Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Phase retrieval via overparametrized nonconvex optimization: nonsmooth amplitude loss landscapes

T0 review · 2 major / 5 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 This paper establishes that the nonsmooth amplitude least-squares loss for phase retrieval has a benign landscape under constant overparametrization: every second-order critical point is a statistically optimal estimate, achieving noise-flo

desk verdict First landscape analysis of the nonsmooth amplitude loss with constant overparametrization, but a quantifier error in the SOCP definition undercuts the main theorems as printed. read the letter →

arxiv 2511.19045 v2 pith:PFIA7MKS submitted 2025-11-24 math.OC math.STstat.TH

Andrew D. McRae This is my paper · ORCID
classification math.OCmath.STstat.TH MSC 90C2649J5290C46
keywords phaseretrievalnonconvexoptimizationoverparametrizationnonsmoothamplitudelosssecond-ordercriticalpointslow-rankmatrixsensingsub-GaussianmeasurementsCut
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 studies a nonconvex, nonsmooth phase-retrieval loss — minimize the average squared difference between measured amplitudes |⟨fi,x⟩| and observations — but over a rank-p variable X with p larger than the true rank, an overparametrization. It proves that, for isotropic sub-Gaussian measurements and n at least a constant multiple of d, every second-order critical point of this amplitude loss lies close to the true signal: the recovery error is bounded by the statistical noise floor, ∥ε∥/√n, with only constant overparametrization. The same guarantee holds for a nonconvex PhaseCut reformulation as a smooth synchronization problem over spheres, and a parallel result covers Gaussian measurements in any Hilbert space with ridge regularization. A sympathetic reader would care because the amplitude loss is the one practitioners prefer, and this shows its global nonconvex landscape is benign with minimal overparametrization, matching guarantees that previously required semidefinite programming or logarithmic rank inflation.

What carries the argument

The engine is Lemma 1, a deterministic second-order-criticality inequality for factored losses Lλ(XX*) = (1/n)∑ ℓ(⟨Ai,XX*⟩,y_i) + λ tr(XX*) with PSD measurement matrices Ai and convex ℓ. It shows that any second-order critical point X must satisfy ∇Lλ(XX*)X = 0 and, for every PSD Z′, 0 ≤ ⟨∇Lλ(XX*), Z′⟩ + (2/(c_F p))∑_{⟨Ai,XX*⟩>0} ℓ″(⟨Ai,XX*⟩,y_i)⟨Ai,XX*⟩⟨Ai,Z′⟩, where c_R=1 and c_C=2. The rank p sits in the denominator of the curvature penalty, which is exactly why overparametrization weakens the nonconvex obstruction and brings the condition close to convex SDP optimality. For the amplitude loss ℓ(b,y)=(√b−y)², the curvature term collapses to a controlled form, yielding Theorem 2; the Phase

What would settle it

Run the two-line check on Lemma 2 using the amplitude loss ℓ(b,y)=(√b−y)² with d=p=2, A1=diag(1,0), A2=diag(0,1), y=(1,1), X=(0,1). This point satisfies the printed limsup definition of a second-order critical point, yet ⟨A1,XX*⟩=0 gives ℓ′(0,1)=−∞, which contradicts the first claim of Lemma 2 and violates the quadratic-ascent inequality used throughout. Reproducing this computation would settle whether the main results hold with the definition as printed.

Watch

Extended reading notes

Core claim

The central claim is Theorem 5: under the measurement model y_i = |⟨f_i,x*⟩| + ε_i with isotropic sub-Gaussian measurement vectors satisfying mild moment and low-coherence conditions, if n ≥ c1 d then with probability at least 1 − c2 n^{−2}, for every p ≥ c3, every second-order critical point X of the nonsmooth amplitude problem (1/2n)∑(√⟨Ai,XX*⟩ − y_i)² satisfies ∥XX* − x*x*∥_* ≤ c4(∥x*∥∥ε∥/√n + ∥ε∥²/n), and the nearest rank-1 factor x̂ satisfies min_{|s|=1}∥x̂ − sx*∥ ≤ c5∥ε∥/√n. The identical conclusion holds when X = F† diag(y) U is the solution map of the PhaseCut formulation evaluated at any second-order critical point U. The theorem says there are no spurious second-order critical poin

Load-bearing premise

The load-bearing premise is that every second-order critical point of the nonsmooth loss satisfies the quadratic-ascent inequality liminf_{t↓0}[f(X+tẊ)−f(X)]/t² ≥ 0 in every direction Ẋ, as used in Lemma 2; the paper's printed definition of second-order critical point uses a weaker limsup, and under the printed definition Lemma 2 is false, so the main theorems depend on the stronger interpretation being the intended one.

Editorial extensions

If this is right

  • Constant-overparametrization nonconvex optimization of the amplitude loss achieves order-optimal sample complexity n=O(d) for sub-Gaussian phase retrieval, matching semidefinite programming guarantees while optimizing over roughly d×p instead of d×d variables.
  • The nonconvex PhaseCut formulation inherits the same landscape guarantee, so a smooth sphere-synchronization problem has no bad second-order critical points either.
  • Ridge-regularized amplitude loss is provably statistically optimal in infinite-dimensional Gaussian phase retrieval, with the usual linear-regression bias term; this is a step toward nonparametric phase retrieval in reproducing kernel Hilbert spaces.
  • The deterministic inequality generalizes beyond amplitude loss, giving landscape control for quartic and Poisson losses (Theorems 1 and 3) in the same factored-matrix framework.

Reading between the lines

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

  • If the stronger second-order-criticality interpretation is retained, the amplitude loss's constant-overparametrization success suggests smoothing the loss (as several nearby works do) may be unnecessary for global landscape guarantees; the unsmoothed objective is already benign.
  • Because the deterministic landscape lemma is separated from the statistical concentration, the same proof template should transfer to other measurement models — coded diffraction patterns, Poisson noise, or non-isotropic covariances — once a stable phaseless injectivity bound is available.
  • The infinite-dimensional result hints that phase retrieval in an RKHS could be solved by the same regularized factored formulation with feature-space inner products, requiring only O(n²) kernel-matrix operations rather than explicit high-dimensional vectors.
  • In practice, constant p means the factorized algorithm stores O(np) measurement-phase variables instead of O(d log d), which could make the amplitude loss the method of choice for large-scale imaging if the constant c3 is small; the paper does not compute c3 numerically.
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 studies the global nonconvex landscape of phase retrieval formulated as overparametrized low-rank semidefinite matrix sensing, with two objectives: the nonsmooth amplitude least-squares loss (8) and a smooth PhaseCut-type reformulation (26). The main deterministic results are a general lemma (Lemma 1) on second-order critical points for losses of the form L_λ(XX^*), specialized to the quartic, amplitude, and Poisson losses (Theorems 1–3), plus a parallel landscape theorem for the PhaseCut formulation (Theorem 4). These are used to prove high-probability recovery guarantees with constant overparametrization: Theorem 5 for finite-dimensional phase retrieval with isotropic sub-Gaussian measurements, and Theorem 6 for Gaussian measurements in a Hilbert space. The claimed statistical error rates are optimal up to constants in the noise level, improving on prior quartic-loss results that required logarithmic overparametrization.

Significance. If the results are correct, the paper makes a substantial contribution: it provides the first global landscape guarantees for the nonsmooth amplitude loss with only constant overparametrization, and it shows that the same techniques cover a smooth PhaseCut formulation. The deterministic inequalities are explicit and parameter-free, and the statistical statements are sharp in their noise dependence. The paper is honest about its limitations, including the suboptimal constant in the non-adversarial noise comparison in Section 5.1. However, the central definition of second-order critical point is not consistent with the proofs, and Theorem 6 omits a necessary condition on the optimization rank. These are load-bearing issues that currently prevent the main theorems from being accepted as proved, although they appear to be fixable without changing the overall strategy.

major comments (2)
  1. [§3 (definition of SOCP) and §3.4 (proof of Lemma 2)] The printed definition of second-order critical point is limsup_{X'→X} [f(X')-f(X)]/||X'-X||^2 ≥ 0. This condition is strictly weaker than the property used in the proof of Lemma 2. It can hold at a strict saddle with both positive- and negative-curvature directions: for f(x,y)=x^2-y^2 at 0, the limsup is +1 along y=0, while along the y direction the ratio is -1. The proof of Lemma 2 replaces the definition with the per-direction condition L((X+t Xdot)(X+t Xdot)^*) - L(XX^*) ≥ o(t^2) for every fixed unit Xdot, i.e. liminf_{t↓0} of the second-order difference along every ray is nonnegative. No argument in §3.4 justifies this replacement. Since Lemma 2 is used to prove Lemma 1 and hence Theorems 1–3 and the nonsmooth branch of Theorem 5, the central landscape results are not proved as stated. The fix is to define SOCP by requiring the liminf to be nonnegative along every unit ray (equivale
  2. [§5.2, Theorem 6] The statement of Theorem 6 imposes no condition on the optimization rank p. The proof, however, explicitly uses p≥c (after equation (47)) and, in the real case, Theorem 2 is only valid for p≥2 because the denominator c_F p − 1 vanishes at p=1. Thus Theorem 6 as written is not valid for p=1 in the real case, and the current proof does not cover arbitrary p. The theorem statement should add a hypothesis such as 'for all p≥c3' (with c3 independent of n,d), consistent with Theorem 5 and with the proof.
minor comments (5)
  1. [§3.4, displayed equation after (18)] The typeset formula contains a stray 'ww' and is unreadable; please correct it.
  2. [§4, after the PhaseCut reformulation] The display 'min_{u∈F^n_1} 1/n ...' contains an extra '1' in the LaTeX source and should be cleaned up.
  3. [§5, first paragraph] Typo: 'we gives' should be 'we give'.
  4. [Lemma 8] The lemma asserts a bound for 'any X that is a first-order critical point of (13)', but first-order criticality of the nonsmooth objective is never defined. The proof uses ∇L_λ(XX^*)X=0, which is a conclusion of Lemma 2 for second-order critical points. Please restate Lemma 8 for the class of points actually used (those satisfying the criticality conditions of Lemma 2) or define the intended first-order notion.
  5. [§2] Typo: 'in the regimen≲dlogd' should be 'in the regime n≲d log d'.

Circularity Check

0 steps flagged · score 0.0 of 10

No material circularity: the main inequalities are derived in-paper; the self-citations are benchmarks or locally re-proved lemmas.

full rationale

The derivation chain is self-contained rather than circular. Lemma 1 is proved directly from Lemma 2; Theorems 1-4 are then obtained by substituting explicit loss functions into Lemma 1 or, in the PhaseCut case, into the in-paper Lemma 3. The PhaseCut lemma is not imported as an unverified self-citation: the paper states 'This is a slight generalization of intermediate results in [38]; we provide a proof at the end of this section' and indeed supplies the proof in Section 4.2. Similarly, the statistical input Lemma 5 is labeled '([34], as stated in [1, Lem. 4])', so the load-bearing concentration/identifiability result originates externally even though [1] restates it. No parameter is fitted to data and then renamed as a prediction: the bounds are deterministic inequalities in terms of the problem data, followed by high-probability statements from concentration lemmas. The only substantive concern raised by a careful reading is not circularity: the printed definition of second-order critical point via a global limsup in Section 3 appears too weak for the per-direction quadratic-lower-bound argument used in the proof of Lemma 2. That is a proof gap or possible quantifier error, not a reduction of the conclusion to its own assumptions. Since circularity is the object of this pass, the appropriate finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No data are fitted anywhere in the paper: all claims are theorem-based. The optimization rank p and the ridge penalty λ are user-chosen hyperparameters (theorems require p ≥ c3 and λ ≥ c3(σ_{d+1} + average tail), respectively). Universal constants c1–c5 are existential, not computed. The mathematical inputs are the measurement-ensemble assumptions (sub-Gaussian entries, Gaussian Hilbert-space measurements with trace-class covariance) and the imported small-ball/nuclear-norm equivalence of Lemma 5, which does the heavy lifting in converting measurement errors into matrix norms.

free parameters (2)
  • optimization rank p = p ≥ c3 (universal constant)
    The overparametrization level p is a free variable of the algorithm; the theorems require it to exceed an unspecified universal constant. Not fitted to data.
  • regularization λ (Theorem 6) = λ ≥ c3(σ_{d+1} + (1/n)Σ_{m>d} σ_m), also λ ≥ cδ
    Regularizer lower bound chosen to control the spectral tail in infinite dimensions; not fitted to data, but a tuning parameter that must be selected using knowledge of the covariance spectrum.
assumptions (5)
  • domain assumption Lemma 5 (Krahmer–Stöger small-ball/nuclear-norm equivalence): for sub-Gaussian w satisfying E|w|²=1, K-subGaussian, and either E|w|⁴>1 or ∥x*∥∞ sufficiently small, with probability ≥ 1−c e^{−c n}, (1/n)Σ|⟨Ai, Z−Z*⟩| ≥ c∥Z−Z*∥* for all PSD Z when n ≥ c1 d.
    Imported from [34] as restated in [1, Lem. 4] and used in the proof of Theorem 5 to convert measurement discrepancy into matrix norm. Load-bearing for the statistical guarantee.
  • domain assumption Isotropic sub-Gaussian measurement model: entries of f_i are i.i.d. zero-mean w with E|w|²=1 and K-subGaussian; |Ew²|<1 if F=C.
    The entire Theorem 5 applies only under this ensemble; moment and flatness conditions are prior-literature conditions ensuring identifiability.
  • domain assumption Gaussian measurements with trace-class covariance Σ in a Hilbert space (Theorem 6), with circular symmetry when F=C.
    Concentration of the empirical covariance (Lemma 9) and restricted eigenvalue conditions require Gaussianity; the result does not extend to sub-Gaussian in infinite dimension.
  • domain assumption Convexity and appropriate differentiability of the loss ℓ(b,υ) in Lemma 1, with lim_{b↓0} ℓ′(b,υ) possibly −∞.
    Framework condition enabling the Taylor-expansion argument; holds for the quartic, amplitude, and Poisson losses considered.
  • standard math Standard concentration inequalities: Lemma 4 (sub-Gaussian covariance) and Lemma 9 (Hilbert-space covariance, Koltchinskii–Lounici).
    Quoted from [48]; used to control the empirical covariance of the measurement vectors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Phase retrieval via overparametrized nonconvex optimization: nonsmooth amplitude loss landscapes." pith.science (2026). https://pith.science/paper/PFIA7MKS

@misc{pith2026251119045,
  author       = {Pith},
  title        = {Pith review of: Phase retrieval via overparametrized nonconvex optimization: nonsmooth amplitude loss landscapes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PFIA7MKS}},
  note         = {Machine review of arXiv:2511.19045}
}
read the original abstract

We study nonconvex optimization for phase retrieval and the more general problem of semidefinite low-rank matrix sensing; in particular, we focus on the global nonconvex landscape of overparametrized versions of the nonsmooth amplitude least-squares loss as well as a smooth reformulation of this loss based on the PhaseCut approach. We first give a general, deterministic result on properties of second-order critical points for a general class of loss functions; we then specialize this result to the nonsmooth amplitude loss and, additionally, prove nearly identical results for a smooth reformulation (similar to PhaseCut) as a synchronization problem over spheres. Finally, we show the usefulness of these tools by proving high-probability landscape guarantees in two settings: (1) phase retrieval with isotropic sub-Gaussian measurements, and (2) phase retrieval in a general (possibly infinite-dimensional) Hilbert space with Gaussian measurements. In both cases, our results give state-of-the-art and statistically optimal guarantees with only a constant amount of overparametrization (in the well-studied case of isotropic sub-Gaussian measurements, such statistical guarantees had previously required greater degrees of overparametrization/relaxation); this demonstrates the potential of overparametrized nonconvex optimization as a principled and scalable algorithmic approach to phase retrieval.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. A Gaussian smoothing-based zeroth-order method for Goldstein second-order stationarity

    math.OC 2026-07 conditional novelty 6.0 of 10

    Introduces the Goldstein second-order δ-subdifferential for C¹,¹ functions and a Gaussian-smoothing cubic-regularization zeroth-order method that provably finds (ε₁, ε₂, δ)-second-order stationary points under a coerc...

Reference graph

Works this paper leans on

49 extracted references · cited by 1 Pith paper

  1. [34]

    Complex phase retrieval from subgaussian measurements,

    F. Krahmer and D. St¨ oger, “Complex phase retrieval from subgaussian measurements,”J. Fourier Anal. Appl., vol. 26, no. 89, 2020

    2020

  2. [1]

    Phase retrieval and matrix sensing via benign and overparametrized nonconvex optimization,

    A. D. McRae, “Phase retrieval and matrix sensing via benign and overparametrized nonconvex optimization,” 2025

    2025

  3. [2]

    On Lipschitz analysis and Lipschitz synthesis for the phase retrieval problem,

    R. Balan and D. Zou, “On Lipschitz analysis and Lipschitz synthesis for the phase retrieval problem,” Linear Algebra Appl., vol. 496, pp. 152–181, 2016

    2016

  4. [3]

    Lipschitz analysis of generalized phase retrievable matrix frames,

    R. Balan and C. B. Dock, “Lipschitz analysis of generalized phase retrievable matrix frames,”SIAM J. Matrix Anal. Appl., vol. 43, no. 3, pp. 1518–1571, 2022. 28

    2022

  5. [4]

    Experimental robustness of Fourier ptychography phase retrieval algorithms,

    L.-H. Yeh, J. Dong, J. Zhong, L. Tian, M. Chen, G. Tang, M. Soltanolkotabi, and L. Waller, “Experimental robustness of Fourier ptychography phase retrieval algorithms,”Opt. Express, vol. 23, no. 26, pp. 33 214–33 240, 2015

    2015

  6. [5]

    Solving systems of random quadratic equations via truncated amplitude flow,

    G. Wang, G. B. Giannakis, and Y. C. Eldar, “Solving systems of random quadratic equations via truncated amplitude flow,”IEEE Trans. Inf. Theory, vol. 64, no. 2, pp. 773–794, 2018

    2018

  7. [6]

    Reshaped Wirtinger flow for solving quadratic system of equations,

    H. Zhang and Y. Liang, “Reshaped Wirtinger flow for solving quadratic system of equations,” in Proc. Conf. Neural Inf. Process. Syst. (NeurIPS), vol. 29, Dec. 2016. [Online]. Available: https: //papers.nips.cc/paper files/paper/2016/hash/83adc9225e4deb67d7ce42d58fe5157c-Abstract.html

    2016

  8. [7]

    Maximum-likelihood estimation in ptychography in the presence of Poisson–Gaussian noise statistics,

    J. Seifert, Y. Shao, R. van Dam, D. Bouchet, T. van Leeuwen, and A. P. Mosk, “Maximum-likelihood estimation in ptychography in the presence of Poisson–Gaussian noise statistics,”Opt. Lett., vol. 48, no. 22, pp. 6027–6030, 2023

    2023

Show all 49 references
  1. [8]

    The local landscape of phase retrieval under limited samples,

    K. Liu, Z. Wang, and L. Wu, “The local landscape of phase retrieval under limited samples,”IEEE Trans. Inf. Theory, vol. 70, no. 12, pp. 9012–9035, 2024

    2024

  2. [9]

    Phase recovery, MaxCut and complex semidefinite programming,

    I. Waldspurger, A. d’Aspremont, and S. Mallat, “Phase recovery, MaxCut and complex semidefinite programming,”Math. Program., vol. 149, pp. 47–81, 2015

    2015

  3. [10]

    The Riemannian Elevator for certifiable distance-based localization,

    T. Halsted and M. Schwager, “The Riemannian Elevator for certifiable distance-based localization,”

  4. [11]

    Phase retrieval with application to optical imaging: A contemporary overview,

    Y. Shechtman, Y. C. Eldar, O. Cohen, H. N. Chapman, J. Miao, and M. Segev, “Phase retrieval with application to optical imaging: A contemporary overview,”IEEE Signal Process. Mag., vol. 32, no. 3, pp. 87–109, 2015

    2015

  5. [12]

    The numerics of phase retrieval,

    A. Fannjiang and T. Strohmer, “The numerics of phase retrieval,”Acta Numer., vol. 29, pp. 125–228, 2020

    2020

  6. [13]

    Phase retrieval: From computational imaging to machine learning: A tutorial,

    J. Dong, L. Valzania, A. Maillard, T.-a. Pham, S. Gigan, and M. Unser, “Phase retrieval: From computational imaging to machine learning: A tutorial,”IEEE Signal Process. Mag., vol. 40, no. 1, pp. 45–57, 2023

    2023

  7. [14]

    A geometric analysis of phase retrieval,

    J. Sun, Q. Qu, and J. Wright, “A geometric analysis of phase retrieval,”Found. Comput. Math., vol. 18, no. 5, pp. 1131–1198, 2018

    2018

  8. [15]

    Nearly optimal bounds for the global geometric landscape of phase retrieval,

    J.-F. Cai, M. Huang, D. Li, and Y. Wang, “Nearly optimal bounds for the global geometric landscape of phase retrieval,”Inverse Probl., vol. 39, no. 7, 2023

    2023

  9. [16]

    Complex dynamics in simple neural networks: Understanding gradient flow in phase retrieval,

    S. Sarao Mannelli, G. Biroli, C. Cammarota, F. Krzakala, P. Urbani, and L. Zdeborov´ a, “Complex dynamics in simple neural networks: Understanding gradient flow in phase retrieval,” inProc. Conf. Neural Inf. Process. Syst. (NeurIPS), vol. 33, Virtual conference, Dec. 2020, pp....

    2020

  10. [17]

    The nonsmooth landscape of phase retrieval,

    D. Davis, D. Drusvyatskiy, and C. Paquette, “The nonsmooth landscape of phase retrieval,”IMA J. Numer. Anal., vol. 40, no. 4, pp. 2652–2695, 2020

    2020

  11. [18]

    Toward the optimal construction of a loss function without spurious local minima for solving quadratic equations,

    Z. Li, J.-F. Cai, and K. Wei, “Toward the optimal construction of a loss function without spurious local minima for solving quadratic equations,”IEEE Trans. Inf. Theory, vol. 66, no. 5, pp. 3242– 3260, 2020

    2020

  12. [19]

    The global landscape of phase retrieval I: Perturbed amplitude models,

    J.-F. Cai, M. Huang, D. Li, and Y. Wang, “The global landscape of phase retrieval I: Perturbed amplitude models,”Ann. Appl. Math., vol. 37, no. 4, pp. 437–512, 2021

    2021

  13. [20]

    Solving phase retrieval with random initial guess is nearly as good as by spectral initializa- tion,

    ——, “Solving phase retrieval with random initial guess is nearly as good as by spectral initializa- tion,”Appl. Comput. Harmon. Anal., vol. 58, pp. 60–84, 2022

    2022

  14. [21]

    The global landscape of phase retrieval II: Quotient intensity models,

    ——, “The global landscape of phase retrieval II: Quotient intensity models,”Ann. Appl. Math., vol. 38, no. 1, pp. 62–114, 2022

    2022

  15. [22]

    Smoothed amplitude flow-based phase retrieval algorithm,

    Z. Xiao, Y. Wang, and G. Gui, “Smoothed amplitude flow-based phase retrieval algorithm,”J. Franklin Inst., vol. 358, no. 14, pp. 7270–7285, 2021. 29

    2021

  16. [23]

    Truncated amplitude flow with coded diffraction patterns,

    H. Li and J. Li, “Truncated amplitude flow with coded diffraction patterns,”Inverse Probl., vol. 41, no. 1, 2025

    2025

  17. [24]

    Phase retrieval via reweighted amplitude flow,

    G. Wang, G. B. Giannakis, Y. Saad, and J. Chen, “Phase retrieval via reweighted amplitude flow,” IEEE Trans. Inf. Theory, vol. 66, no. 11, pp. 2818–2833, 2018

    2018

  18. [25]

    A nonconvex approach for phase retrieval: Reshaped Wirtinger flow and incremental algorithms,

    H. Zhang, Y. Zhou, Y. Liang, and Y. Chi, “A nonconvex approach for phase retrieval: Reshaped Wirtinger flow and incremental algorithms,”J. Mach. Learn. Res., vol. 18, pp. 1–35, 2017. [Online]. Available: https://jmlr.org/papers/v18/16-572.html

    2017

  19. [26]

    Phase retrieval via smoothed amplitude flow,

    Q. Luo, H. Wang, and S. Lin, “Phase retrieval via smoothed amplitude flow,”Signal Process., vol. 177, 2020

    2020

  20. [27]

    Implicit regularization in nonconvex statistical estimation: Gradient descent converges linearly for phase retrieval, matrix completion, and blind deconvolution,

    C. Ma, K. Wang, Y. Chi, and Y. Chen, “Implicit regularization in nonconvex statistical estimation: Gradient descent converges linearly for phase retrieval, matrix completion, and blind deconvolution,” Found. Comput. Math., vol. 20, pp. 451–632, 2020

    2020

  21. [28]

    Noisy phase retrieval from subgaussian measurements,

    H. Peng, D. Han, L. Li, and M. Huang, “Noisy phase retrieval from subgaussian measurements,” 2024

    2024

  22. [29]

    Phase retrieval for sub-Gaussian measurements,

    B. Gao, H. Liu, and Y. Wang, “Phase retrieval for sub-Gaussian measurements,”Appl. Comput. Harmon. Anal., vol. 53, pp. 95–115, 2021

    2021

  23. [30]

    Gradient descent with random initialization: fast global convergence for nonconvex phase retrieval,

    Y. Chen, Y. Chi, J. Fan, and C. Ma, “Gradient descent with random initialization: fast global convergence for nonconvex phase retrieval,”Math. Program., vol. 176, pp. 5–37, 2019

    2019

  24. [31]

    Convergence analysis of Wirtinger flow for Poisson phase retrieval,

    B. Gao, R. Gu, and S. Ma, “Convergence analysis of Wirtinger flow for Poisson phase retrieval,” 2025

    2025

  25. [32]

    Phase retrieval: Stability and recovery guarantees,

    Y. C. Eldar and S. Mendelson, “Phase retrieval: Stability and recovery guarantees,”Appl. Comput. Harmon. Anal., vol. 36, pp. 473–494, 2014

    2014

  26. [33]

    Phase retrieval without small-ball probability assumptions,

    F. Krahmer and Y.-K. Liu, “Phase retrieval without small-ball probability assumptions,”IEEE Trans. Inf. Theory, vol. 64, no. 1, pp. 485–500, 2018

    2018

  27. [35]

    Solving orthogonal group synchronization via convex and low-rank optimization: Tightness and landscape analysis,

    S. Ling, “Solving orthogonal group synchronization via convex and low-rank optimization: Tightness and landscape analysis,”Math. Program., vol. 200, pp. 589–628, 2023

    2023

  28. [36]

    Local geometry determines global landscape in low-rank factorization for synchronization,

    ——, “Local geometry determines global landscape in low-rank factorization for synchronization,” Found. Comput. Math., 2025

    2025

  29. [37]

    Benign landscape for Burer-Monteiro factorizations of MaxCut-type semidefinite programs,

    F. Rakoto Endor and I. Waldspurger, “Benign landscape for Burer-Monteiro factorizations of MaxCut-type semidefinite programs,” 2024

    2024

  30. [38]

    Benign landscapes for synchronization on spheres via normalized Laplacian matri- ces,

    A. D. McRae, “Benign landscapes for synchronization on spheres via normalized Laplacian matri- ces,” 2025

    2025

  31. [39]

    Local and global linear convergence of general low-rank matrix recovery problems,

    Y. Bi, H. Zhang, and J. Lavaei, “Local and global linear convergence of general low-rank matrix recovery problems,” inProc. AAAI Conf. Artif. Intell. (AAAI), vol. 36, no. 9, Virtual conference, Feb. 2022, pp. 10 129–10 137

    2022

  32. [40]

    Sharp global guarantees for nonconvex low-rank recovery in the noisy overparame- terized regime,

    R. Y. Zhang, “Sharp global guarantees for nonconvex low-rank recovery in the noisy overparame- terized regime,”SIAM J. Optim., vol. 35, no. 3, pp. 2128–2154, 2025

    2025

  33. [41]

    Concise complexity analyses for trust region meth- ods,

    F. E. Curtis, Z. Lubberts, and D. P. Robinson, “Concise complexity analyses for trust region meth- ods,”Optim. Lett., vol. 12, pp. 1713–1724, 2018

    2018

  34. [42]

    First-order methods almost always avoid strict saddle points,

    J. D. Lee, I. Panageas, G. Piliouras, M. Simchowitz, M. I. Jordan, and B. Recht, “First-order methods almost always avoid strict saddle points,”Math. Program., vol. 176, pp. 311–337, 2019

    2019

  35. [43]

    On the hardness of meaningful local guarantees in nonsmooth nonconvex optimization,

    G. Kornowski, S. Padmanabhan, and O. Shamir, “On the hardness of meaningful local guarantees in nonsmooth nonconvex optimization,” 2024. 30

    2024

  36. [44]

    The non-convex geometry of low-rank matrix optimization,

    Q. Li, Z. Zhu, and G. Tang, “The non-convex geometry of low-rank matrix optimization,”Inform. Inference., vol. 8, no. 1, pp. 51–96, 2019

    2019

  37. [45]

    Deterministic guarantees for Burer-Monteiro factor- izations of smooth semidefinite programs,

    N. Boumal, V. Voroninski, and A. S. Bandeira, “Deterministic guarantees for Burer-Monteiro factor- izations of smooth semidefinite programs,”Commun. Pure Appl. Math., vol. 73, no. 3, pp. 581–608, 2019

    2019

  38. [46]

    Random design analysis of ridge regression,

    D. Hsu, S. M. Kakade, and T. Zhang, “Random design analysis of ridge regression,”Found. Comput. Math., vol. 14, pp. 569–600, 2014

    2014

  39. [47]

    Sample complexity and effective dimension for regression on manifolds,

    A. D. McRae, J. Romberg, and M. A. Davenport, “Sample complexity and effective dimension for regression on manifolds,” inProc. Conf. Neural Inf. Process. Syst. (NeurIPS), Virtual conference, Dec. 2020, pp. 12 993–13 004. [Online]. Available: https://proceedings.neurips.cc/pape...

    2020

  40. [48]

    Concentration inequalities and moment bounds for sample covari- ance operators,

    V. Koltchinskii and K. Lounici, “Concentration inequalities and moment bounds for sample covari- ance operators,”Bernoulli, vol. 23, no. 1, pp. 110–133, 2017. 31

    2017

  41. [2022]

    Available: https://msl-dev.stanford.edu/papers/halsted riemannian 2022.pdf

    [Online]. Available: https://msl-dev.stanford.edu/papers/halsted riemannian 2022.pdf

    2022

Pith tools

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