acceptodds
Under review as a conference paper at ICLR 2027

Optimal Regret for Single Index Bandits

Abstract

We study the *single-index bandit* problem, where rewards depend on an unknown one-dimensional projection of high-dimensional contexts through an unknown reward function. This model extends linear and generalized linear bandits to a nonparametric setting, and is particularly relevant when the reward function is not known in advance. While optimal regret guarantees are known for monotone reward functions, the general non-monotone case remains poorly understood, with the best known bound being (under standard boundedness, nonzero first-order Stein signal , and Lipschitz assumptions on the reward [kang et al., 2025]). We close this gap by establishing the optimal regret for general single-index bandits. We propose a simple two-phase algorithm, namely, Zooming Single Index Bandit with Upper Confidence Bound (“ZoomSIB-UCB“), that first estimates the projection direction via a normalized Stein estimator. With an adaptive exploration and discretization, it then runs UCB over the resulting one-dimensional bins. This approach achieves a regret of , and improves significantly upon prior work without knowing or a lower bound on its magnitude. We also prove a matching minimax lower bound of , showing that the dependence on is optimal up to logarithmic factors. Our upper and lower bounds together provide a sharp characterization of the regret in single-index bandits in terms of . Empirical results further demonstrate the effectiveness and robustness of our approach.

Primary area:
reinforcement learning
Keywords:
Contextual bandits, Online Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_r1p1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies single index bandits. This is a sequential setting where a learner chooses an action xt∈Xt⊂RK×dx_{t} \in \mathcal{X}_t \subset R^{K \times d}, where Xt\mathcal{X}_t is sampled from a nice distribution pp independently from Xs\mathcal{X}_s for s≤ts \leq t. The learner then observes f(xt⊤θ∗)+ηtf(x_t^\top \theta*)+\eta_t, where ηt\eta_t is zero-mean subgaussian noise, θ∗\theta* satisfies ∥θ∗∥=1\|\theta*\| = 1, and ff is an unknown lipschitz function.

The authors provide a new algorithm with regret optimal in TT and a lower bound that indeed shows that the scaling in TT is optimal. The authors also provide experiments that show that the algorithm is also empirically better.

Critical strengths

The algorithm and corresponding upper bounds are simple and a nice way to tackle the problem setting. The lower bounds compliment these nicely.

Critical weaknesses

Frustratingly, the paper is difficult to read in too many places. I will list some of the places I had trouble with, but I would like to emphasize that since I am an emergency reviewer I did not take the time to list everything.

  • In lines 169 to 172 there seem to be a contradiction around assumptions on μ∗\mu*. In one sentence there is a lower bound on μ∗\mu* but in the next there is not a lower bound. At this point I do not understand what is assumption that is used in the paper.
  • In the derivation of Stein’s Identity p is used. However, I am not sure what p to use here: the distribution of x changes depending on the learner’s choices. However, it is not clear to me if we use just the distribution that generates Xt\mathcal{X}_t or the distribution that depends on the learners choices, for example in which round the learner stops or do we just use the conditional distribution here?
  • In line 261 there is the “with high probability ∥xt∥∞≤L\|x_t\|_\infty \leq L” appears to be unjustified.
  • Assumptions 2 and 3 for no clear reason are not part of the setup and setting section (section 2) and are instead introduced only in section 4. This makes several parts of the paper hard to read.
  • In lines 261-269 new notation is used but not explained. This is only explained in assumptions 2 and 3.
  • And perhaps most frustratingly, big-O notation is used inconsistently. The comparison with prior literature is only in big-O notation. However, the paper ignores several crucial problem parameters when making this comparison in the introduction which do appear in for example Theorem 4.2. This makes it impossible for me to fairly compare to prior literature. I am also curious how the log factors appear in all the bounds.

The assumption on the norm of θ∗\theta*, i.e. ∥θ∗∥=1\|\theta*\| = 1 seems is crucial in the derivation of the results and the design of the algorithm. However, the explanation of why this is a reasonable assumption is non-existent. The sentence “to ensure scaling we adopt the standard normalization ∥θ∗∥=1\|\theta*\| = 1 (see Kang et al. (2025))” is the justification in section 2. However, this does not tell me why this is reasonable. So I went and looked at Kang et al. (2025) and the only explanation I found there was this sentence: “furthermore, we let ∥θ∗∥=1\|\theta*\| = 1 for model identifiability”. This again does not explain to me why this is a reasonable assumption and consequently I do not believe this is a reasonable assumption.

The introduction motivates the setting by saying that assuming that f is known is a strong assumption and sometimes leads to problems due to misspecification. It is important to know here that for the known ff setting there are no assumptions on Xt\mathcal{X}_t other than boundedness. Then the authors proceed by imposing several assumptions on how Xt\mathcal{X}_t is generated. First, it is assumed that Xt\mathcal{X}_t is sampled from some distribution. Second, this distribution needs to be nice, which seems to exclude fixed Xt\mathcal{X}_t. Thirds, the distribution of Xt\mathcal{X}_t needs to be known to the learner. In light of how the paper motivates the SIM bandit setting, I find it sort of problematic to impose all these additional assumptions on the problem to trade away one other assumption. Who is to say that misspecification problems don’t show up in specifying the distribution of Xt\mathcal{X}_t. Furthermore, not allowing for fixed Xt\mathcal{X}_t seems like a significant limitation.

To summarize, I think that the algorithm, lower bound, and upper bound are technically nice, but the writing can be considerably improved, the assumptions are highly limiting and in practical settings probably highly limiting, and the setting with all these assumptions is poorly motivated.

Questions for authors

I suppose I would like to hear a better motivation for the model and the assumptions that you make.

2 comments
AuthorsOpenReview ↗

Presentation, Identifiability, and Sampling Issues (Part 1)

We thank the reviewer for the detailed comments.

Weakness. Assumptions 2 and 3 appear after the algorithm, although their notation is used earlier.

We agree that the previous ordering disrupted the reading flow. We have moved Assumptions 2 and 3 to Section 2 (lines 173–182). All critical parameters, including the score-moment condition, context-tail assumption, reward bounds, and Lipschitz bounds, are now fully defined before they are used by the estimator and algorithm.

Weakness. The high-probability context bound in original line 261 appears unjustified.

The bound follows from the coordinatewise sub-Gaussian assumption and is proved in Lemma C.1. Taking a union bound over the dd coordinates of all KK candidates in all TT rounds gives the stated bound with L=4κlog⁡(8dTK/δ)L=4\kappa\sqrt{\log(8dTK/\delta)}. We have added this explanation in lines 289–291, marked in blue, including the probability 1−δ/81-\delta/8.

Question. Why is ∥∣θ⋆∥∣1=1\||\theta^\star\||_{1}=1 reasonable? What does “for identifiability” mean here?

In our setting, the expected reward is f(x⊤θ⋆)f(x^\top\theta^\star). The learner must learn both the unknown weights θ⋆\theta^\star and the unknown link ff. Because both are unknown, their scales cannot be separated using only the observed rewards.

Consider the example where we multiply the true weights by a constant c>0c>0 to obtain β=cθ⋆\beta=c\theta^\star, and to compensate for this, we define a new link g(u)=f(u/c)g(u)=f(u/c). If we calculate the expected reward using this new pair. g(x⊤β)=f(x⊤(cθ⋆)c)=f(x⊤θ⋆).g(x^\top\beta)=f\left(\frac{x^\top(c\theta^\star)}{c}\right)=f(x^\top\theta^\star). Both pairs (f,θ⋆)(f,\theta^\star) and (g,β)(g,\beta) produce exactly the same expected reward for every context xx. An algorithm therefore cannot uniquely identify their scales from rewards alone. We fix this ambiguity by requiring the absolute values of the weights to sum to one, namely ∥∣θ⋆∥∣1=1\||\theta^\star\||_{1}=1. The link's regularity bounds apply in this normalized representation. We use the ℓ1\ell_{1} norm because it also gives ∣x⊤θ⋆∣≤∥∣x∥∣∞|x^\top\theta^\star|\le\||x\||_\infty, matching our coordinatewise context bound.

We have added this rescaling explanation in lines 188–194, marked in blue.

Weakness. Original lines 169–172 assume μ⋆>0\mu^\star>0 but also say there is no lower bound on μ⋆\mu^\star.

The core assumption is simply that the signal is not exactly zero, μ⋆≠0\mu^\star\ne0. Positivity is a sign convention, not an additional restriction. The statement that μ⋆>0\mu^\star >0 is just a mathematical convention, not a restriction on the model. For example, if we multiply the weights by −1-1, we can compensate by reflecting the link, f~(u)=f(−u)\widetilde f(u)=f(-u). The expected reward remains unchanged: f~(x⊤(−θ⋆))=f(x⊤θ⋆).\widetilde f(x^\top(-\theta^\star))=f(x^\top\theta^\star).

Reflecting the link also reverses its average slope, changing the signal from μ⋆\mu^\star to −μ⋆-\mu^\star. Thus, a model with negative signal has an equivalent representation with positive signal. This is why we take μ⋆>0\mu^\star>0 without loss of generality.

There is no contradiction with allowing arbitrarily small signals. Positivity chooses an orientation; it does not impose a uniform positive lower bound on the magnitude. The algorithm receives neither the signal nor such a lower bound. The revised statement in lines 195–199 makes this distinction explicit.

Question. Which density is used in Stein's identity when the learner chooses the arm and the stopping time?

Here pp is the density from which each candidate feature vector is drawn. The reviewer is right that choosing an arm based on its features can change the selected arm's distribution.

During Phase 1, however, we choose uniformly among the candidates, independently of their feature values. Conditional on the past, the selected feature vector therefore still has density pp. The decision to continue exploring is also made before observing the next round's candidates.

The stopping time is data-dependent. Lemma B.1 gives an estimation bound simultaneously over all exploration sample sizes, allowing us to use it at the selected stopping time. Phase 2 observations update the bin rewards. We have clarified the density-preserving sampling step in lines 219–220, marked in blue.

AuthorsOpenReview ↗

Regret comparison and Model assumptions (Part 2)

Weakness. The comparison suppresses problem parameters, and the logarithmic dependence is unclear.

Our optimality claim concerns the horizon exponent in the stated regime. Kang et al.'s Theorem 3.9 gives expected regret O(d3/8T3/4)O(d^{3/8}T^{3/4}) under Gaussian design and d15=O(T2)d^{15}=O(T^2). Our Corollary 4.3 gives O~((1+(d/∣μ⋆∣)2/3)T2/3)\widetilde O((1+(d/|\mu^\star|)^{2/3})T^{2/3}). Thus, for fixed dimension and nonzero signal, the horizon exponent improves from 3/43/4 to 2/32/3. Our displayed bound has a larger exponent of dd.

The logarithmic factors are stated explicitly alongside the finite-horizon bound in the restatement of Theorem 4.1 in Appendix C. Since our goal is to settle the horizon exponent up to logarithmic factors, we keep the main comparison at that level. Lines 369–375 (in blue) clarify signal dependence and point to the explicit bound.

Question. Why motivate the unknown link while requiring random contexts and a known context distribution? What about fixed contexts and distribution misspecification?

Single-index bandits extend the classical single-index model (SIM) into sequential decision-making. In SIM, both the parameter θ⋆\theta^\star and the link ff are unknown. Specifying ff in advance gives the corresponding generalized linear model (GLM), in which the only unknown component is the parameter. Learning an unknown link introduces an additional nonparametric learning task. Estimators requiring a specified link cannot be applied directly, and their guarantees under general bounded contexts do not automatically extend to our setting.

Existing regret guarantees for single-index bandits also impose assumptions on context generation. Kang et al. (2025), Arya and Song (2025), and Ma and Cai (2025) study stochastic contexts with distributional regularity conditions in their respective formulations. Thus, studying the unknown-link problem under context assumptions follows an established approach. We have added this connection in lines 78–80 of the introduction.

Our work is most closely related to Kang et al. (2025). Their general-link algorithm GSTOR has a T3/4T^{3/4} upper bound, and they explicitly ask whether a O~(T)\widetilde O(\sqrt T) bound might be possible. We study the same stochastic-context formulation and settle its horizon-rate question. Moreover, their general-link theorem assumes Gaussian contexts, whereas ours permits the regular non-Gaussian distributions in Assumptions 2 and 3. We obtain O~(T2/3)\widetilde O(T^{2/3}) regret for fixed dimension and nonzero signal. Theorem 5.1 shows that this exponent is unavoidable even with signal bounded away from zero, when K≍T1/3log⁡TK\asymp T^{1/3}\log T.

  • Fixed and adversarial contexts. Known-link generalized linear bandit methods can handle contexts beyond our stochastic setting. Filippi et al. (2010) study fixed arm-feature vectors, and Sawarni et al. (2024) (Generalized Linear Bandits with Limited Adaptivity), allow adversarially chosen bounded feature vectors. These guarantees assume a known link. Our unknown-link setting includes an additional nonparametric learning task; our Stein estimator requires the stated stochastic-context assumptions and does not cover arbitrary deterministic or adversarial context sequences.

  • Practical relevance and unknown-distribution experiments. Requiring context assumptions does not make the formulation inherently impractical. Potential applications include recommendation, personalized treatment selection, and hyperparameter tuning, as discussed in our response to Reviewer G4Wh. We have added the news recommendation example in lines 73–77. Knowing the score is also natural when candidate feature vectors are generated from a chosen distribution. For observational contexts, however, the known-score requirement is a substantive limitation. An incorrect score can bias the direction estimate. Section 2.1, line 232, states that the density must be known. Remark E.1, lines 1274–1278, explains that estimating the score introduces an additional approximation error. Appendix F.7 examines this practical issue using real context vectors from KDD99 and CoverType with synthetic rewards. Lines 1653–1656 explicitly identify this as a practical approximation and a robustness study beyond the exact known-score setting. These experiments provide empirical evidence that the method can be useful on real context data even when the context score is approximated, supporting the practical relevance of the formulation.

We hope these explanations resolve the concerns about the assumptions and their role in the analysis. We would appreciate reconsideration of the score in light of these clarifications.

Reviewer_G4WhOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper tackles the problem of the single-index bandit in a general non-monotone case. Previously, Kang et al. (2025) presented an explore-then-commit type of algorithm with T3/4T^{3/4} minimax regret. This paper proposes ZoomSIB-UCB, a confidence-bound-based algorithm that achieves an optimal T2/3T^{2/3} regret bound along with a matching lower bound.

The major components of the proposed algorithm are the Stein-score estimator, which provides a high-probability confidence bound for unknown parameters despite the non-monotone nature. In the next phase, they construct bins based on estimated rewards. Considering each bin as an arm, the rest of the algorithm involves UCB treatment.

The authors provide empirical evidence for the performance of their algorithm.

Critical strengths

It is a well-written paper. I did not check the proof line by line, but they all seem okay.

I am not very aware of this field; however, I was surprised to see that no UCB-based algorithm has been developed previously. This field seems pretty unexplored. If this is the first UCB-based algorithm, this is a commendable effort.

Critical weaknesses

The major weakness, as I see it, is that the proposed algorithm also acts like an explore-then-commit strategy. The estimated model is frozen after initial exploration. For phase 1, although they use bins and confidence-based exploration, what is the point of this exploration if θ^\hat{\theta} is not updated? In my understanding, in a usual UCB-type algorithm, the confidence width narrows based on a theoretical gurantee that the true parameter lies in this bound around the estimated parameter. Here, I do not understand how the estimated parameter is frozen and still the confidence width changes in phase 2?

The proposed algorithm is compared only with one baseline. GSTOR.

Since this seems to be a very under-explored field, it would have been nice to motivate it with a practical example.

Questions for authors

Please provide insights regarding the first weakness.

What are some real-world applications of non-monotone single-index bandits?

I might reconsider my score based on the answers by the authors.

1 comment

We thank the reviewer for the positive assessment of the presentation and the opportunity to clarify the algorithm.

Weakness. What does Phase 2 learn if the direction is frozen, and why does its confidence width shrink?

There are two quantities to learn: the direction θ⋆\theta^\star and the reward function ff. Even with the exact direction, the learner must determine, for example, whether an arm with projection 0.20.2 yields a higher reward than one with projection 0.80.8. Since ff is unknown and can be non-monotone, either could be better.

Phase 2 learns these rewards. Arms with nearby projected contexts are grouped into bins. Each selected reward updates its bin's empirical mean and pull count. The statistical confidence width shrinks as 1/nj1/\sqrt{n_{j}}, where njn_{j} counts observations from bin jj. The frozen direction introduces a separate approximation error. Lemma C.5 bounds the difference between an arm's mean reward and its bin-center value by Lf′wL_{f'}w.

Section 3 describes direction updates in lines 283–287 and bin-reward updates in lines 303–312. Steps 13–15 of Algorithm 1 give the latter explicitly. We have clarified in lines 309–312 (in blue) that reward estimates and confidence bounds continue to update, although the direction and bins remain fixed.

GSTOR estimates the direction and link in two exploration blocks, then acts greedily. Our algorithm fixes the direction but continues learning bin rewards and selecting arms optimistically throughout Phase 2. With O~(T1/3)\widetilde O(T^{1/3}) bins, this online learning has statistical cost O~(T2/3)\widetilde O(T^{2/3}).

Weakness. The experiments compare against only GSTOR.

GSTOR is our main comparison because it addresses the same shared, unknown, potentially non-monotone link setting. We have now added the ESTOR of Kang et al. (2025), which provides a O~(T)\widetilde O(\sqrt{T}) guarantee for monotone link functions. We test two increasing links, f(z)=(1+tanh⁡z)/2f(z)=(1+\tanh z)/2 and f(z)=1/2+arctan⁡(z)/πf(z)=1/2+\arctan(z)/\pi, alongside the original three non-monotone links discussed in the paper. We use d=10d=10, K=20K=20; independent standard Gaussian contexts; θ⋆=e1\theta^\star=e_{1}; Gaussian noise with standard deviation 0.50.5, and ten independent runs. We test five horizons from 12,50012{,}500 to 200,000200{,}000.

The mean cumulative pseudo-regret at T=200,000T=200{,}000 is shown below.

LinkESTORZoomSIB-UCBGSTORRandom
Monotone tanh2,418.45,176.122,154.192,870.0
Monotone arctan1,555.35,626.916,581.067,070.4
cos⁡z+sin⁡z\cos z+\sin z276,373.210,401.236,775.4159,769.1
ze−z2ze^{-z^2}104,408.410,446.120,311.683,876.6
sin⁡z+0.3z\sin z+0.3z22,928.518,016.965,567.3288,067.4

ESTOR performs best on both increasing links, showing the benefit of exploiting monotonicity, whereas ZoomSIB-UCB has the lowest mean regret on all three non-monotone links.

ESTOR performs comparatively well on sin⁡z+0.3z\sin z+0.3z, although ZoomSIB-UCB has lower mean regret. These non-monotone tests are outside ESTOR's monotonicity assumption and do not contradict its guarantee.

We have added this comparison in lines 469–475 and Appendix F.1, lines 1437–1457, with the added text in blue. (New) Table 2 reports final regret; (New) Figures 2 and 3 show the regret curve for all five link functions.

Question. What are some real-world applications of non-monotone single-index bandits?

Single-index models are well established in statistics, but their use in contextual bandits is less developed. Potential applications include online recommendation, personalized treatment selection, and hyperparameter tuning. A single-index model represents rewards through an unknown weighted combination of features and an unknown response function.

Online news recommendation provides a clear example. The algorithm recommends an article and observes user clicks. Features are combined into a single number (the index), which then an unknown link function converts into a click probability. Assuming the monotone link dictates that higher indices always yield higher click probabilities. Allowing a non-monotone link removes this restriction and lets the algorithm discover the true relationship directly from user feedback. We have added this example in lines 73–77 of the introduction.

We appreciate the reviewer's willingness to reconsider the score and hope these clarifications resolve the main concerns.

Reviewer_u123OpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an algorithm for the single-index bandit problem that combines Stein-based direction estimation with a data-dependent choice of exploration length and one-dimensional discretization resolution, followed by online learning over the resulting bins using sleeping UCB. This design improves the regret bound from the previous O~(T3/4)\widetilde{O}(T^{3/4}) to O~(T2/3)\widetilde{O}(T^{2/3}). The paper also establishes a matching lower bound in TT, demonstrating the optimality of the time-horizon dependence in the setting considered. Experiments further demonstrate the algorithm's advantages across different reward functions ff.

Critical strengths

The theoretical contribution is substantial and relatively comprehensive: the paper improves the regret rate and provides a corresponding lower bound establishing optimality. The presentation is clear, and the proofs are rigorous and coherent. The experimental evaluation also examines performance under different reward functions ff.

Critical weaknesses

My main question concerns the dependence on μ⋆\mu^\star in Theorem 1. The paper does not appear to discuss in detail either the case μ⋆=0\mu^\star = 0 or the behavior as μ⋆→0\mu^\star \to 0. What does a vanishing μ⋆\mu^\star mean for the problem, and how does the algorithm perform in this regime? Beyond the statement in Lemma 2.3 that the algorithm “learns slowly,” how slow does learning become? Are the algorithm's behavior and the resulting regret bound still consistent with intuition? In particular, can the regret retain the O~(T2/3)\widetilde{O}(T^{2/3}) rate as μ⋆→0\mu^\star \to 0? The authors could also discuss these questions in relation to the lower bound. I believe these points should be clarified in the paper.

Questions for authors

Please check Critical Weaknesses.

3 comments

We thank the reviewer for the positive assessment and for raising the question about weak Stein signals.

Question. What does a small or zero μ⋆\mu^\star mean?

Thanks for pointing this out. The signal μ⋆=E[f′(X⊤θ⋆)]\mu^\star=\mathbb E[f'(X^\top\theta^\star)] is the average signed slope. Stein's identity gives E[YS(X)]=μ⋆θ⋆\mathbb E[YS(X)]=\mu^\star\theta^\star. When μ⋆=0\mu^\star=0, this moment is the zero vector and provides no information about the direction. Rewards can nevertheless vary substantially. For example, with f(z)=cos⁡zf(z)=\cos z and a standard Gaussian index, the positive and negative slopes cancel, giving μ⋆=0\mu^\star=0 although ff is not constant. Kang et al. (2025) also assume a nonzero μ⋆\mu^\star in their setup.

Higher-order methods offer possible ways to address this limitation under additional conditions. Yang et al. (2017) use second-order Stein moments to estimate index directions, including cases where the first-order moment vanishes. This requires a nonzero second-order moment. Under Gaussian covariates, Lee et al. (2024) study learning polynomial single-index targets using Hermite expansions and higher-order information obtained through data reuse. These results concern statistical learning from sampled data. Adapting such methods to our bandit setting would require new estimation guarantees and regret analysis. We leave this extension as future work. Section 7 identifies the removal of the nonzero-signal condition as an open problem. We have added these possible directions in lines 514–516, marked in blue.

Question. How slowly does learning proceed as μ⋆→0\mu^\star\to0? Does the O~(T2/3)\widetilde O(T^{2/3}) rate still hold?

For a fixed problem instance, μ⋆\mu^\star is determined by ff, D\mathcal D, and θ⋆\theta^\star and remains fixed as TT grows. Hence, for every fixed nonzero signal, our bound has O~(T2/3)\widetilde O(T^{2/3}) horizon dependence. The dependence on signal strength is explicit in Theorem 4.1. Its signal-dependent term is O~((d/∣μ⋆∣)2/3T2/3)\widetilde O((d/|\mu^\star|)^{2/3}T^{2/3}). Thus, a smaller signal increases the regret bound. If the signal itself decreases with TT, that dependence must also be included. GSTOR's analysis also has inverse signal dependence (Kang et al., 2025, Theorem 3.9 and Appendix L.2).

Appendix F.4 also examines this behavior experimentally. At T=200,000T=200{,}000, Table 5 reports that the calibrated algorithm's exploration fraction rises from 4.1%4.1\% to 71.6%71.6\% as the estimated signal falls from 0.60700.6070 to 0.01760.0176. We have clarified the quantitative signal dependence in lines 369–375, marked in blue.

Question. How does this relate to the lower bound?

Theorem 5.1 includes hard instances where the signal is bounded below by a positive constant independent of TT. Even in this family, every policy incurs Ω(T2/3)\Omega(T^{2/3}) expected regret on some instance when K≍T1/3log⁡TK\asymp T^{1/3}\log T. The direction is fixed at θ⋆=1\theta^\star=1 throughout the construction, so learning the unknown link already creates this difficulty.

Combined with our upper bound, this establishes the optimal horizon exponent in that regime. Whether the factor ∣μ⋆∣−2/3|\mu^\star|^{-2/3} in our upper bound is necessary remains open. We have distinguished these two questions in Remark 5.2, lines 415–419, marked in blue.

Thanks for the detailed explanation and I do not have any further questions.

AuthorsOpenReview ↗

A Big Thank You!

We sincerely appreciate your time, thoughtful feedback, and positive assessment of our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.