1: \documentclass{emulateapj}

2: %\documentclass[12pt,preprint]{aastex}

3: 

4: %\usepackage{graphicx}

5: %\usepackage{float}

6: %\usepackage{amsmath}

7: %\usepackage{epsfig,floatflt}

8: 

9: 

10: 

11: \begin{document}

12: 

13: \title{Hemispherical power asymmetry in the

14:   three-year\\ Wilkinson Microwave Anisotropy Probe sky maps}

15: 

16: \author{H.\ K.\ Eriksen\altaffilmark{1,2,3,4}, A. J.

17:   Banday\altaffilmark{5}, K.\ M.\ G\'orski\altaffilmark{3,4,6}, F. K.

18:   Hansen\altaffilmark{1,2} and P.\ B.\ Lilje\altaffilmark{1,2}}

19: 

20: \email{h.k.k.eriksen@astro.uio.no}

21: 

22: \altaffiltext{1}{Institute of Theoretical Astrophysics, University of

23:   Oslo, P.O.\ Box 1029 Blindern, N-0315 Oslo, Norway}

24: 

25: \altaffiltext{2}{Centre of Mathematics for Applications, University of

26:   Oslo, P.O.\ Box 1053 Blindern, N-0316 Oslo, Norway}

27: 

28: \altaffiltext{3}{Jet Propulsion Laboratory, 4800 Oak Grove Drive,

29:   Pasadena CA 91109}

30: 

31: \altaffiltext{4}{California Institute of Technology, Pasadena, CA

32:   91125}

33: 

34: \altaffiltext{5}{Max-Planck-Institut f\"ur Astrophysik,

35:   Karl-Schwarzschild-Str.\ 1, Postfach 1317, D-85741 Garching bei

36:   M\"unchen, Germany}

37: 

38: \altaffiltext{6}{Warsaw University Observatory, Aleje Ujazdowskie 4,

39:   00-478 Warszawa, Poland}

40: 

41: %\date{Received - / Accepted -}

42: 

43: \begin{abstract}

44:   We consider the issue of hemispherical power asymmetry in the

45:   three-year WMAP data, adopting a previously introduced modulation

46:   framework. Computing both frequentist probabilities and Bayesian

47:   evidences, we find that the model consisiting of an isotropic CMB

48:   sky modulated by a dipole field, gives a substantially better fit to

49:   the observations than the purely isotropic model, even when

50:   accounting for the larger prior volume. For the ILC map, the

51:   Bayesian log-evidence difference is $\sim1.8$ in favour of the

52:   modulated model, and the raw improvement in maximum log-likelihood

53:   is 6.1. The best-fit modulation dipole axis points toward $(l,b) =

54:   (225^{\circ},-27^{\circ})$, and the modulation amplitude is 0.114,

55:   in excellent agreement with the results from the first-year

56:   analyses. The frequentist probability of obtaining such a high

57:   modulation amplitude in an isotropic universe is $\sim1$\%.  These

58:   results are not sensitive to data set or sky cut.  Thus, the

59:   statistical evidence for a power asymmetry anomaly is both

60:   substantial and robust, although not decisive, for the currently

61:   available data.  Increased sky coverage through better foreground

62:   handling and full-sky and high-sensitivity polarization maps may

63:   shed further light on this issue.

64: \end{abstract}

65: \keywords{cosmic microwave background --- cosmology: observations --- methods: statistical}

66: 

67: \section{Introduction}

68: \label{sec:introduction}

69: 

70: %The initial release of the first-year results from the Wilkinson

71: %Microwave Anisotropy Probe (WMAP) experiment \citep{bennett:2003} was

72: %a triumph for the currently popular inflationary concordance model.

73: %The first analyses unanimously agreed with a flat, isotropic and

74: %homogeneous universe seeded by Gaussian and adiabatic fluctuations, in

75: %excellent agreement with everybody's expectations.  However, in the

76: %months following the February 2003 release scientists all over the

77: %world subjected the raw data to a large number of more stringent

78: %tests, and quickly the picture became more complicated.

79: 

80: While the first-year results from the Wilkinson Microwave Anisotropy

81: Probe (WMAP) experiment \citep{bennett:2003} overall clearly supported

82: the currently popular inflationary cosmological model, describing a

83: flat, isotropic and homogeneous universe seeded by Gaussian and

84: adiabatic fluctuations, a disturbing number of unexpected anomalies on

85: large scales were reported shortly after the public data

86: release. Perhaps the three most important ones were: 1) alignments and

87: symmetry features among low $\ell$ multipoles \citep{de

88: Oliveira-Costa:2004, eriksen:2004a}, 2) an apparent asymmetry in the

89: distribution of fluctuation power in two opposing hemispheres

90: \citep{eriksen:2004b, hansen:2004}, and 3) a peculiar cold spot in the

91: southern hemisphere \citep{vielva:2004, cruz:2005}. All of these

92: features were subsequently studied extensively by independent groups,

93: and all remain unresolved to the present day.

94: 

95: In March 2006 the three-year WMAP results were released, prompting

96: researchers to revisit the anomalies detected in the first-year data

97: \citep{bridges:2006, copi:2006, jaffe:2006, land:2006,

98: martinez-gonzalez:2006}. Of course, considering that already the

99: first-year data were strongly signal-dominated on the scales of

100: interest, it should come as no surprise that most of these analyses

101: concluded with similar results as for the previous data, although

102: different foreground handling could affect some results.

103: 

104: The WMAP team paid particular attention to the question of large-scale

105: power asymmetry in their analyses \citep{hinshaw:2007, spergel:2007a}.

106: Specifically, in an early version of their paper,

107: \citet{spergel:2007a} approached the problem from a semi-Bayesian

108: point of view, by defining a parametric model consisting of an

109: isotropic and Gaussian CMB field modulated by a large-scale function.

110: The power asymmetry anomaly was then addressed by a dipolar modulation

111: field, and the low-$\ell$ alignment anomalies were studied with a

112: quadrupole modulation field. However, due to several issues with this

113: early analysis, several of which were first addressed by the present

114: paper, the authors decided to remove the corresponding section from

115: the final version of their paper \citep{spergel:2007b}.  One example

116: is simple marginalization over non-cosmological monopoles and dipoles

117: components, which was first done by \citet{gordon:2007} in an

118: otherwise identical analysis. A second example was the limited

119: harmonic range considered by \citet{spergel:2007a}. Thus, we present

120: in this Letter the first complete modulation analysis that covers the

121: full range of angular scales presented by \citet{eriksen:2004b}, and

122: that takes into account all known sources of systematics, such as

123: monopole/dipole and foreground marginalization. We also present the

124: first proper computation of the Bayesian evidence for the modulated

125: model.

126: 

127: Following the first report of the power asymmetry, much effort has

128: been spent by theorists on providing possible physical explanations.

129: Examples range from those questioning the very fundamentals of physics

130: and cosmology (e.g., introducing intrinsically inhomogeneous

131: cosmologies -- Moffat 2005 and Jaffe et al.\ 2005; violation of Lorenz

132: invariance -- Kanno \& Soda 2006; or violation of rotational

133: invariance in the very early universe -- Ackermann, Carroll \& Wise

134: 2007) to those essentially considering special cases of established

135: physics (e.g., second-order gravitational effects from local

136: inhomogeneities -- Tomita 2005; the presence of local voids -- Inoue

137: \& Silk 2006; spontaneous isotropy breaking from non-linear response

138: to long-wavelength density fluctuations -- Gordon et al.\

139: 2005). 

140: 

141: \section{Algorithms}

142: \label{sec:analysis}

143: 

144: We now outline the methods used for the analyses presented in the

145: following sections. 

146: 

147: \subsection{Data model and likelihood}

148: 

149: We model the CMB temperature sky maps as

150: \begin{equation}

151:   \mathbf{d}(\hat{\mathbf{n}}) = \mathbf{s}(\hat{\mathbf{n}}) [1 + f(\hat{\mathbf{n}})]

152:   + \mathbf{n}(\hat{\mathbf{n}}),

153: \end{equation}

154: where $\mathbf{s}(\hat{\mathbf{n}})$ is a statistically isotropic and

155: Gaussian random field with power spectrum $C_{\ell}$, $f(\hat{\mathbf{n}})$

156: is a dipole modulation field with amplitude less than unity, and

157: $\mathbf{n}(\hat{\mathbf{n}})$ is instrumental noise. Thus, the modulated signal

158: component is an anisotropic, but still Gaussian, random field, and

159: therefore has a covariance matrix given by

160: \begin{equation}

161: \tilde{\mathbf{S}}(\hat{\mathbf{n}}, \hat{\mathbf{m}}) = [1+f(\hat{\mathbf{n}})]

162: \mathbf{S}(\hat{\mathbf{n}}, \hat{\mathbf{m}})[1+f(\hat{\mathbf{m}})],

163: \end{equation}

164: where

165: \begin{equation}

166: \mathbf{S}(\hat{\mathbf{n}}, \hat{\mathbf{m}}) = \frac{1}{4\pi}\sum_{\ell}

167: (2\ell+1) C_{\ell} P_{\ell}(\hat{\mathbf{n}}\cdot\hat{\mathbf{m}}).

168: \end{equation}

169: 

170: Taking into account instrumental noise and possible foreground

171: contamination, the full covariance matrix is

172: \begin{equation}

173: \mathbf{C}(\hat{\mathbf{n}}, \hat{\mathbf{m}}) =

174: \tilde{\mathbf{S}}(\hat{\mathbf{n}}, \hat{\mathbf{m}}) + \mathbf{N} + \mathbf{F}.

175: \end{equation}

176: The noise and foreground covariance matrices depend on the data

177: processing, and will be described in greater detail in \S \ref{sec:data}. 

178: With these definitions ready at hand, the log-likelihood is given by

179: \begin{equation}

180: -2\log \mathcal{L} = \mathbf{d}^T \mathbf{C}^{-1} \mathbf{d} + \log |\mathbf{C}|,

181: \end{equation}

182: up to an irrelevant constant. 

183: 

184: \subsection{Posterior distributions and choice of parameters}

185: 

186: The posterior distribution $P(\theta|\mathbf{d})$ is a primary goal of

187: any Bayesian analysis, $\theta$ being the set of all free parameters

188: in the model. For the model defined above, the free parameters can be

189: divided into two groups, namely those describing the isotropic CMB

190: covariance matrix or $C_{\ell}$, and those describing the modulation

191: field. Both may be parametrized in a number of different ways, and

192: these choices may affect the outcome of the analysis through different

193: prior definitions.

194: 

195: First, for the isotropic CMB component, we choose to parametrize the

196: power spectrum in terms of a simple two-parameter model with free

197: amplitude $q$ and tilt $n$,

198: \begin{equation}

199: C_{\ell} = q \left(\frac{\ell}{\ell_0}\right)^{n} C_{\ell}^{\textrm{fid}}.

200: \end{equation}

201: Here $\ell_0$ is a pivot multipole and $C_{\ell}^{\textrm{fid}}$ is a

202: fiducial model, in the following chosen to be the best-fit power law

203: spectrum of \citet{hinshaw:2007}. Second, the modulation field

204: $f(\hat{n})$ is parametrized in terms of a direction

205: $\hat{\mathbf{p}}$ and an overall amplitude $A$,

206: \begin{equation}

207: f(\hat{\mathbf{n}}) = A \, \hat{\mathbf{n}} \cdot \hat{\mathbf{p}}.

208: \end{equation}

209: 

210: We use flat priors on all parameters in this paper; the modulation

211: axis is uniform over the sphere, and the amplitude is restricted to $A

212: \le 0.3$. The power spectrum parameters are restricted to $0.5 \le q

213: \le 1.5$ and $-0.5 \le n \le 0.5$. These choices are sufficiently

214: generous to include all non-zero parts of the likelihood.

215: 

216: The posterior distribution,

217: \begin{equation} 

218: P(q, n, A, \hat{\mathbf{p}} | \mathbf{d}) \propto \mathcal{L}(q, n, A,

219:   \hat{\mathbf{p}}) P(q, n, A, \hat{\mathbf{p}}),

220: \end{equation}

221: is then mapped out using a standard Markov Chain Monte Carlo

222: technique. We use a Gaussian proposal density for $q$, $n$, and $A$,

223: and an Euler-matrix based, uniform proposal density for

224: $\hat{\mathbf{p}}$.

225: 

226: \subsection{Bayesian evidence and nested sampling}

227: 

228: In a Bayesian analysis, one is not only interested in the set of

229: best-fit parameter values, but also in the relative probability of

230: competing models. The most direct way of measuring this is through the

231: Bayesian evidence,

232: \begin{equation}

233: E \equiv P(\mathbf{d}|H) = \int P(\mathbf{d}|\theta, H) P(\theta|H) d\theta,

234: \end{equation}

235: which is simply the average likelihood over the prior volume.

236: Typically, one computes this quantity for two competing models, $H_0$

237: and $H_1$, and considers the difference $\Delta \log E = \log E_1 -

238: \log E_0$. If $\Delta \log E > 1$, the evidence for $H_1$ is

239: considered substantial; if $\Delta \log E > 2.5$, it is considered

240: strong.

241: 

242: Traditionally, computation of evidences has been a computational

243: challenge. However, \citet{mukherjee:2006} introduced a method called

244: ``nested sampling'', proposed by \citet{skilling:2004}, to the

245: cosmological community that allows for accurate estimation of the

246: evidence through Monte Carlo sampling. We implemented this for the

247: priors and likelihood described above, and found that it works very

248: well for the problem under consideration.

249: 

250: \subsection{Maximum-likelihood analysis}

251: 

252: We also perform a standard frequentist maximum-likelihood analysis by

253: computing the maximum-likelihood modulation parameters for isotropic

254: Monte Carlo simulations. For these computations, we use a modified

255: version of the evidence code, which we find to be considerably more

256: robust than a simple non-linear search; while the non-linear search

257: algorithms often get trapped in local minima, the nested sampling

258: algorithm always find the correct solution, but of course, at a

259: considerably higher computational expense.

260: 

261: \section{Data}

262: \label{sec:data}

263: 

264: We analyze two versions of the three-year WMAP sky maps in the

265: following; the template-corrected $Q$-, $V$-, and $W$-band maps, and the

266: ``foreground cleaned'' Internal Linear Combination (ILC) map

267: \citep{hinshaw:2007}. All maps are processed as described by

268: \citet{eriksen:2006}: They are first downgraded to

269: HEALPix\footnote{http://healpix.jpl.nasa.gov} resolution

270: $N_{\textrm{side}}=16$, by additional smoothing to a $9^{\circ}$ FWHM

271: Gaussian beam and appropriate pixel window. Second, uniform Gaussian

272: noise of $\sigma_{\textrm{n}} = 1\,\mu\textrm{K}$ is added to each

273: pixel in order to regularize the pixel-pixel covariance matrix. This

274: combination of smoothing and noise level results in a signal-to-noise

275: ratio of unity at $\ell = 40$, and strong noise domination at the

276: Nyquist multipole of $\ell = 47$.

277: 

278: \begin{figure}

279: \plotone{f1.eps}

280: \caption{Posterior distributions for the dipole modulation amplitude,

281:   marginalized over direction and CMB power spectrum.}

282: \label{fig:amplitude}

283: \end{figure}

284: 

285: We use two different sky cuts for our analyses. First, given that the

286: galactic plane is clearly visible in the single frequency data, our

287: first mask is conservatively defined.  This cut is created by

288: expanding the Kp2 mask \citet{hinshaw:2007} by $9^{\circ}$ in all

289: directions, and then manually removing all near-galactic pixels for

290: which any difference map between two channels are clearly larger than

291: noise. In total, 36.3\% of all pixels are rejected by this cut (see

292: figure \ref{fig:modulation_maps}). Second, we also adopt the directly

293: downgraded Kp2 cut used by the WMAP team that removes 12.8\% of all

294: pixels. We use this mask for the ILC map only.

295: 

296: The noise covariance matrix is given by the uniform noise only,

297: $N_{ij} = \sigma_{\textrm{n}}^2 \delta_{ij}$. For completeness, we

298: have also computed the noise covariance from the smoothed instrumental

299: noise for the $V$-band data, but we find that this has no effect on

300: the final results, since its amplitude is far below the CMB signal. It

301: is therefore omitted in the following.

302: 

303: \begin{deluxetable}{lccccc}

304: \tabletypesize{\small}

305: \tablecaption{\label{tab:results}}

306: \tablecomments{Modulation model results. The listed quantities are the

307:   marginal best-fit dipole axis (\emph{second column}) and amplitude

308:   (\emph{third column}); the change in likelihood at the posterior

309:   maximum, $\Delta \log \mathcal{L} = \log \mathcal{L}_{\textrm{mod}}

310:   -\log \mathcal{L}_{\textrm{iso}} $, between the modulated and the

311:   isotropic model (\emph{fourth column}); the Bayesian

312:   evidence difference, $\Delta \log E = \log E_{\textrm{mod}} - \log

313:   E_{\textrm{iso}}$ (\emph{fifth column}); and the

314:   frequentist probability for obtaining a lower maximum-likelihood

315:   modulation amplitude than the observed one, computed from isotropic

316:   simulations (\emph{sixth column}).}

317: \tablecolumns{6}

318: \tablehead{Data  & ($l_{\textrm{bf}}, b_{\textrm{bf}}$) &

319:   $A_{\textrm{bf}}$ & $\Delta \log \mathcal{L}$ & $\Delta \log E$ & $P$}

320: \startdata

321: ILC\tablenotemark{a}  & ($225^{\circ}, -27^{\circ}$)  & $0.114$ & 6.1 & $1.8\pm0.2$ & 0.991 \\

322: ILC\tablenotemark{b}  & ($208^{\circ}, -27^{\circ}$)  & $0.125$ & 6.0 & $1.8\pm0.2$  & 0.991\\

323: $Q$-band\tablenotemark{b} & ($222^{\circ}, -35^{\circ}$)  & $0.124$ & 5.5 & $1.5\pm0.2$ & 0.987 \\

324: $V$-band\tablenotemark{b} & ($205^{\circ}, -19^{\circ}$)  & $0.127$ & 5.6 & $1.5\pm0.2$ & 0.990 \\

325: $W$-band\tablenotemark{b} & ($204^{\circ}, -31^{\circ}$)  & $0.121$ & 5.2 & $1.3\pm0.2$ & $\,\,\,\,\,\,$0.985 

326: %WMAP\tablenotemark{a,c} & ($204^{\circ}, -30^{\circ}$) & $0.103$ & 1.7 & N/A & $\quad$N/A   

327: 

328: %ILC\tablenotemark{a}  & ($-27^{\circ}, 225^{\circ}$)  & $0.114^{+0.031}_{-0.040}$ & 6.1 & $1.8\pm0.2$  \\

329: %ILC\tablenotemark{b}  & ($-27^{\circ}, 208^{\circ}$)  & $0.125^{+0.044}_{-0.048}$ & 6.0 & $1.8\pm0.2$  \\

330: %Q-band\tablenotemark{b} & ($-35^{\circ}, 222^{\circ}$)  & $0.124^{+0.045}_{-0.046}$ & 5.5 & $1.5\pm0.2$   \\

331: %V-band\tablenotemark{b} & ($-19^{\circ}, 205^{\circ}$)  & $0.127^{+0.045}_{-0.049}$ & 5.6 & $1.5\pm0.2$   \\

332: %W-band\tablenotemark{b} & ($-31^{\circ}, 204^{\circ}$)  & $0.121^{+0.044}_{-0.047}$ & 5.2 & $1.3\pm0.2$   \\

333: %Spergel et al\tablenotemark{a}. & ($-30^{\circ}, 204^{\circ}$) &

334: %$0.103\quad\quad\,\,\,\,$ & 1.7 & N/A   

335: \enddata

336: \tablenotetext{a}{Liberal 12.8\% sky cut imposed.}

337: \tablenotetext{b}{Conservative 36.3\% sky cut imposed.}

338: \end{deluxetable}

339: 

340: As an additional hedge against foreground contamination, we

341: marginalize over a set of fixed spatial templates, $\mathbf{t}_i$,

342: through the covariance matrix $\mathbf{F}_i =

343: \alpha_i\mathbf{t}_i\mathbf{t}_i^T$, $\alpha_i \gtrsim 10^3$. Monopole

344: and dipole terms are always included, and one or more foreground

345: templates. For the $V$-band and ILC maps, we follow

346: \citet{hinshaw:2007} and adopt $V$--ILC as our foreground template.

347: For the $Q$-band data, we marginalize over a synchrotron

348: \citep{haslam:1982}, a free-free \citep{finkbeiner:2003}, and a dust

349: \citep{finkbeiner:1999} template individually. Finally, for the

350: $W$-band data, we use the $W$--ILC difference map. However, we have

351: tried various combinations for all maps, and there is virtually no

352: sensitivity to the particular choice, or indeed, to the template at

353: all, due to the conservative sky cut used.

354: 

355: \section{Results}

356: \label{sec:results}

357: 

358: \begin{figure}

359: \plotone{f2.eps}

360: \caption{Posterior distribution for the dipole modulation

361:   axis, shown for the ILC map and 36.3\% sky cut, and marginalized

362:   over power spectrum and amplitude parameters. Grey sky pixels

363:   indicate pixels outside the $2\sigma$ confidence region. The dots

364:   indicate the axis 1) reported by Eriksen et al.\ (2004) in white; 2)

365:   for the ILC map with a 12.8\% sky cut in green; 3) for the $Q$-, $V$-,

366:   and $W$-bands in red, blue and yellow, respectively. The axis reported

367:   by \citet{spergel:2007a} coincides with the $W$-band axis.}

368: \label{fig:modulation_maps}

369: \end{figure}

370: 

371: The results from the analysis outlined above are summarized in Table

372: \ref{tab:results}. For each map, we report the best-fit dipole axis

373: and amplitude, as well as the maximum log-likelihood difference and

374: Bayesian evidence difference for the modulated versus the isotropic

375: model. The errors on the evidence are estimated by performing eight

376: independent analyses for each case, and computing the standard

377: deviation \citep{mukherjee:2006}. We also compute the probability of

378: obtaining a smaller modulation amplitude than the observed one by

379: analyzing 1000 isotropic Monte Carlo simulations.

380: 

381: Starting with the first case in Table \ref{tab:results}, the ILC map

382: cut by a 12.8\% mask, we see that the best-fit modulation axis points

383: toward $(l,b) = (225^{\circ}, -27^{\circ})$, and the corresponding

384: modulation amplitude is 0.114.  The raw likelihood improvement is

385: $\Delta \log \mathcal{L} = 6.1$. The probability of finding such a

386: high modulation amplitude in intrinsically isotropic simulations is

387: $\sim1$\%, and, finally, the improvement in Bayesian evidence is

388: $\Delta \log E = 1.8$. 

389: 

390: Further, these results are not sensitive to data set or sky coverage:

391: Even the Q-band map, which presumably is the least reliable with

392: respect to residual foregrounds, yields a modulation amplitude which

393: is high at the 98.7\% (frequentist) confidence level, and a Bayesian

394: log-evidence improvement of 1.5. This frequency independence is

395: further illustrated in Figure \ref{fig:amplitude}, where we show the

396: marginalized posterior distributions for the modulation amplitudes for

397: each data set. The agreement among data sets is very good.

398: 

399: In Figure \ref{fig:modulation_maps} we show the dipole axis posterior

400: distribution for the ILC map and 36.3\% sky cut.  Superimposed on

401: this, we have also marked the first-year asymmetry axis reported by

402: \citet{eriksen:2004b} [$(l,b) = (237^{\circ}, -10^{\circ})$] in

403: yellow, and also the other axes listed in Table \ref{tab:results}. All

404: agree well within $2\sigma$, and this is another testimony to the

405: excellent stability of the effect with respect to statistical method,

406: data set and overall procedure.

407: 

408: Finally, we note that this model may also partially explain the

409: anomalous cold spot reported by \citet{vielva:2004} and

410: \citet{cruz:2005}: By demodulation the spot would increase its

411: temperature by about 10\%, and although still very cold, it would be

412: significantly less extreme. Similar arguments could possibly also be

413: made for the Bianchi VII$_h$ correlation found by \citet{jaffe:2005}.

414: These issues will be considered further in future work.

415: 

416: \section{Conclusions}

417: \label{sec:conclusions}

418: 

419: A notable power asymmetry between two opposing hemispheres in the

420: first-year WMAP sky maps was reported by \citet{eriksen:2004b}. This

421: feature may be observed as strong fluctuations in the southern

422: ecliptic hemisphere, but virtually no large-scale structure in the

423: northern ecliptic hemisphere (e.g., Hinshaw et al.\ 2007).

424: 

425: In this Letter, we have revisited this issue in the three-year WMAP

426: data, adopting the statistical framework introduced and applied by

427: \citet{spergel:2007a}. With these tools, we find that the evidence for

428: power asymmetry in the WMAP data is very consistent with that

429: initially reported for the first-year maps by \citet{eriksen:2004b},

430: and the WMAP data clearly suggest a dipolar distribution of power on

431: the sky: The best-fit modulation amplitude is roughly 12\% in real

432: space, or about 20\% in terms of power spectra.  The corresponding

433: dipole direction is $(l,b) \sim (225^{\circ}, -27^{\circ})$. All

434: results are independent of data set choices, i.e., frequency channel

435: or sky cut.

436: 

437: However, the statistical evidence for this effect is still only

438: tentative. In frequentist language, the significance is about 99\%,

439: while in Bayesian terms, the log-evidence difference is $\sim 1.5$ to

440: 1.8, corresponding to odds of one to five or six. This

441: is quite comparable to the evidence for $n_{\textrm{s}} \ne 1$ after

442: the three-year WMAP data release, for which the odds are about one to

443: eight in the highest case (Parkinson et al.\ 2006).  Thus, there is

444: still a chance that the effect may be a fluke, and most likely, this

445: will remain the situation until Planck provides new data in some five

446: years. With additional frequency coverage, a better job can be done on

447: foreground treatment, and more sky coverage can be reliably included

448: in the analysis. Second, full-sky and high-sensitivity polarization

449: data should provide valuable insights on the origin of the effect.

450: 

451: \begin{acknowledgements}

452:   HKE acknowledges financial support from the Research Council of

453:   Norway.  Some of the results in this paper have been derived using

454:   the HEALPix (G\'orski et al.\ 1999) software and analysis package.

455:   We acknowledge use of the Legacy Archive for Microwave Background

456:   Data Analysis (LAMBDA). Support for LAMBDA is provided by the NASA

457:   Office of Space Science.  

458: \end{acknowledgements}

459: 

460: \begin{thebibliography}{}

461: 

462: \bibitem[Ackerman et al.(2007)]{ackerman:2007} 

463:   Ackerman, L., Carroll, S.~M., \& Wise, M.~B.\ 2007, [astro-ph/0701357]

464: 

465: \bibitem[Bennett et al.(2003)]{bennett:2003} Bennett, C. L., et al.\

466:   2003, \apjs, 148, 1

467: 

468: \bibitem[Bridges et al.(2006)]{bridges:2006} Bridges, M., McEwen, 

469: J.~D., Lasenby, A.~N., \& Hobson, M.~P.\ 2006, [astro-ph/0605325]

470: 

471: \bibitem[Copi et al.(2006)]{copi:2006} Copi, C.~J., Huterer, D., 

472: Schwarz, D.~J., \& Starkman, G.~D.\ 2006, [astro-ph/0605135]

473: 

474: \bibitem[Cruz et al.(2005)]{cruz:2005} Cruz, M., 

475: Mart{\'{\i}}nez-Gonz{\'a}lez, E., Vielva, P., \& Cay{\'o}n, L.\ 2005, 

476: \mnras, 356, 29 

477: 

478: \bibitem[de Oliveira-Costa et al.(2004)]{de Oliveira-Costa:2004}

479: de Oliveira-Costa, A., Tegmark, M., Zaldarriaga, M., \& Hamilton,

480: A. 2004, \prd, 69, 063516

481: 

482: \bibitem[Eriksen et al.(2004a)]{eriksen:2004a} Eriksen, H.~K., Banday, 

483: A.~J., G{\'o}rski, K.~M., \& Lilje, P.~B.\ 2004a, \apj, 612, 633 

484: 

485: \bibitem[Eriksen et al.(2004b)]{eriksen:2004b} Eriksen, H.~K., Hansen, 

486: F.~K., Banday, A.~J., G{\' o}rski, K.~M., \& Lilje, P.~B.\ 2004b, \apj, 605, 

487: 14 

488: 

489: \bibitem[Eriksen et al.(2006)]{eriksen:2006} Eriksen, H.~K., et al.\ 

490: 2006, ApJ, in press, [astro-ph/0606088]

491: 

492: \bibitem[Finkbeiner et al.(1999)]{finkbeiner:1999} % FDS models for thermal dust emission

493: Finkbeiner, D.P., Davis, M., \& Schlegel, D.J. 1999, ApJ, 524, 867

494: 

495: \bibitem[Finkbeiner(2003)]{finkbeiner:2003} 

496: Finkbeiner, D.~P.\ 2003, \apjs, 146, 407

497: 

498: \bibitem[Gordon et al.(2005)]{gordon:2005} Gordon, C., Hu, W.,

499:   Huterer, D., \& Crawford, T.\ 2005, \prd, 72, 103002

500: 

501: \bibitem[Gordon(2007)]{gordon:2007} Gordon, C.\ 2007, \apj, 656, 636

502: 

503: \bibitem[Hansen et al.(2004)]{hansen:2004} Hansen, F.~K., Banday, 

504: A.~J., \& G{\'o}rski, K.~M.\ 2004, \mnras, 354, 641 

505: 

506: \bibitem[Haslam et al.(1982)]{haslam:1982} % ubiquitous 408 MHz reference

507: Haslam, C.~G.~T., Salter, C.~J., Stoffel, H., \& Wilson, W.\ 1982,

508: \aaps, 47, 1 

509: 

510: \bibitem[Hinshaw et al.(2007)]{hinshaw:2007} Hinshaw, G., et al.\ 

511: 2007, ApJ, in press, [astro-ph/0603451]

512: 

513: \bibitem[Inoue \& Silk(2006)]{inoue:2006} Inoue, K.~T., \& Silk, 

514: J.\ 2006, \apj, 648, 23

515: 

516: \bibitem[Kanno \& Soda(2006)]{kanno:2006} Kanno, S., \& Soda, J.\ 

517: 2006, \prd, 74, 063505 

518: 

519: \bibitem[Jaffe et al.(2005)]{jaffe:2005} Jaffe, T.~R., Banday, 

520: A.~J., Eriksen, H.~K., G{\'o}rski, K.~M., \& Hansen, F.~K.\ 2005, \apjl, 

521: 629, L1 

522: 

523: \bibitem[Jaffe et al.(2006)]{jaffe:2006} Jaffe, T.~R., Banday, 

524: A.~J., Eriksen, H.~K., G{\'o}rski, K.~M., \& Hansen, F.~K.\ 2006, \aap, 

525: 460, 393 

526: 

527: \bibitem[Land \& Magueijo(2006)]{land:2006} Land, K., \& 

528: Magueijo, J.\ 2006, MNRAS, submitted, [astro-ph/0611518]

529: 

530: \bibitem[Mart{\'{\i}}nez-Gonz{\'a}lez et al.(2006)]{martinez-gonzalez:2006} 

531: Mart{\'{\i}}nez-Gonz{\'a}lez, E., Cruz, M., Cay{\'o}n, L., \& Vielva, P.\ 

532: 2006, New Astronomy Review, 50, 875 

533: 

534: \bibitem[Moffat(2005)]{moffat:2005} Moffat, J.~W.\ 2005, Journal of 

535: Cosmology and Astro-Particle Physics, 10, 12 

536: 

537: \bibitem[Mukherjee et al.(2006)]{mukherjee:2006} Mukherjee, P., 

538: Parkinson, D., \& Liddle, A.~R.\ 2006, \apjl, 638, L51 

539: 

540: \bibitem[Parkinson et al.(2006)]{parkinson:2006} Parkinson, D., 

541: Mukherjee, P., \& Liddle, A.~R.\ 2006, \prd, 73, 123523 

542: 

543: \bibitem[Skilling(2004)]{skilling:2004} 

544: Skilling, J. 2004, in AIP Conf.\ Proc.\ 735, Bayesian Inference and

545: Maximum Entropy Methods in Science and Engineering. ed.\ R. Fischer,

546: R. Preuss,\& U. von Toussaint (Melville: AIP), 395

547: 

548: \bibitem[Spergel et al.(2007a)]{spergel:2007a} 

549: Spergel, D.~N., et al.\ 2007a, [astro-ph/0603449]

550: 

551: \bibitem[Spergel et al.(2007b)]{spergel:2007b} 

552: Spergel, D.~N., et al.\ 2007b, ApJ, in press

553: 

554: \bibitem[Tomita(2005)]{tomita:2005} Tomita, K.\ 2005, \prd, 72, 

555: 043526 

556: 

557: \bibitem[Vielva et al.(2004)]{vielva:2004}

558: Vielva, P., Mart\'\i nez-Gonz\'alez, E., Barreiro, R. B., Sanz, J. L., \&

559: Cay\'on, L. 2004, \apj, 609, 22

560: 

561: \end{thebibliography}

562: 

563: \end{document}

564: