arxiv: 2604.20691 · v1 · submitted 2026-04-22 · ⚛️ physics.data-an

Recognition: unknown

Bayesian approach for uncertainty quantification of hybrid spectral unmixing in γ-ray spectrometry

Dinh Triem Phan , J\'er\^ome Bobin , Cheick Thiam , Christophe Bobin

Authors on Pith no claims yet

Pith reviewed 2026-05-09 22:36 UTC · model grok-4.3

classification ⚛️ physics.data-an

keywords gamma-ray spectrometryspectral unmixingBayesian uncertainty quantificationLaplace approximationMarkov Chain Monte Carlocoverage intervalsradionuclide identification

0 comments

The pith

Markov Chain Monte Carlo sampling yields accurate coverage intervals for spectral unmixing estimates under active constraints while Laplace approximation does not.

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

The paper presents two Bayesian approaches to quantify uncertainty in estimates of radionuclide counting rates and spectral deformation parameters from a hybrid unmixing method. It tests Laplace approximation against Markov Chain Monte Carlo sampling on repeated simulations to check if the 95.4 percent long-run success rate for coverage intervals holds. Both approaches succeed without constraints on deformation or counting, but Laplace fails when those constraints activate or background dominates because the posterior becomes non-Gaussian. MCMC continues to deliver the target success rate in those cases.

Core claim

The study shows that the Markov Chain Monte Carlo method provides robust uncertainty quantification for the counting and the variable λ characterizing spectral signatures, maintaining the expected coverage success rate of 95.4% even when constraints related to spectral signatures deformation and counting are active or when background counting dominates, whereas the Laplace approximation deviates due to the non-Gaussian nature of the posterior distribution.

What carries the argument

Bayesian posterior sampling using either Laplace Gaussian approximation or Markov Chain Monte Carlo, applied to the hybrid spectral unmixing model with constraints on spectral deformation.

If this is right

When constraints are inactive both methods achieve coverage close to 95.4 percent.
Laplace approximation deviates under active constraints or dominant background.
Markov Chain Monte Carlo remains accurate regardless of constraint activation.
This distinction supports choosing the appropriate method for reliable radionuclide quantification decisions.

Where Pith is reading between the lines

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

Practitioners facing real gamma spectra with variable backgrounds should default to MCMC for uncertainty estimates to avoid under- or over-coverage.
The results point to a need for diagnostic checks on posterior shape before applying Laplace approximations in constrained inverse problems.
Future work could develop adaptive methods that use Laplace when safe and switch to MCMC otherwise.

Load-bearing premise

The posterior distribution of the counting and spectral signature parameters is sufficiently close to Gaussian that the Laplace approximation accurately captures the coverage intervals even when deformation and counting constraints are active.

What would settle it

Simulate repeated datasets with active spectral deformation constraints and dominant background, compute the empirical coverage success rate of the Laplace intervals, and check whether it differs substantially from 95.4 percent.

Figures

Figures reproduced from arXiv: 2604.20691 by Cheick Thiam, Christophe Bobin, Dinh Triem Phan, J\'er\^ome Bobin.

**Figure 2.** Figure 2: Spectral signatures of 4 radionuclides with 20 mm steel thickness (left) and a simulated γ-spectrum of this mixture including 60Co, 88Y, 99mTc, 137Cs and Bkg at low statistics (right). 8 [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** Figure 3: Schema of CI evaluation. 5. Results of coverage interval evaluation 5.1. Long-run success rate 5.1.1. A mixture of Bkg and 60Co To assess whether the observed LRSR is consistent with the expectation value of 95.4%, an uncertainty bar is defined around the expected value. Each CI either includes the true value or not, which can be modeled as a Bernoulli random variable (1 if it includes the true value, 0 ot… view at source ↗

**Figure 4.** Figure 4: LRSR of 95.4% CI of the LA method of the latent variable λ. The x-axis is the minimum counting (min(a)) between the Bkg and 60Co counting. The y-axis represents the λ value. The expected value of LRSR in the color bar p is 95.4%. The yellow color corresponds to the interval [p−2σp, p+ 2σp], indicating the range where approximately 95.4% (corresponding to two standard deviations of a Gaussian distribution) … view at source ↗

**Figure 5.** Figure 5: Example of simulated spectra (with Poisson noise) and theoretical spectra (without noise) for fours scenarios corresponding to four zones in [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

**Figure 6.** Figure 6: Same as [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

**Figure 7.** Figure 7: Same as [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗

**Figure 8.** Figure 8: A simulated γ-spectrum (left) and λ distribution (right) for the case of non-Gaussian of the posterior distribution, similar to [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗

**Figure 9.** Figure 9: Same as [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗

**Figure 10.** Figure 10: A simulated γ-spectrum (left) and λ distribution (right) for the case of non-Gaussian of the posterior distribution, similar to [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗

**Figure 11.** Figure 11: Same as [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗

**Figure 12.** Figure 12: Mean of TVD of all MC simulations for λ (left) and 60Co (right) for the mixture of Bkg and 60Co [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗

**Figure 13.** Figure 13: Mean of TVD of all MC simulations for λ (left) and 60Co (right) for the mixture of Bkg and four radionuclides. 6. Conclusion In this work, uncertainty quantification of counting and λ which characterizes the spectral signatures has been investigated for the hybrid spectral unmixing algorithm SEMSUN in γ-ray spectrometry with spectral deformation. Given the inverse problem and the constraints involved, t… view at source ↗

read the original abstract

Identifying and quantifying $\gamma$-emitting radionuclides, considering spectral deformation from $\gamma$-interactions in radioactive source surroundings, present a significant challenge in $\gamma$-ray spectrometry. In that context, a hybrid machine learning method has been previously proposed to jointly estimate the counting and spectral signatures of $\gamma$-emitters under conditions of spectral variability. This paper addresses the uncertainty quantification of the estimators (i.e., the counting and the variable $\lambda$ which characterizes the spectral signatures) obtained by this spectral unmixing algorithm. The focus is on the coverage interval, as defined by the GUM, which corresponds closely to a credible interval in the Bayesian framework. Given the inverse problem and the constraints associated with spectral deformation, two Bayesian methods - Laplace approximation and Markov Chain Monte Carlo - have been developed for uncertainty quantification to ensure robust decision-making. The Laplace approximation technique approximates the posterior distribution by a Gaussian distribution, while the Markov Chain Monte Carlo technique samples the posterior distribution. This study evaluates these two methods in terms of precision of coverage interval based on repeated Monte Carlo samples using the long-run success rate. Numerical experiments show that both methods yield similar results close to the expected success rate of 95.4$\%$ when constraints related to spectral signatures deformation and counting are inactive. However, when constraints are active or the background counting significantly dominates other radionuclides, the Laplace approximation method deviates from the expected long-run success rate due to the non-Gaussian posterior distribution. In such cases, the Markov Chain Monte Carlo method still provides robust results.

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

MCMC gives reliable coverage under active constraints in this gamma unmixing setup while Laplace breaks, but the evidence is limited to synthetic runs.

read the letter

The paper applies standard Bayesian uncertainty tools to an existing hybrid spectral unmixing pipeline for gamma-ray spectrometry. It compares Laplace approximation against MCMC for producing credible intervals on radionuclide counts and spectral deformation parameters, then checks coverage via repeated Monte Carlo trials that measure long-run success rate against the nominal 95.4 percent target. When constraints on deformation and counting are inactive, both methods perform similarly and hit the target. When constraints activate or background dominates, Laplace deviates because the posterior is no longer Gaussian, while MCMC remains stable. That targeted comparison under constraints is the concrete new piece, and the validation protocol is a direct and appropriate way to test the coverage claim. The Monte Carlo design itself is clean and reproducible in principle. The main limitation is that all results come from synthetic data. The abstract and stress-test note give no indication of coverage checks on actual measured spectra in the regimes where Laplace fails, so we lack any sense of how often or how badly the non-Gaussianity matters in operational settings. Without the full model equations or data-generation details it is also hard to judge how the constraints are enforced or how sensitive the results are to background levels. This is a narrow but useful incremental contribution for people who already work with gamma-ray unmixing and need practical guidance on when the faster Laplace method is safe. A specialist reader would get a clear, evidence-based reason to prefer MCMC in constrained cases. I would send it to peer review; the validation is honest enough to deserve referee time even if real-data tests would be the obvious next step.

Referee Report

2 major / 2 minor

Summary. The paper develops two Bayesian approaches—Laplace approximation and Markov Chain Monte Carlo—for uncertainty quantification of counting rates and spectral deformation parameters in a previously proposed hybrid machine-learning spectral unmixing algorithm for γ-ray spectrometry. It evaluates coverage properties of the resulting credible intervals via repeated Monte Carlo trials on synthetic data, reporting that both methods achieve the nominal 95.4 % long-run success rate when deformation and counting constraints are inactive, but that the Laplace method deviates (due to non-Gaussian posteriors) once those constraints are active or background dominates, while MCMC remains reliable.

Significance. If the Monte Carlo coverage results generalize, the work supplies concrete guidance on method selection for reliable uncertainty quantification in constrained inverse problems that arise in operational γ-ray spectrometry. The repeated-sampling protocol that directly measures long-run success rate is a methodological strength that allows falsifiable assessment of interval calibration.

major comments (2)

[Section 4] Section 4 (Numerical experiments) and the associated figures: all reported coverage rates are obtained exclusively from synthetic spectra generated under the model; no empirical coverage frequencies are shown for real measured γ-ray spectra under active spectral-deformation or background-dominant regimes, leaving open whether the observed Laplace under-coverage is practically relevant in operational data.
[Section 3.2] Section 3.2 (Laplace approximation) and the constraint definitions in Section 2: the claim that the posterior becomes non-Gaussian when constraints are active is demonstrated only qualitatively in the simulations; the manuscript does not quantify how often or by how much the Gaussian approximation fails (e.g., via posterior skewness or coverage deficit magnitude) across the range of realistic background-to-signal ratios.

minor comments (2)

[Abstract] The abstract states that the coverage interval “corresponds closely to a credible interval,” but the precise mapping between GUM coverage and Bayesian credible intervals under the hybrid unmixing model is not restated in the methods; a short clarifying sentence would help readers.
Notation for the spectral deformation parameter λ and the counting vector is introduced without an explicit table of symbols; adding one would improve readability when comparing the two Bayesian procedures.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the detailed and constructive review. We address each major comment below and indicate planned revisions.

read point-by-point responses

Referee: [Section 4] Section 4 (Numerical experiments) and the associated figures: all reported coverage rates are obtained exclusively from synthetic spectra generated under the model; no empirical coverage frequencies are shown for real measured γ-ray spectra under active spectral-deformation or background-dominant regimes, leaving open whether the observed Laplace under-coverage is practically relevant in operational data.

Authors: We agree that validation on real data would strengthen applicability claims. However, real measured γ-ray spectra lack known ground-truth counting rates and deformation parameters, precluding direct empirical coverage computation. Synthetic data enable the repeated-sampling protocol needed to measure long-run success rates. In revision we will expand Section 4 with a dedicated discussion of this limitation, its implications for operational use, and an illustrative comparison of Laplace and MCMC uncertainty estimates on an example real spectrum. revision: partial
Referee: [Section 3.2] Section 3.2 (Laplace approximation) and the constraint definitions in Section 2: the claim that the posterior becomes non-Gaussian when constraints are active is demonstrated only qualitatively in the simulations; the manuscript does not quantify how often or by how much the Gaussian approximation fails (e.g., via posterior skewness or coverage deficit magnitude) across the range of realistic background-to-signal ratios.

Authors: We accept that a quantitative characterization is needed. The revised manuscript will augment Section 3.2 (and the associated numerical results) with explicit metrics: average posterior skewness and kurtosis, plus tabulated coverage-deficit magnitudes, evaluated across a grid of background-to-signal ratios. These additions will make the conditions under which the Laplace approximation degrades more precise. revision: yes

standing simulated objections not resolved

Direct empirical coverage frequencies on real measured γ-ray spectra, because true parameter values are unavailable in operational data.

Circularity Check

0 steps flagged

Standard Bayesian UQ procedures evaluated on external Monte Carlo ground truth; prior hybrid method referenced but not load-bearing

full rationale

The paper applies the standard Laplace approximation and MCMC sampling to quantify uncertainty in the outputs of a previously proposed hybrid unmixing algorithm. Coverage properties are assessed via repeated Monte Carlo trials that compare estimated intervals against known synthetic truth, which is external to any fitted parameter inside the present work. No equation reduces a claimed prediction to a fitted input by construction, and the referenced prior method supplies the forward model rather than defining the UQ results themselves. The single self-citation is therefore minor and non-load-bearing, yielding only a low circularity score.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard Bayesian modeling assumptions plus the existence of the prior hybrid unmixing algorithm. No new free parameters or invented entities are introduced in the abstract.

axioms (2)

domain assumption The posterior distribution can be approximated by a multivariate Gaussian when constraints are inactive
Invoked to justify the Laplace method; the abstract states this approximation fails when constraints activate.
domain assumption MCMC sampling converges to the true posterior under the given constraints
Standard MCMC assumption required for the robustness claim.

pith-pipeline@v0.9.0 · 5592 in / 1478 out tokens · 82230 ms · 2026-05-09T22:36:20.223502+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

32 extracted references

[1]

Gilmore, Practical gamma-ray spectroscopy, John Wiley & Sons, 2008

G. Gilmore, Practical gamma-ray spectroscopy, John Wiley & Sons, 2008

2008
[2]

Atomic Energy Agency, 1998

A.AC02459866, Radiologicalcharacterizationofshutdownnuclearreactors for decommissioning purposes, Internat. Atomic Energy Agency, 1998

1998
[3]

International Atomic Energy Agency, Combating illicit trafficking in nu- clear and other radioactive material, IAEA Nuclear Security 6 (2007) 3–12

2007
[4]

André, C

R. André, C. Bobin, J. Bobin, J. Xu, A. de Vismes Ott, Metrological approach ofγ-emitting radionuclides identification at low statistics: ap- plication of sparse spectral unmixing to scintillation detectors, Metrologia 58 (1) (2021) 015011

2021
[5]

J. Xu, J. Bobin, A. de Vismes Ott, C. Bobin, Sparse spectral unmixing for activity estimation inγ-ray spectrometry applied to environmental mea- surements, Applied Radiation and Isotopes 156 (2020) 108903

2020
[6]

Malfrait, J

P. Malfrait, J. Bobin, A. de Vismes Ott, Spectral unmixing of multi- temporal data in gamma-ray spectrometry, Nuclear Instruments and Meth- ods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 1045 (2023) 167547

2023
[7]

Malfrait, J

P. Malfrait, J. Bobin, A. de Vismes Ott, Online spectral unmixing in gamma-ray spectrometry, Applied Radiation and Isotopes 201 (2023) 111011

2023
[8]

C. Mano, C. Chapelle, A. D. M. Kabakian, P. Gross, H. Paradis, O. De- laune, L.Patryl, Algorithmdevelopmentforlowlevelradioxenon2dspectra analysis: A first case of study using spectral unmixing for aβ-γdetector, Applied Radiation and Isotopes 203 (2024) 111079

2024
[9]

D. T. Phan, J. Bobin, C. Thiam, C. Bobin, A hybrid machine learning un- mixing method for automatic analysis ofγ-spectra with spectral variability, Nuclear Instruments and Methods in Physics Research Section A: Accel- erators, Spectrometers, Detectors and Associated Equipment 1060 (2024) 169028

2024
[10]

BIPM, IEC, IFCC, ILAC, ISO, IUPAC, IUPAP and OIML, Evaluation of measurement data—Guide to the expression of uncertainty in measure- ment, Bureau International des Poids et Mesures JCGM 100 (2008)

2008
[11]

BIPM, IEC, IFCC, ILAC, ISO, IUPAC, IUPAP and OIML, Evaluation of measurement data-Supplement 1 to the Guide to the Expression of Uncer- tainty in Measurement, Propagation of distributions using a Monte Carlo method (2008)

2008
[12]

Stoudt, A

S. Stoudt, A. Pintar, A. Possolo, Coverage intervals, Journal of Research of the National Institute of Standards and Technology 126 (2021) 1–22. 19

2021
[13]

Elster, Bayesian uncertainty analysis compared with the application of the GUM and its supplements, Metrologia 51 (4) (2014) S159

C. Elster, Bayesian uncertainty analysis compared with the application of the GUM and its supplements, Metrologia 51 (4) (2014) S159

2014
[14]

Giaquinto, L

N. Giaquinto, L. Fabbiano, Examples of S1 coverage intervals with very good and very bad long-run success rate, Metrologia 53 (2) (2016) S65

2016
[15]

Klauenberg, C

K. Klauenberg, C. Elster, Markov chain Monte Carlo methods: an intro- ductory example, Metrologia 53 (1) (2016) S32

2016
[16]

Hall, Evaluating methods of calculating measurement uncertainty, Metrologia 45 (2) (2008) L5

B. Hall, Evaluating methods of calculating measurement uncertainty, Metrologia 45 (2) (2008) L5

2008
[17]

Willink, Probability, belief and success rate: comments on ‘on the mean- ing of coverage probabilities’, Metrologia 47 (3) (2010) 343

R. Willink, Probability, belief and success rate: comments on ‘on the mean- ing of coverage probabilities’, Metrologia 47 (3) (2010) 343

2010
[18]

Bobin, R

J. Bobin, R. C. Gertosio, C. Bobin, C. Thiam, An autoencoder-based model for learning regularizations in unmixing problems, Digital Signal Processing (2023) 104058

2023
[19]

Agostinelli, J

S. Agostinelli, J. Allison, K. a. Amako, J. Apostolakis, H. Araujo, P. Arce, M. Asai, D. Axen, S. Banerjee, G. Barrand, et al., GEANT4—a simulation toolkit, Nuclear instruments and methods in physics research section A: Accelerators, Spectrometers, Detectors and Associated Equipment 506 (3) (2003) 250–303

2003
[20]

D. T. Phan, J. Bobin, C. Thiam, C. Bobin, Automatic identification and quantification ofγ-emitting radionuclides with spectral variability using a hybrid machine learning unmixing method, Radiation Physics and Chem- istry 232 (2025) 112654

2025
[21]

A. W. Van der Vaart, Asymptotic statistics, Vol. 3, Cambridge university press, 2000

2000
[22]

M. J. Beal, Variational algorithms for approximate Bayesian inference, Uni- versity of London, University College London (United Kingdom), 2003

2003
[23]

R. E. Kass, A. E. Raftery, Bayes factors, Journal of the american statistical association 90 (430) (1995) 773–795

1995
[24]

R. A. Fisher, On the mathematical foundations of theoretical statistics, Philosophical transactions of the Royal Society of London. Series A, con- taining papers of a mathematical or physical character 222 (594-604) (1922) 309–368

1922
[25]

R. A. Fisher, Theory of statistical estimation, in: Mathematical proceed- ings of the Cambridge philosophical society, Vol. 22, Cambridge University Press, 1925, pp. 700–725. 20

1925
[26]

Paszke, S

A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala, Pytorch: An imperative style, high- performance deep learning library, in: Advances in Neural Information Processing Syste...

2019
[27]

Metropolis, A

N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, E. Teller, Equation of state calculations by fast computing machines, The journal of chemical physics 21 (6) (1953) 1087–1092

1953
[28]

Duane, A

S. Duane, A. D. Kennedy, B. J. Pendleton, D. Roweth, Hybrid monte carlo, Physics letters B 195 (2) (1987) 216–222

1987
[29]

D. J. MacKay, Information theory, inference and learning algorithms, Cam- bridge university press, 2003

2003
[30]

M. D. Hoffman, A. Gelman, et al., The No-U-Turn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo, J. Mach. Learn. Res. 15 (1) (2014) 1593–1623

2014
[31]

Bingham, J

E. Bingham, J. P. Chen, M. Jankowiak, F. Obermeyer, N. Pradhan, T. Kar- aletsos, R. Singh, P. Szerlip, P. Horsfall, N. D. Goodman, Pyro: Deep universal probabilistic programming, Journal of machine learning research 20 (28) (2019) 1–6

2019
[32]

Marchand, W

É. Marchand, W. E. Strawderman, On Bayesian credible sets, restricted parameter spaces and frequentist coverage (2013). 21

2013