Pith. sign in

REVIEW 3 major objections 5 minor 117 references

Optimal spectral initializers impact on phase retrieval phase transitions -- an RDT view

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Spectral initializers miss the phase-retrieval transition point

desk verdict The optimal spectral overlap formula is old, but the RDT derivation and the flat-region/oversampling observation are worth a look; the practical claim is heuristic and unquantified. read the letter →

arxiv 2506.18279 v1 pith:I444DO2A submitted 2025-06-23 stat.ML cs.ITcs.LGmath.IT

Mihailo Stojnic This is my paper · ORCID
classification stat.MLcs.ITcs.LGmath.IT MSC 60B2062H1290C26
keywords phaseretrievalspectralinitializersrandomdualitytheorytransitionsnon-convexoptimizationparametricmanifoldoverlap-optimalinitializationGaussianmeasurements
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 aims to determine, in the high-dimensional linear regime, whether the best possible spectral initializer can place a descending phase-retrieval algorithm (dPR) in a region from which it converges to the true signal. Using a Random Duality Theory (RDT) program, it derives the overlap-optimal spectral initializer (OptSpin) and shows its starting overlap is $\hat{x}^{({\rm spec})}=\sqrt{1-1/\hat{\gamma}}$, with $\hat{\gamma}$ the solution of a closed-form equation. Placing these overlaps on the parametric manifold ${\mathcal{PM}}(\alpha)$ of the companion paper yields two observations: at the theoretical transition $\alpha=1.4$ the optimal start has overlap about $0.6439$ and lies inside the manifold's flat region, so practical success is fragile; at the safer oversampling $\alpha=1.6$ the overlap is about $0.7055$, lies outside the flat region, and dPR solves phase retrieval. The paper therefore recommends running dPR roughly 10--20% above the theoretical sample-complexity ratio, and reports numerical simulations that match this recommendation.

What carries the argument

The load-bearing objects are the f-spin optimization and the parametric manifold ${\mathcal{PM}}(\alpha)$. The f-spin is the random-primal formulation of the spectral initializer problem, $\xi_s(x)=\max_{\|x\|_2=1,\,Ax=z,\,x^T\bar{x}=x} z^T{\rm diag}(T(A_{:,1}))z$, and a Gaussian comparison argument turns it into a random dual governed by two scalar stationarity equations; solving those equations yields the optimal preprocessing $T(y)=(y-1)/y$ and the overlap formula. On the algorithm side, ${\mathcal{PM}}(\alpha)$ is the curve of optimal objective values versus overlap at fixed norm $c=1$ from the companion paper; superimposing the starting overlap on this curve determines whether the start lies in a flat region, and flat regions are the zones where jitteriness creates traps.

What would settle it

Running the overlap-optimal spectral initializer followed by a descending algorithm at $\alpha=1.4$ with large $n$ (for example, $n=10^4$) and checking whether success probability approaches 1 as $n$ grows would settle the flat-region risk model. A second check is quantitative: measure the empirical distribution of starting overlaps around $0.6439$ and locate the boundary of the flat region on the lifted ${\mathcal{PM}}$ curve, so one can see directly whether the start lies inside it.

Watch

Extended reading notes

Core claim

The central claim is that the theoretically best spectral initializer has a precise, parameter-free overlap curve, and that this curve decides whether a descending phase-retrieval algorithm can practically reach the global optimum. For Gaussian measurements and real signals, the overlap-optimal preprocessing is $T(y)=(y-1)/y$ up to scaling, the optimal starting overlap is $\hat{x}^{({\rm spec})}=\sqrt{1-1/\hat{\gamma}}$ with $\hat{\gamma}$ solving equation (62), and the two numerical anchors are $\hat{x}^{({\rm spec})}\approx0.6439$ at $\alpha=1.4$ and $\hat{x}^{({\rm spec})}\approx0.7055$ at $\alpha=1.6$. The paper argues that the first value falls inside the flat region of ${\mathcal{PM}}(1.4)$, where finite-dimensional jitteriness can trap a descending path, while the second falls outside the flat region of ${\mathcal{PM}}(1.6)$, so the path funnels to the true signal. The practical conclusion is that the reliable threshold for spectral-initialized dPR sits strictly above the asymptotic lifted-RDT transition at $\alpha\approx1.4$.

Load-bearing premise

The central recommendation rests on the companion paper's claim that a descending algorithm succeeds exactly when the parametric manifold has a single funneling point, and that flat regions become dangerous under finite-dimensional jitteriness; if that geometric picture is wrong, the conclusion that $\alpha=1.4$ is fragile fails even though the overlap formula remains correct.

Editorial extensions

If this is right

  • At the lifted-RDT transition $\alpha=1.4$, even the best spectral start lands in a flat region, so the theoretical transition is not the operating point one should use in practice.
  • Increasing $\alpha$ by about 15% to $\alpha=1.6$ moves the optimal starting overlap from roughly $0.6439$ to $0.7055$ and outside the flat region; dPR then solves phase retrieval.
  • The optimal spectral preprocessing for Gaussian measurements takes the explicit form $T(y)=(y-1)/y$, and the RDT derivation reproduces the overlap curve previously obtained by spectral and free-probability methods.
  • Because the RDT program does not rely on eigenvalue machinery, the same derivation extends to structured signals such as sparse, block-sparse, positive, binary, box-constrained, or partially observed signals.

Reading between the lines

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

  • If flat-region jitteriness is the operative failure mode, the practical threshold should drift with dimension: larger $n$ should shrink the jitter and move simulated transitions closer to $\alpha=1.4$, a testable prediction the paper does not make explicitly.
  • The formula $\hat{x}^{({\rm spec})}=\sqrt{1-1/\hat{\gamma}}$ could serve as a cheap diagnostic for other non-convex inverse problems: whenever an initializer's overlap crosses into a flat region of the associated manifold, the same 10-20% oversampling remedy should apply.
  • A quantitative version of the flat-region risk could be obtained by measuring the local curvature or Lipschitz constant of the lifted curve near the initializer overlap; the paper identifies flat regions visually, so an explicit curvature threshold would turn the rule of thumb into a checkable criterion.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript studies the overlap between optimal spectral initializers (OptSpins) and the true signal in real phase retrieval with Gaussian measurements, in the proportional high-dimensional regime alpha = m/n. The authors derive the OptSpin preprocessing function and the resulting starting overlap via Random Duality Theory (RDT), obtaining the closed-form expression \xhat = sqrt(1 - 1/\gamma) with \gamma solving Eq. (62), and they benchmark this formula against the known spectral-method results of [64,78]. The main interpretive claims are about phase transitions of descending phase retrieval algorithms (dPR): at the theoretical phase transition alpha ≈ 1.4, the OptSpin overlap (about 0.6439) falls inside what the companion paper [104] calls a 'flat region' of the parametric manifold PM(α), making practical dPR success fragile; at alpha = 1.6, a roughly 15% oversampled 'safer compression' point, the overlap (about 0.7055) falls outside the flat region, so dPR should succeed. Numerical simulations with n = 300 are reported in support of the alpha = 1.6 recommendation.

Significance. If the overlap formula is taken as the paper's core contribution, the result is solid and valuable: the RDT derivation reproduces the optimal spectral initialization overlaps of [64,78] through a different and potentially more extensible route, and the derivation is internally consistent. The practical phase-transition conclusion, however, is the main advertised message and it is currently a heuristic built on an unquantified 'flat region' concept imported from the same-author companion paper [104]. The numerical evidence does not directly validate the claim because it uses a different objective and reports no trial counts or error bars. The work would be a useful methodological addition if the flat-region criterion were made precise and the simulations were matched to the analyzed objective; in its present form the central practical claim is not yet load-bearing evidence.

major comments (3)
  1. [Section 4, Theorem 2 and Figures 2–3] The 'flat region' is never defined quantitatively, so the central classification of the OptSpin overlap x = 0.6439 as inside the flat region at alpha = 1.4 and x = 0.7055 as outside it at alpha = 1.6 is a visual judgment on plotted curves rather than a measurable criterion. Please provide a formal definition (for example, a threshold on |d√\barφ₀/dx| over an interval) and report the computed flat-region boundaries for both alpha values, together with a sensitivity check with respect to the chosen threshold.
  2. [Section 5, Eq. (70)–(71) and Figure 4] The simulations minimize the squared-magnitude barrier objective fbar(t0;x) = t0‖|A\x|² − |Ax|²‖² + log(1 − ‖x‖²), while the flat-region curves in Figures 2–3 are computed for the non-squared objective (5). Section 5.1 itself concedes that for the squared objective at alpha = 1.4 it is 'even difficult to say ... alpha = 1.4 is indeed the phase transition,' so the claimed agreement of Figure 4 with the alpha = 1.6 prediction is not a quantitative prediction for the simulated objective. Figure 4 also reports no trial counts or error bars. Please either compute flat-region boundaries for the squared objective or simulate the non-squared objective, and in either case report the number of Monte Carlo trials and the success criterion.
  3. [Section 4, Theorem 2] The identification of alpha ≈ 1.4 as the dPR theoretical phase transition, the single-funneling-point condition, and the entire parametric-manifold geometry are imported from the companion paper [104]; the proof of Theorem 2 is only a reference to [104]. Since the paper's practical conclusions depend on that imported geometry, the manuscript should state this dependency explicitly and specify which assertions of [104] are assumed. As it stands, an error or revision in [104] would invalidate the central recommendation even though the overlap formula of Section 3 remains correct.
minor comments (5)
  1. [Title and Abstract] The title contains a typo: 'ph ase' should be 'phase'.
  2. [Section 1, paragraph 1 and Section 1.2] The phrase 'History of these applications is rather reach' should read 'rather rich', and 'global convergence theoretical phase transitions (predicated by the lifted RDT)' should read 'predicted by' rather than 'predicated by'.
  3. [Section 3.1, Eqs. (54)–(62)] The text states that 'strong random duality is in place' immediately after benchmarking Eq. (61)–(62) against [64,78], but no internal argument for strong duality is given; please either supply a proof or state explicitly that the upper bound is inherited from the external optimality result.
  4. [Section 5, Figure 4] Please provide the number of random instances, the noise-free success criterion, and the interpolation method used for the curves in Figure 4; without these details the reported 'excellent agreement' cannot be assessed.
  5. [References] Reference [104] is listed only as '2025, available online at arxiv' without an arXiv identifier or a verifiable URL; this is problematic because Theorem 2 and the flat-region concept are load-bearing for the paper's main claims.

Circularity Check

2 steps flagged · score 4.0 of 10

The OptSpin overlap formula is independently benchmarked, but the flat-region and dPR phase-transition conclusions are imported from same-author companion [104] and remain unquantified.

  1. uniqueness imported from authors [Section 4, paragraph before Figure 2 (Theorem 2 discussion)]
    "In particular, any descending algorithm will converge to the global optimum and therefore solve phase retrieval if the so-called parametric manifold – PM(α) – associated with ξ(c, x) has a single funneling point."

    The single-funnel criterion is not derived in this paper; Theorem 2's proof states it is an 'immediate consequence' of the same author's [104]. The paper then fixes α=1.4 as the theoretical dPR phase transition because [104]'s lifted RDT curve has a single funneling point, and uses that imported threshold to argue that the OptSpin overlap x=0.6439 falls in a dangerous flat region. The central practical conclusion is therefore inherited from a same-author companion result rather than established by the present derivation.

  2. ansatz smuggled in via citation [Section 4, after Figure 2]
    "as pointed out in [104], in practice things are a bit different. Namely, theoretical values rely on strong concentrations which indeed happen in the assumed high-dimensional n → ∞ regime. On the other hand, in practical algorithmic running the values for n are finite and the concentrations will not always be perfect. This basically means that instead of being smooth the shown curve will in practice be slightly jittery."

    The flat-region/local-jitteriness failure model is adopted from companion [104] without any definition of 'flat region' or a quantitative criterion for when a starting overlap is inside versus outside such a region. The paper's key classifications—x=0.6439 as inside and x=0.7055 as outside—are visual judgments on Figures 2 and 3, and the 10-20% 'safer compression' recommendation is exactly this imported ansatz applied to the new overlap values. No equation in the present paper computes a flat-region boundary, so the recommendation cannot be independently checked from the material supplied here.

full rationale

The overlap-optimal spectral initializer derivation in Section 3.1 is not circular: it is a self-contained RDT calculation, and the paper explicitly cross-checks the result against the independent external works [64,78] ('It is not that difficult to check that T̂(·) from (60) precisely matches the optimal choice suggested in [78] and proven for any α in [64]'). That part earns no circularity penalty. The circularity burden is concentrated in the phase-transition interpretation: the PM(α) geometry, the α≈1.4 lifted-RDT transition, and the flat-region/jitteriness risk model are all imported from the same-author companion [104], with Theorem 2's proof delegated to [104] and the flat-region boundary never quantified. Section 5.1 itself concedes that for the squared-magnitude objective the lifted curve is so flat that 'it is even difficult to say that ... α=1.4 is indeed the phase transition', and the numerical support is transferred from non-squared theory by analogy rather than by a quantitative prediction for the simulated squared-magnitude objective. These are load-bearing limitations for the paper's main recommendation, though they do not reduce the overlap formula itself to its inputs. Overall circularity score 4: some self-citation and imported ansatz, with the central overlap result retaining independent content.

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

No data-fitted parameters appear in the derivation; the optimal overlap formula is computed from Gaussian assumptions and matches [64,78]. The main imported assumptions are the companion paper's parametric-manifold geometry and the flat-region jitteriness heuristic. No new physical or algorithmic entities are postulated.

assumptions (6)
  • domain assumption Real, iid standard Gaussian measurement matrix A and unit-norm true signal in the proportional regime alpha = m/n.
    Assumed in Sections 1 and 2; all asymptotic statements rely on this model.
  • standard math Gordon's Gaussian comparison theorem is applicable to the random primal and dual pair.
    Theorem 1's proof invokes Gordon's theorem and Stojnic's generalizations in [99,100].
  • domain assumption Companion paper [104] correctly characterizes PM(alpha) and dPR phase transitions, including the c = 1 single-funneling-point threshold at alpha near 1.4.
    Theorems 2 and 3 in Section 4 are imported from [104] with proof deferred; this is the foundation of the flat-region discussion.
  • ad hoc to paper Flat regions of PM(alpha) cause practical dPR failures because of finite-dimensional jitteriness.
    Asserted in Section 4 around Figures 2 and 3; no mathematical model or quantitative flatness measure is provided.
  • domain assumption The behavior of nonsquared-magnitude objectives is representative of squared-magnitude objectives.
    The simulations use squared magnitudes while the theory uses nonsquared; Section 5 argues qualitative agreement via [104] rather than proof.
  • domain assumption Optimality of the derived preprocessing T over all componentwise preprocessing is established externally in [64], and at the weak threshold in [78].
    Section 3.1 step 4 uses matching to [64] as the strong-duality check; the optimality proof is not reproduced here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimal spectral initializers impact on phase retrieval phase transitions -- an RDT view." pith.science (2026). https://pith.science/paper/I444DO2A

@misc{pith2026250618279,
  author       = {Pith},
  title        = {Pith review of: Optimal spectral initializers impact on phase retrieval phase transitions -- an RDT view},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I444DO2A}},
  note         = {Machine review of arXiv:2506.18279}
}
abstract

We analyze the relation between spectral initializers and theoretical limits of \emph{descending} phase retrieval algorithms (dPR). In companion paper [104], for any sample complexity ratio, $\alpha$, \emph{parametric manifold}, ${\mathcal {PM}}(\alpha)$, is recognized as a critically important structure that generically determines dPRs abilities to solve phase retrieval (PR). Moreover, overlap between the algorithmic solution and the true signal is positioned as a key ${\mathcal {PM}}$'s component. We here consider the so-called \emph{overlap optimal} spectral initializers (OptSpins) as dPR's starting points and develop a generic \emph{Random duality theory} (RDT) based program to statistically characterize them. In particular, we determine the functional structure of OptSpins and evaluate the starting overlaps that they provide for the dPRs. Since ${\mathcal {PM}}$'s so-called \emph{flat regions} are highly susceptible to \emph{local jitteriness} and as such are key obstacles on dPR's path towards PR's global optimum, a precise characterization of the starting overlap allows to determine if such regions can be successfully circumvented. Through the presented theoretical analysis we observe two key points in that regard: \textbf{\emph{(i)}} dPR's theoretical phase transition (critical $\alpha$ above which they solve PR) might be difficult to practically achieve as the ${\mathcal {PM}}$'s flat regions are large causing the associated OptSpins to fall exactly within them; and \textbf{\emph{(ii)}} Opting for so-called ``\emph{safer compression}'' and slightly increasing $\alpha$ (by say $15\%$) shrinks flat regions and allows OptSpins to fall outside them and dPRs to ultimately solve PR. Numerical simulations are conducted as well and shown to be in an excellent agreement with theoretical predictions.

Figures

Figures reproduced from arXiv: 2506.18279 by the authors.

Figure 1
Figure 1. In the following section we discuss how the optimal [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗
Figure 2
Figure 2. φ0 and φ¯ 0 as functions of x; c = 1 and α = 1.4 While the flatness of the manifold and consequential jitteriness are the key factors impacting the practical reaching of the theoretical phase transitions, there are two other things that may have a prominent effect as well. First, the above assumes that the ℓ2 norm of the unknown x is limited while the algorithm is being run. That typically requires utilizing algorit… view at source ↗
Figure 3
Figure 3. φ0 and φ¯ 0 as functions of x; c = 1 and α = 1.6 than say the standard plain gradient. If one is to use just a plain gradient then the above analysis needs to extend to c > 1 scenarios. As [104]’s Section 4 demonstrates, the presented theory then implies that in such scenarios there is no guaranteed sample complexity ratio beyond which the descending algorithms generically converge to the global optimum. In practice… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Simulated and theoretical RDT and lifted RDT phase [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 5
Figure 5. Figure 5: φ (sq) 0 and φ¯ (sq) 0 as functions of x; c = 1 and α = 1.4 and reevaluate the corresponding plain and lifted RDT curves for various α < 1.4 and then determine the critical one for which the corresponding spectral initializers fall within the zone of convergence toward…
Figure 6
Figure 6. Figure 6: φ (sq) 0 and φ¯ (sq) 0 as functions of x; c = 1 and α = 1.6 to practical relevance in a clearer way. We implemented in parallel plain gradient descent and a hybrid combination of plain and log barrier gradient descent algorithms. Even though the simulations were run fo…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

117 extracted references · 19 canonical work pages

  1. [104]

    M. Stojnic. Phase transition of descending phase retrieval algorithms. 2025. available online at arxi v

    work page 2025

  2. [1]

    Achlioptas, A

    D. Achlioptas, A. Coja-Oghlan, and F. Ricci-Tersenghi. On the solution-space geometry of random constraint satisfaction problems. Random Struct. Algorithms , 38(3):251–268, 2011

    2011

  3. [2]

    Ahmed, B

    A. Ahmed, B. Recht, and J. Romberg. Blind deconvolution u sing convex programming. IEEE Trans- actions on Information Theory , 60(3):1711–1732, 2013

    2013

  4. [3]

    Aubin, B

    B. Aubin, B. Loureiro, A. Baker, F. Krzakala, and L. Zdebo rová. Exact asymptotics for phase retrieval and compressed sensing with random generative priors. In Proceedings of Mathematical and Scientific Machine Learning, MSML 2020, 20-24 July 2020, Virtual Confere nce / Princeton, NJ, USA , volume 107 of Proceedings of Machine Learning Research , pages 55–...

    arXiv 2020

  5. [4]

    Bahmani and J

    S. Bahmani and J. Romberg. Phase retrieval meets statist ical learning theory: A flexible convex relaxation. Electronic Journal of Statistics , 11:5254–5281, 2016. 23

    2016

  6. [5]

    Balan, B

    R. Balan, B. Bodmann, P. Casazza, and D. Edidin. Painless reconstruction from magnitudes of frame coefficients. J. Four. Anal. Appl. , 15:488–501, 2009

    2009

  7. [6]

    Balan, P

    R. Balan, P. Casazza, and D. Edidin. On signal reconstruc tion without phase. Applied and Computa- tional Harmonic Analysis , 20(3):343–356, 2006

    2006

  8. [7]

    Baldassi, A

    C. Baldassi, A. Ingrosso, C. Lucibello, L. Saglietti, an d R. Zecchina. Subdominant dense clusters allow for simple learning and high computational performance in n eural networks with discrete synapses. Physical Review letters , 115(12):128101, 2015

    2015

Show all 117 references
  1. [8]

    Baldassi, A

    C. Baldassi, A. Ingrosso, C. Lucibello, L. Saglietti, an d R. Zecchina. Local entropy as a measure for sampling solutions in constraint satisfaction problems. Journal of Statistical Mechanics: Theory and Experiment, (2):021301, 2016

    2016

  2. [9]

    Baldassi, R

    C. Baldassi, R. D. Vecchia, C. Lucibello, and R. Zecchina . Clustering of solutions in the symmetric binary perceptron. Journal of Statistical Mechanics: Theory and Experiment , (7):073303, 2020

    2020

  3. [10]

    A. S. Bandeira, J. Cahill, D. G. Mixon, and A. A. Nelson. S aving phase: Injectivity and stability for phase retrieval. Applied and Computational Harmonic Analysis , 37(1):106–125, 2014

    2014

  4. [11]

    Barbier, F

    J. Barbier, F. Krzakala, N. Macris, L. Miolane, and L. Zd eborová. Optimal errors and phase transi- tions in high-dimensional generalized linear models. In Conference On Learning Theory, COLT 2018, Stockholm, Sweden, 6-9 July 2018 , volume 75 of Proceedings of Machine Learning...

    2018 arXiv

  5. [12]

    A. Bora, A. Jalal, E. Price, and A. G. Dimakis. Compresse d sensing using generative models. In Proceedings of the 34th International Conference on Machine L earning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017 , volume 70 of Proceedings of Machine Learning Research , ...

    2017

  6. [13]

    Bruck and L.G

    Y.M. Bruck and L.G. Sodin. On the ambiguity of the image r econstruction problem. Opt. Comm. , 30:304–308, 1979

    1979

  7. [14]

    O. Bunk, A. Diaz, F. Pfefer, C. David, B. Schmitt, D. K. Sa tapathy, and J. F. Veen. Diffractive imaging for periodic samples: retrieving one-dimensional concent ration profies across microuidic channels. Acta Crystallographica Section A: Foundations of Crystallograph y, 63(4):3...

    2007

  8. [15]

    K. Wang C. Ma, Y. Chi, and Y. Chen. Implicit regularizati on in nonconvex statistical estimation: Gradient descent converges linearly for phase retrieval, m atrix completion, and blind deconvolution. Foundations of Computational Mathematics , pages 1–182, 2018

    2018

  9. [16]

    T T. Cai, X. Li, and Z. Ma. Optimal rates of convergence fo r noisy sparse phase retrieval via thresholded Wirtinger flow. The Annals of Statistics , 44(5):2221–2251, 2016

    2016

  10. [17]

    E. J. Candès, Y. C. Eldar, T. Strohmer, and V. Voroninski . Phase retrieval via matrix completion. SIAM J. Imaging Sci. , 6(1):199–225, 2013

    2013

  11. [18]

    E. J. Candès and X. Li. Solving quadratic equations via p haselift when there are about as many equations as unknowns. Found. Comput. Math. , 14(5):1017–1026, 2014

    2014

  12. [19]

    E. J. Candès, X. Li, and M. Soltanolkotabi. Phase retrie val from coded diffraction patterns. Applied and Computational Harmonic Analysis , 39(2):277–299, 2015

    2015

  13. [20]

    E. J. Candès, X. Li, and M. Soltanolkotabi. Phase retrie val via Wirtinger flow: Theory and algorithms. IEEE Trans. Inf. Theory , 61(4):1985–2007, 2015

    1985

  14. [21]

    E. J. Candès, T. Strohmer, and V. Voroninski. Phaselift : Exact and stable signal recovery from magnitude measurements via convex programming. Comm. Pure Appl. Math. , 66:1241–1274, 2013. 24

    2013

  15. [22]

    Chen and E

    Y. Chen and E. J. Candès. Solving random quadratic syste ms of equations is nearly as easy as solving linear systems. Comm. Pure Appl. Math. , 70(5):822–883, 2017

    2017

  16. [23]

    Conca, D

    A. Conca, D. Edidin, M. Hering, and C. Vinzant. An algebr aic characterization of injectivity in phase retrieval. Applied and Computational Harmonic Analysis , 38(2):346–245, 2015

    2015

  17. [24]

    J. V. Corbett. The Pauli problem, state reconstruction and quantum-real numbers. Rep. Math. Phys. , 57:53–68, 2006

    2006

  18. [25]

    J. C. Dainty and J. R. Fienup. Image recovery: Theory and application. Phase retrieval and image reconstruction for astronomy, 21:231–275, Aug 1987

    1987

  19. [26]

    Daskalakis, D

    C. Daskalakis, D. Rohatgi, and E. Zampetakis. Constant -expansion suffices for compressed sensing with generative priors. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, De cember 6-12, 2020,...

    2020

  20. [27]

    Daude, M

    H. Daude, M. Mezard, T. Mora, and R. Zecchina. Pairs of sa t-assignments in random boolean formulae. Theoretical Computer Science , 393(1):260–279, 2008

    2008

  21. [28]

    Dhifallah, C

    O. Dhifallah, C. Thrampoulidis, and Y. M. Lu. Phase retr ieval via linear programming: Fundamental limits and algorithmic improvements. In 55th Annual Allerton Conference on Communication, Control, and Computing, Allerton 2017, Monticello, IL, USA, October 3- 6, 2017, pages 10...

    2017

  22. [29]

    Dierolf, A

    M. Dierolf, A. Menzel, P. Thibault, P. Schneider, C. M. K ewish, A. Wepf, O. Bunk, and F. Pfeiffer. Ptychographic x-ray computed tomography at the nanoscale. Nature, 477(7314):436–439, 2010

    2010

  23. [30]

    Donoho, A

    D. Donoho, A. Maleki, and A. Montanari. Message-passin g algorithms for compressed sensing. Proc. National Academy of Sciences , 106(45):18914–18919, Nov. 2009

    2009

  24. [31]

    Duadi, O

    H. Duadi, O. Margalit, V. Mico, J. A. Rodrigo, T. Alieva, J. Garcia, and Z. Zalevsky. Digital holography and phase retrieval. In J. Rosen, editor, Source: Holography, Research and Technolo gies. InTech, 2011

    2011

  25. [32]

    Dudeja, M

    R. Dudeja, M. Bakhshizadeh, J. Ma, and A. Maleki. Analys is of spectral methods for phase retrieval with random orthogonal matrices. IEEE Trans. Inf. Theory , 66(8):5182–5203, 2020

    2020

  26. [33]

    Fannjiang and Z

    A. Fannjiang and Z. Zhang. Fixed point analysis of Dougl as-Rachford splitting for ptychography and phase retrieval. SIAM J. Imaging Sci. , 13(2):609–650, 2020

    2020

  27. [34]

    J. R. Fienup. Reconstruction of an object from the modul us of its Fourier transform. Optics letters , 3(1):27–29, Aug 1978

    1978

  28. [35]

    J. R. Fienup. Phase retrieval algorithms: a comparison . Appl. Opt. , 21(15):2758–2769, Aug 1982

    1982

  29. [36]

    D. Gabor. A new microscopic principle. Nature, 161:777778, 1948

    1948

  30. [37]

    Gabor, G

    D. Gabor, G. W. Stroke, D. Brumm, A. Funkhouser, and A. La beyrie. Reconstruction of phase objects by holography. Applied and Computational Harmonic Analysis , 208(516):1159–1162, 1965

    1965

  31. [38]

    Gamarnik

    D. Gamarnik. The overlap gap property: A topological ba rrier to optimizing over random structures. Proceedings of the National Academy of Sciences , 118(41), 2021

    2021

  32. [39]

    Gamarnik and M

    D. Gamarnik and M. Sudan. Limits of local algorithms ove r sparse random graphs. Proceedings of the 5th conference on innovations in theoretical computer scie nce, pages 369–376, 2014

    2014

  33. [40]

    Gamarnik and M

    D. Gamarnik and M. Sudan. Limits of local algorithms ove r sparse random graphs. Ann. Probab., 45(4):2353–2376, 2017

    2017

  34. [41]

    Gamarnik and M

    D. Gamarnik and M. Sudan. Performance of sequential loc al algorithms for the random NAE-K-SAT problem. SIAM Journal on Computing , 46(2):590–619, 2017

    2017

  35. [42]

    R. A. Gerchberg and W. O. Saxton. A practical algorithm f or the determination of phase from image and diffraction plane pictures. Optik, 35:237–246, 1972. 25

    1972

  36. [43]

    Goldstein and C

    T. Goldstein and C. Studer. PhaseMax: Convex phase retr ieval via basis pursuit. IEEE Trans. Inf. Theory, 64(4):2675–2689, 2018

    2018

  37. [44]

    Y. Gordon. On Milman’s inequality and random subspaces which escape through a mesh in Rn. Geometric Aspect of of functional analysis, Isr. Semin. 1986-8 7, Lect. Notes Math , 1317, 1988

    1986

  38. [45]

    Gross, F

    D. Gross, F. Krahmer, and R. Kueng. Improved recovery gu arantees for phase retrieval from coded diffraction patterns. Applied and Computational Harmonic Analysis , 42(1):37–64, 2017

    2017

  39. [46]

    J. Haah, A. W. Harrow, Z. Ji, X. Wu, and N. Yu. Sample-opti mal tomography of quantum states. IEEE Trans. Inf. Theory , 63(9):5628–5641, 2017

    2017

  40. [47]

    P. Hand. Phaselift is robust to a constant fraction of ar bitrary errors. Applied and Computational Harmonic Analysis , 42(3):550–362, 2017

    2017

  41. [48]

    P. Hand, O. Leong, and V. Voroninski. Phase retrieval un der a generative prior. In Advances in Neural Information Processing Systems 31: Annual Conference on Neur al Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada , pages 9154–9164, 2018

    2018

  42. [49]

    Hand and V

    P. Hand and V. Voroninski. An elementary proof of convex phase retrieval in the natural parameter space via the linear program phaseMax. 2016. available onli ne at http://arxiv.org/abs/1611. 03935

    2016

  43. [50]

    R. W. Harrison. Phase problem in crystallography. J. Opt. Soc. Am. A , 10(5):1046–1055, May 1993

    1993

  44. [51]

    Heinosaari, L

    T. Heinosaari, L. Mazzarella, and M. M. Wolf. Quantum to mography under prior information. Com- munications in Mathematical Physics , 318(2):355–374, 2013

    2013

  45. [52]

    N. Hurt. Phase Retrieval and Zero Crossings . Kluwer Academic Publishers, Norwell, MA, 1989

    1989

  46. [53]

    M. Iwen, A. Viswanathan, and Y. Wang. Robust sparse phas e retrieval made easy. Applied and Computational Harmonic Analysis , 42(1):135–142, 2017

    2017

  47. [54]

    Jaganathan, S

    K. Jaganathan, S. Oymak, and B. Hassibi. Sparse phase re trieval: Uniqueness guarantees and recovery algorithms. IEEE Trans. Signal Process. , 65(9):2402–2410, 2017

    2017

  48. [55]

    Jordan and A

    M. Jordan and A. G. Dimakis. Exactly computing the local lipschitz constant of relu networks. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, v irtual, 2020

    2020

  49. [56]

    P. Jung, F. Krahmer, and D. Stroger. Blind demixing and d econvolution at near-optimal rate. IEEE Transactions on Information Theory , 44(2):704–727, 2017

    2017

  50. [57]

    M. V. Klibanov, P. E. Sacks, and A. V. Tikhonravov. The ph ase retrieval problem. Inverse Problems, 11(1):1, feb 1995

    1995

  51. [58]

    Kueng, H

    R. Kueng, H. Rauhut, and U. Terstiege. Low rank matrix re covery from rank one measurements. Applied and Computational Harmonic Analysis , 42(1):88–116, 2017

    2017

  52. [59]

    Q. Lei, A. Jalal, I. S. Dhillon, and A. G. Dimakis. Invert ing deep generative models, one layer at a time. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8- 14, 2019, Vancouver, ...

    2019

  53. [60]

    K.-C. Li. On principal hessian directions for data visu alization and dimension reduction: Another application of Stein’s lemma. J. Am. Stat. Assoc. , 87(420):1025–1039, 1992

    1992

  54. [61]

    X. Li, S. Ling, T. Strohmer, and K. Wei. Rapid, robust, an d reliable blind deconvolution via nonconvex optimization. Applied and computational harmonic analysis , 47(3):893–934, 2017. 26

    2017

  55. [62]

    Li and V

    X. Li and V. Voroninski. Sparse signal recovery from qua dratic measurements via convex programming. SIAM Journal on Mathematical Analysis , 45(5):3019–3033, 2013

    2013

  56. [63]

    Y. M. Lu and G. Li. Spectral initialization for nonconve x estimation: High-dimensional limit and phase transitions. In 2017 IEEE International Symposium on Information Theory, ISIT 2017, Aachen, Germany, June 25-30, 2017 , pages 3015–3019. IEEE, 2017

    2017

  57. [64]

    W. Luo, W. Alghamdi, and Y. M. Lu. Optimal spectral initi alization for signal recovery with applica- tions to phase retrieval. IEEE Trans. Signal Process. , 67(9):2347–2356, 2019

    2019

  58. [65]

    W. Luo, W. Alghamdi, Y. M. Lu, and G. Li. Phase transition s of spectral initialization for high- dimensional non-convex estimation. Information and Inference: A Journal of the IMA , 9(3):5077–541, 2020

    2020

  59. [66]

    J. Ma, R. Dudeja, J. Xu, A. Maleki, and X. Wang. Spectral m ethod for phase retrieval: An expectation propagation perspective. IEEE Trans. Inf. Theory , 67(2):1332–1355, 2021

    2021

  60. [67]

    Maillard, A

    A. Maillard, A. S. Bandeira, D. Belius, I. Dokmanic, and S. Nakajima. Injectivity of relu networks: perspectives from statistical physics. 2023. available on line at http://arxiv.org/abs/2302.14112

    2023 arXiv

  61. [68]

    Maillard, F

    A. Maillard, F. Krzakala, Y. M. Lu, and L. Zdeborová. Con struction of optimal spectral methods in phase retrieval. In Mathematical and Scientific Machine Learning, 16-19 August 2 021, Virtual Conference / Lausanne, Switzerland , volume 145 of Proceedings of Machine Learning Re...

    2021

  62. [69]

    Maillard, B

    A. Maillard, B. Loureiro, F. Krzakala, and L. Zdeborová . Phase retrieval in high dimensions: Statistical and computational phase transitions. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2 020, NeurIPS 202...

    2020

  63. [70]

    Marchesini

    S. Marchesini. Phase retrieval and saddle-point optim ization. JOSA A , 24(10):3289–3296, 2007

    2007

  64. [71]

    Marchesini, Y

    S. Marchesini, Y. C. Tu, and H. Wu. Alternating projecti on, ptychographic imaging and phase syn- chronization. 2014. available online at http://arxiv.org/abs/1402.0550

    2014 arXiv

  65. [72]

    Mezard, T

    M. Mezard, T. Mora, and R. Zecchina. Clustering of solut ions in the random satisfiability problem. Physical Review Letters , 94:197204, 2005

    2005

  66. [73]

    J. Miao, P. S. Charalambous, J. Kirz, and D. Sayre. Exten ding the methodology of x-ray crystallog- raphy to allow imaging of micrometre-sized non-crystallin e specimens. Nature, 400:342–344, 1999

    1999

  67. [74]

    J. Miao, T. Ishikawa, Q. Shen, and T. Earnest. Extending x-ray crystallography to allow the imaging of noncrystalline materials, cells and single protein comp lexes. Annu. Rev. Phys. Chem. , 59:387–410, 2008

    2008

  68. [75]

    R. P. Millane. Phase retrieval in crystallography and o ptics. J. Opt. Soc. Am. A , 7(3):394–411, Mar 1990

    1990

  69. [76]

    R. P. Millane. Recent advances in phase retrieval. In Image Reconstruction from Incomplete Data IV , volume 6316, page 63160E. International Society for Optics and Photonics, SPIE, 2006

    2006

  70. [77]

    D. L. Misell. A method for the solution of the phase probl em in electron microscopy. J. Phys. D: App. Phy., 6(1):L6–L9, 1973

    1973

  71. [78]

    Mondelli and A

    M. Mondelli and A. Montanari. Fundamental limits of wea k recovery with applications to phase retrieval. Found. Comput. Math. , 19(3):703–773, 2019

    2019

  72. [79]

    Montanari

    A. Montanari. Optimization of the Sherrington-Kirkpa trick hamiltonian. In 60th IEEE Annual Sympo- sium on Foundations of Computer Science, FOCS 2019, Baltimo re, Maryland, USA, November 9-12, 2019, pages 1417–1433. IEEE Computer Society, 2019. 27

    2019

  73. [80]

    Netrapalli, P

    P. Netrapalli, P. Jain, and S. Sanghavi. Phase retrieva l using alternating minimization. In Advances in Neural Information Processing Systems , pages 2796–2804, 2013

    2013

  74. [81]

    Netrapalli, P

    P. Netrapalli, P. Jain, and S. Sanghavi. Phase retrieva l using alternating minimization. IEEE Trans. Signal Process., 63(18):4814–4826, 2015

    2015

  75. [82]

    Ohlsson, A

    H. Ohlsson, A. Y. Yang, R. Dong, and S. S. Sastry. Compres sive phase retrieval from squared output measurements via semi-definite programming. In IF AC Proceedings, volume 45, pages 89–94, 2012

    2012

  76. [83]

    Rangan, P

    S. Rangan, P. Schniter, and A. K. Fletcher. Vector appro ximate message passing. In 2017 IEEE International Symposium on Information Theory, ISIT 2017, Aachen, Germany, June 25-30, 2017 , pages 1588–1592. IEEE, 2017

    2017

  77. [84]

    Rodenburg

    J.M. Rodenburg. Ptychography and related diffractive i maging methods. Advances in Imaging and Electron Physics, 150:87–184, 2008

    2008

  78. [85]

    Salehi, E

    F. Salehi, E. Abbasi, and B. Hassibi. A precise analysis of phasemax in phase retrieval. In 2018 IEEE International Symposium on Information Theory, ISIT 2018, Vail, CO , USA, June 17-22, 2018 , pages 976–980. IEEE, 2018

    2018

  79. [86]

    Schniter and S

    P. Schniter and S. Rangan. Compressive phase retrieval via generalized approximate message passing. IEEE Trans. Signal Process. , 63(4):1043–1055, 2015

    2015

  80. [87]

    Schniter, S

    P. Schniter, S. Rangan, and A. K. Fletcher. Vector appro ximate message passing for the generalized linear model. In 50th Asilomar Conference on Signals, Systems and Computers, ACSSC 2016, Pacific Grove, CA, USA, November 6-9, 2016 , pages 1525–1529. IEEE, 2016

    2016

  81. [88]

    Shechtman, Y

    Y. Shechtman, Y. C. Eldar, O. Cohen, H. N. Chapman, J. Mia o, and M. Segev. Phase retrieval with application to optical imaging: A contemporary overview. IEEE Signal Process. Mag. , 32(3):87–109, 2015

    2015

  82. [89]

    Soltanolkotabi

    M. Soltanolkotabi. Structured signal recovery from qu adratic measurements: Breaking sample com- plexity barriers via nonconvex optimization. IEEE Trans. Inf. Theory , 65(4):2374–2400, 2019

    2019

  83. [90]

    M. Stojnic. A framework for perfromance characterizat ion of LASSO algortihms. available online at http://arxiv.org/abs/1303.7291

    arXiv

  84. [91]

    M. Stojnic. Various thresholds for ℓ1-optimization in compressed sensing. available online at http:// arxiv.org/abs/0907.3666

    arXiv

  85. [92]

    M. Stojnic. Block-length dependent thresholds for ℓ2/ℓ1-optimization in block-sparse compressed sens- ing. ICASSP, IEEE International Conference on Acoustics, Signal and Speech Processing, pages 3918– 3921, 14-19 March 2010. Dallas, TX

    2010

  86. [93]

    M. Stojnic. ℓ1 optimization and its various thresholds in compressed sens ing. ICASSP, IEEE Inter- national Conference on Acoustics, Signal and Speech Proces sing, pages 3910–3913, 14-19 March 2010. Dallas, TX

    2010

  87. [94]

    M. Stojnic. Recovery thresholds for ℓ1 optimization in binary compressed sensing. ISIT, IEEE Inter- national Symposium on Information Theory , pages 1593 – 1597, 13-18 June 2010. Austin, TX

    2010

  88. [95]

    M. Stojnic. Towards improving ℓ1 optimization in compressed sensing. ICASSP, IEEE International Conference on Acoustics, Signal and Speech Processing , pages 3938–3941, 14-19 March 2010. Dallas, TX

    2010

  89. [96]

    M. Stojnic. Another look at the Gardner problem. 2013. a vailable online at http://arxiv.org/abs/ 1306.3979

    2013 arXiv

  90. [97]

    M. Stojnic. Regularly random duality. 2013. available online at http://arxiv.org/abs/1303.7295. 28

    2013 arXiv

  91. [98]

    M. Stojnic. Box constrained ℓ1 optimization in random linear systems – asymptotics. 2016. available online at http://arxiv.org/abs/1612.06835

    2016 arXiv

  92. [99]

    M. Stojnic. Fully bilinear generic and lifted random pr ocesses comparisons. 2016. available online at http://arxiv.org/abs/1612.08516

    2016 arXiv

  93. [100]

    M. Stojnic. Generic and lifted probabilistic compari sons – max replaces minmax. 2016. available online at http://arxiv.org/abs/1612.08506

    2016 arXiv

  94. [101]

    M. Stojnic. Fully lifted random duality theory. 2023. available online at http://arxiv.org/abs/ 2312.00070

    2023 arXiv

  95. [102]

    M. Stojnic. Deep relu networks – injectivity capacity upper bounds. 2024. available online at http:// arxiv.org/abs/2412.19677

    2024 arXiv

  96. [103]

    M. Stojnic. Injectivity capacity of relu gates. 2024. available online at http://arxiv.org/abs/2410. 20646

    2024

  97. [105]

    Straziota and L

    D. Straziota and L. Saglietti. Isolating the hard core of phaseless inference: the Phase selection formulation. 2025. available online at http://arxiv.org/abs/2502.04282

    2025 arXiv

  98. [106]

    J. Sun, Q. Qu, and J. Wright. A geometric analysis of pha se retrieval. Found. Comput. Math. , 18(5):1131–1198, 2018

    2018

  99. [107]

    Takahashi and Y

    T. Takahashi and Y. Kabashima. Macroscopic analysis o f vector approximate message passing in a model-mismatched setting. IEEE Trans. Inf. Theory , 68(8):5579–5600, 2022

    2022

  100. [108]

    Y. S. Tan and R. Vershynin. Phase retrieval via randomi zed kaczmarz: Theoretical guarantees. Infor- mation and Inference: A Journal of the IMA , 8(1):97–123, 2019

    2019

  101. [109]

    Thibault, M

    P. Thibault, M. Dierolf, A. Menzel, O. Bunk, C. David, a nd F. Pfeffer. High-resolution scanning x-ray diffraction microscopy. Science, 322(5887):379–382, 2008

    2008

  102. [110]

    C. Vinzant. A small frame and a certificate of its inject ivity. In IEEE International Conference on Sampling Theory and Applications (SampTA) , pages 197–200, 2015

    2015

  103. [111]

    Waldspurger

    I. Waldspurger. Phase retrieval with random Gaussian sensing vectors by alternating projections. IEEE Trans. Inf. Theory , 64(5):3301–3312, 2018

    2018

  104. [112]

    Waldspurger, A

    I. Waldspurger, A. d’Aspremont, and S. Mallat. Phase r ecovery, maxcut and complex semidefinite programming. Math. Program., 149(1-2):47–81, 2015

    2015

  105. [113]

    A. Walther. The question of phase retrieval in optics. Optica Acta: International Journal of Optics , 10(1):41–49, 1963

    1963

  106. [114]

    G. Wang, G. B. Giannakis, and Y. C. Eldar. Solving syste ms of random quadratic equations via truncated amplitude flow. IEEE Trans. Inf. Theory , 64(2):773–794, 2018

    2018

  107. [115]

    K. Wei. Solving systems of phaseless equations via Kac zmarz methods: A proof of concept study. Inverse Problems, 31(12):125008, 2015

    2015

  108. [116]

    G. Yang, B. Dong, B. Gu, J. Zhuang, and O. K. Ersoy. Gerch berg-Saxton and Yang-Gu algorithms for phase retrieval in a nonunitary transform system: a compari son. Applied optics, 33(2):209–218, 1994

    1994

  109. [117]

    Z. Yuan, H. Wang, and Q. Wang. Phase retrieval via spars e Wirtinger flow. Journal of Computational and Applied Mathematics , 355:162–173, 2019. 29

    2019

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.