Pith. sign in

REVIEW 3 major objections 5 minor 70 references

Nonparametric Goodness-of-fit Testing under Covariate Shift

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

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

Pith's one-line read The paper proves that bootstrap confidence balls for a target-domain regression function reach nominal coverage under covariate shift.

desk verdict Real and careful theory — the first nonasymptotic L2(Q) coverage guarantees under covariate shift — but 'sharpness' is an overclaim and the coverage theorem only holds for exactly known density ratio and exact target kernel integrals. read the letter →

arxiv 2608.04860 v2 pith:7O7BTNET submitted 2026-08-05 stat.ME cs.LGmath.STstat.MLstat.TH

Zhen Hou , Dong Xia This is my paper · ORCID
classification stat.MEcs.LGmath.STstat.MLstat.TH MSC 62G0862G1062G0962G20
keywords covariateshiftkernelridgeregressionmultiplierbootstrapgoodness-of-fittestingimportanceweightingconfidencesetsGaussianapproximationtargetpopulationinference
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 asks a practical question: when labelled data come from one population (the source) but the scientific conclusion concerns another (the target), can you still test whether a candidate regression function equals the truth in the target's $L^2(Q)$ metric? The proposed answer is yes, using truncated importance-weighted kernel ridge regression to estimate the regression function and a multiplier bootstrap to calibrate the radius of an $L^2(Q)$ confidence ball. The paper proves that the miscoverage of this ball is bounded by a term of order $\eta+\sqrt{\tau_n}(\Delta_G+\Delta_B+\Delta_{\mathrm{bias}})$, where the three $\Delta$ terms are explicit errors from bias, Gaussian approximation, and bootstrap approximation. Under concrete conditions on the density ratio and the kernel spectrum these reduce to polynomial or near-polynomial rates in the sample size. If the paper is right, a practitioner can invert such a ball to test goodness-of-fit on the target population with asymptotically correct size, even when source and target covariate distributions differ substantially.

What carries the argument

The engine is the truncated importance-weighted kernel ridge estimator $\hat f=k_x^\top(K+n\lambda W^{-1})^{-1}y$, with weight $w_n(x)=\min\{\rho(x),\tau_n\}$, paired with a multiplier bootstrap $B=\sqrt n\, k_x^\top(K+n\lambda W^{-1})^{-1}\mathrm{diag}(\hat\varepsilon)z$ whose $L^2(Q)$-norm is evaluated as $(a^\top K_Q a)^{1/2}$ using the fixed matrix $K_Q=(\langle k_{X_i},k_{X_j}\rangle_{L^2(Q)})$. The proof machinery is operator-theoretic: the unitary isometry $L_Q^{1/2}:\mathrm{Ran}(L_Q)\to H$ generated by the square root of the target integral operator lets the source-side RKHS calculus be compared with the target $L^2(Q)$ geometry. Two compatibility conditions carry the argument: the regularized transfer coverage condition $L_Q\preceq C_1^2(L_{P_w}+\lambda I)$ makes the source risk control the target risk, and the non-degenerate covariance condition prevents the Gaussian covariance from degenerating in the target norm, which is what allows anti-concentration. The total error $\Delta=\Delta_G+\Delta_B+\Delta_{\mathrm{bias}}$ enters the coverage bound multiplied by $\sqrt{\tau_n}$, the anti-concentration factor from Lemma 4.1.

What would settle it

Run Algorithm 1 in the oracle setting with known $\rho$, a compact covariate space, a Gaussian RBF kernel, source $N(0,4)$ and target $N(1,1)$ covariates, and $n=1000$, repeating the coverage experiment in Table 2; if the empirical coverage of the 90% ball deviates from 0.90 by more than the theorem's bound, the Gaussian-coupling or bootstrap-coupling step is wrong. A cheaper targeted check is to test Assumption 4.5 numerically by computing the smallest eigenvalue ratio between the two covariance operators it compares; if that ratio is not bounded below by a constant independent of $n$, the anti-concentration step is unavailable and coverage can break.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is Theorem 4.1: under Assumptions 4.1–4.5, the coverage error of the bootstrap confidence ball $\mathrm{CS}_\alpha(\hat f)$ is bounded by $|\mathbb{P}(f^*\notin \mathrm{CS}_\alpha(\hat f))-\alpha|\lesssim \eta+\sqrt{\tau_n}\Delta$, with $\Delta=\Delta_G+\Delta_B+\Delta_{\mathrm{bias}}$. The proof splits $\Delta$ into the deterministic bias of the truncated IWKRR estimator, the Gaussian approximation of its stochastic fluctuation in the target $L^2(Q)$ geometry, and the multiplier-bootstrap approximation of that Gaussian, and controls each term through operator comparisons between the target integral operator $L_Q$ and the regularized weighted-source operator $L_{P_w}+\lambda I$. Corollaries 4.2–4.4 turn the general bound into explicit rates, for instance $n^{-(\beta-5)(\beta+10)/(6(\beta+3)(2\beta+5))}\log^3 n$ when the density ratio is bounded and the kernel eigenvalues decay like $k^{-\beta}$ with $\beta>5$. The paper is thus establishing that nonparametric target-domain inference under covariate shift is feasible with asymptotically valid coverage, provided the density ratio is known and two kernel-dependent compatibility conditions hold.

Load-bearing premise

The load-bearing premise is that the target-to-source density ratio $\rho(x)$ is known exactly; Section 7 states that estimating $\rho$ is not covered and its effect on the bootstrap approximation and finite-sample coverage remains unclear and challenging.

Editorial extensions

If this is right

  • The inverted test rejects $H_0:\|f_0-f^*\|_{L^2(Q)}=0$ whenever $f_0$ falls outside the bootstrap ball, and Theorem 4.2 shows power tends to 1 for any fixed target-domain separation.
  • With bounded density ratio and polynomial spectral decay $\mu_k\asymp k^{-\beta}$, $\beta>5$, the miscoverage rate is $n^{-(\beta-5)(\beta+10)/(6(\beta+3)(2\beta+5))}\log^3 n$; with exponential spectral decay it is $n^{-1/10}(\log n)^{3/2+1/(2\zeta)}$.
  • A sub-exponential density-ratio tail is handled by a growing truncation level, while a heavy-tailed ratio with a bounded $\theta$-th moment ($\theta>7$) still yields polylogarithmic coverage error when the kernel eigenvalues decay exponentially.
  • The same confidence-ball construction can be used for composite nulls, such as monotonicity of a target regression function, and for two-sample comparisons, as demonstrated in the Survey of Consumer Finances applications.

Reading between the lines

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

  • Editorial inference: if the density ratio is estimated rather than known, the advertised coverage guarantee has no support; Section 7 explicitly leaves this effect unquantified, so a plug-in version is likely to undercover and should be validated by experiment or augmented with a sensitivity analysis.
  • Editorial inference: the explicit rates in Corollaries 4.2–4.4 are conservative; the proof uses piecewise optimizer balancing, so a natural testable extension is to compare empirical miscoverage slopes in simulations with the stated exponents to see whether faster decay is possible under the same assumptions.
  • Editorial inference: because the bootstrap norm is a cheap quadratic form through the fixed matrix $K_Q$, the method's practical cost is dominated by a single kernel ridge solve; this makes it a plausible building block for routine target-domain uncertainty quantification once reliable density-ratio estimates are available.
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 proposes a nonparametric goodness-of-fit procedure under covariate shift: from source-labelled data it builds an L2(Q)-norm confidence ball for the regression function using truncated importance-weighted kernel ridge regression plus a multiplier bootstrap. The main result (Theorem 4.1) bounds the miscoverage by η + sqrt(τ_n)(Δ_G+Δ_B+Δ_bias) under four assumptions, and Corollaries 4.2--4.4 translate this into explicit rates under polynomial or exponential kernel eigenvalue decay and bounded, sub-exponential, or bounded-θ-moment density ratios. The proof chain is explicit: Bahadur representation, Gaussian coupling, bootstrap coupling, and anti-concentration are assembled with operator-theoretic tools. An SCF data analysis and simulations are included.

Significance. If the results hold as stated, the paper provides one of the first nonasymptotic, operator-level coverage guarantees for bootstrap confidence balls in the target distribution under covariate shift. The proof is detailed and uses a coherent RKHS/L2(Q) operator calculus, with couplings due to Zaitsev, Monrad--Philipp, and Pinelis--Sakhanenko deployed carefully; the explicit rates are a genuine contribution. However, the guarantees are oracle: Algorithm 1 and all theorems require exact knowledge of the density ratio ρ and exact target L2(Q) inner products. The paper itself concedes in Section 7 that the effect of estimating ρ is not characterized. Consequently, the practical claim of 'valid inference under covariate shift' is not established in the usual setting where only unlabelled target covariates are observed. The abstract's word 'sharpness' is also unsupported by any lower bound.

major comments (3)
  1. [Corollaries 4.3 and 4.4] The central coverage theorem applies only when the target-to-source density ratio ρ and the target kernel matrix K_Q with entries <k_{X_i},k_{X_j}>_{L2(Q)} are known exactly. Algorithm 1 takes ρ as an input, and the theorem statements contain no term for an estimated density ratio or an estimated target Gram matrix. Section 7 explicitly states that replacing ρ by an estimate leaves the effect on the first-order approximation, bootstrap approximation, and finite-sample coverage 'unclear and challenging'. In the standard covariate-shift setting, where only source and unlabelled target covariates are available, the advertised nominal coverage is therefore not a theorem. This scope gap should be stated prominently in the abstract and introduction, or the paper should extend the analysis to estimated ρ and approximate K_Q (as in the 1200-sample Monte Carlo approximation used in Section 5.1).
  2. [Abstract and Section 1] Corollaries 4.3 and 4.4 claim coverage rates under the sub-exponential or bounded-θ-moment conditions and spectral decay, but the anti-concentration step in Lemma 4.1 and hence Theorem 4.1 requires Assumption 4.5. The proofs of these corollaries verify Assumption 4.4 through Lemma F.29 or Lemma F.31, but they do not verify the covariance lower bound in Assumption 4.5. As written, the stated rates for the sub-exponential and heavy-tailed cases are not fully established. The authors should either add a perturbation lemma proving Assumption 4.5 under the stated density-ratio conditions, or add Assumption 4.5 explicitly to the hypotheses of Corollaries 4.3 and 4.4.
  3. [Corollaries 4.3 and 4.4] The abstract claims 'nonasymptotic validity and sharpness' of the confidence sets, but no lower bound is proved anywhere in the paper. Theorem 4.1 and all corollaries give only upper bounds of the form ≲ on the coverage error. The word 'sharpness' should be removed or replaced by a precise statement such as 'explicit nonasymptotic bounds', unless a matching lower bound is added.
minor comments (5)
  1. [Section 5.1] The symbol ρ is used both for the density ratio function ρ(x) and for the positive lower-bound constant ρ in Assumption 4.5; the two uses are confusing in the same sentence. Consider denoting the lower bound by ρ_min or ρ_0.
  2. [Section 6] The text says the IWKRR estimator 'uses normalized importance weights and truncated weights w_n(x)=min{ρ(x),τ_n}'. The theoretical development uses unnormalized truncated weights; the use of normalized weights in the experiments should be clarified, and any discrepancy between the implemented and analysed weights should be discussed.
  3. [Section 1] In Study 1, empirical distances are evaluated with respect to the empirical measure \widehat{Q} on a held-out target set, while the theory is for the exact L2(Q) norm. The confidence intervals in Figure 5 are therefore intervals for the empirical target distance unless additional uniformity or approximation guarantees are provided; this should be stated explicitly.
  4. [Section 6] Minor typographical issues: 'asourcepopulation' in Section 1 should be 'a source population'; Figure 4 repeats the same panel content; and the caption of Figure 1 uses 'out=' without explanation.
  5. [Section 6] Theorem 4.2 is a simple triangle-inequality argument plus convergence of the estimation error and critical value; it could be stated as a remark to avoid giving it the same weight as the main coverage theorem.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: coverage theorem is conditional on explicit assumptions; the only self-citation is illustrative and non-load-bearing.

full rationale

The paper's central claim is that bootstrap confidence balls for f* in L2(Q) have miscoverage controlled by eta + sqrt(tau_n)Delta (Theorem 4.1). This is not circular: the error terms Delta_G, Delta_B, Delta_bias are not fitted or defined in terms of the final coverage; they are bounded explicitly in Propositions 4.1-4.5 using Assumptions 4.1-4.5, and those assumptions are stated independently and verified through external analytic inputs (Mercer's theorem, Cordes inequality, Gaussian couplings of Zaitsev and Monrad-Philipp, Bernstein inequalities). The regularized transfer coverage condition (Assumption 4.4) and non-degenerate covariance condition (Assumption 4.5) are genuine hypotheses, not restatements of the coverage conclusion. The only self-citation is to 'Assumption 4 in Xia and Yuan [2021]' as an example of a non-degeneracy condition; it plays no role in the proof and is not used to exclude alternatives, so it is not load-bearing. Section 7's caveat that estimation of rho is not covered is an acknowledged scope limitation: the theorems are explicitly conditional on known rho and exact kernel matrix K_Q, and the paper does not claim a coverage theorem for estimated rho. This is a gap in practical applicability, not a circular step. No fitted input is renamed as a prediction, and no known result is repackaged under new coordinates. The derivation is therefore self-contained against its stated assumptions.

Assumptions & free parameters 3 free parameters · 8 assumptions · 0 invented entities

The central claim is a conditional theorem. It rests on strong but transparent assumptions: known density ratio, bounded noise, source condition, effective dimension control, transfer coverage, non-degenerate covariance, and fast spectral decay. The theory does not derive these conditions from the data or from the target conclusion.

free parameters (3)
  • lambda_n (regularization parameter) = n^{-a} in theoretical corollaries; validation-tuned in experiments
    The regularization parameter controls the bias-variance tradeoff and appears in every coverage error bound through Delta_bias and the operator estimates. Its rate-optimal choice is prescribed in Corollaries 4.2 to 4.4, while the experiments select it by validation.
  • tau_n (truncation level) = n^{delta} in theoretical corollaries; validation-tuned in experiments
    The truncation cap on the density ratio stabilizes the estimator but enters the coverage error bound through sqrt(tau_n) Delta. The choice of tau_n is essential and is tuned by validation in the numerical sections.
  • C0 (truncation constant under bounded density ratio) = chosen as a sufficiently large fixed constant C0 >= max(rho_bar, tau0)
    In Corollary 4.2 the truncation level is fixed at C0; the exact value is not data-fitted but is a hand-chosen constant depending on the unknown upper bound rho_bar.
assumptions (8)
  • domain assumption Known density ratio rho(x) = dQ_X/dP_X
    Assumed in Section 2.1 and Algorithm 1; Section 7 explicitly says the effect of estimated density ratios on inference is unresolved.
  • domain assumption Assumption 4.1: uniformly bounded noise |epsilon| <= sigma and E[epsilon^2|X] >= sigma^2
    Used throughout the proof for Bernstein-type bounds and for the boundedness of the influence elements U_i.
  • domain assumption Assumption 4.2: source condition f* = L_{P_{w0}}^s g* for some s in [0,1]
    Controls the smoothness of the regression function relative to the truncated weighted-source integral operator and is needed for bias rates.
  • domain assumption Assumption 4.3: effective dimension condition N_lambda
    Bounds the effective dimension of the kernel under the target measure and appears in all Gaussian and bootstrap approximation errors.
  • domain assumption Assumption 4.4: regularized transfer coverage ||L_Q^{1/2}(L_{P_w}+lambda I)^{-1/2}|| <= C1
    Ensures that the weighted-source regularized risk controls the target L2 norm; verified under bounded, sub-exponential, or bounded-moment density ratio conditions in the appendix.
  • domain assumption Assumption 4.5: non-degenerate covariance and rho(X) >= rho_0 P-almost surely
    Prevents the Gaussian covariance from degenerating in target geometry and is used for anti-concentration; Algorithm 1 does not implement the thinning step described in the text.
  • domain assumption Compact covariate space and bounded kernel sup_x k(x,x) <= kappa^2
    Invoked at the start of Section 2.2 and used repeatedly for point evaluation bounds and trace-class properties of the integral operators.
  • standard math Mercer's theorem and standard operator inequalities
    Used to diagonalize the kernel integral operator and to justify eigenvalue comparisons through the min-max principle and Cordes inequality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Nonparametric Goodness-of-fit Testing under Covariate Shift." pith.science (2026). https://pith.science/paper/7O7BTNET

@misc{pith2026260804860,
  author       = {Pith},
  title        = {Pith review of: Nonparametric Goodness-of-fit Testing under Covariate Shift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7O7BTNET}},
  note         = {Machine review of arXiv:2608.04860}
}
read the original abstract

This paper develops procedures for nonparametric goodness-of-fit testing under covariate shift, where labelled data are drawn from a source population but goodness-of-fit is evaluated for a target population. The distribution mismatch is quantified by either a bounded moment condition or a sub-exponential tail condition on the target-to-source density ratio. Our method combines truncated importance-weighting kernel ridge regression with a multiplier bootstrap to construct confidence sets for the regression function. The truncation stabilizes the importance- weighting kernel ridge regression as well as the bootstrap calibration, making our approach applicable even when the density ratio has heavy tails. We prove nonasymptotic validity and sharpness of the resulting confidence sets under suitable operator compatibility conditions, and establish explicit error rates for coverage probability under specific conditions on the target- to-source density ratio and on the spectral decay of the kernel integral operator. Numerical experiments corroborate our theoretical findings.

Figures

Figures reproduced from arXiv: 2608.04860 by the authors.

Figure 1
Figure 1. truncating large density-ratio weights ρ(x) := (dQ/dP)(x) stabilizes the finite-sample performance of the importance-weighting kernel ridge regression (IWKRR). Source and target dis￾tributions are truncated normal TN(−1, 0.25; [−2, 2]) and TN(1, 0.25; [−2, 2]), respectively. Panel (a) summarizes the validation-tuned errors of KRR, IWKRR, and truncated IWKRR. The true regression function is f∗(x) = |x| . Panel (b) sh… view at source ↗
Figure 2
Figure 2. Bootstrap distribution of the L 2 (Q) error. The figure compares the true sampling distri￾bution of √ n∥fb− f∗∥L2(Q) with the conditional bootstrap distribution of ∥B∥L2(Q) . 5.3 Empirical coverage and size We evaluate the empirical coverage of the feasible confidence ball CSα(fb) with the critical value cˆα,n estimated from 250 multiplier-bootstrap repetitions. Under the null hypothesis f0 = f∗, we record the relat… view at source ↗
Figure 3
Figure 3. Empirical power under fixed alternatives [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figures from the paper (21 more)
Figure 3
Figure 3. Figure 3: Empirical power under fixed alternatives [PITH_FULL_IMAGE:figures/full_fig_p022_3.png]
Figure 5
Figure 5. Figure 5: testing candidate functions for estimating household net worth based on IWKRR [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]
Figure 5
Figure 5. Figure 5: testing candidate functions for estimating household net worth based on IWKRR [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: testing whether the probability of holding stock equity is monotone with respect to the [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 6
Figure 6. Figure 6: testing whether the probability of holding stock equity is monotone with respect to the [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: Distribution of survey weights and their impact in 2022 implicate-1 records. Panel A [PITH_FULL_IMAGE:figures/full_fig_p037_7.png]
Figure 7
Figure 7. Figure 7: Distribution of survey weights and their impact in 2022 implicate-1 records. Panel A [PITH_FULL_IMAGE:figures/full_fig_p038_7.png]
Figure 8
Figure 8. Figure 8: empirical distribution of household net worth with and without re-weighting in Study 1, [PITH_FULL_IMAGE:figures/full_fig_p039_8.png]
Figure 8
Figure 8. Figure 8: empirical distribution of household net worth with and without re-weighting in Study 1, [PITH_FULL_IMAGE:figures/full_fig_p040_8.png]
Figure 9
Figure 9. Figure 9: block-level source–target comparison of covariates used in Study 1. The coordinate with [PITH_FULL_IMAGE:figures/full_fig_p039_9.png]
Figure 9
Figure 9. Figure 9: block-level source–target comparison of covariates used in Study 1. The coordinate with [PITH_FULL_IMAGE:figures/full_fig_p040_9.png]
Figure 10
Figure 10. Figure 10: Coordinate-level source–target comparison for all [PITH_FULL_IMAGE:figures/full_fig_p040_10.png]
Figure 10
Figure 10. Figure 10: Coordinate-level source–target comparison for all [PITH_FULL_IMAGE:figures/full_fig_p041_10.png]
Figure 11
Figure 11. Figure 11: probability of holding stock-equity as a function of usual income. Panel A shows the [PITH_FULL_IMAGE:figures/full_fig_p042_11.png]
Figure 11
Figure 11. Figure 11: probability of holding stock-equity as a function of usual income. Panel A shows the [PITH_FULL_IMAGE:figures/full_fig_p043_11.png]
Figure 12
Figure 12. Figure 12: empirical bootstrap distribution for Study 3 based on [PITH_FULL_IMAGE:figures/full_fig_p043_12.png]
Figure 12
Figure 12. Figure 12: empirical bootstrap distribution for Study 3 based on [PITH_FULL_IMAGE:figures/full_fig_p044_12.png]
Figure 13
Figure 13. Figure 13: equity-share allocation over financial wealth. The upper panel shows the IWKRR [PITH_FULL_IMAGE:figures/full_fig_p044_13.png]
Figure 13
Figure 13. Figure 13: equity-share allocation over financial wealth. The upper panel shows the IWKRR [PITH_FULL_IMAGE:figures/full_fig_p045_13.png]
Figure 14
Figure 14. Figure 14: survey-weighted distribution of the difference [PITH_FULL_IMAGE:figures/full_fig_p046_14.png]
Figure 14
Figure 14. Figure 14: survey-weighted distribution of the difference [PITH_FULL_IMAGE:figures/full_fig_p047_14.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 9 linked inside Pith

  1. [1]

    2017 , author =

    On some extensions of Bernstein's inequality for self-adjoint operators , journal =. 2017 , author =

    2017

  2. [2]

    arXiv preprint arXiv:2302.06578 , year=

    Kernel ridge regression inference , author=. arXiv preprint arXiv:2302.06578 , year=

    arXiv

  3. [3]

    2018 , publisher=

    High-dimensional probability: An introduction with applications in data science , author=. 2018 , publisher=

    2018

  4. [4]

    The Journal of Machine Learning Research , volume=

    A least-squares approach to direct importance estimation , author=. The Journal of Machine Learning Research , volume=. 2009 , publisher=

    2009

  5. [5]

    arXiv preprint arXiv:1206.4650 , year=

    Analysis of kernel mean matching under covariate shift , author=. arXiv preprint arXiv:1206.4650 , year=

    arXiv

  6. [6]

    2013 , publisher=

    The bootstrap and Edgeworth expansion , author=. 2013 , publisher=

    2013

  7. [7]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Statistical inferences of linear forms for noisy matrix completion , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2021 , publisher=

    2021

  8. [8]

    1994 , publisher=

    An introduction to the bootstrap , author=. 1994 , publisher=

    1994

Show all 70 references
  1. [9]

    The annals of statistics , volume=

    Some asymptotic theory for the bootstrap , author=. The annals of statistics , volume=. 1981 , publisher=

    1981

  2. [10]

    1990 , publisher=

    Spline models for observational data , author=. 1990 , publisher=

    1990

  3. [11]

    2013 , publisher=

    The nature of statistical learning theory , author=. 2013 , publisher=

    2013

  4. [12]

    2002 , publisher=

    Learning with kernels: support vector machines, regularization, optimization, and beyond , author=. 2002 , publisher=

    2002

  5. [13]

    Philosophical Transactions of the Royal Society of London

    Functions of positive and negative type, and their connection with the theory of integral equations , author=. Philosophical Transactions of the Royal Society of London. Series A , volume=

  6. [14]

    arXiv preprint arXiv:2303.04020 , year=

    When is Importance Weighting Correction Needed for Covariate Shift Adaptation? , author=. arXiv preprint arXiv:2303.04020 , year=

    arXiv

  7. [15]

    Spectral Theory of Linear Differential Operators and Comparison Algebras , publisher=

    Cordes, Heinz Otto , year=. Spectral Theory of Linear Differential Operators and Comparison Algebras , publisher=

  8. [16]

    Foundations of Computational Mathematics , volume=

    Optimal rates for the regularized least-squares algorithm , author=. Foundations of Computational Mathematics , volume=. 2007 , publisher=

    2007

  9. [17]

    Journal of statistical planning and inference , volume=

    Improving predictive inference under covariate shift by weighting the log-likelihood function , author=. Journal of statistical planning and inference , volume=. 2000 , publisher=

    2000

  10. [18]

    Machine Learning in Non-Stationary Environments: Introduction to Covariate Shift Adaptation , urldate =

    Masashi Sugiyama and Motoaki Kawanabe , publisher =. Machine Learning in Non-Stationary Environments: Introduction to Covariate Shift Adaptation , urldate =

  11. [19]

    IEEE Transactions on knowledge and data engineering , volume=

    A survey on transfer learning , author=. IEEE Transactions on knowledge and data engineering , volume=. 2009 , publisher=

    2009

  12. [20]

    Domain Adaptation for Medical Image Analysis: A Survey , volume=

    Guan, Hao and Liu, Mingxia , year=. Domain Adaptation for Medical Image Analysis: A Survey , volume=. IEEE Transactions on Biomedical Engineering , publisher=

  13. [21]

    Instance Weighting for Domain Adaptation in NLP

    Jiang, Jing and Zhai, ChengXiang. Instance Weighting for Domain Adaptation in NLP. Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics. 2007

    2007

  14. [22]

    The Annals of Statistics , volume=

    Optimally tackling covariate shift in RKHS-based nonparametric regression , author=. The Annals of Statistics , volume=. 2023 , publisher=

    2023

  15. [23]

    Advances in neural information processing systems , volume=

    Learning bounds for importance weighting , author=. Advances in neural information processing systems , volume=

  16. [24]

    arXiv preprint arXiv:0902.3430 , year=

    Domain adaptation: Learning bounds and algorithms , author=. arXiv preprint arXiv:0902.3430 , year=

    arXiv

  17. [25]

    Conference On Learning Theory , pages=

    Marginal singularity, and the benefits of labels in covariate-shift , author=. Conference On Learning Theory , pages=. 2018 , organization=

    2018

  18. [26]

    2025 , eprint=

    Label-Efficient Deep Learning in Medical Image Analysis: Challenges and Future Directions , author=. 2025 , eprint=

    2025

  19. [27]

    2024 , author =

    A comprehensive survey on deep active learning in medical image analysis , journal =. 2024 , author =

    2024

  20. [28]

    Advances in Neural Information Processing Systems , volume=

    Towards a unified analysis of kernel-based methods under covariate shift , author=. Advances in Neural Information Processing Systems , volume=

  21. [29]

    International Conference on Machine Learning , pages=

    A new similarity measure for covariate shift with applications to nonparametric regression , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  22. [30]

    The Annals of Statistics , volume=

    Pseudo-labeling for kernel ridge regression under covariate shift , author=. The Annals of Statistics , volume=. 2026 , publisher=

    2026

  23. [31]

    Bernoulli , volume=

    Local convergence rates of the nonparametric least squares estimator with applications to transfer learning , author=. Bernoulli , volume=. 2024 , publisher=

    2024

  24. [32]

    Anti-concentration and honest, adaptive confidence bands , volume=

    Chernozhukov, Victor and Chetverikov, Denis and Kato, Kengo , year=. Anti-concentration and honest, adaptive confidence bands , volume=. The Annals of Statistics , publisher=

  25. [33]

    The Annals of Statistics , number =

    Victor Chernozhukov and Denis Chetverikov and Kengo Kato , title =. The Annals of Statistics , number =

  26. [34]

    Stochastic Processes and their Applications , volume=

    Empirical and multiplier bootstraps for suprema of empirical processes of increasing complexity, and related Gaussian couplings , author=. Stochastic Processes and their Applications , volume=. 2016 , publisher=

    2016

  27. [35]

    The Annals of Probability , number =

    Victor Chernozhukov and Denis Chetverikov and Kengo Kato , title =. The Annals of Probability , number =

  28. [36]

    Conference on Learning Theory , pages=

    Strong Gaussian approximation for the sum of random vectors , author=. Conference on Learning Theory , pages=. 2022 , organization=

    2022

  29. [37]

    Zaitsev, A. Yu. , title =. Theory of Probability & Its Applications , volume =

  30. [38]

    The Annals of Statistics , author=

    Local and global asymptotic inference in smoothing spline models , volume=. The Annals of Statistics , author=. 2013 , month=

    2013

  31. [39]

    arXiv preprint arXiv:1708.04753 , year=

    Frequentist coverage and sup-norm convergence rate in Gaussian process regression , author=. arXiv preprint arXiv:1708.04753 , year=

    arXiv

  32. [40]

    arXiv preprint arXiv:2501.01610 , year=

    Bootstrap Nonparametric Inference under Data Integration , author=. arXiv preprint arXiv:2501.01610 , year=

    arXiv

  33. [41]

    The Annals of Statistics , volume=

    On nonparametric confidence intervals , author=. The Annals of Statistics , volume=. 1997 , publisher=

    1997

  34. [42]

    The Annals of Statistics , number =

    Christopher Genovese and Larry Wasserman , title =. The Annals of Statistics , number =

  35. [43]

    The Annals of Statistics , number =

    Gin. The Annals of Statistics , number =

  36. [44]

    Tony Cai and Mark G

    T. Tony Cai and Mark G. Low , title =. The Annals of Statistics , number =

  37. [45]

    Mathematical Methods of Statistics , volume=

    Nonparametric confidence set estimation , author=. Mathematical Methods of Statistics , volume=. 2003 , publisher=

    2003

  38. [46]

    The Annals of Statistics , volume=

    Honest confidence regions for nonparametric regression , author=. The Annals of Statistics , volume=. 1989 , publisher=

    1989

  39. [47]

    The Annals of Statistics , volume=

    Adaptive hypothesis testing using wavelets , author=. The Annals of Statistics , volume=. 1996 , publisher=

    1996

  40. [48]

    The Annals of Statistics , number =

    James Robins and Aad van der Vaart , title =. The Annals of Statistics , number =

  41. [49]

    Bernoulli , volume=

    Large ball probabilities, Gaussian comparison and anti-concentration , author=. Bernoulli , volume=. 2019 , publisher=

    2019

  42. [50]

    2012 , publisher=

    Machine learning in non-stationary environments: Introduction to covariate shift adaptation , author=. 2012 , publisher=

    2012

  43. [51]

    international conference on machine learning , pages=

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning , author=. international conference on machine learning , pages=. 2016 , organization=

    2016

  44. [52]

    Journal of Machine Learning Research , volume=

    Covariate Shift Adaptation by Importance Weighted Cross Validation , author=. Journal of Machine Learning Research , volume=

  45. [53]

    2016 23rd International Conference on Pattern Recognition (ICPR) , pages=

    On Regularization Parameter Estimation under Covariate Shift , author=. 2016 23rd International Conference on Pattern Recognition (ICPR) , pages=. 2016 , doi=. 1608.00250 , archivePrefix=

    2016 arXiv

  46. [54]

    2018 24th International Conference on Pattern Recognition (ICPR) , pages=

    Effects of Sampling Skewness of the Importance-Weighted Risk Estimator on Model Selection , author=. 2018 24th International Conference on Pattern Recognition (ICPR) , pages=. 2018 , doi=. 1804.07344 , archivePrefix=

    2018 arXiv

  47. [55]

    2019 IEEE 29th International Workshop on Machine Learning for Signal Processing (MLSP) , pages=

    Robust importance-weighted cross-validation under sample selection bias , author=. 2019 IEEE 29th International Workshop on Machine Learning for Signal Processing (MLSP) , pages=. 2019 , organization=

    2019

  48. [56]

    Annals of the Institute of Statistical Mathematics , volume=

    Direct Importance Estimation for Covariate Shift Adaptation , author=. Annals of the Institute of Statistical Mathematics , volume=

  49. [57]

    Advances in Neural Information Processing Systems , volume=

    Correcting Sample Selection Bias by Unlabeled Data , author=. Advances in Neural Information Processing Systems , volume=

  50. [58]

    Algorithmic Learning Theory , series=

    Sample Selection Bias Correction Theory , author=. Algorithmic Learning Theory , series=. 2008 , publisher=

    2008

  51. [59]

    Advances in Neural Information Processing Systems , volume=

    Conformal Prediction Under Covariate Shift , author=. Advances in Neural Information Processing Systems , volume=

  52. [60]

    The Annals of Statistics , volume=

    Conformal Prediction Beyond Exchangeability , author=. The Annals of Statistics , volume=. 2023 , doi=

    2023

  53. [61]

    arXiv preprint arXiv:2603.05897 , year=

    A Minimax Theory of Nonparametric Regression Under Covariate Shift , author=. arXiv preprint arXiv:2603.05897 , year=

  54. [62]

    2025 , eprint=

    High-Dimensional Importance-Weighted Information Criteria: Theory and Optimality , author=. 2025 , eprint=. doi:10.48550/arXiv.2505.06531 , url=

    2025 doi arXiv

  55. [63]

    Probability Theory and Related Fields , volume=

    Nearby Variables with Nearby Conditional Laws and a Strong Approximation Theorem for Hilbert Space Valued Martingales , author=. Probability Theory and Related Fields , volume=. 1991 , doi=

    1991

  56. [64]

    Review of Economics and Statistics , volume=

    Nonparametric Tests for Treatment Effect Heterogeneity , author=. Review of Economics and Statistics , volume=

  57. [65]

    Journal of the Royal Statistical Society: Series B (Statistical Methodology) , volume=

    Randomization Inference for Treatment Effect Variation , author=. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , volume=

  58. [66]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , pages=

    Don't Stop Pretraining: Adapt Language Models to Domains and Tasks , author=. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , pages=

  59. [67]

    Nature , volume=

    Large Language Models Encode Clinical Knowledge , author=. Nature , volume=. 2023 , doi=

    2023

  60. [68]

    Advances in Neural Information Processing Systems , volume=

    LegalBench: A Collaboratively Built Benchmark for Measuring Legal Reasoning in Large Language Models , author=. Advances in Neural Information Processing Systems , volume=. 2023 , doi=

    2023

  61. [69]

    Theory of Probability and Its Applications , volume=

    Remarks on Inequalities for Large Deviation Probabilities , author=. Theory of Probability and Its Applications , volume=. 1986 , doi=

    1986

  62. [70]

    2023 , howpublished=

    2023

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.