Arxiv:1908.07862V3 [Astro-Ph.CO] 19 Mar 2021 (ISM)

Astronomy & Astrophysics manuscript no. main ©ESO 2021 March 22, 2021

Testing the Analytical Blind Separation method in simulated CMB polarization maps Larissa Santos1, 2, Jian Yao3, Le Zhang3, Shamik Ghosh4, 5, Pengjie Zhang3, 6, 7, Wen Zhao4, 5, Thyrso Villela8, 9, Jiming Chen4, 5, Jacques Delabrouille10, 11, 5

1 Center for Gravitation and Cosmology, College of Physical Science and Technology, Yangzhou University, Yangzhou 225009, China 2 School of Aeronautics and Astronautics, Shanghai Jiao Tong University, Shanghai 200240, China 3 Department of Astronomy, Shanghai Jiao Tong University, Shanghai, 200240, China 4 CAS Key Laboratory for Research in Galaxies and Cosmology, Department of Astronomy, University of Science and Technology of China, Hefei 230026, China 5 School of Astronomy and Space Sciences, University of Science and Technology of China, Hefei, 230026, China 6 IFSA Collaborative Innovation Center, Shanghai Jiao Tong University, Shanghai 200240, China 7 Tsung-Dao Lee Institute, Shanghai 200240, China 8 Divisão de Astrofísica, Instituto Nacional de Pesquisas Espaciais (INPE), 12227-010 - São José dos Campos, SP, Brazil 9 Instituto de Física, Universidade de Brasília, 70919-970 - Brasília, DF, Brazil 10 Laboratoire Astroparticule et Cosmologie (APC), CNRS/IN2P3, Université Paris Diderot, 75205 Paris Cedex 13, France 11 IRFU, CEA, Université Paris Saclay, 91191 Gif-sur-Yvette, France March 22, 2021

ABSTRACT

Context. Multi-frequency observations are needed to separate the CMB from foregrounds and accurately extract cosmological information from the data. The Analytical Blind Separation (ABS) method is dedicated to extracting the CMB power spectrum from multi-frequency observations in the presence of contamination from astrophysical foreground emission and instrumental noise. Aims. In this study, we apply the ABS method to simulated sky maps as could be observed with a future space-borne survey, in order to test the method’s capability for determining the CMB polarization E- and B-mode power spectra. Methods. We present the ABS method performance on simulations for both a full-sky analysis and for an analysis concentrating on sky regions less impacted by Galactic foreground emission. Results. We discuss the origin and minimization of biases in the estimated CMB polarization angular power spectra. We ﬁnd that the ABS method performs quite well for the analysis of full-sky observations at intermediate and small angular scales, in spite of strong foreground contamination. On the largest scales, extra work is still required to reduce biases of various origins and the impact of confusion between CMB E and B polarization for partial- sky analyses. Key words. Cosmology: Cosmic Microwave Background, methods: data analysis

1. Introduction 2002), VSA (Watson et al. 2003), CBI (Mason et al. 2003), ACBAR (Kuo et al. 2004), BEAST (Meinhold et Cosmic Microwave Background (CMB) experiments detect al. 2005; O’Dwyer et al. 2005), SPT (Sievers et al. 2013), a superposition of microwave sky emissions, in which the ACT (Das et al. 2014), POLARBEAR (Polarbear Collab- CMB signal is mixed with signals originating from various oration et al. 2014), SPTpol (Keisler et al. 2015). The astrophysical foreground sources among which, in particu- CMB was also observed from space with Relikt 1 (Strukov lar, diffuse emissions from the Galactic interstellar medium et al. 1992), COBE (Smoot et al. 1992; Mather et al. arXiv:1908.07862v3 [astro-ph.CO] 19 Mar 2021 (ISM). Multi-frequency observations are needed to separate 1994), WMAP (Bennett et al. 2003; Hinshaw et al. 2013) these different emissions on the basis of their different colors and most recently with the Planck space mission (Planck and accurately measure the CMB angular power spectra, Collaboration 2014). The latest results from the Planck enabling the estimation of the cosmological information of satellite achieve a precise measurement not only of tem- interest with adequate precision. perature anisotropies, but also of CMB polarization E- In the past decades, many ground-based and balloon- modes (Planck Collaboration 2016c,d, 2019), which, just borne experiments have been dedicated to CMB ob- like the CMB temperature anisotropies, are primarily gen- servations. We may cite, for instance, Boomerang (de erated by scalar density perturbations in the early universe. Bernardis et al. 2000), MAXIMA (Hanany et al. 2000), As an outcome of these experiments, much cosmologi- Archeops (Benoît et al. 2003), DASI (Halverson et al. cal information has already been extracted from the CMB. Send offprint requests to: L. Santos e-mail: [email protected] This resulted in the establishment of the standard Λ-CDM

Article number, page 1 of 16 A&A proofs: manuscript no. main cosmological model, with tight constraints on the values atic effects, and in particular of those systematic effects that of its main parameters (Planck Collaboration 2018). Re- induce T -E and T -B mixing. We refer the reader to some cently, much attention has been focused on CMB polar- relevant dedicated studies (Kaplan & Delabrouille 2002; Hu ization anisotropies. Polarization E-modes, of even parity, et al. 2003; Rosset et al. 2007; MacTavish et al. 2008; Shi- convey information complementary to that of temperature mon et al. 2008; Keating et al. 2013; Bicep2 Collaboration anisotropies, allowing to lift some of the degeneracies be- 2015; Natoli et al. 2018; Thuong Hoang et al. 2017; Banerji tween cosmological models and parameter sets and to con- et al. 2019). firm the consistency of the global cosmological scenario. Po- The rest of the paper is organised as follows: Section 2 larization B-modes, of odd parity, are of particular interest introduces the state of the art in component separation for as they are expected to probe tensor perturbations due to CMB observations, and gives a short description of the ABS the primordial gravitational waves (Zaldarriaga & Seljak method and how it compares with other approaches. Sec- 1997; Kamionkowski et al. 1997a; Hu & Dodelson 2002), tion 3 discusses possible sources of biases for CMB power potentially generated during a period of cosmic inflation spectrum estimation with ABS. Section 4 describes the gen- (Grishchuk 1975,1976,1977; Starobinsky 1979,1980; Lyth & eration of simulated data for testing ABS, and the analy- Riotto 1999). On small scales, CMB B-modes mostly origi- sis methodology. Results are presented in Section 5, before nate from the gravitational lensing of E-modes on the path we conclude in Section 6. Three appendixes complement of the CMB photons from the last scattering surface to this by discussing a case with an extra foreground compo- the observers (Zaldarriaga & Seljak 1998; Hu 2002; Lewis nent due to polarized anomalous microwave emission (Ap- & Challinor 2006). The level of large-scale additional B- pendix A); an investigation of the effect of E-to-B leakage modes due to inflationary gravitational waves, if any, is correction for partial-sky analysis (Appendix B); and an not known. It’s determination is one of the primary ob- analytic illustration of the origin of biases in a toy-model jectives of upcoming CMB polarization experiments, such case (Appendix C). as WMPol (Levy et al. 2008), QUBIC (Battistelli et al. 2011), BICEP3 (Kang, J. H. et al. 2018), AliCPT (Li et al. 2017), CLASS (Essinger-Hileman T. et al.), Simons Ob- 2. Component separation and CMB observations servatory (Ade et al. 2019), CMB-S4 (Abazajian et al. 2.1. State of the art 2016), as well as next-generation space missions such as CORE (Delabrouille et al. 2017), LiteBIRD (Matsumura et Various methods exist to separate the emission of fore- al. 2014), EPIC/CMBpol (Bock et al. 2009), PIXIE (Kogut ground astrophysical components from the CMB signal in et al. 2011), PRISM (André et al. 2014), PICO (Hanany et multi-frequency CMB observations. They have been widely al. 2019). The ultimate scientific exploitation of the CMB used for the analysis of existing observations, in particu- as a cosmological probe could also be achieved with a pow- lar those of the WMAP and Planck space missions. These erful spectro-polarimetric survey of the microwave sky (De- methods rely on different assumptions about the CMB and labrouille et al. 2019; Basu et al. 2019; Chluba et al. 2019). foreground emission properties, such as frequency depen- Foreground emissions are particularly critical for mea- dence of emission, correlation between frequencies, para- suring CMB B-modes, which even on the cleanest patches metric form of emission law across frequencies and/or an- of sky are fainter than polarized Galactic ISM synchrotron gular power spectrum. They also differ in their optimization and dust emissions (Planck Collaboration 2016a,e, 2018). criteria and practical implementation. To best control the impact of foreground contamination on Some methods assume that each component is charac- experiments which aim at precisely measuring CMB polar- terised by a specific frequency-dependent emission law (i.e., ization, there is a need for data analysis tools to separate scaling of emission as a function of frequency). The func- the CMB signal from its contaminants in a reliable way tional form of this frequency dependence, and the value and characterize any possible errors in the measured angu- of its parameters, may or may not be assumed to be lar power spectra. known a priori. Methods that assume prior information Foreground contamination is lower if one concentrates about the frequency dependence of foreground emission the observations of CMB B-modes in regions of low fore- (and some assumptions about their spatial distributions) ground emission, away, in particular, from the strong emis- include Wiener Filtering (Bunn et al. 1994; Tegmark & sion of the Galactic interstellar medium in the vicinity of Efstathiou 1996; Bouchet et al. 1999) and the Maximum the Galactic plane. However, there is a drawback to this Entropy Method (Hobson et al. 1998). The Gibbs sam- strategy: the measurement of B-modes then suffers from pling approaches proposed and implemented by Jewell et the so-called E-B mixing, which arises from E-B decom- al. (2004); Wandelt et al. (2004); Eriksen et al. (2004); Lar- position on an incomplete sky coverage (Tegmark & de son et al. (2007); Eriksen et al. (2008); Planck Collaboration Oliveira-Costa 2001). The accuracy of the measurement of (2016a) assume instead a functional form for the emission the B-mode angular power spectrum is thus dependent on of each astrophysical emission, and fit the value of its pa- both our ability to decompose the CMB polarization signal rameters in a global optimization approach. However, the with a partial-sky coverage, and on the effectiveness of the performance of such “non-blind” methods highly depends on component separation methods. the reliability of any prior information, in particular on the In this paper, we discuss the estimation of CMB E-mode functional forms of the frequency dependence of foreground and B-mode power spectra in the presence of foreground emission, and possibly on priors on their parameters. contamination using the ABS method (Zhang et al. 2016). As an alternative, several “blind” approaches have been We test ABS both in a full-sky and a part sky analysis developed. Those only assume a structure for the compo- to compare its performance in these two cases. Although nent mixture (e.g. linear superposition), and typically rely an important issue for precise CMB polarization measure- on the statistical independence of components of various ments, we do not discuss the impact of instrumental system- physical origins for their separation in multi-frequency ob-

Article number, page 2 of 16 Larissa Santos et al.: Testing the ABS method in simulated CMB polarization maps servations. Such blind methods include Independent Com- contain a superposition of CMB, foreground emission, and ponent Analysis (ICA) methods, based on the maximiza- noise as follows: tion of some measure of the non-Gaussianity of indepen- obs cmb fore noise dent sources (Baccigalupi et al. 2004); the Spectral Match- d (`, m) = fd (`, m) + d (`, m) + d (`, m) , (1) ing Independent Component Analysis (SMICA), which uses T spectral diversity and decorrelation to identify independent where f = [f1, ..., fNf ] is the “mixing column” (or “mix- components (Delabrouille et al. 2003; Moudden et al. 2005; ing vector”) of the CMB, and dcmb, dfore, and dnoise refer Patanchon et al. 2005; Cardoso et al. 2008; Betoule et al. to CMB, total foreground emission, and noise respectively. 2009); and the (conceptually close) Correlated Component Such a data model can be written independently for each Analysis (CCA) method (Bonaldi et al. 2006). of the T , E and B fields. Without loss of generality, we can Variants of yet another component separation method, use thermodynamic units for the observations, so that the the so-called Internal Linear Combination method (ILC), CMB emission pattern is constant across frequencies, and have been extensively used for the analysis of WMAP fi = 1, i. Alternatively, we can also use units in which the and Planck survey data to obtain foreground-cleaned noise has∀ unit variance, i.e., re-scale the observations by di- obs CMB maps and power spectra (Tegmark & Efstathiou viding the data di (`, m) from each channel i by its noise noise 1996; Tegmark et al. 2003; Bennett et al. 2003; Eriksen level σi (`) (assumed here to be isotropic, and hence in- et al. 2004; Saha et al. 2006; Delabrouille et al. 2009; dependent of m, for simplicity). Basak & Delabrouille 2012, 2013). The method can be Computing the empirical multivariate power spectrum adapted to also extract maps of other astrophysical com- of the observations (in whatever units), we get ponents (Remazeilles et al. 2011a,b). Closely related ap- obs cmb fore noise proaches based on foreground template matching, includ- ij (`) = fifj (`) + ij (`) + ij (`) , (2) ing the Spectral Estimation Via Expectation Maximisa- D D D D tion (SEVEM) (Martínez-González et al. 2003; Fernández- T where the data, and hence also the vector f = [f1, ..., fNf ] , Cobos et al. 2012), have been used in the analysis of data are expressed in the most convenient units (i.e., thermody- from previous CMB surveys. Finally, methods based on namic units, or noise units for noisy observations). obs(`) sparsity rely on the representation of foreground emission Dij with a limited number of coefficients in a redundant basis of represents the cross-band power spectrum of the (polar- i j functions (Bobin, et al. 2007). The component separation ized) observations in the - and -th frequency channels. problem in the context of CMB observations and the com- The three main contributions to the data are the CMB signal, cmb; the auto and cross band power matrix of the parison of performances of various approaches have been D fore foregrounds, ij ; and the noise contribution to the auto discussed in several dedicated papers (Leach et al. 2008; D noise Delabrouille & Cardoso 2009; Dunkley et al. 2009). and cross-spectra, ij . Note that this assumes vanishing correlation betweenD any two of the maps of CMB, foreground emissions, and noise in any of the channels. While 2.2. The ABS method this is a safe assumption from an ensemble average prospec- A method to directly estimate the angular power spectrum tive, in view of the very different physical origins of those of the CMB from the multivariate angular power spectrum contributions to the total signal, for a single finite data set of multi-frequency observations was proposed by Zhang et empirical correlations exist and, as we will see later, impact al. (2016). This method, called Analytical method of Blind the performance of a method such as ABS, which relies on Separation (ABS), shares conceptual elements with SMICA the decorrelation of the main components in the observa- (both methods estimate the CMB angular power spectrum tions. first, before any separation of component maps), as well as In a first step, we subtract from this multivariate spec- with the ILC (see Sec. 2.3) and with GNILC (see Sec. 2.4: trum an estimate of the noise contribution (the ensemble- both methods estimate a subspace of significant compo- average of the instrumental noise spectrum is assumed to nents). Advantages of the ABS method include a simple im- be known). We get: plementation which does not involve heavy computations, obs cmb fore noise (`) = fifj (`) + (`) + δ (`) , (3) and the fact that it does not rely on any assumption about Dij D Dij Dij the characteristics of the foreground components. It uses noise the measured cross-band power between different frequency where now δ ij (`) is the residual of instrumental noise D bands, from which the CMB power spectrum can be solved contribution after noise de-biasing, and is centred around analytically, avoiding multiple parameter fitting procedure. zero. Assuming an uncorrelated instrumental Gaussian noise A description of the mathematical formalism and numer- noise distribution, with zero mean and rms level σD,i for ical techniques was provided by Zhang et al. (2016). The the i-th frequency channel, the properties of the residual ABS method was also tested against simulated tempera- noise are: ture maps observed by the Planck space mission (Yao et δ noise = 0 , al. 2018). As an extension of this work, we test here the Dij ABS method on simulated CMB polarization observations noise 2 1 noise noise (δ ) = σ σ (1 + δij) , (4) with a sensitive future space mission, targeting the recon- Dij 2 i j struction of the E- and B-mode power spectra. We also noise clarify it’s connection to other classical component separa- where δij is the Kronecker symbol, and i we have σ = ∀ i tion methods, and discuss the origin of possible biases in 1 when we work with noise-whitened observations1 (which the power spectrum estimation. The ABS formalism assumes a data model in which the 1 Note that this amounts to changing the units of the input obs observations in Nf different frequency channels, d (`, m), maps, and hence also the CMB mixing column f.

Article number, page 3 of 16 A&A proofs: manuscript no. main we do in the present paper). The multivariate power spec- 2.3. Connection of ABS to ILC trum, for each `, is a semi-positive symmetric matrix. It can be diagonalized as The ILC method estimates a clean CMB component from multi-frequency observations as the linear combination obs(`) = EΛET , (5) Dij T obs −1 cmb T obs f [D ] obs where the columns of E are an orthonormal basis of eigen- db = w d = d . (9) f T [Dobs]−1f vectors, and Λ = diag(λµ) is a diagonal matrix. Zhang et al. (2016) show that the CMB variance for For each `, the variance of the ILC solution is simply the each ` is obtained as power spectrum of the ILC map for that harmonic mode, M+1 !−1 which can be used as an estimator for the CMB power spec- X bcmb = G2 λ−1 , (6) trum. We get D µ µ µ=1 1 cmb = wT Dobsw = . T P b T obs −1 (10) where G = E f, i.e., Gµ = i Eiµfi, and M is the rank D f [D ] f of fore, or the rank of the subspace spanned by foreground Dij obs −1 −1 T components (i.e. the number of component maps that are Using Eq. 5, and noting that [D ] = EΛ E , we get required to represent the total foreground emission in any −1  Nf  of the observation frequency channels as a linear mixture cmb 1 X 2 −1 b = =  Gµλµ  , (11) of those components). The sum in Eq. 6 extends from 1 to D f T EΛ−1ET f M + 1 to account for M dimensions of foreground emission µ=1 and one extra dimension for CMB (which is not in the foreground subspace when the observations comprise enough which is strongly reminiscent of Eqs. 6 and 7 with, how- frequency channels, in frequency bands that break any pos- ever, two differences: the first is in the range of eigenvec- sible degeneracy). The ABS method, however, thresholds tors and eigenvalues which are kept in the summation, as the ILC keeps all Nf dimensions, while ABS keeps only the eigenvalues λµ, to keep only those that are not dominated by noise variance. After thresholding, the power spec- the most significant ones (with signal-to-noise ratio above trum estimate is implemented in practice as some threshold); the second is that, in Eq. 11, the eigenvalues λµ have not been de-biased from the noise. Hence,  −1 λµ≥λcut the power spectrum that is computed directly from the ILC cmb X 2 −1 b =  Gµλµ  , (7) map must be corrected by subtracting the projected noise. D Simple algebra shows that this amounts to subtracting 1 to each eigenvalue on the diagonal of the Λ matrix. We hence which amounts to keeping only in the sum those eigenvalues recover, after noise de-biasing, the exact ABS solution in which are not dominated by noise, i.e., those that contain the case where M + 1 = Nf (i.e., when there is significant significant foreground or CMB signal. Hence, Eq. 7 gives a signal in all dimensions of the diagonalized system). solution that is almost equivalent to Eq. 6. Yao et al. (2018) have shown that the estimate of the 2.4. Connection of ABS to GNILC CMB power spectrum is not very sensitive to the value of the threshold in the range 0.5 . λcut . 1 (when the system We note that the diagonalization and thresholding that are is written for noise-whitened data). In the present paper, performed in the implementation of ABS are very similar we adopt a threshold λcut = 0.5. to those that are performed in the GNILC method of Re- As a final point, we note that the thresholding step in- mazeilles et al. (2011b), with the exception that the lat- troduces an additional technicality. While thresholding al- ter considers that the “noise” term can contain also astro- lows one to discard noise-dominated eigenmodes, it can also physical components of known multivariate spectrum. In discard part of the signal of interest, and hence introduce that case, the implementation of GNILC in harmonic space a bias. This is the case, in particular, when one works in a (instead of needlet space) would be an extension of the low signal-to-noise regime, as is often the case for the obser- harmonic-space ILC which could be used to separate differ- vation of the faint CMB B-modes. To avoid this potential ent astrophysical foreground emissions in harmonic domain, issue, the ABS method proposed by Zhang et al. (2016) in- and share many common features with the ABS method. troduces a “shift parameter” . Before diagonalization and thresholding, the covariance ofS the observations is “shifted” to 3. Biases in the ABS power spectrum estimation obs obs obs eij (`) = ij (`) + fifj . (8) For any set of observations d (`, m), the data in each D D S channel can be expressed as a linear mixture of at most This corresponds to adding to the covariance of the obser- Nf different maps, as follows: vations an extra CMB term with power in CMB units. S For large , this guarantees that the CMB signal is strong dobs(`, m) = Adcomp(`, m), (12) as comparedS to the noise, and thus “lives” in the subspace spanned by the eigenvectors that are not cut-out in the where dcomp(`, m) is the vector of components present in thresholding process, i.e., the set of eigenvectors that are the observations. Note that such a decomposition is always kept in the summation used for estimating the CMB power possible, at least with A equal to the identity matrix, and (Eq. 7). The amount of CMB power that was artificially dcomp(`, m) = dobs(`, m). This trivial decomposition, how- added to the data is then subtracted from the final ABS ever, is of little practical interest. The key ideas behind power spectrum estimate in the last step of the analysis. ABS (and several component separation methods) are that

Article number, page 4 of 16 Larissa Santos et al.: Testing the ABS method in simulated CMB polarization maps there are physical components which are superimposed in not in the “foregrounds subspace” spanned by the eigenvec- each frequency map; that the CMB is uncorrelated with tors of the foreground covariance matrix fore correspond- the other components (foregrounds, and noise if any); and ing to large eigenvalues of foreground power.D that the scaling of the CMB with frequency is known and Residuals of foreground contamination bias high the universal, i.e. does not depend on the direction on the sky CMB power spectrum estimate with ABS. With enough (rigid scaling: the CMB map observed at one frequency is frequency channels and enough sensitivity, this bias can be a simple scaling of the CMB map observed at any other kept within acceptable values. However, the exact require- frequency, when both observations are at the same angular ments are difficult to evaluate without an accurate model resolution). of foreground emission. Consider the Nf -dimensional space of possible frequency scaling laws for the available observations. We can 3.2. The finite sample bias single-out the direction given by the “mixing vector” of the CMB (i.e. the vector f = [1, 1, ..., 1]t when the observations The covariance matrix of the observations obs is com- are expressed in CMB units). We can also decompose the puted using a finite sample of observations.D For this rea- total foreground emission dfore(`, m) as a linear mixture of a son, even if the CMB is theoretically not correlated with set of independent component maps with respective scaling noise and foregrounds, empirical correlations (and anti- vectors ai, i = 1, .., M . correlations) exist. As discussed at length in the appendix { } In the ideal application case of ABS, the observed fre- of Delabrouille et al. (2009), these empirical correlations quency maps are noiseless, the dimension M of the fore- result in a loss of modes in the reconstructed ILC CMB ground space is such that M + 1 Nf , the CMB direction map. The ABS power spectrum estimate, being similar to (or mixing vector) is not in the M≤-dimensional foreground the noise-debiased power spectrum of the ILC map (and if subspace, and the correlation of CMB with foregrounds and all the eigenvalues of the covariance matrix of the obser- noise exactly vanishes. When this is the case, it is possi- vations are above the threshold, equal to it), suffers from ble, in theory, to perfectly separate the CMB signal from the same bias. This bias can be reduced by increasing the foreground signals, and hence get an unbiased CMB power number of modes that are used to compute the covariances, spectrum estimate with ABS. In real observations of a sin- for instance by binning in ` space. gle sky, complexities arise. Empirical correlations between finite-size samples bias low the CMB power spectrum estimate with ABS. 3.1. Residual contamination by foreground emission 3.3. Thresholding biases In practice, foreground emission is a superposition of many distinct processes with different emission laws. Each region The instrumental noise can lead to noise-dominated eigen- obs modes, with eigenvalues λµ 1 in . Rejecting pos- of galactic emission, each extragalactic source, contributes | | . Dij to the observations with its own emission law. Hence, for itive eigenvalues in the summation of Eq. 7 reduces the real observations, the foreground subspace is that of the value of the sum, and hence increases the estimated power fore original data, and ij (`), has rank Nf . The CMB being spectrum, while rejecting negative eigenvalues does the op- in the space spannedD by foreground components, it is im- posite. possible to get a CMB spectrum which would not contain at In the case where there are M < Nf foreground com- least a bit of positive bias from foreground emission power ponents and one CMB component, all of which associated along the CMB direction (i.e. mixing vector). to eigenvalues that are large as compared to those of the This contamination is minimised by increasing the num- noise, one can select the threshold λcut such that all the ber of frequency channels (and optimising their distribution noise-dominated modes are discarded, and contributions in frequency). Most relevant astrophysical emission laws are from negative and positive rejected eigenvalues cancel out smooth functions of frequency.2 If these functions are sam- on average. One must avoid discarding an excess of nega- pled with a sufficient number of frequency bands, it is pos- tive or positive eigenvalues to avoid a thresholding bias on sible to approximately predict the emission in the frequency the CMB power spectrum estimate with ABS. channel(s) most sensitive to the CMB using a linear combination of observations at other frequencies, which serve 3.4. The choice of a shift parameter as foreground monitors. When the error in this “prediction” is smaller than the noise, one can subtract the foreground The free-shift parameter shifts the amplitude of the in- S cmb cmb contamination down to the noise level (and thus get a rea- put CMB power spectrum from (`) to (`) + . D obsD S sonably good foreground-cleaned CMB map). This changes the eigendecomposition of . In partic- D In the ABS formalism, when this is the case, only a sub- ular, it helps keeping the CMB in the subspace that is set of the observations is sufficient to model the foregrounds not rejected by thresholding. In our calculations, we use noise at any frequency. The rank of the covariance matrix of sig- = 1000σD as a baseline, unless specified otherwise. In fore S nificant foregrounds is M < Nf (other eigenvalues of the case with polarized AME, discussed in Appendix A, we D use = 100σnoise. being smaller than the noise variance). A good choice of fre- S D quency channels is when the mixing vector of the CMB is 4. Simulations and analysis pipeline 2 For completeness, we note that emission from a few atomic or molecular lines, such as CO emission, must also be taken 4.1. Simulated data sets into account (Delabrouille et al. 2013; Planck Collaboration et al. 2014); such emissions, however, impact only a subset of the We test the ABS method on simulated data sets which frequency channels used in the observations. could be obtained with a future sensitive CMB polariza-

Article number, page 5 of 16 A&A proofs: manuscript no. main tion space mission. We consider an experiment with ten fre- Table 1. Experimental setup considered in this study. quency bands, the main characteristics of which are listed in Table 1. The experiment frequency range covers the Band center Beam FWHM noise level maximum of CMB emission relative to astrophysical fore- (GHz) (arcmin) (µKCMB-arcmin) grounds, with extra channels at low and high frequency 030 28.3 12.4 to monitor contamination by synchrotron emission, which 043 22.2 7.9 dominates the low-frequency part of the spectrum, and by 075 10.7 4.2 thermal dust, which dominates the higher part (see Fig- 090 9.5 2.8 ure 1). The 10 frequency channels are a subset of those in 108 7.9 2.3 the experimental setup developed for the Probe of Inﬂation 129 7.4 2.1 and Cosmic Origins (PICO) experiment3. 155 6.2 1.8 223 3.6 4.5 268 3.2 3.1 321 2.6 4.2 30 44 70 100 143 217 353 K] 2

µ Synchrotron 10 4.2. CMB E/B decomposition

Sum foregrounds 1 In the ideal case of full-sky coverage, the CMB linear polar- 10 f = 0.83 sky ization ﬁeld, described by the Stokes parameters Q(ˆn) and f = 0.52 sky U(ˆn), can be unambiguously decomposed into E-mode and Thermal dust fsky = 0.27

0 B-mode components (Seljak & Zaldarriaga 1997; Zaldar- 10 CMB riaga & Seljak 1997; Kamionkowski et al. 1997a,b). How- ever, if the polarization field is not measured over the full Rms polarization amplitude [ -1 sky, the decomposition into E and B modes is not unique, 10 since there are modes that satisfy the properties of both E B 10 30 100 300 1000 and modes simultaneously over the observed sky patch Frequency [GHz] (Tegmark & de Oliveira-Costa 2001). This generates confusion between E and B modes, which must be corrected Fig. 1. Frequency dependence and relative levels of the po- for to correctly estimate the E- and B-mode power spec- larized foregrounds and CMB signal (dominated by polarization E-modes). The Planck experiment frequency chan- tra (Bunn et al. 2003; Bunn 2011; Lewis 2000; Cao & Fang nels are explicitly shown in the grey shaded region 2011; Louis et al. 2013; Grain et al. 2009; Smith 2006; Smith (https://www.cosmos.esa.int/web/planck/picture-gallery). & Zaldarriaga 2007; Zhao & Baskaran 2010; Kim & Nasel- sky 2010; Liu et al. 2018; Liu 2018; Liu et al. 2019,b; Liu 2019), and to carefully take into account residuals that may Theoretical CMB power spectra (E- and B-modes) are still be present in the maps (Santos et al. 2017). obtained with the publicly available CAMB software (Lewis For this work, part-sky analyses are made using the et al. 2000), with parameters from the cosmological model UPB77 Planck mask, smoothed with a Gaussian smooth- fitting best Planck observations (Planck Collaboration ing kernel that extends to a distance δc. Specifically, the 2016b). Polarized CMB Q and U maps are generated us- smoothing function is ing the LensPix software (Lewis 2005; Lewis et al. 2011).  δc 1 1 δ 2 We consider two limiting cases for primordial tensor modes,  + erf − δ < δc r = 0 and r = 0.05. For each frequency channel, we add to W (δ) = 2 2 √2σ (13)  the mock observations emission from two polarized fore- 1 δ > δc. ground components (thermal dust and synchrotron), gen- There is theoretically a small remaining discontinuity at erated with the PySM model (Thorne et al. 2017). We use 4 the default d1s1 model, in which the dust emission law is δi = δc and δi = 0, denoted as β, which is : modelled with a single modified blackbody with spectral 1 1 δc index and temperature varying across the sky, and the syn- β = erf 2 . (14) chrotron emission law in each sky pixel is modelled as a 2 − 2 √2σ ◦ −4 power law with, again, spectral index varying across the Here we set δc = 1 and β = 10 , the latter defining the sky. Gaussian instrumental noise, uncorrelated from pixel value of σ. The resulting mask is displayed in Figure 4. to pixel and from channel to channel, at the level indicated We then use with that particular mask the method pro- in Table 1, is also included in the simulations. For all of this posed by Smith (2006) and Smith & Zaldarriaga (2007), to work, we use the HEALPix (Górski et al. 2005) pixelization compute pure-pseudo multipoles coefficients for E and B, scheme at Nside = 1024. avoiding the “E-to-B” leakage. Polarization maps for the Q and U Stokes parameters, as well as E- and B-mode power spectra, for three repre- 4.3. Analysis pipelines sentative frequency bands, without any beam convolution, are displayed in Figures 2 and 3. We apply the ABS method to two different cases. In the first case, we analyze full-sky maps with no masking. This 4 In practice, the sky being pixelised, any mask has unavoidable 3 https://zzz.physics.umn.edu/ipsig/start discontinuities between adjacent pixels.

Article number, page 6 of 16 Larissa Santos et al.: Testing the ABS method in simulated CMB polarization maps

Fig. 2. The total Q (upper panel) and U (lower panel) simulated full-sky maps (without any beam convolution): CMB + synchrotron + thermal dust + noise. The dust and synchrotron are from the PySM model d1s1 (see text). From left to right: 30, 129, and 321 GHz. The scale is in µKCMB.

Fig. 3. The E- and B-mode angular power spectra of the simulated full-sky maps (without any beam convolution), with a comparison of all the simulated components for three frequency channels: CMB, noise, synchrotron and thermal dust. From left to right: 30GHz, 129GHz, and 321GHz. avoids the E-B confusion problem, at the price of strong In the second case, we must modify the procedure to foreground contamination in regions close to the Galactic take into account the impact of the mask. We follow the plane. The second case considers partial-sky observations, steps below: after applying the mask displayed in Figure 4, with reduced foreground contamination but unavoidable E to B leakage. (A) We smooth all maps to a common resolution of 28.30; (B) We apply the apodized Planck 2015 component separation common polarization mask (UPB77 with fsky = In the first case, ABS is applied in the spectral domain 74.68%) to the final Q and U simulations: CMB + after a full-sky analysis of the input maps into harmonic co- foregrounds + noise for each frequency band; efficients. The covariance matrix of the observations, obs, (C) We reconstruct the pure B-mode pseudo-multipoles for is computed ` by `, and the power spectrum is computedD every frequency map, following the method of Smith following the ABS method as described in Section 2.2. and Zaldarriaga;

Article number, page 7 of 16 A&A proofs: manuscript no. main

Fig. 4. The apodized UPB77 mask used for partial-sky analysis in this work.

(D) We then extract the ﬁnal CMB B-mode spectrum using the ABS method; (E) We repeat the procedure for 50 independent realizations of the instrumental noise, keeping the CMB signal and the foregrounds ﬁxed; (F) We calculate the mean and the standard deviation of the estimated B-mode power spectrum based on the results from steps (C) and (D).

5. Results We now present and discuss the results obtained with the ABS method in the recovery of the CMB E- and B-mode obs power spectra, based on the simulated polarization maps Fig. 5. Eigenvalues (labelled as EGV above) of Dij (`) for each made by the fiducial survey specification given int Table 1. of the 10 frequency bands (represented in different symbol col- Means and associated statistical errors are computed from ors), considering CMB, foregrounds (synchrotron and thermal dust) and one noise realization. Eigenvalues from the E-mode the average of 50 simulations with independent noise real- (upper panel) and B-mode (lower panel) are shown. The thresh- izations. old λ˜cut = 1/2 is displayed as an horizontal solid black line. Note õbs that Dij (`) is not strictly positive: the subspace associated to 5.1. The full-sky case with two foreground components the lowest eigenvalues is dominated by noise, and hence corresponding eigenvalues are centered around zero after the subtrac- õbs ˜ The distribution of eigenvalues in the full-sky case, as a tion of the noise covariance from Dij (`). The eigenvalues, λµ, function of multipole `, is displayed for both E and B are shown in absolute value, negative eigenvalues being displayed modes in Figure 5. There are 10 eigenvalues for each multi- as red dots. obs pole, as ij (`) is the covariance of observations in 10 frequency channels.D In spite of the fact that there are only two foreground components and one CMB, we see that at low spectrum. The former effect is likely to be due to empirical multipoles there are more than three significant eigenvalues correlations between CMB and contaminants (foregrounds above threshold. For all `, however, the lowest eigenvalues and noise). The slight excess power at high ell might be are scattered around or below the threshold, showing that due to the thresholding bias discussed in section 3, and il- the simulated microwave sky is completely decomposed into lustrated on a simple case in Appendix C. those eigenmodes. Results for B modes are displayed in Figure 7. For the The result for the full-sky analysis of E modes is shown fainter B-modes, we use a bin width of ∆` = 50 for dis- in Figure 6, which compares the ABS result with the power playing the results. For r = 0 and r = 0.01, we find that spectrum directly obtained from the pure CMB Q and U 0 the difference between the recovered spectrum and the in- maps without foregrounds and noise, but smoothed to 28.3 put spectrum is below 20% for the full multipole range, (dubbed as the “true” power spectrum) for `max = 1050. even though the 1-σ confidence level for the reconstructed We also show the relative error, in percentage, calculated B-mode power spectrum is quite large at the low-` region. cmb real cmb as: b` / ` 1 where b` is the estimator described In the case of r = 0.05, the relative performance is better in EquationD D 6. − D than 10% for all multipoles. In all cases, for ` > 150, the Although the analysis is performed ` by `, we show recovered power spectrum differs from the input spectrum power spectra after binning in `, with bin width ∆` = 20 by less than 5%. At the very lowest `s, however, the errors for the first bin, and ∆` = 10 for the other bins. Overall, all and biases become too large for a successful estimate of the estimated band-powers are within error. However, we the B-mode power spectrum. Considering the relative level see a tendency for low-` estimates to be lower than the input of foreground contamination as compared to the CMB B spectrum, and high-ell estimates to be somewhat above the modes. This is, however, not a big surprise.

Article number, page 8 of 16 Larissa Santos et al.: Testing the ABS method in simulated CMB polarization maps

Fig. 6. Full-sky binned CMB E-mode power spectrum estimated with ABS, for r = 0, and a Gaussian beam of FWHM=28.30. The red curve corresponds to the simulated CMB E-mode spectrum, while the green curve is the ABS estimate. The associated 1-σ statistical errors are also shown as a light green shadow region, computed from 50 simulations with independent realizations of the instrumental noise. The lower part of the ﬁgure shows the relative error in the estimated power rec true spectrum, D` /D` − 1, in per cents. The color and size of the symbols emphasize deviations from 0% error with respect to the input spectrum.

In Figure 8, we show the change in the B-mode spectrum for different values of the shift parameter introduced in Equation 8. We show the recovered spectrumS for r = 0 using four different values of noise noise noise noise 10σD , 100σD , 1000σD , 10000σD . This shows the impact of the choice of on the power spectrum estimation with ABS. An illustrationS of how this choice may impact the estimator in a simple case can be found in Appendix C. As mentioned in Section 3, on the basis of this investigation, we set the nominal value of the shift parameter to = 1000σnoise. S D Finally, we show in Figure 9 the result of a null test, in which we analyse with the exact same pipeline (and same shift parameter and threshold λcut) a simulation with no CMB signal in the Q and U maps, but only foregrounds and Gaussian noise. The null test is important to assess the limitations and possible biases of our estimator, especially for detecting the extremely faint primordial B-mode polarization signal. We find a small, but significant underestimation in the null spectra below ` = 150, in agreement with the other results shown in this section. Overall, the performance of ABS for power spectrum Fig. 7. Full-sky B-mode spectrum for r=0 (top), r=0.01 (mid- estimation in those tests is quite satisfactory, taking into dle), and r=0.05 (bottom). The red curves correspond to the on simulated CMB B-mode spectrum (including lensing B-modes, account the fact that those estimations have been done while the green curves are the ABS estimate. The lower parts of full-sky simulations, with foreground contamination power the figures show the relative error in the estimated power spec- exceeding the CMB power by orders of magnitude, as shown tra. The color and size of the symbols emphasize deviations from in Figure 3. However, although mostly compatible with er- 0% error with respect to the input spectrum. ror bars, E-mode power spectrum estimates in low-` bins are systematically lower than the input spectrum. This is probably due either to the finite sample bias or the thresh- servations (i.e. binning before ABS estimation, to increase olding bias discussed in section 3, or a combination of both, the statistics) and of λcut. We postpone this optimization and is confirmed by the null-case analysis, in which the re- to future work. For B modes, even though the ABS result covered spectrum is systematically negative for low-` bins. seems promising, improvement of the power spectrum re- Reducing these biases would be possible with a careful construction at this low multipole range is necessary for a choice of `-bins to compute the original covariance of the ob- detection of the primordial gravitational wave imprint in

Article number, page 9 of 16 A&A proofs: manuscript no. main

Similar results in the case of a foreground model that includes polarized anomalous microwave emission (AME) are shown in Appendix A.

5.2. The partial-sky case Using the same CMB, synchrotron and dust simulations, and noise realizations as in the previous section, we now consider the E/B decomposition in partial sky for each analyzed frequency band, using both the ABS foreground cleaning approach and the pseudo-power spectrum reconstruction through the Smith and Zaldarriaga method. The result for E-modes (computed for the r = 0 case) is shown in Figure 10. The performance of the ABS method in recovering the E-mode and B-mode power spectra for 30 ` 1050 is comparable with the full-sky case. The difference≤ ≤ between the recovered power spectrum and the Fig. 8. Full-sky reconstructed B-mode power spectrum for r = 0 with different values for the shift parameter S. “true” one is below 17% for the mentioned multipole range, again with a slight underestimation at low multipoles. For the B-modes, shown in Figure 11, we find the relative error to be less than 21% for 30 ` 1050, and less than 15% for 100 ` 1050. ≤ ≤ ≤ ≤ For smaller multipoles (and especially in the first bin), the determination is not very accurate. The uncertainties here originate from both the ABS method and the E-B mixing correction, the performance of which deteriorates rapidly below ` = 50 in our case. In the same way, the results of the E-B mixing correction deteriorate at high `.A more detailed investigation of the effect, with no foreground contamination, can be found in Appendix B. We do not try improving on the E-B mixing correction with more sophisticated approaches for polarization reconstruction on partial sky here. Options are discussed, for instance, by Ferte et al. (2013); Smith (2006); Smith & Zaldarriaga (2007); Ghosh et al. (2020). We leave such extensions of our investigations to future work. We simply note that for partial-sky surveys it is important, especially in the low-` regime, to carefully perform an analysis with optimized power spectrum estimation approaches to extract the primordial B-modes. This is particularly critical here, as the ABS method post-processes the covariance of the observations, and small errors in some of the entries of the covariance matrix (i.e., auto and cross spectra between the different channels) can propagate in unpredictable ways into the ABS eigendecomposition. Finally, in Figure 12, we see the results for the null test in the partial-sky analysis. The biases at low ` that were seen on the full-sky case are not visible, but this is probably mostly by reason of the large error bar induced by the E-B mixing correction.

6. Conclusions

Fig. 9. The null result (foreground + noise only) for the E-mode In this study, we have tested the performance of the ABS (upper panel) and B-mode (lower panel). method to determine CMB E-mode and B-mode power spectra from multi-frequency observations contaminated by both foreground emission and instrumental noise. The ABS estimator has been applied to simulated maps as could the CMB polarization signal. Resolving the spectrum for be observed by a hypothetical future experiment with 10- low `0s, especially for ` < 40, is challenging if we consider frequency bands in the range of 30 to 321 GHz. Taking the ABS foreground separation method in a full-sky ap- into account 50 independent noise realizations, with one proach. We must then consider an analysis in which we simulated CMB signal and two foreground components, we mask the most foreground-contaminated sky regions. have analyzed a set of 50 Q and U maps to evaluate the

Article number, page 10 of 16 Larissa Santos et al.: Testing the ABS method in simulated CMB polarization maps

Fig. 10. Part-sky CMB E-mode power spectrum (binned with ∆l = 50 bands) estimated with ABS (green curve). The red curve is the CMB E-mode input power spectrum. The blue curve, which shows the pseudo-power spectrum computed directly from the CMB sky realization without noise/foregrounds, using the Smith and Zaldarriaga E/B separation method, illus- trates the uncertainties inherent to the correction of the E-B mixing. The 1-σ statistical errors, computed from 50 simulations with independent realizations of the instrumental noise, are shown as the light green shadow region. The relative error, rec real D` /D` − 1, is shown in percentage level (the error in the first bin, which is much larger than the y-axis scale, is not displayed). The symbol colors and sizes illustrate deviations from 0% in respect to the “true” spectrum. performance of the recovery of the E- and B-mode power spectra. We find that the ABS method is able to estimate E B σ Fig. 11. Same as Fig. 10 but for the B-mode power spectrum, both CMB - and -mode power spectra within 1- error for r = 0 (upper panel) and r = 0.05 (lower panel). bars at most scales (50 ` 1050). In the case of full sky, we calculate the ABS estimator≤ ≤ for tensor-to-scalar ratios r = 0, r = 0.01, and r = 0.05. The polarized E-mode power does not allow for masking contaminated regions after fore- spectrum can be recovered within 20% for ` > 30, reaching ground cleaning and before power spectrum estimation. It an agreement with the input spectrum of 5% compared with is plausible that an interaction between errors in the calcu- the input spectra for ` > 150. For the B mode spectrum, − lation of pseudo spectra and the ABS eigenvalue decompo- considering ∆` = 50, we found that the difference between sition is limiting the perfomance of the combination of these the recovered and the input spectrum is also below 20% for two tools. We leave further investigations of those effects to the full multipole range in every tested case. However, it is future work. important to point out that in all the recovered B mode spectra, the 1σ confidence interval is large for low `−’s, and the method therefore provides poor information with re- Acknowledgments spect to the detectability of low values of r. This work was supported by the National Science Even though our results show that a disentanglement Foundation of China (11621303, 11653003, 11773021, of the primordial B-mode from noise and foregrounds is 11433001), the National Basic Research Program of China challenging with ABS on the full sky in the observational (2015CB85701, 2013CB834900), and the National Key configuration that was tested here, the lensing signal can R&D Program of China (2018YFA0404601). Larissa San- be recovered with excellent accuracy, in spite of the strong tos, Wen Zhao, Shamik Ghosh and Jiming Chen are sup- contamination that comes from foreground emission in the ported by NSFC Grants 11773028, 11603020, 11633001, Galactic plane. 11173021, 11322324, 11653002, 11421303 and 11903030, In the case of a partial-sky analysis, combining the ABS by the project of Knowledge Innovation Program of Chi- together with the Smith and Zaldarriaga method to reduce nese Academy of Science, the Fundamental Research Funds the E-to-B leakage, we find that the recovery of the E and for the Central Universities, and the Strategic Priority B power spectra, for most of the scales (30 ` 1050), ≤ ≤ Research Program of the Chinese Academy of Sciences has a relative error below 21%, for both r = 0 and r = 0.05. Grant no. XDB23010200. T.V. acknowledges CNPq Grant However, the results are not very satisfactory in the low-` 308876/2014-8. region (for both E and B). Contrarily to methods that first construct a foreground-cleaned CMB map and then estimate the power spectrum of that map, the ABS method

Article number, page 11 of 16 A&A proofs: manuscript no. main

Fig. 12. The null result (foreground + noise only) for the E- mode (upper panel) and B-mode (lower panel).

Article number, page 12 of 16 Larissa Santos et al.: Testing the ABS method in simulated CMB polarization maps

Appendix A: Including AME polarized emission We repeat our full-sky analysis for a foreground model that includes polarized anomalous microwave emission (AME). Although it is not expected that AME be strongly polarized, Remazeilles et al. (2016) showed it can bias the derived value of the tensor-to-scalar ratio by non-negligible amounts for satellite missions if an 1% level of polarized AME is neglected. We use the AME polarized model (model 2) from the PySM model (Thorne et al. 2017).

Fig. A.2. Reconstructed B-mode power spectrum for r = 0 with diﬀerent values for the shifting parameter S.

point out that an extra foreground increases the deviations of the ABS reconstruction in respect to the “true" CMB for the low and intermediate multipoles. The main limita- tion in these results, again, comes for the very high level of foreground emission as compared to the target CMB signals. This is immediately visible in Figure A.3, which shows the power spectra of the simulated foregrounds, noise, and CMB inputs for three frequency channels as an example. In addition to very strong synchrotron and dust contamination, the AME is seen to be very signiﬁcant at low frequencies for the polarization level that was assumed here.

Appendix B: Note on E/B separation We now look in more detail at the impact of the performance of our E/B separation method with the pseudo-C` estimator on the reconstructed power spectra. For this pur- pose, we perform 300 simulations of pure CMB signal without noise or foreground. The input power spectra used in the simulations and the Gaussian apodized mask are the same as the ones used for the partial-sky analysis of Sec- tion 5.2. We consider, for illustration, the case of r = 0. The resulting E-mode and B-mode spectra are shown in Fig. A.1. B-mode power spectra obtained with ABS for a full- Fig. B.1. While the mean reconstructed spectra above ` = sky analysis with three foreground components including polar- 50 match very well with the input power spectra, both the ized AME, for both r=0 (top panel) and r=0.05 (lower panel). E- and B-mode pseudo-power spectra deviate substantially from the input spectrum. The B-mode power spectrum estimation results for r = We find that for the range 100 ` 1000 the er- 0 and r = 0.05 are shown in Figure A.1. The input power rors for both the E-and B-mode power≤ spectra≤ are consis- spectrum rests inside the ABS recovery error bars, which tent with the cosmic variance limit. This means that the are, however, big for the multipoles at low and intermediate pipeline is optimal in the range 100 ` 1000. But the `. Once more, the spectrum is well recovered for ` > 170, error sharply increases below ` = 100≤and≤ above ` = 1000. with a relative error < 5%. For ` < 170, this deviation It is evident from the calculations presented here that the reaches 50% in the intermediate scales, but remains within Smith-Zaldarriaga method does not perform very well for error bars, which are large for this full-sky analysis. the largest angular scales (ie., ` 50) with the mask we In Figure A.2, we show the test for the shift parameter use here. ≤ (for r = 0). Here, we observe an underestimate of the B- This behavior also explains the partial-sky results shown noise mode spectrum in the first bin already for = 1000σD , previously in Section 5.2. The large uncertainty and devia- noiseS and we choose as our baseline = 100σD . tion of the reconstructed power spectra in the first bin are The results shown in this appendixS agree with the re- mainly due to the effects from the E/B separation method. sults in Section 5, which states that the larger errors at low In other words, the discrepancy observed at low `s for the `0s obscure the disentanglement of the primordial B-mode ABS reconstruction in partial sky would be attributed to signal from foregrounds and noise. It is also important to the E/B separation method. These results also suffer from

Article number, page 13 of 16 A&A proofs: manuscript no. main

Fig. A.3. The B-mode angular power spectra of the simulated full-sky maps (without any beam convolution), with a comparison of all the simulated components for three frequency channels: CMB, noise, synchrotron, thermal dust and AME. From left to right: 30GHz, 129GHz, and 321GHz.

10 2 100 −

4 2 10− 10−

6 ] ] 4 10− 2 2 10− K K µ µ [ [ 8 6 10− EE BB ℓ 10− ℓ D D

10 8 10− 10− Pseudo-estimator recovered Pseudo-estimator recovered CMB only 12 CMB only 10 10− 10− Pseudo-estimator error Pseudo-estimator error Cosmic variance Cosmic variance

102 103 102 103 ℓ ℓ

Fig. B.1. The result for the recovered E-mode (left panel) and B-mode (right panel) spectra for 300 pure CMB realization simulations and r = 0. The red lines shows the input CMB power spectra, the solid green lines the mean of the reconstructed power spectra. The green dashed lines represents the standard deviation of the reconstructed power spectra. The black dashed lines are displaying the cosmic variance similar increase in the error at high-`s, which will have a which is already diagonal. After subtracting the noise co- T T large contribution from the E/B separation method itself. variance, eigenvectors are e1 = (1, 0) and e2 = (0, 1) , 2 with associated eigenvalues λ1 = σs and λ2 = 0. We keep 2 Appendix C: Special case only the first in the sum of Eq. 6, and we get Cb` = σs . In practice however, ABS starts with the empirical co- Consider the very special hypothetical case where the sky variance of the observations, i.e. using estimators is observed with two channels only, one of which measures a perfect CMB signal with no foregrounds and no noise, 2 1 X 2 σbs = x1, (C.5) and the other one measures only white noise, with no CMB 2` + 1 m signal. Then, we have two data sets, a pure CMB signal 2 1 X 2 x1(`, m) = s(`, m), (C.1) σbn = x2, (C.6) 2` + 1 m and a pure noise signal and the cross-covariance term is x2(`, m) = n(`, m). (C.2) 1 X ∗ x1x = ρσnσs, (C.7) Obviously, in this simple case, the spectral estimation prob- 2` + 1 2 b b lem is trivially solved by disregarding the noise channel, and m estimating the CMB power spectrum as where ρ is the empirical correlation due to the finite sample 1 X size. The empirical covariance matrix is, instead of that of Cb` = x1(`, m). (C.3) Eq. C.4, 2` + 1 m 2 σs ρσsσn Let’s assume that we use instead the ABS prescription. obs = b b b . (C.8) 2 2 b ρσ σ σ2 Consider one specific value of `, and denote as σs = x1 D bsbn bn 2 2 h i and σn = x2 , where . denotes the ensemble average. Assuming thath i the noiseh i is not correlated with the CMB After subtracting the noise covariance (assumed to be ∗ known with negligible error), and adding the shift parame- signal, the ensemble average x1x vanishes. The true co- h 2i ter , we get variance matrix of the observations is S 2 2 σs 0 σs + ρσsσn obs = 2 , (C.4) bobs = b S 2b b 2 . (C.9) D 0 σn D ρσsσn σn σn b b b − Article number, page 14 of 16 Larissa Santos et al.: Testing the ABS method in simulated CMB polarization maps

2 2 2 2 Let = ρσsσn/(σs + ), and δ = (σn σn)/(σs + ). For For the lowest ` values, it can happen that the sec- 2 b S b − b S 2 2 σn and σs σn, both and δ are small, with zero ond eigenvalue λ2 (σn σn) is larger than the thresh- S ∼ 2 ' b − mean and standard deviations old λcutσn. When this is the case, the second eigenvalue is included in the sum to estimate cmb. We get, instead of 1 σsσn σ() = 1, (C.10) Eq. C.21, the solution D √2` + 1 (σ2 + ) bs S ρ2σ2 r 2 cmb 2 bn 2 σn σbs 1 2 2 . (C.22) σ(δ) = 1. D ' − (σn σn) 2 (C.11) b − 2` + 1 (σbs + ) S We now have a negative bias, proportional to 1/(2` + 1), With these notations, we get which is due to the empirical correlation between the CMB 2 1 and the noise. It is the ABS equivalent of the ILC bias bobs = (σ + ) . (C.12) D bs S δ discussed by Delabrouille et al. (2009) (notably in the appendix). We note that if we had not subtracted the noise The eigenvalues, and their approximations to second order expectation value from the covariance of the observations in δ and , are (as is the case with the ILC), the power of the CMB map would have been 2 h i (σbs + ) p 2 2 λ1 = S 1 + δ + (1 δ) + 4 cmb 2 2 2 1 2 − σs 1 ρ σs 1 . (C.23) 2 2 D ' b − ' b − 2l + 1 (σ + )(1 + ) (C.13) ' bs S and The multiplicative factor can be compared to that due to the “ILC bias”, which we recall amounts to a multiplicative 2 h i (σbs + ) p 2 2 term of (1 (m 1)/Np), where m is the number of channels λ2 = S 1 + δ (1 δ) + 4 − − 2 − − and Np the number of independent modes used to compute 2 2 the covariance matrix used in the ILC. With m = 2 ob- (σbs + )(δ ). (C.14) ' S − servation channels and with Np = (2` + 1) independent The corresponding eigenvectors are modes used for the computation of the statistics, the ILC bias factor is exactly the factor found in Equation C.23. It 2 is interesting that thresholding the second eigenvalue out e1 p , (C.15) ∝ δ 1 + (1 δ)2 + 42 of the sum cancels this bias, and replaces it with the posi- − − tive bias of Equation C.21, which can, contrarily to the ILC and bias, be reduced by increasing the shift parameter . S 2 While illustrative and enlightening, the naive case dis- e2 p , (C.16) ∝ δ 1 (1 δ)2 + 42 cussed in this appendix does not capture (by far) the full − − − complexity of what happens when the data comprises noisy which after normalization, and to second order in δ and , observations in several channels, contaminated by several can be approximated as non stationary foreground emissions. Numerical simulations, as performed in this paper, are necessary to inves- 1 2 1 2 tigate in more depth the impact of λcut and on subtle e1 − (C.17) S ' biases that may impact the measurement of faint CMB polarization signals. and + δ e2 1 . (C.18) References ∝ 1 + 2 − 2 Abazajian, K. N., Adshead, P., Ahmed, Z., et al. 2016, arXiv e-prints, 2 −1 2 −1 arXiv:1610.02743. Computing G1λ1 and G2λ2 , we get Ade, P., Aguirre, J., Ahmed, Z., et al. 2019, Journal of Cosmology and Astro-Particle Physics, 2019, 56. 2 (1 2 ) André, P., Baccigalupi, C., Banday, A., et al. 2014, JCAP, 2, 006 G2λ−1 − (C.19) 1 1 ' (σ2 + ) Aumont, J., & Macías-Pérez, J. F. 2007, MNRAS, 376, 739 bs S Baccigalupi, C., Perrotta, F., de Zotti, G., et al. 2004, MNRAS, 354, 2 2 55 2 −1 (1 + δ) Banerji R., Patanchon G., Delabrouille J., Hazumi M., Thuong Hoang G2λ2 2 2 . (C.20) ' (δ )(σs + ) D., Ishino H., Matsumura T., 2019, JCAP, 2019, 043 − b S Basak S., Delabrouille J., 2012, MNRAS, 419, 1163 2 2 2 2 When all goes well, λ2 (σbs + )δ = (σbn σn) σn, and Basak S., Delabrouille J., 2013, MNRAS, 435, 18 the second term is neglected∼ in theS ABS solution− (Eq. 6). In Basu, K., et al. 2019, arXiv, arXiv:1909.01592 Battistelli, E., et al. 2011, Astroparticle Physics, 34, 705 that case, keeping only the ﬁrst eigenvalue and eigenvector, Bennett, C. L., Halpern, M., Hinshaw, G., et al. 2003, ApJS, 148, 1 we get (after subtraction of ), Bennett, C. L., Hill, R. S., Hinshaw, G., et al. 2003, ApJS, 148, 97 S Benoît A., et al., 2003, A&A, 399, L19 2ρ2σ2 Betoule M., Pierpaoli E., Delabrouille J., Le Jeune M., Cardoso J.-F., cmb σ2 1 + bn . (C.21) D ' bs (σ2 + ) 2009, A&A, 503, 691 bs S Bicep2 Collaboration, et al., 2015, ApJ, 814, 110 Bobin J., Starck J.-L., Fadili J., Moudden Y., 2007, ITIP, 16, 2662 The second term in the parenthesis is a positive bias, which Bock, J., Aljabri, A., Amblard, A., et al. 2009, arXiv:0906.1188 is reduced by increasing , and, because of the scaling of Bonaldi, A., Bedini, L., Salerno, E., Baccigalupi, C., & de Zotti, G. the variance of ρ with `, isS more important for large scales. 2006, MNRAS, 373, 271

Article number, page 15 of 16 A&A proofs: manuscript no. main

Bouchet, F. R., Prunet, S., & Sethi, S. K. 1999, MNRAS, 302, 663 Liu, H., 2018, A&A, 617, A90 Bunn, E. F., Fisher, K. B., Hoﬀman, Y., et al. 1994, ApJ, 432, L75 Liu, H., Creswell , J., von Hausegger, S., Naselsky, P. Bunn, E. F., Zaldarriaga, M., Tegmark, M., & de Oliveira-Costa, A., 2019,Phys. Rev. D, 100, 023538 2003, Phys. Rev. D, 67, 023501 Liu, H., Creswell , J., Dachlythra, K., 2019, JCAP, 4, 046 Bunn, E. F., 2011, Phys. Rev. D, 83, 083003 Liu, H., 2019, arXiv:1906.10381 Cao, L. & Fang, L. Z. , 2009, ApJ, 706, 1545 Louis, T., Næss, S. , Das, S., Dunkley, J. & Sherwin, B. , MNRAS, Cardoso, J.-F., Martin, M., Delabrouille, J., Betoule, M., & Patan- 2013, 435, 2040 Lyth, D. H. & Riotto, A. 1999 , Phys. Rept. 314, 1 chon, G. 2008, arXiv:0803.1814 MacTavish C. J., et al., 2008, ApJ, 689, 655 Chluba, J., et al. 2019, arXiv, arXiv:1909.01593 Martínez-González, E., Diego, J. M., Vielva, P., & Silk, J. 2003, MN- Di Valentino, E.; Melchiorri, A.; Silk, J. 2019, Nature Astronomy RAS, 345, 1101 Das, S., Louis, T., Nolta, M. R., et al. 2014, JCAP, 4, 014 Mason, B. S., Pearson, T. J., Readhead, A. C. S., et al. 2003, ApJ, de Bernardis, P., Ade, P. A. R., Bock, J. J., et al. 2000, Nature, 404, 591 955 Mather, J. C. et al. 1994, ApJ, 420, 439 Delabrouille, J., Cardoso, J.-F., & Patanchon, G. 2003, MNRAS, 346, Matsumura, T., Akiba, Y., Borrill, J., et al. 2014, Journal of Low 1089 Temperature Physics, 176, 733 Delabrouille J., Cardoso J.-F., 2009, LNP, 665, 159, LNP...665 Meinhold, P. R., Bersanelli, M., Childers,J., et al. 2005, APJS, 158, Delabrouille, J., Cardoso, J.-F., Le Jeune, M., et al. 2009, A&A, 493, 101-108 835 Moudden Y., Cardoso J.-F., Starck J.-L., Delabrouille J., 2005, Delabrouille, J., Betoule, M., Melin, J.-B., et al. 2013, A&A, 553, A96 EJASP, 2005, 484606 Delabrouille J., et al., 2018, JCAP, 2018, 014 Natoli P., et al., 2018, JCAP, 2018, 022 Delabrouille J., et al., 2019, arXiv, arXiv:1909.01591 O’Dwyer, I. J., Bersanelli, M., Childers,J., et al. 2005, APJS, 158, Dunkley, J., 2009, Ap&SS, 1141, 222. 93-100 Efstathiou, G. , 2004, MNRAS, 349, 603. Patanchon G., Cardoso J.-F., Delabrouille J., Vielva P., 2005, MN- Eriksen, H. K., O’Dwyer, I. J., Jewell, J. B., et al. 2004, ApJS, 155, RAS, 364, 1185 227 Planck Collaboration, Ade, P. A. R., Aghanim, N., et al. 2014, A&A, Eriksen H. K., Banday A. J., Górski K. M., Lilje P. B., 2004, ApJ, 571, A16 Planck Collaboration, Adam, R., Ade, P. A. R., et al. 2016, A&A, 612, 633 594, A10 Eriksen H. K., Jewell J. B., Dickinson C., Banday A. J., Górski K. M., Planck Collaboration, Ade, P. A. R., Aghanim, N., et al. 2016, A&A, Lawrence C. R., 2008, ApJ, 676, 10 594, A13 Essinger-Hileman, T. et al., Millimeter, Submillimeter, and Far- Planck Collaboration, Adam, R. et al. 2016, A&A, 594, A1 Infrared Detectors and Instrumentation for Astronomy VII, 2014, Planck Collaboration, Aghanim, N. et al. 2016, A&A, 594, A11 9153, 91531I Planck Collaboration, Adam, R. et al. 2016, A&A, 586, A133 . Fernández-Cobos, R., Vielva, P., Barreiro, R. B., & Martínez- Planck Collaboration, Akrami, Y., Ashdown, M., et al. 2018, arXiv González, E. 2012, MNRAS, 420, 2162 e-prints , arXiv:1801.04945. Ferte, A. , Grain J., Tristram, M. , Stompor, R. 2013, Phys. Rev. D, Planck Collaboration, et al., 2019, arXiv, arXiv:1907.12875 88, 023524 Planck Collaboration, et al., 2018, arXiv, arXiv:1807.06209 Ghosh, S., Delabrouille, J., Zhao, W, & Santos, L., JCAP (in press), Polarbear Collaboration, P. A. R. Ade, Akiba, Y., et al. 2014, ApJ, arXiv:2007.09928 794, 171 Górski, K. M., Hivon, E., Banday, A. J., et al. 2005, ApJ, 622, 759 Remazeilles M., Delabrouille J., Cardoso J.-F., 2011, MNRAS, 410, Grain, J., Tristram, M. & Stompor, R., Phys. Rev. D, 2009, 79, 2481 Remazeilles, M., Delabrouille, J., & Cardoso, J.-F., 2011, MNRAS, 123515 418, 467 Grishchuk, L. P. 1975 , Sov. Phys. JETP, 40, 409; Grishchuk, L. P. Remazeilles, M., Dickinson, C., Eriksen, H. K. K. & Wehus, I. K. 2016 1976, JETP Lett.,23, 293 ; Grishchuk, L. P. 1977 Ann. N. Y. Acad. MNRAS, 458, 2032-2050 Sci., 302, 439 Rosset C., Yurchenko V. B., Delabrouille J., Kaplan J., Giraud- Halverson, N. W., Leitch, E. M., Pryke, C., et al. 2002, ApJ, 568, 38 Héraud Y., Lamarre J.-M., Murphy J. A., 2007, A&A, 464, 405 Hanany, S., Alvarez, M., Artis, E., et al 2019, arXiv:1902.10541 Saha, R., Jain, P., & Souradeep, T. 2006, ApJ, 645, L89 Hanany, S., Ade, P., Balbi, A., et al. 2000, ApJ, 545, L5 Santos, L., Wang, K., Hu, Y., Fang, W. & Zhao, W., 2017, JCAP, 1, Hinshaw, G., Larson, D., Komatsu, E., et al. 2013, ApJS, 208, 19 043 Hobson, M. P., Jones, A. W., Lasenby, A. N., & Bouchet, F. R. 1998, Shimon M., Keating B., Ponthieu N., Hivon E., 2008, PhRvD, 77, MNRAS, 300, 1 083003 Hu, W. & Dodelson, S. 2002, ARA&A, 40, 171 Smith, K. M. 2006 Phys. Rev. D,74, 083002 Hu, W. 2002 , Phys. Rev. D, 65, 023003 Smith, K. M. & Zaldarriaga, M., 2007, Phys. Rev. D, 76, 043001 Hu W., Hedman M. M., Zaldarriaga M., 2003, PhRvD, 67, 043004 Smoot, G. F., Bennett, C. L., Kogut, A. et al., 1992, ApJ, 396, L1-L5 Jewell, J., Levin, S., & Anderson, C. H. 2004, ApJ, 609, 1 Seljak, U. and Zaldarriaga, M. 1996, ApJ469, 437 Kamionkowski, M., Kosowsky, S., & Stebbins, A. 1997, Seljak, U. and Zaldarriaga, M. 1997, Phys. Rev. Lett.78, 2054 Sievers, J. L., Hlozek, R. A., Nolta, M. R., et al. 2013, JCAP, 10, 060 Phys. Rev. Lett., 78, 2058 Starobinsky, A. A. 1979, JETP Lett. 30, 682; Starobinsky, A. A. 1980, Kamionkowski, M., Kosowsky, S., & Stebbins, A. 1997, Phys. Rev. D, Phys. Lett. B, 91, 99S 55, 7368 Strukov, I. A., Brukhanov A. A., Skulachev, D. P., Sazhin, M. V., Kang, J. H. et al. 2018 , Millimeter, Submillimeter, and Far-Infrared 1992, Soviet Astronomy Letters, 18, 153 Detectors and Instrumentation for Astronomy IX, 10708, 107082N Tegmark, M., & Efstathiou, G. 1996, MNRAS, 281, 1297 Kaplan J., Delabrouille J., 2002, AIPC, 609, 209, AIPC..609 Tegmark, M., & de Oliveira-Costa, A. 2001, Phys. Rev. D, 64, 63001. Keating B. G., Shimon M., Yadav A. P. S., 2013, ApJL, 762, L23 Tegmark, M., de Oliveira-Costa, A., & Hamilton, A. J. 2003, Keisler, R., Hoover, S., Harrington, N., et al. 2015, ApJ, 807, 151 Phys. Rev. D, 68, 123523 Kim, J. 2011 A&A, 531, A32 Thorne, B., Dunkley, J., Alonso, D., & Næss, S. 2017, MNRAS, 469, Kim, J. & Naselsky, P., 2010, A&A, 519, A104 2821 Kogut, A., Fixsen, D. J., Chuss, D. T., et al. 2011, JCAP, 7, 025 Thuong Hoang D., et al., 2017, JCAP, 2017, 015 Kuo, C. L., Ade, P. A. R., Bock, J. J., et al. 2004, ApJ, 600, 32 Wandelt, B. D., Larson, D. L., & Lakshminarayanan, A. 2004, Larson, D. L., Eriksen, H. K., Wandelt, B. D., et al. 2007, ApJ, 656, Phys. Rev. D, 70, 083511 653 Wang, Y. F. , Wang, K. and Zhao, W. 2016, Research in Astronomy Leach, S. M., Cardoso, J.-F., Baccigalupi, C., et al. 2008, A&A, 491, and Astrophysics,16, 059 597 Watson, R. A., Carreira, P., Cleary, K., et al. 2003, MNRAS, 341, Levy, A. R.; Leonardi, R.; Ansmann, M., et al. 2008, APJS, 177, 419- 1057 430 Yao, J., Zhang, L., Zhao, Y., Zhang, P., Santos, L. & Zhang, J. 2018, Lewis, A., 2003, Phys. Rev. D, 68, 083509 ApJS, 239,36 Zaldarriaga, M. & Seljak, U. 1997, Phys. Rev. D, 55, 1830 Lewis, A., Challinor, A., & Lasenby, A. 2000, ApJ, 538, 473 Zaldarriaga, M. & Seljak, U. 1998 Phys. Rev. D58, 023003 Lewis, A. and Challinor, A. 2006, Phys. Rept. 429, 1 Zhang, P., Zhang, J., & Zhang, L. 2019, MNRAS, 484, 1616 Lewis, A., 2005, Phys. Rev. D, 71, 083008 Zhao, W. & Baskaran, D. 2010 , Phys. Rev. D, 82, 023001 Lewis, A., Challinor, A. & Hanson, D. 2011 , JCAP, 03, 018 Planck Collaboration, Ade, P. A. R., Aghanim, N., et al. 2014, A&A, Li, H., Li, S.-Y., Liu, Y., et al. 2017, arXiv:1710.03047 571, A13 Liu, H., Creswell , J., Naselsky, P., 2018, JCAP, 5, 059

Article number, page 16 of 16