On the Last-iterate Convergence in Time-varying Zero-sum Games: Extra Gradient Succeeds where Optimism Fails

Last-iterate convergence has received extensive study in two player zero-sum games starting from bilinear, convex-concave up to settings that satisfy the MVI condition. Typical methods that exhibit last-iterate convergence for the aforementioned games include extra-gradient (EG) and optimistic gradient descent ascent (OGDA). However, all the established last-iterate convergence results hold for the restrictive setting where the underlying repeated game does not change over time. Recently, a line of research has focused on regret analysis of OGDA in time-varying games, i.e., games where payoffs evolve with time; the last-iterate behavior of OGDA and EG in time-varying environments remains unclear though. In this paper, we study the last-iterate behavior of various algorithms in two types of unconstrained, time-varying, bilinear zero-sum games: periodic and convergent perturbed games. These models expand upon the usual repeated game formulation and incorporate external environmental factors, such as the seasonal effects on species competition and vanishing external noise. In periodic games, we prove that EG will converge while OGDA and momentum method will diverge. This is quite surprising, as to the best of our knowledge, it is the first result that indicates EG and OGDA have qualitatively different last-iterate behaviors and do not exhibit similar behavior. In convergent perturbed games, we prove all these algorithms converge as long as the game itself stabilizes with a faster rate than $1/t$.

Paper

References (34)

Scroll for more · 22 remaining

Similar papers

Peer review

Reviewer tqyN5/10 · confidence 3/52023-06-18

Summary

This paper considers the problem of unconstrained min-max optimization with bilinear structure. Specifically, this paper considers the setting where the payoff matrix $A_t$ changes over time. The two changing dynamics of $A_t$ considered in this paper is the periodic game and the converging perturbed game. Three strategy dynamics are cosidered in this paper, which is OGDA, extra gradient (EG) and Negative momentum method. In this paper, the authors show that in the period game case, under certain initialization, OGDA provably diverges no matter what learning rate is chosen while EG with a certain choice of learning rate converges to the common NE with exponential rate. In the convergent perturbed game case, all the three types of algorithms converge with a proper choice of learning rate. The analysis is basd on analyzing the eigenvalues of the linear operator in the recursive form of the strategies. Empirical results further verifies their theoretical results.

Strengths

- The min-max optimization with time-varying functions is an important question and has wide real-world applications. - This paper first considers the last-iterate convergence performance for the unconstrained bilinear min-max problem and show that under two different changing dynamics of the bilinear matrix, OGDA performs provably differently and EG always converges. Specifically, the diverging results for OGDA looks intereting to me. - I did not check all the proof details but the general idea of the proof makes sense to me.

Weaknesses

- One issue is whether the optimization of unconstrained min-max optimization with bilinear structure is interesting. Specifically, even when the game is changing over time, one trivial Nash Equilibrium is $(0,0)$. Therefore, all the three algorithms initialized with $(0,0)$ will converge, though the dynamic of these three algorithms may still be interesting with different initializations. Extending the current framework to a slightly general case $x^\top A_ty+b_t^\top x+c_t^\top y$ or an even more general convex-concave function $f(x,y)$ will strengthen the results a lot. - As the authors mentioned, the constrained version of the min-max optimization problem may be more interesting and the dynamic of the strategy can not be expressed in a clean linear recurrsive form as shown in the unconstrained bilinear case.

Questions

- As mentioned in weakness, whether the current result can be extended to a generalized linear case with $f_t(x,y)=x^\top A_ty+b_t^\top x+c_t^\top y$? - Is there a specific reason considering the periodic games and the convergent perturbed games? Can the results generalize to some other type of time-varying games with common NE?

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

3 good

Contribution

3 good

Limitations

See weakness and questions for details.

Reviewer PugT6/10 · confidence 3/52023-07-06

Summary

This paper studies the last-iterate behavior of three different algorithms on two types of time-varying zero-sum games: periodic and convergent perturbed games. For periodic games, the authors prove that EG will converge while OGDA and the negative momentum method could diverge. For convergent perturbed games, all these algorithms converge as long as the error term decays sufficiently fast.

Strengths

1. The paper is well-organized and easy to follow. 2. The different behavior between EG and OGDA seems novel in the literature.

Weaknesses

1. Some results lack an intuitive explanation. 2. Some figures are not well-plotted. Maybe a log-log graph is more proper.

Questions

1. Could the authors provide some examples of periodic and convergent perturbed games? 2. In line 225, the authors say that the key point on the convergence of EG for periodic games is that the iterative matrix is a normal matrix. Could the authors give a more intuitive explanation? Moreover, the update rule of EG only depends on the current point $(x_t, y_t)$, while OGDA and negative momentum method also need to employ the gradient information at the last point $(x_{t-1}, y_{t-1})$. Is this a reason for the different behavior of these methods? 3. For the experiments on Theorem 3.4, the authors only plot the convergent curve of EG. I wonder about the behabior of the other two methods in this case.

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

3 good

Contribution

3 good

Limitations

See Questions.

Reviewer qpgT7/10 · confidence 3/52023-07-07

Summary

This paper studies the problem of learning Nash equilibria in two-player zero-sum bilinear games where the payoff matrix varies with time. When the payoff matrix is a periodic function, it is proven that the extra gradient algorithm converges to a Nash equilibrium, whereas the optimistic gradient descent ascent and negative momentum method could diverge from the equilibrium. Furthermore, it is shown that all the above three algorithms converge to the equilibrium for the convergent perturbed game.

Strengths

* The paper is well-written and studies an interesting problem. * The author provided an instance of periodic zero-sum games where extra gradient algorithms and optimistic gradient descent ascent algorithms behave distinctly differently. * The analysis looks to be sound, and the experimental results seem to bare out the theory.

Weaknesses

* Perhaps a short intuitive explanation of why extra gradient is better than optimistic gradient descent ascent would help readers understand the main results of this paper. * The titles and legends in Figures 1,2,3 and 4 should be enlarged.

Questions

* (Section 2.1) What is the definition of $\mathrm{ker}(\cdot)$? * (Section 3.2) What are the key differences between the extra gradient algorithm and the other two algorithms? * (Section 3.2) For the periodic games with $\lambda_{\ast}=0$, would similar convergence/divergence results hold to Theorem 3.1 and 3.2? * (Section 4) Do the optimistic gradient descent ascent algorithm and negative momentum method empirically converge in the same instance in the experiments for Figure 4?

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

3 good

Contribution

3 good

Limitations

The authors adequately addressed the limitations.

Reviewer oHHw7/10 · confidence 3/52023-07-09

Summary

Over the last few years an extensive literature has studied the last iterate convergence of learning dynamics in zero-sum games, particularly those using optimism and extra-gradient approaches. This paper extends that approach to two classes of unconstrained non-stationary game (periodic and decaying noise) and shows a difference in behavior between the two classes.

Strengths

The problem studied is natural and clearly explained with strong results, including the interesting split in behavior (Theorems 3.1 and 3.2) The approach uses several technical innovations derived from the literature on non-autonomous linear difference systems.

Weaknesses

The particular classes of games studied could be better motivated. Is there a reason to focus on these other than technical convenience? Do they have important applications? What more general but harder class is this a step towards? The conclusion mentions constrained games as future work, nothing about other interesting directions fof future work. Are there other theorems from non-autonomous linear difference systems that seem promising? There is a gap left by Theorems 3.3 and 3.4 about whether the other two dynamics converge under the weaker conditions of Theorem 3.4. Figure 4 gives an experiment on this setting, so it seems like a missed opportunity to not at least show what the other dynamics do on this example. Minor comments: - There is some inconsistency in the way “equilibrium” is used. It is defined on line 98 as what I would instead have called the set of equilibria. However, in other places, like line 252 it seems to be used to refer to a single element of this set. Depending on how this is resolved, some care may be needed in places about whether something is “the” or “a” equilibrium (e.g. on line 216). - Line 178 and 181 the formulas use n but the limits are taken over t - Line 216 the “common Nash equilibrium” is not defined, nor is there an explanation of why it mush exist (presumably because 0 is always an equilibrium?) - Line 305 group[s]

Questions

Please respond regarding the two main 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

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

3 good

Contribution

3 good

Limitations

The conditions to which the results apply are clearly explained.

Reviewer abUR5/10 · confidence 3/52023-07-22

Summary

This work investigates the problem of last-iterate convergence in time-varying zero-sum games. Specifically, the authors study the last-iterate convergence of three kinds of algorithms (OGDA, EG, and negative momentum method) considering two kinds of time-varying games with specific structures, i.e., periodic games and convergent perturbed games. The authors obtained a convergence rate of EG in periodic games while providing a counterexample showing that OGDA and negative momentum method will diverge in this kind of game. In convergent perturbed games, the authors showed that all three kinds of algorithm will converge with a rate dependent on the perturbation $B_t$. Finally, experiments validate the statements proposed by the authors.

Strengths

In general, I think there is no last-iterate convergence rate in time-varying games due to the changing nature of the games. However, with specific problem structures, I believe that last-iterate convergence results can be established, and this work provides a piece of clear evidence. Overall, although the solutions are relatively simple by modeling the learning algorithms as dynamical systems and leveraging some existing results in the control theory, I am satisfied with the motivation and the final results of this work.

Weaknesses

There are no significant weaknesses that affect my rating of the whole paper. In the following, I only list some minor issues. 1. The notion of $T$ in periodic games. The notion of $T$ is commonly used to refer to the total number of rounds in online learning. In this work, the authors use $T$ to represent the number of rounds inside each period, which may lead to unnecessary understandings. I suggest the authors could choose a different notion to denote the period length. 2. The notion of $\text{ker}(\cdot)$. This notion, which appears in Line 98 for the first time, seems not to be defined or given a clear description before. 3. In Line 120, the authors mention that 'Recently, there are also works analyzing the regret behaviors of OGDA under a time-varying setting [1]'. The reference to [1] is not accurate enough. The work of [33], which is published at ICML2022, firstly gives a comprehensive study of the optimistic methods in time-varying zero-sum games by optimizing multiple performance measures simultaneously. I suggest the authors could refer to [33] in the aforementioned statement to make the credits more accurate. 4. The negative momentum method requires the $x$-player to evolve to the next round first to give its gradient $A_{t+1} x_{t+1}$ to the $y$-player. In some cases where the learning procedure is strictly in an online style, this algorithm is NOT applicable since both players must act first to make the game evolve. I suggest the authors could give a discussion on this point in the revised version. Typos: 1. Line 89: 'Section 4' not 'section 4'. 2. Line 188: 'Section 3.3' not 'Section3.3'. 3. The reference to theorems are not unified, e.g., in Line 237, 'theorem 3.2', and in Line 255, 'Theorem (3.1)'.

Questions

I only have a major question on Theorem 3.1. What is the range of $t$? If $t$ is only chosen in $[1,T]$, choosing $t=T$ yields a convergence rate of $O(\lambda_* \text{poly}(T))$, which can be seen as a constant in terms of $T$. As mentioned in the 'Weaknesses' part, a bound linear in $T$ is pretty bad in the standard online learning convention. Theorem 3.1 exhibits an exponential rate only when $t$ can be significantly larger than the period length. Can the authors give some further explanations on this point? In the end, I am very curious about whether the method in this paper, which models the learning algorithms as dynamical systems, can be applied to the constrained case, where the projection operations may bring some unique challenges to the modeling issue. Can the authors give some further explanations on this point?

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

3 good

Contribution

2 fair

Limitations

Please mainly refer to the 'Weaknesses' part.

Reviewer abUR2023-08-11

Thanks for the explanations

Thanks for the detailed feedback, which has fully answered my questions.

Authorsrebuttal2023-08-12

You are very welcome! Thanks a lot for reading and agreeing with the response.

Reviewer PugT2023-08-12

Thanks for the replies.

Thanks for the detailed replies, which have addressed my concerns. I increase my score to 6.

Authorsrebuttal2023-08-12

Thank you for your interest to the work and intriguing questions. We appreciate your boosting to the score, thank you so much!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC