Payoff-based learning with matrix multiplicative weights in quantum games

In this paper, we study the problem of learning in quantum games - and other classes of semidefinite games - with scalar, payoff-based feedback. For concreteness, we focus on the widely used matrix multiplicative weights (MMW) algorithm and, instead of requiring players to have full knowledge of the game (and/or each other's chosen states), we introduce a suite of minimal-information matrix multiplicative weights (3MW) methods tailored to different information frameworks. The main difficulty to attaining convergence in this setting is that, in contrast to classical finite games, quantum games have an infinite continuum of pure states (the quantum equivalent of pure strategies), so standard importance-weighting techniques for estimating payoff vectors cannot be employed. Instead, we borrow ideas from bandit convex optimization and we design a zeroth-order gradient sampler adapted to the semidefinite geometry of the problem at hand. As a first result, we show that the 3MW method with deterministic payoff feedback retains the $\mathcal{O}(1/\sqrt{T})$ convergence rate of the vanilla, full information MMW algorithm in quantum min-max games, even though the players only observe a single scalar. Subsequently, we relax the algorithm's information requirements even further and we provide a 3MW method that only requires players to observe a random realization of their payoff observable, and converges to equilibrium at an $\mathcal{O}(T^{-1/4})$ rate. Finally, going beyond zero-sum games, we show that a regularized variant of the proposed 3MW method guarantees local convergence with high probability to all equilibria that satisfy a certain first-order stability condition.

Paper

References (61)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer UZra7/10 · confidence 4/52023-07-03

Summary

The paper studies the setting of learning in quantum games, in the case where agents only have access to a scalar (for instance, a payoff value) instead of their full individual gradient, and have to update their strategies in order to converge to some equilibrium. The primary contribution is a variant of the well-studied MMW algorithm called 3MW, which matches the equilibrium convergence rate of (full information) MMW in 2-player quantum zero-sum games. Moreover, with a relaxation of 3MW’s information requirements (i.e. that players observe a random realization of their payoff observable), the algorithm converges to equilibrium at O(T^{-1/4}) rate. Finally, in general N-player quantum games, a regularized version of 3MW using estimator 1PE is able to locally converge to Nash equilibria that are variationally stable.

Strengths

The paper answers an important question in the realm of learning in quantum games - in this framework it is quite unreasonable to assume full information of gradients, especially without properly defining how measurements are made, how players access said measurements etc. Hence, in my opinion the bandit framework is very fitting for the quantum setting, and indeed players should be able to update their strategies in a truly ‘online’ setting using only their measured/realized payoff observable. The techniques used in the paper, while not entirely novel, show interesting similarities and differences between classical and quantum settings. Finally, I think the results on variational stability are quite significant, as not only does this give (to my knowledge) the broadest last-iterate convergence result in N-player quantum games yet, it also points to a potential class of FTRL variants using bandit feedback that could potentially also be convergent in the last-iterate sense.

Weaknesses

I think overall this is a strong paper, the only issue to me is that most of the techniques and ideas used in the paper are not particularly different from the classical bandit setting. If there is significant technical novelty, it is not made clear to the reader in the main text. I would be very interested to see what the differences are in the analysis between the quantum and classical setting, if there are any.

Questions

In the experiments, 3MW is compared with MMW and is shown to exhibit decent performance even with the lack of information. Does the practical performance of 3MW deteriorate significantly for games of larger dimension? To what extent can we rely on 3MW in practical settings (i.e. large games, games with larger number of players)?

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

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Limitations are very briefly discussed, and perhaps a longer discussion would improve the paper (if space allows)

Reviewer Kb526/10 · confidence 5/52023-07-04

Summary

This paper studies online learning of quantum algorithms based on the matrix multiplicative weight (MMW) updates. A main technical contribution is to introduce a weaker version called the minimal-information matrix multiplicative weights (3MW), which only requires zeroth-order gradient sampler and can achieve an overall regret of O(T^{-1/4}). In addition, for N-player quantum games, the paper proved a convergence result that 3MW guarantees local convergence with high probability to all equilibria that satisfy a certain first-order stability condition.

Strengths

From my perspective, the most notable strength of this paper is the nice observation that less information is required to make MMW works. Considering that MMW has wide applications in quantum algorithms for machine learning and optimization, I believe that the result may be of general interest. Technically, this is achieved by exploiting ideas from machine learning, in particular bandit optimization, which provides interesting perspective for designing future quantum algorithms.

Weaknesses

The most notable weakness is the claim in intro that the result by Jain and Watrous [21] with O(\sqrt{T}) regret for quantum online learning algorithms remains the tightest known bound for Nash equilibrium learning in quantum games. This is not true, as a recent paper by Gao et al. https://arxiv.org/abs/2304.14197 gave a quantum algorithm with poly-loga regret for zero-sum games. Compared to this paper, the results obtained by this paper is sub-optimal. Another notable weakness is that the dependence in dimension is at least D, and does not present a quantum speedup here. In fact, a main merit of applying MMW into quantum algorithms is that MMW provides fast quantum algorithms for solving relevant problems. An important paper was Brandao and Svore https://arxiv.org/pdf/1609.05537.pdf in FOCS 2017, which can solve semidefinite programs with complexity square-root in D. If the current framework 3MW has at least linear dependence in D, it will prohibit its further applications in quantum computing.

Questions

1. Can the authors further comment on whether the convergence rate of the proposed 3MW algorithm, or in other words the online regret, can be further improved? Given the existence of works such as Gao et al. mentioned above, it would be helpful if the authors can provide a more complete picture here. 2. I’m confused by the sentence “in contrast to classical finite games, quantum games have an infinite continuum of pure states” in the abstract. On classical computers, even if we play games with finite cardinality, we can still choose mixed strategies and the set of probability distributions supported on a finite set can be uncountable. 3. It would in general be helpful if the authors can introduce more on the applications of 3MW. In particular, in MMW the payoff gradient of player i is defined as V_i(X_t). However, in most existing quantum papers, V_i is very simple: the identity function. For instance, in the above Brandao-Svore paper as well as several important follow-up works that solves semidefinite programs with quantum speedup, V_i is the identity function and the update is simply a learning rate \eta * X_t. In all, it’s nice to know that 3MW can be applied with bandit information and for general V_i, but this should probably be further motivated about why we need to consider more complicated V_i’s. 4. In numerical experiments, around Page 29, the authors claimed that the their algorithms achieved a rate of convergence closer O(1/T) instead of the weaker O(1/\sqrt{T}) or O(1/T^{1/4}) bounds. However, I didn’t see detailed evidence for this; it would be helpful if the authors can present more detailed data and plots for demonstrating this phenomenon. Note that this is related to my question 1: if the convergence rate is ~O(1/T), the overall online regret will be poly-log. So it is probably worth closer study about this, both theoretically and experimentally.

Rating

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

2 fair

Limitations

This work is mainly theoretical so I don't think there are notable limitations, but it would work better if the authors can articulate this more clearly.

Reviewer G5vn6/10 · confidence 3/52023-07-16

Summary

This paper studies Quantum Games (mainly in the two-player zero-sum setting) and develops payoff-based learning algorithms with provable convergence-rate guarantees. The learning algorithm combines matrix multiplicative weights (MMW) and zeroth order method for gradient estimation. The overall convergence rate is either $O(1/\sqrt{T})$ or $1/T^{1/4}$, depending the information structure. The authors also provide numerical simulations in the appendix to verify the performance of the proposed algorithm.

Strengths

This paper is very well-written. The appendix is also of high quality.

Weaknesses

(1) The learning algorithm combines MMW with zeroth order method for gradient estimation. Since the convergence rate of MMW was established in the literature, and the zeroth order method is well studied in optimization, the technical novelty is not entirely clear. Are there any significant challenges in the analysis? (2) From Theorem 2, it seems that there are no consequences for choosing $\delta$ to be arbitrarily small. Am I missing anything?

Questions

Do the authors expect the bound to be tight? Are there any known lower bounds for the convergence rate in the literature?

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

The authors did not explicitly point out the limitations of this work.

Reviewer TyoP5/10 · confidence 3/52023-07-22

Summary

The authors consider the quantum games in the bandit setting, which players observe only the payoff value of their actions. In order to facilitate this, the authors extend the online bandit estimation methodology into the quantum-state space. As a result, they are able to realize a duality gap of $\mathcal O(T^{-\frac{1}{4}})$ and $\mathcal O(T^{-\frac{1}{2}})$ for bilinear quantum games.

Strengths

- Originality & Significance: Considering the concept of games within quantum state spaces in the bandit setting is innovative, and the result is nice and matches those found in classical matrix games. - Quality & Clarity: The paper is effectively structured, allowing readers to follow the presented arguments and conclusions with ease.

Weaknesses

- The applications of blinear quantum games are questionable.

Questions

- I would like to inquire about the necessity of the $L$-smooth assumption in the quantum game setting. Specifically, it seems that in classical games or online learning problems, the bandit estimator can be constructed without requiring an $L$-smooth condition. Could the authors elaborate on whether this condition is necessary in the context of quantum games and, if so, could you explain the rationale behind this requirement?

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

N/A

Reviewer Kb522023-08-10

Acknowledgement

I would like to thank the authors for the very detailed reply and the additional numerical experiments. Those provide much clearer understandings, and I'm happy to subsequently increase my score. For the final version, I suggest the authors to: 1) highlight that the setting of this paper is to study classical algorithms for solving quantum games, and make comparison to quantum computing literature in other settings; 2) add the new numerical experiments; and 3) add discussions to my other questions for more clear presentation.

Authorsrebuttal2023-08-12

Thank you for your acknowledgment

Thank you for your quick reply and for upgrading your score, we will update our maunscript according to your detailed suggestions - thanks again for your helpful comments! Regards, The authors

Reviewer G5vn2023-08-13

Acknowledgement of the Rebuttal

Thank the authors for the detailed response. I do not have further questions.

Authorsrebuttal2023-08-21

We are very pleased for answering your questions. Thank you again for your helpful comments and positive evaluation. Regards, The authors

Reviewer UZra2023-08-15

Response to Author Rebuttal

Thank you for your answers to the questions. The explanation for differences between classical and quantum proofs for the bandit setting is quite compelling, and it would be nice to see a subsection summarizing this in the main text if space allows. The experiments added above are also quite helpful for a practical comparison. Finally, I find making it clear that the paper is about classical algorithms for quantum games to be quite imperative - it can be easy to conflate the two settings. Best regards, Reviewer UZra

Authorsrebuttal2023-08-21

Thank you very much for your input and positive evaluation. We will update our manuscript accordingly. Regards, The authors

Reviewer TyoP2023-08-16

I appreciate the authors addressing my concerns regarding the L-smoothness condition and the potential applications of bilinear quantum games. I would like to maintain the score within the acceptance threshold.

Authorsrebuttal2023-08-21

We are very glad for addressing your questions. Thank you again for your remarks and your positive evaluation. Regards, The authors

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC