Nash Regret Guarantees for Linear Bandits

We obtain essentially tight upper bounds for a strengthened notion of regret in the stochastic linear bandits framework. The strengthening -- referred to as Nash regret -- is defined as the difference between the (a priori unknown) optimum and the geometric mean of expected rewards accumulated by the linear bandit algorithm. Since the geometric mean corresponds to the well-studied Nash social welfare (NSW) function, this formulation quantifies the performance of a bandit algorithm as the collective welfare it generates across rounds. NSW is known to satisfy fairness axioms and, hence, an upper bound on Nash regret provides a principled fairness guarantee. We consider the stochastic linear bandits problem over a horizon of $T$ rounds and with set of arms ${X}$ in ambient dimension $d$. Furthermore, we focus on settings in which the stochastic reward -- associated with each arm in ${X}$ -- is a non-negative, $\nu$-sub-Poisson random variable. For this setting, we develop an algorithm that achieves a Nash regret of $O\left( \sqrt{\frac{d\nu}{T}} \log( T |X|)\right)$. In addition, addressing linear bandit instances in which the set of arms ${X}$ is not necessarily finite, we obtain a Nash regret upper bound of $O\left( \frac{d^\frac{5}{4}\nu^{\frac{1}{2}}}{\sqrt{T}} \log(T)\right)$. Since bounded random variables are sub-Poisson, these results hold for bounded, positive rewards. Our linear bandit algorithm is built upon the successive elimination method with novel technical insights, including tailored concentration bounds and the use of sampling via John ellipsoid in conjunction with the Kiefer-Wolfowitz optimal design.

Paper

References (28)

Scroll for more · 16 remaining

Similar papers

Peer review

Reviewer K5vq6/10 · confidence 3/52023-06-30

Summary

In this research, the authors investigate the concept of Nash regret in the context of stochastic linear bandits. Nash regret represents the disparity between the optimal reward and the geometric mean of expected rewards, reflecting fairness considerations across different time instances within the Nash social welfare framework. The study assumes a fixed set of arms denoted as $\mathcal{X}$, and an unknown vector $\theta^*$. When an arm x is chosen, the agent observes a sub-Poisson $r_t$ with a mean of $<x,\theta^*>$. The nash regret is defined as $max_x<x,\theta^*>-(\prod_t \mathbb{E}[<X_t,\theta^*>])^{1/T}$ where $X_t$ is a selected arm at time t. To minimize Nash regret, the authors propose an algorithm that incorporates elimination phases. Firstly, to ensure the lower bound of expected rewards, they employ sampling via the Jon ellipsoid technique, as in Lemma 4. Additionally, they employ a G-optimal design to construct a probability distribution over arms, aiming to minimize the confidence interval width of estimated rewards. By combining these strategies, a sequence of arms is generated, and elimination phases are employed to remove arms with suboptimal mean rewards, utilizing upper confidence bounds (UCB) and lower confidence bounds (LCB) with a doubling time horizon for exploration. To construct the UCB and LCB, a Nash confidence bound is utilized, which contains the mean reward value. Larger mean rewards correspond to larger confidence intervals, as defined in equation (8). The theoretical analysis demonstrates that for a finite number of arms, the proposed algorithm achieves a nash regret of approximately $\tilde{O}(\sqrt{d/T})$. In the case of an infinite number of arms, the regret is $\tilde{O}(d^{5/4}/\sqrt{T})$ with a modified confidence width.

Strengths

- This paper addresses the concept of nash regret in the context of stochastic linear bandits. - It presents a novel algorithm based on elimination phases and provides a detailed regret analysis.

Weaknesses

- The paper does not discuss the computational aspects of finding the G-optimal distribution and sampling via the John ellipsoid method. - It does not include experimental results. - In the case of infinite arms, the achieved regret bound may not be sufficiently tight, as it involves a factor of $d^{5/4}$ instead of just $d$.

Questions

- I'm wondering what is the reason for using an elimination-based algorithm rather than UCB. - wondering what is the reason to consider sub Poisson distribution rather than sub gaussian. - In line 504 (proof of Theorem 1 in appendix), when $\langle x^*,\theta^* \rangle\le \sqrt{d\nu/T}$, what is the reason that the NR bound, which contains $\langle x^\star,\theta^\star \rangle$ in the bound, trivially holds?

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

- This problem deals with a fixed arms set rather than a contextual linear setting. - The regret bound of $d^{5/4}$ in the case of infinite arms may not be tight, and it would be better to elaborate on why the factor of $5/4$ appears. ( $\sqrt{d\log|\mathcal{X}|}$ regret bound for the finite case seems to imply $d$ when we consider $|\mathcal{X}|=2^d$ for infinite case.) - There are no experimental results and discussion on the computational cost, which raises concerns about practical aspects Minor comments: - line 441 in Appendix: $d$-> $\sqrt{d}$ - line 449 in Appendix: $\langle x,\hat{\theta}\rangle \le 20$ --> $\langle x,\hat{\theta}\rangle \ge 20$ - Lemma 7, equation (10) in Appendix: $P()\le 4T^{-1}\log T$ -->$P()\ge 1-4T^{-1}\log T$ ?

Reviewer zbVi6/10 · confidence 3/52023-07-04

Summary

The authors consider stochastic linear bandits, and the goal is to minimize Nash regret, which is a stronger notion than the standard regret, by replacing the arithmetic mean reward of the learner over time with the geometric one in the definition.

Strengths

The paper is well written. I didn't check the proof but I didn't spot any critical errors in the main text, either. The setting, the algorithms, and the proof sketch all make sense to me. I agree that indeed in some scenarios we may need to consider fairness across rounds and the definition of Nash regret in this paper is probably a good notion to consider in this case.

Weaknesses

This setting is rarely known. It doesn't mean this problem is unimportant, but I think it would be better to provide more context for it to better evaluate the contributions. In the related work, the authors only mention that this setting differs from prior works because they all consider fairness across arms while this paper focuses on fairness across rounds. However, I would expect more discussions about this like whether previous techniques can be applied here. Also, there is no discussion about the Nash regret of multi-armed bandits, which is an important special case. Although this setup may not be meaningful in practice, it may provide insights into handling the linear case when designing algorithms or proving regret bounds.

Questions

- The current algorithm requires $T$ as input. Do the authors think this is necessary? Part I in Algorithm 2 is somewhat non-standard for a successive elimination algorithm, so I guess it might be difficult to apply a doubling trick to remove the prior knowledge of $T$. - Does the sub-Poisson assumption include Gaussian as a special case? If not, can the analysis go through under sub-Gaussian?

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

N/A

Reviewer Cioh7/10 · confidence 3/52023-07-04

Summary

The paper extends algorithmic techniques for the linear bandits problem to the Nash Social Welfare objective function which ensures a level of fairness rather than merely the utilitarian welfare. Their work provides essentially optimal regret guarantees as compared to the traditional linear bandits problem (as it is well known that maximizing the geometric mean is necessarily stronger than the arithmetic mean). These results work for both a finite and infinite set of arms with slightly different regret guarantees.

Strengths

The paper takes a novel approach to the linear bandits problem by adapting the objective function to one that ensures some level of fairness across agents in the given system. Moreover, all results appear sound and presented in a nice and compact form within the main text -- the sketches and intuition are clearly depicted with full proofs (not all of which were verified by this reviewer) in the appendix. In addition to near optimal regret guarantees on this modified problem instance, the authors provide various tail bounds that are of independent research to the community.

Weaknesses

The paper lacks a comprehensive discussion on the related work surrounding this problem. Though the problem at hand is novel, it does hurt the submission to not include such context for readers. Specifically, there is considerable work on online NSW maximization which is naturally a similar problem. Moreover, discussion of the linear bandits problem and the best known results there would be very useful (in addition to the brief mention of Nash regret being a stronger guarantee). Some discussion on these results for finite and infinite arms would also potentially alleviate the issue of the gap between upper and lower bounds in the latter setting. As noted by the paper, improving the guarantee for the infinite arm problem setting would naturally strengthen the submission considerably.

Questions

Can you provide a brief sketch of the final theorem (infinite arms) regret guarantee? I understand that the submissions page limit prevents this but it would be helpful to have one in the extended version of the paper. Have you considered the regret guarantees of the contextual bandits problem with a NSW objective? Have you considered the multiplicative regret on this multiplicative objective function as well?

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

4 excellent

Contribution

3 good

Limitations

n/a

Reviewer XR4M6/10 · confidence 3/52023-07-11

Summary

This paper considers the problem minimize 'Nash regret' in linear bandit, where Nash regret here is defined as the difference between the optimum and the geometric mean of expected rewards induced by the bandit algorithm. The stochastic reward in this paper is assumed sub-Poisson. The authors give algorithm to achieve Nash regret of $O(\sqrt{d/T}log(T/|\mathcal{X}|)$ when the set of arms $\mathcal{X}\in \mathbb{R}^d$ is finite and $O(\frac{d^{5/4}}{\sqrt{T}}logT)$ when the set of arms $\mathcal{X}$ is infinite.

Strengths

- This paper exhibits a good technical foundation. Specifically, in the finite-arm case, the upper bound is tightly bounded, even considering logarithmic terms. - The notion of regret explored in this paper is novel and is stronger compared to external regret.

Weaknesses

The paper lacks experimental results. It would be benefical to compare other bandit algorithm with the algorithm in the paper specifically in terms of this kind of Nash regret.

Questions

- The term "Nash regret" has been previously used in [1] and [2], but with a different definition from the one presented in this paper. The current name "Nash regret" in this paper does not capture the characteristic of the geometric mean. - Regarding the definition of Nash regret in this paper, the expectation is taken within the geometric mean instead of first taking the geometric mean and then the expectation. Is there any specific reasons or motivations behind the choice of this order? - In equation (3), the solution obtained provides the optimal sequence of arm pulls. What is the meaning of 'optimal' here? Further explanation and intuition behind equation (3) can be provided by the authors to better understand the optimality criteria and its implications. - In line 439 in the appendix, there's a wrong ref (Algorithm ??) [1] Cui, Qiwen, et al. "Learning in congestion games with bandit feedback." Advances in Neural Information Processing Systems 35 (2022): 11009-11022. [2] Tay, Sebastian Shenghong, et al. "No-regret Sample-efficient Bayesian Optimization for Finding Nash Equilibria with Unknown Utilities." International Conference on Artificial Intelligence and Statistics. PMLR, 2023.

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

2 fair

Contribution

2 fair

Limitations

The authors have partially addressed the limitations of their work, though there is space for improvement (see the section Weaknesses and Questions).

Reviewer 4BdU6/10 · confidence 1/52023-07-27

Summary

This paper proposes an approach based on the Nash regret to extend the stochastic linear bandits framework with fairness and welfare considerations. The Nash regret relies on the geometric mean of expected rewards induced by the linear bandit algorithm, which tends to incentivize uniformly high rewards as opposed to the average regret which promotes cumulatively high rewards. In this setting, the authors derived tight upper bounds on this strengthened notion of regret, thereby providing fairness guarantees as the geometric mean of rewards corresponds to the Nash social welfare (NSW) function. Building upon a novel sampling arms strategy based on John ellipsoid and optimal design, as well as tailored concentration bounds, the authors propose an algorithm that achieves $O\left(\sqrt{\frac{d\nu}{T}}log(T|\chi|)\right)$, for a finite set of arms $\chi$, and $O\left(\frac{d^{5/4}\nu^{1/2}}{\sqrt{T}} log(T)\right)$, in the infinite case.

Strengths

The paper is technically solid and well-motivated. I really like the idea of inherently imposing fairness to the bandit algorithm via the Nash regret. To the best of my knowledge, the authors' result is new and provides valuable insight into incorporating fairness and welfare considerations into bandit algorithms. In addition, the authors propose tail bounds that may be of broader interest.

Weaknesses

Although the paper is technically well-set, the contribution seems incremental. The paper is also suffering from a lack of comparisons with existing approaches. The related work section seems a bit short. Having a richer picture of the literature would be beneficial for understanding the context of the paper’s contribution. Lastly, I feel that the paper could benefit from numerical experiments that support the derived theoretical results.

Questions

What is the computational efficiency of the LinNash algorithm?

Rating

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

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

The limitations of the proposed framework are not entirely clear.

Reviewer K5vq2023-08-11

Thank you for your response

I appreciate your reply. I would like to preserve my score due to its specific focus on sub-poison aspects and the omission of a discussion regarding the dependency on $d$ for infinite arms.

Authorsrebuttal2023-08-11

Apologies for accidentally missing the question. We provide detailed answer below:

Our sincere apologies for having accidentally missed the question on the dependency of $d$ for infinite arms. We hope the following answers change the reviewer's mind regarding the score. ***The regret bound of $d^{5/4}$ in the case of infinite arms may not be tight*** Please note that the factor $\log \left|\mathcal{X}\right|$ is not within the square root in Theorem 1. For $\left|\mathcal{X}\right|=2^d$, our first regret bound for finite arms (Theorem 1) of the form $\sqrt{d}\log \left|\mathcal{X}\right|$ results in a Nash regret that has a scaling factor of $d^{3/2}$. Thus Theorem 2 for infinite arms improves upon the dependence of $d$ in Theorem 1 guarantee when $\left|\mathcal{X}\right|=\omega(2^{d^{3/4}})$. The dependence of $d^{5/4}$ in Theorem 2 is mainly an artifact of our proof techniques that arises from using multiplicative versions of Hoeffding bounds that are crucial to the regret guarantees. We conjecture that this dependence can be removed (as mentioned in Conclusion section) - but this will require a completely novel set of proof techniques. Finally, we emphasize that as stipulated by NeurIPS, we have already acknowledged this limitation in the Conclusion section and Limitations paragraph in Appendix. We will add a more detailed discussion on this in the final manuscript with more available space. ***specific focus on sub-poison aspects*** We urge the reviewer to read the global response. We have provided a detailed equivalence connection between sub-Poisson and sub-Gaussian random variables. In particular, we can prove the following statements (note that positiveness of reward random variables are crucial to the definition of Nash Regret): 1) Bounded, positive random variables are both sub-Gaussian and sub-Poisson (also discussed in Lemma 1). 2) The half-normal random variable (https://en.wikipedia.org/wiki/Half-normal_distribution), with a variance of $\sigma$, is also a $C\sigma$ sub-Poisson random variable, where $C$ is a constant independent of distribution parameters. 3) For many positive sub-Gaussian random variables, including truncated and folded Normal distributions, the sub-Poisson parameter is small.

Authorsrebuttal2023-08-16

Can we expand on the explanations?

Our apologies again for having missed a question in the original response. In the above follow-up comment, we have provided a detailed response on both 1) dependence of $d$ in the case of infinite arms 2) connections between sub-Poisson and sub-Gaussian random variables. Kindly, please let us know if we can expand on the explanations or if there is something that we can clarify further in our response.

Reviewer K5vq2023-08-17

Thank you

Thank you for your response. According to your global rebuttal, sub-gaussian X is $\sigma^2/\mu$ sub poisson. However, $\mu$ is latent, so I'm wondering how the proposed algorithm can handle the sub-gaussian distribution.

Authorsrebuttal2023-08-17

Response

We thank the reviewer for the nice follow-up question. We provide our answer below: In the well-studied case of sub-Gaussians, one needs an upper bound on the sub-Gaussian norm. Along the same lines, our algorithms require only an upper bound on the sub-Poisson norm. (The sufficiency of an upper bound follows from the direct observation that, for any $\nu < \nu’$, a $\nu$ sub-Poisson random variable is also $\nu’$ sub-Poisson.) **Therefore, an upper bound on the stated ratio suffices.** We reiterate that, with relevant upper bounds in hand, our results can be used **as is** for the following well-known variants of the Gaussian random variable: half-normal random variables, positively truncated Gaussians, and rectified Gaussians. Note that under these variants, and as required, the resulting random variables (rewards) are positive.

Reviewer Cioh2023-08-16

Thanks for the response, I have no further questions at this time and will keep my 7 score.

Authorsrebuttal2023-08-18

Thank you for the question

We thank the reviewer for the question. A small $\mu$ doesn't necessarily imply a large sub-Poisson parameter. It is relevant to note that, a positive sub-Gaussian random variable with small mean will tend to have small variance (and, hence, a small sub-Poisson parameter). Though, there can be exceptions. For **all** positive sub-Gaussian random variables, and with an additional factor of $\log^{¼} T$, a Nash regret bound can be obtained as follows: Consider a positive random variable that we observe (on pulling an arm at a certain round) and suppose it has a sub-Gaussian norm (variance proxy) of $\sigma^2$. In that case, by standard concentration bound, with probability at least $1-o(T^{-2})$, the random variable will be bounded from above by $4\sigma\sqrt{\log T}$. Conditioned on this high probability event, the random variable is a $4\sigma\sqrt{\log T}$ sub-Poisson random variable (Lemma 1 in paper for bounded random variables). Then, a union bound leads to boundedness for all observed rewards with high probability. Our entire analysis goes through, and this idea only introduces a additional multiplicative $\log^{¼} T$ factor in the regret bound. Given the interest in sub-Gaussians, we will include this observation in the updated version. However, we reiterate that for natural classes of positive random variables have the same sub-Gaussian and sub-Poisson parameters modulo constants. For these random variables, the additional multiplicative log factor does not even come up.

Reviewer K5vq2023-08-19

Thank you for your response

Thank you for your response. Small mean in subguassian seems to have large variance according to your rebuttal since $\sigma^2/\mu$ sub poisson. Also, I have a concern about whether considering the event of bounded random variables influences the mean of the positive sub-guassian, which makes it difficult to analyze regret.

Authorsrebuttal2023-08-19

Thank you for the follow-up

We understand the reviewer’s interest in sub-Gaussian random variables and provide detailed answers to the raised questions below. However, we also feel that this discussion might be taking away the focus from our contributions. The paper is aimed at establishing regret for sub-Poisson rewards. Furthermore, in and of itself, sub-Poisson is an encompassing and important class. Of course, as future work, one can consider extensions to other distribution classes, such as sub-Gaussian, sub-exponential, and heavy tailed. Here are the responses to the raised questions: 1) Small mean: We reiterate that the sub-Poisson parameter is a ratio of the variance and the mean ($\sigma^2/\mu$) of a positive sub-Gaussian random variable. Hence, the reviewer’s instinct that a small mean will invariably drive up the sub-Poisson parameter is not universally correct. The ratio can continue to be small even with a small denominator. This follows from the intuition that (specifically for **positive sub-Gaussian** random variables) a small mean also suggests a small variance. This intuition is clearly validated in the case of well-studied modifications of standard Gaussians (truncated, restricted, folded Gaussians), which we have pointed out earlier. 2) Mean Shift: Yes, there will be a mean shift after conditioning. Though, the entire proof goes through for identical sub-Gaussian noise. This follows from the fact that the shift, say $\delta$, is identical across all arm pulls. That is, if an arm $x$ initially has mean $<x, \theta^*>$, then it will now have mean $<x, \theta^*> + \delta$. This updated setting can be viewed as yet another linear bandit instance, though now in dimension $d+1$: each arm $x$ is appended with a component equal to one, and $\theta^*$ is appended with a component equal to $\delta$. Everything else follows as in the paper.

Reviewer K5vq2023-08-20

Thank you for your response.

I appreciate your explanation. I understand that the sub-Poisson assumption is versatile enough to encompass various scenarios, including specific sub-Gaussian distributions as you pointed out. (I'm personally interested in understanding how to handle the value of δ, given that an assumption is needed for bounding the norm of θ*.) In addition to the noise distribution assumption, I also have a concern regarding the inherent constraint of the linear model, where there is a predetermined set of arms with 𝒳, as opposed to a contextual linear model. Now that my part of initial concerns have been addressed, I've adjusted my rating to 6.

Authorsrebuttal2023-08-21

Thank you.

We thank the reviewer for updating the score. A quick note: $\delta$ will be a small ($o\left(\frac{\sigma}{T} \right)$). Hence the norm of the new $\theta^*$ will be, at most, the initial one plus a small additive term. For contextual bandits, kindly see the reply to reviewer Cioh.

Reviewer zbVi2023-08-18

I thank the authors for their response. It's my bad not noticing [4] as a highly related work. However, the algorithm looks similar to the one in [4]. It would be good to discuss the common idea better and highlight the key difference and challenges. Overall my concerns were addressed and I increased my score.

Authorsrebuttal2023-08-19

Thank you for your response

We appreciate the reviewer's feedback. We'll include a brief comparison with the algorithm in [4] in the updated version.

Area Chair 69Ca2023-08-21

Dear Reviewer 4BdU, Could you confirm that you have read the authors' rebuttal, and whether your assessment has any change based on the rebuttal? Many thanks. Area Chair

Area Chair 69Ca2023-08-21

Dear Reviewer XR4M, Could you confirm that you have read the authors' rebuttal, and whether your assessment has any change based on the rebuttal? Many thanks. Area Chair

Reviewer XR4M2023-08-21

Thank you for the response. The authors have addressed my concerns. So I will increase my score to 6.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC