First- and Second-Order Bounds for Adversarial Linear Contextual Bandits

We consider the adversarial linear contextual bandit setting, which allows for the loss functions associated with each of $K$ arms to change over time without restriction. Assuming the $d$-dimensional contexts are drawn from a fixed known distribution, the worst-case expected regret over the course of $T$ rounds is known to scale as $\tilde O(\sqrt{Kd T})$. Under the additional assumption that the density of the contexts is log-concave, we obtain a second-order bound of order $\tilde O(K\sqrt{d V_T})$ in terms of the cumulative second moment of the learner's losses $V_T$, and a closely related first-order bound of order $\tilde O(K\sqrt{d L_T^*})$ in terms of the cumulative loss of the best policy $L_T^*$. Since $V_T$ or $L_T^*$ may be significantly smaller than $T$, these improve over the worst-case regret whenever the environment is relatively benign. Our results are obtained using a truncated version of the continuous exponential weights algorithm over the probability simplex, which we analyse by exploiting a novel connection to the linear bandit setting without contexts.

Paper

Similar papers

Peer review

Reviewer pxmm6/10 · confidence 4/52023-06-19

Summary

This paper studies the adversarial linear contextual bandit setting, an online learning problem with $K$ arms, stochastic contexts in $\mathbb{R}^d$, and adversarial rewards (which are assumed to be a linear function of the context for each arm). Previous work proves a worst-case regret guarantee that is $\sqrt{KdT}$. They obtain regret guarantees which instead of scaling with $\sqrt{T}$ can scale with other quantities that are smaller: a certain "second-order" quantity called $V_T$, and a first-order or (small loss) bound of $L_T^\star$. This is accomplished using a new algorithm called ContextEW which is a continuous exponential weight style algorithm, inspired by previous work of [Ito et al. '20]. They show that it achieves the aforementioned regret guarantees and can be efficiently implemented.

Strengths

- The paper is generally well written and explains the problem statement and results well. - The paper provides a new algorithm for adversarial linear contextual bandits which achieves first and second order bounds, a standard object of study in bandit literature.

Weaknesses

- It is a bit unclear from me how much of the work is original, or whether it is a straightforward exercise to modify the results from [Ito et al. '20] and [Neu-Olkhovskaya '20]. Perhaps this can be made clearer in the proof sketch and technical sections to delineate what techniques/results are new. - The discussion on computational complexity could be improved (see questions below). - The key intuitions for the algorithm and analysis are unclear to me (see questions below). - The results rely on the assumption that the distribution of the contexts is log-concave (which is a limitation that comes from building upon previous work).

Questions

1. In the equation after line 131, what exactly is $\mathbb{E} \theta_{t,a}$? The $\theta$ are assumed to be adversarially selected, under some choices of actions selected by a learner, so what does it mean to take the expectation here? 2. What is the intuition for the truncation/rejection sampling step? It is mentioned that it is used to reduce the variance. What would the final bound look like if there was no truncation/rejection sampling step? 3. Typo: in (10), $P_t$ should be $\mathbb{P}_t$ to be consistent with the rest of the notation. 4. Computational complexity: - I don't really understand how one should calculate $\tilde{\Sigma}$. This requires a sampling oracle to $\mathcal{D}$ in order to use the [Lovasz-Vempala '07] results right? My understanding is that one can try to approximate (9) using draws from $\mathcal{D}$. Can the authors explain how to (approximately) calculate $\tilde{\Sigma}_{t,a}$? - A follow up to the previous question is how does the error in sampling/computing the covariances propagate to the final error bound? - One can try to pick $\epsilon = 1/\mathrm{poly}(T)$. Is computing the covariance still "computationally efficient" since the computational complexity seems to scale polynomially with the total length of the problem $T$? 5. At a high level, what is the motivation for studying second-order bounds? When would we expect these to be small in practice? 6. Questions about proof details: - typo: I believe in Lemma C.1 there should not be a subscript of $Q_t$ in line 447. - In the last inequality after line 468, where did the integration over $x$ go? - I'm a bit confused about why the quantity $m_t$ appears in the Proof of Lemma C.4. For example, in the first equality after 495, how does the $-m_t$ appear? I'm not able to follow this. Is there some relationship between $\hat{\theta}_t$ and $m_t$ that is being exploited in this proof? - I didn't follow (38), how is the unbiasedness of $\hat{\theta}_t$ used? 7. It would be nice if the relevant lemmas from [Ito et al. '20] were restated in the appendix, to make the proof self contained. For example, Lemma 6 of their paper is used in the proof.

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

None.

Reviewer zjgE5/10 · confidence 4/52023-06-27

Summary

The submission considers semi-adversarial contextual bandit setting --- where d-dimensional contexts x_t are drawn iid, where some opponent picks K loss functions that are linear in the contexts at the beginning of each round, before the actual context x_t is drawn, and where he learner picks one of the K loss functions and suffers the corresponding linear loss. This setting was introduced by Neu and Olkhovskaya (COLT 2020), who provided 'zero-order' regret bounds of order \sqrt{K d T}. The present submission refines the regret bounds into first-order (for non-ngative losses, as usual) and second-order regret bounds, involving respectively the cumulative loss of the best static policy and the expected sum of instantaneous variances of the rewards obtained by the algorithms (and generalizations thereof). This is achieved by considering a continuous version of exponential weights, à la Cover (1991), and by leveraging techniques introduced by Ito, Hirahara, Soma, and Yoshida (Neurips 2020) for such refined first-order and second-order bounds in the setting of non-contextual adversarial linear bandits. The key to the results is the reduction (18) of this semi-adversarial contextual bandit problem into a linear bandit problem, as underlined in lines 285-288.

Strengths

The submission has a clear technical focus, and combines in a neat way techniques from Neu and Olkhovskaya (COLT 2020) and Ito, Hirahara, Soma, and Yoshida (Neurips 2020). The proof sketch provided on pages 6--8 is well structured. Up to some typos and some unclear pieces of notation, I can only praise the writing.

Weaknesses

My main issue is with the algorithm, which relies on distributions that are difficult to compute exactly---as always with continuous exponential weights (see articles from Vempala and co-authors at the end of the 90s, commenting on Cover's original contributions). Allowing \epsilon--approximations reduces the computational burden (see Section 3.3) but the effect of using these \epsilon--approximations on the regret bounds is not discussed. Would this add some \epsilon T terms? Then, it would probably be wise to pick \epsilon = 1/T, not to kill with this extra term the possible fast rates of convergence yielded by first- and second-order bounds. (Also, it is not clear to me from Section 3.3 whether these \epsilon--errors propagate and get worse and worse as t increases, or if for each t, we may and should compute an \epsilon--approximation from scratch, not based on previous probability distributions.) Another issue would be that I am unsure of what the second-order bound of Theorem 3.2 brings: it is highly general, and no corollary or example for the functions m_t are provided (except, I guess, the choices m_t \equiv 0). More generally, I am unsure of how important it is to get refined bounds with a rather non-practical algorithm based on continuous exponential weights. A more minor issue is about the assumption of log-concavity. It is satisfied by many distributions (see https://en.wikipedia.org/wiki/Logarithmically_concave_function#Log-concave_distributions, and the submission could give examples) but is also not satisfied by other important ones. Typos / Remarks - Lines 96-97: rather cite Cover (1991) or to the very least the monograpu by Cesa-Bianchi and Lugosi (2006), instead of a 2018 reference - Line 123: I believe you refer to it as \pi(X_t) and never use \pi(a|X_t) - Line 131: Yes, this seems clear indeed - Line 134: I would rather have written E_{t-1} instead of E_t - Line 144 / Eq. (4); see also later occurrences like in Th.3.2: Is L^*_T the same as L_T^{\pi^*}? - Algorithm 1: Say that each m_s is a function that is measurable w.r.t. \mathcal{F}_{s-1} - Eq.(6): Somewhere earlier than in lines 291 and following, discuss the number of steps needed to achieve (6), i.e., geometric distribution with a parameter that you anyway bound in the proof---but mention it early, at latest on page 5 or in Section 3.3 - Line 185: Define somewhare what log-concavity of a distribution is, and provide examples - Eq. (18) and comments in lines 285-288: This is the key step and it's somewhat hidden, I'd suggest to put it more in front - Lemma 4.2: Notation is heavy and the result is extremely difficult to read, is there any way to make it smoother? That part and Eq. (19) should also be put more in front. - Line 291: Missing word before 1 ('Algorithm' I guess) - Appendices: Cauchy-Schwarz without a 't'

Questions

Please comment on the main two weaknesses that I listed.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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

1 poor

Limitations

The authors are clear about the limitations, and in particular, on dependency of their regret bounds on K, which does not match, in the worst-case, the known \sqrt{d K T} regret bound. The discussion on page 9 is complete and upfront about the scope of their contribution.

Reviewer QEEf5/10 · confidence 4/52023-06-30

Summary

The authors explore the adversarial linear contextual bandit problem, where the context is i.i.d. drawn from a known fixed log-concave distribution, and the linear loss functions are adaptive adversaries. They provide first adaptive bounds for this setting, including a first-order bound based on the cumulative loss of the optimal policy and a second-order bound based on the squared losses of the algorithm. These bounds demonstrate improvements over the worst-case bound in scenarios where the environment is 'nice'.

Strengths

The authors first provide the adaptive bound for adversarial linear contextual bandits, which improves the worst case bound in benign environments. The computational complexity of the algorithm is commendable, particularly in moderate environments. The results and the addressed problem hold substantial significance and are likely to capture the interest of a broad audience in Neurips. Furthermore, the authors thoroughly discuss future directions, which have the potential to inspire further research aimed at improved adaptive bounds in linear bandits setting.

Weaknesses

I have two primary concerns regarding this work and would like to request the authors to provide additional clarifications. ==Weakness 1== It appears that the results and techniques heavily rely on previous work, such as Ito, S., Hirahara, S., Soma, T., and Yoshida, Y. (2020). Tight first- and second-order regret bounds for adversarial linear bandits, which first introduce adaptive bounds for adversarial linear bandits. In comparison to the adversarial linear bandits setting, the authors in this paper consider the case with i.i.d. contexts and introduce some additional assumptions, such as known context distribution and log-concavity. These assumptions may simplify the problem and restrict the practical application of the proposed algorithms. I suggest the authors emphasize the technical contribution of their paper, as this aspect is not clearly highlighted in the "Techniques" subsection in the Introduction. ==Weakness 2== There does not seem to be any experiments of any type. Neither the first-order bound nor the second-order bound demonstrate strict improvements over the worst-case bound. In other words, they may not be tight enough. In this case, I kindly recommend that the authors provide numerical results, which would enable readers to gain a better understanding of the algorithm's behavior. While numerical evaluations may not significantly enhance the primary theoretical message of the paper, they would improve the overall completeness of the work, considering that linear bandits are closely related to practical applications.

Questions

Please see the points mentioned in the weaknesses above and provide comments on them.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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

2 fair

Limitations

I don't see any issue on societal impact.

Reviewer hPEo6/10 · confidence 4/52023-07-06

Summary

Addressing to the K-armed linear contextual adaptive adversary bandit problem with d-dimensional contexts over T rounds, using a truncated version of the continuous exponential weights algorithm over the probability simplex, the expected regret about the cumulative second moments of the losses is obtained. and then the first order bound is derived.

Strengths

generalizing the results of tight First- and Second-Order Regret Bounds for Adversarial Linear Bandits to First- and Second-Order Bounds for Adversarial Linear Contextual Bandits.

Weaknesses

1 The results of the paper heavily rely on a combination of the recently proposed techniques of Ito et al. (2020) for linear bandits with tools designed by Neu and Olkhovskaya (2020) to deal with the contextual case. 2 The conclusion only applies to the scenario of the joint log-concavity of the context distributions and the exponential weights posterior over the simplex of actions, results may not be generalizable.

Questions

1 if the distribution for contexts is time-vary or need to be estimated from data, what will the situation be like? 2 the analysis is all based on the assumption that they can access the distribution for contexts, which is not the case for practical case. How to handle this difficulty, please give further discussion.

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

2 fair

Limitations

the authors adequately addressed the limitations

Reviewer zjgE2023-08-11

I acknowledge reading the entire thread of reviews and corresponding rebuttals. I confirm my initial evaluation.

Authorsrebuttal2023-08-21

Thank you for the feedback!

Reviewer QEEf2023-08-17

Response to Rebuttal

Thank you for your response and comments. While I acknowledge the paper's interesting contributions, I still consider it borderline for acceptance. Therefore, I would prefer to maintain my current score.

Authorsrebuttal2023-08-21

Thank you for the detailed review!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC