acceptodds
Under review as a conference paper at ICLR 2027

ABC-: Gradient-Informed Inference for Discrete Stochastic Simulators

Abstract

Sequential Monte Carlo approximate Bayesian computation (SMC-ABC) enables likelihood-free Bayesian inference with expensive simulators, by distributing simulators across parallel workers. However, its random-walk moves can become ineffective as the parameter dimension grows and when the particle population is small. Gradient-informed proposals offer a potential solution, but stochastic simulators used in scientific contexts such as population genetics, epidemiology and chemical physics often use discrete distributions, which prevent conventional pathwise differentiation. We introduce ABC-, a Metropolis-adjusted Langevin method that uses discrete stochastic automatic differentiation () to guide particle mutation in SMC-ABC, with unbiased and low-variance gradient estimates of an expected log-kernel ABC objective from a single simulation and without resorting to relaxations or learned surrogates. We prove that drift along an unbiased gradient estimate improves the expected objective at the proposal over a random walk with the same covariance, to leading order, which a biased drift does not guarantee, and that fixing the drift's random numbers within each sweep keeps the Metropolis-Hastings correction exact. On Bernoulli generalised linear models, Gaussian mixture models and a multi-population SIR epidemic model, ABC- reduces posterior-mean error of random-walk SMC-ABC by up to two-thirds and matches the random walk's accuracy with at least an order of magnitude fewer particles. These results demonstrate the potential of discrete stochastic differentiation to improve the accuracy of SMC-ABC for discrete simulators in common scientific applications, where conventional gradient-based proposals cannot be deployed.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Approximate Bayesian Computation, Sequential Monte Carlo, Simulation-based inference

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YhiKOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes ABC-D^2, which modifies the adaptive SMC-ABC sampler by replacing its random-walk mutation with a truncated, preconditioned MALA proposal. The drift is the gradient of a surrogate objective, the expected log Gaussian ABC kernel Fε(θ)=log⁡p0(θ)−E[Δ(θ,U)2]/(2ε2)F_\varepsilon(\theta) = \log p_0(\theta) - \mathbb{E}[\Delta(\theta,U)^2]/(2\varepsilon^2). It is estimated unbiasedly from a single simulated dataset using the stochastic automatic differentiation of Arya et al. (2022), so simulators with discrete randomness can be handled without continuous relaxations, such as Gumbel relaxation. To keep the MH correction exact, all drift randomness within a sweep is seeded from a hash of (t,θ)(t,\theta), which makes the drift a deterministic function of θ\theta. Acceptance uses the standard hard-indicator ABC criterion on MM fresh datasets.

Critical strengths
  • A clean, correct and practical idea addressing a real gap. Gradient-informed proposals have been largely unavailable for simulators with discrete randomness, which are common in ABC application areas. The method changes only the proposal mean within an established sampler, so its validity rests on standard MH and SMC arguments. The per-sweep deterministic seeding is a simple way to use a noisy gradient while keeping the MH ratio exact. The paper is very clearly written and the method would be easy to adopt in practice.

  • The authors give a very clear set of experiments. The experiments cover as reasonable range of models and provide clear evidence of the advantages of using a gradient-based proposal over a zeroth-order proposal.

Critical weaknesses
  • The fact that a Langevin proposal outperforms a random walk is not surprising. Optimal-scaling theory for smooth targets shows MALA’s proposal variance can shrink as d−1/3d^{-1/3} rather than d−1d^{-1} for the random walk, so the gap widens with dimension (Roberts & Rosenthal, 1998; Roberts et al., 1997). Theorem 1 formalises the same intuition for a single proposal. Showing that ABC-D$^2$ beats random-walk SMC-ABC therefore confirms a well-understood effect rather than establishing a new contribution.

  • On the experiments, I think there's an issue with the tuning in RWM. The authors use h=1h=1, Γ=2Σ^\Gamma = 2\hat\Sigma and one MH move per sweep. The 2Σ^2\hat\Sigma scaling comes from importance-sampling perturbation kernels in ABC-PMC (Beaumont et al., 2009). As an MH random walk in d=15d = 15–3030, it is far larger than optimal-scaling theory would suggest. Under a Gaussian approximation to the current target, its acceptance rate is roughly 2Φ(−d/2)2\Phi(-\sqrt{d/2}): about 2.5% at d=10d=10, 1.4% at d=12d=12 and 0.6% at d=15d=15. This is before accounting for pseudo-marginal noise from M=5M=5. The size of the reported advantage may therefore be due to an untuned random walk proposal.

  • Given that the range of the experiments is quite limited, more could have been done with the tools that the authors had available to them. For example, alternative gradient estimators in the same drift. These include finite differences with common random numbers (Meeds et al., 2015), score-function estimators and Gumbel-softmax relaxations. The paper argues against them but never tests them. Also, stronger gradient-free SMC-ABC kernels could have been considered, such as adaptively scaled random walks, multiple MH moves per sweep at matched cost (Drovandi & Pettitt, 2011), or locally adaptive kernels (Filippi et al., 2013).

  • The benchmarks are also quite limited in scope. They are low-dimensional (d≤30d \le 30), and the gains are already diminishing at d=30d = 30, even though they have tractable likelihoods. Each contains a single layer of Bernoulli or binomial randomness and none represents the motivating class (Gillespie-type kinetics, stochastic epidemics, coalescents). In those classes, discrete outcomes drive control flow, and the variance of stochastic-AD perturbations is the important question.

  • The methodological and theoretical contributions are modest, and the work is not situated relative to pseudo-marginal methods. The method composes existing well-known components. Proposition 1 follows directly from the fact that MH with any deterministic state-dependent proposal is reversible.

  • Theorem 1 is correct, but its relevance to sampler efficiency is limited. It concerns the surrogate FF at the proposal, not acceptance, jump distance or posterior accuracy. It holds as h→0h \to 0 with a remainder growing like d\sqrt d, whereas the experiments use h=1h = 1. At h=1h = 1 the truncation is likely active, reducing the drift to a fixed-length step along the preconditioned gradient direction. Its leading term is independent of estimator variance, so it does not distinguish stochastic AD from a high-variance score-function estimator.

Questions for authors
  • Can you rerun the random-walk baseline with proper scaling, for example hth_t adapted per sweep to a target acceptance rate, or h2∝1/dh^2 \propto 1/d? Please tune ABC-D$^2$ similarly, noting that MALA’s optimal scaling is h2∝d−1/3h^2 \propto d^{-1/3}. Please also compare at matched simulator cost, including a random walk with several MH moves per sweep. A plot of error against total simulator calls would be more informative than the particle ratios in Figure 2.

  • How is a "simulator call" counted for the drift? Is a stochastic-AD evaluation with dd perturbations counted as d+1d+1 calls? Please also report wall-clock times including stochastic-AD overhead, since the motivation is expensive simulators.

  • How often is the truncation TδT_\delta active in practice? If it is active in most moves, Theorem 1’s untruncated leading term does not describe the regime actually run.

2 comments

We thank the reviewer for the detailed review. It led to the largest changes in the revision: tuned baselines with optimal-scaling step sizes, alternative estimators in the same drift, stronger gradient-free kernels and an event-driven simulator (general response, G1–G6). We reply in two parts.

W1. A Langevin proposal beating a random walk is expected. We agree that with exact gradients of a smooth target this is classical, and both samplers now use their optimal-scaling step sizes, so the gain does not come from an untuned walk. Our contribution is to make such gradients available, and useful, for simulators with discrete randomness, where pathwise gradients do not exist and stochastic AD gives only unbiased but noisy estimates of the gradient of an ABC objective, a case the classical results do not cover. The new Proposition 2 and Corollary 1 show that averaging these estimates across particles reduces their noise and keeps the gain of Theorem 1, provided the gradient dominates the shrinkage and curvature bias of the average, whereas averaging a biased estimator keeps its bias; App. B.1 confirms this numerically.

W2. Random-walk tuning. We agree that c=2c = 2 with one move per sweep was too large a step for a Metropolis random walk in these dimensions. Both samplers now use the step sizes the reviewer suggests, c=2.382/dc = 2.38^2/d for the random walk and c=1.652d−1/3c = 1.65^2 d^{-1/3} for ABC-D2D^2. The number of moves per sweep is tuned for both (up to 96 for the random walk), and both use early rejection (G1, Sec. 3.1, App. C.2). Against this tuned walk, ABC-D2D^2 reaches a common tolerance with up to 20×, 6.1× and 2.3× fewer simulator calls on the SIR, GLM and Gillespie models (Figs. 2 and 3).

W3. Alternative estimators and stronger kernels.

  • Gradient estimators (App. B.1, Fig. 4). We average up to 16,384 estimates of finite differences with common random numbers, score-function estimators (RLOO, ARM, DisARM) and Gumbel-softmax relaxations, as ABC-D2D^2 averages stochastic AD (GLM, d=30d = 30). The unbiased averages approach the average of exact gradients, whereas the relaxations stop at their bias, and stochastic AD overtakes them between 1,024 and 16,384 estimates. A relaxation also replaces the draws that decide the control flow, such as which reaction of the Gillespie model fires next, whereas stochastic AD differentiates the program as written.
  • Drifts without simulation (App. B.2, Fig. 5). On the SIR model, ABC-D2D^2 needs 2.7–4.1× fewer calls than the prior score alone, and its advantage over a Gaussian fitted to the particles grows to 2.8× with d=30d = 30; on the GLM, the prior score in its drift brings most of the gain.
  • Multiple moves per sweep. KK is tuned for both samplers, and Figure 2 compares the total cost of reaching a common tolerance, so additional moves are paid for.

(continued in part 2)

W4. Scope of the benchmarks. We added an event-driven Gillespie birth-death simulator with dd = 3, 10, 15 and 20 (Sec. 3.3, Algorithm 2), in which discrete reaction draws drive the control flow and the likelihood of the summaries is intractable. ABC-D2D^2 needs 1.17, 1.15, 2.07 and 2.31× fewer simulator calls than the random walk, with 95% intervals above one. On the SIR model, the gains now grow with dimension, from 2.7× fewer calls at d=10d = 10 to 7.6× at d=30d = 30.

W5. Contributions and pseudo-marginal methods. We agree that Proposition 1 rests on the standard result that MH with a proposal density evaluable in both directions is reversible, and the proof says so. Its role is to show how a noisy drift becomes such a proposal: the average is fixed before the moves of each sweep, which make no new gradient estimates. It also covers early rejection. We now discuss the method relative to pseudo-marginal MCMC: an ABC MH step is a pseudo-marginal step (Andrieu & Roberts, 2009), and estimated gradients have been used in exact pseudo-marginal chains (Dahlin et al., 2015; Nemeth et al., 2016). The new theory is Proposition 2 and Corollary 1, on how averaging across particles treats unbiased and biased estimators, and how that affects Theorem 1.

W6. Relevance of Theorem 1. Theorem 1 is a local statement about the expected objective at the proposal. App. A.1 now treats the cap as it is used, growing with the step, δ=34h2d\delta = \tfrac34 h\sqrt{2d}: the leading term becomes the inner product of the gradient with the expected capped drift, which stays non-negative for an unbiased drift with symmetric noise however often the cap is active, and the drift, an average of up to 12(d+1)12(d+1) estimates, is close to symmetric. Variance enters at the next order: for a concave quadratic objective with Hessian −H-H, it lowers the gain by h48tr(HC)\tfrac{h^4}{8}\mathrm{tr}(HC), with CC the covariance of the drift estimate. This is where stochastic AD and a high-variance score-function estimator differ, and where averaging helps (Prop. 2, Cor. 1).

Q1. Proper scaling and matched cost. See W2. Figure 2 now compares the total simulator calls, and CPU time, needed to reach a common tolerance, with W2W_2 at those settings, in place of the particle ratios. We can add plots of W2W_2 against total calls for every confirmed setting to the appendix.

Q2. Counting calls and wall-clock time. Each alternative outcome evaluated by stochastic AD is counted as a whole simulation (on the Gillespie model, as one cell's share of a dataset), in addition to one call for the primal dataset of every estimate (App. C.3), so the counts are upper bounds. We now report CPU time on one thread, including the overhead of stochastic AD and of the averaging: speed-ups of 0.87–5.1 on the SIR model, 0.79–1.3 on the GLM and 1.28–2.64 on the Gillespie model.

Q3. How often the truncation is active. It depends on the model (App. C.2). At the settings of Figure 2, the cap is active for 0.1–24% of proposals on the GLM and 1–99% on the SIR model (89–99% in six of nine settings, 97–100% in the second half of the sweeps). On the Gillespie model, for 97–98% with d≤10d \le 10, 15% with d=15d = 15 and 53% with d=20d = 20. So, as the reviewer suspected, the untruncated leading term does not describe most SIR moves, on which only the direction of the averaged gradient enters the proposal; App. A.1 gives the capped version (W6).

We hope these answers and the revision address your concerns. If so, we would be grateful if you would reconsider your score.

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

This paper introduces a novel method to improve Approximate Bayesian Computation in a particular case of discreteness, called ABC-D^2 . The discreteness appears not in terms of the \theta parameters of the model (to be inferred), which are continuous, but discreteness inside the data model (and so in the simulator), e.g. a Bernoulli label or reaction event.

The novelty is a technique to include gradient-informed proposals in the proposal, because the naive approach would lead to zero-valued gradients. Essentially, the method implements a gradient of a surrogate objective used to shift the proposal mean for SMC-ABC, and the contribution is an improved mutation kernel in SMC.

The paper shows experiments with the following models: Bernoulli GLM, Gaussian mixture models, and a multi-population SIR model. The motivation for this contribution is for scientific applications, including chemical reaction events, and population genetics. The experiments show that for equal-accuracy, there are fewer samples from the simulator required in the examples considered.

The authors contrast with other simulation-based inference (SBI) techniques in the introduction. Note that in their nomenclature, they only call neural SBI techniques "SBI", and do not call "ABC" and SBI method. There is a discussion of relative robustness of ABC versus neural SBI.

Critical strengths

The theoretical grounding of the proposed technique is solid. The combination of Theorem 1 and Proposition 1 is a careful treatment showing the validity of their ABC-D^2 extension.

This, combined with the experiments, makes me confident in the methodology for the examples considered in the paper.

In the limitations section, the authors fairly concede that the method struggles as the dimensionality of the target probability density becomes high (~30).

Critical weaknesses

There are two major weaknesses: (i) a lack of fair comparison with the general SBI literature and the SOTA techniques for this task, and (ii) insufficient metrics to really validate the posterior estimation.

(i) " Although alternatively neural simulation-based inference (SBI) can amortise inference across datasets (Cranmer et al., 2020; Lueckmann et al., 2021), its accuracy may suffer when observation data is different from the simulations used for training, as reported in some biological applications (Wang et al., 2024). Under model misspecification experiments, SBI meth- ods’ accuracy and calibration often deteriorate, while ABC is comparatively robust on some tasks (Cannon et al., 2022), although ABC can converge to a pseudo-true parameter (Frazier et al., 2020). ABC thus remains widely used and well characterised amid rapidly evolving SBI methods."

Given that the authors are proposing an SBI method (or, at least a method with a simulator, in which neural posterior estimation would be the standard alternative), excluding any benchmarks or experiments with neural SBI doesn't seem justifiable. In the above paragraph from the introduction, this is justified solely in terms of model misspecification. I would think this would still require (A) a benchmark of the proposed ABC-D^2 method against current SOTA SBI methods for in-distribution data, and (B) also would still require a validation of the claim the ABC is better suited to model misspecification tasks.

On (B), the papers cited do not fully justify the characterisation of ABC as being robust to model misspecification. In general, if the simulator is misspecified (or, equivalently, the data are out of distribution), both neural-SBI and ABC both target the incorrect distribution. The failure mode is different, as the papers discuss.

And, even if (B) is justified, a comparison with neural SBI would be a minimal expectation.

Further to this, alternatives to ABC gradient methods are also not fully explored.

(ii) The paper used the posterior mean as a metric in the benchmarks. This is unusual, as it is characterises only one aspect of the estimated distribution. The experiments use benchmarks that are common in the SBI literature as baselines, but don't use the metrics from the published literature (e.g. C2ST). This both means the metrics are not telling the full story, but also means we cannot compare with benchmarks from the literature. I understand for low N~10, C2ST is not meaningful, but then it is still not clear that the posterior is actually being characterised.

I would note that if the task is just posterior mean estimation, an standard neural network with a MSE loss would achieve this.

Questions for authors

Is the only reason to prefer ABC, and therefore pursue ABC-D^2, the model mispecification question? And, if not, is there another reason not to benchmark this against neural SBI methods, e.g. NPE with normalising flows, which are likely to be more simulation-efficient at a given accuracy level?

2 comments

We thank the reviewer for the positive assessment and for precise suggestions that shaped the revision. The shared changes are summarised in the general response (G1–G6).

W1. Isolating the contribution; baseline tuning.

  • Other gradient estimators (App. B.1, Fig. 4). At draws of the exact posterior (GLM, d=30d = 30), we average up to 16,384 estimates of stochastic AD, score-function estimators (RLOO, ARM, DisARM), finite differences with common random numbers and Gumbel-softmax relaxations, as ABC-D2D^2 averages its estimates. As Proposition 2 predicts, every unbiased average approaches the average of exact gradients, with an error falling as 1/n1/n, whereas the relaxations stop at their bias. Stochastic AD overtakes them between 1,024 and 16,384 estimates.

  • Drifts without simulation (App. B.2, Fig. 5). We replaced the drift by the prior score alone, or by the score of a Gaussian fitted to the particles. On the SIR model, ABC-D2D^2 needs 2.7–4.1× fewer calls than the prior score, and its advantage over the Gaussian fit grows with dd, from a 0.47× speed-up at d=10d = 10 (the fit is cheaper there) to 1.3× at d=20d = 20 and 2.8× at d=30d = 30. On the GLM (dd = 20 to 30), the prior score, which the drift of ABC-D2D^2 includes, brings most of the gain over the random walk, though the Gaussian fit is always further from the exact posterior.

  • Tuning: see G1 and Q1.

W2. Common fixed tolerance. We agree, and adopted this design throughout (G2). Both samplers run to a common target tolerance ε∗\varepsilon^*, a low quantile of the distance of data simulated at the true parameter, with a sweep cap and no acceptance floor. Figure 2 reports the cost of reaching ε∗\varepsilon^* and the W2W_2 at those settings, which separates cost from accuracy.

Q1. Tuning protocol. Now each sampler uses the optimal-scaling step size for its dimension, the number of moves per sweep is tuned for both, ABC-D2D^2 also over ss, nnewn_{\mathrm{new}} and λ\lambda, and the chosen settings are confirmed on fresh seeds (Sec. 3.1, App. C.2). The particle ratios are replaced by the cost of reaching a common tolerance. Comparisons with adaptively scaled random walks are in our reply to Reviewer YhiK (W3).

Q2. The simultaneous bound. The revision no longer uses this criterion (G2). Both samplers run to a common tolerance and are compared by the cost of reaching it, with W2W_2 to the exact psoteior reported at those settings (G2), so no result depends on the bound or on its 10% margin.

**Q3. What survives at a common tolerance. At a common tolerance, both samplers target the same ABC posterior, so their accuracy can differ only through Monte Carlo error, and the gain of ABC-D2D^2 appears mainly as lower cost. On the SIR model, ABC-D2D^2 needs 2.7–20× fewer calls and is at least as accurate, with a lower W2W_2 in four of the nine settings and a higher one in none. On the GLM, it needs up to 6.1× fewer calls from d=20d = 20, at 0.92–1.2 times the W2W_2 of the random walk, so there the gain is in cost rather than accuracy.

Q4. One simulator call is one simulated dataset (App. C.3): NMNM calls for the initial particles, MM for every simulated proposal, accepted or not, and one for the primal dataset of every gradient estimate. Each alternative outcome evaluated by stochastic AD is counted as a whole simulation (about nyn_y per estimate on the GLM. One cell's share of a dataset on the Gillespie model), so the counts are upper bounds. We also report CPU time.

Q5. Drift variance and discrete events. Yes. On the GLM (App. B.1, Fig. 4), more trials per dataset make the average of stochastic AD with all perturbations more accurate relative to the gradient, and that of pruned stochastic AD less accurate.

We hope these answers address the reviewer's concerns.

I'm afraid with only a few minutes before the deadline, I don't reasonably have time to reasonably respond to the points of the authors.

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

This paper presents ABC-D2, a particle update method for the adaptive SMC-ABC sampler of Del Moral et al. (2012), for simulators with discrete randomness that prevents pathwise differentiation. The update is a Metropolis-adjusted Langevin proposal whose gradient estimate is computed with stochastic automatic differentiation (Arya et al., 2022). The paper shows that, for small step sizes, unbiased gradient estimates improve the expected objective at the proposal compared to a random walk with the same covariance. Moreover, if all seeds are fixed within a sweep, the accept/reject step remains exactly valid. Experiments compare the proposed ABC-D2 with a random walk on a Bernoulli GLM, a Gaussian mixture with a shared binary label, and a multi-population SIR epidemic model with binomial survey data, where the proposed approach offers improvements (error reduction and matching of the random walk's accuracy with fewer particles). The observed gains are reduced at dimension 30.

Critical strengths
  • The method is valid and reusable, and the exactness problem is properly solved, for a real challenge in SBI. Discrete randomness blocks standard gradients and forces samplers to fall back on random walks. The paper obtains a working gradient from stochastic AD, and Proposition 1 shows how to keep the accept/reject step exactly correct despite the noisy gradient. The method can be combined with any SMC-ABC run whose simulator can be differentiated by D2.

  • The evaluation is comprehensive and honest about the observed limitations. It uses three models, multiple dimensions and particle counts, disclosing the simulation cost and the diminishing gains at dimension 30. The results indicate a significant efficiency gain in mid to high dimensions.

Critical weaknesses
  • The comparison does not isolate the contribution, and the baseline's tuning is unclear. The proposed approach is only compared with random walks; the introduction mentions other possibilities, such as finite differences and score-function gradients, and from a practical standpoint one would want to see how the proposed approach compares to those as well. It is also not clear how the baselines were tuned and whether the best parameters have been selected: for random walks, keeping a fixed step size may not be the best option in higher dimensions.

  • Experiments comparing ABC-D2 with random walks at a common fixed tolerance ε\varepsilon, rather than letting the accept/reject-rate floor determine each run's final tolerance, would better support the paper's arguments, by separating how well each sampler represents the target from how small a tolerance it can reach. This concerns the GLM and GMM settings, where both samplers stop at the 1.5% acceptance floor.

Questions for authors
  1. Were hh and δ\delta selected by any tuning protocol, and are they held constant across dimensions by design? If not, can you report the particle-ratio and simulator-call comparisons against a random-walk baseline run at its own best step size (or at a matched acceptance rate)?

  2. Can you provide details about the "one-sided simultaneous 95% upper bound" used to assess equal accuracy? In particular, how it is constructed and how sensitive the conclusions are to the 10% margin?

  3. When both samplers run to a common fixed tolerance ε\varepsilon on a representative subset of settings, how much of the error reduction survives?

  4. Please clarify the cost accounting: what exactly is one "simulator call"?

  5. Is there any relationship between the drift variance and the number of discrete events per simulated dataset?

1 comment

We thank the reviewer for the positive assessment and for precise suggestions that shaped the revision. The shared changes are summarized in the general response (G1–G6).

W1. Isolating the contribution; baseline tuning.

  • Other gradient estimators (App. B.1, Fig. 4). At draws of the exact posterior (GLM, d=30d = 30), we average up to 16,384 estimates of stochastic AD, score-function estimators (RLOO, ARM, DisARM), finite differences with common random numbers and Gumbel-softmax relaxations, as ABC-D2D^2 averages its estimates. As Proposition 2 predicts, every unbiased average approaches the average of exact gradients, with an error falling as 1/n1/n, whereas the relaxations stop at their bias. Stochastic AD overtakes them between 1,024 and 16,384 estimates.

  • Drifts without simulation (App. B.2, Fig. 5). We replaced the drift by the prior score alone, or by the score of a Gaussian fitted to the particles. On the SIR model, ABC-D2D^2 needs 2.7–4.1× fewer calls than the prior score, and its advantage over the Gaussian fit grows with dd, from a 0.47× speed-up at d=10d = 10 (the fit is cheaper there) to 1.3× at d=20d = 20 and 2.8× at d=30d = 30. On the GLM (dd = 20 to 30), the prior score, which the drift of ABC-D2D^2 includes, brings most of the gain over the random walk, though the Gaussian fit is always further from the exact posterior.

  • Tuning: see G1 and Q1.

W2. Common fixed tolerance. We agree, and adopted this design throughout (G2). Both samplers run to a common target tolerance ε∗\varepsilon^*, a low quantile of the distance of data simulated at the true parameter, with a sweep cap and no acceptance floor. Figure 2 reports the cost of reaching ε∗\varepsilon^* and the W2W_2 at those settings, which separates cost from accuracy.

Q1. Tuning protocol. Now each sampler uses the optimal-scaling step size for its dimension, the number of moves per sweep is tuned for both, ABC-D2D^2 also over ss, nnewn_{\mathrm{new}} and λ\lambda, and the chosen settings are confirmed on fresh seeds (Sec. 3.1, App. C.2). The particle ratios are replaced by the cost of reaching a common tolerance. Comparisons with adaptively scaled random walks are in our reply to Reviewer YhiK (W3).

Q2. The simultaneous bound. The revision no longer uses this criterion (G2). Both samplers run to a common tolerance and are compared by the cost of reaching it, with W2W_2 to the exact psoteior reported at those settings (G2), so no result depends on the bound or on its 10% margin.

**Q3. What survives at a common tolerance. At a common tolerance, both samplers target the same ABC posterior, so their accuracy can differ only through Monte Carlo error, and the gain of ABC-D2D^2 appears mainly as lower cost. On the SIR model, ABC-D2D^2 needs 2.7–20× fewer calls and is at least as accurate, with a lower W2W_2 in four of the nine settings and a higher one in none. On the GLM, it needs up to 6.1× fewer calls from d=20d = 20, at 0.92–1.2 times the W2W_2 of the random walk, so there the gain is in cost rather than accuracy.

Q4. One simulator call is one simulated dataset (App. C.3): NMNM calls for the initial particles, MM for every simulated proposal, accepted or not, and one for the primal dataset of every gradient estimate. Each alternative outcome evaluated by stochastic AD is counted as a whole simulation (about nyn_y per estimate on the GLM. One cell's share of a dataset on the Gillespie model), so the counts are upper bounds. We also report CPU time.

Q5. Drift variance and discrete events. Yes. On the GLM (App. B.1, Fig. 4), more trials per dataset make the average of stochastic AD with all perturbations more accurate relative to the gradient, and that of pruned stochastic AD less accurate.

We hope these answers address the reviewer's concerns.

Comments on the paper

1 comment

We thank the three reviewers for careful and constructive reviews. They found the method valid and reusable, with the exactness problem properly solved, its theoretical grounding solid, and the idea clean, correct and practical, addressing a real gap. The main concerns were shared: the tuning of the random-walk baseline, the comparison at an acceptance floor instead of a common tolerance, the posterior-mean metric, the missing neural SBI baseline, untested alternative gradient estimators, and the lack of a simulator whose control flow depends on its discrete draws. We have uploaded a revision that addresses each of them. Section, figure and appendix numbers in our replies refer to it.

G1. Tuned baselines with optimal-scaling step sizes. Both samplers now use the step size that optimal scaling theory prescribes for the dimension: c=2.382/dc = 2.38^2/d for the random walk and c=1.652d−1/3c = 1.65^2 d^{-1/3} for ABC-D2D^2, with proposal covariance cΣtc\Sigma_t. The number of MH moves per sweep is tuned for both (up to 96 for the random walk), and ABC-D2D^2 also over its drift settings. Settings are selected on tuning seeds and confirmed on fresh seeds (Sec. 3.1, App. C.2). Both samplers use early rejection, and the drift cap now grows with the step and the dimension.

G2. A common target tolerance, compared by cost. Every run stops at a common target tolerance ε∗\varepsilon^* or a sweep cap, never at an acceptance floor. Figure 2 compares the simulator calls and CPU time each sampler needs to reach ε∗\varepsilon^*, each at its cheapest confirmed setting that reaches it in at least 90% of its runs, and the accuracy reached there.

G3. A distributional metric. The posterior-mean error is replaced by the sliced Wasserstein distance W2W_2 to the exact posterior (Sec. 3.1).

G4. An event-driven simulator. The Gaussian mixture is replaced by a Gillespie birth-death model whose control flow depends on its discrete reaction draws, with up to 20 parameters (Sec. 3.3, Algorithm 2).

G5. New comparisons (App. B). Averages of score-function, finite-difference and Gumbel-softmax gradient estimates (B.1), drifts that need no simulation (B.2), and neural posterior estimation, NPE (B.3).

G6. Averaged drift and new theory. A single stochastic-AD estimate is noisy, so ABC-D2D^2 now averages the unbiased estimates across particles, the across-particle analogue of the replicate average of the submission, and fixes this average within each sweep. This keeps the MH correction exact without per-point seeding (Prop. 1). The new Proposition 2 and Corollary 1 show that averaging reduces the noise of unbiased estimates but not the bias of other estimators, and that the gain of Theorem 1 survives averaging when the gradient dominates the bias of the average.

Main results against the tuned random walk (Figs. 2 and 3):

  • SIR model: 2.7–20× fewer simulator calls, with 95% intervals above one in all nine settings where both samplers reach ε∗\varepsilon^*, at a similar or lower W2W_2 (lower in four settings, higher in none).
  • GLM: up to 6.1× fewer calls from d=20d = 20, but more calls with d=5d = 5, and a W2W_2 ratio of 0.92–1.2 (higher in three settings with d=30d = 30).
  • Gillespie model: 1.15–2.31× fewer calls and 1.28–2.64× less CPU time, for dd from 3 to 20.
  • In CPU time, the gains on the GLM and the SIR model are smaller: speed-ups of 0.79–1.3 and 0.87–5.1.
  • ABC-D2D^2 needs 2.9–87× less CPU time than NPE to come within 1.25× of the random walk's lowest W2W_2, while NPE is more simulation-efficient, needing 2.7–581× fewer simulations.

We also removed the claim that ABC is more robust than neural SBI under misspecification, and now relate the method to pseudo-marginal MCMC and to optimal scaling theory. We thank the reviewers again for comments that have substantially improved the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.