Arxiv:1207.5344V2 [Astro-Ph.CO] 17 Oct 2012

CMB EB and TB cross-spectrum estimation via pseudospectrum techniques

J. Grain∗ Universit´eParis-Sud 11, Institut d’Astrophysique Spatiale, UMR8617, Orsay, France, F-91405 and CNRS, Orsay, France, F-91405

M. Tristram† CNRS, Laboratoire de l’Accélérateur Linéaire, Université Paris-Sud 11 Bâtiment200, 91898 Orsay Cedex, France

R.Stompor‡ AstroParticule et Cosmologie, Universit´eParis Diderot, CNRS/IN2P3, CEA/Irfu, Obs. de Paris, Sorbonne Paris Cit´e,France

We discuss methods for estimating EB and TB spectra of the cosmic microwave background anisotropy maps covering limited sky area. Such odd-parity correlations are expected to vanish whenever parity is not broken. As this is indeed the case in the standard cosmologies, any evidence to the contrary would have a profound impact on our theories of the early Universe. Such correlations could also become a sensitive diagnostic of some particularly insidious instrumental systematics. In this work we introduce three different unbiased estimators based on the so-called standard and pure pseudo-spectrum techniques and later assess their performance by means of extensive Monte Carlo simulations performed for different experimental configurations. We find that a hybrid approach combining a pure estimate of B-mode multipoles with a standard one for E-mode (or T ) multipoles, leads to the smallest error bars for both EB (or TB respectively) spectra as well as for the three other polarization-related angular power spectra (i.e. EE, BB, and TE). However, if both E and B multipoles are estimated using the pure technique the loss of precision for the EB spectrum is not larger than ∼ 30%. Moreover, for the experimental configurations considered here, the statistical uncertainties –due to sampling variance and instrumental noise– of the pseudo-spectrum estimates is at most a factor ∼ 1.4 for TT , EE, and TE spectra and a factor ∼ 2 for BB, TB, and EB spectra, higher than the most optimistic Fisher estimate of the variance.

PACS numbers: 98.80.-k; 98.70.Vc; 07.05.Kf

I. INTRODUCTION mordial Universe. Due to the presence of a secondary B- mode contribution generated by the gravitational lensing A reliable and complete characterization of the polar- of the CMB photons by the Universe’s large scale struc- ized cosmic microwave background (CMB) anisotropies ture [11], such a picture is not fully correct, however it is is one of the main challenges in observational cosmology, hoped that the latter signal can be accurately subtracted targeted by a large set of ongoing [1, 2], being-deployed permitting a recovery of the primordial B-modes. [3–6] or planned [7, 8] experiments. CMB (linear) polar- In standard cosmology, the physics governing photon ization is completely described by two Stokes parameters, propagation is parity invariant resulting in vanishing odd Q and U, which are mapped over the celestial sphere by parity correlations, hTBi and hEBi. Nevertheless in CMB experiments. In the harmonic domain, the polar- the context of the next generation CMB B-mode exper- ization field can be described either with help of spin-2 iments, the EB and TB spectra will play an important and spin-(−2) or gradient, E, and curl, B, multipoles. role, both for the data characterization and their scien- From the physics point of view, the gradient/curl de- tific interpretation. This is because on the one hand, arXiv:1207.5344v2 [astro-ph.CO] 17 Oct 2012 composition is more natural as it is directly linked to the these odd-parity cross-spectra are comprehensive, end- cosmological perturbations produced in the primordial to-end, null tests of the presence of instrumental and/or universe. Indeed, for symmetry reasons, scalar pertur- astrophysical systematic effects still present in the data bations can produce solely E-mode-like polarization pat- (see e.g. [12, 13]). On the other hand, as some non- terns and therefore the B component of the polarization standard cosmological mechanisms could produce non- field can be viewed as a direct tracer of the primordial vanishing odd-parity cross-spectra, their detection could gravity waves [9, 10] and thus as a window onto the pri- become a smoking gun of such effects with potentially far-reaching consequences for our understanding of the Universe. Examples of such mechanisms include a primordial stochastic magnetic field, which generates TB ∗ [email protected] and EB correlations, if this magnetic field possesses a he- † [email protected] lical component [14–16]. Similar effects can be obtained ‡ [email protected] due to a rotation of the plane of linear polarization of 2 the CMB photons traveling from the last scattering sur- • a standard approach [27, 28] involving the standard face to our detectors. This could result from either the pseudoharmonic coefficients for both the signal maps Faraday rotation induced by interaction with background and no explicit correction for the E-to-B leakage; magnetic fields [17–20] or interactions with pseudoscalar • a pure approach, when both sets of the multipoles are fields [21, 22]. In all these circumstances a robust and re- pure; [23–25]; and liable EB and TB spectrum estimator could be needed • a hybrid approach, where the E multipoles are to constrain, and potentially correct for, such effects. obtained from the standard estimator and the B signal The goal of this paper is to investigate some of the from the pure estimator [23, 24]. possible extensions of the existing power spectrum estimation methods to incorporate the odd-parity power spectra. Hereafter we work within a framework of the Temperature multipoles are always estimated using the pseudospectrum techniques. In this context the polarized standard pseudo-multipole approach, it may therefore CMB multipoles can be calculated with the help of either appear that there are only two possible estimators for the standard or pure estimators, with the latter specifi- the TB spectra corresponding to B multipoles being ei- cally designed to suppress (or to nearly suppress) E-to- ther standard or pure. In fact, the situation is some- B and B-to-E leakages. In this work we use a specific what more complex due to the fact that, as discussed in prescription for calculating those as introduced in [23] Sec. III C., an unbiased estimation of the TB spectrum and later elaborated on in [24, 25], for some alternatives requires treating it together with the corresponding TE see, e.g., [29–31]. The relative merit of using either the spectrum. Consequently, we arrive again at three types standard or the pure estimates of the pseudo-multipoles of estimators of possible interest, which mirror the case has been discussed in the context of the B-mode power of the EB estimators, and are therefore defined by the spectra showing that the pure approach leads to smaller type of multipoles used for E and B signals. We will error bars [23–25]. This can be understood as follows. refer to them as standard, pure, and hybrid estimators. The pure estimator by construction removes some spa- Our paper is organized as follows. In Sec. II, we tial modes, referred to hereafter as ambiguous modes, present our conventions and notation for describing po- which are the source of the leakage between the two polarization fields on the sphere. Section III is devoted to larization components. This would unavoidably lead to the definition of the different pseudo-C` estimators of the some information loss, reflected in an increase of the fi- EB (Sec. III B) & TB spectrum (Sec. III C), with a par- nal variance of the estimated quantity, were not for a ticular emphasis on the computation of the mode-mode gain incurred as a result of removal the E-mode power coupling matrices. The influence of the sky apodization leaked to the B-spectrum and residing in the ambiguous and pixelization scheme is discussed in Sec. IV. Finally, modes. The latter effect offsets the former one at least the performances of each estimator is discussed in Sec. as long as E B as indeed is expected for the CMB V by quantifying the uncertainties of the reconstructed signal. Consequently, a significant gain in precision is EB and TB spectra considered as null tests. In that usually observed as compared with the standard estima- section, we assume a typical sky-coverage and noise of tor, which retains all the modes together with the leaked suborbital experiments, as inspired by the ongoing ebex power. By the same token, for estimating the E-mode experiment [32], and consider a series of mock observa- power spectrum the standard approach is expected to be tions with progressively more realistic features. A simple more efficient as this time extra variance due to the power Fisher analysis estimate of the power spectra uncertain- of the B modes leaked into E is by far subdominant to ties, which provides a lower limit of the variance of the that resulting from the removal of the ambiguous modes. estimators, is used as a theoretical benchmark. Our con- The question, whether to include or not the ambigu- clusions are presented in the last section (Sec. VI) of this ous modes, and whether to do it consistently or only paper. partially, is less straightforward to answer in the context of the two odd-parity power spectra estimation and has Most of the technical details are provided in the ap- not been addressed to date on either a qualitative or pendices of this paper. Appendix A summarizes the quantitative level. This is the problem that is at the main properties of Gaunt integrals and Wigner-3j sym- focus of this work. To address it hereafter we construct bols used in our derivation of the different mode-mode and investigate performance of three alternative EB coupling matrices, while the details of their computation pseudospectrum estimators1: and explicit formulas are provided in Appendix B. In Ap- pendix C, we provide the noise contribution expected to bias the pseudo-C`’s for the three formalisms. In the following of this paper, the term mask will sys- 1 In principle, a fourth alternative is possible: using standard es- tematically denote the binary, pixelized maps assuming timation for the B-modes and the pure estimation for the E- modes. However, such an alternative is a priori disfavored as a two values, 0 or 1, corresponding to the observed (or kept significant amount of information on the E-modes is lost and the in the analysis) or unobserved pixels. The pixel weights B-modes still suffer from E-to-B leakage and therefore will not which can assume any nonzero value will be referred to be considered in this work. sky apodization or window function. 3

X II. CONVENTION AND NOTATION Hereafter we will use a`m,(X = T,E,B) to denote, the full-sky multipoles as in Eq. (5) and referred to them as A. Spin fields on the sphere the CMB multipoles and use X`m to denote partial-sky, pseudo-multipoles defined in the next section. ? Any spin-s field on the sphere satisfying Φs = Φ−s (in the following, s is always assumed to be ≥ 0) can be decomposed using an E and B basis, III. PSEUDOSPECTRUM ESTIMATORS FOR THE EB & TB CROSS-CORRELATION X Φs(~n) = − (E`m + iB`m) sY`m, `m (1) A. Standard and pure pseudo-multipoles s+1 X Φ−s(~n) = (−1) (E`m − iB`m) −sY`m, `m 1. Temperature case where ±sY`m stands for the spin-weighted spherical harmonics obtained by applying the spin-raising and spin- For temperature anisotropies, the pseudo-multipoles lowering operators (∂ and ∂¯) on the standard (spin-0) are defined as spherical harmonics. Z ∆T (T ) ? s T`m = W Y`md~n, (6) 1 s (−1) ¯s 4π T sY`m = ∂ Y`m and −sY`m = ∂ Y`m, (2) N`,s N`,s where W (T ) is a window function applied to the temper- with ature map. Such pseudo-multipoles are given by a con- T s volution of the CMB temperature multipoles a`m with (` + s)! multipoles of the window function, N = . `,s (` − s)! X (T ) T T`m = K`m;`0m0 a`0m0 . (7) ? s+m 0 0 We remind the reader that sY`m = (−1) −sY`(−m). ` m For the special case of a spin-0 and real valued field (like the CMB temperature field), the B component is vanish- (T ) An exact expression for the convolution kernel K`m;`0m0 ing. Moreover, this also applies to any spin-s field built has been derived in [27, 28, 33, 34] and is recalled in Eq. from a spin-0, real-valued field by use of the spin-raising (B14). s operator (i.e. Φs = ∂ Φ0), as can be directly seen by applying the spin-raising or lowering operators on the harmonic representation of the field as in Eq. (1) and 2. Polarized case: Standard pseudo-multipoles using Eq. (2). CMB anisotropies are described by three fields: a spin- 0 field, corresponding to temperature anisotropies T (~n), In the case of the polarization, pseudo-multipoles can and spin-(±2) fields describing the polarization denoted be introduced most straightforwardly by directly adapt- ing the definition from the temperature case, Eq. (6), ±2P . In terms of the Stokes parameters, the two polarization fields are given by P ≡ (Q±iU). The multipole ±2 1 Z decompositions of such anisotropies are given by, (std) ? ? E`m ≡ − W [2P 2Y`m + −2P −2Y`m] d~n, (8) 2 4π X T Z T (~n) = a`m Y`m(~n), (3) (std) i ? ? B`m ≡ W [2P 2Y`m − −2P −2Y`m] d~n. (9) `m 2 4π X P (~n) = − aE ± iaB Y (~n). (4) ±2 `m `m ±2 `m We will refer to an approach based on these definitions as `m the standard estimator. The resulting standard pseudo- Conversely, at least as long as CMB is mapped over the multipoles are then related to the CMB polarization mul- full sky, the T , E, and B multipoles can be reconstructed tipoles by, from the Stokes parameters by applying the following (std) X h diag E off B i scalar product, E`m = K`m,`0m0 a`0m0 + iK`m,`0m0 a`0m0 , (10) Z `0m0 aT ≡ T (~n) Y ? (~n) d~n, (std) X h diag B off E i `m `m B = K 0 0 a`0m0 − iK 0 0 a`0m0 . (11) 4π `m `m,` m `m,` m Z `0m0 E 1 ? ? a ≡ − [2P 2Y +−2 P−2Y ] d~n, (5) `m `m `m diag off 2 4π The convolution kernels, K and K , are functions Z B i ? ? of the Gaunt integrals and their explicit expression are a`m ≡ [2P 2Y`m −−2 P−2Y`m] d~n. 2 4π given in Eq. (B7) of App. B 1. 4

3. Polarized case: Pure pseudo-multipoles to sky pixelization effects,2 and it is prudent and useful to correct for such effects at least on average, what As emphasized in [35], the pseudo-multipoles as de- can be achieved with help of an appropriately calculated fined above suffer from E-to-B leakage. Following [35], off-diagonal block. Such off-diagonal blocks, though not the polarization field on a partial sky can be split into precisely zero, have been shown to be typically much smaller than the diagonal blocks for the pure estimates, three subspaces: the pure E-modes, the pure B-modes off diag and the ambiguous modes, which contain simultaneously i.e., H`m,`0m0 H`m,`0m0 and, more importantly, than information about E and B modes. These are these the off-diagonal blocks of the standard approach, i.e., off off last modes, which are the source of the E-to-B leakage H`m,`0m0 K`m,`0m0 [23, 25]. This is this latter condi- due to impartial sky coverage. Their presence is due to tion, which ensures that the pure estimator typically re- the fact that on the cut-sky the spin harmonics, ±2Y`m, sults in significant improvements over the standard one at are not anymore orthogonal and any alternative basis least as far as B-mode power spectra are concerned. This constructed out of their linear combinations will include emphasizes the fact that an estimator does not have to be some which will be neither of the B- or E- type. strictly pure, i.e., Hoff 6= 0, to provide good estimates of Because, for the CMB the power of the E signal is the B modes. This indeed turns out to be the case with much bigger than that of B, this leakage is a domi- so-called optimized apodizations [24], which lead to the nant factor for the B-modes estimation giving rise to a smallest overall uncertainty of the estimated BB spec- so called E-to-B leakage problem. Nevertheless, by the tra even if not strictly fulfilling the required boundary same token a certain amount of B-mode power is also conditions. present in the E pseudo-multipoles, which occasionally The pure pseudo-multipoles, Eqs. (14) & (15), can be may also become problematic. related to the standard pseudo-multipoles as follows,

The E-to-B leakage problem can be resolved if the (pure) N`,1 1 E`m = E2,`m + 2 E1,`m + E0,`m, (18) ambiguous modes are identified and excluded from the N`,2 N`,2 spectrum estimation procedure or by using some (incom- N 1 plete) set of basis functions, which by construction are B(pure) = B + 2 `,1 B + B . (19) `m 2,`m N 1,`m N 0,`m orthogonal to such modes. Following [23] such a basis `,2 `,2 can be constructed with help of the spherical harmon- Here, Es,`m(Bs,`m), s = 0, 1, 2 is a spin-s multipole of ? s ics and appropriate window functions W and W1 = ∂W type E (B) of a field, W2−s 2P , where Ws ≡ ∂ W both vanishing at the boundaries of the observed sky. are derivatives of the scalar window, W . It is easy to Denoting these basis functions as, (std) (std) show that E2,`m = E`m and B2,`m = B`m , while the two remaining terms, called the counterterms, are there (W ) 1 2 to cancel the ambiguous modes present in the standard 2Y`m ≡ ∂ (WY`m) (12) N`,2 term, what they do prefectly whenever the window ful- 2 fills the required conditions. Each of the spin-s pseudo- (W ) (−1) ¯2 −2Y`m ≡ ∂ (WY`m) , (13) multipoles is related to the CMB multipoles via spin- N`,2 dependent, mode-mode coupling matrices as in Eqs. (10) we can write now expressions for pseudo-multipoles, and (11). Combined together, those lead to expressions diag off which are free of any E/B leakage as, for the convolution kernels, H and H , introduced in Eqs. (16) & (17). The explicit forms for all those ma- Z trices can be found in App. B 1, Eqs. (B7) and (B7). (pure) 1 h (W )? (W )?i E`m ≡ − 2P 2Y`m + −2P −2Y`m d~n,(14) 2 4π i Z h ? ?i (pure) (W ) (W ) 4. Coupling between a`m and pseudo-a`m B`m ≡ 2P 2Y`m −−2 P −2Y`m d~n.(15) 2 4π It is instructive to briefly comment on how the pseudo- As in the standard case these multipoles can be related multipoles are coupled to the CMB and the window func- to the actual CMB multipoles as, tions multipoles. The mode-mode coupling corresponds to the coupling of three angular momenta: the momen- (pure) X h diag E off B i E`m = H`m,`0m0 a`0m0 + iH`m,`0m0 a`0m0 ,(16) tum (`, m) of either the E or B pseudo-multipoles, the 0 0 `0m0 momentum (` , m ) of either the E or B CMB multipoles 00 00 (pure) X h diag B off E i and the momentum (` , m ) of either the E or B compo- B = H a 0 0 − iH a 0 0 .(17) `m `m,`0m0 ` m `m,`0m0 ` m nents of each spin-s window functions. The total number `0m0

We note that in principle by construction H(off) should be identically zero. This would be however only the case 2 More particularly, the Dirichlet and Neumann boundary condi- were the cut-sky a sole source of the leakage. In practice, tions can never be completely fulfilled by W because of pixeliza- other sources of residual leakage exist, most notably due tion. 5 of such angular momentum couplings is eight as any an- For each of the three alternatives, the pseudo-power PP 0 gular momentum can be of either E- or B-type. Each of spectra, C` , are defined as follows these couplings can be classified according to its parity: 1 X (x) (x)? even or odd. For example, the coupling between E , CEE ≡ E E , (22) `m ` 2` + 1 `m `m E (E) m a`0m0 and ws,`00m00 does not vanish only for even values 0 00 + 1 X (y) (y)? of (` + ` + ` ), as it is made of the G quantities (see CBB ≡ B B , (23) ` 2` + 1 `m `m App. A). We call such a coupling even. In contrast, the m B (E) coupling between E`m, a 0 0 and w 00 00 , made of the 1 X h (x) (y)? (x)? (y)i ` m s,` m CEB ≡ E B + E B , (24) G− quantities, is an odd coupling as it is nonzero only ` 2(2` + 1) `m `m `m `m m for odd values of (` + `0 + `00). From this perspective, the pseudo-multipoles can be written using the following where (x) and (y) stands for the standard and the pure matrix notation pseudo-multipoles respectively. Because of the limited sky-coverage and pixel dependent weights, these estima- 2 even odd tors are biased as their average over CMB realizations in- X X X E B gs −igs E`m = a`0m0 a`0m0 odd even volves a mixing between different `-modes and between igs gs `0m0 `00m00 s=0 different polarization type. Those two classes of mix- (E) ! ing are encoded in the mode-mode coupling matrices w 00 00 0 00 000 s,` m PP ,P P P 00P 000 × (B) , (20) M and the unbiased estimator, C , is com- w ``0 ` s,`00m00 puted by solving the following linear system 2 odd even X X X E B −igs −gs  EE,EE EE,BB EE,EB   EE  B = a 0 0 a 0 0 M 0 M 0 M 0 C 0 `m ` m ` m geven −igodd `` `` `` ` 0 0 00 00 s s BB,EE BB,BB BB,EB BB ` m ` m s=0  M``0 M``0 M``0   C`0  = (E) ! EB,EE EB,BB EB,EB EB M 0 M 0 M 0 C 0 w 00 00 `` `` `` ` × s,` m , (21)  EE EE  (B) C` − N` ws,`00m00 BB BB  C` − N`  , EB EB even odd 0 0 C` − N` where both gs and gs depend on (`, m), (` , m ), and (`00, m00), and geven × godd = 0 if they both have the PP 0 s s where N` stands for the noise contribution to the same arguments. This classification is helpful during the pseudo-power spectra, a priori considered as nonzero. computation of the pseudo-power spectra as any term Expressions for the different noise biases under the as- involving the product of an odd coupling times an even sumption of white noise as well as those for the mode- coupling will vanish. The parity nature of each coupling mode coupling matrix depend on a choice of a specific is summarized in Table I. formalism. They are discussed in Appendices C and B 2, respectively, and below we only briefly summarize main E (E) E (B) B (E) B (B) properties of the coupling matrices referring the reader a`0m0 w 00 00 a`0m0 w 00 00 a`0m0 w 00 00 a`0m0 w 00 00 s,` m s,` m s,` m s,` m for all the details there. E`m even odd odd even B`m odd even even odd

TABLE I. Parity classiﬁcation of the coupling between 1. Standard formalism pseudo-multipoles, CMB multipoles and window function multipoles. In this formalism, assuming the same window function is used for estimating the E and the B pseudo-multipoles, the diﬀerent blocks are related to each other as follows,

EE,EE BB,BB diag M``0 = M``0 ≡ M``0 , EE,BB BB,EE off M``0 = M``0 ≡ M``0 , (25) B. Pseudo-power spectra estimators: EB EB,EB diag off M 0 = M 0 − M 0 , cross-correlations `` `` `` and The two types of the pseudo-multipoles defined in the M EE,EB = −M BB,EB = −2M EB,EE ``0 ``0 ``0 (26) previous section permit us to introduce three different EB,BB cross = 2M``0 ≡ M``0 . pseudospectrum estimators. These include a standard estimator in which both E and B pseudo-multipoles used Their explicit expressions are summarized in Eqs. (B9) to construct pseudo-C`’s are standard; a pure estimator and (B10). off diag in which the E and B pseudo-multipoles are pure; and a Though M``0 is smaller than M``0 , it cannot gen- hybrid estimator which combines a standard estimate of erally be neglected as it describes the leakage from E- the E pseudo-multipoles with a pure estimate of the B modes to B-modes, which has to be corrected for, in par- pseudo-multipoles. ticular if the maps to be analyzed cover a tiny amount of 6 the celestial sphere (see for example Fig. 9 of [25]). As C. Pseudo-power spectra estimators: TB cross shown in App. B 2, the M``0 matrix is equal to zero, cross-correlations emphasizing the fact that the EB cross-spectrum is not coupled to the EE and BB auto-spectra. The main rea- The formalism developed in the previous section is eas- son for this is that whenever the window functions con- ily adaptable to the second odd-parity power spectrum, tains only a E-component, as is precisely the case in the TB C` . Because of the E-to-B (and B-to-E) leakage, the standard formalism, E pseudo-multipoles are evenly cou- TB spectrum has to be analyzed with the TE in a unified pled to E multipoles and oddly coupled to B multipoles framework. However, cut-sky effects do not introduce while the B pseudo-multipoles are oddly coupled to E any leakages from temperature to Q or U Stokes param- multipoles and evenly coupled to B multipoles. As a con- eter. As a consequence, there is no leakage from TT , EE, cross sequence, the M``0 block involves product of the even BB, and EB spectra into the TB and TE pseudospec- and odd coupling and thus vanishes. This demonstrates tra and the TB spectrum can be treated independently that the EB cross-spectrum can therefore be treated in- of the EE,BB,EB. dependently of the EE and BB spectra in this formalism. The two pseudospectra are defined as follows

1 X h (x)? (x)i CTE ≡ T E + T ? E , (27) ` 2(2` + 1) `m `m `m `m 2. Pure formalism m 1 X h (y)? (y)i CTB ≡ T B + T ? B , (28) ` 2(2` + 1) `m `m `m `m In the pure formalism, the different mode-mode cou- m pling blocks satisfy the same relations as given in Eqs. from which an unbiased estimator is built by inverting (25) and (26). Nevertheless their respective expressions the usual mode-mode coupling matrix, i.e. are qualitatively and quantitatively different in both these cases, see Eqs. (B11) and (B12). First of all, the T E,T E T E,T B TE TE TE M M C 0 C − N off ``0 ``0 ` = ` ` . pure estimation ensures the M 0 block be small enough T B,T E T B,T B TB TB TB `` M 0 M 0 C`0 C` − N` to substantially reduce amount of E-to-B leakage even in `` `` the cases of small-sky coverage (e.g., see Fig. 9 of [25]). The explicit expressions of the mode-mode coupling ma- cross Second, unlike the standard case, the four blocks M``0 , trix for the three formalisms are provided in App. B 3. which couple the EB cross-spectrum to the EE and BB We observe that in the standard and pure formalisms, T E,T E T B,T B T E,T B T B,T E autospectra, may not vanish. These blocks involve two M``0 = M``0 and M``0 = −M``0 as long types of couplings: i) odd-even couplings corresponding as the same window functions are used to compute the to the EE and BB power spectra of the window func- E and B multipoles. However, the off-diagonal blocks, tions and ii) even-even and odd-odd couplings via the T E,T B M``0 are equal to zero in the standard formalism but cross-correlation of the E-component of the window func- may not vanish in the pure one if the spin-1 and spin-2 tions with their B-component. Though the first type of windows have a nonzero B-component. The fundamental couplings is identically equal to zero, the second class of T reason is that pseudo-a`m are coupled only to the prod- couplings does not vanish if the spin-1 and spin-2 win- uct of the E-component of the window functions applied dows have a nonvanishing B-component. T to the maps times the a`m via an even coupling while in the standard formalism, E(B)-pseudo-multipoles are coupled to B(E)-multipoles via an odd coupling. How- 3. Hybrid formalism ever, the pure formalism introduces an additional even coupling of the E(B)-pseudo-multipoles with the B(E)- The case of the hybrid formalism is more intricate multipoles via the B-component of the nonzero-spin window functions, which translates into nonvanishing off- as it involves the cross-correlation of the pure pseudo- T E,T E multipoles with the standard pseudo-multipoles. De- diagonal blocks. In the hybrid formalism, the M``0 0 T E,T B noting by X, and X either E or B, we find that and M``0 blocks are formally equal to the ones com- the (EE,XX0) blocks are equal to the (EE,XX0) T B,T B puted in the standard formalism while the M``0 and blocks as computed in the standard formalism while the M T B,T E blocks are equal to the ones as computed in the 0 0 ``0 (BB,XX ) blocks are given by the (BB,XX ) blocks pure formalism. as derived in the pure formalism. The expression for the (EB,XX0) blocks is given by Eq. (B13) and in- diag/off volves a cross-product of the matrices K``0 and IV. SKY APODIZATION AND PIXELIZATION diag/off EFFECTS H``0 . As in the case of the pure formalism, the (EB,XX0) blocks generally do not vanish as they contain even-even and odd-odd couplings corresponding to A. Sky apodization the cross-correlation of the E-component of the spin-0 window with the B-components of the spin-1 and spin-2 Applicability of the pure pseudospectrum estimator window functions. depends on our ability to compute sky apodization with 7 the appropriate boundary conditions. Different functions have been proposed. They range from ones derived as a result of an optimization procedure to ones based on some analytic expressions. Their relative merit has been ex- tensively discussed in [25] to show that the pixel-domain optimization scheme proposed in [24] leads to the lowest variance on the power spectrum estimation. The underlying strategy to derive such an optimized sky apodization is to search for window functions, W , which make the pure pseudo-C` as close as possible to the optimal, quadratic power-spectrum estimator (see Sec. V of [24]). An optimized weighting scheme therefore involves a specific sky apodization for each `-band where the power spectrum is to be estimated, according to the signal and noise power aliasing in each band. Such an optimization procedure assumes that noise and signal are known. For the noise and E-modes sufficiently precise assumptions are usually available, e.g., E-mode power spectrum can be recovered precisely enough without any optimization. Though this is not the case with the B-modes, it can be shown [25] that the assumptions about the B-modes power spectrum very weakly affect the results of the power-spectrum estimation using the FIG. 1. The red-dashed curves correspond to the ` = 300 optimized sky apodization. This is due to the fact that column of the (TE,TE) coupling kernels computed in the the optimization process foremost attempts to reduce the standard (left) and pure (right) formalisms. The standard E-to-B leakage and then the noise variance and thus no (TE,TE) block equals the hybrid (TE,TE) block and stan- precise B-mode knowledge is necessary. We note that dard (TB,TB) block. The overlapping black curve shows the in this approach as the derivative relationship is relaxed TE obtained from a map containing TE and TB correlations during the computation, the B-component of the spin- at ` = 300 only. The agreement between the red and black 1 and spin-2 windows becomes nonzero. Therefore, the curves underlines the correctness of the numerical computa- (EB,XX) and (XX,EB) blocks of the mode-mode cou- tion of the different the mode-mode coupling matrices. pling will not vanish. Nevertheless, numerical studies have shown that the B-component of their spin-1 and spin-2 windows is very small. As shown in Appendix B 2, only the cross-spectra of the E-component and B- off-diagonal blocks depends on the sky apodizations component of the spin-weighted window functions are used in the numerical computation of the pure pseudo- involved in the (EB,XX) and (XX,EB) while the au- multipoles. Those blocks are expected to be zero for tospectra of both E- and B-component of those window either the analytic or the harmonic-based optimized win- functions enter in the other blocks. As a consequence, we dow functions but may become nonzero for the pixel- expect the (EB,XX) and (XX,EB) blocks of the mode- based optimized sky apodizations. From a quantitative mode coupling matrix to be much smaller than the other point of view, the off-diagonal blocks share common fea- blocks. tures with the (XX,EB) and (EB,XX) blocks. They in particular depend only on the cross-spectra of the E- Other options for the apodizations considered to component and B-component of window functions and date [25] include i) analytic window functions, and ii) we therefore expect those off-diagonal blocks to be much harmonic-based optimized window functions. In these smaller than the diagonal ones if pixel-based optimized cases, the spin-1 and spin-2 windows are computed by sky apodizations are used. taking numerically the derivative of the spin-0 window. As a consequence, none of the spin-s sky apodizations have a B-component and the (EB,XX) and (XX,EB) blocks of the mode-mode coupling are expected to vanish B. Pixelization impact on the mode-mode coupling within numerical precision in both the pure and the hy- matrices brid formalism. Hence, in these cases the EB spectrum is effectively decoupled from EE and BB, as in the stan- To test our implementation of the mode-mode coupling dard formalism, but in a contrast with the pixel-domain matrices, we compare it to the pseudospectrum obtained apodization cases. for an initial power spectrum with nonzero power only at

Similar arguments apply to the case of mode-mode a single angular frequency, i.e., ∝ δ`,`0 . For example, the coupling for the TB cross-spectrum. In the pure TE pseudospectrum derived from an initial power spec- TE TB and hybrid formalism, the potential cancellation of the trum such as C` = δ`,`0 and C` = 0 is equal to the 8

FIG. 2. The red-dashed curves corresponds to the ` = 300 column of the (EB,EB) coupling kernel for the three formalisms: standard, pure, and hybrid formalisms (from left to right). The overlapping black curve shows the EB pseudospectrum obtained from a map containing EB correlation at ` = 300 only. The agreement between the red and black curves underlines the correctness of the numerical computation of the diﬀerent the mode-mode coupling matrices.

T E,T E elements of the `0 th column of M`,`0 . Similarly, the and EB pseudospectra as derived from a map contain- TB pseudospectrum for such a process is given by the `0 ing correlated T -, B-, and E-modes at ` = 300 only. The T B,T E th column of the M`,`0 block. Such a calculation has overall agreement of red-dashed and black curves pins also an additional advantage as the resulting pseudospec- down the right computation of the mode-mode coupling tra include other effects, which could potentially lead to matrices. The agreement is nevertheless not perfect for nonzero couplings, for instance, such as possible pixeliza- two reasons. First, the equality between the column of tion effects, and which are not taken into account in the the mode-mode coupling and the pseudospectrum from computation of the mode-mode coupling matrices. This a single mode process is exact on average. However, the test has already been applied in [25] to the (EE,EE), black curves displayed on Figs. 1 and 2 are obtained from (BB,BB), and (EE,BB) blocks to show that most of one realization only, explaining the scattering of the black the pixelization effects are carefully taken into account curves around the red ones. Second, we cannot exclude and we will only present our results concerning the re- the influence of residual pixel effects. maining blocks of the full mode-mode coupling kernels. As already explained, the (TB,TE) and (EB,PP ) blocks are nonvanishing (though expected to be small) An example of such tests is shown on Fig. 1 for the case in the peculiar case of the pure and hybrid formalisms if of the (TE,TE), (TB,TB) and in Fig. 2 for (EB,EB) and only if the spin-1 and spin-2 sky apodizations have a blocks in the three proposed formalism (standard, pure nonzero B-component. This is the case of pixel-based op- and hybrid formalisms). The sky coverage adopted here timized sky apodizations. However, the above described is a spherical cap with a radius of ∼11.3 degrees. The tests shows that within the HEALPix scheme, and irre- sky apodizations used are the analytic windows with a 30 spectively of any formalisms, unmodeled pixel effects in- arcminutes apodization length to be applied to temper- duce some nonvanishing (TB,TE) and (EB,PP ) blocks ature and polarized standard pseudo-a`m’s (see [25] for with an amplitude higher than the cut-sky effects ex- details) and the pixel-based optimized windows for polar- pected in the the pure and hybrid formalisms3. However, ized pure pseudo-a`m’s. (Those peculiar sky apodizations the pixel-induced nonvanishing (TB,TE) and (EB,PP ) are shown on Fig. 3.) In this peculiar case, a first sky blocks in the HEALPix scheme remains much smaller apodization is identically applied to compute standard E- than the cut-sky (and modeled) (TE,TE) and (EB,EB) and B-pseudo-multipoles and a second one is identically blocks. We will therefore neglect such effects and system- applied to compute pure E-and B-pseudo-multipoles. As atically set the (TB,TE) and (EB,PP ) blocks to zero already mentioned in the previous section, with such a in the following. weighting, the (TE,TE) block in the standard and hy- As the pixel-induced leakages are not corrected, the fi- brid formalisms equals the (TB,TB) block in the stan- nal estimate of all the angular power spectra but the TT dard formalism, and the (TE,TE) block in the pure for- one is a priori biased as long as the TB and EB correla- malism equals the (TB,TB) block in the pure and hy- tions do not vanish. Fortunately, such cross-correlations brid formalisms. Therefore, only the (TE,TE) block in the standard approach and in the pure approach is depicted in Fig. 1. However, the (EB,EB) block system- 3 atically differs from one formalism to another and such As is already shown in [25] (Sec. IV.A.5) for the specific case of EE and BB spectra , such unmodeled pixelization effects are a block is displayed in Fig. 2 for the three techniques. significantly reduced by using ECP-like pixelization schemes (for In such figures, the red-dashed curves correspond to the which the number of azimuthal grid points remains unchanged ` = 300 column of the computed mode-mode coupling for each constant zenithal ring and therefore providing a nearly while the overlapping black curves stand for TE(TB) exact quadrature over the polar angle). 9 are expected to be zero in the standard cosmological get closer to realistic sky observations, two possible ex- paradigm. Moreover, we have tested that with nonva- tensions of this ideal mock survey are implemented. We nishing TB and EB input power spectra at a level of first consider the effect of more intricate boundaries us- TB 2 EB `(` + 1)C` /(2π) ∼ 0.2 µK and `(` + 1)C` /(2π) ∼ ing a square patch with holes due to e.g. point-sources 0.004 µK2 at ` = 2−1000 (as motivated by the potential removal, and later an implementation of an inhomoge- impact of an helical, primordial magnetic field [15, 16]), neous sky survey as expected for balloon-borne experi- the EE, BB, TE, TB, and EB angular power spectra es- ment. The sky coverage is always taken to be ∼1% and timated on 1% of the sky with inhomogeneous noise (see the RMS of the noise in the inhomogeneous case is set Fig. 8 for the noise distribution) are indeed not biased equal to 5.75 µK-arcmin. The values of both these pa- though residual EB ↔ EE, EB ↔ BB, and TB ↔ TE, rameters are motivated by the ebex experiment [32]. In leakages are not corrected for [40]. addition, in all cases we assume that noise is uncorrelated from pixel to pixel and from one Stokes parameter to another, i.e.,

V. APPLICATIONS TO SMALL-SKY 2 hN (i)N 0 (j)i = σ (i)δ δ 0 . (29) EXPERIMENTS: POWER SPECTRA S S S i,j S,S UNCERTAINTY As a consequence, the TE, TB and EB noise biases vanish (see Appendix C). In this section we discuss the performance and appli- Hereafter, the variance is estimated as the standard cability of the above defined odd-parity power spectra deviation of 500 MC simulations. The input {CMB + estimators in the context of vanishing true TB and EB noise} maps are built using the HEALPix pixelization cross-correlations, i.e., considered as a null test, poten- scheme at a resolution of Nside = 512 corresponding to tially used for search of astrophysical and/or instrumen- a pixel size of ∼ 7 arcmin. The estimated power spectra tal systematic effects. The uncertainty of each estimator are binned with a band width of ∆` = 40 and with the is quantified by computing the power spectrum variance lowest bin starting at ` = 20. The pseudo-a`m for tem- thanks to Monte-Carlo simulations. Our fiducial model perature are estimated applying the arc cosine apodiza- for the input T and E CMB signal is given by the cosmo- tion [25] with an apodization length of 30 arcminutes. logical parameters as constrained by the WMAP 7-year The same apodization is used while calculating the stan- data. The input B-mode is composed of a primordial dard pseudo-multipoles for polarization from the Q and component with a tensor-to-scalar ratio r = 0.05 and U maps. For the pure-pseudo-multipoles of E- and B- a secondary component induced by lensing4. The two modes computation we use a pixel-based, optimized sky input odd-parity power spectra are set to zero (as is pre- apodization to the Q, U maps. dicted in a parity invariant cosmological scenario). As a reference against which to compare the variances Our starting point is an idealized mock survey con- obtained from the MC simulations we use a theoretical – sidered below covering a circular patch of ∼1% of the and optimistic– Fisher estimate of the variance, which for sky area with a level of homogeneous noise set to 5.75 a cross-spectrum of two sets of (I, Q, U) maps, labelled µK-arcmin for the three Stokes parameters. In order to by (A) and (B), reads5

" 2 X Y 1 4π ∆ (A) (B) = CXY + σ2 B−1 B−1 (30) ` X Y ` X(A)Y(B) `,(A) `,(B) (A) (B) Npix (2` + 1)fsky 4π 4π + CXX + σ2 B−2 CYY + σ2 B−2 . ` X(A)X(A) `,(A) ` Y(B)Y(B) `,(B) Npix Npix

X(A)Y(B) h X(A)Y(B) i Here, ∆` ≡ Var C` denotes the variance of a cross-spectrum of X and Y where X and Y 4 (A) (B) Our convention for r follows the WMAP convention: r = stand for either T,E, or B as derived from a map either P (k )/P (k ) with P , the primordial scalar(tensor) power T 0 S 0 S(T) A or B, respectively, N = 12N 2 is the total number spectrum and k = 0.002 Mpc−1 the pivot scale. pix side 0 of pixels, and B stands for a resolution of the map 5 We point out that in [25], a factor 1/2 is missing in front of `,(A) A. Given our assumptions about the noise, Eq. (29), the last term (pure noise term) of their fsky-formulas. Their equation (32) for cross-spectrum should read

v −1 u 2f 2 4 ! B u sky B 2 B σ 1 σ ∆C` = t C` + C` 2 + 4 . (2` + 1) B` 2 B` 10

σ2 = σ2 × δ × δ , where σ2 X(A)Y(B) X(A)X(A) X,Y A,B X(A)X(A) is related to the noise per pixel σ2 (i) via σ2 = pix X(A)X(A) −1 P 2 X(A)Y(B) Nobs i σpix(i). The eﬀective sky fraction, fsky , is deﬁned as

2 Npix  X X(A)Y(B)  Wi  X Y i=1 f (A) (B) = (31) sky N pix 2 X X(A)Y(B) Npix × Wi i=1

X(A)Y(B) X(A) Y(B) where Wi = Wi × Wi is a cross-product of the sky apodizations applied to the maps A and B from which the X- and Y -modes are estimated. As the sky fraction depends on the chosen sky apodizations, which in general will depend on the adopted formalism, on the map itself, and the considered `-bin, the resulting Fisher estimate of the error bars defined in Eq. (30), will also defined on all these factors. As a unique –formalism independent– benchmark, we will therefore use the Fisher variance computed for the maximal value of fsky for each FIG. 3. Analytic sky apodization (to be applied to intensity considered map corresponding to setting the sky apodiza- map and polarization map for standard pseudo-a`m’s esti- tion equal to the binary mask. These theoretical uncer- mate) and spin-0, spin-1, and spin-2 sky apodizations, op- tainties therefore correspond to the lowest Fisher esti- timized for the second bin, ` ∈ (60, 100) (to be applied to mate one may expect given a peculiar sky fraction, not the polarization maps for pure pseudo-a`m’s estimate), in the taking into account specific apodization used in the cal- case of the idealized mock survey. (Only the real parts of the culations of the corresponding pseudospectra. spin-1 and spin-2 windows are shown.) The entire set of operations needed to compute the estimates, i.e., (1) the mode-mode coupling matrices, (2) the six pseudo-cross-spectra, and (3) the six angular cross-power spectra estimates from the noise-debiased here for any internal boundaries, e.g., due to masked pseudo-C`, using CMB maps and sky apodizations as in- point sources. The analytic sky apodization –to be ap- puts, are numerically implemented in the X2pure code. plied to intensity map and polarization maps for standard The code is an extension of the Xspect and Xpol codes pseudo-a`m’s–, and the three spin-weighted sky apodiza- [28] and is based on the (pure)s2hat library –an effi- tion, optimized for the second bin, ` ∈ (60, 100) –and cient massively parallel implementation of spin-weighted to be applied to the polarization maps for pure pseudo- spherical harmonic transforms [36–39]. This code can a`m’s–, are displayed on Fig. 3. also be used in an MC setting allowing the user to sim- For this ideal mock survey, the fraction of the sky ulate, and subsequently analyze, random realizations of is fsky ' 0.96%. For the TT power spectrum (all for- CMB intensity and polarization maps from the six angu- malisms), the EE and TE power spectra (standard and lar power spectra as an input model. hybrid formalisms), and the BB, TB, and EB (standard formalism), the effective sky coverage is constant and equal to 0.92%. For the BB power spectrum (pure A. Results for an ideal mock survey and hybrid formalism), and the EE and EB power spectra (pure formalism), it varies from 0.64% at large an- The idealized mock survey consists of a circular patch gular scales (20 ≤ ` < 60) up to 0.95% at small angular with a radius of ∼11.3 deg., centered at the equator to scales (980 ≤ ` < 1020). Finally, for the TB power reduce pixel effects (see Sec. IV.A.5 of [25]), and with spectrum (pure and hybrid formalisms), the TE power homogeneous noise at a level of 5.75 µK-arcmin. Such spectrum (pure formalism), and the EB power spectrum a survey is ideal due to both the simplifying assumption (hybrid formalism), the sky coverage ranges from 0.78% of noise homogeneity and its simple shape. Indeed, a at large angular scales (20 ≤ ` < 60) up to 0.93% at circular patch exhibits the shortest perimeter for a fixed small angular scales (980 ≤ ` < 1020). value of the sky coverage. As `-to-`0 and E-to-B mix- The MC averaged estimate of the six power spectra ing appears as the result of boundary effects, such shape (i.e. TT,EE,BB,TE,TB, and EB spectra), esti- permits us to minimize different mode-mode couplings mated using the three types of formalism, are displayed for a given sky coverage. In addition, we do not allow on Fig. 4 alongside the theoretical input C`’s. The per- 11

FIG. 4. MC averaged estimation of the six angular power spectra (colored curves) alongside the theoretical input C`’s (black curves) for the case of an ideal mock survey consisting of a spherical cap with homogeneous noise (see Fig. 3). The perfect agreement between the estimated and the theoretical C`’s proves that each of the three formalisms leads to an unbiased estimation.

fect agreement between the estimated and the theoretical the top-left panel of Fig. 5. input shows that each of the formalism indeed leads to an • EE-spectrum: As already suggested in [23] and illus- unbiased estimate of the angular power spectra. As such trated in [26], the standard estimator leads to slightly estimators are built to be unbiased, this essentially con- smaller error bars than the pure estimator. The under- firms the correct numerical implementation of the three lying reason is that, though B-mode leaks into E-mode formalisms in the X2pure code. in the standard approach (thus increasing the sampling The statistical uncertainties of the estimated angular variance), this effect is small and the information lost by power spectra obtained via MC simulations are displayed removing ambiguous mode in the pure approach leads to in Fig. 5 alongside the Fisher estimates of the variance a higher increase of the sampling variance. In addition, computed as explained in the previous section, Eq. the hybrid estimator, being based on the standard (30). We emphasize that except for TT , all the studied pseudo-multipoles for E-modes, performs roughly the formalisms are theoretically different from each other, as same than the standard estimators. The gain obtained for instance they lead to different mode-mode coupling by using the standard or hybrid estimators is clearly matrices, and thus could result in different levels of the seen on the top-middle panel of Fig. 5 at large angular estimates uncertainties. Nevertheless, as the (TE,TB) scales. block is typically negligible, we can expect that the TE • BB-spectrum: The case of BB power spectrum is estimate will be identical in the standard and hybrid precisely the opposite of the EE one. The main source formalisms, and that the TB estimates in the pure and of uncertainties is the extra-sampling variance due to hybrid formalism will coincide. For the EE, BB, and E-modes leaking into B-modes. As clearly illustrated on EB spectra, the mode-mode coupling matrix changes the top-right panel of Fig. 5, both the pure and hybrid from one formalism to another, even if (EE,EB) and estimators perform the same, leading to uncertainties (BB,EB) blocks are neglected. We consequently expect on par of the most optimistic Fisher estimate, while the differences –though the hybrid and standard estimates standard estimator leads to uncertainties greater than BB should be close in the case of the EE spectrum and C` for ` smaller than ∼ 300. the hybrid and the pure estimates of the BB spectrum • TE-spectrum: For the very same reason that standard should also be rather similar as well. Some other and hybrid estimator leads to smaller uncertainties expectations are as follows: than pure estimator for EE spectrum, one expects • TT-spectrum: For such an angular power spectrum, those two approaches to also perform the best for TE the three formalism are equivalent and we expect to cross-spectrum. Results displayed on Fig. 5 (bottom-left reach the same level of uncertainties for all the formal- panel) show significant improvement by using the ism. This is clearly confirmed by the results exhibited in standard or hybrid estimator as compared to the pure 12

FIG. 5. Error bars on the reconstructed angular power spectra for each of the three formalisms (colored curves) alongside the naive (binned) Fisher estimate of such uncertainties. The assumed sky survey is the ideal case consisting of a circular patch with homogeneous noise (see Fig. 3). The error bars are obtained as the standard deviation over 500 MC simulations.

ones especially at large angular scales. • TB-spectrum: This case is conceptually identical to BB power spectrum. The lowest uncertainties are achieved by removing E-modes leaking into B: pure or hybrid formalism are then preferred to get an estimate TB of C` as accurate as possible with respect to statistical error bars, as exhibited on the bottom-middle panel of Fig. 5. • EB-spectrum: Standard estimators, pure estimators and hybrid estimators perform the best respectively for E-pseudo-multipoles only, for B-pseudo-multipoles only and for both E- and B-pseudo-multipoles respectively. Our results show that the gain in using pure B-pseudo-mulipoles is more signiﬁcant than the gain in using standard E-pseudo-multipoles. One then expects the following hierarchy for the error bars: EB EB EB ∆`, standard > ∆`, pure > ∆`, hybrid. Though uncertainties from the standard estimators are clearly higher (see bottom-right panel of Fig. 5), it appears that the pure and hybrid estimators perform the same.

FIG. 6. Same as Fig. 3 for the case of a mock survey with sharpened boundaries and holes.

From these results, we could conclude that the hybrid estimator systematically leads to the lowest uncertainties for the six angular power spectra; an expected conclusion as the hybrid formalism is based on the best pseudo- multipoles computation for both E- and B-modes. 13

B. Toward a realistic survey: Effect of holes and cant noise inhomogeneity of the reconstructed sky maps. sharp boundaries We therefore explore the performance of the TB and EB pseudo-C`’s estimators in the case of an observation mim- In this section we consider a more intricate observed icking a long-duration CMB balloon-borne experiment patch shape consisting of a square patch with internal such as ebex [32]. The resulting sky sampling, i.e. a holes mimicking effects of point source masking [24]. The number of observations per pixel (Fig. 8), has already specific patch used hereafter is depicted in Fig. 6. been used in [25] to assess the performance of the pure 5 The fraction of the sky covered by this patch is fsky ' B-mode estimators. It ranges from 10 at the edges of 7 0.99%. For the TT power spectrum (all formalisms), the patch to more than 2 × 10 in its core after we cut the EE and TE power spectra (standard and hybrid the noisiest pixels in order to ensure convergency of the formalisms), and the BB, TB, and EB (standard for- optimized sky apodization procedure (see [25]). Finally, malism), the effective sky coverage is constant and equal we set the noise per sample such as the average noise to 0.92%. For the BB power spectrum (pure and hy- level for the full patch equals the noise level used in the brid formalism), and the EE and EB power spectra previously-studied homogeneous cases. (pure formalism), it varies from 0.43% at large angular The retained sky fraction is equal to fsky ' 0.89%. scales (20 ≤ ` < 60) up to 0.96% at small angular scales For the TT power spectrum (all formalisms), the EE (980 ≤ ` < 1020). Finally, for the TB power spectrum and TE power spectra (standard and hybrid formalisms), (pure and hybrid formalisms), the TE power spectrum and the BB, TB, and EB (standard formalism, the ef- (pure formalism), and the EB power spectrum (hybrid fective sky coverage is constant and equal to 0.85%. For formalism), the sky coverage ranges from 0.78% at large the BB power spectrum (pure and hybrid formalism), angular scales (20 ≤ ` < 60) up to 0.93% at small angular and the EE and EB power spectra (pure formalism), it scales (980 ≤ ` < 1020). varies from 0.42% at large angular scales (20 ≤ ` < 60) We found MC averaged power spectra in perfect agree- down to 0.29% at small angular scales (980 ≤ ` < 1020). Finally, for the TB power spectrum (pure and hybrid ment with the theoretical input C` showing that power spectra estimators for the three formalisms are unbiased formalisms), the TE power spectrum (pure formalism), as expected. and the EB power spectrum (hybrid formalism), the The error bars on the estimates of the six angular sky coverage ranges from 0.6% at large angular scales power spectra are shown in Fig. 7. As already shown (20 ≤ ` < 60) down to 0.46% at small angular scales for the case of the circular patch, the standard or hy- (980 ≤ ` < 1020). We note that for this inhomogeneous brid estimators gives smaller variance for the EE and noise case, the effective sky coverage decreases for smaller TE power spectra, while the pure or hybrid formalisms scales as the optimized sky apodization converges toward are to be adopted for the BB, TB and EB spectra. One inverse-noise weighting, which is ”naturally” apodized as should however expect some differences: since sharpened a consequence of the scanning strategy. This difference and internal boundaries enhance the E-to-B and B-to- of the apodization lengths can be easily seen by compar- E leakages (i.e. the total power contained in ambiguous ing the left panel of Fig. 8, showing the inverse noise modes is increased), the relative merit of each formalism weighting, with the right top panel of Fig. 9, depicting is supposed to be quantitatively strengthened. Though one of the optimized, large scale apodizations. not obvious for BB and TB spectra (for which the gain Once again, we first checked that each formalism leads in using the pure or the hybrid estimators is already dra- to unbiased estimates for the six power spectra (TT , EE, matic for simple boundaries), it is now clear at large an- BB, TE, TB, EB). gular scales (` ≤ 100) that the error bars on EE and The uncertainties of the estimated angular power spec- 6 TE angular power spectra (see top-middle and bottom- tra are shown in Fig. 10 alongside the Fisher estimates . left panels of Fig. 7) are lower by using the standard or Overall the hybrid formalism leads to the smallest error the hybrid formalism than by using the pure approach. bars and allowing for the inhomogeneous noise does not Moreover, a closer inspection of the EB power spectra change at a qualitative level, the main conclusion drawn seen in the bottom-right of Fig. 7 suggests that, at large earlier based on the homogeneous noise cases. A similar angular scales, ` ≤ 150, the hybrid approach performs conclusion applies to the relative merits of each formal- slightly better than the pure approach, as expected from ism. On the quantitative level, the noise inhomogeneities theoretical considerations. seem to enhance the differences between the techniques, as, for example, can be seen clearly in the case of the EE and TE power spectra in Fig. 10. The larger uncertain- C. Toward a realistic survey: Effect of ties seen in these cases for the pure estimator are due to inhomogeneous sky sampling the fact that sky apodizations employed in the pure esti-

For a realistic small-scales experiment, the observed patch is expected to be rather irregular with the density 6 We note here in passing that the difference of the BB pure spec- of observations per sky area strongly varying from one trum shown here with that shown in Fig. 24 of [25] is due to a pixel to another. This will unavoidably lead to signifi- different noise level assumed erroneously in this earlier work. 14

FIG. 7. Error bars on the reconstructed angular power spectra for each of the three formalisms (colored curves) alongside the naive (binned) Fisher estimate of such uncertainties. The assumed sky survey is a square patch ﬁlled with holes and with homogeneous noise (see Fig. 6).The error bars are obtained as the standard deviation over 500 MC simulations.

FIG. 8. A distribution of sky observation for an observational strategy mimicking the ebex experiment. The complete distribution is shown in linear scale (left panel) for which the density of samples ranges from 1 to ∼ 2 × 107. The noisiest pixels, which lie in the outer part of the observed sky, have been trimmed in order to compute the optimized sky apodizations. This ”well-observed” part is displayed in linear scale (middle panel) and in logarithmic scale (right panel). The density of observations for the conserved pixels ranges from 105 to ∼ 2 × 107.

mation of E-modes are a priori not optimized to account power spectra the approaches already proposed in [23– for the noise variation. 25, 27, 28]. Each technique differs from the other in the way E- and B-pseudo-multipoles are computed from the maps of Stokes parameters. The three approaches are VI. CONCLUSIONS equivalent from the perspective of the TT power spectrum. The first approach, dubbed standard, estimates We have presented three unbiased alternative ap- polarization multipoles by projecting the masked Q and proaches to estimate the six angular power spectra associ- U maps on the standard spherical harmonics of E- and ated to CMB temperature and polarization anisotropies. B-type. This technique does not lose any information Those techniques are based on the so-called pseudo- contained in the observed maps but its variance suffers from the E-to-B and B-to-E leakages. The second ap- C`’s estimators and generalize to the full set of angular 15

hybrid technique should provide the smallest error bars at least as long as the B-mode power is not larger than that of the E-modes. We have developed and implemented those three approaches by providing the set of expressions to efficiently compute the different mode-mode coupling kernels for each of the formalisms, as well as the noise bias for the specific case of uncorrelated noise from pixel to pixel. The overall consistency of those computations has been demonstrated via numerical experiments. Finally, we have assessed the relative performances of the three proposed pseudo-C`’s estimators for the case of small-scales experiments, assuming three different experimental configurations. In all these cases and for all 6 angular power spectra, the hybrid estimator has been found to result in the smallest statistical uncertainties, confirming the theoretical expectations, while the other estimators can produce similar uncertainties in some specific cases. For instance, i) the standard and hybrid formalisms performs equally well for estimating the EE and TE power spectra, ii) the pure and hybrid formalisms performs equally well to estimate the BB and TB, and FIG. 9. Same as Fig. 3 for the case of a realistic mock survey iii) the hybrid technique performs the best (at least at with inhomogeneous noise as displayed in Fig. 8. large angular scales) for the EB power spectra, though the pure approach is on par with the error bars achieved with the hybrid estimator. proach, denoted pure, makes use of the pure-E, pure- B and ambiguous modes decomposition [35] to project the Q and U maps on a E- and B-basis free from any ACKNOWLEDGMENTS leakages. This second technique therefore corrects for E- to-B and B-to-E leakages on any single realization but This research used resources of the National Energy loses some part of the information by removing ambigu- Research Scientific Computing Center, which is sup- ous modes. Finally, in the third approach, called hybrid, ported by the Office of Science of the U.S. Department E-pseudo-multipoles are numerically computed using the of Energy under Contract No. DE-AC02-05CH11231. standard approach while B-pseudo-multipoles are com- Some of the results in this paper have been derived using puted using the pure technique. For the specific case of s2hat [36–39], HEALPix [41] and camb [42] software CMB polarized anisotropies, E-modes are much greater packages. We thank Sam Leach, Will Grainger, Chris than B-modes. From theoretical arguments and previous Cantalupo and Ted Kisner and the ebex team fo provid- numerical experiments, statistical uncertainties on the ing the tools to simulate the realistic experiment study. E-pseudo-multipoles are therefore expected to be mainly JG acknowledges financial support from the Groupement increased by the loss of information while statistical un- d’IntérêtScientifique (GIS) ’consortium Physique des 2 certainties on the B-pseudo-multipoles are expected to Infinis (P2I)’. We also thank the ANR-MIDAS’09 project be mainly increased by E-to-B leakage. As a result, the team for helpful discussions.

Appendix A: Gaunt integrals and Wigner-3j

We provide in this appendix some general deﬁnitions and properties of the Gaunt integrals and Wigner-3j symbols which are useful for the computation of the mode-mode coupling matrices of the pseudo-C` estimators. The general deﬁnition of the Gaunt integrals is

Z `ms G`0m0s0;`00m00s00 ≡ sY`m s0 Y`0m0 s00 Y`00m00 d~n, (A1) 4π 16

FIG. 10. Error bars on the reconstructed angular power spectra for each of the three formalisms (colored curves) alongside the naive (binned) Fisher estimate of such uncertainties. The assumed sky survey is shown in Fig. 9 with inhomogeneous noise as displayed on Fig. 8).The error bars are obtained as the standard deviation over 500 MC simulations.

and it can be expressed by use of the Wigner-3j symbols

r 0 00 0 00 0 00 `ms (2` + 1)(2` + 1)(2` + 1) ` ` ` ` ` ` G 0 0 0 00 00 00 = . (A2) ` m s ;` m s 4π s s0 s00 m m0 m00 Ths is easily generalized to integrals involving complex conjugates of the spin-weighted spherical harmonics by using ? s+m sY`m = (−1) −sY`(−m). The Wigner-3j symbols verify the following symmetry and orthogonality relations 0 00 0 00 ` ` ` 0 00 ` ` ` = (−1)`+` +` (A3) −m −m0 −m00 m m0 m00 and 0 00 0 000 00 X ` ` ` ` ` ` (2` + 1) = δ 00 000 δ 00 000 . (A4) m m0 m00 m m0 m000 ` ,` m ,m mm0 Two linear combinations of Wigner-3j symbols and Gaunt integrals are involved in the expression of the pseudo- multipoles mode-mode coupling matrices. To lighten our expressions, we deﬁned the following quantities: ` `0 `00 ` `0 `00 J (±) (`, `0, `00) = ± , (A5) 2−s −s 2 −2 + s s −2 2 − s and Z (±) 0 00 ? ? G2−s(L, L ,L ) = (sY`m2Y`0m0 −2+sY`00m00 ) ± (−sY`m−2Y`0m0 2−sY`00m00 ) d~n, (A6) 4π r (2` + 1)(2`0 + 1)(2`00 + 1) ` `0 `00 = (−1)s+m J (±) (`, `0, `00), 4π m m0 m00 2−s with L labeling the couple (`, m) in the Gaunt integrals. From the parity relation of the Wigner-3j, it is easily shown (+) 0 00 (−) 0 00 that J2−s = 0 for odd values of (` + ` + ` ) and J2−s = 0 for even values of (` + ` + ` ). Finally, we also deﬁne the following Wigner-3j symbols ` `0 `00 J (T )(`, `0, `00) = (A7) 0 0 0 which is equal to zero for odd values of (` + `0 + `00). 17

Appendix B: Mode-mode coupling matrices

1. Mode-mode coupling for (pure) pseudo-a`m

In this appendix, we present the explicit expressions of the different mode-mode coupling matrices related the pseudo-multipoles to the CMB multipoles. We remind that standard pseudo-a`m’s are defined by Z (std) 1 h 2 ? ¯2 ?i E`m ≡ − W (Q + iU) ∂ Y`m + (Q − iU) ∂ Y`m d~n, (B1) 2N`,2 4π Z (std) i h 2 ? ¯2 ?i B`m ≡ W (Q + iU) ∂ Y`m − (Q − iU) ∂ Y`m d~n, (B2) 2N`,2 4π and pure pseudo-a`m’s are defined by Z (pure) 1 h 2 ? ¯2 ?i E`m ≡ − (Q + iU) ∂ WY`m + (Q − iU) ∂ WY`m d~n, (B3) 2N`,2 4π Z (pure) i h 2 ? ¯2 ?i B`m ≡ (Q + iU) ∂ WY`m − (Q − iU) ∂ WY`m d~n. (B4) 2N`,2 4π We will first provide the expressions for pure pseudo-multipoles as the standard approach can be easily deduced from the pure one.

a. Pure pseudo-multipoles

By expanding the spin-raising(spin-lowering) operators acting on the product W × Y`m and using the deﬁnition of the spin-weighted spherical harmonics, one obtains

Z 2 (pure) 1 X h ? ? i E ≡ − α N (Q + iU) ∂2−sW Y ? + (−1)s (Q − iU) ∂¯2−sW Y ? d~n, (B5) `m 2N s `,s s `m −s `m `,2 4π s=0 Z 2 (pure) i X h ? ? i B ≡ α N (Q + iU) ∂2−sW Y ? − (−1)s (Q − iU) ∂¯2−sW Y d~n, (B6) `m 2N s `,s s `m −s `m `,2 4π s=0 with αs = 1 for s = 0, 2 and αs = 2 for s = 1. The s = 2 term in the above summation over the spin index simply corresponds to the standard multipoles. Therefore, one can ﬁrst compute the mode-mode coupling matrices for the pure case and then keep only the s = 2 term in the summation to recover the mode-mode coupling matrices for the standard estimation. This computation is done by plugging the multipolar decomposition of the (Q ± iU) ﬁelds and of the window function W , i.e.

X (E) W = − w0,`mY`m, `m X Q ± iU = − (aE,`m ± iaB,`m) ±2Y`m. `m

If the derivative relationship between the different spin-weighted window function is completely satisfied, i.e., Ws = ∂sW , the multipolar decomposition of the spin-0 window function W is sufficient as

(E) (B) ws,`m = N`,sw`m and ws,`m = 0.

However, as shown in [25], pixelization on the sphere partially breaks this derivative relationship and the multipolar decomposition of each of the spin-weighted window function has to be considered as independent from the spin-0 ones, i.e.

X (E) (B) Ws = − ws,`m + iws,`m sY`m, `m s+1 X (E) (B) W−s = (−1) ws,`m − iws,`m sY`m. `m 18

By plugging the above multipolar decomposition into Eqs. (B5) and (B6) and using the Gaunt integrals, it is straightforward, though rather long, to show that

(pure) X h diag off i E`m = H`m,`0m0 aE,`0m0 + iH`m,`0m0 aB,`0m0 , `0m0 (pure) X h diag off i B`m = H`m,`0m0 aB,`0m0 − iH`m,`0m0 aE,`0m0 , `0m0 with

2 diag −1 X X s h (E) (+) 0 00 (B) (−) 0 00 i H`m,`0m0 = (−1) αsN`,s w2−s,`00m00 G2−s(L, L ,L ) − iw2−s,`00m00 G2−s(L, L ,L ) , 2N`,2 `00m00 s=0 2 off −1 X X s h (E) (−) 0 00 (B) (+) 0 00 i H`m,`0m0 = (−1) αsN`,s w2−s,`00m00 G2−s(L, L ,L ) − iw2−s,`00m00 G2−s(L, L ,L ) . 2N`,2 `00m00 s=0

We underline that the spin-dependant pseudo-multipoles, Es,`m and Bs,`m, as deﬁned in Eqs. (18) and (19) are also linked to the CMB multipoles via a mode-mode coupling

X h s,diag s,off i Es,`m = K`m,`0m0 aE,`0m0 + iK`m,`0m0 aB,`0m0 , `0m0 X h s,diag s,off i Bs,`m = K`m,`0m0 aB,`0m0 − iK`m,`0m0 aE,`0m0 , `0m0 with

1 X h (E) (+) (B) (−) i Ks,diag = − (−1)sα w G (L, L0,L00) − iw G (L, L0,L00) , `m,`0m0 2 s 2−s,`00m00 2−s 2−s,`00m00 2−s `00m00 1 X h (E) (−) (B) (+) i Ks,off = − (−1)sα w G (L, L0,L00) − iw G (L, L0,L00) . `m,`0m0 2 s 2−s,`00m00 2−s 2−s,`00m00 2−s `00m00

The full mode-mode coupling matrices are then related to the spin-weighted mode-mode coupling matrices by

2 2 1 X 1 X Hdiag = N Ks,diag and Hoff = N Ks,off . `m,`0m0 N `,s `m,`0m0 `m,`0m0 N `,s `m,`0m0 `,2 s=0 `,2 s=0

b. Standard pseudo-multipoles

From the above result, we can easily deduce the mode-mode coupling matrices for standard multipoles by simply keeping the s = 2 terms in the summation over the spin index to get

(std) X h diag off i E`m = K`m,`0m0 aE,`0m0 + iK`m,`0m0 aB,`0m0 , `0m0 (std) X h diag off i B`m = K`m,`0m0 aB,`0m0 − iK`m,`0m0 aE,`0m0 , `0m0 with

1 X (E) (+) Kdiag = K2,diag = − w G (L, L0,L00), `m,`0m0 `m,`0m0 2 0,`00m00 2−s `00m00 (B7) 1 X (E) (−) Koff = K2,off = − w G (L, L0,L00). `m,`0m0 `m,`0m0 2 0,`00m00 2 `00m00

We stress out that for standard estimation of the pseudo-multipoles, only the real-valued, spin-0 window function is used which does not contain any B-component. 19

2. Mode-mode coupling for EB pseudo-C`

We remind that the pseudo-multipoles are related to the CMB multipoles via a mode-mode coupling matrix

(x) X h diag off i E`m = K`m;`0m0 aE,`0m0 + iK`m;`0m0 aB,`0m0 , `0m0 (y) X h diag off i B`m = H`m;`0m0 aB,`0m0 − iH`m;`0m0 aE,`0m0 , `0m0 where (x) and (y) corresponds to either ”standard” or ”pure” estimation. We stress out that the above expressions do not formally depend on the formalism. All the formalism dependancy is encoded in the matrices K and H, which can be set equal to K or H, depending on the values adopted for (x) and (y). In the following, we assume that the polarization ﬁeld is a gaussian and isotropic random ﬁelds

? EE aE,`maE,`0m0 = C` δ`,`0 δm,m0 , ? BB aB,`maB,`0m0 = C` δ`,`0 δm,m0 , ? ? EB aE,`maB,`0m0 = aB,`maE,`0m0 = C` δ`,`0 δm,m0 .

The pseudo-power spectra are deﬁned as follows:

1 X (x) (x)? CEE = E E , `,(xx) 2` + 1 `m `m m 1 X (y) (y)? CBB = B B , `,(yy) 2` + 1 `m `m m 1 X h (x) (y)? (y) (x)?i CEB = E B + B E . `,(xy) 2(2` + 1) `m `m `m `m m

Starting from the deﬁnition of the pseudo-multipoles, it is easily shown that

2 2 D EE E 1 X X diag EE off BB h diag off? i EB C = K C 0 + K C 0 + 2Im K K C 0 , `,(xx) 2` + 1 `m;`0m0 ` `m;`0m0 ` `m;`0m0 `m;`0m0 ` `0 mm0 2 2 D BB E 1 X X off EE diag BB h diag off? i EB C = H C 0 + H C 0 − 2Im H H C 0 , `,(yy) 2` + 1 `m;`0m0 ` `m;`0m0 ` `m;`0m0 `m;`0m0 ` `0 mm0

D EB E 1 X X n h diag diag? off off? i EB C = Re K H − K H C 0 (B8) `,(xy) 2` + 1 `m;`0m0 `m;`0m0 `m;`0m0 `m;`0m0 ` `0 mm0 h diag off? i BB h diag off? i EEo +Im H`m;`0m0 K`m;`0m0 C`0 − Im K`m;`0m0 H`m;`0m0 C`0 .

The form of the above expressions is also formalism-independant as all the speciﬁcities of a given formalism is encoded in the peculiar expressions of the K and H matrices. Without any a priori on the explicit expressions of those matrices, each of the three pseudo-C`’s receives contribution from the three angular power spectra. However, due to the parity properties of Wigner-3j symbols, it will be shown in the following of this appendix that the (EB,PP ) and (PP,EB) blocks are zero as long as the window functions do not contain any B-type component. The mode-mode coupling matrices for the pseudo-power spectra are ﬁnally computed by taking the expression of the K and H matrices valid for the adopted formalism and then performing the summation over the two azimuthal quantum numbers m and m0. This explicit computation is very similar to the temperature case [27, 33, 34]. The standard and pure formalism being a restricted case of the pure formalism7, the computation of the mixing kernels (±) 0 00 is only detailed for this last case. The forthcoming formulas involve the J2−s(`, ` , ` ) quantities and we will remove their `, `0, and `00 dependance to lighten the writings.

7 We remind that the pure pseudo-multipoles are a linear combi- and the standard pseudo-multipoles are given by the spin-2 term nation of three spin-dependant pseudo-multipoles (s = 0, 1, 2) of the pure ones. 20

a. Standard formalism

In the standard formalism, the mode-mode coupling matrices for the pseudospectra reads [43]:

0 2 2` + 1 X (E) (+) M EE,EE = M BB,BB = w J , ``0 ``0 16π 0,`00m00 0 `00m00 0 2 2` + 1 X (E) (−) M EE,BB = M BB,EE = w J , ``0 ``0 16π 0,`00m00 0 `00m00 0 2 2 EB,EB 2` + 1 X (E) (+) (E) (−) M 0 = w 00 00 J − w 00 00 J , (B9) `` 16π 0,` m 0 0,` m 0 `00m00 0 2 2` + 1 X (E) (+) (−) M EE,EB = −M BB,EB = w J J , ``0 ``0 8π 0,`00m00 0 0 `00m00 0 2 2` + 1 X (E) (+) (−) M EB,EE = −M EB,BB = − w J J ``0 ``0 16π 0,`00m00 0 0 `00m00 However, because J (+) vanish for odd values of (` + `0 + `00) while J (−) vanish for even values of (` + `0 + `00), the oﬀ-diagonal blocks relating autospectra to the EB cross-spectrum become null

EE,EB BB,EB M``0 = −M``0 = 0, EB,EE EB,BB (B10) M``0 = −M``0 = 0.

b. Pure formalism

For the (EE,EE) block computation, the expression of Gaunt integrals as functions of Wigner-3j symbols is ﬁrst diag plugged in the expression of H`m,`0m0 to get

0 ! 2 0 00 0 000 EE,EE 2` + 1 X X X p 00 000 ` ` ` ` ` ` M = α α 0 N N 0 (2` + 1)(2` + 1) ``0 16πN 2 s s `,s `,s −m m0 m00 −m m0 m000 `,2 `00m00 `000m000 s,s0=0 h (E) (+) 0 00 (B) (−) 0 00 i × w2−s,`00m00 J2−s(`, ` , ` ) − iw2−s,`00m00 J2−s(`, ` , ` ) ? h (E) (+) 0 000 (B) (−) 0 000 i × w2−s0,`000m000 J2−s0 (`, ` , ` ) − iw2−s0,`000m000 J2−s0 (`, ` , ` ) . We can now ﬁrst use the orthonormalization properties of the Wigner-3j listed in App. A and secondly, the fact that (+) (−) all the terms proportional to J2−sJ2−s0 are equal to zero to get  2 2 2`0 + 1 2 N 1 N EE,EE X X `,s (E) (+) X `,s (B) (−) M``0 =  αs w2−s,`00m00 J2−s + αs w2−s,`00m00 J2−s  . (B11) 16π N`,2 N`,2 `0m00 s=0 s=0 We emphasize that the spin-summation runs from 0 to 2 for the E-component of the window functions but from 0 to 1 for its B-component as only the spin-1 and spin-2 windows may have a nonvanishing B-part. Applying the same reasoning for the (EE,BB) and (EE,EB) blocks leads to  2 2 2`0 + 1 2 N 1 N EE,BB X X `,s (E) (−) X `,s (B) (+) M``0 =  αs w2−s,`00m00 J2−s + αs w2−s,`00m00 J2−s  , 16π N`,2 N`,2 `0m00 s=0 s=0

0 2 EE,EB 2` + 1 X X N`,sN`,s0 n (+) (+) h (E) (B)? i M = α α 0 J J 0 Re w w (B12) ``0 8π s s N 2 2−s 2−s 2−s,`00m00 2−s0,`00m00 `0m00 s,s0=0 `,2 (−) (−) h (B) (E)? io −J2−sJ2−s0 w2−s,`00m00 w2−s0,`00m00 . The other blocks are easily deduced from the above three ones as

BB,BB EE,EE EE,BB BB,EE EB,EB EE,EE EB,EB M``0 = M``0 ,M``0 = M``0 ,M``0 = M``0 − M``0 1 1 M BB,EB = −M EE,EB,M EB,EE = − M EE,EB,M EB,BB = M EE,EB. ``0 ``0 ``0 2 ``0 ``0 2 ``0 21

EE,EB The expression of the M``0 block shows that in the pure formalism, the EB-EE coupling and the EB-BB coupling are not zero (unlike the standard formalism). However, such couplings will vanish if the spin-1 and spin-2 window functions have a vanishing component of type B.

c. Hybrid formalism

EB,P P 0 EE,P P 0 In the hybrid formalism, only the M``0 blocks needs to be computed as the M``0 blocks are equal to the EE,P P 0 BB,P P 0 BB,P P 0 M``0 blocks computed in the standard formalism and the M``0 blocks equal to the M``0 blocks derived EB,P P 0 in the pure formalism. For the remaining M``0 blocks, we found 0 1 EB,EE 2` + 1 X 1 X (+) (+) h (E) (B)? i M``0 = − αsN`,sJ0 J2−sRe w0,`00m00 w2−s,`00m00 , 16π N`,2 `0m00 s=0 0 1 EB,BB 2` + 1 X 1 X (−) (−) h (E)? (B) i M``0 = − αsN`,sJ0 J2−sRe w0,`00m00 w2−s,`00m00 , (B13) 16π N`,2 `0m00 s=0 0 2 EB,EB 2` + 1 X 1 X h (+) (+) (−) (−) i h (E) (E)? i M``0 = αsN`,s J0 J2−s − J0 J2−s Re w0,`00m00 w2−s,`00m00 , 16π N`,2 `0m00 s=0 As is the case for the pure formalism, the EB-EE and EB-BB couplings are not zero if the spin-1 and spin-2 window functions have a nonvanishing B-component.

3. Mode-mode coupling for TB pseudo-C`

The temperature pseudo-multipoles are related to the CMB multipoles via the following coupling [27, 33, 34]:

r 0 00 0 00 X X m (2` + 1)(2` + 1)(2` + 1) (T ) 0 00 ` ` ` (E) T = − (−1) J (`, ` , ` ) w a 0 0 . (B14) `m 4π m m0 m00 0,`00m00 T,` m `0m0 `00m00

T E/B Combining the above pseudo-a`m with pseudo-a`m allows us to compute the mode-mode coupling matrices for the pseudo-power spectra in the diﬀerent formalism. The procedure is similar, though simpler, to the polarization case described in the previous appendix and we only list here our ﬁnal results. First, in such the standard formalism, the mode-mode coupling matrices read 0 2 2` + 1 X (E) (+) M T E,T E = M T B,T B = w J (T )(`, `0, `00)J (`, `0, `00), ``0 ``0 8π 0,`00m00 0 `00m00 (B15) T E,T B T B,T E M``0 = −M``0 = 0. Second, in the pure formalism, the mode-mode coupling matrices are given by 0 T E,T E T B,T B 2` + 1 X X N`,s (T ) (+) h (E) (E)? i M``0 = M``0 = αs J J2−sRe w0,`00m00 w2−s,`00m00 , 8π N`,2 `00m00 s 0 (B16) T E,T B T B,T E 2` + 1 X X N`,s (T ) (+) h (E) (B)? i M``0 = −M``0 = αs J J2−sRe w0,`00m00 w2−s,`00m00 . 8π N`,2 `00m00 s Finally, in the hybrid formalism, the mode-mode coupling matrices read 0 2 2` + 1 X (E) (+) M T E,T E = w J (T )(`, `0, `00)J (`, `0, `00), ``0 8π 0,`00m00 0 `00m00 (B17) T E,T B M``0 = 0, and 0 T B,T B 2` + 1 X X N`,s (T ) (+) h (E) (E)? i M``0 = αs J J2−sRe w0,`00m00 w2−s,`00m00 , 8π N`,2 `00m00 s 0 (B18) T B,T E 2` + 1 X X N`,s (T ) (+) h (E) (B)? i M``0 = − αs J J2−sRe w0,`00m00 w2−s,`00m00 . 8π N`,2 `00m00 s 22

Appendix C: Noise bias

The noise bias in the pseudo-C` estimators diﬀers from one formalism to another. To illustrate this point, we provide here the explicit formulas for such biases assuming that noise is uncorrelated from pixel to pixel but allowing correlation between the diﬀerent measured Stokes parameters, i.e.

0 2 0 hNS(~n)NS0 (~n )i = σSS0 (~n)δ(~n − ~n ) where S, S0 stands for the Stokes parameters, T,Q, or U, and assuming autocorrelation of a given map. To perform such a computation, we recall that [44]

` 4π X ? s−s0 ` −2isγ Y (~n ) 0 Y (~n ) = (−1) D 0 (α, β, γ)e , (C1) 2` + 1 s `m i s `m j s,s m=−`

−1 with α, β, and γ the three Euler angles deﬁned by the rotation multiplication R(α, β, −γ) = R(θi, ϕi, 0)R (θj, ϕj, 0). ` For ~ni = ~nj, those three Euler angles are equal to zero. The Ds,s0 stands for the Wigner rotation matrices and they ` satisfy Ds,s0 (0, 0, 0) = δs,s0 (see [45]). With such an hypothesis about the noise properties and making use of the above addition theorem for spin-weighted spherical harmonics, it is straightforward to derive the noise bias in the three formalisms, as presented below. In the standard formalism ﬁrst, all the noise biases vanish except for the EE and BB autospectra. For such spectra, the noise contribution reads: Z EE BB 1 2 2 2 N` = N` = σQQ + σUU W d~n, (C2) 8π 4π and

EB TE TB N` = N` = N` = 0. (C3) Second, in the pure formalism, the noise biases for autospectra are given by

Z N 2 ! EE BB 1 2 2 2 `,1 2 1 2 N` = N` = σQQ + σUU W + 4 2 |W1| + 2 |W2| d~n, (C4) 8π 4π N`,2 N`,2 while for cross-spectra they reads Z σ2 EB 1 QU 2 N` = 2 Re W2 d~n, 4π 4π N`,2 1 Z W N TE = − σ2 Re[W ] + σ2 Im[W ] d~n, (C5) ` 4π N TQ 2 TU 2 Z4π `,2 TB 1 W 2 2 N` = − σTQIm[W2] − σTU Re[W2] d~n. 4π 4π N`,2 Finally, in the hybrid formalism, the noise biases are easily deduced from the standard and pure formalism computation to be given by Z EE 1 2 2 2 N` = σQQ + σUU W d~n, 8π 4π Z N 2 ! (C6) BB 1 2 2 2 `,1 2 1 2 N` = σQQ + σUU W + 4 2 |W1| + 2 |W2| d~n, 8π 4π N`,2 N`,2 and EB N` = 0, N TE = 0, ` Z (C7) TB 1 W 2 2 N` = − σTQIm[W2] − σTU Re[W2] d~n. 4π 4π N`,2 Unlike the standard calculation, the pure computation induces some noise bias for the two types of odd-parity power spectra. However, the EB noise bias can be set to zero by switching to a hybrid approach. Moreover, the TB 23 and EB noise bias will also vanish if the noise is uncorrelated from one Stokes parameter to another or by invoking cross-spectra between maps coming from diﬀerent experiments or diﬀerent, and thus uncorrelated, detectors. We also stress out that all the noise biases computed above are indeed invariant under any rotation of the polarization reference basis. This can be straightforwardly checked by making use of the spin properties of the Stokes parameters and of the three spin-weighted window functions under a rotation by an angle α: T → T 0 = T, (Q ± iU) → (Q0 ± iU 0) = e±2iα(Q ± iU), 0 ±isα W±s → W±s = e W±s.

[1] BICEP: http://www.astro.caltech.edu/∼lgg/ galupi & R. Stompor, MNRAS 408 2319 (2010) bicep front.htm [27] E. Hivon, K. M. Górski,C. B. Netterfield, B. P. Crill, S. [2] Planck: http://www.esa.int/esaSC/ Prunet, & F. Hansen, Astrophys. J.567 2 (2002) 120398 index 0 m.html [28] M. Tristram, J.F. Macias-Pérez,C. Renault & D. Santos, [3] POLARBEAR: MNRAS 358 833 (2005) http://bolo.berkeley.edu/polarbear/?q=science [29] W. Zhao & D. Baskaran, Phys. Rev. D82 023001 (2010) [4] QUIET: http://quiet.uchicago.edu/ [30] J. Kim & P. Naselsky, Astron. & Astrophys. 519 A104 [5] EBEX: (2010); J. Kim, Astron. & Astrophys. 531 A32 (2011) http://groups.physics.umn.edu/cosmology/ebex/ [31] J. Bowyer, A. Jaffe & D. Novikov, arXiv:1101.0520v3 [6] Spider: [astro-ph.CO] http://cmb.phys.cwru.edu/ruhl lab/spider.html [32] B. Reichborn-Kjennerud, et al., Proc. SPIE Int. Soc. Opt. [7] QUBIC: http://www.qubic-experiment.org/ Eng. 7741 77411C (2010) [8] COrE: http://www.core-mission.org/ [33] M.G. Hauser, & P.J.E. Peebles, Astrophys. J.185 757 [9] M. Zaldarriaga & U. Seljak, Phys. Rev. D55 1830 (1997) (1973) [10] M. Kamionkowski, A. Kosowsky & A. Stebbins, Phys. [34] G. Hinshaw, et al., Astrophys. J. Supp. 148 135 (2003) Rev. Lett. 78 2058 (1997) [35] E.F. Bunn, M. Zaldarriaga, M. Tegmark, & A. de [11] M. Zaldarriaga & U. Seljak, Phys. Rev. D58 023003 Oliveira-Costa, Phys. Rev. D67 023501 (2003) (1998) [36]S 2HAT: [12] W. Hu, M. N. Hedman & M. Zaldarriaga, Phys. Rev. http://www.apc.univ-paris7.fr/APC CS/Recherche/ D67 043004 (2003) Adamis/MIDAS09/software/s2hat/s2hat.html [13] A.P.S. Yadav, M. Su, & M. Zaldarriaga, Phys. Rev. D81 [37] pureS22HAT: 063512 (2010) http://www.apc.univ-paris7.fr/APC CS/Recherche/ [14] L. Pogosian, T. Vachaspati & S. Winitzki, Phys. Rev. Adamis/MIDAS09/software/pures2hat/pureS2HAT.html D65 083502 (2002) [38] I. O. Hupca, J. Falcou, L. Grigori & R. Stompor, Lecture [15] C. Caprini, R. Durrer & T. Kahniasvili, Phys. Rev. D69 Notes in Computer Science, 7155, 355, (2012) 063006 (2004) [39] M. Szydlarski, P. Esterie, J. Falcou, L. Grigori, & R. [16] T. Kahniashvili & B. Ratra, Phys. Rev. D71 103006 Stompor, arXiv:11060159 [c.:DC] (2005) [40] A. Ferté, ”Détection d’un champ magnétique primor- [17] A. Kosowsky & A. Loeb, Astrophys. J.469 1 (1996) dial avec les anisotropies polariséesdu fond diffus cos- [18] A. Kosowsky, T. Kahniashvili, G. Lavrelashvili & B. Ra- mologique”, Master Thesis (in French) of UniversitéParis tra, Phys. Rev. D71 043006 (2005) Sud [19] L. Campanelli, A. D. Dolgov, M. Giannotti & F. L. Vi- [41] K. M. Górski,E. Hivon, A. J. Banday, B. D. Wandelt, F. lante, Astrophys. J.616 1 (2004) K. Hansen, M. Reinecke & M. Bartelmann, Astrophys. [20] C. Scóccola,D. Harari & S. Mollerach, Phys. Rev. D70 J.622 759 (2005) 063003 (2004) [42] A. Lewis, A. Challinor & A. Lasenby, Astrophys. J.538 [21] S. M. Carroll, Phys. Rev. Lett.81 3067 (1998) 473 (2000) [22] A. Lue, L. Wang & M. Kamionkowski, Phys. Rev. [43] A. Kogut et al., Astrophys. J. Supp. 148 161 (2003) Lett.83 1506 (1999) [44] K.-W. Ng & G.-C. Liu, Correlation Functions of CMB [23] K.M. Smith, Phys. Rev. D74 083002 (2006) Anisotropy and Polarization, Int. J. Mod. Phys. D 8 61 [24] K.M. Smith & M. Zaldarriaga, Phys. Rev. D76 043001 (1999) (2007) [45] D.A. Varshalovich, A.N. Moskalev & V.K. Khersonskii, [25] J. Grain, M. Tristram & R. Stompor, Phys. Rev. D79 Quantum Theory of Angular Momentum, (World Scien- 123515 (2009) tific, 1988) [26] F. Stivoli, J. Grain, S. M. Leach, M. Tristram, C. Bacci-