Logarithmic-Regret Quantum Learning Algorithms for Zero-Sum Games

We propose the first online quantum algorithm for solving zero-sum games with $\widetilde O(1)$ regret under the game setting. Moreover, our quantum algorithm computes an $\varepsilon$-approximate Nash equilibrium of an $m \times n$ matrix zero-sum game in quantum time $\widetilde O(\sqrt{m+n}/\varepsilon^{2.5})$. Our algorithm uses standard quantum inputs and generates classical outputs with succinct descriptions, facilitating end-to-end applications. Technically, our online quantum algorithm"quantizes"classical algorithms based on the optimistic multiplicative weight update method. At the heart of our algorithm is a fast quantum multi-sampling procedure for the Gibbs sampling problem, which may be of independent interest.

Paper

Similar papers

Peer review

Reviewer 8qKg7/10 · confidence 3/52023-06-14

Summary

The paper presents an improved quantum online learning algorithm for approximating the Nash equilibrium of a zero-sum game. Notably, this is the first algorithm to achieve $\tilde{O}(1)$ regret with quantum speedup. The presented algorithm can then be applied to linear programming problems using the primal-dual approach. Additional contribution of the paper is a sampling method for Gibbs distribution for obtaining multiple samples with quantum speedups on both the number of samples and the number of possible outcomes.

Strengths

- Multiple novel contributions on quantum algorithms: speedups for multi-Gibbs sampling and logarithmic regret Nash equilibrium approximation - The paper is well-written and easy to follow

Weaknesses

While the presented theoretical results are interesting, I consider the necessity of QRAM as a minor weakness since (to my knowledge) it is still a highly theoretical model whose physical realizability is unclear.

Questions

The presented algorithm requires QRAM to work. How does this compare to the quantum algorithms of regret $\tilde{O}(\sqrt{T})$, i.e., do they too require QRAM? Typos: - Line 285 should probably say "Line 7"

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

I do not consider the work to have any serious limitations other than the ones that follow from the limited practicality of quantum computing (for now).

Reviewer XSuW6/10 · confidence 4/52023-07-05

Summary

The paper studies the online learning quantum algorithms for zero-sum games. It proposes the online quantum algorithm for zero-sum games with near-optimal regret, which makes progress to online learning algorithms in the quantum setting. The algorithm quantizes classical algorithms based on the MMW method and incorporates a fast quantum multi-sampling procedure for the Gibbs sampling problem. By developing a sample-based stochastic version of the optimistic multiplicative weight update method, this method achieves a quadratic improvement over classical algorithms and presents a fast quantum linear programming solver as an application.

Strengths

1. The paper addresses a timely and relevant topic in quantum learning theory, proposing the first online quantum algorithm for zero-sum games with near-optimal regret. 2. The algorithm achieves a quadratic improvement over classical algorithms for computing the approximate Nash equilibrium. 3. The paper presents a fast quantum linear programming solver as an application.

Weaknesses

1. The main algorithm is derived from classical methods, and the explanation of the features/contributions of quantum computing in achieving the speedup or advantage could be further elaborated. 2. As a theory paper, it could benefit from a clear and well-motivated example demonstrating the application of the proposed algorithm. Including such an example would help readers better understand the potential practical implications and usefulness of the work. 3. Could the authors discuss whether this online quantum algorithm could be dequantized? 4. In the abstract, the authors claim that "a fast quantum linear programming solver" but haven't identified under which setting.

Questions

Refer to the part about weaknesses.

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

2 fair

Contribution

3 good

Limitations

NA.

Reviewer UumW6/10 · confidence 4/52023-07-05

Summary

This paper considers the problem of developing quantum, low-regret, algorithms for solving zero-sum games. This paper provides a quantum algorithm which matches the state-of-the-art for solving zero-sum games while improving upon the regret of the associated algorithm. The paper achieves this result by providing a stochastic variant of optimistic mirror descent which they then implement by a quantum algorithm for developing multiple samples from a suitable Gibbs distribution. Additionally, the paper discusses implications and lower bounds for linear programming more broadly.

Strengths

Zero sum games are an important optimization problem in many disciplines. The mere fact that the paper shows how to depart from previous quantum algorithms for solving zero sum games (in using optimistic mirror descent versus mirror descent) while retaining state-of-the-art bounds is interesting and potentially useful for further study of the problem. Additionally, by following this alternative approach the paper obtains improved regret and a type of parallelism in solving the problem. This is an interesting advancement for the study of the problem. Finally, this alternative optimization method gives rise to a different Gibb’s distribution sampling problem than the one considered in prior work. Furthermore, the paper gives an interesting algorithm for solving it which could motivate further study. Finally, the additional discussion of linear programming provided in the paper (though seemingly straightforward) is a nice addition as well.

Weaknesses

It is not clear how substantial the departure from this paper from prior work is, how novel the advancements are, and how complete the comparison of this paper to prior work is. While I think the advancements of the paper are interesting and potentially exciting (as discussed in the strength section), I fear that the paper does not well-articulate them or situate them. For example: * A comparison of the techniques in the Gibb’s sampling algorithm of this procedure and those in prior quantum-zero sum games algorithm does not seem to be provided. This seems particularly important as the techniques were key to recent advances in quantum algorithms for solving zero-sum-games. * It is not clear what exactly the benefits of low regret are over prior work and why previous results couldn’t be used to obtain low-regret (e.g. by only considering a subsequence of the iterates taken). I think the low-regret aspect is interesting and this algorithm is potentially more parallel than prior algorithm’s with state-of-the-art complexity and the sampling problems considered may be simpler than in prior work; however I am not sure that the paper argues this clearly. * It is not clear what obstacles were present to obtaining the result. It would help if the paper clarified whether moving from mirror-descent to optimistic mirror-descent was the main insight and the rest was straightforward (though technical and important) or if further insights were needed. Some additional comments along this line are given “Questions.”

Questions

My main questions / comments are the items in the previous “Weaknesses” section. Beyond what was written there: How does the Gibb’s sampling compare to prior work? Is it simpler in any way? The same? Is a completely different techniques used? Is there a major insight of the paper beyond the use of stochastic optimistic mirror descent? Beyond this, below are some more detailed suggestions, questions, and comments: * Line 1: “for zero-sum games” --> “for solving zero-sum games” * Line 4: “yielding a quadratic improvement” as written, I am concerned this is suggesting that it is the first such improvement. * Line 27: “For, this approximation task, online learning becomes significant” – I am not exactly sure what was meant by this. * Line 33: It might be beneficial to define regret or explain the model more clearly earlier so the improvement over prior work is clearer. (Regret is well-known in general, but when stated for a static optimization problem, the definition is less clear.) * Line 39: I believe zero-sum games were shown to be solvable in $\tilde{O}(1/\epsilon)$ iterations earlier, e.g. “Prox-method with rate of convergence O (1/t) for variational inequalities with Lipschitz continuous monotone operators and smooth convex-concave saddle point problems” by Nemirovski in 2004. I’m less sure about whether this works in regret context but in any event I believe should be mentioned when discussing the history of solving zero-sum games. * Line 81: it might be good to clarify whether this is a new feature of this algorithm or not * Line 130 – 135: it might be good to clarify why prior work could or could not be used for this problem My main questions / comments are the items in the previous “Weaknesses” section. Beyond what was written there: How does the Gibb’s sampling compare to prior work? Is it simpler in any way? The same? Is a completely different techniques used? Is there a major insight of the paper beyond the use of stochastic optimistic mirror descent? Beyond this, below are some more detailed suggestions, questions, and comments: * Line 1: “for zero-sum games” --> “for solving zero-sum games” * Line 4: “yielding a quadratic improvement” as written, I am concerned this is suggesting that it is the first such improvement. * Line 27: “For, this approximation task, online learning becomes significant” – I am not exactly sure what was meant by this. * Line 33: It might be beneficial to define regret or explain the model more clearly earlier so the improvement over prior work is clearer. (Regret is well-known in general, but when stated for a static optimization problem, the definition is less clear.) * Line 39: I believe zero-sum games were shown to be solvable in $\tilde{O}(1/\epsilon)$ iterations earlier, e.g. “Prox-method with rate of convergence O (1/t) for variational inequalities with Lipschitz continuous monotone operators and smooth convex-concave saddle point problems” by Nemirovski in 2004. I’m less sure about whether this works in regret context but in any event I believe should be mentioned when discussing the history of solving zero-sum games. * Line 81: it might be good to clarify whether this is a new feature of this algorithm or not * Line 130 – 135: it might be good to clarify why prior work could or could not be used for this problem

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

2 fair

Contribution

3 good

Limitations

This is primarily a theory paper, and it is unclear how applicable this is. However, the weaknesses and questions discussed reflect limitations for which further discussion might be beneficial.

Reviewer 42Uz6/10 · confidence 4/52023-07-06

Summary

In this paper the authors consider the task of computing the eps-approximate Nash equilibrium of zero-sum games. In particular, for a matrix A in R^{m x n}, the goal is to find the minimum distributions x,y s.t., max_y x^t A y - min_x x^t A y is at most epsilon, given quantum query access to A. In this paper the authors are further concerned with the online learning/regret model, i.e., in each round the players need to choose probability distributions (u,v) and their goal is to pick these distributions (u,v) such that it is as close as possible to the optimal choice of eps-Nash equilibrium. Prior known quantum algorithms for computing the Nash equilibrium ran in time sqrt{m+n}/eps^2.5 but achieved a regret of sqrt{T} for a T round algorithm. In this paper, the authors show how to obtain a regret of O(1) with the same complexity as sqrt{m+n}/eps^2.5.

Strengths

The main strength of this paper is in a way it solves an important question of computing Nash equilibrium with a O(1)-based regret algorithm. Using this, they algorithm obtain a state of the art quantum algorithm for linear programming solvers. I think both of these contributions are interesting on their own right!

Weaknesses

As for weakness, i feel there are a couple of weaknesses: (i) The main techniques used in this paper, aren't very novel. Indeed, they do improve upon prior works and need new tools, but the novelty of the main contributions towards obtaining their speedups doesn't make it a strong contribution. For example, obtaining k copies of a Gibbs state in time sqrt{nk} (compared to the trivial k sqrt{n} algorithm) isn't too surprising (indeed new, but not surprising given Hamoudi's algorithm and also complexity of Grover's search with k marked items). (ii) The speedup to obtain the regret bound is by quantizing the optimistic online algorithm. I think there are some subtelties when quantizing this algorithm, but the main ideas involved aren't novel. Having said the above, i think overall for the ML community, i think the main results and these interesting quantum subroutines above are nice.

Questions

I think the definitions of online learning and regret can be defined much better. In particular, the online learning/regret model for Nash equilibrium is very messily written and I'd have appreciated a better explanation.

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

N/A

Reviewer 8qKg2023-08-11

Reply to Authors

Thank you for addressing my concerns in a satisfying way. While I still consider the necessity of the QRAM as a minor weakness, it then seems like a necessary evil that is hard to get rid of in this case (as is the case with many quantum algorithms).

Reviewer UumW2023-08-14

Response to Rebuttal

Thank you for the reply. I believe adding the information in your rebuttal (and better defining / articulating what exactly the low regret setting / problem means for this paper) would all improve it.

Authorsrebuttal2023-08-15

Thank you!

Thanks for your further comments. We will definitely incorporate our discussion in the rebuttal into future versions of this paper. Regarding defining/articulating what exactly the low regret setting/problem means for this paper, we want to explain more as follows: - Lower regret means fewer interactions between players when they reach an approximate Nash equilibrium in the game setting and fewer rounds for the algorithm to find a good approximation of the optimal value for the linear programming solver. In previous quantum approaches (Refs. [50, 7]), they all have $\tilde O (\sqrt{T})$ regret; that is, the regret would be bounded by the square root of the number of rounds. By contrast, our logarithmic-regret algorithm has a near-constant bound (up to logarithmic factors) of regret. Thus, our algorithm has a good theoretical dependence on the number of rounds for finding the Nash equilibrium of zero-sum games and for finding an approximate optimal value for linear programming problems. We will add these explanations in the revised version of our paper.

Reviewer 42Uz2023-08-15

I thank the authors for their reply. For the time being, I'll keep my rating and keep in mind your comments for future discussions.

Reviewer XSuW2023-08-20

Thanks for the reply that helps me better understand. I have no further questions.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC