Convergence of Adam Under Relaxed Assumptions

In this paper, we provide a rigorous proof of convergence of the Adaptive Moment Estimate (Adam) algorithm for a wide class of optimization objectives. Despite the popularity and efficiency of the Adam algorithm in training deep neural networks, its theoretical properties are not yet fully understood, and existing convergence proofs require unrealistically strong assumptions, such as globally bounded gradients, to show the convergence to stationary points. In this paper, we show that Adam provably converges to $ε$-stationary points with ${O}(ε^{-4})$ gradient complexity under far more realistic conditions. The key to our analysis is a new proof of boundedness of gradients along the optimization trajectory of Adam, under a generalized smoothness assumption according to which the local smoothness (i.e., Hessian norm when it exists) is bounded by a sub-quadratic function of the gradient norm. Moreover, we propose a variance-reduced version of Adam with an accelerated gradient complexity of ${O}(ε^{-3})$.

Paper

Similar papers

Peer review

Reviewer LKzM7/10 · confidence 5/52023-06-22

Summary

The paper removes the Lipschit gradient assumption for the adaptive SGD (ASGD), making the ASGD broader to wide applications. Under the weak assumption, the authors still proved the optimal convergence rate. Moreover, they propose a variance-reduced version with an accelerated complexity. The results are interesting and novel.

Strengths

1. They further relaxed the ($L_0,L_1$) Lipschit gradient assumption, which is more relastic. 2. They develop a new analysis to show that ASGD can be convergent under the weak Lipschit gradient assumption but still with the same rate as existing works on ASGD. The results and techniques are novel. 3. A variance-reduced version of ASGD is proposed with provable acceleration. 4. The authors show that the rates of ASGD and its variance-reduced version is dimension-free under certain cases.

Weaknesses

1. The proof of variance-reduced version is interesting but without any numerical demonstrations. If the authors proposed an accelerated algorithm, it is convening to present some numerics because the theory is not the focus now. 2. In the contribution part, the authors claim that they do not assume bounded gradients. This is because you assume the a.s. bounded noise. which has been studied by previous work [Li and Orabona, 2019].

Questions

There are very recent works on using bounded variance noises [Matthew Faw et al. COLT 2022, Bohang Wang et al. COLT 2023]. Can the results in this paper be extended to that case?

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

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

YES

Reviewer rhLM6/10 · confidence 3/52023-07-05

Summary

The paper proposes a new proof strategy for the convergence of Adam in the non-convex setting. The new analysis relaxes the typical assumptions in the following ways: 1) it assumes relaxed smoothness, where the norm of the Hessian grows sub-quadratically with the gradient norm; 2) it does not require bounded gradients. With a deterministic gradient oracle, the obtained rate is $O(1/\epsilon^2)$. For the stochastic setting, the author shows a convergence rate of $O(1/\epsilon^4)$ with high probability. Furthermore, the author proposes a variance-reduced version of Adam and proves a rate of $O(1/\epsilon^3)$. The above rates in stochastic setting is dimension-dependent and can be dimension-free if the Hessian norm is sub-affine on the gradient norm. The main limitation is that although the bounded gradient assumption is removed, the bounded noise is still needed.

Strengths

1. The paper is well-written and easy to follow. Specifically, the technical challenges and the intuition behind the new analysis are discussed comprehensively. 2. Removing the bounded gradient assumption is generally challenging. Although there are a few works for AdaGrad that address this, existing work targeting the removal of this assumption for Adam only converges to a neighborhood of stationary points. 3. The paper considers the smoothness condition where the norm of the Hessian is bounded by a sub-quadratic function of the gradient norm, which is more general than $(L_0, L_1)$-smoothness.

Weaknesses

1. Although the work relaxes the bounded stochastic gradient assumption commonly required in the analysis of Adam, assuming almost surely bounded noise is still somewhat strong and impractical. Since the proof technique, the contradiction argument, heavily relies on this assumption, it is unclear whether it can be used in more realistic settings. 2. It is difficult to assess the usefulness of the proposed variance reduction method since the convergence result does not improve upon previous methods and there is no experimental evidence. Minor: The result presented for the deterministic setting is not formal. It would be better to have a formal theorem, at least in the appendix.

Questions

1. In Theorem 4.1, the same value of $\beta_{sp}$ and $\beta$ are selected, which differ from conventional choice of $(1 - \beta)^2 < 1 - \beta_{sp}$. Could the author please explain the reason behind this? 2. Why do we have a dependence on $d$ in the convergence result when $\rho \geq 1$ and suddenly become dimension-free when $\rho < 1$? Also, some results for Adam assume the infinity norm of the gradient is upper bounded by a constant [1], which natural leads to dimension dependence. It seems that by assuming bounded 2-norm, the results become dimension-free [2, 3]. Could the author provide further elaborate on this matter? 3. It seems that if we have access to problem-dependent parameters and the initial conditions (e.g., the gradient norm at the first iteration), SGD can also converge with the optimal rate [4]. What is the benefit of adaptive methods (if they also require tuning) compared to SGD in the more generalized smoothness setting? References: 1. Défossez, A., Bottou, L., Bach, F., & Usunier, N. (2020). A simple convergence proof of adam and adagrad. arXiv preprint arXiv:2003.02395. 2. Chen, X., Liu, S., Sun, R., & Hong, M. (2018). On the convergence of a class of adam-type algorithms for non-convex optimization. arXiv preprint arXiv:1808.02941. 3. Guo, Z., Xu, Y., Yin, W., Jin, R., & Yang, T. (2021). A novel convergence analysis for algorithms of the adam family and beyond. arXiv preprint arXiv:2104.14840. 4. Li, H., Qian, J., Tian, Y., Rakhlin, A., & Jadbabaie, A. (2023). Convex and Non-Convex Optimization under Generalized Smoothness. arXiv preprint arXiv:2306.01264.

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

4 excellent

Contribution

3 good

Limitations

The limitations are addressed in the paper.

Reviewer Rvyw6/10 · confidence 4/52023-07-06

Summary

This paper studies the convergence of Adam over non-convex objectives. To begin with, this paper proposes a new non-uniform smoothness condition called $(\rho, L_0,L_1)$ smoothness condition, which generalizes $(L_0,L_1)$ smoothness condition proposed in [Zhang et al. 2019]. The authors then prove the high-probability convergence rate $\mathcal{O}(1/\sqrt{T})$ of Adam under $(\rho, L_0,L_1)$ smoothness condition and affine noise variance assumption. The authors then propose Variance Reduced Adam (VRAdam) by combining the gradient estimation of STORM into Adam, and derive the convergence rate $\mathcal{O}(1/\sqrt[3]{T^2})$ of VRAdam.

Strengths

1. This paper provides the first $O(1/\sqrt{T})$ convergence rate of Adam without the bounded gradient assumption. 2. The constructed stopping time is interesting and can be of independent interest.

Weaknesses

1. Although this paper provides the first $O(1/\sqrt{T})$ convergence rate of Adam without the bounded gradient assumption, the analysis is somewhat restricted because it requires a non-zero $\lambda$ and the convergence rate has a polynomial dependence over $1/\lambda$. However, in the previous analyses of Adam including [Défossez et al., 2020; Zhang et al., 2022, Wang et al., 2022], it is allowed that $\lambda=0$ or the rate has a logarithmic dependence over $1/\lambda$. I treat this as a weakness of this paper because, in practice, $\lambda$ is set to very close to $0$ (for example, $10^{-8}$ as the default value in PyTorch) and may result in very loose bound. 2. As mentioned in this paper, the assumed noise condition is still stronger than the assumptions used for the analysis of other optimizers, for example, affine variance assumption. **Minor Issue**: The statement with respect to [Zhang et al., 2022; Wang et al., 2022] in line 28 is not inproper. In fact, if you pick $\beta_{sq}$ according to $\epsilon$ in [Zhang et al., 2022; Wang et al., 2022] just as this paper does, you can derive the convergence to stationary points (but with a slower rate). Therefore, a proper statement should be "but the convergence rate of [Zhang et al., 2022; Wang et al., 2022] is slower". **References** Defossez et al., A Simple Convergence Proof of Adam and Adagrad, 2019 Zhang et al., Adam Can Converge Without Any Modification On Update Rules, 2022 Wang et al., Provable adaptivity in Adam, 2022

Questions

I will consider increasing my score according to the responses to the questions. 1. Although the studied smoothness assumption is more general, what is the motivation to study it? Specifically, according to [Zhang et al. 2019], $(L_0,L_1)$ smoothness condition seems to be enough to capture the training process of neural networks. 2. The convergence rate has a dependence on the initial gradient norm $\Vert \nabla f(x_1) \Vert $. Does this mean that Adam is sensitive to the initialization? I expect some explanations here.

Rating

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

None

Reviewer d5Nv8/10 · confidence 2/52023-07-10

Summary

This paper studies the convergence of the Adam algorithm. Under a more general local smoothness assumption, the convergence of Adam to stationary points is proved without assuming boundedness of the loss gradient. Here the key technique is to show that the loss gradient along the trajectory is indeed bounded, using a proof by contradiction. A variance-reduced variant of Adam is proposed to achieve an accelerated gradient complexity.

Strengths

The convergence of Adam is definitely an important question, and this paper seems to be a significant contribution by relaxing the bounded gradient assumption. The argument of bounding the gradients along the optimization trajectory looks neat and easy to understand. Overall the result of the paper is solid and novel, and the writing is also easy to follow.

Weaknesses

It seems that the proof doesn't suggest the benefit of the momentum. It would be helpful if the authors can comment on this and provide some insights on the theoretical understanding of the momentum term. It would also be helpful if the authors can summarize the recent results on Adam convergence in the form of a table.

Questions

- Under the current local smoothness condition, how do we compare the performance of Adam and other algorithms like SGD? Does the current result suggest any advantage of Adam?

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

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

The authors have adequately addressed the limitations.

Reviewer hwio7/10 · confidence 3/52023-07-25

Summary

This paper provides convergence results for Adaptive Moment Estimate (Adam) and its variance-reduced variant under generalized smooth and bouned-noise assumptions.

Strengths

This paper mainly studies the convergence of Adaptive Moment Estimate (Adam) under a generalized smooth assumption. The authors drop the commonly used globally bounded gradients assumption and provide a new analysis framework, which is based on a contradiction argument, to show that gradients are bounded along the optimization trajectory. Based on this, the authors could deduce a high probability convergence bound of $\mathcal{O}(\epsilon^{-4})$ for Adam under the generalized smooth assumption. The paper also provides a variance-reduced version of Adam and improves the gradient complexity to $\mathcal{O}(\epsilon^{-3})$. In general, the assumptions are standard in the analysis of adaptive methods including unbiased gradient estimate and bounded noise. The convergence results make sense and are highly valuable since the convergence result of Adam under unbounded gradients still leaves empty, to my best knowledge. As Adam is a widely used optimizer in deep learning field, the theoretical results (like parameter setting in Theorem 4.1 and Theorem 4.3) in this paper could help practitioners to better understand and use the algorithm. The central idea of analysis is inspiring.

Weaknesses

First, the convergence result requires a dedicated tuned step-size and momentum parameters, specifically requiring prior knowledge such as generalized smooth parameter $(L_0, L_1)$ and the noise level $\sigma$. However, in most realistic situations, these parameters could be unknown or hard to obtain. Thus, the result is more valuable in the theoretical aspect but may not be so useful in the experimental aspect. Second, the parameter setting in Theorem 4.1 contradicts the real-one default setting on deep learning packages (such as Pytorch or Tensorflow) where $\beta_1 = 0.9 $ is smaller than $\beta_2 = 0.999$. It may be more persuasive to do some experiments showing the convergence under this new parameter setting. Third, the high probability convergence result in Theorem 6.2 does not provide an optimal rate to the probability margin $\delta$, leaving space for further improvement. In addition, since VRAdam is a new algorithm, it better requires some experiments to show its convergence under the parameter setting in Theorem 6.2 and its faster convergence rate than Adam.

Questions

see the weakness part

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

yes

Reviewer d5Nv2023-08-14

Response to the authors

I thank the authors for their response and their interesting discussion presented here. I don't have further questions.

Reviewer Rvyw2023-08-14

My concerns are addressed by the authors' rebuttal, and I increase my score to 6 as promised.

Reviewer hwio2023-08-16

Thanks for the discussions

Thanks for the interesting discussions. Here are two remaining questions: 1) Regarding the choice of momentum parameters $(\beta,\beta_{\text{sq}})$, does one need some restrictions on the relationship $\beta$ and $\beta_{sq}$ to get the convergence? Worth noting, in [Defossez et al. 2022], they stipulated the condition $0 \le \beta_{sq} < \beta$. 2) There appears to be some confusion among readers regarding the intricate requirements for $\beta, \eta, G$ as outlined in Theorem 4.1. Would it be possible to present a concrete numerical parameter configuration that adheres to these conditions? Or, could the parameter choices be simplified in some specific cases?

Authorsrebuttal2023-08-19

Thanks for the reply

Below we try to answer the two remaining questions in the reply 1. We use a different analysis from existing works like [Defosses et al. 2022]. In Theorem E.1 (in the case of $0\le\rho<1$), $\beta_{sq}$ can be any constant between $0$ and $1$. In Theorem 4.1, we do require $\beta=\beta_{sq}$ for simplicity. However, as we mentioned in the rebuttal, we believe we can completely get rid of this requirement using a more careful analysis, which we will try to make formal in the revision. 2. We find it hard to further simplify the rigorous parameter choices. However, informally speaking, below are the most important requirements, which we will also add in the revision - $G$ is a large enough constant depending on problem-dependent constants and initial gradient norm - $\eta=O(\epsilon^2)$ - $\beta=O(\epsilon^2)$ - $T=O(\epsilon^{-4})$

Reviewer rhLM2023-08-16

Thanks for addressing my questions. I would like to increase my score. I would appreciate it if the author could elaborate more the bounded noise assumption. Regarding the response "The challenging part is that, when noise is heavy-tailed, there is an unignorable probability that $v_t$ becomes very large due to a very large noise in the gradient, which may result in a small learning rate and slow convergence.", recent works for AdaGrad have successfully removed the need for bounded gradients and noise [1, 2]. Specifically, [2] also considers $(L_0, L_1)$-smoothness. Similar to Adam, AdaGrad may also have a large $v_t$ if there is a very large noise in the gradient. I understand the analysis for Adam would be different, but maybe some techniques from these recent works could help removing the bounded noise assumption? **References** [1] Faw, Matthew, et al. "The power of adaptivity in sgd: Self-tuning step sizes with unbounded gradients and affine variance." COLT. 2022. [2] Wang, Bohan, et al. "Convergence of AdaGrad for Non-convex Objectives: Simple Proofs and Relaxed Assumptions." COLT. 2023.

Authorsrebuttal2023-08-19

Thanks for the reply

Thanks for the reply and further comments. We do agree that it is possible that some of the techniques from the two recent works mentioned by the reviewer could help to relax the bounded noise assumption. However, we do not think it is straightforward. One important difference between AdaGrad and Adam is that $v_t$ is non-decreasing for the former but not for the latter. In fact, [2] does rely on such property in their analysis. For example, below is a paragraph taken from their paper (page 3) *"In this paper, we propose a novel auxiliary function $\xi(t)=\frac{\nabla f(w_t)}{\sqrt{v_t}}$ for the convergence analysis of AdaGrad(-Norm), and show the error term can be bounded by $\mathbb{E}^{|\mathcal{F}_t}[\xi(t-1)-\xi(t)]$ (c.f. Lemma 4), which can be reduced by telescoping. As explained in Section 3, such an auxiliary function is rooted in the **non-increasing nature** of the adaptive learning rate $\frac{\eta}{\sqrt{v_t}}$.''* Therefore, we leave it as an interesting and potentially challenging future work.

Reviewer rhLM2023-08-21

Thanks for the explanation. My concerns have been addressed, and I have increased my score.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC