- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper aims to estimate integrals 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 .
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
- 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.
- The REINFORCE variant only requires the Gaussian pre-reflection transition densities and no gradients of , 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
- 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].
- 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).
- The claim that existing diffusion samplers are restricted to 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 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.
- The path-measure perspective [2] and the use of path-space importance weights for diffusion samplers [3] should be cited.
- 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
- The paper reports absolute rather than relative errors. For most Genz targets (Continuous, Gaussian Peak, Product Peak, Corner Peak), decreases by several orders of magnitude with , e.g. for Genz Continuous. The decreasing absolute errors in Fig. 1 and Table 1 therefore seem to largely reflect the scale of , while the relative errors of RDS stay roughly constant in . For instance, FFBQ's errors on all Genz targets with seem to essentially equal (relative errors of ~100%), i.e. its estimates are uninformative, even though its absolute errors appear to improve with dimension.
- 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 , compared to roughly 400 within two standard deviations for the usual funnel with . This also prevents comparison with published results.
- In the hardest settings (Funnel Cube with , Genz Discontinuous with ), all methods have relative errors of 50% or more, so highlighting a best method there is not meaningful.
- Some QMC results look implausible. For the Asian option, the QMC errors are 0.127 () and 0.128 () 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 . 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 with small spread for all , 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.
- 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 for the target restricted to , with a large penalty for trajectories ending outside .
- The dimensions considered are relatively low, whereas recent diffusion samplers have been applied in up to [8].
- For the synthetic targets, mode coverage is only illustrated in . Statistics in higher dimensions showing that all modes are captured would support the sampling claims.
- 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
- Framing the method as an integration method is somewhat unnatural. Since is the normalizing constant of , the paper would be stronger if framed as sampling on bounded domains, with integration following essentially for free.
- 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 , writing suggests a state dependence that is not there; the expression is simply .
Minor points
- Free energies are (up to temperature factors) negative logarithms of .
- is used in the introduction before it is defined.
- Please define (weak convergence?).
- In (19)–(20), denotes a variance increment that already contains the time step. Using the same letter as for the diffusion coefficient is confusing.
- In (22), the notation is not explained and not needed; one can write 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
- 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?
- 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 mentioned in the conclusion?
- Why do you use a reflected variance-exploding rather than a variance-preserving process?
- 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 is used for Digits. Explicit rates for reflected Brownian motion in convex domains are available, e.g. in [9].
- 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?
- 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
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 and leave-one-out baseline 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 , 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 using the penalized target , where 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 . 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.
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 . 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 to be a density, and for a signed no single proposal has zero variance (the best importance proposal is proportional to with variance equal to ). For , the offset works and stays unbiased (), but it multiplies the relative error by , so it helps only when is small compared with . The split 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 for future work.
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. 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.