acceptodds
Under review as a conference paper at ICLR 2027

Numerical Integration on Bounded Domains via Learned Reflected Diffusions

Abstract

We study high-dimensional integration on bounded convex domains. Given a non-negative integrand , the goal is to estimate on . Standard Quasi-Monte Carlo does not adapt its evaluation points to the integrand, while effective importance sampling requires a proposal that captures its concentration. Learned diffusion samplers can supply such a proposal, but existing methods are formulated on and do not respect the domain constraint . We instead build the proposal from reflected dynamics that stay in by construction. A reflected diffusion progressively transforms the target distribution toward the uniform distribution on , and a neural score model is trained to reverse this process. A change-of-measure formula for reflected diffusions yields a training objective without additional boundary terms and an unbiased importance-sampling estimator of . For non-smooth integrands, where reparameterization gradients may be undefined or uninformative, we derive a REINFORCE estimator of the training gradient using Gaussian proposal densities before reflection, without evaluating reflected transition densities or differentiating . We evaluate on numerous synthetic targets and real-world problems, demonstrating strong performance of the proposed method, as well as favorable scaling with the dimension of the problem and the number of points used for the Monte Carlo estimate.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
diffusion, reflected diffusion, numerical integration, importance sampling

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_Y56rOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to estimate integrals Z=∫Dr(x)dxZ=\int_{\mathcal D} r(x)\mathrm dx of non-negative functions over bounded convex domains via diffusion-based sampling and importance sampling in path space. Following previous work on reflected diffusion models in the setting where data from the target measure is available [13], the proposal is a controlled reflected SDE which is related to a reflected variance-exploding SDE started from the uniform distribution as a reference process. Its drift is trained by minimizing the reverse KL divergence to the optimal (zero-variance) path measure. Via Girsanov's theorem this leads to an objective as in the Path Integral Sampler [1] and to an unbiased importance-sampling estimator of ZZ.

Two gradient estimators are considered: pathwise gradients through the simulated SDE (RDS-RP), and a REINFORCE estimator with a leave-one-out baseline based on the Gaussian pre-reflection transitions (RDS-RF). The paper focuses on normalizing-constant estimation but also reports sampling metrics. It compares against QMC, Fast Fourier Bayesian quadrature, a normalizing-flow importance sampler (ZüNIS), and PIS combined with a sigmoid transformation on multiple numerical examples.

Critical strengths
  1. Sampling and integration on domains with boundaries is a relevant problem. To my knowledge, reflected dynamics have not yet been used for data-free diffusion-based samplers, and transferring reflected diffusion models from the data-based setting is a natural and interesting idea.
  2. The REINFORCE variant only requires the Gaussian pre-reflection transition densities and no gradients of rr, so it applies to non-smooth integrands. It seems to improve over the pathwise variant in some examples.
Critical weaknesses

The method essentially combines PIS [1] with reflected dynamics known from diffusion models [13]. Beyond this combination, the contribution is limited, the presentation needs improvement, several claims are overstated, and the numerical results are not convincing.

Novelty and relation to prior work

  1. Several theoretical results are standard and should be presented and cited as such.
    • Proposition 3.1.1 (convergence to the uniform distribution) is a classical property of reflected Brownian motion in convex domains; see e.g. [9], which is also used in the proof.
    • Proposition 3.1.2 (zero-variance proposal) is known [1, 10, 11]. It is also immediate, since both path measures share the same transition kernel and differ only in their initial laws.
    • Theorem 3.3.1 is Girsanov's theorem, and the resulting objective is exactly the PIS objective [1] up to the reflected dynamics and the uniform reference marginal in the terminal cost, rather than an objective with "the same structure"; see also Remark A.3 in [2].
    • Remark 3.3.2 is known as well, see e.g. [2].
    • The SDE representation of the optimal reverse law requires a citation in the main text; the time reversal of diffusions goes back to Nelson [4].
  2. The time-reversal and score perspective is not needed for the method. As in [1, 2], the controlled process can be defined directly and compared to the reference via Girsanov's theorem and the terminal tilt. In fact, the proof of the training identity (Thm. E.2.1) does not use the score representation (11).
  3. The claim that existing diffusion samplers are restricted to Rd\mathbb R^d is overstated, since data-free diffusion samplers on compact domains already exist, e.g. the torus variant of Adjoint Sampling [5] and torsional Boltzmann generators [6]. Moreover, all experiments are carried out on the unit cube, which is a rather special domain due to its product structure and coordinate-wise reflections. Note also that the continuous-time results assume a C∞C^\infty boundary and therefore, strictly speaking, do not cover any of the experiments. It would be good to include at least one example on a different domain, e.g. a smooth convex domain, or the Feynman integral computed directly on the simplex instead of mapping it to the cube. Classical constrained MCMC methods, such as projected [15] or mirror [16] Langevin dynamics, should also be mentioned in the related work.
  4. The path-measure perspective [2] and the use of path-space importance weights for diffusion samplers [3] should be cited.
  5. The REINFORCE estimator with a leave-one-out baseline is closely related to the log-variance divergence [2, 11]. For on-policy samples, the gradient of that divergence corresponds to a score-function estimator with a leave-one-out-type baseline [12], and it does not require gradients of the target either. The difficulty with zero-valued regions discussed in Section 4.2 would therefore not arise for the log-variance loss.

Numerical evaluation

  1. The paper reports absolute rather than relative errors. For most Genz targets (Continuous, Gaussian Peak, Product Peak, Corner Peak), ZZ decreases by several orders of magnitude with dd, e.g. for Genz Continuous. The decreasing absolute errors in Fig. 1 and Table 1 therefore seem to largely reflect the scale of ZZ, while the relative errors of RDS stay roughly constant in dd. For instance, FFBQ's errors on all Genz targets with d≥10d\ge10 seem to essentially equal ZZ (relative errors of ~100%), i.e. its estimates are uninformative, even though its absolute errors appear to improve with dimension.
  2. Funnel Cube is a milder, truncated variant of the standard funnel [14] used in [1, 7]. Its conditional scale varies by a factor of 15 over the whole range of x1x_1, compared to roughly 400 within two standard deviations for the usual funnel with σ=3\sigma=3. This also prevents comparison with published results.
  3. In the hardest settings (Funnel Cube with d=50d=50, Genz Discontinuous with d=20d=20), all methods have relative errors of 50% or more, so highlighting a best method there is not meaningful.
  4. Some QMC results look implausible. For the Asian option, the QMC errors are 0.127 (d=10d=10) and 0.128 (d=120d=120) with a very small spread across seeds, and they do not seem to decrease with the number of samples (Figs. 2 and 6). Since randomized QMC with independent scramblings is unbiased, one would expect the errors to fluctuate across seeds and to decrease with NN. Asian options are, moreover, a standard example where QMC performs well [17]. The same pattern appears for Stripe Cube, where QMC has errors of about 10−310^{-3} with small spread for all dd, and ZüNIS shows almost the same offsets. As RDS agrees well with the reference values on both benchmarks, this suggests a problem in the baseline pipeline, e.g. in how the integrand is evaluated.
  5. The baselines do not reflect the state of the art in diffusion-based sampling; PIS is one of the earliest such methods. On Digits, RDS is worse than the best method reported in [7] in terms of ELBO. Also, a natural additional baseline would be an unconstrained diffusion sampler on Rd\mathbb R^d for the target restricted to D\mathcal D, with a large penalty for trajectories ending outside D\mathcal D.
  6. The dimensions considered are relatively low, whereas recent diffusion samplers have been applied in up to d=2500d=2500 [8].
  7. For the synthetic targets, mode coverage is only illustrated in 2d2d. Statistics in higher dimensions showing that all modes are captured would support the sampling claims.
  8. The paper contains many benchmarks with mixed results. Focusing on three or four examples where the method clearly improves over state-of-the-art alternatives would make the evaluation more convincing.

Presentation

  1. Framing the method as an integration method is somewhat unnatural. Since ZZ is the normalizing constant of π=r/Z\pi=r/Z, the paper would be stronger if framed as sampling on bounded domains, with integration following essentially for free.
  2. Several parts of the presentation could be improved:
    • Importance sampling is used ("Our goal is to construct an importance-sampling estimator") before it has been formally introduced.
    • The reflection map should be formally defined.
    • It is unclear why both inward and outward unit normals are needed.
    • The introduction would benefit from contrasting grid-based and Monte Carlo methods.
    • Since ν≡1/vol(D)\nu\equiv 1/\mathrm{vol}(\mathcal D), writing r(X0)/ν(X0)r(X_0)/\nu(X_0) suggests a state dependence that is not there; the expression is simply vol(D)r(X0)\mathrm{vol}(\mathcal D)r(X_0).

Minor points

  • Free energies are (up to temperature factors) negative logarithms of ZZ.
  • π\pi is used in the introduction before it is defined.
  • Please define μt⇒μ0\mu_t\Rightarrow\mu_0 (weak convergence?).
  • In (19)–(20), gk2g_k^2 denotes a variance increment that already contains the time step. Using the same letter as for the diffusion coefficient g(t)g(t) is confusing.
  • In (22), the notation τ\tau is not explained and not needed; one can write Y∼Q←θY\sim\overleftarrow{\mathbb Q}_\theta or its discrete-time analogue.
  • What is called the reparameterization trick is simply differentiation through the simulated SDE.
  • The error bars in the figures are barely visible.
  • Figures should be vector graphics (Figure 4, for example, is not).

Overall, my current assessment is between a clear and a weak rejection: the method essentially combines PIS with reflected dynamics known from diffusion models, and the numerical evaluation has several issues (see above). I will read the authors' responses carefully during the rebuttal and am open to revising my assessment.

Questions for authors
  1. Why are the domains restricted to be convex? Is convexity only needed for the proofs, and if so, for which results? Non-convex but connected domains should in principle work. Disconnected domains, on the other hand, cannot be handled well with reflections; is that correct?
  2. Why is the integrand restricted to be non-negative? Zero-valued regions are handled by adding a positive offset. Why not treat integrands that are bounded from below in the same way, or via the decomposition r=r+−r−r=r^+-r^- mentioned in the conclusion?
  3. Why do you use a reflected variance-exploding rather than a variance-preserving process?
  4. How large is the error between the terminal law and the uniform distribution for the schedules used in the experiments? This error enters as the irreducible term in Remark 3.3.2, and σ1=0.5\sigma_1=0.5 is used for Digits. Explicit rates for reflected Brownian motion in convex domains are available, e.g. in [9].
  5. What is the effect of the time discretization in practice? The boundary behavior of the controlled process is only approximated by the reflection map, and the continuous-time theory does not cover the implemented scheme. In particular, unbiasedness of (26) relies on the discretized reference step leaving the uniform distribution invariant (Prop. E.5.1), which is only verified for the cube. How would this be ensured for other domains?
  6. Why do you use the REINFORCE estimator rather than the log-variance divergence [2, 11]? The latter includes a baseline directly and does not require gradients of the target either.

References

[1] Q. Zhang, Y. Chen. Path Integral Sampler: a stochastic control approach for sampling. ICLR 2022.

[2] L. Richter, J. Berner. Improved sampling via learned diffusions. ICLR 2024.

[3] J. Chen, L. Richter, J. Berner, D. Blessing, G. Neumann, A. Anandkumar. Sequential Controlled Langevin Diffusions. ICLR 2025.

[4] E. Nelson. Dynamical Theories of Brownian Motion. Princeton University Press, 1967.

[5] A. Havens et al. Adjoint Sampling: Highly scalable diffusion samplers via adjoint matching. ICML 2025.

[6] B. Jing, G. Corso, J. Chang, R. Barzilay, T. Jaakkola. Torsional diffusion for molecular conformer generation. NeurIPS 2022.

[7] D. Blessing, X. Jia, J. Esslinger, F. Vargas, G. Neumann. Beyond ELBOs: A large-scale evaluation of variational methods for sampling. ICML 2024.

[8] D. Blessing, L. Richter, J. Berner, E. Malitskiy, G. Neumann. Bridge Matching Sampler: Scalable sampling via generalized fixed-point diffusion matching. ICML 2026 (arXiv:2603.00530).

[9] J. Loper. Uniform ergodicity for Brownian motion in a bounded convex set. J. Theoret. Probab. 33:22–35, 2020.

[10] C. Hartmann, C. Schütte. Efficient rare event simulation by optimal nonequilibrium forcing. J. Stat. Mech. (2012) P11004.

[11] N. Nüsken, L. Richter. Solving high-dimensional Hamilton–Jacobi–Bellman PDEs using neural networks: perspectives from the theory of controlled diffusions and measures on path space. Partial Differ. Equ. Appl. 2, 48 (2021).

[12] L. Richter, A. Boustati, N. Nüsken, F. J. R. Ruiz, Ö. D. Akyildiz. VarGrad: A low-variance gradient estimator for variational inference. NeurIPS 2020.

[13] A. Lou, S. Ermon. Reflected diffusion models. ICML 2023.

[14] R. M. Neal. Slice sampling. Ann. Statist. 31(3):705–767, 2003.

[15] S. Bubeck, R. Eldan, J. Lehec. Sampling from a log-concave distribution with projected Langevin Monte Carlo. Discrete Comput. Geom. 59:757–783, 2018.

[16] K. Ahn, S. Chewi. Efficient constrained sampling via the mirror-Langevin algorithm. NeurIPS 2021.

[17] P. L'Ecuyer. Quasi-Monte Carlo methods with applications in finance. Finance Stoch. 13(3):307–349, 2009.

6 comments
AuthorsOpenReview ↗

Rebuttal (1/4)

We thank the reviewer for a very detailed and thoughtful review, which contains valuable suggestions that helped us improve the presentation of both theoretical and empirical results, as well as helped us identify and fix several issues in our experiments. Below we address the raised weaknesses and answer the questions in order.

Novelty and relation to prior work

1. We agree that uniform invariance, the zero-variance change of measure, Girsanov’s theorem, the KL chain rule and the time-reversal representation are established principles. The revision makes these connections explicit and adds the suggested references. However, we disagree that our theoretical contribution amounts to applying existing sampler results without further work. PIS [2] treats unconstrained dynamics, and reflected diffusion models [11, 12] learn from target samples. None of these works establishes a reflected sampler trained from evaluations of an unnormalized density, with guarantees for estimating its normalizing constant.

Making this construction rigorous required steps that the cited results do not cover directly. Appendices B–C state which assumptions give weak or strong solutions and which notion of uniqueness, and verify them for our domain and coefficients. For example, Lemma C.1.1 proves the exterior-sphere and cone conditions needed for [9]. Appendix C derives the scheduled reverse SDE from classical reflected time reversal, tracking the noise clock, the boundary regulator and the terminal-time assumptions.

Appendix D shows that the regulator is a measurable function of the reflected trajectory, so the Brownian likelihood ratio is a density between trajectory laws, which underlies the exact training identity and the importance weights. We do not claim a new Girsanov theorem: [12, Theorem 14] proves a reflected version, and we verify its use for our construction, including the random initial law. Lemma D.1.1 states this step for general bounded predictable shifts [13, Theorems 5.22–5.23], which also covers controls evaluated at grid times. The resulting loss is the PIS objective with the reflected reference and no boundary cost.

Finally, transferring the estimation bound of [3, Theorem 5] required establishing the reflected path-density identity. Theorem E.5.3 added in the revision (see answer to Q5) also bounds the error probability of the implemented estimator by the expected discrete objective plus log Z at a fixed grid, and Proposition E.5.1 gives exact unbiasedness at any step size whenever the reference step leaves the uniform law invariant. Together with the gradient estimator based on pre-reflection Gaussian variables, these results make the reflected adaptation usable for data-free integration.

2. We agree that the training objective can be derived without explicitly considering time reversal. Nevertheless, the time-reversal perspective serves several purposes in our work. First, it motivates the score-based parameterization of the learned drift. Second, it highlights that the KL divergence is lower-bounded by initial distribution mismatch (Remark 3.3.2), which also motivates the choice of the noise schedule in practice. This is different from PIS, where both the reference and optimal processes start from the same deterministic point zero, so no mismatch arises. Finally, the reverse-time score representation is used in Theorem E.3.2 to decompose the opposite KL into an initial-distribution mismatch and an integrated squared score error, providing an additional interpretation of the accuracy bound in Theorem 3.3.3. We have clarified these motivations in Section 3.3 and stated that time reversal is not required to derive the training objective.

3. We agree that our claim was too broad and we have corrected it in the paper. Data-free diffusion samplers do exist beyond R^d, but to our knowledge only on manifolds without a boundary. Torsional Boltzmann generators and the torsional variant of Adjoint Sampling diffuse on a flat torus with periodic boundary conditions, and Flow Sampling [4] extends to hyperspheres. Their trajectories wrap around, which requires a periodic target, and their dynamics have no boundary term. Our setting has a genuine boundary and a non-periodic integrand, and the dynamics are confined by reflection, which adds a local-time term and a Neumann boundary condition. Thus, the two lines are separate in the problem they solve and in the theory. The revised introduction states this, and Appendix A now discusses these methods together with constrained MCMC.

We focused on the unit cube because many numerical integration problems are naturally formulated on [0,1]d[0,1]^d. As suggested, we added a Feynman-parameter integral computed directly on the simplex (see Figure 2), and observed that working in simplex coordinates improved integral estimation accuracy. We describe the simplex reflection in Appendix F.2, and will add a proof that it preserves uniform distribution on the simplex.

AuthorsOpenReview ↗

Rebuttal (2/4)

4. We note that we already acknowledge in the Related Work section that the path-space perspective was developed in [1, 2] and that path-space importance weights can be used for normalizing-constant estimation [1, 3]. We have now made these connections more explicit in Section 3, also citing [5]. Appendix A now also cites [6] for combining learned diffusions with path-space weighting and SMC resampling.

5. We thank the reviewer for pointing out this connection. On-policy training with log-variance loss and REINFORCE with trajectory-level scalar baseline essentially give the same gradient estimates, with both not requiring target gradients. However, the discrete-time REINFORCE estimator we use differs in the choice of baselines. The standard log-variance estimator uses a full-trajectory log-ratio with a single leave-one-out baseline per trajectory. In contrast, our estimator exploits the sequential structure by using a separate cost-to-go ckc_k and leave-one-out baseline bkb_k at each time step (Algorithm 1), which generally allows to reduce variance further. We also acknowledge that the log-variance divergence allows off-policy training, which could improve sample efficiency and exploration. Investigating this approach for reflected diffusion samplers is an interesting direction for future work.

Numerical evaluation

6. We agree with the reviewer that using the reported absolute errors made assessing the performance less straightforward. In the revised manuscript, we now report relative errors everywhere.

7. We adapt the standard funnel by shifting and rescaling it to fit within the unit cube. This benchmark is intended to test learning complex density geometry on a bounded domain, rather than to enable direct comparison with the original funnel. To avoid confusion, we renamed it to “Bounded Heteroscedastic“ and clarified its purpose in the manuscript.

8. We have removed the highlighting from these cases in the updated table.

9. We are very grateful to the reviewer for pointing this out. We fully agree that some of the reported results appeared strange, and we did overlook this during the initial submission. We investigated the experimental pipeline, and found a bug in the code that aggregated results from different training runs for the plots, which affected some of the reported results. We also fixed a numerical instability in ZüNIS code that resulted in failed runs on several targets. We carefully reran the affected experiments, revised and updated all the results in the main text and Appendix. QMC and ZüNIS errors no longer appear constant across the number of samples. QMC also now shows great performance on the presented Asian option targets, which aligns with previous literature. This is due to the fact that these targets are actually positive on most of the cube, as well as largely smooth and mostly depend on the average of the Brownian path. We now also present the results on the Asian option problems with a higher value of strike K=130K=130, which makes the target much more sparse, and observe that learning-based approaches become more competitive with QMC. In the light of the updates, we have also revised the discussion of the experimental results in the text.

10. The main goal of our work is to extend the theoretical framework of optimal-control-based diffusion samplers to reflected diffusions. Our training objective follows the same on-policy KL minimization principle as PIS, making it the natural baseline. We focus on a single representative training algorithm to demonstrate how this framework extends to reflected diffusions. Incorporating more recent advances on improved training and parameterization is an interesting direction for future work.

We have also implemented the suggested soft-constraint PIS baseline, which operates on Rd\mathbb R^d using the penalized target rtrain(x)=r(ΠD(x))exp⁡(−λ2∥x−ΠD(x)∥22)r_{\mathrm{train}}(x) = r(\Pi_{\mathcal D}(x))\exp{\left(-\frac{\lambda}{2}\lVert x-\Pi_{\mathcal D}(x)\rVert_2^2\right)}, where ΠD\Pi_{\mathcal D} denotes projection onto the domain (see Appendix F.2 for details). During evaluation, we use the original domain-restricted target, assigning zero weight to samples outside D\mathcal D. We included this comparison in the revised manuscript, see Appendix F.6 for full results.

On Digits, we agree that our method falls behind the best results reported in [14] in terms of ELBO. However, we point out that RDS-RP (−154.5) is on par with the best diffusion sampler reported there (GBS, −154.2) and above DDS (−156.1) and PIS (−173.0). It also outperforms all diffusion samplers reported there in terms of W2, only falling behind FAB. We also note that we never claim state-of-the-art sampling performance in the paper, Digits experiment mainly serves as a controlled ablation: the same objective, gradient estimator, architecture and number of steps, so that only the dynamics differ.

AuthorsOpenReview ↗

Rebuttal (3/4)

11. We agree with the reviewer and now state this as a limitation in the paper. We see scaling RDS further, for example by pairing reflected dynamics with the more scalable training objectives of [15], as a natural next step.

12. In the revised manuscript, we provide additional samples from higher-dimensional synthetic targets to illustrate mode coverage beyond 2D (Figures 10-12), and provide a discussion in the main text.

13. In the submitted manuscript, we deliberately avoided cherry picking and reported comparisons on a wide range of benchmarks, including cases where RDS losses, because they tell practitioners which method is better suited to which problem. However, we agree with the reviewer that the presentation of the results could be improved. We removed some of the least interesting benchmarks (synthetic problems with the smallest dimension d=5, Gaussian Peak Genz since Product Peak is also a unimodal centered synthetic target) and removed FFBQ baseline from the plots, only leaving the results for it in Table 2 for reference, since it struggles to scale to almost any problem with dimension 10 or larger. The text now also contains a more focused discussion of the problems where RDS offers most improvements.

Presentation

14. We would like to respectfully disagree with the reviewer and keep the choice of integration as the primary framing of our paper. It is the problem that motivates this work: every application mentioned in our introduction (free energies, rendering, option pricing, model evidence) needs Z itself. Most of the benchmarks we use in our experiments come from numerical integration literature, and the primary metric of our focus is estimation accuracy. We agree that the two problems are closely linked for RDS. The learned process is the importance proposal, so it is also a sampler for π\pi. However, we already state this in the contributions and Section 3.3 and additionally report sampling metrics on Digits. The converse does not hold in general, since samplers such as MCMC do not always provide Z estimates.

15. Regarding the additional points on the presentation, as well as mentioned minor points, we addressed the majority of them in the updated paper text following the suggestions by the reviewer. We kept the “reparameterization trick” term since it is a very standard term that is widely used in generative modeling and variational inference literature. We use the inward normal in the SDE because reflection pushes trajectories into the domain, and the outward normal for the usual Fokker–Planck flux convention. The normals differ only by sign. Reversing the normal leaves the zero-flux condition unchanged, so a single convention would also suffice.

Questions

Q1. We agree that the method should extend to sufficiently regular connected non-convex domains. Pilipenko [9, Theorems 2.5.2–2.5.4] requires uniform exterior-sphere and cone conditions, together with suitable coefficient regularity, rather than convexity. In Lemma C.1.1, we used convexity to verify these conditions for the smooth domain already assumed in our paper. Convexity is, however, an explicit hypothesis of Loper [8, Lemma 1(1)–(2)], which we invoke in Proposition 3.1.1 (formal version C.3.1) for uniform invariance and mixing. We need it to apply that particular result, without claiming it is necessary for mixing in general. The discrete gradient and importance-weighting arguments do not use convexity, although unbiasedness requires a reference update preserving the uniform law. Finally, on disconnected domains, continuous reflection cannot move particles between components. Importance weighting can still correct their relative contributions when all components are covered, but the variance may be large.

Q2. We develop the methodology for non-negative targets since the sampler view needs π=r/Z\pi = r/Z to be a density, and for a signed rr no single proposal has zero variance (the best importance proposal is proportional to ∣r∣|r| with variance equal to (∫D∣r(x)∣dx)2−Z2\left(\int_D |r(x)| dx\right)^2 - Z^2). For r≥−cr \ge -c, the offset works and stays unbiased (Z=Zc−c⋅vol(D)Z = Z_c - c \cdot vol(D)), but it multiplies the relative error by (Z+c⋅vol(D))/Z(Z + c \cdot vol(D))/Z, so it helps only when c⋅vol(D)c \cdot vol(D) is small compared with ZZ. The split r=r+−r−1r = r^{+} - r^{-1} avoids this and lets each part reach zero variance. Both parts have zero-valued regions, which RDS-RF handles. However, this comes at an additional cost of training two samplers for separate targets. That is why we decided to leave empirically quantifying the trade-offs between different approaches to integrating signed rr for future work.

Reviewer_y3DROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper considers the problem of computing integrals Z=∫Dr(x)dxZ = \int_D r(x) dx on a bounded domain DD. The idea is to learn to sample from the normalised probability π=r(x)/Z\pi = r(x)/Z. Writing ZZ as an expectation over samples of π\pi leads to a zero-variance proposal: for ν\nu the uniform probability on D\mathcal{D} it holds Z=Eν[r(x)ν(x)]=Eπ[r(x)ν(x)ν(x)π(x)]Z = E_{\nu}[\frac{r(x)}{\nu(x)}] = E_{\pi}[\frac{r(x)}{\nu(x)}\frac{\nu(x)}{\pi(x)}], where r(x)ν(x)ν(x)π(x)=Z\frac{r(x)}{\nu(x)}\frac{\nu(x)}{\pi(x)} = Z. However, samples from π\pi aren’t readily available. Hence the paper uses time-reversals to sample from π\pi instead. Since the domain D\mathcal{D} is bounded, they use reflected diffusion processes that stay on the domain.

Critical strengths

I found this paper well-written and reasonably easy to follow: although it is quite technical the paper still manages to provide a good level of details, where the ideas are clearly presented. I did not find any mistakes in the maths, and find the paper to be sound. Moreover, I think the idea of using time reversals to compute integrals on bounded domains is interesting and useful. Overall, I thought it was a good paper.

Critical weaknesses
  1. Evaluation: In evaluating the method, the paper provides a graph of error against the number of samples used for the integral computation. However, this does not take training or sampling trajectories into account. I think it would be fair to include a comparison that includes this especially against the Quasi Monte-Carlo method, for example by timing the methods including training and sampling.
  2. Baselines: I think it would also be informative to include a comparison against another couple of classical methods, namely Annealed Importance Sampling and/or Sequential Monte-Carlo which are similar in spirit to the proposed method, with paths of densities between a base density and rr.

Summary: Overall, I think the paper is well-written with an interesting contribution. I don’t have any major weaknesses, except for including a fairer comparison against classical methods. Therefore, I give the paper a weak accept.

Minor:

  • The text in Figures 1 and 2 is very small and hard to read.
  • Tau in Eq. (22) is undefined
  • Line 188: the line spacing is wrong
  • Line 190: Perhaps change “a target π\pi” to “the target π\pi” to make it clear that π\pi was already defined, and this doesn’t hold for any π\pi.
  • Line 212: What’s ℓˉ\bar{\ell} defined as?
Questions for authors

How does the total time of the method compare to the others?

2 comments

We thank the reviewer for the time spent reviewing our work and the positive feedback.

Computational cost. We would like to clarify that our evaluation protocol, described in Appendix F.2, accounts for the training cost of neural-network-based methods. Specifically, the training budget of neural samplers (PIS, ZüNIS, RDS) is aligned with the number of target evaluations, so different checkpoints are evaluated throughout the training run.This allows us to compare methods under approximately matched computational budgets rather than considering only the performance of fully trained models.

Nevertheless, we acknowledge that neural-network-based methods incur additional computational costs from training and trajectory generation. To provide a more complete comparison, we have added wall-clock time measurements, including both training and sampling costs for learning-based approaches (see Figure 7 in Appendix). As expected, training-free methods such as QMC and AIS are substantially faster. Among learning-based approaches, ZüNIS generally has lower computational costs, as it does not require simulating diffusion trajectories. The diffusion-based methods have comparable wall-clock times, with reflected samplers incurring slight additional overhead from applying the reflection operator.

These results highlight the computational overhead of learning-based samplers. However, such methods may offer better estimation accuracy on challenging integrands where QMC performs poorly, compensating for their higher cost. For example, on high-dimensional tasks where the mass is concentrated near a number of different modes of target distribution, like GMM Cube, neural samplers that learn to find these modes outperform QMC by a wide margin.

We have added the discussion alongside the wall-clock plots in Appendix F.4.

Additional baseline. Following the reviewer's suggestion, we have implemented annealed importance sampling (AIS) [1] with constrained Metropolis-adjusted Langevin transitions, using the boundary-rejection approach of [2, Algorithm 2] (see Appendix F.2 for details). Proposals outside the cube are rejected, while feasible proposals undergo the standard Metropolis–Hastings correction, ensuring invariance of each intermediate annealed distribution. We have added this baseline to the experimental comparison in the revised manuscript. For full experimental results see Appendix F.6. It performs well on some targets, but struggles with concentrated high-dimensional distributions such as GMM Cube and Genz Continuous.

Manuscript corrections. We thank the reviewer for identifying these issues and have addressed all of them in the revised manuscript. We also clarify that ℓ←\overleftarrow{\ell} denotes the corresponding quantity associated with the time-reversed process.

References:

[1] Nea et all. Annealed Importance Sampling. Statistics and Computing, 2001.
[2] Zhu et al. Convergence Rates for the Constrained Sampling via Langevin Monte Carlo. Entropy, 2023.

I thank the authors for their response. This has addressed my concerns particularly by including a wall clock comparison of the trained and classical methods which I think gives a fair comparison. After reading the other reviews and rebuttals I maintain my positive score.

Reviewer_rkVjOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The authors present and analyze a numerical integration framework for quantities over bounded convex domains based on generative modeling via reflected diffusion. The manuscript contains a convergence analysis of the terminal distribution of the RVE SDE solution and existence and uniqueness of the model denoising dynamics. Based on a log-RN derivative representation of the model reverse path distribution w.r.t. some reference distribution, an importance sampling estimator Z^\widehat{Z} for Z=∫Dr(z)dzZ = \int_{\mathcal{D}} r(z) dz is constructed. Theorem 3.3.3 controls the deviation of this estimator in terms of the approximation quality of the model reverse path distribution. For the implementation of the training scheme of the diffusion model, a gradient reparametrization through a REINFORCE gradient estimator is defined. Numerical experiments on Genz functions, GMM Cube tests, elliptic two-loop integrals, Asian option pricing and the Digits benchmark show the proposed method among the methods with lowest test error, often with improved scaling to higher dimenstions.

Critical strengths

S1: The paper presents an interesting approach to numerical integration by leveraging generative models. It is a well-done application of the reflected diffusion framework with overall compelling numerical results across a broad set of applications with a sensible selection of baselines.

S2: The paper is overall pleasant to read and, I believe, quite accessible to large parts of the community despite the technicalities.

Critical weaknesses

I do not see a lot in the wrong with this paper but what perhaps should be addressed are parts of the evaluation:

In Fig.3, the evaluation on the Digits dataset, only part of the metrics are presented, Wasserstein and ELBO. The benchmark seems to define further metrics which are omitted and I wonder why. Also, the results in Fig. 2 show that with increasing sample size, error does not always decrease monotonically. This phenomenon is not discussed and could just be due to seeding in which case it would be interesting to see multi-seed evaluation and uncertainty bars/corridors in the plot or an ablation.

Questions for authors
  • What are the authors' motivations for the metric selection in Figure 3? Can they also report the other Digits metrics (MMD, EUBO, partition function error)?
  • Have the authors investigated multi-seed variance (training) for their models? What is the impact of this on monotonicity in Figure 2 and how large are the variances?
4 comments

We thank the reviewer for the time spent reviewing our work and the positive feedback.

Metrics on Digits. We initially reported Wasserstein-2 distance and ELBO because they are standard and widely used metrics for evaluating diffusion-based samplers, capturing sample quality and approximation quality of a partition function, respectively. To provide a more comprehensive evaluation, we have extended the results in the revised manuscript to include the remaining benchmark metrics from (Blessing et al., 2024): MMD, EUBO, and partition function estimation error (see Table 1 in Appendix). Both RDS-RP and RDS-RF beat PIS across 5 out of 6 metrics, so the conclusions drawn from the results remain consistent.

Multi-seed evaluation. We agree that the non-monotonic behavior in Figure 2 may partly reflect stochastic variability in training and estimation. To investigate this, we conducted additional experiments with three independent training random seeds on the 50-dimensional Stripe Cube and 20-dimensional GMM Cube benchmarks (see Figure 8 in Appendix).

Firstly, we would like to clarify that our evaluation protocol (Appendix F.2) accounts for the training cost, aligning it with the number of target evaluations, so for each sample size, a model from a separate training checkpoint is evaluated. Thus, non-monotonicity is partially explained by the varying accuracy of the learned proposal as the training progresses.

We observe more stable training and a nearly monotonic decrease in integration error for numerically well-behaved targets, such as Stripe Cube. In contrast, GMM Cube exhibits larger fluctuations, likely due to the substantial variation in reward magnitudes and extremely small values outside the mixture modes. Such fluctuations could potentially be mitigated through gradient-norm clipping or adjustments to optimization hyperparameters, such as the learning rate.

Dear authors,

thank you for your reply and addressing my questions!

I agree that the extended results from Table 1 further support the RDS methods. Do you have any intuition why the Δlog⁡Zr\Delta \log Z_{\mathrm{r}} metric behaves differently and shows a different ranking than all others?

Regarding multi-seed, I appreciate the additional results shown. Switching to mean relative error seems like a sensible decision. Do the Plots in Figure 8 show the averaged results? If so, what is the meaning of the vertical errors here and in the original plots?

We thank the reviewer for follow-up questions and positive feedback.

Regarding the metric ranking, similar differences between ELBO/EUBO and importance-weighted metrics were also observed by Blessing et al. (2024). One possible explanation is that ELBO and EUBO average log importance weights, whereas importance-weighted estimates involve the logarithm of averaged weights. The latter are more sensitive to rare trajectories with unusually large importance weights, which can sometimes change the ranking of methods, particularly at finite sample sizes.

Regarding multi-seed evaluation, Figure 8 shows relative errors averaged over three independent training runs, with error bars representing the standard deviation across training seeds. In contrast, for all learning-based approaches the error bars in the main figures represent the standard deviation across multiple independent evaluations of one model (computed via Equation 26 in case of RDS) at given checkpoints of one training run. We will clarify this distinction in the figure captions to avoid confusion.

Thank you again for your insights and responsiveness. I think the figure 8 caption deserves clarification but with the additional information provided it is easy to follow. My concerns and questions were resolved and I happily maintain my initial positive assessment of the paper.

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

The paper proposes reflected diffusion samplers for numerical integration on bounded domains. A learned diffusion starts from a uniform distribution and uses reflection to keep samples inside the domain. The authors derive importance weights for estimating the integral and introduce two training variants to handle integrands rr with different properties, including smoothness and discontinuities. Experiments compare these methods with existing baselines on synthetic benchmarks and real-world integration problems.

Critical strengths

This paper addresses an important problem with potential applications beyond the settings studied. The method keeps samples inside the domain and provides computable importance weights for solving problems where much of the function r's mass lies near the boundary.

The proposed RDS-RF is particularly interesting approach for dealing with non differentiable target function. And the authors show empirically its advantages.

Critical weaknesses

A more careful analysis of the additional computational effort spent for the proposed method with training free methods (QMC) would be helpful to evaluate the method.

Comparisons with soft constraint penalties would better establish when reflection provides a practical advantage. Such comparisons should retain valid importance weights and account for invalid or rejected proposals.

Questions for authors
  • Would RDS-RF retain its advantage on Genz Discontinuous and Asian Option dataset if the PIS baseline were also trained using REINFORCE ?

  • Two alternative baselines came to mind while reviewing this work. Could the authors discuss whether they are feasible for this problem? (1) a soft-constraint approach that penalizes leaving the domain during training with a cost function, and (2) a rejection based approach that repeatedly resamples an out-of-domain step until it falls inside the domain? The first comparison would help determine whether the evaluated tasks benefit from hard constraint enforcement or whether a soft penalty is sufficient.

3 comments

We thank the reviewer for the time spent reviewing our work and the positive feedback.

Computational cost comparison. We would like to clarify that our evaluation protocol, described in Appendix F.2, accounts for the training cost of neural-network-based methods. Specifically, the training budget of neural samplers (PIS, ZüNIS, RDS) is aligned with the number of target evaluations, so different checkpoints are evaluated throughout the training run.This allows us to compare methods under approximately matched computational budgets rather than considering only the performance of fully trained models.

Nevertheless, we acknowledge that neural-network-based methods incur additional computational costs from training and trajectory generation. To provide a more complete comparison, we have added wall-clock time measurements, including both training and sampling costs for learning-based approaches (see Figure 7 in Appendix). As expected, training-free methods such as QMC and AIS are substantially faster. Among learning-based approaches, ZüNIS generally has lower computational costs, as it does not require simulating diffusion trajectories. The diffusion-based methods have comparable wall-clock times, with reflected samplers incurring slight additional overhead from applying the reflection operator.

These results highlight the computational overhead of learning-based samplers. However, such methods may offer better estimation accuracy on challenging integrands where QMC performs poorly, compensating for their higher cost. For example, on high-dimensional tasks where the mass is concentrated near a number of different modes of target distribution, like GMM Cube, neural samplers that learn to find these modes outperform QMC by a wide margin.

We have added the discussion alongside the wall-clock plots in Appendix F.4.

Q1. We evaluated PIS trained with REINFORCE on Asian options and Genz Discontinuous, the results are presented in the plots in Appendix F.6. This variant indeed improves upon standard PIS on these environments due to the same reason RDS-RF improves upon RDS-RP (see Section 4.2). Across 6 targets (two additional Asian option variations were added based on the discussion with Reviewer Y56r), RDS-RF showed better performance than PIS-RF on 4, the same performance on 1, and slightly worse performance on 1.

Q2, additional baselines. We have implemented the soft-constraint PIS baseline, which operates on Rd\mathbb R^d using the penalized target rtrain(x)=r(ΠD(x))exp⁡(−λ2∥x−ΠD(x)∥22)r_{\mathrm{train}}(x) = r(\Pi_{\mathcal D}(x))\exp{\left(-\frac{\lambda}{2}\lVert x-\Pi_{\mathcal D}(x)\rVert_2^2\right)}, where ΠD\Pi_{\mathcal D} denotes projection onto the domain (see Appendix F.2 for details). During evaluation, we use the original domain-restricted target, assigning zero weight to samples outside D\mathcal D.

We additionally consider annealed importance sampling (AIS) [1] with constrained Metropolis-adjusted Langevin transitions, following the boundary-rejection approach of [2, Algorithm 2] (see Appendix F.2 for details). Proposals outside the cube are rejected, while feasible proposals undergo the standard Metropolis–Hastings correction, ensuring invariance of each intermediate annealed distribution.

We run these baselines on all benchmarks for integral estimation and extend our plots in Appendix F.6. The additional experiments show that soft-constraint PIS is competitive on some problems, e.g., on GMM Cube at d=20d=20, but its accuracy deteriorates on several higher-dimensional targets. For example, on Stripe Cube at d=50d=50, its relative error is approximately 33%, compared with below 0.01% for both RDS variants. AIS also performs well on some targets, but struggles with concentrated high-dimensional distributions such as GMM Cube and Genz Continuous. These results support the practical benefit of reflection on various benchmarks.

We have added a detailed description of these baselines in Appendix F.2.

References:

[1] Nea et all. Annealed Importance Sampling. Statistics and Computing, 2001.
[2] Zhu et al. Convergence Rates for the Constrained Sampling via Langevin Monte Carlo. Entropy, 2023.

Thank you for your reply and additional results. For computational cost my understanding is for some applications (high-dimensional), learned method like the one proposed by the author performs better though quite costly (Fig 7). And if we look at Figure 7 at roughly the same time cost, AIS is getting close to RDS with more samples? I think this analysis is worth pointing out in the main paper. Also could the author explain more why PIS with soft constraint perform worse at higher dimension?

We thank the reviewer for the follow-up questions and feedback.

We acknowledge that AIS has lower computational cost compared to RDS, and additionally examined if it can produce more accurate estimates at higher sample sizes. Below we report the average relative errors on several targets in dimension d=20d=20:

MethodProduct PeakGMM CubeBounded HeteroscedasticStripe Cube
RDS-RP (N=221N=2^{21})2.4e-4 ± 1.1e-41.3e-2 ± 6.7e-31.7e-2 ± 2.0e-25.9e-5 ± 3.6e-5
AIS (N=221N=2^{21})5.0e-3 ± 3.3e-34.5e-1 ± 2.8e-14.8e-1 ± 8.3e-23.1e-4 ± 2.2e-4
AIS (N=223N=2^{23})3.1e-3 ± 2.1e-34.1e-1 ± 7.6e-24.7e-1 ± 3.6e-26.1e-4 ± 4.7e-4
AIS (N=225N=2^{25})2.6e-3 ± 1.4e-33.9e-1 ± 4.7e-23.5e-1 ± 3.7e-25.0e-4 ± 1.3e-4
AIS (N=227N=2^{27})1.5e-3 ± 1.6e-42.6e-1 ± 9.7e-22.6e-1 ± 8.7e-29.6e-5 ± 6.7e-5

At N=227N=2^{27} samples, AIS catches up to the performance of RDS with N=221N=2^{21} on Stripe Cube (error is higher but std intervals intersect), but still falls behind on other tasks.

AIS with Langevin transitions can struggle to accurately adapt to complex geometries. On Bounded Heteroscedastic, for example, the geometric annealing path keeps particles at the wide end of the funnel until latest steps, and entering the neck in the few late steps is challenging. Reliance on local moves can also lead to problems with accurately capturing well-separated modes. Barriers between modes form early during the annealing, so particles cannot rebalance between modes later. Importance weights absorb the shift in mode masses, which further increases the estimate variance. In the constrained setting, boundary rejection further limits particle movement along trajectories. When a proposal leaves the cube, the particle remains at its current position even as the annealing schedule advances. Consequently, many transitions provide no movement toward the next intermediate distribution, making it harder for particles to track the changing target. RDS does not rely on local moves: its drift is trained on whole trajectories and can distribute mass early in the generative process, moving freely across the domain. It also handles keeping the points inside the domain through reflection, so the are no rejected transitions to limit the movement.

PIS with soft constraint (PIS-penalty in the paper) degrades with dimension for two reasons. First, a sample lies in D\mathcal{D} only if all dd coordinates do. At a fixed penalty strength λ\lambda each coordinate leaks a fixed share of mass outside, so the fraction of the penalized target inside D\mathcal{D} decays exponentially in dd. Samples outside D\mathcal{D} contribute zero during the evaluation, so even a perfectly trained PIS-penalty can produce high-variance estimates (since it uses a different target for training). Increasing λ\lambda to keep the portion of the target inside the domain fixed makes the target increasingly stiff near ∂D\partial \mathcal{D}, which is more difficult to learn and can lead to numerical instability. Second, in high dimensions most of the volume of D\mathcal{D} lies close to ∂D\partial \mathcal{D} (for the cube, 87% within distance 0.02 of a face at d=50d = 50), exactly where the penalized sampler must reproduce a sharp drop in density, so the errors there produce heavy-tailed weights for the samples inside D\mathcal{D}. RDS avoids both: reflection keeps every sample in D\mathcal{D}, so the boundary is handled by construction rather than exclusively learned. We kept λ\lambda fixed in the rebuttal experiments for PIS-penalty due to time limitations, but tuning it could potentially improve the results for this approach.

We will include both discussions in the revision of the paper.

Comments on the paper

2 comments

We thank all reviewers again for their time spent reviewing our work and valuable comments. We updated the paper PDF, which includes requested additional experiments and presentational changes, addressing concerns, questions and suggestions. We summarize the changes below:

  1. New baselines: AIS (Lxuc, y3DR), PIS with soft penalty outside domain (Lxuc, Y56r), PIS trained with REINFORCE on zero-valued targets (Lxuc).
  2. Wall-clock time measurements (Lxuc, y3DR).
  3. Additional metrics on DIGITS (rkVj).
  4. Ablation of multi-seed training variance (rkVj).
  5. Reporting relative errors instead of absolute (Y56r).
  6. Feynman integral (Elliptic Two-Loop) results for RDS computed directly on the simplex (Y56r).
  7. Ablation of the number of discretization steps (Y56r).
  8. Additional sample visualizations (Y56r).
  9. Following the comment of Y56r, we identified several issues in the experimental pipeline, reran the affected experiments, revised and updated all the results in the main text and Appendix.
  10. Revised Introduction and Related work following the suggestions by Y56r.
  11. Extended discussion of experimental results (Y56r).
  12. Added missing references and more explicit connections with existing theoretical results (Y56r).
  13. Added Theorem E.5.3 bounding the error probability through discrete-time loss (Y56r).
  14. Various presentational improvements and minor fixes (y3DR, Y56r).
AuthorsOpenReview ↗

Summary of the discussion

We thank the AC and all reviewers for their time and engagement. We especially thank Reviewer Y56r for their detailed review that helped us improve our paper.

Reviewers rkVj, y3DR, Lxuc. rkVj (additional Digits metrics, multi-seed ablation) and y3DR (wall-clock comparison, AIS baseline) confirmed that their concerns were resolved and kept their positive assessments. For Lxuc, we added wall-clock measurements, soft-constraint PIS, AIS, and PIS trained with REINFORCE, and answered the follow-up questions on AIS at larger sample sizes and on why the penalty baseline degrades with dimension.

Reviewer Y56r

To make going through the revision easier, the table below maps each point of the review to where it is addressed in the revised PDF (numbering follows the review).

PointChangeWhere
W9Fixed result-aggregation bug and ZüNIS instability; reran affected experiments (details in Rebuttal 2/4)Figs. 1–2, Table 2, App. F.6
W6Relative errors throughoutAll plots, Table 2
W10Soft-constraint PIS baseline; AIS baselineSec. 4, App. F.2, F.6
W3Elliptic Two-Loop computed directly on the simplex; claim on existing samplers corrected; constrained MCMC and torus/sphere samplers discussedFig. 2 (RDS simplex-RP), App. F.2; Sec. 1, App. A
W1, W4Previous results and suggested references attributed explicitlySec. 3, App. A
W2Clarified that time reversal is not necessary for the objective, and its roleSec. 3.3
W5, Q6Relation to the log-variance divergenceRebuttal 2/4; will be added to the paper
W7, W8, W13Funnel renamed Bounded Heteroscedastic; highlighting removed where all methods fail; benchmark set trimmedApp. F.1, App. F6, Table 2
W12Mode coverage beyond 2DFigs. 10–12
Q4Choice of σ1\sigma_1 via MMD to uniformFig. 5, App. F.2
Q5Theorem E.5.3 (discrete-time accuracy bound); discretization ablationApp. E.5; Fig. 6, App. F.3
W11, Q2Stated as limitationsSec. 5

After the fixes, RDS achieves the lowest mean error in 17 of 25 settings in Table 2 (with RDS-RF performing better on discontinuous targets and targets with zero-valued regions), while QMC remains best where mass is spread evenly across the boundary (Stripe Cube).

Additional points.

  • Novelty. We agree, and now state explicitly, that the objective is the PIS objective on reflected dynamics. The contribution is making diffusion samplers work on domains with a boundary. To our knowledge the first data-free sampler in this setting, with the path-space results that justify the unbiased weights and accuracy bound, and a REINFORCE estimator that needs neither reflected transition densities nor target gradients.
  • Framing (W14). The motivating applications and most benchmarks require ZZ itself.
  • Baselines (W10). PIS is the controlled comparison (same objective, architecture and steps; only the dynamics differ). On Digits, RDS-RP is on par with the best diffusion sampler (GBS) in Blessing et al. (2024) in terms of ELBO and better in terms of W2.

For the final version, we will also add the simplex-reflection invariance proof, the clarified Figure 8 caption, the log-variance discussion, and the AIS-cost and PIS-penalty discussions requested by Lxuc.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.