1: %\documentclass[manuscript]{aastex}
2: %\documentclass[preprint2]{aastex}
3: %\documentclass[12pt,preprint]{aastex}
4: \documentstyle[emulateapj,pstricks]{article}
5:
6: \newcommand{\Maxima}{{\sc Maxima}}
7: \newcommand{\etal}{{\it et al.}}
8: \newcommand{\DD}{{\mathcal{D}}}
9: \newcommand{\EE}{{\mathcal{E}}}
10:
11: \begin{document}
12:
13: \title{The Trispectrum of the 4 Year $COBE$-DMR data}
14:
15: \author {M.~Kunz\altaffilmark{1}, A.J.~Banday\altaffilmark{2},
16: P.G.~Castro\altaffilmark{1},
17: P.G.~Ferreira\altaffilmark{1}, K.M.~G\'{o}rski\altaffilmark{3,4}}
18: \altaffiltext{1}{Astrophysics, University of Oxford, Keble Road,
19: Oxford OX1 3RH, UK}
20: \altaffiltext{2}{Max Planck Institut fuer Astrophysik, Karl-Schwarzschildstr.
21: 1, Postfach 13 17, D-85741 Garching, Germany}
22: \altaffiltext{3}{European Southern Observatory, Garching, Germany}
23: \altaffiltext{4}{Warsaw University Observatory,
24: Aleje Ujazdowskie 4, 00-478 Warszawa, Poland.}
25:
26: \begin{abstract}
27: We propose an estimator for the trispectrum of a
28: scalar random field on a sphere, discuss its geometrical
29: and statistical properties, and outline its implementation.
30: By estimating the trispectrum of the 4 year $COBE$-DMR data
31: (in HEALPix pixelization) we find new evidence of a non-Gaussian
32: signal associated with a known systematic effect.
33: We find that by removing data from the sky maps for those periods of
34: time perturbed by this effect, the amplitudes of the trispectrum coefficients
35: become completely consistent with predictions for a Gaussian sky.
36: These results reinforce the importance of
37: statistical methods based in harmonic space for quantifying
38: non-Gaussianity.
39: \end{abstract}
40:
41: \keywords{cosmic microwave background --- cosmology: observations}
42:
43: \section{Introduction}
44: The Cosmic Microwave Background (CMB) is the cleanest window
45: on the origin of structure in the very early universe. A complete
46: description of the statistical properties of cosmological fluctuations
47: at a redshift $z\simeq1000$ affords us an essential insight
48: into those processes which may have seeded the formation of
49: galaxies.
50: In a Gaussian theory of structure formation, such as the currently favored
51: model of Inflation, the power spectrum contains all the possible information
52: about the fluctuations. Any higher
53: order moment can subsequently be described in terms of it. However, if
54: the theory is non-Gaussian (as expected for structure formation theories
55: due to local effects from primordial
56: phase transitions or more generally from non-linear processes),
57: then there will be deviations from the simple Gaussian expressions
58: for the higher order moments.
59: Such behavior can serve as a powerful discriminator
60: between different models of structure formation.
61:
62: Most analyses of CMB data to-date have focused on the
63: angular power spectrum and its sensitivity to various parameters
64: of cosmological theories.
65: Some work has been done on the estimation of the three
66: point correlation function and its analogue in spherical harmonic
67: space, with intriguing results (Heavens 1998, Ferreira,
68: Magueijo and G\'{o}rski 1998, Magueijo 2000, Banday, Zaroubi
69: and G\'{o}rski 2000). It is the purpose of this
70: letter to propose a method for estimating the four point
71: spectrum, the {\it trispectrum}, and to apply it to the
72: $COBE$ 4 year DMR data. This work complements the recent
73: work of Hu (2001) where some of the properties of the angular
74: trispectrum of the CMB are discussed.
75:
76: The outline of this letter is as follows. In Section \ref{est} we
77: construct a set of orthonormal estimators and describe their properties
78: for a Gaussian random field. In Section \ref{res} we apply the
79: estimators to the $COBE$ 4 year DMR data. We show that we
80: detect the non-Gaussian signal found in Ferreira, Magueijo $\&$ G\'{o}rski
81: (1998) and that it can be explained by the arguments presented
82: in Banday, Zaroubi \& G\'{o}rski (2000), and in particular that this
83: is a manifestation of a known systematic effect.
84: We therefore conclude that
85: the $COBE$ 4 year data is consistent with a Gaussian cosmological signal.
86: In Section \ref{dis} we summarize our results.
87:
88:
89: \section{The estimator}
90: \label{est}
91: In this section we wish to construct a set of quantities for
92: estimating the trispectrum of a random field on the sphere.
93: The temperature anisotropy in a given direction on the celestial
94: sphere, $T({\bf n})$, can be expanded in
95: terms of spherical harmonic functions, $Y_{\ell m}({\bf n})$:
96: \begin{equation}
97: T({\bf n})=\sum_{\ell m}a_{\ell m}Y_{\ell m}({\bf n})
98: \end{equation}
99: For any theory of structure formation, the $a_{\ell m}$ coefficients are
100: a set of random variables; we shall restrict ourselves to
101: theories which are statistically homogeneous and isotropic.
102: In this case we can define the power spectrum $C_\ell$ of the
103: temperature anisotropies by $\langle a_{\ell m}
104: a^*_{\ell' m'}\rangle=C_{\ell}\delta_{\ell \ell'}$.
105:
106: We now seek to construct a set of tensors that are geometrically
107: independent, describe their statistical properties
108: for a Gaussian random field and then discuss the practical
109: issue of their implementation.
110: Given a set of $a_{\ell m}$ we wish to find the index structure of
111: the set of four point correlators such that (1) they are
112: rotationally invariant (2) they form a complete basis
113: (preferably orthonormal) of the whole space of admissible
114: four-point correlators and
115: (3) they satisfy the appropriate symmetries under interchanges of
116: $m$- and $\ell$-values. We shall restrict ourselves to the case in which
117: $\ell_1=\ell_2=\ell_3=\ell_4=\ell$. Furthermore, throughout this
118: section we keep $\ell$ fixed. We determine the tensor ${\cal T}$
119: such that
120: \begin{equation}
121: \langle a_{\ell m_1} a_{\ell m_2} a_{\ell m_3} a_{\ell m_4}\rangle
122: =\sum_{a=0}^{n}T_{\ell;a}{\cal T}^{a;\ell}_{m_1m_2m_3m_4} \label{4mom}
123: \end{equation}
124: where $n$=${\rm int}(\ell/3)$ (due to reflection, permutation and rotational
125: symmetry). The $T_{\ell;a}$ values are then the components of
126: the trispectrum which we wish to estimate.
127: The explicit form of the
128: ${\cal T}$ are
129: \begin{eqnarray}
130: &&{\cal T}^{a;\ell}_{m_1m_2m_3m_4}=\sum_{\alpha=0}^{\ell}
131: {\cal L}^{a\alpha}_{\ell}
132: {\bar {\cal T}}^{\alpha;\ell}_{m_1m_2m_3m_4} \label{llmatrix} \\
133: &&{\bar {\cal T}}^{\alpha;\ell}_{m_1m_2m_3m_4}=
134: \sum_{M=-2\alpha}^{2\alpha}(-1)^M
135: \left( \begin{array}{ccc} \ell & \ell & 2\alpha \\
136: m_1 & m_2 & M \end{array} \right)\times\nonumber\\
137: &&\left( \begin{array}{ccc} 2\alpha & \ell & \ell \\
138: -M & m_3 & m_4 \end{array} \right) + \mbox{inequiv. permutations}
139: \label{tens}
140: \end{eqnarray}
141: where the matrices in parentheses are the Wigner 3-J symbols. The
142: ${\bar {\cal T}}^{\alpha;\ell}$ are not orthogonal and satisfy
143: \begin{eqnarray}
144: {\bar {\cal T}}^{\alpha;\ell}_{m_1m_2m_3m_4}{\bar {\cal T}}^{\alpha;\ell}_{m_1m_2m_3m_4}=\frac{3}{4\alpha+1}\delta_{\alpha\beta}+
145: 6
146: \left\{ \begin{array}{ccc} \ell & \ell & 2\alpha \\
147: \ell & \ell & 2\beta \end{array} \right\} \nonumber
148: \end{eqnarray}
149: (where summation over the $m_i$ is assumed) which has rank $n+1$.
150: The matrix ${\cal L}_\ell$ in (\ref{llmatrix})
151: is a rectangular matrix (with a triangular
152: sub-block) with $n+1$ columns and $\ell+1$ rows. It is constructed
153: through a Gram-Schmidt procedure by subtracting for each $\alpha$
154: (starting from $\alpha=0$)
155: the projection onto all $a'<a$ and then normalizing the result.
156: The $\alpha=0$ (and hence $a=0$) tensor is proportional to the
157: Gaussian contribution.
158: This can be easily seen given that for $\alpha=0$ the Wigner 3J
159: symbols are simply Kronecker $\delta$ symbols in the corresponding indices.
160: The remaining $a>0$ terms contain therefore no Gaussian signal
161: and quantify the non-Gaussian part of the trispectrum.
162:
163: The ${\cal T}$ are orthonormal and can be used to construct an
164: estimator for $T_a$ from a realization of $a_{\ell m}$:
165: \begin{equation}
166: {\hat T}_{\ell;a}={\cal T}^{a;\ell}_{m_1m_2m_3m_4}a_{\ell m_1} a_{\ell m_2} a_{\ell m_3} a_{\ell m_4} \label{tnull}
167: \end{equation}
168: For a Gaussian random field we expect
169: $\sigma^2[{\hat T}_{\ell;0}]\gg\sigma^2[{\hat T}_{\ell;a}]$ for
170: $a>0$, where $\sigma^2[A]$ denotes the variance of the random
171: variable $A$ and ${\hat T}_{\ell;0}$ is simply the square of the minimum
172: variance estimator of the $C_\ell$. One finds that
173: $\langle{\hat T}_{\ell;a}\rangle=0$ and $\sigma^2[{\hat T}_{\ell;a}]=
174: 24 C_\ell^4$ for all $a>0$.
175:
176: To show that the
177: ${\hat T}_{\ell;a}$ constitute a family of minimum variance estimators
178: we construct a linear combination of the estimators:
179: \begin{equation}
180: {\cal T}^{\ell}_{m_1m_2m_3m_4}=\sum_{a=0}^{n}c_a{\cal T}^{a;\ell}_{m_1m_2m_3m_4}
181: \end{equation}
182: and minimize the function
183: \begin{eqnarray}
184: \sigma^2_\ell[c_a,\lambda]=\langle({\cal T}^{\ell}_{m_1m_2m_3m_4}a_{\ell m_1} a_{\ell m_2} a_{\ell m_3} a_{\ell m_4})^2\rangle \nonumber \\
185: -\langle{\cal T}^{\ell}_{m_1m_2m_3m_4}a_{\ell m_1}
186: a_{\ell m_2} a_{\ell m_3} a_{\ell m_4}\rangle^2 \nonumber \\
187: -\lambda C^2_\ell ({\cal T}^{\ell}_{m_1m_2m_3m_4}
188: {\cal T}^{\ell}_{m_1m_2m_3m_4}-1)
189: \end{eqnarray}
190: where summation over all $m_i$ is implied. The last term, a Lagrange
191: multiplier, ensures that ${\cal T}^{\ell}$ is normalized. We solve
192: $\partial_{c_a} \sigma^2_\ell[c_a,\lambda]=\partial_\lambda \sigma^2_\ell[c_a,\lambda]=0$
193: to find a set of two equations
194: \begin{equation}
195: (24I+72A_\ell)^{ab}c_b+\lambda c_b=0 \quad \mathrm{and} \quad c^2=1
196: \end{equation}
197: where
198: $$
199: A^{ab}_\ell={\cal T}^{a;\ell}_{m_1m_2m_am_a}{\cal T}^{b;\ell}_{m_1m_2m_bm_b} .
200: $$
201: This is an eigenvector equation where, for a given eigenvector ${\bf c}$,
202: the eigenvalue $\lambda$ will give the expected variance of the
203: estimator.
204: Of the $n+1$ eigenvalues, one is large and
205: has an eigenvector proportional to ${\hat T}_{\ell;0}$.
206: The remaining eigenvalues
207: have an amplitude $\lambda=24$ and each eigenvector is
208: a ${\hat T}_{\ell;a}$ for $a>0$.
209:
210:
211: Note that we can relate our parameterization to the one proposed in Hu (2001);
212: If we reexpress equation (\ref{4mom}) as
213: \begin{equation}
214: \langle a_{\ell m_1} a_{\ell m_2} a_{\ell m_3} a_{\ell m_4}\rangle
215: =\sum_{\alpha=0}^{\ell}{\bar T}_{\ell;\alpha}{\bar {\cal T}}^{\alpha;\ell}_{m_1m_2m_3m_4}
216: \end{equation}
217: where ${T}_{\ell;a}={\cal L}^{a\alpha}_{\ell}{\bar T}_{\ell;\alpha}$
218: then $Q^{\ell \ell}_{\ell \ell}$ as defined in equation 15 of
219: Hu (2001) can be written as
220: \begin{eqnarray}
221: Q^{\ell \ell}_{\ell \ell}(2\alpha)={\bar T}_{\ell;\alpha}+
222: 2(4\alpha+1)\sum_{\beta}
223: \left\{ \begin{array}{ccc} \ell & \ell & 2\alpha \\
224: \ell & \ell & 2\beta \end{array} \right\}{\bar T}_{\ell;\beta} .
225: \end{eqnarray}
226:
227: The numerical implementation of these estimators is more involved
228: than for the bispectrum. If we omit the numerous symmetries,
229: we have to consider for each $\ell$
230: a set of up to $8\ell^3$ Wigner 3J symbols (compared to
231: just one for the bispectrum). There are reasonably fast ways
232: for constructing the Wigner 3J symbols (Schulten and Gordon 1976) but the number
233: of operations per estimator scales as ${\cal O}(\ell^6)$.
234: For repeated computations of the estimators (eg. in Monte Carlo studies),
235: this can partially be avoided by
236: storing the precomputed estimators in a lookup table, with
237: the amount of memory required
238: scaling as ${\cal O}(\ell^4)$.
239:
240: Clearly, to be able to estimate the trispectrum
241: on small angular scales, approximate methods must be developed
242: to make the procedure computationally feasible.
243: However, the ability to constrain non-Gaussianity on large angular
244: scales is in any case more important physically for two reasons;
245: the ratio of the non-Gaussian to the Gaussian signal will
246: in general be higher for lower moments, and
247: the signal to noise is better for low $l$.
248: To understand these points, let us assume a source for non-Gaussianity which
249: leads to approximately scale invariant moments of the gravitational
250: potential on arbitrary scales. i.e. $\langle \Phi(R)^N\rangle$ is
251: constant for any $R$,
252: where $\Phi(R)$ is the gravitational potential within a ball of
253: radius $R$ and $\langle \cdots \rangle$ denotes the ensemble average.
254: This might be expected from a primordial source with no preferred
255: scale such as inflation (Komatsu \& Spergel 2000)
256: or from an active source where the only
257: scale is set by the horizon today (Durrer {\it et al} 2000).
258: Current observations of the CMB certainly favor such
259: scale-invariant descriptions of the potential.
260: One then expects the moment of order $N$ of the $a_{\ell m}$ to scale
261: as $\ell^{2(1-N)}$. This signal will be competing against the
262: fluctuations due to the disconnected (or Gaussian) part, which is
263: proportional to $N!\ell^{-(2N+1)/2}$, the former therefore dominating
264: for $N>2$. Since the power spectrum for white noise has constant amplitude,
265: the signal to noise as a function of scale will have the same form as
266: the scale invariant power spectrum itself, therefore being
267: larger for smaller $l$, ie. larger angular scales.
268:
269: \section{Results}
270: \label{res}
271: As an application of the formalism described in Section
272: \ref{est}, we estimate the trispectrum of the
273: coadded 53 and 90 GHz $COBE$-DMR 4 year sky maps in HEALPix format
274: (G\'{o}rski {\it et al} 1999). The resolution of the maps
275: is $N_{\mbox{side}}=64$ or $49152$ pixels.
276: We do not extend our analysis beyond $\ell_{\mbox{max}}=20$
277: since the signal to noise is poor for higher $l$.
278: Hence the maximal number of independent non-Gaussian estimators
279: for the trispectrum is ${\rm int}(\ell_{\mbox{max}}/3)=6$.
280: We set the pixels in the extended Galactic
281: cut (Banday {\it et al} 1997) to zero and subtract the residual monopole and
282: dipole of the resulting map. After convolving the maps
283: with spherical harmonics to extract a set of
284: $a_{lm}$'s for $l \leq 20$ we then apply equation (\ref{tnull}).
285: To validate our software,
286: we have estimated the bispectrum of the $COBE$-DMR 4 year sky data
287: repixelized in the HEALPix format (for convenience denoted by {\sc ec})
288: and reproduced the results of Ferreira, Magueijo \& G\'{o}rski (1998),
289: and in particular the strong non-Gaussian
290: signal present at $\ell=16$. When an equivalent map, from which
291: that part of the DMR time stream contaminated by the
292: `eclipse effect\footnote{The \lq eclipse effect' was an orbitally
293: modulated signal which took place for approximately two months every year
294: around the June solstice when the $COBE$ spacecraft repeatedly flew through the
295: Earth's shadow.}'
296: is removed (denoted {\sc nec}), is subsequently analyzed
297: we also reproduce the results of Banday, Zaroubi
298: \& G\'{o}rski 2000, namely that the non-Gaussian signal is no longer detected.
299: For our subsequent analysis we will present the trispectra of
300: {\it both} the {\sc ec} and {\sc nec} data.
301:
302: One of our primary concerns is to compare our results with
303: the assumption that the CMB sky measured by $COBE$-DMR is Gaussian. To do
304: so, we generate 10000 full-sky
305: maps at the same resolution using a scale invariant power spectrum normalized to
306: $Q_{rms-PS}=18\mu$K (G\'{o}rski {\it et al} 1998).
307: We convolve each map with the DMR beam and add uncorrelated
308: pixel noise with rms amplitude $\sigma_n = 15.95 \mathrm{mK} / \sqrt{N_{obs}}$,
309: (where $N_{obs}$ is the number of times a given pixel was observed);
310: we then subject the synthetic map to the same procedure
311: as the original data.
312:
313: \vbox{\vskip -0.cm \hskip -0.5cm\epsfxsize=9.cm\epsfbox{f1.eps}}
314: \vskip 0.1cm
315: { \small
316: F{\scriptsize IG}.~1.--- The six estimators
317: of the normalized trispectrum applied to the
318: {\sc ec} data (circles) and the {\sc nec} data (crosses).
319: 95\% of all simulated Gaussian
320: skies lie within the solid lines. Although removing the `eclipse'
321: data changes the noise properties, we find that the Gaussian confidence
322: limits essentially remain unchanged.
323: \label{fig2}}
324: \vskip 0.25cm
325: %\begin{figure}
326: %\epsscale{0.6}
327: %\plotone{f1.eps}
328: %\caption{The six estimators
329: %of the normalized trispectrum applied to the
330: %{\sc ec} data (circles) and the {\sc nec} data (crosses).
331: %95\% of all simulated Gaussian
332: %skies lie within the solid lines. Although removing the `eclipse'
333: %data changes the noise properties, we find that the Gaussian confidence
334: %limits essentially remain unchanged.
335: %\label{fig2}}
336: %\end{figure}
337:
338: Figure 1 shows the trispectra of the DMR data together
339: with Gaussian 95\% confidence limits.
340: Instead of the ``raw'' estimator (\ref{tnull}) we prefer to use the
341: normalized trispectrum,
342: $\tau^{(a)}_\ell={\hat T}_{\ell;a}/{\hat C}^2_{\ell}$ for $a>1$ (where
343: ${\hat C}_{\ell}=\frac{1}{2\ell+1}\sum_{m}|a_{\ell m}|^2$),
344: thus effectively removing the dependence on the power spectrum.
345: This prevents fluctuations in the power spectrum from introducing
346: spurious signals and from masking real non-Gaussianities.
347: Figure 1 shows that in this case, most values fall within the $95\%$
348: confidence lines and demonstrate the scatter expected for a
349: Gaussian random field.
350:
351: Of particular interest is the value of the normalized $\tau^{(3)}$ at
352: $\ell=16$ in figure 1. One finds that $99.9 \%$ of the Gaussian
353: models in the {\sc ec} case have a smaller $\tau^{(3)}$ than the
354: measured one. This is clearly a manifestation of the non-Gaussianity found
355: in Ferreira, Magueijo \& G\'{o}rski (1998) which is highly localized
356: in $\ell$ space. However, if we estimate $\tau^{(3)}$ for the {\sc nec}
357: we find that it falls comfortably within the $95\%$ confidence limits.
358: This leads us to believe that this detection of non-Gaussianity results
359: from the `eclipse effect',
360: consistent with the hypothesis of Banday, Zaroubi \& G\'{o}rski (1999).
361:
362: \vbox{\vskip -0.cm \hskip -0.5cm\epsfxsize=9.cm\epsfbox{f2.eps}}
363: \vskip 0.1cm
364: { \small
365: F{\scriptsize IG}.~2.--- The $\chi^2$ distribution of the
366: Gaussian models (histogram) and the actual data value (dotted line)
367: for the {\sc ec} (top graphs) and {\sc nec} (bottom graphs) datasets.
368: The left two graphs show $\tau^{(3)}$ which contains the main
369: contribution to the non-Gaussian signal and the right graphs
370: show the total $\chi^2$ over all six non-Gaussian estimators,
371: $\tau^{(1)}$ to $\tau^{(6)}$.
372: \label{fig3}}
373: \vskip 0.25cm
374:
375: Let us now construct a goodness of fit for our statistic. In
376: Ferreira, Magueijo \& G\'{o}rski (1998), a modified $\chi^2$ was
377: constructed which took into account the non-Gaussian distribution
378: of each method: as above, the distribution of each estimator for
379: a Gaussian sky was constructed and used as an approximate
380: likelihood function to evaluate the goodness of fit. One
381: shortcoming of such a method was that correlations between
382: the estimates for different $\ell$s were discarded.
383: To include them, we use the Gaussian ensemble of data sets to derive the
384: expectation values $<>_G$ and the covariance matrix $C$ for both
385: the power spectrum, $C_\ell$, and all seven trispectrum
386: estimators, $\tau^{(0)}_\ell$ to $\tau^{(6)}_\ell$. We
387: proceed to calculate the $\chi^2$ value for the estimator
388: $\EE$ and the data set $\DD$,
389: \begin{equation}
390: \chi^2 [\EE,\DD] \equiv \sum_{\ell,\ell'}
391: \left(\left< \EE_\ell \right>_{\mathrm{G}} - \EE[\DD]_\ell\right)
392: C^{-1}_{\ell\ell'}
393: \left(\left< \EE_{\ell'} \right>_{\mathrm{G}} - \EE[\DD]_{\ell'}\right) ,
394: \end{equation}
395: using as data sets the {\sc ec} data and the {\sc nec} data.
396: Finally we use another 10000 Gaussian realizations to estimate the
397: expected distribution of the $\chi^2$ for both the {\sc ec} and the
398: {\sc nec} data.
399:
400: %\begin{figure}
401: %\epsscale{0.6}
402: %\plotone{f2.eps}
403: %\caption{The $\chi^2$ distribution of the
404: %Gaussian models (histogram) and the actual data value (dotted line)
405: %for the {\sc ec} (top graphs) and {\sc nec} (bottom graphs) datasets.
406: %The left two graphs show $\tau^{(3)}$ which contains the main
407: %contribution to the non-Gaussian signal and the right graphs
408: %show the total $\chi^2$ over all six non-Gaussian estimators,
409: %$\tau^{(1)}$ to $\tau^{(6)}$. \label{fig3}}
410: %\end{figure}
411:
412: For all normalized non-Gaussian trispectrum estimators ($\tau^{(1)}$
413: to $\tau^{(6)}$) we find that 94\% of the Gaussian models
414: have a smaller $\chi^2$ than the {\sc ec} data as can be
415: seen in figure 2.
416: As expected the main contribution to the $\chi^2$ for the
417: {\sc ec} data stems from $\tau^{(3)}$ at $\ell = 16$;
418: indeed, this is the only
419: normalized trispectrum estimator which exhibits any significant
420: non-Gaussianity, in this case at about 99.9\%.
421: If we use the {\sc nec} data, the detection
422: vanishes. In this case, 60\% of all Gaussian models have a lower
423: $\chi^2$ when computed over all six trispectrum estimators
424: (83\% for $\tau^{(3)}$ alone). Hence the {\sc nec} data
425: is compatible with Gaussianity.
426:
427:
428:
429:
430:
431: \section{Discussion}
432: \label{dis}
433: In this paper, we have derived an estimator for the trispectrum
434: of a scalar random field on the sphere. Application of this
435: estimator, normalized by the power spectrum
436: (a procedure adopted in Ferreira, Magueijo \& G\'{o}rski, 1998
437: for the bispectrum, see also Komatsu {\it et al} 2002 for
438: a detailed discussion), to the $COBE$-DMR data
439: provides evidence for non-Gaussianity at the 94\% confidence level.
440: As in the case of the bispectrum, the signal is mainly
441: present in the $\ell=16$ multipole (and the $\tau^{(3)}$ estimator here).
442: However, when data is excluded to correct for the `eclipse effect',
443: the non-Gaussian behavior is removed, allowing us to conclude that
444: the non-Gaussianity present in the uncorrected sky maps is not
445: cosmological in origin.
446:
447: The detection of a signal that is so strongly localized
448: in $\ell$ space provides convincing support to our contention
449: that the trispectrum is an important and sensitive probe
450: of non-Gaussianity in the frequency (scale) domain. It affords
451: complementary information to the bispectrum since it is an even
452: moment, and, despite the higher computational effort required,
453: has the obvious advantage in that it can probe all values of
454: $\ell$, not just the even ones.
455:
456: Interestingly enough, from a theoretical perspective,
457: there may be some possible sources of non-Gaussianity
458: for which the trispectrum provides a far more sensitive test
459: than the bispectrum. In many
460: cases a given moment of the $a_{\ell m}$s can be expressed as the
461: projection of a cosmological field. If that field is vector-like
462: in nature (as in the case of the Doppler effect or the Ostriker-Vishniac effect
463: and its non-linear extensions), any
464: odd moment may suffer from the Sunyaev-Kaiser cancellation, where
465: the integral of a given wavenumber, $k$, over a smoothly varying projection
466: function with width $\sigma$ tends to suppress the moment by a factor
467: of order $1/(\sigma k)^2$ (Sunyaev 1978, Kaiser 1985, Scannapieco 2000).
468: For even moments one can always construct a
469: scalar component which will not be subject to this cancellation.
470: Such a tool will be of great use in the analysis of the data sets
471: from the MAP and Planck Surveyor satellites.
472:
473:
474:
475: \section{Acknowledgments}
476:
477: We gratefully acknowledge use of the HEALPix software package
478: in this publication (see {\tt http://www.eso.org/science/healpix/}).
479: We thank James Binney, Carlos Contaldi, Michael Joyce, Janna Levin
480: and Jo\~ao Magueijo
481: for stimulating discussions. MK acknowledges
482: financial support from the Swiss National Science Foundation
483: under contract 83EU-062445. PGC is supported by the Funda\c{c}\~{a}o
484: Ci\^encia e Tecnologia. PGF thanks the Royal Society for
485: support.
486:
487: \begin{references}
488: \reference{bgal} A.J.~Banday {\it et al}, Astrophys. J. {\bf 475}, 393 (1997).
489: \reference{bzg} A.J.~Banday, S.~Zaroubi \& K.M.~G\'{o}rski, Astrophys. J.
490: {\bf 533}, 575 (2000).
491: \reference{d} R.~Durrer, R.~Juszkiewicz,M.~Kunz \& J.P.~Uzan Phys.Rev.
492: {\bf D62} 021301 (2000).
493: \reference{fmg} P.G.~Ferreira, J.~Magueijo \& K.M.G\'{o}rski,
494: Astrophys. J. {\bf 503}, L1 (1998).
495: \reference{} K.M. G\'{o}rski, E.Hivon \& B.D.Wandelt, {\it Evolution of
496: Large Scale Structure: from Recombination to Garching} ed. A.J.Banday,
497: R.Sheth \& L.N. Da Costa, PrintPartners Ipskamp, NL, 37 (1999).
498: \reference{} K.M. G\'{o}rski {\it et al}, Astrophys. J. {\bf 464}, L11 (1998).
499: \reference{heavens} A.F.Heavens, Mon.Not.R.Astron.Soc. {\bf 299}, 805 (1998).
500: \reference{} W.Hu, {\tt astro-ph/0105117} (2001).
501: \reference{kaiser} N. Kaiser, Astroph. J. {\bf 282}, 374 (1984).
502: \reference{komatsu} E.~Komatsu \& D.~N.~Spergel Phys. Rev. {\bf D63} 063002
503: (2001).
504: \reference{} E. Komatsu {\it et al}, Astrophys. J. {\bf 566}, 19 (2002).
505: \reference{mag}J.~Magueijo, Astrophys.J. {\bf 528}, L57 (2000).
506: \reference{scannapiecco} E. Scannapieco, Astrophys. J. {\bf 540}, 20 (2000).
507: \reference{sg} K. Schulten and R.G. Gordon,
508: Computer Phys Comm {\bf 11}, 269 (1976).
509: \reference{sunyaev} R.A. Sunyaev, {\it IAU sumposium 79, Large Scale
510: Structure of the Universe} ed. M.S.Longair and J. Einasto, Dordrecht:Reidel,
511: 393 (1979).
512: \end{references}
513:
514: {\it Note added in proof} -- E. Komatsu investigates the trispectrum of the
515: $COBE$ DMR data in his Ph. D. thesis. His conclusions agree with ours, namely
516: that the $COBE$ data is consistent with Gaussian initial fluctuations.
517:
518: \end{document}
519: