acceptodds
Under review as a conference paper at ICLR 2027

Theoretical Guarantees for SMC-Guided Diffusion Sampling

Abstract

Post-hoc conditioning of pretrained diffusion models can be addressed using Sequential Monte Carlo (SMC) methods. By evolving an interacting particle system, SMC-guided diffusion samplers combine unconditional reverse-diffusion dynamics with sequential reweighting to approximate conditional distributions. Nevertheless, even in the infinite-particle limit, the implemented sampler may differ from the ideal conditional target because of errors in the diffusion model, its numerical implementation, and the guidance mechanism. We characterize how these local errors propagate through forward-smoothing kernels, which jointly account for the reverse dynamics and the remaining conditioning information. This yields non-asymptotic error bounds that capture both finite-particle fluctuations and approximation errors arising from initialization, numerical integration, score approximation, and potential design. In doing so, we extend stability guarantees for diffusion models to the conditional setting. Finally, we apply our framework to several state-of-the-art SMC-guided diffusion algorithms, providing a unified theoretical perspective on their approximation mechanisms and sources of error.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Diffusion models, Sequential Monte Carlo, conditional sampling, score-based generative models, Feynman–Kac models, particle methods.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qYWuOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper develops non-asymptotic guarantees for SMC-based conditional sampling with pretrained diffusion models. The analysis separates the final error into a finite-particle M−1/2M^{-1/2} term and approximation errors due to initialization, score estimation, numerical discretization, and approximate guidance potentials. These local errors are propagated through forward-smoothing kernels using a Harris-type forgetting argument on a noncompact state space. The framework is illustrated on Gaussian-mixture conditional sampling problems and connected to several existing diffusion-SMC methods.

Critical strengths

The main results provide a unified error decomposition, which separates finite-particle error from approximation errors due to the score matching error, discretization error, initialization error, and imperfect guidance potentials.

Critical weaknesses

First, the reviewer finds that the main bound presented in the paper turns out to be a bit difficult to be interpreted quantitatively. Although the result is non-asymptotic, the forgetting factors Γℓ,N\Gamma_{\ell,N} depend on Harris minorization constants that can become very small in high dimension, while the prefactor 1/bN1/b_N can correspondingly become large. The paper does not quantify the resulting dependence on dimension, number of SMC steps, or guidance strength, even in the simple case of Gaussian mixtures. Relatedly, Assumptions D.1--D.2 impose moment and normalization conditions directly on the particle system. The results can be further strengthened to clarify when these follow from more primitive conditions on the terms like inaccurate potentials.

Furthermore, to the best of the reviewer's knowledge, SMC-guided diffusion sampling with pretrained score functions have been studied extensively in a lot of previous work under the "inference-time scaling" framework, such as [1-6], but the current version of the manuscript's literature review section is missing almost all of them. Moreover, a few recent work like [7-9] have also devoted to study the theoretical aspects of SMC-based inference-time scaling under the setting of both diffusion models and large language models (LLMs). The authors are strongly encouraged to revise the manuscript by including a comprehensive discussion on these related work to clarify the main novelty and contribution of this paper.

Questions for authors
  1. Would it be possible for the authors to provide a quantitative characterization of Γℓ,N\Gamma_{\ell,N} or the underlying contraction constants as a function of dimension NN and guidance strength in the Gaussian-mixture setting? This would help determine whether the forgetting bound is informative beyond establishing finiteness.

  2. Can Assumption D.2 be possibly verified from Assumptions 3.1, 3.3, and 3.5 for a common class of potentials, such as Gaussian likelihoods? If not, please make it clear in the article that the particle bound is conditional on such additional non-degeneracy assumption.

  3. How does Theorem 3.6 compare with the theoretical guarantee in existing work like [6,7,8]? The authors are strongly encouraged to cite these related work and discuss the main differences to further enhance quality of the current manuscript.

References:

[1] Skreta, M., Akhound-Sadegh, T., Ohanesian, V., Bondesan, R., Aspuru-Guzik, A., Doucet, A., ... & Neklyudov, K. (2025). Feynman-kac correctors in diffusion: Annealing, guidance, and product of experts. arXiv preprint arXiv:2503.02819.

[2] Uehara, M., Zhao, Y., Wang, C., Li, X., Regev, A., Levine, S., & Biancalani, T. (2025). Inference-time alignment in diffusion models with reward-guided generation: Tutorial and review. arXiv preprint arXiv:2501.09685.

[3] He, J., Hernández-Lobato, J. M., Du, Y., & Vargas, F. (2025). Rne: a plug-and-play framework for diffusion density estimation and inference-time control. arXiv e-prints, arXiv-2506.

[4] Wang, C., Wang, W., Ren, Y., Blanchet, J., & Lu, Y. (2026, May). Simple Approximation and Derivative Free Inference-Time Scaling for Diffusion Models via Sequential Monte Carlo on Path Measures. In Forty-third International Conference on Machine Learning.

[5] Ren, Y., Gao, W., Ying, L., Rotskoff, G., & Han, J. (2026, April). Driftlite: Lightweight drift control for inference-time scaling of diffusion models. In International Conference on Learning Representations (Vol. 2026, pp. 76664-76702).

[6] Chen, H., Ren, Y., Min, M. R., Ying, L., & Izzo, Z. (2025). Solving inverse problems via diffusion-based priors: An approximation-free ensemble sampling approach. arXiv preprint arXiv:2506.03979.

[7] Howard, S., & Nüsken, N. (2026). A Mean-Field Framework for Inference-Time Distributional Control of Diffusion Models. arXiv preprint arXiv:2608.08770.

[8] Delgadino, M. G., Motsch, S., Parulekar, A., Porteous, W., & Shakkottai, S. (2026). Diffusion-Based Posterior Sampling: A Feynman-Kac Analysis of Bias and Stability. arXiv preprint arXiv:2605.06538.

[9] Zhu, Y., & Lu, Y. On the Power of (Approximate) Reward Models for Inference-Time Scaling: Sequential Monte Carlo and Beyond. In Forty-third International Conference on Machine Learning.

3 comments
AuthorsOpenReview ↗

Main response

We thank the reviewer for these questions and the references.

Q1. We agree that Harris bounds can be conservative in high dimension. However, dimension and guidance strength alone do not determine the constants: score dissipativity, choice of the Lyapunov function and observation structure also matter. That said, Lyapunov drift and local minorization are classical tools for establishing stability of Feynman–Kac semigroups on noncompact state spaces (Whiteley, 2012; Douc et al., 2014). Our approach builds on this theory, adapting the quantitative Harris argument of Hairer and Mattingly (2008) to guided forward-smoothing kernels.

The dependence on these parameters is partly explicit in our results. Proposition 3.2 shows how stronger dissipativity improves the proposal drift coefficient. Equation (43) expresses the guided contraction factors in terms of drift and minorization constants. In particular, smaller contraction factors lead to stronger attenuation of earlier errors, whereas the prefactor bN−1b_N^{-1} arises from converting the terminal weighted norm into the VV-weighted norm. Guidance also affects these coefficients through the continuation weights.

We will clarify that the dimension enters the constants through Lyapunov moments, score regularity and minorization coefficients, and consider adding a Gaussian or Gaussian-mixture example to make these dependencies more concrete. We acknowledge that our results do not establish dimension-efficient guarantees in general. Obtaining sharper dimension-dependent estimates remains an important direction for future research, particularly for Harris-type analyses on noncompact state spaces. Finally, to our knowledge, references [6–9] do not provide polynomial-in-dimension guarantees jointly controlling particle fluctuations and initialization, score, discretization and guidance-potential errors in diffusion-SMC. We discuss their distinct targets and error criteria in our separate comment.

Q2. We agree that D.2 does not follow from Assumptions 3.1, 3.3 and 3.5 alone. D.1 and D.2 are explicitly stated among the hypotheses of Theorem 3.6. Their detailed formulations were placed in Appendix D for space reasons. The finite-particle bound is explicitly conditional on these additional integrability requirements.

We have, nevertheless, provided concrete sufficient conditions in our responses to reviewer prsc. First, Assumption 3.1, a uniform positive lower bound on the implemented potentials suffices at a fixed horizon (see our main response). More generally, a uniform positive lower bound is not necessary: D.1–D.2 can also be verified for potentials satisfying a Gaussian lower envelope under compatible exponential-moment conditions (see our comment “Technical details for D.1–D.2 under a Gaussian lower envelope”). This includes certain untruncated Gaussian likelihoods. The Gaussian-envelope criterion makes explicit a trade-off between guidance concentration, proposal contraction, the moment order pp, and the Gaussian initialization variance.

We will make the role of these assumptions more prominent in the main text and incorporate the sufficient conditions and their proof into the revised manuscript.

Q3 We thank the reviewer for highlighting this recent and relevant literature, which will undoubtedly strengthen the contextualization of our work. While we acknowledge that these references deserve a more thorough discussion, we believe that our theoretical contributions are complementary and address questions not covered by these works. We therefore view this as an opportunity to improve the literature review rather than a limitation of the theoretical contribution itself.

Recent inference-time scaling methods exploit additional computational resources at generation time to improve the quality of samples produced by pretrained generative models. We acknowledge the relevance of these contributions and we will discuss them in the revised manuscript. Our objective is complementary. Rather than proposing a new inference-time scaling strategy, we investigate the mathematical guarantees of SMC-guided diffusion sampling, establishing how accurately the resulting particle system approximates the conditional distribution.

This distinction is important because, even with an infinite number of particles, the implemented algorithm may fail to recover the desired conditional distribution due to initialization, score approximation, numerical discretization or guidance-potential errors. Conversely, even when these components are individually accurate, a finite number of interacting particles introduces an additional Monte Carlo approximation error.

We provide a detailed comparison with references [6–9] in a separate comment, highlighting the differences in theoretical guarantees, error criteria, assumptions and proof techniques. We will incorporate this discussion into the revised manuscript.

AuthorsOpenReview ↗

On the various references

Regarding the theoretical contributions of references [6-9], we believe that their guarantees are different. Despite some overlap, the settings, distances considered, proof techniques, errors tackled are genuinely different. To our knowledge, none of the cited works propagates both particle fluctuations and model approximation errors through the same unified perturbation analysis on noncompact state spaces and with potentials that are not bounded away from zero. None of the cited works provides a joint non-asymptotic terminal sampling error bound explicitly accounting for the time discretization of the underlying backward SDE. Moreover, our approximation criteria reflect the conditioning problem: score error is evaluated along exact reverse trajectories initialized from ideal selected distributions and potential mismatch is measured under the ideal filtering distribution. Thus, small average score error under unconditional diffusion marginals need not guarantee comparable accuracy in regions emphasized by conditioning. We detail the differences below.

[6] Chen et al. establish approximation guarantees and particle consistency for their specific AFDPS construction based on PDE evolution, whereas our framework provides a joint non-asymptotic error analysis for general Feynman–Kac selection–mutation schemes, without prescribing a particular guidance mechanism. They establish separate approximation bounds (Theorem 4.1) and particle consistency (Theorem 4.2) for their continuous-time algorithm. These guarantees cannot be directly combined into a joint non-asymptotic terminal error bound as Theorem 4.1 and Theorem 4.2 do not use the same metrics. Moreover, the particle result provides no explicit convergence rate in the particle count. Compared to our approach they do not have a common unifying forward-smoothing stability mechanism governing both particle and approximation errors. Moreover, their theoretical guarantees do not account for numerical discretization or resampling errors. Regarding potentially large controls we highlight that their constant can be exponentially large (see def Cy,T(3)C_{y,T}^{(3)} equation C.41). Similarly, they define Cy(2)=e−Cy(1)/Z(y)C_y^{(2)}=e^{-C_y^{(1)}}/Z(y). For a standard Gaussian prior and likelihood proportional to e−τ∥x∥2/2e^{-\tau\|x\|^2/2}, this coefficient equals (1+τ)d/2(1+\tau)^{d/2}. Thus, their posterior approximation bound is not uniformly dimension-free either.

[7] Howard and Nüsken. Their contribution is distribution-level reward steering: the reward depends on the evolving measure, and weighted McKean–Vlasov dynamics track an implicitly defined tilted distribution. Theorem 3.4 establishes particle consistency for bounded smooth observables. Our setting instead uses prescribed state potentials and gives a non-asymptotic LpL^p bound separating finite-particle fluctuations from initialization, score, discretization and potential errors. We do not claim to subsume their measure-dependent reward framework.

[8] Delgadino et al. This work specifically analyzes the DPS construction rather than a finite-particle SMC approximation. Its Theorem 1 gives a Feynman–Kac formula relating the DPS output density to the target posterior, characterizing bias that can persist even with an exact prior score. It also studies numerical instability of DPS updates and bias from early guidance stopping. Our result addresses a different question: how finite-particle fluctuations and model approximation errors propagate through SMC selection–mutation steps. Our stability estimates concern forgetting through forward-smoothing kernels, rather than numerical stability of DPS updates.

[9] Zhu and Lu study sampling from a reward-tilted pretrained model. Under uniform reward-ratio and multiplicative Bellman-error bounds, they quantify the particle and computational complexity needed to attain a prescribed total-variation accuracy. Importantly, their intermediate twists telescope to the specified terminal reward, so imperfect intermediate guidance need not introduce a persistent target bias.

Our analysis additionally distinguishes the implemented diffusion model from the ideal reverse diffusion associated with the data distribution. We therefore control both particle fluctuations around the implemented flow and its discrepancy from the desired target, including initialization, learned-score, numerical-discretization and potential errors. These latter errors can persist as the particle count increases.

Their SMC complexity theorem controls the total-variation distance of the averaged output law E[ηTM]\mathbb{E}[\eta_T^M], whereas we bound the LpL_p error of random empirical estimates, allowing unbounded observables controlled by a Lyapunov function. This provides a complementary perspective to their explicit complexity analysis under Bellman-error assumptions; we do not claim that our assumptions subsume theirs or that our theorem supplies comparable horizon-complexity guarantees.

Reviewer_qYWuOpenReview ↗

Response

The reviewer would like to thank the authors for the detailed response and clarifications. The reviewer would double check the responses and let the authors know in case there is any further question. If there will be an updated version of the manuscript, the reviewer would make sure to read it as well and revise the review accordingly if needed.

Reviewer_C1XfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the theory of Sequential Monte Carlo (SMC) methods on top of a pretrained diffusion sampler. The analysis decomposes the total error into the finite-particle error and the approximation bias and provide upper bounds for each. Numerical experiments are performed on the Gaussian-mixture model.

Critical strengths
  1. The topic is relevant. Given the many recent inference-time sampling methods, it is important to understand the underlying theoretical guarantees.
  2. The proof is largely sound, though some implicit dependencies are omitted or not captured, which do not affect the theorem as shown.
Critical weaknesses
  1. The assumptions need more explanations and justifications. For example, Assumption 3.1 is quite abstract, and it is not clear if some example distributions and/or practical data would satisfy this (and, it is claimed to be "standard" yet without any citations). Also, Assumptions D.1 and D.2, needed for the main theorem, are put only in the appendix without any justification.
  2. The interpretation of theoretical results, primarily the main theorem, is not sufficient. Essentially, it only shows explicit dependence on MM but that on all other system parameters remains implicit. It is therefore unclear how the bound could provide practical insight for those parameters.
Questions for authors
  1. Please address my two comments in the weaknesses above.
  2. From the main theorem, how do you arrive at the claim that "the bound yields a discretization contribution of order h\sqrt{h}"? Your bound has an explicit summation over NN, which could depend on the discretization scheme.
  3. For the experiment, it is claimed that it "illustrates Monte Carlo decay compatible with the M−1/2M^{-1/2} scale." How to see that scale, particularly the exponent?
  4. Could you elaborate a little bit on how your result is compared to another SMC theoretical study in [1]?

[1] Youheng Zhu and Yiping Lu, On the Power of (Approximate) Reward Models for Inference-Time Scaling, 2026.

4 comments
AuthorsOpenReview ↗

Response to the weaknesses

We thank the reviewer for identifying aspects of our assumptions and results requiring clarification.

W1. Assumption 3.1 imposes dissipativity and linear growth on the learned score. Dissipativity is a standard condition for establishing Lyapunov drift and controlling tail behavior (Hairer and Mattingly 2011; Strasman et al., 2026). These assumptions accommodate various non-Gaussian distributions, including suitable finite and infinite Gaussian mixtures, strongly log-concave distributions with linearly growing scores, and non-convex densities with quadratic confinement. While Strasman et al. (2026) establish this property for exact scores, here we assume the learned score preserves this behavior. Linear growth is a technical integrability condition, implied for instance by global spatial Lipschitz continuity. We will add examples and clarify the applicability and limitations of these assumptions for trained networks and real data.

D.1 controls moments for local particle fluctuations, while D.2 controls inverse continuation normalizers. Neither follows from positivity and boundedness above of the potentials alone. With Gaussian initialization and Assumption 3.1, a uniform positive lower bound on the implemented potentials suffices at a fixed horizon (see our response to reviewer prsc). More generally, potentials may vanish at infinity: D.1–D.2 can still be verified under a Gaussian lower envelope and compatible exponential-moment conditions. We provide precise sufficient conditions and their proof in our separate comment to prsc, “Technical details for D.1–D.2 under a Gaussian lower envelope.” We will clarify these assumptions and their scope in the revised manuscript.

W2. We agree that the statement makes the dependence on the particle count MM more immediately visible than the dependence on the other parameters. We will expand the discussion to distinguish the explicit error terms from the stability and normalization coefficients governing their propagation. The bound has the structure

\le \frac{C_{MC}}{\sqrt M} + C_{init}\,\varepsilon_{init} + \sum_{k=1}^{N}w_{k,N} \left( C_k^{discr}\sqrt h + C_k^{net}\varepsilon_k^{\mathrm{score}} + C_k^{pot}\varepsilon_k^{\mathrm{pot}} \right), \quad w_{k,N}=\Gamma_{k,N}\Lambda_{k,N}.$$ This separates two ways in which the parameters affect accuracy: through the magnitude of the local errors, and through the sensitivity of the guided flow to those errors. First, increasing $M$ reduces particle fluctuations around the implemented flow, while initialization, score, discretization and potential errors require separate improvements. The bound therefore identifies an accuracy floor that increasing the particle count alone cannot remove. For numerical refinement, the $\sqrt h$ contribution is interpreted at a fixed selection grid, subject to uniform control of the propagation coefficients, as clarified in the next comment response. Second, score accuracy is measured along exact reverse trajectories initialized from the ideal selected distributions. Thus, the relevant score error depends on the conditioning problem: an unconditional training error alone need not characterize accuracy in the regions emphasized by the guidance. Similarly, the potential-error term measures the discrepancy between normalized ideal and implemented potentials under the ideal filtering distribution. Third, the coefficients describe how these local errors affect the terminal distribution. The forgetting factor $$\Gamma_{k,N} = b_N^{-1}\prod_{j=k+1}^{N}\alpha_{j,N}, \quad k<N$$ captures contraction through the remaining guided kernels. Its constituent constants depend on proposal drift, local minorization and continuation weights. Consequently, guidance affects both local approximation accuracy and error propagation. Stronger conditioning may reduce the continuation weights and worsen the resulting bounds, even when the proposal dynamics remain stable. The sufficient conditions discussed in our response on D.1–D.2 to prsc make one such dependence explicit through a compatibility condition between guidance concentration, proposal contraction and exponential moments. We acknowledge that the theorem does not provide a sharp, universal dependence on every system parameter, nor an optimal prescription for choosing the guidance strength or selection schedule. Its practical contribution is to identify the distinct sources of error, the distributions under which approximation accuracy matters, and the stability quantities that must be controlled when changing the sampler. We will make these implications and limitations explicit.
AuthorsOpenReview ↗

Response to the questions

Q2. The summation is over SMC selection stages, not individual Euler updates. As distinguished in Section 2.1, we fix the selection grid tk=kT/Nt_k=kT/N and refine the numerical grid within each selection interval, keeping the learned score and potentials fixed. With JJ Euler updates per interval, h=T/(NJ)h=T/(NJ), so h→0h\to0 corresponds to J→∞J\to\infty with NN fixed. Increasing the number of selection stages, or resampling at every Euler step, is not necessary to improve the numerical approximation of the diffusion proposals.

Importantly, the proposal drift and local discretization constants are already explicitly uniform: Proposition 3.2 gives drift constants independent of JJ and hh under the step-size condition (11), while Lemma E.4 gives the local contribution CkdiscrhC_k^{discr}\sqrt h with CkdiscrC_k^{discr} independent of numerical refinement.

The fixed-duration blockwise drift and minorization estimates also give uniform control of the continuation-normalizer terms and hence of the guided propagation coefficients. In particular, Γk,NΛk,N≤Bk,N,\Gamma_{k,N}\Lambda_{k,N}\le B_{k,N}, with Bk,NB_{k,N} independent of hh. We will make the tracking of these uniform bounds explicit in Appendices C.3 and E.2 of the revised manuscript. As a consequence, the discretization contribution satisfies

\Gamma_{k,N}\Lambda_{k,N}C_k^{discr} \le \sqrt h\sum_{k=1}^{N}B_{k,N}C_k^{discr} =O(\sqrt h),$$ since $N$ is fixed. The implied constant may depend on the fixed selection grid, diffusion coefficients, learned score and guidance potentials, but is independent of the Euler refinement. If selection instead occurs at every Euler step, then $N$ increases as $h$ decreases and the growing sum requires additional control. This is a different refinement regime; our stated discretization order concerns a fixed selection grid. **Q3.** Figure 5 in Appendix F.3 presents the particle-error experiment on log–log axes, complementing the linear-scale plot in the main text. On these axes, an error proportional to $M^{-1/2}$ corresponds to a straight line of slope $-1/2$. The exact-reverse-kernel baseline is the most informative curve for this comparison. It uses the true initialization, exact reverse kernels and exact Doob-bridge weights, thereby isolating finite-particle Monte Carlo error. The theoretical $M^{-1/2}$ bound in (17) concerns particle error relative to the implemented Feynman–Kac flow, whereas the plotted empirical $L^2$ error is measured relative to the ideal posterior. The Euler curves therefore also contain approximation bias and can flatten as this bias becomes dominant; they need not maintain a slope of $-1/2$ throughout the plotted range. We agree that displaying the curves on logarithmic axes alone does not establish the exponent. We will explicitly reference Figure 5 in the main discussion and add an $M^{-1/2}$ reference line to facilitate the comparison. Our statement is that the experiment illustrates behavior compatible with the theoretical particle-error bound, not that it independently establishes an exact asymptotic rate. **Q4.** We thank the reviewer for pointing us to this relevant work. Both papers use a Feynman--Kac formulation to study guided sampling with approximate intermediate guidance, but address complementary questions. Zhu and Lu study sampling from a reward-tilted pretrained model. Under uniform reward-ratio and multiplicative Bellman-error bounds, they quantify the particle and computational complexity needed to attain a prescribed total-variation accuracy. Importantly, their intermediate twists telescope to the specified terminal reward, so imperfect intermediate guidance need not introduce a persistent target bias. Our analysis additionally distinguishes the implemented diffusion model from the ideal reverse diffusion associated with the data distribution. We therefore control both particle fluctuations around the implemented flow and its discrepancy from the desired target, including initialization, learned-score, numerical-discretization and potential errors. These latter errors can persist as the particle count increases. The error criteria also differ. Their SMC complexity theorem controls the total-variation distance of the averaged output law $\mathbb{E}[\eta_T^M]$, whereas we bound the $L_p$ error of random empirical estimates, allowing unbounded observables controlled by a Lyapunov function. Thus, their particle-count dependence and our $M^{-1/2}$ fluctuation bound concern different quantities. Our stability analysis uses proposal drift and local minorization transferred to forward-smoothing kernels on non-compact spaces, with additional particle-integrability conditions. This provides a complementary perspective to their explicit complexity analysis under Bellman-error assumptions; we do not claim that our assumptions subsume theirs or that our theorem supplies comparable horizon-complexity guarantees. We will add this comparison to the revised manuscript.

Thank you for your response. I decided to keep my current score, for reasons below.

  1. Even D.1 and D.2 can be verified under a Gaussian lower envelop, this seems away from practice. For example, how can image (or other high-dimensional) data be guaranteed with that envelop?

  2. I understand the error decomposition, but my concern regarding the practical insight the theory brings is still not sufficiently addressed. What's the point, for example, to identify these sources of error?

1. We would like to clarify that the Gaussian lower-envelope condition concerns the implemented guidance potentials, not the underlying data distribution. For example, linear-Gaussian observation likelihoods can satisfy such an envelope independently of whether the prior is a distribution of image or represents complex high-dimensional data.

2. Our contribution goes beyond identifying error sources: we establish a joint non-asymptotic LpL^p bound controlling the error between the practical particle estimator and the desired target, separating finite-particle Monte Carlo fluctuations from deterministic approximation bias, with explicit bounds on the underlying local errors. We quantify how both propagate through the guided dynamics, explaining why increasing the particle count alone cannot guarantee accuracy relative to the target. Moreover, our approximation criteria account for regions emphasized by conditioning, while the forgetting coefficients quantify the attenuation of local errors.

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

The paper provides a unified framework to analyse the error for approximate diffusion-based SMC to sample from a tilted distribution when both the model and its implementation are approximated. A key point in the paper is the distinction between particle error and approximation bias. In particular, even when infinitely many particles are used, the implemented sampler may deviate from the desired target because it uses an approximate initialisation, a learned score, discretised dynamics, and approximate guidance potentials. The paper introduces an ideal Feynman-Kac flow reaching the target, an implemented flow incorporating these approximations, and a finite-particle approximation of that implemented flow. Its main technical ingredient is a forgetting result for guided dynamics. Errors propagate through forward-smoothing kernels, which incorporate both the diffusion transitions and all remaining guidance. Under dissipativity and growth assumptions on the learned score, suitable numerical step sizes and conditions on the potentials, the authors transfer Lyapunov drift and local minorisation properties of the proposals to these guided kernels. Finally a simple Gaussian mixture experiment illustrates the analysis.

Critical strengths
  • The paper addresses a meaningful gap between particle convergence and target accuracy. THe authors highlight that convergence of an SMC algorithm as the particle count increases does not establish convergence to the desired conditional data distribution when the diffusion model itself is approximate. Theorem 3.6 explicitly separates the difference sources of error and show that increasing the number of particles cannot eliminate the inherited bias for the diffusion model.

  • The assumptions under which the analysis is done cover realistic cases. In particular, the framework is applicable to non-compact spaces and unbounded observables.

  • The error terms concerns accuracy in regions which are actually selected by conditioning. The score error is evaluated along exact reverse trajectories initialised from the ideal selected distribution rather than solely under unconditional diffusion marginals. This captures an important issue: an unconditional score model can perform well on average while being inaccurate in regions emphasised by a particular observation or reward.

Critical weaknesses
  • The tools and analysis used rely substantially on Strasman et al. (2026). This later paper already develops Harris-based forgetting and uses it to propagate initialisation, discretisation and score errors through a weighted telescoping argument. The main addition seems to be the guided Feynman-Kac dynamics.

  • Assumption D.2 requires a substantial condition on the random particle system itself, rather than deriving its stability from the score and potential assumptions. It may hold in some settings but strict positivity and boundedness of the potentials does not necessarily establish it.

Questions for authors
  1. Could you clarify what are the main distinctions and technical novelty compared with Strasman et al. (2026)?

  2. Assumption D.2 appears to require strong lower-tail control of the random continuation normalisers, which does not follow directly from positivity and boundedness of the potentials. Could you provide sufficient conditions for D.2 in a concrete diffusion-SMC setting? Additionally, is it possible to quantify how its constants depends on observation noise or guidance strength?

3 comments
AuthorsOpenReview ↗

Main response

We thank the reviewer for their careful reading and for highlighting the importance of separating particle error from approximation bias.

Q1 Establishing forgetting and stability for unconditional SGMs, as developed in Strasman et al. (2026), provides an essential foundation for our analysis of SMC diffusion samplers. These properties are classical tools in SMC theory: they quantify how local perturbations propagate through successive selection–mutation steps and underpin bounds on the accumulation of particle errors; see Del Moral (2004, Chapters 4 and 7) and Cappé, Moulines and Rydén (2005, Chapter 9). Our contribution is to extend this foundation to guided diffusion samplers, where sequential reweighting requires establishing stability for kernels incorporating all remaining guidance, and to use this stability to control both approximation bias and finite-particle fluctuations.

Our main technical contribution is to transfer Lyapunov drift and local minorization from the proposals to the forward-smoothing kernels, using continuation-dependent Lyapunov functions. Crucially, this works on non-compact spaces with strictly positive, bounded potentials that need not be uniformly bounded below.

This result enables a unified bound for (i) finite-particle fluctuations around the implemented Feynman–Kac flow and (ii) its bias relative to the ideal target, including initialization, discretization, learned-score and guidance-potential errors. The score error is evaluated along exact reverse trajectories initialized from ideal selected distributions, reflecting regions emphasized by conditioning rather than only unconditional diffusion marginals. Similarly, the guidance-potential error is measured in L2L^2 under the ideal filtering distribution, comparing normalized ideal and approximate potentials. Thus, both errors reflect accuracy in regions relevant to the conditional sampling problem and not only to the unconditional problem.

Guided Feynman–Kac dynamics are indeed the central extension, but establishing their stability and using it to control both approximation bias and particle error require additional analysis. We will revise the presentation to distinguish these contributions more clearly from the tools inherited from Strasman et al. (2026).

Q2 We agree that D.2 does not follow from strict positivity and boundedness above of the potentials. This condition can be restrictive and difficult to verify in practical diffusion-SMC applications, particularly under strong conditioning.

Assumption D.2 is a sufficient integrability condition used by our proof to control the random denominators arising when local particle errors are propagated through the normalized Feynman--Kac flow. We do not claim that it is necessary for convergence of the sampler. Its role is to make the error bound explicit and separate the contributions of particle fluctuations, model approximation and guidance.

There is precedent in SMC theory for obtaining informative error bounds under restrictive sufficient conditions. For example, Del Moral, Doucet and Singh (2015, Section 3) establish particle stability under strong mixing and boundedness assumptions, explicitly acknowledge their practical limitations, and discuss their results as a qualitative guide beyond the proven setting. Their assumptions differ from D.2, but illustrate the same distinction between sufficient conditions for a quantitative proof and necessary conditions for algorithmic stability.

A concrete sufficient condition for D.2 is 0<aj≤gj(x):=G^j(x)∥G^j∥∞≤1.0<a_j\le g_j(x):=\frac{\widehat G_j(x)}{\lVert\widehat G_j\rVert_\infty}\le1.

Let δℓ,N=∏j=ℓN−1aj\delta_{\ell,N}=\prod_{j=\ell}^{N-1}a_j. Then

K_\ell=\sup_{M\ge1}\lVert\eta_\ell^{M,\theta}[V]\rVert_{L^{2p}}.$$ For a fixed horizon, $K_\ell<\infty$ follows from suitable initial polynomial moments and uniform proposal moment bounds; stable linear Gaussian diffusion updates with bounded reward potentials provide a concrete example. The bound makes explicit a possible deterioration under stronger conditioning. For $\lambda\ge0$ and bounded rewards $r_j$, taking $g_j(x)=\exp[\lambda(r_j(x)-\sup r_j)]$ gives $$\delta_{\ell,N}^{-1}=\exp\left[\lambda\sum_{j=\ell}^{N-1}\operatorname{osc}(r_j)\right].$$ For Gaussian observation costs capped at $B_j$, the corresponding factor is $$\delta_{\ell,N}^{-1}=\exp\left[\frac{\sum_{j=\ell}^{N-1}B_j}{2\sigma^2}\right].$$ Thus the bound can deteriorate with stronger guidance or smaller observation noise. Uncapped Gaussian likelihoods require additional conditions (see our separate technical comment). We expect the qualitative error decomposition to remain informative under weaker conditions. Establishing comparable quantitative guarantees beyond D.2, however, requires additional analysis and is not proved in the present paper. We will clarify this distinction and add the sufficient conditions above.
AuthorsOpenReview ↗

Technical details for D.1-D.2 under a Gaussian lower envelope.

It is also possible to verify D.1–D.2 without requiring the potentials to be uniformly bounded away from zero. For the VP diffusion, fix the selection grid, p≥1p\ge1, V(x)=∥x∥2V(x)=\lVert x\rVert^2, and ϕ0θ=π∞=N(0,α−1Id)\phi_0^\theta=\pi_\infty=\mathcal N(0,\alpha^{-1}I_d), α>0\alpha>0. Assume the drift of Proposition 3.2, with λˉ=max⁡jλjQ<1\bar\lambda=\max_j\lambda_j^Q<1. Suppose a,δ,c,C>0a,\delta,c, C>0 such that 4pa1−λˉ<δ<α2,\frac{4pa}{1-\bar\lambda}<\delta<\frac{\alpha}{2}, and, for every k=0,…,N−1k=0,\ldots,N-1,

\frac{\widehat g_k}{\lVert\widehat g_k\rVert_\infty}\le1, \qquad Q_{k+1}^{\theta,h} e^{\delta V}\le Ce^{(\delta-a)V}.$$ Then D.1 (with $q=2p$) and D.2 hold. These conditions make explicit the trade-off between potential decay, proposal contraction, the moment order $p$, and the variance of the Gaussian initialization. They are sufficient, not necessary, for D.1–D.2. **Verification of D.1.** For any probability measure $\mu$ with $\mu[e^{\delta V}]<\infty$, $$\tau_{\widehat g_k}(\mu)[ Q_{k+1}^{\theta,h} e^{\delta V}] \le\frac Cc \frac{\mu[e^{\delta V}e^{-aV}]}{\mu[e^{-aV}]} \le\frac Cc\mu[e^{\delta V}].$$ The last inequality follows from the nonpositive covariance of the increasing function $e^{\delta V}$ and the decreasing function $e^{-aV}$ of $V$. Since $\pi_\infty[ e^{\delta V} ] <\infty$ and $(1+V)^{q}\le C_{q,\delta} e^{\delta V}$ for some $C_{q,\delta}>0$, an induction argument yields D.1. **Verification of D.2.** Set $r_N=0$ and $$r_k=a+\lambda_{k+1}^Qr_{k+1}, \qquad r_k\le\frac{a}{1-\bar\lambda}.$$ Using $\widetilde\beta_{N|N}^\theta=1$, the recursion $\widetilde\beta_{k|N}^\theta=\widetilde g_k Q_{k+1}^{\theta,h} \widetilde\beta_{k+1|N}^\theta$, Jensen's inequality and proposal drift, backward induction, with $c_{N,N}=1$, gives $$\widetilde\beta_{k|N}^\theta \ge cc_{k+1,N}e^{-aV} Q_{k+1}^{\theta,h} e^{-r_{k+1}V} \ge cc_{k+1,N}e^{-r_{k+1}K_{k+1}^Q}e^{-r_kV} =:c_{k,N}e^{-r_kV}.$$ Consequently, $$\mu[ \widetilde\beta_{k|N}^\theta ]^{-1} \le c_{k ,N}^{-1}e^{r_k \mu[V]}.$$ Applying this bound to both $\phi_\ell^{\theta,M}$ and $\eta_\ell^{M,\theta}$, the random variable inside the norm defining $\mathcal A_{\ell,N}$ in D.2 is bounded by $$C_{\ell,N}(2+\eta_\ell^{M,\theta}[V]) \exp\left\\{ r_\ell\left( \phi_\ell^{\theta,M}[V] +\eta_\ell^{M,\theta}[V] \right) \right\\}.$$ Since $4pr_\ell<\delta$, the exponential moment estimates established for D.1, together with Jensen's and Cauchy--Schwarz inequalities, imply $\mathcal A_{\ell,N}<\infty$ which is D.2.

Thanks for your detailed reply. I'm still a bit concerned about assumption D.2. I will discuss with the other reviewers. Thanks again

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.