Pith. sign in

REVIEW 3 major objections 4 minor 41 references

Spectral Algorithms in Misspecified Regression: Convergence under Covariate Shift

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Weighted spectral algorithms converge at minimax-optimal rates under covariate shift, including for misspecified target functions outside the RKHS, with truncation handling unbounded density ratios.

desk verdict First unified spectral-algorithm analysis under covariate shift and misspecification, with likely-correct rates; fix the filter-domain gap before accepting. read the letter →

arxiv 2509.05106 v1 pith:PXK7NBNI submitted 2025-09-05 stat.ML cs.LG

Ren-Rui Liu , Zheng-Chu Guo This is my paper · ORCID
classification stat.MLcs.LG MSC 62G0868T05
keywords spectralalgorithmscovariateshiftimportanceweightingkernelridgeregressionmodelmisspecificationreproducingHilbertspaceconvergenceratestruncateddensityratio
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

Under covariate shift, training data come from a source distribution while predictions are evaluated on a target distribution with a different input density. Standard kernel regression must be reweighted by the density ratio w = dρ_T/dρ_S, but existing theory mostly assumes this ratio is bounded and that the true regression function lies in the RKHS. This paper proves that a weighted spectral estimator—any regularized inverse of the empirical operator built from filter functions—converges to the true regression function at rates governed by smoothness r, eigenvalue decay β, embedding index α₀, and the tail order p of w. When w is bounded the L2 rate is minimax-optimal; when w is unbounded, truncating w at D = n^ν recovers near-optimal rates, and both results extend to misspecified targets outside the RKHS.

What carries the argument

The argument runs on four pieces: (1) filter functions g_λ satisfying two spectral bounds—bounded regularized inverse and qualification—which unify kernel ridge regression, early-stopped gradient descent, and spectral cutoff; (2) the weighted empirical integral operator L̂_K = (1/n)Σ w(x_i)K(·,x_i)K(·,x_i)* and its truncated version; (3) interpolation spaces [H]^γ together with the embedding index α₀, which quantify misspecification and RKHS capacity; and (4) a four-term error decomposition J₁J₂(J₃+J₄) whose factors are bounded by Cordes-type operator inequalities and Bernstein concentration under the moment condition on w.

What would settle it

Fix a kernel with t_j ≍ j^{−β} and a target f_ρ ∈ [H]^{2r}. Generate source and target covariate distributions whose density ratio satisfies Assumption 1 in one scenario and violates it in another (for instance, a Pareto ratio with tail exponent just above 1/(p(m−1)) for some m ≥ 2). Run the weighted spectral estimator with λ = n^{−s} and plot ‖f̂ − f_ρ‖_{L2} against n. If both scenarios show the same n^{−r/(2r+1/β)} slope, the assumption is not load-bearing; if the heavy-tailed scenario degrades or the estimator becomes ill-defined, the stated rates are confirmed as conditional on Assumption

Watch

Extended reading notes

Core claim

The paper's central claim is a single rate bound: under Assumptions 1–4, with λ = n^{−s} and s chosen from (r, α₀, β, p), the weighted spectral estimator satisfies ‖f̂_{z,λ} − f_ρ‖_{[H]^γ} = O(n^{−s(r−γ/2)} log(6/δ)) with probability 1−δ, for 0 ≤ γ ≤ min{2r,1}. In the bounded-ratio case p = ∞, this gives the L2 rate n^{−r/(2r+1/β)} when 2r > α₀, matching the known minimax lower bound; for 2r ≤ α₀ the rate is n^{−r/(α₀+ε+1/β)}. For unbounded ratios, the truncated estimator with w† = min{w, D}, D = n^ν, achieves rates arbitrarily close to the optimal rate as the moment order m grows, including in the misspecified regime r < 1/2. The proof splits the error into an approximation term controlled

Load-bearing premise

The moment condition in Assumption 1 requires the density ratio to have exponential-type tail decay—each m-th moment bounded by (1/2)m! L^{m−2}σ²—and every concentration inequality in the appendix relies on it; if the true ratio has heavier tails, the stated rates, including the truncation theorem, do not follow.

Editorial extensions

If this is right

  • For bounded density ratios and 2r > α₀, the L2 convergence rate n^{−r/(2r+1/β)} matches the no-shift minimax lower bound, so importance weighting costs nothing in the exponent.
  • For unbounded density ratios, truncating at D = n^{1/(p(m−1)+1)} gives rates within any ε of the optimal rate once m is chosen large enough, and the same truncation works for misspecified targets.
  • The bound holds in every interpolation norm [H]^γ, including the RKHS norm (γ=1) and L2 norm (γ=0), so one theorem yields both function-space and prediction-error guarantees.
  • The filter-function framework covers kernel ridge regression, early-stopped gradient descent, and spectral cutoff simultaneously, transferring the rates across algorithms automatically.
  • The finite-sample threshold makes explicit how sample complexity grows with tail order p, embedding norm M_α, and effective dimension N(λ), enabling concrete complexity comparisons.

Reading between the lines

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

  • The threshold α₀ in Theorem 2.1 is weaker by 1/β than the no-shift threshold α₀ − 1/β; the authors attribute this to the loss of uniform boundedness of the empirical operator under weighting, suggesting a sharper threshold may be provable for bounded w.
  • Because the analysis assumes the density ratio is known exactly, a natural extension is to plug in an estimated ratio; the moment condition would need an extra estimation-error term, and the truncation level D could be chosen adaptively from an estimated tail order p.
  • The γ > 0 interpolation rates imply control of the RKHS norm of the error, which combined with the embedding index suggests uniform sup-norm rates could be derived by taking γ near α₀ + ε, a step the paper does not take.
  • Corollary 2 identifies a concrete kernel property—embedding index attaining 1/β, as with uniformly bounded eigenfunctions or Sobolev-type embeddings—under which unbounded covariate shift becomes asymptotically free, a checkable condition for practitioners.
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. This paper analyzes weighted spectral algorithms for nonparametric regression in an RKHS under covariate shift, allowing the target regression function to be misspecified (outside the RKHS). Importance weights are used to correct the distribution mismatch, and convergence is measured in the interpolation norms [H]^gamma. Under a moment condition on the density ratio, a source condition, an eigenvalue decay condition, and an embedding-index condition, Theorem 2.1 gives rates of the form n^{-s(r-\gamma/2)} log(6/\delta). Corollary 1 obtains minimax-optimal L2 rates for bounded density ratios when 2r > alpha_0, Corollary 2 shows near-optimal rates when alpha_0 = 1/beta, and Theorem 2.2 introduces a truncation scheme for unbounded density ratios, with near-optimal rates stated in Corollary 3. The proofs use the standard bias-variance decomposition, operator norm bounds, and Bernstein-type concentration inequalities in Hilbert space.

Significance. If the technical gaps identified below are repaired, this is a valuable contribution: it extends spectral-algorithm theory under covariate shift from the well-specified setting to the misspecified setting, incorporates the embedding index to capture the influence of the kernel and measure, and covers both RKHS-norm and L2-norm rates in one framework. The paper makes no fitted-parameter claims and the rates follow from explicit assumptions; the proofs are detailed and the comparison with prior work is helpful. The authors also candidly acknowledge the suboptimal threshold T = alpha_0 in Theorem 2.1, which strengthens confidence in the presentation. The main load-bearing concerns are local and fixable, but they currently prevent the theorems from being fully correct as stated.

major comments (3)
  1. [Definition 1 and Remark after (4)] The filter functions g_lambda are defined only on [0, kappa^2], but the weighted empirical operator \hat L_K can have norm larger than kappa^2 with non-negligible probability even when w is bounded. Lemma A.9 only gives ||\hat L_K - L_K|| <= eps with high probability, hence ||\hat L_K|| <= kappa^2 + eps, not ||\hat L_K|| <= kappa^2. For example, if L_K is rank one with K(x,x)=1, then ||\hat L_K|| is the sample mean of w and P(||\hat L_K|| > 1) tends to 1/2 for bounded w with variance. Thus g_lambda(\hat L_K) is not defined on a non-negligible event, and the 1-delta bounds in Theorems 2.1 and 2.2 do not follow as stated. The w.l.o.g. assertion in the Remark is unsupported. This is fixable by extending Definition 1 to all of [0,infty)---the three standard filters listed in Section 1 already satisfy the conditions there---or by proving a one-sided high-probability bound on ||\hat L_K||.
  2. [Assumption 2 and Proposition 4.4] The sentence after Assumption 2 claims that the boundedness condition ||f_rho||_infty <= G 'additionally guarantees |y| <= G holds rho_T-a.e.' This is not true: the conditional mean being bounded does not imply the response variable is bounded. The proof of Proposition 4.4, however, uses |y| <= G to obtain the uniform bound in Eq. (15), and this bound is then used in the concentration argument for J_3. Without an explicit assumption that |y| <= G almost surely (or that Y is bounded), the proof of Theorem 2.1 is incomplete. This is easily repaired by adding such a boundedness assumption, but as written the implication is incorrect.
  3. [Theorem 2.1 / Proposition 4.2] In the regime 2r <= alpha_0, Theorem 2.1 states that any epsilon > 0 is allowed and then sets alpha = alpha_0 + epsilon/2. But Proposition 4.2 requires alpha in (alpha_0, 1] when 2r <= alpha_0, and the embedding constant M_alpha is defined only for alpha <= 1. If epsilon > 2(1-alpha_0), then alpha > 1 and the proof does not apply. The theorem should restrict epsilon to (0, 2(1-alpha_0)) in this regime, or choose alpha = min(alpha_0 + epsilon/2, 1) and adjust the rate calculation accordingly. This does not affect the main minimax claim, but it is a gap in the theorem as stated.
minor comments (4)
  1. [Section 2, Eq. (7) and (8)] The constants C_N and c_N from Lemma A.1 are used in (7) and (8) but not explicitly defined in the main text. Please define them alongside the eigenvalue decay assumption.
  2. [Corollary 1] The phrase 'minimax optimal' should be qualified as 'optimal up to logarithmic factors' since the stated bound contains log(6/delta) while the standard minimax lower bound does not include such a factor.
  3. [Proof of Proposition 4.5] Minor typo: 'the proof is complete by by combining' should read 'by combining'.
  4. [Theorem 2.2 and Proposition 4.6] In the proof of Theorem 2.2, the authors state that M_1 can be replaced by kappa. This replacement should be made explicit in the statement of Proposition 4.6 or in the condition (S2), since the theorem statement uses only kappa.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the convergence rates follow from stated assumptions and standard operator-concentration lemmas, with self-citations only contextual.

full rationale

Walking the derivation chain: Theorem 2.1 is obtained by combining Proposition 4.1 (approximation error) and Proposition 4.2 (estimation error). The approximation bound is a direct functional-calculus consequence of the filter qualification (3) and the source condition f_rho = L_K^r u_rho. The estimation bound is split into J1-J4, each controlled by: (i) Cordes-type operator inequalities (Lemmas A.3, A.4, A.6), (ii) Bernstein inequalities for Hilbert-space random variables (Lemma A.2, cited to Caponnetto–de Vito), and (iii) concentration bounds for the weighted empirical operator (Lemmas A.8, A.9) whose moment hypotheses are exactly Assumption 1. No parameter is fitted to a subset of the data and then renamed a prediction; the rates are algebraic simplifications of these bounds under lambda = n^{-s}. Assumption 1 is adopted from prior work [16], but it is an input assumption, not derived from the target theorem. Self-citations ([9], [10], [18]) appear only in related-work comparisons and table entries, not as load-bearing justifications of the rates. The embedding-index and minimax-comparison results are cited from external work ([14], [40]). One non-circular caveat should be flagged: the Remark after (4) asserts one may assume ||hat L_K|| <= kappa^2 without loss of generality because of Lemma A.9; Lemma A.9 only gives ||hat L_K - L_K|| <= epsilon, hence ||hat L_K|| <= kappa^2 + epsilon, not <= kappa^2. This is a technical gap in the WLOG reduction, not a circularity, because the filter-domain issue does not make the claimed rates equivalent to the assumptions by construction. There is therefore no circular step scoring above 0.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The paper introduces no fitted parameters or invented entities. It relies on standard mathematical background and domain assumptions that define the statistical problem. The moment condition on the density ratio is the most restrictive premise and is taken from prior work.

assumptions (6)
  • domain assumption Assumption 1: density ratio w satisfies moment condition with constants p, L, sigma
    Strong tail control of the density ratio; used in all Bernstein concentration bounds (e.g., Lemma A.8, A.9). If w has heavier tails, the stated rates do not hold.
  • domain assumption Assumption 2: source condition f_rho = L_K^r u_rho with ||f_rho||_inf <= G
    Defines regularity of the target function; central for bias-variance tradeoff. Includes misspecification for r < 1/2.
  • domain assumption Assumption 3: eigenvalues of L_K decay polynomially with exponent beta > 1
    Controls effective dimension; both lower and upper bounds needed for matching minimax lower bounds.
  • domain assumption Assumption 4: embedding index alpha_0 in [1/beta, 1) exists
    Used to bound sup norm of interpolation spaces; crucial for the misspecified case. Existence is not guaranteed for all kernels.
  • standard math Mercer decomposition and RKHS embedding properties
    Background from Cucker-Zhou and Steinwart-Scovel; used to define interpolation spaces and the operator representation.
  • standard math Bernstein inequality for Hilbert space valued random variables (Lemma A.2)
    Used throughout to derive high-probability bounds from moment conditions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spectral Algorithms in Misspecified Regression: Convergence under Covariate Shift." pith.science (2026). https://pith.science/paper/PXK7NBNI

@misc{pith2026250905106,
  author       = {Pith},
  title        = {Pith review of: Spectral Algorithms in Misspecified Regression: Convergence under Covariate Shift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PXK7NBNI}},
  note         = {Machine review of arXiv:2509.05106}
}
read the original abstract

This paper investigates the convergence properties of spectral algorithms -- a class of regularization methods originating from inverse problems -- under covariate shift. In this setting, the marginal distributions of inputs differ between source and target domains, while the conditional distribution of outputs given inputs remains unchanged. To address this distributional mismatch, we incorporate importance weights, defined as the ratio of target to source densities, into the learning framework. This leads to a weighted spectral algorithm within a nonparametric regression setting in a reproducing kernel Hilbert space (RKHS). More importantly, in contrast to prior work that largely focuses on the well-specified setting, we provide a comprehensive theoretical analysis of the more challenging misspecified case, in which the target function does not belong to the RKHS. Under the assumption of uniformly bounded density ratios, we establish minimax-optimal convergence rates when the target function lies within the RKHS. For scenarios involving unbounded importance weights, we introduce a novel truncation technique that attains near-optimal convergence rates under mild regularity conditions, and we further extend these results to the misspecified regime. By addressing the intertwined challenges of covariate shift and model misspecification, this work extends classical kernel learning theory to more practical scenarios, providing a systematic framework for understanding their interaction.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 38 canonical work pages

  1. [1]

    Optimal learning rates for kernel conjugate gradient regression,

    G. Blanchard and N. Kr¨ amer, “Optimal learning rates for kernel conjugate gradient regression,” inAdvances in Neural Information Processing Systems, vol. 23, Curran Associates, Inc., 2010 (cit. on p. 41)

    work page 2010

  2. [2]

    Optimal rates for regularization of statistical inverse learning problems,

    G. Blanchard and N. M¨ ucke, “Optimal rates for regularization of statistical inverse learning problems,”Foun- dations of Computational Mathematics, vol. 18, no. 4, pp. 971–1013, 2018 (cit. on pp. 12, 14)

    work page 2018

  3. [3]

    What is the effect of importance weighting in deep learning?

    J. Byrd and Z. Lipton, “What is the effect of importance weighting in deep learning?” InProceedings of the 36th International Conference on Machine Learning, vol. 97, PMLR, 2019, pp. 872–881 (cit. on p. 13)

    work page 2019

  4. [4]

    Stochastic gradient descent for two-layer neural networks,

    D. H. Cao, Z. C. Guo, and L. Shi, “Stochastic gradient descent for two-layer neural networks,”arXiv preprint,

    work page

  5. [5]

    Optimal rates for the regularized least-squares algorithm,

    A. Caponnetto and E. de Vito, “Optimal rates for the regularized least-squares algorithm,”Foundations of Computational Mathematics, vol. 7, no. 3, pp. 331–368, 2007 (cit. on pp. 7, 8, 12, 40)

    work page 2007

  6. [6]

    H. O. Cordes,Spectral theory of linear differential operators and comparison algebras. Cambridge University Press, 1987 (cit. on p. 41)

    work page 1987

  7. [7]

    Cucker and D

    F. Cucker and D. X. Zhou,Learning theory: An approximation theory viewpoint. Cambridge University Press, 2007 (cit. on pp. 6, 7)

    work page 2007

  8. [8]

    Regularization of inverse problems,

    H. W. Engl and R. Ramlau, “Regularization of inverse problems,” inEncyclopedia of Applied and Computational Mathematics. Springer Berlin Heidelberg, 2015, pp. 1233–1241 (cit. on p. 4)

    work page 2015

Show all 41 references
  1. [9]

    Spectral algorithms for functional linear regression,

    J. Fan, Z. C. Guo, and L. Shi, “Spectral algorithms for functional linear regression,”Communications on Pure and Applied Analysis, vol. 23, no. 7, pp. 895–915, 2024 (cit. on pp. 12, 14)

    2024

  2. [10]

    Spectral algorithms under covariate shift,

    J. Fan, Z. C. Guo, and L. Shi, “Spectral algorithms under covariate shift,”arXiv preprint, 2025. eprint:2504. 12625(stat.ML) (cit. on pp. 2, 5, 13, 14)

    2025

  3. [11]

    Deep nonparametric quantile regression under covariate shift,

    X. D. Feng, X. He, Y. L. Jiao, L. C. Kang, and C. X. Wang, “Deep nonparametric quantile regression under covariate shift,”Journal of Machine Learning Research, vol. 25, no. 385, pp. 1–50, 2024 (cit. on p. 13)

    2024

  4. [12]

    Towards a unified analysis of kernel-based methods under covariate shift,

    X. D. Feng, X. He, C. X. Wang, C. Wang, and J. N. Zhang, “Towards a unified analysis of kernel-based methods under covariate shift,” inAdvances in Neural Information Processing Systems, vol. 36, Curran Associates, Inc., 2023, pp. 73 839–73 851 (cit. on pp. 2, 4, 10, 13, 14)

    2023

  5. [13]

    The clinician and dataset shift in artificial intelligence,

    S. G. Finlayson et al., “The clinician and dataset shift in artificial intelligence,”New England Journal of Medicine, vol. 385, no. 3, pp. 283–286, 2021 (cit. on p. 2)

    2021

  6. [14]

    Sobolev norm learning rates for regularized least-squares algorithms,

    S. Fischer and I. Steinwart, “Sobolev norm learning rates for regularized least-squares algorithms,”Journal of Machine Learning Research, vol. 21, no. 205, pp. 1–38, 2020 (cit. on pp. 8, 9, 12, 41)

    2020

  7. [15]

    On a regularization of unsupervised domain adaptation in rkhs,

    E. R. Gizewski et al., “On a regularization of unsupervised domain adaptation in rkhs,”Applied and Compu- tational Harmonic Analysis, vol. 57, pp. 201–227, 2022 (cit. on pp. 2, 4, 13, 14)

    2022

  8. [16]

    When is importance weighting correction needed for covariate shift adaptation?

    D. Gogolashvili, M. Zecchin, M. Kanagawa, M. Kountouris, and M. Filippone, “When is importance weighting correction needed for covariate shift adaptation?”arXiv preprint, 2023. eprint:2303.04020(stat.ML) (cit. on pp. 2, 5, 6, 10, 13, 14)

    2023 arXiv

  9. [17]

    Gradient descent for robust kernel-based regression,

    Z. C. Guo, T. Hu, and L. Shi, “Gradient descent for robust kernel-based regression,”Inverse Problems, vol. 34, no. 6, p. 065 009, 2018 (cit. on pp. 8, 12). 37

    2018

  10. [18]

    Learning theory of distributed spectral algorithms,

    Z. C. Guo, S. B. Lin, and D. X. Zhou, “Learning theory of distributed spectral algorithms,”Inverse Problems, vol. 33, no. 7, p. 074 009, 2017 (cit. on pp. 12, 14)

    2017

  11. [19]

    Early stopping for iterative regularization with general loss functions,

    T. Hu and Y. W. Lei, “Early stopping for iterative regularization with general loss functions,”Journal of Machine Learning Research, vol. 23, no. 339, pp. 1–36, 2022 (cit. on p. 12)

    2022

  12. [20]

    Correcting sample selection bias by unlabeled data,

    J. Y. Huang, A. Gretton, K. Borgwardt, B. Sch¨ olkopf, and A. Smola, “Correcting sample selection bias by unlabeled data,” inAdvances in Neural Information Processing Systems, vol. 19, MIT Press, 2006 (cit. on p. 13)

    2006

  13. [21]

    A short survey on importance weighting for machine learning,

    M. Kimura and H. Hino, “A short survey on importance weighting for machine learning,”Transactions on Machine Learning Research, 2024 (cit. on p. 13)

    2024

  14. [22]

    Optimal rates for regularized conditional mean embed- ding learning,

    Z. Li, D. Meunier, M. Mollenhauer, and A. Gretton, “Optimal rates for regularized conditional mean embed- ding learning,” inAdvances in Neural Information Processing Systems, vol. 35, Curran Associates, Inc., 2022, pp. 4433–4445 (cit. on p. 9)

    2022

  15. [23]

    Towards a unified analysis of random fourier features,

    Z. Li, J.-F. Ton, D. Oglic, and D. Sejdinovic, “Towards a unified analysis of random fourier features,” in Proceedings of the 36th International Conference on Machine Learning, vol. 97, PMLR, 2019, pp. 3905–3914 (cit. on p. 12)

    2019

  16. [24]

    Optimal rates for spectral algorithms with least-squares regression over hilbert spaces,

    J. H. Lin, A. Rudi, L. Rosasco, and V. Cevher, “Optimal rates for spectral algorithms with least-squares regression over hilbert spaces,”Applied and Computational Harmonic Analysis, vol. 48, no. 3, pp. 868–890, 2020 (cit. on pp. 8, 12–14)

    2020

  17. [25]

    Spectral algorithms for supervised learning,

    L. Lo Gerfo, L. Rosasco, F. Odone, E. de Vito, and A. Verri, “Spectral algorithms for supervised learning,” Neural Computation, vol. 20, no. 7, pp. 1873–1897, 2008 (cit. on pp. 12, 14)

    2008

  18. [26]

    Optimally tackling covariate shift in rkhs-based nonparametric regression,

    C. Ma, R. Pathak, and M. J. Wainwright, “Optimally tackling covariate shift in rkhs-based nonparametric regression,”The Annals of Statistics, vol. 51, no. 2, pp. 738–761, 2023 (cit. on pp. 2, 4, 10, 13, 14)

    2023

  19. [27]

    On regularized radon-nikodym differentiation,

    D. H. Nguyen, W. Zellinger, and S. Pereverzyev, “On regularized radon-nikodym differentiation,”Journal of Machine Learning Research, vol. 25, no. 266, pp. 1–24, 2024 (cit. on p. 13)

    2024

  20. [28]

    Qui˜ nonero-Candela, M

    J. Qui˜ nonero-Candela, M. Sugiyama, A. Schwaighofer, and N. D. Lawrence,Dataset shift in machine learning. MIT Press, 2008 (cit. on p. 2)

    2008

  21. [29]

    Early stopping for non-parametric regression: An optimal data- dependent stopping rule,

    G. Raskutti, M. J. Wainwright, and B. Yu, “Early stopping for non-parametric regression: An optimal data- dependent stopping rule,”Journal of Machine Learning Research, vol. 15, no. 11, pp. 335–366, 2011 (cit. on p. 12)

    2011

  22. [30]

    Generalization properties of learning with random features,

    A. Rudi and L. Rosasco, “Generalization properties of learning with random features,” inAdvances in Neural Information Processing Systems, vol. 30, Curran Associates, Inc., 2017 (cit. on p. 12)

    2017

  23. [31]

    Improving predictive inference under covariate shift by weighting the log-likelihood function,

    H. Shimodaira, “Improving predictive inference under covariate shift by weighting the log-likelihood function,” Journal of Statistical Planning and Inference, vol. 90, no. 2, pp. 227–244, 2000 (cit. on pp. 2, 3, 13)

    2000

  24. [32]

    Mercer’s theorem on general domains: On the interaction between measures, kernels, and rkhss,

    I. Steinwart and C. Scovel, “Mercer’s theorem on general domains: On the interaction between measures, kernels, and rkhss,”Constructive Approximation, vol. 35, no. 3, pp. 363–417, 2012 (cit. on p. 6)

    2012

  25. [33]

    Sugiyama, T

    M. Sugiyama, T. Suzuki, and T. Kanamori,Density ratio estimation in machine learning. Cambridge University Press, 2012 (cit. on p. 13)

    2012

  26. [34]

    V. N. Vapnik,Statistical learning theory. Wiley, 1998 (cit. on p. 2). 38

    1998

  27. [35]

    Risk bounds for the regularized least-squares algorithm with operator-valued kernels,

    E. de Vito and A. Caponnetto, “Risk bounds for the regularized least-squares algorithm with operator-valued kernels,” Massachusetts Institute of Technology Computer Science and Artificial Intelligence Laboratory, Tech. Rep. MIT-CSAIL-TR-2005-031, 2005 (cit. on p. 3)

    2005

  28. [36]

    Learning from examples as an inverse problem,

    E. de Vito, L. Rosasco, A. Caponnetto, U. de Giovannini, and F. Odone, “Learning from examples as an inverse problem,”Journal of Machine Learning Research, vol. 6, no. 30, pp. 883–904, 2005 (cit. on p. 4)

    2005

  29. [37]

    Robust learning under uncertain test distributions: Relating covariate shift to model misspecification,

    J. F. Wen, C.-N. Yu, and R. Greiner, “Robust learning under uncertain test distributions: Relating covariate shift to model misspecification,” inProceedings of the 31st International Conference on Machine Learning, vol. 32, PMLR, 2014, pp. 631–639 (cit. on p. 13)

    2014

  30. [38]

    Understanding the role of importance weighting for deep learning,

    D. Xu, Y. T. Ye, and C. W. Ruan, “Understanding the role of importance weighting for deep learning,” in International Conference on Learning Representations, 2021 (cit. on p. 13)

    2021

  31. [39]

    Estimating unbounded density ratios: Applications in error control under covariate shift,

    S. T. Xu, Z. Yu, and J. Huang, “Estimating unbounded density ratios: Applications in error control under covariate shift,”arXiv preprint, 2025. eprint:2504.01031(stat.ML) (cit. on p. 13)

    2025 arXiv

  32. [40]

    On the optimality of misspecified spectral algorithms,

    H. B. Zhang, Y. C. Li, and Q. Lin, “On the optimality of misspecified spectral algorithms,”Journal of Machine Learning Research, vol. 25, no. 188, pp. 1–50, 2024 (cit. on pp. 8–10, 12–14). 39 A Appendix This appendix presents auxiliary lemmas referenced in Section 4. Throughou...

    2024

  33. [2024]

    eprint:2407.07670(stat.ML) (cit. on p. 12)

    arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.