Accelerated Zeroth-order Method for Non-Smooth Stochastic Convex Optimization Problem with Infinite Variance

In this paper, we consider non-smooth stochastic convex optimization with two function evaluations per round under infinite noise variance. In the classical setting when noise has finite variance, an optimal algorithm, built upon the batched accelerated gradient method, was proposed in (Gasnikov et. al., 2022). This optimality is defined in terms of iteration and oracle complexity, as well as the maximal admissible level of adversarial noise. However, the assumption of finite variance is burdensome and it might not hold in many practical scenarios. To address this, we demonstrate how to adapt a refined clipped version of the accelerated gradient (Stochastic Similar Triangles) method from (Sadiev et al., 2023) for a two-point zero-order oracle. This adaptation entails extending the batching technique to accommodate infinite variance -- a non-trivial task that stands as a distinct contribution of this paper.

Paper

Similar papers

Peer review

Reviewer zmkR6/10 · confidence 3/52023-06-25

Summary

This paper proposes a novel gradient-free (zeroth-order) clipped version of stochastic similar triangles method for solving non-smooth stochastic convex optimization problem under a much weaker infinite variance assumption. The derived iteration and oracle complexity bounds are optimal in both convex and strongly convex case.

Strengths

The problem addressed in this paper is well-motivated, the write-up is easy to follow, and the novelty and contribution are easy to identify.

Weaknesses

1. While the paper is a valuable theoretical contribution, the addition of experimental results would enhance the overall work by demonstrating the feasibility and effectiveness of this method in practical applications. 2. The absence of a Conclusion section or any definitive end to the work makes it feel like a work in progress.

Questions

Line 118: Should the upper bound of the second inequality be a notation $\sigma_B^{\alpha}$?

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

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

None

Reviewer zmkR2023-08-13

I appreciate the authors' response, which adequately resolved my concerns. Due to the addition of the experimental and conclusion chapters, I will raise my score to 6.

Authorsrebuttal2023-08-14

Thank you for the response

We are glad that the reviewer's concerns were resolved and are grateful for raising the score.

Reviewer M3Ug5/10 · confidence 3/52023-07-05

Summary

In this papers, the authors build upon the work that has been done in [28] and adjust the proposed algorithms there for zero-order oracles rather than the gradient oracles. The goal is to optimize non-smooth stochastic convex optimization problems with infinite variance.

Strengths

The paper does a good job when it comes to introducing the notions they have used with clarity. The organization of the paper helps the reader to understand the concepts. The quality of the write up and the technical contributions look solid.

Weaknesses

- Minor typo in the abstract: ajust --> adjust. - "We emphasis (--> emphasize) that this generalization requires an extension of the batching technique to (--> for) infinite variance." The emphasis should be emphasize in this sentence and I believe for is more suitable than to. - The motivation about why we should be interested in such problems is only explained in citation numbers ([30, 6]) with no mention of what these motivating examples actually are. You have 9 pages of space without the references and you are not utilizing all of it. I suggest using the remaining space to attract more attention to your paper by mentioning concrete examples. This would help with the visibility of your paper as well. - "• can be generalized for saddle-point problems (based on [28]) and one-point feedback [13]. We leave it for future work." This should not been mentioned in the contributions since this paper has not actually made this contributions yet. You can add a conclusion and/or future work paragraph or section in the end mention it there. - Similarly, related work could have a section of its own. - My main concern about this paper is that, even though adjusting the method from [28] for two-point zero-order oracle is nontrivial, it feels a bit too incremental for a venue like NeurIPS. Also, the trade-off between using weaker or stronger assumptions can be discussed further in order to convince the reader on the advantages and the disadvantages of the zeroth-order methods and where to use which.

Questions

- In which cases we should utilize the proposed accelaerated zeroth-order method for non-smooth stochastic convex optimization problems with infinite variance? - It is not very clear to me why the paper considers the case when \delta = 0 on page 6. Could you please elaborate more on that? - Could you please list some instances where non-smooth stochastic convex optimization problems show up and why we should be interested in them?

Rating

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

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

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

The advantages and the disadvantages of the proposed method is not discussed in detail except for the advantage of enabling weaker assumptions.

Reviewer FKLk7/10 · confidence 4/52023-07-07

Summary

This paper proposed and analyzed a zeroth-order method for non-smooth stochastic optimization under heavy-tailed noise and adversarial noise, by combining ball-averaging-based smoothing technique (to tackle non-smoothness) and gradient clipping technique (to tackle heavy-tailed/adversarial noise). This generalizes previous results where $L_\infty$ or $L_2$ boundness of noise is assumed. There are also several technical improvements over previous results, including proving a high-probability bound instead of a bound in expectation.

Strengths

1. The topic is important. Gradient free method for stochastic optimization is a popular field of research, and it certainly helps to handle the heavy-tailed/adversarial noisy setting which is common in practice. 2. I have briefly gone through the proof and have no doubt on its correctness. 3. Though not proved in the paper, the theoretical bound seems close to being tight.

Weaknesses

1. The presentation is concise, but maybe at the cost of some necessary clarity. In Eqn. (2) it's not specified what distribution should $\bf e$ follow (should be the uniform distribution in the unit _ball_), which leaves the entire equation undefined. It becomes even more confusing that later in (3) $\bf e$ is used with a different meaning, denoting a random vector uniformly distributed on the unit _sphere_! 2. It may also improve clarity to discuss the order of the theoretical bounds and compare them with previous results. In particular, it would be helpful to discuss whether the dependence of $\alpha$ is optimal.

Questions

I have no additional questions other than the ones discussed in `Weaknesses`.

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

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

2 fair

Contribution

3 good

Limitations

Yes.

Reviewer nSeP7/10 · confidence 3/52023-07-07

Summary

This paper presented derivative-free methods for the optimization of stochastic convex functions with a potentially infinite variance noise. Here the level of noise is defined in terms of the boundedness of modulus of a Hölder-type continuous condition. The main technique is to adopt a gradient clipping to the two-point estimation of the gradient of the randomized smoothed function. For some of their results, they also claim the attained bounds are rate optimal. The presentation and organization is of very clear and relatively easy to follow.

Strengths

Overall the contribution is well-motivated, and fits into the flurry of recent development of methods for problem with infinite noise variance. The paper is well written, and the proofs are intuitive and relatively easy to follow.

Weaknesses

I only have some minor comments and questions: * L113: it might be better to use {\xi_i}_i and {e_i}_i as the input of the function g^B. * L119: where is the \sigma_B used in the statement of Lemma 3? * the second line of Section 2: you might need to assume g \neq 0. * L126: where is the distribution D_k defined? * L139: could you elaborate on why the first term is optimal using the lower bound from [4]? Do you need some assumption on batch-size B here to illustrate rate optimality? * L152: it might be better to emphasize that w.p. 1-\beta, *for any* 1 \leq t \leq N, ..., if the boundedness of iteration holds uniformly. * Question (out of curiosity): is the bound in L190 optimal? Or, is there any lower bound for such a so-called maximum allowable noise level? * L224: proof -> prove.

Questions

See above.

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

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

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Yes.

Reviewer 6zia6/10 · confidence 3/52023-07-09

Summary

This paper provides high probability bounds for the convergence of gradient-free methods on convex and strongly-convex functions when the noise in the gradient oracle has infinite variance. An oracle provides $f(x,\xi)$, a noisy evaluation of the function $f$ at point $x\in \mathbb{R}^d$ by the oracle, where $\xi$ is the noise variable. For the same noise $\xi$, the function is $M_2(\xi)$-Lipschitz, where $M_2(\xi)$ quantifies the noise level. If the noise variance is finite, ($\mathbb{E}[M_2(\xi)^2] < \infty$), [1] provides optimal iteration and oracle complexity for convergence of in expectation. The primary technique used in [1] is a batched accelerated gradient method which uses smoothing. For a fixed constant $\tau>0$ which defines the smoothing level, smoothing computes an approximate gradient of $f$ at point $x\in \mathbb{R}^d$ as, $$ g^B(x) = \frac{d}{2B \tau }\sum_{i=1}^B (f(x + \tau e_i, \xi_i) - f(x - \tau e_i, \xi_i))e_i $$ Here, $e_i$ are sampled uniformly from the unit sphere, $\xi_i$ are the noise variables of the oracle and $B$ is the batch size. On expectation, the smoothed gradient is the gradient of $ \mathbb{E}_{e,\xi}[f(x + \tau e, \xi)] $. For small value of $\tau$, this approximation is close to $f$. Further, even if the function $f$ is non-smooth but Lipschitz, smoothing makes $\frac{\sqrt{d}M_2}{\tau}$ smooth. This paper extends this technique to the infinite noise variance setting, $\mathbb{E}[M_2(\xi)^\alpha] < M_2^\alpha$ for some $\alpha \in (1,2]$ by applying clipping. Specifically, the technique of clipped Stochastic Similar Triangles used for handling heavy tailed noise in smooth optimization is extended to the non-smooth case by the above smoothing procedure. For convex Lipschitz functions, the iteration complexity and oracle complexity is $\frac{\sqrt{d}^{1/4}}{\epsilon}$ and $\left(\frac{\sqrt{d}}{\epsilon}\right)^\frac{\alpha}{\alpha-1}$. For $\mu$-strongly convex and lipshcitz functions, the corresponding bounds are $\frac{d^{1/4}}{(\mu\epsilon)^{-1/2}}$ and $\left(\frac{d}{\mu\epsilon}\right)^{\frac{\alpha}{\alpha-1}}$. These rates are shown to be optimal in $\epsilon$. Further, when if the oracle provides a corrupted value of $f$ with an additive corruption of $\lvert\delta(x)\rvert \leq \Delta$, the authors derive the maximum possible values of $\Delta$ such that the convergence rates for both smooth and non-smooth settings are unaffected by the corruption. **References** 1. Gasnikov et al. The power of first-order smooth optimization for black-box non-smooth problems. ICML 2022. 2. Sadiev et al. High-Probability Bounds for Stochastic Optimization and Variational Inequalities: the Case of Unbounded Variance. ICML 2023.

Strengths

- **Interesting Problem Setting**: Heavy-tailed noise is a significant problem which violates the commonly used bounded variance assumption in stochastic optimization. The authors extend the solution of clipping to handle it for derivative-free methods. - **Convergence Rates** : These are the first convergence rates for derivate-free optimization under heavy-tailed noise. Further, the rates are high probability bounds instead of in expectation. Additionally, for both cases, convex and strongly convex, the obtained rates are optimal in terms of error $\epsilon$. - **Thorough literature review** : The authors thoroughly review existing results in derivative-free methods and clipping.

Weaknesses

- **Presentation** : The paper seems to be missing an introduction and experiment section. Although the paper is theoretical, the proposed algorithm, clipped-SSTM with two-point feedback, is new and should have been tested on at least synthetic problems. - **Lack of a motivating example** : The authors do not provide a motivating example which justifies the heavy-tailed noise in gradient-free settings.

Questions

- Are there any lower bounds for oracle and iteration complexity in the adversarial corruption case in terms of $\Delta$? - Are there methods other than clipping to handle unbounded variance in stochastic optimization?

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

3 good

Presentation

1 poor

Contribution

3 good

Limitations

-

Reviewer FKLk2023-08-12

I appreciate the authors' response and will keep my score unchanged.

Authorsrebuttal2023-08-14

Thank you for the response

We thank the reviewer for checking our response and for the very positive evaluation.

Reviewer M3Ug2023-08-17

I thank the authors for carefully responding to my concerns.

Authorsrebuttal2023-08-19

Thank you for the response

We thank the reviewer for checking our response and for the positive rating.

Reviewer 6zia2023-08-18

Response

Thanks for providing a detailed response to all of my questions. The motivating example and numerical experiments seem nice and the authors should include it in the final version of the draft. I'm increasing my score based on this.

Authorsrebuttal2023-08-19

Thank you for the response

We are very grateful to the reviewer for raising the score. We will definitely include motivating examples and numerical experiments in the final version.

Reviewer nSeP2023-08-18

I appreciate the thorough responses, and I'll maintain my current score.

Authorsrebuttal2023-08-19

Thank you for the response

We thank the reviewer for checking our response and for the very positive evaluation.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC