Appendix A.4 shows that, for rendering, the training cost is recovered within about half of a single frame.

Q1. Is there a mathematical proof that minimizing the loss gives c≈Ic \approx I? Yes. We prove the stronger result that c=Ic=I is the exact, unique minimizer of the population loss (Appendix A.1, The integral is the unique minimizer). Since the divergence term integrates to zero, the loss separates into a field-fitting term and (c−I)2(c-I)^2. Thus, for any fixed field, the unique minimizing constant is the true integral II. Finite-sample optimization approximates this population optimum.

Reviewer_DgB5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper learns parameterized integrals directly from integrand samples. It represents the integrand as a divergence term with zero integral plus a network predicting the integral. Six applications compare this approach with regression on Monte Carlo integral estimates under equal training-time budgets.

Critical strengths

The construction is simple and produces a differentiable integral predictor. Gains over the tested Monte Carlo baseline are useful, particularly for difficult integrands such as rendering and Fourier transforms. The paper also reports cases where the baseline catches up.

Critical weaknesses
  1. The advance over existing integration methods is unclear. The zero-integral-plus-constant decomposition comes from Ott et al. (2023), and neural control variates already address parametric integration (Müller et al., 2020). Section 4.3 also interprets the proposed term as a control variate. Moreover, single-sample Monte Carlo regression already has the integral as its population squared-loss minimizer. The contribution therefore needs to be established through improved efficiency, rather than access to integrand-only supervision itself.

  2. The comparisons do not establish a substantial practical advantage. The main baseline uses plain Monte Carlo targets. The experiments cover one- and two-dimensional integrals, but do not compare with quasi-Monte Carlo or suitable quadrature targets, or a neural control-variate method. The equal two-hour training budget is useful, but does not show when training an integral predictor becomes cheaper than direct numerical integration at matched accuracy.

Questions for authors
  1. What is the main methodological advantage over Ott et al. and existing parametric neural control variates, and which result demonstrates it?

  2. Was one sample per target included in the Monte Carlo sweep? Does the advantage persist with quasi-Monte Carlo targets under a comparable computational budget?

  3. At matched accuracy, how many queries are needed to recover the training cost compared with direct numerical integration on a representative application?

3 comments
AuthorsOpenReview ↗

Response to Reviewer DgB5

We thank the reviewer for the thoughtful review. Our contribution lies in the combination of both integrand-only supervision and efficiency. Learning from pointwise integrand evaluations removes the need for numerically estimated targets, and it also yields more accurate models under the same training budget and cheaper inference at query time. To validate both claims, we have added comparisons against antiderivative-based methods, integral supervision with quasi-Monte Carlo and quadrature targets, and neural control variates, together with an analysis of the number of queries needed to amortize training. The current revision reports these results on three representative applications in Appendix A.3, while the remaining applications are still training. In the final version, we will include all applications and move the antiderivative, quasi-Monte Carlo, and quadrature baselines into the main comparison of Section 4.

Q1(a). Advantage over Ott et al. (2023). The key advance is amortization across parametric integrals. Ott et al. use a related zero-integral decomposition but learn the integral of a single fixed integrand. Our method learns the integral as a function of the integrand parameters (Section 3.3) and the integration domain (Section 3.4). One trained network then answers any new query with a single evaluation, which enables the six parametric applications in Section 4.

Q1(b). Advantage over neural control variates. This is a different setting. Neural control variates (Müller et al., 2020; Li et al., 2024) do not return the integral directly. They apply their learned control variate at inference time and estimate the residual by Monte Carlo sampling for every query, where each sample requires an integrand evaluation and an additional network evaluation. Their cost per query therefore grows with the target accuracy. In our method, the divergence term acts as a control variate only during training (Section 4.3). Once trained, our method then answers each query with a single network evaluation and no sampling. We nevertheless gladly followed the reviewer's suggestion and conducted a new experiment (Fig. 7, Appendix A.3), which measures how many residual samples per query the neural control variate of Li et al. needs to match our accuracy. Without residual samples, its MSE is 25 times higher than ours on the Fourier transform, and its PSNR is 2.3 dB lower on bilateral filtering. At matched accuracy, it requires about 200 times as many evaluations per query as our method on bilateral filtering and about 8,000 times as many on the Fourier transform.

Q2(a). Was one sample per target included in the Monte Carlo sweep? Yes. Our sweep over the split between the number of targets and the number of samples per target (Section 4.1) includes one sample per target. This setting has high target variance and is never the best split, with 2.2 to 2.9 times the integral error of the best split on the Fourier transform, rendering, and bilateral filtering. We therefore report the best split for each application (Table 3, Appendix A.2).

Q2(b). Does the advantage persist with quasi-Monte Carlo targets? Yes. In new experiments (Fig. 6, Appendix A.3), we add integral supervision with scrambled Sobol (quasi-Monte Carlo) and quadrature targets, under the same per-iteration budget and training time. Quasi-Monte Carlo targets improve on Monte Carlo targets in most settings, but our method still achieves the lowest error on the Fourier transform, up to an order of magnitude below the strongest baseline, and the highest PSNR on bilateral filtering and rendering at every budget. In the final version, we will extend this comparison to all applications and move it into Section 4 and Fig. 4.

Q3(a). How many queries are needed to recover the training cost? Our method pays the integrand-evaluation cost once during training, then answers subsequent queries with a single network evaluation at inference time. Direct numerical integration, in contrast, incurs sampling costs at every query. Let EtrainE_{\mathrm{train}} be the total integrand evaluations during training and NaccN_{\mathrm{acc}} the evaluations that numerical integration needs per query to match our accuracy at inference. The training cost is recovered after Q⋆=Etrain/NaccQ^\star = E_{\mathrm{train}}/N_{\mathrm{acc}} queries (Appendix A.4, Table 4).

Q3(b). On a representative application? We take rendering as the example. Our model is trained with 1.1×1061.1 \times 10^6 integrand evaluations. Even at the optimistic lower bound of one integrand evaluation per pixel (Nacc≥1N_{\mathrm{acc}} \geq 1), far too few to match our accuracy in practice, the training cost is recovered after about half of a single 1920×10801920 \times 1080 frame. In interactive rendering at 60 frames per second, numerical integration spends at least 7.5×1097.5 \times 10^9 integrand evaluations per minute, and the cost recurs every minute. Our training cost is less than 0.02% of this.

Thank you for the additional experiments and clarifications. The stronger baselines provide better evidence of the method’s effectiveness. However, I remain unconvinced that the demonstrated improvements represent a sufficiently substantial advance over existing approaches. I therefore maintain my rating of 2.

We thank the reviewer for acknowledging the stronger evidence, and we briefly summarize what we consider the advance.

Method. Ours is the only approach that learns parametric integrals from integrand evaluations alone and answers each query with a single network evaluation. Integral supervision needs numerical targets, antiderivatives need multiple evaluations per query, and neural control variates must sample at every query.

Results. Across six applications, we outperform all baselines in most settings. On complex integrands, we lead at every budget, often by a substantial margin, precisely where learned integrals are needed most. Neural control variates need 200–8,000x more evaluations per query to reach our quality, and on rendering our training cost is recovered within 0.1% of a single frame.

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

Summary: This paper introduces a way of creating neural surrogates for parametric integration problems of the form I(p,q)=∫Ω(q)f(x;p)  dx.I(p,q) = \int_{\Omega(q)} f(x; p) \; dx. Instead of learning a supervised learning problem (p,q)↦I(p,q)(p, q) \mapsto I(p, q), which requires evaluating ff many times to obtain even a single training data pair; the authors instead learn the integral directly from a decomposition of the integrand: f(x;p)=f0(x;p)+I(p)f(x; p) = f_0(x; p) + I(p) where ∫Ω(q)f0(x;p)  dx=0\int_{\Omega(q)} f_0(x; p) \; dx = 0 and I(p)I(p) does not depend on xx. Then the desired integral is simply I(p)I(p). (This is a simplified form that does not depend on qq, the authors handle irregular domains using a change of variables).

The authors parameterize f0(x;p)f_0(x; p) with the particular form ∇x⋅[w(x)g(x;p)]\nabla_x \cdot [w(x)g(x; p)], where ww is a fixed function vanishing on the boundary ∂Ω(q)\partial \Omega(q) and gg is a learned function. The form of the learned network is fΘ=∇x⋅[w(x)gΘ(x;p)]+IΘ(p),f_\Theta = \nabla_x \cdot [w(x)g_\Theta(x; p)] + I_\Theta(p),

where gΘ(x;p)g_\Theta(x; p) and IΘ(p)I_\Theta(p) are learned neural fields. Then the network can be trained on only pointwise evaluations of ff: L=Ep,x[∥fΘ(x,p)−f(x,p)∥2].L = E_{p, x} \left[ \| f_\Theta(x,p) - f(x, p) \|^2 \right].

Contributions: A method for training neural networks for integration tasks that does not require a training dataset with evaluated integrals, which may be expensive to compute with high accuracy.

Validation: The authors evaluate their proposed method on 6 integration tasks across a variety of dimensions and application domains. By training directly on pointwise evaluations of ff rather than Monte-Carlo estimates of I(p,q)I(p,q), the authors demonstrate faster time-to-quality in training measured by the number of evaluations of ff.

Critical strengths
  1. The article is well-written and pleasant to read. There is a clear central idea, which sounds implausible at first but becomes inevitable by the end. The ideas are explained clearly. and progressively.
  2. The idea is broadly and immediately applicable. Furthermore, it seems obvious that the idea can be tailored and extended by other researchers.
Critical weaknesses

I believe a bit more mathematical rigor is called for in Section 3, given the subject matter. For example

  1. I would prefer to see the mathematical spaces for ff, gg, II, etc. specified exactly
  2. Are we guaranteed that a decomposition of the form (8) exists for every ff in the chosen space? (This does not follow simply from the Divergence Theorem).
  3. The statement on line 260: ("We require T_q to be a diffeomorphism with a tractable Jacobian determinant"). What exactly does "tractable" mean here? What is the space of domains Ω\Omega for which your method is applicable?

I would be fine if some of these mathematical details are added to appendices; they do not have to be in the main text itself.

Preliminary decision summary: I believe the revisions I am asking for are relatively straightforward. Assuming they are provided, I will rate it "weak accept". The paper presents its ideas extremely well, and I have been enriched by the time I spent reading it.

Questions for authors

Please see my questions in "Critical Weaknesses"

2 comments
AuthorsOpenReview ↗

Response to Reviewer tQmN

We thank the reviewer for the positive assessment and precise questions. We address all three points in the revised paper, in Section 3.4 and Appendix A.1.

Q1. Function spaces. We now explicitly specify the function spaces and regularity assumptions in Appendix A.1 (Spaces). In particular, for each parameter, the integrand ff is square-integrable over the unit hypercube. The taper ww is fixed, positive in the interior, and vanishes on the boundary. The learned field g\mathbf{g} is differentiable, and the integral field is constant with respect to the integration variable.

Q2. Does the decomposition exist for every ff? Yes. We establish this in Appendix A.1 (Existence of decomposition). Choosing the constant term as the integral of ff leaves a zero-mean remainder. The classical result of Bogovskii (1979) guarantees that this remainder admits a divergence representation by a vector field with zero boundary trace. This establishes existence for every integrand in the stated function space, and a finite network approximates the decomposition arbitrarily well.

Q3. Meaning of "tractable" and admissible domains. By tractable we meant that the Jacobian determinant is available in closed form. We now state this directly in Section 3.4 and Appendix A.1. Our method applies to any domain that is the image of the unit hypercube under a diffeomorphism with a Jacobian determinant available in closed form, including boxes, intervals, line segments, balls, and spheres.

Thank you for integrating my feedback, this addresses my concerns. I maintain my positive view of the paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank the reviewers for their constructive feedback. We have uploaded a revision with all changes marked in blue. In summary, we have provided:

  • Mathematical details (tQmN, xAe1): function spaces, existence of the decomposition, and the integral as the unique minimizer of our loss (Appendix A.1).
  • New baselines (DgB5, xAe1): antiderivative methods and integral supervision with quasi-Monte Carlo and quadrature targets (Fig. 6), and a neural control variate analysis (Fig. 7), in Appendix A.3.
  • Break-even analysis (DgB5): for rendering, the training cost is recovered within half a frame (Appendix A.4).
  • Clarifications (tQmN, DgB5): the Jacobian determinant must be available in closed form (Section 3.4), and our control variate acts only during training, whereas neural control variates sample the residual at inference (Sections 2 and 4.3).
  • Related work (xAe1): corrected discussion of Li et al. (2024) (Section 2, Table 1).

Fig. 6 currently covers the Fourier transform, bilateral filtering, and rendering, and the new baselines for the remaining applications are still training. In the final version, we will include all applications and move the new baselines into Section 4.

AuthorsOpenReview ↗

Updated General Response (Final Revision)

We thank the reviewers again for their constructive feedback. As announced in our first response, the new baselines have now finished training for all six applications, and we have uploaded our final revision. Below, we summarize all changes relative to the original submission, which are marked in blue.

  • Mathematical details (tQmN, xAe1). We specify the function spaces, prove that the decomposition exists, and show that the integral is the unique minimizer of our loss (Appendix A.1).
  • Additional baselines on all six applications (DgB5, xAe1). Fig. 4 compares our method against an antiderivative baseline and against integral supervision with Monte Carlo, quasi-Monte Carlo, and quadrature targets (Sections 4.1 and 4.2). Across all baselines, our method achieves the highest PSNR or lowest MSE in most settings, particularly at small budgets. Our advantage is largest on harder integrands, such as the Fourier transform, bilateral filtering, and rendering. On low-variance integrands such as motion blur, integral supervision overtakes our method at large budgets, which we report explicitly.
  • Comparison with neural control variates (DgB5, xAe1). Neural control variates address a different setting, as they correct a learned integral with residual samples for every query at inference. We nevertheless compare against the method of Li et al. (2024) (Appendix A.3, Fig. 6). To match our accuracy, it needs about 200, 900, and 8,000 times as many evaluations per query as our method on bilateral filtering, rendering, and the Fourier transform, highlighting the efficiency of our method.
  • Break-even analysis (DgB5). We report after how many queries our one-time training cost is recovered compared to numerical integration. On rendering, direct quasi-Monte Carlo integration needs about 800 samples per pixel to match our accuracy, so our training cost is recovered after about 1,400 queries, less than 0.1% of a single 1920 x 1080 frame (Appendix A.4).
  • Framing (xAe1, DgB5). The abstract, introduction, and contributions now state our advantage explicitly. Our method learns from integrand evaluations alone and is efficient in both training and inference. During training, it requires no numerical targets and achieves lower error under the same training time. At inference, it answers each query with a single network evaluation, whereas antiderivative methods require 2d2^d evaluations and neural control variates additionally sample the residual for every query.
  • Clarifications (tQmN, DgB5). The Jacobian determinant must be available in closed form (Section 3.4). Our control variate acts only during training, whereas neural control variates sample the residual at inference (Sections 2 and 4.3).
  • Related work (xAe1). We corrected the discussion of Li et al. (2024), who extend antiderivative-based integration to parameterized domains, and added the number of network evaluations per query at inference as another distinguishing factor to Table 1 (Section 2, Table 1).

Figure references in our earlier responses refer to the first revision. The former Fig. 6 is now part of Fig. 4, as the new baselines have moved from the appendix into the main paper. The former Fig. 7, the neural control variate comparison, is now Fig. 6.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.