Pith. sign in

REVIEW 3 major objections 5 minor 55 references

Landscape Analysis of Simultaneous Blind Deconvolution and Phase Retrieval via Structured Low-Rank Tensor Recovery

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 For the surrogate of simultaneous blind deconvolution and phase retrieval, every critical point is either the global optimum or a non-strict saddle.

desk verdict A correct and novel landscape classification for the x∘x∘h population risk, but the paper's claim to characterize the original BDPR problem rests on an unproven TRIP bridge. read the letter →

arxiv 2509.10834 v1 pith:3B7AQQE2 submitted 2025-09-13 eess.SP cs.ITmath.IT

Xiao Liang , Zhen Qin , Zhihui Zhu , Shuang Li This is my paper · ORCID
classification eess.SPcs.ITmath.IT MSC 15A6965K1090C2694A12
keywords blinddeconvolutionphaseretrievaltensorfactorizationsensinggeometriclandscaperank-onerecoveryRiemanniangradientdescent
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 analyzes simultaneous blind deconvolution and phase retrieval (BDPR), where one must recover both an unknown signal x and source-shape coefficients h from intensity-only blurred measurements. Because the sensing tensor of the original problem is too complicated to analyze directly, the authors study a surrogate: recovering the rank-one partial-symmetric tensor T* = x*∘x*∘h* from linear measurements. For the population risk of this factorization, they prove that all critical points are either the global optimum x=±x*, h=h* or non-strict saddles x⊥x*, h=0, so there are no spurious local minima. They then prove that Riemannian gradient descent with a spectral initialization converges linearly under a tensor restricted isometry property, and degrades gracefully under noise. If the surrogate faithfully captures the original problem, these results provide the first convergence guarantees for a class of BDPR algorithms that previously had none.

What carries the argument

The load-bearing object is the partial-symmetric rank-one CP tensor x∘x∘h, in which two modes share the factor x. The machinery has four parts: the population risk f(x,h)=1/2||x∘x∘h−T*||$_F^{2}$ over the unit sphere, whose hybrid Riemannian Hessian at critical points takes the closed form in equation (7) and immediately shows positive semidefiniteness; the factor distance dist²(x,h)=min_{a=±1} 2||T*||_F²||x−ax*||²+||h−h*||² together with Lemma IV.1 relating it to the tensor reconstruction error; the tensor restricted isometry property (TRIP) for CP rank at most r, which preserves inner products under the linear sensing map; and a hybrid Riemannian gradient descent update with a spectral initialization obtained from the SVD of the mode-1 unfolding of the adjoint applied to the measurements. These ingredients convert a nonconvex tensor recovery problem into one where all critical points are classified and descent contracts the distance metric at a linear rate.

What would settle it

Compute the restricted isometry constant of the structured operator (3) numerically: draw many rank-one tensors x∘x∘h, evaluate deviations of (1/m)||A(x∘x∘h)||² from ||x∘x∘h||², and check whether they stay below the theory's δ when m=O(N+K); a violating tensor, or a large gap between the structured and Gaussian phase transitions of RGD, would break the surrogate bridge.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the factorized objective f(x,h)=1/2||x∘x∘h−T*||$_F^{2}$ with ||x||_2=1 has a benign landscape despite the repeated factor x. Theorem IV.1 characterizes every critical point: the global minima are x=±x*, h=h*, and the only other critical points are x⊥x*, h=0, where the hybrid Riemannian Hessian is positive semidefinite. This means first-order methods that keep h away from zero cannot be trapped away from the truth. For the tensor sensing surrogate (12), the paper shows that under TRIP with rank up to 5, the spectral initializer lands within a controlled neighborhood, and RGD contracts the distance metric by a factor of 1−O(µ) per iteration; with noise, convergence stops at a floor proportional to γ√((N+K)/m). The authors present these results as principled guidance for the original BDPR problem because the underlying tensor x*∘x*∘h has the same structure, while explicitly deferring a direct landscape analysis of the structured sensing tensor (3) and a proof of TRIP for it to future work.

Load-bearing premise

The load-bearing premise is that the structured sensing tensors of the original BDPR problem obey the same tensor restricted isometry condition as the random subgaussian surrogate, so the surrogate's benign landscape carries over; the paper leaves that proof to future work.

Editorial extensions

If this is right

  • Because Theorem IV.1 rules out spurious local minima, any first-order method that reaches a point with h≠0 will converge to the global optimum up to the sign ambiguity.
  • With subgaussian sensing tensors, m=Ω(N+K) measurements suffice for the TRIP prerequisite, so RGD with spectral initialization is sample-efficient in the problem's ambient dimension (Theorem V.1 and Theorem V.3).
  • In the noisy case, RGD converges linearly down to a noise floor of order γ√((N+K)/m), meaning measurement noise degrades the estimate gracefully rather than derailing convergence (Theorem V.5).
  • Lemma IV.1 transfers the linear contraction in factor distance to linear contraction in tensor Frobenius error, so the guarantee applies to the actual reconstruction error.
  • The structured sensing experiments show the same linear convergence pattern, which the paper offers as empirical evidence that the surrogate analysis guides the original BDPR problem even though the theory does not yet cover it.

Reading between the lines

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

  • If the structured operator (3) is later shown to satisfy TRIP, Theorems V.2-V.5 would immediately supply recovery guarantees for the original BDPR problem; the missing link is purely a verification of the isometry constant.
  • The same Hessian calculation should extend to other factor-sharing low-rank templates, such as tensors of the form x∘y∘y or higher-order symmetric slices, which would give a unified proof tool for repeated-factor tensor factorization landscapes.
  • A quantitative test of surrogacy would compare the phase transition of RGD on the structured tensors (3) with the Gaussian phase transition: if the required number of measurements tracks N+K with similar constants, the surrogate captures the original geometry.
  • The benign saddle structure suggests that even without exact TRIP, relaxed isometry conditions or batch-stochastic variants of RGD may still work, so the practical guidance of the paper is likely stronger than its theorems.
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 paper studies the simultaneous blind deconvolution and phase retrieval (BDPR) problem through a structured low-rank tensor recovery surrogate. The target tensor is T* = x* ∘ x* ∘ h*, and the authors analyze the population risk f(x,h) = 1/2 ||x∘x∘h − T*||_F^2 on the unit sphere. Theorem IV.1 classifies all critical points as either the global optima (±x*, h*) or non-strict saddles of the form x⊥x*, h=0. Theorem IV.2 proves local linear convergence of a hybrid Riemannian gradient descent (RGD) scheme for the factorization problem. The paper then extends the analysis to the tensor sensing problem (12) under a tensor restricted isometry property (TRIP): Theorems V.2–V.5 give spectral initialization accuracy, local linear convergence, and a noise-robust variant. Numerical experiments cover both Gaussian sensing tensors and the structured sensing tensors (3) arising from the original BDPR model.

Significance. The critical-point analysis in Section IV is clean and clearly derived, and the local convergence proofs give explicit contraction rates and noise floors, which are useful and falsifiable. If the TRIP-based theorems hold, the paper provides a rigorous landscape analysis for rank-one partial-symmetric tensor sensing, a model that is more general than the original BDPR problem. The main weakness is that the connection to BDPR is conditional: the TRIP for the structured sensing operator (3) is explicitly left to future work, and Section VI-B admits that the structured experiments are outside the current theory. The paper is therefore stronger as a standalone tensor-sensing analysis than as a landscape analysis of BDPR itself; the abstract and conclusion overstate the equivalence between the surrogate and the original problem.

major comments (3)
  1. [V (after Theorem V.1), VI-B, Abstract] The load-bearing bridge from the surrogate to the original BDPR problem is unproven. Theorem V.3 and Theorem V.5 require A to satisfy the TRIP with r=5, but Theorem V.1 supplies TRIP only for i.i.d. subgaussian sensing tensors. Immediately after Theorem V.1, the paper states that the formal proof of TRIP for the structured sensing operator (3) is left as future work, and Section VI-B states that the structured experiments are outside the current theory. Despite this, the abstract and conclusion claim that the surrogate "equivalently characterizes the original BDPR problem" and provides "principled guidance" for it. If the structured operator (3) fails TRIP, or satisfies it only with a much larger measurement complexity than the subgaussian bound (11), none of the convergence theorems transfers to BDPR. The authors should either prove an appropriate TRIP for (3), or explicitly reframe the contribution as an analysis of generic rank-one partial-symmetric tensor sensing and remove the BDPR-equivalence claims.
  2. [Appendix B, Lemma B.3 (used in Appendices D, E, F, G)] Lemma B.3 is stated for TRIP with r=2 and for "any CP format tensors X1, X2". As stated, this is false when X1±X2 has CP rank larger than 2, because TRIP with r=2 gives no control over such pairs. The proof of Theorem V.2 invokes "Lemma B.3 with r=3", and the proof of Theorem V.3 invokes "Lemma B.3 with r=5", so the paper implicitly relies on a rank-aware polarization bound of the form: if A satisfies TRIP with r ≥ rank(X1) + rank(X2), then |(1/m)<A(X1),A(X2)> − <X1,X2>| ≤ δ_r ||X1||_F ||X2||_F. The lemma must be restated in that general form and the rank budget verified in each application. Without this correction, Theorem V.2 and the convergence proofs built on it are not fully supported as written.
  3. [V-B (Theorems V.4 and V.5), Appendix G] The noisy theorems assume only that e has i.i.d. entries with mean zero and variance γ², but the proofs use the concentration bound (47) cited from [54, eq. (D.6)], which requires at least sub-Gaussian (or bounded) noise entries. Finite variance alone does not give the exponentially small failure probability needed to control the spectral initializer and the iterates. The authors should either strengthen the noise assumption to sub-Gaussian entries, or supply a proof under the stated second-moment assumption. As written, the noise-robustness guarantee is not fully justified.
minor comments (5)
  1. [Appendix E, after equation (40)] In the displayed bound for ||c2||_2, the second inequality reads ||c2−b2||_2 + ||c2||_2; it should be ||c2−b2||_2 + ||b2||_2. The numerical bound is unchanged, but the typo obscures the argument.
  2. [Appendix G, proof of Theorem V.5] In the display after equation (54), "3−15δ_c" should be "3−15δ_r"; this appears to be a typographical error.
  3. [VI-B, structured experiments] The text says the authors "replace ||T*||_F^2 with ||T_t||_F^2 in the RGD updates (8)". This modification is not covered by Theorems IV.2, V.3, or V.5, which use the fixed quantity ||T*||_F^2. A remark explaining why this algorithmic change is benign would improve clarity.
  4. [Theorem V.5 statement] The noise floor is written as O((5(N+K)+5^3)/m γ²), while the proof tracks terms of the form O(sqrt((5(N+K)+5^3)/m) γ) before applying Young's inequality. Please clarify the exact constants and how the final expression is obtained.
  5. [Section IV-A, after Theorem IV.1] The sentence that any first-order method producing h≠0 "will converge to the second class of critical points" is informal: gradient descent can in principle converge to a non-strict saddle or move slowly along the flat directions unless additional conditions are imposed. This statement should be softened or given a precise qualification.
Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities and fits no numerical constants to data. The theoretical constants are analytic. The main unpaid assumptions are the TRIP condition for the surrogate sensing model and the unproven bridge from that surrogate to the structured BDPR operator.

assumptions (5)
  • domain assumption The sensing operator A satisfies TRIP with CP rank r=5 and constant δ_r ≤ 4/15 in the noiseless case and δ_r ≤ 3/15 in the noisy case.
    Assumed in Theorems V.3 and V.5. Theorem V.1 establishes TRIP only for i.i.d. subgaussian sensing tensors, not for the structured BDPR operator (3).
  • ad hoc to paper The unstructured subgaussian tensor sensing problem (12) preserves the essential structural features of the original BDPR tensor recovery problem.
    Stated in the introduction and conclusion; no quantitative sense of 'preserves' is given, and the paper admits the structured TRIP is unproved.
  • domain assumption The theoretical results are derived in the real-valued setting, and the extension to the complex domain is asserted without proof.
    Section IV states 'we focus on the real-valued setting for clarity... all results naturally extend to the complex domain.' The original BDPR problem is complex-valued.
  • domain assumption The unknown source shape s lies in a known low-dimensional subspace s = Bh with K ≪ N.
    Inherited from [19] in Section III; this assumption is needed for the tensor reformulation to be low-rank.
  • standard math Standard subgaussian concentration results are used for the TRIP sample complexity and the noise bounds.
    Theorem V.1 invokes [48, Theorem 2]; Appendix F uses [54, eqn (D.6)].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Landscape Analysis of Simultaneous Blind Deconvolution and Phase Retrieval via Structured Low-Rank Tensor Recovery." pith.science (2026). https://pith.science/paper/3B7AQQE2

@misc{pith2026250910834,
  author       = {Pith},
  title        = {Pith review of: Landscape Analysis of Simultaneous Blind Deconvolution and Phase Retrieval via Structured Low-Rank Tensor Recovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3B7AQQE2}},
  note         = {Machine review of arXiv:2509.10834}
}
read the original abstract

This paper presents a geometric analysis of the simultaneous blind deconvolution and phase retrieval (BDPR) problem via a structured low-rank tensor recovery framework. Due to the highly complicated structure of the associated sensing tensor, directly characterizing its optimization landscape is intractable. To address this, we introduce a tensor sensing problem as a tractable surrogate that preserves the essential structural features of the target low-rank tensor while enabling rigorous theoretical analysis. As a first step toward understanding this surrogate model, we study the corresponding population risk, which captures key aspects of the underlying low-rank tensor structure. We characterize the global landscape of the population risk on the unit sphere and show that Riemannian gradient descent (RGD) converges linearly under mild conditions. We then extend the analysis to the tensor sensing problem, establishing local geometric properties, proving convergence guarantees for RGD, and quantifying robustness under measurement noise. Our theoretical results are further supported by extensive numerical experiments. These findings offer foundational insights into the optimization landscape of the structured low-rank tensor recovery problem, which equivalently characterizes the original BDPR problem, thereby providing principled guidance for solving the original BDPR problem.

Figures

Figures reproduced from arXiv: 2509.10834 by the authors.

Figure 1
Figure 1. Convergence behavior of RGD for the noiseless Gaussian tensor sensing problem. (a) The loss function [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Successful recovery rates of RGD under varying [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Convergence behavior of RGD for the noiseless structured tensor sensing problem. (a) The loss function [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Successful recovery rates under varying numbers [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Gaussian tensor sensing with noise: the loss function [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Structured tensor sensing with noise: the loss function [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 50 canonical work pages

  1. [1]

    Blind deconvolution using con- vex programming,

    A. Ahmed, B. Recht, and J. Romberg, “Blind deconvolution using con- vex programming,”IEEE Transactions on Information Theory, vol. 60, no. 3, pp. 1711–1732, 2014

    work page 2014

  2. [2]

    Gespar: Efficient phase retrieval of sparse signals,

    Y . Shechtman, A. Beck, and Y . C. Eldar, “Gespar: Efficient phase retrieval of sparse signals,”IEEE Transactions on Signal Processing, vol. 62, no. 4, pp. 928–938, 2014

    work page 2014

  3. [3]

    Neural blind deconvolution using deep priors,

    D. Ren, K. Zhang, Q. Wang, Q. Hu, and W. Zuo, “Neural blind deconvolution using deep priors,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

    work page 2020

  4. [4]

    A neural network approach for the blind deconvolution of turbulent flows,

    R. Maulik and O. San, “A neural network approach for the blind deconvolution of turbulent flows,”Journal of Fluid Mechanics, vol. 831, p. 151–181, 2017. SEPTEMBER 2025 10

    work page 2017

  5. [5]

    Feature-based phase retrieval wavefront sensing approach using machine learning,

    G. Ju, X. Qi, H. Ma, and C. Yan, “Feature-based phase retrieval wavefront sensing approach using machine learning,”Opt. Express, vol. 26, pp. 31767–31783, Nov 2018

    work page 2018

  6. [6]

    Blind image deconvolution using deep generative priors,

    M. Asim, F. Shamshad, and A. Ahmed, “Blind image deconvolution using deep generative priors,”IEEE Transactions on Computational Imaging, vol. 6, pp. 1493–1506, 2020

    work page 2020

  7. [7]

    Relaxation of align- ment errors and phase calibration in computational frequency-diverse imaging using phase retrieval,

    O. Yurduseven, T. Fromenteze, and D. R. Smith, “Relaxation of align- ment errors and phase calibration in computational frequency-diverse imaging using phase retrieval,”IEEE Access, vol. 6, pp. 14884–14894, 2018

    work page 2018

  8. [8]

    Blind deconvolution using a normalized sparsity measure,

    D. Krishnan, T. Tay, and R. Fergus, “Blind deconvolution using a normalized sparsity measure,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 233–240, 2011

    work page 2011

Show all 55 references
  1. [9]

    Understanding blind deconvolution algorithms,

    A. Levin, Y . Weiss, F. Durand, and W. T. Freeman, “Understanding blind deconvolution algorithms,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 33, no. 12, pp. 2354–2367, 2011

    2011

  2. [10]

    Normalized blind deconvolution,

    M. Jin, S. Roth, and P. Favaro, “Normalized blind deconvolution,” in Proceedings of the European Conference on Computer Vision (ECCV), September 2018

    2018

  3. [11]

    Understanding and evaluating blind deconvolution algorithms,

    A. Levin, Y . Weiss, F. Durand, and W. T. Freeman, “Understanding and evaluating blind deconvolution algorithms,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1964–1971, 2009

    1964

  4. [12]

    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 Processing Magazine, vol. 32, no. 3, pp. 87–109, 2015

    2015

  5. [13]

    The numerics of phase retrieval,

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

    2020

  6. [14]

    The phase retrieval problem,

    L. Taylor, “The phase retrieval problem,”IEEE Transactions on Anten- nas and Propagation, vol. 29, no. 2, pp. 386–391, 1981

    1981

  7. [15]

    Phase retrieval via Wirtinger flow: Theory and algorithms,

    E. J. Cand `es, X. Li, and M. Soltanolkotabi, “Phase retrieval via Wirtinger flow: Theory and algorithms,”IEEE Transactions on Information The- ory, vol. 61, no. 4, pp. 1985–2007, 2015

    1985

  8. [16]

    Class-specific blind deconvolutional phase retrieval under a generative prior,

    F. Shamshad and A. Ahmed, “Class-specific blind deconvolutional phase retrieval under a generative prior,”arXiv preprint arXiv:2002.12578, 2020

    2002 arXiv

  9. [17]

    Simultaneous phase retrieval and blind deconvolution via convex programming,

    A. Ahmed, A. Aghasi, and P. Hand, “Simultaneous phase retrieval and blind deconvolution via convex programming,”Journal of Machine Learning Research, vol. 20, no. 157, pp. 1–28, 2019

    2019

  10. [18]

    Blind deconvolution meets phase retrieval in optical wireless communications,

    M. Fu and Y . Shi, “Blind deconvolution meets phase retrieval in optical wireless communications,” in2019 IEEE 90th Vehicular Technology Conference (VTC2019-Fall), pp. 1–5, 2019

    2019

  11. [19]

    Simultaneous blind deconvolution and phase retrieval with tensor iterative hard thresholding,

    S. Li, G. Tang, and M. B. Wakin, “Simultaneous blind deconvolution and phase retrieval with tensor iterative hard thresholding,” inInternational Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2977–2981, IEEE, 2019

    2019

  12. [20]

    Structured local optima in sparse blind deconvolution,

    Y . Zhang, H.-W. Kuo, and J. Wright, “Structured local optima in sparse blind deconvolution,”IEEE Transactions on Information Theory, vol. 66, no. 1, pp. 419–452, 2020

    2020

  13. [21]

    Global geometry of multichannel sparse blind de- convolution on the sphere,

    Y . Li and Y . Bresler, “Global geometry of multichannel sparse blind de- convolution on the sphere,”Advances in Neural Information Processing Systems, vol. 31, 2018

    2018

  14. [22]

    The nonsmooth landscape of blind deconvolution,

    M. D ´ıaz, “The nonsmooth landscape of blind deconvolution,”arXiv preprint arXiv:1911.08526, 2019

    1911 arXiv

  15. [23]

    Solving random quadratic systems of equations is nearly as easy as solving linear systems,

    Y . Chen and E. Cand`es, “Solving random quadratic systems of equations is nearly as easy as solving linear systems,” inAdvances in Neural Information Processing Systems, pp. 739–747, 2015

    2015

  16. [24]

    A geometric analysis of phase retrieval,

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

    2018

  17. [25]

    On the optimization landscape of tensor decompo- sitions,

    R. Ge and T. Ma, “On the optimization landscape of tensor decompo- sitions,”Advances in Neural Information Processing Systems, vol. 30, 2017

    2017

  18. [26]

    Optimization landscape of tucker decomposi- tion,

    A. Frandsen and R. Ge, “Optimization landscape of tucker decomposi- tion,”Mathematical Programming, vol. 193, no. 2, pp. 687–712, 2022

    2022

  19. [27]

    Local and global convergence of general burer- monteiro tensor optimizations,

    S. Li and Q. Li, “Local and global convergence of general burer- monteiro tensor optimizations,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 36, pp. 10266–10274, 2022

    2022

  20. [28]

    Landscape analysis of an improved power method for tensor decomposition,

    J. Kileel, T. Klock, and J. M Pereira, “Landscape analysis of an improved power method for tensor decomposition,”Advances in Neural Information Processing Systems, vol. 34, pp. 6253–6265, 2021

    2021

  21. [29]

    Guaranteed nonconvex factorization approach for tensor train recovery,

    Z. Qin, M. B. Wakin, and Z. Zhu, “Guaranteed nonconvex factorization approach for tensor train recovery,”Journal of Machine Learning Research, vol. 25, no. 383, pp. 1–48, 2024

    2024

  22. [30]

    A scalable factorization approach for high-order structured tensor recovery,

    Z. Qin, M. B. Wakin, and Z. Zhu, “A scalable factorization approach for high-order structured tensor recovery,”arXiv preprint arXiv:2506.16032, 2025

    2025 arXiv

  23. [31]

    Tensor decomposition for signal processing and machine learning,

    N. D. Sidiropoulos, L. De Lathauwer, X. Fu, K. Huang, E. E. Papalex- akis, and C. Faloutsos, “Tensor decomposition for signal processing and machine learning,”IEEE Transactions on Signal Processing, vol. 65, no. 13, pp. 3551–3582, 2017

    2017

  24. [32]

    Tensorizing neural networks,

    A. Novikov, D. Podoprikhin, A. Osokin, and D. P. Vetrov, “Tensorizing neural networks,”Advances in Neural Information Processing Systems, vol. 28, 2015

    2015

  25. [33]

    Tensor fusion network for multimodal sentiment analysis,

    A. Zadeh, M. Chen, S. Poria, E. Cambria, and L.-P. Morency, “Tensor fusion network for multimodal sentiment analysis,” inProceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 1103–1114, 2017

    2017

  26. [34]

    Towards quantum machine learning with tensor networks,

    W. Huggins, P. Patil, B. Mitchell, K. B. Whaley, and E. M. Stoudenmire, “Towards quantum machine learning with tensor networks,”Quantum Science and Technology, vol. 4, no. 2, p. 024001, 2019

    2019

  27. [35]

    Tensor decompositions and applications,

    T. G. Kolda and B. W. Bader, “Tensor decompositions and applications,” SIAM Review, vol. 51, no. 3, pp. 455–500, 2009

    2009

  28. [36]

    A new trunca- tion strategy for the higher-order singular value decomposition,

    N. Vannieuwenhoven, R. Vandebril, and K. Meerbergen, “A new trunca- tion strategy for the higher-order singular value decomposition,”SIAM Journal on Scientific Computing, vol. 34, no. 2, pp. A1027–A1052, 2012

    2012

  29. [38]

    Quantitative phase- amplitude microscopy i: optical microscopy,

    E. Barone-Nugent, A. Barty, and K. Nugent, “Quantitative phase- amplitude microscopy i: optical microscopy,”Journal of Microscopy, vol. 206, no. 3, pp. 194–203, 2002

    2002

  30. [39]

    Phase retrieval and differential phase-contrast imaging with low-brilliance X-ray sources,

    F. Pfeiffer, T. Weitkamp, O. Bunk, and C. David, “Phase retrieval and differential phase-contrast imaging with low-brilliance X-ray sources,” Nature Physics, vol. 2, no. 4, pp. 258–261, 2006

    2006

  31. [40]

    Contrast in the electron spectroscopic imaging mode of a tem: Iv. thick specimens imaged by the most-probable energy loss,

    L. Reimer, R. Rennekamp, I. Fromm, and M. Langenfeld, “Contrast in the electron spectroscopic imaging mode of a tem: Iv. thick specimens imaged by the most-probable energy loss,”Journal of Microscopy, vol. 162, no. 1, pp. 3–14, 1991

    1991

  32. [41]

    Born and E

    M. Born and E. Wolf,Principles of optics: electromagnetic theory of propagation, interference and diffraction of light. Elsevier, 2013

    2013

  33. [42]

    The landscape of non-convex empirical risk with degenerate population risk,

    S. Li, G. Tang, and M. B. Wakin, “The landscape of non-convex empirical risk with degenerate population risk,” inAdvances in Neural Information Processing Systems, pp. 3502–3512, 2019

    2019

  34. [43]

    Landscape correspondence of empirical and population risks in the eigendecomposition problem,

    S. Li, G. Tang, and M. B. Wakin, “Landscape correspondence of empirical and population risks in the eigendecomposition problem,” IEEE Transactions on Signal Processing, vol. 70, pp. 2985–2999, 2022

    2022

  35. [44]

    Compressed sensing,

    D. L. Donoho, “Compressed sensing,”IEEE Transactions on Informa- tion Theory, vol. 52, no. 4, pp. 1289–1306, 2006

    2006

  36. [45]

    Robust uncertainty principles: Ex- act signal reconstruction from highly incomplete frequency information,

    E. J. Cand `es, J. Romberg, and T. Tao, “Robust uncertainty principles: Ex- act signal reconstruction from highly incomplete frequency information,” IEEE Transactions on Information Theory, vol. 52, no. 2, pp. 489–509, 2006

    2006

  37. [46]

    An introduction to compressive sampling,

    E. J. Cand `es and M. B. Wakin, “An introduction to compressive sampling,”IEEE Signal Processing Magazine, vol. 25, no. 2, pp. 21–30, 2008

    2008

  38. [47]

    Iterative hard thresholding for low CP-rank tensor models,

    R. Grotheer, S. Li, A. Ma, D. Needell, and J. Qin, “Iterative hard thresholding for low CP-rank tensor models,”Linear and Multilinear Algebra, vol. 70, no. 22, pp. 7452–7468, 2022

    2022

  39. [48]

    Low rank tensor recovery via iterative hard thresholding,

    H. Rauhut, R. Schneider, and Z. Stojanac, “Low rank tensor recovery via iterative hard thresholding,”Linear Algebra and its Applications, vol. 523, pp. 220–262, 2017

    2017

  40. [49]

    Quantum state tomography for matrix product density operators,

    Z. Qin, C. Jameson, Z. Gong, M. B. Wakin, and Z. Zhu, “Quantum state tomography for matrix product density operators,”IEEE Transactions on Information Theory, 2024

    2024

  41. [50]

    Nonconvex low-rank symmetric tensor completion from noisy data,

    C. Cai, G. Li, H. V . Poor, and Y . Chen, “Nonconvex low-rank symmetric tensor completion from noisy data,”Advances in Neural Information Processing Systems, vol. 32, 2019

    2019

  42. [51]

    Provable tensor-train format tensor completion by Riemannian optimization,

    J.-F. Cai, J. Li, and D. Xia, “Provable tensor-train format tensor completion by Riemannian optimization,”Journal of Machine Learning Research, vol. 23, no. 123, pp. 1–77, 2022

    2022

  43. [52]

    Weakly convex optimization over Stiefel manifold using Riemannian subgradient-type methods,

    X. Li, S. Chen, Z. Deng, Q. Qu, Z. Zhu, and A. Man-Cho So, “Weakly convex optimization over Stiefel manifold using Riemannian subgradient-type methods,”SIAM Journal on Optimization, vol. 31, no. 3, pp. 1605–1634, 2021

    2021

  44. [53]

    Tensor-train decomposition,

    I. V . Oseledets, “Tensor-train decomposition,”SIAM Journal on Scien- tific Computing, vol. 33, no. 5, pp. 2295–2317, 2011

    2011

  45. [54]

    An optimal statistical and computational framework for generalized tensor estimation,

    R. Han, R. Willett, and A. R. Zhang, “An optimal statistical and computational framework for generalized tensor estimation,”The Annals of Statistics, vol. 50, no. 1, pp. 1–29, 2022. SEPTEMBER 2025 11 APPENDIXA DERIVATION OF THERIEMANNIANHESSIAN AT THE CRITICALPOINTS OFPROBLEM(...

    2022

  46. [55]

    Proof of(24)by induction:First note that (24) holds att= 0by initialization

    This establishes local linear conver- gence. Proof of(24)by induction:First note that (24) holds att= 0by initialization. Suppose it holds att=t ′, so that∥h t′∥2 2 ≤ 9∥T ⋆∥2 F 4 . By invoking (35), we then have dist2(xt′+1,ht′+1)≤dist 2(xt′,ht′). Hence, (24) also holds att=t ...

    2025

  47. [56]

    APPENDIXF PROOF OFTHEOREMV.4 Proof.We begin by establishing a fundamental probabilistic property for the noise term

    This completes the proof. APPENDIXF PROOF OFTHEOREMV.4 Proof.We begin by establishing a fundamental probabilistic property for the noise term. Since the tensor±Pr i=1 xi◦xi◦hi can be viewed as a Tucker decomposition with multilinear ranks(r,r,r), we have 1 m mX i=1 * eiAi, rX ...

    2025

Pith tools

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