Achieving Near-Optimal Convergence for Distributed Minimax Optimization with Adaptive Stepsizes

In this paper, we show that applying adaptive methods directly to distributed minimax problems can result in non-convergence due to inconsistency in locally computed adaptive stepsizes. To address this challenge, we propose D-AdaST, a Distributed Adaptive minimax method with Stepsize Tracking. The key strategy is to employ an adaptive stepsize tracking protocol involving the transmission of two extra (scalar) variables. This protocol ensures the consistency among stepsizes of nodes, eliminating the steady-state error due to the lack of coordination of stepsizes among nodes that commonly exists in vanilla distributed adaptive methods, and thus guarantees exact convergence. For nonconvex-strongly-concave distributed minimax problems, we characterize the specific transient times that ensure time-scale separation of stepsizes and quasi-independence of networks, leading to a near-optimal convergence rate of $\tilde{\mathcal{O}} \left( \epsilon ^{-\left( 4+\delta \right)} \right)$ for any small $\delta>0$, matching that of the centralized counterpart. To our best knowledge, D-AdaST is the first distributed adaptive method achieving near-optimal convergence without knowing any problem-dependent parameters for nonconvex minimax problems. Extensive experiments are conducted to validate our theoretical results.

Paper

References (62)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer KXvx6/10 · confidence 3/52024-07-10

Summary

The paper presents D-AdaST, a distributed adaptive minimax method designed to address non-convergence issues in nonconvex-strongly-concave (NC-SC) minimax problems caused by inconsistent locally computed adaptive stepsizes in distributed settings. The method incorporates a stepsize tracking mechanism, which ensures consistency across local stepsizes and maintains time-scale separation, achieving near-optimal convergence rates comparable to centralized counterparts.

Strengths

The paper addresses a significant challenge in distributed minimax optimization by introducing a method that achieves near-optimal convergence rates. The theoretical contributions are well-founded, and the experimental validation is comprehensive.

Weaknesses

Please consider my questions for this part.

Questions

1- How does D-AdaST compare to other existing adaptive minimax methods in terms of computational time efficiency? 2-The current analysis focuses on nonconvex-strongly-concave minimax problems. How would the theoretical framework extend to other problem classes, such as nonconvex-nonconcave or convex-concave problems? What modifications or additional assumptions would be necessary? 3-In a distributed setting, communication delays and asynchrony can pose significant challenges. How does the theoretical analysis account for these factors? Are there any bounds or guarantees provided for performance in asynchronous or delayed communication settings?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

They discussed the limitations of their work in terms of assumptions and main results.

Reviewer Nhkx6/10 · confidence 3/52024-07-11

Summary

This paper proposed a decentralized stochastic first-order method for nonconvex minimax optimization. For nonconvex-strongly-concave setting, the proposed D-AdaST has the convergence rate of $O(\epsilon^{-4+\delta})$ for any small $\delta>0$.

Strengths

see questions

Weaknesses

see questions

Questions

The proposed method does not require the knowledge of the parameters $L$, $\mu$ and total iterations number, which is nice in empirical. I have some questions as follows: 1. This paper states D-AdaST achieves the near-optimal convergence rate, while the definition of optimality is unclear, e.g., problem setting and algorithm class. I’m not sure whether this statement is correct. Specifically, Assumption 2 requires the second-order Lipschitz continuous, while the existing lower for solving nonconvex-strongly-concave minimax problem by first-order methods only considers the first-order Lipschitz continuous. The author should clarify how to fill this gap and explain how to achieve the lower bound in the setting of D-AdaST more clearly. 2. Assumption 3 introduce the upper bound $C$ for stochastic gradient, which should be included in the final convergence rate. 3. Can you provide some explanation why the small $\delta$ cannot be avoided in result? What happens if we try to take $\delta\to0$? 4. The section of introduction describes the problem setting with constraints on both $x$ and $y$, while the algorithm design and convergence analysis looks only consider the constraint on $y$. 5. Is it possible to increase the batch-size of stochastic gradient to reduce the total communication rounds? Typically, the communication rounds in decentralized optimization can be smaller than the computation rounds, e.g., Chen et al., 2024.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer vU1y6/10 · confidence 2/52024-07-13

Summary

The authors introduced a new distributed adaptive minimax method, D-AdaST, to address the issue of non-convergence in nonconvex-strongly-concave minimax problems caused by inconsistencies in locally computed adaptive stepsizes. D-AdaST employs an efficient adaptive stepsize tracking protocol that ensures time-scale separation and consistency of stepsizes among nodes, thereby eliminating steady-state errors. Extensive experiments on both real-world and synthetic datasets validate their theoretical findings across various scenarios.

Strengths

Originality and Significance: The paper introduces a novel Distributed Adaptive Minimax Optimization Algorithm with a Stepsize Tracking Protocol. This approach is significant in addressing the non-convergence issues that arise when transitioning adaptive single-machine algorithms to a distributed environment. The authors provide a thorough theoretical analysis, which could facilitate the development of similar methods or modifications to existing single-machine algorithms. Quality and Clarity: The paper is well-structured and clear, with thoroughly explained assumptions and convergence rates, making it easy to understand.

Weaknesses

The experimental section has some deficiencies, as it only includes a GAN experiment on the CIFAR-10 dataset. I believe it would be beneficial to supplement the paper with results from additional datasets or other minimax optimization problems to provide a more comprehensive evaluation.

Questions

Please see weakness.

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

The paper employs general assumptions for distributed minimax optimization problems; therefore, it is unnecessary to discuss the limitations.

Reviewer rQb76/10 · confidence 2/52024-07-13

Summary

The paper introduces a method for distributed minimax optimization, for the scenario that various "agents" each hold part of a data set locally, and aim to coordinate to find the minimax solution of some criterion. Here, the criterion consists of the average of local cost functions, which are assumed to be smooth but non-convex-strongly-concave. Each agent is assumed to able to compute a stochastic gradient at each iteration, which can be communicated across the neighboring agents, who collectively communicate according to a given, known graph structure. In the nondistributed setting, known gradient-based methods depend on hyper-parameter tuning, in particular in choosing the stepsizes. When such a method is properly adaptive, it finds the right stepsize without apriori knowledge based on continuous adjustment per iteration. The authors construct counterexamples showing that directly applying adaptive methods designed for centralized problems results in non-convergence in distributed setting. The authors then propose an adaptive stepsize tracking protocol that involves transmitting two extra scalar variables to ensure consistency among stepsizes of different nodes. The authors give theoretical guarantees, matching (nearly) the convergence rate in the non-distributed setting. Furthermore, the authors exhibit their method on various real-world datasets to further underline their theoretical findings.

Strengths

* The article is well written: the authors explain the problem clearly and the algorithm is clearly outlined. * The result and formulation of Theorem 1 are insightful, and the synthetic example of (6) is simple yet very instructive. Together they outline the problem to be overcome in Section 2.2 nicely. * The authors provide an (almost) rate optimal theoretical guarantee for their proposed method. * The chosen real-world settings are both interesting and illustrative. * The proposed method is attractive in terms of communication complexity, with only a couple of scalars being required on top of the gradients.

Weaknesses

* Whilst I think the problem that the authors address is interesting, and within their scope they provide satisfying answers, I also think that this scope is rather limited. There are many things to investigate here in terms of e.g. network topology, communication efficiency, the influence of various noise sources. The authors stay within a set of rather stringent assumptions and do not capture the effect of these problem characteristics fully. * In high-dimensional settings, which the authors aim to address, achieving zero bias for stochastic gradients is not always feasible. Additionally, confirming the absence of bias in the gradient estimator is challenging to verify in practice. Does the method still work without this assumption? What if the bias is very small? * Similarly, the uniform bound on the gradient seems rather strong. Is this really required by the analysis?

Questions

Some of the figures are rather difficult to read when printed on paper as they are rather small, but were readable digitally. * Is there a reason the constant $C$ does not appear in e.g. (13)? * I do not see the point of Corollary 1. As I understand it, this is an upper bound for D-TiAda. The authors say that this result is provided for "proper comparison". How does Corollary 1 provide any comparison if it is only an upper bound? Isn't Theorem 1 to provide the comparison that the authors are after?

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

2

Limitations

Within the scope of the article, I think limitations are appropriately addressed.

Reviewer Nhkx2024-08-07

Thanks for your detailed rebuttal. I am satisfied with most of your response, while there are still some questions should be addressed. My main consideration is the applicability of the lower bound provided by Li et al. [6]. Noticing that their Definition 1 does not include second-order Lipschitz continuity. The lower bound of $\Omega(\epsilon^{-4})$ in their Theorem 2 is based on the function $f^{{\rm nc}-{\rm sc}-{\rm sg}}$. The proof of this theorem only verifies its gradient is Lipschitz continuous. However, the Lipschitz continuity of its Hessian is unclear. Therefore, we should check if the construction of Li et al. [6] holds the second-order Lipschitz continuity shown in Assumption 2. If the author can address this issue, I will increase my overall rating. Otherwise, the statement on the optimality should be avoided in revision (I will still keep borderline accept). **Some minor comments:** For Q4, I recommend to unify the presentation of constrained/unconstrained setting on the variable. For Q5, I think reduce the communication complexity is possible even if there is no average smoothness assumption. At least, this can be address in nonconvex minimization, e.g., [18] Yucheng Lu and Christopher De Sa. Optimal Complexity in Decentralized Training. ICML 2021.

Authorsrebuttal2024-08-09

Response to Reviewer Nhkx

Thank you for your prompt response and valuable comments. We have carefully considered your remaining concerns and have responded in detail as follows: >__Q6:__ My main consideration is the applicability of the lower bound provided by Li et al. [6]. Noticing that their Definition 1 does not include second-order Lipschitz continuity. The lower bound of $\varOmega \left( \epsilon ^{-4} \right)$ in their Theorem 2 is based on the function $f^{\text{sc-nc-sg}}$. The proof of this theorem only verifies its gradient is Lipschitz continuous. However, the Lipschitz continuity of its Hessian is unclear. Therefore, we should check if the construction of Li et al. [6] holds the second-order Lipschitz continuity shown in Assumption 2. If the author can address this issue, I will increase my overall rating. Otherwise, the statement on the optimality should be avoided in revision (I will still keep borderline accept). __Response__ We have carefully checked the construction of the hard examples for obtaining the complexity lower bound in [6] and verified that they satisfy second-order Lipschitz continuity on $y$ as assumed in Assumption 2. Specifically, we recall the hard instance in the stochastic setting in [6] (c.f., Eqs. (10) and (11)) as follows: $$ \bar{f}^{\text{nc-sc-sg}}\left( \boldsymbol{x},\boldsymbol{z};\bar{\boldsymbol{y}} \right) =-\Psi \left( 1 \right) \Phi \left( x_1 \right) +\sum_{i=2}^T{\left[ \Psi \left( -z_i \right) \Phi \left( -x_i \right) -\Psi \left( z_i \right) \varPhi \left( x_i \right) \right]}+\sum_{i=1}^{T-1}{\left[ c_1x_{i}^{2}+c_2z_{i+1}^{2} \right]}+\sum_{i=1}^{T-1}{h^{\text{sg}}\left( x_i,z_{i+1};\bar{\boldsymbol{y}}^{\left( i \right)} \right)}, $$ with $$ h^{\text{sg}}\left( x,z;\boldsymbol{y} \right) =\frac{C}{n}\left[ -\frac{1}{2}\boldsymbol{y}^{\text{T}}\left( \frac{1}{n^2}I_n+A \right) \boldsymbol{y}+\boldsymbol{b}_{x,z}^{\text{T}}\boldsymbol{y} \right] $$ and $\boldsymbol{b}_{x,z}=x\boldsymbol{e}_1-\frac{1}{2}z\boldsymbol{e}_n$, where $\boldsymbol{x}$ and $\boldsymbol{z}$ are variables to minimize, $\boldsymbol{y}$ is the variable to maximize, $\Psi \left( \cdot \right)$ and $\Phi \left( \cdot \right)$ are component functions, $c_1$, $c_2$ and $C$ are constants, $A$ is a positive semi-definite matrix (refer to Section 4 in [6] for detailed definitions). It should be noted that the assumption of second-order Lipschitz continuity of this work is only required to hold for $y$ (c.f., Eq. (11) in Assumption 2), and the terms related to $y$ in the objective function $\bar{f}^{\text{nc-sc-sg}}$ are only in $h^{\text{sg}}$, which is twice differentiable. It is not difficult to verify that $\nabla _{x,z;\boldsymbol{y}}^{2}h^{\mathrm{sg}}=-\left( \frac{1}{n^2}I_n+A \right)$ and $\nabla _{\boldsymbol{yy}}^{2}h^{\mathrm{sg}}=\mathrm{diag}\left\\{ 1,0,\cdots,0,-1/2 \right\\}$ are constant matrices and thus both are Lipschitz continuous. This indicates that $\bar{f}^{\text{nc-sc-sg}}$ is second-order Lipschitz continuous for $y$, satisfying Eq. (11) in Assumption 2. Therefore, we believe that the lower bound given by Li et al. [6] is applicable to the function class considered in our work. Combined with the previous discussion, we will carefully revise our statement about "near-optimal" to avoid any possible confusion. >__Q7:__ For Q4, I recommend to unify the presentation of constrained/unconstrained setting on the variable. __Response__ Thank you for your suggestion. We will do a proper revision of the presentation of the decision variable domains. >__Q8:__ For Q5, I think reduce the communication complexity is possible even if there is no average smoothness assumption. At least, this can be address in nonconvex minimization, e.g., > * [18] Yucheng Lu and Christopher De Sa. Optimal Complexity in Decentralized Training. ICML 2021. __Response__ We agree with the reviewer that reducing the communication complexity is possible even without relying on the average smoothness assumption. As mentioned by the reviewer, without this assumption, Lu et al. [18] combined local updates with the Factorized Consensus method (c.f., Lemma 1 in [18]) and Accelerated Gossip protocol (c.f., Algorithm 3), yielding reduced communication complexity. However, in order to obtain a computational complexity of $\varOmega \left( \epsilon ^{-3} \right)$ as in [12], this assumption is required in most of the literature based on the variance reduction methods. Otherwise, without this assumption, the sampling complexity obtained by Lu et al. [18] matches another lower bound $\varOmega \left( \epsilon ^{-4} \right)$. Having noted the characteristics and limitations of these methods, we believe they are all worthy of further investigation to improve the communication efficiency of distributed algorithms, with or without this average smoothness assumption.

Reviewer Nhkx2024-08-09

Thanks for your very detailed response. The response to Q7 and Q8 should be involved into the revision. Since my questions have been well addressed, I will raise my overall rating.

Authorsrebuttal2024-08-09

Thanks to Reviewer Nhkx

We are grateful for the reviewer's engagement and recognition. Thanks for spending your time carefully reviewing our paper and responding to our clarifications. We will incorporate the necessary revisions, taking into account all the points we have discussed. We believe that these changes will enhance the quality of our paper.

Reviewer rQb72024-08-07

I appreciate the detailed response by the authors. I wish to maintain my score.

Authorsrebuttal2024-08-09

Thanks to Reviewer rQb7

We appreciate the acknowledgment of the reviewer. Thanks for spending your time carefully reviewing our paper and responding to our clarifications.

Authorsrebuttal2024-08-10

Thanks to Reviewer KXvx

We appreciate the reviewer's acknowledgement. Thank you for taking the time to carefully review our paper and respond to our clarifications.

Reviewer vU1y2024-08-13

Thank you for the additional information regarding the experiment. I will accordingly adjust my score.

Authorsrebuttal2024-08-14

Thanks to Reviewer vU1y

Thanks for the thoughtful acknowledgement and consideration of our responses. We sincerely appreciate the reviewer's valuable time in reviewing our paper.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC