Gaussian Approximation and Multiplier Bootstrap for Polyak-Ruppert Averaged Linear Stochastic Approximation with Applications to TD Learning

In this paper, we obtain the Berry-Esseen bound for multivariate normal approximation for the Polyak-Ruppert averaged iterates of the linear stochastic approximation (LSA) algorithm with decreasing step size. Moreover, we prove the non-asymptotic validity of the confidence intervals for parameter estimation with LSA based on multiplier bootstrap. This procedure updates the LSA estimate together with a set of randomly perturbed LSA estimates upon the arrival of subsequent observations. We illustrate our findings in the setting of temporal difference learning with linear function approximation.

Paper

Similar papers

Peer review

Reviewer uNZU5/10 · confidence 4/52024-07-02

Summary

The paper presents advancements in the theoretical understanding of the linear stochastic approximation (LSA) algorithm. It establishes the Berry–Esseen bound for the normal approximation of Polyak-Ruppert averaged iterates, achieving an optimal rate with an aggressive step size of $\alpha_k \approx k^{-1/2}$. Additionally, it demonstrates the non-asymptotic validity of confidence intervals using a novel multiplier bootstrap procedure, marking a first in this domain. The practical utility of these theoretical results is showcased through applications in temporal difference (TD) learning for reinforcement learning.

Strengths

1. **Theoretical Advancements**: The paper makes theoretical contributions by establishing the Berry–Esseen bound for the normal approximation of Polyak-Ruppert averaged iterates. Though some previous works have done this in special cases before, this work differs from them in providing a tighter bound. 2. **Good Clarity:** The paper is easy to follow and well-written. The proof seems correct (not checked very carefully).

Weaknesses

1. **Strong Assumptions**: The requirement that $\epsilon(z)$ is uniformly bounded is a strong assumption that might limit the applicability of the results. Relaxing this condition to weaker ones, such as finite moments, could enhance the paper's generality. 2. **Missing References**: There are some missing references, see the limitations. 3. **Discussion on Lower Bounds**: The paper focuses on deriving upper bounds but lacks a discussion on the tightness and potential lower bounds. Addressing this aspect could provide a more comprehensive understanding of the bounds' efficacy and limitations.

Questions

See the Limitations.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

1. **Missing References on Statistical Inference**: Statistical inference for nonlinear stochastic approximation has been considered by [1*] and [2*], which is highly related to this manuscript but hasn’t been cited. Note that [1*] provides a Berry-Esseen-like bound for the whole trajectory rather than the averaged iterates. These references could be cited after [36] in line 116. - [1*] Li, Xiang, Jiadong Liang, and Zhihua Zhang. "Online statistical inference for nonlinear stochastic approximation with Markovian data." arXiv preprint arXiv:2302.07690 (2023). - [2*] Li, Xiang, et al. "A statistical analysis of Polyak-Ruppert averaged Q-learning." International Conference on Artificial Intelligence and Statistics. PMLR, 2023. 2. **Assumption on Bounded $\epsilon(z)$**: Theorem 1 requires that $\epsilon(z)$ is uniformly bounded, which is a strong condition. Is it possible to relax this condition to a weaker one, such as $\epsilon(z)$ only having a finite order of moments (such as the fourth order moment or smaller)? 3. **Tightness of Derived Upper Bounds**: The paper provides upper bounds, but it would be insightful to discuss the tightness of these bounds. Are there any thoughts or conjectures regarding potential lower bounds? #### Minor Corrections: 1. **Figure 1**: The last subfigure should be labeled as (c).

Reviewer ew5L7/10 · confidence 3/52024-07-12

Summary

The present paper studies linear stochastic approximation with martingale difference noise and diminishing step-sizes. The authors obtain Berry-Esseen bounds for the parameter sequence with Polyak-Ruppert averaging as well as a generalization of finite-time bounds for estimation confidence intervals for parameters in LSA. The obtained Berry-Esseen bounds are illustrated by a TD learning numerical example.

Strengths

To the best of the reviewer's knowledge, both of the contributions are novel. In particular, I believe that this is the first Berry-Esseen bound type bound to be obtained for general linear SA, which is is exciting to see. The assumptions, contributions and approach to analysis are objectively identified. The authors also did a great job in providing discussions/remarks/intuition for their results. The paper is well-written but some proofreading is recommended.

Weaknesses

Although the authors did a good job in outlining the scope and contributions of the paper, the analysis and main text are hard to follow given the number of symbols and equations. It is easy for a reader to get lost/distracted midway and it is very hard to keep track of the definitions of each of the terms I understand that this is an issue with theory papers like this, but would encourage the authors to move unnecessary terms or inequalities to the appendix (e.g. (9) in A3. The exact lower bound adds very little in the main text in my opinion. Its definition could have been postponed to the Appendix.) The numerical experiments are also weak since many of the plots are not related to the contributions of the paper itself. Plot c seems to be the only one directly related to the theorems of the paper, but it still does not illustrate the theory that well. Maybe including a plot of C k^{-1/4} to plot (b) where C is a constant for comparison could help in inferring convergence rates. Also, there is a mistake in the label of Figure 1. Subfigure (b) is mentioned twice.

Questions

- Could the authors clarify if using Polyak Ruppert averaging is necessary to obtain such a Berry-Esseen bound? It would be exciting to see bounds for unaveraged estimates as well. -Could the authors run the experiment supporting figure (c) for longer to see if the curve with \gamma = 1/2 willnot continue to go upwards eventually?

Rating

7

Confidence

3

Soundness

2

Presentation

2

Contribution

3

Limitations

The authors clearly identified the limitations of their results through a clear list of assumptions

Authorsrebuttal2024-08-12

We thank the referee for their comments. We will include a discussion on the Berry-Esseen result for last iterate as well as a longer run for simulations.

Reviewer DkFj6/10 · confidence 3/52024-07-25

Summary

## Overview Let $Z, Z_1, \dots, Z_n$ be i.i.d. random elements with a common distribution $\pi$ over $\mathbf{Z}$. Given $A : \mathbf{Z} \to \mathbb{R}^{d\times d}$ and $b : \mathbf{Z} \to \mathbb{R}^d$, the goal of the LSA procedure is to find the unique solution $\theta^\star$ of $$\mathbb{E}\left(A(Z)\theta^\star - b(Z)\right) = 0. $$ Given a decreasing sequence of step sizes $\alpha_k$ and a starting point $\theta_0$, the standard LSA is given by $$\theta_k = \theta_{k-1} - \alpha_k(A(Z_k)\theta_{k-1} - b(Z_k)) $$ and the Polyak-Ruppert averaged LSA is given by $$\overline{\theta_n} = \frac{1}{n} \sum_{k=n}^{2n-1} \theta_k. $$ The authors provide Berry-Essen-type bounds for the Gaussian approximation of $\sqrt{n} \left( \overline{\theta_n} - \theta^\star \right)$ and for the corresponding multiplicative bootstrap process. Namely, for the Gaussian approximation result they upper bound the quantity $$\rho_n = \sup_\text{B convex} \left| \mathbb{P}\left( \sqrt{n} \left( \overline{\theta_n} - \theta^\star \right) \in B \right) - \mathbb{P}\left( \Sigma_\infty^\frac{1}{2} \eta \in B \right)\right| $$ where $\eta \sim \mathcal{N}(0,I_d)$. And for the bootstrap approximation result they upper bound $$\rho_n^b = \sup_\text{B convex} \left| \mathbb{P}\left( \left.\sqrt{n} \left( \overline{\theta_n^b} - \overline{\theta_n} \right) \in B \right| Z_1, \dots, Z_{2n} \right) - \mathbb{P}\left( \sqrt{n} \left( \overline{\theta_n} - \theta^\star \right) \in B \right)\right| $$ where $\overline{\theta_n^b}$ are obtained from a multiplier bootstrap process. It is interesting to notice that this multiplier process can be evaluated online without keeping a history in memory. ## Overall proof arguments To provide the Gaussian approximation result the authors write $$\theta_n - \theta^\star = (I - \alpha_n A(Z_n))(\theta_{n-1} - \theta^\star) - \alpha_n \varepsilon(Z_n) $$ where $$\varepsilon(z) = A(z)\theta^\star - b(z). $$ The authors observe that $\varepsilon(Z_n)$ can be viewed as a noise, which is assumed to be bounded. Meanwhile, the operator $ I - \alpha_n A(Z_n) $ is a random perturbation around $ I - \alpha_n \mathbb{E}A(Z) $, which is shown to act as a contraction in an appropriate norm, provided that $-\mathbb{E}A(Z)$ is Hurwitz. Thus, when one take the PR average the noise is expected to behave as the sum of i.i.d. noises and the contraction term must shrink. This is formally done writing the PR average in the form of Theorem 2.1 [reference 60 of the paper] and bounding the terms given by the latter theorem. The proof the bootstrap approximation result follows the standard practice in this literature: - First, conditionally on the sample, a Gaussian approximation is obtained relating the bootstrap to a Gaussian with its covariance. - Second, a Gaussian comparison theorem relates the latter Gaussian with the desired Gaussian, with some concentration results being used to bound their difference in high probability. ## Main claims and observations The authors show that taking a step size of order $\frac{1}{\sqrt{k}}$ yields the best possible convergence rate in their bounds. They provide empirical evidence that this rate is optimal. They also claim to be the first ones to fully provide a non-asymptotic bootstrap approximation result.

Strengths

The paper seems to be the first to provide non-asymptotic Gaussian and bootstrap approximation bounds for LSA. Their assumptions are quite mild and are in line with the ones made in similar papers in other domains. For instance, assumption A.2 is similar in nature to the boundness assumptions and the strong-covariance assumption made in [A]. The paper is mathematically sound and poses interesting research directions. I'm particularly curious about the convergence rate of $n^{-\frac{1}{4}}$ suggested by their theoretical results and supported by their experiment. The proposed application to policy evaluation in RL is also interesting. Finally, I point out that the code on the supplementary material was easy to reproduce. [A] Chernozhukov, Victor, Denis Chetverikov, and Yuta Koike. "Nearly optimal central limit theorem and bootstrap approximations in high dimensions." The Annals of Applied Probability 33.3 (2023): 2374-2425.

Weaknesses

The main claim of the paper is that their bounds suggest an optimal convergence rate of $n^{-\frac{1}{4}}$ when taking $\alpha_k = \frac{c}{\sqrt{k}}$. The bottleneck of this convergence rate comes from an application of Cauchy-Schwarz inequality, the boundness of $\varepsilon$ and the MSE bound on $D$ given by Theorem 1. Meanwhile, the authors also provide experimental evidence that this convergence rate is indeed optimal (although not considering gamma values below 0.5 in Figures 1 or 3). A counter-example or a deeper discussion on this convergence rate would be beneficial to the paper: is it an artifact of the proof or is there reason to believe it cannot be improved?

Questions

Some observations: - Line 54 has a typo in "Berry-Essee". - In Equation (15) the $\ell$ can be removed from $\theta_{k}^{b,\ell}$ to enhance clarity. It is only used to explain that to evaluate the probability in practice one must run several samples of the bootstrap process. - Line 263 has a typo in "date". - The Equation after line 303 is using $\phi$ instead of $\varphi$. - Line 307 asks for a sequence of "TD(0) updates", what does the "(0)" stands for? - Figure 1 lacks x and y labels. The legend can also be improved for clarity.

Rating

6

Confidence

3

Soundness

4

Presentation

2

Contribution

3

Limitations

The authors adequately addressed the limitations and the potential impact of their work.

Reviewer DkFj2024-08-09

I thank the authors for their rebuttal. The comment made in the rebuttal on the optimal $n^{1/4}$-rate obtained in [Bolthausen, 1982] for the martingale Berry-Essen should be on the main text. I wonder if the authors can find an example matching their upper bound (thus showing it is optimal), maybe drawing inspiration from the example in Section 6 of [Bolthausen, 1982]. I thank the authors for the explanation of the meaning of TD(0), it should also be included on the revised text.

Authorsrebuttal2024-08-12

Yes, we will include the corresponding discussion on the optimal rates in martingale CLT in the revised version of the text, as well as a comment on TD(0). Fetching the example provided in [Bolthausen, 1982] into the LSA paradigm is not immediate, but we work in this direction, and, of course, if we succeed to construct such a lower bound, we will include it in the final text.

Reviewer ew5L2024-08-11

I appreciate the author's responses. I believe that it would be beneficial to include some discussion/remarks on the final version about the extension of their results for estimates without PR as in the response provided by the authors. I apologize for not taking enough time to fully grasp the experiments in the paper, but I understand them now. Thank you for providing a longer run.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC