Compressed ’CMB-lite’ Likelihoods Using Automatic Differentiation

L. Balkenhol 1,∗ · 2024

Read Compressed ’CMB-lite’ Likelihoods Using Automatic Differentiation by L. Balkenhol 1,∗ (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-00826.txt.

Introduction

Balkenhol shrinks a cosmic data likelihood into a fast lite form in a 2024 paper. The cosmic microwave background is the faint glow left from the early universe. Squeezing multi band measures into clean band powers makes later fits faster and easier to read, but the squeeze itself is costly and noisy.

The new build uses automatic differentiation, a tool that tracks how outputs shift with inputs. One fit and one curvature check now do the whole job in about a minute on a home computer. Tests on a recent telescope data set match the full version closely, with tiny shifts in best fit values. The lite set and a worked example are shared openly.

Catalog record

CREATORL. Balkenhol 1,∗
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-00826

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Compressed ’CMB-lite’ Likelihoods Using Automatic Differentiation L. Balkenhol 1,∗ 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 T ​ T / T ​ E / E ​ E T\!T/T\!E/E\!E 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 ​ σ <0.1\,\sigma , where σ \sigma is the width of the multi-frequency posterior, and the inferred parameter error bars match to within < 10 % <10\% . We publicly release the SPT-3G 2018 T ​ T / T ​ E / E ​ E T\!T/T\!E/E\!E lite likelihood and a python notebook showing its construction on the candl website. \defcitealias dunkley13D13 \journalinfo The Open Journal of Astrophysics 1 Introduction Observations of the cosmic microwave background (CMB) anisotropies form a cornerstone of modern cosmology. Data from recent and contemporary experiments, such as the Planck satellite ( Planck Collaboration et al., 2020a ; Planck Collaboration et al., 2020b ; Planck Collaboration et al., 2020c ) , the South Pole Telescope (SPT) ( Carlstrom et al., 2011 ; Dutcher et al., 2021 ; Balkenhol et al., 2023 ; Pan et al., 2023 ; Ge et al., 2024 ) , the Atacama Cosmology Telescope (ACT) ( Kosowsky, 2003 ; Aiola et al., 2020 ; Choi et al., 2020 ; Madhavacheril et al., 2024 ; Qu et al., 2024 ) , and the BICEP/Keck experiments ( Keating et al., 2003 ; Staniszewski et al., 2012 ; Ade et al., 2021 ) allow us to test the \Lambda (cosmological constant) ( \mathrm{CDM} ) cold-dark-matter model across a wide range of angular scales in temperature and polarisation and search for new physics. These experiments typically supply us with measurements of the CMB anisotropy power spectra at different observational frequencies binned into band powers; from N frequency channels \geq N(N-1)/2 multi-frequency spectra are constructed. At the likelihood level, one then fits for the common CMB signal based on a cosmological model and the varying foreground contamination in the multi-frequency band powers along with any systematic effects. The ensemble of foreground and systematic parameters is referred to as nuisance parameters. While, as their name suggests, the nuisance parameters are not of primary interest, they must be included in a likelihood analysis to obtain the correct results on cosmological parameters. However, nuisance parameters can be more numerous and hence slow down a Markov Chain Monte Carlo (MCMC) analysis by increasing the dimensionality of the parameter space to be explored and by slowing down the evaluation time of the likelihood due to the additional components in the data model. While different methods have been proposed to compress the multi-frequency data into a shorter vector containing only the information of interest, the most common strategy for CMB power spectrum analyses is the construction of a so-called lite likelihood, which contains a set of foreground-marginalised CMB-only band powers. This allows for faster cosmological analyses primarily through the reduction of the number of nuisance parameters. The CMB-lite framework was first introduced by Dunkley et al. (2013) (hereafter 13 ) and has been applied to ACT ( 13 ; Choi et al. (2020) ), SPT ( Calabrese et al., 2013 ) , Planck ( Planck Collaboration et al., 2016 ) , and most recently BICEP/Keck data ( Prince et al., 2024 ) . Though lite likelihoods are advantageous once at hand, their construction is computationally expensive. As set out by 13 , it typically involves sampling a high-dimensional parameter space ( \mathcal{O}(10^{2}) ) with an alternating Gibbs-sampling technique. This procedure naturally produces a covariance matrix with sampling noise, which must be reduced sufficiently or mitigated in another way to avoid adverse effects in subsequent cosmological analyses ( Hartlap et al., 2007 ; Dodelson & Schneider, 2013 ; Percival et al., 2022 ; Balkenhol & Reichardt, 2022 ) . Recently, the application of automatic differentiation in cosmology has seen a surge in popularity (see Campagne et al. (2023) for a series of example applications). In this work, we show that by using the JAX-friendly ( Bradbury et al., 2018 ) , differentiable CMB likelihood library candl ( Balkenhol et al., 2023 ) we can construct lite likelihoods in circa one minute on a personal computer and eliminate the sampling noise in the covariance matrix. We apply this procedure to the SPT-3G 2018 T\!T/T\!E/E\!E data set ( Balkenhol et al., 2023 ) and benchmark the performance of the lite likelihood against its multi-frequency counterpart. The lite likelihood and code to construct CMB-only band powers is made publicly available on the candl website. 1 1 1 https://github.com/Lbalkenhol/candl We begin with background on the CMB-lite framework and the candl library in Section 2 . In Section 3 we outline our technique for the construction of CMB-lite likelihoods and subsequently apply it to the SPT-3G 2018 T\!T/T\!E/E\!E data set in Section 4 . We share our conclusions in Section 5 . 2 Background We briefly review the basics of CMB likelihood inference and outline the CMB-lite framework as introduced by 13 . A CMB power spectrum analysis typically yields multi-frequency band powers C^{\mu\nu,\mathrm{data}}_{b} , i.e. the binned cross power spectrum between observations in frequency bands \mu and \nu . The weights of individual multipole moments of the power spectrum \ell in a band power bin b are given by the window function W_{b\ell} . We usually also have access to the covariance of the multi-frequency band powers, \Sigma_{bb^{\prime}} . We index binned band powers with subscripts b and b^{\prime} , the unbinned power spectrum with \ell , and use repeated indices to imply summation. We formulate model band powers C_{b}^{\mu\nu,\mathrm{model}}(\phi) to describe the data, where the \phi parameters contain CMB band powers C_{b}^{\mathrm{CMB}} , as well as nuisance parameters \theta , \phi=(C_{b}^{\mathrm{CMB}},\theta) . The nuisance parameters enter a frequency-dependent foreground model C_{b}^{\mu\nu,\mathrm{FG}}(\theta) and calibration factors y^{\mu\nu}(\theta) . Explicitly, our data model is: \displaystyle C_{b}^{\mu\nu,\mathrm{model}}(\phi)=y^{\mu\nu}(\theta)\bigg( \displaystyle A^{\mu\nu}C_{b}^{\mathrm{CMB}} \displaystyle+C_{b}^{\mu\nu,\mathrm{FG}}(\theta)\bigg). (1) Here, A^{\mu\nu} is sometimes referred to as the mapping-matrix ( 13 ) or the design-matrix ( Mocanu et al., 2019 ; Planck Collaboration et al., 2020c ; Dutcher et al., 2021 ; Balkenhol et al., 2023 ) ; this matrix of zeros and ones connects the multi-frequency estimates of the same CMB band power in the data vector to the corresponding single element in C_{b}^{\mathrm{CMB}} . Finally, to perform a cosmological analysis we must assume a cosmological model, i.e. a prescription that produces predictions for the CMB power spectrum C^{\mathrm{CMB}}_{\ell}(\psi) for a set of cosmological parameter values \psi , which we can

Read the complete text