THE ASTROPHYSICAL JOURNAL, 567 : 2È17, 2002 March 1 V

( 2002. The American Astronomical Society. All rights reserved. Printed in U.S.A.

MASTER OF THE COSMIC MICROWAVE BACKGROUND ANISOTROPY POWER SPECTRUM :

A FAST METHOD FOR STATISTICAL ANALYSIS OF LARGE AND COMPLEX
COSMIC MICROWAVE BACKGROUND DATA SETS
ERIC HIVON,1,2 KRZYSZTOF M. GO RSKI,3,4 C. BARTH NETTERFIELD,5 BRENDAN P. CRILL,1
SIMON PRUNET,6 AND FRODE HANSEN7
Received 2001 May 17 ; accepted 2001 October 17

ABSTRACT
We describe a fast and accurate method for estimation of the cosmic microwave background (CMB)
anisotropy angular power spectrumÈMonte Carlo Apodized Spherical Transform Estimator
(MASTER). Originally devised for use in the interpretation of the BOOMERANG experimental data,
MASTER is both a computationally efficient method suitable for use with the currently available CMB
data sets (already large in size, despite covering small fractions of the sky, and a†ected by inhomoge-
neous and correlated noise) and a very promising application for the analysis of very large future CMB
satellite mission products.
Subject headings : cosmic microwave background È methods : statistical
On-line material : color Ðgures

1. INTRODUCTION involve computations of complexity DN3 and become

During the past decade since the ground-breaking dis-
covery of the cosmic microwave background (CMB) radi-
ation anisotropy by the COBE satellite (Smoot et al. 1992),
numerous successful measurements of microwave sky struc-
tures have provided us with the data for powerful tests
of the current cosmological paradigm and created an
unprecedented opportunity to estimate key parameters of
the candidate theoretical models of the universe.
Recent ground-based and balloon-borne experiments
with improved sky coverage, angular resolution, and noise
performance (see de Bernardis et al. 2000, Hanany et al.
2000, Padin et al. 2001, Ja†e et al. 2001, Lee et al. 2001,
Halverson et al. 2001, Pryke et al. 2001, and references
therein for some of the most recent experiments and their
interpretation) have both given us a taste of what future
satellite missions Microwave Anisotropy Probe (MAP)8 and
Planck9 should accomplish and revealed the growing chal-
lenges that we will have to meet in the analysis of the forth-
coming CMB data sets.
In the currently favored structure formation model of
inÑation-induced, Gaussian-distributed curvature pertur-
bations, all the statistical information contained in a CMB
map can be summarized in its angular power spectrum C .
General maximum likelihood methods for extracting Cl
from a N -pixel map with nonuniform coverage and corre-l
pix (Gorski 1994, 1997 ; Bond 1995 ; Tegmark &
lated noise
Bunn 1995 ; Bond, Ja†e, & Knox 1998 ; Borrill 1999b)
1 Observational Cosmology, MS 59-33, Caltech, Pasadena, CA 91125.
2 IPAC, MS 100-22, Caltech, Pasadena, CA 91125.
3 European Southern Observatory, Garching bei MuŽnchen, Germany.
4 Warsaw University Observatory, Warsaw, Poland.
5 Department of Physics and Astronomy, University of Toronto, 60 St.
George Street, Toronto, ON M5S 3H8, Canada.
6 CITA, University of Toronto, 60 St. George Street, Toronto, ON M5S
3H8, Canada.
7 MPA, Garching, Germany.
8 See http ://map.gsfc.nasa.gov/.
9 See http ://astro.estec.esa.nl/Planck/.
2
pix104 maps pro-
prohibitively CPU expensive for the N [
pix
duced by current experiments. With the presently antici-
pated computer performance, such methods appear totally
impractical for application to the N [ 106 maps expected
pix 1999a). Hence, there
from the future space missions (Borrill
is a well-recognized need for faster, more economical, and
accurate C extraction methods, which should enable a
l
correct cosmological interpretation of the CMB anisotropy
observations.
In this paper we introduce and discuss a new method for
fast estimation of the CMB anisotropy angular power spec-
trum from Ñuctuations observed on a limited area of the
sky. This method is based on a direct spherical harmonic
transform (SHT) of the available map and allows one to
incorporate a description of the particular properties of a
given CMB experiment, including the survey geometry,
scanning strategy, instrumental noise behavior, and pos-
sible non-Gaussian and/or nonstationary events that can
occur during the data acquisition. The estimated power
spectrum is a†ected by the unwanted contribution of the
instrumental noise and the e†ects of any necessary alter-
ation of either the recorded data stream (such as high-pass
Ðltering) or the raw map of the observed region of the sky,
which are introduced during the data analysis. These e†ects
are calibrated in Monte Carlo (MC) simulations of the
modeled observation and analysis stage of the experiment
and can then be removed or corrected for in the estimated
power spectrum. The harmonic mode-mode coupling
induced by the incomplete sky coverage is described ana-
lytically by the SHT of the sky window and corrected for in
order to obtain an unbiased estimate of the C . We refer to
l Spherical
this method as MASTER (Monte Carlo Apodized
Transform Estimator).
NetterÐeld et al. (2002) described an application of this
method in the extraction of the CMB angular power spec-
trum C , for 75 \ l \ 1025, from the sky map (analyzed
l
region comprising D1.8% of the sky covered with 57,000
pixels of 7@ size) made by coadding four frequency channel
data of the 1998/1999 Antarctic long duration balloon Ñight
 MASTER ANALYSIS OF CMB DATA SETS
of the BOOMERANG experiment (BOOMERANG-LDB).
The Ðrst derivation of the CMB anisotropy spectrum from
the same data (de Bernardis et al. 2000) involved the
MADCAP method (Borrill 1999b) applied to a smaller
subset of the data (one frequency channel, D1% of the sky
covered with D8000 pixels of 14@ size). The MADCAP
approach is too CPU intensive for repeated applications to
the new, enlarged subset of the BOOMERANG data, and,
hence, the MASTER approach was the method of choice
for extraction of the high-l angular power spectrum of the
CMB anisotropy.
Other fast methods have recently been proposed for esti-
mation of the angular spectrum of the CMB anisotropy.
Szapudi et al. (2001) advocate the use of the 2 point corre-
lation function for extraction of the angular power spec-
trum from the CMB maps. The computational demands of
this method scale quadratically, DN2 , with the size of data
pix
set (that may be improved to DN log N ). In the same
pix pix
way as in the case of MASTER, the e†ects of the noise and
correlations of the derived C are quantiÐed by MC simula-
l p p
tions (although the demonstrated applications involved
only the case of a uniform white noise). Dore, Knox, & Peel
(2001a) proposed a hierarchical implementation of the usual
quadratic C estimator with a computational scaling pro-
portional tol N2 that may be reduced to DN (with a
large prefactor) pix
pix
at the price of additional approximations.
Experiment speciÐc techniques have also been proposed :
Oh, Spergel, & Hinshaw (1999) described a fast power spec-
trum extraction technique designed for usage with the MAP
satellite data. Their method scales like DN2 with the size
of the pixelized map and takes advantage pix of uncorrelated
pixel noise with approximate axisymmetric distribution on
the sky. Wandelt (2000) advocates the use of the set of rings
as a compressed form of the Planck data set from which to
extract optimally the C in the presence of correlated noise.
The applicability of thisl approach is limited by its assump-
tion of the symmetry of the scanning strategy.
This paper is organized as follows : In ° 2 we describe how
a data stream of observations is reduced to a CMB Ñuctua-
tion map and how the angular pseudoÈpower spectrum C3
is extracted from such a map by SHT. In ° 3 we show howl
an unbiased estimate of the true underlying power spec-
trum can be recovered from the C3 with the aid of the MC
simulations. The tests of the l method on simulated
BOOMERANG-LDB observations are described in ° 4,
and the application of the method is discussed in ° 5.
2. FROM TIME-ORDERED DATA TO PSEUDOÈPOWER
SPECTRUM
Single-dish CMB experiments produce for each detector
a time-ordered data stream (TOD) describing at regular
time intervals the direction of observation and the sky tem-
perature as measured through the instrumental beam. We
assume that the beam is known, that it is close to isotropic
in the main lobe, that the side lobes are negligible, and that
the pointing at each time is known to an accuracy better
than the size of the main lobe of the beam. We also assume
that the response of the detectors to the sky temperature
Ñuctuations is calibrated beforehand. Exceptions to these
assumptions will be addressed in ° 3.6. We will Ðnally
assume that all the TOD samples a†ected by transient
events, such as cosmic-ray hits, have been Ñagged out and
that in order to preserve the TOD continuity necessary to
the Ðltering operations (performed in Fourier space), the
3
resulting gaps are Ðlled with fake data having the same
statistical properties as the genuine observations (see, e.g.,
Prunet et al. 2000 ; Stompor et al. 2000 for details).
2.1. From T OD to Sky Map
The data produced by each detector at a time t can be
modeled as
d \P * ]n , (1)
t tp p t
where * is the sky temperature, which we assume to be
p
pixelized and smoothed with the instrument beam, P is the
tp
pointing matrix, p is the pixel index, and n is the instrumen-
t
tal noise.
If the TOD noise is Gaussian distributed with a known
correlation function N { \ Sn(t)n(t@)T, the optimal solution
tt
for the sky map
m \ (Ps N~1 P )~1Ps N~1 d (2)
p pt tt{ t{p{ pt tt{ t{
minimizes the residual noise in the pixelized map, * -m
(Lupton 1993 ; Wright 1997 ; Tegmark 1997). While being
completely general, this procedure is impractical for very
long TOD streams because of the required inversion of the
large matrix N { . A simpliÐcation is possible under the
assumption of the tt TOD noise being piece-wise stationary
and its correlation matrix representable as circulant, N { \
N(t [ t@). Equation (2) can then be solved either directly tt
(with a computational scaling of DN3 ) or by using iter-
pix
ative methods as discussed by Wright (1997) or Natoli et al.
(2001 ; in which case the computation time is dominated by
Fourier space convolutions of the TOD corresponding to
the product N~1 d in eq. [2]). Iterative approaches scale
like N N log tt{ Nt{ , where N is the number of time
iterandq N isq the numberq of iterations. N depends
samples,
on the requirediter accuracy of the Ðnal map, and iterit is on the
order of a few tens in the case of a conjugate gradient
method of linear system solution (Natoli et al. 2001).
If the TOD noise properties are not known beforehand,
however, as is generally the case, equations (1) and (2) can
be solved iteratively together. This returns at each time step
an estimate of the noise stream, n(t), and, hence, of the noise
time power spectrum. The required computational scaling
involves a somewhat larger N (Ferreira & Ja†e 2000 ;
Prunet et al. 2000 ; Stompor et al.iter2000 ; Dore et al. 2001b).
Since the MASTER method requires repetitive TOD
simulations, processing, and mapmaking, the iterative solu-
tion of equation (2) can be too time-consuming for practical
applications. Therefore, to avoid the necessity to iterate, we
use a suboptimal, fast mapmaking method involving the
high-pass Ðltering of the TOD stream, which improves the
long timescale behavior of the noise and reduces the strip-
ing of the resulting map (see ° 3.2). The map solution is now
m \ [N (p)]~1 ; Ps f (t [ t@)d { , (3)
p obs pt t
tt@
where N (p) 4 Ps P is the number of observations in the
pixel p, obs pt tp the high-pass Ðlter. The computa-
and f denotes
tional scaling is now reduced to N log N . Clearly, equa-
q (2) if the
q TOD noise is
tion (3) is only equivalent to equation
white, i.e., N { \ N d (t [ t@) (in which case the Ðlter
tt to0 f Dirac
\d
would be reduced (t), i.e., no Ðltering would be
Dirac
applied). While the application of the high-pass Ðlter
reduces the long-term noise correlations, it degrades the
CMB signal at low frequencies (see Fig. 1) and a†ects the
4 HIVON ET AL. Vol. 567

FIG. 1.ÈSimulation of the BOOMERANG-LDB experiment and application of MASTER to extract the CMB angular power spectrum. The oval
contour on the maps shows the ellipse (distorted by projection) within which the power spectrum is estimated ( f \ 1.8% of the sky). T op left panel : A
random realization of the CMB sky from the theoretical model described by the power spectrum shown in the top right sky panel (red line). Middle left panel : A
noiseless map of the same region of the sky made from the TOD with actual BOOMERANG-LDB pointing and processed with the 100 mHz high-pass Ðlter
(see text). Bottom left panel : The di†erence between the two CMB sky maps shown above, which shows the component of the sky signal lost owing to the
combination of BOOMERANG scanning and data processing. Middle right panel : A simulation of the same BOOMERANG CMB sky map with the
instrumental noise included. Bottom right panel : Integration time per pixel for the actual scanning of the BOOMERANG-LDB channel B150A ; the average
integration time is about 500 s deg~2. T op right panel : The input power spectrum smoothed by the beam and pixel window function (red line) ; the average
angular power spectrum of the instrumental noise (black line) ; the pseudo-C directly measured on the sky map shown in the middle right panel and divided
l
by f (orange line) ; the binned MASTER estimate of the full-sky power spectrum after removal of noise contribution and correction of the e†ect of the
sky
high-pass Ðltering and mode-mode coupling (blue histogram).

resulting angular power spectrum derived from the Ðltered with

map solution equation (3). This e†ect is quantiÐed and cor-
rected for with the MC simulations and analysis involving l
*T (u) \ ; ; a Y (u) . (5)
the Ðltered mapmaking technique applied to the simulated lm lm
l;0 m/~l
TODs of the pure CMB signal. This procedure will be dis-
If the CMB temperature Ñuctuation *T is assumed to be
cussed in detail later on.
Gaussian distributed, each a is an independent Gaussian
2.2. From Sky Map to PseudoÈPower Spectrum lm
deviate with
A scalar Ðeld *T (u) deÐned over the full sky can be
Sa T \ 0 , (6)
decomposed in spherical harmonic coefficients lm

a \
P du*T (u)Y * (u) , (4)
and
lm lm Sa a*{ {T \ d { d {SC T , (7)
lm l m ll mm l
No. 1, 2002 MASTER ANALYSIS OF CMB DATA SETS 5

where SC T 4 Cth is speciÐed by the theory of primordial (not necessarily uniform) white noise. Under these assump-
l l tions, P(C3 o Cth) can be used to perform a maximum likeli-
perturbations and parameterized accordingly, and d is the
l
Kronecker symbol. In the absence of noise, an unbiased hood Ðt of the cosmological parameters to the observed
estimator of Cth is given by data set. F. Hansen et al. (2002, in preparation) extend this
l approach by using the full pseudo-C3 covariance matrix. We
1 l l
C\ ; o a o2 . (8) will now build, starting from the measured C3 , and under
l 2l ] 1 lm l
m/~l more general conditions on the noise properties and shape
C is s2-distributed with the mean equal to Cth, l \ 2l ] 1 of the observing window, a new estimator of the full-sky
l
degrees l of freedom (dof ), and a variance of 2Cth2/l.
l power spectrum that can be compared directly to Cth.
l
In the case of CMB measurements, the temperature Ñuc- 3. FROM PSEUDOÈPOWER SPECTRUM TO FULL-SKY
tuations cannot be measured over the full sky, either POWER SPECTRUM ESTIMATOR
because of ground obscuration or galactic contamination,
for example, and a position-dependent weighting W (u) can The pseudoÈpower spectrum C3 rendered by the direct
l
also be applied to the measured data, for instance, in order SHT of a partial sky map (eq. [13]), is clearly di†erent from
to reduce the edge e†ects or to downweight the noisy pixels. the full-sky angular spectrum C , but their ensemble aver-
l
If f represents the sky fraction over which the weighting ages can be related by
sky
applied is nonzero, then SC3 T \ ; M {SC {T , (14)

f w\
1 P
duWi(u) (9)
l
l@
ll l
where M { describes the mode-mode coupling resulting
sky i 4n
4n from the llcut sky. As described in Appendix A, this kernel
is the ith moment of the arbitrary weighting scheme. The depends only on the geometry of the cut sky and can be
window function can also be expanded in spherical har- expressed simply in terms of the power spectrum W of the
l and
monics with the coefficients w \ / duW (u)Y * (u), and with spatial window applied to the survey (see eqs. [A31]
lm lm [A14] for the spherical and planar geometry, respectively).
a power spectrum
The e†ect of the instrumental beam, experimental noise,
1 and Ðltering of the TOD stream can be included as follows :
W\ ; o w o2 , (10)
l 2l ] 1 lm
m SC3 T \ ; M { F { B2SC T ] SN3 T , (15)
for which W \ 4nf 2 w2 and £ W (2l ] 1) \ l ll l l{ l{ l
l/0 sky 1 lz0 l l@
4nf w . where B is a window function describing the combined
Askysky2 temperature Ñuctuation map *T (u) on which a smoothing
l e†ects of the beam and Ðnite pixel size, SN3 T is
window W (u) is applied can be decomposed in spherical l
the average noise power spectrum, and F is a transfer func-
harmonics coefficients l
tion that models the e†ect of the Ðltering applied to the data

a8 \
P du*T (u)W (u)Y * (u) (11)
stream or to the maps. The determination of each of these
terms will be described below.
lm lm
It is often assumed that the s2 distribution of C on
B ) ; *T (p)W (p)Y * (p) , (12) the full-sky can be generalized l/2l`1
to cut sky observationsl by
p lm
p scaling l to the number of dof e†ectively available. Given
where the integral over the sky is approximated by a dis- the large value of l, the central limit theorem is also invoked
crete sum over the pixels that make the map, with an indi- to further simplify this to a Gaussian of the same mean and
vidual surface area ) . variance. From these successive (and excessive) simpliÐca-
The pseudoÈpowerp spectrum C3 can be deÐned as tions we will only retain, as a rule of thumb, that the rms of
C averaged over a range *l is approximately
C3 \
1
l 2l ] 1
l
; o a8 o2 .
m/~l
lm
(13)
l
C
*C B C ]
DS
N(l) 2
, (16)
The computation of equation (12) for each (l, m) up to l l B2(l) l
l
l\l performed on an arbitrary pixelization of the sphere with
maxscale as N l2 . However, if one uses an adequate
would
pix to
max exploit the symmetries of spherical w2
layout of the pixels l \ (2l ] 1)*lf 2, (17)
harmonics such as, for example, the equidistant cylindrical l sky w
4
projection (ECP ; Muciaccia, Natoli, & Vittorio1997), where the factor w2/w accounts for the loss of modes
induced by the pixel 2weighting.
HEALPix (Gorski, Hivon, & Wandelt 199810), or Igloo 4
We will show in ° 4 how this
(Crittenden & Turok 1998), this computation actually compares to the results of MC simulations.
scales like N1@2 l2 . In our implementation of the MASTER
pix application
max 3.1. Mode-Mode Coupling Kernel
method, after of the window function on the
map, the program ANAFAST from the package HEALPix The resolution in l of the measured power spectrum is
was used to compute the pseudoÈpower spectrum. ultimately determined by the extent of the observed area of
Wandelt, Hivon, & Gorski (2000) showed that the mar- the sky, its geometrical shape, and the pixel weighting W (u)
ginalized likelihood P(C3 o Cth, N) of the pseudo-C3 for a applied to the survey (see Hobson & Magueijo 1996 ;
l and a given noise covariancel
given underlying theory Cth N Tegmark 1996). Although we only tested numerically the
can be computed analytically in O(N1@2 l2 ) operations, method on a circular or elliptically shaped window, nothing
pix max function and
under conditions of axisymmetric sky window prevents the use of a more complex window, especially for a
pixel-starved experiment with a nonconvex survey area, for
10 See also http ://www.eso.org/kgorski/healpix/. which a well-designed apodization could help to improve
6 HIVON ET AL.
the achievable spectral resolution. W (u) can also be
designed to downweight the noisy pixels in order to
improve the accuracy of the recovered spectrum at large l,
where the noise is dominant, but such a weighting scheme
generally increase the error bars at low l. We will show in ° 4
how the choice of window proÐle changes the estimated C
l peak
spectrum error bars.
3.2. T OD Filter T ransfer Function
The transfer function F introduced in equation (15)
describes the e†ect of any l Ðltering applied to the TOD
stream or to the map on the angular power spectrum. A
speciÐc example of the latter is the removal of parallel
stripes (of unknown origin) extending along a direction dif-
ferent from the scanning direction observed in some chan-
nels of the BOOMERANG-LDB data (NetterÐeld et al.
2002 ; C. R. Contaldi et al. 2002, in preparation). The Ðlter-
ing of the TOD has broader applications, however, and can
take the form of a high-pass Ðltering that serves several
purposes, as follows :
1. Reduce the contribution of the low-frequency noise
(1/f noise) to the map, especially if the scanning strategy
and/or the mapmaking technique used do not optimize the
removal of these modes ;
2. reduce the scan or spin synchronous noise, which may
appear at the scan frequency and its harmonics ;
3. remove from the signal the contribution from the large
scale anisotropies, which are poorly constrained on an
incomplete sky survey, and are likely to contaminate the
estimated power spectrum at all the smaller scales.
It should be noted that the validity of equation (14) for
any sky window relies on the fact that the statistical proper-
ties of the full-sky temperature Ñuctuations are isotropic, as
implied by equation (7). This assumption is broken by the
high-pass Ðltering of the data stream, which creates a pre-
ferred direction on the sky. Therefore, the introduction of
the single scalar function F in equation (15) should be seen
as a simplifying Ansatz for la more complex reality. We can
numerically test the validity of this Ansatz by showing for
instance that F depends weakly on the choice of the under-
l
lying power spectrum.
Given an input CMB power spectrum Cth, a number,
N(s) , of noise-free full-sky realizations of this lspectrum can
beMCsimulated (using, for example, the program SYNFAST
from HEALPix). These maps are then ““ observed ÏÏ using the
actual scanning strategy, the resulting model TODs project-
ed back on the sky using equation (3), and their individual
power spectra C3 are extracted using equations (12) and (13).
Equation (15)l can then be applied to the average SC3 T
l MC
of these measurements to determine the transfer function. In
order to avoid inverting the kernel M { , which can be ill-
ll
conditioned as it corresponds to a convolution operator,
this system can be solved iteratively. In the case of a scan-
ning experiment, such as BOOMERANG, Appendix B
shows how a model of parallel scan can render a Ðrst-order
solution F(0). In the more general cases, a starting solution
could be F(0) \ SC3 T /( f w B2 Cth). If SC3 T
is replaced by l its running
l MC average
sky 2 l S(SC
l 3 T ) (typically
l MC
l MC
computed over *l \ 50 points), it appears that one itera-
tion is sufficient to obtain a stable estimate of the transfer
function
S(SC3 T ) [ ; @ M { F(0) B2SC T
F \ F(0) ] l MC l ll l{ l{ l{ . (18)
l l B2SC T f w
l l sky 2
Vol. 567
We used equation (19) to compute the transfer function in
MC simulations of BOOMERANG-LDB observations on
an elliptically shaped region that comprises 1.8% of the sky
(see ° 4 for details) for two di†erent input power spectra, the
Ðrst one corresponding to a Ñat universe with a Ðrst peak at
l ^ 220, and the other one corresponding to an open
universe with l ^ 420. Figure 2 shows that the resulting
peak
F is almost identical, with a discrepancy smaller than 1% in
l
the range 50 \ l \ 1000. This justiÐes the use of the simple
Ansatz (15) as a model of the e†ect of the time Ðltering on
the angular power spectrum and demonstrates that the
determination of the transfer function can be done nearly
independently of any assumptions about the actual CMB
power spectrum.
Using approximation (16) one expects that the error dF
l
done on the MC estimation of the transfer function
decreases as
dF
lB
S 2 1
. (19)
F (2l ] 1)*lf JN(s)
l sky MC
So if *lf ^ 1 and l [ 100, an estimate of the transfer
sky than 1% can be obtained in N(s) [ 100 rea-
function better
MC
lizations.
The computation of the transfer function F is required
l (eq. [3]),
because of the Ðltered mapmaking technique used
which alters the signal in low-frequency modes. Even if a
more sophisticated mapmaking were used, the map
obtained is generally not an unbiased representation of the
true sky because of various systematic e†ects present in the
data requiring a preprocessing of the TOD or of the map
and the computation of F might still be necessary.
l di†erent beams are analyzed
If several detectors with
simultaneously, the realization of the same sky with a di†er-
ent smearing can be used as an input to the TOD simula-
FIG. 2.ÈTransfer function of the measured angular power spectrum
corresponding to the sharp high-pass Ðltering at 100 mHz of the
BOOMERANG-LDB TOD is shown here derived for two di†erent input
CMB anisotropy power spectra. The upper panel shows the two input
spectra corresponding to a Ñat universe (green solid line) and an open
universe (red dashed line). The bottom panel shows the respective transfer
functions, which di†er by less than 1% up to l \ 1000. The black long-
dashed line shows the analytical prediction for a toy model of parallel
scans (eqs. [B14] and [B15]). [See the electronic version of the Journal for a
color version of this Ðgure.]
No. 1, 2002 MASTER ANALYSIS OF CMB DATA SETS 7

tion of each detector. In the analysis of the coadded map an The two operators above are deÐned for the Ñat-band, dis-
approximate e†ective beam window function B(l) can be joint bins. They can be easily modiÐed to account for an
modeled as the weighted average of all individual beams l-dependent weighting within each bin (for example,
(see Wu et al. 2001). If the actual e†ective beam (which may designed to enhance the less noisy multipoles) without
vary from pixel to pixel according to the integration time of changing the rest of the discussion.
each detector in each pixel) were di†erent from this model, Rewriting equation (15) as
the discrepancy would be reÑected in the transfer function SC3 T \ K {SC {T ] SN3 T , (22)
F and would therefore be taken into account in the C l ll l l
extraction. l we then look for a solution to
In the assumption that the experiment is designed for, P K {SC {T \ P (SC3 T [ SN3 T) . (23)
and actually achieves, an observation free of any systematic bl ll l bl l l
and long-term drift, the computation of F would not be This system has l unknowns for n equations. We
l max bins
necessary, speeding up signiÐcantly the power spectrum cal- seek the solution such that SC T \ l(l ] 1)SC T/2n is a
culation process. piece-wise constant. If we replacel SC T by Q Pl {SC {T \
Q SC T, then l lb bl l
3.3. Noise Power Spectrum lb b
From the point of view of modeling the CMB experiment, SC T \ K~1{ P { (SC3 T [ SN3 T) , (24)
b bb b l l l
the simplest form of instrumental noise is the stationary, where
white, Gaussian process. Reality, however, is usually not as
K \P K {Q{ { ,
simple ; the actual experimental noise is often nonsta- bb{ bl ll l b
tionary, ““ colored,ÏÏ sometimes non-Gaussian, and corre- \ P M { F { B2{ Q { { . (25)
bl ll l l l b
lated to some internal variables of the instrument, such as An unbiased estimator CΠof the whole sky power spec-
its acceleration, the cold plate temperature, the orientation b
trum is then given by
relative to the Sun, to the balloon, or to the ground. For
many of those reasons the noise often cannot be efficiently CΠ\ K~1{ P { (C3 [ SN3 T ) . (26)
b bb b l l l MC
averaged out. However, if these noises can be modeled to a An estimator of the noise ““ on the sky ÏÏ can also be intro-
reasonable accuracy, they can be included, at little or no duced
extra cost, in the MC pipeline described here, and their
e†ect on the measured C can be assessed and possibly NŒ \ K~1{ P { SN3 T . (27)
l b bb b l l MC
removed. Figure 3 shows the kernel K { and its inverse for the
As mentioned in °2.1 an estimate of the noise time corre- conÐguration described in ° 4. Itbbshould be noted that the
lation function N { and its time-power spectrum can be binning of the l space is performed at the last stage of the
extracted from thettactual data stream. Using this informa- analysis and can be chosen after the MC simulations are
tion, a fake Gaussian ““ noise stream ÏÏ can be simulated and done.
projected on the sky with the actual scanning strategy using
equation (3), with the same high-pass Ðltering f. The power 3.5. Covariance Matrix of the Estimated Power Spectrum
spectrum N3 (l) of the resulting noise map is extracted accor- In order to be able to extract the cosmological parame-
ding to equations (12) and (13), and the whole process is ters from the power spectrum estimated as described above,
reproduced as many times as necessary to obtain a MC one needs to know the errors on each C , and the corre-
b is contained in
estimate of the average noise angular power spectrum, lations between the bins. This information
SN3 (l) T. If several detectors are analyzed simultaneously the covariance matrix of the estimated power spectrum,
MC noise is known to be correlated, these correlations which can be estimated as follows. A smooth interpolation
and their
can be included in the noise stream simulations. of CΠis used as the underlying CMB power spectrum, and a
new bset of MC simulations, including both signal and noise,
3.4. Estimated Power Spectrum
as well as all the experiment peculiarities, is generated,
In order to reduce the correlations of the C induced by analyzed, and reduced in the same way as the real data. This
l resulting
the cut sky and also to reduce the errors on the generates a set of binned power spectrum estimators MCΠN.
power spectrum estimator, it is convenient to bin the power b
The elements of the correlation matrix are deÐned as
spectrum in l. The slowly varying ““ Ñattened ÏÏ spectrum
C 4 l(l ] 1)C /2n is a preferable candidate for such binning. C { \ S(CΠ[ SCΠT )(CΠ{ [ SCΠ{T )T . (28)
l a set of n l bins, indexed by b, with respective bound- bb b b MC b b MC MC
For The error bars on CΠare then given by the square root of
aries l(b) \ l(b)bins\ l(b`1), one can deÐne the binning oper- b C:
the diagonal elements of
ator Plow
high low
as follows
*CΠ\ C1@2 . (29)
04 1 l(l ] 1)
P 5 2n l(b`1) [ l(b)
, if 2 ¹ l(b) ¹ l \ l(b`1)
low low
b bb

bl
060 ,
low low
otherwise
(20) 3.6. T he Algorithm
Assuming that for a given CMB experiment the instru-
mental beam and the time domain power spectrum of the
and the binned power spectrum is C \ P C . The recipro-
b bl linterpolation) instrumental noise are known, the process of estimation of
cal operator (corresponding to a piece-wise
then reads the full-sky power spectrum from the noisy observations of
CMB temperature Ñuctuations can be summarized.
0 4 2n
5 l(l ] 1)
, if 2 ¹ l(b) ¹ l \ l(b`1)
low low
The required tools are as follows :
Q
0
lb
60 , otherwise .
(21) (a) A simulation facility for generation of the random
realizations of the CMB sky (e.g., synfast from HEALPix) ;
8 HIVON ET AL.
FIG. 3.ÈBinned power spectrum coupling kernel K { and its inverse (absolute values shown ; negative elements are in green) for an elliptically shaped
top-hat sky window covering 1.8% of the sky. Bin widthbbis *l \ 50, except for the last bin, for which *l \ 150. The diagonal elements and the l \ 200 and
700 rows of both matrices are shown in the bottom panels.
(b1) a software model of the experiment that simulates
observations of the sky using the appropriate scanning
strategy (eq. [1]) and generates the model CMB signal
TOD streams ;
(b2) a noise simulator that can generate random realiza-
tions of the noise with an appropriate power spectrum ;
possible non-Gaussian noise features or cross-correlations
between detectors should be added at this stage ;
(b3) a fast map-making facility, which implements equa-
tion (3), and accounts for any alterations of the observed
TOD stream and/or the produced map ;
(c) a software to compute the pseudoÈpower spectrum
(eqs. [12] and [13]) from a given apodized cut sky map (e.g.,
ANAFAST from HEALPix).
After the choice of the sky window apodization function
is made, the procedure of estimation of the power spectrum
involves the following steps :
1. Equation (A31) is used to evaluate the C coupling
l
kernel M { , which accounts for the e†ects of limited sky
ll
coverage and apodization.
2. A number N(s) of noise-free MC simulations of the
MC
observed TOD [produced with (a) and (b1), projected on
Vol. 567
b1
the sky with (b3), and analyzed with (c)] are used to estimate
the transfer function F of any Ðltering that is applied to the
actual TOD stream. l
3. A number N(n) of pure noise MC simulations of the
TODs [made withMC(b1) and (b2), and then projected and
analyzed with (b3) and (c)] are used to estimate the angular
power spectrum SN3 T of the noise projected on the sky ;
MC TOD is converted into a map using
4. The experimental
(b3) and its pseudoÈpower spectrum C3 is obtained with (c).
l an estimate of the
5. Next a set of l-bins is deÐned, and
underlying full-sky binned power spectrum is computed
using equation (26) ; according to equation (24) this is an
unbiased estimatorÈSCŒ T \ SC T ; this will be demon-
b BOOMERANG
b
strated in ° 4 with simulated observations
in which the input spectrum is known.
6. The covariance matrix C { (eq. [28]) is computed
bb
from N(s`n) simulations of the whole experiment, and the
MC
error bars on the binned power spectrum are obtained from
its diagonal elements (eq. [29]).
We assumed the physical beam to be close enough to
axisymmetric so that its smoothing e†ect on the tem-
perature map is independent of the payload attitude along
No. 1, 2002 MASTER ANALYSIS OF CMB DATA SETS 9

the line of sight. If it were not the case, a direct integration
of the temperature over the beam would have to be per-
formed for each time sample in equation (1). This operation
could be very intensive for extremely structured beams,
unless some symmetries in the scanning strategy, such as the
one expected for satellite missions, allow for a fast convolu-
tion implementation (Wandelt & Gorski 2000 ; Challinor et
al. 2000).
On the other hand, the e†ect of a pointing inaccuracy,
whether it is axisymmetric or not, can easily be included in
the method by modifying equation (1) accordingly.
3.7. Computational Scaling
The algorithm is maximally parallelizable as each MC
cycle of the algorithm can be performed on a separate CPU.
Therefore, the total time required for completion of the
estimation of the power spectrum from the single detector
CMB observations, with N \ N(s) ] N(n) ] N(s`n)
MC MC MC MC
cycles run on n processors, is given by
CPU
N mapmaking and a larger number of MC cycles to improve
T \ MC T , (30)
total n MCcycle
CPU
with T \ T ] T ] T , where (a), (b), and (c) refer
to theMCcycle
CPU time (a) consumption
(b) (c) by the simulation tools
described in ° 3.6 (CMB map synthesis, observation simula-
tion and mapmaking, and map analysis, respectively). In the
case of joint multidetector analysis, stage (b) has to be
repeated for each detector, whereas stage (a) only has to be
repeated for each di†erent beam. In our own implementa-
tion (see ° 4 for detailed speciÐcations), the MASTER
method is executed on personal computers (PCs) equipped
with 850 MHz AMD Athlon CPUs ; the following per-
formance is achieved (on each processor) :
observations, return N ^ 5.6 ] 109 TOD points, which
q
will be used to make a HEALPix full-sky map with N \ 5
pix
] 107 pixels (of average size of 1@.7) and which in turn
should allow us to estimate the angular power spectrum of
the sky signals up to l \ 3000. With these speciÐcations,
max
and assuming N ^ 1.5 ] 107 (one day), the CPU time
FFT
required for execution of our current implementation of
MASTER becomes
N
T ^ 24 days MC
A BA B 8
(34)
total 300 n
CPU
on currently available PCs. Since dedicated computer
servers with 32 or more processors at least twice as fast as
the PCs that we used are already presently available, the
quoted total execution time for the MASTER method can
be reduced to about 3 days. The extra speed-up of the CPUs
between now and the launch of Planck (a factor of D16
according to MooreÏs law) should enable the simultaneous
analysis of several Planck channels, with more sophisticated
the power spectrum estimation accuracy, in a total time of a
few days. This means that the MASTER approach is fully
practical from the point of view of demands related to scien-
tiÐc analysis of the Planck data.
4. MASTER TESTS ON SIMULATED BOOMERANG
OBSERVATIONS
4.1. Simulations of CMB Observations during the L ong
Duration Balloon Flight of BOOMERANG Experiment
The MASTER method was tested for application to the
extraction of the CMB anisotropy power spectrum from the
data collected by the BOOMERANG-LDB experiment.
These CMB observations comprise a total of about

T \ T ^ 300 s
N A
pix
1@2 l B A B
max
2
, (31)
50 ] 106 samples that cover about 4.4% of the sky and were
acquired at two di†erent azimuthal sky scan rates of 1¡ and
(a) (c) 3 ] 106 1300
2¡ s~1 (for more details on the BOOMERANG-LDB Ñight,
and see Crill 2000 and B. Crill et al. 2002, in preparation).

T ^ 300 s
A
N
q
BC log N
FFT
D , (32)
We have modeled these CMB observations of the simu-
lated CMB sky with a scanning pattern identical to that of
(b) 5 ] 107 log (5 ] 105) one of the BOOMERANG-LDB channels. The instrumen-
tal noise generated in this channel was assumed to be
where N is the number of pixels over the whole sky at the
pix resolution (not in the cut sky map), N is the Gaussian with a time power spectrum identical to the one
chosen map
q measured during the observations. The characteristic fea-
number of time samples in the TOD set used, and N is
FFT
the typical number of time samples on which the required
tures of this noise power spectrum include a D1/f behavior
at low frequency, a knee-frequency of about 100 mHz, a
fast Fourier transforms (FFTs) are computed. This leads to
white noise level of 130 kK s1@2, a series of lines located at
an overall CPU time requirement of
N
T \ 0.5 day MC
A BA B 8
. (33)
the harmonics of the scanning frequency (8 mHz at 1¡ s~1
and 16 mHz at 2¡ s~1), and some microphonic artifacts at 8
Hz (see Fig. 4). The angular resolution of this channel was
total 300 n
CPU FWHM B10@, and we used the actual measured beam
These timing estimates could easily be improved if some proÐle in the calculations. The high-pass Ðltering applied
repetitive tasks, such as the translation from sky coordi- during the process of mapmaking (eq. [3]) was a sharp cut
nates to pixel index in (b), or the evaluation of Legendre o† at 100 mHz for the 1¡ s~1 scan rate and 200 mHz for the
polynomials in (a) and (c) were precomputed and stored on 2¡ s~1 scan rate.
disc. The sky maps were pixelized using HEALPix with 7@
N.B. The most daunting challenge currently foreseen for pixels (N \ 512). The power spectrum was computed
CMB anisotropy analysis arises in the process of reduction side of the data on an elliptically shaped region of
from a subset
and interpretation of the data that will be collected during semiaxes a \ 20¡ and b \ 12¡, which covers f \ 1.82% of
the ESA Planck mission (currently scheduled for launch in the sky and is centered on the best-observedskyregion of the
2007). Let us recompute our CPU time requirements to BOOMERANG-LDB Ñight. As shown in Figure 1, the sky
match the parameters that describe some of the Planck coverage in this area, and therefore the noise per pixel, is
High Frequency Instrument speciÐcations. One channel of nonuniform. The number of observations per pixel varies
the HFI with a beam resolution of 5@ should, during 1 yr of between 50 and 1510, with an average of 370, and a stan-
10 HIVON ET AL. Vol. 567

very low. This demonstrates that estimator (26) is not

biased. The error bars obtained from the MC simulations
(eq. [29]) can be compared to the ““ naive ÏÏ ones (eq. [16])
corrected for the transfer function F . This e†ect decreases
the e†ective number of modes at eachl l to
w2
lf \ (2l ] 1)*lf 2F , (35)
b b sky w lb
4
which renders the new analytical error estimate

*CΠ\ (Cth ] NΠ)
S 2
. (36)
b b b lf
b
Figure 5 shows that the MC error bars are almost identical
to these analytical estimates. Figure 5 also shows the spec-
trum estimate bin-bin correlation matrix, renormalized
such that the diagonal is unity. This matrix is diagonally
FIG. 4.ÈSpectral densities of the noise measured for one of the dominated, and except for the Ðrst few bins, the o†-diagonal
BOOMERANG-LDB channels is shown for the two scanning speeds (1¡ elements are at most 10% as large as the diagonal elements.
and 2¡ s~1) of the actual CMB observations. These spectra were used to The distribution of the MASTER CŒ estimators measured
generate the simulated observations. A sharp high-pass Ðlter was applied
in an l-bin b deÐnes the likelihood P(CŒ b o Cth) of measuring
to the measured and simulated TODs at 100 and 200 mHz, respectively.
CΠgiven the input power spectrum Cth (afterb likelihood mar-
[See the electronic version of the Journal for a color version of this Ðgure.] b
ginalization over all the other bins). This distribution was
estimated from 1352 MC simulations and is illustrated in
dard deviation of 120. The CMB dipole was not included in
Figure 6. The histograms show the simulated distribution
the sky simulations as the high-pass Ðltering used in map-
together with the scaled s2 model, where l (given by eq.
[35]) is indicated on each lpanel, and a Gaussian distribu-
making reduces its rms residual variation in the observed
region of the sky to less than 0.3 kK, negligible compared to
tion with the same mean and variance, which becomes
the rms amplitude of D150 kK of the intrinsic small-scale
undistinguishable from the s2 curve for large l. At all l-
CMB Ñuctuations.
values, and especially for small values of l, the s2 model,
The input CMB anisotropy spectrum used for the CMB
which has no free parameters, gives a much better descrip-
sky model was chosen to Ðt the results of the joint
tion of the actual distribution than the Gaussian.
MAXIMA-BOOMERANG data analysis (Ja†e et al. 2001) :
To estimate the cosmological parameters, one seeks a
) \ 1, " \ 0.7, h \ .82, ) h2 \ 0.03, and n \ 0.975.
b s theoretical model Cth deÐned by these parameters, for
The number of MC simulations performed to estimate l
which the likelihood P(Cth o CΠ) constructed for a given data
set is maximized. Using theb Bayes theorem, this can be
the noise, the high-pass Ðlter related transfer function, and
the C statistics was, respectively, N(n) \ 450, N(s) \ 250,
and Nl (s`n) \ 1350. Such high numbers
MC of MC simulations
MC rewritten as
MC P(Cth o CΠ) P P(CΠo Cth)P(Cth) , (37)
are not necessary in practice but were executed here to test
b b
accurately for possible biases and to measure to a good where P(Cth) is the prior on the cosmological parameters.
precision the statistical distribution of the C estimates. The term P(CΠo Cth) is often assumed to be a Gaussian func-
Three di†erent apodizations of the analyzed lsky region, b CŒ , whereas its dependence on the theory
tion of the data
described in Table 1, were used for power spectrum b
Cth can be approximated by an o†set lognormal function
estimation. (Bond, Ja†e, & Knox 2000). We see, however, from Figure 6
4.2. Statistics of Binned Spectrum Estimates that for experiments with small sky coverage, the likelihood
P(CΠo Cth) has to be described at small l as s2 function of
CΠtob avoid biasing the power spectrum estimation and the
Figure 5 shows the MC averages and errors on the esti-
mated CΠ, computed with the binwidth *l \ 50 for the three b
b apodizations used. cosmological parameters extracted from it.
sky window
Clearly, the average recovered power spectrum is very 4.3. Monte Carlo Convergence of the Power Spectrum
close to the input one both at small l, where the signal is Estimator
dominant, and at large l, where the signal-to-noise ratio is We checked, in the MC simulations described above, that
the estimators for the transfer function, the noise power
TABLE 1
spectrum (on the sky), and the accuracy of the error on
APODIZATIONS OF THE ANALYZED SKY REGION CΠconverge with the number of MC cycles, respectively, as
b
follows :
S S
Apodization W [r ¹ o(/)] w2/w
2 4
Top hat . . . . . . . 1 1 dF 100 400
l D a10~2 , (38)
Cosine . . . . . . . . cos [nr/2o(/)] 0.514 F N(s) l
Gaussian . . . . . . exp M[(1/2)[3r/o(/)]2N 0.223 l MC
with a between 0.6 and 0.9, depending on the sky window
NOTE.ÈApodization applied to the elliptically used,
shaped region of the sky used for extraction of the
CMB anisotropy power spectrum ; o(/) \ a/[1 ] (a2/
b2 [ 1) sin2 /]1@2 is the radius of the ellipse in a direc-
dNŒ
b D 0.6 ] 10~2
S S S
100 400 50
, (39)
tion /, and w \ / duW(u)i/4nf . NΠN(n) l *l
i sky b MC b
No. 1, 2002 MASTER ANALYSIS OF CMB DATA SETS 11

FIG. 5.ÈResults of the test of accuracy of our method of estimation of the CMB power spectrum are shown in application to the simulated
BOOMERANG-LDB data (with known input power spectrum) with three di†erent sky window apodizations (top to bottom : top hat, cosine, and Gaussian).
We used an ensemble of 1352 MC simulations. L eft panels : The input power spectrum is shown as a solid black line, and the black histogram shows its
bin-averaged values. The dashed lines show the average spectra of the simulated noise. The red histogram shows our MASTER ensemble mean values of the
estimated bin-averaged power spectra, which are in perfect agreement (to an error on the mean) with the input bin-averaged theory. For each spectral bin
three error bars are shown, all centered vertically on the mean values of the MASTER estimates, SCΠT , and, for clarity, spread around the l-centers of the
bins. The central (red) error bars show the rms values of the MASTER ensemble of estimates, whileb the MC right- and left-shifted (green and blue, respectively)
ones show the theoretical error estimates based on eqs. (35) and (36) derived, respectively, without and with an inclusion of the e†ect of the TOD high-pass
Ðlter related spectral transfer function F . Right panels : Absolute value of normalized binned spectrum correlation matrix C {/(C C { {)1@2 (see text). Elements
with value in [[0.2, 0] are represented inl green, those with value in [0, 0.2] in red, and those larger than 0.2 in black. bb bb b b

and and can be as low as a few percent for N(s) D N(n) D 100 in
MC CΠ. InMCthe noise-
d(*CΠ)
b D 0.1
100 S
. (40)
the signal-dominated regime where NΠ>
b
dominated regime, the ratio of the method induced
b
uncer-
*CΠN(s`n) tainty dCΠto the statistical error *CΠis
b MC b b
The contribution of the MC based estimation of F and NΠto
the error on the recovered power spectrum is
dCŒ
b ^ 0.1
100S (42)
*CΠN(n)
b MC
dCΠdF dNΠNΠand can be brought below 10% in about 100 MC simula-
b\ l] b b (41)
CΠF NΠCΠtions and analysis of pure noise TOD streams. Similarly,
b l b b
from equation (40), estimates of the individual statistical
12 HIVON ET AL. Vol. 567

FIG. 6.ÈStatistical distribution of the MASTER estimates of the binned full-sky CMB anisotropy power spectrum are shown for the bins centered at
l \ 50, 200, and 750. The results of application of three di†erent sky window apodizations, the top hat, Cosine, and Gaussian are shown in top to bottom
panels, respectively. The abscissa shows the values of the binned spectrum estimates, CΠ, centered and normalized with the theoretical values used in the
b
MASTER simulations : x \ (CΠ[ Cth)/*CΠ(see eq. [36]). The histogram shows the distribution derived from the MASTER simulations, the (blue) solid line
b b b
is derived from the s2 model where l is given by eq. (35), and the red dashed line a Gaussian of same mean and variance. [See the electronic version of the
l
Journal for a color version of this Ðgure.]

errors of each binned *CΠbetter than 10% can be obtained
with a few hundred MC bcycles. We have seen, however, in
° 4.2 that it is possible to predict analytically these errors
with great accuracy.
5. CONCLUSIONS
We have introduced a MASTER method for rapid esti-
mation of the angular power spectrum of the CMB anisot-
ropy from the modern CMB data sets. The method is based
on direct SHT of the observed area of the sky and MC
simulations of the relevant details of observations and
data processing. We demonstrated in an application of
the MASTER method to simulated observations of
BOOMERANG-LDB that this method renders unbiased
estimates of the CMB power spectrum, with error bars very
close to optimal. MC calibration of the MASTER method
requires generation and analysis of [1000 independent
simulated realizations of a given CMB experiment. We
demonstrated that CPU time requirements of the
MASTER approach permit successful analysis of the largest
CMB data sets that exist at the present time using very
modest computer facilities, for example, an inexpensive PC
farm.
Because of the combination of the unsophisticated map-
making that we used and the aggressive high-pass Ðltering
applied to the BOOMERANG data stream in our numeri-
cal tests of MASTER, the estimated power spectrum at the
lowest multipoles (l \ 100) has relatively large error bars.
We are currently investigating the use of a more sophisti-
cated mapmaking algorithm to try to improve this situ-
ation. Ultimately, however, in the large sky coverage
experiments, such as MAP or Planck, the power spectrum
at low l can be analyzed with fully Ñedged likelihood tech-
niques if a coarsened pixelization is used.
The CMB dipole provides a convenient way to calibrate
the detector response for experiments observing a large
enough fraction of the sky such as BOOMERANG or the
satellite missions. But, as already noted, because of the
aggressive high-pass Ðltering applied to the data stream,
such a calibration technique cannot be applied directly on
the unsophisticated maps used by the fastest implementa-
tion of MASTER. It should be stressed, however, that
MASTER can be used even if the absolute calibration of the
detectors is not known, all the analysis then being done in
the detector units (such as volts), and the noise and signal
power spectra are expressed in the same units. The conver-
sion to CMB temperature can be done in the Ðnal stage,
using a dipole calibration obtained from optimal maps.
Possible improvements of the method include the model-
ization of speciÐc systematic e†ects, the use of more sophis-
ticated mapmaking techniques, and the extension to CMB
polarization measurements. A head-to-head comparison
No. 1, 2002 MASTER ANALYSIS OF CMB DATA SETS 13

with optimal, and more costly, methods, such as MADCAP
will be presented elsewhere (e.g., C. R. Contaldi et al. 2002,
in preparation).
E. H. would like to thank O. Dore for stimulating dis-
cussions, J. Ruhl for useful comments on the manuscript, A.
Lange for continuous encouragement while this technique
was developed, and all the BOOMERANG team for pro-
viding such a stimulating environment. We thank A. J.
Banday for help with creating the acronym for our method
and for his careful reading of the manuscript. We acknow-
ledge the use of HEALPix, CMBFAST, and FFTW.

APPENDIX A

MODE-MODE COUPLING KERNEL

In this appendix we compute the mode-mode coupling kernel resulting from the cut-sky analysis, in both the planar
geometry case, where the mathematics involved may be more familiar, and on the sphere.

A1. ANALYSIS ON THE PLANE

A scalar Ðeld *T (r) deÐned on the plane (or deÐned on the sphere and projected on a tangent plane) can be decomposed in
Fourier coefficients as follows :

a(k) \
P dr*T (r)e~2ink Õ r , (A1)

and

*T (r) \
P dka(k)e2ink Õ r . (A2)

If *T is the homogeneous, isotropic, Gaussian-distributed temperature Ñuctuation, each a(k) is an independent Gaussian
random variable with
Sa(k)T \ 0 , (A3)
and
Sa(k)a*(k@)T \ d(k [ k@)SC(k)T \ d(k [ k@)SC(k)T , (A4)
where d is the Dirac delta function.
The Fourier coefficients derived on a weighted plane are then

a8 (k) \
P dr*T (r)W (r)e~2ink Õ r , (A5)

\
P dk@a(k@)K { [W ] . (A6)
kk
If we write W (r) \ / dkw(k)e2ink Õ r, the coupling kernel reads

K 4
P dre2ink 1Õ rW (r)e~2ink 2 Õ r (A7)
k1 k2

\
P P
dk w(k ) dre2inr Õ (k 1~k 2`k 3)
3 3

\
P dk w(k )d(k [ k ] k ) . (A8)
3 3 1 2 3
If the polar coordinates of the vector k are (k , h ), the following useful property of the Dirac delta function can be
i i i
demonstrated :
PP dh dh d(k ] k ] k ) \
P dh d(k [ o k ] k o )/k
1 2 1 2 3 2 1 2 3 1
\ 2nJ(k , k , k ) , (A9)
1 2 3
where the function J is deÐned as follows :
2
J(k , k , k ) \ (2k2 k2 ] 2k2 k2 ] 2k2 k2 [ k4 [ k4 [ k4)~1@2 (A10)
1 2 3 n 1 2 1 3 2 3 1 2 3
14 HIVON ET AL. Vol. 567

for o k [ k o \ k \ k ] k , and J \ 0 otherwise. It follows that

2 3 1 2 3
P dk J(k , k , k ) \ n /(k2`k3)22d(k2)J(k , k , k )
1 1 2 3 (k2~k3) 1 1 2 3

\2
P 1
du(1 [ u2)~1@2
~1
\ 2n , (A11)
and J(k , k , 0) \ d(k [ k )/k .
1 2 1 2 1
SC3 T 4
1 P dh Sa8 (k )a8 *(k )T , (A12)
k1 2n 1 1 1

\
1 P P P dh dk dk Sa(k )a*(k )TK [W ]Kp [W ]
2n 1 2 3 2 3 k1 k2 k1k3

\
1 P P Pdh dk SC T dh o K [W ] o 2
2n 1 2 k2 2 k1 k2

\ 2n
PP k dk k dk SC TW(k )J(k , k , k ) ,
2 2 3 3 k2 3 1 2 3
where W(k) \ / dhw(k)w(k)*/2n.
This equation can be rewritten as follows

SC3 T \
P k dk M SC T , (A13)
k1 2 2 k1 k2 k2
where the coupling kernel is given by

M
P
\ 2n k dk W(k )J(k , k , k ) . (A14)
k 1 k2 3 3 3 1 2 3

If C(k ) \ N (the white noise spectrum), then

2
SC3 T \ 2nN k dk W(k ) ,
P (A15)
k1 3 3 3
where we used equation (A11).
A2. ANALYSIS ON THE SPHERE
A scalar Ðeld *T (u) deÐned on the sphere and weighted with an arbitrary window function W (u) can be expanded in
spherical harmonics as follows :

a8 \
P du*T (u)W (u)Y * (u) (A16)
lm lm

\ ; a { { duY { {(u)W (u)Y * (u)

P (A17)
lm lm lm
l@m@
\ ; a { { K { {[W ] , (A18)
l m lml m
l@m@
where the kernel K describes the mode-mode coupling resulting from the sky weighting. If W is z-axis azimuthally symmetric,
K { { \ K { d { . (See Wandelt et al. 2000 for an analytical calculation of K { in the case when W is a top-hat window.)
lml mthat a8 lml
Note is ma linear
mm combination of Gaussian variables and is therefore Gaussian
lml m as well, but the a8 is not independent.
lm reads
If we use the series representation of the window function, W (u) \ £ w Y (u), the coupling kernel

K 4
P duY
lm lm lm
(u)W (u)Y * (u) (A19)
l1 m1 l2 m2 l1 m1 l2 m2

\ ; w
P duY (u)Y (u)Y * (u)
l3 m3 l1 m1 l3 m3 l2 m2
l3 m3
\ ; w ([1)m2 1 2
C
(2l ] 1)(2l ] 1)(2l ] 1) 1@2
3
D
l3 m3 4n
l3 m3
( l l l )( l l l )
] t 1 2 3 tt 1 2 3 t, (A20)
: 0 0 0 ;: m1 [m2 m3 ;
No. 1, 2002 MASTER ANALYSIS OF CMB DATA SETS 15

where we introduced the Wigner 3j symbol (or Clebsch-Gordan coefficient)

( l1 l2 l3 )
t t.
: m1 m2 m3 ;
Several properties of the 3j symbol will prove useful. This scalar object describes the coupling of three angular momentum
vectors (whose squared moduli are l (l ] 1), and projections on the same axis are m , for i \ 1, 2, 3) such that the total angular
i i i
momentum vanishes. The matrix
( l1 l2 l3 )
t t
: m1 m2 m3 ;
is nonzero only if the triangle relation
ol [l o¹l ¹l ]l (A21)
1 2 3 1 2
is satisÐed, and
m ]m ]m \0 . (A22)
1 2 3
The orthogonality relations of the Wigner symbols read

; (2l ] 1)( l l l )( l l l )
3 t 1 2 3 tt 1 2 3 t\d d
m1 m1{ m2 m2{
, (A23)
l3 m3 m
: 1 m m ;: m@ m@ m ;
2 3 1 2 3
; ( l1 l l )( l l l@ ) 1
tm 2 3 tt 1 2 3 t\ d d d(l , l , l ) , (A24)
m1 m2 : 1 m m ;: m m m @ ; l3 l 3{ m 3 m3{ 1 2 3 2l ]1
2 3 1 2 3 3
where d(l , l , l ) \ 1 when the triangular relation (A21) is satisÐed, and d(l , l , l ) \ 0 otherwise. Finally, several recursive or
1 2relations
3 1 2 3of the latter is
closed form can be used to compute the 3 [ j symbols. A useful example
( l1 l2 l3 )
t t \ ([1)L@2
C
(L [ 2l ) !(L [ 2l ) !(L [ 2l ) ! 1@2
1 2 3
D (L /2) !
(A25)
:0 0 0; (L ] 1) ! (L /2 [ l ) !(L /2 [ l ) !(L /2 [ l ) !
1 2 3
for even L 4 l ] l ] l (or equal to 0 for odd L ). When L ? 2l with i \ 1 , 2, 3 (i.e., when the triangle l , l , l is large and
1 i limit
2the 33 [ j symbol with m \ 0 has the asymptotic 1 2 3
close to equilateral)
i
( l1 l2 l3 ) 2
t t2 ] [L (L [ 2l1)(L [ 2l2)(L [ 2l3)]~1@2 .
:0 0 0; n
\ J(l , l , l ) , (A26)
1 2 3
where J is introduced in equations (A9) and (A10). (See Edmonds 1957 for further details on Wigner symbols.)
The ensemble-averaged power spectrum of the random scalar Ðeld *T (u) on the sphere computed with an arbitrary
weighting function W (u) can be represented as follows :
1 l1
SC3 T 4 ; Sa8 a8 T, (A27)
l1 2l ] 1 l1 m1 l1* m1
1 m1/~l1
1 l1
\ ; ; ; Sa a TK [W ]K * [W ]
2l ] 1 l2 m2 l3* m3 l1 m1 l2 m2 l1 m1 l3 m3
1 m1/~l1 l2 m2 l3 m3
1 l1 l2
\ ; ; SC T ; o K [W ] o 2 . (A28)
2l ] 1 l 2 l1 m1 l2 m2
1 m1/~l1 l2 m2/~l2
Upon substituting the kernel expansion in terms of Wigner symbols (A20) and reordering the sums, this expression expands to
2l ] 1
SC3 T \ ; SC T 2 ; ; w w [(2l ] 1)(2l ] 1)]1@2
l1 l2 4n l3 m3 l4* m4 3 4
l2 l3 m3 l4 m4
( l l l )( l l l ) (l l l )( l l l )
] t 1 2 3 tt 1 2 4 t ; t 1 2 3 tt 1 2 4 t, (A29)
: 0 0 0 ;: 0 0 0 ;m1 m2 : m1 [m2 m3 ;: m1 [m2 m4 ;
which can be remarkably simpliÐed with the aid of both the orthogonality relation of the Wigner symbols (A24), and the
deÐnition (10) of the power spectrum of the window function, W . The Ðnal expression reads
l
SC3 T \ ; M SC T , (A30)
l1 l1 l2 l2
l2
with
2l ] 1 (l l l )
M \ 2 ; (2l ] 1)W t 1 2 3 t2 . (A31)
l1 l2 4n 3 l3:0 0 0;
l3
16 HIVON ET AL. Vol. 567

The Wigner symbols can be numerically computed from (A25) or from any equivalent recurrence relations. Equation (A30)
expresses the ensemble averaged angular power spectrum measured with an arbitrary window on the sky for the statistically
homogeneous and isotropic Ñuctuations described by an arbitrary ensemble averaged power spectrum over the full sky.
If the input power spectrum is constant, SC T \ N, corresponding to the white noise distributed over the sky, equation
l
(A30) can be simpliÐed using the orthogonality relation (A23), and the measured windowed power spectrum is also a constant
SC3 T \ Nf w . (A32)
l sky 2

APPENDIX B

TRANSFER FUNCTION FOR PARALLEL SCANS

In the case of BOOMERANG, or of any scanning survey, a crude estimate of the transfer function corresponding to the
high-pass Ðltering of the TOD, F(0) (see eq. [19]), can be obtained by assuming that to Ðrst order the scans are parallel and
performed at uniform angular speed. In such a case the Ðltering of the TOD will alter the sky signal in a highly anisotropic
way. If the surveyed area is small enough (D20¡ in each direction), the tangent plane approach is sufficient to model the
survey. Hence, the map can be decomposed in plane waves

*T (x, y) \
P a(k , k )ei(kx x`ky y) , (B1)
x y
with k \ k cos h and k \ k sin h. The a(k) are zero-mean Gaussian variables with a variance
x y
Sa(k)a(k@)*T \ SP(k)Td(k [ k@) , (B2)
and the map power spectrum is given by

P(k) \
1 P dha(k)a(k)* . (B3)
2n
If the scan is performed along the x-axis, the map obtained from the Ðltered TOD is

*T (x, y) \
P a(k , k ) f (k )ei(kx x`ky y) , (B4)
filt x y x
and its power spectrum is

P (k) \
1 P dha(k)a(k)*f (k )2 . (B5)
filt 2n x
The ensemble-averaged Ðltered power spectrum is

SP (k)T \ SP(k)T
1 P dhf (k cos h)2 . (B6)
filt 2n
Hence, the e†ect of the TOD Ðltering on the power spectrum of the map amounts to a simple transfer function given by (l D k)

F(0) \
1 P dhf (k cos h)2 . (B7)
l 2n
If the scan is performed at an azimuthal speed v at an elevation h , and the high-pass Ðlter applied to the data has a
az el
Gaussian form
f (l) \ 1 [ e~(l@lc)2@2 , (B8)
then
n@2
F(0) \ 2/n
P
dh(1 [ e~(l cos h@lc)2@2)2 , (B9)
l
0
where l \ 2nl /(v cos h ). The asymptotic forms of F(0) are
c c az el
F(0) ]
l
3 l 4
,
AB l>l , (B10)
l 32 l c
c
J8 [ 1 l
F(0) ] 1 [ c, l?l . (B11)
l Jn l c
No. 1, 2002 MASTER ANALYSIS OF CMB DATA SETS 17

On the other hand, if the high-pass Ðlter is chosen in the form of a sharp cut at the frequency l
c
f (l) \ 0 , l\l , (B12)
c
f (l) \ 1 , lºl , (B13)
c
then
F(0) \ 0 , l\l , (B14)
l c
F(0) \ 1 [ 2/n sin~1 (l /l) , lºl . (B15)
l c c

REFERENCES
Bond, J. R. 1995, Phys. Rev. Lett., 74, 4369
Bond, J. R., Ja†e, A. H., & Knox, L. 1998, Phys. Rev. D, 57, 2117
ÈÈÈ. 2000, ApJ, 533, 19
Borrill, J. 1999a, in AIP Conf. Proc. 476, 3 K Cosmology EC-TMR Conf.,
ed. L. Maiani, F. Melchiorri, & N. Vittorio (Woodbury : AIP), 277
ÈÈÈ. 1999b, in Proc. 5th European SGI/Cray MPP Workshop,
Bologna
Challinor, A., Fosalba, P., Mortlock, D., Ashdown, M., Wandelt, B., &
Gorski, K. 2000, preprint (astro-ph/0008228)
Crill, B. P. 2000, Ph.D. thesis, California Inst. Tech.
Crittenden, R. G., & Turok, N. G. 1998, preprint (astro-ph/9806374)
de Bernardis, P., et al. 2000, Nature, 404, 955
Dore, O., Knox, L., & Peel, A. 2001a, preprint (astro-ph/0104443)
Dore, O., Teyssier, R., Bouchet, F. R., Vibert, D., & Prunet, S. 2001b, A&A,
374, 358
Edmonds, A. R. 1957, Angular Momentum in Quantum Mechanics
(Princeton : Princeton Univ. Press)
Ferreira, P. G., & Ja†e, A. H. 2000, MNRAS, 312, 89
Gorski, K. M. 1994, ApJ, 430, L85
ÈÈÈ. 1997, in Microwave Background Anisotropies, ed. F. R. Bouchet,
R. Gispert, B. Guiderdoni, & J. Tran Thanh Van (Gif-sur-Yvette : Edi-
tions Frontières), 77
Gorski, K. M., Hivon, E., & Wandelt, B. D. 1998, in Evolution of Large-
Scale Structure, ed. A. J. Banday, R. K. Sheth, & L. Da Costa (Garching :
ESO), 37
Halverson, N. W., et al. 2001, preprint (astro-ph/0104489)
Hanany, S., et al. 2000, ApJ, 545, L5
Hobson, M. P., & Magueijo, J. 1996, MNRAS, 283, 1133
Ja†e, A. H., et al. 2001, Phys. Rev. Lett., 86, 3475
Lee, A. T., et al. 2001, ApJ, 561, L1
Lupton, R. 1993, Statistics in Theory and Practice (Princeton : Princeton
Univ. Press)
Muciaccia, P. F., Natoli, P., & Vittorio, N. 1997, ApJ, 488, L63
Natoli, P., de Gasperis, G., Gheller, C., & Vittorio, N. 2001, A&A, 372, 346
NetterÐeld, C. B., et al. 2002, ApJ, (astro-ph/0104460)
Oh, S. P., Spergel, D. N., & Hinshaw, G. 1999, ApJ, 510, 551
Padin, S., et al. 2001, ApJ, 549, L1
Prunet, S., NetterÐeld, B., Hivon, E., & Crill, B. 2000, preprint (astro-
ph/0006052)
Pryke, C. L., Halverson, N. W., Leitch, E., Kovac, J., Carlstrom, J. E.,
Holzapfel, W. L., & Dragovan, M. 2001, preprint (astro-ph/0104490)
Smoot, G. F., et al. 1992, ApJ, 396, L1
Stompor, R., et al. 2000, preprint (astro-ph/0012418)
Szapudi, I., Prunet, O., Pogosyan, D., Szalay, A. S., & Bond, J. R. 2001,
ApJ, 548, L115
Tegmark, M. 1996, MNRAS, 280, 299
ÈÈÈ. 1997, ApJ, 480, L87
Tegmark, M., & Bunn, E. F. 1995, ApJ, 455, 1
Wandelt, B. D. 2000, preprint (astro-ph/0012416)
Wandelt, B. D., & Gorski, K. M. 2000, preprint (astro-ph/0008227)
Wandelt, B. D., Hivon, E., & Gorski, K. M. 2000, preprint (astro-ph/
0008111)
Wright, E. L. 1996, preprint (astro-ph/9612105)
Wu, J. H. P., et al. 2001, ApJS, 132, 1