Pith. sign in

REVIEW 4 major objections 4 minor 58 references

Group zero-norm regularized robust loss minimization: proximal MM method and statistical error bound

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A proximal majorization-minimization algorithm for group zero-norm regularized robust loss minimization is proved to converge, and, under restricted eigenvalue conditions, its limit point is shown to lie within $O(\sqrt{r})$ of the ground…

desk verdict A genuine group-wise extension of the zero-norm PMM framework, but the headline statistical bounds contain a factor-2 gap between statement and proof and depend on an unproven condition on the algorithm's limit point. read the letter →

arxiv 2505.23294 v1 pith:UNSEYU43 submitted 2025-05-29 math.OC

Ling Liang , Shujun Bi This is my paper
classification math.OC MSC 90C2690C3062J0765K05
keywords groupzero-normsparsityrobustlossproximalmajorization-minimizationdifference-of-convexsurrogateKurdyka-ŁojasiewiczpropertyrestrictedeigenvalueconditionsemismoothNewtonmethod
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 proposes a proximal majorization-minimization (PMM) algorithm for minimizing a robust loss plus a group zero-norm penalty, an objective that is nonconvex, discontinuous, and generally NP-hard. It proves that, under the Kurdyka-Łojasiewicz property of the potential function, the generated sequence converges globally to a critical point and, when the KL exponent is $1/2$, locally at an R-linear rate. For the linear observation model with a design matrix satisfying a restricted eigenvalue condition, it derives statistical error bounds of order $O(\sqrt{r})$ between the algorithm's iterates, and their cluster point, and the true coefficient vector with $r$ nonzero groups. These bounds extend earlier element-wise zero-norm results to group sparsity with a general linear mapping $g(x)=Bx$. Subproblems are solved by a proximal dual semismooth Newton method, and numerical experiments on synthetic and benchmark data report faster computation than a proximal alternating direction method of multipliers.

What carries the argument

The central object is the family $\mathcal{L}$ of convex functions $\varphi$ satisfying $\varphi(t^*)=0$, $\varphi(1)=1$, and the induced $\psi$ on $[0,1]$, from which the surrogate penalty $\rho^{-1}\psi^*(\rho\|g_{J_i}(x)\|)$ is built. The identity doing the work is the convexity inequality (13) that majorizes the surrogate by a weighted $\ell_{2,1}$-type term with weights $w_{i,\rho}(x)=(\psi^*)'(\rho\|g_{J_i}(x)\|)$; the proximal MM iteration then solves strongly convex subproblems of the form (16). In the statistical analysis, the load-bearing mechanism is the restricted-eigenvalue cone $\mathcal{C}(S,1.5r)$ and the off-support weight condition $\max_{i\notin S} w_i \le 1/2$, which together force the off-support part of the error into the on-support part (Lemmas 7 and 8).

What would settle it

Build a synthetic instance with sample size above the theorem's threshold, a design matrix satisfying the restricted eigenvalue condition on $\mathcal{C}(S,1.5r)$, and a known ground-truth vector with $r$ nonzero groups; run the PMM algorithm with $\lambda$ in the stated interval. If the cluster point satisfies the off-support condition but its distance to the ground truth exceeds the theorem's displayed bound, the theorem is false; if the cluster point violates the off-support condition, the instance lies outside the stated guarantee and the bound is not required to hold.

Watch

Extended reading notes

Core claim

The paper establishes that a tractable DC surrogate of the group zero-norm regularized problem can be minimized by a proximal MM scheme whose iterates converge to a critical point, and that, under a restricted eigenvalue condition on the design matrix, both the iterates and their limit point satisfy an explicit statistical error bound against the true solution. Theorem 1 proves global convergence under the Kurdyka-Łojasiewicz property and an R-linear local rate when the KL exponent is $1/2$. Theorems 3 and 5, for squared strongly convex and piecewise linear robust losses respectively, bound the distance from the cluster point to the truth by expressions of the form $O(\sqrt{r})$ whenever the sample size exceeds an explicit threshold, the regularization parameter $\lambda$ lies in an interval, and the cluster point has all off-support group norms below $2a/(\rho(a+1))$. The proof relies on the convexity of $\psi^*$ to majorize the surrogate; Lemmas 7 and 8 then show the off-support part of the error is controlled by the on-support part, and the restricted eigenvalue condition converts that control into the announced bound.

Load-bearing premise

The statistical error bound holds only when the algorithm's own limit point satisfies $\|\bar{x}_{J_i}\| \leq 2a/(\rho(a+1))$ for every group outside the true support, a condition on the output of the algorithm rather than on the design matrix or the noise.

Editorial extensions

If this is right

  • Group-sparse high-dimensional regression with robust losses now has an implementable proximal MM method whose limit point carries a provable $O(\sqrt{r})$ recovery guarantee under restricted eigenvalue conditions.
  • The explicit interval for $\lambda$ in Theorems 3 and 5 offers a principled range for regularization tuning, and the range widens as the sample size grows.
  • The local R-linear convergence rate for KL exponent $1/2$ gives a predictable worst-case number of iterations near the solution.
  • Because the framework allows a general linear map $g(x)=Bx$, the same convergence and error analysis covers structured sparsity patterns beyond plain group selection.
  • The proximal dual semismooth Newton subproblem solver turns the theoretical iteration into a practical algorithm, and the reported experiments indicate that it is substantially faster than pADMM on the tested problems.

Reading between the lines

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

  • One could replace the cluster-point condition with a verifiable stopping criterion, making the guarantee checkable during the run rather than after the fact.
  • A testable direction is whether the same cone argument extends to overlapping groups, since the matrix $B$ already encodes inter-group dependencies but the proof of Lemmas 7 and 8 uses disjoint groups.
  • The constant in the $O(\sqrt{r})$ bound depends on the surrogate parameters $a$ and $\rho$; a natural extension is a data-dependent rule for choosing them inside the family $\mathcal{L}$.
  • On runs where the limit point's off-support group norms exceed $2a/(\rho(a+1))$, the theorem's guarantee is not active; tracking that quantity would give a practical diagnostic for when the promised bound applies.
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 manuscript proposes an inexact proximal Majorization-Minimization (PMM) algorithm for group zero-norm regularized robust loss minimization, formulated through an equivalent DC surrogate problem. It proves global convergence to a critical point under a Kurdyka-Łojasiewicz assumption, with a local R-linear rate when the KL exponent is 1/2. The main claimed contribution is a set of statistical error bounds (Theorems 2–5) showing that PMM iterates and their limit points are close to the ground-truth solution under restricted eigenvalue conditions, together with a proximal dual semismooth Newton solver for the subproblems and numerical comparisons against pADMM on synthetic and UCI data. The paper extends the element-wise results of [57] to group sparsity.

Significance. If the statistical bounds were established, the paper would make a useful contribution: it would provide a convergent algorithm with quantitative error control for an NP-hard group-sparse robust regression problem, and the numerical experiments indicate practical efficiency. The convergence analysis in Section III follows standard KL machinery and appears credible; the DC-surrogate equivalence is imported from published prior work rather than derived here. However, the central statistical theorems are not established as written. They depend on unverified conditions on the algorithm's active sets and on its limit point, and the proof of Theorem 3 contains a factor-of-two inconsistency between the stated assumption and the condition actually used. The paper's headline error-bound claims are therefore currently conditional and need substantial repair.

major comments (4)
  1. [Section IV, Theorem 3 (proof after 'According to the assumption on x̄')] Theorem 3 states the cluster-point condition ∥x̄_Ji∥ ≤ 2a/(ρ(a+1)) for i∉S, but its proof asserts 'namely ∥x̄_Ji∥ ≤ a/(ρ(a+1))'. For Example 2, formula (19) gives w_i ≤ 1/2 if and only if ρ∥x_Ji∥ ≤ 1; the stated threshold only yields ρ∥x_Ji∥ ≤ 2a/(a+1), which is strictly larger than 1 for every a>1. For instance, with a=4 and ρ=2, the stated assumption allows ∥x̄_Ji∥=0.8, for which w_i=1. Consequently the proof cannot establish S_{k-1}≡S and max_{i∉S} w_i^{k-1}≤1/2, and the error bound of Theorem 3 is not proved as stated.
  2. [Section IV, Theorem 5 (proof, same argument)] Theorem 5 repeats the same factor-of-two gap: the stated condition ∥x̄_Ji∥ ≤ 2a/(ρ(a+1)) does not imply the weight condition needed in Lemma 8, because the proof requires ∥x̄_Ji∥ ≤ a/(ρ(a+1)) to conclude ρ∥x^k_Ji∥≤1 and hence w_i^{k-1}≤1/2. Since both Theorem 3 and Theorem 5 are the paper's headline limit-point error bounds, the central statistical guarantee of the abstract is not established as written.
  3. [Section IV, Theorems 3 and 5 (assumption on the limit point)] Even after correcting the threshold, the results are conditional on an algorithmic event that Section III does not guarantee. Theorem 1 only shows convergence to a critical point; nothing in Section III shows that the limit point satisfies an off-support group-norm bound or that S_{k-1}⊇S for all sufficiently large k. The paragraph preceding Theorem 3 explicitly assumes this smallness, and Remark 7-like discussion does not prove it. Thus the abstract's claim that the paper derives statistical error bounds for the PMM-generated iterates and limit points overstates what is proved. The authors should either prove such a support-recovery or off-support bound under Assumption 1, or explicitly present Theorems 3 and 5 as conditional results and qualify the abstract and conclusions accordingly.
  4. [Section IV, Theorem 3 proof (passage involving G(ξ^k) and G(μx*))] The proof invokes Theorem 2 with a λ satisfying c(∥G(μx*)∥∞ + 2L_ϑ∥A∥), but Theorem 2's lower bound involves ∥G(ξ^k)∥∞. The proof only establishes ∥G(ξ^k)∥∞ ≤ (3/2)∥G(μx*)∥∞ for large k, so the stated lower bound can be too small. The λ interval and the final error estimate in Theorem 3 (and similarly Theorem 5) must include this factor or otherwise justify the passage. This is a second constant gap in the central statistical argument.
minor comments (4)
  1. [Section IV, Theorem 2] The displayed λ interval contains 'pκρ', which should almost certainly be 'nκρ'; the interval is also typeset with ambiguous parentheses and should be rewritten with clear delimiters.
  2. [Section II/IV notation] The norm ∥A∥_{2,∞} is used in Theorems 2 and 4 but is not defined in the notation section; please define it or replace it with the spectral norm ∥A∥ where that is what is meant.
  3. [Section IV, Lemma 8] In Lemma 8, the set J_k is defined using ε_i while the noise is elsewhere denoted ϖ_i; also the expressions '2n^{-1} eρ| ∥AJ· | ∥1' contain unexplained '|' and should be clarified.
  4. [Section IV, Theorem 5 statement] The threshold condition contains 'eρ| ∥AJ· | ∥1' with a stray vertical bar; please correct the typo and ensure all occurrences of this term are consistently typeset.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the derivation chain is conditional and self-contained; self-citations are to external published results.

full rationale

The paper's convergence analysis (Section III) is self-contained: Lemma 4 establishes a sufficient decrease using the proximal MM construction, Lemma 5 shows cluster points are critical via subdifferential calculus, and Theorem 1 follows by standard KL arguments. The statistical bounds (Section IV) are stated under explicit conditions—Assumption 1 (restricted eigenvalue), S_{k-1} contains S, max off-support weights at most 1/2, and a smallness condition on the cluster point in Theorems 3 and 5. These are hypotheses, not fitted parameters; the bounds are obtained by algebraic inequalities, not by construction from the quantities being bounded. The equivalence to the DC surrogate and the local-optimality interpretation are cited from [4] and [57], prior published work with overlapping authors; under the stated rules this is external support rather than a self-citation loop, because the cited theorems are parameter-free and do not assume the present results. Two non-circular caveats are noted: (i) the proofs of Theorems 3 and 5 use the stronger off-support threshold a/(rho(a+1)) while the theorem statements give 2a/(rho(a+1)), a factor-of-2 mismatch that appears to leave the stated bound unproven for a>1; (ii) the cluster-point smallness condition is an assumption on the algorithm's own limit point, so the abstract's claim to derive error bounds is conditional on that assumption. These are rigor and limitation concerns, not circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central results rest on (i) the usual KL/definability assumptions for convergence, (ii) a restricted eigenvalue condition on the design matrix, (iii) a squared-strong-convexity/Lipschitz assumption on the robust loss, and (iv) an ad hoc condition on the algorithm's limit point (off-support group norms below a threshold). The method introduces four user-selected hyperparameters (λ, ρ, a, µ) that enter the error bounds but are not fitted to data in the theoretical claims. No new entities are postulated.

free parameters (4)
  • regularization parameter λ = chosen per dataset (e.g., γ1, γ2 in Tables I-IV)
    Weights the group zero-norm penalty; the statistical bounds require λ to lie in an interval that depends on unknown quantities such as ∥G(µx∗)∥∞.
  • surrogate parameter ρ = 2 (synthetic), max(1,6/∥x0∥∞) (real data)
    Controls the DC surrogate via φ_ρ; appears in the threshold 2a/(ρ(a+1)) in the cluster-point assumption.
  • scale parameter a = 4 or 6 in experiments
    Defines the surrogate φ in Examples 2/3 and the constant c; affects the cluster-point condition and the error bound.
  • proximal/coercivity parameter µ = 1e-8
    Added to the quadratic term to ensure coercivity; enters the error bounds and the λ interval.
assumptions (6)
  • domain assumption The potential function Ψ_{ρ,ν,µ} has the KL property, with exponent 1/2 at the limit point for R-linear rate.
    Invoked in Theorem 1 and Section III-B; standard but not verified for the specific problem.
  • domain assumption Θ_{ρ,ν,µ} is definable in an o-minimal structure over a real field.
    Used in Remark 4(a) to conclude Ψ is a KL function; holds for semi-algebraic ϕ, but the surrogate with ϑ might not be semi-algebraic.
  • domain assumption Assumption 1: restricted eigenvalue condition 1/(2n)∥Ax∥² ≥ κ∥x∥² on the cone C(S,1.5r).
    Central for the statistical error bounds; standard in high-dimensional statistics but not guaranteed for given design.
  • ad hoc to paper The cluster point x satisfies ∥x_{J_i}∥ ≤ 2a/(ρ(a+1)) for i∉S.
    This is the load-bearing assumption for Theorems 3 and 5, imposed on the algorithm's own limit point.
  • ad hoc to paper At the relevant iteration k, the active set of the previous iterate contains the true support and off-support weights are ≤1/2.
    Required by Lemma 7/8 and Theorems 2/4; not proven to hold until the cluster-point condition is imposed.
  • domain assumption The functions ϑ and θ satisfy Assumptions 2/3 (squared strong convexity, Lipschitz, subgradient boundedness).
    Used in Lemmas 6-8; restrict the loss functions covered to squared strongly convex and piecewise linear-type losses.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Group zero-norm regularized robust loss minimization: proximal MM method and statistical error bound." pith.science (2026). https://pith.science/paper/UNSEYU43

@misc{pith2026250523294,
  author       = {Pith},
  title        = {Pith review of: Group zero-norm regularized robust loss minimization: proximal MM method and statistical error bound},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UNSEYU43}},
  note         = {Machine review of arXiv:2505.23294}
}
read the original abstract

This study focuses on solving group zero-norm regularized robust loss minimization problems. We propose a proximal Majorization-Minimization (PMM) algorithm to address a class of equivalent Difference-of-Convex (DC) surrogate optimization problems. First, we present the core principles and iterative framework of the PMM method. Under the Kurdyka-{\L}ojasiewicz (KL) property assumption of the potential function, we establish the global convergence of the algorithm and characterize its local (sub)linear convergence rate. Furthermore, for linear observation models with design matrices satisfying restricted eigenvalue conditions, we derive statistical estimation error bounds between the PMM-generated iterates (including their limit points) and the ground truth solution. These bounds not only rigorously quantify the approximation accuracy of the algorithm but also extend previous results on element-wise sparse composite optimization from reference [57]. To efficiently implement the PMM framework, we develop a proximal dual semismooth Newton method for solving critical subproblems. Extensive numerical experiments on both synthetic data and the UCI benchmark demonstrate the superior computational efficiency of our PMM method compared to the proximal Alternating Direction Method of Multipliers (pADMM).

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 58 canonical work pages

  1. [4]

    SHAOHUAPAN, LINGLIANG, YULANLIU, Local optimality for stationary points of group zero-norm regularized problems and equivalent surrogates[J], Optimization: A Journal of Mathematical Pro- gramming and Operations Research 2023, 72(9): 2311-2343

    work page 2023

  2. [57]

    D. D. ZHANG, S. H. PAN, S. J. BI, D. F. SUN, Zero-norm regularized problems: equivalent surrogates, proximal MM method and statistical error bound[J], Computational Optimization and Applications, 2023

    work page 2023

  3. [1]

    P. TANG, C. WANG, D. SUN,ANDK. C. TOH, A sparse semismooth newton based proximal majorization-minimization algorithm for nonconvex square-root-loss regression problems, 2019

    work page 2019

  4. [2]

    ATTOUCH, J

    H. ATTOUCH, J. BOLTE, P. REDONT, A. SOUBEYRAN, Proximal alternating minimization and projection methods for nonconvex problems: an approach based on the Kurdyka-Łojasiewicz inequality[J], Mathematics of Operations Research, 2010, 35: 438-457

    work page 2010

  5. [3]

    ATTOUCH, J

    H. ATTOUCH, J. BOLTE, On the convergence of the proximal algorithm for nonsmooth functions involving analytic features[J], Mathematical Programming: Series A and B, 2009, 116(1-2): 5-16

    work page 2009

  6. [5]

    M. E. AHSEN, M. VIDYASAGAR, Error bounds for compressed sensing algorithms with group sparsity: A unified approachy[J], Applied and Computational Harmonic Analysis, 2017, 43(2): 212-232

    work page 2017

  7. [6]

    P. S. BRADLEY, O. L. MANGASARIAN, Feature selection via concave minimization and support vector machines[J], In Proceeding of interna- tional conference on machine learning ICML, 1998

    work page 1998

  8. [7]

    A. BECK, N. HALLAK, Optimization problems involving group sparsity terms[J], Mathematical Programming, 2019, 178: 39-67

    work page 2019

Show all 58 references
  1. [8]

    CHARTRAND, Exact reconstruction of sparse signals via nonconvex minimization, IEEE Signal Processing Letters, 14(2007): 707-710

    R. CHARTRAND, Exact reconstruction of sparse signals via nonconvex minimization, IEEE Signal Processing Letters, 14(2007): 707-710

    2007

  2. [9]

    X. J. CHEN, F. M. XU ANDY. Y. YE, Lower bound theory of nonzero entries in solutions of ℓ2-ℓp minimization, SIAM Journal on Scientific Computing, 32(2010): 2832-2852

    2010

  3. [10]

    W. DENG, W. T. YIN ANDY. ZHANG, Group sparse optimization by alternating direction method, Proceedings of SPIE the International Society for Optical Engineering, vol. 8858

  4. [11]

    FAN ANDR

    J. FAN ANDR. LI, Variable selection via nonconcave penalized likelihood and its oracle properties, Journal of the American Statistical Association, 9(2001): 1348-1360

    2001

  5. [12]

    X. GUO, H. ZHANG, Y. WANG ANDJ. L. WU, Model selection and estimation in high dimensional regression models with group SCAD, Statistics & Probability Letters, 103(2015): 86-92

    2015

  6. [13]

    Y. HU, C. LI, K. MENG, J. QIN ANDX. YANG, Group sparse optimization via lp,q regularization, Journal of Machine Learning Re- search, 18(2017): 1-52

    2017

  7. [14]

    Q. LING, Z. W. WEN ANDW. T. YIN, Decentralized jointly sparse signal recovery by reweighted lq minimization, IEEE Transactions on Signal Processing, 61(2013): 1165-1170

    2013

  8. [15]

    MEIER, S.VANDEGEER ANDP

    L. MEIER, S.VANDEGEER ANDP. B ¨UHLMANN, The group lasso for logistic regression, Journal of the Royal Statistical Society Series B, 70(2008): 53-71

    2008

  9. [16]

    RINALDI, F

    F. RINALDI, F. SCHOEN ANDM. SCIANDRONE, Concave programming for minimizing the zero-norm over polyhedral sets, Computation Opti- mization and Applications, 46(2010): 467-486

    2010

  10. [17]

    YANG ANDH

    Y. YANG ANDH. ZOU, A fast unified algorithm for solving group-lasso penalize learning problems, Statistical Computation, 15(2015): 1129- 1141

    2015

  11. [18]

    C. H. ZHANG, Nearly unbiased variable selection underminimax concave penalty, The Annals of Statistics, 38(2010): 894-942

    2010

  12. [19]

    BOLTE, S

    J. BOLTE, S. SABACH, M. TEBOULLE, Proximal alternating linearized minimization for nonconvex and nonsmooth problems[J], Mathematical Programming, 2014, 146(1-2): 459-494

    2014

  13. [20]

    C. C. CHANG, C. J. LIN, LIBSVM: A library for support vector machines[J], ACM Transactions on Intelligent Systems and Technology, 2007

    2007

  14. [21]

    M. A. DAVENPORT, M. F. DUARTE, Y. C. ELDAR, G. KUTYNIOK, Introduction to compressed sensing[J], Preprint, 2011: 1-68

    2011

  15. [22]

    M. F. DUARTE, Y. C. ELDAR, Structured compressed sensing: from theory to applications[J], IEEE Transactions on Signal Processing, 2011, 59(9): 4053-4085

    2011

  16. [23]

    DAVID, Sparse group fused lasso for model segmentation: a hybrid approach[J], Advances in Data Analysis and Classification, 2021, 15: 625-671

    D. DAVID, Sparse group fused lasso for model segmentation: a hybrid approach[J], Advances in Data Analysis and Classification, 2021, 15: 625-671

    2021

  17. [24]

    Sparse and Redundant Representations: from Theory to Ap- plications in Signal and Image Processing[J], Springer, Berlin, 2010

    Elad M.. Sparse and Redundant Representations: from Theory to Ap- plications in Signal and Image Processing[J], Springer, Berlin, 2010

    2010

  18. [25]

    Y. C. ELDAR, M. MISHALI, Robust recovery of signals from a structured union of subspaces[J], IEEE Transactions on Information Theory, 2009, 55: 5302-5316

    2009

  19. [26]

    J. Q. FAN, L. XUE, H. ZOU, Strong oracle optimality of folded concave penalized estimation[J], Annals of Statistics, 2012, 42(3): 819-849

    2012

  20. [27]

    P. GONG, J. YE, C. ZHANG, Multi-stage multi-task feature learning, Journal of Machine Learning Research[J], 2013, 14: 2979-3010

    2013

  21. [28]

    HUANG, T

    J. HUANG, T. ZHANG, The benefit of group sparsity[J], The Annals of Statistics, 2010, 38: 1978-2004

    2010

  22. [29]

    HUANG, J

    L. HUANG, J. JIA, B. YU, B. G. CHUN, P. MANIATIS, M. NAIK, Predicting execution time of computer programs using sparse polynomial regression[J], in Advances in Neural Information Processing Systems, 2010: 883-891

    2010

  23. [30]

    J. B. HIRIART-URRUTY, J. J. STRODIOT, V. H. NGUYEN, Generalized Hessian matrix and second-order optimality conditions for problems with C1,1 data[J], Applied Mathematics and Optimization, 1984, 11(1): 169-180

    1984

  24. [31]

    A. D. IOFFE, J. V. OUTRATA, On metric and calmness qualification conditions in subdifferential calculus[J], Set-Valued Analysis, 2008, 16: 199-227. JOURNAL OF LATEX CLASS FILES, VOL. , NO. , 24

    2008

  25. [32]

    JENATTON, J

    R. JENATTON, J. AUDIBERT, F. BACH, Structured variable selection with sparsity-inducing norms[J], Journal of Machine Learning Research, 2011, 12(10): 2777-2824

    2011

  26. [33]

    G. Y. LI, T. K. PONG, Calculus of the exponent of Kurdyka-Łojasiewicz inequality and its applications to linear convergence of first-order methods[J], Foundations of Computational Mathematics, 2018, 18: 1199-1232

    2018

  27. [34]

    LICHMAN, UCI Machine Learning Repository[J], http://archive.ics.uci.edu/ml/ datasets.html

    M. LICHMAN, UCI Machine Learning Repository[J], http://archive.ics.uci.edu/ml/ datasets.html

  28. [35]

    Y. L. LIU, S. J. BI, S. H. PAN, Equivalent Lipschitz surrogates for zero-norm and rank optimization problems[J], Journal of Global Opti- mization, 2018, 72(4): 679-704

    2018

  29. [36]

    X. D. LI, D. F. SUN, K. C. TOH, A highly efficient semismooth Newton augmented Lagrangian method for solving Lasso problems[J], SIAM Journal on Optimization, 2018, 28: 433-458

    2018

  30. [37]

    Y. LI, B. NAN,ANDJ. ZHU, Multivariate sparse group lasso for the multivariate multiple linear regression with an arbitrary group structurey[J], Biometrics, 2015, 71(2): 354-363

    2015

  31. [38]

    P. L. LOH ANDM. J. WAINWRIGHT, Regularized M-estimators with nonconvexity: Statistical and algorithmic theory for local optima[C], Journal of Machine Learning Research, 2015, 16: 559-616

    2015

  32. [39]

    MEIER, S.VANDEGEER, P

    L. MEIER, S.VANDEGEER, P. B ¨UHLMANN, The group lasso for logistic regression[J], Journal of the Royal Statistical Society Series B, 2008, 70: 53-71

    2008

  33. [40]

    OBOZINSKI, B

    G. OBOZINSKI, B. TASKAR, M. I. JORDAN, Joint covariate selection for grouped classification[J], Statistics and Computing, 2010, 20: 231-252

    2010

  34. [41]

    POIGNARD, Asymptotic theory of the adaptive sparse group lasso[J], Annals of the Institute of Statistical Mathematics, 2018: 1-32

    B. POIGNARD, Asymptotic theory of the adaptive sparse group lasso[J], Annals of the Institute of Statistical Mathematics, 2018: 1-32

    2018

  35. [42]

    N. RAO, C. COX, R. NOWAK,ANDT. T. ROGERS, Sparse overlapping sets lasso for multitask learning and its application to fmri analysis[J], in Advances in neural information processing systems, 2013: 2202-2210

    2013

  36. [43]

    N. RAO, R. NOWAK, C. COX,ANDT. ROGERS, Classification with the sparse group lassoy[J], IEEE Transactions on Signal Processing, 2015, 64(2): 448-463

    2015

  37. [44]

    R. T. ROCKAFELLAR, Augmented Lagrangians and applications of the proximal point algorithm in convex programming[J], Mathematics of Operations Research, 1976, 1: 97-116

    1976

  38. [45]

    R. T. ROCKAFELLAR, Convex Analysis[M], Princeton University Press, 1970

    1970

  39. [46]

    R. T. ROCKAFELLAR, R. J. B. WETS, Variational Analysis[M], Springer, 1998

    1998

  40. [47]

    STOJNIC, F

    M. STOJNIC, F. PARVARESH, B. HASSIBI, On the reconstruction of block-sparse signals with an optimal number of measurements[J], IEEE Transactions on Signal Processing, 2009, 57: 3075-3085

    2009

  41. [48]

    STEINHAEUSER, A

    K. STEINHAEUSER, A. BANERJEE, S. CHATTERJEE,ANDA. GAN- GULY, Sparse group lasso: Consistency and climate applications[J], in Proceedings of the 2012 SIAM International Conference on Data Mining, SIAM, 2012: 47-58

    2012

  42. [49]

    J. A. TROPP, S. J. WRIGHT, Computational methods for sparse solution of linear inverse problems[J], Proceedings of the IEEE, 2010, 98(6): 948-958

    2010

  43. [50]

    Y. Q. WU, S. H. PAN, S. J. BI, Kurdyka-Łojasiewicz property of zero-norm composite functions[J], Journal of Optimization Theory and Applications, 2021, 188: 94-112

    2021

  44. [51]

    P. R. YU, G. Y. LI, T. K. PONG, Kurdyka-Łojasiewicz exponent via inf-projection[J], Foundations of Computational Mathematics, 2022, 22: 1171-1217

    2022

  45. [52]

    M. YUAN, Y. LIN, Model selection and estimation in regression with grouped variables[J], Journal of The Royal Statistical Society Series B, 2006, 68: 49-67

    2006

  46. [53]

    X. T. YUAN, X. LIU, S. YAN, Visual classification with multitask joint sparse representation[J], IEEE Transactions on Image Processing, 2012, 21: 4349-4360

    2012

  47. [54]

    ZHANG, Z

    J. ZHANG, Z. GHAHRAMANI, Y. YANG, Flexible latent variable models for multi-task learning[J], Machine Learning, 2008, 73: 221-242

    2008

  48. [55]

    X. Y. ZHAO, D. F. SUN, K. C. TOH, A Newton-CG augmented Lagrangian method for semidefinite programming[J], SIAM Journal on Optimization, 2010, 20: 1737-1765

    2010

  49. [56]

    Y. J. ZHANG, N. ZHANG, D. F. SUN, K. C. TOH, An efficient Hessian based algorithm for solving large-scale sparse group Lasso problems[J], Mathematical Programming, 2020, 179: 223-263

    2020

  50. [58]

    C. H. ZHANG, T. ZHANG, A general theory of concave regularization for highdimensional sparse estimation problems[J], Statistical Science, 2012, 27: 576-593

    2012

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.