1: % aa.dem
2: % AA vers. 6.1, LaTeX class for Astronomy & Astrophysics
3: % demonstration file
4: % (c) Springer-Verlag HD
5: % revised by EDP Sciences
6: %-----------------------------------------------------------------------
7: %
8: %\documentclass[referee]{aa} % for a referee version
9: %\documentclass[onecolumn]{aa} % for a paper on 1 column
10: %\documentclass[longauth]{aa} % for the long lists of affiliations
11: %\documentclass[rnote]{aa} % for the research notes
12: %\documentclass[letter]{aa} % for the letters
13: %
14: \documentclass[structabstract,usenatbib]{aa}
15: %\documentclass[tradiabstract]{aa} % for the abstract without structuration
16: % (traditional abstract)
17: %
18: %CHM
19: %\newcommand{\vt}{\mbox{\bf {T}}}
20: \newcommand{\vtcmb}{\mbox{\bf {T}}}
21: %\newcommand{\vm}{\mbox{\bf {M}}}
22: \newcommand{\vnn}{\mbox{N}}
23: \newcommand{\vn}{\mbox{\bf {n}}}
24: \newcommand{\vf}{\mbox{\bf {F}}}
25: \newcommand{\1}{\'\i }
26: \newcommand{\vx}{\mbox{\bf {x}}}
27: \newcommand{\vy}{\mbox{\bf {y}}}
28: \newcommand{\vk}{\mbox{\bf {k}}}
29: \newcommand{\vs}{{\bf {s}}}
30: \newcommand{\vm}{{\bf {m}}}
31: \newcommand{\vt}{{\bf {t}}}
32: \newcommand{\vkt}{\hat{ \mbox{\bf {k}}}}
33: \newcommand{\vzt}{\hat{ \mbox{\bf {z}}}}
34: \newcommand{\vq}{\mbox{\bf {p}}}
35: \newcommand{\va}{\mbox{\bf {a}}}
36: \newcommand{\vp}{\mbox{\bf {p}}}
37: \newcommand{\vrv}{\mbox{\bf {r}}}
38: \newcommand{\vu}{\mbox{\bf {u}}}
39: \newcommand{\vv}{\mbox{\bf {v}}}
40: \newcommand{\valpha}{{\bf{\alpha}}}
41: \newcommand{\lmax}{l_{max}}
42: \newcommand{\lmin}{l_{min}}
43: \newcommand{\vnh}{\hat{\mbox{\bf {n}}}}
44: \newcommand{\np}{\hat{\mbox{\bf {n}}}'}
45: \newcommand{\m}{\mbox{\bf {m}}}
46: %\newcommand{\vv}{\mbox{\bf {v}}}
47: \newcommand{\tonetwo}{\theta_{12}}
48: \newcommand{\ctonetwo}{\cos{\theta_{12}}}
49: \newcommand{\gsim}{\; ^{>}_{\sim}\;}
50:
51: \newcommand{\dthreek}{\frac{d\vk}{(2\pi)^3}}
52: \newcommand{\dthreeq}{\frac{d\vq}{(2\pi)^3}}
53: \newcommand{\dthreeu}{\frac{d\vu}{(2\pi)^3}}
54: \newcommand{\dthreekrad}{\frac{(4\pi)^2\;k^2dk}{(2\pi)^3}}
55: \newcommand{\dthreekr}{k^2dk}
56: \newcommand{\parone}{\frac{\partial {\bar n}(M_1,z_1)}{\partial \delta}\biggr|_{\delta =0}}
57: \newcommand{\partwo}{\frac{\partial {\bar n}(M_2,z_2)}{\partial \delta}\biggr|_{\delta =0}}
58: \newcommand{\parzero}{\frac{\partial n(M,r)}{\partial \delta}\biggr|_{\delta =0}}
59: \newcommand{\intrc}{\int_{0}^{r_{LSS}}}
60: \newcommand{\taud}{\dot{\tau}}
61: \usepackage{multirow}
62: \def\plotone#1{\centering \leavevmode
63: \epsfxsize=\columnwidth \epsfbox{#1}}
64: \def\plotancho#1{\includegraphics[width=18cm]{#1}}
65:
66: %
67: %
68: \usepackage{graphicx}
69: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
70: \usepackage{txfonts}
71: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
72: %
73: \begin{document}
74: %
75: \title{Implementation of a Fourier Matched Filter in CMB Analyses. Application
76: to ISW Studies.}
77:
78: %\subtitle{I. Overviewing the $\kappa$-mechanism}
79:
80: \author{C. Hern\'andez--Monteagudo
81: %\inst{1}%\fnmsep%\thanks{Just to show the usage
82: %of the elements in the author field}
83: }
84:
85: \institute{Max Planck Institut f\"ur Astrophysik,
86: Karl Schwarzschild Str.1, D-85741,
87: Garching bei M\"unchen, Germany\\
88: \email{chm@mpa-garching.mpg.de}
89: }
90:
91: \date{Received; accepted}
92:
93: % \abstract{}{}{}{}{}
94: % 5 {} token are mandatory
95:
96: \abstract % context heading (optional)
97: {}
98: %leave it empty if necessary %{ }
99: % aims heading (mandatory)
100: {Implement a matched filter (MF)
101: cross-correlation algorithm in multipole space and compare it
102: to the standard Angular Cross Power Spectrum (ACPS) method. Apply both
103: methods on a Integrated Sachs Wolfe (ISW) - Large Scale Structure (LSS)
104: cross correlation scenario and study how sky masks influence the multipole
105: range where the cross correlation signal arises and its comparison to
106: theoretical predictions. }
107: % methods heading (mandatory)
108: {The MF requires the inversion of a multipole covariance matrix
109: that, under non-full sky coverage ($f_{sky}<1$), is generally non-diagonal
110: and singular. We choose a Singular Value Decomposition (SVD) approach
111: that enables the identification of those modes carrying most of
112: the information from those more likely to introduce numerical noise, (that
113: are dropped from the analysis). We compare the MF to the ACPS in ISW-LSS
114: Monte Carlo simulations, focusing on the effect that a limited sky coverage
115: has on the cross-correlation results. }
116: % results heading (mandatory)
117: {
118: Within the data model $\vs = \vt + \alpha \vm$ where $\vt$ is Gaussian
119: noise and $\vm$ is a known filter,
120: we find that the MF performs comparatively better than the ACPS for smaller
121: values of $f_{sky}$ and scale dependent (non-Poissonian) noise fields. In
122: the context of ISW studies both methods are comparable, although the MF
123: performs slightly more sensitively under more restrictive masks (smaller
124: values of $f_{sky}$). A preliminary analytical study of the ISW--LSS cross
125: correlation signal to noise (S/N) ratio shows that most of it should be
126: found in the very large scales (50\% of the S/N at $l<10$, 90\% at
127: $l<40-50$), and this is confirmed by Monte Carlo simulations. The
128: statistical significance of our cross-correlation statistics reaches its
129: maximum when considering $l\in [2,l_{max}]$, with $l_{max} \in[5,40]$ for
130: all values of $f_{sky}$ observed, despite of the smoothing and power
131: aliasing that aggressive masks introduce in Fourier space. This
132: $l$-confinement of the ISW-LSS cross correlation should
133: enable a safe distinction from other secondary effects arising at smaller
134: (higher $l$-s) angular scales. }
135: % conclusions heading (optional), leave it empty if necessary
136: {}
137:
138: \keywords{(Cosmology) : cosmic microwave background, Large Scale Structure of the Universe
139: }
140: \titlerunning{Matched Filter in Multipole Space and ISW Studies}
141:
142: \maketitle
143: %
144: %________________________________________________________________
145: %CHM
146:
147:
148:
149: %CHM
150:
151:
152: \section{Introduction}
153:
154: Auto and cross-correlation analyses are crucial in the study of the Cosmic
155: Microwave Background (CMB) anisotropies. This is due not only to the fact that
156: the theory can only predict statistical properties of the intrinsic intensity
157: and polarization anisotropies (and hence auto-correlation tests must be
158: conducted in order to compare theory to observations, see \cite{hudodelson}
159: for a review), but also due to the presence of secondary anisotropies and
160: foreground emission that add up to the measurements in the microwave
161: range. These other components must be identified and separated from the
162: intrinsic ones generated at the surface of last scattering, and therefore
163: cross-correlation analyses to other data sets probing the sources of this
164: secondary emission must be carried out. This has been done practically for all
165: CMB experiments, from COBE data (\cite{cobe1,cobe2}) all the way to WMAP data
166: (\cite{wmap1,spergel06}). These cross correlation techniques may be either
167: based in real space (like the angular two point correlation function), in
168: Fourier space (like the Auto and Cross Angular Power Spectrum), or in wavelet
169: space (\cite{laura,larson}).\\
170:
171: In the linear theory that characterizes the intensity and polarization
172: anisotropies of the CMB, predictions are done in the Fourier space of the 2D
173: sphere, that is, in multipole space. In this space the statistical covariance
174: matrices between different modes are particularly simple, and so is the
175: comparison of theory to observations. It is in this space where theoretical
176: expectations for other secondary effects present in the CMB are also
177: displayed, and where the constraints on the cosmological parameters are set
178: (e.g., \cite{dunkley,acbar}). However, there are two practical issues that
179: tend to complicate this theory to data comparison: the presence of Cosmic
180: Variance in the large angular scales (that is, the sample variance due to
181: having only one single sky to look at) and the coupling of different Fourier
182: modes whenever {\em not} the entire sky is subject to analysis (as it happens
183: in practice for current and future CMB and LSS surveys like ACT (\cite{act}),
184: SPT (\cite{spt}), DUNE\footnote{{\tt http://www.dune-mission.net/}},
185: SNAP\footnote{{\tt http://snap.lbl.gov/}} etc). These two effects are of
186: particular relevance in the study of the Integrated Sachs-Wolfe (ISW) effect
187: (\cite{critturok96}): the ISW arises in the large angular scales, and since
188: its frequency dependence is identical to that of the intrinsic CMB
189: fluctuations, it must be identified via cross-correlation tests to Large Scale
190: Structure surveys that are likely to cover only a fraction of the sky.\\
191:
192: In this work we generalize the matched filter cross-correlation method to
193: multipole space in the context of CMB studies. We compare it to the standard
194: Angular Cross Power Spectrum in different scenarios, and show that the former
195: is either equivalent or superior to the latter. We also perform this
196: comparison in Monte Carlo simulations of the ISW effect, with similar
197: results. The method is developed in Section (2), whereas a first comparison to
198: the Angular Cross Power Spectrum is given in Section (3). A detailed analysis
199: of the signal to noise ratio of ISW cross-correlation measurements is provided
200: in Section (4), where the matched filter method is again compared to the
201: Angular Cross Power Spectrum. Finally, in Section (5) we discuss our results
202: and conclude.
203:
204: \section{The matched filter method}
205:
206: \subsection{A brief description}
207: \label{sec:method}
208:
209: Our first goal is to estimate the level of presence of some known signal $\vm$
210: in some measured data array $\vs$, which is therefore decomposed as $\vs = \vt
211: + \alpha \vm$. We shall assume that $\vt$ is a Gaussian vector (which will be
212: regarded as {\em noise}) whose covariance matrix $\bf{C}$ is known. Given
213: the Gaussian assumption, $\bf{C}$ completely characterizes $\vt$. As shown in,
214: e.g., \cite{alphamethod}, the minimization of the quantity
215: \begin{equation}
216: \chi^2 \equiv \sum_{i,j} (\vs - \alpha \vm)_i ({\bf C}^{-1} )_{i,j}(\vs - \alpha \vm)_j,
217: \label{eq:chisq1}
218: \end{equation}
219: yields the following estimates for $\alpha$ and its formal error:
220: \begin{equation}
221: \hat{\alpha} = \frac{\vt^t {\bf C}^-1\vm}{\vm^t {\bf C}^{-1}\vm}, \;\;\;\;\,
222: \hat{\sigma}_{\alpha}^2 = \frac{1}{\vm^t {\bf C}^{-1}\vm}.
223: \label{eq:mf1}
224: \end{equation}
225: Note that the superscript $t$ denotes {\it transpose}. The difficulty usually
226: lies in the inversion of the covariance matrix for long data arrays $\vt$
227: and/or for close-to-singular covariance matrices ${\bf C}$. The first scenario
228: was already addressed in \cite{metals06}, where this technique was applied in
229: separated subsets of data, and then the covariance among different subsets was
230: computed separately. Here, we shall also consider the case where ${\bf C}$ is
231: singular or close to singular.\\
232:
233: Indeed, the use of the matched filter is very extended in CMB analyses
234: (e.g., \cite{jal,tsz1,tsz2,hansen}), but it has been mostly restricted to real
235: space. In works like that of \cite{hansen} it was also implemented in Fourier
236: (multipole) space, but only after approximating the covariance matrix as
237: diagonal, assumption that we shall avoid here.\\
238:
239: \subsection{The covariance matrix in multipole space}
240: \label{sec:cov}
241: We will focus our analyses on real signals defined on 2D spheres. These are usually decomposed
242: on an spherical harmonic basis as follows:
243: \begin{equation}
244: \vs (\vnh ) = \sum_{l=l_{min}}^{l_{max}}\sum_{m=-l}^{l} a_{l,m} Y_{l,m}(\vnh ),
245: \label{eq:sp1}
246: \end{equation}
247: with $\vnh$ denoting a direction on the sky (or a position in the sphere). If
248: $\vs$ is real, then the multipole coefficients verify $a_{l,-m} = (-1)^m
249: a_{l,m}^*$, with the symbol "$*$" denoting {\it complex conjugate}. This
250: limits the number of degrees of freedom per $l$ to $2l+1$. The $m=0$ multipole
251: is by definition real, so the $2l$ remaining degrees of freedom can be
252: assigned to the real and imaginary parts of the $a_{l,m}$ multipoles with
253: magnetic number ranging from $m=1$ to $m=l$. I.e., for a given multipole $l$,
254: the multipole array $a_{l,m}$ will be decomposed into a $2l+1$ dimensioned
255: array $(u_0^l,u_1^l,...,u_l^l,v_0^l,..,v_l^l)$, where $u_0^l \equiv a_{l,0}$,
256: $u_1^l,...,u_l^l $ contain the real parts of $a_{l,m=1,l}$, and
257: $v_1^l,...,v_l^l$ the imaginary ones. Since we will simultaneously consider
258: all multipoles $l \in [l_{min},l_{max}]$, we define the multipole array
259: \begin{equation}
260: \va = (\vu_0, \vu, \vv),
261: \label{eq:va_def}
262: \end{equation}
263: where
264: \begin{eqnarray}
265: \vu_0 & \equiv & (u_0^{l_{min}},u_0^{l_{min}+1},...,u_0^{l_{max}}), \\
266: \vu & \equiv & (u_1^{l_{min}},...,u_{l_{min}}^{\lmin},...,u_1^{\lmax},...,u_{\lmax}^{\lmax}),
267: \label{eq:us}
268: \end{eqnarray}
269: and
270: \begin{equation}
271: \vv \equiv (v_1^{l_{min}},...,v_{l_{min}}^{\lmin},...,v_1^{\lmax},...,v_{\lmax}^{\lmax}).
272: \label{eq:vv}
273: \end{equation}
274:
275: The dimension of $\vu$ and $\vv$ is given by $n_{l,2}=
276: \lmax(\lmax+1)/2+\lmax+1 - ( \lmin(\lmin-1)+\lmin) $, whereas the dimension of
277: $\vu_0$ is simply $n_{l,1} = \lmax - \lmin + 1$, so the total dimension of
278: $\va$ reads $n_l = n_{l,1} + 2\times n_{l,2}$. If $\vs (\vnh)$ is an
279: isotropic, Gaussian distributed signal over the {\em whole} sphere
280: ($f_{sky}=1$), then the correlation matrix of the $a_{l,m}$ coefficients is
281: diagonal: $\langle a_{l,m} (a_{l',m'})^*\rangle = C_l \delta_{l,l'}
282: \delta_{m,m'}$. (Note that due to isotropy there is no dependence on
283: $m$). Likewise, we have that $({\bf C})_{i,j} \equiv \langle a_i a_j \rangle$
284: is diagonal in such case. This fact makes the inversion of the covariance
285: matrix in equation (\ref{eq:chisq1}) trivial. Let us now relax the assumption on
286: having $\vs (\vnh)$ defined over the full sphere. In an astrophysical context,
287: if some parts of the sky are lacking data, i.e., if $f_{sky} < 1$, the
288: covariance matrix ${\bf C}$ will no longer be diagonal and for large enough
289: $l_{max}$ it will also be singular. A traditional matrix inversion is likely
290: to either fail or provide inaccurate results. (Note that the accuracy of the
291: inversion can be tested by running Monte Carlo simulations and comparing the
292: dispersion of the recovered $\hat{\alpha}$'s with the actual prediction of
293: equation (\ref{eq:mf1})).
294:
295: In these circumstances, we perform a SVD decomposition of the covariance matrix,
296: \begin{equation}
297: {\bf C} = {\bf R}^t {\bf \Lambda} {\bf R},
298: \label{eq:svd1}
299: \end{equation}
300: where ${\bf \Lambda}$ is a diagonal matrix (containing the eigenvalues of
301: ${\bf C}$) and ${\bf R}$ is a rotation orthogonal matrix (${\bf R}^t{\bf R} =
302: {\bf I}$). Note that since ${\bf C}$ is symmetric and positive definite, the
303: eigenvalues should all be positive\footnote{In practice, we find that for
304: dense and close to singular cases of ${\bf C}$, some eigenvalues (of small
305: absolute value) were negative.}. The SVD decomposition sorts the eigenvectors
306: according to the magnitude of the eigenvalues, so the first eigenvectors are
307: those containing more information about ${\bf C}$, whereas the last ones are
308: the most likely to introduce numerical noise. Note that there is always some
309: numerical error in our estimates of the covariance matrix, since it is
310: computed through a finite number of Monte Carlo realizations (10,000 in this
311: work). Therefore, this decomposition provides a way to rotate the vector $\va$
312: into its principal modes, and permits distinguishing those having most of the
313: information from those being more affected by numerical error (which can be
314: safely {\em projected out} of the analysis). In practice, we neglected all
315: eigenvectors whose eigenvalues were smaller that a given fraction $\epsilon$
316: of the first (largest) eigenvalue. (For most cases, the choice $\epsilon =
317: 10^{-8}$ yielded optimal results). The inversion of ${\bf C}$ after the SVD
318: decomposition becomes straightforward, enabling an easy implementation of the
319: matched filter as given by equation (\ref{eq:mf1}). Note that, unlike in
320: \cite{gs_gorski} or \cite{mortlock}, we are not worried in building a new set
321: of orthonormal functions in the patch of the sky under analysis, nor we
322: attempt to perform component separation (\cite{bouchet}). In all those works,
323: the techniques used were in some way close to ours, but their goals were
324: different.\\
325:
326: As we shall see below, we may be interested in applying the matched filter in
327: {\em different} $l$-bins. One can readily find that, given the
328: outcome of the matched filter in two different $l$-bins
329: %[\lmin^1,\lmax^1]$ and $[\lmin^2,\lmax^2]$,
330: $\hat{\alpha}_p$ and $\hat{\alpha}_q$, their covariance is given by
331: \begin{equation}
332: ({\bf \tilde{C}})_{p,q}\equiv \langle \hat{\alpha}_p \hat{\alpha}_q \rangle -
333: \langle\hat{\alpha}_p\rangle\langle \hat{\alpha}_q \rangle = \frac{\vm_p^t
334: {\bf C}_{pp}^{-1} {\bf C}_{pq} {\bf C}^{-1}_{qq} \vm_q}{(\vm_p^t
335: {\bf C}_{pp}^{-1} \vm_p)(\vm_q^t {\bf C}_{qq} \vm_q)}.
336: \label{eq:covalpha}
337: \end{equation}
338: Here, ${\bf C}_{pp}$ and ${\bf C}_{qq}$ denote the covariance
339: matrices of the noise $\vt$ for $l$-bins $p$ and $q$, respectively, whereas
340: ${\bf C}_{pq}$ is the covariance matrix for the noise in different $l$
341: bins\footnote{Note that, according to our notation, $(\va)_i = a_i$ denotes
342: the $i$-th component of the array $\va$, where $\va_p$ denotes the $p$-th
343: array of some larger group of arrays. Same for matrices: $({\bf C})_{i,j}$
344: denotes the array element in the $i$-th row and $j$-th column, not to be
345: confused with ${\bf C}_{pq}$ which denotes the covariance matrix computed
346: from arrays $\va_p$ and $\va_q$.}: ${\bf C}_{pq} = \langle \vt_p \vt_q
347: \rangle$. For a set of $l$-bins we shall obtain a vector of measured
348: $\hat{\alpha}_p$-s, whose combined $\chi^2$ will be given by
349: \begin{equation}
350: \chi^2 \left[{\hat{\alpha}}\right] = \sum_{p,q} \hat{\alpha}_p ({\bf \tilde{C}^{-1}})_{p,q} \hat{\alpha}_q
351: \label{eq:chisq2}
352: \end{equation}
353: An overall detection level for a given set of $l$-bins and $\hat{\alpha}$-s
354: will be provided by this $\chi^2$ statistic. Another statistic providing the level of detection is
355: the variance weighted average for the $\hat{\alpha}_p$'s, here defined as $\hat{\beta}$:
356: \begin{equation}
357: \hat{\beta} \equiv \frac{ \sum_p {\left( \hat{\alpha}_p / \hat{\sigma}_{\alpha_p}^2\right)} } {\sum_p{1/\hat{\sigma}_{\alpha_p}^2}}.
358: \label{eq:owa}
359: \end{equation}
360: We will show below that, in ISW studies, the distribution of the
361: $\hat{\alpha}_p$-s will be very close to Gaussian, and therefore Gaussian will
362: also be the distribution of $\hat{\beta}$. (Note that the matched filter method, as
363: defined from a minimization of the statistic given in equation
364: (\ref{eq:chisq1}), is only optimal if the noise is actually Gaussian
365: distributed). The diagonal elements of the matrix ${\bf \tilde{C}}$ can be
366: computed via equation (\ref{eq:covalpha}) or via numerical simulations: the
367: agreement is very good (down to a few percent, compatible to the number of
368: realizations). However, this agreement is not satisfactory for the non-diagonal
369: elements when working under aggressive masks: in this case (and also
370: when computing the dispersion of the statistic $\hat{\beta}$) we shall use the
371: results obtained from 10,000 Monte Carlo simulations. This assures a fair
372: estimation of the correlation between different $\alpha_p$ estimates and hence
373: an accurate estimation of the overall $\chi^2$ statistic.
374:
375: \begin{figure*}
376: \centering
377: \plotancho{./fig_masks_light.eps}
378: \caption[fig:masks]{Three different masks used in this paper:
379: {\it Left}: Mask corresponding to the product of the SDSS-DR4 mask times the Kp0 mask used in \cite{hinshaw}
380: {\it Middle}: Mask assigned to the future PAU-BAO galaxy survey times the Kp0 mask.
381: {\it Right}: Mask of the NVSS times the Kp0 mask. }
382: \label{fig:masks}
383: \end{figure*}
384:
385:
386: \section{Comparison to the Angular Cross Power Spectrum}
387: \label{sec:comp}
388:
389: In this Section we shall compare the matched filter (as defined above) to the
390: Angular Cross Power Spectrum (hereafter ACPS) method. This comparison will be
391: made within the model motivated in the previous Section,
392: \begin{equation}
393: \vs = \vt + \alpha\vm,
394: \label{eq:model}
395: \end{equation}
396: and will be restricted to the large angular scales. This choice is motivated
397: by the study of the Integrated Sachs-Wolfe (ISW) effect that follows in
398: subsequent Sections of the paper, and that is typically restricted to $l<50$.
399: When studying small scales, one has to be careful with the SVD decomposition,
400: which might fail for too large matrices. The matched filter in real space is,
401: in most of those occasions, more adequate.
402: \begin{table*}
403: \begin{centering}
404: %\scriptsize
405: %
406: \begin{tabular}{||c||c|c|c|c|c|c||}
407: \hline
408: \hline
409: & \multicolumn{2}{|c|}{$l_{max}=15$} & \multicolumn{4}{|c|}{$l_{max}=50$}\\
410: \hline\hline
411: & MF & ACPS & \multicolumn{2}{|c|}{MF} & \multicolumn{2}{|c|}{ACPS}\\
412: \hline
413: & \multicolumn{2}{|c|}{$\langle \hat{\alpha}\rangle / \sigma_{\hat{\alpha}}$} & $\langle \chi^2_N\rangle $ & $\langle \hat{\beta} \rangle / \sigma_{\hat{\beta}}$ & $\langle \chi^2_N\rangle $ & $\langle \hat{\beta} \rangle / \sigma_{\hat{\beta}}$ \\
414: \hline
415: All sky & $25.70$&$16.50$ &2,312 & 192&2,316&192\\
416: \hline
417: SDSS DR4 & $34.68$ & $6.44$ & 271 & 63& 201& 55\\
418: \hline
419: \hline
420: \end{tabular}
421: %\normalsize
422:
423: \medskip
424: \caption[tab:tab1]{Under two different masks (SDSS-DR4 mask and all sky), we
425: compare the performance of the matched filter (MF) and the ACPS when trying to
426: estimate the amplitude $\alpha$ from a data set $\vs = \vt + \alpha\vm$ given
427: $\vs$, $\vm$ and the power spectrum of the Gaussian noise $\vt$, which is
428: taken from a CMB power spectrum. In case {\it (i)} we consider a single
429: $l$-bin containing all multipoles in $l\in[2,15]$, whereas in case {\it (ii)}
430: we consider 16 $l$-bins: $l\in
431: [2,3],[4,5],[6,8],[9,14],[15,25],[26,28],[29,31],[32,34],[35,37],[38,40],[41,43],[44,45],[46,47],[48,49]$
432: and $[50,51]$. We are quoting the results for the $\hat{\alpha}$ statistic in
433: case {\it (i)}, and for the $\chi^2$ and $\hat{\beta}$ statistics in case
434: {\it (ii)}. Note that, in this case, the $\chi^2$ statistic has been
435: normalized by the number of degrees of freedom, i.e., the number of $l$-bins.
436: }
437: \label{tab:tab1}
438: \end{centering}
439: \end{table*}
440:
441: The Angular Cross Power Spectrum (hereafter ACPS) can be viewed as a
442: simplification of the matched filter presented here, where the covariance
443: matrix is approximated by a diagonal matrix with identical non zero
444: elements. Following the notation of Section \ref{sec:method}, the estimate of
445: $\alpha$ provided by this method is given by
446: \begin{equation}
447: \hat{\alpha}_{ACPS} = \frac{\sum_{l,m} {m_{l,m} (s_{l,m})^*}}{\sum_{l,m}{|m_{l,m}|^2}
448: },
449: \label{eq:acps1}
450: \end{equation}
451: where $s_{l,m}$ and $m_{l,m}$ are the Fourier multipoles of the signals $\vs$
452: and $\vm$, respectively. Note that we shall refer to these signals in Fourier
453: space, and thus the vectors $\vs$ and $\vm$ will contain the components of the
454: $s_{l,m}$ and $m_{l,m}$ multipoles as explained in Section \ref{sec:cov}.
455:
456: In order to compare this method to the matched filter, we have to define $\vt$
457: and $\vm$ and build $\vs$ according to equation (\ref{eq:model}).
458: Throughout this paper, we shall not use real data but only Gaussian realizations generated from
459: a given cosmological model.
460: For $\vt$,
461: we choose CMB realizations for which the ISW contribution has been
462: subtracted. I.e., we simulate the Fourier multipoles $t_{l,m}$-s from an
463: angular power spectrum computed using a modified version of the CMBFAST code
464: (\cite{cmbfast}) with a cosmological parameter set equal to that given in
465: \cite{spergel06}: $\Omega_c=0.1994$, $\Omega_{\Lambda}=0.759$,
466: $\Omega_b=0.0416$, $n_s=0.958$, $\sigma_8=0.75$ and $\tau= 0.089$. This will be
467: the reference cosmological model hereafter. The
468: template $\vm$ is a Gaussian realization of a projection of the linear density
469: field as computed from the matter power spectrum obtained with the same
470: cosmological parameters. This density field is placed within a shell centered
471: at $z=0.8$ with a total width of $\Delta z \sim 0.8$, i.e., the redshift range
472: where ISW contribution is maximal (this will be addressed in detail in Section
473: (\ref{sec:isw})). To each realization of $\vt$ we added the component
474: $\alpha\vm$ (for a given choice of $\alpha$, $\alpha=10^{-3}$), and applied
475: the two methods. All maps were convolved with a Gaussian PSF of
476: FWHM$=2{\degr}$. Let us remark that by this exercise we do not attempt to
477: simulate ISW observations, but simply test the two methods in the context of
478: equation (\ref{eq:model}).
479:
480: In this comparison, we applied both the matched filter and the ACPS under two
481: different masks: the first one assumes full sky coverage ($f_{sky}=1$),
482: whereas the second one adopts the mask provided by the fourth data release of
483: Sloan Digital Sky Survey (SDSS-DR4, \cite{dr4}) combined with the Kp0 mask
484: used in WMAP data analyses, (\cite{hinshaw}; see left panel of
485: Fig. (\ref{fig:masks})). In Table (\ref{tab:tab1}) we display the results
486: after applying both methods to an ensemble of 10,000 simulations. We consider
487: two scenarios: {\it (i)} a unique $l$-bin limited to $\lmax=15$, $l\in
488: [2,15]$, and {\it (ii)} a set of 16 $l$-bins, ranging from $\lmin = 2$ to
489: $\lmax = 50$. We always find that the estimates of $\tilde{\alpha}$ are
490: unbiased for both methods\footnote{A mask in real space involves a convolution of
491: different $\alpha$ values in Fourier space, which may generate a bias if $\alpha$ is not
492: constant versus $l$.The case considered in this Section observes a constant $\alpha$ for every multipole, but in subsequent sections this will not be the case and a bias will appear.},
493: and that the matched filter provides an estimate of
494: the dispersion of $\tilde{\alpha}$ (see equation (\ref{eq:mf1})) that actually
495: agrees with the value recovered from the Monte Carlo simulations. In case {\it
496: (i)} we obtain that the matched filter works better than the ACPS under the
497: two masks considered. Note that the noise signal in these analyses corresponds
498: to the CMB (after having the ISW component subtracted), and that therefore
499: the noise power spectrum scales as $\langle | t_{l,m} |^2 \rangle\propto
500: l^{-2}$. On the other hand, the angular power spectrum of our density
501: template scales roughly as $ \langle | m_{l,m} |^2 \rangle\propto const $ at
502: $l<50$. This means that for the low-$l$ range considered in {\it (i)}, the
503: matched filter is going to weight more the high-$l$ end multipoles: this
504: effect makes this method superior to the ACPS (which weights all multipoles
505: equally) in the all sky case. For the SDSS-DR4 mask, the matched filter also
506: accounts for the coupling among mutipoles, and this enlarges the difference
507: between the two methods. One obvious question that arises is: how can the
508: matched filter perform better under the SDSS-DR4 mask than in the all sky
509: case? The low $l$ modes are {\em degenerate} under the SDSS-DR4 mask, that is,
510: they are not orthonormal as for $f_{sky}=1$ and are decomposed onto other
511: modes corresponding to smaller angular scales. I.e., the noisiest (low $l$)
512: modes under the full sky mask are not {\em eigenmodes} anymore and they are
513: partially dropped from the analysis (the matched filter method handles 176
514: different modes under the SDSS-DR4 mask, as opposed to 252 modes if
515: $f_{sky}=1$). This means that under the SDSS-DR4 mask we have a different
516: statistic (since it handles a different number of degrees of freedom) that is
517: more concentrated in angular scales where the noise amplitude is
518: smaller. This provides this new statistic a better S/N ratio.
519:
520: In case {\it (ii)} and $f_{sky}=1$, it turns out that, given the scaling of
521: the power spectra of $\vm$ and $\vt$, most of the S/N ratio is in the few
522: higher $l$ bins, centered at multipoles $l = 45, 46, ..., 50$. Indeed, these
523: last multipoles are dominating the sums in equation (\ref{eq:acps1}). For these
524: few high-$l$ bins carrying most of the information, the change of the noise
525: properties is very small, and therefore each of these bins is roughly
526: equivalent to the rest. The weighting applied by the matched filter introduces
527: very little difference, and both methods perform similarly, yielding almost
528: identical values of $\chi^2$. But again, under the SDSS-DR4 mask the coupling
529: among multipoles is observed by the matched filter, and this
530: introduces a difference between the two methods. However, it is clear that the
531: matched filter proves comparatively better when when a single $l$-bin is considered,
532: (case {\it (i)}).
533: %
534:
535: \section{Application to ISW Studies}
536:
537:
538: The Integrated Sachs Wolfe (ISW) effect arises as a consequence of a late time
539: variation of the gravitational potentials in the large scales. If there is a
540: net change in the depth of the potential wells while they are being crossed by
541: CMB photons, then this radiation field will experience a gravitational
542: red/blueshift. \cite{critturok96} pointed out that gravitational
543: potentials should be traced by the Large Scale Structure (LSS), and proposed
544: the cross-correlation of CMB maps to LSS surveys to unveil this
545: signal. However, in most plausible models the time variation of the potentials
546: occurs at late times (or low redshifts, $z< 2$), and the angle subtended by
547: the linear scales for which the ISW effect is important is rather large
548: ($\theta > 3-5\degr$). This means that there will be room for relatively {\em
549: few} independent ISW spots on the sky, i.e., this effect will be considerably
550: limited by Cosmic Variance. Further, it is in this large angular range where
551: the Galactic emission is more important, and errors in its subtraction might
552: be more relevant. For this reason, it becomes necessary the use of masks that
553: project out regions where this galactic contamination is large and cannot be
554: removed accurately. Furthermore, when doing a cross-correlation analysis
555: between CMB maps and LSS maps, the latter may not likely cover the whole sky,
556: but also be restricted to a given limited region. In this context, it becomes
557: particularly important the implementation of as sensitive as possible
558: cross-correlation tools that are able to handle optimally the limitations
559: imposed by the sky masks.
560:
561: \subsection{The S/N ratio in ISW Studies}
562: \label{sec:isw}
563: \begin{figure}
564: \centering
565: %\plotone{./fig_cls.eps}
566: \includegraphics[width=6.cm,height=11.cm]{./fig_cls.eps}
567: \caption[fig:cls]{ {\it (a)} The dot-dashed line displays the cross power
568: spectrum of the projected density field through a shell centered at $z=0.8$
569: (dashed line) times the ISW component (thin solid line). The thick solid line
570: displays the total CMB angular power spectrum. {\it (b)} Normalized signal to
571: noise ratio (as given by equation (\ref{eq:s_l})) versus multipole $l$: the
572: solid line corresponds to a density probe in a shell centered at $z=
573: 0.8$, whereas for the dotted and dashed lines the central redshifts are 0.4
574: and 1.3, respectively. In both panels, the width of each density shell is
575: equal to 20\% of the comoving distance to the central shell redshift.}
576: \label{fig:cls}
577: \end{figure}
578:
579:
580: In this Subsection we briefly describe the ISW -- LSS cross-correlation in the
581: frame of the WMAP3 cosmogony. Unlike \cite{marian}, we refrain from
582: addressing the dependence of this correlation under different Dark Energy
583: models. We concentrate on a single LSS survey, and search for its optimal
584: redshift in terms of ISW detection. We study the amount of S/N that arises in
585: those cases, and the angular scales where it is generated. This sets the
586: scenario for our cross-correlation method comparison.
587:
588: The ISW -- LSS cross correlation arises from the fact that both LSS probes and
589: gravitational potentials are tracing the underlying matter density field.
590: The expression for the temperature anisotropies introduced by a
591: gravitational blue/redshift reads (\cite{sw67,enrique})
592: \begin{equation}
593: \left( \frac{\delta T}{T_0} \right)_{ISW} (\vnh ) = \frac{-2}{c^3} \int dt\;
594: \dot{\phi} (\vnh, t).
595: \label{eq:isw_def}
596: \end{equation}
597: The symbol $\dot{\phi}$ denotes the time derivative of the gravitational
598: potential $\phi$. In linear theory, this expression can be rewritten as
599: (e.g., \cite{cooray02})
600: \[
601: a_{l,m}^{ISW} = (-i)^l (4\pi) \int \frac{d\vk}{(2\pi)^3}\;
602: Y_{l,m}^{\star}(\hat{\vk}) \;\times
603: \]
604: \begin{equation}
605: \phantom{xxxxxxxx} \int dr\; j_l (kr) \frac{-3\Omega_mH_0^2}{k^2} \;\frac{d(D/a)}{dr} \delta_{\vk}.
606: \label{eq:alm_isw0}
607: \end{equation}
608: The multipole coefficients $a_{l,m}^{ISW}$ are related to the ISW temperature
609: anisotropies by equation (\ref{eq:sp1}). In this equation, $r$ denotes
610: comoving distance, $k$ comoving wavevector, $j_l(x)$ denotes the spherical
611: Bessel function of order $l$, $H_0$ is the Hubble parameter, $a(r)$ is the
612: scale factor and $D(r)$ is the standard linear growth factor. The 3D Fourier
613: mode of the density contrast is denoted by $\delta_{\vk}$. Note that for an
614: Einstein-de Sitter Universe ($D=a$) the whole integral vanish. In a similar
615: way, the multipole coefficients for the angular number density of a matter
616: density probe (which will be taken to be galaxies in what follows) read
617: \[
618: a_{l,m}^{g} = (-i)^l (4\pi) \int \frac{d\vk}{(2\pi)^3}\;
619: Y_{l,m}^{\star}(\hat{\vk}) \;\times
620: \]
621: \begin{equation}
622: \phantom{xxxxxxxxxxxxx} \int dr\; j_l (kr)\; r^2\; n_g(r)b(r,k)\;D(r) \;
623: \delta_{\vk},
624: \label{eq:alm_rho}
625: \end{equation}
626: with $n_g(r)$ the average number density of galaxies. The bias function
627: $b(r,k)$ accounts for usual probes of the LSS actually being biased tracers of
628: the underlying mass distribution ($b > 1$). This expression neglects the
629: presence of shot (Poisson) noise in the galaxy number. Note that the
630: coordinate $r$ here is being taking as a look-back time coordinate, equivalent
631: to conformal time or redshift ($z$). For the sake of clarity, in what follows
632: we shall use $z$ as look-back time coordinate. We have that a given galaxy
633: survey will probe the redshift range given by the product $\Pi(z)\equiv
634: b\;N_g$, with $N_g(z)\equiv r^2(z)\;n_g(z)D(z)$. Note that, at this stage, we
635: are ignoring the $k$ dependence of the bias function $b$. For simplicity, we
636: shall rewrite equations (\ref{eq:alm_isw0},\ref{eq:alm_rho}) as
637: \begin{equation}
638: a_{l,m}^{ISW,g} = (-i)^l (4\pi) \int \frac{d\vk}{(2\pi)^3}\;
639: Y_{l,m}^{\star}(\hat{\vk}) \;\times \Delta_l^{ISW,g}(k,z),
640: \label{eq:redef}
641: \end{equation}
642: with the $\Delta_l^{ISW,g}(k,z)$ being referred to as transfer functions
643: of the ISW and the galaxy fields, respectively. In
644: real space, the cross correlation function between ISW temperature
645: anisotropies and LSS probes reads
646: \begin{equation}
647: C_{ISW\otimes g}(\theta ) = \sum_l \frac{2l+1}{4\pi} C_l^{ISW\otimes g} P_l(\cos \theta ),
648: \label{eq:ccf}
649: \end{equation}
650: with $C_l^{ISW\otimes g}$ the cross power spectrum,
651: \begin{equation}
652: C_l^{ISW\otimes g} = \left( \frac{2}{\pi}\right)\int \;k^2dk\; \Delta_l^{ISW}
653: \Delta_l^{g}\; P_m(k),
654: \label{eq:cps_cl}
655: \end{equation}
656: and $P_m(k)$ is the linear matter power spectrum.
657: The symbol "$\otimes$" denotes cross correlation. The theory makes actual
658: predictions on this cross-power spectrum, and its covariance matrix is
659: diagonal if $f_{sky}=1$, and in general is simpler than that of the
660: correlation function, (see the detailed analysis of \cite{cabre}). As an
661: example, we show a cross power spectrum (dot-dashed line) in the top panel
662: of Fig. (\ref{fig:cls}). The total CMB contribution (for the chosen
663: $\Lambda$CDM model) is given by the thick solid line, and its ISW component is
664: displayed by the thin solid line. The angular power spectrum corresponding to
665: the projection of the galaxy field whose window function $\Pi(z)$ is centered
666: at $z= 0.8$ is displayed by the dashed line. The total width in redshift space
667: is roughly $\Delta z \approx 0.80$, so it is a thick shell. The cross power
668: spectrum peaks at scales at around $l\sim 30-50$, but its amplitude at $l\sim
669: 200$ is roughly equal to that at $l\sim 2$. This might suggest that the there
670: is so much cross-correlation signal at large ($l<10$) as in small ($l>100$)
671: scales, but indeed most (90\%) of the signal comes from the large scales
672: ($l<40$) if $f_{sky}=1$, as we will show below. Note that the amplitude of the
673: density power spectrum (and hence the cross power spectrum) are taken
674: arbitrary. The signal-to-noise (S/N) ratio for the measurement of a given
675: multipole $l$ of the cross power spectrum can be computed once one takes into
676: account that the $a_{l,0}$ multipoles are real defined, and that the real and
677: complex components of the $a_{l,m}$ ($m > 0$) multipoles, besides being
678: equivalent and independent, must satisfy the constraint for the total
679: amplitude $\langle | a_{l,m} |^2 \rangle = C_l$. We obtain
680: \begin{equation}
681: \left( \frac{S}{N}\right)^2_l = \frac{f_{sky}\left(C_l^{ISW\otimes g}\right)^2(l+1)^2}{\left[C_l^{CMB}C_l^{g} + (C_l^{ISW\otimes g})^2\right](l/2+1)}.
682: \label{eq:s2nl1}
683: \end{equation}
684: In this equation, $C_l^{CMB}$ is the CMB angular power spectrum and $C_l^g$ is the LSS probe auto power spectrum.
685: The quantity
686: \begin{equation}
687: s_l \equiv \sqrt{ \frac{\sum_{l'=2}^{l'=l} {(S/N)_{l'}^2}}{\sum_{l'=2}^{l'=l_{max}}{(S/N)_{l'}^2}}}
688: \label{eq:s_l}
689: \end{equation}
690: is displayed versus $l$ in the bottom panel of Fig. (\ref{fig:cls}). I.e.,
691: this figure shows the ratio of the signal to noise ratio contained below some
692: given $l$. The thick solid line corresponds to the galaxy survey centered at
693: $z= 0.8$ mentioned above, whereas the dotted line for a galaxy survey with a
694: window function centered at $z= 0.4$. The dashed line corresponds to a case
695: where the galaxy survey is probed at $z= 1.3$. In all cases we are taking a
696: shell width equal to 20\% of the comoving distance to the peak of the window
697: function $\Pi$, and we are assuming that $f_{sky}=1$. We see that regardless
698: where $\Pi (z)$ peaks, {\em practically half of the total signal is contained
699: at multipoles $l<10$, whereas its 90\% fraction is typically contained at
700: $l<40-50$}. (Had we considered thinner shells -width equal to 2\% of the
701: comoving distance-, then all those shells below $z=0.8$ would have still
702: shown a pattern very close to that given by the solid line). This suggests
703: that by dropping all multipoles above $l=50$ (or by neglecting scales smaller
704: than $\theta \approx$ 3\degr -- 4\degr) one should recover practically the
705: same ISW detection significance. This sets a useful consistency check, given
706: the number of other physical phenomena (Rees-Sciama effect (\cite{rs}),
707: kinetic Sunyaev-Zel'dovich effect (\cite{ksz}), intrinsic source emission,
708: etc) that arise at smaller angular scales and that, a priori, correlate with
709: the spatial position of LSS probes.
710:
711: \begin{figure}
712: \centering
713: %\plotone{./s2n_vs_z.eps}
714: \includegraphics[width=6.cm,height=6.cm]{./s2n_vs_z.eps}
715: \caption[fig:s2nvsz]{Total signal to noise ratio of the cross power spectrum detection as given by the numerator of equation (\ref{eq:s_l}) with respect to the central redshift of the density probe shell. The thick solid line displays the case when the shell containing the density probes has a width equal to 20\% of the comoving distance to the shell. For the thin solid line, this width is only 2\%.}
716: \label{fig:s2nvsz}
717: \end{figure}
718:
719: In Fig. (\ref{fig:s2nvsz}) we show how the total signal to noise ratio
720: depends on the central redshift for the galaxy survey window function
721: $\Pi(z)$. For the thick solid line, the width is taken to be roughly 20\% of
722: the comoving distance to the central redshift. In redshift space, it implies
723: a width of $\Delta z \approx 0.7$ for central redshift $z=0.3$, $\Delta z
724: \approx 0.84$ for central redshift $z=0.8$, and $\Delta z \approx 0.99 $ for
725: central redshift $z=1.3$. The thin solid line observes a width of only 2\%
726: the comoving distance to the central redshift, and this translates into
727: $\Delta z \approx 0.07, \;0.08$ and $0.1$ for central redshifts $z=0.3,\;
728: 0.8$ and $1.3$ respectively. We should obtain the larger detection levels for
729: thick shells (large $\Delta z$'s) and $z= 0.8$, and this motivates our choice
730: of a thick survey centered at this redshift\footnote{It has been noted
731: elsewhere (e.g. \cite{afshordi}) that by combining different LSS surveys at
732: different redshifts one can obtain larger S/N ratios. We shall avoid that
733: discussion here and focus our method comparison on one single survey.}. Note
734: that our assumption that the bias is independent of scale might not be
735: accurate, but it has less impact in the large scales (low $l$s) where most of
736: the effect is coming from. We do not expect significant changes after
737: introducing a scale dependent bias in our galaxy survey description, although
738: we shall address this issue in detail when applying our method to real CMB
739: and LSS data. Note that, {\it a priori}, this method can be applied the same on
740: multiple redshift shells, and is affected in exactly the same way than the ACPS by
741: realistic aspects such as the redshift or scale dependence of the bias, survey incompleteness, etc.
742:
743: We next compare the performance of the matched filter to that of the ACPS. We
744: use one Gaussian realization of our chosen density shell, and compute a single Gaussian
745: realization of an ISW map compatible to it. If $a_{l,m}^{g}$ are the Fourier multipoles of our
746: density 2D template, then they can be related to those of a {\it compatible} ISW map via (e.g., \cite{cabre})
747: \begin{equation}
748: a_{l,m}^{ISW} = \alpha_l a_{l,m}^{g} + \beta_{l,m}= \frac{C_l^{ISW\otimes g}}{C_l^{g}}\; a_{l,m}^{g} + \beta_{l,m}.
749: \label{eq:alm_isw}
750: \end{equation}
751: The Gaussian signal $\beta_{l,m}$ is the part of the ISW component that is uncorrelated
752: to the LSS 2D template, verifying $\langle | \beta_{l,m}|^2 \rangle =
753: C_l^{ISW} - (C_l^{ISW\otimes g})^2/C_l^g$, where $C_l^{g}$ denotes the angular power
754: spectrum of the LSS probe. Note that this is a correct way to express the ISW field in terms of the galaxy density field as long as both fields are Gaussian and their are completely determined by the first and second order momenta. The ratio $C_l^{ISW\otimes g}/C_l^{g}$ is explicitly
755: identified with $\alpha_l$, which is precisely the output of our matched
756: filter technique. According to the theory, $\alpha_l$ shows a strong
757: dependence on $l$, and therefore our method must be applied in separate
758: $l$-bins. Gaussian simulations of the CMB were built from the addition of our
759: {\em fixed} ISW template plus realizations of a CMB angular power spectrum for
760: which the ISW component had been subtracted, just as for the $\vt$ component
761: in Section (\ref{sec:comp}). The realizations from the modified CMB angular
762: power spectrum were computed upto a maximum multipole $l=160$, and convolved
763: with a Gaussian beam of 2\degr of FWHM. The fixed ISW and the LSS maps were
764: also convolved with the same PSF, and all maps were produced under the HEALPix\footnote{{\tt http://www.healpix.jpl.nasa.gov}} (\cite{healpix}) resolution
765: parameter $N_{side}=64$.
766:
767:
768:
769: \subsection{Performance under Different Masks}
770:
771: \begin{table*}
772: \begin{centering}
773: \begin{tabular}{||c|c|c|c|c|c|c|c|c|c|c|c|c||}
774: \hline
775: \hline
776: & & \multicolumn{2}{|c|}{$l_{max}=5$} & \multicolumn{2}{|c|}{$l_{max}=14$} & \multicolumn{2}{|c|}{$l_{max}=31$} & \multicolumn{2}{|c|}{$l_{max}=40$} & \multicolumn{2}{|c|}{$l_{max}
777: =51$ } \\
778: \hline
779: & & MF & ACPS & MF & ACPS & MF & ACPS & MF & ACPS & MF & ACPS \\
780: \hline
781: \multirow{2}{*}{SDSS-DR4} &$\langle \chi^2_N \rangle$ & 1.79 & 1.31 & 2.04 & 1.56 & 1.61 & 1.48 & 1.46 & 1.40 & 1.33 & 1.30\\ & $\langle \hat{\beta}\rangle / \sigma_{\hat{\beta}}$ &1.24 & 0.65 & 1.87 & 1.35 & 1.74 & 1.70 & 1.57 & 1.61 & 1.46 & 1.22\\
782: \hline
783: \multirow{2}{*}{PAU-BAO} &$\langle \chi^2_N \rangle$ & 2.93 & 2.02 & 2.94 & 2.45 & 2.12 & 2.29 & 1.87 & 2.02 & 1.61 & 1.74\\ & $\langle \hat{\beta}\rangle / \sigma_{\hat{\beta}}$ &1.69 & 1.28 & 2.48 & 2.22 & 1.93 & 2.33 & 1.90 & 2.33 & 1.69 & 1.98\\
784: \hline
785: \multirow{2}{*}{NVSS} &$\langle \chi^2_N \rangle$ & 6.27 & 7.03 & 5.30 & 6.04 & 4.15 & 4.71 & 3.47 & 3.91 & 2.75 & 3.06\\ & $\langle \hat{\beta}\rangle / \sigma_{\hat{\beta}}$ &3.24 & 3.37 & 3.51 & 3.70 & 3.44 & 3.62 & 3.43 & 3.56 & 3.19 & 3.19\\
786: \hline
787: \hline
788: \end{tabular}
789: \medskip
790: \caption[tab:tab2]{Comparison of the matched filter (MF) to the
791: ACPS in the context of ISW studies. We quantify the sensitivity of each method
792: by two statistics: $\chi^2$ (which has been normalized by the number of
793: degrees of freedom) and $\hat{\beta}$, for different choices of $l_{max}$. In
794: total, we considered 21 $l$-bins: $l\in$
795: [2,3],[4,5],[6,8],[9,14],[15,25],[26,28],[29,31],[32,34],[35,37],[38,40],[41,43],[44,45],[46,47],[48,49],[50,51],[52,53],[54,55],[56,57],[58,59] and [60,61]. }
796: \label{tab:tab2}
797: \end{centering}
798: \end{table*}
799:
800:
801: \begin{figure*}
802: \centering
803: \plotancho{./fig_alphas_log_lin.eps}
804: \caption[fig:alphas]{Recovered values of $\alpha_l = C_l^{ISW\otimes g}/C_l^{g}$
805: with the matched filter (filled circles) and the ACPS (filled triangles) under
806: the three masks considered: {\it (a)} SDSS-DR4 $\times$ Kp0, {\it (b)} PAU-BAO
807: $\times$ Kp0 and {\it(c)} NVSS $\times$ Kp0.
808: %Our choice for the density and
809: %ISW templates is such that full sky analyses should yield $\alpha_l$ estimates {\em exactly} on the solid lines. Note that mask-induced aliasing introduces a low bias in $\alpha_l$ estimates at low $l$-s under the most
810: %aggressive masks.
811: }
812: \label{fig:alphas}
813: \end{figure*}
814:
815: \begin{figure*}
816: \centering
817: \plotancho{./plot_s2n_2.eps}
818: \caption[fig:s2n]{Recovered signal to noise ratio for each multipole bin with the matched filter (filled circles) and the ACPS (filled triangles), under the three masks considered in Fig. (\ref{fig:alphas}).
819: }
820: \label{fig:s2n}
821: \end{figure*}
822:
823: The simulated maps were cross-correlated to the projected density map
824: (hereafter denoted as $\vm$) with both the ACPS and the matched filter
825: methods, according to the multipole decomposition given in Section \ref{sec:cov}.
826: Their performance was compared under three different masks shown in
827: Fig. (\ref{fig:masks}): the left panel shows the mask corresponding to the sky
828: coverage of the fourth data release of the Sloan Digital Sky Survey (SDSS-DR4,
829: \cite{dr4}). This mask was multiplied by the Kp0 mask used in the analysis of
830: WMAP data (\cite{hinshaw}), and therefore the combined mask observes a bit
831: less than 10\% of the total sky. In the middle panel we consider the fraction
832: of the sky covered by the upcoming survey PAU-BAO \footnote{{\tt
833: http://www.ice.csic.es/research/PAU/PAU-welcome.html}}. This survey is planned
834: to cover $\sim$ 10,000 square degrees of the celestial northern hemisphere,
835: and in this work we have assumed that it is limited to the region
836: $b>20\degr$ {\it outside} the Kp0 mask, in such a way that $f_{sky} \simeq 0.26$.
837: Finally, the right hand
838: side panel displays the product of the Kp0 mask with the mask corresponding to
839: the NVSS survey (\cite{nvss}). In this case, $f_{sky}\simeq 0.65$.
840:
841: A total of 10,000 simulations were run in order to perform the method
842: comparison, and results are given in Table (\ref{tab:tab2}). The sensitivity
843: of both methods is measured by the $\chi^2$ and the $\hat{\beta}$ statistics
844: for different choices of the maximum multipole $l_{max}$ considered in the
845: analyses. We see that, unlike in the previous Section (where the matched
846: filter was in general significantly more sensitive under the data model $\vs =
847: \vt + \alpha\vm$), in this ISW context both filters perform very
848: similarly. There seems to be however a slighter better sensitivity of the
849: matched filter under the most aggressive masks, but the difference is small
850: (at least in terms of the output of the $\chi^2$ and $\hat{\beta}$
851: statistics). We remark that the ACPS is the Legendre transform of the angular
852: cross correlation function, and that both methods should, a priori, show
853: similar sensitivities. Of the two statistics considered in Table
854: (\ref{tab:tab2}), $\hat{\beta}$ provides the largest significance of the ISW
855: detection (its distribution is very close to Gaussian, and the number of
856: sigmas yield smaller {\it chance probabilities}), but is remarkable that for
857: both of them this significance finds a maximum at $l\sim5-40$, and then
858: starts dropping again. {\em According to this result, by merely observing
859: multipoles below $l=40$ we should recover practically all the ISW-LSS cross
860: correlation significance}. This points in the direction of our $s_l$
861: estimates in Subsection (\ref{sec:isw}), which predicted most of the S/N ratio
862: to be confined at low $l$-s, and should prevent confusion with other secondary
863: effects (giving rise to correlations of other nature such as Rees-Sciama
864: effect, kinetic Sunyaev-Zel'dovich effect, etc).
865:
866: One advantage of the correlation methods implemented here is that they conduct
867: analyses in {\em separate} ranges of angular scales, enabling a direct
868: comparison with theoretical predictions. This is explicitly shown in Figs.
869: (\ref{fig:alphas}) and (\ref{fig:s2n}). In all panels of Fig.
870: (\ref{fig:alphas}), solid lines display the theoretical prediction for the
871: correlation coefficient $\alpha_l = C_l^{ISW\otimes g}/C_l^{g}$ versus multipole
872: $l$. Our choice for the density and
873: ISW templates is such that full sky analyses should yield $\alpha_l$ estimates {\em exactly} on the solid lines. Filled circles and triangles display average estimates of $\alpha_l$ for
874: the matched filter and the ACPS methods, respectively. Error bars denote the
875: rms scatter for each of them. Let us first note that, according to the error
876: bars displayed in Fig. (\ref{fig:alphas}) , most of the information seems
877: again to be restricted to the large angular scales ($l<30-40$), as it has been
878: quoted above.
879:
880: Let us now address the issue of the impact of the mask on the methods' output.
881: A clear low bias can be seen in the estimates of $\alpha_l$
882: for low $l$-s, specially under the SDSS-DR4 and PAU-BAO masks. This is a
883: direct effect of the mask: a multiplication of the full sky map by the actual
884: mask in real space translates into a convolution in Fourier space, which
885: involves a wider range of multipoles the smaller the mask is. Therefore, the
886: $\alpha_l$ estimates for the SDSS-DR4 mask will be the result of an {\it
887: average or smoothing} of the $\alpha_l$ values in a wide space of $l$
888: multipoles. Since at low $l$-s, the values of $\alpha_l$ are falling steeply,
889: the convolution will provide a value that is smaller than the actual
890: theoretical value, as it is displayed in the left panel of
891: Fig. (\ref{fig:alphas}).
892:
893: At the same time, the mask introduces another bias in the large $l$ range, for
894: which the actual recovered values of $\alpha_l$ are above the theoretical
895: expectation. This is showing the fact that the aliasing introduced by the mask
896: is shifting some large scale (low $l$) power into the small (large $l$)
897: angular range. We attempted to quantify this aliasing by performing the
898: following exercise: we generated one ISW map by using multipoles restricted to
899: the range $l\in [2, 10]$. We multiplied this map by the SDSS-DR4 mask used in
900: this work, and computed the power spectrum of the resulting map. We measured
901: the aliased variance contained above some $l_{min}>10$ by using $\sigma^2
902: [l_{min}] = \sum_{l_{min}}(2l+1)/(4\pi) C_l$. We found that about 30\% of the
903: total rms\footnote{The total rms was computed by taking $l_{min}=2$, and
904: was less than 10\% off the estimate obtained from the map in real space. } was
905: contained above $l_{min}=30$. I.e., the ISW power aliasing from large to small
906: scales is indeed significant. Let us remark as well that this effect is more present for the
907: matched filter $\alpha_l$ estimates, as we shall discuss next.
908:
909: A direct visual comparison of the two methods can be found in Fig.
910: (\ref{fig:s2n}), where the S/N ratio for each multipole bin is shown. The
911: matched filter (solid circles) performs more accurately than the ACPS method
912: (filled triangles), specially under the SDSS-DR4 and PAU-BAO masks, although
913: this has a limited impact on the final detection significance quoted by the $\chi^2$ and $\hat{\beta}$
914: statistics (see next Section).
915:
916: \section{Discussion and Conclusions}
917:
918: In order to assess the sensitivity of the two methods to the ISW, we have
919: defined two different statistics: the $\chi^2$ statistic uses a quadratic
920: combination of the method's output (in a similar way as in \cite{tegmark}),
921: and the $\hat{\beta}$ statistic, which instead is linear in the $\hat{\alpha_l}$'s and
922: for our purposes is Gaussian distributed. Both statistics pick up the
923: information of the cross-correlation in different ways. The $\chi^2$ statistic
924: is more sensitive to the presence of ISW at the very large angular scales
925: only, and rapidly gets degraded as smaller angular scales are considered,
926: i.e., it seems to be particularly affected by the inclusion of modes that have
927: a low S/N ratio. On the other hand, the $\hat{\beta}$ statistic is more
928: sensitive to the actual signal to noise ratio even at scales where
929: such ratio is below unity, and, as mentioned above, seems to be more efficient
930: in terms of detection of the ISW-LSS cross-correlation.
931:
932: This connects to the multipole or range where most of the correlation is arising.
933: Of the two methods, the matched filter seems to
934: be more confined in $l$-space than the ACPS when looking at the output of the
935: $\hat{\beta}$ statistic: it quotes the maximum
936: significance at $l_{max}=14$ and always drops at larger $l_{max}$-s, whereas
937: the ACPS seems to peak at around $l_{max}=30-40$. In this case, the exception is the
938: NVSS-like survey, for which ${\hat \beta}$ yields the maximum detection
939: significance at $l_{max}=14$. This different behavior
940: suggests that aliasing induced by SDSS-DR4 and PAU-BAO masks is indeed
941: shifting some S/N into the smaller scales (larger $l$-s), but this effect is
942: not perceptible beyond $l_{max}=30-40$. This is also visible in
943: Figs.(\ref{fig:alphas} and \ref{fig:s2n}): under the most aggressive masks,
944: there is less information in the first $l$-bins ($l\in [2,3], [4,5]$), whereas
945: for the NVSS-like survey these contain the largest values of the S/N
946: ratio.
947:
948: These two figures also show that, in almost every $l$-bin, the matched
949: filter $\alpha_l$ estimates seem to be more accurate than the ACPS method, and
950: that, at the same time, they seem to be more affected by aliasing. These two
951: facts are connected: the matched filter tends to pick up the signal from those
952: modes having largest S/N ratio within each $l$-bin. For small $f_{sky}$ and
953: moderately high $l$-bins, these modes are actually aliased components of low
954: $l$ modes whose power has been shifted by the mask into smaller scales. This
955: makes the matched filter provide more accurate $\alpha_l$ values in these high
956: $l$-bins, but these estimates are actually highly correlated to those found at
957: lower $l$-s. This limits the amount of information that high-$l$ bins actually
958: add, and partially explains the high bias of the $\alpha_l$ estimates at large
959: $l$-s in the left panel of Fig.(\ref{fig:alphas}).
960:
961: In this work, we have generalized the implementation of the matched filter
962: into the Fourier space of the 2D sphere, and applied it in the context of CMB
963: analyses and ISW studies. The matched filter provides a tool to estimate the
964: level of presence of some template $\vm$ in some measured signal $\vs$
965: containing a noise component $\vt$. This tool is optimized for the case of
966: $\vt$ being isotropic and Gaussian distributed, and hence is particularly
967: suited for cross-correlation tests where the CMB is the background (noise)
968: signal. In Fourier (or multipole) space, the correlation properties of the CMB
969: are particularly simple (specially, but not only, in the full sky case
970: $f_{sky}=1$). For $f_{sky}<1$, the covariance matrix of the CMB multipoles can
971: be inverted via a SVD approach: this permits simultaneously identifying those
972: Fourier modes containing more information and dropping those other modes that
973: introduce numerical error. After all modes have been sorted in terms of their
974: S/N ratio, the matched filter algorithm weights them accordingly in order to
975: produce an optimal (minimum variance) output for the cross-correlation
976: test. We have compared this method with the standard Angular Cross Band Power
977: Spectrum (ACPS), and found the the matched filter to be either superior or
978: equivalent to the ACPS.
979:
980: In the context of ISW analyses, the matched filter provides estimates of the
981: level of cross-correlation of CMB maps with LSS probes at separate multipole
982: ranges, and this enables a direct and clean comparison to theoretical
983: predictions . When applying both the matched filter and the ACPS methods to
984: three mock surveys having distinct values of $f_{sky}$, we find that both
985: methods perform similarly (the matched filter is slightly more sensitive under
986: aggressive masks, the ACPS more accurate under the NVSS mask). The masks
987: introduce some power aliasing from large into small angular scales, but this
988: does not prevent most of the S/N ratio of the ISW-LSS cross correlation from
989: being confined into the large angular scales ($l<40$). This $l$-confinement
990: may result particularly useful when distinguishing this effect from other
991: secondary anisotropies that, while tracing the LSS distribution, arise at
992: smaller angular scales.
993:
994: Natural extensions of this work involve large angle component separation in future CMB maps
995: when tracers or templates for the signal to be distinguished are available, such as galactic
996: or extragalactic radio, synchrotron or dust maps, large scale 1/f noise component, local kSZ or tSZ contributions, etc.
997:
998: %A natural extension of this work is the construction of a ISW template to be
999: %cross-correlated with real CMB data with the matched filter. For that, it is
1000: %required having a detailed description of the redshift and scale dependence of
1001: %the function $\Pi(z,k) = N_g(z)\; b(z,k)$, and will be the goal of a
1002: %forthcoming work. Other natural applications of this matched filter are
1003: %related to large scale galactic or extragalactic foreground extraction, or
1004: %known 1/f noise component removal.
1005:
1006:
1007: \begin{acknowledgements}
1008:
1009: I am grateful to Tony Banday for carefully reading the manuscript. I acknowledge the use of the HEALPix (\cite{healpix}) package and the LAMBDA\footnote{{\tt http://lambda.gsfc.nasa.gov}} data base.
1010: \end{acknowledgements}
1011:
1012: \begin{thebibliography}{99}
1013:
1014: \bibitem[Afshordi(2004)]{afshordi} Afshordi, N.\ 2004, PhRvD, 70,
1015: 083536
1016:
1017: \bibitem[Bennett et al.(1996)]{cobe2} Bennett, C.~L., et al.\
1018: 1996, ApJL, 464, L1
1019:
1020: \bibitem[Bouchet et al.(1999)]{bouchet} Bouchet, F.~R., Prunet,
1021: S., \& Sethi, S.~K.\ 1999, \mnras, 302, 663
1022:
1023: \bibitem[Cabr{\'e} et al.(2007)]{cabre} Cabr{\'e}, A.,
1024: Fosalba, P., Gazta{\~n}aga, E., \& Manera, M.\ 2007, MNRAS, 381, 1347
1025:
1026: \bibitem[Cay{\'o}n et al.(2000)]{laura} Cay{\'o}n, L., et
1027: al.\ 2000, MNRAS, 315, 757
1028:
1029: \bibitem[Ruhl et al.(2004)]{spt} Ruhl, J., et al.\ 2004,
1030: SPIE, 5498, 11
1031:
1032: \bibitem[Condon et al.(1998)]{nvss} Condon, J.~J., Cotton,
1033: W.~D., Greisen, E.~W., Yin, Q.~F., Perley, R.~A., Taylor, G.~B.,
1034: \& Broderick, J.~J.\ 1998, AJ, 115, 1693
1035:
1036:
1037: \bibitem[Cooray(2002)]{cooray02} Cooray, A.\ 2002, PhRvD, 65,
1038: 103510
1039:
1040: \bibitem[Crittenden \& Turok (1996)]{critturok96} Crittenden, R.~G.,
1041: \& Turok, N.\ 1996, Physical Review Letters, 76, 575
1042:
1043: \bibitem[Douspis et al.(2008)]{marian} Douspis, M., Castro,
1044: P.~G., Caprini, C., \& Aghanim, N.\ 2008, ArXiv e-prints, 802, arXiv:0802.0983
1045:
1046: \bibitem[Dunkley et al.(2008)]{dunkley} Dunkley, J., et al.\
1047: 2008, ArXiv e-prints, 803, arXiv:0803.0586
1048:
1049: \bibitem[Eisenstein et al.(2001)]{dr4} Eisenstein, D.~J.,
1050: et al.\ 2001, AJ, 122, 2267
1051:
1052: \bibitem[Gorski(1994)]{gs_gorski} Gorski, K.~M.\ 1994, ApJL,
1053: 430, L85
1054:
1055: \bibitem[G\'orski et al.(1996)]{alphamethod} G\'orski, K.~M., Banday,
1056: A.~J., Bennett, C.~L., Hinshaw, G., Kogut, A., Smoot, G.~F., \& Wright,
1057: E.~L.\ 1996, ApJL, 464, L11
1058:
1059: \bibitem[G\'orski et al.(2005)]{healpix} G\'orski, K.M., E. Hivon, A.J.
1060: Banday, B.D. Wandelt, F.K. Hansen, M. Reinecke, \& M. Bartelmann, \ 2005
1061: ApJ, 622, 759
1062:
1063: \bibitem[Hansen et al.(2005)]{hansen} Hansen, F.~K.,
1064: Branchini, E., Mazzotta, P., Cabella, P.,
1065: \& Dolag, K.\ 2005, MNRAS, 361, 753
1066:
1067: \bibitem[Hern{\'a}ndez-Monteagudo et al.(2006)]{metals06}
1068: Hern{\'a}ndez-Monteagudo, C., Verde, L., \& Jimenez, R.\ 2006, ApJ, 653, 1
1069:
1070: \bibitem[Hern{\'a}ndez-Monteagudo
1071: \& Rubi{\~n}o-Mart{\'{\i}}n(2004)]{tsz1} Hern{\'a}ndez-Monteagudo, C., \& Rubi{\~n}o-Mart{\'{\i}}n, J.~A.\ 2004, MNRAS, 347, 403
1072:
1073: \bibitem[Hern{\'a}ndez-Monteagudo et al.(2004)]{tsz2}
1074: Hern{\'a}ndez-Monteagudo, C., Genova-Santos, R.,
1075: \& Atrio-Barandela, F.\ 2004, ApJL, 613, L89
1076:
1077: \bibitem[Hinshaw et al.(2003)]{hinshaw} Hinshaw, G., et al.\
1078: 2003, ApJSS 148, 135
1079:
1080: %\bibitem[Ho et al.(2008)]{ho} Ho, S., Hirata, C.~M.,
1081: %Padmanabhan, N., Seljak, U.,
1082: %\& Bahcall, N.\ 2008, ArXiv e-prints, 801, arXiv:0801.0642
1083:
1084: \bibitem[Hu \& Dodelson(2002)]{hudodelson} Hu, W., \& Dodelson, S.\ 2002, ARA\&A, 40, 171
1085:
1086: \bibitem[Kosowsky(2003)]{act} Kosowsky, A.\ 2003, New
1087: Astronomy Review, 47, 939
1088:
1089: \bibitem[Larson et al.(2005)]{larson} Larson, G.~J., Bunn,
1090: E.~F., Kasliwal, V.,
1091: \& McCann, M.\ 2005, Bulletin of the American Astronomical Society, 37, 1429
1092:
1093: \bibitem[Martinez-Gonzalez et al.(1990)]{enrique}
1094: Martinez-Gonzalez, E., Sanz, J.~L., \& Silk, J.\ 1990, ApJL, 355, L5
1095:
1096: \bibitem[Mortlock et al.(2002)]{mortlock} Mortlock, D.~J.,
1097: Challinor, A.~D., \& Hobson, M.~P.\ 2002, \mnras, 330, 405
1098:
1099: \bibitem[Rees \& Sciama(1968)]{rs} Rees, M.~J., \& Sciama, D.~W.\ 1968, Nature, 217, 511
1100:
1101: \bibitem[Reichardt et al.(2008)]{acbar} Reichardt, C.~L., et
1102: al.\ 2008, ArXiv e-prints, 801, arXiv:0801.1491
1103:
1104: \bibitem[Rubi{\~n}o-Mart{\'{\i}}n et al.(2000)]{jal}
1105: Rubi{\~n}o-Mart{\'{\i}}n, J.~A., Atrio-Barandela, F.,
1106: \& Hern{\'a}ndez-Monteagudo, C.\ 2000, ApJ, 538, 53
1107:
1108: \bibitem[Sachs \& Wolfe(1967)]{sw67} Sachs, R.~K., \& Wolfe, A.~M.\ 1967, ApJ, 147, 73
1109:
1110: \bibitem[Sanz \& et al.(1999)]{jlsanz} Sanz, J., \& et al.\ 1999, Evolution of Large Scale Structure : From Recombination to Garching, 53
1111:
1112: \bibitem[Seljak \& Zaldarriaga(1996)]{cmbfast} Seljak, U., \& Zaldarriaga, M.\ 1996, ApJ, 469, 437
1113:
1114: \bibitem[Smoot et al.(1992)]{cobe1} Smoot, G.~F., et al.\
1115: 1992, ApJL, 396, L1
1116:
1117: \bibitem[Bennett et al.(2003)]{wmap1} Bennett, C.~L., et al.\
1118: 2003, ApJSS, 148, 1
1119:
1120: \bibitem[Spergel et al.(2007)]{spergel06} Spergel, D.~N., et al.\
1121: 2007, ApJSS, 170, 377
1122:
1123: \bibitem[Sunyaev \& Zeldovich(1972)]{ksz} Sunyaev, R.~A.,
1124: \& Zeldovich, Y.~B.\ 1972, Comments on Astrophysics and Space Physics, 4,
1125: 173
1126:
1127: \bibitem[Tegmark(1997)]{tegmark} Tegmark, M.\ 1997, PhRvD, 55,
1128: 5895
1129:
1130: \end{thebibliography}
1131:
1132:
1133: %\appendix
1134:
1135: %\section{}
1136:
1137: %\vspace{.2cm}
1138:
1139:
1140:
1141: %\label{lastpage}
1142:
1143: \end{document}
1144:
1145: