acceptodds
Under review as a conference paper at ICLR 2027

Quantum Heavy-Tailed Bandits: Lower Bounds and Improved Dimension Dependence

Abstract

We study stochastic multi-armed and linear bandits with heavy-tailed rewards under quantum access. The learner chooses one action at a time and accesses a quantum procedure representing that action’s reward distribution, which it can run forwards and backwards to estimate the expected reward. Assume that the reward associated with each action satisfies , where . For arms and a budget of oracle queries, with at least a constant multiple of , we achieve regret and prove a lower bound matching up to logarithmic factors. This improves the linear dependence on in Wu et al. (2023) whenever . For -dimensional linear bandits with , we obtain the same exponent in , with dimension dependence for finite action sets and for general compact action sets. The finite-action bound depends only logarithmically on the number of actions; the compact-action result assumes access to routines for experimental design and linear optimization. These results improve the quadratic dimension dependence of Wu et al. (2023) over their established range and extend regret guarantees to the full range . The improvements come from jointly allocating estimation queries across actions and reconstructing linear rewards from a small set of representative actions. We also establish worst-case lower bounds for linear bandits and extend the analysis to bounded central moments. Numerical simulations illustrate the theoretical trade-offs.

Primary area:
learning theory
Keywords:
Quantum Computing, Heavy-tailed Rewards, Bandits

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_2XfnOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies stochastic multi-armed and linear bandits with coherent quantum access to each arm's reward distribution. Actions are selected classically, and both forward and inverse reward-oracle calls count toward the horizon and incur the selected action's regret. Under E∣Ya∣1+α≤1\mathbb E|Y_a|^{1+\alpha}\le 1, where α∈(0,1]\alpha\in(0,1], the paper proves a multi-armed regret bound of O~(K2α/(1+α)T(1−α)/(1+α)+Klog⁡T),\widetilde O\left(K^{2\alpha/(1+\alpha)}T^{(1-\alpha)/(1+\alpha)}+K\log T\right), with a lower bound matching the leading polynomial dependence. For linear bandits, it improves the leading dimension dependence to d1/2+α/(1+α)d^{1/2+\alpha/(1+\alpha)} for finite action sets and d1/2+2α/(1+α)d^{1/2+2\alpha/(1+\alpha)} for compact action sets. The compact case requires additional design and optimization oracles.

The algorithms use phased elimination with existing quantum mean estimators. The improvements come from allocating a shared query budget across active arms and, in the linear case, using experimental designs to estimate rewards more efficiently. The lower bounds reduce learning to distinguishing distributions that differ on rare outcomes, while allowing the learner to stop querying adaptively. The paper also considers central-moment assumptions. The results are supported mainly by proofs, with synthetic experiments that emulate the estimators.

Critical strengths
  1. The multi-armed upper and lower bounds match in their leading polynomial dependence on the number of arms and the horizon. This gives a sharper characterization than the previous guarantees and is the strongest reason for my positive assessment. The lower bound also allows adaptive query allocation and stopping, which is important for the stated optimality claim.
  2. The linear-bandit results make useful use of the action geometry. Allocating queries according to an experimental design and controlling the resulting estimation errors leads to a concrete improvement in dimension dependence. The paper therefore contributes more than an application of a faster scalar estimator to an existing bandit algorithm.
Critical weaknesses

I do not see a weakness that warrants rejection. My main reservation is that the linear-bandit bounds still leave gaps in dimension dependence, especially for compact action sets. The lower-bound construction based on informative rays does not fully capture the difficulty introduced by the geometry of the action set.

The technical novelty is also somewhat limited: the scalar estimators, elimination approach, and experimental-design tools are established, with the new work mainly in how queries are allocated and how the errors are analyzed. The experiments emulate the estimators and illustrate the consequences of the query bounds; they do not validate a quantum implementation. These limitations keep me from giving a stronger recommendation, but they do not undermine the main results.

Questions for authors

None essential to my recommendation. The remaining dimension gaps are a natural direction for future work, but I do not think they need to be resolved for this paper to be accepted.

1 comment

We thank the reviewer for the careful summary and positive assessment. Three changes address the stated reservations.

  1. Dimension gaps and the ray construction. The new Remark 4.2 in the main text states the gaps and their sources: d(1−α)/(2(1+α))d^{(1-\alpha)/(2(1+\alpha))} for finite sets, which vanishes at α=1 for poly(d)\mathrm{poly}(d) actions, and d\sqrt d for compact sets. For compact sets, the concurrent work of Liu et al. (2026b, arXiv:2611.15237) proves a matching Ω(d3/2)\Omega(d^{3/2}) lower bound for bounded rewards, so our compact-action bound is optimal in dd at α=1 and the gap remains only for α<1. The new Proposition C.8 shows that a hypercube instance, the type that separates linear from multi-armed bandits classically (Tajdini et al. 2025), yields only the d-armed rate under the stopped-query hybrid argument, because the hybrid bound is linear in the per-query perturbation whereas the classical KL divergence is quadratic. This explains why the ray construction does not capture the geometry and why a different construction, such as the spherical cap of Liu et al. (2026b), is needed.

  2. Technical novelty. A new "Positioning" paragraph in Section 1, together with Remark 4.4, states which components are inherited and which are new.

    • Inherited: the scalar estimators, the elimination template, the small-support design and, at α=1, the design-matched allocation of Liu et al. (2026a).
    • New: the heavy-tailed allocation exponents and (1+α)(1+\alpha)-moment aggregation, the raw-moment unbiased interface, the stopped-query lower bounds, and the central-moment extension.

    We also position our bounds against the bounded-reward results at α=1 (Table 1), which our algorithms recover as a special case.

  3. Experiments. Section 5 now states that the experiments compare certified query-complexity rates under emulated estimators and make no claim about a hardware implementation.

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

The paper studies quantum bandits with heavy tailed rewards assuming the rewards have finite moment 1+α1 + \alpha for some 1<α≤21 < \alpha \le 2. For multi-armed bandits, they prove a better regret upper bound in terms number of arms (K2α1+αK^\frac{2\alpha}{1 + \alpha}) by a sharper analysis of a prior algorithm, and they prove a matching lower bound using a rare-reward hard instance. For linear bandits, they propose a new phased elimination algorithm achieving regret with better dimension dependence than prior work (d2d^2 to d1+5α2(1+α)d^\frac{1 + 5\alpha}{2(1 + \alpha)}). They also extend these regret bounds to finite centered moment setting.

Critical strengths
  • The heavy tailed quantum multi armed bandits is almost fully characterized with matching upper and lower bounds in terms of number of arms and horizon. The analysis achieves the sharper regret by retaining the shared query budget across arm instead of bounding each arm separately.
  • For linear bandit setting, better dimension regret terms is achieved using a new elimination algorithm, and also prove an upper bound bound for finite action set setting.
Critical weaknesses
  • The lower bound given for linear bandit setting is equal to multi-armed bandit lower bound, and no gap between the two setting's hardness is shown unlike the known sub-gaussian setting.
Questions for authors
  • Algorithm proposed in [1] for linear bandits is UCB based, Can shaper analysis prove the same improved regret bounds similar to the multi armed bandit setting? Was there a technical challenge preventing this?
  • Have the authors considered the hard instance achieving the classical lower bound for linear bandits in [2] for proving a stronger lower bound for quantum linear bandits, separating the linear and MAB settings? Was there a technical challenge preventing such an extension unlike the MAB lower bound?

[1] Wu, Yulian, Chaowen Guan, Vaneet Aggarwal, and Di Wang. "Quantum heavy-tailed bandits." arXiv preprint arXiv:2301.09680 (2023).

[2] Tajdini, Artin, Jonathan Scarlett, and Kevin Jamieson. "Improved regret bounds for linear bandits with heavy-tailed rewards." arXiv preprint arXiv:2506.04775 (2025).

1 comment

We thank the reviewer for the positive assessment and the two questions. A small clarification: the moment order is 1+α1+\alpha with α∈(0,1]\alpha\in(0,1], where α=1 is the finite-variance case.

Weakness / Q2: A lower bound beyond the d-armed embedding. We carried out this extension; see Proposition C.8 (new Appendix C.4), summarized in the new Remark 4.2. Under the stopped-query hybrid argument, an instance of the same type as Theorem 1 of [2] (independent signs on the dd coordinates, rare-event Bernoulli rewards, every action informative about all coordinates) gives exactly the rate of the d-armed embedding, Ω(u1/(1+α)d2α/(1+α)T(1−α)/(1+α))\Omega(u^{1/(1+\alpha)}d^{2\alpha/(1+\alpha)}T^{(1-\alpha)/(1+\alpha)}).

The reason is a difference between classical and quantum information. In this instance, each wrong coordinate sign costs regret of order g/dg/\sqrt d per round.

  • Classically, the per-sample KL divergence about one coordinate is quadratic in the perturbation, of order g2/dg^2/d. Information is diluted by 1/d1/d but regret only by 1/d1/\sqrt d, and this mismatch produces the extra dα/(1+α)d^{\alpha/(1+\alpha)} in [2].
  • The quantum hybrid bound is linear in the oracle perturbation, of order g/dg/\sqrt d per query. Information and regret are diluted by the same factor, and the d\sqrt d cancels.

Hence instances of this type cannot separate linear from multi-armed bandits under the hybrid argument; a separation requires a different construction or technique. The concurrent work of Liu et al. (2026b, arXiv:2611.15237) confirms this: at α=1 they prove Ω(d3/2log⁡(1+T/d3/2))\Omega(d^{3/2}\log(1+T/d^{3/2})) for compact action sets, using a spherical cap on which gaps are quadratic in the distance to the optimal action and a posterior Fisher-information argument. Since bounded rewards belong to our class, this also shows that our compact-action upper bound is optimal in dd at α=1. We expect their technique to extend to heavy tails with our rare-event oracles and to reduce the compact-action gap for α<1 to the exponent of the classical gap (Appendix C.4); we leave this to future work.

Two further facts, now in Remark 4.2:

  • For finite action sets with n=poly(d)n=\mathrm{poly}(d) actions there is no polynomial separation at α=1: the linear-bandit upper bound O~(dlog⁡Tlog⁡(nT))\tilde O(d\log T\log(nT)) (Theorem C.6; Liu et al. 2026a) matches the d-armed lower bound Ω(dlog⁡(T/d))\Omega(d\log(T/d)) up to logarithmic factors.
  • Classically, the finite-action lower bound of [2] exceeds the d-armed rate by a (log⁡n/log⁡d)α/(1+α)(\log n/\log d)^{\alpha/(1+\alpha)} factor, which is polynomial in dd only when nn is exponential in dd; this is consistent with the first fact.

Q1: Sharper analysis of the UCB-type algorithm of [1]. Partly.

  • For general action sets, we expect so. The compact-action rate needs only high-probability support estimates of accuracy ϵ/d\epsilon/\sqrt d; at α=1, Liu et al. (2026a, Appendix C) obtain the same d3/2d^{3/2} from the UCB-type analysis of Hikima et al. (2024).
  • For finite action sets, the improvement to d(1+3α)/(2(1+α))d^{(1+3\alpha)/(2(1+\alpha))} relies on independent unbiased estimates on a fixed design support, aggregated through their (1+α)(1+\alpha)-moment (von Bahr–Esseen). High-probability support estimates incur a d\sqrt d loss from worst-case error propagation through the prediction weights, as Liu et al. (2026a, Section 5) show for their QMC-based variant at α=1, and a confidence-ellipsoid analysis works with high-probability radii in the same way.

In this sense the finite-action part of the improvement is algorithmic. The distinction is now stated in Remark 4.4, while Remark 3.3 states that the multi-armed improvement, by contrast, is purely analytical.

Reviewer_R9W6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies stochastic multi-armed and linear bandits with bounded (1+α)(1+\alpha)-moment rewards under quantum reward-oracle access, where each forward or inverse oracle invocation counts toward the horizon and incurs regret. For multi-armed bandits, it establishes matching upper and lower bounds, up to logarithmic factors, with leading regret O~(K2α/(1+α)T(1−α)/(1+α))\widetilde O(K^{2\alpha/(1+\alpha)}T^{(1-\alpha)/(1+\alpha)}) under a unit moment bound. For linear bandits, design-based phased elimination yields leading dimension factors d1/2+α/(1+α)d^{1/2+\alpha/(1+\alpha)} for finite action sets and d1/2+2α/(1+α)d^{1/2+2\alpha/(1+\alpha)} for compact action sets, with the latter assuming design and linear-optimization oracles. The paper also provides linear-bandit lower bounds and extensions to bounded central moments.

The main validation is theoretical. Synthetic experiments illustrate the consequences of the assumed estimation rates using emulated quantum estimates charged their theoretical query costs.

Critical strengths
  1. The paper establishes matching upper and lower regret bounds, up to logarithmic factors, for quantum heavy-tailed multi-armed bandits. Theorems 3.1 and 3.2 match the polynomial dependence on the moment scale, number of arms, and horizon. The lower-bound argument accommodates adaptive queries and data-dependent stopping while charging inverse calls consistently with the upper-bound model. Establishing this matching bounds is the paper’s strongest contribution, even though the upper-bound improvement uses an existing estimator and a sharper analysis.

  2. The paper improves the previously reported quadratic dimension dependence of quantum linear-bandit regret bounds and extends the guarantees to the full moment range α∈(0,1]\alpha\in(0,1]. The finite-action method combines design-matched query allocation, unbiased estimates, and the von Bahr–Esseen inequality to control reconstruction error. The compact-action method obtains uniform reconstruction from a small design support. These results improve the previously reported quadratic dimension dependence over its established range and extend guarantees to α<1/3\alpha<1/3. This constitutes a useful theoretical advance under the stated oracle assumptions.

Critical weaknesses
  1. The extent of the new algorithmic contribution and the remaining linear-bandit gaps need clearer positioning. Remark 3.3 explains that the multi-armed improvement comes from retaining the shared query budget, leaving open whether the same analysis also sharpens the existing algorithm of Wu et al. (2023). The finite-action result also needs a direct comparison with Liu et al. (2026), particularly on bounded-reward instances at α=1\alpha=1. Moreover, the linear lower bounds embed a dd-armed instance, with the infinite-action construction extending it along dd rays. Comparing the leading polynomial terms leaves dimension gaps of d(1−α)/(2(1+α))d^{(1-\alpha)/(2(1+\alpha))} for finite actions and d\sqrt d for compact actions. Remark C.2 acknowledges the limitation, but it should be prominent in the main discussion. These points affect how much the paper establishes beyond improved upper bounds.

  2. Some important proof and validation details remain incomplete. Equation (24) states the estimator’s error-moment guarantee using O~α\widetilde O_\alpha, whereas Lemma C.5 applies it with a constant CαC_\alpha and the allocation in equation (28) does not display the corresponding logarithmic overhead. The paper should explain whether these quantities are nominal accuracy parameters or actual expected query budgets, and propagate the overhead consistently through the timeout analysis. Appendix B.2 also states the crucial shared-budget exploration bound without a complete derivation.

Questions for authors
  1. Lemma C.3 hides logarithmic factors in the estimator’s query cost, but these factors do not appear in the query allocation in equation (28). How does the algorithm account for these additional queries, and does the stated regret bound still hold after including them?

  2. Does applying the shared-budget argument in Remark 3.3 to the batch-UCB algorithm of Wu et al. (2023) already yield the improved K2α/(1+α)K^{2\alpha/(1+\alpha)} dependence? Please distinguish the contribution of the new policy from that of the sharper analysis and matching lower bound.

  3. How does Theorem C.6 compare with Liu et al. (2026) on their common bounded-reward setting at α=1\alpha=1? Which parts of the allocation and reconstruction analysis are new, and which extend that earlier framework to heavier tails?

2 comments

We thank the reviewer for the careful reading. All changes are summarized in our general comment.

W1 / Q2: Source of the K-improvement. We agree, and the revision says so (Remark 3.3). The K2α/(1+α)K^{2\alpha/(1+\alpha)} rate comes from the analysis, not the elimination policy. The proof of Theorem 4 of Wu et al. (2023) already records the global constraint ∑i(2Mi+1−1)≤T\sum_i(2^{M_i+1}-1)\le T and the per-arm bound R(T;i)≲u1/(1+α)(2Mi)(1−α)/(1+α)R(T;i)\lesssim u^{1/(1+\alpha)}(2^{M_i})^{(1-\alpha)/(1+\alpha)}; its last step bounds each 2Mi2^{M_i} by TT separately. Applying Jensen's inequality to the sum instead gives K2α/(1+α)T(1−α)/(1+α)K^{2\alpha/(1+\alpha)}T^{(1-\alpha)/(1+\alpha)} for their algorithm. Our multi-armed contribution is therefore this analysis step together with the matching lower bound (Theorem 3.1), whose stopped-query hybrid argument covers adaptive queries, inverse calls and data-dependent stopping. We use elimination only because it is the template shared with the linear-bandit algorithms.

W1 / Q3: Comparison with Liu et al. (2026a) at α=1. Bounded rewards satisfy E∣Y∣2≤1\mathbb E|Y|^2\le1, so Theorem C.6 at α=1 applies to their setting. It gives O~(dlog⁡Tlog⁡(nT))\tilde O(d\log T\log(nT)), the same rate as their LV-G-Elim; we claim no improvement at α=1, and Table 1 now shows this.

Algorithm 2 is the heavy-tailed generalization of LV-G-Elim. It inherits from their framework the small-support design, the allocation qi∝π(ai)q_i\propto\sqrt{\pi(a_i)}, the shared timeout and the action-wise median. New for α<1 are:

  • the allocation exponents in (28), calibrated to the q−2αq^{-2\alpha} error-moment rate;
  • the aggregation step, which replaces the variance calculation by the von Bahr–Esseen inequality and L1+α≤L2L^{1+\alpha}\le L^2 monotonicity under the design weights (Lemma C.5);
  • the unbiased raw-moment interface (Lemma C.3) and the central-moment version (Appendix D). Liu et al. (2026a, Remark 5.5) leave even the bounded-variance extension open.

For compact sets, our d3/2d^{3/2} at α=1 improves the d2d^2 bound of Wan et al. (2023) by d\sqrt d and matches the d3/2d^{3/2} bound obtained from Hikima et al. (2024) (Liu et al. 2026a, Appendix C); Remark 4.4 now states both comparisons. Our improvement over the best prior bound is for α<1, where we improve the d2d^2 of Wu et al. (2023) over its range α∈[1/3,1] and extend the guarantee to α<1/3.

W1: Lower-bound gaps. The new Remark 4.2 in the main text states the gaps explicitly.

  • Finite action sets: the gap is d(1−α)/(2(1+α))d^{(1-\alpha)/(2(1+\alpha))}. It vanishes at α=1 for n=poly(d)n=\mathrm{poly}(d) and has exactly the exponent of the gap between the classical bounds of Tajdini et al. (2025).
  • Compact action sets: the gap is d\sqrt d. At α=1 it is closed by the concurrent work of Liu et al. (2026b, arXiv:2611.15237), who prove Ω(d3/2log⁡(1+T/d3/2))\Omega(d^{3/2}\log(1+T/d^{3/2})) for bounded rewards; since bounded rewards belong to our class, our compact-action bound is minimax optimal in dd at α=1. The gap remains open for α<1.

Proposition C.8 (new Appendix C.4) shows that a hypercube instance of the type used classically gives only the d-armed rate under the hybrid argument; see also our reply to Reviewer fxwW.

W2 / Q1: Logarithmic overhead. The revision makes the role of each quantity explicit: the qℓ,iq_{\ell,i} in (28) are nominal accuracy parameters, and the expected query count carries a separate polylogarithmic factor Λ\Lambda. Lemma C.3 now states:

  • for a nominal budget qq, the estimator's error satisfies Ee=0\mathbb E e=0 and E∣e∣1+α≤Cαu q−2α\mathbb E|e|^{1+\alpha}\le C_\alpha u\,q^{-2\alpha};
  • it uses at most Λq\Lambda q queries in expectation, where Λ\Lambda is the polylogarithmic factor hidden in the query bound of Luo et al. (2026, Theorem 3.7). Their error bound carries no logarithmic factor, so the overhead affects only the query count.

The moment bound (30) uses CαC_\alpha exactly. The overhead enters only the query count: the one-replicate cost (31) is Uℓ=Λ∑iqℓ,iU_\ell=\Lambda\sum_iq_{\ell,i}, the timeout 6Uℓ6U_\ell includes it, and inverting the budget relation in Theorem C.6 produces the factor (HLΛ)2α/(1+α)(H_L\Lambda)^{2\alpha/(1+\alpha)}, i.e., the displayed (log⁡(nT))2α/(1+α)(\log(nT))^{2\alpha/(1+\alpha)} times a polylogarithmic factor. The stated regret bound is therefore unchanged. The same convention is applied to (49) and Lemma D.3; (36) and (51) use deterministic-cost estimators whose confidence logarithms are displayed in (17) and (42). This is the same accounting as in LV-G-Elim, whose timeout (Liu et al. 2026a, (5.1)) also includes the logarithmic overhead of the estimator explicitly.

W2: Appendix B.2. The derivation is now complete. For each action aa, let Na=∑ℓ≤LaqℓN_a=\sum_{\ell\le L_a}q_\ell be its total number of queries, so ∑aNa≤T\sum_aN_a\le T.

  • Since ϵℓqℓ\epsilon_\ell q_\ell grows geometrically in ℓ\ell, the exploration regret of aa is ≲u1/(2α)ϵLa−(1−α)/(2α)\lesssim u^{1/(2\alpha)}\epsilon_{L_a}^{-(1-\alpha)/(2\alpha)}.
  • The lower bound qℓ≥c u1/(2α)ϵℓ−(1+α)/(2α)q_\ell\ge c\,u^{1/(2\alpha)}\epsilon_\ell^{-(1+\alpha)/(2\alpha)} on the cap converts this into ≲u1/(1+α)Na(1−α)/(1+α)\lesssim u^{1/(1+\alpha)}N_a^{(1-\alpha)/(1+\alpha)}.
  • Jensen's inequality for x↦x(1−α)/(1+α)x\mapsto x^{(1-\alpha)/(1+\alpha)} then gives ∑aNa(1−α)/(1+α)≤K2α/(1+α)T(1−α)/(1+α)\sum_aN_a^{(1-\alpha)/(1+\alpha)}\le K^{2\alpha/(1+\alpha)}T^{(1-\alpha)/(1+\alpha)}.

The exploitation term 4TϵL4T\epsilon_L uses the budget test (phase L+1L+1 does not fit), which gives ϵL≲u1/(1+α)(K/T)2α/(1+α)\epsilon_L\lesssim u^{1/(1+\alpha)}(K/T)^{2\alpha/(1+\alpha)} when T≳Klog⁡TT\gtrsim K\log T; the cases T≲Klog⁡TT\lesssim K\log T and L=0L=0 are absorbed in the additive and leading terms.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

Summary of revisions

We thank all reviewers for their feedback. A revised PDF has been uploaded; all changes are in blue, and the additional main-text page is used for the first three items. Because of a new remark in Section 4.1, the former Theorem 4.2 and Remark 4.3 are now Theorem 4.3 and Remark 4.4; equation numbers are unchanged.

  1. [All] Section 1: a "Positioning" paragraph separating inherited components from new ones. Table 1: prior bounds at α=1 (bounded rewards) from Liu et al. (2026a), Hikima et al. (2024) and the concurrent work (Nov 2026) of Liu et al. (2026b, arXiv:2611.15237).
  2. [R9W6, fxwW, 2Xfn] New Remark 4.2: the remaining dimension gaps and their sources, including the concurrent Ω(d3/2)\Omega(d^{3/2}) lower bound of Liu et al. (2026b), which shows that our compact-action bound is optimal in dd at α=1. New Proposition C.8 (Appendix C.4): a hypercube instance yields the same rate as the d-armed embedding under the hybrid argument.
  3. [R9W6, fxwW] Remark 3.3: the K-improvement is analytical and holds for the batched UCB of Wu et al. (2023). Remark 4.4: comparison with Hikima et al. (2024) and Liu et al. (2026a) at α=1, and the distinction between UCB-type and design-based analyses.
  4. [R9W6] Appendix B.2: complete derivation of the exploration and exploitation bounds. Lemma C.3, (24), (28), (31), (49) and Lemma D.3: the logarithmic overhead of the unbiased estimator is explicit and propagated through the timeout and budget analysis.
  5. [2Xfn] Section 5: the scope of the emulated experiments is stated explicitly.
  6. [R9W6] Appendix A: precise statement of the results of Liu et al. (2026a,b) and Hikima et al. (2024).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.