Pith. sign in

REVIEW 3 major objections 5 minor 49 references

Contextual Online Pricing with (Biased) Offline Data

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Biased offline pricing data, used under a known bias bound, never worsens regret and can beat the √T online benchmark.

desk verdict The paper is a real step forward for contextual pricing with biased offline data, but the stress test is right: Case 3 of Theorem 1's proof needs an empirical lower bound that the paper never states, so the central guarantee is not yet established as written. read the letter →

arxiv 2507.02762 v1 pith:IJGNV2O2 submitted 2025-07-03 cs.LG stat.ML

Yixuan Zhang , Ruihao Zhu , Qiaomin Xie This is my paper · ORCID
classification cs.LGstat.ML
keywords contextualonlinepricingbiasedofflinedataregretboundsoptimisminthefaceofuncertaintyoffline-to-onlinelearningpriceelasticityminimaxlowerstochasticlinearbandits
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks whether a firm that already holds historical pricing logs—possibly collected from a different, shifted market—can use them to reduce the regret of contextual online pricing, and it answers yes, with caveats. Its central claim is that if the firm knows an upper bound $V$ on how far the offline demand parameters are from the online ones, an optimism-based policy attains regret no worse than the purely online $\tilde{O}(\sqrt{T})$ benchmark and, when bias and data dispersion are favorable, strictly better rates. In the scalar price-elasticity case the rate is $\tilde{O}(d_1\sqrt{T} \wedge (V^2T + d_1T/(\lambda_{\min}(\hat{\Sigma})+(N\wedge T)\delta^2)))$, where $\delta^2$ measures how far the offline data's implicit pricing rule sits from the unknown online optimum. A generalized version achieves the analogous worst-case optimal rate $\tilde{O}((d_1+d_2)\sqrt{T} \wedge (V^2T + (d_1+d_2)T/\lambda_{\min}(\hat{\Sigma})))$ when elasticity is vector-valued. If the bias bound is unknown, a third algorithm still guarantees sublinear regret and beats pure online whenever the true bias is small, and a companion impossibility result shows that without any bias information no policy can beat $\sqrt{T}$ uniformly.

What carries the argument

The carrying object is the three-ellipsoid confidence set used by CO3, the intersection of an online Gram ellipsoid, a combined offline-online Gram ellipsoid, and an unweighted Euclidean ball around the combined least-squares estimate. The online ellipsoid preserves the pure-online $\tilde{O}(\sqrt{T})$ bound; the other two let the offline data shrink the parameter estimate, and the combined-Gram ellipsoid paired with the optimistic pricing rule drives prices away from the offline rule when $\delta^2$ is large, accelerating both exploration and exploitation. The quantity $\delta^2$ is constructed from the empirical offline pricing rule $\hat{p}(x,y)=\hat{A}^\top x / y$ with $\hat{A}=\hat{\Sigma}_{x,x}^{-1}\hat{\Sigma}_{x,y}$, which generalizes the classical average price in the no-context case. A preliminary offline test decides whether simply charging $\hat{p}$ is already sufficient, which is how the algorithm adapts to the unknown $\delta^2$.

What would settle it

Take the scalar case with $d_1=1$, $T=N=1000$, $\lambda_{\min}(\hat\Sigma)\approx T$, $\delta^2 = T^{-1/2}$, and declared $V^2=T^{-1/2}$, so Theorem 1 predicts regret $\tilde{O}(\sqrt{T})$; if repeated simulations of CO3 under these exact parameters show regret growing faster than $c\sqrt{T}\log T$ for any fixed $c$, the claimed upper bound is contradicted.

Watch

Extended reading notes

Core claim

The paper's core claim is that biased offline data can be leveraged in contextual online pricing without sacrificing the $\sqrt{T}$ safety guarantee, provided the learner holds a bias bound $V \ge \|\theta'_*-\theta_*\|$. For the scalar-elasticity case ($d_2=1$), the paper identifies $\delta^2 = \mathbb{E}_{x,y}[(\hat{p}(x,y)-p^*_{\theta_*}(x,y))^2]$ as the natural measure of how far the offline data lie from the unknown online optimum, where $\hat{p}(x,y)$ is the ratio of an OLS-fitted baseline-price coefficient vector to the scalar elasticity feature. The CO3 algorithm, using a three-ellipsoid confidence set and an offline preliminary test, is shown to attain $\tilde{O}(d_1\sqrt{T} \wedge (V^2T + d_1T/(\lambda_{\min}(\hat{\Sigma})+(N\wedge T)\delta^2)))$ regret, and matching lower bounds show the dependence on $\delta^2$, $V$, $N$, and $\lambda_{\min}(\hat{\Sigma})$ is minimax optimal up to dimension factors. In the general setting, GCO3 attains $\tilde{O}((d_1+d_2)\sqrt{T} \wedge (V^2T + (d_1+d_2)T/\lambda_{\min}(\hat{\Sigma})))$, and RCO3 handles unknown bias with a $T^\alpha$ test phase, yielding $\tilde{O}(T^\alpha + V_{\mathrm{true}}^2T)$ regret when the true bias is small and $\tilde{O}(\sqrt{T})$ otherwise. The same machinery transfers to stochastic linear bandits with biased offline data.

Load-bearing premise

The guarantees assume the firm knows a bias bound $V$ that actually covers the true shift, $\|\theta'_*-\theta_*\| \le V$; if the true bias exceeds $V$, the confidence set can exclude $\theta_*$ and the stated bounds no longer hold, while a very loose $V$ makes the $V^2T$ term dominate and adds nothing over the $\sqrt{T}$ baseline.

Editorial extensions

If this is right

  • With a small bias bound ($V^2 \in O(T^{-1/2})$) and sufficiently dispersed offline data, CO3 and GCO3 achieve regret strictly below the purely online $\tilde{O}(\sqrt{T})$ rate, so historical logs shorten the expensive exploration phase.
  • The bounds have the form of a minimum between the pure-online rate and an offline-improved rate, so the algorithms inherit a never-worse safety property: irrelevant or badly shifted logs cannot increase regret beyond what an online-only policy would incur.
  • Setting $V=0$ recovers the existing OPOD (unbiased offline data) guarantee while relaxing the assumptions on the offline data, and the general bound is the first for contextual pricing with either biased or unbiased offline logs.
  • When the bias bound is unknown, RCO3 trades a $T^\alpha$ test phase for a guarantee that is always sublinear and improves on online-only pricing whenever $V_{\mathrm{true}}^2 \lesssim T^{-\alpha}$.
  • The same two-ellipsoid confidence set yields analogous regret bounds for stochastic linear bandits with biased offline data, subsuming the earlier finite-action result.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The $\delta^2$ definition suggests a practical pre-deployment diagnostic: fit $\hat p$ from historical logs, estimate $\delta^2$ on a held-out context sample, and use the size of $\delta^2/(V^2+1/\lambda_{\min}(\hat\Sigma))$ to decide whether to trust offline exploitation or explore online, mirroring the algorithm's internal test.
  • The bound implies a bias-versus-dispersion trade-off: a log collected under many different prices (large $\lambda_{\min}(\hat\Sigma)$) can be worth more than an unbiased log from a single fixed price, which could change how firms design historical price experiments.
  • By transferring verbatim to stochastic linear bandits, the machinery likely extends to other offline-to-online decision problems—recommendation, clinical dosing, ad bidding—where logged policies are stale; the practical question becomes estimating $\|\theta'_*-\theta_*\|$ and the log's dispersion, exactly the two quantities the regret bound depends on.
  • An open extension the authors leave implicit is a fully adaptive version that re-estimates the bias periodically rather than in one test phase, potentially recovering the instance-dependent $\delta^2$ rate without the $T^\alpha$ worst-case cost.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper studies contextual online pricing with biased offline data, where the firm has access to an offline dataset generated under a potentially different demand parameter. For the scalar price-elasticity case, it defines an instance-dependent quantity \delta^2 measuring the distance between the offline pricing strategy and the unknown online optimum, and proposes the CO3 algorithm with a three-ellipsoid confidence set and an offline test phase, claiming a regret bound of \tilde{O}(d_1\sqrt{T} \wedge (V^2T + d_1T/(\lambda_{\min}(\hat\Sigma)+(N\wedge T)\delta^2))). For general elasticity, it proposes GCO3 with a two-ellipsoid confidence set and a worst-case bound \tilde{O}((d_1+d_2)\sqrt{T} \wedge (V^2T + (d_1+d_2)T/\lambda_{\min}(\hat\Sigma))). It also provides lower bounds, a robust variant RCO3 when the bias bound V is unknown, numerical experiments, and an extension to stochastic linear bandits.

Significance. If the stated results hold, this is the first tight characterization of contextual online pricing with biased offline data, extending prior work on unbiased offline data [6, 28] and the K-armed bandit setting [9]. The identification of \delta^2 as an instance-dependent complexity measure and the three-ellipsoid confidence set are genuinely novel. The lower-bound arguments using multivariate van Trees and change-of-measure are detailed and provide useful techniques. The paper also gives a first robust algorithm (RCO3) with sub-linear regret guarantees without a known bias bound. However, the proof of the central upper bound (Theorem 1) contains a nontrivial gap in the switching-test argument, and the proof of Theorem 3 is omitted; these issues need to be fixed before the results can be considered established.

major comments (3)
  1. [Appendix E, proof of Theorem 1, Case 3] The lower bound on the empirical offline loss is unjustified. The chain after '(i)' lower-bounds \sum_{n=1}^N (\hat{p}(\hat{x}_n,\hat{y}_n) - p^*_{\theta_*}(\hat{x}_n,\hat{y}_n))^2 by a multiple of N(V^2 + (d_1+\log T)/\lambda_{\min}(\hat\Sigma)) directly from the assumption \delta^2 \ge K N \cdots (V^2 + \cdots). But \delta^2 is an expectation over (x,y), and Lemma 10 gives only an upper bound of the empirical sum by N\cdot\text{const}\cdot\delta^2, never a lower bound. A lower bound can be obtained by writing p^*-\hat{p} = ((\hat{A}+\alpha_*/(2\beta_*))^\top x)/y and using Assumption 2, but this argument is not present. Without it, Case 3 does not rule out the scenario where the test condition holds while \delta^2 is large, which would lead to linear regret. Additionally, the proof invokes inequality (18), which requires \lambda_{\min}(\hat\Sigma) \ge \sqrt{T}; this condition is not implied by the Case 3 assumption \max\{V^2, 1/\lambda_{\min}(\hat\Sigma)\} \le \sqrt{T}.
  2. [Algorithm 1, Line 1] The offline test in Line 1 uses the threshold N x_{\max}^2 y_{\max}^2/(y_{\min}^2 \lambda_{\min}(E[xx^\top])) \max\{V^2, 1/\lambda_{\min}(\hat\Sigma)\}, which requires knowledge of \lambda_{\min}(E[xx^\top]) and other distributional constants. These are not included in the algorithm input list (which only specifies offline data, \lambda, confidence radii, and V), and they are generally unknown because the online feature distribution is not observed before time 0. The paper does not explain how the algorithm obtains these constants or how to replace them with data-dependent estimates. As stated, CO3 is not implementable.
  3. [Section 4.1, Theorem 3] The proof of Theorem 3 is omitted entirely. The text states 'Because the steps are nearly identical, the proof is omitted' and refers to Appendices E.1.1 and E.1.2, but those appendices prove specific parts of Lemma 13 for the CO3 algorithm, not the general two-ellipsoid GCO3 bound with dimension d_1+d_2. A central theorem of the paper must have a complete proof, either in the main text or the appendix; this omission makes the general-case claim unverified.
minor comments (5)
  1. [Appendix E.1.1] The phrase 'the last equality holds' after bounding the per-round regret should be 'the last inequality holds', since the preceding display is an inequality chain.
  2. [Lemma 2 proof] There is a typo: 'the largest possible largest c' should read 'the largest possible c'.
  3. [Theorem 5 statement] The phrase 'the optimal choices of \alpha is \alpha \in (\max\{0,1-\beta\}, 1/2)' is imprecise: within this interval there is a trade-off between the two regret regimes, and no single value is uniformly optimal, as the proof itself notes. The text should say 'admissible choices' rather than 'optimal choices'.
  4. [Checklist item 5] The NeurIPS checklist claims open access to data and code, but no code or data repository is provided anywhere in the paper or supplementary material.
  5. [Figure 2] The caption says '2-sigma error bars' but the text says 'shaded bands indicate 2-sigma error bars'; please clarify whether the shaded regions represent standard errors or standard deviations across the 20 trials, and state the calculation method for the bands.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the regret bounds are instance-dependent statements proved from the model and offline data, not restatements of fitted inputs or self-citations.

full rationale

The paper's central claims are regret bounds whose dependence on δ^2, V, λ_min(Σ̂), and N comes from the analysis, not from fitting. The quantity δ^2 is defined as E[(p̂ - p*_θ*)^2] with p̂ an OLS-based empirical policy; it is an instance-dependent property of the unknown θ*, and the CO3/GCO3 algorithms never need to know it. The offline test in Algorithm 1 is a genuine decision rule, and the 'small δ^2 implies charge p̂' direction is supported by Lemma 10, while the opposite direction is argued through the three-ellipsoid confidence set. There is no step where a parameter is fitted to data and then renamed a prediction: the confidence radii are chosen via standard concentration arguments, and V is an assumed known bound rather than an estimated quantity. Lower bounds (Theorems 2 and 4) use van Trees and two-point change-of-measure arguments that are independent of the upper-bound proofs. Self-citation is not load-bearing: the closest algorithmic comparisons are to [6], [9], and [28], none of which share authors with this paper; the only author-overlapping reference, [4], is a background citation for the linear demand model and plays no role in the derivation chain. The reviewer's noted concern in Appendix E, Case 3 (that a lower bound on the empirical offline loss in terms of δ^2 is asserted without an explicit lemma, since Lemma 10 only gives an upper bound), is a proof-rigor gap rather than circularity: it does not make the claimed regret bound equivalent to its inputs. Similarly, the omission of the proof of Theorem 3 ('the proof is omitted' because steps are 'nearly identical') is an exposition/completeness issue, not a circular reduction. Overall, the derivation chain is self-contained with respect to the stated assumptions and does not reduce to its own conclusions.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central regret bounds rest on the linear demand model, Assumptions 1 and 2, and a known bias bound V; the robust variant relaxes only the last at the cost of weaker bounds. No new physical entities are introduced. The RCO3 exponent alpha is a hand-chosen tradeoff parameter.

free parameters (1)
  • RCO3 test-length exponent alpha = alpha in (max{0,1-beta}, 1/2); experiments use alpha = 1/4
    In Theorem 5, the test phase length is T' = Theta(T^alpha). Alpha trades off worst-case regret T^{1-alpha} against small-bias regret T^alpha + V^2 T, and is chosen by hand rather than learned from data.
assumptions (5)
  • domain assumption Linear demand model: D_t = alpha^T_* x_t + beta^T_* y_t p_t + epsilon_t, with sub-Gaussian noise (Eq. 1).
    The entire pricing and regret analysis, including the definition of the optimal price p* = -alpha^T x / (2 beta^T y), is built on this structural model.
  • domain assumption Assumption 1: compact parameter and feature sets, positive definite E[x x^T] and E[y y^T], and bounds l_alpha, u_alpha, l_beta, u_beta such that p* stays in [l,u].
    Used for strong concavity of revenue, Lipschitzness of p*, and to ensure the confidence ellipsoids have non-vacuous radii.
  • domain assumption Assumption 2: lambda_min(Sigma_hat_xx) >= c N for the offline baseline features.
    Needed in Lemmas 9 and 10 to transfer offline dispersion into the delta-squared-dependent regret terms; if offline features are degenerate, the improvement disappears.
  • domain assumption Bias bound V is known and satisfies V >= ||theta'_* - theta_*|| for Theorems 1 through 4.
    The confidence sets and all non-robust regret bounds are parameterized by V; underestimating V breaks the guarantee, and overestimating it removes the benefit of offline data.
  • standard math Van Trees inequality, Freedman's inequality, matrix Chernoff, and standard concentration results are correct as invoked.
    Used in lower-bound proofs (Appendix F) and high-probability events (Lemmas 5 and 6); no alternative derivation is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Contextual Online Pricing with (Biased) Offline Data." pith.science (2026). https://pith.science/paper/IJGNV2O2

@misc{pith2026250702762,
  author       = {Pith},
  title        = {Pith review of: Contextual Online Pricing with (Biased) Offline Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IJGNV2O2}},
  note         = {Machine review of arXiv:2507.02762}
}
abstract

We study contextual online pricing with biased offline data. For the scalar price elasticity case, we identify the instance-dependent quantity $\delta^2$ that measures how far the offline data lies from the (unknown) online optimum. We show that the time length $T$, bias bound $V$, size $N$ and dispersion $\lambda_{\min}(\hat{\Sigma})$ of the offline data, and $\delta^2$ jointly determine the statistical complexity. An Optimism-in-the-Face-of-Uncertainty (OFU) policy achieves a minimax-optimal, instance-dependent regret bound $\tilde{\mathcal{O}}\big(d\sqrt{T} \wedge (V^2T + \frac{dT}{\lambda_{\min}(\hat{\Sigma}) + (N \wedge T) \delta^2})\big)$. For general price elasticity, we establish a worst-case, minimax-optimal rate $\tilde{\mathcal{O}}\big(d\sqrt{T} \wedge (V^2T + \frac{dT }{\lambda_{\min}(\hat{\Sigma})})\big)$ and provide a generalized OFU algorithm that attains it. When the bias bound $V$ is unknown, we design a robust variant that always guarantees sub-linear regret and strictly improves on purely online methods whenever the exact bias is small. These results deliver the first tight regret guarantees for contextual pricing in the presence of biased offline data. Our techniques also transfer verbatim to stochastic linear bandits with biased offline data, yielding analogous bounds.

Figures

Figures reproduced from arXiv: 2507.02762 by the authors.

Figure 1
Figure 1. Piecewise regret bound as a function of exact bias square [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Performances of CO3, GCO3, and RCO3 compared with baseline algorithms. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 42 canonical work pages

  1. [9]

    Leveraging (biased) information: Multi-armed bandits with offline data

    Wang Chi Cheung and Lixing Lyu. Leveraging (biased) information: Multi-armed bandits with offline data. arXiv preprint arXiv:2405.02594, 2024. 2, 5, 6, 7, 8, 12, 13, 40

    arXiv 2024

  2. [1]

    Improved algorithms for linear stochastic bandits

    Yasin Abbasi-Yadkori, Dávid Pál, and Csaba Szepesvári. Improved algorithms for linear stochastic bandits. Advances in neural information processing systems, 24, 2011. 5

    work page 2011

  3. [2]

    Personalized dynamic pricing with machine learning: High- dimensional features and heterogeneous elasticity

    Gah-Yi Ban and N Bora Keskin. Personalized dynamic pricing with machine learning: High- dimensional features and heterogeneous elasticity. Management Science, 67(9):5549–5568,

    work page

  4. [3]

    High-probability regret bounds for bandit online linear optimization

    Peter Bartlett, Varsha Dani, Thomas Hayes, Sham Kakade, Alexander Rakhlin, and Ambuj Tewari. High-probability regret bounds for bandit online linear optimization. In Proceedings of the 21st Annual Conference on Learning Theory-COLT 2008, pages 335–342. Omnipress, 2008. 19

    work page 2008

  5. [4]

    Meta dynamic pricing: Transfer learning across experiments

    Hamsa Bastani, David Simchi-Levi, and Ruihao Zhu. Meta dynamic pricing: Transfer learning across experiments. Management Science, 68(3):1865–1881, 2022. 3, 4

    work page 2022

  6. [5]

    Robust wasserstein profile inference and applications to machine learning

    Jose Blanchet, Yang Kang, and Karthyek Murthy. Robust wasserstein profile inference and applications to machine learning. Journal of Applied Probability, 56(3):830–857, 2019. 4

    work page 2019

  7. [6]

    Online pricing with offline data: Phase transition and inverse square law

    Jinzhi Bu, David Simchi-Levi, and Yunzong Xu. Online pricing with offline data: Phase transition and inverse square law. In International Conference on Machine Learning, pages 1202–1210. PMLR, 2020. 2, 3, 5, 6, 9, 12, 13, 14, 24

    work page 2020

  8. [7]

    Domain adaptation for offline rein- forcement learning with limited samples

    Weiqin Chen, Sandipan Mishra, and Santiago Paternain. Domain adaptation for offline rein- forcement learning with limited samples. arXiv preprint arXiv:2408.12136, 2024. 12

    work page arXiv 2024

Show all 49 references
  1. [8]

    Data-pooling reinforcement learning for personal- ized healthcare intervention

    Xinyun Chen, Pengyi Shi, and Shanwen Pu. Data-pooling reinforcement learning for personal- ized healthcare intervention. arXiv preprint arXiv:2211.08998, 2022. 4

    2022 arXiv

  2. [10]

    Feature-based dynamic pricing

    Maxime C Cohen, Ilan Lobel, and Renato Paes Leme. Feature-based dynamic pricing. Manage- ment Science, 66(11):4921–4943, 2020. 1

    2020

  3. [11]

    Off-dynamics reinforcement learning: Training for transfer with domain classifiers

    Benjamin Eysenbach, Swapnil Asawa, Shreyas Chaudhari, Sergey Levine, and Ruslan Salakhut- dinov. Off-dynamics reinforcement learning: Training for transfer with domain classifiers. arXiv preprint arXiv:2006.13916, 2020. 12

    2006 arXiv

  4. [12]

    A tail inequality for quadratic forms of subgaussian random vectors

    Daniel Hsu, Sham Kakade, and Tong Zhang. A tail inequality for quadratic forms of subgaussian random vectors. 2012. 17

    2012

  5. [13]

    Dynamic pricing with an unknown demand model: Asymp- totically optimal semi-myopic policies

    N Bora Keskin and Assaf Zeevi. Dynamic pricing with an unknown demand model: Asymp- totically optimal semi-myopic policies. Operations research, 62(5):1142–1167, 2014. 2, 13 10

    2014

  6. [14]

    Meta-thompson sampling

    Branislav Kveton, Mikhail Konobeev, Manzil Zaheer, Chih-wei Hsu, Martin Mladenov, Craig Boutilier, and Csaba Szepesvari. Meta-thompson sampling. In International Conference on Machine Learning, pages 5884–5893. PMLR, 2021. 2, 13

    2021

  7. [15]

    Bandit algorithms

    Tor Lattimore and Csaba Szepesvári. Bandit algorithms. Cambridge University Press, 2020. 17, 27, 40

    2020

  8. [16]

    Dynamic pricing with external information and inventory constraint

    Xiaocheng Li and Zeyu Zheng. Dynamic pricing with external information and inventory constraint. Management Science, 70(9):5985–6001, 2024. 4

    2024

  9. [17]

    On the prior sensitivity of thompson sampling

    Che-Yu Liu and Lihong Li. On the prior sensitivity of thompson sampling. In International Conference on Algorithmic Learning Theory, pages 321–336. Springer, 2016. 2, 13

    2016

  10. [18]

    Dynamic pricing with demand covariates

    Sheng Qiang and Mohsen Bayati. Dynamic pricing with demand covariates. arXiv preprint arXiv:1604.07463, 2016. 3

    2016 arXiv

  11. [19]

    Hybrid transfer reinforcement learning: Provable sample efficiency from shifted-dynamics data

    Chengrui Qu, Laixi Shi, Kishan Panaganti, Pengcheng You, and Adam Wierman. Hybrid transfer reinforcement learning: Provable sample efficiency from shifted-dynamics data. arXiv preprint arXiv:2411.03810, 2024. 12

    2024 arXiv

  12. [20]

    Online learning with predictable sequences

    Alexander Rakhlin and Karthik Sridharan. Online learning with predictable sequences. In Conference on Learning Theory, pages 993–1019. PMLR, 2013. 2, 12

    2013

  13. [21]

    Multi-armed bandit problems with history

    Pannagadatta Shivaswamy and Thorsten Joachims. Multi-armed bandit problems with history. In Artificial intelligence and statistics, pages 1046–1054. PMLR, 2012. 2, 12

    2012

  14. [22]

    Bayesian decision-making under misspecified priors with applications to meta-learning

    Max Simchowitz, Christopher Tosh, Akshay Krishnamurthy, Daniel J Hsu, Thodoris Lykouris, Miro Dudik, and Robert E Schapire. Bayesian decision-making under misspecified priors with applications to meta-learning. Advances in neural information processing systems, 34:26382– 26394...

    2021

  15. [23]

    An introduction to matrix concentration inequalities

    Joel A Tropp et al. An introduction to matrix concentration inequalities. Foundations and Trends® in Machine Learning, 8(1-2):1–230, 2015. 39

    2015

  16. [24]

    Introduction to the non-asymptotic analysis of random matrices

    Roman Vershynin. Introduction to the non-asymptotic analysis of random matrices. arXiv preprint arXiv:1011.3027, 2010. 19

    2010 arXiv

  17. [25]

    Leveraging offline data in online reinforcement learning

    Andrew Wagenmaker and Aldo Pacchiano. Leveraging offline data in online reinforcement learning. In International Conference on Machine Learning, pages 35300–35338. PMLR, 2023. 2, 12

    2023

  18. [26]

    Taking a hint: How to leverage loss predictors in contextual bandits? In Conference on Learning Theory, pages 3583–3634

    Chen-Yu Wei, Haipeng Luo, and Alekh Agarwal. Taking a hint: How to leverage loss predictors in contextual bandits? In Conference on Learning Theory, pages 3583–3634. PMLR, 2020. 2, 12

    2020

  19. [27]

    Combining offline causal inference and online bandit learning for data driven decision

    Li Ye, Yishi Lin, Hong Xie, and John Lui. Combining offline causal inference and online bandit learning for data driven decision. arXiv preprint arXiv:2001.05699, 2020. 2, 12

    2001 arXiv

  20. [28]

    Advancements in Management Science: Applications to Online Retail, Healthcare, and Non-Profit Fundraising

    Chen Wen Sabrina Zhai. Advancements in Management Science: Applications to Online Retail, Healthcare, and Non-Profit Fundraising. PhD thesis, Massachusetts Institute of Technology,

  21. [29]

    Warm-starting contextual bandits: Robustly combining supervised and bandit feedback

    Chicheng Zhang, Alekh Agarwal, Hal Daumé III, John Langford, and Sahand N Negahban. Warm-starting contextual bandits: Robustly combining supervised and bandit feedback. arXiv preprint arXiv:1901.00301, 2019. 2, 12, 13 11 A Additional notations This section summarizes the notat...

    1901 arXiv

  22. [32]

    t−1X s=1 ∆α⊤ t x 2 # − Ex,y

    studies contextual bandits but does not improve on the standard online regret and require a restrictive offline log (see also [9, Appendix A.1]). By extending our Algorithm 2 and analysis to the stochastic linear bandit with biased offline data, we obtain regret upper bounds t...

  23. [33]

    λmin( ˆΣ) ∈ O( √ T ) and δ2 ∈ O(T −1/2), then for any admissible policy π ∈ Πo, there exists (θ′ ∗, θ∗) ∈ Θ† × Θ† satisfying ∥θ − ˆθ∥ ≤V and Ex,y[(ˆp(x, y) − p∗ θ∗ (x, y))2] ∈ Θ(δ2) such that Rπ θ′∗,θ∗ (T ) ∈ Ω √ T (log T )λ0 ! . We first define two vectors of offline demand p...

  24. [34]

    Guidelines: • The answer NA means that the abstract and introduction do not include the claims made in the paper

    Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: The main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope. Guidelin...

  25. [35]

    Limitations

    Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: This paper discusses the limitations of the work performed by the authors. Guidelines: • The answer NA means that the paper has no limitation while t...

  26. [36]

    Guidelines: • The answer NA means that the paper does not include theoretical results

    Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [Yes] 41 Justification: This paper provides the full set of assumptions and a complete (and correct) proof for ea...

  27. [37]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and da...

  28. [38]

    Guidelines: • The answer NA means that paper does not include experiments requiring code

    Open access to data and code 42 Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: This paper provides open...

  29. [39]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: This paper specifies all the ...

  30. [40]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes] Justification: This paper reports error bars suitably and correct...

  31. [41]

    All experiments can be conducted on a personal computer

    Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the com- puter resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: This paper provides suf...

  32. [42]

    Guidelines: • The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics

    Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: The research conducted in this paper conforms, in every respect, with the NeurI...

  33. [43]

    Guidelines: • The answer NA means that there is no societal impact of the work performed

    Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [NA] Justification: There is no societal impact of this work performed. Guidelines: • The answer NA means that there is no soc...

  34. [44]

    Guidelines: • The answer NA means that the paper poses no such risks

    Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] Justification: This paper pose...

  35. [45]

    Guidelines: • The answer NA means that the paper does not use existing assets

    Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [NA] Justification: This paper does n...

  36. [46]

    Guidelines: • The answer NA means that the paper does not release new assets

    New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [NA] Justification: This paper does not release new assets. Guidelines: • The answer NA means that the paper does not release new assets....

  37. [47]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? A...

  38. [48]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

  39. [49]

    Answer: [NA] Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components

    Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the ...

    2025

  40. [2021]

    1, 2, 3, 4, 6, 8, 13, 14

  41. [2024]

    1, 2, 3, 5, 6, 9, 12, 13, 14, 18, 31

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.