Pith. sign in

REVIEW 5 minor 1 cited by

Compressed 'CMB-lite' Likelihoods Using Automatic Differentiation

T0 review · 0 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read A CMB-lite likelihood can be assembled in about a minute on a laptop by replacing sampling-based reconstruction with one minimisation and one Hessian evaluation, producing a sampling-noise-free covariance and recovering the reference…

desk verdict Clean AD-based CMB-lite construction that actually works on SPT-3G data; the Laplace-approximation caveat is real but empirically tested. read the letter →

arxiv 2412.00826 v2 pith:JTIFYXHC submitted 2024-12-01 astro-ph.CO

classification astro-ph.CO
keywords CMB-litelikelihoodcompressionautomaticdifferentiationcosmicmicrowavebackgroundbandpowersforegroundmarginalisationcovarianceestimationSPT-3G
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

CMB experiments observe the sky at several frequencies, so their power-spectrum likelihoods carry both cosmological information and nuisance parameters describing foregrounds and calibration. The CMB-lite idea is to first compress those multi-frequency measurements into a shorter vector of foreground-marginalised, CMB-only band powers, which can then be analysed quickly with any cosmological model. This paper claims that the expensive and noisy Markov-chain reconstruction step can be replaced by a single minimisation of the reconstruction likelihood followed by one evaluation of its Hessian, using automatic differentiation, so the whole compressed likelihood is built in about a minute on a personal computer with no sampling noise in its covariance. Applied to the 2018 SPT-3G temperature and polarisation data, the resulting lite likelihood recovers the cosmological constraints of the full multi-frequency likelihood for Lambda CDM and two extensions, with best-fit shifts below 0.1 sigma and error bars matching to better than 10 percent. If it holds, this makes lite likelihood construction cheap enough to be rerun routinely for robustness tests and for the much larger, lower-noise data sets now being planned.

What carries the argument

The object that carries the argument is the Gaussian reconstruction likelihood $\mathcal{L}_{\rm recon}(\mathbf{C}^{\rm CMB},\boldsymbol{\theta})$ over CMB-only band powers and nuisance parameters, together with the identity connecting its Hessian at the best-fit point to the covariance: the negative inverse Hessian of a Gaussian likelihood is its covariance, equivalently the Hessian is the negative Fisher matrix. Because the likelihood is written in a differentiable form, its gradient and Hessian can be evaluated exactly and cheaply; a truncated-Newton minimiser finds the best fit, and the Hessian at that point gives the sampling-noise-free covariance of the compressed band powers. This replaces the alternating Gibbs-sampling scheme originally used for CMB-lite construction, which is what removes the computational cost and the sampling noise.

What would settle it

Construct a synthetic multi-frequency data set whose reconstruction posterior is visibly non-Gaussian, for instance low signal-to-noise band powers or strong degenerate foregrounds, then compare parameter constraints from the Hessian-based lite likelihood against those from a long, converged MCMC sampling of the same reconstruction likelihood; if best-fit shifts or error-bar ratios exceed $0.1\sigma$ and $10\%$, the core Gaussian assumption fails.

Watch

Extended reading notes

Core claim

The central claim is that the CMB-lite reconstruction, the step that turns multi-frequency band powers into foreground-marginalised CMB-only band powers and their covariance, can be performed from one minimisation and one Hessian evaluation rather than from Markov-chain Monte Carlo sampling. Writing the reconstruction likelihood over both the CMB band powers and the nuisance parameters, the author minimises the joint likelihood to obtain the compressed band powers, then evaluates the Hessian at the best-fit point; the negative inverse Hessian is the covariance of the compressed band powers, with nuisance marginalisation folded in. The paper demonstrates this on the differentiable SPT-3G 2018 TT/TE/EE likelihood: the compressed data vector shrinks from 728 to 123 bins, the nuisance parameters from 33 to two, likelihood evaluation is about four times faster, and marginalised posteriors for $\Lambda$CDM, $\Lambda$CDM+$N_{\rm eff}$, and $\Lambda$CDM+$A_L$ agree with the reference likelihood to within $<0.1\sigma$ in best-fit shifts and $<10\%$ in error bars. One thousand Gaussian mock realisations show that the observed shifts are consistent with the expected realisation-dependent bias.

Load-bearing premise

The load-bearing premise is that the reconstruction likelihood is sufficiently Gaussian near its best fit that the negative inverse Hessian there is an unbiased covariance for the CMB-only band powers, including nuisance marginalisation; the paper validates this on the SPT-3G data and Gaussian mocks but does not give a general bound on the approximation error.

Editorial extensions

If this is right

  • Lite likelihood construction goes from a high-dimensional sampling problem to a local optimisation problem, so it can be repeated on a laptop in about a minute for different data cuts, foreground models, or robustness tests.
  • The covariance of the CMB-only band powers no longer carries sampling noise, so subsequent cosmological analyses avoid the need for sampling-noise corrections in this step.
  • Because the compressed data vector has 123 bins and two nuisance parameters instead of 728 bins and 33 nuisance parameters, MCMC exploration of cosmological parameters is faster and the likelihood is easier to interpret.
  • Parameter constraints from the lite likelihood are claimed to match the full multi-frequency likelihood within $<0.1\sigma$ best-fit shifts and $<10\%$ error bars for the three models tested, with the realisation-dependent offset shown to be statistically normal.
  • The recipe should transfer to future multi-frequency experiments with many channels, where the reduction in data-vector length and nuisance dimensionality is larger.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the Gaussian (quadratic) approximation of the reconstruction likelihood degrades in low-signal-to-noise regimes or with strongly non-Gaussian foregrounds, the inverse-Hessian covariance could understate the true uncertainty; a direct comparison with a converged MCMC reconstruction on such a case would bound the error.
  • The same single-minimisation recipe could be applied to any differentiable likelihood, potentially producing 'lite' versions of other nuisance-heavy cosmological data vectors such as galaxy clustering or lensing measurements.
  • Because no cosmological model enters the reconstruction, the compressed band powers and covariance form a model-independent data product that can be reused for many cosmological models without rebuilding.
  • A natural testable extension is to replace the fixed Hessian covariance with a Laplace-corrected or non-Gaussian covariance for cases where the realisation-dependent bias exceeds the quoted thresholds.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

0 major / 5 minor

Summary. The paper presents an automatic-differentiation-based implementation of the CMB-lite framework. Using the differentiable SPT-3G 2018 TT/TE/EE likelihood from the candl library, the author minimizes the reconstruction likelihood once and evaluates its Hessian to obtain foreground-marginalised CMB-only band powers and their covariance. The resulting lite likelihood is benchmarked against the multi-frequency likelihood for ΛCDM, ΛCDM+Neff, and ΛCDM+AL, finding best-fit shifts below 0.1σ and posterior-width ratios within 10%. The author also performs scatter-free mock reconstructions and 1000 Gaussian realisations to quantify systematic and realisation-dependent biases, and publicly releases the lite likelihood and a construction notebook.

Significance. If the result holds, this is a practical and timely contribution: it reduces lite-likelihood construction from expensive MCMC sampling to about a minute on a personal computer and removes sampling noise from the covariance estimate. The paper is transparent about its approximations, including the Laplace approximation for the covariance, the treatment of frequency-dependent window functions, the omission of beam terms in the lite likelihood, and the representation of super-sample lensing through an additive covariance term. The empirical validation is a strength: scatter-free mocks, 1000 Gaussian realisations, three cosmological models, and public code. The main limitation is that no general error bound is given for the Laplace approximation, but the paper does not overclaim beyond the SPT-3G 2018 data set and the tests support the central claim for that application.

minor comments (5)
  1. [Abstract and Section 1] The text contains spacing/formatting artifacts such as 'TT /TE/E E' and 'Elikelihood'; these should be corrected to 'TT/TE/EE' and 'EE likelihood'.
  2. [Section 3, step 2] Please specify explicitly whether the Hessian is evaluated for the log-likelihood or the log-posterior, since priors on nuisance parameters such as the calibration parameters affect the resulting covariance matrix.
  3. [Section 4] The phrase 'systemic bias' appears twice (in the scatter-free mock paragraph and in Figure 4); the intended term is 'systematic bias'.
  4. [Figure 1 caption] There is a typo: 'Mulitpole' should be 'Multipole'.
  5. [Table 1] The header 'ln(1010A)' omits the subscript 's'; it should read 'ln(10^10 A_s)' to match the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the lite likelihood is a direct compression of the multi-frequency likelihood via minimisation and Hessian evaluation, and the validation is an internal consistency check rather than a circular prediction.

full rationale

The derivation chain is self-contained. The lite likelihood is constructed from the multi-frequency likelihood by (1) minimising the reconstruction likelihood over CMB band powers and nuisance parameters to obtain C_lite, and (2) taking the negative inverse Hessian at the best fit as Sigma_lite. No parameter is fitted to the validation targets; the posterior comparison in Section 4 is a consistency check between two representations of the same data, not an independent prediction, so a mismatch would indicate compression error rather than a circular step. The Hessian step is an explicitly acknowledged Laplace/Gaussian approximation (Section 3), not an assumption that already contains the claimed parameter recovery. Self-citations to candl and the SPT-3G 2018 likelihood are instrumental data/software provenance and are not used to license the central claim; the accuracy of the compression is demonstrated directly against the reference likelihood and 1000 Gaussian mocks. No equation is defined in terms of its output, no fitted input is relabelled as a prediction, and no uniqueness theorem is imported. The internal-benchmark nature of the validation limits the strength of the evidence but does not constitute circularity.

Assumptions & free parameters 0 free parameters · 8 assumptions · 0 invented entities

No new free parameters or invented entities are introduced by the method; the reconstructed band powers are outputs fitted to SPT-3G data, and the retained calibration parameters come from the reference likelihood. The central claim relies on standard CMB modeling assumptions and on several dataset-specific approximations that are explicitly stated and tested.

assumptions (8)
  • domain assumption The data model of Eq. (1) (CMB band powers, foreground model, calibration factors) and the Gaussian likelihood of Eq. (2) describe the multi-frequency SPT-3G band powers.
    Inherited from the reference SPT-3G likelihood; not re-derived in this paper.
  • standard math The Hessian of the reconstruction likelihood at the best-fit point equals the negative Fisher matrix, so its inverse is the covariance of the CMB-only band powers.
    Section 3; requires the likelihood to be locally Gaussian, which the paper tests with mocks but does not prove in general.
  • ad hoc to paper Differences in window functions across frequencies are small enough that a covariance-weighted combination is a valid single set of CMB-only band powers.
    Section 4, first bullet; the paper acknowledges that no unique set of CMB-only band powers exists.
  • ad hoc to paper The model- and frequency-dependent beam term can be included in the reconstruction covariance and ignored in the lite likelihood.
    Section 4, second bullet; justified only for this data set.
  • ad hoc to paper The super-sample lensing contribution can be represented as an additive contribution to the band power covariance matrix.
    Section 4, fifth bullet; the magnitude of this contribution is not specified.
  • domain assumption The Planck-based Gaussian prior on tau (mean 0.054, width 0.0074) is appropriate.
    Section 4; same prior as Balkenhol et al. 2023.
  • ad hoc to paper The aberration effect can be treated as part of the reconstructed CMB signal rather than a separate nuisance parameter.
    Section 4, fourth bullet; no unique band-power level mapping exists.
  • ad hoc to paper The calibration split into internal and external parameters, with Tcal and Ecal retained in the lite likelihood, avoids long-range correlations.
    Section 4, third bullet; follows the suggestion of D13.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Compressed 'CMB-lite' Likelihoods Using Automatic Differentiation." pith.science (2026). https://pith.science/paper/JTIFYXHC

@misc{pith2026241200826,
  author       = {Pith},
  title        = {Pith review of: Compressed 'CMB-lite' Likelihoods Using Automatic Differentiation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JTIFYXHC}},
  note         = {Machine review of arXiv:2412.00826}
}
abstract

The compression of multi-frequency cosmic microwave background (CMB) power spectrum measurements into a series of foreground-marginalised CMB-only band powers allows for the construction of faster and more easily interpretable 'lite' likelihoods. However, obtaining the compressed data vector is computationally expensive and yields a covariance matrix with sampling noise. In this work, we present an implementation of the CMB-lite framework relying on automatic differentiation. The technique presented reduces the computational cost of the lite likelihood construction to one minimisation and one Hessian evaluation, which run on a personal computer in about a minute. We demonstrate the efficiency and accuracy of this procedure by applying it to the differentiable SPT-3G 2018 TT/TE/EE likelihood from the candl library. We find good agreement between the marginalised posteriors of cosmological parameters yielded by the resulting lite likelihood and the reference multi-frequency version for all cosmological models tested; the best-fit values shift by $<0.1\,\sigma$, where $\sigma$ is the width of the multi-frequency posterior, and the inferred parameter error bars match to within $<10\%$. We publicly release the SPT-3G 2018 TT/TE/EE lite likelihood and a python notebook showing its construction at https://github.com/Lbalkenhol/candl .

Figures

Figures reproduced from arXiv: 2412.00826 by the authors.

Figure 1
Figure 1. — Foreground-marginalised CMB-only band powers (blue) based on the SPT-3G 2018 TT /TE/EE multi-frequency data (grey). Though the construction of the CMB-only data vector does not require a cosmological model, the data visually follow the best-fit ΛCDM model of the multi-frequency likelihood shown in black. The CMB-lite band powers visibly remove the foreground contamination in the temperature power spectrum. in prin… view at source ↗
Figure 2
Figure 2. — Ratio of the error bars of the lite likelihood to two reference cases (σ ref b ): the coadd (blue) and the 150 GHz auto￾spectrum (grey). As foregrounds are already marginalised over in the CMB-lite band powers, error bars are larger than the two reference cases at high ℓ for TT. For TE and EE we see that the lite likelihood improves on the 150 GHz auto-spectrum on all scales due to the combination of multi-frequen… view at source ↗
Figure 3
Figure 3. — Marginalised posteriors for ΛCDM parameters obtained from the full multi-frequency (grey filled contours) and lite (orange dashed line contours) SPT-3G 2018 TT /TE/EE likelihoods (68% and 95% confidence levels). The constraints match well, with only a small offset in ln (1010As) of < 10% of the width of the multi-frequency posterior visually discernable. from the multi-frequency band power covariance matrix based … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: — The difference in best-fit ΛCDM parameters of the lite likelihood compared to the multi-frequency likelihood for 1000 mock band power realisations (black histogram) normalised by the width of the constraints of the multi-frequency likelihood (1σ). We indicate the bia…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Alleviating prior dependencies for DESI DR1 clustering fits through reparameterization

    astro-ph.CO 2026-07 unverdicted novelty 6.0 of 10

    Jeffreys prior over EFTofLSS coefficients mitigates projection effects in DESI DR1 power spectrum multipole fits, recentering posteriors for late-time expansion parameters.

Reference graph

Works this paper leans on

38 extracted references · 25 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ade, P. A. R., Ahmed, Z., Amiri, M., et al. 2021, Phys. Rev. Lett., 127, 151301

  2. [2]

    Aiola, S., Calabrese, E., Maurin, L., et al. 2020, J. Cosmology Astropart. Phys., 2020, 047

  3. [3]

    Constructing Exact Confidence Regions on Parameter Manifolds of Non-Linear Models

    Arutjunjan, R., Schaefer, B. M., & Kreutz, C. 2022, arXiv e-prints, arXiv:2211.03421

  4. [4]

    2023, Phys

    Balkenhol, L., Dutcher, D., Spurio Mancini, A., et al. 2023, Phys. Rev. D, 108, 023510

  5. [5]

    & Reichardt, C

    Balkenhol, L. & Reichardt, C. L. 2022, MNRAS, 512, 4394

  6. [6]

    2018, JAX: composable transformations of Python+NumPy programs

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

  7. [7]

    A., Battaglia, N., et al

    Calabrese, E., Hlozek, R. A., Battaglia, N., et al. 2013, Phys. Rev. D, 87, 103012

  8. [8]

    2023, The Open Journal of Astrophysics, 6, 15

    Campagne, J.-E., Lanusse, F., Zuntz, J., et al. 2023, The Open Journal of Astrophysics, 6, 15

Show all 38 references
  1. [9]

    E., Ade, P

    Carlstrom, J. E., Ade, P. A. R., Aird, K. A., et al. 2011, PASP, 123, 568

  2. [10]

    K., Hasselfield, M., Ho, S.-P

    Choi, S. K., Hasselfield, M., Ho, S.-P. P., et al. 2020, J. Cosmology Astropart. Phys., 2020, 045 CMB-S4 Collaboration, Abazajian, K., Addison, G., et al. 2019, arXiv e-prints, arXiv:1907.04473

  3. [11]

    & Schneider, M

    Dodelson, S. & Schneider, M. D. 2013, Phys. Rev. D, 88, 063537

  4. [12]

    Dunkley, J., Calabrese, E., Sievers, J., et al. 2013, J. Cosmology Astropart. Phys., 7, 25

  5. [13]

    Dutcher, D., Balkenhol, L., Ade, P. A. R., et al. 2021, Phys. Rev. D, 104, 022003

  6. [14]

    2024, arXiv e-prints, arXiv:2411.06000

    Ge, F., Millea, M., Camphuis, E., et al. 2024, arXiv e-prints, arXiv:2411.06000

  7. [15]

    2007, A&A, 464, 399

    Hartlap, J., Simon, P., & Schneider, P. 2007, A&A, 464, 399

  8. [16]

    F., Seikel, M., Nord, B

    Heavens, A. F., Seikel, M., Nord, B. D., et al. 2014, MNRAS, 445, 1687

  9. [17]

    W., Sayre, J

    Henning, J. W., Sayre, J. T., Reichardt, C. L., et al. 2018, ApJ, 852, 97

  10. [18]

    Hoffman, M. D. & Gelman, A. 2011, arXiv e-prints, arXiv:1111.4246

  11. [19]

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

  12. [20]

    2014, Phys

    Jeong, D., Chluba, J., Dai, L., Kamionkowski, M., & Wang, X. 2014, Phys. Rev. D, 89, 023003

  13. [21]

    2001, SciPy: Open source scientific tools for Python, [Online; accessed 2014-10-22]

    Jones, E., Oliphant, T., Peterson, P., et al. 2001, SciPy: Open source scientific tools for Python, [Online; accessed 2014-10-22]

  14. [22]

    G., Ade, P

    Keating, B. G., Ade, P. A. R., Bock, J. J., et al. 2003, in Polarimetry in Astronomy. Edited by Silvano Fineschi. Proceedings of the SPIE, Volume 4843., 284–295

  15. [23]

    The Cosmic Microwave Background and its Polarization,

    Kosowsky, A. 2003, in the proceedings of the workshop on “The Cosmic Microwave Background and its Polarization,” New Astronomy Reviews, ed. S. Hanany & K. A. Olive (Elsevier)

  16. [24]

    2000, ApJ, 538, 473

    Lewis, A., Challinor, A., & Lasenby, A. 2000, ApJ, 538, 473

  17. [25]

    S., Qu, F

    Madhavacheril, M. S., Qu, F. J., Sherwin, B. D., et al. 2024, ApJ, 962, 113

  18. [26]

    2014, Phys

    Manzotti, A., Hu, W., & Benoit-L´ evy, A. 2014, Phys. Rev. D, 90, 023003

  19. [27]

    M., Crawford, T

    Mocanu, L. M., Crawford, T. M., Aylor, K., et al. 2019, J. Cosmology Astropart. Phys., 2019, 038

  20. [28]

    Nash, S. G. 1984, SIAM Journal on Numerical Analysis, 21, 770

  21. [29]

    & Wright, S

    Nocedal, J. & Wright, S. J. 2006, Numerical Optimization (New

  22. [30]

    Pan, Z., Bianchini, F., Wu, W. L. K., et al. 2023, Phys. Rev. D, 108, 122005

  23. [31]

    J., Friedrich, O., Sellentin, E., & Heavens, A

    Percival, W. J., Friedrich, O., Sellentin, E., & Heavens, A. 2022, MNRAS, 510, 3207

  24. [32]

    & Spurio Mancini, A

    Piras, D. & Spurio Mancini, A. 2023, The Open Journal of Astrophysics, 6, 20 Planck Collaboration, Aghanim, N., Akrami, Y., et al. 2020a, A&A, 641, A1 Planck Collaboration, Aghanim, N., Akrami, Y., et al. 2020b, A&A, 641, A6 Planck Collaboration, Aghanim, N., Akrami, Y., et al...

  25. [33]

    2024, arXiv e-prints, arXiv:2403.00085

    Prince, H., Calabrese, E., & Dunkley, J. 2024, arXiv e-prints, arXiv:2403.00085

  26. [34]

    J., Sherwin, B

    Qu, F. J., Sherwin, B. D., Madhavacheril, M. S., et al. 2024, ApJ, 962, 112

  27. [35]

    2014, MNRAS, 441, 1831 Simons Observatory Collaboration, Ade, P., Aguirre, J., et al

    Sellentin, E., Quartin, M., & Amendola, L. 2014, MNRAS, 441, 1831 Simons Observatory Collaboration, Ade, P., Aguirre, J., et al. 2019, J. Cosmology Astropart. Phys., 2019, 056 Spurio Mancini, A., Piras, D., Alsing, J., Joachimi, B., & Hobson, M. P. 2022, MNRAS, 511, 1771

  28. [36]

    W., Amiri, M., et al

    Staniszewski, Z., Aikin, R. W., Amiri, M., et al. 2012, Journal of Low Temperature Physics, 167, 827

  29. [37]

    & Lewis, A

    Torrado, J. & Lewis, A. 2021, J. Cosmology Astropart. Phys., 2021, 057 van der Walt, S., Colbert, S., & Varoquaux, G. 2011, Computing in Science Engineering, 13, 22

  30. [38]

    E., et al

    Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020, Nature Methods, 17, 261 This paper was built using the Open Journal of As- trophysics LATEX template. The OJA is a journal which provides fast and easy peer review for new papers in the astro-ph section of the arXiv, mak...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.