E. Komatsu et al- First-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Tests of Gaussianity

8/3/2019 E. Komatsu et al- First-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Tests of Gaussianity

1/16

FIRST-YEAR WILKINSON MICROWAVE ANISOTROPY PROBE (WMAP)1 OBSERVATIONS:TESTS OF GAUSSIANITY

E. Komatsu,2 A. Kogut,3 M. R. Nolta,4 C. L. Bennett,3 M. Halpern,5 G. Hinshaw,3 N. Jarosik,4

M. Limon,3,6 S. S. Meyer,7 L. Page,4 D. N. Spergel,2 G. S. Tucker,3,6,8 L. Verde,2,9

E. Wollack,3 and E. L. Wright10

Received 2003 February 11;accepted 2003 May13

ABSTRACT

We present limits to the amplitude of non-Gaussian primordial fluctuations in the WMAP 1 yr cosmicmicrowave background sky maps. A nonlinear coupling parameter, fNL, characterizes the amplitude of aquadratic term in the primordial potential. We use two statistics: one is a cubic statistic which measures phasecorrelations of temperature fluctuations after combining all configurations of the angular bispectrum. Theother uses the Minkowski functionals to measure the morphology of the sky maps. Both methods find theWMAPdata consistent with Gaussian primordial fluctuations and establish limits, 58 < fNL < 134, at 95%confidence. There is no significant frequency or scale dependence of fNL. The WMAPlimit is 30 times betterthan COBE and validates that the power spectrum can fully characterize statistical properties of CMBanisotropy in the WMAPdata to a high degree of accuracy. Our results also validate the use of a Gaussiantheory for predicting the abundance of clusters in the local universe. We detect a point-source contribution tothe bispectrum at 41 GHz, bsrc 9:5 4:4 105 lK

3 sr2, which gives a power spectrum from pointsources of csrc 15 6 103 lK2 sr in thermodynamic temperature units. This value agrees well withindependent estimates of source number counts and the power spectrum at 41 GHz, indicating that bsrcdirectly measures residual source contributions.

Subject headings: cosmic microwave background cosmology: observations early universe galaxies: clusters: general large-scale structure of universe

1. INTRODUCTION

The Gaussianity of the primordial fluctuations is a keyassumption of modern cosmology, motivated by simplemodels of inflation. Statistical properties of the primordialfluctuations are closely related to those of the cosmic micro-wave background (CMB) radiation anisotropy; thus, a

measurement of non-Gaussianity of the CMB is a direct testof the inflation paradigm. If CMB anisotropy is Gaussian,then the angular power spectrum fully specifies thestatistical properties. Recently, Acquaviva et al. (2002) andMaldacena (2002) have calculated second-order perturba-tions during inflation to show that simple models basedupon a slowly rolling scalar field cannot generate detectablenon-Gaussianity. Their conclusions are consistent with pre-vious work (Salopek & Bond 1990, 1991; Falk, Rangarajan,& Srednicki 1993; Gangui et al. 1994). Inflation models thathave significant non-Gaussianity may have some complex-

ity such as non-Gaussian isocurvature fluctuations (Linde& Mukhanov 1997; Peebles 1997; Bucher & Zhu 1997), ascalar-field potential with features (Kofman et al. 1991;Wang & Kamionkowski 2000), or curvatons (Lyth &Wands 2002; Lyth, Ungarelli, & Wands 2002). Detection ornondetection of non-Gaussianity thus sheds light on thephysics of the early universe.

Many authors have tested the Gaussianity of CMB aniso-tropy on large angular scales ($7) (Kogut et al. 1996;Heavens 1998; Schmalzing & Gorski 1998; Ferreira,Magueijo, & Gorski 1998; Pando, Valls-Gabaud, & Fang1998; Bromley & Tegmark 1999; Banday, Zaroubi, &Gorski 2000; Contaldi et al. 2000; Mukherjee, Hobson,& Lasenby 2000; Magueijo 2000; Novikov, Schmalzing, &Mukhanov 2000; Sandvik & Magueijo 2001; Barreiro et al.2000; Phillips & Kogut 2001; Komatsu et al. 2002; Komatsu2001; Kunz et al. 2001; Aghanim, Forni, & Bouchet 2001;Cayon et al. 2003), on intermediate scales ($1) (Park et al.2001; Shandarin et al. 2002), and on small scales ($100)(Wu et al. 2001; Santos et al. 2002; Polenta et al. 2002).So far there is no evidence for significant cosmologicalnon-Gaussianity.

Most of the previous work only tested the consistencybetween the CMB data and simulated Gaussian realizationswithout having physically motivated non-Gaussian models.They did not, therefore, consider quantitative constraintson the amplitude of possible non-Gaussian signals allowedby the data. On the other hand, Komatsu et al. (2002),Santos et al. (2002), and Cayon et al. (2003) derivedconstraints on a parameter characterizing the amplitude ofprimordial non-Gaussianity inspired by inflation models.The former and the latter approaches are conceptually dif-ferent; the former does not address how Gaussian the CMBdata are or the physical implication of the results.

1 WMAPis theresult of a partnership between PrincetonUniversityandtheNASA Goddard Space FlightCenter.Scientific guidance is provided bythe WMAPScience Team.

2 Department of Astrophysical Sciences, Peyton Hall, PrincetonUniversity, Princeton, NJ 08544; [email protected].

3 NASA Goddard Space Flight Center, Code 685, Greenbelt,MD 20771.

4 Department of Physics, Jadwin Hall, Princeton, NJ 08544.5 Department of Physics and Astronomy, University of British

Columbia, Vancouver, BC V6T 1Z1, Canada.6 NationalResearchCouncil (NRC) Fellow.7 Departments of Astrophysics and Physics, EFI, and CfCP, University

of Chicago,Chicago, IL 60637.8 Department of Physics, BrownUniversity, Providence, RI 02912.9 Chandra Fellow.10 Department of Astronomy, UCLA, P.O. Box 951562, Los Angeles,

CA 90095-1562.

The Astrophysical Journal Supplement Series, 148:119134, 2003September

# 2003.The American AstronomicalSociety. All rightsreserved.Printedin U.S.A.

119


2/16

In this paper, we adopt the latter approach and constrainthe amplitude of primordial non-Gaussianity in the WMAP1 yrsky maps.

Some previous work all had roughly similar sensitivity tonon-Gaussian CMB anisotropy at different angular scales,because the number of independent pixels in the maps aresimilar, i.e., 40006000 for COBE (Bennett et al. 1996),QMASK (Xu, Tegmark, & de Oliveira-Costa 2002), andMAXIMA (Hanany et al. 2000) sky maps. Polenta et al.(2002) used about 4 104 pixels from the BOOMERanGmap (de Bernardis et al. 2000), but found no evidence fornon-Gaussianity. The WMAP provides about 2:4 106

pixels (outside the Kp0 cut) uncontaminated by the Galacticemission (Bennett et al. 2003a), achieving more than 1 orderof magnitude improvement in sensitivity to non-GaussianCMB anisotropy.

This paper is organized as follows. In x 2, we describe ourmethods for measuring the primordial non-Gaussianityusing the cubic (bispectrum) statistics and the Minkowskifunctionals, and we present the results of the measurementsof the WMAP1 yr sky maps. Implications of the results forinflation models and the high-redshift cluster abundance arethen presented. In x 3, we apply the bispectrum to individualfrequency bands to estimate the point-source contributionto the angular power spectrum. The results from theWMAP data are then presented, and also comparisonamong different methods. In x 4, we present summary of ourresults. In Appendix A, we test our cubic statistics for theprimordial non-Gaussianity using non-Gaussian CMB skymaps directly simulated from primordial fluctuations. InAppendix B, we test our cubic statistic for the point sourcesusing simulated point-source maps. In Appendix C, we cal-culate the CMB angular bispectrum generated from featuresin a scalar-field potential.

2. LIMITS ON PRIMORDIAL NON-GAUSSIANITY

2.1. The WMAP1 yr Sky Maps

We use a noise-weighted sum of the Q1, Q2, V1, V2,W1, W2, W3, and W4 maps. The maps are created in theHEALPix format with nside 512 (Gorski, Hivon, &Wandelt 1998), having the total number of pixels of12 nside2 3; 145; 728. We do not smooth the maps to acommon resolution before forming the co-added sum. Thispreserves the independence of noise in neighboring pixels, atthe cost of complicating the effective window function forthe sky signal. We assess the results by comparing theWMAPdata to Gaussian simulations processed in identicalfashion. Each CMB realization draws a sample from theCDM cosmology with the power-law primordial powerspectrum fit to the WMAP data (Hinshaw et al. 2003b;Spergel et al. 2003). The cosmological parameters are inTable 1 of Spergel et al. (2003) (we use the best-fit WMAPdata only parameters). We copy the CMB realization andsmooth each copy with the WMAPbeam window functionsof the Q1, Q2, V1, V2, W1, W2, W3, and W4 (Page et al.2003a). We then add independent noise realizations to eachsimulated map and co-add weighted by Nobs=20, where theeffective number of observations Nobs varies across the sky.The values of the noise variance per Nobs,

20, are tabulated

in Table 1 of Bennett et al. (2003b).We use the conservative Kp0 mask to cut the Galactic

plane and known point sources, as described in Bennett

et al. (2003a), retaining 76.8% of the sky (2,414,705 pixels)for the subsequent analysis. In total 700 sources are maskedon the 85% of the sky outside the Galactic plane in all bands;thus, the number density of masked sources is 65.5 sr1. TheGalactic emission outside the mask has visible effects onthe angular power spectrum (Hinshaw et al. 2003b). Sincethe Galactic emission is highly non-Gaussian, we need toreduce its contribution to our estimators of primordial non-Gaussianity. Without foreground correction, both thebispectrum and the Minkowski functionals find strongnon-Gaussian signals. We thus use the foreground templatecorrection given in x 6 of Bennett et al. (2003c) to reduceforeground emission to negligible levels in the Q, V, and Wbands. The method is termed as an alternative fittingmethod, which uses only the Q, V, and W band data. Thedust component is separately fitted to each band withoutassuming spectrum of the dust emission (three parameters).We assume that the free-free emission has a 2.15 spectrum,and the synchrotron has a 2.7 spectrum. The amplitude ofeach component in the Q band is then fitted across threebands (two parameters).

2.2. Methodology2.2.1. Model for Primordial Non-Gaussianity

We measure the amplitude of non-Gaussianity inprimordial fluctuations parameterized by a nonlinear cou-pling parameter, fNL (Komatsu & Spergel 2001). Thisparameter determines the amplitude of a quadratic termadded to the Bardeen curvature perturbations (H inBardeen 1980), as

x Lx fNL 2Lx h

2Lxi

; 1

where L are Gaussian linear perturbations with zero mean.Although the form in equation (1) is inspired by simpleinflation models, the exact predictions from those inflationmodels are irrelevant to our analysis here because the pre-dicted amplitude offNL is much smaller than our sensitivity;however, this parameterization is useful to find quantitativeconstraints on the amount of non-Gaussianity allowed bythe CMB data. Equation (1) is general in that fNL param-eterizes the leading-order nonlinear corrections to . Wediscuss the possible scale-dependence in Appendix C.

Angular bispectrum analyses found fNLj j < 1500 (68%)from the COBE DMR 53+90 GHz co-added map(Komatsu et al. 2002) and fNLj j < 950 (68%) from theMAXIMA sky map (Santos et al. 2002). The skewness mea-sured from the DMR map smoothed with filters, called theSpherical Mexican Hat wavelets, found fNLj j < 1100 (68%)(Cayon et al. 2003), although they neglected the integratedSachs-Wolfe effect in the analysis, and therefore under-estimated the cosmic variance of fNL. BOOMERanG didnot measure fNL in their analysis of non-Gaussianity(Polenta et al. 2002). The rms amplitude of is given byh2i1=2 h2Li

1=2 1 f2NLh2Li

. Since h2i1/2 measured on

the COBE scales through the Sachs-Wolfe effect ish2i1=2 3hDT2i1=2=T 3:3 105 (Bennett et al. 1996),one obtainsf2NLh

2Li < 2:5 10

3 from the COBE68% con-straints; thus, we already know that the contribution fromthe nonlinear term to the rms amplitude is smaller than0.25%, and that to the power spectrum is smaller than 0.5%.This amplitude is comparable to limits on systematic errorsof the WMAP power spectrum (Hinshaw et al. 2003a) and

120 KOMATSU ET AL. Vol. 148


3/16

needs to be constrained better in order to verify the analysisof the power spectrum.

2.2.2. Method 1: The Angular Bispectrum

Our first method for measuring fNL is a cubic statistic that combines nearly optimally all configurations of theangular bispectrum of the primordial non-Gaussianity(Komatsu, Spergel, & Wandelt 2003). The bispectrum

measures phase correlations of field fluctuations. We com-pute the spherical harmonic coefficients am of temperaturefluctuations from

am

Zd2nnMnn

DTnn

T0Ymnn ; 2

where Mnn is a pixel-weighting function. Here Mnn is theKp0 sky cut where Mnn takes 0 in the cut region and 1otherwise. We filter the measured am in -space and trans-form it back to compute two new maps, Ar; nn and Br; nn,given by

Ar; nn Xmax

2X

m

rb~CC

amYmnn ; 3

Br; nn Xmax2

Xm

rb~CC

amYmnn : 4

Here ~CC Cb2 N, where C is the CMB anisotropy, Nisthe noise bias, and b is the beam window function describ-ing the combined smoothing effects of the beam (Page et al.2003a) and the finite pixel size. The functions r and rare defined by

r 2

Zk2dkgTkjkr ; 5

r 2

Zk2dkPkgTkjkr ; 6where r is the comoving distance. These two functions con-stitute the primordial angular bispectrum and correspondto r f1NLb

NL r and r b

L r in the notation of

Komatsu & Spergel (2001). We compute the radiationtransfer function gTk with a code based upon CMBFAST(Seljak & Zaldarriaga 1996) for the best-fit cosmologicalmodel of the WMAP1 yr data (Spergel et al. 2003). We alsouse the best-fit primordial power spectrum of, Pk. Wethen compute the cubic statistic for the primordial non-Gaussianity, Sprim, by integrating the two filtered maps overr as (Komatsu et al. 2003)

Sprim m13 Z4r2dr Zd2nn

4Ar; nnB2r; nn ; 7

where the angular average is done on the full sky regardlessof sky cut, and m3 4

1 Rd2nnM3nn is the third-order

moment of the pixel-weighting function. When the weight isonly from a sky cut, as is the case here, we have m3 fsky,i.e., m3 is the fraction of the sky covered by observations(Komatsu et al. 2002). Komatsu et al. (2003) show that Bisa Wiener-filtered map of the underlying primordial fluctua-tions, . The other map A combines the bispectrum con-figurations that are sensitive to nonlinearity of the form inequation (1). Thus, Sprim is optimized for measuring theskewness of and picking out the quadratic term inequation (1).

Finally, the nonlinear coupling parameterfNL is given by

fNL Xmax

123

Bprim123

2

C1C2C3

" #1Sprim ; 8

where Bprim123 is the primordial bispectrum (Komatsu &Spergel 2001) multiplied by b1 b2 b3 and computed for

fNL 1 and the best-fit cosmological model. This equation is

used to measure fNL as a function of the maximum multipolemax. The statistic Sprim takes only N3=2 operations to com-pute without loss of sensitivity, whereas the full bispectrumanalysis takes N5=2 operations. It takes about 4 minutes on16 processors of an SGI Origin 300 to compute fNL from asky map at the highest resolution level, nside 512. Wemeasure fNL as a function ofmax. Since there is little CMBsignal compared with instrumental noise at > 512, we shalluse max 512 at most; thus, nside 256 is sufficient, speed-ing up evaluations of fNL by a factor of 8 as the computa-tional timescales as nside3. The computation takes only 30 sat nside 256. Note that since we are eventually fitting fortwo parameters, fNL and bsrc (see x 3), we include covariancebetween these two parameters in the analysis. The covariance

is, however, small (see Fig. 8 in Appendix A).While we use uniform weighting for Mnn, we couldinstead weight by the inverse noise variance per pixel,Mnn N1nn; however, this weighting scheme is subopti-mal at low where the CMB anisotropy dominates overnoise so that the uniform weighing is more appropriate. Formeasuring bsrc, on the other hand, we shall use a slightlymodified version of the N1 weighting, as bsrc comesmainly from small angular scales where instrumental noisedominates.

2.2.3. Method 2: The Minkowski Functionals

Topology offers another test for non-Gaussian features inthe maps, measuring morphological structures of fluctua-tion fields. The Minkowski functionals (Minkowski 1903;Gott et al. 1990; Schmalzing & Gorski 1998) describe theproperties of regions spatially bounded by a set of contours.The contours may be specified in terms of fixed temperaturethresholds, DT=, where is the standard deviation ofthe map, or in terms of the area. Parameterization of con-tours by threshold is computationally simpler, while param-eterization by area reduces correlations between theMinkowski functionals (Shandarin et al. 2002). We use a

joint analysis of the three Minkowski functionals [areaA, contour length C, and genus G] explicitlyincluding their covariance; consequently, we work in the

simpler threshold parameterization.The Minkowski functionals are additive for disjointregions on the sky and are invariant under coordinate trans-formation and rotation. We approximate each Minkowskifunctional using the set of equal-area pixels hotter or colderthan a set of fixed temperature thresholds. The fractionalarea

A 1

A

Xi

ai N

Ncut9

is thus the number of enclosed pixels, N, divided by thetotal number of pixels on the cut sky, Ncut. Here ai isthe area of an individual spot, and A is the total area of the

No. 1, 2003 WMAP: TESTS OF GAUSSIANITY 121


4/16

pixels outside the cut. The contour length

C 1

4A

Xi

Pi 10

is the total perimeter length of the enclosed regions Pi, whilethe genus

G 1

2A

Nhot Ncold 11

is the number of hot spots, Nhot, minus the number of coldspots, Ncold. We calibrate finite pixelization effects by com-paring the Minkowski functionals for the WMAP data toMonte Carlo simulations.

The WMAP data are a superposition of sky signal andinstrument noise, each with a different morphology. TheMinkowski functionals transform monotonically(although not linearly) between the limiting cases of asky signal with no noise and a noise map with no sky sig-nal. Unlike spatial analyses such as Fourier decomposi-tion, different regions of the sky cannot be weighted bythe signal-to-noise ratio, nor does the noise averagedown over many pixels. The choice of map pixelizationbecomes a trade-off between resolution (favoring smallerpixels) versus signal-to-noise ratio (favoring larger pixels).We compute the Minkowski functionals at nside 16through 256 (3072 to 786,432 pixels on the full sky). Weuse the WMAP Kp0 sky cut to reject pixels near theGalactic plane or contaminated by known sources. Thecut sky has 1433 pixels at resolution nside 16 and666,261 pixels at nside 256.

We compute the Minkowski functionals at 15 thresholdsfrom 3.5 to +3.5 and compare each functional to thesimulations using a goodness-of-fit statistic,

2

X12FiWMAP hF

isimi1

112

FiWMAP hFisimi2 ; 12

where FiWMAP is a Minkowski functional from the WMAPdata (the index i denotes a kind of functional), hFisimi isthe mean from the Monte Carlo simulations, and12 is thebin-to-bin covariance matrix from the simulations.

2.3. Monte Carlo Simulations

Monte Carlo simulations are used to estimate the statisti-cal significance of the non-Gaussian signals. One kind ofsimulation generates Gaussian random realizations ofCMB sky maps for the angular power spectrum, windowfunctions, and noise properties of the WMAP 1 yr data.This simulation quantifies the uncertainty arising fromGaussian fields, or the uncertainty in the absence ofnon-Gaussian fluctuations. The other kind generates non-Gaussian CMB sky maps from primordial fluctuations ofthe form of equation (1) (see Appendix A for our methodfor simulating non-Gaussian maps). This simulation quan-tifies the uncertainty more accurately and consistently in the

presence of non-Gaussian fluctuations.In principle, one should always use the non-Gaussian

simulations to characterize the uncertainty in fNL; how-ever, the uncertainty estimated from the Gaussianrealizations is good approximation to that from the non-Gaussian ones as long as fNLj j < 500. Our non-Gaussiansimulations verify that the distribution of fNL and bsrcaround the mean is the same for Gaussian and non-

Gaussian realizations (see Fig. 8 in Appendix A for anexample of fNL 100). The Gaussian simulations havethe advantage of being much faster than the non-Gaussian ones. The former takes only a few seconds tosimulate one map, whereas the latter takes 3 hours on asingle processor of an SGI Origin 300. Also, simulatingnon-Gaussian maps at nside 512 requires 17 GB ofphysical memory. We therefore use Gaussian simulationsto estimate the uncertainty in measured fNL and bsrc.

2.4. Limits to Primordial Non-Gaussianity

Figure 1 shows fNL measured from the Q+V+W co-added map using the cubic statistic (eq. [8]), as a function ofthe maximum multipole max. We find the best estimate of

fNL 38 48 (68%) for max 265. The distribution offNLis close to a Gaussian, as suggested by Monte Carlo simula-tions (see Fig. 8 in Appendix A). The 95% confidence inter-val is 58 < fNL < 134. There is no significant detection of

fNL at any angular scale. The rms error, estimated from 500Gaussian simulations, initially decreases as / 1max,althoughfNL for max 265 has a smaller error than that formax 512 because the latter is dominated by the instru-

mental noise. Since all the pixels outside the cut region areuniformly weighted, the inhomogeneous noise in the map(pixels on the ecliptic equator are noisier than those on thenorth and south poles) is not accounted for. This leads to anoisier estimator than a minimum variance estimator. Theconstraint on fNL for max 512 will improve with moreappropriate pixel-weighting schemes (Heavens 1998; Santoset al. 2002). The simple inverse noise (N1) weighting makesthe constraints much worse than the uniform weighting, asit increases errors on large angular scales where the CMBsignal dominates over the instrumental noise. The uniformweighting is thus closer to optimal. Note that for the powerspectrum, one can simply use the uniform weighting tomeasure C at small and the N1 weighting at large . For

Fig. 1.Nonlinear coupling parameter fNL as a function of the maxi-mum multipole max, measured from the Q+V+W co-added map using thecubic (bispectrum) estimator (eq. [8]). The best constraint is obtained frommax 265. The distribution is cumulative, so that the error bars at eachmax are not independent.



5/16

the bispectrum, however, this decomposition is not simple,as the bispectrum B123 measures the mode coupling from1 to 2 and 3 and vice versa. This property makes it difficultto use different weighting schemes on different angularscales. The first column of Table 1 shows fNL measured in

the Q, V, and W bands separately. There is no significantband-to-band variation, or a significant detection in anyband.

Figure 2 shows the Minkowski functionals atnside 128 (147,594 high-latitude pixels, each 280 indiameter). The gray band shows the 68% confidenceregion derived from 1000 Gaussian simulations. Table 2shows the 2-values (eq. [12]). The data are in excellentagreement with the Gaussian simulations at all resolu-

tions. The individual Minkowski functionals are highlycorrelated with each other (e.g., Shandarin et al. 2002).We account for this using a simultaneous analysis of allthree Minkowski functionals, replacing the 15 elementvectors FiWMAP; and hF

isim;i in equation (12) (the index i

TABLE 1

The Nonlinear Coupling Parameter, the Reduced

Point-SourceAngularBispectrum, and the

Point-Source Angular Power Spectrum

(Positive Definite) by Frequency Band

Band fNL

bsrc(105lK3 sr2)

csrc(103lK2 sr)

Q.......................... 51 61 9.5 4.4 15 6

V.......................... 42 63 1.1 1.6 4.5 4W......................... 37 75 0.28 1.3 . . .Q+V+W ............. 38 48 0.94 0.86 . . .

Notes.The error bars are68%. Thetabulated valuesare fortheKp0 mask, while the Kp2 maskgives similar results.

Fig. 2.Left panels show the Minkowski functionals for WMAP data(filled circles) at nside 128(280 pixels). Thegray band shows the68% con-fidence interval for the GaussianMonte Carlo simulations. The right panelsshow the residuals between the mean of the Gaussian simulations and theWMAP data. The WMAP data are in excellent agreement with theGaussian simulations.

TABLE 2

2 for Minkowski Functionals

nside

PixelDiameter

(deg)

Minkowski

Functional

WMAP

2 f(>WMAP)a

256 ............ 0.2 Genus 15.9 0.57128 ............ 0.5 Genus 10.7 0.79

64.............. 0.9 Genus 15.7 0.44

32.............. 1.8 Genus 18.7 0.2616.............. 3.7 Genus 16.8 0.22

256 ............ 0.2 Contour 9.9 0.93128 ............ 0.5 Contour 9.9 0.8364.............. 0.9 Contour 14.6 0.54

32.............. 1.8 Contour 12.8 0.5816.............. 3.7 Contour 11.9 0.67

256 ............ 0.2 Area 17.4 0.50

128 ............ 0.5 Area 10.9 0.7464.............. 0.9 Area 11.9 0.6632.............. 1.8 Area 21.9 0.12

16.............. 3.7 Area 15.7 0.33

Notes.

2

computed using Gaussian simulations. There are 15degrees of freedom.a Fraction of simulations with 2 greater than the value from the

WMAPdata.

Fig. 3.Limits to fNL from 2 fit of the WMAP data to the non-Gaussian models (eq. [1]). The fit is a joint analysis of the three Minkowskifunctionals at 280 pixel resolution. There are 44 degrees of freedom. TheMinkowski functionals show no evidence for non-Gaussian signals in theWMAPdata.



6/16

denotes each Minkowski functional) with 45 element vec-tors F F1; F2; F3 area, contour, genus and usingthe covariance of this larger vector as derived from thesimulations. We compute 2 for values fNL 0 to 1000,comparing the results from WMAP to similar 2-valuescomputed from non-Gaussian realizations. Figure 3shows the result. We find a best-fit value fNL 22 81(68%), with a 95% confidence upper limit fNL < 139, inagreement with the cubic statistic.

2.5. Implications of the WMAP Limits on fNL

2.5.1. Inflation

The limits on fNL are consistent with simple inflationmodels: models based on a slowly rolling scalar fieldtypically give fNLj j $ 102 101 (Salopek & Bond 1990,1991; Falk et al. 1993; Gangui et al. 1994; Acquaviva etal. 2002; Maldacena 2002), 3 to 4 orders of magnitudebelow our limits. Measuring fNL at this level is difficultbecause of the cosmic variance. There are alternativemodels which allow larger amplitudes of non-Gaussianity in the primordial fluctuations, which weexplore below.

A large fNL may be produced when the following condi-tion is met. Suppose that is given by x, where is a transfer function that converts x to andx x1 x2 Ox3 denotes a fluctuating fieldexpanded into a series of xi fixi1x1 with f1 1.Then, fNL 1f2. Inflation predicts the amplitude of xi

and the form offi, which eventually depends upon the scalarfield potential; thus, xi would be of order H=mplanck

i (His the Hubble parameter during inflation) for H < mplanck,and the leading-order term is H=mplanck $ 105. In thisway suppresses the amplitude of fluctuations, allowinga larger amplitude for H=mplanck $ 1051. What does thismean? If H $ 102mplanck, then $ 103 and fNL $ 103f2.The amplitude of f

NLis thus large enough to detect for

f2e0:1. This suppression factor, , seems necessary for oneto obtain a large fNL in the context of the slow-roll inflation.The suppression also helps us to avoid a fine-tuning prob-lem of inflation models, as it allows H=mplanck to be oforder slightly less than unity (which one might thinknatural) rather than forcing it to be of order 105.

Curvatons proposed by Lyth & Wands (2002) provide anexample of a suppression mechanism. A curvaton is a scalarfield, , having mass, m, that develops fluctuations, ,during inflation with its energy density, V, tinycompared to that of the inflation field that drives inflation.After inflation ends, radiation is produced as the inflationdecays, generating entropy perturbations between andradiation, S

=

3

4

=

. When H decreases to

become comparable to m, oscillations ofat the bottom ofV give m22. In the limit of cold inflation forwhich = is nearly zero, one finds S = 2= =2. As long as survives after the productionof S, the curvature perturbation is generated as 12 S x

1 12 x12, where x1 = (i.e., f2

12).

The generation of continues until decays, and is essen-tially determined by a ratio of to the total energy density,, at the time of the decay. Lyth et al. (2002) numericallyevolved perturbations to find 25 at the time of thedecay. The smaller the curvaton energy density is, the lessefficient the S to conversion becomes (or the more effi-cient the suppression becomes). The small thus leads to

the large fNL, asfNL 1f2 541 (i.e., fNL is always posi-

tive in this model). Assuming the curvaton exists and isentirely responsible for the observed CMB anisotropy, ourlimits on fNL imply > 9 103 at the time of the curva-ton decay. (However, the lower limit to does not meanthat we need the curvatons. This constraint makes senseonly when the curvaton exists and is entirely producing theobserved fluctuations.)

Features in an inflation potential can generate significantnon-Gaussian fluctuations (Kofman et al. 1991; Wang &Kamionkowski 2000), and it is expected that measurementsof non-Gaussianity can place constrains on a class of thefeature models. In Appendix C, we calculate the angularbispectrum from a sudden step in a potential of the form inequation (C2). This step is motivated by a class of super-gravity models yielding the steps as a consequence of succes-sive spontaneous symmetry-breaking phase transitions ofmany fields coupled to the inflaton (Adams, Ross, & Sarkar1997; Adams, Cresswell, & Easther 2001). One step generatestwo distinct regions in space where fNLj j is very large: a pos-itive fNL is predicted at < f, while a negative fNL at > f,where f is the projected location of the step. Our calculationssuggest that the two regions are separated in by less than afactor of 2, and one cannot resolve them without knowing f.The average of many modes further smears out the signals.The averaged fNL thus nearly cancels out to give only smallsignals, being hidden in our constraints in Figure 1. Peiriset al. (2003) argue that some sharp features in the WMAPangular power spectrum producing large 2-values may arisefrom features in the inflation potential. If this is true, thenone may be able to see non-Gaussian signals associated withthe features by measuring the bispectrum at the scales of thesharp features of the power spectrum.

2.5.2. Massive Cluster Abundance at High Redshift

Massive halos, like clusters of galaxies at high redshift,are such rare objects in the universe that their abundance issensitive to the presence of non-Gaussianity in the distribu-tion function of primordial density fluctuations. Severalauthors have pointed out the power of the halo abundanceas a tool for finding primordial non-Gaussianity (Lucchin &Matarrese 1988; Robinson & Baker 2000; Matarrese, Verde,& Jimenez 2000; Benson, Reichardt, & Kamionkowski2002); however, the power of this method is extremely sensi-tive to the accuracy of the mass determinations of halos. Itis necessary to go to redshifts of ze1 to obtain tight con-straints on primordial non-Gaussianity, as constraints fromlow and intermediate redshifts appear to be weak (Koyama,Soda, & Taruya 1999; Robinson, Gawiser, & Silk 2000) (seealso Figs. 4 and 5). Because of the difficulty of measuringthe mass of a high-redshift cluster the current constraintsare not yet conclusive (Willick 2000). The limited numberof clusters observed at high redshift also limits the currentsensitivity. In this section, we translate our constraints on

fNL from the WMAP 1 yr CMB data into the effects on themassive halos in the high-redshift universe, showing theextent to which future cluster surveys would see signaturesof non-Gaussian fluctuations.

We adopt the method of Matarrese et al. (2000) to calcu-late the dark-matter halo mass function dn=dM for a given

fNL, using the CDM with the running spectral index modelbest-fit to the WMAPdata and the large-scale structure data.This set of parameters is best suited for the calculations of the

124 KOMATSU ET AL.


7/16

Fig. 4.Limits to theeffectof theprimordial non-Gaussianityon thedark-matterhalo mass function dn=dMasa functionofz. Theshaded area representsthe95% constraint on theratio of thenon-Gaussian dn=dMto theGaussian one.

Fig. 5.Sameas Fig. 4 butfor thedark-matter halo numbercounts dN=dz as a function of thelimitingmass Mlim of a survey


8/16

cluster abundance. The parameters are in the rightmostcolumn of Table 8 of Spergel et al. (2003). We calculate

dn

dM 2

m0M dP=dMj j

; 13

where

m0 2:775 1011mh

2 M Mpc3

3:7 1010 M Mpc3

is the present-day mean mass density of the universe, PM; zis the probability for halos of mass Mto collapse at redshiftz,and dP=dMis given by

dP

dM

Z10

d

2

d2

dMsin

3

dl3dM

cos

e

22=2 ; 14

where the angleh is given by c 3l3=6, and cz isthe threshold overdensity of spherical collapse (Lacey & Cole1993; Nakamura & Suto 1997). The variance of the mass fluc-tuations as a function of z is given by 2M; z D2z2M; 0, where Dz is the growth factor of linear

density fluctuations,

2M; 0

Z10

dkk1F2MkD2k ;

D2k 221k3Pk

is the dimensionless power spectrum of the Bardeen curva-ture perturbations, FMk gkTkWkRM a filterfunction, gk 23 k=H0

21m0 is a conversion factor from

to density fluctuations, Tk is the transfer function of lineardensity perturbations, Wx 3j1x=x the spherical top-hat window smoothing density fields, and RM 3M=4m0

1=3 is the spherical top-hat radius enclosing amass M. The skewnessl3M; z D

3zl3M; 0, where

l3M; 0 6fNL

Z10

dk1

k1FMk1D

2k1

Z10

dk2

k2FMk2D

2k2

Z10

dlFM

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

k21 k22 2k1k2l

q ; 15

arises from the primordial non-Gaussianity. We use a MonteCarlo integration routine called vegas (Press et al. 1992)to evaluate the triple integral in equation (15). It follows fromequation (15) that a positivefNL gives a positivel3, positivelyskewed density fluctuations. Also this dn=dM reduces to thePress-Schechter form (Press & Schechter 1974) in the limit of

fNL ! 0. Although the Press-Schechter form predicts signifi-cantly fewer massive halos than N-body simulations (Jenkinset al. 2001), we assume that a predicted ratio of the non-Gaussian dn=dM to the Gaussian dn=dM is still reasonablyaccurate, as the primordial non-Gaussianity does not affectthe dynamics of halo formations which causes the differencebetween the Press-Schechter form ofdn=dM and the N-bodysimulations.

Figure 4 shows the WMAP constraints on the ratio ofnon-Gaussian dn=dM to the Gaussian one, as a function ofM and z. We find that the WMAP constraint on fNLstrongly limits the amplitude of changes in dn=dM due tothe non-Gaussianity. At z 0, dn=dM is changed by no

more than 20% even for 4 1015 M clusters. The numberof clusters that would be newly found at z 1 forM < 1015 M should be within

4010% of the value predicted

from the Gaussian theory. At z 3, however, much largereffects are still allowed: dn=dM can be increased by up to afactor of 2.5 for 2 1014 M.

Predictions for actual cluster surveys are made clearer bycomputing the source number counts as a function ofz,

dNdz

dVdz

Z1

Mlim

dM dndM

; 16

where Vz is the comoving volume per steradian, and Mlimis the limiting mass that a survey can reach. In practice Mlimwould depend on z due to, for example, the redshift dim-ming of X-ray surface brightness; however, a constant Mlimturns out to be a good approximation for surveys of theSunyaev-Zeldovich (SZ) effect (Carlstrom, Holder, & Reese2002). Figure 5 shows the ratio for dN=dz as a function ofzand Mlim. A source-detection sensitivity of Slim 0:5 Jyroughly corresponds to Mlim 1:4 1014 M (Carlstromet al. 2002), for which dN=dz should follow the prediction ofthe Gaussian theory out to z 1 to within 10%, but dN=dzat z 3 can be increased by up to a factor of 2. As Mlimincreases, the impact on dN=dz rapidly increases.

The SZ angular power spectrum CSZ is so sensitive to 8that we can use CSZ to measure 8 (Komatsu & Kitayama1999). The sensitivity arises largely from massive(M > 1014 M) clusters at z $ 1. From this fact one mightargue that CSZ is also sensitive to the primordial non-Gaussianity. We use a method of Komatsu & Seljak (2002)with dn=dM replaced by equation (13) to compute CSZ forthe WMAP limits on fNL. We find that C

SZ should follow

the prediction from the Gaussian theory to within 10% for100 < < 10000. This is consistent with CSZ being primar-ily sensitive to halos at z $ 1, where the effect on dN=dz isnot too strong (see Fig. 5). Since CSZ

/ 7

8bh

2 (Komatsu& Seljak 2002), 8 can be determined from CSZ to within 2%accuracy at a fixed bh using the Gaussian theory. The cur-rent theoretical uncertainty in the predictions of CSZ is afactor of 2 in CSZ (10% in 8), still much larger than theeffect of the non-Gaussianity.

3. LIMITS TO RESIDUAL POINT SOURCES

3.1. Point-Source Angular Power Spectrum and Bispectrum

Radio point sources distributed across the sky generatenon-Gaussian signals, giving a positive bispectrum, bsrc(Komatsu & Spergel 2001). In addition, the point sourcescontribute significantly to the angular spectrum on smallangular scales (Tegmark & Efstathiou 1996), contaminatingthe cosmological angular power spectrum. It is thus impor-tant to understand how much of the measured angular powerspectrum is due to sources. We constrain the source contribu-tion to the angular power spectrum, csrc, by measuring bsrc.Komatsu & Spergel (2001) have shown that WMAP candetect bsrc even after subtracting all (bright) sources detectedin the sky maps. Fortunately, there is no degeneracy between

fNL and bsrc, as shown later in Appendix A.In this section we measure the amplitude of non-

Gaussianity from residual point sources that are fainterthan a certain flux threshold, Sc, and left unmasked in thesky maps. The bispectrum bsrc is related to the number of



9/16

sources brighter than Sc per solid angle N> Sc:

bsrcSc

ZSc0

dSdN

dSgS3

N> ScgSc3

3

ZSc0

dS

SN> SgS3 ; 17

where g is a conversion factor from Jy sr1 to lK whichdepends on observing frequency as

g 24:76 Jy lK1 sr11sinh x=2=x22 ;

x h=kBT0 =56:78 GHz

for T0 2:725 K (Mather et al. 1999), and dN=dSis the dif-ferential source count per solid angle. The residual pointsources also contribute to the point-source power spectrumcsrc as

csrcSc

ZSc0

dSdN

dSgS2

N> ScgSc2

2

ZSc0

dS

SN> SgS2 : 18

By combining equation (17) and (18) we find a relationbetween bsrc and csrc,

csrcSc bsrcScgSc1

ZSc0

dS

SbsrcSgS

1 :

19

We can use this equation combined with the measured bsrc asa function ofSc to directly determine csrc as a function ofSc,

without relying on any extrapolations. When the sourcecounts obey a power law like dN=dS / S, one findsbsrcS / S4; thus, brighter sources contribute more to theintegral in equation (19) than fainter ones as long as > 3,which is the case for fluxes of interest. Bennett et al. (2003a)have found 2:6 0:2 for S 2 10 Jy in the Q band.Below 1 Jy, becomes even flatter (Toffolatti et al. 1998),implying that one does not have to go down to the very faintend to obtain reasonable estimates of the integral. In practice,we use equation (17) with N> S of the Toffolatti et al.(1998, hereafter T98) model at 44 GHz to computebsrcS < 0:5 Jy, inserting it into the integral to avoid missingfaint sources and underestimating the integral.

3.2. Measurement of the Point-Source Angular Bispectrum

The reduced point-source angular bispectrum, bsrc, ismeasured by a cubic statistic for point sources (Komatsuet al. 2003),

Sps m13

Zd2nn

4D3nn ; 20

where the filtered map Dnn is given by

Dnn Xmax2

Xm

b~CC

amYmnn : 21

This statistic is even quicker ($100 times) to compute than

Sprim (eq. [7]), as it involves only one integral over nn andonly one filtered map. This statistic also retains the samesensitivity to the point-source non-Gaussianity as the fullbispectrum analysis. The cubic statistic Sps gives bsrc as

bsrc 3

2

Xmax123

Bps123 2

C1C2C3

" #1Sps ; 22

whereB

ps

123 is the point-source bispectrum for bsrc 1(Komatsu & Spergel 2001) multiplied by b1 b2 b3 . While theuniform pixel-weighting outside the Galactic cut was usedforfNL, we use here Mnn 2CMB Nnn

1 where

2CMB 41X

2 1Cb2

is the variance of CMB aniostropy and Nnn is the varianceof noise per pixel which varies across the sky. This weightingscheme is nearly optimal for measuring bsrc as the signalcomes from smaller angular scales where noise dominates.The factor of 2CMB approximately takes into account thenonzero contribution to the variance from CMB aniso-tropy. This weight reduces uncertainties of bsrc by 17%,23%, and 31% in the Q, V, and W bands, respectively, com-pared to the uniform weighting. We use the highest resolu-tion level, nside 512, and integrate equation (22) up tomax 1024. In Appendix B, it is shown that this estimatoris optimal and unbiased as long as very bright sources,which have contributions to ~CC too large to ignore, aremasked. We cannot include csrc in the filter, as it is what weare trying to measure using bsrc.

The filled circles in the left panels of Figure 6 representbsrc measured in the Q (top panel) and V (bottom panel)bands. We have used source masks for various flux cuts, Sc,defined at 4.85 GHz to make these measurements. (Themasks are made from the GB6+PMN 5 GHz sourcecatalog.)

We find that bsrc increases as Sc: the brighter the sourcesunmasked, the more non-Gaussianity is detected. On theother hand, one can make predictions for bsrc using equation(17) for a given N> S. Comparing the measured values ofbsrc with the predicted values from N> S of T98 (dashedlines) at 44 GHz, one finds that the measured values aresmaller than the predicted values by a factor of 0.65. Thesolid lines show the predictions multiplied by 0.65. Botherrors in the T98 predictions and a nonflat energy spectrumof sources easily cause this factor. (If sources have a nonflatspectrum like S / , where 6 0, then Sc at the Q or Vband is different from that at 4.85 GHz.) Bennett et al.(2003a) find that the majority of the radio sources detectedin the Q band have a flat spectrum, 0:0 0:2. Our valuefor the correction factor matches well the one obtained fromthe WMAP source counts for 210 Jy in the Q band(Bennett et al. 2003a).

Equation (18) combined with the measured bsrc is used toestimate the point-source angular power spectrum csrc. Theright panels of Figure 6 show the estimated csrc as filledcircles. These estimates agree well with predictions fromequation (18) with N> S of T98 multiplied by a factorof 0.65 (solid lines). For Sc 1 Jy at the Q band,ccsrc 19 5 103 lK

2 sr and matches well the valueestimated from the WMAP source counts at the same fluxthreshold (Bennett et al. 2003a), which corresponds to thesolid lines in the figure. At V band, ccsrc 5 4



10/16

103 lK2 sr. Here the hat denotes that these values do notrepresent csrc for the standard source mask used by Hinshawet al. (2003b) for estimating the cosmological angular powerspectrum. Since the standard source mask is made of severalsource catalogs with different selection thresholds, it is diffi-cult to clearly identify a mask flux cut. We give the standardmask an effective flux cut threshold at 4.85 GHz by com-paring bsrc measured from the standard source mask (Fig. 6,shaded areas; see the second column of Table 1 for actualvalues) with those from the GB6+PMN masks defined at4.85 GHz. The measurements agree when Sc 0:75 Jy inthe Q band. Using this effective threshold, one expects csrcfor the standard source mask as csrc 15 6 103 lK2 sr in the Q band. This value agrees with the excesspower seen on small angular scales, 15:5 1:7 103 lK2 sr (Hinshaw et al. 2003b), as well as the valueextrapolated from the WMAPsource counts in the Q band,15:0 1:4 103 lK2 sr (Bennett et al. 2003a). In the Vband, csrc 4:5 4 103 lK

2 sr.The source number counts, angular power spectrum, and

bispectrum measure the first-, second-, and third-ordermoments of dN=dS, respectively. The good agreementamong these three different estimates of csrc indicates thevalidity of the estimate of the effects of the residual pointsources in the Q band. There is no visible contribution tothe angular power spectrum from the sources in the V andW bands. We conclude that our understanding of theamplitude of the residual point sources is satisfactory for

the analysis of the angular power spectrum not to becontaminated by the sources.

4. CONCLUSIONS

We use cubic (bispectrum) statistics and the Minkowskifunctionals to measure non-Gaussian fluctuations in theWMAP 1 yr sky maps. The cubic statistic (eq. [7]) and theMinkowski functionals place limits on the nonlinearcoupling parameter fNL, which characterizes the amplitudeof a quadratic term in the Bardeen curvature perturbations(eq. [1]). It is important to remove the best-fit foregroundtemplates from the WMAP maps in order to reduce thenon-Gaussian Galactic foreground emission. The cubicstatistic measures phase correlations of temperature fluctua-tions to find the best estimate of fNL from the foreground-removed, weighted average of Q+V+W maps as fNL 38 48 (68%) and 58 < fNL < 134 (95%). The Minkowskifunctions measure morphological structures to find

fNL 22 81 (68%) and fNL < 139 (95%), in good agree-ment with the cubic statistic. These two completely differentstatistics give consistent results, validating the robustness ofour limits. Our limits are 2030 times better than the pre-vious ones (Komatsu et al. 2002; Santos et al. 2002; Cayonet al. 2003) and constrain the relative contribution from thenonlinear term to the rms amplitude of to be smaller than2 105 (95%), much smaller than the limits on systematicerrors in the WMAP observations. This validates that the

Fig. 6.Point-source angular bispectrum bsrc and power spectrum csrc. The left panels show bsrc in the Q (top panel) and V bands (bottom panel). Theshaded areas show measurements from the WMAPskymaps with the standard sourcecut, while thefilledcircles show those with fluxthresholds Sc defined at4.85 GHz. Thedashed lines show predictions from eq. (17) with N> S modeled by Toffolattiet al. (1998), while the solid lines arethosemultiplied by 0.65 tomatch the WMAPmeasurements. The rightpanels show csrc. Thefilledcircles arecomputed from themeasured bsrc substituted into eq.(19). Thelinesare fromeq. (18). The errorbars are not independent,because the distribution is cumulative.



11/16

angular power spectrum can fully characterize statisticalproperties of the WMAPCMB sky maps. We conclude thatthe WMAP 1 yr data do not show evidence for significantprimordial non-Gaussianity of the form in equation (1).Our limits are consistent with predictions from inflationmodels based upon a slowly rolling scalar field,fNLj j 102 101. The span of all non-Gaussian models,

however, is large, and there are models which cannot beparameterized by equation (1) (e.g., Bernardeau & Uzan2002b, 2002a). Other forms such as multifield inflationmodels and topological defects will be tested in the future.

The non-Gaussianity also affects the dark-matter halomass function dn=dM, since the massive halos at high red-shift are sensitive to changes in the tail of the distributionfunction of density fluctuations. Our limits show that thenumber of clusters that would be newly found at z 1 forM < 1015 M should be within

4010% of the value predicted

from the Gaussian theory. At higher redshifts, however,much larger effects are still allowed. The number countsdN=dz at z 3 with the limiting mass of 3 1014 M canbe reduced by a factor of 2, or increased by more than a fac-tor of 3. Since the SZ angular power spectrum is primarilysensitive to massive halos at z $ 1, where the impact of non-Gaussianity is constrained to be within 10%, a measurementof8 from the SZ angular power spectrum is changed by nomore than 2%. Our results on dn=dM derived in this papershould be taken as the current observational limits to non-Gaussian effects on dn=dM. In other words, this is theuncertainty that we currently have in dn=dM when theassumption of Gaussian fluctuations is relaxed.

The limits on fNL will improve as the WMAP satelliteacquires more data. Monte Carlo simulations show that the4 yr data will achieve 95% limit of 80. This value will furtherimprove with a more proper pixel-weighting function thatbecomes the uniform weighting in the signal-dominatedregime (large angular scales) and becomes the N1 weight-ing in the noise-dominated regime (small angular scales).

There is little hope of testing the expected levels offNL 102 101 from simple inflation models, but somenonstandard models can be excluded.

We have detected non-Gaussian signals arising from theresidual radio point sources left unmasked at the Q band,

characterized by the reduced point-source angular bispec-trum bsrc 9:5 4:4 105 lK

3 sr2, which, in turn, givesthe point-source angular power spectrum csrc 15 6 103 lK2 sr. This value agrees well with those from thesource number counts (Bennett et al. 2003a) and the angularpower spectrum analysis (Hinshaw et al. 2003b), giving usconfidence on our understanding of the amplitude of theresidual point sources. Since bsrc directly measures csrc with-out relying on extrapolations, any CMB experiments thatsuffer from the point-source contamination should use bsrcto quantify csrc to obtain an improved estimate of the CMBangular power spectrum for the cosmological-parameterdeterminations.

Hinshaw et al. (2003b) found that the best-fit power spec-trum to the WMAPtemperature data has a relatively large2-value, corresponding to a chance probability of 3%.While still acceptable fit, there may be missing componentsin the error propagations over the Fisher matrix. Since theFisher matrix is the four-point function of the temperaturefluctuations, those missing components (e.g., gravitationallensing effects) may not be apparent in the bispectrum, thethree-point function. The point-source non-Gaussianitycontributes to the Fisher matrix by only a negligibleamount, as it is dominated by the Gaussian instrumentalnoise. Non-Gaussianity in the instrumental noise due to the1=f striping may have additional contributions to the Fishermatrix; however, since the Minkowski functionals, whichare sensitive to higher order moments of temperature fluctu-ations and instrumental noise, do not find significant non-Gaussian signals, non-Gaussianity in the instrumental noiseis constrained to be very small.

The WMAP mission is made possible by the support ofthe Office of Space Sciences at NASA Headquarters and bythe hard and capable work of scores of scientists, engineers,technicians, machinists, data analysts, budget analysts,managers, administrative staff, and reviewers. L. V. is sup-ported by NASA through Chandra Fellowship PF2-30022issued by the Chandra X-ray Observatory Center, which isoperated by the Smithsonian Astrophysical Observatory foran on behalf of NASA under contract NAS8-39073.

APPENDIX A

SIMULATING COSMIC MICROWAVE BACKGROUND SKY MAPS FROM PRIMORDIAL FLUCTUATIONS

In this Appendix, we describe how to simulate CMB sky maps from generic primordial fluctuations. As a specific example,we choose to use the primordial Bardeen curvature perturbations x, which generate CMB anisotropy at a given position ofthe skyDTnn as (Komatsu et al. 2003)

DTnn T0X

m

Ymnn

Zr2drmrr ; A1

where lmr is the harmonic transform ofx at a given comoving distance r xj j, mr R

d2nnr; nnYmnn, and rwas defined previously (eq. [5]). We can instead use isocurvature fluctuations or a mixture of the two. Equation (A1) suggeststhat r is a transfer function projecting x onto DTnn through the integral over the line of sight. Since r is just amathematical function, we precompute and store it for a given cosmology, reducing the computational time of a batch ofsimulations. We can thus use or extend equation (A1) to compute DTnn for generic primordial fluctuations.

We simulate CMB sky maps using a non-Gaussian model of the form in equation (1) as follows. (1) We generate ~Lk as aGaussian random field in Fourier space for a given initial power spectrum Pk and transform it back to real space to obtainLx. (2) We transform from Cartesian to spherical coordinates to obtainLr; nn, compute its harmonic coefficients mr,and obtain a temperature map of the Gaussian part DTnn by integrating equation (A1). (3) We repeat this procedure for2Lx V

1x Rd

3x2Lx to obtain a temperature map of the non-Gaussian part DT2 nn. (4) By combining these two



12/16

temperature maps, we obtain non-Gaussian sky maps for any values offNL,

DTnn DTnn fNLDT2 nn : A2

We do not need to run many simulations individually for different values of fNL, but run only twice to obtain DTnn andDT2 nn for a given initial random number seed. Also, we can combineDTnn for one seed withDT2 nn for the other to makerealizations for a particular kind of two-field inflation models. We can apply the same procedure to isocurvature fluctuationswith or withoutx correlations.

We need the simulation box of the size of the present-day cosmic horizon size Lbox 2c0, where 0 is the present-day

conformal time. For example, Lbox $ 20 h1

Gpc is needed for a flat universe with m 0:3, whereas we need spatialresolution of at least $20 h1 Mpc to resolve the last-scattering surface accurately. From this constraint the number of gridpoints is at least Ngrid 10243, and the required amount of physical memory to store x is at least 4.3 GB. Moreover, whenwe simulate a sky map having 786 432 pixels at nside 256, we need 1.6 GB to store a field in spherical coordinates r; nn,where the number of r evaluated for Ngrid 10243 is 512. Since our algorithm for transforming Cartesian into sphericalcoordinates requires another 1.6 GB, in total we need at least 7.5 GB of physical memory to simulate one sky map.

We have generated 300 realizations of non-Gaussian sky maps with Ngrid 10243 and nside 256. It takes 3 hours on oneprocessor of SGI Origin 300 to simulate DTnn and DT2 nn. We have used six processors to simulate 300 maps in one week.Figure 7 shows the one-point probability density function (PDF) of temperature fluctuations measured from simulatednon-Gaussian maps (without noise and beam smearing) compared with the rms scatter of Gaussian realizations. We find itdifficult for the PDF alone to distinguish non-Gaussian maps of fNLj j < 500 from Gaussian maps, whereas the cubic statisticSprim (eq. [8]) can easily detectfNL 100 in the same data sets.

We measure fNL on the simulated maps using Sprim to see if it can accurately recover fNL. Similar tests show the Minkowskifunctionals to be unbiased and able to discriminate different fNL values at levels consistent with the quoted uncertainties. We

also measure the point-source angular bispectrum bsrc to see if it returns null values as the simulations do not contain pointsources. We have included noise properties and window functions in the simulations. Figure 8 shows histograms of fNL andbsrc measured from 300 simulated maps of fNL 100 (solid lines) and fNL 0 (dashed lines). Our statistics find correct valuesfor fNL and find null values for bsrc; thus, our statistics are unbiased, and fNL and bsrc are orthogonal to each other as pointedout by Komatsu & Spergel (2001).

Fig. 7.One-point PDF of temperature fluctuations measured from simulated non-Gaussian maps (noise and beam smearing are not included). From thetop left to the bottom right panel the solid lines show the PDF for fNL 100, 500, 1000, and 3000, while the dashed lines enclose the rms scatter of Gaussianrealizations (i.e.,fNL 0).



13/16

APPENDIX B

POWER OF THE POINT-SOURCE BISPECTRUMIn this Appendix, we test our estimator for bsrc and csrc using simulated Q-band maps of point sources, CMB, and detector

noise. The 44 GHz source count model of T98 was used to generate the source populations. The total source count in eachrealization was fixed to 9043, the number predicted by T98 to lie between Smin 0:1 Jy and Smax 10 Jy. By generatinguniform deviates u 2 0; 1 and transforming to flux Svia

u N> Smin N> S

N> Smin N> Smax; B1

we obtain the desired spectrum. The sources were distributed evenly over the sky and convolved with a Gaussian profileapproximating the Q-band beam. Flux was converted to peak brightness using the values in Table 8 of Page et al. (2003b). TheCMB and noise realizations were not varied between realizations. The goal in this Appendix is to prove that our estimator forbsrc works well and is very powerful in estimating csrc.

The left panel of Figure 9 compares the measured bsrc from simulated maps with the expectations of the simulations. Black,dark gray, and light gray indicate three different realizations of point sources. The measurements agree well with theexpectations at Sc < 1:75 Jy. They, however, show significant scatter at Sc > 1:75 Jy, because our filter for computing bsrc (eq.[21]) does not include contribution from csrc to ~CC, making the filter less optimal in the limit of too many unmasked pointsources. We can see from the figure that csrc at Sc > 2 Jy is comparable to or larger than the noise power spectrum for the Qband, 54 103 lK2 sr.

Fortunately, this is not a problem in practice, as we can detect and mask those bright sources which contribute significantlyto ~CC. The residual point sources that we cannot detect (therefore we want to quantify using bsrc) should be hidden in the noisehaving only a small contribution to ~CC. In this faint-source regime bsrc works well in measuring the amplitude of residual pointsources, offering a promising way for estimating csrc. The right panel of Figure 9 compares csrc estimated from bsrc (eq. [17])with the expectations. The agreement is good for Sc < 1:75 Jy, proving that estimates of csrc from bsrc are unbiased andpowerful. Since bsrc measures csrc directly, we can use it for any CMB experiments that suffer from the effect of residual pointsources. While we have considered the bispectrum only here, the fourth-order moment may also be used to increase oursensitivity to the point-source non-Gaussianity (Pierpaoli 2003).

Fig. 8.Distribution of the nonlinear coupling parameter fNL (left panel) and the point-source bispectrum bsrc (right panel) measured from 300 simulatedrealizations of non-Gaussian maps for fNL 100 (solid line) and fNL 0 (dashed line). The simulations include noise properties and window functions of theWMAP1 yr data but do notinclude point sources.



14/16

APPENDIX C

THE ANGULAR BISPECTRUM FROM A POTENTIAL STEP

A scalar-field potential V with features can generate large non-Gaussian fluctuations in CMB by breaking the slow-rollconditions at the location of the features (Kofman et al. 1991; Wang & Kamionkowski 2000). We estimate the impact of thefeatures by using a scale-dependentfNL,

fNL 5

24G

@2 ln H

@2

; C1

which is calculated from a nonlinear transformation between the curvature perturbations in the comoving gauge and thescalar-field fluctuations in the spatially flat gauge (Salopek & Bond 1990, 1991). This expression does not assume the slow-rollconditions. Although this expression does not include all effects contributing to fNL during inflation driven by a single field(Maldacena 2002), we assume that an order-of-magnitude estimate can still be obtained.

A sharp feature in V at f produces a significantly scale-dependent fNL near f through the derivatives of H in

equation (C1). We illustrate the effects of the steps using the potential features proposed by Adams et al. (2001),

V 1

2m2

2 1 c tanh f

d

!; C2

which has a step in V at f with the height c and the slope d1. Adams et al. (1997) show that the steps are created by a classof supergravity models in which symmetry-breaking phase transitions of many fields in flat directions gravitationally coupledto continuously generate steps in V every 1015 e-folds, giving a chance for a step to exist within the observable region ofV.

It is instructive to evaluate equation (C1) combined with equation (C2) in the slow-roll limit, @2 ln H=@2 12 @2 ln V=@2.

For cj j5 1, one obtains

fNL 5

24G

1

2

c

d2tanh x

cosh2 x

; C3

Fig. 9.Testing the estimator for the reduced point-source bispectrum bsrc (eq. [22]). The left panel shows bsrc measured from a simulated map includingpoint sources and properties of the WMAPsky map at the Q band, as a function of flux cut Sc (filled circles). Black, dark gray, and light gray indicate threedifferent realizations of point sources. The solid line is the expectation from the input source number counts in the point-source simulation. The right panelcompares the power spectrum csrc estimated from bsrc with the expectation. The error bars are not independent, because the distribution is cumulative. Thebehaviorfor Sc > 2 Jy shows thecumulativeeffectof sources with brightness comparable to theinstrumentnoise (see text in AppendixB).



15/16

where x f=d. The first term corresponds to a standard, nearly scale-independent prediction giving 7:4 103 at 3mplanck , while the second term reveals a significant scale dependence. The function tanh x= cosh

2 x is a symmetric oddfunction about x 0 with extrema of0.385 at x 0:66. The picture is the following: as rolls down V from a positivex > 0:66, gets accelerated at x 0:66, reaches constant velocity at x 0, decelerates at x 0:66, and finally reaches slowroll at x < 0:66. The ratio of the second term in equation (C3) to the first at the extrema is 0:385c=d2. For example,c 0:02 and =d 300 (i.e., d 0:01mplanck) make the amplitude of the second term 700 times larger than the first, givingfNLj j 5 at the extrema. Despite the slow-roll conditions having a tendency to underestimate fNL, it is possible to obtainfNLj j > 1. Neglecting the first term in equation (C3) and converting for k, one obtains

fNLk 5c24Gd2

hstepk 5c24Gd2

tanh xkcosh2 xk

; C4

where xk d1@=@ln kfk=kf 1 d1 _=Hfk=kf 1 for k kf5kf. The slow-roll approximation gives

xk 4Gfd1k=kf 1. Finally, following the method of Komatsu & Spergel (2001), we obtain the reduced bispectrum

of a potential step model, bstep123 , as

bstep123

25c

24Gd2

Z10

r2dr

1 r2 rstep3

r 2 permutations

; C5

where r is given by equation (6), and

step r

2

Zk2dkhstepkgTkjkr : C6

The amplitude is thus proportional to c=d2

: a bigger (larger c) and steeper (smaller d) step gives a larger bispectrum. The steep-ness affects the amplitude more, because the non-Gaussianity is generated by breaking the slow-roll conditions.Since bstep123 linearly scales as c for a fixed d, we can fit for c by using exactly the same method as for the scale-independent

fNL, but with r in equation (3) replaced by step r. The exact form of the fitting parameter in the slow-roll limit is

5c=24Gd2. A reason for the similarity between the two models in methods for the measurement is explained as follows.Komatsu et al. (2003) have shown that Bnn; r (eq. [4]) is a Wiener-filtered, reconstructed map of the primordial fluctuationsnn; r. Our cubic statistic (eq. [7]) effectively measures the skewness of the reconstructed field, maximizing the sensitivity tothe primordial non-Gaussianity. One of the three maps comprising the cubic statistic is, however, not Bnn; r, but Ann; r givenby equation (3). This map defines what kind of non-Gaussianity to look for, or more detailed form of the bispectrum. For thepotential step case, Astepnn; r made of

step r picks up the location of the step to measure 5c=24Gd

2 near kf, while for theform in equation (1), Ann; r explores all scales on equal footing to measure the scale-independentfNL.

The distinct features in k space are often smeared out in space via the projection. This effect is estimated from equation(C6) as follows. The function hstepk near kf is accurately approximated by hstepk 0:385sin2xk, which has a period ofDk 42Gfdkf. On the other hand, the radiation transfer function gTk behaves as jkr, where r is the comoving

distance to the photon decoupling epoch, and gTkjkr behaves as j2 kr (the integral is very small when r 6 r). Theoscillation period of this part is thus Dk =r for kr > . A ratio of the period of hstepk to that of gTkjkr is then

estimated as 4Gfdrkf f=3d=0:01mplanckf=3mplanck, where f kfr is the angular wave number of the locationof the step. We thus find that hstepk oscillates much more slowly than the rest of the integrand in equation (C6) for f41.

What does it mean? It means that the results would look as if there were two distinct regions in space where fNL isvery large: a positive fNL is found at < f and a negative one at > f. The estimated location is=f 1 0:664Gfd 1 0:2d=0:01mplanckf=3mplanck; thus, the positive and negative regions are separated in byonly 40%, making the detection difficult when many modes are combined to improve the signal-to-noise ratio. The twoextrema would cancel out to give only small signals. In other words, it is still possible that non-Gaussianity from a potentialstep is hidden in our measurements shown in Figure 1. Note that the cancellation occurs because of the point symmetry ofhstepk about k kf. If the function has a knee instead of a step, then the cancellation does not occur and there would be asingle region in space where fNLj j is large (Wang & Kamionkowski 2000). Note that our estimate in this Appendix was basedupon equation (C3), which uses the slow-roll approximations. While instructive, since the slow-roll approximations breakdown near the features, our estimate may not be very accurate. One needs to integrate the equation of motion of the scalar field

to evaluate equation (C1) for more accurate estimations of the effect.

REFERENCES

Acquaviva, V., Bartolo, N., Matarrese, S., & Riotto, A. 2002, preprint(astro-ph/0209156)

Adams, J.,Cresswell, B.,& Easther, R. 2001, Phys. Rev. D, 64,123514Adams,J. A., Ross, G.G., & Sarkar, S.1997,Nucl. Phys. B, 503,405Aghanim, N.,Forni, O.,& Bouchet, F. R. 2001, A&A, 365, 341Banday, A. J.,Zaroubi,S., & Gorski, K. M. 2000,ApJ, 533, 575Bardeen, J. M. 1980, Phys. Rev. D, 22, 1882Barreiro, R. B., Hobson, M. P., Lasenby, A. N., Banday, A. J., Go rski,

K. M.,& Hinshaw, G. 2000, MNRAS,318,475Bennett,C. L., etal. 1996, ApJ,464, L1

. 2003a,ApJS, 148, 1. 2003b,ApJS, 148, 97Bennett,C. L., etal. 2003c,ApJ, 583,1Benson, A. J.,Reichardt, C.,& Kamionkowski, M. 2002, MNRAS, 331, 71

Bernardeau, F., & Uzan, J.-P. 2002a,Phys. Rev. D, 67,121301. 2002b, Phys. Rev. D, 66,103506Bromley, B. C.,& Tegmark, M. 1999, ApJ, 524, L79Bucher, M., & Zhu,Y. 1997, Phys. Rev.D, 55, 7415Carlstrom, J. E., Holder, G. P.,& Reese,E. D. 2002, ARA&A, 40,643Cayon, L., Martinez-Gonzalez, E., Argueso, F., Banday, A. J., & Gorski,

K. M. 2003, MNRAS, 339, 1189Contaldi, C. R., Ferreira, P. G., Magueijo, J., & Go rski, K. M. 2000, ApJ,

534, 25de Bernardis, P.,et al.2000,Nature,404, 955Falk, T.,Rangarajan, R.,& Srednicki, M. 1993, ApJ, 403, L1Ferreira,P. G.,Magueijo, J.,& Gorski, K.M. 1998, ApJ,503, L1Gangui, A., Lucchin, F., Matarrese, S., & Mollerach, S. 1994, ApJ, 430,

447



16/16

Gorski, K. M., Hivon, E., & Wandelt, B. D. 1998, in Evolutionof Large-Scale Structure: From Recombination to Garching ed. A. J.Banday, R. K. Sheth, & L. A. N. da Costa (Enschede: PrintPartners Ipskamp), 37

Gott, J. R. I., Park, C., Juszkiewicz, R., Bies, W. E., Bennett, D. P.,Bouchet, F. R.,& Stebbins,A. 1990,ApJ, 352, 1

Hanany, S.,et al.2000, ApJ, 545, L5Heavens, A. F. 1998, MNRAS, 299, 805Hinshaw, G. F.,et al.2003a, ApJS, 148, 63

. 2003b,ApJS, 148, 135Jenkins, A., Frenk,C. S., White, S. D. M., Colberg,J. M., Cole, S., Evrard,

A. E., Couchman,H. M. P.,& Yoshida, N. 2001,MNRAS, 321, 372

Kofman, L., Blumenthal, G. R., Hodges, H., & Primack, J. R. 1991, inASP Conf. Ser. 15, Large-Scale Structures and Peculiar Motions in theUniverse, ed. D. W. Latham & L. A. N. da Costa (San Francisco: ASP),339

Kogut, A., Banday, A. J., Bennett, C. L., Gorski, K. M., Hinshaw, G.,Smoot,G. F., & Wright, E.I. 1996, ApJ,464, L5

Komatsu, E. 2001, Ph.D. thesis, Tohoku Univ.Komatsu,E., & Kitayama, T. 1999,ApJ, 526, L1Komatsu,E., & Seljak, U. 2002, MNRAS, 336, 1256Komatsu,E., & Spergel,D. N. 2001, Phys. Rev. D, 63, 63002Komatsu, E., Spergel, D. N., & Wandelt, B. D. 2003, ApJ, submitted

(astro-ph/0305189)Komatsu, E., Wandelt, B. D., Spergel, D. N., Banday, A. J., & Go rski,

K.M. 2002, ApJ,566, 19Koyama, K.,Soda, J.,& Taruya, A. 1999, MNRAS, 310, 1111Kunz, M., Banday, A. J., Castro, P. G., Ferreira, P. G., & Go rski, K. M.

2001,ApJ, 563, L99Lacey,C., & Cole, S. 1993, MNRAS, 262, 627Linde,A., & Mukhanov,V. 1997,Phys.Rev.D, 56,535

Lucchin, F.,& Matarrese, S. 1988,ApJ, 330, 535Lyth, D. H.,Ungarelli,C., & Wands, D. 2002, Phys. Rev. D, 67,023503Lyth, D.H., & Wands,D. 2002, Phys. Lett. B, 524,5Magueijo, J. 2000, ApJ, 528, L57Maldacena, J. 2002, JHEP, 5, 13Matarrese, S.,Verde, L., & Jimenez, R. 2000, ApJ, 541, 10

Mather, J. C., Fixsen, D. J., Shafer, R. A., Mosier, C., & Wilkinson, D. T.1999, ApJ, 512, 511

Minkowski, H. 1903, Math. Ann., 57,447Mukherjee, P.,Hobson, M. P.,& Lasenby, A. N. 2000, MNRAS,318,1157Nakamura, T. T., & Suto,Y. 1997, Prog. Theor.Phys., 97, 49Novikov, D.,Schmalzing, J., & Mukhanov, V. F. 2000,A&A, 364, 17Page, L.,et al.2003a, ApJS,148, 39

. 2003b, ApJ, 585, 566Pando,J., Valls-Gabaud,D., & Fang, L. Z. 1998,Phys. Rev. Lett.,81, 4568Park, C.,Park, C.,Ratra, B., & Tegmark, M. 2001, ApJ, 556, 582Peebles,P. J.E. 1997, ApJ,483, L1Peiris, H.,et al.2003, ApJS,148, 213

Phillips,N. G.,& Kogut,A. 2001, ApJ, 548, 540Pierpaoli, E. 2003,ApJ, 589, 58Polenta, G.,et al.2002, ApJ, 572, L27Press,W. H.,& Schechter, P. 1974, ApJ, 187, 425Press, W. H., Teukolsky, S. A., Vetterling, W. T., & Flannery, B. P. 1992,

Numerical Recipes in C (2d ed.; Cambridge: CambridgeUniv. Press)Robinson,J., & Baker,J. E. 2000,MNRAS, 311, 781Robinson,J., Gawiser, E.,& Silk, J. 2000, ApJ, 532, 1Salopek,D. S., & Bond, J.R. 1990, Phys. Rev.D, 42, 3936

. 1991,Phys. Rev. D, 43,1005Sandvik,H. B.,& Magueijo, J. 2001, MNRAS, 325, 463Santos, M. G.,et al.2003,MNRAS, 341, 623Schmalzing, J.,& Gorski, K. M. 1998, MNRAS, 297, 355Seljak, U.,& Zaldarriaga,M. 1996,ApJ, 469, 437Shandarin, S. F., Feldman, H. A., Xu, Y., & Tegmark, M. 2002, ApJS,

141,1Spergel, D. N.,et al.2003, ApJS, 148, 175Tegmark, M.,& Efstathiou, G. 1996, MNRAS, 281, 1297Toffolatti, L., Argueso Gomez, F., de Zotti, G., Mazzei, P., Franceschini,

A.,Danese,L., & Burigana, C. 1998,MNRAS, 297, 117(T98)Wang, L.,& Kamionkowski, M. 2000, Phys.Rev. D, 61, 63504Willick, J. A. 2000, ApJ, 530, 80Wu,J. H.,et al.2001, Phys. Rev. Lett.,87, 251303Xu, Y., Tegmark, M., & de Oliveira-Costa, A. 2002, Phys. Rev. D, 65,

83002

134 KOMATSU ET AL.

Documents

E. Komatsu et al- First-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Tests of Gaussianity