Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Multilook Coherent Imaging: Theoretical Guarantees and Algorithms

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper proves the first mean-squared-error bound for maximum-likelihood multilook coherent imaging under the deep image prior hypothesis, and pairs it with a practical projected-gradient algorithm.

desk verdict First MSE bound for multilook coherent imaging under a deep image prior is a real contribution, but the main theorem overclaims its regime: the proof only goes through for m/n ≤ 1/5, not m<n. read the letter →

arxiv 2505.23594 v1 pith:V3FS5LOM submitted 2025-05-29 stat.ML cs.LGeess.IV

classification stat.MLcs.LGeess.IV
keywords multilookcoherentimagingspecklenoisedeepimagepriormaximumlikelihoodestimationmeansquarederrorprojectedgradientdescentbaggingNewton-Schulzalgorithm
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

Multilook coherent imaging captures several noisy measurement sets, or looks, of the same scene, and the multiplicative speckle noise makes reconstruction hard when there are fewer measurements than pixels. The paper aims to give a quantitative guarantee for the maximum-likelihood estimator under the deep image prior hypothesis, which says natural images lie in the range of a small untrained network. It establishes the first upper bound on the mean squared error of that estimator, showing explicitly how the error depends on the number of looks, the number of network parameters, the number of measurements, and the image dimension. On the algorithmic side, it develops projected gradient descent variants that use a fast approximate matrix inverse and a bagged ensemble of deep-image-prior projections to make the method work in practice. If the guarantee holds, it tells an imaging engineer how many looks are worth collecting before the measurement geometry becomes the limiting factor.

What carries the argument

The load-bearing object is the negative log-likelihood $f(x)=\log\det\Sigma(x)+\frac{1}{L\sigma_w^2}\sum_{\ell=1}^L y_\ell^\top \Sigma(x)y_\ell$, with $\Sigma(x)=(AX^2A^\top)^{-1}$, minimized over the range of the deep image prior. The proof compares $\hat{\Sigma}=\Sigma(\hat{x})$ with $\Sigma_o=\Sigma(x_o)$ through the gap in the expected likelihood, then bounds the stochastic fluctuation of the quadratic form around its mean. A $\delta$-net over the parameter ball $B_k(0,x_{\max}\sqrt{n/k})$ lets the fluctuation be controlled uniformly, and the 1-Lipschitz property transfers closeness in parameter space to closeness in image space. The argument is carried by concentration bounds for quadratic forms, a decoupling inequality for off-diagonal terms, singular-value bounds on $A$, and a covering-number bound on the parameter set.

What would settle it

Take a synthetic image that exactly lies in the range of a 1-Lipschitz network satisfying the parameter-ball condition, generate $L$ looks with iid Gaussian speckle through an iid Gaussian $A$ with $m<n$, and compute the global maximizer of the likelihood by a dense search over the small parameter space. If the normalized squared error repeatedly exceeds $C_1\left[\frac{n}{m^2}\frac{k\log n}{L}+\frac{\sqrt{k\log n}}{m}\right]$, or if increasing $L$ far beyond the predicted threshold continues to push the error below the $\sqrt{k\log n}/m$ floor, the theorem's rate is false.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem III.1: when the measurement matrix $A$ has iid $N(0,1)$ entries, $m<n$, and the true signal lies in the range of a network $g_\theta(u)$ that is 1-Lipschitz in $\theta$, satisfies $g_0(u)=0$, and has parameters in a ball of radius $x_{\max}\sqrt{n/k}$, the maximum-likelihood estimate $\hat{x}$ obeys $\frac{1}{n}\|\hat{x}-x_o\|_2^2 \le C_1\left[\frac{n}{m^2}\frac{k\log n}{L}+\frac{\sqrt{k\log n}}{m}\right]$ with probability at least $1-C_2(e^{-m/2}+e^{-Ln/8}+e^{-C_3 k\log n}+e^{k\log n-n/2})$. The first term decays like $1/L$ and reflects the parametric gain from averaging independent looks, while the second term does not decay with $L$, matching the intuition that a fixed measurement matrix provides only $m(m+1)/2$ informative quadratic measurements of the signal. The proof avoids the expected-likelihood step that loosened earlier bounds and instead controls the fluctuation of the empirical likelihood uniformly over a fine covering of the parameter set.

Load-bearing premise

The bound assumes the true image is produced exactly by an untrained neural network whose parameters change the image by at most the same amount and lie in a ball of radius proportional to $\sqrt{n/k}$, and that the optimizer finds the true global minimum of a non-convex likelihood.

Editorial extensions

If this is right

  • As $L\to\infty$ the first error term vanishes, leaving a floor of order $\sqrt{k\log n}/m$ that no amount of additional looks can remove for a fixed undersampled sensing matrix.
  • For fixed $m,n,L$, richer image classes with larger parameter count $k$ receive a larger error bound, matching the intuition that more complex image sets are harder to distinguish from speckle noise.
  • In the single-look case $L=1$, the bound is tighter than previous speckle compressed-sensing bounds in the regime $k\log n/m$ small.
  • The predicted scaling gives a concrete sample-size rule: the first error term becomes negligible when the number of looks satisfies $L\gg n k\log n/m^2$.
  • Algorithmically, the paper shows empirically that a single Newton-Schulz step per gradient iteration maintains the performance of exact matrix inversion while cutting the per-iteration computational cost substantially.

Reading between the lines

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

  • A testable extension the paper does not pursue is whether the same two-term MSE decomposition appears in other quadratic forward models, suggesting that any fixed measurement geometry imposes a floor of order $\sqrt{k\log n}/m$ on recovery from multiplicative noise.
  • Because the theorem concerns the global minimizer, the bagging heuristic can be read as an attempt to make local optima behave like the global one; one could test whether the gap between the projected gradient output and the theoretical bound shrinks as the number of bagged projections grows.
  • The bound's growth with $k$ offers a model-selection principle: choose the smallest network parameter count that still represents the image class, balancing approximation quality against the error penalty in the guarantee.
  • If real images are only approximately in the network's range, the guarantee would need an additional approximation-error term; quantifying that bias would make the theorem directly applicable to natural image datasets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies multilook coherent imaging under the model y_l = A X_o w_l with speckle noise, in the undersampled regime m < n, and with a deep image prior (DIP) on the unknown image. Its main theoretical claim, Theorem III.1, is a high-probability bound on the normalized mean squared error of the maximum likelihood estimator: under Lipschitzness of g_theta(u) in theta and a parameter-ball assumption, the paper asserts (1/n)||xhat - x_o||_2^2 <= C1( n/m^2 * k log n / L + sqrt(k log n)/m ). The proof uses a comparison between the empirical and expected log-likelihood, a delta-net over the DIP parameter space, concentration inequalities, and algebraic bounds relating the likelihood gap to Hilbert-Schmidt norms of A(X^2 - X_o^2)A^T. On the algorithmic side, the paper proposes projected gradient descent with Newton-Schulz approximate matrix inversion and a Bagged-DIP projection step, and reports experiments on eight 128x128 images at sampling ratios m/n = 0.125, 0.25, 0.5 and numbers of looks L = 1,...,128.

Significance. If Theorem III.1 were established in the full regime stated, it would be the first MSE control for DIP-based maximum likelihood estimation in undersampled multilook coherent imaging, and its explicit dependence on k, L, m, and n would be a valuable contribution. The L-to-infinity floor sqrt(k log n)/m and the term n k log n / (L m^2) give concrete, falsifiable predictions about the benefit of additional looks. The paper also ships code and a reproducible experimental protocol, and the proof is detailed, reusing published lemmas from [11] rather than fitting constants to the conclusion. However, the proof currently supports only a restricted sampling regime, and the DIP parameter-ball assumption is stated inconsistently, so the central claim needs substantial revision before the result can be taken as stated.

major comments (3)
  1. [Section III-D, Eq. (16)-(17)] The theorem states m < n, but the proof's data-independent bound on lambda_max is explicitly derived only for m/n <= 1/5. For 1/5 < m/n < 1, the ratio (sqrt(n)+2sqrt(m))^2/(sqrt(n)-2sqrt(m))^2 is not uniformly bounded, and for m/n >= 1/4 the denominator is non-positive, so the event E4 in (16) provides no positive lower bound on sigma_min(A). Since Lemma VI.7 then yields factors lambda_min^{-2}(AA^T) lambda_max^{-4}(AA^T) that can grow with n, no n-independent constant C1 follows in this regime. The theorem should either be restricted to m/n <= 1/5 or the proof must be extended; as written, the central guarantee is not proved for the full stated range m < n.
  2. [Section III-A, parameter-ball definition] The DIP parameter-ball radius is defined inconsistently: the text first sets r_Theta = x_max sqrt(n), then concludes Theta = B_k(0, x_max sqrt(n/k)), and Theorem III.1 uses the latter. With Lipschitz constant 1 and g_0(u)=0, the radius x_max sqrt(n/k) implies ||g_theta(u)||_2 <= x_max sqrt(n/k) for every theta in the ball, while the pixel assumptions on x_o imply ||x_o||_2 >= x_min sqrt(n). For k > (x_max/x_min)^2 these two inequalities cannot both hold, so the stated DIP hypothesis is infeasible for such k unless one of the assumptions is relaxed. Please clarify the intended radius and state the feasibility condition relating k, x_min, and x_max.
  3. [Section III-D, Step 2 after Eq. (19)] The proof states that conditioned on the event E4 in (16), one has sigma_max(A) <= sqrt(n)+sqrt(m) and sigma_min(A) >= sqrt(n)-sqrt(m). However, the event written in (16) has sigma_max(A) <= sqrt(n)+2sqrt(m) and sigma_min(A) >= sqrt(n)-2sqrt(m). The stronger bounds with sqrt(m) are not consequences of E4 as defined, and applying Lemma VI.2 with t=0 gives no probability guarantee. This is a gap in the derivation of the constant c used in the choice of R in (20); the event or the constants need to be corrected.
minor comments (5)
  1. [Theorem III.1] The probability expression contains the term e^{k log n - n/2}; for this to be a valid high-probability statement, the manuscript should state the implicit condition k log n <= c n for some constant c, otherwise the displayed lower bound on the probability can be meaningless.
  2. [Abstract and introduction] The theorem is stated and proved for a real-valued, noiseless-look model with sigma_z = 0. The abstract and introduction should carry this qualification explicitly rather than describing the result as applying to the complex-valued coherent imaging model of Eq. (1).
  3. [Appendix A] The appendix title contains a typo: 'Caculation of the likelihood function' should be 'Calculation of the likelihood function'.
  4. [Remark III.5 and Section V-B] There are minor typos in the text: 'matirx' in Remark III.5 and 'simulaion' in Section V-B should be corrected.
  5. [Section III-A] The sentence 'This gives a lower bound for r_Theta' is misleading: the argument gives an upper bound on the required radius only if one knows the theta that represents x_o, and it does not justify setting r_Theta equal to x_max sqrt(n). The derivation should be rewritten to state the assumption directly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem III.1 is a genuine derivation from stated assumptions; the self-cited lemmas are independent published support.

full rationale

Theorem III.1 is derived from the stated DIP and Lipschitz assumptions through a chain of concentration inequalities and linear-algebra bounds; no step fits a parameter to data and then renames the fit as a prediction. The proof reuses Lemmas VI.6 and VI.7 from the authors' earlier paper [11], but these are published, parameter-free deterministic matrix inequalities whose statements do not include the MSE bound; under the review rules they qualify as independent support and do not raise the circularity score. Lemma VI.11 is proved in the paper from decoupling and Hanson-Wright concentration, and the remaining proof steps are algebraic manipulation of the resulting bounds. The algorithmic contributions (Bagged-DIP and Newton-Schulz) are empirical and are compared against prior work; they are not inputs to the theoretical claim. One non-circular limitation should be noted: in equation (17) the data-independent bound on lambda_max is asserted only for m/n <= 1/5, and Step 2 requires n >= 2m, so the proof as written establishes the claimed bound only in that undersampling regime rather than for the full stated range m < n. That is a correctness/completeness gap, not a circularity, and therefore does not change the circularity score.

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

The central theorem is conditional on the DIP hypothesis, a Lipschitz parameterization, a noiseless real model, Gaussian sensing, and an exact global minimizer; the proof also requires an m/n constraint absent from the theorem. Empirical claims depend on several hand-set hyperparameters. No new physical entities are introduced.

free parameters (5)
  • DIP parameter-ball radius r_Theta = x_max sqrt(n/k), with text also writing x_max sqrt(n)
    Chosen to relate the Lipschitz bound on g_theta to image bounds; the two radius statements in Section III-A are inconsistent, and the covering arguments depend on this radius.
  • Universal constants C1, C2, C3, c, c2, c3 = unspecified positive constants
    Theorem III.1 is non-quantitative: the bound is up to constants chosen in the proof, such as c2 set to be suitably large, so the claimed rates cannot be evaluated numerically without these values.
  • PGD learning rate mu = 0.001 for L<=8, 0.01 otherwise
    Set by hand in Section V-B; the performance of the algorithm depends on it.
  • Newton-Schulz trigger threshold delta_x = 0.12
    Chosen from a small simulation in Appendix C, Table IV, so that exact inverses are used only until the iterate stops changing; it affects stability and speed.
  • Bagged-DIP patch sizes and training iterations = patch sizes 32, 64, 128; iterations in Table III
    Three patch scales and their DIP training budgets are manual choices; the reported bagging gain is specific to these settings.
assumptions (5)
  • domain assumption DIP hypothesis: natural images can be embedded in the range of untrained networks with substantially fewer parameters than pixels.
    Defines the set C; cited from [9,10], not proven, and the theorem is conditional on it.
  • ad hoc to paper g_theta is Lipschitz in theta with constant 1, and g_0(u)=0.
    Introduced in Section III-A; not guaranteed by the DIP hypothesis and needed for the covering arguments.
  • domain assumption The true image xo is real-valued, 0 < x_min <= x_o,i <= x_max, and the additive noise sigma_z is 0.
    Section III-A restricts theory to the noiseless real case; the abstract model includes additive noise and complex phases.
  • ad hoc to paper A has iid N(0,1) entries and the proof requires m/n <= 1/5 even though Theorem III.1 states m<n.
    The singular-value event E4 and inequality (17) only control the relevant quantities when m/n <= 1/5; the theorem statement does not include this condition.
  • domain assumption xo lies in C and the global minimizer of f_L over C is attained by xhat.
    Used in the inequality f(Sigmahat) <= f(Sigma_o) at Eq. (9); the non-convex optimization is not guaranteed to be solved exactly by the proposed PGD.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multilook Coherent Imaging: Theoretical Guarantees and Algorithms." pith.science (2026). https://pith.science/paper/V3FS5LOM

@misc{pith2026250523594,
  author       = {Pith},
  title        = {Pith review of: Multilook Coherent Imaging: Theoretical Guarantees and Algorithms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V3FS5LOM}},
  note         = {Machine review of arXiv:2505.23594}
}
read the original abstract

Multilook coherent imaging is a widely used technique in applications such as digital holography, ultrasound imaging, and synthetic aperture radar. A central challenge in these systems is the presence of multiplicative noise, commonly known as speckle, which degrades image quality. Despite the widespread use of coherent imaging systems, their theoretical foundations remain relatively underexplored. In this paper, we study both the theoretical and algorithmic aspects of likelihood-based approaches for multilook coherent imaging, providing a rigorous framework for analysis and method development. Our theoretical contributions include establishing the first theoretical upper bound on the Mean Squared Error (MSE) of the maximum likelihood estimator under the deep image prior hypothesis. Our results capture the dependence of MSE on the number of parameters in the deep image prior, the number of looks, the signal dimension, and the number of measurements per look. On the algorithmic side, we employ projected gradient descent (PGD) as an efficient method for computing the maximum likelihood solution. Furthermore, we introduce two key ideas to enhance the practical performance of PGD. First, we incorporate the Newton-Schulz algorithm to compute matrix inverses within the PGD iterations, significantly reducing computational complexity. Second, we develop a bagging strategy to mitigate projection errors introduced during PGD updates. We demonstrate that combining these techniques with PGD yields state-of-the-art performance. Our code is available at https://github.com/Computational-Imaging-RU/Bagged-DIP-Speckle.

Figures

Figures reproduced from arXiv: 2505.23594 by the authors.

Figure 1
Figure 1. PSNR (averaged over 8 images) versus iteration count is depicted for four DIP models fitted to both clean (left panel) and noisy [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Newton-Schulz approximation compared with computing exact inverse for all interations, the rest of the curves correspond to stopping [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. (Left) We compare a Bagged-DIP with three sophisticated DIP estimates, where [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: As is clear from the figure, each DIP block is composed of the following components: • Up sample: This unit increases the height and width of the datacube that receives by a factor of 2. To interpolate the missing elements, it uses the simple bilinear interpolation. He…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Minimax Theory of Likelihood-Based Deep Learning for Speckle Regression

    math.ST 2026-07 conditional novelty 6.0 of 10

    Deep ReLU maximum-likelihood estimators are near-minimax optimal for speckle regression, achieving the same rates as additive-noise regression up to log factors.

Reference graph

Works this paper leans on

54 extracted references · 47 canonical work pages · cited by 1 Pith paper

  1. [11]

    Compressed sensing in the presence of speckle noise,

    W. Zhou, S. Jalali, and A. Maleki, “Compressed sensing in the presence of speckle noise,”IEEE Transactions on Information Theory, vol. 68, no. 10, pp. 6964–6980, 2022

  2. [1]

    Compressed sensing,

    D. L. Donoho, “Compressed sensing,”IEEE Transactions on information theory, vol. 52, no. 4, pp. 1289–1306, 2006

  3. [2]

    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

  4. [3]

    Compressive phase retrieval,

    M. L. Moravec, J. K. Romberg, and R. G. Baraniuk, “Compressive phase retrieval,” inWavelets XII, vol. 6701, pp. 712–722, SPIE, 2007

  5. [4]

    Minimum complexity pursuit for universal compressed sensing,

    S. Jalali, A. Maleki, and R. G. Baraniuk, “Minimum complexity pursuit for universal compressed sensing,” vol. 60, pp. 2253–2268, Apr. 2014

  6. [5]

    Compressive phase retrieval via generalized approximate message passing,

    P. Schniter and S. Rangan, “Compressive phase retrieval via generalized approximate message passing,”IEEE Transactions on Signal Processing, vol. 63, no. 4, pp. 1043–1055, 2014

  7. [6]

    Using black-box compression algorithms for phase retrieval,

    M. Bakhshizadeh, A. Maleki, and S. Jalali, “Using black-box compression algorithms for phase retrieval,”IEEE Transactions on Information Theory, vol. 66, no. 12, pp. 7978–8001, 2020

  8. [7]

    A tutorial on speckle reduction in synthetic aperture radar images,

    F. Argenti, A. Lapini, T. Bianchi, and L. Alparone, “A tutorial on speckle reduction in synthetic aperture radar images,”IEEE Geoscience and remote sensing magazine, vol. 1, no. 3, pp. 6–35, 2013

Show all 54 references
  1. [8]

    Experimental validation of model-based digital holographic imaging using multi-shot data,

    T. Bate, D. O’Keefe, M. F. Spencer, and C. J. Pellizzari, “Experimental validation of model-based digital holographic imaging using multi-shot data,” inUnconventional Imaging and Adaptive Optics 2022, vol. 12239, pp. 83–94, SPIE, 2022

  2. [9]

    Deep image prior,

    D. Ulyanov, A. Vedaldi, and V . Lempitsky, “Deep image prior,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 9446–9454, 2018. 24

  3. [10]

    Deep decoder: Concise image representations from untrained non-convolutional networks,

    R. Heckel and P. Hand, “Deep decoder: Concise image representations from untrained non-convolutional networks,” inInternational Conference on Learning Representations, 2018

  4. [12]

    C. L. Lawson and R. J. Hanson,Solving least squares problems. SIAM, 1995

  5. [13]

    Denoising and regularization via exploiting the structural bias of convolutional generators,

    R. Heckel and M. Soltanolkotabi, “Denoising and regularization via exploiting the structural bias of convolutional generators,” in International Conference on Learning Representations, 2019

  6. [14]

    Bagging predictors,

    L. Breiman, “Bagging predictors,”Machine learning, vol. 24, pp. 123–140, 1996

  7. [15]

    Iterative berechung der reziproken matrix,

    G. Schulz, “Iterative berechung der reziproken matrix,”ZAMM-Journal of Applied Mathematics and Mechanics/Zeitschrift f ¨ur Angewandte Mathematik und Mechanik, vol. 13, no. 1, pp. 57–59, 1933

  8. [16]

    Techniques for speckle noise removal,

    J. S. Lim and H. Nawab, “Techniques for speckle noise removal,” inApplications of speckle phenomena, vol. 243, pp. 35–45, SPIE, 1980

  9. [17]

    Speckle filtering of sar images: a comparative study between complex-wavelet-based and standard filters,

    L. Gagnon and A. Jouan, “Speckle filtering of sar images: a comparative study between complex-wavelet-based and standard filters,” in Wavelet Applications in Signal and Image Processing V, vol. 3169, pp. 80–91, SPIE, 1997

  10. [18]

    Speckle denoising by variant nonlocal means methods,

    Y . Tounsi, M. Kumar, A. Nassim, F. Mendoza-Santoyo, and O. Matoba, “Speckle denoising by variant nonlocal means methods,”Applied optics, vol. 58, no. 26, pp. 7110–7120, 2019

  11. [19]

    Attention-guided cnn for image denoising,

    C. Tian, Y . Xu, Z. Li, W. Zuo, L. Fei, and H. Liu, “Attention-guided cnn for image denoising,”Neural Networks, vol. 124, pp. 117–129, 2020

  12. [20]

    Sunet: swin transformer unet for image denoising,

    C.-M. Fan, T.-J. Liu, and K.-H. Liu, “Sunet: swin transformer unet for image denoising,” in2022 IEEE International Symposium on Circuits and Systems (ISCAS), pp. 2333–2337, IEEE, 2022

  13. [21]

    Phase-error estimation and image reconstruction from digital-holography data using a bayesian framework,

    C. J. Pellizzari, M. F. Spencer, and C. A. Bouman, “Phase-error estimation and image reconstruction from digital-holography data using a bayesian framework,”JOSA A, vol. 34, no. 9, pp. 1659–1669, 2017

  14. [22]

    Optically coherent image reconstruction in the presence of phase errors using advanced-prior models,

    C. J. Pellizzari, M. F. Spencer, and C. A. Bouman, “Optically coherent image reconstruction in the presence of phase errors using advanced-prior models,” inLong-range imaging III, vol. 10650, pp. 68–82, SPIE, 2018

  15. [23]

    Coherent plug-and-play: digital holographic imaging through atmospheric turbulence using model-based iterative reconstruction and convolutional neural networks,

    C. J. Pellizzari, M. F. Spencer, and C. A. Bouman, “Coherent plug-and-play: digital holographic imaging through atmospheric turbulence using model-based iterative reconstruction and convolutional neural networks,”IEEE Transactions on Computational Imaging, vol. 6, pp. 1607–1621, 2020

  16. [24]

    Solving coherent-imaging inverse problems using deep neural networks: an experimental demonstration,

    C. J. Pellizzari, T. J. Bate, K. P. Donnelly, and M. F. Spencer, “Solving coherent-imaging inverse problems using deep neural networks: an experimental demonstration,” inUnconventional Imaging and Adaptive Optics 2022, vol. 12239, pp. 57–65, SPIE, 2022

  17. [25]

    Multilook compressive sensing in the presence of speckle noise,

    X. Chen, Z. Hou, C. Metzler, A. Maleki, and S. Jalali, “Multilook compressive sensing in the presence of speckle noise,” inNeurIPS 2023 Workshop on Deep Learning and Inverse Problems, 2023

  18. [26]

    Algorithmic guarantees for inverse imaging with untrained network priors,

    G. Jagatap and C. Hegde, “Algorithmic guarantees for inverse imaging with untrained network priors,”Advances in neural information processing systems, vol. 32, 2019

  19. [27]

    Deep learning techniques for inverse problems in imaging,

    G. Ongie, A. Jalal, C. A. Metzler, R. G. Baraniuk, A. G. Dimakis, and R. Willett, “Deep learning techniques for inverse problems in imaging,”IEEE Journal on Selected Areas in Information Theory, vol. 1, no. 1, pp. 39–56, 2020

  20. [28]

    Accelerated mri with un-trained neural networks,

    M. Z. Darestani and R. Heckel, “Accelerated mri with un-trained neural networks,”IEEE Transactions on Computational Imaging, vol. 7, pp. 724–733, 2021

  21. [29]

    One-dimensional deep image prior for time series inverse problems,

    S. Ravula and A. G. Dimakis, “One-dimensional deep image prior for time series inverse problems,” in2022 56th Asilomar Conference on Signals, Systems, and Computers, pp. 1005–1009, IEEE, 2022

  22. [30]

    Practical phase retrieval using double deep image priors,

    Z. Zhuang, D. Yang, F. Hofmann, D. Barmherzig, and J. Sun, “Practical phase retrieval using double deep image priors,”arXiv preprint arXiv:2211.00799, 2022

  23. [31]

    Blind image deblurring with unknown kernel size and substantial noise,

    Z. Zhuang, T. Li, H. Wang, and J. Sun, “Blind image deblurring with unknown kernel size and substantial noise,”International Journal of Computer Vision, pp. 1–30, 2023

  24. [32]

    Deepred: Deep image prior powered by red,

    G. Mataev, P. Milanfar, and M. Elad, “Deepred: Deep image prior powered by red,” inProceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pp. 0–0, 2019

  25. [33]

    Compressed sensing with deep image prior and learned regularization,

    D. Van Veen, A. Jalal, M. Soltanolkotabi, E. Price, S. Vishwanath, and A. G. Dimakis, “Compressed sensing with deep image prior and learned regularization,”arXiv preprint arXiv:1806.06438, 2018

  26. [34]

    Plug-and-play image restoration with deep denoiser prior,

    K. Zhang, Y . Li, W. Zuo, L. Zhang, L. Van Gool, and R. Timofte, “Plug-and-play image restoration with deep denoiser prior,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 10, pp. 6360–6376, 2021

  27. [35]

    A plug-and-play deep image prior,

    Z. Sun, F. Latorre, T. Sanchez, and V . Cevher, “A plug-and-play deep image prior,” inICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 8103–8107, IEEE, 2021

  28. [36]

    Deep random projector: Accelerated deep image prior,

    T. Li, H. Wang, Z. Zhuang, and J. Sun, “Deep random projector: Accelerated deep image prior,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 18176–18185, 2023

  29. [37]

    Introduction to compressed sensing.,

    M. A. Davenport, M. F. Duarte, Y . C. Eldar, and G. Kutyniok, “Introduction to compressed sensing.,” 2012

  30. [38]

    Compressed sensing using generative models,

    A. Bora, A. Jalal, E. Price, and A. G. Dimakis, “Compressed sensing using generative models,” inInternational conference on machine learning, pp. 537–546, PMLR, 2017

  31. [39]

    Solving inverse problems via auto-encoders,

    P. Peng, S. Jalali, and X. Yuan, “Solving inverse problems via auto-encoders,”IEEE Journal on Selected Areas in Information Theory, vol. 1, no. 1, pp. 312–323, 2020

  32. [40]

    Plugin-cs: A simple algorithm for compressive sensing with generative prior,

    B. Joshi, X. Li, Y . Plan, and O. Yilmaz, “Plugin-cs: A simple algorithm for compressive sensing with generative prior,” inNeurIPS 2021 Workshop on Deep Learning and Inverse Problems, 2021

  33. [41]

    Provable compressed sensing with generative priors via langevin dynamics,

    T. V . Nguyen, G. Jagatap, and C. Hegde, “Provable compressed sensing with generative priors via langevin dynamics,”IEEE Transactions on Information Theory, vol. 68, no. 11, pp. 7410–7422, 2022

  34. [42]

    Is speckle noise more challenging to mitigate than additive noise?,

    R. Malekian and A. Maleki, “Is speckle noise more challenging to mitigate than additive noise?,”arXiv preprint arXiv:2409.16585, 2024. 25

  35. [43]

    Correction to

    W. Zhou, S. Jalali, and A. Maleki, “Correction to” compressed sensing in the presence of speckle noise”,”IEEE Transactions on Information Theory, 2024

  36. [44]

    Simultaneous analysis of lasso and dantzig selector,

    P. J. Bickel, Y . Ritov, and A. B. Tsybakov, “Simultaneous analysis of lasso and dantzig selector,” 2009

  37. [45]

    How well can we estimate a sparse vector?,

    E. J. Candes and M. A. Davenport, “How well can we estimate a sparse vector?,”Applied and Computational Harmonic Analysis, vol. 34, no. 2, pp. 317–323, 2013

  38. [46]

    Shalev-Shwartz and S

    S. Shalev-Shwartz and S. Ben-David,Understanding machine learning: From theory to algorithms. Cambridge university press, 2014

  39. [47]

    J. B. Conway,A course in functional analysis, vol. 96. Springer, 2019

  40. [48]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”arXiv preprint arXiv:1412.6980, 2014

  41. [49]

    Randomized quasi-newton updates are linearly convergent matrix inversion algorithms,

    R. M. Gower and P. Richt ´arik, “Randomized quasi-newton updates are linearly convergent matrix inversion algorithms,”SIAM Journal on Matrix Analysis and Applications, vol. 38, no. 4, pp. 1380–1409, 2017

  42. [50]

    Convergence rate improvement of richardson and newton-schulz iterations,

    A. Stotsky, “Convergence rate improvement of richardson and newton-schulz iterations,”arXiv preprint arXiv:2008.11480, 2020

  43. [51]

    Reconnet: Non-iterative reconstruction of images from compressively sensed measurements,

    K. Kulkarni, S. Lohit, P. Turaga, R. Kerviche, and A. Ashok, “Reconnet: Non-iterative reconstruction of images from compressively sensed measurements,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 449–458, 2016

  44. [52]

    Non-asymptotic theory of random matrices: extreme singular values,

    M. Rudelson and R. Vershynin, “Non-asymptotic theory of random matrices: extreme singular values,” inProceedings of the International Congress of Mathematicians 2010 (ICM 2010) (In 4 Volumes) Vol. I: Plenary Lectures and Ceremonies Vols. II–IV: Invited Lectures, pp. 1576–1602,...

  45. [53]

    Minimum complexity pursuit for universal compressed sensing,

    S. Jalali, A. Maleki, and R. G. Baraniuk, “Minimum complexity pursuit for universal compressed sensing,”IEEE Transactions on Information Theory, vol. 60, no. 4, pp. 2253–2268, 2014

  46. [54]

    Cameraman

    V . De la Pena and E. Gin ´e,Decoupling: from dependence to independence. Springer Science & Business Media, 2012. 26 APPENDIX A. Caculation of the likelihood function The aim of this section is to derive the loglikelihood for our model, yℓ =AXw ℓ +z ℓ,forℓ= 1, . . . , L, wher...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.