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

2: \documentclass[twocolumn]{emulateapj}

3: \usepackage{natbib}

4: \usepackage{epsfig}

5: \usepackage{bm}

6: 

7: %\usepackage{hyperref}

8: %\bibliographystyle{apj_hyperref}

9: \providecommand{\eprint}[1]{\href{http://arxiv.org/abs/#1}{#1}}

10: \newcommand{\ISBN}[1]{\href{http://cosmologist.info/ISBN/#1}{ISBN: #1}}

11: \providecommand{\adsurl}[1]{\href{#1}{}}

12: 

13: \begin{document}

14: 

15: \newcommand{\clee}{C_{\ell}^{EE}}

16: \newcommand{\xef}{x_e^{\rm fid}}

17: \newcommand{\xet}{x_e^{\rm true}}

18: \newcommand{\dz}{\Delta z}

19: \newcommand{\zmax}{z_{\rm max}}

20: \newcommand{\zmin}{z_{\rm min}}

21: \newcommand{\zmid}{z_{\rm mid}}

22: \newcommand{\lcdm}{$\Lambda$CDM}

23: \newcommand{\wmap}{\emph{WMAP}}

24: 

25: 

26: 

27: \title{Model-independent constraints on reionization from large-scale CMB polarization}

28: %\shorttitle{Model-independent constraints on reionization}

29: 

30: \author{Michael J. Mortonson$^{1,2}$ and Wayne Hu$^{1,3}$}

31: \affil{$^{1}$Kavli Institute for Cosmological Physics, 

32: Enrico Fermi Institute, University of Chicago, Chicago, IL 60637\\

33: $^{2}$Department of Physics,  University of Chicago, Chicago, IL 60637\\

34: $^{3}$Department of Astronomy and Astrophysics, University of Chicago, Chicago, IL 60637

35: }

36: 

37: 

38: \begin{abstract}

39: On large angular scales, the polarization of the CMB contains 

40: information about the evolution of the average ionization during the epoch

41: of reionization.   

42:  Interpretation of the polarization 

43: spectrum usually requires the assumption of a fixed functional form for the

44: evolution, e.g.\ instantaneous reionization.

45: We develop a model-independent method where a small set of

46: principal components 

47:  completely encapsulate the 

48: effects of reionization on the large-angle $E$-mode polarization

49: for any reionization history 

50: within an adjustable range in redshift. Using Markov Chain Monte Carlo 

51: methods, 

52: we apply this approach to both the 3-year \wmap\ data and 

53: simulated future data.  \wmap\ data constrain two principal components 

54: of the reionization history, approximately corresponding to the 

55: total optical depth and the difference between the contributions to the 

56: optical depth at high and low redshifts.  

57: The optical depth is consistent with the constraint found in previous 

58: analyses of \wmap\ data that assume instantaneous reionization, with 

59: only slightly larger uncertainty due to the expanded set of models. 

60: Using the principal component approach, \wmap\ data also place a 

61: 95\% CL upper limit of 0.08 on the contribution to the optical depth from 

62: redshifts $z>20$. 

63: With improvements in 

64: polarization sensitivity and foreground modeling,

65: approximately five of the principal components can ultimately be measured. 

66: Constraints on the principal components, which probe the entire reionization 

67: history, can test models of reionization, 

68: provide model-independent constraints on the optical depth, and 

69:  detect signatures of high-redshift reionization.

70: \end{abstract}

71: 

72: 

73: \keywords{cosmic microwave background --- cosmology: theory --- large-scale structure of universe}

74: 

75: 

76: % =====================================================

77: \section{Introduction}

78: \label{sec:intro}

79: 

80: 

81: The amplitude of the $E$-mode component of 

82: the cosmic microwave background (CMB)

83: polarization on large scales provides the current 

84: best constraint on the Thomson scattering optical depth to 

85: reionization, $\tau$.  

86: Using the first three years of data from the \emph{Wilkinson Microwave 

87: Anisotropy Probe} (\wmap) and making the simple assumption 

88: that the universe was reionized instantaneously, 

89: \cite{Speetal06} find $\tau = 0.09 \pm 0.03$. 

90: Theoretical studies suggest 

91: that the process of reionization 

92: was too complex to be well described 

93: as a sudden transition~\citep[e.g.,][]{BarLoe01}.  

94: Previous studies have examined how the constraint on $\tau$ depends on 

95: the evolution of the globally-averaged ionized fraction during 

96: reionization, $x_e(z)$, for a variety of specific theoretical scenarios. 

97: If the assumed form of $x_e(z)$ is incorrect, 

98: the estimated value of $\tau$ can be biased; this bias can be lessened by 

99: considering a wider variety of reionization histories at the expense of 

100: increasing the uncertainty in $\tau$~\citep{Kapetal03,Holetal03,Coletal05}.

101: 

102: The angular scales on which CMB polarization from reionization is correlated 

103: depend on the horizon size at the redshift of the free electrons: 

104: the higher the redshift, the higher the multipole, $\ell$ 

105: \citep[e.g.,][]{Zal97,HuWhi97a}. 

106: Varying $x_e(z)$ changes the relative contributions to the polarization 

107: coming from different redshifts, and therefore changes the shape of the 

108: large-scale $E$-mode angular power spectrum, $\clee$. 

109: Because of this dependence, measurements of the low-$\ell$ $E$-mode 

110: spectrum should place at least weak constraints on the global reionization 

111: history in addition to the constraint on the total optical depth. 

112: Recent studies suggest 

113: that \wmap\ data provide little information about $x_e(z)$ beyond $\tau$ 

114: \citep[e.g.,][]{LewWelBat06}, 

115: but it is worth asking what we can ultimately expect to 

116: learn about reionization from CMB polarization.

117: 

118: \citet{HuHol03} proposed using a principal component decomposition

119: of the ionization 

120: history to quantify the information contained in the large-scale $E$-mode 

121: polarization.  

122: The effect of any ionization history on the $E$-modes can be completely 

123: described by a small number of eigenmode parameters, unlike

124: a direct discretization of $x_{e}(z)$ in redshift bins.

125:  Here we extend the methods of \cite{HuHol03} 

126: using Markov Chain Monte Carlo techniques to find constraints on the 

127: principal components of $x_e(z)$ using both the 3-year \wmap\ observations 

128: and simulated future data.

129: 

130: Analytic studies and simulations indicate that reionization is an 

131: inhomogeneous process, and this inhomogeneity is expected to contribute 

132: to the small-scale CMB temperature and polarization anisotropies 

133: \citep[e.g.,][]{Hu00,Ilietal06a,MorHu07,Doretal07}. 

134: Here we focus on the large-scale $E$-modes only ($\ell \lesssim 100$) 

135: where such effects can be neglected, so we only consider  

136: the evolution of the globally-averaged ionized fraction as a function of 

137: redshift.

138: 

139: In the following section, we describe the principal component method 

140: for parameterizing the ionization history and show that the effects of 

141: $x_e(z)$ on large-scale $E$-mode polarization can be encapsulated in a 

142: small set of parameters. The method allows $x_e(z)$ to be a free 

143: function of redshift that is not bounded by physical considerations, so 

144: in \S~\ref{sec:phys} we derive limits that can be placed on the 

145: model parameters to eliminate most of the unphysical models where 

146: $x_e>1$ or $x_e<0$. We outline several ways to apply the principal 

147: component approach to constrain the reionization history with 

148: large-scale $E$-mode data in \S~\ref{sec:appl}. Using Markov Chain 

149: Monte Carlo methods, we examine some of these applications in more detail 

150: in \S~\ref{sec:mcmc} using both 3-year \wmap\ data and simulated 

151: future CMB polarization data. We summarize our findings and conclude 

152: in \S~\ref{sec:discuss}.

153: 

154: 

155: 

156: % =====================================================

157: \section{Ionization History Eigenmodes}

158: \label{sec:eigenmodes}

159: 

160: Models with similar total optical depth but different ionization histories

161: can produce markedly different predictions for the $E$-mode power spectrum.

162: In an instantaneous reionization scenario, the contribution to the optical 

163: depth from $x_e(z)$ is concentrated at the lowest redshifts possible 

164: for a given $\tau$, and the $E$-mode power spectrum for such a model is 

165: sharply peaked on large scales, at $\ell \lesssim 10$.  The main effect of 

166: shifting some portion of the reionization history to higher redshifts 

167: while keeping $\tau$ fixed is to reduce the $E$-mode power on the largest 

168: scales and increase it on smaller scales. 

169: 

170: This redistribution of power is illustrated in Figure~\ref{fig:cl} by the ionization histories 

171: and $\clee$ for two models, one with nearly instantaneous reionization 

172: and the other with comparable optical depth but with $x_e(z)$ 

173: concentrated at higher $z$.  In general, a flatter large-scale $E$-mode 

174: spectrum with power extending out to $\ell \sim 10$-20 is a sign of 

175: a large ionized fraction at high redshift.  However, there is not 

176: a one-to-one correspondence between $\ell$ and $z$; the ionized fraction 

177: within any particular narrow redshift bin affects $\clee$ over a 

178: wide range of angular scales.

179: Ionized fractions in adjacent redshift bins have highly correlated 

180: effects on $\clee$, which makes it difficult to extract from $E$-mode 

181: polarization data a constraint on $x_e$ at a specific redshift.

182: 

183: 

184: % ****************************************

185: \begin{figure}

186: \centerline{\psfig{file=f1.eps, width=3.0in}}

187: \caption{$E$-mode polarization angular power spectra and 

188: ionization histories (\emph{inset}) for a nearly-instantaneous 

189: reionization model with optical depth 

190: $\tau = 0.105$ (\emph{thick curves}) and an extended, 

191: ``double'' reionization history with $\tau = 0.090$ (\emph{thin}). 

192: Points with error bars represent the 3-year \wmap\ data from \cite{Pagetal06}.

193: \vskip 0.25cm}

194: \label{fig:cl}

195: \end{figure}

196: % ****************************************

197: 

198: 

199: As suggested by \cite{HuHol03}, one can use principal components of 

200: the reionization history as the model parameters instead of $x_e$ in 

201: redshift bins. These components are defined to have uncorrelated 

202: contributions to the $E$-mode power; since each has a unique effect on 

203: $\clee$, the amplitudes of the components can be inferred from 

204: measurements of the large-scale power spectrum. Principal component analysis 

205: of $x_e(z)$ also indicates which components can be determined best from 

206: the data. In this section, we describe the principal component method and 

207: introduce the notation that we will use throughout the paper.

208: 

209: Consider a binned ionization history $x_e(z_i)$, $i\in \{1,2,\ldots,N_z\}$, 

210: with redshift bins of width $\dz$ spanning $\zmin \leq z \leq \zmax$, where 

211: $z_1=z_{\rm min}+\dz$ and $z_{N_z}=z_{\rm max}-\dz$ so that 

212: $N_z+1=(z_{\rm max}-z_{\rm min})/\Delta z$. 

213: Throughout this paper we assume that the ionized fraction is 

214: $x_e\approx 1$ for redshifts $z\leq z_{\rm min}$ and $x_e\approx 0$ 

215: at $z\geq z_{\rm max}$. 

216: We take $z_{\rm min}=6$, consistent with observations of 

217: quasar spectra~\citep{FanCarKea06}. 

218: 

219: The principal components of $x_e(z_i)$ are eigenfunctions 

220: of the Fisher matrix, computed by taking derivatives of 

221:  $\clee$ with respect to $x_e(z_i)$:

222: \begin{equation}

223: F_{ij}=\sum_{\ell=2}^{\ell_{\rm max}}\left(\ell+\frac{1}{2}\right)

224:  \frac{\partial \ln \clee}{\partial x_e(z_i)}

225:  \frac{\partial \ln \clee}{\partial x_e(z_j)} \,,

226: \label{eq:fisher1}

227: \end{equation}

228: assuming full sky coverage and neglecting noise.

229: Since $x_e(z_i)$ only significantly contributes to 

230: the $E$-mode spectrum at small $\ell$, 

231: we typically truncate the sum in equation~(\ref{eq:fisher1}) 

232: at $\ell_{\rm max}=100$

233: where $\clee$ is dominated by the first acoustic peak.

234: The derivatives are evaluated 

235: at a fiducial reionization history,

236: $\xef(z_i)$. Following \cite{HuHol03} we typically choose 

237: $\xef(z_i)$ to be constant during reionization, although other 

238: functional forms  may be used.

239: 

240: Since the effects on $\clee$ of $x_e(z_i)$ in 

241:  adjacent redshift bins are highly 

242: correlated, the Fisher matrix contains large off-diagonal elements. 

243: The principal components $S_{\mu}(z_i)$ are the eigenfunctions of $F_{ij}$,

244: \begin{equation}

245: F_{ij}=(N_z+1)^{-2}\sum_{\mu=1}^{N_z}

246:  S_{\mu}(z_i)\sigma^{-2}_{\mu}S_{\mu}(z_j),

247: \label{eq:fisher2}

248: \end{equation}

249: where the factor $(N_z+1)^{-2}$ is included so that the eigenfunctions and 

250: their amplitudes have certain convenient properties. The inverse eigenvalues, 

251: $\sigma^2_{\mu}$, give the estimated variance 

252: of each principal component eigenmode from the measurement of low-$\ell$ 

253: $E$-modes. We order the modes so that the best-constrained 

254: principal components (smallest $\sigma^2_{\mu}$) have the lowest values of 

255: $\mu$, starting at $\mu=1$.

256: The noise level and other characteristics of an experiment 

257: can be included in the construction of the eigenfunctions, but since the 

258: effect on $S_{\mu}(z)$ is small we always use the noise-free 

259: eigenfunctions here.

260: 

261: The eigenfunctions satisfy the orthogonality and completeness

262: relations

263: \begin{eqnarray}

264: \int _{\zmin}^{\zmax} dz

265: ~S_{\mu}(z)S_{\nu}(z)&=&(z_{\rm max}-z_{\rm min})\delta_{\mu\nu}\,,

266: \label{eq:orthog1}\\ 

267: \sum_{\mu=1}^{N_z} S_{\mu}(z_i)S_{\mu}(z_j)&=& (N_z+1) \delta_{ij}\,.

268: \label{eq:orthog2}

269: \end{eqnarray}

270: The normalization 

271: of $S_{\mu}(z)$ is chosen so that the eigenfunctions 

272: are independent of bin width as $\dz\rightarrow 0$.

273: In equation (\ref{eq:orthog1})

274: and elsewhere in this paper where there are sums over redshift, 

275: we assume the continuous limit, 

276: replacing $\sum_i \Delta z$ by $\int dz$.

277: As long as the bin width 

278: is chosen to be sufficiently small, the final results we obtain are 

279: independent of the redshift binning. We adopt $\Delta z=0.25$ as the 

280: default bin width.

281: 

282: 

283: 

284: The three lowest-variance eigenfunctions for two different fiducial models 

285: are shown in Figure~\ref{fig:eigfn}. The lowest eigenmode ($\mu=1$) is an 

286: average of the ionized fraction over the entire redshift range, weighted at 

287: high $z$.  The $\mu=2$ mode can be thought of as a difference between 

288: the amount of ionization at high $z$ and at low $z$, and higher modes 

289: follow this pattern with weighted averages of $x_e(z)$ that oscillate with 

290: higher and higher frequency in redshift. Eigenfunctions of fiducial models 

291: with different values of $\zmax$ have similar shapes with the redshift 

292: axis rescaled according to the width of $(\zmax-\zmin)$.

293: The eigenfunctions are mostly insensitive to the choice of ionized 

294: fraction in the constant-$x_e$ fiducial histories.

295: 

296: % ****************************************

297: \begin{figure}

298: \centerline{\psfig{file=f2.eps, width=3.0in}}

299: \caption{Eigenfunctions for fiducial models with $\zmax=30$ and constant 

300: $x_e=0.15$ (\emph{top}), and $\zmax=20$ and $x_e=0.3$ (\emph{bottom}). 

301: In each case, the first three modes are shown: $\mu=1$ (\emph{thick}), 

302: 2 (\emph{medium}) and 3 (\emph{thin}). For both fiducial models, $\zmin=6$ 

303: and $\Delta z=0.25$.

304: \vskip 0.25cm}

305: \label{fig:eigfn}

306: \end{figure}

307: % ****************************************

308: 

309: 

310: 

311: An arbitrary reionization history can be represented in terms of the 

312: eigenfunctions as

313: \begin{equation}

314: x_e(z)=\xef(z)+\sum_{\mu}m_{\mu}S_{\mu}(z),

315: \label{eq:mmutoxe}

316: \end{equation}

317: where the amplitude of eigenmode $\mu$ for a perturbation 

318: $\delta x_e(z)\equiv x_e(z)-\xef(z)$ is

319: \begin{equation}

320: m_{\mu}=\frac{1}{\zmax-\zmin} 

321:   \int _{\zmin}^{\zmax} dz~S_{\mu}(z)\delta x_e(z).

322: \label{eq:xetommu}

323: \end{equation}

324: [Note that our conventions for the normalization of 

325: $m_{\mu}$ and $S_{\mu}(z)$ differ from those of \cite{HuHol03} by factors 

326: of $(N_z+1)^{1/2}$.] 

327: Any global ionization history $x_e(z)$ over the range $z_{\rm min}<z<\zmax$ 

328: is completely specified by a set of mode amplitudes $m_{\mu}$.

329: 

330: If perturbations to the fiducial history are small, $\delta x_e(z) \ll 1$, 

331: then the mode amplitudes are uncorrelated, with covariance matrix 

332: $\langle m_{\mu}m_{\nu}\rangle=\sigma_{\mu}^2 \delta_{\mu\nu}$. 

333: For a fixed fiducial model, however, arbitrary reionization histories 

334:  generally have $\delta x_e\sim 1$, in which case the amplitudes 

335: of different modes can become correlated as we discuss in \S~\ref{sec:fisher}.

336: 

337: 

338: The main advantage of using the principal component eigenmodes of $x_e(z)$ 

339: instead of some other parameterization is 

340: that most of the information relevant for large-scale $E$-modes is 

341: contained in the first few modes. This means that if one constructs $x_e(z)$ 

342: from equation~(\ref{eq:mmutoxe}) keeping only the first few terms in the sum 

343: over $\mu$, then the $E$-mode spectrum of the resulting ionization history 

344: will closely match that of $x_e(z)$ with all modes included in the sum. 

345: The effect of each eigenmode on $\clee$ becomes smaller as $\mu$ increases, 

346: as shown in Figure~\ref{fig:zmaxa} for the first two modes. 

347: \cite{HuHol03} demonstrated that for a specific fiducial ionization 

348: history $\xef(z)$ and assumed true history, 

349: only the first three modes of $x_e(z)$ are needed 

350: to produce $\clee$ indistinguishable from the true $E$-mode spectrum. 

351: 

352: % ****************************************

353: \begin{figure}

354: \centerline{\psfig{file=f3.eps, width=3.0in}}

355: \caption{Change in $\ln \clee$ per unit $m_{\mu}$

356: for fiducial $x_e(z)$ with $\zmax=30$ (\emph{solid}) and 20 (\emph{dashed}), 

357: showing the effect on $\clee$ of the first two principal components, 

358: $\mu=1$ (\emph{thick}) and $2$ (\emph{thin}). 

359: \vskip 0.25cm

360: }

361: \label{fig:zmaxa}

362: \end{figure}

363: % ****************************************

364: 

365: 

366: The ionization histories and corresponding $E$-mode spectra in 

367: Figure~\ref{fig:clcomp} demonstrate this completeness for a fairly 

368: extreme model in which the first ten eigenmodes all have 

369: significant amplitudes.  Even in the simplest case where a single eigenmode 

370: is used in place of the original $x_e(z)$, the error in $\clee$ is 

371: only $\sim 10\%$.  With 3-5 modes, the error is a few percent or less at

372: all multipoles and safely smaller than the cosmic variance of 

373: \begin{equation}

374: {\Delta \clee \over \clee} = \sqrt{2 \over 2 \ell +1} \,.

375: \end{equation}

376: The top panel of Figure~\ref{fig:clcomp} shows that this completeness 

377: does not extend to the ionization history itself: $x_e(z)$ constructed from 

378: as many as five eigenmodes is a poor approximation to the full 

379: reionization history.

380: 

381: % ****************************************

382: \begin{figure}

383: \centerline{\psfig{file=f4.eps, width=3.0in}}

384: \caption{Test of completeness in $\clee$ for ionization histories with 

385: truncated sets of eigenfunctions. The base history (\emph{top panel, dashed}) 

386: is the sum of the first 10 eigenfunctions 

387: with comparable values of $|m_{\mu}|$ for 

388: each mode. Solid curves show $x_e(z)$ (\emph{top}) and the error in $\clee$ 

389: (\emph{bottom}) for histories that retain the first one (\emph{thin curves}), 

390: three (\emph{medium}), and five (\emph{thick}) eigenmodes. 

391: The fiducial model used for the eigenmode decomposition has constant 

392: $x_e=0.15$ from $\zmin=6$ up to $\zmax=30$.

393: \vskip 0.25cm}

394: \label{fig:clcomp}

395: \end{figure}

396: % ****************************************

397: 

398: 

399: 

400: From this and other similar tests on the completeness in $\clee$ 

401: of the lowest-variance eigenmodes we conclude that the first 3-5 modes contain 

402: essentially all of the information about the reionization history that 

403: is relevant for  large-scale $E$-mode polarization.

404: This fact is particularly useful for constraining the global 

405: reionization history with CMB polarization data using 

406: Markov Chain Monte Carlo techniques. Since the number of parameters that 

407: must be added to a Monte Carlo chain  

408: to describe an arbitrary $x_e(z)$ is relatively small, we can obtain 

409: constraints from the data that are independent of assumptions about 

410: the reionization history with minimal added computational 

411: expense~\citep{HuHol03}.

412: The exact number of modes required varies depending on the true 

413: ionization history and the fiducial history, so when analyzing data it is 

414: a good idea to check that the results do not change significantly when 

415: the next modes are included in the sum in equation~(\ref{eq:mmutoxe}).

416: 

417: 

418: The main caveat to this model independence is that in practice we must 

419: set some maximum redshift for histories in any particular chain of 

420: Monte Carlo samples, ignoring any 

421: contribution to the observed low-$\ell$ $\clee$ from ionization at $z>\zmax$. 

422: Since the eigenfunctions of the ionization history are stretched in redshift 

423: as $\zmax$ increases (Fig.~\ref{fig:eigfn}), 

424: at higher $\zmax$ more modes are needed to accurately 

425: represent any particular feature in $x_e(z)$. For example, take the true 

426: ionization history to be instantaneous reionization at $z=11.5$. 

427: Figure~\ref{fig:zmaxb} shows the error in $\clee$ if we truncate the 

428: eigenmode sum of equation~(\ref{eq:mmutoxe}) 

429: at $N$ modes using fiducial histories 

430: with $\zmax=30$ and $\zmax=20$. For each fiducial model, the error decreases 

431: as the number of modes in the sum increases, but the error at fixed $N$ is 

432: larger for $\zmax=30$ than $\zmax=20$. 

433: The requirement of retaining a larger set of parameters as $\zmax$ 

434: increases makes it less practical to study models with significant 

435: reionization at extremely high redshifts 

436: [$\zmax \gtrsim 100$; e.g., \cite{NasChi04,KasKawSug04}], but even for 

437: $\zmax$ as high as $\sim 40$ the number of eigenmodes needed 

438: is reasonably small ($N\lesssim 5$).

439: 

440: 

441: % ****************************************

442: \begin{figure}

443: \centerline{\psfig{file=f5.eps, width=3.0in}}

444: \caption{Error in $\clee$ from $x_e(z)$ constructed 

445: from the first $N$ eigenmodes of fiducial models with $\zmax=30$ 

446: (\emph{solid curves}) and 20 (\emph{dashed}), taking $N=3$ (\emph{thin}) 

447: and $N=5$ (\emph{thick}). The true $x_e(z)$ and $\clee$ 

448: are assumed to be those of an instantaneous reionization model with $\tau=0.09$.

449: \vskip 0.25cm}

450: \label{fig:zmaxb}

451: \end{figure}

452: % ****************************************

453: 

454: 

455: While theories of reionization provide useful priors on $\zmax$ in 

456: the context of specific models, it would be better to be able to constrain 

457: $\zmax$ empirically by measuring the $E$-mode power accurately up 

458: to $\ell \sim 100$. 

459: The current 3-year \wmap\ polarization data have 

460: high enough signal-to-noise to be useful for parameter constraints only at 

461: $\ell < 24$ \citep{Pagetal06}, so their sensitivity to 

462: high-$z$ reionization is limited.  

463: Given that ionization at some redshift generates polarization out

464: to a maximum $\ell$, 

465:  \wmap\ data can still place weak bounds on the total optical depth 

466: contribution above a certain redshift, even if it cannot distinguish whether

467: these contributions arise from redshifts above a chosen $\zmax$

468: as we discuss in \S~\ref{sec:wmap}.

469: The values we choose here for $\zmax$ are 

470: partly influenced by the fact that few theoretical 

471: reionization scenarios predict an ionized 

472: fraction at $z\gtrsim 30$ that would significantly affect large-scale $\clee$. 

473: Future data should better constrain $\clee$ at higher multipoles, 

474: allowing useful limits to be placed on $\zmax$ from the data alone. 

475: 

476: 

477: The need to consider a limited range in redshift is shared by other 

478: methods, for example those that constrain binned $x_e(z_i)$ instead of 

479: the eigenmode amplitudes \citep{LewWelBat06}. As already mentioned, the 

480: principal component approach has the unique advantages that results can be

481: made independent of bin width and that relatively few extra parameters 

482: are required.  However, this approach also has a unique difficulty in that 

483: the physicality of the reionization history 

484: [$0\leq x_e(z)\leq 1$] is not built in to the method.   Hence constraints derived

485: from measurements of the eigenmodes can be weaker than those from 

486: a method that enforces physicality.

487: We can, however, place some prior constraints on the set of eigenmode 

488: amplitudes that must be satisfied by any physical model, as we discuss in 

489: the next section.

490: 

491: 

492: % =====================================================

493: \section{Priors from Physicality}

494: \label{sec:phys}

495: 

496: Although the actual ionized fraction must be between 0 and 1 (neglecting 

497: helium reionization  and the small residual ionized fraction after 

498: recombination), there is nothing in the construction of $x_e(z)$ from 

499: the eigenmodes in equation~(\ref{eq:mmutoxe}) that ensures that 

500: the ionized fraction will 

501: obey these limits. Whether or not $x_e$ has a physical value at a 

502: particular redshift depends on the amplitudes of all of the 

503: principal components. 

504: Even for a physical ionization history, 

505:  the truncated sum up to mode $N$, 

506: \begin{equation}

507: x_e^{(N)}(z)\equiv \xef(z)+\sum_{\mu=1}^N m_{\mu} S_{\mu}(z),

508: \label{eq:trunc}

509: \end{equation}

510: is not necessarily bounded by 0 and 1 at all redshifts. 

511: While formally it is possible to evaluate $\clee$ for reionization 

512: histories with unphysical values of $x_e$, we would like to 

513: eliminate as much as possible those models for which 

514: the full sum of the eigenmodes, 

515: $x_e(z)=\lim_{N\to\infty}x_e^{(N)}(z)$, is unphysical.

516: 

517: 

518: We find the largest and smallest values of 

519: $m_{\mu}$ that are consistent with $x_e(z)\in [0,1]$ for all $z$ using 

520: the definition of $m_{\mu}$ in equation~(\ref{eq:xetommu}). 

521: We are free to choose $\xef(z)\in [0,1]$ so that  

522:  $-\xef(z)\leq\delta x_e(z)\leq 1-\xef(z)$, where the lower limit 

523: is strictly negative or zero and the upper limit is positive or zero. 

524: For a particular mode $\mu$, 

525: the choice of $\delta x_e(z)$ that maximizes $m_{\mu}$ is

526: \begin{equation}

527: \delta x_e^{\rm max}(z)=\left\{ \begin{array}{rl}

528: -\xef(z), & S_{\mu}(z)\leq 0 \\

529: 1-\xef(z), & S_{\mu}(z)>0.

530: \end{array}\right.

531: \label{eq:dxemax}

532: \end{equation}

533: Using this in equation~(\ref{eq:xetommu}) gives an upper limit on $m_{\mu}$. 

534: Similarly, a lower limit can be obtained by reversing the signs of the 

535: inequalities in equation~(\ref{eq:dxemax}).  

536: The resulting physicality bounds  

537: are $m_{\mu}^{(-)}\leq m_{\mu}\leq m_{\mu}^{(+)}$, where

538: \begin{equation}

539: m_{\mu}^{(\pm)} = \int _{\zmin}^{\zmax} dz 

540: \frac{S_{\mu}(z)[1-2\xef(z)]\pm |S_{\mu}(z)|}{2(\zmax-\zmin)}.

541: \label{eq:physbounds}

542: \end{equation}

543: If $m_{\mu}$ violates these bounds for any $\mu$, the reionization history 

544:  is guaranteed to be unphysical for some 

545: range in redshift. The opposite is not true, however: even if all $m_{\mu}$ 

546: satisfy equation~(\ref{eq:physbounds}), $x_e(z)$ may still be unphysical 

547: for some $z$.

548: 

549: The parameter space that physical models may occupy is restricted 

550: further by an inequality that must be satisfied by all eigenmodes 

551: simultaneously. Assume for simplicity that the fiducial model has 

552: a constant ionized fraction, $\xef\in [0,1]$, for $\zmin<z<\zmax$. 

553: Any physical reionization history $x_e(z)$ must satisfy

554: \begin{equation}

555: \int _{\zmin}^{\zmax} dz [x_e(z)-\xef]^2 \leq (\zmax-\zmin) f,

556: \label{eq:phys2}

557: \end{equation}

558: where $f\equiv \max[(\xef)^2,(1-\xef)^2]$.

559: Using equation~(\ref{eq:mmutoxe}), the left side of the inequality can also be 

560: written

561: \begin{eqnarray}

562: \int _{\zmin}^{\zmax} dz [x_e(z)-\xef]^2&=&\int _{\zmin}^{\zmax} dz

563:  \left[\sum_{\mu} m_{\mu} S_{\mu}(z) \right]^2 \nonumber\\

564: &=& (\zmax-\zmin) \sum_{\mu} m_{\mu}^2 \,,

565: \label{eq:phys3}

566: \end{eqnarray}

567: where the second line follows from the orthogonality of the eigenfunctions 

568: (eq.~[\ref{eq:orthog1}]). Comparing equations~(\ref{eq:phys2}) 

569: and~(\ref{eq:phys3}) we obtain a constraint on the sum of the 

570: squares of the mode amplitudes,

571: \begin{equation}

572: \sum_{\mu} m_{\mu}^2 \leq f,

573: \label{eq:physbounds2}

574: \end{equation}

575: where $0.25\leq f\leq 1$, depending on the value of $\xef$. 

576: As with the physicality bounds of equation~(\ref{eq:physbounds}), 

577: this 

578: upper limit is a necessary but not sufficient condition for physicality.

579: 

580: 

581: Since in practice we can only constrain a limited set of eigenmodes, 

582: the uncertainty in modes higher than the first few prevents us from 

583: simply excluding all models where $x_e<0$ or $x_e>1$ at any redshift 

584: because the higher modes can have a significant effect on $x_e(z)$. 

585: However, as shown in \S~\ref{sec:eigenmodes}, the higher modes do not 

586: affect the polarization power spectrum since the 

587: high-frequency oscillations in 

588: redshift of higher modes are averaged out.  Similarly, such eigenmodes 

589: have a small effect on the optical depth from a sufficiently large 

590: range in redshift. For example, the optical depth from $15<z<30$ due to 

591: modes $\mu>5$ subject to the physicality constraints of this section 

592: can be no larger than $\sim 0.01$, and is likely to be smaller 

593: for realistic reionization scenarios. Because of this, we assume that 

594: we can place priors on the \emph{optical depth} that correspond to $0 \le x_e \le 1$ 

595:  over the relevant range in redshift. Reionization histories 

596: with an unphysical 

597: optical depth over a large range in redshift are considered to be 

598: unphysical models since the addition of higher modes can not perturb the 

599: optical depth enough to give it a physical value.

600: 

601: 

602: % =====================================================

603: \section{Applications of the Principal Component Method}

604: \label{sec:appl}

605: 

606: Once constraints on the principal components of the reionization history 

607: have been obtained from CMB polarization data, there are several ways 

608: to use those constraints to place limits on 

609: observables such as $\tau$ or to test 

610: theories of reionization \citep{HuHol03}. 

611: We describe some possible applications in this section, 

612: and in \S~\ref{sec:mcmc} we put these ideas into practice using 

613: the 3-year \wmap\ data and simulated future data.

614: 

615: As mentioned in \S~\ref{sec:intro}, the constraint on the total optical 

616: depth to reionization depends on the assumed model for $x_e(z)$. 

617: The principal component method allows us to explore 

618: all globally-averaged ionization 

619: histories within a chosen redshift range, $\zmin<z<\zmax$. 

620: For a given set of eigenmode amplitudes, $\{m_{\mu}\}$, 

621: equation~(\ref{eq:mmutoxe}) yields the corresponding ionization history 

622: which can then be integrated to find the optical depth between any two

623: redshifts $z_{1}$ and $z_{2}$,

624: \begin{equation}

625: \tau(z_1,z_2) = 0.0691(1-Y_p)\Omega_b h

626: \int_{z_1}^{z_2} dz \frac{(1+z)^2}{H(z)/H_0}x_e(z).

627: \label{eq:tauz1z2}

628: \end{equation}

629: In particular, the total optical depth to reionization is

630: \begin{equation}

631: \tau = \tau(0,\zmin)+\tau(\zmin,\zmax),

632: \label{eq:tau}

633: \end{equation}

634: where $\tau(0,\zmin)\approx 0.04$ for $\zmin=6$. 

635: The principal component approach provides a model-independent way to 

636: constrain $\tau$, so we expect the results to be unbiased and the uncertainty 

637: in $\tau$ to accurately reflect the present uncertainty about 

638: $x_e(z)$.  Moreover, the information about $\tau(\zmin,\zmax)$

639:  is encapsulated in the first few eigenmodes, i.e. the 

640: truncated ionization history of equation~(\ref{eq:trunc}), 

641: $x_e^{(N)}(z)$, with $N=3$-5 for typical fiducial models.

642: 

643: 

644: The total optical depth and the first principal component amplitude, $m_1$, 

645: are similar quantities in that they are both averages of $x_e(z)$ weighted 

646: at high $z$. As we show in the next section, CMB $E$-mode polarization data 

647: can constrain higher eigenmodes as well ($\mu \geq 2$). 

648: In particular, the second mode should be the next best constrained 

649: quantity since it is constructed to have the smallest variance after $m_1$.

650: Since $m_2$ is related to the difference in $x_e$ between  

651: high redshift and low redshift (see Fig.~\ref{fig:eigfn}), 

652: we might guess that besides total $\tau$ the data are mainly 

653: constraining the fraction of optical depth coming from 

654: high redshift versus low redshift.  

655: Given any set of mode amplitudes defining an ionization history, we can 

656: compute a low-$z$ optical depth, $\tau(\zmin,\zmid)$, and a high-$z$ 

657: optical depth, $\tau(\zmid,\zmax)$, for some choice of intermediate 

658: redshift, $\zmid$.

659: 

660: Note that if either of the redshift ranges $[\zmin,\zmid]$ or 

661: $[\zmid,\zmax]$ is too narrow, constraints on the partial optical depths  

662: will be influenced by the physicality priors 

663: (since $0\leq x_e\leq 1$ sets limits 

664: on $\tau$ in redshift intervals as discussed in \S~\ref{sec:phys}) 

665: and by the uncertainty in modes higher than those included in the chain, 

666: which have greater effect on the optical depth in narrower redshift 

667: intervals. For these reasons, $\zmid$ should be chosen to be not too close 

668: to either $\zmin$ or $\zmax$.

669: 

670: Given an appropriately chosen value of $\zmid$, constraints on the 

671: principal components can be converted into constraints on the  

672: optical depths at high and low redshift. These partial optical depths 

673: are observables in the sense that high-$z$ and low-$z$ 

674: optical depth affect the 

675: large-scale $E$-modes over different ranges of multipoles.

676: For example, compare the two reionization models 

677: in Figure~\ref{fig:cl}. Both have similar total optical depth, but in 

678: one case $\tau$ only comes from $z<15$, resulting in more power at 

679: $\ell < 8$ and less power at $\ell > 8$ than the other model in which 

680: the optical depth primarily comes from $z>15$.

681: 

682: Besides learning about the relative amount of ionization at $z<\zmid$ 

683: and $z>\zmid$, the partial optical depth constraints also provide a 

684: way to empirically set the maximum redshift, $\zmax$. 

685: If some set of data are found to place a tight upper bound 

686: on $\tau(\zmid,\zmax)$ 

687: for fairly conservative (i.e. large) values of $\zmid$ and $\zmax$, then 

688: for analyses of future data this value of $\zmid$ can be used as a new, lower 

689: value for $\zmax$ since 

690: optical depth from higher redshifts is small. 

691: This approach assumes that there is 

692: essentially no reionization earlier than the original $\zmax$, but as long 

693: as this initial maximum redshift is taken to be large the presence or 

694: absence of high-$z$ ionization can be tested in this way over a wide 

695: range of redshifts. We explore this idea further in \S~\ref{sec:mcmc}.

696: 

697: 

698: While we do not examine constraints on specific reionization models in this 

699: paper, the principal component approach is well suited to model testing. 

700: Consider a model of the global reionization history, $x_e(z;\bm{\theta})$, 

701: parameterized by $\bm{\theta}$. This could be a simple toy model (for example, 

702: instantaneous reionization where $\bm{\theta}$ is a single parameter, 

703: the redshift of reionization) or a more physical model where $\bm{\theta}$ 

704: might include parameters that govern properties of the ionizing sources. 

705: For a particular choice of $\xef(z)$, the principal 

706: component amplitudes of the reionization model follow from 

707: equation~(\ref{eq:xetommu}), giving a set of mode amplitudes that 

708: depend on the model parameters, $\{m_{\mu}(\bm{\theta})\}$. 

709: Then constraints on the mode amplitudes from CMB polarization data 

710: (defined using the same fiducial history) can be mapped to 

711: constraints on the parameters of the reionization model. 

712: 

713: Applying this method to a model that has as one of its parameters a 

714: maximum redshift allows one to obtain constraints on $\zmax$ within a 

715: class of theoretical models.

716: Although obtaining good constraints on model parameters does not necessarily

717: imply validation of the model class, this procedure  provides a 

718: straightforward method by which different model 

719: classes can tested and falsified within a single analysis.

720: 

721: 

722: Generating Monte Carlo chains to find constraints on 

723: $\{m_{\mu}\}$ as described in the next section can be a somewhat 

724: time-consuming process, but it only needs to be done once 

725: per redshift range, after which any model of the global 

726: reionization history within this range $\zmin<z<\zmax$

727: can be tested using the same parameter chains. 

728: In cases where constraints from the data turn out 

729: to be close to Gaussian, the covariance matrix of the principal components 

730: can be used in place of the full Monte Carlo chains, reducing the 

731: amount of information needed for model testing from $\sim 10^5$ numbers 

732: in chains of Monte Carlo samples to only 

733: $N(N+3)/2$ numbers when $N$ eigenmodes are included in the chains 

734: [$N$ mean values plus $N(N+1)/2$ entries in the covariance matrix, 

735: $\langle m_{\mu}m_{\nu}\rangle$].

736: However, near-Gaussian constraints on $\{m_{\mu}\}$ are likely to be 

737: possible only for certain realizations of future data, as discussed in 

738: the next section.

739: 

740: 

741: 

742: 

743: % =====================================================

744: \section{Markov Chain Monte Carlo Constraints on Eigenmodes}

745: \label{sec:mcmc}

746: 

747: We find reionization constraints from CMB polarization data using 

748: Markov Chain Monte Carlo (MCMC) techniques to explore the principal 

749: component parameter space \citep[see e.g.][]{Chretal01,KosMilJim02,Dunetal04}. 

750: Chains of Monte Carlo samples are generated using the publicly available 

751: code CosmoMC\footnote{http://cosmologist.info/cosmomc/} \citep{LewBri02}, 

752: which includes the code CAMB \citep{Lewetal00} for computing the 

753: theoretical angular power spectrum at each point in the $\{m_1,\ldots,m_N\}$ 

754: parameter space. We have modified both codes to allow 

755: specification of an arbitrary reionization history calculated from a set 

756: of mode amplitudes using equation~(\ref{eq:mmutoxe}). 

757: 

758: The principal component amplitudes are the only parameters 

759: allowed to vary in the chains.

760: Since nearly all the information in $\clee$ from 

761: reionization is contained within the first few eigenmodes of $x_e(z)$, 

762: we include the modes $\{m_1,\ldots,m_N\}$ in each chain with $3\leq N\leq 5$. 

763: We use only the $E$-mode polarization data for parameter constraints, 

764: and assume that the values of the 

765: standard \lcdm\ parameters (besides $\tau$) are fixed by measurements of the 

766: CMB temperature anisotropies. 

767: This leads to a slight underestimate of the error on $\tau$ as we discuss 

768: later, but to a good approximation the effect of $x_e(z)$ on the large-scale 

769: $E$-modes is independent of the other parameters. 

770: 

771: Specifically, we take  

772: $\Omega_bh^2=0.0222$, $\Omega_ch^2=0.106$, $100\theta=1.04$ (corresponding 

773: to $h=0.73$), $A_s e^{-2\tau}=1.7 \times 10^{-9}$, and $n_s=0.96$, 

774: consistent with results from the most recent version of the \wmap\ 3-year 

775: likelihood code. When computing the optical depth to reionization we 

776: take the primordial helium fraction to be $Y_p=0.24$. 

777: We also assume that $\zmin=6$, so that the optical depth contributed at 

778: lower redshifts is fixed at $\tau(\zmin)\approx 0.04$. 

779: The remaining total optical depth from reionization,

780: $\tau(\zmin,\zmax)$, is determined by the 

781: values of $\{m_{\mu}\}$ for each sample in the chains.

782: The default bin width for our 

783: fiducial models $\xef(z)$ is $\dz=0.25$, which is small enough that 

784: numerical effects related to binning should be negligible. 

785: 

786: To get accurate results from MCMC analysis, 

787: it is important to make sure that the 

788: parameter chains contain enough independent samples 

789: covering a sufficient volume of parameter space so that the density of 

790: the samples converges to the actual posterior 

791: probability distribution. 

792: For each scenario that we study, we run 4 separate chains 

793: until the Gelman and Rubin convergence statistic $R$, 

794: corresponding to the ratio of the variance of parameters between chains to 

795: the variance within each chain, satisfies 

796: $R-1<0.01$ \citep{GelRub92,BroGel98}. The convergence diagnostic of 

797: \cite{RafLew92} is used to determine how much each chain must be thinned 

798: to obtain independent samples.  Both of these statistics and other 

799: diagnostic measures are computed automatically by CosmoMC.

800: 

801: Since $x_e(z)$ during reionization must match onto $x_e\approx 1$ at 

802: $z=z_{\rm min}$ and $x_e\approx 0$ at $z=\zmax$, there are often sharp 

803: transitions at these redshifts since nothing in 

804: equation~(\ref{eq:mmutoxe}) forces $x_e(z)$ to satisfy these 

805: boundary conditions. 

806: To avoid problems with the time integration in CAMB, we 

807: smooth the reionization history by convolving $x_e(z)$ with a Gaussian of 

808: width $\sigma_z=0.5$.

809: 

810: As described in previous sections, we assume that $0\leq x_e\leq 1$ 

811: between $\zmin$ and $\zmax$. This assumption neglects helium reionization, 

812: which can make $x_e$ slightly larger than unity, and the small 

813: residual ionized fraction remaining after recombination that 

814: prevents $x_e$ from ever being exactly zero. 

815: These are relatively small effects, 

816: especially since we are not placing constraints on $x_e(z)$ directly but 

817: rather on weighted averages of $x_e$ over redshift. The residual 

818: ionized fraction at $z>\zmax$ is accounted for in the Monte 

819: Carlo exploration of reionization histories.

820: 

821: In \S~\ref{sec:wmap}, we examine the current 

822: constraints from the 3-year \wmap\ data. 

823: We then provide forecasts for principal component constraints with idealized, 

824: cosmic variance-limited, simulated data in \S~\ref{sec:cvdata}.

825: In each case the likelihood 

826: computation includes only the $E$-mode polarization data, up to $\ell=100$ for 

827: simulated data and $\ell=23$ for \wmap ; the likelihood code for \wmap\ 

828: does not use $\clee$ at smaller scales due to low signal-to-noise.

829: At multipoles $\ell \gtrsim 100$ the global reionization history only affects 

830: the amplitude of the angular power spectra, which we fix by setting 

831: $A_s e^{-2\tau}$ constant in the Monte Carlo chains.

832: A comparison of the MCMC results with the Fisher matrix 

833: approximation follows in \S~\ref{sec:fisher}.

834: 

835: 

836: 

837: 

838: % --------------------------------------------

839: \subsection{Constraints from \wmap}

840: \label{sec:wmap}

841: 

842: In our analysis of \wmap\ data we use the 3-year \wmap\ likelihood code, 

843: with settings chosen so that likelihoods include only contributions from 

844: the low-$\ell$ $E$-mode polarization.  In this regime, the code 

845: computes model likelihoods with 

846: a pixel-based method instead of using the angular power 

847: spectrum~\citep{Pagetal06}.

848: Since the maximum redshift 

849: at which there is still a significant ionized fraction is uncertain, we 

850: generate Monte Carlo chains using different fiducial models with 

851: $10\leq \zmax \leq 40$. 

852: To avoid possible bias due to neglecting the possibility of high-redshift 

853: reionization, we focus here on the results obtained using 

854: the more conservative values of $\zmax=30$ and 40.

855: 

856: The MCMC constraints on the first three principal components from \wmap\ are 

857: shown in Figure~\ref{fig:mcmcwmap} using fiducial models with $\zmax=30$ and 

858: $\zmax=40$. The marginalized constraints are plotted within the physicality 

859: bounds of equation~(\ref{eq:physbounds}), 

860: $m_{\mu}^{(-)}\leq m_{\mu}\leq m_{\mu}^{(+)}$, so the size of the 

861: contours inside each box gives an idea of the constraining power of the 

862: data within the space of potentially physical models. 

863: For both fiducial histories, the data 

864:  place a strong upper limit on $m_1$ and weakly constrain $m_2$. 

865: It is important to note that 

866: although the parameters $\{m_1,m_2,m_3\}$ have the same names in both the

867: $\zmax=30$ and $\zmax=40$ plots, they are defined with respect to different 

868: fiducial models and so we do not expect the contours in the left and right 

869: plots in Figure~\ref{fig:mcmcwmap} to agree exactly. 

870: The qualitative similarity 

871: between the contours simply reflects the fact that the eigenfunctions 

872: for different $\zmax$ have similar shapes (Figure~\ref{fig:eigfn}).

873: 

874: 

875: 

876: % ****************************************

877: \begin{figure*}

878: \centerline{\psfig{file=f6a.eps, width=3.0in}

879: \psfig{file=f6b.eps, width=3.0in}}

880: \caption{Marginalized 2D contours 

881: (\emph{dark shading}: 68\% CL; \emph{light}: 95\%)  for

882: the first three principal components of $x_e(z)$ from MCMC analysis of 

883: 3-year \wmap\ $E$-mode polarization data, 

884: using fiducial reionization histories with $\zmax=30$ (\emph{left}) 

885: and $\zmax=40$ (\emph{right}). Note that the mode amplitudes $m_{\mu}$ 

886: are defined differently in the left and right panels since the 

887: fiducial models are different.

888: \vskip 0.25cm

889: }

890: \label{fig:mcmcwmap}

891: \end{figure*}

892: % ****************************************

893: 

894: 

895: 

896: Since $m_1$ and $\tau$ are both averages of $x_e$ from $\zmin$ to $\zmax$ 

897: with more weight at high $z$ than low $z$, the strong constraint on 

898: $m_1$ mostly reflects the ability of the data to constrain the total 

899: optical depth to reionization. Indeed, $m_1$ and $\tau$ are strongly 

900: correlated in the Monte Carlo chains, while the correlations between 

901: $\tau$ and higher principal components of $x_e(z)$ are weaker.

902: 

903: Relative to the physicality bounds, the constraints 

904: in Figure~\ref{fig:mcmcwmap} are 

905: stronger for the $\zmax=40$ chains than for $\zmax=30$; this is 

906: to be expected since the physicality bounds on $\{m_{\mu}\}$ permit 

907: a greater variety of ionization histories, and in particular a 

908: wider range in $\tau$, when $\zmax$ is larger.

909: The range of eigenmode amplitudes allowed by 

910: the limits of equation~(\ref{eq:physbounds}) 

911: is nearly independent of $\zmax$, but the effect of a 

912: unit-amplitude principal component on $\clee$ increases with $\zmax$ 

913: as illustrated in Figure~\ref{fig:zmaxa}.

914: 

915: As described in \S~\ref{sec:appl}, principal component constraints 

916: from MCMC yield model-independent constraints on the total optical 

917: depth. For each Monte Carlo sample of the principal components we compute 

918: $\tau$ using equation~(\ref{eq:tau}). The constraints on $\tau$ are 

919: listed in Table~\ref{tab:tau} for 

920: various values of $\zmax$ and $N$, the number of modes in the Monte Carlo 

921: chains. 

922: The uncertainty in $\tau$ from the Monte Carlo chains 

923: is slightly underestimated because 

924: we fix all cosmological parameters that are not directly related to 

925: reionization. To estimate the effect of this assumption, we 

926: compare constraints on $\tau$ in two cases where instantaneous 

927: reionization is assumed instead of using the model-independent 

928: principal component approach: 

929: one in which the non-reionization parameters are 

930: fixed as in the rest of our Monte Carlo chains, and the other, from the 

931: 3-year \wmap\ analysis of \cite{Speetal06}, in which these  parameters 

932: are allowed to vary. These constraints, listed in rows 4 and 5 of 

933: Table~\ref{tab:tau}, suggest that fixing \lcdm\ parameters besides $\tau$ 

934: reduces $\sigma_{\tau}$ by $\sim 10\%$.

935: 

936: 

937: \begin{table}

938: \caption{Constraints on total optical depth to reionization.}

939: \begin{center}

940: \begin{tabular}{llcclc}

941: \hline

942: \hline

943: & Use PCs &  &  & Fix other & \\

944: Data & of $x_e(z)$?\tablenotemark{a} & ~$\zmax$~ & ~$N$~ & parameters?\tablenotemark{b} & $\tau$ \\

945: \hline

946: WMAP3 & Yes & 20 & 3 & Yes & $0.098\pm0.025$ \\

947: WMAP3 & Yes & 30 & 5 & Yes & $0.106\pm0.027$ \\

948: WMAP3 & Yes & 40 & 5 & Yes & $0.107\pm0.029$ \\

949: WMAP3 & No  & -- & -- & Yes & $0.096\pm0.027$ \\

950: WMAP3 & No  & -- & -- & No & $0.089\pm0.030$\tablenotemark{c} \\

951: \hline

952: CV\tablenotemark{d} & Yes & 20 & 3 & Yes & $0.103\pm0.004$ \\

953: CV & Yes & 30 & 5 & Yes & $0.108\pm0.005$ \\

954: CV & No & -- & -- & Yes & $0.108\pm0.003$ \\

955: \hline

956: \tablenotetext{a}{If not using principal components, reionization is assumed to occur instantaneously.}

957: \tablenotetext{b}{i.e. $\Omega_bh^2$, $\Omega_ch^2$, $h$, $A_s e^{-2\tau}$, and $n_s$. Other non-reionization parameters are always fixed.}

958: \tablenotetext{c}{From \cite{Speetal06}.}

959: \tablenotetext{d}{The true history assumed here is instantaneous reionization 

960: with $\tau=0.105$, using the cosmic variance-limited realization of $\clee$ in the left 

961: panel of Figure~\ref{fig:clcv}.}

962: \end{tabular}

963: \end{center}

964: \label{tab:tau}

965: \end{table}

966: 

967: 

968: 

969: % ****************************************

970: \begin{figure}

971: \centerline{\psfig{file=f7.eps, width=3.0in}}

972: \caption{\emph{Left panel}: 

973: Marginalized 2D contours (68 and 95\%) from a $\zmax=30$ \wmap\ 

974: chain of Monte Carlo samples 

975: (\emph{thick curves}), lines of constant $\tau=\{0.06,0.09,0.12\}$ 

976: (\emph{dashed lines}), and the set of 

977: one-parameter instantaneous reionization models 

978: (\emph{thin curve}) plotted in the $m_1-m_2$ plane. 

979: Higher eigenmodes ($\mu\geq 3$), which only weakly influence $\tau$, 

980: are fixed to $m_{\mu}=0$ for the purposes of plotting the 

981: constant-$\tau$ lines.

982: \emph{Right panel}: posterior probability of the optical depth from 

983: 3-year \wmap\ data, $P(\tau)$, for 

984: arbitrary $x_e(z)$ with $\zmin=6$, $\zmax=30$, and $N=5$ principal 

985: components (\emph{thick curve}), and for the one-parameter 

986: instantaneous reionization model with a $\zmin=6$ prior (\emph{thin curve}).

987: \vskip 0.25cm

988: }

989: \label{fig:tauinst}

990: \end{figure}

991: % ****************************************

992: 

993: 

994: 

995: In all cases where we fix the other 

996: cosmological parameters, the uncertainty in $\tau$ is similar 

997: regardless of whether we use principal components to explore a variety 

998: of reionization histories or restrict the analysis to 

999: instantaneous $x_e(z)$. This is somewhat surprising, since in general 

1000: one would expect that expanding the model space for $x_e(z)$ would 

1001: increase the estimated error on $\tau$. 

1002: The physicality priors may be responsible for reducing $\sigma_{\tau}$ 

1003: slightly --- top-hat priors on $\{m_{\mu}\}$ induce 

1004: a prior on $\tau$ that is flat over a certain range 

1005: but falls off approximately linearly at the edges --- but 

1006: even after accounting for priors, 

1007: the optical depth constraint is robust to replacing the 

1008: instantaneous reionization assumption with a model-independent analysis. 

1009: 

1010: The insensitivity of the constraint on $\tau$ to the set of models 

1011: considered is partly due to the fact that the degeneracy in the 

1012: eigenmode constraints is aligned in the direction of constant $\tau$, 

1013: as shown in Figure~\ref{fig:tauinst} in the $m_1-m_2$ plane. 

1014: The set of instantaneous reionization models, plotted as a curve 

1015: in Figure~\ref{fig:tauinst}, cuts across the \wmap\ constraints on more 

1016: general models in a region of high posterior 

1017: probability where the distribution of samples varies slowly along 

1018: lines of constant $\tau$. This large overlap between the general 

1019: reionization histories favored by the data and the line of 

1020: instantaneous models is the main reason why the probability distributions 

1021: $P(\tau)$, plotted in the right panel of Figure~\ref{fig:tauinst}, 

1022: are similar for the two classes of models. [The sharp cutoff at low $\tau$ 

1023: in the instantaneous reionization $P(\tau)$ comes from our 

1024: $\zmin=6$ prior.] 

1025: Note that the fact that the 

1026: instantaneous reionization curve passes through the middle of 

1027: the 68\% confidence region also indicates that models with rapid 

1028: reionization are not at all disfavored by the 3-year \wmap\ data.

1029: We use two methods to find the posterior distribution 

1030: for $\tau$ in the instantaneous reionization case: one is the usual 

1031: approach of varying the optical depth (or reionization redshift) in a 

1032: Monte Carlo chain, and the other involves computing $P(\tau)$ for 

1033: a subset of samples from chains in which principal components of $x_e(z)$ 

1034: are varied, selecting only those samples with $\{m_{\mu}\}$ values close 

1035: to the 1D instantaneous reionization curve. Both approaches produce 

1036: consistent probability distributions; the former method is used for 

1037: $P(\tau)$ plotted in Figure~\ref{fig:tauinst}.

1038: 

1039: 

1040: 

1041: The weak $m_2$ constraint visible in Figure~\ref{fig:mcmcwmap} suggests 

1042: that in addition to determining the total optical depth, 

1043: \wmap\ data may provide useful limits on high-$z$ and low-$z$ 

1044: optical depth as defined in the previous section. 

1045: The 68 and 95\% posterior probability contours for the 

1046: partial optical depths from \wmap\ 

1047: are shown in Figure~\ref{fig:tauwmap} for 

1048: $\zmid = 20$ and $\zmax = \{30,40\}$. 

1049: The $x_e=0$ physicality prior cuts off the contours at $\tau(z_1,z_2)=0$; 

1050: in both panels, the upper limits set by $x_e=1$ are outside the plotted area, 

1051: so the contours are not strongly influenced by those priors.

1052: The error ``ellipses'' are narrowest in the 

1053: direction along which the total optical depth is constrained, as shown 

1054: by dashed lines of constant total optical depth at $\tau=0.06$ and 

1055: $\tau=0.12$ [approximately the upper and lower 1 $\sigma$ limits from 

1056: the 3-year \wmap\ analysis of \citet{Speetal06}].

1057: 

1058: 

1059: % ****************************************

1060: \begin{figure}

1061: \centerline{\psfig{file=f8.eps, width=3.0in}}

1062: \caption{Constraints on optical depth during reionization divided between 

1063: two wide ranges in redshift, $\tau(6,20)$ and $\tau(20,\zmax)$, for 

1064: $\zmax=30$ (\emph{left}) and 40 (\emph{right}). 

1065: The contours are drawn at 68\% (\emph{dark gray}) and 95\% (\emph{light gray}) 

1066: 2D marginalized posterior probability. 

1067: Dashed lines in each panel are lines of constant total optical depth: 

1068: $\tau=0.12$ for the upper line, and $\tau=0.06$ for the lower line. 

1069: \vskip 0.25cm

1070: }

1071: \label{fig:tauwmap}

1072: \end{figure}

1073: % ****************************************

1074: 

1075: 

1076: Comparison of the two panels in Figure~\ref{fig:tauwmap} 

1077: reveals that the choice of $\zmax$ 

1078: does not have a large effect on constraints on the

1079:  optical depth above and below $z=20$.  This suggests

1080: that $\tau(20,\zmax)$ should be interpreted as $\tau(z>20)$.  

1081: We provide further

1082: justification for this interpretation in \S~\ref{sec:cvdata}.

1083: 

1084: The 95\% upper limits on the high-redshift optical depth, 

1085: $\tau(z>20)$, 

1086: are $0.076$ and $0.078$ for $\zmax=30$ and 40, respectively, 

1087: after marginalizing over all other parameters.

1088: This is not a particularly strong constraint, since 

1089: this result is only marginally inconsistent with all of the optical depth from 

1090: reionization coming from $z>20$, but with future data it should be 

1091: possible to either reduce the upper limit on high-$z$ optical depth 

1092: or to detect the presence of a substantial ionized fraction at high redshift 

1093: (see \S~\ref{sec:cvdata}). 

1094: 

1095: Models of reionization can be tested by computing the 

1096: principal components of 

1097: proposed ionization histories and comparing with constraints on 

1098: $\{m_{\mu}\}$ from Monte Carlo chains.  

1099: The \wmap\ constraints in Figure~\ref{fig:mcmcwmap} 

1100:  are non-Gaussian, in part 

1101: because the physicality priors on $\{m_{\mu}\}$ intersect the 

1102: posterior probability contours where the likelihood is large. 

1103: Because the constraints do not have a simple Gaussian form, it appears 

1104: that the full parameter chains are necessary for accurate model testing, 

1105: although the viability of various models can be estimated by comparing 

1106: their eigenmode amplitudes with marginalized constraints such as 

1107: those in Figure~\ref{fig:mcmcwmap}, keeping in mind that models favored 

1108: by marginalized constraints could be disfavored in the full $N$-D 

1109: parameter space.

1110: 

1111: 

1112: 

1113: 

1114: % --------------------------------------------

1115: \subsection{Cosmic variance-limited data}

1116: \label{sec:cvdata}

1117: 

1118: To forecast how well principal components of the reionization history could be 

1119: measured by future CMB polarization experiments, we repeat the analysis 

1120: of \S~\ref{sec:wmap} using simulated realizations of the full-sky, 

1121: noiseless $E$-mode angular power spectrum instead of \wmap\ data. 

1122: Any real experiment will of course involve sky cuts, 

1123: noise, foregrounds, and other complications, so the results 

1124: presented here represent an optimistic limit on what we can learn about 

1125: the global reionization history 

1126: from low-$\ell$ $E$-mode polarization. At the end of this section, we 

1127: estimate how much the constraints might be degraded from this idealized 

1128: case for an experiment with characteristics similar to those proposed 

1129: for the \emph{Planck} satellite.

1130: 

1131: We generate simulated realizations of $\clee$ 

1132: drawn from $\chi^2$ distributions with cosmic variance determined by 

1133: the theoretical angular power spectra that we compute using CAMB. 

1134: For the $j$th sample in a chain, the likelihood 

1135: including only $E$-mode polarization is

1136: \begin{equation}

1137: -\ln L_{(j)}=\sum_{\ell}\left(\ell+\frac{1}{2}\right)

1138: \left(\frac{\hat{C}_{\ell}^{EE}}

1139: {C_{\ell(j)}^{EE}}+\ln\frac{C_{\ell(j)}^{EE}}

1140: {\hat{C}_{\ell}^{EE}}-1\right),

1141: \label{eq:like}

1142: \end{equation}

1143: where $\hat{C}_{\ell}^{EE}$ is the spectrum of the simulated data and

1144: $C_{\ell(j)}^{EE}$ is the theoretical spectrum calculated with the 

1145: parameter values at step $j$ in the chain.

1146: 

1147: We have run Monte Carlo chains for multiple realizations of 

1148: $\clee$ drawn from spectra computed assuming a variety of ``true'' 

1149: reionization histories, $\xet(z)$.

1150: %

1151: % The ability of the simulated data to 

1152: %constrain the principal components tends to be similar regardless 

1153: %of the choice of $\xet(z)$, so 

1154: We start by taking  the 

1155: true history to be a model with nearly instantaneous reionization and 

1156: $\tau=0.105$.  As a contrasting model for comparison we also use 

1157: an extended, double reionization model with $\tau=0.090$. 

1158: Figure~\ref{fig:cl} shows $\clee$ and $x_e(z)$ for 

1159: each of these models. 

1160: 

1161: Since parameter constraints derived using 

1162: a single draw of $\clee$ from the underlying power spectrum 

1163: may contain features unique to that realization, we generated Monte Carlo 

1164: chains for 10 random realizations of  

1165: the instantaneous reionization 

1166: power spectrum. Two of these realizations are 

1167: plotted as points in Figure~\ref{fig:clcv}, along with the 

1168: theoretical spectrum with cosmic variance bands. The thick curves 

1169: in Figure~\ref{fig:clcv} are spectra of the best-fit models 

1170: from the Monte Carlo chains for each realization. 

1171: The overall best-fit models are similar for 

1172: the two realizations and both agree closely with the theoretical $\clee$. 

1173: The dashed curve in the right panel of Figure~\ref{fig:clcv} shows an 

1174: ``alternative'' best-fit model that we discuss later.

1175: 

1176: 

1177: 

1178: % ****************************************

1179: \begin{figure*}

1180: \centerline{\psfig{file=f9a.eps, width=3.0in}

1181: \psfig{file=f9b.eps, width=3.0in}}

1182: \caption{$E$-mode spectra of the best-fit models from two Monte Carlo 

1183: chains (\emph{thick curves}), each using a different realization of 

1184: simulated cosmic variance-limited data (\emph{points}).  The 

1185: realizations are drawn from the same $\clee$ of the instantaneous 

1186: reionization model in Figure~\ref{fig:cl}, 

1187: plotted as thin curves with shaded regions showing 

1188: the $1~\sigma$ cosmic variance limits.

1189: The dashed curve in the right panel shows the best fit model within the 

1190: high-$m_2$ peak in the posterior probability (see Fig.~\ref{fig:mcmccv}).

1191: \vskip 0.25cm}

1192: \label{fig:clcv}

1193: \end{figure*}

1194: % ****************************************

1195: 

1196: 

1197: % ****************************************

1198: \begin{figure*}

1199: \centerline{\psfig{file=f10a.eps, width=3.0in}

1200: \psfig{file=f10b.eps, width=3.0in}}

1201: \caption{Marginalized 2D contours (68 and 95\%) for 

1202: the first three principal components with two different 

1203: realizations of cosmic variance-limited data simulated using the 

1204: instantaneous reionization history. The fiducial model has 

1205: maximum redshift $\zmax=20$, and only the three modes shown here 

1206: are varied in the parameter chains. The cross in each panel marks the 

1207: location of the principal components of $\xet(z)$.

1208: \vskip 0.25cm}

1209: \label{fig:mcmccv}

1210: \end{figure*}

1211: % ****************************************

1212: 

1213: 

1214: The 2D marginalized MCMC constraints on principal components for 

1215: these two realizations are shown in Figure~\ref{fig:mcmccv}.

1216: The plotted regions are bounded by the physical top-hat priors on $\{m_{\mu}\}$

1217: as in Figure~\ref{fig:mcmcwmap}, but the eigenmodes are different here 

1218: with $\zmax=20$ instead of 30 or 40 for the fiducial history. 

1219: Since we have constructed $\xet(z)$ to have no ionization at $z\gtrsim 15$, 

1220: we know that $\zmax=20$ should be a large enough maximum redshift; 

1221: this choice of $\zmax$ also reflects the fact that improved empirical 

1222: determination of $\zmax$ will allow it to be set at lower redshift 

1223: for future data analysis, assuming that high-$z$ reionization is 

1224: not detected. 

1225: 

1226: 

1227: The contours in Figure~\ref{fig:mcmccv} show that cosmic variance-limited 

1228: data can constrain models within the physically allowed parameter space 

1229: much better than current data can. The chains used for the results 

1230: shown here vary the first 3 eigenmodes; by running chains with $N>3$ we 

1231: find that cosmic variance-limited data can provide $2~\sigma$ constraints 

1232: that are tighter than the physicality bounds for the first 4 eigenmodes 

1233: of a $\zmax=20$ fiducial model and for at least 5 eigenmodes when $\zmax=30$.

1234: 

1235: 

1236: 

1237: Although the only difference in the MCMC setup for the left and right panels 

1238: of Figure~\ref{fig:mcmccv} is the realization 

1239: of $\clee$ (Fig.~\ref{fig:clcv}), 

1240: there are some significant differences in the resulting 

1241: eigenmode constraints, particularly in the $m_2-m_3$ plane. This kind of 

1242: variation is typical among the realizations of simulated data, and 

1243: is related to whether features in the theoretical $\clee$ remain intact 

1244: with the inclusion of cosmic variance. The relevant features in the 

1245: spectrum are the small oscillations in the low-power ``valley'' in 

1246: $\clee$ on scales $10\lesssim \ell \lesssim 30$. In the example shown here, 

1247: the feature that matters most is the bump in the theoretical spectrum at 

1248: $\ell \approx 13$. In the left panel of Figure~\ref{fig:clcv}, the 

1249: $\clee$ realization retains this bump, while in the right panel the 

1250: bump is washed out by cosmic variance. The result is that the left 

1251: realization is better able to pick out those reionization models 

1252: in Monte Carlo chains that 

1253: reproduce the true $\clee$, leading to the tight constraints on the left 

1254: side of Figure~\ref{fig:mcmccv}. The realization in the right panel of 

1255: Figure~\ref{fig:clcv} lacks this important ``fingerprint'' for identifying 

1256: models that match the theoretical $\clee$, so the data allow a wider 

1257: variety of models as reflected in the constraints on the right side of 

1258: Figure~\ref{fig:mcmccv}.  In this case, there are two best-fit models 

1259: corresponding to the two $68\%$ contours separated in the 

1260: $m_2$ direction. The spectrum of the 

1261: overall best-fit model (with smaller $m_2$), which is close to the 

1262: theoretical spectrum, is 

1263: plotted as a thick solid curve in the right panel of Figure~\ref{fig:clcv}, 

1264: while the other best-fit model (at larger $m_2$) is plotted as a dashed 

1265: curve. These two models have significantly different spectra, but 

1266: because of the scatter in the random draw of $\clee$ they are each able 

1267: to fit the data better at some multipoles and worse at others in such a way 

1268: that the high-$m_2$ best-fit model is only a slightly worse fit than  

1269: the low-$m_2$ best-fit model.

1270: 

1271: 

1272: Among the realizations of simulated data that we explored with MCMC 

1273: methods, constraints like those in the right panel of Figure~\ref{fig:mcmccv} 

1274: are a fairly extreme case, occurring about 

1275: 10-20\% of the time. In general the realizations form a sort of 

1276: continuum between the two presented here, with about half showing some 

1277: displacement of the $95\%$ contour towards larger $m_2$ but 

1278: not as much as in the second realization that we show as 

1279: an example.

1280: 

1281: The extent to which differences between realizations show up in 

1282: principal component constraints depends on $\xet(z)$. For example, 

1283: the $\clee$ for the double reionization model (Fig.~\ref{fig:cl}) 

1284: have more pronounced oscillations at $\ell\sim 10$-$30$ than the 

1285: instantaneous reionization spectrum, so they are not as easily 

1286: erased by cosmic variance or noise and the resulting principal component 

1287: constraints tend to be more consistent from one realization to the next 

1288: and more like the left panel of Figure~\ref{fig:mcmccv}.

1289: 

1290: 

1291: The differences between constraints from various realizations of simulated 

1292: data suggest the interesting possibility that our ability to learn about 

1293: the reionization history from large scale $E$-mode polarization may 

1294: ultimately depend on the luck of the draw of $\clee$ at our particular 

1295: vantage point.  In an unlucky

1296:  draw, cosmic variance can distort subtle features in the power 

1297: spectrum in a way that would limit constraints on reionization 

1298: eigenmodes to be worse than we might expect from the Fisher approximation 

1299: or a more typical draw. The good news is that 

1300: even a realization like the one in the right panels of Figures~\ref{fig:clcv}

1301: and~\ref{fig:mcmccv} would permit constraints that are far better than 

1302: what is currently possible.  It is also important to note that although such 

1303: constraints are weaker than in the best-case scenario, 

1304: they are still consistent with the true 

1305: parameter values and in general would not lead us to rule out the 

1306: true reionization history based on a principal component analysis 

1307: of the data. 

1308: 

1309: 

1310: As with the principal components of $x_e(z)$, MCMC analysis indicates that 

1311: constraints on the total optical depth to reionization would be  

1312: greatly improved with cosmic variance-limited data: 

1313: $\sigma_{\tau}\approx 0.005$, down a factor of six from the 3-year 

1314: \wmap\ value, $\sigma_{\tau}\approx 0.03$. 

1315: The constraints on 

1316: $\tau$ using the realization of instantaneous reionization $\clee$ 

1317: in the left panel of Figure~\ref{fig:clcv} are listed in Table~\ref{tab:tau} 

1318: for two fiducial models with $\zmax=20$ and 30. 

1319: Optical depth constraints from the other draw of $\clee$ are similar, 

1320: so estimates of $\tau$ and $\sigma_{\tau}$ appear not to be biased 

1321: for atypical realizations. It makes sense that differences in realizations 

1322: do not significantly affect $\tau$ since the optical depth constraint 

1323: comes mainly from the peak in $\clee$ on the largest scales whereas 

1324: differences in principal component constraints between realizations 

1325: arise from the low-power part of the spectrum at higher $\ell$.

1326: 

1327: The model-independent constraints on $\tau$ are neither biased nor 

1328: significantly weaker relative to  

1329: the constraint obtained assuming instantaneous reionization 

1330: (bottom row in Table~\ref{tab:tau}), as is true for the \wmap\ data. 

1331: This should not be surprising since we have assumed that $\xet(z)$ is 

1332: an instantaneous reionization model. 

1333: However, if $\xet(z)$ is in fact very different from the instantaneous 

1334: model, the estimate of $\tau$ obtained under the assumption of 

1335: instantaneous reionization will be incorrect. 

1336: For example, if we take $\xet(z)$ to be the double reionization 

1337: model of Figure~\ref{fig:cl}, the model-independent approach yields 

1338: an optical depth constraint consistent with the true value of 

1339: $\tau=0.090$. On the other hand, MCMC analysis restricted to 

1340: instantaneous reionization histories gives a significantly biased 

1341: estimate, $\tau=0.135\pm 0.005$. In this case, the 1D curve of instantaneous 

1342: reionization models in the space of eigenmodes never 

1343: lies near the principal component values favored by the data, so any 

1344: analysis that only considers such models would never find a 

1345: good fit to the data.  This illustrates the importance of a model-independent 

1346: approach: although results from current data may not be significantly 

1347: affected by the assumption of a specific form of $x_e(z)$, future 

1348: data will be sensitive enough to the reionization history that 

1349: the choice of a specific model can greatly 

1350: impact the results \citep{Holetal03,Coletal05}.

1351: 

1352: 

1353: 

1354: % ****************************************

1355: \begin{figure*}

1356: \centerline{\psfig{file=f11a.eps, width=3.0in}

1357: \psfig{file=f11b.eps, width=3.0in}}

1358: \caption{Same as Figure~\ref{fig:tauwmap} for simulated 

1359: cosmic variance-limited data using the two realizations 

1360: of instantaneous reionization $\clee$ shown in 

1361: Figure~\ref{fig:clcv}. Crosses indicate the 

1362: partial optical depths for the true reionization history. 

1363: The redshifts chosen for $\zmid$ and $\zmax$ are also lower here than  

1364: for the \wmap\ analysis, reflecting the expectation that 

1365: future data may permit better 

1366: empirical constraints on $\zmax$.

1367: \vskip 0.25cm}

1368: \label{fig:taucv1}

1369: \end{figure*}

1370: % ****************************************

1371: 

1372: 

1373: % ****************************************

1374: \begin{figure}

1375: \centerline{\psfig{file=f12.eps, width=3.0in}}

1376: \caption{Same as Figure~\ref{fig:taucv1}, but 

1377: assuming that $\xet(z)$ is a double reionization history 

1378: with $\zmax=23$ (thin curves in Figure~\ref{fig:cl}). 

1379: Crosses mark the true partial optical depths for this model; 

1380: note that unlike in Figure~\ref{fig:taucv1}, $\tau(15,20)\neq \tau(15,30)$ 

1381: for the expected values here since $\xet(z)$ has a nonzero 

1382: ionized fraction between redshifts 20 and 30.

1383: \vskip 0.25cm}

1384: \label{fig:taucv2}

1385: \end{figure}

1386: % ****************************************

1387: 

1388: 

1389: 

1390: Though constraints on the contributions to $\tau$ from high and low $z$ 

1391: are currently limited to weak upper bounds, 

1392: future polarization data will likely enable more definitive detections 

1393: of high-$z$ reionization if it is present, or tighter upper limits if absent.

1394: The left and right panels of Figure~\ref{fig:taucv1} 

1395: show these constraints from Monte Carlo chains using the 

1396: $\clee$ plotted in the left and right panels of Figure~\ref{fig:clcv}, 

1397: respectively. The true ionization history has zero ionized fraction at 

1398: $z>15$, and the MCMC analysis places a 95\% CL upper limit 

1399: on $\tau(z>15)$ of $\sim 0.03$.

1400: Figure~\ref{fig:taucv2} shows MCMC constraints on the same partial 

1401: optical depths using simulated data where the true model is instead 

1402: taken to be the extended, double reionization history plotted as 

1403: a thin curve in the inset of Figure~\ref{fig:cl}. In this case, most 

1404: of $\tau(\zmin,\zmax)$ comes from $z>15$, and using the principal 

1405: components of $x_e(z)$ the high-$z$ optical depth can be detected at 

1406: a high level of confidence and measured to an accuracy of $\sim 0.01$.

1407: Note that the constraints in both Figures~\ref{fig:taucv1} 

1408: and~\ref{fig:taucv2} are consistent with the \wmap\ constraints 

1409: in Figure~\ref{fig:tauwmap}.

1410: 

1411: 

1412: As with \wmap\ data, most of the 

1413: constraints from simulated data in Figures~\ref{fig:taucv1} 

1414: and~\ref{fig:taucv2} do not appear to depend strongly on the 

1415: choice of $\zmax$. For the instantaneous reionization model used for 

1416: Figure~\ref{fig:taucv1}, this is to be expected since we know that the 

1417: simulated $\clee$ have no contribution from $z \gtrsim 20$. 

1418: Each set of contours in Figure~\ref{fig:taucv1} is consistent with 

1419: the values of $\tau(6,15)$ and $\tau(15,\zmax)$ for $\xet(z)$. 

1420: The constraints in the second panel in Figure~\ref{fig:taucv1} 

1421: extend to larger high-$z$ optical depth than those in the first panel 

1422: because the effect of the physicality priors is not as strong at 

1423: larger $\zmax$, as noted in \S~\ref{sec:wmap}; the leftmost 

1424: $\zmax=20$ contours would stretch to larger $\tau(15,20)$ and 

1425: smaller $\tau(6,15)$ if we did not apply the physicality bounds.

1426: The two panels on the right side of Figure~\ref{fig:taucv1} show 

1427: weaker constraints on the partial optical depths because 

1428: the principal component constraints for that realization of $\clee$ 

1429: are weaker.

1430: 

1431: In the extended, double reionization model used for 

1432: Figure~\ref{fig:taucv2}, reionization 

1433: actually starts at $z=23$, so the choice of $\zmax=20$ is not 

1434: appropriate for this model. The result of such an error is that 

1435: the constraint on $\tau(15,20)$ is inconsistent with the true 

1436: value, marked by a cross in the left panel of Figure~\ref{fig:taucv2}. 

1437: The $\zmax=30$ constraints in the right panel are consistent with 

1438: the expected value, although not in perfect agreement due to 

1439: cosmic variance. It is interesting that although the value chosen for 

1440: $\zmax$ in the left panel is too low for this model, the 

1441: constraints on $\tau(z>15)$ for $\zmax=20$ and $\zmax=30$ are still consistent 

1442: with each other.  

1443: This supports the interpretation of constraints on $\tau(\zmid,\zmax)$

1444: as $\tau(z>\zmid)$ in \S~\ref{sec:wmap}  

1445: 

1446: 

1447: While the cosmic variance-limited simulated data are useful for 

1448: determining how well the reionization history could possibly be 

1449: constrained by large-scale $E$-mode polarization measurements, it is 

1450: also interesting to ask how well we can do with future experiments 

1451: that fall somewhat short of the idealized case that we have considered 

1452: so far. In particular, the upcoming \emph{Planck} satellite is 

1453: expected to improve our knowledge of the large-scale $E$-mode spectrum 

1454: substantially \citep{Planck}; what does this imply for constraints on $x_e(z)$?

1455: To estimate what might be possible with \emph{Planck} data, 

1456: we assume that after subtracting foregrounds 

1457: a single foreground-free frequency channel remains 

1458: for constraining the low-$\ell$ 

1459: $E$-mode polarization. We take this to be the 143 GHz channel with 

1460: a white noise power level of $w_P^{-1/2}=81~\mu$K$'$ and 

1461: beam size $\theta_{\rm FWHM}=7.1'$, and we assume that the sky coverage is 

1462: $f_{\rm sky}=0.8$ after cutting out the Galactic 

1463: plane~\citep{Planck,Albetal06}. We compute the likelihood of Monte Carlo 

1464: samples using the routines provided in CosmoMC 

1465: and analyze parameter chains with the principal component method as 

1466: described for \wmap\ and cosmic variance-limited data.

1467: 

1468: For various choices of $\xef(z)$ and $\xet(z)$ we find that going from 

1469: full-sky cosmic variance-limited data to \emph{Planck}-like data 

1470: increases the uncertainty in $x_e(z)$ principal components, $\sigma_{\mu}$,

1471: and in the optical depth, $\sigma_{\tau}$, by roughly a factor of two.

1472: Based on these results, it should be possible to constrain about  

1473: three eigenmodes at $\sim 2~\sigma$ 

1474: within the space of physical models, and we expect 

1475: $\tau$ to be determined to an accuracy of about $\pm 0.01$.

1476: This is consistent with previous estimates of the \emph{Planck} 

1477: optical depth uncertainty when considering more general reionization 

1478: models than instantaneous reionization \citep[e.g.,][]{Holetal03,HuHol03}.

1479: 

1480: 

1481: 

1482: % --------------------------------------------

1483: \subsection{Fisher matrix predictions versus MCMC results}

1484: \label{sec:fisher}

1485: 

1486: 

1487: The Fisher matrix analysis of \S~\ref{sec:eigenmodes} predicts that 

1488: the principal components of $x_e(z)$ should be uncorrelated, with 

1489: errors given by $\sigma_{\mu}$ from equation~(\ref{eq:fisher2}) 

1490: (in the idealized limit of full sky, noiseless observations). 

1491: These characteristics rely on the assumption that the difference between 

1492: the true and fiducial reionization histories is small: 

1493: $\delta \xet(z)=\xet(z)-\xef(z) \ll 1$. Clearly this assumption will 

1494: not hold for any single fiducial history if one wants to consider 

1495: a variety of possible true histories, and the consequence is that 

1496: $\langle m_{\mu}m_{\nu} \rangle$ differs from $\sigma_{\mu}^2 \delta_{\mu\nu}$.

1497: 

1498: 

1499: % ****************************************

1500: \begin{figure}

1501: \centerline{\psfig{file=f13.eps, width=3.0in}}

1502: \caption{Monte Carlo constraints in the $m_1$-$m_2$ plane from the 

1503: left panel of Figure~\ref{fig:mcmccv} (68 and 95\% contours) compared 

1504: with the estimated constraints from the Fisher matrix 

1505: (\emph{shaded contours}).

1506: \vskip 0.25cm

1507: }

1508: \label{fig:fisher2d}

1509: \end{figure}

1510: % ****************************************

1511: 

1512: 

1513: As an example of the difference between the Fisher approximation and 

1514: the full Monte Carlo analysis, in Figure~\ref{fig:fisher2d} 

1515: the marginalized constraints in the $m_1-m_2$ plane from the 

1516: left panel of Figure~\ref{fig:mcmccv} 

1517: are compared with the Fisher matrix error ellipses, centered on 

1518: the true $(m_1,m_2)$ 

1519: for the same choices of $\xet(z)$ (instantaneous reionization 

1520: with $\tau=0.105$) and $\xef(z)$ (constant $x_e=0.15$ out to $\zmax=20$). 

1521: The MCMC constraints have correlated values of $m_1$ and $m_2$ with

1522: somewhat larger uncertainties than the Fisher matrix $\{\sigma_{\mu}\}$. 

1523: The eigenfunctions, $S_{\mu}(z)$, are constructed to have orthogonal 

1524: effects on $\clee$ in 

1525: the vicinity of the fiducial model, but for large $\delta x_e(z)$  

1526: the orthogonality breaks down.  Fortunately the errors on $\tau$ 

1527: remain largely unaffected because the

1528: $m_{1}-m_{2}$ correlation is oriented 

1529: such that the increased errors

1530: are along the direction of constant total $\tau$ (see Figure~\ref{fig:tauinst}).

1531: 

1532: One can compute new eigenfunctions 

1533: by evaluating ${\partial \ln \clee}/{\partial x_e(z)}$ at $\xet(z)$ 

1534: and see that their dependence on redshift 

1535: differs from the original $S_{\mu}(z)$. 

1536: It is possible to decorrelate the principal components by 

1537: diagonalizing the covariance matrix from the Monte Carlo chains 

1538: and rotating the eigenmodes into the new basis. 

1539: However, this procedure is unnecessary for most applications; 

1540: the most important property of the principal components for 

1541: constraining $x_e(z)$ is that the first few modes form a complete basis 

1542: for $\clee$ on large scales, accurate within cosmic variance limits. 

1543: Completeness of the eigenmodes holds true even if they are not

1544: exactly orthogonal.

1545: 

1546: Finally, note that if the true ionization history and the draw of $\clee$ 

1547: permit constraints on principal components similar to those in the 

1548: left panel of Figure~\ref{fig:mcmccv}, then these constraints may be 

1549: close enough to Gaussian so that the covariance matrix would be 

1550: sufficient to describe the information in the $\clee$ for 

1551: finding constraints on reionization models. However, if what we see looks 

1552: more like the right panel of Figure~\ref{fig:mcmccv} then the full 

1553: chains of Monte Carlo samples may be necessary for further applications 

1554: even if we can measure the $E$-mode polarization perfectly.

1555: 

1556: 

1557: 

1558: 

1559: % =====================================================

1560: \section{Discussion}

1561: \label{sec:discuss}

1562: 

1563: Observations of the large-scale $E$-mode polarization of the CMB in 

1564: the near future are expected to yield new information 

1565: about the spatially-averaged reionization history of the universe. 

1566: The principal components of the reionization history are a 

1567: promising tool for extracting as much of that information as possible 

1568: from the data. We have shown that the principal component method 

1569: can be usefully applied to real, currently available data, and 

1570: forecasts from simulated data suggest that there is room to 

1571: substantially improve constraints on the reionization history 

1572: using this method as measurements of the large-scale $E$-modes improve.

1573: 

1574: We find that the key features of the principal component analysis 

1575: put forward by \cite{HuHol03} continue to apply when we go from 

1576: the Fisher matrix approximation to an exploration of the full 

1577: likelihood surface using Markov Chain Monte Carlo methods. 

1578: For fairly conservative choices of the maximum redshift of reionization 

1579: ($\zmax \sim 30$-40), only the first five principal components at most 

1580: are needed for a complete representation of the $E$-mode angular 

1581: power spectrum to within cosmic variance. To account for 

1582: arbitrary reionization histories in the analysis of CMB data, only a 

1583: few additional parameters must be included in chains of Monte Carlo 

1584: samples if those parameters are taken to be the lowest-variance 

1585: principal components of $x_e(z)$. 

1586: 

1587: Specific models of reionization can be tested easily by computing 

1588: their eigenmode amplitudes and comparing with constraints on the 

1589: eigenmodes from the data. Constraints on derived parameters, 

1590: such as total optical depth or the optical depth from a certain 

1591: range in redshift, represent other applications of the 

1592: MCMC constraints on principal components of $x_e(z)$.

1593: 

1594: Often, estimates of the optical depth to reionization 

1595: are computed assuming instantaneous reionization or some other simple 

1596: form for $x_e(z)$. Here we extend the analysis of the 3-year \wmap\ 

1597: data to allow a more general set of models of the global 

1598: reionization history. We find that expanding the model space does not 

1599: significantly widen the uncertainty in $\tau$ beyond the 

1600: instantaneous reionization value of $\sigma_{\tau} = 0.03$. 

1601: Robust $\tau$ constraints are important for tests of the dark energy

1602: based on the growth of structure since they control the uncertainty

1603: on the amplitude of the initial spectrum.

1604: 

1605: Moreover, even with current data the principal component constraints 

1606: are beginning to show the possibility of determining properties of 

1607: reionization in addition to $\tau$. By comparing the optical depth from 

1608: high $z$ with that from low $z$, we obtain an upper limit on the 

1609: contribution to the optical depth from high redshift: 

1610: $\tau(z>20) < 0.08$ at 95\% confidence, assuming that there is no 

1611: significant episode of reionization at $z\gg 40$.

1612: 

1613: Due to the limitations of noise and foreground contamination, 

1614: only the first two eigenmodes of the reionization 

1615: history, $m_1$ and $m_2$, can be determined with present 

1616: polarization data to any reasonable degree of 

1617: accuracy. Constraints on these two modes come primarily from the 

1618: main, broad peak in $E$-mode power at low $\ell$ from reionization. 

1619: 

1620: As measurements of the $E$-mode polarization improve, for example from 

1621: additional \wmap\ data or through planned future experiments such as 

1622: \emph{Planck}, better knowledge of the low-power ``trough'' in $\clee$ 

1623: between the main reionization peak and the first acoustic peak 

1624: should enable constraints on the third and higher principal components, 

1625: up to about $m_5$ for near cosmic variance-limited data. 

1626: Since constraints on these higher modes rely on the ability to identify 

1627: subtle features in the trough of $\clee$, the ultimate accuracy to 

1628: which the eigenmodes can be determined may depend on whether or not 

1629: the necessary features are well reproduced in the particular 

1630: random draw of $\clee$ that is available to us. However, even if we 

1631: are unlucky enough to have a realization in which some of the important 

1632: features of the spectrum are washed out by randomness, it should still 

1633: be possible to measure several of the principal components to 

1634: better accuracy than is currently possible. Knowledge of the $\mu\geq 3$ 

1635: eigenmodes of $x_e(z)$, along with improved 

1636: constraints on $m_1$ and $m_2$, will 

1637: allow more stringent tests of reionization models and a better 

1638: understanding of the global reionization history.

1639: 

1640: 

1641: 

1642: 

1643: \acknowledgements {\it Acknowledgments}:  We thank Cora Dvorkin,

1644: Gilbert Holder, Dragan Huterer, Hiranya Peiris, and Jochen Weller

1645: for useful discussions.  MJM was supported by a 

1646: National Science Foundation Graduate Research Fellowship. 

1647: WH was supported by the KICP

1648: through the grant NSF PHY-0114422, the DOE through 

1649: contract DE-FG02-90ER-40560 and the David and

1650: Lucile Packard Foundation. 

1651: 

1652: 

1653: %\bibliographystyle{apj_hyperref}

1654: \bibliographystyle{apj}

1655: %\bibliography{ms}

1656: 

1657: \begin{thebibliography}{27}

1658: \expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi

1659: 

1660: \bibitem[{{Albrecht} {et~al.}(2006){Albrecht}, {Bernstein}, {Cahn}, {Freedman},

1661:   {Hewitt}, {Hu}, {Huth}, {Kamionkowski}, {Kolb}, {Knox}, {Mather}, {Staggs},

1662:   \& {Suntzeff}}]{Albetal06}

1663: {Albrecht}, A., {Bernstein}, G., {Cahn}, R., {Freedman}, W.~L., {Hewitt}, J.,

1664:   {Hu}, W., {Huth}, J., {Kamionkowski}, M., {Kolb}, E.~W., {Knox}, L.,

1665:   {Mather}, J.~C., {Staggs}, S., \& {Suntzeff}, N.~B. 2006, ArXiv Astrophysics

1666:   e-prints astro-ph/0609591

1667: 

1668: \bibitem[{{Barkana} \& {Loeb}(2001)}]{BarLoe01}

1669: {Barkana}, R. \& {Loeb}, A. 2001, \physrep, 349, 125

1670: 

1671: \bibitem[{{Brooks} \& {Gelman}(1998)}]{BroGel98}

1672: {Brooks}, S.~P. \& {Gelman}, A. 1998, Journal of Computational and Graphical

1673:   Statistics, 7, 434

1674: 

1675: \bibitem[{Christensen {et~al.}(2001)Christensen, Meyer, Knox, \&

1676:   Luey}]{Chretal01}

1677: Christensen, N., Meyer, R., Knox, L., \& Luey, B. 2001, Class. Quant. Grav.,

1678:   18, 2677

1679: 

1680: \bibitem[{{Colombo} {et~al.}(2005){Colombo}, {Bernardi}, {Casarini}, {Mainini},

1681:   {Bonometto}, {Carretti}, \& {Fabbri}}]{Coletal05}

1682: {Colombo}, L.~P.~L., {Bernardi}, G., {Casarini}, L., {Mainini}, R.,

1683:   {Bonometto}, S.~A., {Carretti}, E., \& {Fabbri}, R. 2005, \aap, 435, 413

1684: 

1685: \bibitem[{{Dor{\'e}} {et~al.}(2007){Dor{\'e}}, {Holder}, {Alvarez}, {Iliev},

1686:   {Mellema}, {Pen}, \& {Shapiro}}]{Doretal07}

1687: {Dor{\'e}}, O., {Holder}, G., {Alvarez}, M., {Iliev}, I.~T., {Mellema}, G.,

1688:   {Pen}, U.-L., \& {Shapiro}, P.~R. 2007, ArXiv Astrophysics e-prints

1689:   astro-ph/0701784

1690: 

1691: \bibitem[{Dunkley {et~al.}(2005)Dunkley, Bucher, Ferreira, Moodley, \&

1692:   Skordis}]{Dunetal04}

1693: Dunkley, J., Bucher, M., Ferreira, P.~G., Moodley, K., \& Skordis, C. 2005,

1694:   Mon. Not. Roy. Astron. Soc., 356, 925

1695: 

1696: \bibitem[{Fan {et~al.}(2006)Fan, Carilli, \& Keating}]{FanCarKea06}

1697: Fan, X.-H., Carilli, C.~L., \& Keating, B. 2006, Ann. Rev. Astron. Astrophys.,

1698:   44, 415

1699: 

1700: \bibitem[{{Gelman} \& {Rubin}(1992)}]{GelRub92}

1701: {Gelman}, A. \& {Rubin}, D.~B. 1992, Statistical Science, 7, 457

1702: 

1703: \bibitem[{{Holder} {et~al.}(2003){Holder}, {Haiman}, {Kaplinghat}, \&

1704:   {Knox}}]{Holetal03}

1705: {Holder}, G.~P., {Haiman}, Z., {Kaplinghat}, M., \& {Knox}, L. 2003, \apj, 595,

1706:   13

1707: 

1708: \bibitem[{{Hu}(2000)}]{Hu00}

1709: {Hu}, W. 2000, \apj, 529, 12

1710: 

1711: \bibitem[{{Hu} \& {Holder}(2003)}]{HuHol03}

1712: {Hu}, W. \& {Holder}, G.~P. 2003, \prd, 68, 023001

1713: 

1714: \bibitem[{{Hu} \& {White}(1997)}]{HuWhi97a}

1715: {Hu}, W. \& {White}, M. 1997, \apj, 479, 568

1716: 

1717: \bibitem[{{Iliev} {et~al.}(2006){Iliev}, {Pen}, {Richard Bond}, {Mellema}, \&

1718:   {Shapiro}}]{Ilietal06a}

1719: {Iliev}, I.~T., {Pen}, U.-L., {Richard Bond}, J., {Mellema}, G., \& {Shapiro},

1720:   P.~R. 2006, New Astronomy Review, 50, 909

1721: 

1722: \bibitem[{{Kaplinghat} {et~al.}(2003){Kaplinghat}, {Chu}, {Haiman}, {Holder},

1723:   {Knox}, \& {Skordis}}]{Kapetal03}

1724: {Kaplinghat}, M., {Chu}, M., {Haiman}, Z., {Holder}, G.~P., {Knox}, L., \&

1725:   {Skordis}, C. 2003, \apj, 583, 24

1726: 

1727: \bibitem[{{Kasuya} {et~al.}(2004){Kasuya}, {Kawasaki}, \&

1728:   {Sugiyama}}]{KasKawSug04}

1729: {Kasuya}, S., {Kawasaki}, M., \& {Sugiyama}, N. 2004, \prd, 69, 023512

1730: 

1731: \bibitem[{Kosowsky {et~al.}(2002)Kosowsky, Milosavljevic, \&

1732:   Jimenez}]{KosMilJim02}

1733: Kosowsky, A., Milosavljevic, M., \& Jimenez, R. 2002, Phys. Rev., D66, 063007

1734: 

1735: \bibitem[{{Lewis} \& {Bridle}(2002)}]{LewBri02}

1736: {Lewis}, A. \& {Bridle}, S. 2002, \prd, 66, 103511

1737: 

1738: \bibitem[{{Lewis} {et~al.}(2000){Lewis}, {Challinor}, \& {Lasenby}}]{Lewetal00}

1739: {Lewis}, A., {Challinor}, A., \& {Lasenby}, A. 2000, \apj, 538, 473

1740: 

1741: \bibitem[{{Lewis} {et~al.}(2006){Lewis}, {Weller}, \& {Battye}}]{LewWelBat06}

1742: {Lewis}, A., {Weller}, J., \& {Battye}, R. 2006, \mnras, 373, 561

1743: 

1744: \bibitem[{{Mortonson} \& {Hu}(2007)}]{MorHu07}

1745: {Mortonson}, M.~J. \& {Hu}, W. 2007, \apj, 657, 1

1746: 

1747: \bibitem[{{Naselsky} \& {Chiang}(2004)}]{NasChi04}

1748: {Naselsky}, P. \& {Chiang}, L.-Y. 2004, \mnras, 347, 795

1749: 

1750: \bibitem[{{Page} {et~al.}(2006){Page}, {Hinshaw}, {Komatsu}, {Nolta},

1751:   {Spergel}, {Bennett}, {Barnes}, {Bean}, {Dor\'{e}}, {Halpern}, {Hill},

1752:   {Jarosik}, {Kogut}, {Limon}, {Meyer}, {Odegard}, {Peiris}, {Tucker}, {Verde},

1753:   {Weiland}, {Wollack}, \& {Wright}}]{Pagetal06}

1754: {Page}, L., {Hinshaw}, G., {Komatsu}, E., {Nolta}, M.~R., {Spergel}, D.~N.,

1755:   {Bennett}, C.~L., {Barnes}, C., {Bean}, R., {Dor\'{e}}, O., {Halpern}, M.,

1756:   {Hill}, R.~S., {Jarosik}, N., {Kogut}, A., {Limon}, M., {Meyer}, S.~S.,

1757:   {Odegard}, N., {Peiris}, H.~V., {Tucker}, G.~S., {Verde}, L., {Weiland},

1758:   J.~L., {Wollack}, E., \& {Wright}, E.~L. 2006, ArXiv Astrophysics e-prints

1759:   astro-ph/0603450

1760: 

1761: \bibitem[{Raftery \& Lewis(1992)}]{RafLew92}

1762: Raftery, A.~E. \& Lewis, S.~M. 1992, in Bayesian Statistics, ed. J.~M. Bernado

1763:   (OUP), 765

1764: 

1765: \bibitem[{{Spergel} {et~al.}(2006){Spergel}, {Bean}, {Dor\'{e}}, {Nolta},

1766:   {Bennett}, {Hinshaw}, {Jarosik}, {Komatsu}, {Page}, {Peiris}, {Verde},

1767:   {Barnes}, {Halpern}, {Hill}, {Kogut}, {Limon}, {Meyer}, {Odegard}, {Tucker},

1768:   {Weiland}, {Wollack}, \& {Wright}}]{Speetal06}

1769: {Spergel}, D.~N., {Bean}, R., {Dor\'{e}}, O., {Nolta}, M.~R., {Bennett}, C.~L.,

1770:   {Hinshaw}, G., {Jarosik}, N., {Komatsu}, E., {Page}, L., {Peiris}, H.~V.,

1771:   {Verde}, L., {Barnes}, C., {Halpern}, M., {Hill}, R.~S., {Kogut}, A.,

1772:   {Limon}, M., {Meyer}, S.~S., {Odegard}, N., {Tucker}, G.~S., {Weiland},

1773:   J.~L., {Wollack}, E., \& {Wright}, E.~L. 2006, ArXiv Astrophysics e-prints

1774:   astro-ph/0603449

1775: 

1776: \bibitem[{{The Planck Collaboration}(2006)}]{Planck}

1777: {The Planck Collaboration}. 2006, ArXiv Astrophysics e-prints astro-ph/0604069

1778: 

1779: \bibitem[{{Zaldarriaga}(1997)}]{Zal97}

1780: {Zaldarriaga}, M. 1997, \prd, 55, 1822

1781: 

1782: \end{thebibliography}

1783: 

1784: \end{document}

1785: