Gaussian Surrogates for Poisson Imaging: Some Theoretical and Empirical Results

arxiv: 2602.17274 · v2 · submitted 2026-02-19 · 📡 eess.IV · stat.ML

Gaussian Surrogates for Poisson Imaging: Some Theoretical and Empirical Results

Alexandra Spitzer , Lorenzo Baldassari , Valentin Derbanot , Ivan Dokmani\'c This is my paper

Pith reviewed 2026-05-15 20:56 UTC · model grok-4.3

classification 📡 eess.IV stat.ML

keywords Poisson noiseGaussian surrogateslow-dose imagingmean squared errorcomputed tomographyinverse problemsregularization

0 comments p. Extension

The pith

Gaussian surrogate objectives achieve mean squared error comparable to Poisson MAP in low-dose Poisson imaging problems.

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

The paper examines whether reconstruction objectives based on Gaussian approximations can match the mean squared error performance of objectives derived from the exact Poisson likelihood in imaging inverse problems. In a simplified diagonal model, the unregularized Poisson maximum-likelihood estimator shows large errors at low doses due to instability, while adding regularization improves results. Two Gaussian quadratic surrogates—one heteroscedastic based on normal approximation and one homoscedastic yielding a linear estimator—are shown to reach similar mean squared error levels in this low-dose setting. Computed tomography experiments indicate that the pattern holds in realistic imaging scenarios, suggesting that simpler Gaussian objectives may serve as practical alternatives.

Core claim

In a stylized diagonal model under Poisson noise, both the heteroscedastic Gaussian quadratic objective and the homoscedastic Gaussian quadratic objective produce mean squared error values comparable to those of the regularized Poisson MAP estimator in the low-dose regime, even though the surrogates depart from the Poisson likelihood; numerical CT experiments support extension of this behavior beyond the model.

What carries the argument

The stylized diagonal model that isolates per-component estimation under Poisson noise to compare mean squared error of Poisson MAP versus Gaussian quadratic surrogates.

If this is right

Unregularized Poisson maximum likelihood incurs large mean squared error at low dose due to instability.
Regularization stabilizes Poisson MAP performance in the low-dose regime.
Heteroscedastic and homoscedastic Gaussian surrogates both reach comparable mean squared error without using the Poisson likelihood.
Numerical CT results indicate the findings extend beyond the diagonal model.

Where Pith is reading between the lines

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

For tasks evaluated by mean squared error, the computational simplicity of linear Gaussian estimators could be preferred over exact Poisson objectives.
The result may encourage use of Gaussian surrogates in other Poisson-noise modalities where speed matters more than exact likelihood modeling.
If the diagonal-model analysis generalizes, similar surrogate comparisons could be tested in non-diagonal settings with structured noise.

Load-bearing premise

The stylized diagonal model captures the essential instability and regularization behavior of real imaging inverse problems.

What would settle it

A low-dose CT experiment in which either Gaussian surrogate objective produces substantially higher mean squared error than Poisson MAP would contradict the central claim.

read the original abstract

In imaging inverse problems with Poisson-distributed measurements, it is common to use objectives derived from the Poisson likelihood. But performance is often evaluated by mean squared error (MSE), which raises a practical question: how much does a Poisson objective matter for MSE, even at low dose? We analyze the MSE of Poisson and Gaussian surrogate reconstruction objectives under Poisson noise. In a stylized diagonal model, we show that the unregularized Poisson maximum-likelihood estimator can incur large MSE at low dose, while Poisson MAP mitigates this instability through regularization. We then study two Gaussian surrogate objectives: a heteroscedastic quadratic objective motivated by the normal approximation of Poisson data, and a homoscedastic quadratic objective that yields a simple linear estimator. We show that both surrogates can achieve MSE comparable to Poisson MAP in the low-dose regime, despite departing from the Poisson likelihood. Numerical computed tomography experiments indicate that these conclusions extend beyond the stylized setting of our theoretical analysis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Gaussian surrogates match Poisson MAP on MSE in the low-dose diagonal case and CT tests, but the independent-measurement model leaves open how well this holds when real operator coupling drives the regularization needs.

read the letter

The central result is that both the heteroscedastic and homoscedastic Gaussian surrogates reach MSE levels comparable to regularized Poisson MAP under Poisson noise, while the plain Poisson MLE degrades sharply at low counts. The paper derives the explicit MSE expressions for each estimator in the diagonal setting and shows how regularization stabilizes the Poisson case without the surrogates needing the exact likelihood form. The CT phantom experiments then indicate the same ordering persists in a more applied reconstruction task. That separation of behaviors and the direct MSE comparisons are the useful pieces here; they give a concrete basis for deciding when a simpler quadratic objective is acceptable if the goal is MSE rather than likelihood fidelity. The derivations stay grounded in the noise model and chosen objectives without circular fitting. The main limitation is the diagonal forward operator. It removes the spatial correlations and singular-value structure that make actual CT operators ill-conditioned, so the amount of regularization required in practice often comes from the null space and noise amplification rather than just per-pixel variance. The experiments use a single phantom and a fixed regularization schedule with limited detail on parameter selection across dose levels or run-to-run variability, which makes it harder to judge how sensitive the comparability is to those choices. For someone working on approximate objectives for Poisson imaging pipelines, the paper supplies a clear incremental comparison that could be cited when arguing for Gaussian surrogates on MSE grounds. The math is transparent for the model it uses and the experiments are relevant, so it deserves a serious referee to verify the derivations and ask for more varied controls on the imaging operator and tuning.

Referee Report

2 major / 3 minor

Summary. The paper claims that, for Poisson-noisy imaging inverse problems, both a heteroscedastic Gaussian surrogate (normal approximation to Poisson) and a homoscedastic quadratic surrogate can achieve MSE comparable to Poisson MAP in the low-dose regime, even though they depart from the Poisson likelihood. This is shown analytically in a stylized diagonal forward model that separates the behaviors of unregularized Poisson MLE, regularized Poisson MAP, and the two surrogates; numerical CT experiments are offered as evidence that the conclusion extends to a realistic operator.

Significance. If the central claim holds, the work supplies a clean theoretical separation between likelihood choice and MSE performance under Poisson noise, together with reproducible numerical support. This could justify simpler linear or quadratic surrogates in low-dose CT pipelines where computational cost or implementation simplicity matters, while still preserving near-optimal MSE.

major comments (2)

[§3] §3 (diagonal-model MSE derivations): the analytic comparability between the two Gaussian surrogates and Poisson MAP is proved only under the assumption of a diagonal forward operator. Because this removes all pixel coupling, singular-value decay, and null-space components that drive ill-posedness and regularization needs in real CT, the load-bearing claim that the surrogates remain reliable substitutes rests on an untested extrapolation from the stylized setting.
[§5] §5 (CT experiments): the reported reconstructions use a single phantom, a fixed regularization schedule, and no tabulated sensitivity of MSE to the choice of λ across dose levels. Without these controls it is impossible to determine whether the observed MSE parity is robust or an artifact of a particular λ that happens to favor the surrogates.

minor comments (3)

[Abstract, §1] The abstract and §1 should explicitly state that the theoretical results are confined to the diagonal case and that the CT study is empirical confirmation rather than a proof of generality.
[Table 1] Table 1 (or equivalent) should report the exact λ values used for each method and dose level, together with a brief description of how they were selected.
[Figures 4-6] Figure captions for the CT results should include the number of independent noise realizations and the precise definition of the displayed error bars.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for the constructive review and the recommendation for minor revision. We address the major comments point by point below.

read point-by-point responses

Referee: [§3] §3 (diagonal-model MSE derivations): the analytic comparability between the two Gaussian surrogates and Poisson MAP is proved only under the assumption of a diagonal forward operator. Because this removes all pixel coupling, singular-value decay, and null-space components that drive ill-posedness and regularization needs in real CT, the load-bearing claim that the surrogates remain reliable substitutes rests on an untested extrapolation from the stylized setting.

Authors: We agree that the MSE derivations in §3 are obtained under a diagonal forward operator, which removes pixel coupling, singular-value decay, and null-space effects that characterize realistic CT. This stylized setting was deliberately chosen to permit closed-form MSE expressions that isolate the contribution of the noise model versus regularization. The numerical experiments in §5, performed with a non-diagonal CT operator, are intended to provide empirical evidence that the observed MSE comparability extends beyond the diagonal case. In the revision we will add an explicit paragraph in the discussion section acknowledging the limitations of the diagonal model and clarifying the supporting role of the CT experiments. revision: partial
Referee: [§5] §5 (CT experiments): the reported reconstructions use a single phantom, a fixed regularization schedule, and no tabulated sensitivity of MSE to the choice of λ across dose levels. Without these controls it is impossible to determine whether the observed MSE parity is robust or an artifact of a particular λ that happens to favor the surrogates.

Authors: We thank the referee for noting the limited controls in the experimental section. The current results employ a standard phantom and a single regularization parameter chosen to yield comparable bias-variance trade-offs across the four estimators. To demonstrate robustness, the revised manuscript will include an additional figure (or appendix table) that reports MSE as a function of λ for each method at multiple dose levels. This will allow readers to verify that the reported parity is not an artifact of the particular λ selected. revision: yes

Circularity Check

0 steps flagged

Derivation chain is self-contained with no circular reductions

full rationale

The paper derives explicit MSE expressions for Poisson ML, Poisson MAP, and the two Gaussian surrogate estimators directly from the Poisson noise model and the chosen quadratic objectives inside the stylized diagonal forward operator. These derivations use only the stated statistical assumptions and do not reduce any claimed prediction to a fitted constant or to a self-referential definition. Regularization parameters are introduced as external tuning knobs rather than being solved for inside the same equations. The CT experiments are presented as numerical confirmation that the stylized conclusions are not contradicted by a more realistic operator; they are not invoked to prove the theoretical claims by construction. No self-citation chain, uniqueness theorem, or ansatz smuggling appears in the load-bearing steps.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The analysis rests on the standard Poisson measurement model and the choice of MSE as the evaluation metric; no new entities are introduced and the only free parameter is the regularization strength in the MAP estimator.

free parameters (1)

regularization parameter
Tuned to stabilize the Poisson MAP estimator at low dose; its specific value is chosen externally rather than derived from the model.

axioms (2)

domain assumption Measurements follow independent Poisson distributions
Standard statistical model for photon-counting imaging data.
domain assumption Mean squared error is the relevant performance metric
Explicitly adopted even though the reconstruction objective may differ.

pith-pipeline@v0.9.0 · 5468 in / 1449 out tokens · 26376 ms · 2026-05-15T20:56:18.412561+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 1 internal anchor

[1]

Gaussian Surrogates for Poisson Imaging: Some Theoretical and Empirical Results

INTRODUCTION When photon or electron counts are low, the Poisson noise model is standard [1, 2, 3] and it is natural to optimize a Pois- son likelihood instead of a Gaussian quadratic data term for reconstruction [4, 5, 6]. But likelihood and reconstruction er- ror are different objectives. Maximizing the correct likeli- hood does not guarantee minimal MS...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[2]

, ym)∈N m mod- eled as independent Poisson counts yj ∼ P (sAx⋆)j , j= 1,

A STYLIZED THEORETICAL ANALYSIS We consider measurementsy= (y 1, . . . , ym)∈N m mod- eled as independent Poisson counts yj ∼ P (sAx⋆)j , j= 1, . . . , m,(1) whereP(λ)denotes the Poisson distribution with parameter λ,A:R n →R m is a forward operator such that(Ax⋆)j ≥0 for allj= 1, . . . , m,x ⋆ ∈R n is the unknown signal, ands > 0is the dose. A key quanti...

work page
[3]

EXPERIMENTS We now explore numerically whether these conclusions ex- tend to a more general problem of 2D parallel-beam CT. 3.1. Experimental setting We reconstruct a ground-truth imagex ⋆ ∈ X + supported on the inscribed circleC(see Fig. 1), withn= 256×256. We simulate 2D parallel-beam CT with 180 equispaced angles in [0, π)using ASTRA-cuda [23]. The noi...

work page
[4]

CONCLUSION Our analysis shows a counterintuitive result: when recon- structing from Poisson-distributed measurements, the choice between Poisson and Gaussian likelihood objectives matters less for MSE than proper regularization—even at very low dose where one would expect Gaussian approximations to fail. The theoretical analysis, while conducted in a styl...

work page
[5]

ACKNOWLEDGMENTS The authors would like to thank Jeremy Cohen, V oichita Maxim and Thibaut Modrzyk for valuable discussions re- garding Poisson inverse problems. V .D. is supported by the Agence National de la Recherche (ANR) and the Minist`ere de l’Enseignement Sup´erieur et de la Recherche. Calculations were performed at sciCORE (https://scicore.unibas.ch/)

work page
[6]

Multiscale analysis of photon-limited astronomical images,

Rebecca Willett, “Multiscale analysis of photon-limited astronomical images,” inStatistical Challenges in Mod- ern Astronomy IV, 2007, vol. 371, p. 247

work page 2007
[7]

Image deblurring with Pois- son data: from cells to galaxies,

Mario Bertero, Patrizia Boccacci, Gabriele Desider `a, and Giuseppe Vicidomini, “Image deblurring with Pois- son data: from cells to galaxies,”Inverse Problems, vol. 25, no. 12, pp. 123006, 2009

work page 2009
[8]

Inverse problems with Poisson data: statistical regularization theory, ap- plications and algorithms,

Thorsten Hohage and Frank Werner, “Inverse problems with Poisson data: statistical regularization theory, ap- plications and algorithms,”Inverse Problems, vol. 32, no. 9, pp. 093001, 2016

work page 2016
[9]

Penalized maximum-likelihood image reconstruction using space- alternating generalized em algorithms,

Jeffrey A Fessler and Alfred O Hero, “Penalized maximum-likelihood image reconstruction using space- alternating generalized em algorithms,”IEEE Transac- tions on Image Processing, vol. 4, no. 10, pp. 1417– 1429, 1995

work page 1995
[10]

Richardson–Lucy algorithm with total variation regularization for 3D confocal micro- scope deconvolution,

Nicolas Dey, Laure Blanc-Feraud, Christophe Zimmer, Pascal Roux, Zvi Kam, Jean-Christophe Olivo-Marin, and Josiane Zerubia, “Richardson–Lucy algorithm with total variation regularization for 3D confocal micro- scope deconvolution,”Microscopy research and tech- nique, vol. 69, no. 4, pp. 260–266, 2006

work page 2006
[11]

Mul- tiview attenuation estimation and correction,

Valentin Debarnot, Jonas Kahn, and Pierre Weiss, “Mul- tiview attenuation estimation and correction,”Journal of Mathematical Imaging and Vision, vol. 61, no. 6, pp. 780–797, 2019

work page 2019
[12]

Three-dimensional imaging by deconvolution microscopy,

James G McNally, Tatiana Karpova, John Cooper, and Jos´e Angel Conchello, “Three-dimensional imaging by deconvolution microscopy,”Methods, vol. 19, no. 3, pp. 373–385, 1999

work page 1999
[13]

This is SPIRAL-TAP: Sparse Poisson intensity reconstruction algorithms—theory and prac- tice,

Zachary T Harmany, Roummel F Marcia, and Re- becca M Willett, “This is SPIRAL-TAP: Sparse Poisson intensity reconstruction algorithms—theory and prac- tice,”IEEE Transactions on Image Processing, vol. 21, no. 3, pp. 1084–1096, 2011

work page 2011
[14]

Multiscale photon-limited spectral image re- construction,

Kalyani Krishnamurthy, Maxim Raginsky, and Rebecca Willett, “Multiscale photon-limited spectral image re- construction,”SIAM Journal on Imaging Sciences, vol. 3, no. 3, pp. 619–645, 2010

work page 2010
[15]

Bayesian-based iterative method of image restoration,

William Hadley Richardson, “Bayesian-based iterative method of image restoration,”Journal of the optical society of America, vol. 62, no. 1, pp. 55–59, 1972

work page 1972
[16]

An iterative technique for the rectifica- tion of observed distributions,

Leon B Lucy, “An iterative technique for the rectifica- tion of observed distributions,”Astronomical Journal, vol. 79, pp. 745, 1974

work page 1974
[17]

Maximum likelihood from incomplete data via the EM algorithm,

Arthur P Dempster, Nan M Laird, and Donald B Rubin, “Maximum likelihood from incomplete data via the EM algorithm,”Journal of the royal statistical society: se- ries B (methodological), vol. 39, no. 1, pp. 1–22, 1977

work page 1977
[18]

Maximum likelihood recon- struction for emission tomography,

LA Shepp and Y Vardi, “Maximum likelihood recon- struction for emission tomography,”IEEE transactions on medical imaging, vol. 1, no. 2, pp. 113–122, 1982

work page 1982
[19]

EM reconstruc- tion algorithms for emission and transmission tomogra- phy,

Kenneth Lange and Richard Carson, “EM reconstruc- tion algorithms for emission and transmission tomogra- phy,”Journal of Computer Assisted Tomography, vol. 8, no. 2, pp. 306–316, 1984

work page 1984
[20]

Stopping rule for the MLE algorithm based on statistical hypothesis test- ing,

Eugene Veklerov and Jorge Llacer, “Stopping rule for the MLE algorithm based on statistical hypothesis test- ing,”IEEE Transactions on Medical Imaging, vol. 6, no. 4, pp. 313–319, 2007

work page 2007
[21]

Bayesian reconstructions from emission tomography data using a modified EM algorithm,

Peter J Green, “Bayesian reconstructions from emission tomography data using a modified EM algorithm,”IEEE transactions on medical imaging, vol. 9, no. 1, pp. 84– 93, 1990

work page 1990
[22]

Total varia- tion regulated EM algorithm,

VY Panin, GL Zeng, and GT Gullberg, “Total varia- tion regulated EM algorithm,” in1998 IEEE Nuclear Science Symposium Conference Record. 1998 IEEE Nu- clear Science Symposium and Medical Imaging Confer- ence (Cat. No. 98CH36255). IEEE, 1998, vol. 3, pp. 1562–1566

work page 1998
[23]

Accurate EM-TV algorithm in PET with low SNR,

Alex Sawatzky, Christoph Brune, Frank Wubbeling, Thomas Kosters, Klaus Schafers, and Martin Burger, “Accurate EM-TV algorithm in PET with low SNR,” in2008 IEEE Nuclear Science Symposium Conference Record, 2008, pp. 5133–5137

work page 2008
[24]

The transformation of poisson, binomial and negative-binomial data,

Francis J Anscombe, “The transformation of poisson, binomial and negative-binomial data,”Biometrika, vol. 35, no. 3/4, pp. 246–254, 1948

work page 1948
[25]

Simple formula for the distortions in a gaus- sian representation of a poisson distribution,

LJ Curtis, “Simple formula for the distortions in a gaus- sian representation of a poisson distribution,”American Journal of Physics, vol. 43, no. 12, 1975

work page 1975
[26]

Analysis of an approximate model for Poisson data reconstruction and a related discrepancy principle,

A Stagliano, P Boccacci, and M Bertero, “Analysis of an approximate model for Poisson data reconstruction and a related discrepancy principle,”Inverse Problems, vol. 27, no. 12, pp. 125003, 2011

work page 2011
[27]

Learning Weighted Least Squares Data Term for Pois- son Image Deconvolution,

Abhijit Singh, Emmanuel Soubies, and Caroline Chaux, “Learning Weighted Least Squares Data Term for Pois- son Image Deconvolution,” inICASSP 2025 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing. IEEE, 2025, pp. 1–5

work page 2025
[28]

The ASTRA Toolbox: A plat- form for advanced algorithm development in electron tomography,

Wim van Aarle, Willem Jan Palenstijn, Jan De Been- houwer, Thomas Altantzis, Sara Bals, K. Joost Baten- burg, and Jan Sijbers, “The ASTRA Toolbox: A plat- form for advanced algorithm development in electron tomography,”Ultramicroscopy, vol. 157, 2015

work page 2015
[29]

LoDoPaB-CT, a benchmark dataset for low-dose computed tomography reconstruc- tion,

Johannes Leuschner, Maximilian Schmidt, Daniel Otero Baguer, and Peter Maass, “LoDoPaB-CT, a benchmark dataset for low-dose computed tomography reconstruc- tion,”Scientific Data, vol. 8, no. 1, pp. 109, 2021. Supplementary Materials for ”Gaussian surrogates do well on Poisson inverse problems” A. PROOFS OF SECTION 2 A.1. Proof of Proposition 2.1 A.1.1. Pe...

work page 2021

[1] [1]

Gaussian Surrogates for Poisson Imaging: Some Theoretical and Empirical Results

INTRODUCTION When photon or electron counts are low, the Poisson noise model is standard [1, 2, 3] and it is natural to optimize a Pois- son likelihood instead of a Gaussian quadratic data term for reconstruction [4, 5, 6]. But likelihood and reconstruction er- ror are different objectives. Maximizing the correct likeli- hood does not guarantee minimal MS...

work page internal anchor Pith review Pith/arXiv arXiv 2026

[2] [2]

, ym)∈N m mod- eled as independent Poisson counts yj ∼ P (sAx⋆)j , j= 1,

A STYLIZED THEORETICAL ANALYSIS We consider measurementsy= (y 1, . . . , ym)∈N m mod- eled as independent Poisson counts yj ∼ P (sAx⋆)j , j= 1, . . . , m,(1) whereP(λ)denotes the Poisson distribution with parameter λ,A:R n →R m is a forward operator such that(Ax⋆)j ≥0 for allj= 1, . . . , m,x ⋆ ∈R n is the unknown signal, ands > 0is the dose. A key quanti...

work page

[3] [3]

EXPERIMENTS We now explore numerically whether these conclusions ex- tend to a more general problem of 2D parallel-beam CT. 3.1. Experimental setting We reconstruct a ground-truth imagex ⋆ ∈ X + supported on the inscribed circleC(see Fig. 1), withn= 256×256. We simulate 2D parallel-beam CT with 180 equispaced angles in [0, π)using ASTRA-cuda [23]. The noi...

work page

[4] [4]

CONCLUSION Our analysis shows a counterintuitive result: when recon- structing from Poisson-distributed measurements, the choice between Poisson and Gaussian likelihood objectives matters less for MSE than proper regularization—even at very low dose where one would expect Gaussian approximations to fail. The theoretical analysis, while conducted in a styl...

work page

[5] [5]

ACKNOWLEDGMENTS The authors would like to thank Jeremy Cohen, V oichita Maxim and Thibaut Modrzyk for valuable discussions re- garding Poisson inverse problems. V .D. is supported by the Agence National de la Recherche (ANR) and the Minist`ere de l’Enseignement Sup´erieur et de la Recherche. Calculations were performed at sciCORE (https://scicore.unibas.ch/)

work page

[6] [6]

Multiscale analysis of photon-limited astronomical images,

Rebecca Willett, “Multiscale analysis of photon-limited astronomical images,” inStatistical Challenges in Mod- ern Astronomy IV, 2007, vol. 371, p. 247

work page 2007

[7] [7]

Image deblurring with Pois- son data: from cells to galaxies,

Mario Bertero, Patrizia Boccacci, Gabriele Desider `a, and Giuseppe Vicidomini, “Image deblurring with Pois- son data: from cells to galaxies,”Inverse Problems, vol. 25, no. 12, pp. 123006, 2009

work page 2009

[8] [8]

Inverse problems with Poisson data: statistical regularization theory, ap- plications and algorithms,

Thorsten Hohage and Frank Werner, “Inverse problems with Poisson data: statistical regularization theory, ap- plications and algorithms,”Inverse Problems, vol. 32, no. 9, pp. 093001, 2016

work page 2016

[9] [9]

Penalized maximum-likelihood image reconstruction using space- alternating generalized em algorithms,

Jeffrey A Fessler and Alfred O Hero, “Penalized maximum-likelihood image reconstruction using space- alternating generalized em algorithms,”IEEE Transac- tions on Image Processing, vol. 4, no. 10, pp. 1417– 1429, 1995

work page 1995

[10] [10]

Richardson–Lucy algorithm with total variation regularization for 3D confocal micro- scope deconvolution,

Nicolas Dey, Laure Blanc-Feraud, Christophe Zimmer, Pascal Roux, Zvi Kam, Jean-Christophe Olivo-Marin, and Josiane Zerubia, “Richardson–Lucy algorithm with total variation regularization for 3D confocal micro- scope deconvolution,”Microscopy research and tech- nique, vol. 69, no. 4, pp. 260–266, 2006

work page 2006

[11] [11]

Mul- tiview attenuation estimation and correction,

Valentin Debarnot, Jonas Kahn, and Pierre Weiss, “Mul- tiview attenuation estimation and correction,”Journal of Mathematical Imaging and Vision, vol. 61, no. 6, pp. 780–797, 2019

work page 2019

[12] [12]

Three-dimensional imaging by deconvolution microscopy,

James G McNally, Tatiana Karpova, John Cooper, and Jos´e Angel Conchello, “Three-dimensional imaging by deconvolution microscopy,”Methods, vol. 19, no. 3, pp. 373–385, 1999

work page 1999

[13] [13]

This is SPIRAL-TAP: Sparse Poisson intensity reconstruction algorithms—theory and prac- tice,

Zachary T Harmany, Roummel F Marcia, and Re- becca M Willett, “This is SPIRAL-TAP: Sparse Poisson intensity reconstruction algorithms—theory and prac- tice,”IEEE Transactions on Image Processing, vol. 21, no. 3, pp. 1084–1096, 2011

work page 2011

[14] [14]

Multiscale photon-limited spectral image re- construction,

Kalyani Krishnamurthy, Maxim Raginsky, and Rebecca Willett, “Multiscale photon-limited spectral image re- construction,”SIAM Journal on Imaging Sciences, vol. 3, no. 3, pp. 619–645, 2010

work page 2010

[15] [15]

Bayesian-based iterative method of image restoration,

William Hadley Richardson, “Bayesian-based iterative method of image restoration,”Journal of the optical society of America, vol. 62, no. 1, pp. 55–59, 1972

work page 1972

[16] [16]

An iterative technique for the rectifica- tion of observed distributions,

Leon B Lucy, “An iterative technique for the rectifica- tion of observed distributions,”Astronomical Journal, vol. 79, pp. 745, 1974

work page 1974

[17] [17]

Maximum likelihood from incomplete data via the EM algorithm,

Arthur P Dempster, Nan M Laird, and Donald B Rubin, “Maximum likelihood from incomplete data via the EM algorithm,”Journal of the royal statistical society: se- ries B (methodological), vol. 39, no. 1, pp. 1–22, 1977

work page 1977

[18] [18]

Maximum likelihood recon- struction for emission tomography,

LA Shepp and Y Vardi, “Maximum likelihood recon- struction for emission tomography,”IEEE transactions on medical imaging, vol. 1, no. 2, pp. 113–122, 1982

work page 1982

[19] [19]

EM reconstruc- tion algorithms for emission and transmission tomogra- phy,

Kenneth Lange and Richard Carson, “EM reconstruc- tion algorithms for emission and transmission tomogra- phy,”Journal of Computer Assisted Tomography, vol. 8, no. 2, pp. 306–316, 1984

work page 1984

[20] [20]

Stopping rule for the MLE algorithm based on statistical hypothesis test- ing,

Eugene Veklerov and Jorge Llacer, “Stopping rule for the MLE algorithm based on statistical hypothesis test- ing,”IEEE Transactions on Medical Imaging, vol. 6, no. 4, pp. 313–319, 2007

work page 2007

[21] [21]

Bayesian reconstructions from emission tomography data using a modified EM algorithm,

Peter J Green, “Bayesian reconstructions from emission tomography data using a modified EM algorithm,”IEEE transactions on medical imaging, vol. 9, no. 1, pp. 84– 93, 1990

work page 1990

[22] [22]

Total varia- tion regulated EM algorithm,

VY Panin, GL Zeng, and GT Gullberg, “Total varia- tion regulated EM algorithm,” in1998 IEEE Nuclear Science Symposium Conference Record. 1998 IEEE Nu- clear Science Symposium and Medical Imaging Confer- ence (Cat. No. 98CH36255). IEEE, 1998, vol. 3, pp. 1562–1566

work page 1998

[23] [23]

Accurate EM-TV algorithm in PET with low SNR,

Alex Sawatzky, Christoph Brune, Frank Wubbeling, Thomas Kosters, Klaus Schafers, and Martin Burger, “Accurate EM-TV algorithm in PET with low SNR,” in2008 IEEE Nuclear Science Symposium Conference Record, 2008, pp. 5133–5137

work page 2008

[24] [24]

The transformation of poisson, binomial and negative-binomial data,

Francis J Anscombe, “The transformation of poisson, binomial and negative-binomial data,”Biometrika, vol. 35, no. 3/4, pp. 246–254, 1948

work page 1948

[25] [25]

Simple formula for the distortions in a gaus- sian representation of a poisson distribution,

LJ Curtis, “Simple formula for the distortions in a gaus- sian representation of a poisson distribution,”American Journal of Physics, vol. 43, no. 12, 1975

work page 1975

[26] [26]

Analysis of an approximate model for Poisson data reconstruction and a related discrepancy principle,

A Stagliano, P Boccacci, and M Bertero, “Analysis of an approximate model for Poisson data reconstruction and a related discrepancy principle,”Inverse Problems, vol. 27, no. 12, pp. 125003, 2011

work page 2011

[27] [27]

Learning Weighted Least Squares Data Term for Pois- son Image Deconvolution,

Abhijit Singh, Emmanuel Soubies, and Caroline Chaux, “Learning Weighted Least Squares Data Term for Pois- son Image Deconvolution,” inICASSP 2025 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing. IEEE, 2025, pp. 1–5

work page 2025

[28] [28]

The ASTRA Toolbox: A plat- form for advanced algorithm development in electron tomography,

Wim van Aarle, Willem Jan Palenstijn, Jan De Been- houwer, Thomas Altantzis, Sara Bals, K. Joost Baten- burg, and Jan Sijbers, “The ASTRA Toolbox: A plat- form for advanced algorithm development in electron tomography,”Ultramicroscopy, vol. 157, 2015

work page 2015

[29] [29]

LoDoPaB-CT, a benchmark dataset for low-dose computed tomography reconstruc- tion,

Johannes Leuschner, Maximilian Schmidt, Daniel Otero Baguer, and Peter Maass, “LoDoPaB-CT, a benchmark dataset for low-dose computed tomography reconstruc- tion,”Scientific Data, vol. 8, no. 1, pp. 109, 2021. Supplementary Materials for ”Gaussian surrogates do well on Poisson inverse problems” A. PROOFS OF SECTION 2 A.1. Proof of Proposition 2.1 A.1.1. Pe...

work page 2021