Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

Wild refitting for black box prediction

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Single wild refit bounds a black-box predictor's excess risk.

desk verdict A clever one-refit method for optimism estimation that is undermined by an unquantified, self-referential pilot error term in the main theorems. read the letter →

arxiv 2506.21460 v2 pith:WF3ZELMY submitted 2025-06-26 stat.ML cs.LGmath.STstat.TH

classification stat.MLcs.LGmath.STstat.TH MSC 62G0862J0762G09
keywords wildrefittingblack-boxpredictionexcessriskupperboundRademachersymmetrizationnonparametricregressionbootstrapplug-and-playdenoisingestimationerror
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 introduces a procedure called wild refitting that produces a high-probability upper bound on the instance-wise mean-squared prediction error of a penalized least-squares estimator, using only a single dataset and black-box access to the fitting method. The procedure computes residuals, multiplies them by random signs, rescales them by a noise factor, and refits the same method on the recentered problem; the resulting wild optimism can then be used to upper-bound the true optimism, and hence the excess risk. The advantage over hold-out, cross-validation, and bootstrap methods is that it is non-asymptotic, makes just one refitting call, and tolerates heteroskedastic and heavy-tailed noise. The paper also shows how to upper-bound the unknown estimation error in terms of a quantity called wild complexity, so the bound can be made fully computable.

What carries the argument

The central objects are the wild optimism and the wild complexity. The wild optimism is the Rademacher-symmetrized inner product between the residual-based noise vector and the displacement of the refit from the original predictor, and the wild complexity Wn(r) is its supremum over a radius-r ball around the predictor; Lemma 1 shows the two coincide for the refit whose distance from the original predictor equals r. The argument is carried by firm non-expansiveness of the estimator around the regression function, which makes projections onto convex sets admissible, and by concentration inequalities for convex Lipschitz functions of Rademacher variables, which give the exponential tails.

What would settle it

Take a deep plug-and-play denoiser on a fixed image, draw a set of perturbation vectors u, and numerically check inequality (11a); if any perturbation yields a squared refit distance noticeably larger than the inner product, the estimation-error bound in Theorem 2 and the resulting excess-risk upper bound lose their stated foundation.

Watch

Extended reading notes

Core claim

The central claim is that, for fixed design with symmetric independent noise, the true optimism of a penalized least-squares estimator is bounded by the wild optimism from a single refit, provided the wild noise scale is chosen so the refit lies at exactly twice some radius that dominates the estimation error. More precisely, Theorem 1 gives Opt*(fhat) ≤ gOpt(f_rho) + H_n(t) + A_n(ef) with probability at least 1 − 4 exp(−$t^{2}$), where the deviation term decays like t times the sup-norm of the noise divided by the square root of the sample size, and the pilot error term is small whenever the recentering function is a good estimate of the regression function. Theorem 2 closes the loop by showing the estimation error itself can be upper-bounded by a computable envelope built from the wild complexity, so the procedure yields a self-contained, non-asymptotic, high-probability upper bound on the instance-wise excess risk. The proof rests on Rademacher symmetrization of residuals and on sharp concentration properties of convex Lipschitz functions of Rademacher variables, in contrast to the asymptotic distributional guarantees of the classical wild bootstrap.

Load-bearing premise

The load-bearing premise is that the black-box estimator is firmly non-expansive around the true regression function: for any perturbation of the response vector, the squared distance between the two refits must be no larger than the inner product of that perturbation with their difference.

Editorial extensions

If this is right

  • One black-box call to the fitting method is enough to obtain a non-asymptotic upper bound on the excess risk, so risk estimation no longer requires repeated refitting as in a full bootstrap.
  • Model selection can be performed by minimizing the wild-refitting upper bound over tuning parameters, and the paper demonstrates this on nuclear-norm-constrained matrix recovery and on deep plug-and-play image denoisers.
  • The bound holds for heteroskedastic and heavy-tailed noise, with the noise entering only through its sup-norm, so the method remains meaningful when only a finite polynomial moment is available.
  • The estimation-error bound can be optimized over the wild noise scale by balancing the two terms in the envelope max(r, Wn(r)/r), giving a concrete rule for choosing rho.
  • If the estimator is firmly non-expansive and the pilot error term is controlled, the wild-refitting upper bound is valid for a single realization of the random signs, rather than only in distribution or on average.

Reading between the lines

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

  • Beyond the paper, the same recentering and symmetrization scheme could be extended to random-design settings by first conditioning on the realized covariates and then averaging over their distribution, though the fixed-design analysis here does not cover that averaging step.
  • The firm non-expansiveness condition is verified for convex projections but not for the deep plug-and-play denoiser used in the demonstrations, so a direct numerical check of inequality (11a) on those denoisers would be a natural test of whether the stated guarantee transfers to that setting.
  • Choosing the recentering function ef differently from the fitted predictor fhat could reduce the pilot error term A_n(ef), suggesting a two-stage wild refitting scheme in which a preliminary pilot estimate is used to shrink that residual term before the main refit.
  • The method's reliance on least-squares structure suggests that analogous wild-refitting bounds might be derivable for other convex loss-based M-estimators, with the optimism identity replaced by a Bregman-divergence analogue.
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

4 major / 4 minor

Summary. The paper introduces 'wild refitting,' a black-box procedure for producing non-asymptotic, high-probability upper bounds on the instance-wise excess risk of penalized least-squares estimators. The procedure forms residuals from a recentering pilot estimator, multiplies them by Rademacher signs and a scale rho, refits the original method to these wild responses, and uses the resulting wild optimism to bound the true optimism. The main theoretical results are Theorem 1, which bounds true optimism by wild optimism plus a deviation term and a pilot error term A_n(ef), and Theorem 2, which uses the same machinery to bound the estimation error ||fhat - fdagger||_n. Numerical demonstrations are given for low-rank NRSFM recovery, plug-and-play image denoising with deep priors, and kernel sketching.

Significance. The idea is attractive: a single black-box refit with Rademacher-symmetrized residuals, combined with sharp concentration for Lipschitz functions of Rademacher variables, could give finite-sample upper bounds under heteroskedastic and heavy-tailed noise. If the advertised bound were fully explicit, the paper would be a useful contribution to risk estimation for complex estimators where Stein-type or cross-validation methods are not available. The paper also gives an explicit algorithmic recipe and empirically illustrates the method on several nontrivial problems. The main current weakness is that the central theorems contain an unquantified pilot error term, which for the default pilot choice is proportional to the square root of the very excess risk being bounded; this prevents the stated results from being the closed-form high-probability upper bounds promised in the abstract and in Section 1.2.

major comments (4)
  1. [Theorem 1, Eq. (14)-(15b)] The advertised upper bound is not explicit because the pilot error A_n(ef) is never bounded. For the default and simplest choice ef = fhat, the supremum in (15b) over the full ball B_{2r}(fhat) is exactly 2r ||fhat - f*||_n, so the theorem reduces to Opt*(fhat) <= wild optimism + H_n(t) + 2r ||fhat - f*||_n. Since the excess risk is E = ||fhat - f*||_n^2, the RHS depends on sqrt(E), yielding an implicit inequality that the paper neither solves nor supplements with conditions under which the pilot error is dominated. Section 3.2 only says that 'we can expect' the pilot error to be dominated when ef is a reasonable estimate; this is a heuristic, not a theorem. The abstract's claim of a high-probability upper bound on prediction error is therefore not established by the stated results.
  2. [Theorem 2, Eq. (16a)-(16b)] The estimation-error bound suffers from the same unquantified A_n(ef) term and is also not a closed-form bound because brn^2 appears on both sides. In (16a), the term 6||w||_infty/t * brn^2 is retained without any condition ensuring t large enough relative to ||w||_infty to make the coefficient less than 1; the assumption t >= 3 alone does not suffice. Thus the statement that one can 'compute upper bounds on the true error' by varying rho in Section 3.4, as illustrated in Figure 3, is not justified by the displayed inequalities unless additional control on A_n and on ||w||_infty/t is supplied.
  3. [Section 3.1.1, assumption (11a)] The central assumption of firm non-expansiveness is not verified for the estimators used in the experiments. Eq. (10) restricts the theory to hard constraint sets C (indicator penalties), and the paper explicitly defers Lagrangian regularization to future work. Yet Section 2.4 uses Tikhonov and TV penalties of the form P_q(f) = sum |f(x_{i+1}) - f(x_i)|^q, and Section 4.2 uses a plug-and-play deep denoiser. For these procedures, (11a) is neither proved nor numerically checked, and Lemma 5 and Theorem 2 depend directly on it. Consequently, the numerical demonstrations in Figures 5-8 are not covered by the theory as written.
  4. [Section 4] The experiments report MSE curves for different rho but never report the value of the pilot error A_n(ef), never check whether the radius condition ||fhat - fdagger||_n <= r is met, and do not assess coverage of the stated 1 - 4e^{-t^2} guarantee. The statements 'consistent with our theory' in Section 4.2 are therefore not supported by the numerical evidence in the paper.
minor comments (4)
  1. [Abstract and Section 1] There are typographical errors: 'computionally' in the abstract and 'block-box' in the abstract and Section 1.2; these should be corrected.
  2. [Section 5.2.3] In the last sentence of the proof of Lemma 4, the phrase 'noise scale rho chosen to ensure ||f^D_rho - fhat||_n = 2rho' should read '= 2r', to match the lemma statement.
  3. [Section 3.2] The term H_n(t) in Eq. (15a) is said to scale as t||w||_infty / sqrt(n), but the displayed expression has 2||w||_infty t / sqrt(n) multiplied by (3r + ||fdagger - f*||_n); the scaling is fine, but the constant is slightly larger than the text suggests.
  4. [Section 4.1] The text says wild noise scales rho in {1.0, 1.1, 2.2} were used, while Figure 5 is described as showing rho in {1.0, 1.1, 1.2}; this apparent mismatch should be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the core inequalities are genuine high-probability bounds, not fitted predictions or self-citation chains.

full rationale

The paper's derivation chain is not circular. Theorem 1 is proved as a high-probability inequality: the true optimism is bounded by the wild optimism plus a deviation term H_n(t) and a pilot-error remainder A_n(tilde f). The wild optimism is computed from the same data and the same Rademacher signs that produce the refit, but the theorem is a statement over the Rademacher randomness and does not assume the target error on the right-hand side. Lemma 1 ('wild optimism equals wild complexity at the refit radius') is a proved identity, not an ansatz or a relabeling. Lemma 4 controls the ideal wild complexity by the computable wild optimism plus the pilot-error term via the triangle inequality; this is a valid remainder, not an assumed conclusion. The self-citations in the paper are to standard concentration results (Wainwright's book, reference [28]) and to an application used only for illustration (reference [30]); neither is load-bearing in a way that reduces the argument to the authors' prior work. There is no fitted parameter renamed as a prediction: the noise scale rho is analyzed, not fitted to make the bound hold, and the experiments vary rho and compare to true MSE rather than claiming a tuned rho is predicted. The one limitation that should be flagged explicitly is the treatment of the pilot-error term in Section 3.2: the paper says only 'as long as the pilot function tilde f is a reasonable estimate of f* ... we can expect that the pilot error is dominated by gOpt', which is a heuristic rather than a theorem. Moreover, for the default choice tilde f = f_hat, Eq. (15b) reduces to A_n = 2r ||f_hat - f*||_n, so the right-hand side of Theorem 1 contains sqrt(E) through the very quantity being bounded. This makes the stated bound implicit rather than a closed-form, computable upper bound unless the quadratic inequality is solved or A_n is otherwise controlled, and Theorem 2 inherits the same remainder. That is a genuine correctness/utility gap, but it is not circularity: the theorem is still a valid probabilistic inequality, and the pilot-error term is explicitly presented as a remainder rather than silently assumed away. Therefore the circularity score is 0.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central derivation rests on one tuning parameter (rho), one pilot choice (ef), and five substantive assumptions: symmetric independent noise, firm non-expansiveness, indicator-penalty structure, convexity for the simplified bound, and fixed design. No new particles, forces, or unobserved entities are introduced; the wild refit is a procedure, not an entity.

free parameters (2)
  • Wild noise scale rho = rho in {1.0, 1.15, 1.30} (or {1.0, 1.1, 1.2}) per experiment
    The upper-bound guarantee holds only for a suitable rho; experiments select rho by trying a small grid and reporting which value gives an upper bound, rather than applying the theoretical balance condition.
  • Pilot/recentering estimator ef = ef = fhat in the default instantiation
    The unquantified pilot error A_n(ef) depends on this choice; no optimality criterion or estimation procedure for ef is provided.
assumptions (6)
  • domain assumption Noise vector (w1,...,wn) is conditionally independent with each wi symmetric about zero (eq. 11b).
    Used in the proof of Lemma 2 to represent wi as epsilon_i times |wi|; the sentence about losing a factor of two without symmetry is asserted but not proved.
  • domain assumption The estimator M is firmly non-expansive around f* (eq. 11a).
    Central to Lemma 5 and Theorem 2; holds for convex projections, but is not established for the plug-and-play or Lagrangian methods used in the demonstrations.
  • domain assumption The penalty P is the indicator of a compact convex set C (eq. 10).
    Lemma 1 and the main theorems are proved for this constrained form; Section 6 says Lagrangian regularizers are future work.
  • domain assumption The function class F is convex for the strengthened bound (16b).
    Lemma 7 uses convexity of F to prove concavity of the wild complexity W_n, which is needed for the simplified bound.
  • domain assumption Fixed design: covariates are conditioned upon and all norms are empirical over the given x_i.
    The entire analysis is conditional on the covariates; Section 6 lists random design as future work.
  • standard math Concentration inequality for convex Lipschitz functions of Rademacher variables (Theorem 3.24 of reference [28]).
    Imported as a black box in Lemma 2 and Lemma 6; this is standard, but the paper does not reprove it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Wild refitting for black box prediction." pith.science (2026). https://pith.science/paper/WF3ZELMY

@misc{pith2026250621460,
  author       = {Pith},
  title        = {Pith review of: Wild refitting for black box prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WF3ZELMY}},
  note         = {Machine review of arXiv:2506.21460}
}
abstract

We describe and analyze a computionally efficient refitting procedure for computing high-probability upper bounds on the instance-wise mean-squared prediction error of penalized nonparametric estimates based on least-squares minimization. Requiring only a single dataset and black box access to the prediction method, it consists of three steps: computing suitable residuals, symmetrizing and scaling them with a pre-factor $\rho$, and using them to define and solve a modified prediction problem recentered at the current estimate. We refer to it as wild refitting, since it uses Rademacher residual symmetrization as in a wild bootstrap variant. Under relatively mild conditions allowing for noise heterogeneity, we establish a high probability guarantee on its performance, showing that the wild refit with a suitably chosen wild noise scale $\rho$ gives an upper bound on prediction error. This theoretical analysis provides guidance into the design of such procedures, including how the residuals should be formed, the amount of noise rescaling in the wild sub-problem needed for upper bounds, and the local stability properties of the block-box procedure. We illustrate the applicability of this procedure to various problems, including non-rigid structure-from-motion recovery with structured matrix penalties; plug-and-play image restoration with deep neural network priors; and randomized sketching with kernel methods.

Figures

Figures reproduced from arXiv: 2506.21460 by the authors.

Figure 1
Figure 1. Top row: illustration of the family of functions f ∗ γ for different choices of the smoothing parameter γ ≥ 0. Middle row: Corresponding function estimate fb with Tikhonov regularization. Bottom row: corresponding function estimates with TV regularization. We then applied both the Tikhonov and the TV estimates to observations from the model (8) for the γ-convolved function f ∗ γ over a range of convolution strengths… view at source ↗
Figure 2
Figure 2. Panels (a) and (b): Comparison of the true MSE E(fb) and wild estimates at noise scales ρ ∈ {1.0, 1.2, 1.4} for the Tikhonov estimate (panel (a)) and the TV estimate (panel (b)). (c) Plots of the MSE obtained by an oracle (red diamonds) that can choose optimally between the two estimators. (d) Comparison of model selection using the wild estimates to the oracle error. In panel (c) of [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 3
Figure 3. The family of inequalities defined by the functions (19) can be used to compute upper bounds on the true error rbn = ∥fb− f †∥n. Illlustrated for two different prediction problems: (a) Randomized sketching in kernel regression [30]. (b) Non-rigid shape-from-motion (Section 4.1). Observe that the bound (18) holds for each choice of noise scale ρ; consequently, we can op￾timize this choice so as to obtain the tightest… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Comparison of reconstructed frames to ground truth for motion capture data (“Pick￾up” sequence from the CMU motion capture dataset [5]). Reconstructions obtained by minimizing Frobenius norm subject to a nuclear norm constraint (20b) with three different choices of the…
Figure 5
Figure 5. Figure 5: Plots of the true MSE (red solid line with diamonds) versus the log radius (log r), and comparison with wild refitting estimates of the MSE for three different choices of the noise scale ρ ∈ {1.0, 1.1, 1.2}. Consistent with our theory, a sufficiently large value of ρ l…
Figure 6
Figure 6. Figure 6: Performance of wild refitting for MSE estimation for plug-and-play denoising for Gaussian noise. Top plots: Denoised images fbγ as the smoothing parameter is decreased from left to right. Panel (a): Plots of the true MSE and wild refitting estimates versus the smoothin…
Figure 7
Figure 7. Figure 7: Performance of wild refitting for MSE estimation for plug-and-play denoising with Student-t noise with d = 6 degrees of freedom. Top plots: Denoised images fbγ as the smoothing parameter is decreased from left to right. Panel (a): Plots of the true MSE and wild refitti…
Figure 8
Figure 8. Figure 8: Performance of wild refitting for MSE estimation for plug-and-play denoising with a heterogeneous form of noise: the middle strip has high variance, with the outer strips having very low variance. Top plots: Denoised images fbγ as the smoothing parameter is decreased f…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Upper Confidence Bounds for the Prediction Error of Kernel Ridge Regression via Gaussian Refitting

    stat.ME 2026-07 accept novelty 8.0 of 10

    Gaussian-refit movement quantiles plus a bias term give a computable, rate-optimal upper confidence bound on the realized prediction error of kernel ridge regression under symmetric noise.

  2. Optimizing the Preconditioner: A Black-box Online-to-Nonconvex Conversion with Static Regret Minimization Oracles

    cs.LG 2026-07 conditional novelty 7.0 of 10

    An OCO algorithm with only O(√T) static regret, pluggable as a preconditioner selector, recovers the classical O(1/√T) stationarity rate on smooth stochastic nonconvex problems and the O(T^{-2/7}) rate on nonsmooth ones.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages · cited by 2 Pith papers

  1. [1]

    Akhter, Y

    I. Akhter, Y. Sheikh, S. Khan, and T. Kanade. Trajectory space: A dual representation of nonrigid structure from motion. IEEE Trans. PAMI, 33(7):1442–1456, 2011

  2. [2]

    Arlot and A

    S. Arlot and A. Celisse. A survey of cross-validation procedures for model selection. Statistics Surveys, 4:40–79, 2010

  3. [3]

    Bates, T

    S. Bates, T. Hastie, and R. Tibshirani. Cross-validation: What does it estimate and how well does it do it? Journal of the American Statistical Association , 119(546):1434–1445, 2024. 30

  4. [4]

    Bregler, A

    C. Bregler, A. Hertzmann, and H. Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 690–696, 2000

  5. [5]

    Cmu graphics lab motion capture database

    Carnegie Mellon University. Cmu graphics lab motion capture database. http://mocap.cs. cmu.edu/, 2003

  6. [6]

    S. H. Chan, X. Wang, and O. A. Elgendy. Plug-and-play admm for image restoration: Fixed- point convergence and applications. IEEE Transactions on Computational Imaging , 3(1):84– 98, 2016

  7. [7]

    Y. Dai, H. Li, and M. He. A simple prior-free method for non-rigid structure-from-motion factorization. International Journal of Computer Vision , 107(2):101–122, 2014

  8. [8]

    B. Efron. The estimation of prediction error: Covariance penalties and cross-validation. Jour- nal of the American Statistical Association , 99(467):619–632, 2004

Show all 30 references
  1. [9]

    Efron and R

    B. Efron and R. J. Tibshirani. An Introduction to the Bootstrap. Chapman & Hall/CRC, New York, 1993

  2. [10]

    S. Geisser. The predictive sample reuse method with applications. Journal of the American Statistical Association, 70(350):320–328, 1975

  3. [11]

    Graikos, V

    A. Graikos, V. Papyan, and J. Mairal. Diffusion models as plug-and-play priors. In Advances in Neural Information Processing Systems , volume 35, pages 29213–29226, 2022

  4. [12]

    Hiriart-Urruty and C

    J. Hiriart-Urruty and C. Lemar´ echal.Fundamentals of convex analysis . Springer-Verlag, New York, 2001

  5. [13]

    Hurault, A

    S. Hurault, A. Leclaire, and N. Papadakis. Proximal denoiser for convergent plug-and-play op- timization with nonconvex regularization. In Proceedings of the 39th International Conference on Machine Learning , pages 9483–9505. PMLR, 2022

  6. [14]

    U. S. Kamilov, C. A. Bouman, G. T. Buzzard, and B. Wohlberg. Plug-and-play methods for integrating physical and learned models in computational imaging. IEEE Signal Processing Magazine, 40(1):85–97, 2023

  7. [15]

    M. Ledoux. The Concentration of Measure Phenomenon . Mathematical Surveys and Mono- graphs. American Mathematical Society, Providence, RI, 2001

  8. [16]

    K.-C. Li. Resampling residuals: Bootstrap methods for regression models. Journal of the American Statistical Association, 84(406):881–889, 1989

  9. [17]

    R. Y. Liu. Bootstrap procedures under some non-i.i.d. models. Annals of Statistics, 16(4):1696– 1708, 1988

  10. [18]

    E. Mammen. Bootstrap and wild bootstrap for high dimensional linear models. Annals of Statistics, 21(1):255–285, 1993

  11. [19]

    Bootstrap, wild bootstrap, and asymptotic normality

    Enno Mammen. Bootstrap, wild bootstrap, and asymptotic normality. Probability Theory and Related Fields, 93(4):439–455, 1992. 31

  12. [20]

    R. T. Rockafellar. Convex Analysis. Princeton University Press, Princeton, 1970

  13. [21]

    Rockafellar and R

    R.T. Rockafellar and R. J-B Wets. Variational Analysis , volume 317. Springer Science & Business Media, 2009

  14. [22]

    C. Stein. Estimation of the mean of a multivariate normal distribution. Annals of Statistics , 9(6):1135–1151, 1981

  15. [23]

    M. Stone. Cross-validatory choice and assessment of statistical predictions. Journal of the Royal Statistical Society. Series B , 36(2):111–147, 1974

  16. [24]

    Tibshirani and R

    R. Tibshirani and R. J. Tibshirani. A bias correction for the minimum error rate in cross- validation. Annals of Applied Statistics , 3(2):822–829, 2009

  17. [25]

    van de Geer

    S. van de Geer. Empirical Processes in M-Estimation . Cambridge University Press, 2000

  18. [26]

    A. W. van der Vaart and J. Wellner. Weak Convergence and Empirical Processes . Springer- Verlag, New York, NY, 1996

  19. [27]

    S. V. Venkatakrishnan, C. A. Bouman, and B. Wohlberg. Plug-and-play priors for model based reconstruction. In 2013 IEEE Global Conference on Signal and Information Processing , pages 945–948. IEEE, 2013

  20. [28]

    M. J. Wainwright. High-dimensional statistics: A non-asymptotic viewpoint . Cambridge Uni- versity Press, Cambridge, UK, 2019

  21. [29]

    C. F. J. Wu. Jackknife, bootstrap and other resampling methods in regression analysis. Annals of Statistics , 14(4):1261–1295, 1986

  22. [30]

    Y. Yang, M. Pilanci, and M. J. Wainwright. Randomized sketches for kernels: Fast and optimal non-parametric regression. Annals of Statistics , 45(3):991–1023, 2017. 32

Pith tools

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