Pith. sign in

REVIEW 3 major objections 4 minor 28 references

Pre-Trained AI Model Assisted Online Decision-Making under Missing Covariates: A Theoretical Perspective

T0 review · 3 major / 4 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 The paper claims that a pre-trained model used to fill missing covariates carries a measurable price, and that under missing-at-random observations this price can be calibrated away online, leaving only an irreducible noise term.

desk verdict The model elasticity idea is a genuinely clean way to frame imputation error in bandits, but the paper's headline calibration rate (Theorem 5.3) does not follow from the proof as written. read the letter →

arxiv 2507.07852 v1 pith:5B4JZYS7 submitted 2025-07-10 cs.LG stat.ML

Haichen Hu , David Simchi-Levi This is my paper · ORCID
classification cs.LGstat.ML MSC 68T0562L0562G05
keywords contextualbanditsmissingcovariatespre-trainedmodelimputationelasticityatrandomorthogonalstatisticallearningdoublyrobustregressionregretbounds
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

This paper studies a sequential decision-maker who observes only part of the context each round, with some covariate missing. Rather than discarding the round, the decision-maker asks a pre-trained AI model to fill in the missing value. The paper's central claim is that this shortcut has a measurable price, and that under a missing-at-random mechanism the price can be paid down online. It introduces model elasticity, the worst-case sensitivity of the reward class to imputation error, and shows that without calibration the regret carries a linear-in-horizon term proportional to the square root of elasticity. Under missing at random, a calibration procedure built on orthogonal statistical learning and doubly robust regression replaces that linear term with statistically decaying terms plus an irreducible covariate-noise term, so a good pre-trained model yields near-optimal regret.

What carries the argument

The load-bearing objects are: (1) model elasticity, a worst-case squared reward gap between true and imputed covariates, which converts imputation quality into a regret penalty; (2) inverse gap weighting, which turns per-epoch estimation error into a regret bound; (3) local Rademacher complexity and its critical radius, which set the oracle estimation error of the reward class; and (4) a two-step calibration procedure that first estimates the missingness probability $e^*$ and then regresses on the centered class $\mathcal{G}_{\delta_0} = \{g\in\mathcal{G} : \|g-\tilde{g}\|_2 \le \delta_0\}$ with inverse-probability weights, using cross-fitting and a Taylor expansion of the doubly robust loss to control localization error. The machinery turns imputation quality into a quantity that can be improved by data rather than a fixed cost.

What would settle it

Run Algorithm 3 on a parametric reward class with a pre-trained model whose true error is $2\delta_0$ while the algorithm is fed $\delta_0$: if the bound holds, the calibrated term $\delta_0^{d/(d+2)}T^{(d-2)/(d+2)}$ must appear in the regret; the observed linear-in-$T$ behavior would instead match Theorem 4.6's uncalibrated rate. A second test replaces the bounded noise $\eta$ in $z^*=g^*(x)+\eta$ with a heavy-tailed draw of infinite second moment and checks whether the linear term stays at $L_Z\omega_0 T$.

Watch

Extended reading notes

Core claim

The paper establishes oracle inequalities and regret bounds for contextual bandits with missing covariates imputed by a pre-trained model $\tilde{g}$. Its main positive result is that the per-round cost of imputation is governed by model elasticity $\mathcal{E}_F^P(\tilde{g}) = \sup_{f\in F, a\in A}\mathbb{E}[(f(x,z^*,a)-f(x,\tilde{g}(x),a))^2]$; in the general MNAR setting the regret is $O\big(\sqrt{K}(\sum \text{oracle terms} + \sqrt{\mathcal{E}_F^P(\tilde{g})}\,T)\big)$. Under MAR, the pre-trained model can be sequentially calibrated: with a known bound $\delta_0$ on $\|\tilde{g}-g^*\|_2$ and bounded covariate noise of second moment $\omega_0$, Algorithm 3 achieves a regret in which the $\sqrt{\mathcal{E}_F^P(\tilde{g})}\,T$ term is replaced by calibration-error terms, and for parametric reward and missingness classes the rate is $\tilde{O}(T^{1/2} + \delta_0^{d/(d+2)}T^{(d-2)/(d+2)} + \omega_0 T)$. The retained linear term $L_Z\omega_0 T$ is shown to be irreducible because even perfect calibration cannot remove the noise $\eta$ in $z^* = g^*(x) + \eta$.

Load-bearing premise

The entire calibration guarantee rests on the decision-maker knowing a constant $\delta_0$ that truthfully upper-bounds the $\mathbb{L}_2$ error between the pre-trained model and the true covariate function, plus a bound $\omega_0$ on the covariate noise; misspecified $\delta_0$ excludes the true model from the localized class or bloats it, and heavy-tailed noise breaks the irreducible linear term.

Editorial extensions

If this is right

  • Without calibration, every unit of imputation error costs linear regret: the term $\sqrt{\mathcal{E}_F^P(\tilde{g})}\,T$ cannot be avoided under MNAR by this plug-in approach.
  • Under MAR, calibration removes that linear term, leaving only the irreducible noise term $L_Z\omega_0 T$.
  • For parametric reward and missingness classes, regret becomes $\tilde{O}(T^{1/2} + \delta_0^{d/(d+2)}T^{(d-2)/(d+2)} + \omega_0 T)$, so a more accurate pre-trained model (smaller $\delta_0$) strictly shrinks the middle term.
  • The dominant estimation term is controlled by the hardest of the reward, missingness, and covariate function classes, so the framework degrades gracefully when any one class is non-parametric.
  • Model elasticity provides a single metric that ranks pre-trained models for decision-making across different missingness mechanisms.

Reading between the lines

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

  • Model elasticity could be used as a pre-deployment selection criterion: among candidate imputers, one could pick the model minimizing an empirically estimated $\mathcal{E}_F^P(\tilde{g})$ rather than minimizing prediction error alone.
  • The irreducibility of $\omega_0 T$ suggests a budget trade-off: when covariate noise is large, improving the imputer has limited value and investment should shift to higher-quality raw covariates.
  • The bound's $1/\epsilon_0^2$ dependence on the minimum observation probability implies that very sparse covariates amplify calibration error; a testable extension is to compare regret across different missingness rates.
  • The calibration scheme is modular and could be plugged into other base algorithms, such as Thompson sampling with generation of missing data, to remove their imputation bias as well.
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 / 4 minor

Summary. The paper studies sequential contextual decision-making when some covariates are missing but a pre-trained model is available to impute them. The authors introduce a quantity called “model elasticity”, defined as the worst-case squared reward discrepancy between using the true covariate and the imputed one, and show via a local Rademacher complexity/IGW analysis that the regret of their plug-in algorithm (Algorithm 1, PRIMO) decomposes into a standard oracle estimation term plus a term proportional to the square root of the model elasticity times the horizon. Under a missing-at-random (MAR) assumption, they propose a calibration step (Algorithms 2 and 3, PRIMO-Cal) based on orthogonal/doubly robust regression, and claim that the linear elasticity term can be replaced by statistically decaying calibration terms plus an unavoidable noise term linear in the horizon.

Significance. The paper addresses a timely and practically relevant question: how to use pre-trained imputation models in online decision-making and how to quantify their effect on regret. The notion of model elasticity is a clean conceptual device, and the use of local Rademacher complexities, peeling arguments, and inverse gap weighting is appropriate for the type of guarantee the authors seek. The MNAR result (Theorem 4.6) is a plausible and useful decomposition, and the appendix contains a substantial amount of proof machinery. If the calibration analysis were correct, the paper would make a meaningful theoretical contribution to surrogate-assisted contextual bandits. However, the central calibration theorem has a serious exponent error and the statement, proof, and discussion disagree about the rates, so the main advertised improvement is not established as written.

major comments (3)
  1. [Theorem 5.3 and Appendix C] The claimed calibration rate does not follow from the stated proof. Theorem 5.2 bounds ||\hat g - g^*||_2^2 by (1/\epsilon_0^2)(\delta_0^{2d/(d+2)} N^{-2/(d+2)} + \cdots). In the proof of Theorem 5.3 this enters through \sqrt{E_F^P(\hat g)} \lesssim L_Z ||\hat g - g^*||_2, so the epoch-s contribution should involve N^{-1/(d+2)}, not N^{-2/(d+2)}. With N=|D_{s-2}|=2^{s-3}, the correct decay is L_Z \epsilon_0^{-1} \delta_0^{d/(d+2)} 2^{-(s-3)/(d+2)}, not the displayed 2^{-2(s-3)/(d+2)} inside \Upsilon_s. Combining with the epoch weight 2^{s-1} from Lemma B.4 and summing gives a regret contribution of order \delta_0^{d/(d+2)} T^{(d+1)/(d+2)}, not \delta_0^{d/(d+2)} T^{(d-2)/(d+2)}. The theorem statement, the proof, and the interpretation paragraph also disagree among themselves about the exponents: \Upsilon_s uses \delta_0^{d/(d+1)}, the regret display uses \delta_0^{d/(d+2)}, and the discussion advertises T^{(d-2)/(d+2)}. This is load-bearing because the claimed superiority of Algorithm 3 over Algorithm 1 rests on this rate.
  2. [Assumption 5.1(A4) and Algorithm 2] The calibration procedure requires the decision-maker to know a constant \delta_0 such that ||\tilde g - g^*||_2 \le \delta_0, and the class G_{\delta_0} is defined using this constant. This is an explicit assumption, but it is load-bearing for the calibration claim. If the true distance exceeds \delta_0, the true function g^* may be excluded from G_{\delta_0}, and the ERM in Algorithm 2 is no longer estimating g^*, so the regret analysis would need an additional misspecification term. If \delta_0 is enlarged to be safe, the critical radius q_{\delta_0,N} \asymp \delta_0^{d/(d+2)} N^{-1/(d+2)} degrades polynomially. The manuscript does not discuss how \delta_0 is obtained in practice or how the guarantee degrades under misspecification, so the statement that calibration “significantly improves” regret is conditional on an unverified oracle quantity.
  3. [Theorem 5.2, d=2 case] In the computation of the critical radius q_{\delta_0,n}, the proof says “without loss of generality, we assume that d \ne 2” and then evaluates the Dudley integral using 1/(1-d/2). The case d=2 is not handled. Since Assumption 5.1(A3) allows any d>0, the d=2 case should be treated separately, typically with an additional logarithmic factor. This is a local gap, but it must be closed because the statement of Theorem 5.2 and the subsequent regret rates cover all d>0.
minor comments (4)
  1. [Theorem 5.3] The notation in the displayed regret bound is hard to read: the exponents such as “-2(s-3)/(d+2)+s-1” and “s2^{s-3}2^{s-1}” are ambiguous. Please use explicit parentheses and consistent superscript formatting, and ensure that \Upsilon_s has the same exponent as the regret display.
  2. [Assumption 5.1(A1)] The condition “G is 1-uniformly bounded” is used in the proofs, but the norm with respect to which the bound holds is not defined. Please specify whether this is a uniform sup-norm bound or an L2 bound, since the concentration arguments rely on it.
  3. [Section 4.1] The algorithm description in prose says “the s+1 th epoch is twice as long as the previous one,” but the pseudocode and proofs use \beta_s = 2^s. This is consistent, but the wording “twice as long” should be “exponentially longer” to match the schedule.
  4. [Appendix B, Lemma B.3] In the display after “By some algebra”, the factor 2n/N is rewritten as 2\sqrt{n}/\sqrt{N}; the implicit inequality used is that n/N \le \sqrt{n/N}, which should be stated explicitly for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the regret bounds are self-contained decompositions, and model elasticity is a defined quantity rather than a fitted or predicted input.

full rationale

The paper's derivation chain is self-contained in the sense relevant to circularity. Model elasticity in Definition 4.4 is defined directly as sup_{f,a} E[(f(x,z*,a)-f(x,g~(x),a))^2], and this same quantity appears as a term in Theorems 4.5 and 4.6. That is a definitional decomposition of the regret bound, not a fitted parameter being relabeled as a prediction: the paper never estimates E_F^P(g~) from data and then claims to predict the regret from that fit. The remaining components of Theorems 4.5 and 4.6 (critical radii, fluctuation terms, and the IGW epoch summation) are derived from the ERM analysis and Lemma B.4. In the MAR section, Theorem 5.2 is a standard localized ERM/orthogonal-learning bound built on Assumption 5.1, and Theorem 5.3 combines it with Theorem 4.5 and Lemma B.4 by Lipschitz composition; no step defines the calibrated estimator in terms of the regret it is supposed to explain. Self-citations to Simchi-Levi and Xu (2021), Qian et al. (2024), and Hu et al. (2025) are used for standard realizability and IGW-to-regret conversion results; Lemma B.4 is stated explicitly and comes from an independent published source, so the central regret conversion does not reduce to an unverified self-citation. The skeptical observation that the advertised rate T^{(d-2)/(d+2)} in the interpretation paragraph is not what the displayed Theorem 5.3 bound yields, and that taking the square root of the Theorem 5.2 bound changes the epoch exponent, is an arithmetic/correctness defect in the stated rates rather than a circular reduction of the result to its inputs. Accordingly, no circular step is identified and the circularity score is 0.

Assumptions & free parameters 7 free parameters · 6 assumptions · 2 invented entities

The central results depend on a long list of known constants (delta_0, omega_0, tau, epsilon_0, lambda, L_Z, d) and on structural realizability assumptions on F, G, and T. The only genuinely new construct is model elasticity, which is definitional. The paper is transparent about these assumptions, but any practical deployment would need to estimate or tune most of them. The calibration procedure itself does not introduce a new physical or empirical entity, only a new analytical quantity.

free parameters (7)
  • delta_0 = assumed known
    Assumption 5.1(A4): the decision-maker knows an upper bound on ||g~ - g*||_2. This prior error bound controls the localized class G_delta0 and appears in the calibrated regret bound of Theorem 5.3. If it is misspecified, the guarantees degrade or fail.
  • omega_0 = assumed known
    Assumption 5.1(A2): the second moment bound of the covariate noise eta. It creates the irreducible linear term omega_0 * T in the regret of Theorem 5.3. The paper argues this is unavoidable, but the value must be known to the algorithm and appears linearly in the bound.
  • tau = assumed known
    Assumption 5.1(A2): bound on the covariate noise eta. It appears in the calibration error bound, the definition of phi_{1+tau}, and in the regret terms involving log factors.
  • epsilon_0 = assumed known
    Assumption 5.1(B): lower bound on the missingness probability e*(x) >= epsilon_0. It appears as a 1/epsilon_0^2 factor in the calibration bound and in the regret, and is required for the inverse-probability weighting to be stable.
  • lambda = assumed known
    Bounded reward noise in [-lambda, lambda]. Appears throughout the oracle inequality and regret bounds.
  • L_Z = assumed known
    Assumption 5.1(C): Lipschitz constant of all reward functions with respect to the covariate. It scales the calibration error contribution and the irreducible omega_0 * T term.
  • d = assumed known
    Covering number exponent of the class G in Assumption 5.1(A3). Determines the rate of the localized critical radius q_{delta_0,n} and appears in the regret bound.
assumptions (6)
  • domain assumption Realizability: the true reward function f* belongs to a known convex and star-shaped class F (Assumption 3.1).
    This is a standard realizability assumption in contextual bandits, but it means the analysis does not cover misspecified reward models.
  • domain assumption Missing at random: the missingness indicator b depends only on the observed context x, with e*(x) in a known class T (Assumption 5.1(B)).
    This is the structural assumption that enables calibration. The paper explicitly leaves MNAR calibration for future work.
  • domain assumption The covariate is additive: z* = g*(x) + eta with g* in G and bounded zero-mean noise (Assumption 5.1(A2)).
    Used throughout Section 5 to expand the loss into ||g - g*||^2 + E[eta^2]. If the covariate model is mis-specified, the calibration bound does not hold.
  • domain assumption The pre-trained model is in the same class G and close to g*: g~ in G and ||g~ - g*||_2 <= delta_0 (Assumption 5.1(A4)).
    This is the load-bearing premise for the calibration regret guarantee. If the pre-trained model is outside G or delta_0 is unknown, the localized class construction fails.
  • domain assumption Independent and identically distributed data in each epoch, including the imputed covariates and the policy distribution (Section 4.2).
    The oracle inequalities and the IGW lemma rely on the data being i.i.d. within an epoch. In Algorithm 3 the calibrated imputer changes between epochs, and the paper assumes this does not break the i.i.d. property, which is plausible but not fully proven.
  • standard math Standard empirical process tools: local Rademacher complexity bounds, peeling lemma, Talagrand's inequality (Appendix A).
    These are standard mathematical tools used to derive the oracle inequalities.
invented entities (2)
  • Model elasticity E_F^P(g~)
    purpose: Summarizes the sensitivity of the reward function class to imputation error in a single number that multiplies T in the uncalibrated regret bound.
    It is a definitional object: sup over f in F and a in A of the expected squared difference between f(x, z*, a) and f(x, g~(x), a). The paper provides no method to estimate it from data and no falsifiable prediction attached to it. Its usefulness is as an analytic quantity, not an empirically measurable constant.
  • G_delta0 centered class
    purpose: Restricts the calibration regression to functions that are close to the pre-trained model g~, enabling the localized Rademacher bound q_{delta_0,n}.
    This is a construction that depends on the prior bound delta_0. It is an auxiliary device for the proof, not a separate empirical entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pre-Trained AI Model Assisted Online Decision-Making under Missing Covariates: A Theoretical Perspective." pith.science (2026). https://pith.science/paper/5B4JZYS7

@misc{pith2026250707852,
  author       = {Pith},
  title        = {Pith review of: Pre-Trained AI Model Assisted Online Decision-Making under Missing Covariates: A Theoretical Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5B4JZYS7}},
  note         = {Machine review of arXiv:2507.07852}
}
read the original abstract

We study a sequential contextual decision-making problem in which certain covariates are missing but can be imputed using a pre-trained AI model. From a theoretical perspective, we analyze how the presence of such a model influences the regret of the decision-making process. We introduce a novel notion called "model elasticity", which quantifies the sensitivity of the reward function to the discrepancy between the true covariate and its imputed counterpart. This concept provides a unified way to characterize the regret incurred due to model imputation, regardless of the underlying missingness mechanism. More surprisingly, we show that under the missing at random (MAR) setting, it is possible to sequentially calibrate the pre-trained model using tools from orthogonal statistical learning and doubly robust regression. This calibration significantly improves the quality of the imputed covariates, leading to much better regret guarantees. Our analysis highlights the practical value of having an accurate pre-trained model in sequential decision-making tasks and suggests that model elasticity may serve as a fundamental metric for understanding and improving the integration of pre-trained models in a wide range of data-driven decision-making problems.

Figures

Figures reproduced from arXiv: 2507.07852 by the authors.

Figure 1
Figure 1. Pre-trained model assisted sequential decision-making with missing covariates. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 17 canonical work pages

  1. [2]

    On one hand, we have σ2(H(t)) = sup ||f−f∗||2≤t PD ( (f (x,z,a)−f∗(x,z,a))4 ) ≤ 4PD ( (f (x,z,a)−f∗(x,z,a))2 ) ≤ 4t2

    to get with probability at least 1−δ, ZN (t)≤ 2E[ZN (t)] + √ σ2(H(t)) √ 2 log(1/δ) N + 16 log(1/δ) N . On one hand, we have σ2(H(t)) = sup ||f−f∗||2≤t PD ( (f (x,z,a)−f∗(x,z,a))4 ) ≤ 4PD ( (f (x,z,a)−f∗(x,z,a))2 ) ≤ 4t2. E[ZN (t)] = E [ sup ||f∗−f||2≤t ⏐⏐⏐||f∗−f||2 N−||f−f∗||2 2 ⏐⏐⏐ ] ≤ 2E [ sup ||f∗−f||2≤t ⏐⏐⏐⏐⏐ 1 N N∑ i=1 εi (f∗(xi, ˜zi,ai)−f (xi, ˜zi,a...

    work page 2021

  2. [4]

    Automatic Doubly Robust Forests

    URL https://arxiv.org/abs/2412.07184. Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whit- ney Newey, and James Robins. Double/debiased machine learning for treatment and causal parameters. Technical report,

    work page Pith review arXiv

  3. [7]

    Statistically valid post-deployment monitoring should be standard for ai-based digital health

    Pavel Dolin, Weizhi Li, Gautam Dasarathy, and Visar Berisha. Statistically valid post-deployment monitoring should be standard for ai-based digital health. arXiv preprint arXiv:2506.05701 ,

    arXiv

  4. [8]

    Imputation Strategies for Rightcensored Wages in Longitudinal Datasets

    URL https://arxiv.org/abs/2502.12967. James D Dziura, Lori A Post, Qing Zhao, Zhixuan Fu, and Peter Peduzzi. Strategies for dealing with missing data in clinical trials: from design to analysis. The Yale journal of biology and medicine, 86(3):343,

    work page Pith review arXiv

  5. [12]

    Linear Bandits with Partially Observable Features

    Wonyoung Kim, Sungwoo Park, Garud Iyengar, Assaf Zeevi, and Min-hwan Oh. Linear bandits with partially observable features. arXiv preprint arXiv:2502.06142 ,

    work page Pith review arXiv

  6. [13]

    Concentration around the mean for maxima of empirical processes

    T Klein and E Rio. Concentration around the mean for maxima of empirical processes. arXiv preprint math/0506594,

    work page Pith review arXiv

  7. [15]

    LLM aided semi-supervision for Extractive Dialog Summarization

    Nishant Mishra, Gaurav Sahu, Iacer Calixto, Ameen Abu-Hanna, and Issam H Laradji. Llm aided semi-supervision for extractive dialog summarization. arXiv preprint arXiv:2311.11462 ,

    work page Pith review arXiv

  8. [16]

    Synthetic data generation using large language models: Advances in text and code

    Mihai Nadas, Laura Diosan, and Andreea Tomescu. Synthetic data generation using large language models: Advances in text and code. arXiv preprint arXiv:2503.14023 ,

    arXiv

Show all 28 references
  1. [17]

    Offline oracle-efficient learning for contextual mdps via layerwise exploration-exploitation tradeoff

    Jian Qian, Haichen Hu, and David Simchi-Levi. Offline oracle-efficient learning for contextual mdps via layerwise exploration-exploitation tradeoff. arXiv preprint arXiv:2405.17796 ,

    arXiv

  2. [19]

    Does synthetic data generation of llms help clinical text mining? arXiv preprint arXiv:2303.04360 ,

    Ruixiang Tang, Xiaotian Han, Xiaoqian Jiang, and Xia Hu. Does synthetic data generation of llms help clinical text mining? arXiv preprint arXiv:2303.04360 ,

    arXiv

  3. [20]

    Handling censoring and censored data in survival analysis: a standalone systematic literature review

    Anthony Joe Turkson, Francis Ayiah-Mensah, and Vivian Nimoh. Handling censoring and censored data in survival analysis: a standalone systematic literature review. International journal of mathematics and mathematical sciences , 2021(1):9307475,

    2021

  4. [21]

    Zikun Ye, Hema Yoganarasimhan, and Yufeng Zheng

    URL https: //arxiv.org/abs/2412.09364. Zikun Ye, Hema Yoganarasimhan, and Yufeng Zheng. Lola: Llm-assisted online learning algorithm for content experiments. Marketing Science,

    arXiv

  5. [22]

    The application of large language models in recommendation systems

    Peiyang Yu, Zeqiu Xu, Jiani Wang, and Xiaochuan Xu. The application of large language models in recommendation systems. arXiv preprint arXiv:2501.02178 ,

    arXiv

  6. [23]

    Zhang, Tiffany Tianhui Cai, Hongseok Namkoong, and Daniel Russo

    Kelly W. Zhang, Tiffany Tianhui Cai, Hongseok Namkoong, and Daniel Russo. Contextual thomp- son sampling via generation of missing data, 2025a. URL https://arxiv.org/abs/2502.07064. Weizhou Zhang, Chen Li, Hanzhang Qin, Yunbei Xu, and Ruihao Zhu. Thompson sampling for repeated...

  7. [24]

    Consider a countable, σ-uniformly bounded function class G such that E[g(X)] = 0 for all g. Then for any τ >0, we have with probability at least 1−δ, sup g∈G /radicaltp/radicalvertex/radicalvertex√1 N N∑ i=1 g(Xi)2 =||PN||G≤ (1 +τ )E[||PN||G] + √ σ2(G)· √ 2 log(1/δ) N + (3 + 1...

    2024

  8. [25]

    There is a function Q(r,t) increasing in the first argument r and Q(2r,t)≤ 2Q(r,t) for all r≥s, where s is some scalar

    Given a function classF , with norm||·|| and some em- pirical process{Vn(f ) :f∈F} , consider ar-localized supremum of the formZN (r) := supf∈F ,||f||≤rVN (f ). There is a function Q(r,t) increasing in the first argument r and Q(2r,t)≤ 2Q(r,t) for all r≥s, where s is some scal...

    2021

  9. [27]

    data D ={(xi,ai,ri)}n i=1 where E[ri|xi,ai] = f∗(xi,ai)

    Assume that we are given an offline regression oracle RegOff and i.i.d. data D ={(xi,ai,ri)}n i=1 where E[ri|xi,ai] = f∗(xi,ai). With probability at 33 least 1−δ, it returns ˆf :X×A→ R such that Ex,a [( ˆf (x,a)−f∗(x,a) )2] ≤ Estδ(n) for some number Estδ(n). Then, define epoch...

    1966

  10. [28]

    Applying Dudley’s integral bound, we have Rn(t,Gδ0) ≲ inf α>0 { 4α + 12√n ∫t α √ logN (ε,Gδ0,L2(P))dε }

    ≲ (δ0 ϵ )d. Applying Dudley’s integral bound, we have Rn(t,Gδ0) ≲ inf α>0 { 4α + 12√n ∫t α √ logN (ε,Gδ0,L2(P))dε } . Using the bound logN (ε,Gδ0,L2(P)) ≲ (δ0 ε )d , we get √ logN (ε,Gδ0,L2(P)) ≲ (δ0 ε )d/2 . Substituting into the Dudley integral (Koltchinskii, 2011): ∫δ0 α (δ...

    2011

  11. [1996]

    Contextual online decision making with infinite-dimensional functional regression

    Haichen Hu, Rui Ai, Stephen Bates, and David Simchi-Levi. Contextual online decision making with infinite-dimensional functional regression. arXiv preprint arXiv:2501.18359 ,

    arXiv

  12. [2005]

    Oracle inequalities in empirical risk minimization and sparse recovery prob- lems: ´Ecole D’ ´Et´ e de Probabilit´ es de Saint-Flour XXXVIII-2008, volume

    Vladimir Koltchinskii. Oracle inequalities in empirical risk minimization and sparse recovery prob- lems: ´Ecole D’ ´Et´ e de Probabilit´ es de Saint-Flour XXXVIII-2008, volume

    2008

  13. [2017]

    Automatic debiased machine learning for covariate shifts

    Victor Chernozhukov, Michael Newey, Whitney K Newey, Rahul Singh, and Vasilis Srygkanis. Automatic debiased machine learning for covariate shifts. arXiv preprint arXiv:2307.04527 ,

  14. [2019]

    Byoungwook Jang, Julia Nepper, Marc Chevrette, Jo Handelsman, and Alfred O Hero

    URL https://arxiv.org/abs/ 1901.07031. Byoungwook Jang, Julia Nepper, Marc Chevrette, Jo Handelsman, and Alfred O Hero. High dimensional stochastic linear contextual bandit with missing covariates. In 2022 IEEE 32nd International Workshop on Machine Learning for Signal Process...

    1901 arXiv

  15. [2020]

    Dynamic pricing in the linear valuation model using shape constraints

    Daniele Bracale, Moulinath Banerjee, Yuekai Sun, Kevin Stoll, and Salam Turki. Dynamic pricing in the linear valuation model using shape constraints. arXiv preprint arXiv:2502.05776 ,

    arXiv

  16. [2021]

    Contextual bandit with missing rewards

    Djallel Bouneffouf, Sohini Upadhyay, and Yasaman Khazaeni. Contextual bandit with missing rewards. arXiv preprint arXiv:2007.06368 ,

    2007 arXiv

  17. [2022]

    missing at ran- dom

    URL https://arxiv.org/abs/2204.01959. Shaun Seaman, John Galati, Dan Jackson, and John Carlin. What is meant by “missing at ran- dom”?

    arXiv

  18. [2023]

    org/abs/1901.09036

    URL https://arxiv. org/abs/1901.09036. Dylan J Foster, Claudio Gentile, Mehryar Mohri, and Julian Zimmert. Adapting to misspecification in contextual bandits. Advances in Neural Information Processing Systems , 33:11478–11489,

    1901 arXiv

  19. [2024]

    Natalie Collina, Ira Globus-Harris, Surbhi Goel, Varun Gupta, Aaron Roth, and Mirah Shi

    URL https://arxiv.org/abs/2411.19791. Natalie Collina, Ira Globus-Harris, Surbhi Goel, Varun Gupta, Aaron Roth, and Mirah Shi. Collaborative prediction: Tractable information aggregation via agreement. arXiv preprint arXiv:2504.06075,

    arXiv

  20. [2025]

    Junyu Cao, Ruijiang Gao, and Esmaeil Keyvanshokooh

    URL https://arxiv.org/abs/2405.19466. Junyu Cao, Ruijiang Gao, and Esmaeil Keyvanshokooh. Hr-bandit: Human-ai collaborated linear recourse bandit. arXiv preprint arXiv:2410.14640 ,

    arXiv

Pith tools

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