Quantifying the global parameter tensions between ACT, SPT and Planck

Will Handley1, 2, 3, ∗ and Pablo Lemos4, † 1Astrophysics Group, Cavendish Laboratory, J.J.Thomson Avenue, Cambridge, CB3 0HE, UK 2Kavli Institute for Cosmology, Madingley Road, Cambridge, CB3 0HA, UK 3Gonville & Caius College, Trinity Street, Cambridge, CB2 1TA, UK‡ 4Department of Physics and Astronomy, University College London, Gower Street, London, WC1E 6BT, UK (Dated: Submitted 16th July 2020) The overall cosmological parameter tension between the Atacama Cosmology Telescope 2020 fourth data release (ACT) and Planck 2018 data within the concordance cosmological model is quantified using the Suspiciousness statistic to be 2.6σ. Between ACT and the South Pole Telescope (SPT) we find a tension of 2.4σ, and 2.8σ between ACT and Planck+SPT combined. While it is unclear whether the tension is caused by statistical fluctuations, systematic effects or new physics, caution should be exercised in combining these cosmic microwave background datasets in the context of the ΛCDM standard model of the universe.

INTRODUCTION the primordial power spectrum ns are mildly discrepant, but it is always possible in a high dimensional parameter As cosmological datasets increase in quantity and qual- space that such discrepancies occur by chance and are ity, so does our capacity to use them to pin down the unremarkable. properties of our universe [1]. The error bars on the In this letter we discuss how this tension is quan- measurements of cosmological parameters have narrowed tified rigorously using the global Suspiciousness statis- over recent years and discrepancies between datasets (or tic [18, henceforth H19], and find that ACT is in mild- “tensions”) have begun to emerge. Whilst this is most to-moderate tension with Planck and SPT, at a similar stark when examining differing observations of the Hub- or greater level to that found in weak lensing data. We ble parameter between early and late time cosmological place ACT’s own global tension analysis in the context of probes [2–4], other more minor tensions arguably exist in the tensions literature, and extend it by considering SPT clustering parameters between weak lensing and the cos- data and further emphasise the perils of focussing too mic microwave background (CMB) [5, 6] and in cosmic closely on lower-dimensional views onto the cosmological curvature between the CMB and CMB lensing/Baryon constraints. Acoustic Oscillations [7–9]. When a substantial tension occurs, it may indicate METHODOLOGY either a systematic error in how either or both of the datasets have been gathered and analysed, or more ex- citingly may hint at evidence for new physics if extensions Quantifying tension between high dimensional poste- or modifications to our concordance model can bring the rior distributions is a non-trivial problem, even under the inferred parameters back into alignment. approximation of a Gaussian distribution. This has led In the case of the “Hubble tension” where a single ob- to a large number of papers describing methods to quan- vious cosmological parameter such as the present day ex- tify tension in high dimensional problems [for reviews, see 10, 11]. Working in a Bayesian framework, as most pansion rate H0 is discrepant by ∼ 5σ, there is little doubt that something is fundamentally wrong. The other cosmological analyses do, arguably the most natural way tensions are more subtle, in that they are only visible in to quantify tension is using the Bayes Ratio [19], defined complicated combinations of the parameters. As shown as the ratio of the probability that the two datasets are by Fig. 1, in modern cosmology, error bars on the param- described by a single set of parameters, to the probability that they are described by separate sets of parameters arXiv:2007.08496v4 [astro-ph.CO] 19 Apr 2021 eters of our universe are represented by high-dimensional Bayesian probability distributions. Visualising a “dis- P (A, B) ZAB tance” between these degrees of belief is challenging, and R = = , (1) P (A)P (B) ZAZB in recent years a good deal of theory has been developed for defining a variety of metrics of discrepancy [10, 11]. where P represent a probability, we have omitted the de- The latest Atacama Cosmology Telescope (ACT) data pendence of both probabilities on an underlying model, release 4 [12–14] represents the most recently acquired such as ΛCDM, and Z is the Bayesian Evidence. Fur- CMB data, with two other measurements of the CMB thermore, we have assumed that both data sets are in- power spectrum across a wide range of multipoles being dependent, an assumption that we further comment on provided by the Planck satellite [15, 16], and the South later. High values of R correspond to concordance, and Pole Telescope (SPT) [17]. By eye it is clear that in the low values are indicative of discordance, with R often in- ACT data some parameters such as the spectral tilt of terpreted on a Jeffreys’ scale [20, 21]. The main issue of 2 this tension metric, in particular for the analysis of cos- and covariance µ and Σ, then the Suspiciousness is: mological data sets, is that it is easily proven that R is d χ2 proportional to the prior volume of shared parameters. log S = − , (5) 2 2 Therefore, R cannot be used for analyses that use delib- 2 −1 erately flat and wide uninformative priors, such as the χ = (µA − µB)(ΣA + ΣB) (µA − µB). (6) analyses of Planck, Dark Energy Survey [DES, 21], Kilo This may be turned into a tension probability via the Degree Survey [KiDS, 6], ACT, SPT, etc. without the ar- survival function of the chi-squared distribution bitrary width of this prior affecting tension assessment. ∞ A more detailed interpretation of this discussion can be Z xd/2−1e−x/2 p = dx, (7) found in H19. 2d/2Γ(d/2) 2 Motivated by this, H19 defined a new statistic, the χ Suspiciousness which keeps all the desired properties and calibrated using a σ-tension by analogy with the of Eq. (1), but corrects for this undesired dependence Gaussian case using the inverse of the complementary on the prior volume. To do so, we divide the Bayes Ra- error function: √ tio in two components: Information and Suspiciousness. σ(p) = 2erfc−1(1 − p). (8) The Information is defined as: Note that, while several methods to quantify tension have been proposed in recent years, they are often built to log I = DA + DB − DAB, (2) recover Eq. (7) and Eq. (8) in the case of Gaussian poste- where D is the Kullback-Leibler divergence [22]. The rior distributions. Therefore, if this work were performed Information contains the dependence on the prior vol- using tension metrics such as Monte-Carlo Parameter ume, therefore by removing it, we obtain a statistic Shifts [26], Parameter Shifts in Update Form [27], or that does not depend on it, but is composed of well- EigenTension [28], we would expect to obtain very simi- defined Bayesian and information theoretic quantities lar, if not the same results, under the Gaussian approx- and is therefore covariantly insensitive to reparameterisa- imation used in this work. This is also equivalent to tion of the space. Therefore, we define the Suspiciousness the multivariate measure of tension used in the ACT pa- as: per [12]. It should be noted that alternative measures of tension R have also been defined and explored that are specialised S = . (3) I for the case when two datasets are correlated [26, 29]. In particular, [23] extended the formalism described in In the language of priors, the Suspiciousness may be in- this section to the case of correlated data sets. Applying terpreted as the most cautious Bayes Ratio R correspond- this to the case of CMB datasets such as Planck, ACT, ing to the narrowest possible priors that do not signifi- SPT and WMAP (which are correlated by virtue of their cantly alter the shape of the posteriors [23]. measuring the same sky) will form the subject of a future A significant innovation to the field which we highlight paper. here, first noted in the appendix F.3 of [24] and explored DATA in detail in [25] is that since log Z = hlog LiP − D, the suspiciousness can be computed from MCMC chains via In this work we analyse the three latest CMB data sets, Planck, SPT and ACT. As with all cosmological analyses, log S = hlog LABiPAB − hlog LAiPA − hlog LBiPB . (4) when considering combining or comparing them at the likelihood level we implicitly assume that the datasets are This observation means that so long as one has pos- independent, even though this may not strictly true. Ex- terior samples for each of the datasets run separately amining the effect of relaxing this assumption will form and in combination, one may compute the suspicious- the subject of future work. It should also be noted that ness without explicitly computing the Bayesian evidence. the prior treatment for τ is different across the three col- However, it should be noted that in non-CMB applica- laborations, one of the aims of future work will be to treat tions only a portion of the parameters are constrained, this in a consistent manner for all three cases. The ACT resulting in hypersurface-like posteriors which are ex- analysis uses a CMB-derived prior for τ, so there is cor- tremely challenging for traditional posterior samplers, relation between posteriors for the τ parameter. A more but present little challenge for nested samplers. complete analysis could adjust the tension in either direc- If the posteriors are such that we may approximate tion, since correlations in τ act to reduce the dimensional- them in the cosmological parameters by a Gaussian (an ity to less than d = 6, increasing the tension [18, 29, 30], approximation which is reasonably justified as shown by but since τ is a degeneracy-breaking parameter it can Fig. 1), as derived in H19, if the d-dimensional posterior have dramatic effects in moving the relative locations of distributions are Gaussian in the parameters with means posteriors, increasing or reducing tension. 3

Planck information on large scales, and as a consequence pro- The Planck mission [31] was a space observatory that duces parameter constraints that are weaker than those Planck measured the CMB for four years between 2009 and from . SPT [17] reports constraints that differ Planck 2013. Planck observed the sky in nine frequencies, from ’s, in particular when only SPTpol’s high between 30 and 857 GHz, with the goal of detecting multipoles are used, but the significance of this reported Planck both temperature and polarization anisotropies, and ac- discrepancy is not quantified. [31] used a pa- curately removing foreground effects. Planck measured rameter difference statistic, and found no evidence for the power spectrum of temperature anisotropies in multi- statistical inconsistencies between the two analyses. Cu- poles ` ∈ (2, 2508), and for E-mode polarization in mul- riously, performing an AL analysis on SPTpol yields a ± tipoles ` ∈ (2, 1996), providing the most powerful con- value lower than one, AL = 0.81 0.14 straints in the parameters of the ΛCDM cosmological Atacama Cosmology Telescope model to date. Beyond the already mentioned tensions in H0 cosmic Finally, we use the Atacama Cosmology Telescope curvature, and with weak lensing; the most puzzling (ACT) posterior samples2 from Data Release 4 (DR4), aspect of the Planck analysis is arguably the AL pa- which used 6000 square degrees at 98 and 150 GHz to 1 rameter . AL was introduced for internal consistency produce power spectrum for temperature and polariza- checks [32], and can smooth the peak of the Planck power tion extending to ` = 4000. Their results by eye ap- spectrum. Planck [31] reports a value AL = 1.180±0.065 pear to be in tension with Planck, and ACT [12] report a for the combination of temperature and polarization, global tension with Planck consistent with that recovered meaning that the data seems to prefer more smoothing of in this paper. the peaks than the best fit ΛCDM cosmology provides. While it has been discussed that this could be caused by a statistical fluctuations, especially since the signifi- RESULTS cance is lower for different versions of the likelihood [33], it has also been hypothesised that it could be a hint of Our results are summarised in Tab. I and Fig. 1. new physics [8], although no theoretical model that pro- When the Suspiciousness tension quantification tech- duces this effect exists in the literature. It is important to niques are applied to the ACT data products in compar- point out that, while this effect is similar to that of CMB ison with the Planck baseline, we find a tension prob- lensing, Planck lensing measurements [34] are compatible ability of p = 0.86%, with a corresponding Gaussian- with AL = 1. calibrated tension of 2.63σ. This level of discrepancy Throughout this letter we use the Planck is generally termed mild-to-moderate, and is compara- legacy archive chains derived using the baseline ble with some of the larger tensions found between weak TTTEEE+low`+lowE+lensing, and have confirmed lensing and CMB data [H19, 6]. that our conclusions are insensitive to excluding the The degree of discrepancy between Planck and ACT is lensing portion of the likelihood. consistent with the level of tension reported by ACT [12]. It is important to note that a global tension quantifi- South Pole Telescope cation such as Suspiciousness does not depend on any specific direction choice in parameter space, nor on the We make use of the South Pole Telescope measure- choice of parameters. It also naturally takes into ac- ments of temperature and polarization from the 500 count the effect that in having d = 6 parameters, it is square degree analysis of their SPTpol instrument [17]. not improbable that some would be in strong marginal This analysis used data at 150 GHz to produce power tension by chance. One can make this point explicit by spectra for the E-mode polarization (EE) and the computing an artificial parameter t defined as the linear temperature-E-mode cross-spectrum (TE). It should be combination of the other parameters θ which maximises noted that the TE and EE have been identified as be- tension. In the Gaussian case this “maximum tension ing in disagreement, so strictly this internal combination parameter” may be computed as should also be viewed with suspicion. The main advan- T −1 tage of SPTpol with respect to Planck is its higher res- t ∝ (µA − µB) (ΣA + ΣB) θ, (9) olution, which allows it to measure much smaller scales, covering a multipole range ` ∈ (50, 8000). However, be- and by construction will have a one-dimensional cause of its smaller sky coverage, SPTpol cannot obtain marginalised tension of χ√, which√ in the case of consis- tency takes the value t ∼ d ± 1/ 2. Maximum tension

1 Often known as ‘lensing parameter’ or Alens, but we will refrain but these names as we believe they can be misleading 2 phy-act1.princeton.edu/public/zatkins/ACTPol_lcdm_1.txt 4

SPT ACT Planck

0.0240 2 h

b 0.0225 Ω

0.0210

0.120 2 h c

Ω 0.105

1.044 MC

θ 1.040 100 1.036

0.10 τ 0.05

3.15 ) s A

10 3.00 ln(10 1.08

1.02 s n

0.96

0.550 t 0.545

0.02100.02250.0240 0.105 0.120 1.036 1.040 1.044 0.05 0.10 3.00 3.15 0.96 1.02 1.08 0.545 0.550 2 2 10 Ωbh Ωch 100θMC τ ln(10 As) ns t

FIG. 1. Measurements of the six parameters of the concordance ΛCDM model using data from the South Pole Telescope (SPT, blue), the Atacama Cosmology Telescope (ACT, orange) and the Planck satellite (green). Plots along the diagonal show one-dimensional marginalised probability distributions normalised to equal height, below the diagonal show iso-probability contours containing 68% and 95% of the 2d marginal probability mass, and above the diagonal show samples drawn from the full probability distribution. Of the six cosmological parameters, visually ACT stands out in tension from the other two most clearly 2 in the ns − Ωbh plane. We can artificially emphasise this further by computing and plotting the linear combination coordinate 2 2 10 of maximum tension between ACT and Planck t = −Ωbh + 0.022Ωch + 34θMC − 0.092τ + 0.05ln(10 As) + 0.067ns, which by construction will have a tension of χ = 4.15σ. Marginalised plots can therefore over-emphasise tension by ignoring the other active coordinates, but the headline statistics in Tab. I are derived from considering the entire distribution as a whole. Plot produced under anesthetic [35] 5

2 coordinates will be discussed in greater detail in an up- Dataset combination χ p tension log S coming work [36]. ACT vs Planck 17.2 0.86% 2.63σ −5.60 Marginalised one and two-dimensional projections of ACT vs SPT 15.4 1.77% 2.37σ −4.68 the posteriors and the Planck-ACT tension coordinate Planck vs SPT 9.1 16.82% 1.38σ −1.55 are summarised in Fig. 1. It is easy for the eye to be ACT vs Planck+SPT 18.4 0.52% 2.79σ −6.22 drawn to certain projections where the marginalised ten- ACT+SPT vs Planck 12.2 5.81% 1.90σ −3.09 sion is large, but as the maximum tension coordinate ACT+Planck vs SPT 10.3 11.09% 1.59σ −2.17 demonstrates, these can be misleading. We emphasise that the Suspiciousness synthesises all of the posterior TABLE I. Global tensions between CMB datasets. For each information correctly into a single intuitive summary pairing of datasets we report the χ2 value calculated using statistic. Eq. (6), the corresponding tension probability p from Eq. (7) Comparing ACT with SPT [37], we find a slightly lower that such datasets would be this discordant by (Bayesian) chance, a conversion into a Gaussian-equivalent tension using mild-to-moderate tension of 2.37σ (p = 1.8%). Interest- Eq. (8) and finall the Suspiciousness from Eq. (5). Addition ingly, comparing SPT with Planck we find no significant signs in the left column indicate combining the datasets at evidence for tension (p = 16.8%), in contradiction with the likelihood level, and combinations below the line should some of the historical literature [17], and in agreement be viewed with suspicion on account of their discordance re- with [31]. ported above the line. Since SPT and Planck are consistent, we may con- fidently combine these datasets. In the absence of a ACT vs Planck tension metric p tension log S full pipeline run, we combine the Gaussian posterior True Suspiciousness Eq. (4) 0.57% 2.76σ −6.10 approximations using Eqs (14)–(20) from H19. This Gaussian approximation Eq. (5) 0.86% 2.63σ −5.60 Planck+SPT combination is 2.79σ in tension with ACT (p = 0.52%), well into the “moderate” regime. TABLE II. We compare the tension computed using the Since ACT is in mild-to-moderate tension with both full non-Gaussian expression from Eq. (4), and the tension Planck and SPT, we should be suspicious of combin- computed via the Gaussian approximation. Note that in ing it with either, but when we do, as in the final two both cases, for this application since all parameters are well- constrained, all that is required are publicly available MCMC rows of Tab. I, we find no significant evidence for ten- chains. sion, although still higher than when comparing Planck and SPT. 2 In Tab. I we also report the χ values for each data view statistical fluctuations at this level as a very unsat- combination, and the suspiciousness log S for reference. isfactory explanation. As log S can be regarded as the most conservative value The general view (or hope) of many members of the log R can take by adjusting priors, it is interesting that cosmological community at the moment is that the cause all values are negative, reflecting the fact that all of the of all of these tensions is likely a combination of (b) and tension probabilities are a little low, when one would tra- (c), and before anyone can claim any kind of new physics ditionally expect p to be uniformly distributed in a fre- we need to get a stronger handle on the systematics in quentist sense, and in general for d = 6 one would expect many of our cosmological probes. positive values of log S 58% of the time. In Tab. II we compare the full non-Gaussian tension As mentioned earlier, this analysis can and will be im- evaluated using Eq. (4) and find the Gaussian approxi- proved by using a full pipeline of evidences and KL diver- mation to be a slight underestimate of the tension. gences computed using nested sampling [38], as well as using techniques that are specialised for dealing with cor- related datasets. However, we would like to draw prac- CONCLUSIONS titioners’ attention in particular to Eq. (4), which allows In general the causes of tension can be one of three them to compute the Suspiciousness using only MCMC things: (a) statistical fluctuation (b) systematics in at chains. least one of the experiments (c) evidence for new physics. In this letter we do not seek to pass judgement on any Given that we confidently launch manned space mis- of the Planck, ACT, or SPT analyses. Indeed, it could be sions with higher failure rates than these tensions3, as argued that given the quality of all three analyses, it is Bayesians we should be very concerned that our CMB more likely that these discrepancies indicate a problem measurements are in this much disagreement, so should with the underlying cosmology, rather than any of the independent pipelines. Combined with the many other tensions emerging between other datasets, the discrep- ancy quantified in this work lends credence to the possi- 3 “Bet someone’s life” probabilities can be computed using data bility that before long we may yet see a paradigm shift from http://www.spacelaunchreport.com/logyear.html in our understanding of the universe. 6

ACKNOWLEDGEMENTS ogy Astropart. Phys., 2020(12):045, December 2020. doi: 10.1088/1475-7516/2020/12/045. [14] ACT Collaboration. The Atacama Cosmology Telescope: WH thanks Gonville & Caius College for their support arcminute-resolution maps of 18 000 square degrees of the via a Research Fellowship. PL thanks STFC & UCL for microwave sky from ACT 2008–2018 data combined with their support via a STFC Consolidated Grant. Many Planck. J. Cosmology Astropart. Phys., 2020(12):046, thanks are accorded to Lukas Hergt for invaluable con- December 2020. doi:10.1088/1475-7516/2020/12/046. [15] Planck Collaboration. Planck 2018 results. V. CMB tributions to the anesthetic package, and to Erminia power spectra and likelihoods. A&A, 641:A5, Septem- Calabrese and Daan Meerburg for comments on an early ber 2020. doi:10.1051/0004-6361/201936386. draft. [16] Planck Collaboration. Planck 2018 results. VI. Cosmo- logical parameters. A&A, 641:A6, September 2020. doi: 10.1051/0004-6361/201833910. [17] SPT Collaboration. Measurements of the Temperature and E-mode Polarization of the CMB from 500 Square ∗ [email protected] Degrees of SPTpol Data. ApJ, 852(2):97, January 2018. † [email protected] doi:10.3847/1538-4357/aa9ff4. ‡ https://www.kicc.cam.ac.uk/directory/wh260 [18] Will Handley and Pablo Lemos. Quantifying tensions in [1] Douglas Scott. The standard model of cosmology: A cosmological parameters: Interpreting the DES evidence skeptic’s guide. Proc. Int. Sch. Phys. Fermi, 200:133– ratio. Phys. Rev. D, 100(4):043504, August 2019. doi: 153, 2020. doi:10.3254/ENFI200007. 10.1103/PhysRevD.100.043504. [2] George Efstathiou. H0 revisited. MNRAS, 440(2):1138– [19] Phil Marshall, Nutan Rajguru, and Anˇze Slosar. 1152, May 2014. doi:10.1093/mnras/stu278. Bayesian evidence as a tool for comparing datasets. [3] Edvard M¨ortselland Suhail Dhawan. Does the Hub- Phys. Rev. D, 73(6):067302, March 2006. doi: ble constant tension call for new physics? J. Cosmol- 10.1103/PhysRevD.73.067302. ogy Astropart. Phys., 2018(9):025, September 2018. doi: [20] H. Jeffreys. Theory of Probability. International 10.1088/1475-7516/2018/09/025. series of monographs on physics. Clarendon Press, [4] Adam G. Riess, Stefano Casertano, Wenlong Yuan, Lu- 1939. URL https://books.google.co.uk/books?id= cas M. Macri, and Dan Scolnic. Large Magellanic Cloud Bu1rAAAAIAAJ. Cepheid Standards Provide a 1% Foundation for the De- [21] DES Collaboration. Dark Energy Survey year 1 re- termination of the Hubble Constant and Stronger Evi- sults: Cosmological constraints from galaxy clustering dence for Physics beyond ΛCDM. ApJ, 876(1):85, May and weak lensing. Phys. Rev. D, 98(4):043526, August 2019. doi:10.3847/1538-4357/ab1422. 2018. doi:10.1103/PhysRevD.98.043526. [5] Marian Douspis, Laura Salvati, and Nabila Aghanim. [22] S. Kullback and R. A. Leibler. On information and suf- On the tension between Large Scale Structures and ficiency. Ann. Math. Statist., 22(1):79–86, 03 1951. doi: Cosmic Microwave Background. arXiv e-prints, art. 10.1214/aoms/1177729694. URL https://doi.org/10. arXiv:1901.05289, January 2019. 1214/aoms/1177729694. [6] KiDS Collaboration. KiDS+VIKING-450: Cosmic shear [23] Pablo Lemos, Fabian K¨ohlinger, Will Handley, Ben- tomography with optical and infrared data. A&A, 633: jamin Joachimi, Lorne Whiteway, and Ofer La- A69, January 2020. doi:10.1051/0004-6361/201834878. hav. Quantifying Suspiciousness within correlated data [7] Will Handley. Curvature tension: evidence for a closed sets. MNRAS, 496(4):4647–4653, August 2020. doi: universe. Phys. Rev. D, art. L041301, February 2021. 10.1093/mnras/staa1836. doi:10.1103/PhysRevD.103.L041301. [24] KiDS Collaboration. KiDS-1000 Cosmology: Multi- [8] Eleonora Di Valentino, Alessandro Melchiorri, and probe weak gravitational lensing and spectroscopic Joseph Silk. Planck evidence for a closed Universe and a galaxy clustering constraints. A&A, 646:A140, February possible crisis for cosmology. Nature Astronomy, 4:196– 2021. doi:10.1051/0004-6361/202039063. 203, February 2020. doi:10.1038/s41550-019-0906-9. [25] Lukas T. Hergt, Will J. Handley, Michael P. Hobson, [9] George Efstathiou and Steven Gratton. The evidence for and Anthony N. Lasenby. Bayesian evidence for the a spatially flat Universe. MNRAS, 496(1):L91–L95, July tensor-to-scalar ratio r and neutrino masses mν : Effects 2020. doi:10.1093/mnrasl/slaa093. of uniform vs logarithmic priors. arXiv e-prints, art. [10] Tom Charnock, Richard A. Battye, and Adam Moss. arXiv:2102.11511, February 2021. Planck data versus large scale structure: Methods to [26] Marco Raveri, Georgios Zacharegkas, and Wayne Hu. quantify discordance. Phys. Rev. D, 95(12):123535, June Quantifying concordance of correlated cosmological data 2017. doi:10.1103/PhysRevD.95.123535. sets. Phys. Rev. D, 101(10):103527, May 2020. doi: [11] Weikang Lin and Mustapha Ishak. A Bayesian inter- 10.1103/PhysRevD.101.103527. pretation of inconsistency measures in cosmology. arXiv [27] Marco Raveri and Wayne Hu. Concordance and discor- e-prints, art. arXiv:1909.10991, September 2019. dance in cosmology. Phys. Rev. D, 99(4):043506, Febru- [12] ACT Collaboration. The Atacama Cosmology Telescope: ary 2019. doi:10.1103/PhysRevD.99.043506. DR4 maps and cosmological parameters. J. Cosmology [28] Youngsoo Park and Eduardo Rozo. Concordance cos- Astropart. Phys., 2020(12):047, December 2020. doi: mology? MNRAS, 499(4):4638–4645, December 2020. 10.1088/1475-7516/2020/12/047. doi:10.1093/mnras/staa2647. [13] ACT Collaboration. The Atacama Cosmology Tele- [29] Fabian K¨ohlinger,Benjamin Joachimi, Marika Asgari, scope: a measurement of the Cosmic Microwave Back- Massimo Viola, Shahab Joudaki, and Tilman Tr¨oster. ground power spectra at 98 and 150 GHz. J. Cosmol- A Bayesian quantification of consistency in correlated 7

data sets. MNRAS, 484(3):3126–3153, April 2019. doi: 10.1051/0004-6361/201833886. 10.1093/mnras/stz132. [35] Will Handley. anesthetic: nested sampling visualisation. [30] Will Handley and Pablo Lemos. Quantifying di- The Journal of Open Source Software, 4:1414, May 2019. mensionality: Bayesian cosmological model complexi- doi:10.21105/joss.01414. ties. Phys. Rev. D, 100(2):023512, July 2019. doi: [36] W Handley, L Lau, P Lemos, and W Barker. Maximum 10.1103/PhysRevD.100.023512. tension coordinates with cosmological applications. In [31] Planck Collaboration. Planck 2018 results. I. Overview preparation, 2020. and the cosmological legacy of Planck. A&A, 641:A1, [37] Eleonora Di Valentino, Silvia Galli, Massimiliano Lat- September 2020. doi:10.1051/0004-6361/201833880. tanzi, Alessandro Melchiorri, Paolo Natoli, Luca Pagano, [32] Erminia Calabrese, AnˇzeSlosar, Alessandro Melchiorri, and Najla Said. Tickling the CMB damping tail: George F. Smoot, and Oliver Zahn. Cosmic mi- Scrutinizing the tension between the Atacama Cos- crowave weak lensing data as a test for the dark uni- mology Telescope and South Pole Telescope experi- verse. Phys. Rev. D, 77(12):123531, June 2008. doi: ments. Phys. Rev. D, 88(2):023501, July 2013. doi: 10.1103/PhysRevD.77.123531. 10.1103/PhysRevD.88.023501. [33] George Efstathiou and Steven Gratton. A Detailed De- [38] John Skilling. Nested sampling for general bayesian com- scription of the CamSpec Likelihood Pipeline and a Re- putation. Bayesian Anal., 1(4):833–859, 12 2006. doi: analysis of the Planck High Frequency Maps. arXiv e- 10.1214/06-BA127. URL https://doi.org/10.1214/ prints, art. arXiv:1910.00483, October 2019. 06-BA127. [34] Planck Collaboration. Planck 2018 results. VIII. Grav- itational lensing. A&A, 641:A8, September 2020. doi: