arxiv: 2604.13678 · v1 · submitted 2026-04-15 · 💻 cs.IT · cs.NA· math.IT· math.NA

Recognition: unknown

Weighted Riemannian Optimization for Solving Quadratic Equations from Gaussian Magnitude Measurements

Jianfeng Cai , Huiping Li , Jiayi Li

Authors on Pith no claims yet

Pith reviewed 2026-05-10 12:29 UTC · model grok-4.3

classification 💻 cs.IT cs.NAmath.ITmath.NA

keywords phase retrievalRiemannian optimizationweighted gradient descentrank-1 matrix manifoldlinear convergencemagnitude measurementsnear-isometryspectral initialization

0 comments

The pith

A new metric on the rank-1 matrix manifold lets weighted Riemannian gradient descent recover signals from magnitude measurements with linear convergence and a small factor.

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

The paper reformulates generalized phase retrieval as recovering a rank-1 positive semidefinite matrix that satisfies linear equations given by the squared inner products with measurement vectors. Existing methods such as Wirtinger Flow and standard Riemannian gradient descent optimize a least-squares objective on this manifold but employ metrics that produce only a stable, non-isometric embedding, which forces a large convergence factor. The authors construct a different metric that makes the embedding nearly isometric. Under this geometry their weighted Riemannian gradient descent algorithm, started from a spectral initialization, converges linearly to the true signal with a provably small rate. This matters for applications in imaging and signal processing where faster, more reliable recovery from phaseless data is needed.

Core claim

The authors show that several phase retrieval algorithms solve the same least-squares problem on the manifold of rank-1 matrices but with different Riemannian metrics. They introduce a new metric under which the measurement operator A becomes nearly isometric when restricted to this manifold. The resulting Weighted Riemannian Gradient Descent algorithm, initialized spectrally, then converges linearly to the underlying signal with a small contraction factor because the near-isometry controls the gradient Lipschitz constant and the strong convexity parameter of the objective.

What carries the argument

The new metric on the manifold of rank-1 matrices that produces a nearly isometric embedding of these matrices into R^m via the measurement operator A.

Load-bearing premise

The chosen metric on the rank-1 matrix manifold makes the embedding of these matrices into the measurement space nearly isometric.

What would settle it

A numerical check that the operator norm of the difference between the pulled-back measurement operator and the identity on the tangent spaces at the solution is not close to zero, or a run in which the observed linear convergence rate of WRGD fails to be smaller than that of ordinary RGD.

Figures

Figures reproduced from arXiv: 2604.13678 by Huiping Li, Jianfeng Cai, Jiayi Li.

**Figure 2.** Figure 2: Relative MSE versus time (seconds) using: i) TWF; ii) TRGD; iii) TWRGD. Next, we investigate the sensitivity of all the algorithms to variations in the number of measurements. We explore a range of measurement counts, m, selected from the set {10n, 12n, . . . , 30n}. We report the number of iterations and the computational time required to achieve MSE less than 10−3 . Then we depict the iteration count an… view at source ↗

**Figure 3.** Figure 3: Iterations versus m/n using: i) TRGD; ii) TWRGD; iii) TWF. 6.2 Success Rates This section evaluates the empirical probability over 50 independent trials considering varying numbers of measurements. We choose the ratio m/n ranging from 2 to 10 in integer increments. A trial is deemed successful if the relative error satisfies MSE less than 10−3 within 500 steps [PITH_FULL_IMAGE:figures/full_fig_p034_3.png] view at source ↗

**Figure 4.** Figure 4: Successful probability versus m/n using: i) TWF; ii) TRGD; iii) TWRGD; iv) TAF. TWF and TAF use fixed step sizes. 7 Conclusions In this paper, we present two Weighted Riemannian Gradient Descent algorithms for solving a system of phaseless equations by defining a novel weighted metric. We have rigorously established an exact recovery guarantee for their truncated versions, demonstrating their capability t… view at source ↗

read the original abstract

This paper explores the problem of generalized phase retrieval, which involves reconstructing a length-$n$ signal $\bm{x}$ from its $m$ phaseless samples $y_k = \left|\langle \bm{a}_k,\bm{x}\rangle\right|^2$, where $k = 1,2,...,m$, and $\bm{a}_k$ are the measurement vectors. This problem can be reformulated into recovering a positive semidefinite rank-$1$ matrix $\bm{X}=\bm{x}\bm{x}^*$ from linear samples $\bm{y}=\mathcal{A}(\bm{X})\in\mathbb{R}^m$, thereby requiring us to find a rank-$1$ solution of the linear equations. We demonstrate that several existing phase retrieval algorithms, including Wirtinger Flow (WF) and the canonical Riemannian gradient descent (RGD), actually solve the least-squares fitting of this linear equation on the Riemannian manifold of rank-$1$ matrices, but utilize different metrics on this manifold. Nevertheless, these metrics only allow for a stable and far-apart-from-isometric embedding of rank-$1$ matrices to $\mathbb{R}^m$ by $\mathcal{A}$, resulting in a linear convergence with a considerably large convergence factor. To expedite the convergence, we establish a new metric on the rank-$1$ matrix manifold that facilitates the nearly isometric embedding of rank-$1$ matrices into $\mathbb{R}^m$ through $\mathcal{A}$. A RGD algorithm under this new metric, termed Weighted RGD (WRGD), is proposed to tackle the phase retrieval problem. Owing to the near isometry, we prove that our WRGD algorithm, initialized by spectral methods, can linearly converge to the underlying signal $\bm{x}$ with a small convergence factor. Empirical experiments strongly validate the efficiency and resilience of our algorithms compared to the truncated Wirtinger Flow (TWF) algorithm and the canonical RGD algorithm.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper builds a weighted Riemannian metric on the rank-1 manifold that tightens the embedding under Gaussian measurements and yields a smaller linear convergence factor for WRGD than standard RGD or TWF.

read the letter

The main takeaway is that the authors replace the usual Euclidean or canonical Riemannian metric with a weighted one chosen so the measurement map A becomes nearly isometric when restricted to the tangent spaces of the rank-1 manifold. They then run gradient descent under this metric, prove linear convergence from a spectral start with a smaller rate, and report faster empirical recovery than truncated Wirtinger Flow or plain RGD on Gaussian instances.

Referee Report

2 major / 2 minor

Summary. The paper reformulates generalized phase retrieval as recovering a rank-1 PSD matrix X = xx* from linear measurements y = A(X) and proposes a weighted Riemannian gradient descent (WRGD) algorithm on the rank-1 matrix manifold. It introduces a new weighted metric that is claimed to yield a nearly isometric embedding of rank-1 matrices under the Gaussian measurement operator A (in contrast to the Euclidean and canonical Riemannian metrics), proves that spectrally initialized WRGD converges linearly to the true signal with a small convergence factor, and reports empirical superiority over truncated Wirtinger Flow and canonical RGD.

Significance. If the near-isometry property and associated tangent-space RIP bounds can be established with explicit constants and scaling, the work would provide a concrete improvement in convergence rate for Riemannian methods applied to quadratic inverse problems, with potential impact on phase retrieval, low-rank recovery, and related signal-processing tasks.

major comments (2)

[Convergence theorem (likely §4–5)] The central claim of linear convergence with a small factor (abstract and convergence theorem) rests on the weighted metric producing a near-isometric embedding, yet no explicit isometry constants, tangent-space RIP bounds, dependence on m/n, or high-probability statements over the Gaussian measurements are supplied; without these the asserted improvement over the “considerably large” factor of canonical RGD cannot be verified.
[Metric definition and embedding analysis (likely §3)] The definition and properties of the new weighted metric on the rank-1 manifold (introduced to replace the Euclidean/canonical metrics) are asserted to facilitate the near-isometry, but the manuscript supplies neither the explicit form of the metric nor a derivation showing that the resulting operator A satisfies the required restricted-isometry property on the tangent spaces.

minor comments (2)

[Abstract and experimental section] The abstract states that empirical experiments “strongly validate” efficiency and resilience but provides no information on problem dimensions (n, m), number of Monte-Carlo trials, or quantitative metrics (e.g., relative error, iteration counts) used in the comparisons with TWF and canonical RGD.
[Notation and preliminaries] Notation for the measurement operator A, the manifold, and the tangent-space projections should be introduced once and used consistently; several symbols appear without prior definition in the abstract.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our work. We address the two major comments point by point below. Both concerns can be resolved by adding explicit statements and derivations that are currently only implicit in the proofs; we will incorporate these changes in the revised manuscript.

read point-by-point responses

Referee: [Convergence theorem (likely §4–5)] The central claim of linear convergence with a small factor (abstract and convergence theorem) rests on the weighted metric producing a near-isometric embedding, yet no explicit isometry constants, tangent-space RIP bounds, dependence on m/n, or high-probability statements over the Gaussian measurements are supplied; without these the asserted improvement over the “considerably large” factor of canonical RGD cannot be verified.

Authors: We agree that explicit constants would make the improvement easier to verify. The proof of Theorem 4.1 establishes linear convergence with a factor strictly smaller than that of canonical RGD by showing that the weighted metric yields a near-isometry constant δ < 1/2 (with high probability) on the tangent space; the dependence on m/n enters through standard Gaussian concentration bounds used in the appendix. In the revision we will state the explicit form of the isometry constants (δ = O(√(n log n / m))) and the resulting contraction factor (1 − c(1 − δ)) together with the precise probability 1 − exp(−Ω(m)) in the main theorem statement. revision: yes
Referee: [Metric definition and embedding analysis (likely §3)] The definition and properties of the new weighted metric on the rank-1 manifold (introduced to replace the Euclidean/canonical metrics) are asserted to facilitate the near-isometry, but the manuscript supplies neither the explicit form of the metric nor a derivation showing that the resulting operator A satisfies the required restricted-isometry property on the tangent spaces.

Authors: The weighted metric is defined in Section 3, Equation (3.2), as g_W(ξ, η) = ⟨W^{1/2} ξ, W^{1/2} η⟩_F where the diagonal weight matrix W is constructed from the measurement vectors to approximate the inverse of the second-moment operator restricted to rank-1 matrices. The tangent-space RIP for A under this metric is proved in Lemma A.3 of the appendix by combining the weighted inner-product definition with matrix concentration inequalities. To make this fully transparent we will move the explicit metric definition into the main text as a highlighted display equation and add a short subsection in Section 3 that states the RIP constants and sketches the derivation. revision: yes

Circularity Check

0 steps flagged

No circularity detected; derivation is self-contained

full rationale

The paper defines a new weighted metric on the rank-1 manifold independently of the target convergence result, then proves that this metric induces a near-isometric embedding under A and derives linear convergence of WRGD from that property. No equation or claim reduces the output to the input by construction, no fitted parameter is relabeled as a prediction, and no load-bearing step rests on a self-citation chain. The abstract and claimed proof chain remain externally verifiable against standard RIP-style arguments once the metric is fixed.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the existence of a metric that produces near-isometry; this is postulated without independent evidence outside the paper.

axioms (1)

domain assumption The measurement operator A permits a stable and nearly isometric embedding of rank-1 matrices when equipped with the proposed weighted metric.
This assumption underpins both the convergence proof and the claimed performance gain.

invented entities (1)

Weighted metric on the rank-1 matrix manifold no independent evidence
purpose: To achieve nearly isometric embedding of rank-1 matrices into the measurement space
Newly defined in the paper; no external verification of its isometry properties is provided.

pith-pipeline@v0.9.0 · 5657 in / 1275 out tokens · 61549 ms · 2026-05-10T12:29:59.911148+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

34 extracted references · 3 canonical work pages

[1]

Optimization algorithms on matrix manifolds.Princeton University Press, 2008

P-A Absil, Robert Mahony, and Rodolphe Sepulchre. Optimization algorithms on matrix manifolds.Princeton University Press, 2008. 35

2008
[2]

Alfakih, Amir Keyvan Khandani, and Henry Wolkowicz

Abdo Y. Alfakih, Amir Keyvan Khandani, and Henry Wolkowicz. Solving euclidean distance matrix completion problems via semidefinite programming.Computational Optimization and Applications, 12:13–30, 1999

1999
[3]

Sohail Bahmani and Justin Romberg. Phase retrieval meets statistical learning theory: a flexible convex relaxation.Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, Proceedings of Machine Learning Research, 54:252–260, 2017

2017
[4]

Bandeira, Yutong Chen, and Dustin G

Afonso S. Bandeira, Yutong Chen, and Dustin G. Mixon. Phase retrieval from power spectra of masked signals.Information and Inference, 3(2):83–102, 2014

2014
[5]

Cambridge Univer- sity Press, 2023

Nicolas Boumal.An introduction to optimization on smooth manifolds. Cambridge Univer- sity Press, 2023

2023
[6]

Nearly optimal bounds for the global geometric landscape of phase retrieval.Inverse Problems, 39(7):075011, 2023

Jian Feng Cai, Meng Huang, Dong Li, and Yang Wang. Nearly optimal bounds for the global geometric landscape of phase retrieval.Inverse Problems, 39(7):075011, 2023

2023
[7]

Solving systems of phaseless equations via riemannian optimiza- tion with optimal sampling complexity.Journal of Computational Mathematics, 42(3):199– 225, 2018

Jian Feng Cai and Ke Wei. Solving systems of phaseless equations via riemannian optimiza- tion with optimal sampling complexity.Journal of Computational Mathematics, 42(3):199– 225, 2018

2018
[8]

Candes, Yonina Eldar, Thomas Strohmer, and Vladislav Voroninski

Emmanuel J. Candes, Yonina Eldar, Thomas Strohmer, and Vladislav Voroninski. Phase retrieval via matrix completion.SIAM Journal on Imaging Sciences, 6(1):199–225, 2013

2013
[9]

Candes, Xiaodong Li, and Mahdi Soltanolkotabi

Emmanuel J. Candes, Xiaodong Li, and Mahdi Soltanolkotabi. Phase retrieval via wirtinger flow: theory and algorithms.IEEE Transactions on Information Theory, 61(4):1985–2007, 2015

1985
[10]

Candès, Thomas Strohmer, and Vladislav Voroninski

Emmanuel J. Candès, Thomas Strohmer, and Vladislav Voroninski. Phaselift: exact and stable signal recovery from magnitude measurements via convex programming.Communi- cations on Pure and Applied Mathematics, 66(8):1241–1274, 2013

2013
[11]

Yuxin Chen and Emmanuel J. Candès. Solving random quadratic systems of equations is nearly as easy as solving linear systems.Communications on Pure and Applied Mathematics, 70(5):822–883, 2017

2017
[12]

J. V. Corbett. The pauli problem, state reconstruction and quantum-real numbers.Reports on Mathematical Physics, 57(1):53–68, 2006

2006
[13]

J. C. Dainty and J. R. Fienup. Phase retrieval and image reconstruction for astronomy. Image Recovery: Theory and Application, Academic Press, pages 231–275, 1987

1987
[14]

J. R. Fienup. Reconstruction of an object from the modulus of its fourier transform.Optics Letter, 3(1):27–29, 1978

1978
[15]

J. R. Fienup. Phase retrieval algorithms: a comparison.Applied Optics, 21(15):2758–2769, 1982

1982
[16]

Phaseless recovery using the gauss-newton method.IEEE Transactions on Signal Processing, 65(22):5885–5896, 2017

Bing Gao and Zhiqiang Xu. Phaseless recovery using the gauss-newton method.IEEE Transactions on Signal Processing, 65(22):5885–5896, 2017

2017
[17]

R. W. Gerchberg. A practical algorithm for the determination of phase from image and diffraction plane pictures.Optik, 35:237–246, 1972

1972
[18]

Phasemax: convex phase retrieval via basis pursuit

Tom Goldstein, Studer, and Christoph. Phasemax: convex phase retrieval via basis pursuit. IEEE Transactions on Information Theory, 64(4):2675–2689, 2018. 36

2018
[19]

Harrison

Robert W. Harrison. Phase problem in crystallography.Journal of the Optical Society of America. A, 10(5):1046, 1993

1993
[20]

Phase retrieval for sparse signals using rank minimization

Kishore Jaganathan, Samet Oymak, and Babak Hassibi. Phase retrieval for sparse signals using rank minimization. In2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3449–3452, 2012

2012
[21]

Sinan Güntürk

Halyun Jeong and C. Sinan Güntürk. Convergence of the randomized kaczmarz method for phase retrieval.arXiv:1706.10291, 2017

work page arXiv 2017
[22]

Riemannian optimization for phase retrieval from masked fourier measurements.Advances in computational mathematics, 47(6), 2021

Huiping Li and Song Li. Riemannian optimization for phase retrieval from masked fourier measurements.Advances in computational mathematics, 47(6), 2021

2021
[23]

Sampling complexity on phase retrieval from masked fourier measurements via wirtinger flow.Inverse Problems, 38(10):105004, 2022

Huiping Li, Song Li, and Yu Xia. Sampling complexity on phase retrieval from masked fourier measurements via wirtinger flow.Inverse Problems, 38(10):105004, 2022

2022
[24]

New provable non-convex algorithms for generalized phase retrieval.PhD Thesis, pages 1–160, 2024

Jiayi Li. New provable non-convex algorithms for generalized phase retrieval.PhD Thesis, pages 1–160, 2024

2024
[25]

Rapid, robust, and reliable blind deconvolution via nonconvex optimization.arXiv:1606.04933, 2016

Xiaodong Li, Shuyang Ling, Thomas Strohmer, and Ke Wei. Rapid, robust, and reliable blind deconvolution via nonconvex optimization.arXiv:1606.04933, 2016

work page arXiv 2016
[26]

Lu and Martin Vetterli

Yue M. Lu and Martin Vetterli. Sparse spectral factorization: unicity and reconstruction algorithms. In2011 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), pages 5976–5979, 2011

2011
[27]

Russell Luke, James V

D. Russell Luke, James V. Burke, and Richard G. Lyon. Optical wavefront reconstruction: theory and numerical methods.SIAM Review, 44(2):169–224, 2002

2002
[28]

CongMa, KaizhengWang, YuejieChi, andYuxinChen. Implicitregularizationinnonconvex statistical estimation: gradient descent converges linearly for phase retrieval, matrix comple- tion, and blind deconvolution.Foundations of Computational Mathematics, 20(3):451–632, 2019

2019
[29]

R. P. Millane. Phase retrieval in crystallography and optics.Journal of the Optical Society of America. A, 7(3):394, 1990

1990
[30]

Phase retrieval using alternating minimization.IEEE Transactions on Signal Processing, 63(18):4814–4826, 2015

Praneeth Netrapalli, Prateek Jain, and Sujay Sanghavi. Phase retrieval using alternating minimization.IEEE Transactions on Signal Processing, 63(18):4814–4826, 2015

2015
[31]

Phase retrieval via randomized kaczmarz: theoretical guarantees.arXiv:1706.09993, 2018

Yan Shuo Tan and Roman Vershynin. Phase retrieval via randomized kaczmarz: theoretical guarantees.arXiv:1706.09993, 2018

work page arXiv 2018
[32]

Phase recovery, maxcut and complex semidefinite programming.Mathematical Programming, 149(1–2):47–81, 2015

Irène Waldspurger, Alexandre D’aspremont, and Stéphane Mallat. Phase recovery, maxcut and complex semidefinite programming.Mathematical Programming, 149(1–2):47–81, 2015

2015
[33]

Giannakis, and Yonina C

Gang Wang, Georgios B. Giannakis, and Yonina C. Eldar. Solving systems of random quadratic equations via truncated amplitude flow.IEEE Transactions on Information The- ory, 64(2):773–794, 2017

2017
[34]

u∗ t q∗ t ∥qt∥2 # . andW t is a semi-definite positive matrix with rank at most2. Thus, Zt+1 =H 1(Wt) = h ut qt ∥qt∥2 i H1 ∥zt∥2 2 + αtθt 2 αt∥qt∥2 αt∥qt∥2 0

Ke Wei, Jian Feng Cai, Tony F. Chan, and Shingyu Leung. Guarantees of riemannian opti- mization for low rank matrix recovery.SIAM Journal on Matrix Analysis and Applications, 37(3):1198–1222, 2016. 37 8 Appendix 8.1 Useful Lemmas In this section, we provide some useful lemmas for supporting our results. Lemma 8.1(Lemma 7, [22]).For anyx,z∈C n obeyingRe(z ...

2016