CamSpec Likelihood: Planck High-ℓ Analysis
- CamSpec Likelihood is a high-ℓ Planck CMB likelihood framework that uses masked-sky pseudo-Cl cross-spectra to infer parameters from TT, TE, and EE power spectra.
- It employs advanced cleaning techniques, beam and transfer-function corrections, and calibration adjustments to accurately process both temperature and polarization data.
- Its evolving implementations from PR3 to PR4/NPIPE and variants like CamSpec-NPIPE-lite enable robust cosmological inference and serve as a consistency check against other high-ℓ likelihoods.
Searching arXiv for recent and foundational papers on CamSpec likelihood. CamSpec Likelihood is a high- Planck CMB likelihood framework for temperature and polarization power spectra, used to infer cosmological parameters from , , and angular power spectra measured in Planck high-frequency maps. In its PR4/NPIPE realization, it is described as a likelihood built from masked-sky pseudo- cross-spectra of Planck HFI maps, with analytic covariance matrices, beam and transfer-function corrections, dust cleaning, calibration corrections, and a compact nuisance model; in the broader literature, it also appears as an external high- likelihood combined with distinct low- likelihood constructions for reionization, inflationary-feature, and curvature analyses (Rosenberg et al., 2022). The term does not denote a single immutable object: the literature distinguishes earlier “clean” CamSpec releases, PR3 and PR4/NPIPE implementations, and nuisance-marginalized compressed forms such as CamSpec-NPIPE-lite (Hazra et al., 2021).
1. Historical placement and scope
CamSpec has been used to analyse Planck temperature and polarization maps of the cosmic microwave background since the first Planck data release, and one dedicated methodological paper presents “a detailed description of the CamSpec likelihood pipeline and a reanalysis of the Planck high frequency maps” (Efstathiou et al., 2019). In that presentation, CamSpec is framed as a high-, power-spectrum-based Planck likelihood for , , and 0, designed to account for sky masking, detector-set combinations, noise, beams, calibration, Galactic foregrounds, unresolved extragalactic foregrounds, and polarization-specific instrumental effects (Efstathiou et al., 2019).
Later work makes clear that CamSpec is also a moving target within the Planck ecosystem. A PR4/NPIPE realization, denoted PR4_12.6, was constructed from Planck PR4 NPIPE maps, while an updated PR3 realization, PR3_12.6, was retained for comparison (Rosenberg et al., 2022). A further distinction became standard in comparative cosmology papers: “PR3 reference cosmology” usually meant Legacy maps with Plik, whereas “PR4 reference cosmology” usually meant NPIPE maps with CamSpec (Jense et al., 10 Oct 2025). This implies that “CamSpec likelihood” can refer either to a specific public Planck high-1 likelihood release or, more generally, to a family of related high-2 likelihood pipelines.
A concise version map is therefore useful.
| Variant | Maps | Characterization |
|---|---|---|
| PR3_12.6 | Planck 2018 maps | Updated PR3 CamSpec for comparison |
| PR4_12.6 | PR4 / NPIPE maps | New high-3 CamSpec likelihood |
| CamSpec-NPIPE-lite | NPIPE maps | Nuisance-marginalized compressed likelihood |
This multiplicity matters because later papers often use CamSpec as an already established likelihood choice rather than reconstructing it. In reionization analyses, for example, CamSpec is adopted as the external high-4 likelihood that complements a newly built low-5 polarization likelihood; those works do not build, modify, or validate CamSpec itself (Genesini et al., 23 Mar 2026).
2. Data inputs, spectra, and multipole coverage
In the PR4/NPIPE implementation, the data entering CamSpec are Planck HFI frequency maps at 100, 143, 217, 353, and 545 GHz, with 100, 143, and 217 GHz as the main cosmology channels, 353 GHz as the dust template for 6 and 7 cleaning, and 545 GHz as the dust template for 8 cleaning (Rosenberg et al., 2022). For NPIPE, the independent data split is given by A/B detector-set maps rather than half-mission maps (Rosenberg et al., 2022).
The spectra retained for cosmology are highly specific. In 9, the likelihood uses coadded 0, 1, and 2 spectra. In 3 and 4, it uses all 5 cross-spectra among 100, 143, and 217 GHz (Rosenberg et al., 2022). The multipole ranges are likewise explicit: 6 7 from 8 to 2000; 9 and 0 1 from 2 to 2500; and 3 from 4 to 2000 (Rosenberg et al., 2022). Although figures may display binned spectra, the PR4 likelihood itself is unbinned (Rosenberg et al., 2022).
Because the high-5 likelihood starts at 6, it is supplemented by low-7 Planck likelihoods, specifically Commander 8 for 9 and SimAll 0 for 1 (Rosenberg et al., 2022). This pattern recurs across the literature. Curvature analyses using CamSpec PR4 define “CamSpec” as 2 spectra at 3, based on Planck-PR4 NPIPE CMB maps, combined with SimAll and Commander below 4 (Specogna et al., 30 Sep 2025). Inflation-feature studies similarly use CamSpec only for the high-5 block and pair it with official low-6 Planck likelihoods (Hazra et al., 2021).
The 2019 CamSpec reanalysis emphasized a statistically powerful configuration using 143 and 217 GHz in temperature and polarization over 80% of the sky (Efstathiou et al., 2019). Later feature-model papers adopted the corresponding “clean CamSpec” framing, describing a version that excludes the 100 GHz temperature auto-spectrum, cleans the 143 and 217 GHz channels using the 545 GHz channel, and uses sky fractions up to 80% in temperature and polarization (Hazra et al., 2021).
3. Likelihood construction and nuisance structure
At high level, CamSpec is a Gaussian likelihood over measured spectra. In the PR4/NPIPE description, the likelihood compares a data vector to a model consisting of CMB plus residual foregrounds, corrected by calibration parameters (Jense et al., 10 Oct 2025). A compact expression given for the theory spectra is
7
where 8 is the cosmological CMB spectrum, 9 is the residual foreground term, and 0 captures calibration or polarization-efficiency corrections (Jense et al., 10 Oct 2025).
The PR4 paper presents the same construction in operational terms as the standard CamSpec high-1 Gaussian likelihood in the vector of deconvolved spectra,
2
with 3 the data vector of mask-deconvolved, beam-corrected cross-spectra, 4 the theory plus residual foreground plus calibration model, and 5 the covariance matrix (Rosenberg et al., 2022). CamSpec uses the pseudo-6 method on masked skies, a coupling matrix to deconvolve mask-induced mode coupling, and beam transfer functions 7 from QuickPol (Rosenberg et al., 2022).
Foreground treatment is a defining feature. The PR4/NPIPE implementation is described as foreground-cleaned rather than heavily parametric (Rosenberg et al., 2022). Dust cleaning is performed with high-frequency templates: 545 GHz for 8, 353 GHz for 9 and 0 (Rosenberg et al., 2022). In 1, after 545 cleaning, the remaining residual foregrounds are modeled phenomenologically as a power law for each retained 2 spectrum,
3
with separate amplitude and slope parameters for 4, 5, and 6 (Rosenberg et al., 2022). In the later NPIPE-lite paper, the residual 7 foreground model is written equivalently as
8
again with one amplitude and one slope for each retained 9 frequency pair (Jense et al., 10 Oct 2025).
The nuisance sector is intentionally compact. In the PR4 realization there are three calibration-like nuisance parameters: 0, 1, and 2, with Gaussian priors 3, 4, and 5 (Rosenberg et al., 2022). The 6 residual power-law nuisance parameters have flat priors 7 and 8 (Rosenberg et al., 2022). A distinctive simplification is that no polarization foreground nuisance parameters are included in the PR4 likelihood, because dust cleaning is treated as effective enough that all 9 and 0 spectra can be coadded without explicit residual polarization foreground modeling (Rosenberg et al., 2022).
This compact nuisance structure is one reason CamSpec is often described as having a smaller nuisance sector than Plik. In one feature-analysis comparison, the complete CamSpec 1 likelihood is said to have 9 nuisance parameters, 6 foreground and 3 calibration, whereas the corresponding Plik setup can involve many more nuisance degrees of freedom (Hazra et al., 2021).
4. PR4/NPIPE, map-level cleaning, and compressed “lite” forms
The PR4/NPIPE transition changed both the map input and the practical role of CamSpec. The lower noise of NPIPE led to approximately 10% tighter constraints, with smaller error bars and shifts toward 2CDM values for beyond-3CDM parameters including 4 and 5 (Rosenberg et al., 2022). The PR4 paper also introduced a covariance correction relative to an earlier clean CamSpec release: foreground contributions had been left in the covariance despite using foreground-cleaned spectra, slightly overestimating high-6 errors; PR3_12.6 fixed that issue, and PR4_12.6 inherited the correction (Rosenberg et al., 2022).
A later development was the release of a nuisance-marginalized dataset and CamSpec-NPIPE-lite likelihood (Jense et al., 10 Oct 2025). This compressed product is the CamSpec analogue of Plik-lite and is motivated by three linked goals: isolating the CMB-only constraining power of CamSpec-NPIPE after marginalizing over foreground nuisance parameters, enabling correct combination with external small-scale CMB datasets such as ACT DR6 or SPT, and avoiding incorrect use of a naively truncated full multi-frequency CamSpec likelihood when Planck multipole ranges are cut (Jense et al., 10 Oct 2025).
For CamSpec-NPIPE, the full high-7 likelihood consists of five unbinned spectra: 8, 9, and 0 in 1, plus one coadded 2 and one coadded 3 spectrum (Jense et al., 10 Oct 2025). The TE and EE spectra are formed by inverse-noise-variance weighting of all cross-spectra between 100, 143, and 217 GHz, using only cross-spectra to avoid auto-spectrum noise bias (Jense et al., 10 Oct 2025). A standard cut range for joint analyses with ACT or SPT is
4
implemented as camspecnpipe-lite-cut (Jense et al., 10 Oct 2025).
The lite product is built by Gibbs sampling over CMB bandpowers and nuisance parameters. The released Gaussian lite likelihood is
5
with
6
so that the released lite likelihood depends only on 7 (Jense et al., 10 Oct 2025). One explicit result of this marginalization is that the 8 covariance at 9 becomes highly correlated; the paper emphasizes that visible high-00 01 deviations from the best-fitting 02CDM curve are then encoded in the covariance rather than immediately interpretable as a cosmological anomaly (Jense et al., 10 Oct 2025).
5. Role in cosmological parameter inference
A recurrent scientific role of CamSpec is as the high-03 anchor that turns a low-04 or reionization-focused analysis into a full cosmological parameter inference. The relevant degeneracy is
05
since at small angular scales re-scattering suppresses the primordial scalar anisotropies by a factor 06, and primary CMB data then measure mainly the product of the primordial amplitude 07 with that suppression factor (Genesini et al., 23 Mar 2026). In this logic, low-08 09 constrains 10 directly through the reionization bump, while CamSpec supplies the high-11 information that fixes the rest of the acoustic-scale 12CDM physics and the amplitude combination 13 (Genesini et al., 23 Mar 2026).
This complementarity is explicit in the ELiCA analysis. Combining ELiCA with the Planck low-14 temperature likelihood and the CamSpec high-15 likelihood gives
16
at 68% confidence (Genesini et al., 23 Mar 2026). The paper stresses that CamSpec is not rebuilt there; it is adopted as an already established high-17 temperature and polarization likelihood in order to extend a two-parameter low-18 analysis to the full six-parameter 19CDM fit (Genesini et al., 23 Mar 2026).
An earlier optical-depth study used the same pattern with the SRoll2 low-20 momento likelihood. Its preferred final estimate of 21 came from combining the low-multipole SRoll2 momento 22 likelihood with the high-23 CamSpec v12.5HM 24 likelihood, giving
25
in a full six-parameter 26CDM MCMC (Belsunce et al., 2021). That paper is explicit that CamSpec supplies the high-27 information needed to relax externally imposed constraints on 28 and to perform a genuine six-parameter exploration (Belsunce et al., 2021).
Beyond 29, CamSpec underlies standard Planck-only parameter determinations. For PR4_12.6 30, one paper reports
31
32
(Rosenberg et al., 2022). The same analysis reports single-parameter extensions such as 33, 34, 35, and 36 (Rosenberg et al., 2022).
6. Comparisons with Plik, consistency tests, and interpretive debates
CamSpec is often compared with the official Planck high-37 likelihood Plik, and these comparisons are scientifically consequential rather than purely bookkeeping. In the dedicated 2019 CamSpec reanalysis, the six-parameter 38CDM model is reported to provide an excellent fit to the Planck data, with no evidence for statistically significant internal tensions in the 39, 40, and 41 spectra computed for different frequency combinations (Efstathiou et al., 2019). That paper also argues that the tendencies for Planck temperature power spectra to favor 42 and positive spatial curvature are caused by statistical fluctuations in the temperature power spectra in the multipole range 43, and that 44 differs from unity at no more than the 45 level in the statistically most powerful likelihood (Efstathiou et al., 2019).
A distinct line of analysis uses Gaussian processes to compare residual structures across Planck, ACT, and SPT. There, CamSpec is treated as the latest CamSpec NPIPE PR4_v12.6 likelihood, a recent reanalysis of the Planck data with differences in the treatment of polarisation, calibration, and systematic corrections, significantly more sky fraction, and the tightest constraints in terms of cosmological parameters (Calderón et al., 2023). Using CamSpec best-fit spectra as the mean function, that study finds CamSpec broadly self-consistent in 46 and 47, but reports a notable 48 preference for extra uncorrelated variance and interprets this as a possible slight underestimation of the covariance matrix in the CamSpec 49 data, especially at low 50 (Calderón et al., 2023). The same work finds that disagreements between CamSpec and Plik, or between CamSpec and ACT, are mainly visible in 51 residuals rather than 52 (Calderón et al., 2023).
These results do not amount to a general rejection of CamSpec. A later comparison of Planck likelihood choices instead concludes that cosmological parameters from different Planck sky maps and likelihood pipelines are very similar over the Planck multipole range retained in combination with ground-based observations, and that constraints on extended cosmological models become completely insensitive to the choice of Planck maps and likelihood once other CMB datasets are added (Jense et al., 10 Oct 2025). In particular, the additional constraining power from PR3 to PR4 is said to come from polarization at all scales and from temperature at multipoles above 1500 (Jense et al., 10 Oct 2025). This suggests that many headline differences between CamSpec and Plik are concentrated in specific Planck-only configurations and become subdominant in standard joint analyses with ACT or SPT.
A recurrent misconception is therefore that CamSpec and Plik represent wholly incompatible cosmologies. The comparative literature does not support that characterization. It instead emphasizes modest but real shifts linked to mapmaking, sky fraction, cleaning strategy, and nuisance marginalization, with many of those shifts shrinking when the Planck contribution is restricted to the large and intermediate scales usually retained in modern joint analyses (Jense et al., 10 Oct 2025).
7. Use in nonstandard-model and feature searches
CamSpec has also been important in testing whether apparent departures from smooth 53CDM behavior survive changes in likelihood construction. In inflationary-feature analyses, “clean CamSpec” v12.5HMcln is often preferred because it is unbinned and therefore more suitable for sharp oscillatory structures that can be averaged out in binned likelihoods (Hazra et al., 2021). One such study uses CamSpec specifically for oscillatory primordial-feature models, while reserving binned Plik for featureless or broad-suppression models, and concludes that the complete slow-roll baseline potential is moderately preferred against potentials that generate features (Hazra et al., 2021).
A closely related multi-field standard-clock analysis treats CamSpec 2020 as the “statistically most powerful” alternative to Plik, with 80% sky coverage, temperature cleaning with 545 GHz maps, polarization cleaning with 353 GHz maps, and a much smaller nuisance sector than Plik in the full 54 case (Braglia et al., 2021). In that work, CamSpec supports some of the same low-, medium-, and high-frequency feature candidates found with Plik, but generally with smaller average 55, and one dramatic Plik high-frequency candidate is not reproduced by CamSpec (Braglia et al., 2021). This use of CamSpec is methodological as much as inferential: agreement across Plik and CamSpec is treated as evidence of robustness, while failure to reproduce a candidate in CamSpec is treated as evidence that the candidate is likely spurious (Braglia et al., 2021).
The same pattern appears in a damped-oscillation inflation study. There, the authors compare Plik unbinned and CamSpec clean likelihoods and find that the best-fit candidates match closely in parameter space, but the fit improvements are systematically smaller for CamSpec clean than for Plik-bin1 (Antony et al., 2021). The paper explicitly summarizes this as an indication that the extent of improvement is less in CamSpec clean than in Plik-bin1 for all candidates (Antony et al., 2021).
CamSpec also enters curvature reassessments. A closed-universe analysis using CamSpec PR4 with Commander and SimAll reports that a gauge-invariant closed-inflation spectrum shifts the curvature constraint toward spatial flatness, from the standard CamSpec phenomenological result
56
to
57
reducing the preference for closed geometry to about 58 (Specogna et al., 30 Sep 2025). That paper attributes part of the difference between CamSpec and Plik curvature behavior to a less pronounced lensing anomaly in CamSpec PR4 (Specogna et al., 30 Sep 2025).
Taken together, these applications show that CamSpec has become more than an internal Planck alternative. It functions as a robustness benchmark for claims about reionization, feature models, curvature, and extended-parameter fits. The most consistent theme is not that CamSpec always moves results in one direction, but that its specific mapmaking, cleaning, masking, and nuisance choices alter the balance between statistical power and susceptibility to localized high-59 structures. In that sense, CamSpec likelihood is best understood as a mature, compact, and repeatedly stress-tested high-60 Planck likelihood family whose scientific significance lies both in the cosmological constraints it produces directly and in the cross-checking role it plays against other likelihood constructions such as Plik.