Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

This paper presents CLiENT, a neural network that emulates the cosmological likelihood directly from parameters, achieving credible intervals within 0.1σ of the true likelihood using fewer than 20,000 evaluations.

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 ↗

A neural-network tool that directly emulates cosmological likelihood functions, achieving ~0.1σ-accurate credible intervals with roughly 20,000 likelihood evaluations.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Useful likelihood-emulation tool with a real precision claim, but the headline numbers need qualification: the evaluation count is off and the Planck reference chains are not validated. the 5 major comments →

arxiv 2512.17509 v2 pith:TBVB6GSZ submitted 2025-12-19 astro-ph.CO astro-ph.IM

CLiENT: A new tool for emulating cosmological likelihoods using deep neural networks

Luca Janken , Steen Hannestad , Thomas Tram , Andreas Nygaard This is my paper
classification astro-ph.CO astro-ph.IM
keywords cosmological likelihood emulationneural networkactive learningMarkov Chain Monte Carloauto-differentiable surrogatePlanck CMB dataparameter inferenceBayesian evidence
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

The paper introduces CLiENT, a neural network that learns the likelihood function itself—rather than cosmological observables that feed into it—so that the trained surrogate can be used for parameter inference in place of the full pipeline. The core idea is to gather training points with a density proportional to a tempered version of the true likelihood, concentrating the network's capacity on the parameter regions that shape posterior credible intervals. Across four test cases, including a 29-dimensional non-Gaussian analytic likelihood and two configurations of the Planck CMB likelihood, the emulator recovers credible intervals to better than about 0.1σ and single-point χ² errors around 0.3–0.5, using fewer than 20,000 full likelihood evaluations. If the claim holds, expensive cosmological inference and model comparison become dramatically cheaper, and the fully differentiable surrogate enables gradient-based statistical methods that are otherwise inaccessible.

Core claim

CLiENT directly emulates the log-likelihood as a function of all cosmological and nuisance parameters, trained on a cloud of points whose density is proportional to L^{1/T_T} with a tunable temperature T_T. The active-learning scheme uses a k-nearest-neighbour density estimate to accept new training points only when the local density is below the tempered likelihood target, which makes the training data concentrate where the posterior is shaped. A relative-error loss with a scale set by a Wilson–Hilferty approximation of the χ² distribution ensures the network is accurate near the best fit. The trained surrogate reproduces the posterior credible intervals of the full Planck likelihood to wit

What carries the argument

The central mechanism is the tempered training-point selection: at each iteration, candidate parameters are proposed by an MCMC run over the current surrogate, and accepted only when the ratio of the k-nearest-neighbour density estimate to the tempered true likelihood is below a threshold. This drives the training cloud's density to follow L^{1/T_T}, so the network learns the likelihood only where it matters. A scale-relative MSRE loss penalizes relative error in χ², with the scale epsilon chosen via a Wilson–Hilferty approximation of the χ² distribution. The network uses a learnable activation function with trainable slope and transition width, in a five-layer, 512-neuron architecture.

Load-bearing premise

The training-point selection assumes that a k-nearest-neighbour density estimate in 27–29 dimensions gives an accurate local density, so that the acceptance criterion ρ_p / L(θ_p)^{1/T_T} < c yields training points with density proportional to L^{1/T_T}; if this proxy fails in high dimensions or along narrow non-Gaussian directions, the promised 0.1σ credible-interval accuracy is not guaranteed.

What would settle it

Run the training pipeline on a synthetic likelihood with multiple well-separated modes in, say, 30 dimensions, using the same hyperparameters. If the accepted training points do not populate each mode in proportion to L^{1/T_T}, or if the resulting credible intervals deviate from the true likelihood by more than about 0.1σ, the central claim fails. Additionally, vary k from 5 to 50 on the Planck likelihood and check whether the reported Δχ² and credible intervals remain stable; a strong dependence on k would indicate that the density estimate is fragile.

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

If this is right

  • A full MCMC parameter inference for a Planck-scale likelihood can be replaced by a one-time emulator training of roughly 20,000 evaluations, cutting the compute time from hours to minutes.
  • Because the surrogate is differentiable end-to-end, gradient-based inference techniques such as Hamiltonian Monte Carlo, variational inference, and maximum-likelihood optimization become directly applicable to the likelihood surface.
  • Bayesian evidence estimation becomes more feasible, since the smooth emulator can be inexpensively evaluated in the tails if the training data cover them.
  • The method is likelihood-agnostic and works on any continuous likelihood, so it can be applied to other cosmological surveys and to non-cosmological problems without modification.
  • Profile likelihoods can be computed rapidly by optimizing the differentiable surrogate, which is otherwise expensive with a non-differentiable likelihood.

Where Pith is reading between the lines

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

  • The k-nearest-neighbour density estimate is a crude proxy; in very high dimensions or with severe funnel geometries the acceptance criterion may misallocate training points, which the authors themselves observe as slower convergence along non-Gaussian directions.
  • The 0.1σ credible-interval claim is based on the worst single parameter's credible metric; a stricter test would compare the full 1D marginal KL divergences or the joint posterior, which may be more sensitive.
  • One could stress-test the method by varying the k in the density estimate and the temperature T_T, and by using a likelihood with multiple separated modes, to see whether the accepted training set still follows the target density L^{1/T_T}.
  • The cost advantage over observable emulators is clearest when the likelihood itself is the bottleneck; if the likelihood is cheap but observables are expensive, the balance may shift.
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

5 major / 5 minor

Summary. The manuscript introduces CLiENT, a deep neural network framework that directly emulates the cosmological likelihood function as a map from model and nuisance parameters to log-likelihood, bypassing the usual two-step observable emulation plus likelihood evaluation. Training points are acquired iteratively: a surrogate network is trained, used to run an MCMC, and new points are accepted according to a criterion designed to make the training density proportional to a tempered version of the true likelihood. The method is tested on two synthetic likelihoods (a 27-dimensional Gaussian and a 29-dimensional case with two non-Gaussian directions) and on Planck 2018 data for ΛCDM and ΛCDM+(Neff,s,ms). The paper claims credible-interval recovery to better than 0.1σ (as defined by its own metric) with fewer than 2×10^4 likelihood evaluations, and single-point emulator precision better than Δχ²∼0.5, concluding that the approach is competitive with observable emulators while offering a fully differentiable surrogate.

Significance. If the quantitative claims hold, CLiENT is a useful contribution to cosmological parameter inference: direct likelihood emulation removes the need for repeated Boltzmann-solver evaluations and provides a surrogate that is completely auto-differentiable, enabling gradient-based methods. The paper's strengths include a public code release, a clear description of the acquisition scheme, and benchmarking on both analytic and real likelihoods. However, the headline claims contain internal inconsistencies and rest on reference chains whose convergence is not demonstrated. These issues must be resolved before the performance numbers can be taken at face value.

major comments (5)
  1. [Abstract; §7; Table 1; Fig. 7 caption] The claim of 'less than 2×10^4 function evaluations' is contradicted by the paper's own numbers. Table 1 and the Fig. 7 caption report 22,728 total true-likelihood evaluations for the ΛCDM run, and Fig. 9 reports 22,024 for the sterile run. These exceed the claimed threshold. The paper should either revise the headline claim, clarify the counting (e.g., if only accepted points count, 5000+12654=17654, but the total evaluations still exceed 20k for ΛCDM), or demonstrate a run that satisfies the stated bound.
  2. [§6.2] The 0.1σ credible-interval recovery for the real Planck data is measured against MontePython/class reference chains, but the paper provides no convergence diagnostics for these references—no chain lengths, effective sample sizes, or Gelman-Rubin statistics. The paper itself states in §6.2 that the 0.2σ deviation in the sterile case is 'most likely not entirely due to lack of precision of the emulator, but rather a lack of convergence of the class-based run.' This admission undermines the benchmark: if the reference is underconverged, the 0.1σ figure for ΛCDM could be artificially good, and the 0.2σ could be largely noise. Please report convergence diagnostics for the reference chains, or restrict the headline precision claim to the analytic tests where the true likelihood is known.
  3. [§3.3–3.4] The acceptance criterion in Eq. (3.12) relies on a k-nearest-neighbour density estimate (k=20) in 27–29 dimensions. The paper does not verify that the accepted training points actually follow the intended density q(θ)∝L^{1/T_T}. This is a load-bearing assumption for the method's generality. The slower convergence observed for the non-Gaussian directions (§6.1) is consistent with the kNN proxy breaking down in high-dimensional funnel regions. A direct test (comparing the empirical density of the training cloud to L^{1/T_T}, or a sensitivity study in k) would substantially strengthen the paper.
  4. [§4.1; Abstract; §7] The '0.1σ' credible-interval claim is ambiguous. The metric defined in Eq. (4.4) is a fractional deviation of the credible-interval width, i.e., ΔCM_i is the sum of absolute deviations of upper and lower bounds divided by the reference interval width. This is not a σ unit. For non-Gaussian posteriors (e.g., the sterile case), a 10% width error is not equivalent to 0.1σ. The abstract and conclusions should either define the metric explicitly or use a genuine σ-based comparison, such as the difference in posterior standard deviations.
  5. [§6.2; §7; Abstract] The single-point emulator precision claim 'better than Δχ²∼0.5' is not representative of all test cases. In §6.2 the authors report that for the ΛCDM+(Neff,s,ms) case the single-point precision 'does not exceed Δχ²∼1', while §7 states a 'typical precision of Δχ²∼0.3−0.5'. The abstract's unconditional 'better than Δχ²∼0.5' is therefore overstated. Please reconcile the range and qualify the claim per test case.
minor comments (5)
  1. [§3.4; §6] The typical candidate count is given as N_cand=10000 in §3.4, but §6 says 5000 candidate points per iteration. Please make these consistent.
  2. [Fig. 10 and surrounding text] The sentence after Fig. 10 is incomplete: '...and for κ≳38 the complementary error function underflows...' The text is garbled and should be rewritten.
  3. [§6.1, figure captions] The text states 'a total of around 2×10^4 are used' for the synthetic runs, but the figure captions report 19,334 and 17,094 evaluations. Using exact numbers would avoid ambiguity.
  4. [Reproducibility statement] The public code is a strength, but specifying a version/commit hash and the exact environment would further improve reproducibility.
  5. [§3.3, Eq. (3.9)] The notation ρ_p ∼ k/V_k ∼ r_k^{-d} is dimensionally loose; since the authors themselves call it a 'crude estimate', a brief remark on the constants and the range of validity would help.

Circularity Check

0 steps flagged

No significant circularity: validation is against independent true likelihoods; self-citations are methodological only.

full rationale

CLiENT's central claim is that a network trained on true likelihood values at adaptively selected points reproduces credible intervals from the true likelihood. That claim is checked against external benchmarks: for the synthetic cases the exact analytic likelihood (eq. 2.1) is used, and for Planck the references are MontePython/class runs (figs 7 and 9), not quantities derived from CLiENT. Hyperparameters are tuned on a synthetic likelihood whose covariance is taken from Planck and then applied unchanged to Planck ('All runs presented in this work were performed with essentially identical hyperparameter settings'), so no fitted value is reused to define the success metric. The self-citations to CONNECT ([6]) are implementation borrowings — activation choice ('same choice of activation function as in the connect framework'), proposal generation ('essentially identical to what is used in connect'), and a similar Gelman-Rubin stopping rule — none of which is the load-bearing evidence for the 0.1σ/Δχ² claims. The paper itself flags a validation caveat in §6.2: for ΛCDM+(Neff,s,ms) the 0.2σ deviation is 'most likely not entirely due to lack of precision of the emulator, but rather a lack of convergence of the class-based run with which we compare'; this is a benchmark-quality concern, not a circularity. Likewise Table 1's 22,728 evaluations versus the abstract's '<2×10^4' is an internal-consistency/accuracy concern about the efficiency claim, not a definitional tautology. No step in the derivation reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The method's performance depends on a set of hand-chosen hyperparameters (κ, T_T, T_MCMC, nstd, k, network size, convergence thresholds). None are fitted to the target posteriors; they are tuned on the 27D synthetic Gaussian and then fixed across all runs. The main domain assumptions are that the kNN density estimate captures local density in 27–29 dimensions and that the emcee chains on the surrogate are converged.

free parameters (8)
  • κ (loss width) = 3
    Set to 3 standard deviations; balances accuracy near best-fit vs tails (§3.1).
  • Training temperature T_T = 7
    Controls training density via q ∝ L^{1/T_T}; hand-tuned (§3.3).
  • MCMC temperature T_MCMC = 7
    Temperature used when sampling surrogate with emcee (§3.4).
  • Prior restriction width nstd = 10
    Sampling hypercube ±10σ around fiducial (§3.4).
  • k-NN density estimator k = 20
    Number of neighbours for local density estimate; authors find results k-independent for 5≤k≤50 (§3.3).
  • Network architecture = 5 hidden layers × 512 neurons + trainable Alsing β/γ
    Chosen to match CONNECT-style capacity (§6).
  • Training schedule = 5000 epochs max, batch 128, lr 1e-4, patience 250
    Chosen to ensure convergence without overfitting (§6).
  • Convergence thresholds = N_ESS=50, δτ=0.05, R−1<0.01
    Stopping criteria for emcee and training iterations (§3.4, §5).
axioms (6)
  • standard math Wilson–Hilferty approximation (Eq. A.9) is accurate for κ in 1–10 and n=27/29.
    Used to set the ε in the loss function from κ (§3.1, Appendix A).
  • domain assumption kNN density with k=20 approximates local density in 27–29 dimensions (Eq. 3.9).
    No proof; empirical claim in §3.3.
  • domain assumption Acceptance criterion in Eq. 3.12 yields training density proportional to L^{1/T_T}.
    Heuristic derivation, §3.3–3.4.
  • domain assumption Emcee chains on the surrogate converge and provide good candidate proposals.
    Convergence criteria used but no guarantee (§3.4).
  • domain assumption The trained NN generalizes within the ±10σ box sufficiently to represent the likelihood.
    Neural network generalization assumption (§6).
  • domain assumption Reference MCMC runs (class + MontePython/Cobaya) are converged.
    Used as ground truth; authors note possible non-convergence for ΛCDM+sterile case (§6.2).

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CLiENT: A new tool for emulating cosmological likelihoods using deep neural networks." pith.science (2026). https://pith.science/paper/TBVB6GSZ

@misc{pith2026251217509,
  author       = {Pith},
  title        = {Pith review of: CLiENT: A new tool for emulating cosmological likelihoods using deep neural networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TBVB6GSZ}},
  note         = {Machine review of arXiv:2512.17509}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Cosmological emulation of observables such as the Cosmic Microwave Background (CMB) spectra and matter power spectra have become increasingly common in recent years because of the potential for saving computation time in connection with cosmological parameter inference or model comparison. In this paper we present CLiENT (Cosmological Likelihood Emulator using Neural networks with TensorFlow), a new method which circumvents the computation of observables in favour of directly emulating the likelihood function for a data set given a model parameter vector. We find that the method is competitive with observable emulators in terms of the required number of function evaluations, but has the distinct advantage of producing a surrogate likelihood which is completely auto-differentiable. Using less than $2 \times 10^4$ function evaluations CLiENT typically achieves credible intervals within better than $0.1 \sigma$ of those obtained using the true likelihood and single-point emulator precision better than $\Delta \chi^2 \sim 0.5$ across relevant regions in parameter space.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Posterior sampling in the Age of Emulators

    astro-ph.IM 2026-06 unverdicted novelty 5.0

    Compares MH, MALA, HMC, NUTS, and AIES on differentiable likelihood emulators for ΛCDM and sterile-neutrino models, finding MALA and MH competitive in wall time despite NUTS needing fewer samples.

Reference graph

Works this paper leans on

22 extracted references · 16 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Conservative Constraints on Early Cosmology: an illustration of the Monte Python cosmological parameter inference code,

    B. Audren, J. Lesgourgues, K. Benabed, and S. Prunet, “Conservative Constraints on Early Cosmology: an illustration of the Monte Python cosmological parameter inference code,” JCAP02(2013) 001,arXiv:1210.7183 [astro-ph.CO]

    Pith/arXiv arXiv 2013

  2. [2]

    Cobaya: Code for Bayesian Analysis of hierarchical physical models,

    J. Torrado and A. Lewis, “Cobaya: Code for Bayesian Analysis of hierarchical physical models,”JCAP05(2021) 057,arXiv:2005.05290 [astro-ph.IM]

    Pith/arXiv arXiv 2021

  3. [3]

    Efficient computation of CMB anisotropies in closed FRW models,

    A. Lewis, A. Challinor, and A. Lasenby, “Efficient computation of CMB anisotropies in closed FRW models,”Astrophys. J.538(2000) 473–476,arXiv:astro-ph/9911177

    Pith/arXiv arXiv 2000

  4. [4]

    The Cosmic Linear Anisotropy Solving System (CLASS) II: Approximation schemes,

    D. Blas, J. Lesgourgues, and T. Tram, “The Cosmic Linear Anisotropy Solving System (CLASS) II: Approximation schemes,”JCAP07(2011) 034,arXiv:1104.2933 [astro-ph.CO]

    Pith/arXiv arXiv 2011

  5. [5]

    CosmoPower: emulating cosmological power spectra for accelerated Bayesian inference from next-generation surveys,

    A. Spurio Mancini, D. Piras, J. Alsing, B. Joachimi, and M. P. Hobson, “CosmoPower: emulating cosmological power spectra for accelerated Bayesian inference from next-generation surveys,”Mon. Not. Roy. Astron. Soc.511(2022) no. 2, 1771–1788,arXiv:2106.03846 [astro-ph.CO]

    Pith/arXiv arXiv 2022

  6. [6]

    CONNECT: a neural network based framework for emulating cosmological observables and cosmological parameter inference,

    A. Nygaard, E. B. Holm, S. Hannestad, and T. Tram, “CONNECT: a neural network based framework for emulating cosmological observables and cosmological parameter inference,” JCAP05(2023) 025,arXiv:2205.15726 [astro-ph.IM]. – 20 –

    Pith/arXiv arXiv 2023

  7. [7]

    OLÉ — Online Learning Emulation in cosmology,

    S. Günther, L. Balkenhol, C. Fidler, A. R. Khalife, J. Lesgourgues, M. R. Mosbech, and R. K. Sharma, “OLÉ — Online Learning Emulation in cosmology,”JCAP09(2025) 059, arXiv:2503.13183 [astro-ph.CO]

    arXiv 2025

  8. [8]

    Fast cosmological parameter estimation using neural networks,

    T. Auld, M. Bridges, M. P. Hobson, and S. F. Gull, “Fast cosmological parameter estimation using neural networks,”Mon. Not. Roy. Astron. Soc.376(2007) L11–L15, arXiv:astro-ph/0608174

    Pith/arXiv arXiv 2007

  9. [9]

    Pico: Parameters for the Impatient Cosmologist,

    W. A. Fendt and B. D. Wandelt, “Pico: Parameters for the Impatient Cosmologist,”Astrophys. J.654(2006) 2–11,arXiv:astro-ph/0606709

    Pith/arXiv arXiv 2006

  10. [10]

    Power spectrum emulators from neural networks and tree-based methods,

    A. Lazanu, “Power spectrum emulators from neural networks and tree-based methods,”Phys. Dark Univ.51(2026) 102186,arXiv:2506.07514 [astro-ph.CO]

    arXiv 2026

  11. [11]

    The Coyote Universe Extended: Precision Emulation of the Matter Power Spectrum,

    K. Heitmann, E. Lawrence, J. Kwan, S. Habib, and D. Higdon, “The Coyote Universe Extended: Precision Emulation of the Matter Power Spectrum,”Astrophys. J.780(2014) 111, arXiv:1304.7849 [astro-ph.CO]. [12]EuclidCollaboration, M. Knabenhanset al., “Euclid preparation: II. The EuclidEmulator – A tool to compute the cosmology dependence of the nonlinear matte...

    Pith/arXiv arXiv 2014

  12. [13]

    Accelerating Large-Scale-Structure data analyses by emulating Boltzmann solvers and Lagrangian Perturbation Theory,

    G. Aricò, R. E. Angulo, and M. Zennaro, “Accelerating Large-Scale-Structure data analyses by emulating Boltzmann solvers and Lagrangian Perturbation Theory,”arXiv:2104.14568 [astro-ph.CO]

    Pith/arXiv arXiv

  13. [14]

    The Mira–Titan Universe – IV. High-precision power spectrum emulation,

    K. R. Moran, K. Heitmann, E. Lawrence, S. Habib, D. Bingham, A. Upadhye, J. Kwan, D. Higdon, and R. Payne, “The Mira–Titan Universe – IV. High-precision power spectrum emulation,”Mon. Not. Roy. Astron. Soc.520(2023) no. 3, 3443–3458,arXiv:2207.12345 [astro-ph.CO]

    Pith/arXiv arXiv 2023

  14. [15]

    Neural likelihood estimators for flexible gravitational wave data analysis,

    L. Negri and A. Samajdar, “Neural likelihood estimators for flexible gravitational wave data analysis,”arXiv:2509.17606 [astro-ph.HE]. [16]PlanckCollaboration, N. Aghanimet al., “Planck 2018 results. V. CMB power spectra and likelihoods,”Astron. Astrophys.641(2020) A5,arXiv:1907.12875 [astro-ph.CO]. [17]BOSSCollaboration, S. Alamet al., “The clustering of...

    arXiv 2018

  15. [18]

    The 6dF Galaxy Survey: Baryon Acoustic Oscillations and the Local Hubble Constant,

    F. Beutler, C. Blake, M. Colless, D. H. Jones, L. Staveley-Smith, L. Campbell, Q. Parker, W. Saunders, and F. Watson, “The 6dF Galaxy Survey: Baryon Acoustic Oscillations and the Local Hubble Constant,”Mon. Not. Roy. Astron. Soc.416(2011) 3017–3032, arXiv:1106.3366 [astro-ph.CO]

    Pith/arXiv arXiv 2011

  16. [19]

    The clustering of the SDSS DR7 main Galaxy sample – I. A 4 per cent distance measure at z= 0.15,

    A. J. Ross, L. Samushia, C. Howlett, W. J. Percival, A. Burden, and M. Manera, “The clustering of the SDSS DR7 main Galaxy sample – I. A 4 per cent distance measure at z= 0.15,”Mon. Not. Roy. Astron. Soc.449(2015) no. 1, 835–847,arXiv:1409.3242 [astro-ph.CO]

    Pith/arXiv arXiv 2015

  17. [20]

    The distribution of chi-square,

    E. B. Wilson and M. M. Hilferty, “The distribution of chi-square,”Proceedings of the National Academy of Sciences of the United States of America17(1931) no. 12, 684–688

    1931

  18. [21]

    SPECULATOR: Emulating Stellar Population Synthesis for Fast and Accurate Galaxy Spectra and Photometry,

    J. Alsing, H. Peiris, J. Leja, C. Hahn, R. Tojeiro, D. Mortlock, B. Leistedt, B. D. Johnson, and C. Conroy, “SPECULATOR: Emulating Stellar Population Synthesis for Fast and Accurate Galaxy Spectra and Photometry,”The Astrophysical Journal Supplement Series249(2020) 5, arXiv:1911.11778 [astro-ph.IM]

    Pith/arXiv arXiv 2020

  19. [22]

    emcee: The MCMC Hammer,

    D. Foreman-Mackey, D. W. Hogg, D. Lang, and J. Goodman, “emcee: The MCMC Hammer,” Publ. Astron. Soc. Pac.125(2013) 306–312,arXiv:1202.3665 [astro-ph.IM]. – 21 –

    Pith/arXiv arXiv 2013

  20. [23]

    On Information and Sufficiency,

    S. Kullback and R. A. Leibler, “On Information and Sufficiency,”The Annals of Mathematical Statistics22(1951) no. 1, 79–86

    1951

  21. [24]

    Fast and effortless computation of profile likelihoods using CONNECT,

    A. Nygaard, E. B. Holm, S. Hannestad, and T. Tram, “Fast and effortless computation of profile likelihoods using CONNECT,”JCAP11(2023) 064,arXiv:2308.06379 [astro-ph.CO]

    Pith/arXiv arXiv 2023

  22. [25]

    Markov chain Monte Carlo method and its application,

    S. Brooks, “Markov chain Monte Carlo method and its application,”Journal of the Royal Statistical Society: Series D (The Statistician)47no. 1, 69–100, https://rss.onlinelibrary.wiley.com/doi/pdf/10.1111/1467-9884.00117. https://rss.onlinelibrary.wiley.com/doi/abs/10.1111/1467-9884.00117. – 22 –

    arXiv

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