Pith. sign in

REVIEW 2 major objections 5 minor 2 cited by

nDspec is a Python library that models X-ray spectra, variability power, and energy-dependent lags in one operator-based framework, demonstrated on a bright black hole binary with clear diagnosis of where simple models fail.

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 ↗

nDspec is a modular Python framework for forward-modelling multi-dimensional X-ray data, demonstrated on spectral-timing fits to a NICER observation of a black hole X-ray binary.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A solid, honest software-release paper for a genuinely missing modular spectral-timing framework; the demo shows what works and what doesn't, and the linearity caveat is standard and stated, not fatal. the 2 major comments →

arxiv 2512.10615 v2 pith:KXTW6HP7 submitted 2025-12-11 astro-ph.HE astro-ph.IM

nDspec: a new Python library for modelling multi-dimensional datasets in X-ray astronomy

Matteo Lucchini , Benjamin Ricketts , Phil Uttley , Daniela Huppenkothen This is my paper
classification astro-ph.HE astro-ph.IM
keywords multi-dimensional X-ray dataspectral-timingcross spectrumimpulse response formalismforward modellingblack hole X-ray binariesNICERPython astronomy software
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.

The reading

nDspec is the alpha release of a Python library that aims to supply what the X-ray astronomy software ecosystem lacks: a single modular framework in which astronomers forward-model the multi-dimensional datasets modern telescopes produce — time-averaged spectra, variability power spectra, and cross spectra (energy-dependent lags) — instead of analysing each dimension separately. Its design treats instrument-response convolution, Fourier transforms, and model-data comparison as composable operators, so the same model component can predict several observables and parameters can be constrained jointly across datasets. The paper demonstrates the library on a bright NICER observation of the black hole X-ray binary MAXI J1820+070: the included phenomenological models fit the power spectrum (reduced χ² ≈ 1.05) and the energy- and frequency-dependent phase lags (reduced χ² ≈ 1.76), while the full cross-spectrum fit fails (reduced χ² ≈ 2.99), with residuals pointing to a missing relativistic reverberation component. If the framework works as claimed, it lowers the barrier to joint spectral-timing modelling and provides a common base for future extensions such as polarimetry, non-Gaussian likelihoods, and more advanced samplers. The authors are explicit that the cross-spectrum machinery rests on a linearity assumption whose breakdown is beyond the present scope.

Core claim

nDspec's central claim is that the linear impulse-response formalism, embedded in an operator-based software design, is sufficient to build a general-purpose framework for multi-dimensional X-ray modelling. The time-dependent flux in an energy band is written f(E,t) = s(t) ⊛ g(E,t): a driving variability signal convolved with an energy- and time-dependent impulse response. In the Fourier domain the cross spectrum between any band and a reference band becomes C(E,ν) = P(ν) G(E,ν) G*(E_ref,ν) — all energy dependence of timing products is carried by the transfer function G(E,ν), while the driving power spectrum P(ν) acts as a frequency-dependent weight. The library chains operator classes that

What carries the argument

Two elements carry the argument. First, the linear impulse response (IRF) formalism: the system's response to a delta-function flash is g(E,t), and its Fourier transform G(E,ν) is the transfer function; the convolution theorem turns time-domain convolution into Fourier-domain multiplication, so cross spectra are computed as C(E,ν) = P(ν) G(E,ν) G*(E_ref,ν). Second, the operator framework: each transformation applied to a model — folding through the detector response matrix, Fourier transformation, power/cross-spectrum construction, comparison with data — is a composable class, so a single model, whether defined in the time domain, as a transfer function, or directly as a cross spectrum, can

Load-bearing premise

The cross-spectrum machinery stands on the assumption that source variability is linear: flux in every energy band is a convolution of one driving signal with a fixed impulse response, so the intrinsic coherence is exactly one and higher-order Fourier products vanish. If a source's variability is significantly nonlinear, predicted lags and cross-spectrum amplitudes will be wrong no matter how good the model components are — a limitation the paper itself acknowledges as beyond

What would settle it

Measure the intrinsic coherence of a bright accreting source over a wide range of Fourier frequencies and energies and look for a significant dip below unity where the transfer function is well constrained, or for detectable bispectral power. Because the formalism fixes coherence at unity, such a detection would make any full cross-spectrum fit systematically wrong at those frequencies, independent of the impulse response chosen. A cheaper version: generate a synthetic dataset from a known nonlinear process (multiplicative or squared variability), fit it with the library, and show that no line

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

If this is right

  • The same model component can be fitted simultaneously to a time-averaged spectrum, a power spectrum, and a lag spectrum, pulling constraints from all three observables rather than analysing each dimension in isolation.
  • The demonstration quantifies what current phenomenological models can and cannot do: power spectrum (reduced χ² = 1.05) and energy-dependent lags (reduced χ² = 1.76) fit well, while the full cross spectrum (reduced χ² = 2.99) shows structured residuals at soft energies and near 6 keV, localizing the missing physics to relativistic reverberation.
  • Because the Fourier transform can be computed on arbitrary logarithmic grids through a sinc-function decomposition, reverberation models requiring fine time resolution at short delays can be evaluated without uniform oversampling.
  • The operator design allows new dimensions — polarimetry, multi-wavelength data, phase-resolved spectra — to be added as new operators and model components, and later releases plan non-Gaussian likelihoods and more advanced sampling methods.

Where Pith is reading between the lines

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

  • The joint pattern of residuals — near 6 keV in both the time-averaged spectrum and the cross-spectrum modulus — suggests the cross-spectrum modulus is a sharper test of reverberation models than phase lags alone; a plausible next step is to make full modulus-plus-phase fits the benchmark for new physical models rather than lag-energy spectra.
  • The linearity assumption draws the framework's boundary: any source with intrinsic coherence measurably below unity, or with significant bispectral power, falls outside what the current cross-spectrum class can represent; introducing a coherence parameter or a nonlinear impulse-response treatment would extend the formalism without redesigning the operator chain.
  • Because all energy dependence of timing products flows through the transfer function G(E,ν), testing competing geometries (a lamppost corona versus extended Comptonization, for instance) reduces to substituting different impulse responses — a comparison that the current phenomenological components only hint at, but that the operator chain already supports.
  • The paper's repeated warnings about energy-rebinning and unfolding amount to a methodological recommendation: for multi-dimensional products such as rms and lag-energy spectra, forward-folding through the full instrument response is the safe default, and the library is designed around that path.
Share X Bluesky LinkedIn Reddit HN

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

2 major / 5 minor

Summary. This paper introduces nDspec, an open-source Python library for forward-modelling multi-dimensional X-ray data. The library currently supports time-averaged spectra, Fourier power spectra, and cross spectra (phase lags, modulus, complex products), using an operator-based architecture: instrument response folding, Fourier transforms (FFT and sinc decomposition), and inference via χ² optimization and emcee MCMC. Unit tests verify response folding against Xspec to 1e-6. The demonstration on a NICER observation of MAXI J1820+070 includes: a 3-Lorentzian fit to the power spectrum (χ²/dof=1.05), a diskbb+powerlaw fit to the time-averaged spectrum (χ²/dof=2.49 with residuals attributed to reflection), and a pivoting-powerlaw + reverberation model fit to energy-dependent phase lags (χ²/dof=1.76). A full modulus+phase cross-spectrum fit is also attempted and fails (χ²/dof=2.99), which the authors interpret as missing relativistic reverberation. Limitations, including the linear IRF assumption and the lack of non-Gaussian likelihoods, are acknowledged.

Significance. The paper addresses a genuine gap: existing X-ray modelling packages are largely 1D or instrument-specific, whereas nDspec provides a modular framework where models, responses, Fourier transforms, and inference are composable. If the software performs as claimed, it will enable joint spectral-timing-polarimetric modelling and is likely to become a useful community tool. Strengths include a reproducibility package (Zenodo), open-source code, and machine-checked agreement with Xspec for response folding. The honest reporting of the failed full cross-spectrum fit is also a strength: it demonstrates that the framework can identify model inadequacy rather than forcing an overparameterized fit. The alpha-release limitations (Gaussian likelihood only, no Xspec wrapper, no polarimetry) are clearly stated and are not obstacles to the underlying design.

major comments (2)
  1. [Sec. 3.4, phase renormalization] The text contains a direct contradiction. It states 'We begin as previously with least-squares optimization, without enabling the phase re-normalization... The results are shown in Fig.6 and 7; ... χ2/d.o.f.=397.92/226=1.76.' Two paragraphs later: 'we found that enabling the phase renormalization ... improved the quality of all fits without affecting the other parameters, so in the manuscript we only report fits in which it has been enabled.' This makes it impossible to determine whether the reported lag fit (and the corresponding parameters in Table 3) were obtained with or without the six extra phase parameters. Please clarify the exact sequence and state in the figure/table captions which configuration is being reported. This is needed for reproducibility of the central demonstration.
  2. [Sec. 2.2 / Sec. 3.4, linearity assumption] The cross-spectrum machinery is built on the linear IRF formalism (Eqs. 2–4), which assumes unity intrinsic coherence and zero bispectrum. The paper acknowledges this, but the full cross-spectrum demonstration does not test it. The fit fails with χ2/dof=2.99 and the residuals are interpreted as missing relativistic reverberation. An equally plausible interpretation is that the source variability is nonlinear or intrinsically incoherent, in which case Eq. (4) would be invalid irrespective of the model. I recommend computing the intrinsic coherence of the data (e.g., from the averaged cross spectrum and the two power spectra) or fitting a direct Fourier-domain cross-spectrum model as a control. This would either support the reverberation interpretation or reveal a more fundamental limitation. This is a suggestion to strengthen the interpretation; it does not invalidate the software claim,
minor comments (5)
  1. [Eq. (4)] The notation S²(ν) should be |S(ν)|²; the power spectrum is |S(ν)|², not S(ν)², since S(ν) is complex. The accompanying text 'S²(ν) is, by definition, the power spectrum' is therefore incorrect.
  2. [Sec. 2.2, near Eq. (3)] The sentence describing 'the transfer function S(E,ν)' should read G(E,ν); in Eq. (3), S is the transform of the driving signal and G is the transfer function.
  3. [Table 4] The entry for s_t reports '0+*;* −0.001'; please clarify the notation for a parameter pinned at a limit.
  4. [Fig. 7 and Fig. 9 captions] The captions describe 'the same model shown in fig.6' and 'fig.8'; this is redundant and can be confusing when figures are viewed separately.
  5. [Sec. 2.3.1] The text states that only Gaussian uncertainties are supported. It may be worth noting explicitly that χ² fitting of cross-spectrum modulus and phase treats those estimates as Gaussian, which is an approximation for finite integration lengths.

Circularity Check

0 steps flagged

No significant circularity: nDspec is a forward-modelling software framework whose claims are demonstrated on independent NICER data, with fits and failures reported transparently.

full rationale

No circular step found. nDspec's central claim is that it provides a modular framework for forward-modelling multi-dimensional X-ray data, demonstrated by fitting NICER observations of MAXI J1820+070. These are genuine fits: the power spectrum (chi2/dof=1.05), time-averaged spectrum (2.49), lag spectra (1.76) and full cross spectrum (2.99) are evaluated against the data, and the poor full cross-spectrum fit is reported rather than concealed. The linear impulse response formalism in Sec. 2.2 (Eqs. 2-4) is an explicit modelling assumption, not a derived prediction, and its limitation is openly stated: 'A full treatment of non-linear impulse response functions is beyond the scope of this paper.' The modulus renormalization constants m1-m6 are fitted nuisance parameters, not predictions, and the paper notes their degeneracy with normalization parameters. Self-citations, including Uttley & Malzac (2025) for the sinc transform and phenomenological models, provide algorithms or standard constructs rather than forcing conclusions; the response folding is unit-tested against Xspec to one part in a million, and a reproduction package is provided. Nothing in the paper reduces a claimed result to its own input by construction.

Axiom & Free-Parameter Ledger

13 free parameters · 7 axioms · 0 invented entities

No new physical entities (particles, forces, dimensions) are postulated. The library's model components (pivoting_pl, bbody_bkn, gauss_fred, etc.) are phenomenological and explicitly described as such, not claims of new physical objects. The demo relies on a substantial set of free parameters, but these are standard for spectral-timing fits and do not affect the core software claim.

free parameters (13)
  • gamma_0 = 4.9e-2 (posterior median)
    Fractional amplitude of photon-index variability in pivoting_pl at nu0=0.2 Hz; fitted to NICER lag spectra (Table 3).
  • s_gamma = -7e-3
    Frequency-scaling slope of gamma in pivoting_pl; fitted to lag spectra (Table 3).
  • phi_0 = -1.9
    Initial phase between power-law normalization and index changes; fitted to lag spectra (Table 3).
  • s_phi = 2.2
    Frequency-scaling slope of phi; fitted to lag spectra (Table 3).
  • A0 = 6.7 (posterior median, cross-spectrum fit)
    Pivoting power-law normalization freed in the cross-spectrum fit (Table 4).
  • R0 = 115 (posterior median)
    Disc irradiation IRF normalization in bbody_bkn component (Table 4).
  • kT0 = 0.272 keV
    Initial temperature of the disc IRF in bbody_bkn (Table 4).
  • s2 = -2.10 (posterior median)
    Decay slope of the disc IRF broken power-law in time (Table 4).
  • st = 0 (pinned to limit)
    Temperature time-scaling slope; parameter hits its prior bound (Table 4).
  • tbrk (fixed) = 0.01 s
    IRF rise timescale fixed by hand because unconstrained by data (Section 3.4).
  • s1 (fixed) = 4
    IRF rise slope fixed by hand because unconstrained by data (Section 3.4).
  • Phi_1..Phi_6 = approx -0.011, 0.007, 0, 0.011, 0.016, 0.014 rad
    Per-frequency-bin phase renormalization constants to correct instrument-calibration phase artifacts (Table 4).
  • m_1..m_6 = approx 0.80, 1.38, 0.99, 0.73, 0.84, 1.10
    Per-frequency-bin modulus renormalization constants when fitting the full cross spectrum (Table 4).
axioms (7)
  • domain assumption Flux can be described by a linear impulse response: f(E,t)=s(t) g(E,t) (Eq. 2), so intrinsic coherence is unity and higher-order Fourier products vanish.
    Invoked in Section 2.2 as the basis for cross-spectrum modelling; nonlinear variability is explicitly deferred.
  • standard math Convolution theorem: the Fourier transform of a convolution is a product (Eq. 3).
    Core step in Section 2.2 for converting time-domain IRFs to the Fourier domain.
  • domain assumption Instrument response matrices accurately map incident flux to detector channels and are non-invertible, necessitating forward-folding.
    Basis of the ResponseMatrix class (Section 2.1); unit tests claim agreement with Xspec to 1e-6.
  • domain assumption Gaussian (chi-square) likelihood is adequate for the spectral-timing products supported in the alpha release.
    Stated in Section 2.3 as the only supported statistic; Cash and Whittle likelihoods are promised in future releases.
  • domain assumption Instrumental background is negligible for the bright NICER observation used in the demo.
    Section 3.1: background is omitted due to the >=1e4 cts/s count rate.
  • domain assumption The sinc-function decomposition (Uttley & Malzac 2025) provides correct Fourier transforms on arbitrary frequency/time grids.
    Used in Section 2.2 as an alternative to FFTW; the method is cited from a co-authored paper and not re-derived.
  • domain assumption tbabs (Wilms et al. 2000) correctly models interstellar absorption for the demo fits.
    Used in Sections 3.3 and 3.4 for time-averaged, lag, and cross-spectrum fits.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of nDspec: a new Python library for modelling multi-dimensional datasets in X-ray astronomy." pith.science (2026). https://pith.science/paper/KXTW6HP7

@misc{pith2026251210615,
  author       = {Pith},
  title        = {Pith review of: nDspec: a new Python library for modelling multi-dimensional datasets in X-ray astronomy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KXTW6HP7}},
  note         = {Machine review of arXiv:2512.10615}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The current fleet of X-ray telescopes produces a wealth of multi-dimensional data, allowing us to study sources in time, photon energy, and polarisation. At the same time, it has become increasingly clear that progress in our physical understanding will only come from studying these sources in multiple dimensions simultaneously. Enabling multi-dimensional studies of X-ray sources requires new theoretical models predicting these datasets, new methods to analyse them, and, crucially, a software framework to combine data, models, and methods efficiently. However, the current ecosystem of software packages developed for X-ray data analysis does not provide the flexibility for advanced modelling of multi-dimensional datasets. In this paper, we introduce nDspec, a new python-based library designed to allow users to seamlessly model one- and multi-dimensional datasets common to X-ray astronomy. Unlike most other libraries, it is designed as a flexible, modular, and extensible framework capable of accommodating multi-dimensional data and able to connect to a range of different inference tools and algorithms. Here we focus on modelling timing and spectral-timing data as a function of both Fourier frequency and energy, in addition to limited support for time-averaged spectra. We discuss design philosophy and current features, and showcase an example use case by characterising a NICER observation of a black hole X-ray binary. We also highlight plans for extensions to other dimensions and new features, such as the inclusion of polarimetry and the improved statistical methods for Bayesian inference.

Figures

Figures reproduced from arXiv: 2512.10615 by Benjamin Ricketts, Daniela Huppenkothen, Matteo Lucchini, Phil Uttley.

Figure 1
Figure 1. Figure 1: The NICER instrument response with different binning. Top left: the response in its native resolution, 3451 physical energy bins×1501 channels (roughly containing 100 channels per keV). Top right: the response re-binned over channels down to 30 geometrically-spaced energy channels. Bottom left: the response re-binned over energies by a factor 30. Bottom right: the response re-binned over energies by a fact… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of nDspec folding a constant model. The blue, or￾ange, green and red lines correspond to the top left, top right, bottom left and bottom right panels, respectively. As shown by the green line, re-binning over energies introduces distortions in the spectrum if the number of energy bins is not larger than the number of channels. On the other hand, re-binning over channels (orange line) or over bot… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of a physical model before folding with the Swift?XRT response matrix (blue line), and of the same model folded and then un-folded in nDspec (orange line). The two agree over most of the energy range, but there is a noticeable discrepancy at energies below ≈ 1 keV. response matrix is not strictly invertible, the deconvolution of the intrinsic spectrum with the instrument response necessarily req… view at source ↗
Figure 4
Figure 4. Figure 4: Data (black) together with the median of the posterior distribu￾tion (red solid line) and random draws from the posterior (pink lines). In addition to the data and total model, we also show the individual Lorentzian components for the posterior median only. While the contri￾bution of the first Lorentzian is easily separated, the second and third overlap. The residuals are calculated using the posterior med… view at source ↗
Figure 5
Figure 5. Figure 5: The time-averaged spectrum (black) with the median of the pos￾terior distribution (red) and random draws from the posterior (pink). The posteriors are very narrow, such that the posterior draws signifi￾cantly overlap with each other and the posterior median as to be nearly invisible. The residuals are calculated using the posterior median only. Reflection features clearly visible in the residuals; modellin… view at source ↗
Figure 6
Figure 6. Figure 6: Energy-dependent time lags, showing the data as a function of both Fourier frequency and energy in units of phase (left plot), the posterior median model in the same units (middle plot), and the residuals in units of ∆χ (right plot). Our phenomenological model reproduces the data very well at all energies and Fourier frequencies. 0.00 0.05 0.10 Lag (s) 0.2-0.4 Hz 0.4-0.9 Hz 0.9-1.8 Hz 1.8-3.7 Hz 3.7-7.7 Hz… view at source ↗
Figure 7
Figure 7. Figure 7: The same model shown in fig.6, but using a traditional one￾dimensional plot to represent the data, posterior median model and residuals. sϕ frequency scaling parameter for the phase between the vari￾ability in Γ and A Ecut low energy cutoff in the powerlaw emission R0 disc irradiation IRF normalization kT0 disc IRF initial temperature tbrk disc IRF rise timescale s1 disc IRF rise slope s2 disc IRF decay sl… view at source ↗
Figure 8
Figure 8. Figure 8: Full cross spectrum, shown as a two-dimensional plot. The top row is the data as a function of Fourier frequency and energy, the middle row shows the posterior median model in the same units, and the bottom row shows the residuals in units of ∆χ. The left column shows the modulus (in units of absolute rms) and the right column shows the phase (in units of radians). The model shows strong structured residua… view at source ↗
Figure 9
Figure 9. Figure 9: The same model shown in fig.8, but using a traditional one-dimensional plot to represent the data, posterior median model and residuals. The left column shows the modulus in units of absolute rms, and the right column shows the phase in units of radians. Parameter least-χ 2 Posterior A0 6.2 6.7 +0.1 −0.1 γ0 8.5 · 10−2 8.5 +0.3 −0.3 · 10−2 sγ −1.7 · 10−2 −1.70+0.03 −0.03 · 10−2 ϕ0 −1.9 −1.95+0.07 −0.06 sϕ 2… view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. Energy-resolved pulse profiles of Vela X-1: cross-calibrating XMM-Newton and NuSTAR to trace spectral features

    astro-ph.HE 2026-05 unverdicted novelty 5.0

    Pulsed fraction spectra from XMM-Newton and NuSTAR on Vela X-1 agree within 5% after instrumental corrections and reveal emission-line and cyclotron features whose strength varies with orbital absorption.

  2. Multi-wavelength outburst activity from EP J174942.2-384834: a very faint X-ray transient discovered by Einstein Probe

    astro-ph.HE 2026-05 accept novelty 5.0

    EP J174942.2-384834 is classified as a very faint X-ray transient black hole candidate based on its hard X-ray spectra, optical/UV brightening correlated with X-rays, and lack of radio emission.

Reference graph

Works this paper leans on

55 extracted references · 2 linked inside Pith · cited by 2 Pith papers

  1. [1]

    & Uttley, P

    Arévalo, P. & Uttley, P. 2006, MNRAS, 367, 801

    2006

  2. [2]

    Arnaud, K. A. 1996, in Astronomical Society of the Pacific Conference Series, V ol. 101, Astronomical Data Analysis Software and Systems V , ed. G. H. Jacoby & J. Barnes, 17 Astropy Collaboration, Price-Whelan, A. M., Lim, P. L., et al. 2022, ApJ, 935, 167 Astropy Collaboration, Price-Whelan, A. M., Sip˝ocz, B. M., et al. 2018, AJ, 156, 123 Astropy Collab...

    1996

  3. [3]

    2024, Journal of Open Source Software, 9, 7389

    Bachetti, M., Huppenkothen, D., Stevens, A., et al. 2024, Journal of Open Source Software, 9, 7389

    2024

  4. [4]

    D., et al

    Baldini, L., Bucciantini, N., Lalla, N. D., et al. 2022, SoftwareX, 19, 101194

    2022

  5. [5]

    2022, MNRAS, 515, 2099

    Bellavita, C., García, F., Méndez, M., & Karpouzas, K. 2022, MNRAS, 515, 2099

    2022

  6. [6]

    2018, JAX: composable transforma- tions of Python+NumPy programs

    Bradbury, J., Frostig, R., Hawkins, P., et al. 2018, JAX: composable transforma- tions of Python+NumPy programs

    2018

  7. [7]

    2021, The Journal of Open Source Software, 6, 3001

    Buchner, J. 2021, The Journal of Open Source Software, 6, 3001

    2021

  8. [8]

    2023, Statistics Surveys, 17, 169

    Buchner, J. 2023, Statistics Surveys, 17, 169

    2023

  9. [9]

    & Boorman, P

    Buchner, J. & Boorman, P. 2023, in Handbook of X-ray and Gamma-ray Astro- physics, 150

    2023

  10. [10]

    2014, A&A, 564, A125

    Buchner, J., Georgakakis, A., Nandra, K., et al. 2014, A&A, 564, A125

    2014

  11. [11]

    & Stella, L

    Campana, S. & Stella, L. 1995, MNRAS, 272, 585

    1995

  12. [12]

    & Young, A

    Chainakun, P. & Young, A. J. 2017, MNRAS, 465, 3965

    2017

  13. [13]

    2020, Proceedings of the National Academy of Science, 117, 30055 De Marco, B., Ponti, G., Petrucci, P

    Cranmer, K., Brehmer, J., & Louppe, G. 2020, Proceedings of the National Academy of Science, 117, 30055 De Marco, B., Ponti, G., Petrucci, P. O., et al. 2017, MNRAS, 471, 1475 de Vries, M. N., Wise, M. W., Huppenkothen, D., et al. 2018, MNRAS, 478, 4010 Dupourqué, S., Barret, D., Diez, C. M., Guillot, S., & Quintin, E. 2024, A&A, 690, A317

    2020

  14. [14]

    2025, arXiv e-prints, arXiv:2507.15461

    Ewing, M., Ingram, A., Rankin, J., & Muleri, F. 2025, arXiv e-prints, arXiv:2507.15461

    Pith/arXiv arXiv 2025

  15. [15]

    & Hobson, M

    Feroz, F. & Hobson, M. P. 2008, MNRAS, 384, 449

    2008

  16. [16]

    P., & Bridges, M

    Feroz, F., Hobson, M. P., & Bridges, M. 2009, MNRAS, 398, 1601

    2009

  17. [17]

    P., Cameron, E., & Pettitt, A

    Feroz, F., Hobson, M. P., Cameron, E., & Pettitt, A. N. 2019, The Open Journal of Astrophysics, 2, 10

    2019

  18. [18]

    W., Lang, D., & Goodman, J

    Foreman-Mackey, D., Hogg, D. W., Lang, D., & Goodman, J. 2013, PASP, 125, 306

    2013

  19. [19]

    2001, in Society of Photo-Optical In- strumentation Engineers (SPIE) Conference Series, V ol

    Freeman, P., Doe, S., & Siemiginowska, A. 2001, in Society of Photo-Optical In- strumentation Engineers (SPIE) Conference Series, V ol. 4477, Astronomical Data Analysis, ed. J.-L. Starck & F. D. Murtagh, 76–87

    2001

  20. [20]

    & Calderhead, B

    Girolami, M. & Calderhead, B. 2011, Journal of the Royal Statistical Society Series B: Statistical Methodology, 73, 123

    2011

  21. [21]

    Gunderson, S. J. & Huenemoerder, D. P. 2025, AJ, 169, 2

    2025

  22. [22]

    R., Millman, K

    Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357

    2020

  23. [23]

    Hunter, J. D. 2007, Computing in Science & Engineering, 9, 90

    2007

  24. [24]

    L., et al

    Huppenkothen, D., Bachetti, M., Stevens, A. L., et al. 2019, apj, 881, 39

    2019

  25. [25]

    2022, in Handbook of X-ray and Gamma-ray Astrophysics, ed

    Ingram, A. 2022, in Handbook of X-ray and Gamma-ray Astrophysics, ed. C. Bambi & A. Sangangelo, 113

    2022

  26. [26]

    J., Poutanen, J., & Krawczynski, H

    Ingram, A., Maccarone, T. J., Poutanen, J., & Krawczynski, H. 2015, ApJ, 807, 53

    2015

  27. [27]

    2016, MNRAS, 461, 1967

    Ingram, A., van der Klis, M., Middleton, M., et al. 2016, MNRAS, 461, 1967

    2016

  28. [28]

    Ingram, A. R. & Maccarone, T. J. 2017, MNRAS, 471, 4206

    2017

  29. [29]

    S., Raassen, A

    Kaastra, J. S., Raassen, A. J. J., de Plaa, J., & Gu, L. 2025, SPEX X-ray spectral fitting package

    2025

  30. [30]

    2001, MNRAS, 327, 799

    Kotov, O., Churazov, E., & Gilfanov, M. 2001, MNRAS, 327, 799

    2001

  31. [31]

    2022, arXiv e-prints, arXiv:2206.09972

    Krawczynski, H., Muleri, F., Dov ˇciak, M., et al. 2022, arXiv e-prints, arXiv:2206.09972

    Pith/arXiv arXiv 2022

  32. [32]

    2024, A&A, 686, A259

    Lascar, J., Bobin, J., & Acero, F. 2024, A&A, 686, A259

    2024

  33. [33]

    2023, ApJ, 951, 19

    Lucchini, M., Mastroserio, G., Wang, J., et al. 2023, ApJ, 951, 19

    2023

  34. [34]

    2018, MNRAS, 475, 4027

    Mastroserio, G., Ingram, A., & van der Klis, M. 2018, MNRAS, 475, 4027

    2018

  35. [35]

    2020, MNRAS, 498, 4971

    Mastroserio, G., Ingram, A., & van der Klis, M. 2020, MNRAS, 498, 4971

    2020

  36. [36]

    2021, MNRAS, 507, 55 Méndez, M., Peirano, V ., García, F., et al

    Mastroserio, G., Ingram, A., Wang, J., et al. 2021, MNRAS, 507, 55 Méndez, M., Peirano, V ., García, F., et al. 2024, MNRAS, 527, 9405

    2021

  37. [37]

    2022, MNRAS, 511, 255

    Nathan, E., Ingram, A., Homan, J., et al. 2022, MNRAS, 511, 255

    2022

  38. [38]

    2025, LMFIT: Non-Linear Least- Squares Minimization and Curve-Fitting for Python

    Newville, M., Otten, R., Nelson, A., et al. 2025, LMFIT: Non-Linear Least- Squares Minimization and Curve-Fitting for Python

    2025

  39. [39]

    2014, MNRAS, 440, 2882

    Rapisarda, S., Ingram, A., & van der Klis, M. 2014, MNRAS, 440, 2882

    2014

  40. [40]

    J., Huppenkothen, D., Lucchini, M., et al

    Ricketts, B. J., Huppenkothen, D., Lucchini, M., et al. 2025, MNRAS, 538, 1096

    2025

  41. [41]

    E., Choudhury, D., Salmi, T., et al

    Riley, T. E., Choudhury, D., Salmi, T., et al. 2023, Journal of Open Source Soft- ware, 8, 4977

    2023

  42. [42]

    C., et al

    Rogantini, D., Costantini, E., Gallo, L. C., et al. 2022, MNRAS, 516, 5171

    2022

  43. [43]

    Shakura, N. I. & Sunyaev, R. A. 1973, A&A, 500, 33

    1973

  44. [44]

    Stevens, A. L. & Uttley, P. 2016, MNRAS, 460, 2796

    2016

  45. [45]

    L., Uttley, P., Altamirano, D., et al

    Stevens, A. L., Uttley, P., Altamirano, D., et al. 2018, ApJ, 865, L15

    2018

  46. [46]

    M., Fabian, A

    Uttley, P., Cackett, E. M., Fabian, A. C., Kara, E., & Wilkins, D. R. 2014, A&A Rev., 22, 72

    2014

  47. [47]

    & Malzac, J

    Uttley, P. & Malzac, J. 2025, MNRAS, 536, 3284

    2025

  48. [48]

    2011, MNRAS, 414, L60 van Paradijs, J

    Uttley, P., Wilkinson, T., Cassatella, P., et al. 2011, MNRAS, 414, L60 van Paradijs, J. & Lewin, H. G. 1986, A&A, 157, L10

    2011

  49. [49]

    2010, MNRAS, 402, 307

    Vaughan, S. 2010, MNRAS, 402, 307

    2010

  50. [50]

    J., & Castro, D

    Vink, J., Patnaude, D. J., & Castro, D. 2022, ApJ, 929, 57

    2022

  51. [51]

    E., et al

    Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020, Nature Methods, 17, 261

    2020

  52. [52]

    2022, ApJ, 930, 18

    Wang, J., Kara, E., Lucchini, M., et al. 2022, ApJ, 930, 18

    2022

  53. [53]

    2021, ApJ, 910, L3

    Wang, J., Mastroserio, G., Kara, E., et al. 2021, ApJ, 910, L3

    2021

  54. [54]

    R., Cackett, E

    Wilkins, D. R., Cackett, E. M., Fabian, A. C., & Reynolds, C. S. 2016, MNRAS, 458, 200

    2016

  55. [55]

    2000, ApJ, 542, 914 Article number, page 13 of 14 A&A proofs:manuscript no

    Wilms, J., Allen, A., & McCray, R. 2000, ApJ, 542, 914 Article number, page 13 of 14 A&A proofs:manuscript no. aanda Appendix A: Posterior distributions Here we report the posterior distributions for all the models. Fig..1 shows the corner plot for the fit of the power spectrum. Fig..2 shows the corner plot for the fit of the time-averaged spec- trum. Fig...

    arXiv 2000

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.