arxiv: 2605.11980 · v1 · submitted 2026-05-12 · 🌌 astro-ph.IM · astro-ph.SR

Recognition: no theorem link

Marginal multi-object multi-frame blind deconvolution

A. Asensio Ramos (IAC , ULL)

Authors on Pith no claims yet

Pith reviewed 2026-05-13 03:52 UTC · model grok-4.3

classification 🌌 astro-ph.IM astro-ph.SR

keywords blind deconvolutionmarginal estimationsolar imagingatmospheric turbulence correctionmulti-frame reconstructionlog-determinant regularizationimage restorationground-based astronomy

0 comments

The pith

Marginalizing over objects yields a more regularized estimator for solar blind deconvolution.

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

The paper develops a marginal estimator for the multi-object multi-frame blind deconvolution problem that arises in high-resolution ground-based solar imaging. Instead of jointly estimating the object and the atmospheric aberrations, the method integrates out the objects to produce a new objective function. This marginalization supplies extra regularization that accounts for uncertainty in the object, which stops the algorithm from fitting noise as high-order aberrations. The resulting estimator is also far less sensitive to the hyperparameters that control the object's power spectral density, so those parameters can be optimized once and the method deployed without manual tuning. Implementation is straightforward: existing pipelines need only the addition of a log-determinant term to their merit function.

Core claim

The paper claims that marginalizing over the observed objects in the multi-object multi-frame blind deconvolution problem produces a reconstruction method whose merit function contains an extra log-determinant term. This term naturally incorporates object uncertainty into the regularization and thereby prevents the algorithm from erroneously assigning noise to high-order aberrations. The marginal estimator is also much less sensitive to the hyperparameters that set the object's power spectral density, which allows those hyperparameters to be optimized and yields a plug-and-play procedure that requires no manual tuning. The approach is accessible because it can be added to existing blind-deco

What carries the argument

Marginalization over the observed objects, which adds a log-determinant term to the traditional merit function.

If this is right

The reconstruction avoids assigning noise to high-order aberrations because object uncertainty is folded into the regularization.
Hyperparameters controlling the object's power spectral density can be optimized once, enabling plug-and-play use without manual tuning.
Existing blind deconvolution pipelines can be updated by adding only the log-determinant term to the merit function.
The method provides more stable contrast control across varying atmospheric conditions.

Where Pith is reading between the lines

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

The marginal approach could support fully automated pipelines that process large volumes of solar data without per-image tuning.
Similar marginalization might be applied to other multi-frame inverse problems in astronomy where the object is uncertain.
The log-determinant term could be combined with learned priors to further speed up reconstruction.

Load-bearing premise

Marginalizing over the objects remains tractable and the log-determinant term supplies enough regularization without bias or extra approximations.

What would settle it

Compare joint and marginal reconstructions on the same simulated solar images that contain known turbulence and noise, then check whether the marginal version assigns measurably less noise to high-order aberrations.

Figures

Figures reproduced from arXiv: 2605.11980 by A. Asensio Ramos (IAC, ULL).

**Figure 1.** Figure 1: Reconstruction of the quiet Sun CRISP dataset for different values of the hyperparameter K controlling the PSD of the object prior. The contrast of the reconstructions is very sensitive to the choice of K. gives a reconstruction more in the line of what we expect for the quiet Sun without the need of tuning any hyperparameter. 3.3. Comparison with MOMFBD [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗

**Figure 2.** Figure 2: Comparison of the joint and marginal estimators obtained with torchmfbd with the results of MOMFBD. rows display the joint and marginal estimators, respectively, while the last row shows the results obtained with MOMFBD. The contrast quoted in each panel has been obtained in the whole image. The joint reconstruction is computed for su = 100, while the MOMFBD reconstructions are obtained with the default se… view at source ↗

**Figure 3.** Figure 3: Reconstruction of the quiet Sun HiFI dataset for different number of frames M and different values of the constant PSD su. The last row displays the reconstruction obtained with the marginal estimator. dataset, which has a large number of frames per burst. The results of the joint reconstruction with different values of su and different number of frames M are displayed in the first three rows of [PITH_FUL… view at source ↗

**Figure 4.** Figure 4: Azimuthally averaged power spectra of the reconstructions for different number of images and modes, correspoding to the results shown in [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗

**Figure 5.** Figure 5: Reconstruction of the quiet Sun HiFI dataset for different number of modes N and different values of the constant PSD su. The last row displays the reconstruction obtained with the marginal estimator. when using M = 50, all reconstructions produce a similar power spectrum, with less artifacts close to the diffraction limit. 3.5. Number of modes The number of modes used to parameterize the PSFs also has a s… view at source ↗

**Figure 6.** Figure 6: Loss and standard deviation of the residuals as a function of the number of modes. improve if enough information is present in the observed frames. However, the higher order modes are less certain and the contrast becomes out of control if the PSD of the object prior is not properly tuned. This is shown in [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗

read the original abstract

High-resolution ground-based solar imaging relies heavily on multi-object multi-frame blind deconvolution to correct for atmospheric turbulence. However, the traditional joint maximum likelihood estimation methods in which object and the atmospheric aberrations are estimated together face some problems. In this paper, we introduce a marginal estimator for the multi-object multi-frame blind deconvolution problem. By employing a framework to marginalize over the observed objects, we develop a reconstruction method that offers several distinct advantages over joint estimation. First, the marginalization provides enhanced regularization that naturally accounts for object uncertainty, successfully preventing the reconstruction algorithm from erroneously assigning noise to high-order aberrations. Second, the marginal estimator yields more contrast control, as it is much less sensitive to the hyperparameters dictating the power spectral density (PSD) of the object. This robustness allows these hyperparameters to be optimized, enabling a ``plug-and-play'' deployment that removes the need for manual tuning. Finally, we demonstrate that the proposed method is accessible and simple to implement, requiring only the addition of a log-determinant term to the traditional merit function. With minimal modifications required for existing blind deconvolution pipelines, the estimator has been fully integrated into the open-source torchmfbd package for its use by the solar physics community.

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

The paper claims that marginalizing over objects in multi-object multi-frame blind deconvolution adds a log-determinant term that improves regularization and reduces hyperparameter sensitivity compared to joint estimation.

read the letter

The paper's main contribution is a marginal estimator for multi-object multi-frame blind deconvolution in solar imaging. Instead of jointly estimating the objects and aberrations, it marginalizes over the objects and adds a log-determinant term to the traditional merit function. The authors argue this supplies natural regularization that accounts for object uncertainty and prevents fitting noise to high-order aberrations. They also claim the approach is far less sensitive to the object's power spectral density hyperparameters, which would let those be optimized automatically for a plug-and-play workflow without manual tuning. The method is presented as simple to implement and has already been added to the open-source torchmfbd package. That last part is genuinely useful for adoption in solar physics pipelines. The work targets a real practical bottleneck in high-resolution ground-based solar observations, and the focus on minimal changes to existing code is a clear strength. The soft spots are in the missing technical details. The abstract gives no derivation, no complexity analysis, and no numerical tests or comparisons, so it is not possible to verify whether the marginalization stays tractable once multiple objects and frames are involved or whether the log-determinant term actually avoids bias. The multi-object coupling could introduce cross terms in the covariance that complicate the closed-form marginal the stress-test note flags. If the full paper contains clean math and solid validation against joint methods, those issues would be settled; from the description alone they remain open. This paper is for solar physicists and image-reconstruction specialists who already work with blind deconvolution. A reader comfortable with merit-function methods would get the most out of it. It deserves a serious referee to check the derivations and any empirical results.

Referee Report

3 major / 1 minor

Summary. The paper claims to develop a marginal estimator for the multi-object multi-frame blind deconvolution problem in solar imaging by marginalizing over the objects. This leads to a modified merit function that includes a log-determinant term, purportedly offering enhanced regularization, greater robustness to object PSD hyperparameters, and simplicity of implementation compared to joint estimation methods. The method is integrated into the torchmfbd package.

Significance. If the marginalization yields a tractable, bias-free log-determinant regularization as described, this work could provide a practical improvement for high-resolution ground-based solar imaging by reducing the need for manual hyperparameter tuning and preventing over-fitting of noise to aberrations. The open-source integration is a notable strength for reproducibility and adoption.

major comments (3)

[§3] The derivation of the marginal likelihood must explicitly demonstrate that the covariance matrix for the multi-frame multi-object observations permits an exact closed-form log-determinant without additional approximations, as any hidden approximation would undermine the claimed simplicity and advantages.
[§4.1] The assertion that the log-determinant term naturally accounts for object uncertainty and prevents noise assignment to high-order aberrations requires a specific proof or numerical test showing no systematic bias is introduced in the aberration estimates; this is load-bearing for the central claim of superiority over joint estimation.
[Results section] Comparative experiments demonstrating reduced sensitivity to PSD hyperparameters (e.g., across a grid of values) and improved contrast control are necessary to substantiate the 'plug-and-play' advantage; without them, the robustness claim is not yet supported.

minor comments (1)

[Abstract] The abstract states the method is 'accessible and simple' but would benefit from a one-sentence indication of the computational overhead of the log-determinant term.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. These have highlighted areas where additional rigor and experiments will strengthen the manuscript. We address each major comment below and commit to the corresponding revisions.

read point-by-point responses

Referee: [§3] The derivation of the marginal likelihood must explicitly demonstrate that the covariance matrix for the multi-frame multi-object observations permits an exact closed-form log-determinant without additional approximations, as any hidden approximation would undermine the claimed simplicity and advantages.

Authors: Under the Gaussian object and noise assumptions stated in Section 3, the covariance of the stacked multi-frame observations takes the form Σ = A K Aᵀ + σ²I, where A encodes the multi-frame PSF operators acting on the shared object and K is the object covariance. This admits an exact closed-form log-determinant via the matrix determinant lemma applied to the low-rank update structure induced by the common object across frames; no further approximations are introduced. We will revise §3 to include the explicit algebraic steps showing this structure and the resulting determinant expression. revision: yes
Referee: [§4.1] The assertion that the log-determinant term naturally accounts for object uncertainty and prevents noise assignment to high-order aberrations requires a specific proof or numerical test showing no systematic bias is introduced in the aberration estimates; this is load-bearing for the central claim of superiority over joint estimation.

Authors: We agree that a direct demonstration of unbiased aberration recovery is necessary. In the revised manuscript we will add a controlled numerical test (new subsection or appendix) that injects known aberrations into simulated multi-object data, recovers them with both the marginal and joint estimators, and reports the bias and variance of the aberration coefficients. The test will confirm that the marginal estimator suppresses spurious high-order power without shifting the low-order estimates. A brief analytic argument based on the marginalization integral will also be included. revision: yes
Referee: [Results section] Comparative experiments demonstrating reduced sensitivity to PSD hyperparameters (e.g., across a grid of values) and improved contrast control are necessary to substantiate the 'plug-and-play' advantage; without them, the robustness claim is not yet supported.

Authors: We accept that quantitative sensitivity results are required to support the plug-and-play claim. The revised Results section will contain a new figure and accompanying table that sweep the object PSD hyperparameters (power-law index and amplitude) over a representative grid, reporting Strehl ratio, image contrast, and aberration residual norms for both estimators on the same data sets. These experiments will demonstrate the markedly flatter performance surface of the marginal estimator. revision: yes

Circularity Check

0 steps flagged

Marginalization derivation is self-contained via standard Bayesian integration

full rationale

The paper derives its marginal estimator by integrating the joint likelihood over the object variables, which under standard Gaussian object priors produces an additional log-determinant term in the merit function. This is a direct mathematical consequence of marginalization and does not reduce to a self-definition, fitted parameter renamed as prediction, or load-bearing self-citation. The claimed regularization benefits and robustness to PSD hyperparameters follow from the properties of the marginal likelihood itself rather than being imposed by construction. No equations or sections in the text exhibit a reduction where the result is equivalent to its inputs by definition, and the approach is presented as a modification of existing blind deconvolution pipelines without hidden approximations that would create circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available; therefore the precise free parameters, axioms, and any invented entities cannot be audited. The method relies on standard marginalization and the addition of a log-determinant term whose derivation details are not shown.

axioms (1)

domain assumption Marginalization over object variables is valid and tractable for the blind deconvolution likelihood
Central to the claimed regularization benefit; abstract does not specify the form of the marginalization or any approximations used.

pith-pipeline@v0.9.0 · 5513 in / 1184 out tokens · 106870 ms · 2026-05-13T03:52:25.766832+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · 1 internal anchor

[1]

Astrophys.703, A269

Asensio Ramos, A., L¨ ofdahl, M.G., D´ ıaz Baso, C., Kuckein, C., Esteban Pozuelo, S.: 2025, torchmfbd: A flexible multi-object, multi-frame blind deconvolution code.Astron. Astrophys.703, A269. DOI. ADS. Blanco, L., Mugnier, L.M.: 2011, Marginal blind deconvolution of adaptive optics retinal images.Opt. Express19, 23227. DOI.URL. SOLA: marginal.tex; 13 M...

work page 2025
[2]

DOI. ADS. Fitzgerald, M.P., Graham, J.R.: 2006, Speckle Statistics in Adaptively Corrected Images.The Astrophysical Journal637,

work page 2006
[3]

Gonsalves, R.A.: 1982, Phase retreival and diversity in adaptive optics.Optical Engineering 21,

DOI.URL. Gonsalves, R.A.: 1982, Phase retreival and diversity in adaptive optics.Optical Engineering 21,

work page 1982
[4]

DOI. ADS. Harris, C.R., Millman, K.J., van der Walt, S.J., Gommers, R., Virtanen, P., Cournapeau, D., Wieser, E., Taylor, J., Berg, S., Smith, N.J., Kern, R., Picus, M., Hoyer, S., van Kerkwijk, M.H., Brett, M., Haldane, A., Fern´ andez del R´ ıo, J., Wiebe, M., Peterson, P., G´ erard- Marchant, P., Sheppard, K., Reddy, T., Weckesser, W., Abbasi, H., Gohl...

work page arXiv 2020
[5]

Jang, E., Gu, S., Poole, B.: 2016, Categorical Reparameterization with Gumbel-Softmax.arXiv e-prints, arXiv:1611.01144. DOI. ADS. Kingma, D.P., Ba, J.: 2014, Adam: A Method for Stochastic Optimization.CoRR abs/1412.6980.URL. Kleint, L., Berkefeld, T., Esteves, M., Sonner, T., Volkmer, R., Gerber, K., Kr¨ amer, F., Grassin, O., Berdyugina, S.: 2020, GREGOR...

work page internal anchor Pith review Pith/arXiv arXiv 2016
[6]

DOI. ADS. Labeyrie, A.: 1970, Attainment of Diffraction Limited Resolution in Large Telescopes by Fourier Analysing Speckle Patterns in Star Images.Astron. Astrophys.6,

work page 1970
[7]

In: Bones, P.J., Fiddy, M.A., Millane, R.P

L¨ ofdahl, M.G.: 2002, Multi-frame blind deconvolution with linear equality constraints. In: Bones, P.J., Fiddy, M.A., Millane, R.P. (eds.)Image Reconstruction from Incomplete Data II,Proc. SPIE4792,

work page 2002
[8]

DOI. ADS. L¨ ofdahl, M.G., Scharmer, G.B.: 1994, Wavefront sensing and image restoration from focused and defocused solar images.Astron. Astrophys.s107,

work page 1994
[9]

Astrophys.653, A68

L¨ ofdahl, M.G., Hillberg, T., de la Cruz Rodr´ ıguez, J., Vissers, G., Andriienko, O., Scharmer, G.B., Haugan, S.V.H., Fredvik, T.: 2021, SSTRED: Data- and metadata-processing pipeline for CHROMIS and CRISP.Astron. Astrophys.653, A68. DOI. ADS. Modestov, M., Khomenko, E., Vitas, N., de Vicente, A., Navarro, A., Gonz´ alez-Morales, P.A., Collados, M., Fel...

work page 2021
[10]

DOI. ADS. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: 2019, PyTorch: An Imperative Style, High-Performance Deep Learning Library. In: Wallach, H....

work page 2019
[11]

21 Asensio Ramos Paxman, R.G., Shulz, T.J., Fienup, J.R.: 1992, Phase-Diverse Speckle Interferometry

SOLA: marginal.tex; 13 May 2026; 1:10; p. 21 Asensio Ramos Paxman, R.G., Shulz, T.J., Fienup, J.R.: 1992, Phase-Diverse Speckle Interferometry. In:Signal Recovery and Synthesis IV,OSA Technical Digest Series11,

work page 2026
[12]

Astrophys.675, A93

Quintero Noda, C., Khomenko, E., Collados, M., Ruiz Cobo, B., Gafeira, R., Vitas, N., Rem- pel, M., Campbell, R.J., Pastor Yabar, A., Uitenbroek, H., Orozco Su´ arez, D.: 2023, A study of the capabilities for inferring atmospheric information from high-spatial-resolution simulations.Astron. Astrophys.675, A93. DOI. ADS. Rimmele, T.R., Warner, M., Keil, S....

work page 2023
[13]

DOI. ADS. Scharmer, G.B.: 2006, Comments on the optimization of high resolution Fabry-P´ erot filter- graphs.Astron. Astrophys.447,

work page 2006
[14]

In:SO- LARNET IV: The Physics of the Sun from the Interior to the Outer Atmosphere,

Scharmer, G.: 2017, SST/CHROMIS: a new window to the solar chromosphere. In:SO- LARNET IV: The Physics of the Sun from the Interior to the Outer Atmosphere,

work page 2017
[15]

Scharmer, G.B., Narayan, G., Hillberg, T., de la Cruz Rodriguez, J., L¨ ofdahl, M.G., Kiselman, D., S¨ utterlin, P., van Noort, M., Lagg, A.: 2008, CRISP Spectropolarimetric Imaging of Penumbral Fine Structure.Astrophys. J. Lett.689, L69. Scharmer, G.B., Sliepen, G., Sinquin, J.-C., L¨ ofdahl, M.G., Lindberg, B., S¨ utterlin, P.: 2024, The 85-electrode AO...

work page 2008
[16]

DOI. ADS. van Noort, M., Rouppe van der Voort, L., L¨ ofdahl, M.G.: 2005, Solar Image Restoration By Use Of Multi-frame Blind De-convolution With Multiple Objects And Phase Diversity.Sol. Phys.228,

work page 2005
[17]

Vicente Ar´ evalo, A., Asensio Ramos, A., Esteban Pozuelo, S.: 2022, Accelerating Non-LTE Synthesis and Inversions with Graph Networks.Astrophys. J.928,

work page 2022
[18]

DOI. ADS. von der L¨ uhe, O.: 1993, Speckle imaging of solar small scale structure. I - Methods.Astron. Astrophys.268,

work page 1993
[19]

SOLA: marginal.tex; 13 May 2026; 1:10; p. 22

work page 2026