Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Contextual Online Decision Making with Infinite-Dimensional Functional Regression

T0 review · 3 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A single algorithm learns whole outcome distributions online and ties regret to spectral decay.

desk verdict Novel framework with a genuinely appealing spectral-decay connection, but the main regret theorem is not proven due to two load-bearing technical errors and a bandit-reduction mismatch. read the letter →

arxiv 2501.18359 v1 pith:ST7FWFYK submitted 2025-01-30 stat.ML cs.LG

Haichen Hu , Rui Ai , Stephen Bates , David Simchi-Levi This is my paper · ORCID
classification stat.MLcs.LG MSC 62G0868Q3247B0668W27
keywords contextualbanditsfunctionalregressioncumulativedistributionfunctionsonlinedecision-makingutilityregreteigenvaluedecayintegraloperatorsinversegapweighting
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 tries to establish that every contextual online decision problem—bandit reward maximization, sequential hypothesis testing, risk control, quantile calibration—can be handled by one framework that learns the whole conditional outcome distribution instead of a particular statistic. The proposed functional regression oracle estimates the unknown coefficient function $\theta^*$ in an infinite-dimensional CDF basis, and the paper proves that the decay rate of the design integral operator's eigenvalue sequence controls both the regression error and the final utility regret. Under a $\gamma$-dominating eigendecay condition, the regret after $T$ rounds is $\tilde{O}\bigl(T^{(3\gamma+2)/(2(\gamma+2))}\bigr)$, which recovers the optimal $\tilde{O}(\sqrt{T})$ contextual-bandit rate when $\gamma=0$ and still gives sublinear $\tilde{O}(T^{5/6})$ regret when no decay-rate information is available. If these bounds hold, practitioners would not need separate algorithms for different objectives; they could plug the relevant Lipschitz functional into the same CDF-learning oracle.

What carries the argument

The central object is the design integral operator $U_D$ with kernel $K_{x,a}(w,r)=\int_S \phi(x,a,w,s)\phi(x,a,r,s)\,dm(s)$; the algorithm retains only eigenvalues above the truncation level $\varepsilon_*=n^{-2/(\gamma+2)}$. The pseudo-inverse of this truncated operator turns noisy indicator observations into the solution of a finite-rank least-squares problem for the coefficient function, and the $\gamma$-dominating eigendecay assumption lets the analysis bound the truncated determinant $\sum_{i\le N_\varepsilon}\log(1+\lambda_i(U_D))$ by $O(\varepsilon^{-\gamma}\log(1+\varepsilon^\gamma n))$. A uniform lower bound on $K_{x,a}$ is then used to pass from the empirical design norm to the population $L^2$ CDF error, and a batched inverse-gap-weighting schedule converts that error into epoch-wise regret with only $O(\log T)$ oracle calls.

What would settle it

Construct an instance satisfying the realizability and Lipschitz assumptions with localized CDF basis functions $\phi(x,a,w,s)$ supported on tiny disjoint intervals, so that $K_{x,a}(w,r)=0$ for neighboring $w\neq r$ while polynomial eigenvalue decay still holds; in that instance the constant in Theorem 3.10 blows up as $\eta\to 0$, and a simulation of Algorithm 2 should fail to reproduce the claimed $\tilde{O}(T^{(3\gamma+2)/(2(\gamma+2))})$ regret rate.

Watch

Extended reading notes

Core claim

The core discovery is that the spectral decay of the design integral operator $U_D=\sum_{j=1}^n L_{x_j,a_j}$, defined through the kernels $K_{x,a}(w,r)=\int_S \phi(x,a,w,s)\phi(x,a,r,s)\,dm(s)$, is the quantity that governs learnability of the infinite-dimensional coefficient function $\theta^*$. The algorithm computes the spectral decomposition of $U_D$, truncates it at $\varepsilon_* = n^{-2/(\gamma+2)}$, applies the pseudo-inverse of the truncated operator to the empirical indicator data, solves the induced least-squares problem in the truncated span, and projects the solution onto the set of nonnegative, unit-mass, bounded coefficient functions. The oracle inequality bounds the design-norm error by a truncated functional determinant, and a concentration argument over the kernel family converts that into an expected squared $L^2$ error for the estimated CDF. Feeding these estimates into a batched inverse-gap-weighting policy yields the utility regret bound of Theorem 4.2 and shows that the method recovers $\tilde{O}(\sqrt{T})$ for finite-dimensional or exponentially decaying spectra.

Load-bearing premise

The load-bearing premise is that for every context, action, and pair of basis indices the overlap kernel $K_{x,a}(w,r)$ is bounded below by a fixed positive number $\eta$, which excludes localized basis functions with disjoint supports; if this fails, the ratio step that converts empirical operator concentration into the population error bound collapses.

Editorial extensions

If this is right

  • For finite-dimensional or exponentially decaying spectra, the general framework matches the established optimal contextual-bandit regret $\tilde{O}(\sqrt{T})$ up to logarithmic factors.
  • A single implementation can switch from reward maximization to risk control or hypothesis testing just by changing the Lipschitz functional $T$ and the basis family, without redesigning the algorithm.
  • With no prior knowledge of the eigendecay order, the algorithm still guarantees sublinear $\tilde{O}(T^{5/6})$ regret, so it can be deployed before the spectrum is understood.
  • The algorithm calls the functional regression oracle only $O(\log T)$ times, amortizing the expensive spectral computation over exponentially growing epochs.
  • The numerical degenerate-kernel method gives a concrete way to compute the eigenvalue sequence that the regret bound depends on, making the framework implementable.

Reading between the lines

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

  • The paper leaves open whether a weaker condition—such as control of the empirical-to-population operator ratio on a low-effective-dimensional subspace rather than pointwise on $\Omega\times\Omega$—could replace Assumption 3.9 and cover wavelet or spline basis families with local support.
  • The authors themselves suspect the $\gamma$-exponent is not minimax optimal; a sharper lower bound might be expressed through the truncated determinant or effective dimension rather than the single exponent $\gamma$.
  • The fixed-design oracle inequality suggests the regression step is insensitive to non-i.i.d. contexts, so testing the same oracle inside an adversarial contextual-bandit algorithm could extend the framework beyond stochastic contexts.
  • Because the eigenvalue sequence can be computed numerically, the predicted regret exponent is directly testable on synthetic basis families by measuring actual regret for several values of $\gamma$.
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 proposes a unified framework for contextual online decision making in which the decision objective is a known Lipschitz functional of the outcome CDF F*(x,a,·). It models each CDF as an integral of an unknown coefficient function θ* against a context-dependent family of CDF basis functions, and it introduces a functional regression oracle FuncReg based on spectral truncation and pseudo-inversion of a design integral operator. The oracle is embedded in a batched inverse-gap-weighting policy. The main theoretical claims are a fixed-design oracle inequality for ||θ_D−θ*||_{U_D} (Theorem 3.6), a data-independent refinement (Lemma 3.7 and Corollary 3.8), a random-design L2 error bound for the estimated CDF (Theorem 3.10), and a utility regret bound of order T^{(3γ+2)/(2(γ+2))} under a γ-dominating eigenvalue decay assumption (Theorem 4.2), with the √T rate recovered in exponential/finite-dimensional limiting cases. A degenerate-kernel numerical method for computing the required eigenvalues is presented in Appendix E.

Significance. If the results were correct, the paper would contribute a genuinely useful bridge between spectral properties of design integral operators and regret in contextual decision making, and it would unify bandits, hypothesis testing, and risk control under one algorithm with O(log T) oracle calls. The paper also makes a concrete algorithmic proposal and includes a numerical scheme, which are assets. However, two load-bearing proof steps are invalid as written: the minimax argument in Lemma 3.7 and the pointwise-to-quadratic-form transfer in Theorem 3.10. Because Theorems 4.1 and 4.2 inherit these steps, the central regret bound is not established by the manuscript.

major comments (3)
  1. [Appendix C, proof of Lemma 3.7] The proof of Lemma 3.7 applies Sion's minimax theorem with the domain {α ∈ S_k : ||α||=1}, but this unit sphere is not convex, so the theorem cannot be used. The subsequent swap of max over L and min over α is therefore not justified. Moreover, the displayed inequality λ_k(U_D) ≤ n max_{x,a} λ_k(L_{x,a}) is not a consequence of the assumptions as argued; for sums of positive operators, the k-th eigenvalue of a sum can exceed n times the largest k-th eigenvalue of the summands. Since Lemma 3.7 supplies the data-independent bound Eδ(n) used in Corollary 3.8 and in Theorem 3.10, the derivation of the claimed regret rate is unsupported. Additionally, in the line preceding (C.6), the summation changes from Σ τ_i^γ to Σ τ_i; only the former is controlled by Assumption 2.8, so the displayed bound does not follow even from the earlier steps.
  2. [Appendix C, proof of Theorem 3.10] The central transfer step in Theorem 3.10 is invalid. The proof establishes, with high probability, sup_{w,r} |Q(K_{w,r})/Q_n(K_{w,r}) − 1| ≤ δ, and then replaces Q(K(w,r)) by (Q(K(w,r))/Q_n(K(w,r))) Q_n(K(w,r)) inside the double integral defining ⟨θ, U* θ⟩. But pointwise closeness of the ratio does not imply the quadratic-form inequality ⟨θ, U* θ⟩ ≤ (1+δ)⟨θ, U_D θ⟩ for every θ, because θ(w)θ(r) is not pointwise nonnegative and the ratio is not a constant factor dominating Q_n. A two-point example with the CDF basis φ_a(s)=1{s≥a} shows the failure: with Q_n = [[0.8,0.7],[0.7,0.7]] and Q = [[0.8,0.69],[0.69,0.69]], one has sup |Q/Q_n − 1| ≤ 0.0143, yet for θ=(1,−1), ⟨θ,Qθ⟩=0.11 while (1+0.0143)⟨θ,Q_nθ⟩≈0.1014. The kernels can be smoothed to satisfy Assumption 2.2, so the issue is not merely an artifact of the discontinuous basis. Since Theorem 4.1 and Theorem 4.2 rely on this empirical-to-population transfer, the regret bound is not established.
  3. [Section 3.2, Assumption 3.9] Assumption 3.9 requires K_{x,a}(w,r) ≥ η > 0 uniformly over all w,r ∈ Ω. This excludes many natural basis families, including localized indicator or bump bases with disjoint supports, and it even fails at the upper endpoint for the simple CDF basis φ_a(s)=1{s≥a} used in the paper's own motivating examples. The uniform lower bound η appears in the constant of Theorem 3.10 and hence in Theorem 4.2, so the scope of the main result is much narrower than the 'universal framework' claimed in the abstract and introduction. The authors should either restrict the claims accordingly or supply a nontrivial family of CDF bases satisfying Assumption 3.9 along with a discussion of why those bases are representative.
minor comments (4)
  1. [Algorithm 1] Algorithm 1 states ε* = n^{−γ/(γ+2)}, while Definition 3.3, Theorem 3.6, and the surrounding text use ε* = n^{−2/(γ+2)}; the two choices lead to different bias terms and should be reconciled.
  2. [Assumption 2.8 and abstract] The abstract and Section 1.2 state that the √T rate is obtained by setting γ=0, but Assumption 2.8 is stated for 0<γ≤1. The limiting case γ=0 should be defined explicitly, for example by viewing it as the limit of the bound as γ→0 rather than as an allowed value of the parameter.
  3. [Section 2.1, regret definition] In the definition of regret, the notation T(F^*_{x,a,s}) mixes the function F^* with one of its arguments s; the functional T should be applied consistently to the CDF F^*(x,a,·).
  4. [Section 2.2, footnote] The footnote after Assumption 2.8 says that under mild conditions the assumption can be proved and cites Carrijo and Jordão, but no proof or precise condition is given in the manuscript. Either state the conditions and prove the claim, or remove the footnote to avoid suggesting a result the paper does not contain.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the regret bound is derived from explicit spectral assumptions via self-contained oracle inequalities, and the cited IGW reduction lemma is independent support.

full rationale

The paper's central regret bound (Theorem 4.2) is derived, not fitted: Theorem 3.6 bounds fixed-design regression error in the U_D norm with a martingale argument; Lemma 3.7 converts the data-dependent eigenvalue sum into the gamma-dependent rate using Assumption 2.8; Theorem 3.10 transfers the oracle bound to population L2 CDF error under i.i.d. sampling; and Theorem 4.1 converts L2 error to utility error via Lipschitzness of T. The truncation threshold epsilon* = n^{-2/(gamma+2)} is chosen by balancing the n*epsilon*M bias term against the variance term, which is a standard bias-variance tradeoff rather than a fit to the claimed rate. The parameters gamma, M, s0, eta are assumed inputs and are not selected to match the theorem statement. The only self-citations are (i) inspiration from Simchi-Levi and Xu [2021] and Qian et al. [2024] for the batched IGW policy, and (ii) Lemma D.1 quoted from Simchi-Levi and Xu for the regret-to-oracle reduction; both are independent, parameter-free reduction results whose assumptions do not include the target regret rate, so they do not make the argument circular. The paper explicitly leaves the minimax lower bound as an open question, which is an honest limitation rather than a circular closure. The skeptical counterexample to Theorem 3.10 concerns the validity of a pointwise-to-quadratic-form proof step; that is a correctness gap, not a definitional or fitted equivalence, and is not counted as circularity here.

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

The central claim rests on the mixture model for CDFs, the Lipschitz and lower-bound assumptions on the basis and kernel, and the γ-dominating eigendecay with convexity. No new physical or mathematical entities are postulated; the design integral operator is a construction, not an invented entity. The key free parameters are the assumed constants γ, M, s0, η, and L0, none of which are fitted to data.

free parameters (5)
  • γ (eigendecay exponent)
    Assumed in Assumption 2.8; controls truncation threshold ε* and the regret rate; when unknown, set to 1 in Corollary 4.3, yielding T^(5/6).
  • M (L2 norm bound for θ*)
    Assumption 2.1 bounds ||θ*||_{L2} ≤ M; appears in the bias term n^{γ/(γ+2)} M and the final regret bound.
  • s0 (sum of τ_i^γ)
    Assumption 2.8 states Σ τ_i^γ ≤ s0; appears in the regret bound via Eδ(n); not needed for algorithm implementation.
  • η (uniform lower bound of kernel K)
    Assumption 3.9 requires K_{x,a}(w,r) ≥ η > 0; used to convert empirical operator concentration into a ratio bound; no example basis family is shown to satisfy this.
  • L0 (Lipschitz constant of basis)
    Assumption 2.2; appears in the uniform concentration constant C(d,L0,δ,A,η).
assumptions (5)
  • domain assumption F*(x,a,s) = ∫_Ω θ*(w) φ(x,a,w,s) dν(w) with θ* ≥ 0, ∫θ* dν = 1, and ||θ*|| ≤ M
    Assumption 2.1; defines the functional regression model and the constraint set C.
  • domain assumption Basis functions are L0-Lipschitz in w: |φ(x,a,w,s) - φ(x,a,r,s)| ≤ L0 ||w-r||
    Assumption 2.2; needed for entropy bounds and the uniform concentration argument in Theorem 3.10.
  • ad hoc to paper The set {L_{x,a}} is convex and has a γ-dominating eigenvalue sequence with Σ τ_i^γ ≤ s0
    Assumption 2.8; convexity is used in Lemma 3.7 via Sion's minimax theorem, and the eigendecay assumption drives the regret rate.
  • ad hoc to paper K_{x,a}(w,r) = ∫_S φ(x,a,w,s) φ(x,a,r,s) dm(s) ≥ η > 0 uniformly
    Assumption 3.9; strong and not instantiated; enables the empirical-to-population operator comparison in Theorem 3.10.
  • domain assumption The utility functional T is L-Lipschitz with respect to ||·||_{L2(S,m)}
    Assumption 2.5; converts CDF estimation error into utility error in Theorem 4.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Contextual Online Decision Making with Infinite-Dimensional Functional Regression." pith.science (2026). https://pith.science/paper/ST7FWFYK

@misc{pith2026250118359,
  author       = {Pith},
  title        = {Pith review of: Contextual Online Decision Making with Infinite-Dimensional Functional Regression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ST7FWFYK}},
  note         = {Machine review of arXiv:2501.18359}
}
abstract

Contextual sequential decision-making problems play a crucial role in machine learning, encompassing a wide range of downstream applications such as bandits, sequential hypothesis testing and online risk control. These applications often require different statistical measures, including expectation, variance and quantiles. In this paper, we provide a universal admissible algorithm framework for dealing with all kinds of contextual online decision-making problems that directly learns the whole underlying unknown distribution instead of focusing on individual statistics. This is much more difficult because the dimension of the regression is uncountably infinite, and any existing linear contextual bandits algorithm will result in infinite regret. To overcome this issue, we propose an efficient infinite-dimensional functional regression oracle for contextual cumulative distribution functions (CDFs), where each data point is modeled as a combination of context-dependent CDF basis functions. Our analysis reveals that the decay rate of the eigenvalue sequence of the design integral operator governs the regression error rate and, consequently, the utility regret rate. Specifically, when the eigenvalue sequence exhibits a polynomial decay of order $\frac{1}{\gamma}\ge 1$, the utility regret is bounded by $\tilde{\mathcal{O}}\Big(T^{\frac{3\gamma+2}{2(\gamma+2)}}\Big)$. By setting $\gamma=0$, this recovers the existing optimal regret rate for contextual bandits with finite-dimensional regression and is optimal under a stronger exponential decay assumption. Additionally, we provide a numerical method to compute the eigenvalue sequence of the integral operator, enabling the practical implementation of our framework.

Figures

Figures reproduced from arXiv: 2501.18359 by the authors.

Figure 1
Figure 1. Data D drives the functional regression and decision making. Example 2.3 (Contextual Bandits). Define the functional as T : F → R S xdF(x), which is a known functional in terms of F. Given any context x, F ∗ (x, a, s) is the conditional reward distribution of arm a given context x. Consequently, we recover the contextual bandit problem. ♮ δa is the dirac function of a, i.e., δa(x) = 1 only if x=a, othwise, δa(x) = 0… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

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

    cs.LG 2025-07 unverdicted novelty 6.0 of 10

    The paper introduces model elasticity to bound the regret of contextual bandits with AI-imputed missing covariates, and shows that MAR-based calibration removes the dominant linear regret term.

Reference graph

Works this paper leans on

18 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [1]

    Learning smooth functions in high dimensions: from sparse polynomials to deep neural networks

    Ben Adcock, Simone Brugiapaglia, Nick Dexter, and Sebastian Moraga. Learning smooth functions in high dimensions: from sparse polynomials to deep neural networks. arXiv preprint arXiv:2404.03761 ,

    arXiv

  2. [2]

    For any self-adjoint positive Hilbert-Schmidt integral operator U , it holds that λj(U ) = sj(U ) for ∀ j. Proof. We denote the integral kernel of U as K. Because U is self-adjoint positive Hilbert-Schmidt integral operator, then for ∀ θ, U (θ) = ∞X i=1 λi(U ) ⟨θ, ei⟩ ei. Therefore, U (U (θ)) = ∞X i=1 λ2 i ⟨θ, ei⟩ ei. Because U ∗ = U , we finish the proof...

    work page 2012

  3. [15]

    A Useful Math Theorems Theorem A.1 (Theorem 8.2 in Gohberg et al. [2012]). Let positive operator U (θ)(w) = R Ω K(w, r)θ(r)dν(r) be self-adjoint. If the kernel K(w, r) is continuous and satisfies the Lipschitz condition, |K(w, r1) − K(w, r2)| ≤C||s2 − s1||, then ∞X j=1 λj(U ) < ∞ i.e., U is a trace class operator. Theorem A.2 (Theorem 8.1 in Gohberg et al...

    work page 2012

  4. [16]

    log( 1 δ ) + 1 2 NεX i=1 log(1 + λi) # . Combined with Equation (C.5), it indicates that with probability 1 − δ, ||θD,ε − θ∗||UD ≤ vuut2(1 + 1/(nε))

    Therefore, M k n≥k≥0 is a non-negative supermartingale with a uniform expectation upper bound. On the other hand, we can directly calculate M n. We use wi to denote ⟨Wn, ei⟩ for simplicity. Then, it holds that Mn(β) = exp ⟨β, Wn⟩ −1 2 ||β||2 UD = exp ( NεX i=1 wiζi − 1 2 NεX i=1 λiζ 2 i ) , and M n = Z RNε exp ( NεX i=1 wiζi − 1 2 NεX i=1 λiζ 2 i + ζ 2 i ...

    work page 1994

  5. [17]

    Define piecewise polynomial space associated with Pr S r h = n u : [0, 1] → R : u Ik ∈ Pr, 1 ≤ k ≤ n o

    Denote Nh = nr. Define piecewise polynomial space associated with Pr S r h = n u : [0, 1] → R : u Ik ∈ Pr, 1 ≤ k ≤ n o . Find the Gauss point set (zeros set) of Lr(t) in [0 , 1], Br = {y1 · · ·, yr}. Let A = ∪n k=1fk(Br) = {ωi,k = fk(yi) : 1 ∈ [r], k∈ [n]} and li(x) be the Lagrange polynomials of degree r − 1 with respect to y1 · · ·, yr such that li(yj) ...

    work page 2012

  6. [1958]

    Risk-Aware Algorithms for Adversarial Contextual Bandits

    Wen Sun, Debadeepta Dey, and Ashish Kapoor. Risk-aware algorithms for adversarial contextual bandits. arXiv preprint arXiv:1610.05129 ,

    work page Pith review arXiv

  7. [1972]

    Peer Expectation in Robust Forecast Aggregation: The Possibility/Impossibility

    13 Yuqing Kong. Peer expectation in robust forecast aggregation: The possibility/impossibility. arXiv preprint arXiv:2402.06062,

    work page Pith review arXiv

  8. [1991]

    Online multivalid learning: Means, moments, and prediction intervals

    Varun Gupta, Christopher Jung, Georgy Noarov, Mallesh M Pai, and Aaron Roth. Online multivalid learning: Means, moments, and prediction intervals. arXiv preprint arXiv:2101.01739 ,

    arXiv

Show all 18 references
  1. [2001]

    Bandits for online calibration: An application to content moderation on social media platforms

    Vashist Avadhanula, Omar Abdul Baki, Hamsa Bastani, Osbert Bastani, Caner Gocmen, Daniel Haimovich, Darren Hwang, Dima Karamshuk, Thomas Leeper, Jiayuan Ma, et al. Bandits for online calibration: An application to content moderation on social media platforms. arXiv preprint ar...

    arXiv

  2. [2002]

    Numerical methods for solving fredholm integral equations of second kind

    S Saha Ray and PK Sahu. Numerical methods for solving fredholm integral equations of second kind. In Abstract and Applied Analysis , volume 2013, page 426916. Wiley Online Library,

    2013

  3. [2003]

    Learning linear operators: Infinite-dimensional regres- sion as a well-behaved non-compact inverse problem

    Mattes Mollenhauer, Nicole M¨ ucke, and TJ Sullivan. Learning linear operators: Infinite-dimensional regres- sion as a well-behaved non-compact inverse problem. arXiv preprint arXiv:2211.08875 ,

    arXiv

  4. [2014]

    Effective dimension of exp-concave optimization

    Naman Agarwal and Alon Gonen. Effective dimension of exp-concave optimization. arXiv preprint arXiv:1805.08268,

    arXiv

  5. [2016]

    Survey on applications of multi-armed and contextual bandits

    12 Djallel Bouneffouf, Irina Rish, and Charu Aggarwal. Survey on applications of multi-armed and contextual bandits. In 2020 IEEE Congress on Evolutionary Computation (CEC) , pages 1–8. IEEE,

    2020

  6. [2017]

    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

  7. [2021]

    Truthful information elicitation from hybrid crowds

    Qishen Han, Sikai Ruan, Yuqing Kong, Ao Liu, Farhad Mohsin, and Lirong Xia. Truthful information elicitation from hybrid crowds. arXiv preprint arXiv:2107.10119 ,

    arXiv

  8. [2022]

    A survey on contextual multi-armed bandits

    Li Zhou. A survey on contextual multi-armed bandits. arXiv preprint arXiv:1508.03326 ,

    arXiv

  9. [2023]

    Functional linear regression of cumulative dis- tribution functions

    Qian Zhang, Anuran Makur, and Kamyar Azizzadenesheli. Functional linear regression of cumulative dis- tribution functions. arXiv preprint arXiv:2205.14545 ,

    arXiv

  10. [2024]

    Trace ideals with applications (lecture notes)

    Aleksey Kostenko. Trace ideals with applications (lecture notes). Nikola B Kovachki, Samuel Lanthaler, and Andrew M Stuart. Operator learning: Algorithms and analysis. arXiv preprint arXiv:2402.15715 ,

    arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.