A Finite-Particle Convergence Rate for Stein Variational Gradient Descent

We provide the first finite-particle convergence rate for Stein variational gradient descent (SVGD), a popular algorithm for approximating a probability distribution with a collection of particles. Specifically, whenever the target distribution is sub-Gaussian with a Lipschitz score, SVGD with n particles and an appropriate step size sequence drives the kernel Stein discrepancy to zero at an order 1/sqrt(log log n) rate. We suspect that the dependence on n can be improved, and we hope that our explicit, non-asymptotic proof strategy will serve as a template for future refinements.

Paper

Similar papers

Peer review

Reviewer awG87/10 · confidence 2/52023-07-03

Summary

This paper provides an analysis of the convergence rate of finite-sample Stein Variational Gradient Descent (SVGD) for sub-Gaussian targets with Lipschitz scores. In contrast to previous works such as Liu 2017, Duncan et al. 2019 and Korba et al. 2020, the presented results offer convergence guarantees that hold for finite samples and rely on weaker assumptions. The authors present two key results regarding the discretization error of finite-sample SVGD compared to infinite-sample SVGD, measured in terms of Wasserstein (Theorem 1) and KSD (Theorem 2), respectively. These results are then combined to establish a finite-sample bound on the KSD error between a finite-sample approximation and the target distribution (Theorem 3). By carefully selecting step sizes, the authors demonstrate that this error decays at a rate of $1 / \sqrt{\log \log n}$ (Corollary 2).

Strengths

**Quality**: This paper provides a rigorous study of the convergence of finite-sample SVGD and delivers clear and well-supported results. The main findings (Theorem 1, 2, 3) build upon existing results (Lemma 1 and 4) but require sophisticated combinations and detailed arguments. The authors provide extensive discussions and overview the proof strategy, which appears reasonable and comprehensive (although I did not examine the proofs in depth). **Novelty**: This paper offers a crucial contribution by bridging the gap between the existing convergence guarantees of infinite-sample SVGD and the practical implementation of finite-sample approximations, which is currently lacking in the literature. Whilst the established convergence rate is slow, this work stands out as the first to provide an explicit, non-asymptotic guarantee for SVGD approximations under reasonably mild assumptions on the target distribution. This, in my view, is the primary novelty and significance of this paper. **Clarity**: The paper demonstrates clear motivation and objectives. Despite the technical nature of the content, the exposition is highly comprehensible. The intuitive explanations and remarks provided before and after each main theorem are particularly valuable in aiding understanding.

Weaknesses

**Slow convergence rate**: As mentioned above and in the paper, the established rate $1 / \sqrt{\log \log n}$ is notably slow, suggesting that the bound (3) is likely to be very loose. Although the authors acknowledge the potential of the presented proof strategy as a starting point for refining the bounds, it is not entirely clear whether or not or how the proof could be adapted to achieve an improved bound. Including discussions on potential avenues for enhancing the convergence rate would be valuable and insightful.

Questions

1. There are a few places that say “the Algorithm 2 outputs $\mu_r^n = SVGD(\mu_0^n, r)$…”, e.g. L166. Since this output is based on a discrete measure $\mu_0^n$, should it be “the **Algorithm 1** outputs …” instead? 2. COuld you elaborate on the suggested step size scheme in Corollary 2? Specifically, what is the dependence on the sample size $n$ and on the dimension $d$? How easy is it to construct the upper bounds $(\overline{w}_{0, n}, \overline{A}, \overline{B}, \overline{C})$ in practice? 3. As also mentioned in the paper, the established rate $1 / \sqrt{\log \log n}$ is very slow. Could you provide some insights into which parts of the proof strategy could possibly have led to this slow rate? Is the bound (3) tight? How does this convergence rate compare with empirical performance in numerical simulations?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The assumptions on the target distribution, its score function and the positive definite kernel are summarised in Section 2, accompanied by interpretations and discussions on their connections to related literature. Limitations are also discussed in Section 9. Overall, I think the discussions on assumptions and limitations are adequately covered.

Reviewer bq3E5/10 · confidence 1/52023-07-04

Summary

In this work, the authors present a novel analysis of finite-particle Stein Variational Gradient Descent (SVGD) and derive a unified convergence bound for this algorithm. The convergence bound provides an explicit measure of how close the finite-particle SVGD algorithm gets to its target. To establish this convergence bound, the authors first introduce a bound on the discretization error of the 1-Wasserstein distance between the finite-particle and continuous SVGD. They make certain assumptions that are commonly satisfied in SVGD applications and compatible with Kernelized Stein Discrepancy (KSD) weak convergence control. Overall, this work contributes to the understanding of finite-particle SVGD and provides a unified convergence bound that quantifies the algorithm's convergence to its target. The derived bounds enable better control and evaluation of the accuracy of finite-particle SVGD in practical applications.

Strengths

The strengths are: * The authors have proved the first unified convergence bound and rate for finite-particle SVGD. * They show that SVGD with n-particles drives the KSD to zero at an order $1/ \sqrt{\log\log(n)}$ rate.

Weaknesses

In my perspective, this paper would be more suitable for an optimization journal, as it would provide an environment where the technical contributions of the study can undergo a more comprehensive evaluation through an extended rebuttal cycle.

Questions

None.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

None.

Reviewer igw18/10 · confidence 2/52023-07-04

Summary

The authors provide the first convergence guarantee for finite particle Stein Variational Gradient Descent (SVGD). Although I am not an expert on this topic, I believe this problem remained open for a long time, and it should be the first of many finite particle results to come. While $(\log\log n)^{-1/2}$ is a fairly slow rate, I believe we should overlook the rate, but consider the significance of the technical leap taken by the authors to achieve a finite particle result at all. For this reason, I will recommend accept fro this paper.

Strengths

1. This is the first finite particle convergence guarantee for SVGD. 2. The contents are well organized and presented. 3. The proof is concise and clean.

Weaknesses

N/A

Questions

Given that I am not an expert on this subject, I would like the authors to clarify a couple of questions for me. 1. What was the main conceptual challenge in establishing a finite particle guarantee, and how did this work overcome it? For me Theorem 3 reads like a bit of magic, and the desired result just appears without much intuition. I would like to understand how this came to be. 2. What do the authors perceive as the next challenge preventing improvements to this result? Similar to the previous question, I don't quite see on an intuitive level where the $log log n$ dependence came about, which the authors also believe can be improved. I would appreciate the authors can elaborate further on this topic.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

4 excellent

Limitations

N/A

Reviewer g2wP6/10 · confidence 3/52023-07-08

Summary

This work studies the non-asymptotic convergence rate of Stein variational gradient descent (SVGD), an algorithm for approximating a target probability distribution with a collection of particles. This work presents a finite-particle convergence rate for SVGD, which provides a measure of how quickly the algorithm converges to the target distribution with a finite number of particles. The convergence rate formula drives the kernel Stein discrepancy to zero at an order 1/√log log n rate, but the authors suspect that the dependence on n can be improved and hope that their proof strategy will serve as a template for future refinements.

Strengths

This work provides the first finite-particle convergence rate for Stein variational gradient descent (SVGD), which is a popular algorithm for approximating a probability distribution with a collection of particles. This work presents an explicit, non-asymptotic proof strategy for the convergence rate formula --- which drives the kernel Stein discrepancy to zero at an order $1/\sqrt{\log \log n}$ rate providing a measure of how quickly the algorithm converges to the target distribution with a finite $n$ particles. Prior to this work, relatively little is known about SVGD's non-asymptotic approximation quality, despite that SVGD has demonstrated promising results for various inferential tasks. The authors also claims that it serves as a template for future refinements to improve the dependence on $n$. I highly believe in this, due to the soundness of technical tools the work adopted. The authors provide a thorough discussion of the assumptions and conditions required for the convergence rate formula to hold, which helps to clarify the limitations and applicability of the formula. Finally, this work also includes a comprehensive list of references to related work, which provides a useful starting point for further research on SVGD and related algorithms.

Weaknesses

Despite its success in providing the first non-asymptotic convergence rate, this work assumes a level of familiarity with the mathematical concepts and notation for experts, which may make it difficult for readers without a strong background in probability theory and optimization to follow. In addition, this work does not provide any experimental results or comparisons with other algorithms to demonstrate the practical usefulness of the convergence rate formula, and does not provide any specific information on how the dependence on $n$ can be improved, which may limit its usefulness for researchers looking to optimize the performance of SVGD. Lastly, this work focuses exclusively on SVGD and does not provide any insights or comparisons with other algorithms for approximating probability distributions, which may limit its broader relevance to the field.

Questions

I am at an educative level but quite enjoy reading on this topic. I quite like the generic / infinite-particle continuum manner SVGD Algorithm 2 is written instead of the $n$-particle SVGD Algorithm 1. I wondered if this continuous-time approximation is the critical reason for the success of the non-asymptotic rate first established by the authors.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

This paper contributes as a theoretical work and does not raise negative social impacts.

Reviewer igw12023-08-12

Response

Thank you for the reply. I believe my questions are answered. While I am not an expert on this subject, and it's hard for me to justify raising the score, I would like to see this paper accepted given what I understand about it now. Therefore I will raise my score to 8, mostly in context of other reviews that are too pessimistic in my opinion and for non-technical reasons. I hope the AC will consider evaluating this work on a more fair benchmark, especially given this is the first finite particle guarantee.

Reviewer g2wP2023-08-17

Thank you for your informative response, especially the clarification of infinite-particle continuum limit. Indeed I believe this work should not be made obsolete due to its providing the first non-asymptotic convergence rate "as a significant leap". I have raised my score from 5 to 6 accordingly.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC