Pith. sign in

REVIEW 2 major objections 4 minor 1 cited by

Two-Stage Robust Sparse Gradient Methods for Regression Under Heavy-Tailed Designs

T0 review · 2 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read The paper claims that in heavy-tailed sparse regression, noise tails determine the estimation error while design tails determine the sample complexity, and its RIGHT algorithm matches both minimax limits.

desk verdict Novel and worth engaging, but the main theorem's no-sample-splitting proof has a trajectory-dependence hole; the decoupling claim is currently unsupported as written. read the letter →

arxiv 2601.05669 v2 pith:EOYUJXSZ submitted 2026-01-09 stat.ME

Kaiyuan Zhou , Xiaoyu Zhang , Wenyang Zhang , Di Wang This is my paper
classification stat.ME MSC 62J0762F35
keywords robustregressionheavy-taileddistributionsmedian-of-meansiterativehardthresholdinggradientdescentminimaxoptimalitysamplecomplexitysparseestimation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that high-dimensional sparse regression under heavy-tailed covariates and noise splits into two separable difficulties: the tail of the noise sets the best possible estimation error, while the tail of the design sets how many samples an iterative algorithm needs before it can start contracting. The proposed RIGHT algorithm — iterative hard thresholding driven by coordinate-wise median-of-means gradient estimates — is claimed to match the minimax lower bounds for both effects, working under only (2+2λ)-th design moments and (1+δ)-th noise moments, and without requiring the population risk to exist. A companion SRCG condition replaces the usual restricted strong convexity, so convergence is argued from the geometry of the gradient field rather than the curvature of a risk surface. A two-stage initialization via a robust Dantzig selector is also proposed. The same framework is extended to logistic and multi-response regression, where the bounded logistic gradient is claimed to remove the design-tail sample-complexity barrier.

What carries the argument

RIGHT replaces the empirical gradient in Iterative Hard Thresholding with a coordinate-wise median-of-means gradient estimator: split the data into K blocks, average the gradient within each block, and take the median coordinate-wise, with K ≍ log p. The convergence proof rests on two conditions: SRCG (Sparsity-Restricted Correlated Gradient), which says the population gradient aligns with the error direction and provides curvature even when the population risk is infinite, and SRS (Sparsity-Restricted Stability), which bounds the gradient estimator's deviation on sparse coordinate sets by φ‖θ−θ*‖₂ + γ. The hard-thresholding projection keeps iterates sparse, so both conditions only need to h

What would settle it

On a synthetic linear regression with heavy-tailed t-design (λ<1) and heavy-tailed t-noise (δ<1), run RIGHT without sample splitting across a grid of n and at every iteration record max_{|S|≤2s+s*} ‖g(θ_t)_S − E∇L(θ_t)_S‖₂ along with the iteration error. If the empirical error slope does not follow (log p/n)^{δ/(1+δ)} once n passes the claimed threshold, or if the trajectory hits a point where the fixed-θ SRS bound is violated by a constant factor, the central claim and the no-splitting proof are not supported.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that heavy-tailed sparse regression has two separate minimax axes. In linear regression, for design tail index λ ∈ (0,1] and noise tail index δ ∈ (0,1], RIGHT attains ‖θ_T − θ*‖₂ ≲ κ₋⁻¹ M_eff,δ √(2s+s*) (log p/n)^{δ/(1+δ)} once n exceeds a threshold that grows with M_{x,2+2λ}^{1/λ} s^{(1+λ)/(2λ)} log p. The matching lower bounds assert that the noise exponent δ/(1+δ) is unavoidable for accuracy and the n^{−λ/(1+λ)} multiplicative gradient-estimation rate is unavoidable for any gradient-based iterative method. The paper further claims the same machinery covers logistic regression at the standard √(s log p/n) rate under only finite second desi

Load-bearing premise

The load-bearing premise is that the sparsity-restricted stability of the MoM gradient holds uniformly for every iterate along the data-dependent optimization path; the proof only establishes it pointwise for fixed θ, and the union-bound argument used to extend it to the whole trajectory does not control the dependence of later iterates on the data.

Editorial extensions

If this is right

  • If Theorem 4.5 holds, the required sample size can be read off the design tail λ while the achievable precision is read off the noise tail δ, instead of treating heavy tails as one combined obstacle.
  • Robust gradient estimation can replace restricted strong convexity: the SRCG condition allows convergence guarantees even when the population risk is infinite and the noise has only slightly more than one moment.
  • The claimed no-sample-splitting analysis implies the full dataset can be reused at every iteration, avoiding the sample-efficiency loss of earlier robust gradient methods.
  • For logistic regression, the bounded gradient implies standard n^{−1/2} rates under only second design moments, formalizing the intuition that classification is more robust to leverage points.
  • For multi-response regression, the decoupling persists with sample complexity (m+log p) rather than m log p, which is favorable when the number of covariates is large.

Reading between the lines

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

  • Editorial inference: if the decoupling is real, practitioners could separately assess the noise tail to set the target accuracy and the design tail to size the sample, rather than estimating a joint heavy-tail model.
  • Editorial inference: the proof gap flagged in the reader pass — SRS is shown pointwise for fixed θ but not uniformly on the data-dependent trajectory — suggests the honest version of the method may need fresh samples for the refinement stage, which would change the stated sample-complexity constants.
  • Editorial inference: the SRCG/SRS modular recipe could transfer to other first-order heavy-tailed problems, such as sparse generalized linear models with unbounded losses, as long as a suitable robust gradient oracle exists.
  • Editorial inference: a testable extension is to compare RIGHT with a sample-split version on heavy-tailed designs; if the no-splitting theory is unsound, the split version should not be worse at the same total sample size.
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

2 major / 4 minor

Summary. The paper proposes RIGHT, an iterative hard-thresholding algorithm based on coordinate-wise median-of-means (MoM) gradient estimates, for high-dimensional sparse regression with heavy-tailed covariates and noise. The main theoretical claims are: (i) the estimation error is governed by the noise tail index δ, with rate O((log p/n)^{δ/(1+δ)}); (ii) the sample complexity required for algorithmic stability is governed by the design tail index λ; (iii) this decoupling is minimax optimal; and (iv) these guarantees hold without sample splitting and without requiring a finite population risk. Results are developed for linear, logistic, and multi-response regression, with matching lower bounds and numerical experiments. The central theorem is Theorem 4.5, which states that under sparsity-restricted moment and eigenvalue conditions, RIGHT attains the claimed rate with high probability.

Significance. If the proof of Theorem 4.5 were valid, the decoupling result would be notable and practically relevant: the algorithm is simple, the moment assumptions are substantially weaker than existing RSC-based analyses, and the paper ships code and empirical slope validations that are genuine, not curve-fitted constants. The lower-bound framework also gives useful phase-transition intuition. However, the central proof has a load-bearing gap: the uniform SRS bound along the data-dependent trajectory is not established, and the no-sample-splitting claim rests on an invalid union bound. The two-stage methodology and the decoupling narrative depend on this missing piece, so the main convergence guarantees are currently unsupported as written.

major comments (2)
  1. [Appendix B, proof of Theorem 4.5; Remark 4.7; Section 1.2] The no-sample-splitting claim is not proved. Proposition 4.3 establishes the SRS condition only for a fixed θ, with probability 2p exp(−K/8). In the proof of Theorem 4.5 the authors state that because θ_t varies with t, they need SRS uniformly for all t, and then take a union bound over t, choosing K so that each fixed-θ failure probability is p^{−ξ}/T. This union bound is invalid: θ_t is a function of the same data used to compute the MoM gradients at later steps, so the event {SRS fails at θ_t} is not controlled by the fixed-θ probability bound. Conditioning on θ_t changes the distribution of the blocks. Consequently the contraction condition ϕ<1/c2 is not guaranteed along the trajectory, and the sample complexity (4.1), the rate (4.2), and the statements in Section 1.2 and Remark 4.7 are not supported. A valid proof needs either a uniform SRS bound over an s-sparse ball (e.g., via a n
  2. [Section 5.2, Theorem 5.2 and Remark 5.3] The claimed algorithmic sample-complexity lower bound is not established as stated. Theorem 5.2 lower-bounds the error of estimating the population gradient G(P,θ) at a fixed θ. This is a statistical lower bound for the gradient-estimation sub-problem; it does not by itself imply that every first-order iterative method requires n ≳ (s*)^{(1+λ)/(2λ)} to converge, because the adaptive nature of the iterates and the information accumulated across iterations are not modeled. Remark 5.3 acknowledges this limitation, but the surrounding text ("fundamental sample complexity bottleneck", "nearly optimal for stability") goes beyond what the theorem proves. Either formalize a class of algorithms and prove an optimization lower bound, or soften the claim to a statement about the sub-problem.
minor comments (4)
  1. [Abstract vs. Remark 4.7] The abstract states that the method uses "delayed sample splitting" and reserves fresh batches for refinement, while Section 1.2 and Remark 4.7 explicitly claim the analysis avoids sample splitting and reuses the full dataset at every step. This is a substantive inconsistency that should be resolved in revision.
  2. [Appendix E, proof of Theorem 5.6, Step 2] The KL divergence for a heavy-tailed location shift should scale as nν^{(1+δ)/δ}, consistent with Lemma E.3, not nν^{1+δ} as written. The final rate δ/(1+δ) is correct, but the intermediate formula needs correction.
  3. [Theorem 4.17] The probability statement reads "with probability at least 1−p^{ξ}"; it should be 1−p^{−ξ} for ξ>0, matching Theorem 4.5.
  4. [Section 6.2, Table in Figure 3] The theoretical slopes for λ=0.09 and λ=0.19 are reported as −0.083 and −0.160, respectively, while the formula λ/(1+λ) gives −0.0826 and −0.1600. The table is consistent, but the caption or text might clarify that the empirical slopes are obtained via log-log linear regression over a finite range, so small deviations are expected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the main rates are derived from moment and eigenvalue assumptions, and the empirical slope studies are genuine validations.

full rationale

The paper's central derivation chain is self-contained. The deterministic Theorem 3.5 states a contraction result conditional on SRCG and SRS; Propositions 4.3 and 4.4 verify those conditions directly from Assumptions 4.1 and 4.2, not from the target error rate. The additive noise floor γ in Proposition 4.3 is a proved high-probability concentration bound for the MoM gradient, so the final n^{-δ/(1+δ)} rate in Theorem 4.5 is a consequence rather than an assumed input. Likewise, the sample-complexity condition (4.1) follows by imposing the proved multiplicative term ϕ < 1/c2, which is derived from the same SRS bound. In Section 6, the empirical slopes are obtained from independent simulations and compared to the theoretical exponents; they are not fitted constants renamed as predictions. The minimax lower bounds in Section 5 use standard Fano/packing constructions over the same moment class; constructing hard instances inside the assumed class is normal minimax methodology and does not make the upper bounds circular. There is no load-bearing self-citation: the overlapping-author reference (Zhang et al., 2025) is only contextual. The only notable concern is a proof-level issue in the union bound over data-dependent iterates in Appendix B, but that is a correctness/martingale gap, not a circular reduction of the conclusion to the inputs. Therefore no circularity step is present.

Assumptions & free parameters 4 free parameters · 8 assumptions · 0 invented entities

The statistical theory rests on sparse moment and eigenvalue assumptions on the unknown distribution, standard concentration lemmas, and an implicit uniform trajectory stability condition that is not proven. No new physical entities are introduced; SRCG is a regularity condition rather than an invented object.

free parameters (4)
  • Number of MoM blocks K = K ≍ log p (or c log p; c grid-searched from {0.5,1,2} in experiments)
    Number of blocks balances coordinate-wise concentration and union-bound probability. It is a hand-set tuning parameter, not estimated from the model, but it enters the rates in Theorems 4.5, 4.12, and 4.17.
  • Step size η = η = 2a η0 with η0 ∈ (0,1]
    Choice of step size enters the contraction factor and the stability condition φ < 1/c2. It is chosen by hand within a permissible interval.
  • Sparsity level s = s > c0 s*; s = 10 in simulations
    Hard-thresholding level must exceed the true sparsity by a constant factor; in practice selected via cross-validation, and in theory set to a multiple of the unknown s*.
  • Truncation thresholds τ_x, τ_yx = τ_x ≍ (n M_{x,2+2λ}/log d)^{1/(1+λ)}, τ_yx ≍ (n M_{yx,1+υ}/log d)^{1/(1+υ)}
    Used by the Robust Dantzig Selector initializer. They depend on unknown moment constants and are tuned in practice, though the theory states the scaling.
assumptions (8)
  • domain assumption Assumption 4.1: sparsity-restricted moments M_{x,2+2λ,s+s*} < ∞ and M_{ε,1+δ} < ∞
    Defines the heavy-tailed regime for linear regression: design has (2+2λ)-th moments, noise has (1+δ)-th moments.
  • domain assumption Assumption 4.2: sparse restricted eigenvalue condition on Σ_x
    Needed to establish SRCG for linear and multi-response regression and to control contraction.
  • domain assumption Assumption 4.9: M_{x,2,2(s+s*)} < ∞ for logistic regression
    Only finite second moments of the design are assumed in the logistic case.
  • domain assumption Assumption 4.15: sparsity-restricted moments for multi-response regression
    Analog of Assumption 4.1 for the matrix-valued response model.
  • ad hoc to paper Uniform SRS along the data-dependent trajectory
    Needed for Theorem 4.5 and the no-sample-splitting claim. Proposition 4.3 proves SRS only pointwise, and the proof's union bound over t does not control θ_t's dependence on the data.
  • standard math Lemma B.2: concentration inequality for sample means under (1+λ)-th moments
    Imported from Bubeck, Cesa-Bianchi and Lugosi (2013); used to derive MoM gradient bounds.
  • standard math Lemma A.1: near non-expansiveness of hard thresholding
    Restated from Li et al. (2016); used in the deterministic contraction proof.
  • ad hoc to paper Lemma E.3: existence of heavy-tailed location families with KL ≈ Δ^{(1+δ)/δ}
    Constructed for the lower bounds. The proof is a bulk/tail approximation and is not fully rigorous as written.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Two-Stage Robust Sparse Gradient Methods for Regression Under Heavy-Tailed Designs." pith.science (2026). https://pith.science/paper/EOYUJXSZ

@misc{pith2026260105669,
  author       = {Pith},
  title        = {Pith review of: Two-Stage Robust Sparse Gradient Methods for Regression Under Heavy-Tailed Designs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EOYUJXSZ}},
  note         = {Machine review of arXiv:2601.05669}
}
read the original abstract

We study high-dimensional sparse regression under simultaneous heavy-tailed covariates and noise. Heavy-tailed data affect sparse optimization in two different ways: extreme covariates can destabilize the gradient field during global localization, while heavy-tailed noise limits the final statistical accuracy during local refinement. Motivated by this two-phase structure, we propose two-stage RIGHT, a robust sparse first-order method based on coordinate-wise median-of-means (MoM) gradient estimation and delayed sample splitting. The MoM gradient estimator is computationally simple, compatible with hard-thresholded updates, and admits phase-adaptive concentration bounds whose rates depend on the current localization radius. Delayed splitting reuses data during global localization and reserves fresh batches for the shorter refinement stage, reducing the sample-splitting cost. The theoretical results reveal a decoupled rate structure: the design-tail index controls gradient stability and sample complexity, whereas the noise-tail index controls the final statistical rate. We also provide phase-wise lower-bound benchmarks showing that the design-driven localization barrier is intrinsic. Extensive simulation experiments and real data analysis showcase the efficacy of the proposed method over existing competitors.

Figures

Figures reproduced from arXiv: 2601.05669 by the authors.

Figure 1
Figure 1. Decoupled phase transitions in estimation error rate and sample complexity. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Estimation error rates of the RIGHT across different [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
Figure 3
Figure 3. Convergence rates of gradient estimation error across different [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Linear regression performance comparison. [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]
Figure 5
Figure 5. Figure 5: Logistic regression performance comparison. [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]
Figure 6
Figure 6. Figure 6: Multi-response regression performance comparison. [PITH_FULL_IMAGE:figures/full_fig_p023_6.png]
Figure 7
Figure 7. Figure 7: Q-Q plot of responses and histogram of features kurtosis from data set riboflavin. [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]
Figure 8
Figure 8. Figure 8: QQ-plot of residuals from RIGHT fit on real data. [PITH_FULL_IMAGE:figures/full_fig_p051_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Statistical Inference on Gradient Flows

    math.ST 2026-05 unverdicted novelty 7.0 of 10

    Proves uniform CLT for gradient flows in ERM and constructs an algorithm-aware, inversion-free covariance estimator for asymptotically valid time-uniform confidence intervals.

Reference graph

Works this paper leans on

44 extracted references · 1 linked inside Pith · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION format.url url empty "" url if FUNCTION article output.bibitem format.authors "author" output.check author format.key output output.year.check new.block format.title "title" output.check new.block crossref missing format.jour.vol output format.article.crossref output.nonnull format.pages output if ne...

  2. [2]

    , Matias , Yossi Y

    barticle [author] Alon , Noga N. , Matias , Yossi Y. Szegedy , Mario M. ( 1999 ). The space complexity of approximating the frequency moments . Journal of Computer and System Sciences 58 137--147 . barticle

    1999

  3. [3]

    Chernozhukov , Victor V

    barticle [author] Belloni , Alexandre A. Chernozhukov , Victor V. ( 2011 ). l_1 -penalized quantile regression in high-dimensional sparse models . The Annals of Statistics 39 82--130 . barticle

    2011

  4. [4]

    barticle [author] Bickel , Peter J P. J. , Ritov , Ya'acov Y. Tsybakov , Alexandre B A. B. ( 2009 ). Simultaneous analysis of L asso and D antzig selector . The Annals of Statistics 37 1705--1732 . barticle

    2009

  5. [5]

    , Cesa-Bianchi , Nicol \`o N

    barticle [author] Bubeck , S \'e bastien S. , Cesa-Bianchi , Nicol \`o N. Lugosi , G \'a bor G. ( 2013 ). Bandits with heavy tail . IEEE Transactions on Information Theory 59 7711--7717 . barticle

    2013

  6. [6]

    , Kalisch , Markus M

    barticle [author] B \"u hlmann , Peter P. , Kalisch , Markus M. Meier , Lukas L. ( 2014 ). High-dimensional statistics with a view toward applications in biology . Annual Review of Statistics and Its Application 1 255--278 . barticle

    2014

  7. [7]

    Tao , Terence T

    barticle [author] Cand \`e s , Emmanuel E. Tao , Terence T. ( 2007 ). The D antzig selector: statistical estimation when p is much larger than n . The Annals of Statistics 35 2313--2351 . barticle

    2007

  8. [8]

    , Li , Yuanxin Y

    binproceedings [author] Chi , Yuejie Y. , Li , Yuanxin Y. , Zhang , Huishuai H. Liang , Yingbin Y. ( 2019 ). Median-truncated gradient descent: A robust and scalable nonconvex approach for signal estimation . In Compressed Sensing and Its Applications: Third International MATHEON Conference 2017 237--261 . Birkh \"a user , Cham . binproceedings

    2019

Show all 44 references
  1. [9]

    , Lerasle , Matthieu M

    barticle [author] Devroye , Luc L. , Lerasle , Matthieu M. , Lugosi , G \'a bor G. Oliveira , Roberto I. R. I. ( 2016 ). Sub- G aussian mean estimators . The Annals of Statistics 44 2695--2725 . barticle

    2016

  2. [10]

    , Li , Quefeng Q

    barticle [author] Fan , Jianqing J. , Li , Quefeng Q. Wang , Yuyan Y. ( 2017 ). Estimation of high dimensional mean regression in the absence of symmetry and light tail assumptions . Journal of the Royal Statistical Society Series B: Statistical Methodology 79 247--265 . barticle

    2017

  3. [11]

    , Wang , Weichen W

    barticle [author] Fan , Jianqing J. , Wang , Weichen W. Zhu , Ziwei Z. ( 2021 ). A shrinkage principle for heavy-tailed data: High-dimensional robust low-rank matrix recovery . The Annals of Statistics 49 1239--1266 . barticle

    2021

  4. [12]

    , Liu , Han H

    barticle [author] Fan , Jianqing J. , Liu , Han H. , Sun , Qiang Q. Zhang , Tong T. ( 2018 ). I-LAMM for sparse learning: Simultaneous control of algorithmic complexity and statistical error . The Annals of Statistics 46 814--841 . barticle

    2018

  5. [13]

    Nordhausen , Klaus K

    barticle [author] Filzmoser , Peter P. Nordhausen , Klaus K. ( 2021 ). Robust linear regression for high-dimensional data: An overview . Wiley Interdisciplinary Reviews: Computational Statistics 13 e1524 . barticle

    2021

  6. [14]

    , Tibshirani , Robert R

    bbook [author] Hastie , Trevor T. , Tibshirani , Robert R. Wainwright , Martin M. ( 2015 ). Statistical learning with sparsity: The L asso and generalizations . CRC Press , New York . Monographs on Statistics and Applied Probability, 143 . bbook

    2015

  7. [15]

    Sabato , Sivan S

    barticle [author] Hsu , Daniel D. Sabato , Sivan S. ( 2016 ). Loss minimization and parameter estimation with heavy tails . Journal of Machine Learning Research 17 1--40 . barticle

    2016

  8. [16]

    barticle [author] Huber , Peter J P. J. ( 1964 ). Robust estimation of a location parameter . The Annals of Mathematical Statistics 35 73--101 . barticle

    1964

  9. [17]

    Kar , Purushottam P

    barticle [author] Jain , Prateek P. Kar , Purushottam P. ( 2017 ). Non-convex optimization for machine learning . Foundations and Trends in Machine Learning 10 142--363 . barticle

    2017

  10. [18]

    , Tewari , Ambuj A

    binproceedings [author] Jain , Prateek P. , Tewari , Ambuj A. Kar , Purushottam P. ( 2014 ). On iterative hard thresholding methods for high-dimensional M -estimation . In Advances in Neural Information Processing Systems 685--693 . binproceedings

    2014

  11. [19]

    barticle [author] Jerrum , Mark R. M. R. , Valiant , Leslie G. L. G. Vazirani , Vijay V. V. V. ( 1986 ). Random generation of combinatorial structures from a uniform distribution . Theoretical Computer Science 43 169--188 . barticle

    1986

  12. [20]

    Lerasle , Matthieu M

    barticle [author] Lecu \'e , Guillaume G. Lerasle , Matthieu M. ( 2020 ). Robust machine learning by median-of-means: Theory and practice . The Annals of Statistics 48 906--931 . barticle

    2020

  13. [21]

    , Arora , Raman R

    barticle [author] Li , Xingguo X. , Arora , Raman R. , Liu , Han H. , Haupt , Jarvis J. Zhao , Tuo T. ( 2016 ). Nonconvex sparse learning via stochastic optimization with progressive variance reduction . arXiv preprint arXiv:1605.02711 . barticle

    2016 arXiv

  14. [22]

    , Chi , Yuejie Y

    barticle [author] Li , Yuanxin Y. , Chi , Yuejie Y. , Zhang , Huishuai H. Liang , Yingbin Y. ( 2020 ). Non-convex low-rank matrix recovery with arbitrary outliers via median-truncated gradient descent . Information and Inference: A Journal of the IMA 9 289--325 . barticle

    2020

  15. [23]

    ( 2017 )

    barticle [author] Loh , Po-Ling P.-L. ( 2017 ). Statistical consistency of regularized M -estimators . The Annals of Statistics 45 2233--2258 . barticle

    2017

  16. [24]

    ( 2025 )

    barticle [author] Loh , Po-Ling P.-L. ( 2025 ). A theoretical review of modern robust statistics . Annual Review of Statistics and Its Application 12 477--496 . barticle

    2025

  17. [25]

    , Pontil , Massimiliano M

    barticle [author] Lounici , Karim K. , Pontil , Massimiliano M. , van de Geer , Sara S. Tsybakov , Alexandre B. A. B. ( 2011 ). Oracle inequalities and optimal inference under group sparsity . The Annals of Statistics 39 2164--2204 . barticle

    2011

  18. [26]

    Mendelson , Shahar S

    barticle [author] Lugosi , G \'a bor G. Mendelson , Shahar S. ( 2019 ). Mean estimation and regression under heavy-tailed distributions: A survey . Foundations of Computational Mathematics 19 1145--1190 . barticle

    2019

  19. [27]

    ( 2007 )

    bbook [author] Massart , Pascal P. ( 2007 ). Concentration inequalities and model selection: \'E cole d' \'E t \'e de Probabilit \'e s de Saint-Flour XXXIII-2003 . Lecture Notes in Mathematics 1896 . Springer , Berlin, Heidelberg . bbook

    2007

  20. [28]

    ( 2015 )

    barticle [author] Minsker , Stanislav S. ( 2015 ). Geometric median and robust estimation in B anach spaces . Bernoulli 21 2308--2335 . barticle

    2015

  21. [29]

    , Ndaoud , Mohamed M

    barticle [author] Minsker , Stanislav S. , Ndaoud , Mohamed M. Wang , Lang L. ( 2024 ). Robust and tuning-free sparse linear regression via square-root S lope . SIAM Journal on Mathematics of Data Science 6 428--453 . barticle

    2024

  22. [30]

    barticle [author] Negahban , Sahand N S. N. , Ravikumar , Pradeep P. , Wainwright , Martin J M. J. Yu , Bin B. ( 2012 ). A unified framework for high-dimensional analysis of M -estimators with decomposable regularizers . Statistical Science 27 538--557 . barticle

    2012

  23. [31]

    bbook [author] Nemirovsky , Arkadii S. A. S. Yudin , David B. D. B. ( 1983 ). Problem complexity and method efficiency in optimization . Wiley , New York . bbook

    1983

  24. [32]

    , Wainwright , Martin J

    barticle [author] Obozinski , Guillaume G. , Wainwright , Martin J. M. J. Jordan , Michael I. M. I. ( 2011 ). Support union recovery in high-dimensional multivariate regression . The Annals of Statistics 39 1--47 . barticle

    2011

  25. [33]

    , Suggala , Arun Sai A

    barticle [author] Prasad , Adarsh A. , Suggala , Arun Sai A. S. , Balakrishnan , Sivaraman S. Ravikumar , Pradeep P. ( 2020 ). Robust estimation via robust gradient estimation . Journal of the Royal Statistical Society Series B: Statistical Methodology 82 601-627 . barticle

    2020

  26. [34]

    , Zhou , Wen-Xin W.-X

    barticle [author] Sun , Qiang Q. , Zhou , Wen-Xin W.-X. Fan , Jianqing J. ( 2020 ). Adaptive H uber regression . Journal of the American Statistical Association 115 254--265 . barticle

    2020

  27. [35]

    barticle [author] Tan , Kean Ming K. M. , Sun , Qiang Q. Witten , Daniela D. ( 2023 ). Sparse reduced rank H uber regression in high dimensions . Journal of the American Statistical Association 118 2383--2393 . barticle

    2023

  28. [36]

    ( 1996 )

    barticle [author] Tibshirani , Robert R. ( 1996 ). Regression shrinkage and selection via the L asso . Journal of the Royal Statistical Society Series B: Statistical Methodology 58 267--288 . barticle

    1996

  29. [37]

    ( 2018 )

    bbook [author] Vershynin , Roman R. ( 2018 ). High-dimensional probability: An introduction with applications in data science . Cambridge Series in Statistical and Probabilistic Mathematics 47 . Cambridge University Press , Cambridge . bbook

    2018

  30. [38]

    bbook [author] Wainwright , Martin J. M. J. ( 2019 ). High-dimensional statistics: A non-asymptotic viewpoint 48 . Cambridge university press , Cambridge . bbook

    2019

  31. [39]

    , Li , Guodong G

    barticle [author] Wang , Hansheng H. , Li , Guodong G. Jiang , Guohua G. ( 2007 ). Robust regression shrinkage and consistent variable selection through the LAD - L asso . Journal of Business & Economic Statistics 25 347--355 . barticle

    2007

  32. [40]

    Tsay , Ruey S R

    barticle [author] Wang , Di D. Tsay , Ruey S R. S. ( 2023 ). Rate-optimal robust estimation of high-dimensional vector autoregressive models . The Annals of Statistics 51 846--877 . barticle

    2023

  33. [41]

    , Peng , Bo B

    barticle [author] Wang , Lan L. , Peng , Bo B. , Bradic , Jelena J. , Li , Runze R. Wu , Yunan Y. ( 2020 ). A tuning-free robust and efficient approach to high-dimensional regression . Journal of the American Statistical Association 115 1700--1714 . barticle

    2020

  34. [42]

    Lin , Yi Y

    barticle [author] Yuan , Ming M. Lin , Yi Y. ( 2006 ). Model selection and estimation in regression with grouped variables . Journal of the Royal Statistical Society Series B: Statistical Methodology 68 49--67 . barticle

    2006

  35. [43]

    , Wang , Di D

    barticle [author] Zhang , Xiaoyu X. , Wang , Di D. , Li , Guodong G. Sun , Defeng D. ( 2025 ). Robust gradient descent estimation for tensor models under heavy-tailed distributions . arXiv preprint arXiv:2412.04773 . barticle

    2025

  36. [44]

    Zhou , Wenjing W

    binproceedings [author] Zhu , Ziwei Z. Zhou , Wenjing W. ( 2021 ). Taming heavy-tailed features by shrinkage . In International Conference on Artificial Intelligence and Statistics 3268--3276 . PMLR . binproceedings

    2021

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.