Pith. sign in

REVIEW 3 major objections 4 minor 121 references

Plug-and-play inverse problems require physics-aware denoisers for guaranteed recovery.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 05:16 UTC pith:NTCOZ4AG

load-bearing objection Solid, useful theory paper on PnP with MMSE denoisers, but the advertised recovery guarantees require a contraction condition that is much more restrictive than the "mild assumptions" language suggests. the 3 major comments →

arxiv 2607.29525 v1 pith:NTCOZ4AG submitted 2026-07-31 math.OC cs.ITmath.IT

Physics Matters in PnP: Recovery Guarantees with the MMSE and NN Denoisers

Tobias Wolf , Jalal Fadili , Jin Guo , Roy Y. He This is my paper
classification math.OC cs.ITmath.IT MSC 65J2262F1568T07
keywords plug-and-playMMSE denoiserforward-backward splittingrecovery guaranteesWasserstein distanceinverse problemsneural networksnoise covariance matching
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper proves that in the forward-backward splitting version of Plug-and-Play (PnP), the denoiser must be designed in tandem with the forward model, not chosen independently. The authors analyze an iteration where the standard gradient step is replaced by a flexible operator B, and the denoiser is an MMSE estimator specialized to Gaussian noise with possibly degenerate, non-diagonal covariance. They derive pointwise and Wasserstein recovery bounds that depend on how closely the denoiser's training noise matches the noise transformed by B. The findings imply that conventional choices like B=A^T tacitly assume the denoiser is agnostic to the physics, which harms recovery; instead, the covariance of the training noise should be set to match γ²BΣ_E B^T.

Core claim

The central claim is that recovery guarantees for PnP hold only when the denoiser's noise model is compatible with the forward operator: the covariance of the MMSE denoiser's training noise should equal the covariance of the observation noise after transformation by B. This makes the Bures metric between the two Gaussian noises vanish, leaving a residual dominated by the operator norm of B and the noise level. The paper also establishes that MMSE denoisers are cocoercive (hence firmly nonexpansive up to scaling), which justifies a common constraint in training.

What carries the argument

The iteration x_{k+1} = D(x_k + γB(y - A x_k)) with MMSE denoiser D. The load-bearing quantity is the Lipschitz constant of D, bounded by M²/λmin(ΣΥ), where M bounds the signal and ΣΥ is the denoiser's noise covariance. The condition ΣΥ = γ²BΣ_EB^T makes the noise-mismatch term vanish, and the contraction factor q = (M²/λmin(ΣΥ)) sqrt(1 - κ(BA)^{-2}) governs all bounds.

Load-bearing premise

The error bounds only contract if the contraction factor q = (M²/λmin(ΣΥ)) sqrt(1 - κ(BA)^{-2}) is less than 1, which requires either a large denoiser noise covariance relative to the signal bound, or a nearly isotropic preconditioned operator BA.

What would settle it

Compute q for a simple ill-conditioned inverse problem (e.g., a diagonal A with condition number 100, B = A^T, M fixed) and a denoiser trained with isotropic noise of variance σ². For σ² small enough that M²/σ² > 1 and κ(BA) large, q exceeds 1, so the pointwise bound (23) does not contract; run the iteration and show the error does not converge to the bound, contradicting the theorem's hypotheses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the denoiser's training noise is chosen to match the forward model, PnP-FBS has a pointwise recovery error bounded by terms involving noise amplification, denoiser bias, and sampling error.
  • In Wasserstein distance, the optimal choice ΣΥ = γ²BΣ_EB^T yields a residual of order (1-q)^{-1} γ ||B|| sqrt(λmax(Σ_E)) sqrt(r).
  • MMSE denoisers are cocoercive, providing a theoretical justification for firm nonexpansiveness in PnP.
  • Results extend to neural network denoisers with width and depth bounds depending on the manifold dimension of the signal support.
  • The conventional choice B = A^T implicitly assumes the denoiser is physics-agnostic, which is suboptimal.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The contraction condition q<1 may be stringent for ill-conditioned problems; for a fixed conditioning, the signal-to-noise ratio of the training data must be low enough to maintain stability. This suggests that for highly ill-conditioned forward operators, the denoiser's training noise must be large, potentially limiting resolution.
  • The framework suggests a design principle for PnP: choose B as a preconditioner that balances the conditioning of BA and the noise matching, rather than using B=A^T.
  • One could test the bounds empirically by training denoisers with covariance matched to BΣ_EB^T and comparing recovery quality against physics-agnostic denoisers, especially for degenerate noise (e.g., missing cone in tomography).
  • The cocoercivity result may be used to regularize neural-network training to enforce convergence guarantees in practice.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper studies the forward-backward splitting (FBS) variant of Plug-and-Play (PnP) in which the proximal operator is replaced by an MMSE denoiser designed for zero-mean Gaussian noise with possibly degenerate covariance Σ_Υ, and the gradient of the data fidelity is preconditioned by a matrix B. The main theoretical results are: (i) regularity properties of the MMSE denoiser, including a Jacobian formula, a Lipschitz bound L ≤ M²/λ_min(Σ_Υ), monotonicity, cocoercivity (for isotropic noise), and stability with respect to the prior measure; (ii) pointwise and Wasserstein recovery bounds for the PnP-FBS iteration, stating that if either L≤1 or a near-isotropic condition on BA holds, then the error is controlled by a contraction factor q plus a term involving the Bures metric between γBE and Υ; (iii) an extension to neural-network denoisers with approximation guarantees. The paper concludes that the denoiser cannot be chosen independently of the forward model.

Significance. If correct, the paper provides a rigorous framework for designing physics-aware denoisers in PnP and identifies the Bures-metric mismatch as the key quantity controlling distributional recovery. The proofs are detailed and largely self-contained; the cocoercivity result for MMSE estimators (Proposition 4.10) and the derivative identity (40) are useful contributions. The Wasserstein analysis with the explicit Bures-metric term (Remark 3.7) is a genuine step forward. However, the advertised 'mild assumptions' and the claim that nonexpansiveness is not required are substantially overstated: the contraction conditions (20)/(21) essentially reduce to nonexpansiveness or to BA being near the identity, which excludes typical ill-conditioned inverse problems. This does not invalidate the mathematics, but it materially narrows the scope of the headline claim.

major comments (3)
  1. [§3.2, Theorems 3.2 and 3.6, conditions (20)-(21)] The factor q = (M²/λ_min(Σ_Υ))√(1−κ(BA)^{-2}) must be <1 for the bounds (23) and (30) to be meaningful. Since Lemma 4.3 gives L := M²/λ_min(Σ_Υ) as a Lipschitz bound for the MMSE denoiser, condition (20) is exactly L≤1, i.e. nonexpansiveness. Condition (21), equivalently κ(BA) < (1−L^{-2})^{-1/2}, permits L>1 only when κ(BA) is very close to 1 (for L=2, κ<1.15; for L=1.1, κ<2.4). Thus the 'non-contractive' regime is essentially the near-isotropic regime. For the ill-conditioned, non-injective forward operators targeted in the introduction, neither condition will hold in general, and the bounds are vacuous because (1−q)^{-1} is undefined when q≥1. The abstract and Remark 3.3(i) present these as 'mild assumptions'; this is misleading. Please revise the claims and state explicitly the restricted regime in which the guarantees apply.
  2. [Lemma 5.1 and Eq. (23)-(24)] The pointwise recovery bound contains the term 2M(N−1)exp(−ε/(2λ_max(Σ_Υ))). Here ε is the separation gap between the squared distances of the nearest and second-nearest samples; for a typical continuous distribution this gap tends to 0 as N increases (often at rate 1/N), so the term need not vanish and may even grow with N. The phrase 'there exists ε>0' is not quantitative. The modification in Remark 5.2 replaces it with an extra nonvanishing √(λ_max(Σ_Υ))ε term. Hence Theorem 3.2 does not demonstrate consistency of the pointwise error as N grows. Please provide a quantitative treatment of ε or state the bound with an explicit, controlled extra term and discuss the resulting asymptotics.
  3. [Theorem 3.11, Eq. (36)-(37)] The neural-network recovery bound contains a √N ε term, where ε is the uniform approximation error of the identity on the support manifold S. This term grows with the sample size N, so the bound degrades as more training data are used, unless ε is chosen to decay with N. But then, by Lemma 3.10, the required network width scales like ε^{-d_S}, which is polynomial in N (or worse). The current statement does not make this trade-off explicit and gives the misleading impression that the NN bound is a convergence-in-N result. Please optimize ε as a function of N and state the resulting rate, or clarify the non-asymptotic nature of the bound.
minor comments (4)
  1. [Abstract] There are missing spaces in the typeset text ('Wefurtherdeviate', 'thatthedenoiser'), likely a formatting issue; please fix.
  2. [Remark 3.3(i)] The sentence 'This condition imposes that BA cannot be too ill-conditioned' refers to condition (20), which controls M²/λ_min(Σ_Υ) rather than the conditioning of BA. Reword to avoid confusion.
  3. [Lemma 3.1] The proof states that T maps co(supp(μ_X)) 'onto' itself; by Brouwer it suffices that it maps into a compact convex set. Please correct 'onto' to 'into'.
  4. [Section 3.2.2] The notation for seκ(BA), seq, and seγ is typographically heavy and easy to confuse with the un-tilded versions. Consider renaming these quantities for readability.

Circularity Check

0 steps flagged

No circular derivation; minor non-load-bearing self-citation only.

full rationale

The derivation chain is self-contained. Theorem 3.2 is built from Lemma 4.3 (Jacobian identity giving the Lipschitz bound L ≤ M²/λ_min(eΣΥ)), Lemma 5.1 (denoiser bias bound), and an explicit contraction estimate with ρ(γ)² = 1 − 2γλ_min(BA;cone) + γ²λ_max(BA)². Conditions (20) and (21) are exactly the sufficient conditions q < 1; they are assumptions of the theorem, not conclusions smuggled in. When q ≥ 1 the bounds are vacuous, which is a limitation but not circularity. Theorem 3.6 repeats the argument in Wasserstein space; the Bures-metric term is a genuine computed distance W₂((γB)_#μ_E, μ_Υ), and the choice ΣΥ = γ²BΣ_EB^T is a deliberate design choice that zeroes this term, not a fitted parameter disguised as a prediction. The paper's 'physics matters' claim is a consequence of the explicit inequality: the denoiser's training noise Υ appears in the bound through ∥γBE − Υ∥, but this term is derived, not assumed as the conclusion. The only notable self-citation is [60] (with co-author Fadili) for the example B = A^TΣ_E^{-1/2}; this is used as a comparison/design example, while the theorems are stated for general B satisfying (A-5)–(A-7), so the citation is not load-bearing. External approximation results in the NN section are also used as genuine inputs. No circular step meets the evidentiary standard of the review rules.

Axiom & Free-Parameter Ledger

3 free parameters · 10 axioms · 0 invented entities

The paper adds no new physical entities. Its central claim rests on a small set of modeling assumptions (Gaussianity, compactness, support/range inclusions, conical injectivity) plus standard external theorems from convex analysis, optimal transport, and neural-network approximation. The free parameters are genuine design choices (B, Σ_Υ, ε), not fitted constants; the main hidden burden is the contraction condition q<1, which is the practical bottleneck.

free parameters (3)
  • Σ_Υ (denoiser noise covariance)
    User-chosen covariance of the Gaussian noise Υ used to define the MMSE denoiser. It is not fitted to data but is a free design input; λ_min(eΣΥ) enters every bound and the contraction condition (20).
  • B (preconditioning matrix)
    User-chosen n×m matrix in the iteration (6); must satisfy Im(B)⊂V, BA symmetric PSD and conical injectivity. The 'physics-aware' recommendation is to choose B and ΣΥ together; no data fitting.
  • ε (NN approximation tolerance)
    In Theorem 3.11/Lemma 3.10, ε is a user-chosen accuracy for approximating the identity on the signal support; it controls the network width W*_ε and appears as √N ε in the bound, creating a trade-off.
axioms (10)
  • domain assumption A-1: observation noise E is zero-mean Gaussian with positive definite covariance Σ_E.
    Used throughout; in particular for the exact Bures-metric formula for W2((γB)#μ_E, μ_Υ) in Theorem 3.6.
  • domain assumption A-2: signal support X is compact, with ∥x∥≤M for all x∈X.
    Compactness gives the Lipschitz bound, uniform constants, and fixed-point arguments; it excludes unbounded signals such as natural images without normalization.
  • domain assumption A-3: Υ is zero-mean Gaussian with possibly degenerate covariance Σ_Υ.
    Defines the MMSE denoiser's internal noise model; the degenerate/non-diagonal case is the paper's extension beyond i.i.d. Gaussian denoisers.
  • domain assumption A-4: co(supp μ_X) ⊂ Im(Σ_Υ).
    The denoiser noise must span the signal support. Strong restriction when Σ_Υ is low-rank; without it the MMSE denoiser's Lipschitz/stability analysis breaks.
  • domain assumption A-5: Im(B) ⊂ Im(Σ_Υ).
    Ensures iterates remain in the subspace V=Im(Σ_Υ); central to the 'physics-aware' design and to Lemma 3.1.
  • domain assumption A-6: BA is symmetric and positive semidefinite.
    Makes the descent part of the iteration a gradient step for a convex quadratic; excludes many reasonable preconditioners.
  • domain assumption A-7 / A-7': conical injectivity of BA on tangent cones of co(supp μ_X).
    Equivalent to λ_min(BA;cone)>0, needed for the contraction factor q<1 and for Lemma 3.4. Stronger A-7' is required for uniform Wasserstein bounds.
  • domain assumption A-8: supp(μ_X) is a d_S-dimensional manifold.
    Only needed for the NN approximation Lemma 3.10 and Theorem 3.11; not used in the MMSE-only results.
  • standard math External fixed-point, monotone-operator, and empirical-Wasserstein results (Brouwer/Kakutani, Minty, [11,114], [80]).
    Brouwer/Kakutani for existence of fixed points, Minty for maximal monotonicity in Prop 4.11, [11,114] for empirical Wasserstein rates, [80] for ReLU approximation of the identity on manifolds.
  • domain assumption For the high-probability bounds, μ_X has a density w.r.t. volume measure on V that is bounded away from zero.
    Used in Lemma 5.3 to obtain uniform N^{-1/r} covering rates; not required for the deterministic pointwise bound (23).

pith-pipeline@v1.3.0-daily-deepseek · 38243 in / 18296 out tokens · 182423 ms · 2026-08-03T05:16:42.083674+00:00 · methodology

0 comments
read the original abstract

We investigate the forward-backward-splitting version of the Plug and Play (PnP) method for linear ill-posed problems with MMSE estimators as denoisers. In contrast to existing literature, we consider estimators which are specialized for (degenerate) Gaussian noise with possibly non-diagonal covariance matrices. We further deviate from the classical iteration by replacing parts of the descent step with a linear operator that relates the observation noise to that of the MMSE estimator. Under mild assumptions, we derive several properties of the denoiser and prove recovery guarantees of the iteration both pointwise and in the Wasserstein distance of the underlying probability distributions. Crucially, our analysis shows that the denoiser cannot be chosen in a physics-agnostic way, that is, independently of the forward model. We extend our results to the case where the MMSE denoiser is parametrized by a neural network and derive the corresponding recovery bounds.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

121 extracted references · 8 linked inside Pith

  1. [1]

    InverseProblems, 33(12):124007, 2017

    JonasAdlerandOzanÖktem.Solvingill-posedinverseproblemsusingiterativedeepneuralnetworks. InverseProblems, 33(12):124007, 2017

    2017

  2. [2]

    Aliprantis and Kim C

    Charalambos D. Aliprantis and Kim C. Border. Infinite Dimensional Analysis. SpringerLink. Springer-Verlag, Berlin, Heidelberg, third edition edition, 2006. Literaturverz. S. 667 - 680

    2006

  3. [3]

    Solving inverse problems using data-driven models.ActaNumerica, 28:1–174, May 2019

    Simon Arridge, Peter Maass, Öktem Ozan, and Carola-Bibiane Schönlieb. Solving inverse problems using data-driven models.ActaNumerica, 28:1–174, May 2019

    2019

  4. [4]

    A variational approach to removing multiplicative noise

    Gilles Aubert and Jean-François Aujol. A variational approach to removing multiplicative noise. SIAMJournal on AppliedMathematics, 68(4):925–946, January 2008

    2008

  5. [5]

    DifferentialInclusions: Set-ValuedMapsandViabilityTheory

    Jean-PierreAubinandArrigoCellina. DifferentialInclusions: Set-ValuedMapsandViabilityTheory. Springer Berlin Heidelberg, 1984

    1984

  6. [6]

    ConvexAnalysisandMonotoneOperatorTheoryinHilbertSpaces

    H.BauschkeandP.L.Combettes. ConvexAnalysisandMonotoneOperatorTheoryinHilbertSpaces. Springer, 2nd edition, 2017

    2017

  7. [7]

    Society for Industrial and Applied Mathematics, October 2017

    Amir Beck.First-Order Methods in Optimization. Society for Industrial and Applied Mathematics, October 2017

    2017

  8. [8]

    Modern regularization methods for inverse problems.Acta numerica, 27:1–111, 2018

    Martin Benning and Martin Burger. Modern regularization methods for inverse problems.Acta numerica, 27:1–111, 2018

    2018

  9. [9]

    Deep unfolding of a proximal interior point method for image restoration.Inverse Problems, 36(3):034005, 2020

    Carla Bertocchi, Emilie Chouzenoux, Marie-Caroline Corbineau, Jean-Christophe Pesquet, and Marco Prato. Deep unfolding of a proximal interior point method for image restoration.Inverse Problems, 36(3):034005, 2020. 38

    2020

  10. [10]

    Ontikhonovfunctionals penalizedby bregmandistances

    IsmaelBleyer andAntonio Leitao. Ontikhonovfunctionals penalizedby bregmandistances. CUBO, 11:099–115, 12 2009

    2009

  11. [11]

    On the mean speed of convergence of empirical and oc- cupationmeasuresinWassersteindistance

    Emmanuel Boissard and Thibaut Le Gouic. On the mean speed of convergence of empirical and oc- cupationmeasuresinWassersteindistance. Annalesdel’I.H.P.Probabilitésetstatistiques,50(2):539– 563, 2014

    2014

  12. [12]

    Tweedie moment projected diffusions for inverse problems.arXiv preprint arXiv:2310.06721, 2023

    Benjamin Boys, Mark Girolami, Jakiw Pidstrigach, Sebastian Reich, Alan Mosca, and O Deniz Aky- ildiz. Tweedie moment projected diffusions for inverse problems.arXiv preprint arXiv:2310.06721, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    Learning firmly nonexpansive operators

    Kristian Bredies, Jonathan Chirinos-Rodriguez, and Emanuele Naldi. Learning firmly nonexpansive operators. arXiv, 2024

    2024

  14. [14]

    Total generalized variation.SIAM Journal on ImagingSciences, 3(3):492–526, January 2010

    Kristian Bredies, Karl Kunisch, and Thomas Pock. Total generalized variation.SIAM Journal on ImagingSciences, 3(3):492–526, January 2010

    2010

  15. [15]

    Buskulic, M

    N. Buskulic, M. J. Fadili, and Y. Quéau. Implicit regularization of the deep inverse prior trained with inertia. InAndreasHauptmann,BangtiJin,andCarola-BibianeSchönlieb,editors, MachineLearning Solutions for InverseProblems: Part A, volume 26 ofHandbook of Numerical Analysis, chapter 6. Elsevier, 2025

    2025

  16. [16]

    Convergence and recovery guarantees of unsuper- visedneuralnetworksforinverseproblems

    Nathan Buskulic, Jalal Fadili, and Yvain Quéau. Convergence and recovery guarantees of unsuper- visedneuralnetworksforinverseproblems. JournalofMathematicalImagingandVision,pages1–22, 2024

    2024

  17. [17]

    NathanBuskulic,JalalFadili,andYvainQuéau.Recoveryguaranteesofunsupervisedneuralnetworks for inverse problems trained with gradient descent.EUSIPCO, 2024

    2024

  18. [18]

    Convergence guarantees of overparametrized wide deepinverseprior

    Nathan Buskulic, Yvain Quéau, and Jalal Fadili. Convergence guarantees of overparametrized wide deepinverseprior. In International ConferenceonScaleSpaceandVariationalMethodsinComputer Vision, pages 406–417, 2023

    2023

  19. [19]

    Image recovery via total variation minimization and related problems.NumerischeMathematik, 76:167–188, 1997

    Antonin Chambolle and Pierre-Louis Lions. Image recovery via total variation minimization and related problems.NumerischeMathematik, 76:167–188, 1997

    1997

  20. [20]

    Diffusion posterior sampling for general noisy inverse problems.arXivpreprint arXiv:2209.14687, 2022

    Hyungjin Chung, Jeongsol Kim, Michael T Mccann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems.arXivpreprint arXiv:2209.14687, 2022

    Pith/arXiv arXiv 2022

  21. [21]

    Improving diffusion models for inverse problems using manifold constraints.Advances in Neural Information Processing Systems, 35:25683–25696, 2022

    Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints.Advances in Neural Information Processing Systems, 35:25683–25696, 2022

    2022

  22. [22]

    Consistent diffusion meets tweedie: Training exact ambient diffusion models with noisy data.arXivpreprint arXiv:2404.10177, 2024

    Giannis Daras, Alexandros G Dimakis, and Constantinos Daskalakis. Consistent diffusion meets tweedie: Training exact ambient diffusion models with noisy data.arXivpreprint arXiv:2404.10177, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Deep generative models and inverse problems.Mathematical Aspects of DeepLearning, 400, 2022

    Alexandros G Dimakis. Deep generative models and inverse problems.Mathematical Aspects of DeepLearning, 400, 2022. 39

    2022

  24. [24]

    D.L. Donoho. Compressed sensing. IEEE Transactions on Information Theory, 52(4):1289–1306, April 2006

    2006

  25. [25]

    Regularising inverse problems with generative machine learning models.Journal of MathematicalImagingandVision, 66(1):37–56, 2024

    Mag Duff, Neill Campbell, and Matthias J Ehrhardt. Regularising inverse problems with generative machine learning models.Journal of MathematicalImagingandVision, 66(1):37–56, 2024

    2024

  26. [26]

    Vincent Poor, and Shlomo Shamai Shitz

    Alex Dytso, H. Vincent Poor, and Shlomo Shamai Shitz. A general derivative identity for the condi- tionalmeanestimatoringaussiannoiseandsomeapplications.In 2020IEEEInternationalSymposium onInformationTheory (ISIT), pages 1183–1188, 2020

    2020

  27. [27]

    Plug-and-playimagereconstructionisaconvergentregulariza- tion method.IEEETransactionsonImageProcessing, 33:1476–1486, 2024

    AndreaEbnerandMarkusHaltmeier. Plug-and-playimagereconstructionisaconvergentregulariza- tion method.IEEETransactionsonImageProcessing, 33:1476–1486, 2024

    2024

  28. [28]

    ProximalLangevinsamplingwith inexact proximal mapping.SIAMJournal on ImagingSciences, 17(3):1729–1760, 2024

    MatthiasJEhrhardt,LorenzKuger,andCarola-BibianeSchönlieb. ProximalLangevinsamplingwith inexact proximal mapping.SIAMJournal on ImagingSciences, 17(3):1729–1760, 2024

    2024

  29. [29]

    Regularization of inverse problems.MathematicsanditsApplications, 375, 1996

    HW Engl. Regularization of inverse problems.MathematicsanditsApplications, 375, 1996

    1996

  30. [30]

    Esposito

    R. Esposito. On a relation between detection and estimation in decision theory.Information and Control, 12(2):116–120, 1968

    1968

  31. [31]

    Falconer

    K. Falconer. Fractal Geometry: Mathematical Foundations and Applications. John Wiley & Sons, 2nd edition, 2003

    2003

  32. [32]

    On a formula for thel2 wasserstein metric between measures on euclidean and hilbert spaces.MathematischeNachrichten, 147:185–203, 1990

    Matthias Gelbrich. On a formula for thel2 wasserstein metric between measures on euclidean and hilbert spaces.MathematischeNachrichten, 147:185–203, 1990

    1990

  33. [33]

    Stein’sunbiasedriskestimate and hyv\" arinen’s score matching.arXivpreprint arXiv:2502.20123, 2025

    SulagnaGhosh,NikolaosIgnatiadis,FredericKoehler,andAmberLee. Stein’sunbiasedriskestimate and hyv\" arinen’s score matching.arXivpreprint arXiv:2502.20123, 2025

    arXiv 2025

  34. [34]

    The split bregman method for l1-regularized problems.SIAM Journal on ImagingSciences, 2(2):323–343, January 2009

    Tom Goldstein and Stanley Osher. The split bregman method for l1-regularized problems.SIAM Journal on ImagingSciences, 2(2):323–343, January 2009

    2009

  35. [35]

    Learningfastapproximationsofsparsecoding

    KarolGregorandYannLeCun. Learningfastapproximationsofsparsecoding. In Proceedingsofthe 27thinternationalconferenceoninternationalconferenceonmachinelearning,pages399–406,2010

    2010

  36. [36]

    Shouldpenalizedleastsquaresregressionbeinterpretedasmaximumaposteriori estimation? IEEE TransactionsonSignal Processing, 59(5):2405–2410, May 2011

    RémiGribonval. Shouldpenalizedleastsquaresregressionbeinterpretedasmaximumaposteriori estimation? IEEE TransactionsonSignal Processing, 59(5):2405–2410, May 2011

    2011

  37. [37]

    priors" and

    Rémi Gribonval and Pierre Machart. Reconciling "priors" and "priors" without prejudice? EURASIPJournal onAdvancesinSignal Processing, 2013(1), 12 2013

    2013

  38. [38]

    Agem: Solvinglinearinverseproblemsviadeeppriors and sampling.Advancesin NeuralInformationProcessingSystems, 32, 2019

    BichuanGuo,YuxingHan,andJiangtaoWen. Agem: Solvinglinearinverseproblemsviadeeppriors and sampling.Advancesin NeuralInformationProcessingSystems, 32, 2019

    2019

  39. [39]

    Shamai, and S

    Dongning Guo, S. Shamai, and S. Verdu. Mutual information and minimum mean-square error in gaussian channels.IEEE TransactionsonInformationTheory, 51(4):1261–1282, 2005

    2005

  40. [40]

    Estimationingaussiannoise: Proper- tiesoftheminimummean-squareerror

    DongningGuo,YihongWu,ShlomoS.Shitz,andSergioVerdú. Estimationingaussiannoise: Proper- tiesoftheminimummean-squareerror. IEEETransactionsonInformationTheory,57(4):2371–2385, 2011. 40

    2011

  41. [41]

    Agenerativevariationalmodelforinverseproblemsinimaging

    AndreasHabringandMartinHoller. Agenerativevariationalmodelforinverseproblemsinimaging. SIAMJournal on Mathematicsof DataScience, 4(1):306–335, 2022

    2022

  42. [42]

    Universal bayes consistency in metric spaces.Ann.Statist., 49(4):2129–2150, August 2021

    Steve Hanneke, Aryeh Kontorovich, Sivan Sabato, and Roi Weiss. Universal bayes consistency in metric spaces.Ann.Statist., 49(4):2129–2150, August 2021

    2021

  43. [43]

    Hatsell and L

    C. Hatsell and L. Nolte. Some geometric properties of the likelihood ratio (corresp.). IEEE TransactionsonInformationTheory, 17(5):616–618, 1971

    1971

  44. [44]

    MachineLearningSolutions forInverseProblems: Part A, volume 26 ofHandbookof NumericalAnalysis

    AndreasHauptmann,BangtiJin,andCarola-BibianeSchönlieb,editors. MachineLearningSolutions forInverseProblems: Part A, volume 26 ofHandbookof NumericalAnalysis. Elsevier, 2025

    2025

  45. [45]

    Convergent regularizationininverseproblemsandlinearplug-and-playdenoisers

    AndreasHauptmann,SubhadipMukherjee,Carola-BibianeSchönlieb,andFerdiaSherry. Convergent regularizationininverseproblemsandlinearplug-and-playdenoisers. FoundationsofComputational Mathematics, 25(4):1087–1120, June 2025

    2025

  46. [46]

    Denoising diffusion probabilistic models.Advancesin neuralinformationprocessing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advancesin neuralinformationprocessing systems, 33:6840–6851, 2020

    2020

  47. [47]

    Convergentplug-and-play methods for image inverse problems with explicit and nonconvexdeep regularization

    Samuel Hurault. Convergentplug-and-play methods for image inverse problems with explicit and nonconvexdeep regularization. PhD thesis, Institut de Mathématiques de Bordeaux, 2023

    2023

  48. [48]

    GradientStepDenoiserforconvergentPlug- and-Play

    SamuelHurault,ArthurLeclaire,andNicolasPapadakis. GradientStepDenoiserforconvergentPlug- and-Play. arXiv:2110.03220[cs, eess,math], October 2021. arXiv: 2110.03220

    Pith/arXiv arXiv 2021

  49. [49]

    Proximaldenoiserforconvergentplug-and- play optimization with nonconvex regularization.arXivpreprint, 2022

    SamuelHurault,ArthurLeclaire,andNicolasPapadakis. Proximaldenoiserforconvergentplug-and- play optimization with nonconvex regularization.arXivpreprint, 2022

    2022

  50. [50]

    Estimationofnon-normalizedstatisticalmodelsbyscorematching

    AapoHyvärinenandPeterDayan. Estimationofnon-normalizedstatisticalmodelsbyscorematching. Journal of MachineLearning Research, 6(4), 2005

    2005

  51. [51]

    Deep convolutional neuralnetworkforinverseproblemsinimaging

    Kyong Hwan Jin, Michael T McCann, Emmanuel Froustey, and Michael Unser. Deep convolutional neuralnetworkforinverseproblemsinimaging. IEEEtransactionsonimageprocessing,26(9):4509– 4522, 2017

    2017

  52. [52]

    McCann, Emmanuel Froustey, and Michael Unser

    Kyong Hwan Jin, Michael T. McCann, Emmanuel Froustey, and Michael Unser. Deep convolutional neuralnetworkforinverseproblemsinimaging. IEEETransactionsonImageProcessing,26(9):4509– 4522, 2017

    2017

  53. [53]

    Stochastic mirror descent method for linear ill-posed problems in banach spaces.InverseProblems, 39(6):065010, May 2023

    Qinian Jin, Xiliang Lu, and Liuying Zhang. Stochastic mirror descent method for linear ill-posed problems in banach spaces.InverseProblems, 39(6):065010, May 2023

    2023

  54. [54]

    Stochastic solutions for linear inverse problems using the prior implicit in a denoiser.Advancesin Neural Information Processing Systems, 34:13242–13254, 2021

    Zahra Kadkhodaie and Eero Simoncelli. Stochastic solutions for linear inverse problems using the prior implicit in a denoiser.Advancesin Neural Information Processing Systems, 34:13242–13254, 2021

    2021

  55. [55]

    Plug-and-play methods for integrating physical and learned models in computational imaging: Theory, algorithms, and applications.IEEESignal ProcessingMagazine, 40(1):85–97, 2023

    Ulugbek S Kamilov, Charles A Bouman, Gregery T Buzzard, and Brendt Wohlberg. Plug-and-play methods for integrating physical and learned models in computational imaging: Theory, algorithms, and applications.IEEESignal ProcessingMagazine, 40(1):85–97, 2023. 41

    2023

  56. [56]

    Accurate image superresolution using very deep convolutional networks.2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1646–1654, 2016

    Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image superresolution using very deep convolutional networks.2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1646–1654, 2016

    2016

  57. [57]

    Total deep variation for linear inverse problems

    Erich Kobler, Alexander Effland, Karl Kunisch, and Thomas Pock. Total deep variation for linear inverse problems. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 7549–7558, 2020

    2020

  58. [58]

    Bayesian imaging using plug & play priors: when Langevin meets tweedie.SIAM Journal onImagingSciences, 15(2):701–737, 2022

    Rémi Laumont, Valentin De Bortoli, Andrés Almansa, Julie Delon, Alain Durmus, and Marcelo Pereyra. Bayesian imaging using plug & play priors: when Langevin meets tweedie.SIAM Journal onImagingSciences, 15(2):701–737, 2022

    2022

  59. [59]

    On maximum a posteriori estimation with plug & play priors and stochastic gradient de- scent

    Rémi Laumont, Valentin De Bortoli, Andrés Almansa, Julie Delon, Alain Durmus, and Marcelo Pereyra. On maximum a posteriori estimation with plug & play priors and stochastic gradient de- scent. Journal of MathematicalImagingandVision, 65(1):140–163, 2023

    2023

  60. [60]

    Leterme, A

    H. Leterme, A. Tersenov, J. Fadili, and J.-L. Starck. A plug-and-play approach with fast uncertainty quantification for weak-lensing mass mapping.Astronomy& Astrophysics, 2025

    2025

  61. [61]

    NETT: solving inverse problems with deep neural networks.InverseProblems, 36(6):065005, June 2020

    Housen Li, Johannes Schwab, Stephan Antholzer, and Markus Haltmeier. NETT: solving inverse problems with deep neural networks.InverseProblems, 36(6):065005, June 2020

    2020

  62. [62]

    Recovery analysis for plug-and- play priors using the restricted eigenvalue condition.Advances in Neural Information Processing Systems, 34:5921–5933, 2021

    Jiaming Liu, Salman Asif, Brendt Wohlberg, and Ulugbek Kamilov. Recovery analysis for plug-and- play priors using the restricted eigenvalue condition.Advances in Neural Information Processing Systems, 34:5921–5933, 2021

    2021

  63. [63]

    Outersemicontinuityofpositivehullmappingswithapplication tosemi-infiniteandstochasticprogramming

    DinhTheLucandRogerJ.-B.Wets. Outersemicontinuityofpositivehullmappingswithapplication tosemi-infiniteandstochasticprogramming. SIAMJournalonOptimization,19(2):700–713,January 2008

    2008

  64. [64]

    Adversarialregularizersininverseprob- lems

    SebastianLunz,OzanÖktem,andCarola-BibianeSchönlieb. Adversarialregularizersininverseprob- lems. Advancesin neuralinformationprocessing systems, 31, 2018

    2018

  65. [65]

    Elsevier, 1999

    Stéphane Mallat.A wavelettourof signalprocessing. Elsevier, 1999

    1999

  66. [66]

    An empirical bayes estimator of the mean of a normal population.Bull

    Koichi Miyasawa et al. An empirical bayes estimator of the mean of a normal population.Bull. Inst. Internat. Statist, 38(181-188):1–2, 1961

    1961

  67. [67]

    Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing.IEEESignal ProcessingMagazine, 38(2):18–44, 2021

    Vishal Monga, Yuelong Li, and Yonina C Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing.IEEESignal ProcessingMagazine, 38(2):18–44, 2021

    2021

  68. [68]

    Data-driven convex regularizersforinverseproblems.In ICASSP2024-2024IEEEInternationalConferenceonAcoustics, SpeechandSignalProcessing (ICASSP), pages 13386–13390

    S Mukherjee, S Dittmer, Z Shumaylov, S Lunz, O Öktem, and C-B Schönlieb. Data-driven convex regularizersforinverseproblems.In ICASSP2024-2024IEEEInternationalConferenceonAcoustics, SpeechandSignalProcessing (ICASSP), pages 13386–13390. IEEE, 2024

    2024

  69. [69]

    End-to-end reconstructionmeetsdata-drivenregularizationforinverseproblems

    Subhadip Mukherjee, Marcello Carioni, Ozan Öktem, and Carola-Bibiane Schönlieb. End-to-end reconstructionmeetsdata-drivenregularizationforinverseproblems. AdvancesinNeuralInformation ProcessingSystems, 34:21413–21425, 2021. 42

    2021

  70. [70]

    Learned reconstruction methods with convergence guarantees: a survey of concepts and applications

    Subhadip Mukherjee, Andreas Hauptmann, Ozan Öktem, Marcelo Pereyra, and Carola-Bibiane Schönlieb. Learned reconstruction methods with convergence guarantees: a survey of concepts and applications. IEEESignal ProcessingMagazine, 40(1):164–182, 2023

    2023

  71. [71]

    Chaudhury

    Pravin Nair and Kunal N. Chaudhury. Averaged deep denoisers for image regularization.Journal of MathematicalImagingandVision, 66(3):362–379, March 2024

    2024

  72. [72]

    Multiscale hierarchical decomposition of im- ageswithapplicationstodeblurring,denoising,andsegmentation

    Suzanne Nezzar, Eitan Tadmor, and Luminita Vese. Multiscale hierarchical decomposition of im- ageswithapplicationstodeblurring,denoising,andsegmentation. CommunicationsinMathematical Sciences, 6(2):281–307, 01 2008

    2008

  73. [73]

    An analytic theory of convo- lutional neural network inverse problems solvers.arXivpreprint, 2026

    Minh Hai Nguyen, Quoc Bao Do, Edouard Pauwels, and Pierre Weiss. An analytic theory of convo- lutional neural network inverse problems solvers.arXivpreprint, 2026

    2026

  74. [74]

    Training-free stein diffusion guidance: Posterior correction for sampling beyond high-density regions.arXiv preprint arXiv:2507.05482, 2025

    Van Khoa Nguyen, Lionel Blondé, and Alexandros Kalousis. Training-free stein diffusion guidance: Posterior correction for sampling beyond high-density regions.arXiv preprint arXiv:2507.05482, 2025

    Pith/arXiv arXiv 2025

  75. [75]

    Metzler, Richard G

    Gregory Ongie, Ajil Jalal, Christopher A. Metzler, Richard G. Baraniuk, Alexandros G. Dimakis, and Rebecca Willett. Deep Learning Techniques for Inverse Problems in Imaging.IEEE Journal on SelectedAreasin InformationTheory, pages 39–56, May 2020

    2020

  76. [76]

    Aniterativeregularization methodfortotalvariation-basedimagerestoration

    StanleyOsher,MartinBurger,DonaldGoldfarb,JinjunXu,andWotaoYin. Aniterativeregularization methodfortotalvariation-basedimagerestoration. MultiscaleModeling&Simulation,4(2):460–489, January 2005

    2005

  77. [77]

    Palomar and S

    D.P. Palomar and S. Verdu. Gradient of mutual information in linear vector gaussian channels.IEEE TransactionsonInformationTheory, 52(1):141–154, 2006

    2006

  78. [78]

    MAP estimation with denoisers: Convergence rates and guarantees

    Scott Pesme, Giacomo Meanti, Michael Arbel, and Julien Mairal. MAP estimation with denoisers: Convergence rates and guarantees. InThe Thirty-ninth Annual Conference on Neural Information ProcessingSystems, 2026

    2026

  79. [79]

    Optimalapproximationofpiecewisesmoothfunctionsusing deep relu neural networks.NeuralNetworks, 108:296–330, December 2018

    PhilippPetersenandFelixVoigtlaender. Optimalapproximationofpiecewisesmoothfunctionsusing deep relu neural networks.NeuralNetworks, 108:296–330, December 2018

    2018

  80. [80]

    Deep neural networks can stably solve high-dimensional, noisy, non-linear inverse problems

    Andrés Felipe Lerma Pineda and Philipp Christian Petersen. Deep neural networks can stably solve high-dimensional, noisy, non-linear inverse problems. Analysis and Applications, 21(01):49–91, 2023

    2023

Showing first 80 references.