Pith. sign in

REVIEW 2 major objections 5 minor 48 references

Robust Sparse Phase Retrieval: Statistical Guarantee, Optimality Theory and Convergent Algorithm

T0 review · 2 major / 5 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 Minimizing a Huber-loss plus ℓ1/2-penalty objective recovers sparse phase-retrieval signals at rate O_P(√(p log n/n)), and the associated MM iteration converges linearly under a spectral gap condition.

desk verdict Solid robust sparse phase retrieval paper with a real constant-factor error in the fixed-point/algorithm theorems that needs correcting before the convergence results can be trusted. read the letter →

arxiv 2505.23273 v1 pith:5IDS3QAN submitted 2025-05-29 math.OC

Jun Fan , Ailing Yan , Xianchao Xiu , Wanquan Liu This is my paper
classification math.OC MSC 90C2690C3062F3594A12
keywords phaseretrievalHuberlossℓ1/2-normregularizationmajorization-minimizationWirtingercalculusfixedpointinclusionconvergenceratesparserecovery
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

The paper tries to show that a single estimator—minimizing the average Huber loss on squared-magnitude measurements plus an ℓ1/2-norm penalty—solves sparse phase retrieval robustly in both real and complex settings. If the consistency theorem is right, one can recover a sparse signal from noisy or outlier-corrupted intensity measurements at error rate O_P(√(p log n/n)), with the dimension p allowed to grow slowly with n. It also claims that every global minimizer must satisfy a fixed-point inclusion, which is the correct optimality condition for complex-valued problems where lifting to the real domain changes the objective. On the algorithmic side, the paper proves that an MM iteration converges as a whole sequence and converges linearly under a spectral gap condition. The practical payoff is a method that handles bounded noise, Laplace noise, and outliers, and runs fast enough for image-scale problems.

What carries the argument

The central object is the composite objective $F(x)=f(x)+\lambda\|x\|_{1/2}^{1/2}$, with $f$ the averaged Huber loss on squared-magnitude residuals and $\|x\|_{1/2}^{1/2}$ the nonconvex $\ell_{1/2}$ quasi-norm penalty. The proof is carried by the half-thresholding operator $H_\mu$ (the closed-form solution of $\min_v \|v-\xi\|^2+\mu\|v\|_{1/2}^{1/2}$), the MM surrogate $F_\tau(x,y)=f(y)+2\operatorname{Re}\langle g(y),x-y\rangle+\frac{1}{2\tau}\|x-y\|^2+\lambda\|x\|_{1/2}^{1/2}$, and the global stability and inlier restricted-eigenvalue inequalities that control the quadratic form. In the complex case the Wirtinger gradient $g$ replaces the ordinary gradient, and the same machinery yields the fixed-point inclusion $x\in H_{\lambda\tau}(x-2\tau g(x))$ by showing $F_\tau$ majorizes $F$ on a bounded level set.

What would settle it

Compute, for a real-valued instance with unit-norm Gaussian measurement vectors, symmetric sub-exponential noise, and a sparse signal satisfying the minimum-magnitude condition, the actual minimizer error for increasing $n$; if $\min\{\|\hat{x}-x^\diamond\|,\|\hat{x}+x^\diamond\|\}$ does not decay at the claimed $O_P(\sqrt{p\log n/n})$ rate, Theorem 3.2 would be refuted. A cheaper check targets the linear-rate claim: evaluate both sides of (5.8) at the limit point of Algorithm 1 on the Type-II experiments; if the inequality fails while the observed error still decays linearly, the spectral gap condition is not necessary.

Watch

Extended reading notes

Core claim

The paper's central claim is that the regularized Huber phase retrieval estimator $\hat{x}$ minimizing $F(x)=\frac{1}{n}\sum_{i=1}^{n} h_\alpha(|\langle a_i,x\rangle|^2-b_i)+\lambda\|x\|_{1/2}^{1/2}$ is consistent for real-valued sparse signals: under $C_1$-stability of the measurement vectors with $C_1>1/2$, an inlier restricted-eigenvalue condition, symmetric sub-exponential noise, and suitable choices of $\alpha$ and $\lambda$, one has $\min\{\|\hat{x}-x^\diamond\|,\|\hat{x}+x^\diamond\|\}=O_P(r_n)$ with $r_n=\sqrt{2}\lambda\|x^\diamond\|_0^{1/2}/(C_2|x^\diamond|_{\min}^{1/2}\|x^\diamond\|_2)$, which is at most $O_P(\sqrt{p\log n/n})$ under the stated conditions. For the complex case, the paper proves that every global minimizer satisfies the fixed-point inclusion $x\in H_{\lambda\tau}(x-2\tau g(x))$ for $\tau\in(0,\min\{1,1/L\}]$, where $H$ is the half-thresholding operator of the $\ell_{1/2}$-norm and $g$ is the Wirtinger gradient of the Huber term. This fixed-point inclusion is then used as the template for Algorithm 1, an MM iteration with Armijo line search whose whole sequence converges, and linearly so when the spectral gap conditions (5.8) or (5.9) hold. A key methodological point is that lifting a complex problem to twice the real dimension changes the objective, so fixed points of the lifted problem are not guaranteed to be optimal for the original complex problem.

Load-bearing premise

The load-bearing premise is that the measurement vectors spread out enough over all directions and enough clean measurements survive the outliers; if either fails, the claimed consistency rate has no foundation.

Editorial extensions

If this is right

  • Real-valued sparse phase retrieval with bounded noise, Laplace noise, or outliers can be solved by a single nonconvex estimator that is consistent at the rate $O_P(\sqrt{p\log n/n})$ when the dimension grows no faster than $o(n/\log n)$.
  • Every global minimizer of the complex objective satisfies an explicit fixed-point inclusion, so an algorithm that lands on a fixed point is certified against the original complex problem rather than a lifted real surrogate.
  • Algorithm 1 converges as a whole sequence, not only along subsequences, and at least sublinearly; under the spectral gap conditions it converges linearly, giving a theoretical rate for the practical iteration.
  • The numerical experiments indicate that the same method achieves higher success rates and lower image relative errors than several comparison methods at sub-second running times in the reported settings.

Reading between the lines

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

  • An extension the paper leaves open is the conjectured high-dimensional bound $O_P(\sqrt{\|x^\diamond\|_0\log(p/\|x^\diamond\|_0)/n})$ for $p>n$; proving that would make the method sample-efficient in the regime where $n$ is much smaller than $p$.
  • The fixed-point inclusion suggests a convergence certificate that the paper does not test: check whether an output $x$ satisfies $x\approx H_{\lambda\tau}(x-2\tau g(x))$ for some admissible $\tau$; such a check could be applied to any black-box phase retrieval solver.
  • The Remark 4 non-equivalence warning applies beyond $\ell_{1/2}$: for any $\varrho\in[0,1]$, the separable penalty on real and imaginary parts is not the same as the penalty on the complex modulus, so complex-domain algorithms should verify optimality in the original variables.
  • The spectral gap conditions (5.8)-(5.9) are sufficient but their empirical frequency is not measured; one could test on the paper's own Type-I/II datasets whether the inequality holds at the limit point, which would tell whether the linear-rate theorem is the explanation for the observed fast convergence.
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 / 5 minor

Summary. The paper proposes and analyzes the estimator (1.2), which combines the Huber loss with ℓ_{1/2}-norm regularization for sparse phase retrieval, covering both real- and complex-valued signals. Its main claims are: a consistency rate O_P(r_n) for the real-valued estimator under stability and restricted-eigenvalue conditions (Theorem 3.2); a fixed-point-inclusion necessary condition for global minimizers derived through an MM surrogate and Wirtinger calculus (Theorem 4.4); whole-sequence convergence of an MM algorithm with an Armijo-type line search (Theorem 5.5); and a linear convergence rate under spectral gap conditions (Theorem 5.6). Numerical experiments compare the method against several robust phase retrieval algorithms on simulated signals and images.

Significance. If correct as stated, the paper would be a useful contribution: it brings a robust loss together with ℓ_{1/2} regularization, addresses the complex-valued case directly rather than through a real-dimensional lifting that is known to be non-equivalent (Remark 4), and gives explicit convergence-rate statements for the MM iteration. The proofs are detailed and self-contained, and the conditions on stability, restricted eigenvalues, and the spectral gap are stated explicitly. However, the optimality and convergence results are currently compromised by an internal parameter mismatch in the MM surrogate, and the consistency theorem contains an undefined symbol. These are substantial but appear to be fixable within the manuscript's scope.

major comments (2)
  1. [Section 4, Theorem 4.4, Eq. (4.3); also Algorithm 1 and Theorems 5.3/5.5] The fixed-point inclusion uses the wrong proximal parameter. For Fτ(x,y)=f(y)+2Re(⟨g(y),x−y⟩)+(1/(2τ))∥x−y∥²+λ∥x∥_{1/2}^{1/2}, completing the square gives Fτ(x,y)=f(y)+(1/(2τ))∥x−y+2τg(y)∥²+λ∥x∥_{1/2}^{1/2}−2τ∥g(y)∥², so minimizing Fτ over x is equivalent to minimizing ∥x−(y−2τg(y))∥²+2λτ∥x∥_{1/2}^{1/2}. By Lemma 4.1, the minimizer is therefore H_{2λτ}(y−2τg(y)), not H_{λτ}(y−2τg(y)). The proof of Theorem 4.4 itself writes the penalty 2λτ before invoking Lemma 4.1, and then concludes (4.3) with H_{λτ}. Consequently Eq. (4.3), Step 1 of Algorithm 1, and the conclusions of Theorems 5.3(iii) and 5.5 that accumulation points and the whole sequence satisfy (4.3) are unsupported as written. This error is load-bearing for the optimality characterization and for the convergence analysis; it should be fixed globally, either by replacing H_{λτ} with H_{2λτ} and adjusting the line-search constants, or by redefining the auxiliary function so that the half-threshold parameter is genuinely λτ.
  2. [Section 3, Theorem 3.2, Eq. (3.6) and proof Step 3] The symbol ∥x⋄∥_{1/2}^{0} is never defined. Section 2.1 defines ∥x⋄∥0 as the sparsity count and ∥x⋄∥_{1/2}^{1/2} as the ℓ_{1/2} quasi-norm, but condition (3.6) and the definition of r_n use ∥x⋄∥_{1/2}^{0}, and the same symbol appears in Step 3 of the proof. If the intended quantity is ∥x⋄∥_0^{1/2}, the manuscript should state this and replace all occurrences, because the existence of a λ satisfying the two inequalities in (3.6) and the stated rate O_P(r_n) depend on the correct definition.
minor comments (5)
  1. [Section 4, proof of Theorem 4.4, real-valued case] The displayed mean value expansion f(x)=f(y)+⟨∇f(y),x−y⟩+⟨∇f(y+θ(x−y)),x−y⟩ is not the standard mean value theorem; the correct form is f(y)+⟨∇f(ξ),x−y⟩ for some ξ on the segment. The subsequent cancellation suggests this is a typo, but it should be corrected.
  2. [Section 5.1, Lemma 5.1] The index j is defined using log_α, whereas the line search uses γβ^j; this appears to be a typo for log_β. The range notation j∈[j,0] should also be clarified.
  3. [Section 5.3, proof of Theorem 5.6] The letter ε1 is used both for the constant appearing in conditions (5.8)/(5.9) and for the support-separation radius 1/2 min{1,min_j |x*_j|}, while ε2 is used both as a radius and as a tolerance in displays around (5.15)-(5.17). Renaming these quantities would make the linear-rate proof substantially easier to verify.
  4. [Appendix A, proof of Lemma 3.1] The variable t1 is introduced without being set; the implication leading to P((1/n)Σ(|εi|−E|εi|)>t_n)≤exp{−nt_n²/(2σ²)} only goes through if t1=t_n. Please state this explicitly.
  5. [Section 6, Tables 3-6] The tables would be easier to interpret if the noise-level parameter η were defined consistently across all noise types and if the text noted which entries are not statistically distinguishable; several entries for Type-III noise already show that the proposed method is not uniformly best, which is acknowledged in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the statistical bound is stated in terms of the true signal and explicit conditions, the MM fixed-point inclusion is derived rather than assumed, and the self-citations are backed by full in-appendix proofs.

full rationale

This paper's derivation chain is self-contained in the sense that matters for circularity. The consistency theorem (Theorem 3.2) bounds min{∥x̂−x⋄∥, ∥x̂+x⋄∥} by O_P(r_n) with r_n built from the true signal norm ∥x⋄∥, sparsity ∥x⋄∥_0, the stability constant C1, the inlier restricted-eigenvalue constant C2, and the regularization parameter λ restricted by condition (3.6); no parameter is fit to the data that the theorem claims to predict. The probabilistic engine behind the guarantee, Lemma 3.1, is proved in full in Appendix A using Bernstein inequalities and a covering net; the citation to the authors' prior work [48] says only 'similar to', and the argument is reproduced, so it is not load-bearing. Likewise, Lemma 5.1 on the Armijo index is asserted 'Analog to the proof of Lemma B.3 in [42]', but Appendix B supplies a complete proof, so the self-citation ([42] is by co-author J. Fan) carries no weight. The half-thresholding operator in Lemma 4.1 is imported from external references [23, 41], and the Wirtinger Taylor lemma from [31]. The fixed-point inclusion (4.3) is derived from the MM surrogate F_τ(x,y) rather than assumed, and the linear-rate hypotheses (5.8)/(5.9) are explicit conditions whose plausibility is argued in Remark 5, not outputs of a fit. Using the authors' own L1/2-LAD [20] as an experimental baseline is normal comparison practice and does not support any theoretical claim. One non-circularity caveat is required for completeness: the skeptic's H_{2λτ} versus H_{λτ} observation is correct — completing the square in F_τ and applying Lemma 4.1 yields the exact MM minimizer H_{2λτ}(y−2τg(y)), whereas (4.3) and Algorithm 1 use H_{λτ}(·). That is an internal mathematical inconsistency affecting the optimality and convergence proofs as written; it is a correctness issue, not an equivalence of inputs and outputs, so it does not raise the circularity score under the hard rules.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central results depend on standard assumptions about the sensing vectors and noise, plus a spectral gap condition that is not empirically checked. There are no invented physical entities. The free parameters λ, α, and ε1 are hyperparameters or constants in the theoretical conditions, not fit during the derivation.

free parameters (3)
  • λ (regularization parameter) = cross-validated in experiments
    Balances Huber loss and sparsity; in theory subject to (3.6) which depends on true sparsity and minimum signal magnitude.
  • α (Huber parameter) = 1.345 for noise, 0.1345 for outliers
    Sets the quadratic-to-linear transition in the Huber loss; in theory lower bounded by (3.2), which involves E|ε1|, unknown in practice.
  • ε1 (bandwidth in Theorem 5.6) = small positive, unspecified
    Introduced to define the index sets around the Huber threshold in (5.8)/(5.9); no practical selection rule given.
assumptions (6)
  • domain assumption Measurement vectors a_i are unit norm (Section 3)
    Used in stability and in the proof of Lemma 3.1; the general case would need re-scaling.
  • domain assumption Stability condition (3.1) holds with C1 > 1/2
    Core seed of the consistency lower bound; not verified for arbitrary sensing designs.
  • domain assumption Restricted eigenvalue condition (3.3) on the inlier set I_in^0
    Needed in Steps 2 and 3 of Theorem 3.2 to localize the estimator; presumed for low corruption rates.
  • domain assumption Noise ε_i i.i.d. sub-exponential, symmetric about 0
    Used in Lemma 3.1 to get E h'_α(ε)=0 and to control deviations.
  • ad hoc to paper Spectral gap condition (5.8)/(5.9)
    Assumed for the linear convergence rate; only argued plausibly in Remark 5, not verified.
  • standard math Semialgebraicity of the objective (Lemma 5.4)
    Used to invoke the KL property; holds for the constructed functions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Sparse Phase Retrieval: Statistical Guarantee, Optimality Theory and Convergent Algorithm." pith.science (2026). https://pith.science/paper/5IDS3QAN

@misc{pith2026250523273,
  author       = {Pith},
  title        = {Pith review of: Robust Sparse Phase Retrieval: Statistical Guarantee, Optimality Theory and Convergent Algorithm},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5IDS3QAN}},
  note         = {Machine review of arXiv:2505.23273}
}
abstract

Phase retrieval (PR) is a popular research topic in signal processing and machine learning. However, its performance degrades significantly when the measurements are corrupted by noise or outliers. To address this limitation, we propose a novel robust sparse PR method that covers both real- and complex-valued cases. The core is to leverage the Huber function to measure the loss and adopt the $\ell_{1/2}$-norm regularization to realize feature selection, thereby improving the robustness of PR. In theory, we establish statistical guarantees for such robustness and derive necessary optimality conditions for global minimizers. Particularly, for the complex-valued case, we provide a fixed point inclusion property inspired by Wirtinger derivatives. Furthermore, we develop an efficient optimization algorithm by integrating the gradient descent method into a majorization-minimization (MM) framework. It is rigorously proved that the whole generated sequence is convergent and also has a linear convergence rate under mild conditions, which has not been investigated before. Numerical examples under different types of noise validate the robustness and effectiveness of our proposed method.

Figures

Figures reproduced from arXiv: 2505.23273 by the authors.

Figure 1
Figure 1. Success rates versus n/p for signal examples: (a)-(c) are the results for real-valued cases, (d)-(f) are the results for complex-valued cases [PITH_FULL_IMAGE:figures/full_fig_p027_1.png] view at source ↗
Figure 2
Figure 2. Relative errors versus iterations with n = 6p for signal examples: (a)-(c) are the results for real-valued cases, (d)-(f) are the results for complex-valued cases. • Type-III: Outliers [18], each measurement bi is corrupted by U(0, ∥xtrue∥ 2 ) with Bernoulli(η) and η = 0.1. 6.2 Signal Examples In this study, we fix p = 128 and s = 12. Assume that ai ∈ Hp are generated from i.i.d. standard normal distributions, the s… view at source ↗
Figure 3
Figure 3. Tested images: (a) HI with the pixel size 72 [PITH_FULL_IMAGE:figures/full_fig_p029_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 47 canonical work pages

  1. [1]

    Grohs, S

    P. Grohs, S. Koppensteiner, M. Rathmair, Phase retrieval: uniqueness and stability, SIAM Review 62 (2) (2020) 301–350

    work page 2020

  2. [2]

    Shechtman, Y

    Y. Shechtman, Y. C. Eldar, O. Cohen, H. N. Chapman, J. Miao, M. Segev, Phase retrieval with application to optical imaging: A contemporary overview, IEEE Signal Processing Magazine 32 (3) (2015) 87–109. 34

    work page 2015

  3. [3]

    Diederichs, F

    B. Diederichs, F. Filbir, P. R¨ omer, Wirtinger gradient descent methods for low-dose poisson phase retrieval, Inverse Problems 40 (12) (2024) 125030

    work page 2024

  4. [4]

    H. Tu, H. Liu, T. Pan, W. Xie, Z. Ma, F. Zhang, P. Xu, L. Wu, O. Xu, Y. Xu, et al., Deep empirical neural network for optical phase retrieval over a scattering medium, Nature Communications 16 (1) (2025) 1369

    work page 2025

  5. [5]

    D. L. Donoho, Compressed sensing, IEEE Transactions on Information Theory 52 (4) (2006) 1289–1306

    2006

  6. [6]

    Shechtman, A

    Y. Shechtman, A. Beck, Y. C. Eldar, GESPAR: Efficient phase retrieval of sparse signals, IEEE Transactions on Signal Processing 62 (4) (2014) 928–938

    work page 2014

  7. [7]

    Z. Yang, L. F. Yang, E. X. Fang, T. Zhao, Z. Wang, M. Neykov, Misspecified nonconvex statistical optimization for sparse phase retrieval, Mathematical Programming 176 (1) (2019) 545–571

    work page 2019

  8. [8]

    J.-F. Cai, J. Li, J. You, Provable sample-efficient sparse phase retrieval initialized by truncated power method, Inverse Problems 39 (7) (2023) 075008

    work page 2023

Show all 48 references
  1. [9]

    Jaganathan, S

    K. Jaganathan, S. Oymak, B. Hassibi, Sparse phase retrieval: Convex algorithms and limitations, in: 2013 IEEE International Symposium on Information Theory, IEEE, 2013, pp. 1022–1026

    2013

  2. [10]

    E. J. R. Pauwels, A. Beck, Y. C. Eldar, S. Sabach, On fienup methods for sparse phase retrieval, IEEE Transactions on Signal Processing 66 (4) (2017) 982–991

    2017

  3. [11]

    Bolte, S

    J. Bolte, S. Sabach, M. Teboulle, Y. Vaisbourd, First order methods beyond convexity and lip- schitz gradient continuity with applications to quadratic inverse problems, SIAM Journal on Optimization 28 (3) (2018) 2131–2151

    2018

  4. [12]

    Chang, S

    H. Chang, S. Marchesini, Y. Lou, T. Zeng, Variational phase retrieval with globally convergent preconditioned proximal algorithm, SIAM Journal on Imaging Sciences 11 (1) (2018) 56–93

    2018

  5. [13]

    Kanzow, P

    C. Kanzow, P. Mehlitz, Convergence properties of monotone and nonmonotone proximal gradient methods revisited, Journal of Optimization Theory and Applications 195 (2) (2022) 624–646

    2022

  6. [14]

    X. Jia, C. Kanzow, P. Mehlitz, Convergence analysis of the proximal gradient method in the presence of the Kurdyka–Lojasiewicz property without global lipschitz assumptions, SIAM Journal on Optimization 33 (4) (2023) 3038–3056

    2023

  7. [15]

    Takahashi, A

    S. Takahashi, A. Takeda, Approximate bregman proximal gradient algorithm for relatively smooth nonconvex optimization, Computational Optimization and Applications 90 (1) (2025) 227–256

    2025

  8. [16]

    Y. Duan, C. Wu, Z.-F. Pang, H. Chang,ℓ 0-regularized variational methods for sparse phase retrieval, arXiv preprint arXiv:1612.02538 (2016)

    2016 arXiv

  9. [17]

    Zhang, Y

    H. Zhang, Y. Chi, Y. Liang, Median-truncated nonconvex approach for phase retrieval with outliers, IEEE Transactions on Information Theory 64 (11) (2018) 7287–7310

    2018

  10. [18]

    Zhang, D

    Q. Zhang, D. Liu, F. Hu, A. Li, H. Cheng, Median momentum reweighted amplitude flow for phase retrieval with arbitrary corruption, Journal of Modern Optics 68 (7) (2021) 374–381

    2021

  11. [19]

    Jiang, H

    X. Jiang, H. C. So, X. Liu, Robust phase retrieval with outliers, in: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2020, pp. 5320–5324

    2020

  12. [20]

    L. Kong, A. Yan, Y. Li, J. Fan,ℓ 1/2-regularized least absolute deviation method for sparse phase retrieval, Pacific Journal of Optimization 18 (1) (2022) 213–232

    2022

  13. [21]

    E. J. Candes, X. Li, M. Soltanolkotabi, Phase retrieval via wirtinger flow: Theory and algorithms, IEEE Transactions on Information Theory 61 (4) (2015) 1985–2007. 35

    2015

  14. [22]

    D. S. Weller, A. Pnueli, G. Divon, O. Radzyner, Y. C. Eldar, J. A. Fessler, Undersampled phase retrieval with outliers, IEEE Transactions on Computational Imaging 1 (4) (2015) 247–258

    2015

  15. [23]

    Z. Xu, X. Chang, F. Xu, H. Zhang,l 1/2-regularization: A thresholding representation theory and a fast solver, IEEE Transactions on Neural Networks and Learning Systems 23 (7) (2012) 1013–1027

    2012

  16. [24]

    Z. Li, F. Nie, J. Bian, D. Wu, X. Li, Sparse PCA viaℓ 2,p-norm regularization for unsupervised feature selection, IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (4) (2023) 5322–5328

    2023

  17. [25]

    Z. Li, P. Chai, Z. Yang, J. Tian, X. Zhu, H. Ouyang, Enhancedℓ 1/2 regularization via dual iterative half thresholding for azimuthal mode identification, Measurement 241 (2025) 115658

    2025

  18. [26]

    M. Chao, Y. Lu, J. Jian, X. Xu, An efficient regularized pr splitting type algorithm for two-block nonconvex linear constrained programs inℓ 1/2 regularized compressed sensing problems, Journal of Computational and Applied Mathematics 453 (2025) 116145

    2025

  19. [27]

    J. Fan, Q. Li, Y. Wang, Estimation of high dimensional mean regression in the absence of sym- metry and light tail assumptions, Journal of the Royal Statistical Society: Series B (Statistical Methodology) 79 (1) (2017) 247–265

    2017

  20. [28]

    J. Xie, S. Liu, J. Chen, J. Jia, Huber loss based distributed robust learning algorithm for random vector functional-link network, Artificial Intelligence Review 56 (8) (2023) 8197–8218

    2023

  21. [29]

    Tyralis, G

    H. Tyralis, G. Papacharalampous, N. Dogulu, K. P. Chun, Deep huber quantile regression net- works, Neural Networks 187 (2025) 107364

    2025

  22. [30]

    P. J. Huber, E. M. Ronchetti, Robust statistics, John Wiley & Sons, 2011

    2011

  23. [31]

    J. Sun, Q. Qu, J. Wright, A geometric analysis of phase retrieval, Foundations of Computational Mathematics 18 (2018) 1131–1198

    2018

  24. [32]

    B. S. Mordukhovich, Variational analysis and applications, Vol. 30, Springer, 2018

    2018

  25. [33]

    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-Lojasiewicz inequality, Mathematics of Operations Research 35 (2) (2010) 438–457

    2010

  26. [34]

    Y. C. Eldar, S. Mendelson, Phase retrieval: Stability and recovery guarantees, Applied and Com- putational Harmonic Analysis 36 (3) (2014) 473–494

    2014

  27. [35]

    J. C. Duchi, F. Ruan, Solving (most) of a set of quadratic equalities: Composite optimization for robust phase retrieval, Information and Inference: A Journal of the IMA 8 (3) (2019) 471–529

    2019

  28. [36]

    E. J. Candes, T. Strohmer, V. Voroninski, Phaselift: Exact and stable signal recovery from magni- tude measurements via convex programming, Communications on Pure and Applied Mathematics 66 (8) (2013) 1241–1274

    2013

  29. [37]

    L. Lei, P. J. Bickel, N. El Karoui, Asymptotics for high dimensional regression m-estimates: fixed design results, Probability Theory and Related Fields 172 (2018) 983–1079

    2018

  30. [38]

    Y. Chi, Y. M. Lu, Y. Chen, Nonconvex optimization meets low-rank matrix factorization: An overview, IEEE Transactions on Signal Processing 67 (20) (2019) 5239–5269

    2019

  31. [39]

    Huang, J

    J. Huang, J. L. Horowitz, S. Ma, Asymptotic properties of bridge estimators in sparse high- dimensional regression models (2008)

    2008

  32. [40]

    J. Fan, L. Kong, L. Wang, N. Xiu, Variable selection in sparse regression with quadratic measure- ments, Statistica Sinica 28 (3) (2018) 1157–1178. 36

    2018

  33. [41]

    J. Zeng, J. Fang, Z. Xu, Sparse SAR imaging based onℓ 1/2 regularization, Science China Infor- mation Sciences 55 (2012) 1755–1775

    2012

  34. [42]

    J. Fan, L. Kong, L. Wang, N. Xiu, Variable selection in sparse regression with quadratic measure- ments, Statistica Sinica 28 (3) (2018) 1157–1178

    2018

  35. [43]

    Y. Wu, S. Pan, X. Yang, A regularized Newton method for norm composite optimization problems, SIAM Journal on Optimization 33 (3) (2023) 1676–1706

    2023

  36. [44]

    Attouch, J

    H. Attouch, J. Bolte, On the convergence of the proximal algorithm for nonsmooth functions involving analytic features, Mathematical Programming 116 (2009) 5–16

    2009

  37. [45]

    Hiriart-Urruty, Mean value theorems in nonsmooth analysis, Numerical Functional Analysis and Optimization 2 (1) (1980) 1–30

    J. Hiriart-Urruty, Mean value theorems in nonsmooth analysis, Numerical Functional Analysis and Optimization 2 (1) (1980) 1–30

    1980

  38. [46]

    J.-F. Cai, Y. Jiao, X. Lu, J. You, Sample-efficient sparse phase retrieval via stochastic alternating minimization, IEEE Transactions on Signal Processing 70 (2022) 4951–4966

    2022

  39. [47]

    M. J. Wainwright, High-dimensional statistics: A non-asymptotic viewpoint, Vol. 48, Cambridge university press, 2019

    2019

  40. [48]

    J. Fan, J. Sun, A. Yan, S. Zhou, An oracle gradient regularized Newton method for quadratic measurements regression, Applied and Computational Harmonic Analysis (2025) 101775. 37

    2025

Pith tools

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