acceptodds
Under review as a conference paper at ICLR 2027

A Splitting Method for SDE Terminal-Law Estimation

Abstract

In many settings involving stochastic differential equations (SDEs), including diffusion based generative AI, our aim is to accurately generate samples from a distribution. Typically, this is done by generating i.i.d. samples of the SDE. Given a fixed simulation budget, a reasonable way to gain efficiency may be to instead generate a tree of paths through appropriately split partial paths. This suggests improved performance, but one worries about the injected dependence. In this paper, we study this issue comprehensively. We develop a general method to construct the splitting design and analyze its performance for a broad class of error measures, giving insight into the elegant underlying structure in the problem. For the Kolmogorov-Smirnov (KS) distance and maximum mean discrepancy (MMD), we identify the limiting errors of the associated empirical distributions as the simulation budget increases to infinity. We characterize a splitting strategy motivated by a corresponding asymptotic optimization problem. Overall, we observe a 12-26% improvement in mean KS over i.i.d. samples in many settings. For CIFAR-10 and FFHQ experiments, our method reduces the MMD by 9-15%.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
stochastic differential equations, Monte Carlo methods, variance reduction, diffusion models, splitting estimators

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_5zW8OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
3(+1)
Read review
Paper summary

The paper introduces a splitting method for estimating SDE terminal distributions under a fixed simulation budget. The authors derive covariance and limit results for the empirical distribution, optimize a variance-based allocation, and construct a mixture of valid trees. A two-phase procedure estimates the required quantities and is evaluated on SDEs and diffusion models.

Overall, the paper addresses an important problem in simulation-based distribution estimation. The motivation is clear, and the theoretical analysis provides useful insights into the dependence created by shared trajectories.

However, the practical advantage of the full allocation procedure and several aspects of its empirical validation need further clarification.

Critical strengths
  1. Addresses an important problem: how to estimate a terminal distribution more accurately under a fixed simulation budget by sharing partial trajectories.
  2. The theoretical analysis explicitly accounts for sample dependence and provides covariance formulas, limit theorems, and an approximation guarantee for converting relaxed allocations into valid trees.
  3. The evaluation covers multiple SDE and diffusion models, different budgets and split schedules, and comparisons with independent sampling, simpler splitting strategies, and an OU oracle.
Critical weaknesses
  1. The paper emphasizes accurate terminal-distribution estimation. However, the KS allocation minimizes a worst-case variance surrogate rather than the expected KS error itself. Although this distinction is acknowledged, the experiments provide limited explanation of when the surrogate accurately ranks different splitting strategies.
  2. The practical advantage over simpler splitting strategies is not consistently demonstrated. Learned-c is often competitive on CIFAR-10 and FFHQ; in Table 8, with nine split points and budget 100k, its MMD reduction is 12.7%, compared with 8.8% for the full method.
  3. The practical procedure relies on estimated variance contributions and, for KS, a finite query set, whose errors remain outside the theoretical guarantees. Although the appendix provides implementation details, sensitivity to the pilot budget and query count is not evaluated, leaving the robustness of the learned allocation unclear.
  4. The OU oracle comparison is useful, but the claim of close performance needs qualification. At 19 split points, the learned KS reduction is 18.8%, compared with 22.8% for the oracle. The source of this gap is not investigated, and the comparison is limited to the largest budget.
Questions for authors
  1. Could the authors compare the variance surrogate with measured KS across candidate allocations, to clarify whether the stronger performance of some Uniform-c baselines comes from surrogate mismatch or variance-estimation error?
  2. Could the authors provide sensitivity results for the pilot budget and query count, and clarify when the full allocation provides a meaningful advantage over Learned-c?
  3. Could the authors explain the gap in the 19-split OU oracle comparison and extend this comparison to smaller budgets to better evaluate the two-phase procedure?
3 comments

Thank you for the careful review and questions.

W1 / Q1. Does the surrogate rank allocations like KS? On OU process, where the variances are exact, we studied two gaps: between the bounds of Proposition 3, and between KS and the surrogate.

The expected KS error is hard to compute, so we cannot optimize it directly. Proposition 3 bounds it above and below, up to a logarithmic factor, and our allocation minimizes the lower bound. We also minimized the upper bound. After rounding, the dyadic trees produced by the two allocations are identical in all six settings (9, 19 and 39 splits, B=100B = 100k and 5M), and hence the results are exactly the same. Here, the log factor does not change the allocation.

For eight splitting strategies, the surrogate and KS orders agree in 26 of 28 pairs (Spearman 0.93) at both budgets. Hence, we believe the two rank allocations in nearly the same order.

W1 / Q1. Why can Uniform-c do better? Since the surrogate ranks allocations well, the main remaining cause is noise in the pilot's variance estimates. To confirm this, we reran Langevin and EDM with a much larger pilot, not counted in the budget. KS reduction over i.i.d. (%), higher is better:

TaskB, splitsLearnedLearned-cBest Uniform-c
EDM100k, 923.123.922.8
EDM100k, 1925.921.821.8
EDM1M, 922.521.620.3
EDM1M, 1922.723.219.6
Langevin100k, 922.020.624.9
Langevin100k, 1923.424.625.0
Langevin1M, 923.724.124.0
Langevin1M, 1926.625.025.0

With the larger pilot, the learned methods match or beat the best Uniform-c in every EDM case and at B=1B = 1M on Langevin, and come within 3 points at B=100B = 100k on Langevin.

W2 / Q2. When does the full allocation beat Learned-c? Note that the full allocation is strictly more expressive than Learned-c, but Learned-c only has to find a single parameter. Hence, when the pilot estimates are noisy, Learned-c can do better. As the budget grows and more effort is spend on learning the allocations, the full allocation becomes the better choice.

We now score MMD on Inception-v3 and DINOv2 features for the image models, as is standard for images (revised Tables 10 and 11). Here, the full allocation beats Learned-c in 43 of 48 settings. In the setting you cite (9 splits, 100k), the two are within 0.3 points (full allocation 21.9% vs Learned-c 21.6% with Inception-v3, and 21.1% vs 21.4% with DINOv2).

The full allocation scheme is complex to understand and can be sensitive to noise at low budgets. Our aim with developing Learned-c was to come up with a simpler and more robust method that requires estimating fewer parameters. Our experiments show that the simpler method can also be highly competitive, and this should be seen as an additional contribution to a practitioner.

W3 / Q2. Pilot budget and query count. The pilot uses B1=CB0.66B_1 = C B^{0.66} of the budget BB. As suggested, for each task we tried three values of CC, with 9, 19 and 39 split points and 500 runs per setting. The KS reduction of the learned allocation over i.i.d. (%) is:

Task (values of CC)SplitsB=100B = 100kB=200B = 200k
OU (2.5 / 5 / 10)913.5 / 13.4 / 12.213.7 / 14.1 / 11.8
OU1917.0 / 15.9 / 17.116.2 / 16.6 / 15.6
OU3918.8 / 18.3 / 16.419.9 / 18.9 / 17.6
Langevin (2.5 / 5 / 10)918.7 / 18.9 / 15.619.6 / 20.5 / 17.9
Langevin1922.5 / 19.8 / 17.722.1 / 20.3 / 19.6
Langevin3923.9 / 22.0 / 18.825.3 / 24.6 / 20.0
EDM (5 / 10 / 20)9-17.1 / 12.0 / 12.3-2.0 / 18.8 / 15.4
EDM1916.6 / 17.5 / 12.220.2 / 18.5 / 16.4
EDM3918.7 / 17.2 / 12.921.4 / 21.7 / 16.5

Every setting beats i.i.d. except EDM with 9 split points and the smallest pilot, at both budgets. There the pilot is too small, so its estimates are noisy and the full allocation does worse than i.i.d. In that setting, Learned-c, with its single parameter, still gives 18–21% (see W2 for why). Otherwise, the results change by only a few points across pilot sizes.

We did not sweep the query count, which was too expensive to run in the rebuttal period.

W4 / Q3. The OU oracle gap. Our two phase algorithm differs from the oracle in two ways: it estimates the variances from a pilot, and it uses 1,024 instead of 10,200 query thresholds. To isolate these errors, we replaced the estimated variances with the exact ones, which closes most of the gap. The KS reduction rises from 15.9% to 20.6% at B=100B = 100k (oracle 24.0%) and from 18.3% to 20.8% at 5M (oracle 22.7%). The smaller remaining gap comes from the oracle's denser grid. Hence the algorithm is close to the oracle.

We are grateful for your careful review. We believe the new experiments address the questions you raised, and we hope you will consider raising your score in light of them. We are happy to clarify anything further.

Thank you for the detailed response and the additional experiments. The new results have addressed most of my main concerns. In particular, the agreement between surrogate and empirical KS rankings provides useful support for the proposed allocation objective. The expanded comparisons with Learned-c further clarify the benefits of the full allocation, while the exact-variance experiments offer a convincing explanation of the gap to the oracle. The pilot-budget analysis also helps explain the sensitivity observed in some settings. I am therefore raising my overall rating to 3.

Thank you for your reassessment of our paper. We appreciate your time and consideration and would be happy to discuss any further questions or concerns.

Reviewer_o6seOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

In a number of applications (e.g., diffusion generative modeling), one samples from a probability distribution by simulating an SDE. Given an infinite budget, one can approximate that distribution (or measure any property of it) arbitrarily well by considering an arbitrarily large number of samples. But in practice, we may have a finite computational budget, so it may be beneficial to, as much as possible, 'reuse' computational effort. In particular, instead of considering NN completely independent trajectories, we could instead consider trajectories with shared components, since by reusing segments, we can effectively increase the number of samples while keeping computational costs manageable.

Given a 'splitting' procedure, where trajectories are 'branched' to generate many distinct samples from the same initial condition, how should the splitting be done to minimize error given a fixed computational budget? The authors consider this problem and prove a number of mathematical results. They also show, in some simple experiments, that splitting can indeed improve error metrics given a finite budget.

Critical strengths

The problem the authors consider is potentially interesting. Strategies for increasing the effective number of samples it is possible to obtain from a given computational budget are broadly interesting, for example, to people working on diffusion models. I haven't seen an extensive discussion of splitting in that setting.

The mathematical content of the paper is detailed, and although I did not study it in great detail (something that is difficult to do during the review period), it is probably technically correct.

Critical weaknesses

The biggest problem with this paper is that it is very difficult to understand. Why are the authors studying this problem? What does one gain from studying their results in detail? Does their mathematical analysis provide a practical benefit, or not really at present? What are the significant parts of their analysis---and to what extent are their proofs 'routine' (but obviously very complex) stochastic analysis exercises?

I do not think that the Introduction or Sec. 2 convince a reader not already invested in this idea that the results of this paper are worth thinking about in detail. The authors should be much more concrete regarding the motivation and practical consequences of their work. For example, the toy tasks considered in their experiments are far enough from what people generally care about that it is difficult to evaluate them.

Not unrelatedly, the main text is extremely technical and difficult to follow. As much as possible, the authors should convey only the most important parts of their argument and the relevant intuition, and put statements of inessential results or assumptions in the appendices.

The second major problem of the paper is that it is not clear why the authors made various technical choices that strongly influence their results. One major example of this is that their main text analysis focuses on the KS statistic, which is cumbersome and poorly behaved in D>1D > 1 dimensions. But this statistic is completely impractical for analyzing diffusion models, which generally operate in high dimensions; this is weird given that diffusion models are one of the motivating domains the authors mention. The authors do perform an analysis that uses a different error metric (MMD), but it seems conceptually different and it's buried in the appendices. Why?

More broadly, because many of the specific choices and techniques the authors mention are not explained or well motivated (Neyman allocation, PAVA, KS vs MMD...) it is difficult to make sense of their results. It does not help that their experiments appear to involve a variety of nontrivial assumptions not included in the theoretical analysis.

Questions for authors
  1. Why does the main text focus on the KS statistic rather than MMD?
  2. Are there specific downstream tasks/applications where this theory is useful? How could they be used by people that care about diffusion models, for example?
  3. Why assume the nkn_k are integers, and then do the analysis assuming they're real, and then make them integers again in the 'rounding' section? Why not just assume they're integers the entire time?
  4. Are the differences described in the experiments section 'significant'? Not in the statistical sense; would they affect any downstream application?
3 comments

Thank you for the review. We agree that the paper did not motivate the problem well and was too technical, and we revised it accordingly.

W1. Motivation, readability and novelty. We added a paragraph at the end of the introduction that states the results and main contributions and moved statements not needed for the main argument to the appendix (see the general response).

Novelty. While there is a lot of work on estimating expectations of terminal distributions of an SDE, there is limited work on allocating splitting effort across split times to estimate the entire terminal distribution. To the best of our knowledge, the problem we consider is new (Anderson & Ehlert consider a related problem restricted to discrete reaction networks where splitting is allowed only once), and we build a comprehensive framework for SDEs where splitting is allowed by an arbitrary amount in continuous time. Working at this level of generality leads to non-trivial analysis.

Moreover, some technical steps are not standard. In the CLT, the number of samples per root changes with the budget BB, so the estimators at different budgets are in different spaces. To take the limit, we have to place them in a common space and show that the properties we need hold uniformly in BB. This required careful work and is not a routine application of known results. Apart from such steps, the analysis uses well-known tools (empirical-process limits, Neyman allocation, isotonic regression). However, we emphasize that combining them to obtain clean insights is a novel contribution.

W2. "Toy tasks". We agree that our benchmarks were simple. Hence, we added experiments showing that generating samples with our method improves downstream applications too (see Q2/Q4 below).

W3 / Q1. Why KS. KS is indeed cumbersome in higher dimensions, but statistically it is a natural choice. It implies convergence in distribution, and is equivalent to it when the terminal distribution is continuous. For bounded random variables, it upper bounds the Wasserstein-1 distance. Furthermore, the analysis for MMD follows from the analysis for KS. We have now moved the MMD analysis to the main text (see the general response).

W4. Neyman allocation and PAVA. Neyman allocation and PAVA are not design choices, infact one of our contributions is to bring out the elegant underlying structure of the optimal solution, which is characterized by Neyman allocation and PAVA.

Q2 / Q4. Downstream use. In many settings, knowing the entire probability distribution of a phenomena is useful (example, in weather forecasting), since that conveys a complete picture. Furthermore, many perfomance measures are functions of the distribution (e.g., tail risk measures), and estimating the distribution efficiently leads to better estimations of these performance measures. We added the following experiments in Section 7, where the samples are used for a downstream task:

Electricity demand planning. A diffusion model generates scenarios of the next day's electricity demand, and we use them to plan the generation. We measure the regret, which is the extra cost of following this plan instead of the optimal one. Splitting reduces the regret by 28% at 50k denoiser calls, 56% at 500k and 65% at 1M.

Tail-risk estimation. We simulate many 100-dimensional SDEs used in financial modeling. The splitting allocation is designed to reduce the MMD of the terminal distribution, which falls by 45–54%. The RMSE of the estimated mean payoff falls by 20–48%, and that of its tail risk (the 95% CVaR) by 35–57%. This highlights that there are also large gains in the downstream estimates.

Q3. Why relax and then round? Our analysis works with integer branching from the start, so the theory itself goes through without any relaxation. The difficulty is the optimization: choosing the best integer branching factors is a hard combinatorial problem. We therefore first solve the continuous problem exactly (Lemma 2, Theorem 2) and then approximate its solution with trees we can implement. Further, a corollary of Proposition 4 is that, if the variances are known, our allocation is within 6% of the best integer allocation.

We sincerely thank you for your thoughtful and constructive feedback, which has helped us improve the manuscript. We hope that our revisions have adequately addressed your concerns and would be grateful if you could reconsider your assessment. We would be happy to address any further questions or comments.

Thank you again for your feedback. We have made substantial revisions in response to your concerns, focusing on improving presentation and including new experiments to demonstrate the practical value of our framework. We hope these changes address your main concerns, and would be very grateful if you could reconsider your assessment and score. We are happy to discuss any remaining questions.

Thank you for the careful response. These changes and additions are helpful. I confess to still being a bit confused about the paper, but it's possible that this is for trivial reasons (i.e., not my area or don't understand basic things). For now, to recognize your efforts, I will raise my score a little. I'll think carefully during the next phase. I think I just need time to go over things more carefully.

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

The paper studies splitting estimators for estimating the terminal distribution function F(x) = P(X_T≤ x) of an SDE under a fixed simulation budget: particles are copied at several split times so that paths sharing a common history can be reused, thereby reducing Monte Carlo variance. The authors first provide a covariance decomposition of the discrete splitting estimator, splitting the variance into an initial-state term and covariance contributions from each transition. In the single-threshold case, they prove that the optimal allocation is a Neyman allocation with a monotonicity constraint, which can be solved by PAVA. They then establish a continuous-time limit: using a martingale representation, they characterize the covariance density g_x,y(t), prove the covariance limit and a functional CLT for the splitting estimator, and give the convergence rate of the KS error under a bias–variance trade-off. When KS is used as the criterion, the authors transform the allocation problem of "worst-case variance over thresholds" into a designer–adversary minimax problem, and prove that the optimal splitting function is given by the PAVA projection of the covariance density; the finite-budget version is approximated by a mixture of exact dyadic trees via dyadic rounding, with a constant-factor approximation. Finally, they propose a practical two-phase algorithm and validate it on OU, Langevin, EDM-GMM, DDPM CIFAR-10, and LDM FFHQ.

Critical strengths

The theory is complete and the explanation is clear. The authors decompose the covariance of the splitting estimator into contributions from each transition step, and in continuous time use a martingale representation to obtain the covariance density; the single-threshold optimal allocation is a PAVA-based Neyman allocation, while the full-CDF/KS case is transformed via minimax into a mixture of PAVA projections. The theory is complete, yields an algorithm, and explains the experimental phenomenon of "why one should split in the middle."

The method is general and the experiments cover real diffusion models. The framework does not depend on a specific SDE, and corresponding allocations are provided for both KS and MMD; experiments range from 2D OU/Langevin to EDM-GMM, DDPM CIFAR-10, and LDM FFHQ, and report confidence intervals and different split schedules. Tables 5/8/9 show that the learned allocation stably outperforms independent simulation under most budgets.

Critical weaknesses

The practical gain is limited, and the fixed-c baseline is not weaker than the learned allocation in some settings. In Table 7, for EDM with 9 splits and a 100k budget, Uniform-c = 1.5 reduces KS by 22.8%, while the learned allocation reduces it by only 12.5%; at 200k/500k budgets, Uniform-c = 1.5 is also often comparable to or better than the learned allocation. On CIFAR-10/FFHQ, the gap between Learned-c and the learned allocation is usually within the confidence interval. The authors acknowledge that a fixed c must be chosen in advance and is task-dependent, but to support the claim that "the learned allocation is a better design," they should more systematically compare against a Uniform-c tuned with the same pilot budget and clarify on which tasks the automatic nature of the learned allocation brings substantial benefits. This is the main weakness affecting the acceptance decision.

There is a clear gap between the theoretical guarantees and the implementation, and the actual optimality of the KS surrogate objective is not verified. The KS upper and lower bounds differ by a logarithmic factor, so the surrogate objective may deviate from the true KS optimum; the finite query set, MLP regression, variance-estimation error, and PAVA projection in the two-phase procedure are all outside the theory. The experiments also do not report the gap between the learned allocation and the oracle on non-OU tasks (only OU has an oracle).

Questions for authors

(1) Can the authors add a baseline on EDM, CIFAR-10, and FFHQ that tunes Uniform-c by grid/Bayesian search with the same pilot budget as the learned allocation? In Table 7, Uniform-c = 1.5 outperforms the learned allocation at low budgets; do the authors attribute this to pilot estimation error, surrogate-objective bias, or the possibility that a fixed c is already close to optimal for that task? Please explain. (2) The KS upper and lower bounds differ by a logarithmic factor, and the surrogate objective ignores this factor. Can the authors provide an experimental gap between the learned allocation and the true KS optimum (or an approximate oracle with a denser query set)? On non-OU tasks, how large is the impact of the finite query set and MLP estimation error on the allocation? (3) Does the total budget of the two-phase algorithm include the overhead of Phase 1 pilot, MLP training, Frank–Wolfe optimization, and dyadic mixture construction? Compared with independent simulation, what are the memory and parallel efficiency of splitting? (4) The image experiments only report MMD, and CIFAR-10/FFHQ do not compare Uniform-c. Can the authors add KS or other distribution distances (e.g., sliced W1, precision/recall) as well as Uniform-c results? Could the choice of random Fourier feature dimension and bandwidth in MMD affect the ranking of the learned allocation? (5) The theoretical splitting function r(t) is nondecreasing with r(0+) = 1, but the actual dyadic rounding produces a mixture of piecewise-constant trees. Proposition 5 requires Bh_B^² → ∞ and fixed mixture weights; in the actual two-phase procedure, Phase 1 and Phase 2 reuse data and the weights depend on the pilot. Does this break the asymptotic framework? Can the authors provide a finite-sample bound on the rounding error? (6) What is the substantive difference between this paper and existing multilevel splitting, Giles & Haji-Ali’s path branching, and Anderson & Ehlert’s distributional estimation, specifically for "terminal CDF/KS allocation"? Can the authors provide a direct experimental comparison or a clearer boundary of applicability?

2 comments

Thank you for the careful review.

W1 / Q1. Uniform-c vs the learned allocation. The baseline you suggest is already in the paper: Learned-c chooses c from the same pilot and budget, without seeing the test metric.

To check whether the gap comes from noise in the pilot, we reran Langevin and EDM with a much larger pilot (100×100\times the default on Langevin, 5×5\times on EDM), then spent the full budget on splitting. KS reduction over i.i.d. (%), higher is better:

TaskB, splitsLearnedLearned-cBest Uniform-c
EDM100k, 923.123.922.8
EDM100k, 1925.921.821.8
EDM1M, 922.521.620.3
EDM1M, 1922.723.219.6
Langevin100k, 922.020.624.9
Langevin100k, 1923.424.625.0
Langevin1M, 923.724.124.0
Langevin1M, 1926.625.025.0

With the larger pilot, the learned methods match or beat the best Uniform-c in every EDM case and at B=1B = 1M on Langevin. At B=100B = 100k on Langevin they come within 3 points. So the gap is mostly due to pilot noise, not to the surrogate.

We could not run a Uniform-c grid on CIFAR-10 and FFHQ, since they are large models and were too expensive to run within the rebuttal period with the compute available to us.

W2 / Q2. The log factor and approximations. The expected KS error is hard to compute, so we cannot optimize it directly. Proposition 3 bounds it above and below, up to a logarithmic factor, and our allocation minimizes the lower bound. To better understand this gap, on OU process, where the variances are exact, we also minimized the upper bound. After rounding, the dyadic trees produced by the two allocations are identical in all six settings (9, 19 and 39 splits, B=100B = 100k and 5M), and hence the results are exactly the same. Here, the log factor does not change the allocation, suggesting that our surrogate objective's allocation essentially minimizes the KS objective as well.

Moreover, PAVA is not an approximation, since Lemma 2 and Theorem 2 show that the optimum is exactly a PAVA projection. That leaves the finite query set and the variance estimation, which includes the MLP error. To separate them on OU, we used the exact variances with 1,024 queries, where the KS reduction went from 15.9% to 20.6% at B=100B = 100k (oracle 24.0%) and from 18.3% to 20.8% at 5M (oracle 22.7%). The only remaining difference is the oracle's denser grid (10,200 vs 1,024 thresholds), and hence our procedure is close to the oracle.

Q3. Budget and implementation efficiency. Yes, the pilot is counted in B, and its samples are reused. The other steps (MLP fitting, the allocation solve and the mixture construction) are not counted in B, but the end-to-end wall-clock overhead is only 0.7–2.2% on CIFAR-10 and FFHQ.

Q4. Other metrics and MMD settings. The KS distance is impractical to implement in large dimensions. Even though our analysis carries over to most other metrics, it does not extend easily to sliced W1. Instead, we measured the MMD in two other feature spaces, Inception-v3 and DINOv2, while the splitting profile was designed for the model's own output space. In these spaces, splitting still lowers the MMD by 17–26% in every setting. The choice of bandwidth also does not change this. Scoring each of the 8 bandwidths separately, splitting beats i.i.d. sampling in all 144 comparisons on CIFAR-10 and in 134 of 144 on FFHQ.

Q5. Proposition 5, data reuse and rounding. Proposition 5 shows that a mixture of dyadic trees keeps the covariance limit, the functional CLT and the KS limits. It applies to our algorithm after conditioning on the pilot, which fixes the weights. Wrong pilot estimates make the allocation less efficient but do not break the framework, and reused pilot samples do not matter in the limit, since their weight B1/BB_1/B goes to 0. Proposition 4 adds a finite-budget guarantee. At every budget, with known variances, rounding to dyadic trees costs at most a factor 2/(eln⁡2)≈1.062/(e \ln 2) \approx 1.06 in worst-case variance.

Q6. Prior work. Anderson & Ehlert split only once, near the end, to estimate a distribution on a discrete state space, while we split at many times and choose how much to split at each. Multilevel splitting, MLMC and Giles & Haji-Ali's path branching have a different intent: they focus on estimating the expectation of a random variable, such as one rare-event or option probability. Our framework focuses on generating samples from a distribution of interest.

If one wants to estimate an expectation, our method can be used on top of Giles & Haji-Ali's path branching to further improve their performance.

We appreciate your encouraging evaluation and valuable suggestions. We hope our responses and revisions have addressed the points you raised, and we would be grateful if you reassess your rating in light of these clarifications. We are happy to answer any further questions.

Thank you again for your careful review and constructive feedback. We hope our responses and additional results have addressed your concerns, and would be grateful if you could reconsider your assessment in light of them. We are happy to clarify any remaining points.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.