Pith. sign in

REVIEW 2 minor 25 references

On the conditional equivalence of phase retrieval algorithms

T0 review · 0 major / 2 minor · reviewed 2026-06-27 · grok-4.3

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

Pith's one-line read The Gerchberg-Saxton magnitude replacement step is exactly a unit gradient descent update on the amplitude least-squares loss.

desk verdict The GS magnitude replacement is exactly unit GD on amplitude loss, a direct algebraic identity that bridges classical and differentiable phase retrieval. read the letter →

arxiv 2606.07257 v1 pith:Q3G3CLKA submitted 2026-06-05 physics.optics cs.LG

Jakob Schroeder , Andreas Döpp This is my paper · ORCID
classification physics.opticscs.LG
keywords phaseretrievalGerchberg-Saxtongradientdescentamplitudeleast-squaresdifferentiableimagingcomputationalopticsBayesianupdate
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 shows that the magnitude replacement operation in the Gerchberg-Saxton algorithm matches a single gradient descent step of size one when the objective is the amplitude least-squares loss. This identity is derived by direct comparison of the two update rules under the stated conditions. A reader cares because it removes the apparent separation between classical projection-based phase retrieval and modern gradient-based methods in computational imaging. The same identity also yields two probabilistic readings: the loss as negative log-likelihood under Gaussian amplitude noise, and each step as a Bayesian update with the propagated field as prior.

What carries the argument

The conditional mathematical identity between the Gerchberg-Saxton magnitude replacement operation and a unit-step gradient descent update on the amplitude least-squares loss.

What would settle it

Direct numerical evaluation of the gradient of the amplitude least-squares loss followed by a unit step, compared against the field produced by magnitude replacement under identical conditions, would show whether the two operations coincide.

Watch

Extended reading notes

Core claim

The GS magnitude replacement step is exactly a unit gradient descent step on an amplitude least-squares loss. This equivalence enables seamless integration of classical phase retrieval with differentiable physics pipelines. Two complementary probabilistic interpretations follow: globally, the amplitude loss is the negative log-likelihood under Gaussian amplitude noise; locally, each projection step arises as a Bayesian update with the propagated field as prior. The local view supplies qualitative guidance for relaxation in iterative phase retrieval.

Load-bearing premise

The identity holds only when the objective is exactly the amplitude least-squares loss and the descent step size is precisely one.

Editorial extensions

If this is right

  • Classical phase retrieval methods become native components of automatic-differentiation pipelines without custom implementation.
  • The amplitude least-squares loss functions as the canonical objective that unifies projection and gradient approaches.
  • Probabilistic readings allow relaxation parameters in iterative algorithms to be interpreted as prior strengths.
  • Hybrid algorithms can alternate or combine projection steps with gradient steps while preserving the same loss surface.

Reading between the lines

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

  • The equivalence suggests that any phase retrieval variant using a different loss could be rewritten as a modified gradient step, opening a route to new algorithms.
  • Implementation in machine-learning frameworks becomes direct: the magnitude replacement can be replaced by an autograd call with fixed step size.
  • The Bayesian-update view may extend to other noise models, yielding analogous projection steps for Poisson or other intensity statistics.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

0 major / 2 minor

Summary. The manuscript establishes that the magnitude replacement step in the Gerchberg-Saxton (GS) algorithm is mathematically identical to a single unit-step gradient descent update on the amplitude least-squares loss L = ½‖|z| − a‖². Direct Wirtinger differentiation yields ∇L = (|z| − a)(z/|z|), so the update z − ∇L algebraically simplifies to a·(z/|z|). The paper additionally supplies two probabilistic readings: the amplitude loss as negative log-likelihood under Gaussian amplitude noise, and each projection as a local Bayesian update with the propagated field as prior. These observations are presented as enabling integration of classical phase retrieval into differentiable physics pipelines.

Significance. The result supplies a parameter-free, direct algebraic bridge between a classical iterative method and modern gradient-based optimization, allowing GS steps to be dropped unchanged into autodiff frameworks. The probabilistic interpretations supply qualitative guidance for step-size relaxation without introducing new parameters or data fits. Because the identity follows immediately once the loss and step size are fixed, the contribution is self-contained and falsifiable by direct substitution.

minor comments (2)
  1. The abstract states the equivalence but does not explicitly name the Wirtinger gradient or the precise loss function; adding one sentence with the expression for ∇L would improve immediate readability for readers outside phase retrieval.
  2. Notation for the complex field z and measured amplitude a is introduced without a dedicated symbols table; a short table in §2 would eliminate any ambiguity when the same symbols appear in the probabilistic sections.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive review and recommendation to accept the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; direct algebraic identity

full rationale

The paper's core claim is that the Gerchberg-Saxton magnitude replacement equals a unit gradient-descent step on the amplitude least-squares loss L = ½‖|z| − a‖². This is established by explicit Wirtinger differentiation yielding ∇L = (|z| − a)(z/|z|), so that z − ∇L algebraically simplifies to a·(z/|z|). The identity is unconditional once the loss and step size are fixed; it does not invoke fitted parameters, self-citations as load-bearing premises, uniqueness theorems, or ansatzes smuggled from prior work. No step in the derivation chain reduces to its own inputs by construction. The result is therefore self-contained mathematical verification rather than circular re-labeling.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

No free parameters, invented entities, or ad-hoc axioms are introduced in the abstract; the claim rests on standard properties of gradient descent and projections.

assumptions (1)
  • standard math Standard properties of gradient descent and alternating projections in complex Hilbert space
    The identity is asserted to follow from these background facts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the conditional equivalence of phase retrieval algorithms." pith.science (2026). https://pith.science/paper/Q3G3CLKA

@misc{pith2026260607257,
  author       = {Pith},
  title        = {Pith review of: On the conditional equivalence of phase retrieval algorithms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q3G3CLKA}},
  note         = {Machine review of arXiv:2606.07257}
}
read the original abstract

Phase retrieval - recovering a complex-valued field from intensity measurements - is typically solved using variants of the Gerchberg-Saxton (GS) algorithm, understood as alternating projections between measurement planes. Meanwhile, modern computational imaging increasingly relies on gradient-based optimization and automatic differentiation. Here we show that these two approaches are mathematically identical: the GS magnitude replacement step is exactly a unit gradient descent step on an amplitude least-squares loss. This equivalence enables seamless integration of classical phase retrieval with differentiable physics pipelines. We further identify two complementary probabilistic interpretations of this equivalence: globally, the amplitude loss is the negative log-likelihood under Gaussian amplitude noise; locally, each projection step arises as a Bayesian update with the propagated field as prior. The local view provides qualitative guidance for relaxation in iterative phase retrieval.

Figures

Figures reproduced from arXiv: 2606.07257 by the authors.

Figure 1
Figure 1. FIG. 1. Amplitude loss versus iteration for GS (alternating [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Amplitude loss versus iteration for three different [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. FIG. 4. Phase reconstruction with an overcomplete dictionary [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 4 canonical work pages

  1. [1]

    Phase Retrieval with Application to Optical Imaging

    Y. Shechtman, Y. C. Eldar, O. Cohen, H. N. Chapman, J. Miao, and M. Segev, Phase Retrieval with Application to Optical Imaging (2014), arXiv:1402.7350 [cs]

    work page Pith review arXiv 2014

  2. [2]

    J. Miao, T. Ishikawa, I. K. Robinson, and M. M. Mur- nane, Beyond crystallography: Diffractive imaging using coherent x-ray light sources, Science348, 530 (2015)

    2015

  3. [3]

    Marinelli, M

    A. Marinelli, M. Dunning, S. Weathersby, E. Hemsing, D. Xiang, G. Andonian, F. O’Shea, J. Miao, C. Hast, 7 and J. B. Rosenzweig, Single-Shot Coherent Diffraction Imaging of Microbunched Relativistic Electron Beams for Free-Electron Laser Applications, Physical Review Let- ters110, 094802 (2013)

    2013

  4. [4]

    J. R. Fienup, Phase-retrieval algorithms for a compli- cated optical system, Applied Optics32, 1737 (1993)

    1993

  5. [5]

    Vorndran, J

    S. Vorndran, J. M. Russo, Y. Wu, S. A. Pelaez, and R. K. Kostuk, Broadband Gerchberg-Saxton algorithm for freeform diffractive spectral filter design, Optics Ex- press23, A1512 (2015)

    2015

  6. [6]

    Smartsev, A

    S. Smartsev, A. Liberman, I. A. Andriyash, A. Cav- agna, A. Flacco, C. Giaccaglia, J. Kaur, J. Monzac, S. Tata, A. Vernier, V. Malka, R. Lopez-Martens, and J. Faure, Simple few-shot method for spectrally resolv- ing the wavefront of an ultrashort laser pulse (2024), arXiv:2307.15799 [physics]

    work page arXiv 2024

  7. [7]

    Bakkali Taheri, I

    F. Bakkali Taheri, I. V. Konoplev, G. Doucas, P. Bad- doo, R. Bartolini, J. Cowley, and S. M. Hooker, Electron bunch profile reconstruction based on phase-constrained iterative algorithm, Physical Review Accelerators and Beams19, 032801 (2016)

    2016

  8. [8]

    Longitudinal Bunch Diagnostics using Coherent Transition Radiation Spectroscopy

    B. Schmidt, S. Wesch, T. K¨ ovener, C. Behrens, E. Hass, S. Casalbuoni, and P. Schm¨ user, Longitudinal Bunch Diagnostics using Coherent Transition Radiation Spec- troscopy (2018), arXiv:1803.00608 [physics]

    work page Pith review arXiv 2018

Show all 25 references
  1. [9]

    R. W. Gerchberg, Holography without Fringes in the Electron Microscope, Nature240, 404 (1972)

    1972

  2. [10]

    J. R. Fienup, Phase retrieval algorithms: A comparison, Applied Optics21, 2758 (1982)

    1982

  3. [11]

    D. R. Luke, Relaxed averaged alternating reflections for diffraction imaging, Inverse Problems21, 37 (2004)

    2004

  4. [12]

    Sidorenko, O

    P. Sidorenko, O. Kfir, Y. Shechtman, A. Fleischer, Y. C. Eldar, M. Segev, and O. Cohen, Sparsity-based super- resolved coherent diffraction imaging of one-dimensional objects, Nature Communications6, 8209 (2015)

    2015

  5. [13]

    Chang, Y

    H. Chang, Y. Lou, M. K. Ng, and T. Zeng, Phase Re- trieval from Incomplete Magnitude Information via To- tal Variation Regularization, SIAM Journal on Scientific Computing38, A3672 (2016)

    2016

  6. [14]

    Kandel, S

    S. Kandel, S. Maddali, M. Allain, S. O. Hruszkewycz, C. Jacobsen, and Y. S. G. Nashed, Using automatic dif- ferentiation as a general framework for ptychographic re- construction, Optics Express27, 18653 (2019)

    2019

  7. [15]

    A. Wong, B. Pope, L. Desdoigts, P. Tuthill, B. Norris, and C. Betters, Phase retrieval and design with auto- matic differentiation: tutorial, JOSA B38, 2465 (2021)

    2021

  8. [16]

    M. Du, S. Kandel, J. Deng, X. Huang, A. Demortiere, T. T. Nguyen, R. Tucoulou, V. D. Andrade, Q. Jin, and C. Jacobsen, Adorym: a multi-platform generic X-ray image reconstruction framework based on automatic dif- ferentiation, Optics Express29, 10000 (2021)

    2021

  9. [17]

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams, Learning representations by back-propagating errors, na- ture323, 533 (1986)

    1986

  10. [18]

    Levi and H

    A. Levi and H. Stark, Image restoration by the method of generalized projections with application to restoration from magnitude, J. Opt. Soc. Am. A1, 932 (1984)

    1984

  11. [19]

    H. H. Bauschke, P. L. Combettes, and D. R. Luke, Phase retrieval, error reduction algorithm, and Fienup variants: A view from convex optimization, Journal of the Optical Society of America A19, 1334 (2002)

    2002

  12. [20]

    E. J. Cand` es, X. Li, and M. Soltanolkotabi, Phase Re- trieval via Wirtinger Flow: Theory and Algorithms, IEEE Transactions on Information Theory61, 1985 (2015)

    1985

  13. [21]

    Z. Wei, W. Chen, C.-W. Qiu, and X. Chen, Conju- gate gradient method for phase retrieval based on the Wirtinger derivative, Journal of the Optical Society of America A34, 708 (2017)

    2017

  14. [22]

    R. M. Neal and G. E. Hinton, A View of the Em Algo- rithm that Justifies Incremental, Sparse, and other Vari- ants, inLearning in Graphical Models, edited by M. I. Jordan (Springer Netherlands, Dordrecht, 1998) pp. 355– 368

    1998

  15. [23]

    Howard, N

    S. Howard, N. Weisse, J. Schrder, C. Barbero, B. Alonso, . Sola, P. Norreys, and A. Dpp, Sparse reconstruction of wavefronts using an over-complete phase dictionary, Op- tics Express33, 12939 (2025), publisher: Optica Pub- lishing Group

    2025

  16. [24]

    D. P. Kingma and J. Ba, Adam: A Method for Stochastic Optimization (2017), arXiv:1412.6980 [cs]

    2017 arXiv

  17. [25]

    Moulanier, L

    I. Moulanier, L. T. Dickson, F. Massimo, G. Maynard, and B. Cros, Fast laser field reconstruction method based on a GerchbergSaxton algorithm with mode decomposi- tion, Journal of the Optical Society of America B40, 2450 (2023)

    2023

Pith tools

Reviewed June 27, 2026 · model on record in the stance chip above.