EURASIP Journal on Applied Signal Processing
Applications of Signal Processing in Astrophysics and Cosmology
Guest Editors: Ercan E. Kuruoglu and Carlo Baccigalupi
EURASIP Journal on Applied Signal Processing Applications of Signal Processing in Astrophysics and Cosmology
EURASIP Journal on Applied Signal Processing Applications of Signal Processing in Astrophysics and Cosmology
Guest Editors: Ercan E. Kuruoglu and Carlo Baccigalupi
Copyright © 2005 Hindawi Publishing Corporation. All rights reserved.
This is a special issue published in volume 2005 of “EURASIP Journal on Applied Signal Processing.” All articles are open access articles distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Editor-in-Chief Marc Moonen, Belgium
Senior Advisory Editor K. J. Ray Liu, College Park, USA Associate Editors Gonzalo Arce, USA Arden Huang, USA King N. Ngan, Hong Kong Jaakko Astola, Finland Jiri Jan, Czech Douglas O’Shaughnessy, Canada Kenneth Barner, USA Søren Holdt Jensen, Denmark Antonio Ortega, USA Mauro Barni, Italy Mark Kahrs, USA Montse Pardas, Spain Jacob Benesty, Canada Thomas Kaiser, Germany Wilfried Philips, Belgium Kostas Berberidis, Greece Moon Gi Kang, Korea Vincent Poor, USA Helmut Bölcskei, Switzerland Aggelos Katsaggelos, USA Phillip Regalia, France Joe Chen, USA Walter Kellermann, Germany Markus Rupp, Austria Chong-Yung Chi, Taiwan Lisimachos P. Kondi, USA Hideaki Sakai, Japan Satya Dharanipragada, USA Alex Kot, Singapore Bill Sandham, UK Petar M. Djurić, USA C.-C. Jay Kuo, USA Dirk Slock, France Jean-Luc Dugelay, France Geert Leus, The Netherlands Piet Sommen, The Netherlands Frank Ehlers, Germany Bernard C. Levy, USA Dimitrios Tzovaras, Greece Moncef Gabbouj, Finland Mark Liao, Taiwan Hugo Van hamme, Belgium Sharon Gannot, Israel Yuan-Pei Lin, Taiwan Jacques Verly, Belgium Fulvio Gini, Italy Shoji Makino, Japan Xiaodong Wang. USA A. Gorokhov, The Netherlands Stephen Marshall, UK Douglas Williams, USA Peter Handel, Sweden C. Mecklenbräuker, Austria Roger Woods, UK Ulrich Heute, Germany Gloria Menegaz, Italy Jar-Ferr Yang, Taiwan John Homer, Australia Bernie Mulgrew, UK
Contents
Editorial, Ercan E. Kuruoglu and Carlo Baccigalupi Volume 2005 (2005), Issue 15, Pages 2397-2399
Separation of Correlated Astrophysical Sources Using Multiple-Lag Data Covariance Matrices, L. Bedini, D. Herranz, E. Salerno, C. Baccigalupi, E. E. Kuruoğlu, and A. Tonazzini Volume 2005 (2005), Issue 15, Pages 2400-2412
Adapted Method for Separating Kinetic SZ Signal from Primary CMB Fluctuations, Olivier Forni and Nabila Aghanim Volume 2005 (2005), Issue 15, Pages 2413-2425
Detection of Point Sources on Two-Dimensional Images Based on Peaks, M. López-Caniego, D. Herranz, J. L. Sanz, and R. B. Barreiro Volume 2005 (2005), Issue 15, Pages 2426-2436
Blind Component Separation in Wavelet Space: Application to CMB Analysis, Y. Moudden, J.-F. Cardoso, J.-L. Starck, and J. Delabrouille Volume 2005 (2005), Issue 15, Pages 2437-2454
Analysis of the Spatial Distribution of Galaxies by Multiscale Methods, J-L. Starck, V. J. Martínez, D. L. Donoho, O. Levi, P. Querre, and E. Saar Volume 2005 (2005), Issue 15, Pages 2455-2469
Cosmological Non-Gaussian Signature Detection: Comparing Performance of Different Statistical Tests, J. Jin, J.-L. Starck, D. L. Donoho, N. Aghanim, and O. Forni Volume 2005 (2005), Issue 15, Pages 2470-2485
Time-Scale and Time-Frequency Analyses of Irregularly Sampled Astronomical Time Series, C. Thiebaut and S. Roques Volume 2005 (2005), Issue 15, Pages 2486-2499
Restoration of Astrophysical Images—The Case of Poisson Data with Additive Gaussian Noise, H. Lantéri and C. Theys Volume 2005 (2005), Issue 15, Pages 2500-2513
A Data-Driven Multidimensional Indexing Method for Data Mining in Astrophysical Databases, Marco Frailis, Alessandro De Angelis, and Vito Roberto Volume 2005 (2005), Issue 15, Pages 2514-2520
Virtually Lossless Compression of Astrophysical Images, Cinzia Lastri, Bruno Aiazzi, Luciano Alparone, and Stefano Baronti Volume 2005 (2005), Issue 15, Pages 2521-2535
Astrophysical Information from Objective Prism Digitized Images: Classification with an Artificial Neural Network, Emmanuel Bratsolis Volume 2005 (2005), Issue 15, Pages 2536-2545 Multiband Segmentation of a Spectroscopic Line Data Cube: Application to the HI Data Cube of the Spiral Galaxy NGC 4254, Farid Flitti, Christophe Collet, Bernd Vollmer, and François Bonnarel Volume 2005 (2005), Issue 15, Pages 2546-2558
Adaptive DFT-Based Interferometer Fringe Tracking, Edward Wilson, Ettore Pedretti, Jesse Bregman, Robert W. Mah, and Wesley A. Traub Volume 2005 (2005), Issue 15, Pages 2559-2572
Technique for Automated Recognition of Sunspots on Full-Disk Solar Images, S. Zharkov, V. Zharkova, S. Ipson, and A. Benkhalil Volume 2005 (2005), Issue 15, Pages 2573-2584
On-board Data Processing to Lower Bandwidth Requirements on an Infrared Astronomy Satellite: Case of Herschel-PACS Camera, Ahmed Nabil Belbachir, Horst Bischof, Roland Ottensamer, Franz Kerschbaum, and Christian Reimers Volume 2005 (2005), Issue 15, Pages 2585-2594 EURASIP Journal on Applied Signal Processing 2005:15, 2397–2399 c 2005 Hindawi Publishing Corporation
Editorial
Ercan E. Kuruoglu Istituto di Scienza e Tecnologie dell’Informazione “A. Faedo” (ISTI), Area della ricerca CNR di Pisa, via G. Moruzzi 1, 56124 Pisa, Italy Email: [email protected] Carlo Baccigalupi Scuola Internazionale Superiore di Studi Avanzati (SISSA/ISAS), via Beirut 4, 34014 Trieste, Italy Email: [email protected]
We live in an epoch where the frontiers of our investigation formidable challenge for signal processing. We need state-of- and comprehension of fundamental physics depend largely the-art techniques that can analyse, summarise, and extract on the light coming from the sky, that is, on the study of the necessary information from this ocean of data. galactic and extra-galactic radiation. Watching the sky, in Tocontinue with the example above, the microwave sky is principle, we have access to the highest energies conceiv- dominated by the CMB radiation, but several processes con- able, generated by the laws of nature in extreme conditions, tribute to the total emission, coming for instance from all the such as nearby black holes or even close to the origin of processes occurring along the line of sight, such as the emis- the universe itself. For example, in the microwave band, sion from other galaxies or clusters of those, as well as from the extra-Galactic radiation is dominated by a markedly the diffuse gas in our own Galaxy. Each of these processes are isotropic component, obeying a black body spectrum char- most relevant in different contexts in astrophysics and cos- acterized by a temperature of about 2.726 Kelvin. That is mology. Recently, the astrophysics field has benefited a great the relic of the Big Bang, originated just 300 000 years af- deal from the rich research work going on source separation ter the initial starting point of the universe. This radiation, in the signal processing field. Source separation aims at the namely the cosmic microwave background (CMB) radia- recovery of the various different components from the multi- tion, today is the most important observable we have to ac- band observations exploiting the differences between them, cess the mysterious physics of the Big Bang itself. The lat- induced by their independent physical origins. ter is telling us about the unknown fundamental interac- Despite the mutual interest, the two disciplines suffer tions and particles, the physics of spacetime, and the na- from lack of a common publication ground, implying that ture of quantum gravity, and represents the only way to ad- the results produced in one of them are not immediately vis- dress those issues in physics today. Electronics hardware tech- ible in the other. The aim of the present issue is to provide nology has reached in these very recent years the capabil- a unified platform that would strengthen the bridge between ity to study the tiniest details of the CMB, carrying the im- signal processing and astrophysics and cosmology and enable age of the primordial stage of cosmic geometry, structure, the sharing of information. We would like to provide astro- and composition. Such a fantastic challenge is ongoing in physicists and cosmologists with a spectrum of the most ad- this very moment, while several CMB detectors are operating vanced signal processing techniques and the signal process- and advanced probes are being designed for the forthcoming ing community an exposure to various vital real problems in decades. analysing astrophysics data that await solution. Finally, our Many breakthroughs in physics are made possible by aim is to provide a reference for present and future literature, the use of the most advanced data analysis techniques. The in the widest possible context, accounting for various appli- present datasets obtained in astrophysical and cosmologi- cations and algorithms proposed. Indeed, as the reader may cal observations are huge, and cover the entire electromag- see, the topics we collected range from solar physics, thus on netic spectrum, dealing with very different processes, from the scale of stars, to the reconstructuon of the most ambi- gamma and X-rays of the high-energy astrophysics of com- tious signal from the Big Bang, with the reconstruction of pact stars or black holes, to the microwave and infrared emis- the CMB pattern on all sky. The methods presented in the sion from the whole large-scale universe. This variety of the issue range from transform domain analysis of such wavelets observational techniques and signals to deal with represents a to data mining techniques. 2398 EURASIP Journal on Applied Signal Processing
We start the special issue with four papers addressing the Lastly, to add a flavour of implementation issues of dig- problem of separating components in astrophysical radiation ital signal processing systems for astrophysical tasks we in- maps which is a very hot problem due to the recent availabil- cluded a work by Belbachir et al. who describe a DSP sys- ity of WMAP satellite radiation maps and the future Planck tem for infrared astronomy that implements a combination satellite mission. The first paper on the problem is by Bedini of lossy and lossless compression. et al. who consider the separation of components which are We would like to thank all of the authors who contributed mutually dependent, therefore differing from classical ICA to this issue for their very interesting work. We are very grate- approaches. Therefore the paper also houses novelties from a ful to the referees who gave their time and energy in the re- signal processing point of view. The second paper in the se- view process and contributed immensely to the value of this quel, by Forni and Aghanim, considers a more specific prob- issue. Finally, we would like to thank the Editor-in-Chief Dr. lem, namely the separation of kinetic SZ signal from pri- Moonen and the EURASIP JASP staff for helping us at every mary CMB fluctuations. The next paper, by Lopez-Caniego´ stage of the creation of this issue. We do hope that the reader et al., provides instead a technique based on Bayesian detec- will find the issue useful, interesting, and inspirational. tion theory for the separation of point sources from the rest of the astrophysical radiation map. Moudden et al., in con- Ercan E. Kuruoglu trast to these works, consider separation in the transform do- Carlo Baccigalupi main. In particular, they propose a new separation method in the wavelet space. A related paper, by Starck et al., again in the wavelet methodological frame follows: it analyses the spatial dis- Ercan E. Kuruoglu was born in Ankara, Turkey, in 1969. He obtained his B.S. and tribution of the galaxies by multiscale methods. Jin et al. ff M.S. degrees both in electrical and elec- also concentrate their e orts on the analysis of the statisti- tronics engineering from Bilkent University cal distribution of the CMB utilising wavelet transform, with in 1991 and 1993, respectively. He com- the particular aim of detecting non-Gaussianity. Continuing pleted his graduate studies with M.Phil. and with transform domain analysis techniques, closely related to Ph.D. degrees in information engineering waveletideas,ThiebautandRoquespresenttime-frequency at the Cambridge University, in the Signal analysis techniques applied to irregularly sampled astronom- Processing Laboratory, in 1995 and 1998, ical time series. respectively. Upon graduation from Cam- Astrophysical images are obtained by mechanisms far bridge, he joined the Xerox Research Center in Cambridge as a per- from perfect and the images are corrupted with noise and manent member of the Collaborative Multimedia Systems Group. are blurred by the cameras’ limited resolution and may be After two years in Xerox, he won an ERCIM fellowship which he spent in INRIA-Sophia Antipolis, France, and IEI CNR, Pisa, distorted by nonlinearities in the cameras. Lanteri´ and Theys Italy. In January 2002, he joined ISTI-CNR, Pisa, as a permanent provide a technique for the restoration of the images in the member. His research interests are in statistical signal processing case of Poisson data and additive Gaussian noise. and information and coding theory with applications in image Recent satellite missions have provided us with vast processing, astronomy, telecommunications, intelligent user inter- amounts of data. Therefore, it is of paramount importance faces, and bioinformatics. He is currently on the Editorial Board to build efficient indexing methods and equally design data of Digital Signal Processing and an Associate Editor for the IEEE mining methods to recover information from big databases. Transactions on Signal Processing. He was the Guest Editor for a To this end, Frailis et al. propose a multidimensional index- special issue of the Signal Processing Journal on “Signal Process- ing method. Again due to the vast amount of data available, ing with Heavy-tailed Distributions,” December 2002. He is the it is important to be able to store them in an economic way. Special Sessions Chair for EURASIP European Signal Processing Conference, EUSIPCO 2005, and is the Technical Chair for EU- Lastri et al. suggest a new compression technique which is SIPCO 2006. He is also a Member of the IEEE Technical Com- virtually lossless. mittee on Signal Processing Theory and Methods. He has more Bratsolis tackles the problem of classification of astro- than 50 publications and holds 5 US, European, and Japanese physical images and his paper gives a flavour of the ap- patents. plication of neural-network-based techniques to the prob- lem. Flitti et al. deal with the problem of 3D segmentation Carlo Baccigalupi is currently an Assistant and present a new technique which they apply on real data. Professor at SISSA/ISAS. He is a member Wilson et al. present the adaptive DFT-based interferometer of the Planck and EBEx cosmic microwave tracking algorithm they have designed. background (CMB) polarization experi- While most of the works presented here are on cosmo- ment. In Planck, he is leading the work- logical problems of dimensions, surely solar science, in which ing group on component separation, and in EBEx he is responsible for the control of the important developments take place, is also in the interest area foreground polarized contamination to the of the issue. In particular, Zharkov et al. present a detailed CMB radiation. He is the author of about 40 system for the automated recognition of sunspots. They use papers on refereed international scientific a combination of elaborate morphological operators which review, on topics ranging from the theory of gravity to CMB data make the paper interesting also from an image processing analysis. He is teaching linear cosmological perturbations and CMB point of view. anisotropies courses for the Astroparticle Ph.D. course at SISSA. Editorial 2399
He is involved in long-term international projects. The most im- portant ones are the Long-term Space Astrophysics funded by NASA for the duration of five years, on component separation on COBE, WMAP, and future CMB experiments, and a one-year mer- cator professorship to be carried out in the University of Heidelberg in the academic year 2005/2006. EURASIP Journal on Applied Signal Processing 2005:15, 2400–2412 c 2005 Hindawi Publishing Corporation
Separation of Correlated Astrophysical Sources Using Multiple-Lag Data Covariance Matrices
L. Bedini Istituto di Scienza e Tecnologie dell’Informazione, CNR, Area della Ricerca di Pisa, via G. Moruzzi 1, 56124 Pisa, Italy Email: [email protected]
D. Herranz Istituto di Scienza e Tecnologie dell’Informazione, CNR, Area della Ricerca di Pisa, via G. Moruzzi 1, 56124 Pisa, Italy Email: [email protected]
E. Salerno Istituto di Scienza e Tecnologie dell’Informazione, CNR, Area della Ricerca di Pisa, via G. Moruzzi 1, 56124 Pisa, Italy Email: [email protected]
C. Baccigalupi International School for Advanced Studies, via Beirut 4, 34014 Trieste, Italy Email: [email protected]
E. E. Kuruoglu˘ Istituto di Scienza e Tecnologie dell’Informazione, CNR, Area della Ricerca di Pisa, via G. Moruzzi 1, 56124 Pisa, Italy Email: [email protected]
A. Tonazzini Istituto di Scienza e Tecnologie dell’Informazione, CNR, Area della Ricerca di Pisa, via G. Moruzzi 1, 56124 Pisa, Italy Email: [email protected]
Received 8 June 2004; Revised 18 October 2004
This paper proposes a new strategy to separate astrophysical sources that are mutually correlated. This strategy is based on second- order statistics and exploits prior information about the possible structure of the mixing matrix. Unlike ICA blind separation ap- proaches, where the sources are assumed mutually independent and no prior knowledge is assumed about the mixing matrix, our strategy allows the independence assumption to be relaxed and performs the separation of even significantly correlated sources. Besides the mixing matrix, our strategy is also capable to evaluate the source covariance functions at several lags. Moreover, once the mixing parameters have been identified, a simple deconvolution can be used to estimate the probability density functions of the source processes. To benchmark our algorithm, we used a database that simulates the one expected from the instruments that will operate onboard ESA’s Planck Surveyor Satellite to measure the CMB anisotropies all over the celestial sphere. Keywords and phrases: statistical, image processing, cosmic microwave background.
1. INTRODUCTION sion from galactic dust, the galactic synchrotron and free- free emissions. If one is only interested in estimating the Separating the individual radiations from the measured sig- CMB anisotropies, the interfering signals can just be treated nals is a common problem in astrophysical data analysis [1]. as noise, and reduced by suitable cancellation procedures. As an example, in cosmic microwave background anisotropy However, the foregrounds have an interest of their own, and surveys, the cosmological signal is normally combined with it could be useful to extract all of them from multichannel foreground radiations from both extragalactic and galactic data, by exploiting their different emission spectra. sources, such as the Sunyaev-Zeldovich effects from clusters Some authors [2, 3]havetriedtoextractanumberof of galaxies, the effect of the individual galaxies, the emis- individual radiation data from measurements on different Separation of Correlated Astrophysical Sources 2401 frequency channels, assuming that the physical mixture sumption, and to pursue identification by optimisation of model is perfectly known. Unfortunately, such an assump- a suitable function. A further advantage of this strategy is tion is rather unrealistic and could overconstrain the prob- that the relevant correlation coefficients between pairs of lem, thus leading to unphysical solutions. Attempts have sources can also be estimated. In our particular case, more- been made to avoid this shortcoming by introducing crite- over, being able to parametrise the mixing matrix allows us ria to evaluate a posteriori the closeness to reality of the mix- to substantially reduce the number of unknowns. This per- ture model and allowing individual sources to be split into mits to improve the performance of our technique. We will separate templates to take spatial parameter variability into show that a very fast model learning algorithm can be de- account [4, 5]. vised by matching the theoretical and the observed covari- A class of techniques capable of estimating the source sig- ance matrices, even if all the cross-covariances are nonnegli- nals as well as identifying the mixture model has recently gible. been proposed in astrophysics [6, 7, 8, 9]. In digital signal The paper is organised as follows. In Section 2,wefor- processing, these techniques are referred to as blind source malise the problem and introduce the relevant notation. separation (BSS) and rely on statistical assumptions on the In Section 3, we describe how the mixing matrix can be source signals. In particular, mutual independence and non- parametrised in our case. In Sections 4 and 5, we describe the Gaussianity of the source processes are often required [10]. methods we used to learn the mixing model and to estimate This totally blind approach, denoted as independent com- the original sources, respectively. In Section 6,wepresent ponent analysis (ICA), has already given promising results, some experimental results, with both stationary and nonsta- proving to be a valid alternative to assuming a known data tionary noises. In the final section, we give some remarks and model. On the other hand, most ICA algorithms do not per- future directions. mit to introduce prior information. Since all available in- formation should always be used, semiblind techniques are being studied to make astrophysical source separation more 2. PROBLEM STATEMENT flexible with respect to the specific knowledge often available As usual [2, 6],weassumethateachradiationprocess in this type of problem [11]. Moreover, the independence as- s˜c(ξ, η, ν) from the microwave sky has a spatial pattern sumption is not always justified; if there is evidence of cor- sc(ξ, η) that is independent of its frequency spectrum Fc(ν): relation between pairs of sources, it should be made possible to take this information into account, thus abandoning the strict ICA approach. s˜c(ξ, η, ν) = sc(ξ, η)Fc(ν). (1) The first blind technique proposed to solve the separation problem in astrophysics [6] was based on ICA, and allowed Here, ξ and η are angular coordinates on the celestial sphere, simultaneous model identification and signal estimation to and ν is frequency. The total radiation observed in a certain be performed. The independence requirement was fulfilled direction at a certain frequency is given by the sum of a num- by taking the statistics of all orders into account, as in all ICA ber N of signals (processes, or components) of the type (1), methods presented in the literature (see, e.g., [10, 12, 13]). where subscript c has the meaning of a process index. As- The problem of estimating all the model parameters and suming that the effects of the telescope beam on the angu- source signals cannot be solved by just using second-order lar resolution at different measurement channels have been statistics, since these are only able to enforce uncorrelation. equalised (see [16]), the observed signal at M different fre- However, this has been done in special cases, where ad- quencies can be modelled as ditional hypotheses on the spatial correlations or, equiva- lently, on the spectra of the individual signals are assumed [9, 14, 15]. As will be clear in the following, within the x(ξ, η) = As(ξ, η)+n(ξ, η), (2) framework of any noisy linear mixture model, the data co- variance matrix at a particular lag is related to the source where x ={xd, d = 1, ..., M} is the M-vector of the observa- covariance matrix at the same lag, the mixing matrix, and tions, d being a channel index, A is an M × N mixing matrix, the noise covariance matrix. If there is a sufficient num- s ={sc, c = 1, ..., N} is the N-vector of the individual source ber of lags for which the source covariance matrices are processes, and n ={nd, d = 1, ..., M} is the M-vector of in- not null, then it is possible to identify the model parame- strumental noise. The elements of A are related to the source ters by estimating the data covariance matrices from the ob- spectra and to the frequency responses through the following served data. Indeed, if we know the noise covariance matrix, formula: we are able to write a number of relationships from which the unknown parameters can be estimated. This is what is done by the second-order blind identification (SOBI) algo- adc = Fc(ν)bd(ν)dν,(3) rithmpresentedin[15]. SOBI, however, relies on joint di- agonalization of covariance matrices at different lags, which is only applicable in the case of uncorrelated source signals. where bd(ν) is the instrumental frequency response in the dth In our approach, we assumed that the mixing matrix can be measurement channel, which is normally known very well. If parametrised. This allows us to relax the independence as- we assume that the source spectra are constant within the 2402 EURASIP Journal on Applied Signal Processing passbands of the different channels, (3)canberewrittenas In the present approach, we found experimentally that, if a noise covariance map is known, even nonstationary noise can be treated. adc = Fc νd bd(ν)dν. (4)
Frequency-dependent telescope beams The element adc is thus proportional to the spectrum of the cth source at the center frequency νd of the dth channel. The model assumed in (2) is valid if the telescope radiation The separation problem consists in estimating the source patterns are the same in all the frequency channels. As the vector s from the observed vector x. Several estimation al- beams are frequency-dependent, a way to tackle the problem gorithms have been derived assuming a perfect knowledge is to preprocess the observed data in order to equalise the res- of the mixing matrix. As already said, however, this ma- olution on all the measurement channels, as in [16]. This also trix is related to both the instrumental frequency responses, changes the autocorrelation function of each noise process, ff which are known, and the emission spectra Fc(ν), which are but in a way that can be exactly evaluated. A di erent way to normally unknown. For this reason, relying on an assumed tackle the problem has been to approach it in the frequency mutual independence of the source processes sc(ξ, η), some domain [2, 9]. Also in these cases, the validity of the solution blind separation algorithms have been proposed [6, 7, 17], relies on a number of simplifiying assumptions, such as the which are able to estimate both the mixing matrix and the perfect circular symmetry of the telescope beams. Moreover, source vector. Assuming that the source signals are mutually the actual capability of extrapolating the spectrum at spatial independent, the MN mixing coefficients can be estimated frequencies where reduced information is available has still to by finding a linear mixture that, when applied to the data be assessed, especially in the cases where the signal-to-noise vector, nullifies the cross-cumulants of all orders. If, how- ratio is particularly low. ever, some prior information allows us to reduce the num- ber of unknowns, the identification problem can be solved Structure of the source covariance matrices by only using second-order statistics. This is the case with In the Planck experiment, the sources of interest are the our approach, which is based on a parametrisation of ma- CMB signal and the foregrounds. While no correlation is ex- trix A.Thisapproach,describedinSection 4, does not need a pected between the CMB signal and foregrounds, some sta- strict mutual independence assumption. Logically, any blind tistical dependence between pairs of foregrounds has to be separation algorithm is divided into two phases: using the taken into account. The off-diagonal entries of the source co- notation introduced here, the estimation of A will be re- variance matrices related to pairs of correlated sources will ferred to as system identification (or model learning), and the thus be nonzero, whereas all the remaining off-diagonal ele- estimation of s will be referred to as source separation.In ments will be zero. When it is known that some of the cross- this paper, we first address aspects related to learning, and covariances are close to zero, these can be kept fixed at zero, then give some details on source separation strategies de- thus further reducing the total number of unknowns. For in- rived from standard reconstruction procedures. Before de- stance, in a 3 × 3 case, if we assume the following structure scribing our algorithm in detail, we recall here some applica- for the source covariance matrix at zero shift: bility issues. σ 00 Source and noise processes 11 Cs(0, 0) = 0 σ22 σ23 ,(5) To estimate the covariance matrices from the available data, 0 σ32 σ33 the source and the noise processes must necessarily be as- sumed stationary. While CMB satisfies this assumption, the foregrounds are not stationary all over the celestial sphere. this means that we assume zero or negligible correlations be- This assumption can be made for small sky patches. How- tween sources 1 and 2 and sources 1 and 3, and the remaining cross-covariance σ23 = σ32 betweensources2and3isanun- ever, depending on the particular sky scanning strategy, noise σ is normally nonstationary, even within small patches, and known of the problem, along with the autocovariances ii. can also be autocorrelated. The noise covariance function Note that, for the typical scaling ambiguity of the blind iden- tification problem, the absolute values of both the diagonal should be known for any shift and for any angular coordi- ff τ ψ nate in the celestial sphere. Provided that the noise nonsta- and o -diagonal elements of matrices Cs( , )havenophys- tionarity and cross-correlation between sources can be ne- ical significance, while, by calculating ratios of the type glected, various methods are available, both in space and fre- 2 quency domains, to estimate samples of the noise covari- σij ance function or, equivalently, of noise spectrum [9]. Tack- ,(6) σiiσjj ling the space-variant nature of the noise process is difficult, and no simple method has been proposed so far to this pur- pose. In [11] the noise variance at each pixel is assumed to be we can actually estimate the correlation coefficients between known and a method is proposed to estimate the mixing ma- different sources, whatever the values of the individual co- trix and the probability density function of each component. variances. Separation of Correlated Astrophysical Sources 2403
3. PARAMETRISATION OF THE MIXING MATRIX 4. A SECOND-ORDER IDENTIFICATION ALGORITHM While in a general source separation problem the elements Let us consider the source and noise signals in (2) as realisa- adc are totally unknown, in our case we have some knowl- tions of two stationary vector random processes. The covari- edge about them. In fact, the integral in (4) is related to ance matrices of these processes are, respectively, known instrumental features and to the emission spectra of the single source processes, on which we do have some T Cs(τ, ψ) = s(ξ, η) − µs s(ξ + τ, η + ψ) − µs , knowledge. As an example, if the observations are made in (10) T the microwave and millimeter-wave range, the dominant ra- Cn(τ, ψ) = n(ξ, η) − µn n(ξ + τ, η + ψ) − µn , diations are the cosmic microwave background, the galactic dust, the free-free emission and the synchrotron (see [18]). where · denotes expectation under the appropriate joint Another significant signal comes from the extragalactic point probability, µs and µn are the mean vectors of processes s and sources. It is not possible to treat the point sources as a sin- n, respectively, and the superscript T means transposition. gle signal to be separated from the others on the basis of its As usual, the noise process is assumed signal-independent, emission spectrum, since each source has its own spectrum. white, and zero-mean, with known variances. Thus, for both Since the brightest point sources are the ones that affect more τ and ψ equal to zero, Cn is a known diagonal matrix whose strongly the study of the CMB [19], the usual approach is to elements are the noise variances in all the measurement remove them from the data before separating the other fore- channels, whereas for any τ or ψ different from zero Cn is grounds. Bright resolved point sources can be removed by us- the null M × M matrix. ing some of the specific techniques proposed in the literature As already proved [15, 22], covariance matrices, that [19, 20, 21]. Faint unresolved point sources are usually con- is, second-order statistics, permit blind separation to be sidered as an additional noise term in (2)(referredtoas“con- achieved when the sources show a spatial structure, namely, fusion noise” in the radio astronomy literature). For simplic- when they are spatially correlated. Thus, the mutual indepen- ity, we will not consider extragalactic point sources in our test dence requirement of ICA can be replaced by an equivalent examples. Moreover, although other sources (such as SZ and requirement on the spatial structure of the signal, and the free-free) could be taken into account, in our experiments identifiability of the system is assured. In other words, find- we only considered the synchrotron and dust foregrounds, ing matrices A and Cs is generally not possible from covari- which are the most significant in the Planck frequency range. ances at zero shift alone; to identify the mixing operator, ei- The emission spectrum of the cosmic microwave back- ther higher-order statistics or the covariance matrices at sev- ground is perfectly known, being a blackbody radiation. In eral nonzero shift pairs (τ, ψ) must be taken into account. Of terms of antenna temperature, it is course, this is also a requirement on the sources, since if the τ ψ ν˜2 exp(ν˜) covariance matrices are null for any pair ( , ), identification Fcmb(ν) = ,(7)is not possible. This aspect will become clearer below. ν − 2 exp(˜) 1 Let us now see our approach to system identification. By where ν˜ is the frequency in GHz divided by 56.8. From (4) exploiting (2), the covariance of the observed data can be and (7), the column of A related to the CMB radiation is thus written as known up to an unessential scale factor. For the synchrotron T radiation, we have Cx(τ, ψ) = x(ξ, η) − µx x(ξ + τ, η + ψ) − µx (11) T −ns = s τ ψ n τ ψ Fsyn(ν) ∝ ν . (8) AC ( , )A + C ( , ),
Thus, the column of A related to synchrotron only depends where Cx(τ, ψ)can be estimated from on a scale factor and the spectral index ns. For the thermal galactic dust, we have 1 T Cx(τ, ψ) = x(ξ, η) − µx x(ξ + τ, η + ψ) − µx , Np ν¯m+1 ξ,η F (ν) ∝ ,(9) dust exp(ν¯) − 1 (12) where ν¯ = hν/kTdust, h is the Planck constant, k is the Boltz- where Np is the number of pixels. Equation (11)providesa mann constant, and Tdust is the physical dust temperature. If number of independent nonlinear relationships that can be we assume a uniform temperature value, the frequency law used to estimate both A and Cs. Obviously, this possibility (9), that is, the column of A related to dust emission, only does not rely on mutual independence between the source depends on a scale factor and the parameter m. signals, as required by the ICA approach: the only require- The above properties enable us to describe the mixing ment is having a sufficient number of nonzero covariance matrix by means of just a few parameters. As an example, if matrices. In other words, spatial structure can be used in the we assume to have a perfectly known source spectrum (such place of mutual independence as a basis for model learning as the one of CMB) and N − 1 sources with one-parameter and signal separation. As assumed in the previous section, spectra, the number of unknowns in the identification prob- in this particular application the number of unknowns is re- lem is N − 1 instead of NM. duced by parametrising the mixing matrix. This allows us to 2404 EURASIP Journal on Applied Signal Processing solve the identification problem from the relationships made lated, the unknowns to be determined are 4 + 5 + Ns · 6, by available by (11) by only using the zero-shift covariance ma- using a maximum of M(M +1)/2+Ns · M2 equations. This trix, even if some of the sources are cross-correlated. We in- means that in this case, as soon as M = 4, the number of in- vestigated this possibility in [23]. In a general case, matrices dependent equations is larger than the number of unknowns A and Cs(τ, ψ)canbeestimatedfrom even for Ns = 0. Γ, Σ(:, :) = arg min A(Γ)Cs Σ(τ, ψ) AT(Γ) τ,ψ 5. SIGNAL SEPARATION STRATEGY (13) − Cx(τ, ψ) − Cn(τ, ψ) . Model learning is only the first step in solving the problem of source separation. Although, in principle, one could sim- The minimisation is performed over vectors Γ and Σ(:, :), ply use multichannel inverse filtering to recover the source where Γ is the vector of all the parameters defining A (pos- maps, this approach is not feasible in practice, for the pres- sibly consisting of all the matrix elements), and Σ(:, :) is the ence of noise. In our treatment, the data are assumed to be an ergodic process, in order to be able to evaluate its statistics vector containing all the unknown elements of matrices Cs for every shift pair. The matrix norm adopted is the Frobe- from the available sample. This entails a space invariant noise nius norm. Our present strategy to find the minimiser in (13) process. The estimation of the individual source maps should is to perform a stochastic minimisation in Γ, considering that be made on the basis of all the products of the learning stage. In our case, we have estimates of the mixing matrix and of Cs(Σ(τ, ψ)), for each (τ, ψ), can be calculated exactly once A(Γ) is fixed. A more accurate minimisation strategy is now the source covariance matrices at several shift pairs. In the being studied. hypothesis of stationary noise, we could exploit this infor- From the above scheme, it is clear that for each indepen- mation to implement a multichannel Wiener filter for source reconstruction. If the noise is not stationary, a generalized dent element of the matrices Cx(τ, ψ)wehaveanindepen- dent equation for the estimation of vector Γ and of all the vec- Kalman filter should be used. Our point here is on model learning, and thus we do not address the separation issues in tors Σ(τ, ψ). Since for (τ, ψ) = (0, 0) matrix Cx is symmetric, for zero shift we have M(M +1)/2 independent equations. detail. We only observe that a possible Bayesian separation scheme would make use of the source probability densities, For any other shift pair, Cx is a general matrix and thus, pro- videdthatitisnotzero,wehaveM2 additional independent and these can be estimated from our mixing matrix. Indeed, let us assume that our learning procedure has given a good equations. If Ns is the total number of nonzero shift pairs generating nonzero data covariance matrices, we thus have a estimate of A.LetB be its Moore-Penrose generalised inverse. In our case, we have M ≥ N, thus, as is known, total number of M(M+1)/2+Ns ·M2 = M[(2Ns +1)M+1]/2 independent equations. The number of unknowns is at most = T −1 T. NM+N(N +1)/2+Ns ·N2, in the case where all the elements B A A A (14) of A are unknown and all the source covariance matrices are full, that is, all the sources at any shift are correlated to each From (2), we have other. Note that, in this worst case situation, if it is M = N, N2 we always have more unknowns than equations, indepen- Bx = s + Bn. (15) dently of Ns.AssoonaswehaveM>N, there are always a number of nonzero shift pairs for which we have more in- LetusdenoteeachoftheN rows of B as an M-vector bi, dependent equations than unknowns to be estimated. This i = 1, ..., N, and consider the generic element yi of the N- observation gives an idea of the amount of information we vector Bx, have available for our estimation problem. The number of independent equations affects the behaviour of the nonlin- y = T · = s T · = s n . ear optimization landscape in (13). Qualitatively, we can af- i : bi x i + bi n : i + ti (16) firm that the more independent equations we have, the more well-posed the optimization problem will be. In particular, The probability density function of yi, p(yi), can be esti- it is likely that, in absence of any prior information about mated from bi and the data record x(ξ, η), while the prob- n p n the structure of A and Cs(τ, ψ), having a number of observed ability density function of ti , ( ti ), is a Gaussian, whose channels equal to the number of sources always leads to in- parameters can be easily derived from Cn and bi. The pdf of y p s p n sufficient information, independently of the number of shift i is the convolution between ( i)and ( ti ): pairs chosen. If, instead, the number of the available obser- p y = p s ∗ p n . vations is larger than the number of sources, the possibility i i ti (17) of estimating the unknowns relies on the number of shift pairs for which the data covariance matrices are nonzero. The From this relationship, p(si)canbeestimatedbydeconvo- availability of prior information, as in the application con- lution. As is well known, deconvolution is normally an ill- sidered here, can of course alleviate these requirements. For posed problem and, as such, it lacks a stable solution. In our example, if we have a 4 × 4 mixing matrix only depending case, we can regularise it by enforcing smoothness, positivity, on four parameters and only two sources significantly corre- and the normalisation condition for pdfs. Separation of Correlated Astrophysical Sources 2405
Any Bayesian estimation approach should exploit the The results from learning are the mixing matrix and the knowledge of the source densities to regularise the solution, source covariance matrices at the shift pairs chosen. From but these are normally unknown. In the case examined here, the estimate of the mixing matrix, it is also possible to de- the source distributions can be efficiently estimated as sum- rive the marginal source densities, by using relationships (16) marised above, and the computational cost of otherwise ex- and (17). As already mentioned, the estimates of the mix- pensive Bayesian algorithms can be reduced. As an exam- ing matrix and of the source covariance matrices are very ple, in [11], the source densities are modelled as mixtures robust against noise. Conversely, the estimates of the source of Gaussians, and the related parameters are estimated by distributions by means of (16)and(17) are more sensitive an independent factor analysis approach (see [24, 25]). The to noise. To obtain satisfactory results, it is necessary to rely method we propose here could well be used to fix the source on regularization methods; the choice of regularization pa- densities, thus reducing the overall cost of the identification- rameters, however, is known to be critical. In our case, we separation task. selected them empirically, by checking the smoothness of the From (15), it can be seen that the generalised inverse so- solutions. lution is already an estimate of the sources, since it is com- Our separation results are all derived from the applica- posed of the original source vectors corrupted by amplified tion of the Moore-Penrose pseudoinverse of the estimated noise. Thus, a simple source estimation strategy could be first mixing matrix, followed by a classical Wiener filtering on to apply (15) and then to reduce the influence of noise by fil- each output image. From this processing, estimates of the tering the result. In next section, we show some experimental source maps are obtained. Also, estimated source power results obtained by pseudoinversion of the estimated mix- spectra can be obtained from either the maps or the source ing matrix, followed by Wiener filtering of each individual autocorrelation matrices. In particular, the results we show source. This strategy would be strictly valid with stationary here are derived from the unfiltered pseudoinverse solutions, noise and high signal-to-noise ratio, however, interesting re- showing that, although the reconstructed images are heavily sults have been found even with strong nonstationary noise. affected by noise, the derived power spectra can be corrected Multichannel Wiener filtering for stationary noise and an ex- for the theoretical noise spectrum and thus estimated quite tended Kalman filter for the nonstationary case are now be- accurately. ing developed. The results presented here will all be related to a single data record, derived from a simulated 15◦ × 15◦ sky patch centered at 40◦ galactic longitude and 0◦ galactic latitude. 6. EXPERIMENTAL RESULTS It is to be noted that in such a patch, located on the galac- In this section, we present some results from our extensive tic plane, the measured data will be affected by strong fore- experimentation with the method described above. Our data ground interference, thus making the problem very difficult were drawn from a data set that simulates the one expected to solve. Indeed, many separation approaches experimented from Planck (see the Planck homepage).1 The source maps so far simply fail in proximity of the galactic plane, and they we considered were the CMB anisotropy, the galactic syn- are normally applied after masking the all-sky data in the chrotron, and thermal dust emissions over the four mea- high-interference regions. Here, the dust emission is stronger surement channels centred at 30 GHz, 44 GHz, 70 GHz, and than CMB, and separation is strictly necessary if CMB is to be 100 GHz. The test data maps have been generated by ex- distinguished from the foregrounds. Our method performed tracting several sky patches at different galactic coordinates very well with these data, and all the relevant parameters were from the simulated database, scaling them exactly accord- satisfactorily estimated even with the strongest noise com- ing to formulas (7), (8), and (9), generating the mixtures ponents. The noise standard deviation we adopted in the for the channels chosen, and adding realisations of Gaussian, case shown here is 30% the standard deviation of CMB at signal-independent, white noise. Several noise levels have 100 GHz. The noise level in the other channels has been sim- been used, from a ten percent to more than one hundred per- ply obtained by scaling the level at 100 GHz in accordance cent of the CMB standard deviation. The range chosen con- with the expected Planck sensitivity at those frequencies. For tains noise levels within the Planck specifications. Although each patch considered, we tried different noise levels, up to our method would be only suited for uniform noise, we also more than 100% of the CMB level at 100 GHz, and for each tried to apply it to data corrupted by nonuniform noise, and noise level, we performed a Monte Carlo simulation with obtained promising results. hundreds of different noise realizations. This analysis is not Within this section, we will divide the results obtained in reported in detail here, but we can say that no significant bias model learning from the results in separation, and the cases has been found in the results. with stationary noise from those with nonstationary noise. It is to remark that, at high galactic latitudes, the CMB ra- In these latter cases, knowledge of a noise variance map is diation is dominant at our frequencies, and the foregrounds assumed, and the additional problem arises of choosing the are well below the noise level assumed in our experiments. appropriate noise covariance matrix. Thus, the CMB is almost the only measured radiation, and is estimated very well with all the assigned signal-to-noise ratios. Conversely, as expected with these noise levels, the 1http://astro.estec.esa.nl/SA-general/Projects/Planck/. foregrounds cannot be estimated correctly. Assuming much 2406 EURASIP Journal on Applied Signal Processing
(a) (b) (c)
Figure 1: Source maps from a 15◦ × 15◦ patchcenteredat0◦ galactic latitude and 40◦ galactic longitude, at 100 GHz: (a) CMB; (b) synchrotron; (c) thermal dust. lower noise levels, our method, as other techniques such as in Figure 3. The typical elapsed times per run were a few min- ICA (see [6]), allows the foregrounds to be estimated satis- utes on a 2 GHz CPU computer, with a Matlab interpreted factorily. code. In the case described here, we estimated ns = 2.8985 In Figure 1, we show the three source maps we used in the and m = 1.7957, corresponding to the mixing matrix situation described above. In this figure and in all the others shown here, the grayscale is linear with black correspond- 111 ing to the maximum image value. We assigned the sources 1.1353 2.8118 0.5494 A = . (20) s1 to CMB, s2 to synchrotron, and s3 to dust, and the sig- 1.2241 10.8009 0.2473 nals x1, x2, x3,andx4 to the measurement channels at 100, 1.2570 32.7775 0.1267 70, 44, and 30 GHz, respectively. Therefore, the first, second, and third columns of the mixing matrix will be related to As a quality index for our estimation, we adopted the matrix T −1 −1 T −1 CMB, synchrotron, and dust, respectively, and the first, sec- Q = (A Cn A) (A Cn Ao), which, in the ideal case, should ond, third, and fourth rows of the mixing matrix will be re- be the N × N identity matrix I. In the present case, we have lated to the 100 GHz, 70 GHz, 44 GHz, and 30 GHz channels, respectively. The mixing matrix, Ao,hasbeenderivedfrom . − . − . 1 0000 0 0074 0 0013 (7), (8), and (9) with spectral indices ns = 2.9andm = 1.8 Q = 0.0000 1.0020 0.0000 . (21) (see, e.g., [26, 27]): 0.0000 0.0054 1.0013 111 − The Frobenius norm of matrix Q I should be zero in the 1.1353 2.8133 0.5485 case of perfect model learning. In this case, it is 0.0096. Ao = . (18) 1.2241 10.8140 0.2464 These results have been found by considering 25 uni- 1.2570 32.8359 0.1260 formly distributed shift pairs, with 0 ≤ τ ≤ 20 and 0 ≤ ψ ≤ 20. As a synthetic index for the quality of the reconstructed In Figure 2, we show the data maps for stationary noise. source covariance matrices, we adopted a matrix E,where Also, note that the case examined does not fit the ICA as- each element is the relative error in the same covariance ele- sumptions. For example, the normalized source covariance ment, averaged over all the pairs (τ, ψ): matrix at zero shift is τ ψ − τ ψ . . . 1 Csi,j ( , ) Csi,j ( , ) 1 0000 0 1961 0 0985 Ei,j = , (22) Ns +1 C i j (τ, ψ) Cs(0, 0) = 0.1961 1.0000 0.6495 , (19) τ,ψ s , 0.0985 0.6495 1.0000 where Cs are the estimated source covariance matrices. Of where a significant correlation, of the order of 65%, can be course, matrix (22) is only defined when all the denomina- observed between the dust and synchrotron maps. tors are nonzero. A more accurate analysis of the results can For the data described above, we ran our learning al- be made from the element-by-element comparison of the es- gorithm for 500 different noise realisations; for each run, timated and the original matrices, but we do not report these 10 000 iterations of the minimisation procedure described results here. For the case shown above, we have in the previous section were performed. The unknown pa- ns m . . . rameters were the spectral indices and , and all the el- 0 0274 0 0392 0 0496 ements of matrices Cs(τ, ψ).Thecostdefinedin(13), as a E = 0.0472 0.0170 0.0120 . (23) function of the iteration number in a particular run, is shown 0.0917 0.0125 0.0050 Separation of Correlated Astrophysical Sources 2407
(a) (b)
(c) (d)
Figure 2: Noisy data maps at (a) 100 GHz; (b) 70 GHz; (c) 44 GHz; (d) 30 GHz.
×10−3 uate more quantitatively the results of the whole learning- . 1 8 separation procedure, we compared the power spectrum of 1.6 the CMB map with the one of the reconstructed map. This comparison is shown in Figure 6, where we also show the 1.4 possibility of correcting the reconstructed spectrum for the . n 1 2 known theoretical spectrum of the noise component t1 ,ob- 1 tained as in (16). As can be seen, the reconstructed spectrum is very similar to the original within a multipole l = 2000. 0.8 Strictly speaking, our algorithm could not be applied 0.6 to nonstationary processes. However, let us assume that the original sources are stationary, and the noise is nonstationary . 0 4 but still spatially white and uncorrelated. This means that its 0.2 pixel-dependent covariance matrices, Rn(τ, ψ; ξ, η), are zero τ ψ 0 for any nonzero shift pair ( , ). We tried our method on 0 2000 4000 6000 8000 10000 nonstationary data, by assuming to know Rn(0, 0; ξ, η), and using a constant covariance matrix given by Figure 3: Norm of the residual in (13) as a function of the iteration number. 1 Cn(0, 0) = Rn(0, 0; ξ, η). (24) Np ξ,η
The reconstructed probability density functions of the The nonstationary data were obtained from a spatial tem- source processes, estimated from (16)and(17), are shown plate of noise standard deviations expected for typical Planck in Figure 4. observations, shown in Figure 7. The actual standard devi- We separated the sources by multiplying the data ma- ations were adjusted so as to obtain the average signal-to- trix by the Moore-Penrose generalised inverse, as in (15), and noise ratios desired for the different channels. The separa- then by applying a Wiener filter to the results thus obtained. tion results for a case where these SNRs were the same as in As already said, this is not the best choice reconstruction al- the above stationary case are shown in Figure 8, where the gorithm at all, especially when the data are particularly noisy degradation in the reconstruction is apparent in the regions and the noise is not stationary. However, the results we ob- where the noise is stronger. The results, in terms of recon- tained are visually very good, as shown in Figure 5.Toeval- tructed power spectra, are perfectly comparable to the ones 2408 EURASIP Journal on Applied Signal Processing
0.07 0.035
0.06 0.03
0.05 0.025
0.04 0.02
0.03 0.015
0.02 0.01
0.01 0.005
0 0 −1 −0.500.51 −0.100.10.20.3
Real source density functions Real source density functions Estimated source density functions Estimated source density functions
(a) (b)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0 −2 −10 1 2
Real source density functions Estimated source density functions
(c)
Figure 4: Real (dotted) and estimated (solid) source density functions for (a) CMB, (b) synchrotron, and (c) dust.
(a) (b) (c)
Figure 5: Wiener-filtered estimated maps: (a) CMB; (b) synchrotron; (c) dust. Separation of Correlated Astrophysical Sources 2409
0.1 0.05 0.09 0.04 0.08 0.07 0.03 0.06 ∗ l ∗ l . . C 0 02 C 0 05 0.04 0.01 0.03 0.02 0 0.01 0 −0.01 0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 2500 Multipole l Multipole l
Estimated CMB power spectra Estimated CMB power spectra Real CMB power spectra Real CMB power spectra Theoretical power spectrum
(a) (b)
Figure 6: (a) Real (dotted) and estimated (solid) CMB power spectra. The dashed line represents the theoretical power spectrum of the n noise component t1 in (16), evaluated from the noise covariance and the Moore-Penrose pseudoinverse of the estimated mixing matrix. (b) Real (dotted) and estimated (solid) CMB power spectra, corrected for theoretical noise.
1 The Frobenius norm of matrix Q − I is now 0.0736, that is, 0.9 slightly worse than for the above stationary case.
0.8 7. CONCLUDING REMARKS 0.7 By exploiting the spatial structure of the sources, we devel- 0.6 oped an identification and separation algorithm that is able 0.5 to exploit any available information on possible structure of the mixing matrix and the source covariance matrices. This . 0 4 can include the fully blind approach and the case exempli- 0.3 fied here, where the mixing matrix is known to only depend on two parameters. The identification task is performed by 0.2 a simple optimization strategy, while the proper separation can be faced by different approaches. We experimented the Figure 7: Map of noise standard deviations used to generate simplest one, but we are also developing more accurate tech- nonstationary data. niques, especially suited to treat nonstationary noise on the data. Our method is suitable to work directly with all-sky exemplified in Figure 6. The estimated spectral indices were maps, but it could be necessary to apply it to small patches, ns = 2.8885 and m = 1.7881, corresponding to the mixing as is shown in the above experimental section, to cope with matrix the expected variability of the spectral indices and the noise variances in different sky regions. 111 It has been observed that it does not make sense to try 1.1353 2.8018 0.5509 source separation in those regions where the foreground A = . (25) 1.2241 10.7128 0.2488 emissions are much smaller than CMB and well below the 1.2570 32.3861 0.1279 noise level. In any case, the CMB angular power spectrum has always been estimated fairly well up to a multipole l = Theaverageerroroncovariancematricesisinthiscase 2000, irrespective of the galactic latitude. The estimation of the source densities has also given good results. Source . . . separation by our method has been particularly interesting 0 0158 0 1165 0 1930 E = 0.1163 0.0331 0.0254 . (26) with data from low galactic latitudes, where the foreground 0.2440 0.0261 0.0144 variance is often higher than the one of the CMB signal. 2410 EURASIP Journal on Applied Signal Processing
(a) (b) (c) Figure 8: Wiener-filtered estimated maps from nonstationary data: (a) CMB; (b) synchrotron; (c) dust.
Note that many separation strategies, both blind and non- Franc¸ois Bouchet (IAP) for setting up and distributing the blind, have failed their goal in this region of the celestial database. Extensive use of the HEALPix scheme (Hierarchi- sphere. As an example, WMAP data analysis (see [28]) was cal, Equal Area and iso-Latitude Pixelisation of the sphere, often performed by using pixel intensity masks that exclude http://www.eso.org/science/healpix), by Krysztof M. Gorski´ the brightest sky portion from being considered. Another et al., has been made throughout this work. interesting feature of our method is that significant cross- correlations between pairs of foregrounds can be straigh- REFERENCES forwardly taken into account. Recently, some methods for a completely blind separation of correlated sources have been [1] M. Tegmark, D. J. Eisenstein, W. Hu, and A. de Oliveira-Costa, proposed in the literature (see, e.g., [29]). Their effective- “Foregrounds and forecasts for the cosmic microwave back- ground,” Astrophysical Journal, vol. 530, no. 1, pp. 133–165, ness in astrophysical map separation has not been proved yet. 2000. Moreover, they have a high computational complexity. [2] M. P. Hobson, A. W. Jones, A. N. Lasenby, and F. R. Bouchet, Recently [9], a frequency-domain implementation of the “Foreground separation methods for satellite observations of method in [15] has been proposed. This method allows to the cosmic microwave background,” Monthly Notices of the take antenna beam effects into account straightforwardly by Royal Astronomical Society, vol. 300, no. 1, pp. 1–29, 1998. including the effect of the antenna transfer functions in the [3] F. R. Bouchet, S. Prunet, and S. K. Sethi, “Multifrequency model. It also permits to introduce prior information about Wiener filtering of cosmic microwave background data with polarization,” Monthly Notices of the Royal Astronomical Soci- the entries of the mixing matrix and the spatial power spec- ety, vol. 302, no. 4, pp. 663–676, 1999. tra of the components. An open problem is the extension of [4]A.W.Jones,M.P.Hobson,P.Mukherjee,andA.N.Lasenby, these methods to the case of correlated sources. A possible ex- “The effect of a spatially varying Galactic spectral index on the tended method might be implemented in the space or in the maximum entropy reconstruction of Planck Surveyor satellite frequency domain according to convenience. Another prob- data,” Astrophysical Letters & Communications,vol.37,no.3- lem that is still open with the expected Planck data is the dif- 6, pp. 369–375, 2000. ferent resolution of the data maps in some of the measure- [5]R.B.Barreiro,M.P.Hobson,A.J.Banday,etal.,“Foreground separation using a flexible maximum-entropy algorithm: an ment channels. The identification part of our method can application to COBE data,” Monthly Notices of the Royal As- work with maps whose resolution has been degraded in or- tronomical Society, vol. 351, no. 2, pp. 515–540, 2004. der to be the same in all the channels. The result would be [6] C. Baccigalupi, L. Bedini, C. Burigana, et al., “Neural net- an estimate of the mixing matrix, which can be used in any works and the separation of cosmic microwave background nonblind separation approach with channel-dependent reso- and astrophysical signals in sky maps,” Monthly Notices of the lution, such as maximum entropy [2]. However, the possible Royal Astronomical Society, vol. 318, no. 3, pp. 769–780, 2000. [7] D. Maino, A. Farusi, C. Baccigalupi, et al., “All-sky astrophysi- asymmetry of the telescope beam patterns should be taken cal component separation with Fast Independent Component into account in verifying this possibility. Analysis (fastica),” Monthly Notices of the Royal Astronomical Society, vol. 334, no. 1, pp. 53–68, 2002. ACKNOWLEDGMENTS [8] C. Baccigalupi, F. Perrotta, G. De Zotti, et al., “Extracting cos- mic microwave background polarization from satellite astro- This work has been partially supported by the Italian Space physical maps,” Monthly Notices of the Royal Astronomical So- Agency, under Contract ASI/CNR 1R/073/01. D. Herranz ciety, vol. 354, no. 1, pp. 55–70, 2004. is supported by the European Community’s Human Po- [9] J. Delabrouille, J.-F. Cardoso, and G. Patanchon, “Multidetec- tential Programme under Contract HPRN-CT-2000-00124 tor multicomponent spectral matching and applications for CMBNET. The authors adopted the simulated sky tem- cosmic microwave background data analysis,” Monthly Notices of the Royal Astronomical Society, vol. 346, no. 4, pp. 1089– plates provided by the Planck Technical Working Group ff 1102, 2002. 2.1 (di use component separation). In particular, the au- [10] A. Hyvarinen¨ and E. Oja, “Independent component analysis: thors are grateful to Martin Reinecke (MPA), Vlad Stol- algorithms and applications,” Neural Networks, vol. 13, no. 4- yarov (Cambridge), Andrea Moneti, Simon Prunet, and 5, pp. 411–430, 2000. Separation of Correlated Astrophysical Sources 2411
[11] E. E. Kuruoglu,˘ L. Bedini, M. T. Paratore, E. Salerno, and A. [29] A. K. Barros, “Dependent component analysis,” in Advances in Tonazzini, “Source separation in astrophysical maps using in- Independent Component Analysis, M. Girolami, Ed., Springer, dependent factor analysis,” Neural Networks, vol. 16, no. 3-4, New York, NY, USA, July 2000. pp. 479–491, 2003. [12] P. Comon, “Independent component analysis, a new con- cept?” Signal Processing, vol. 36, no. 3, pp. 287–314, 1994. L. Bedini graduated cum laude in elec- [13] J.-F. Cardoso, “Blind signal separation: statistical principles,” tronic engineering from the University of Proc. IEEE, vol. 86, no. 10, pp. 2009–2025, 1998. Pisa, Italy, in 1968. Since 1970, he has been [14] L. Tong, R.-W. Liu, V. C. Soon, and Y.-F. Huang, “Indetermi- a researcher of the Italian National Re- nacy and identifiability of blind identification,” IEEE Trans. search Council, Istituto di Scienza e Tec- Circuits Syst., vol. 38, no. 5, pp. 499–509, 1991. nologie dell’Informazione, Pisa, Italy. His [15] A. Belouchrani, K. Abed-Meraim, J.-F. Cardoso, and E. interests have been in modelling, identifica- Moulines, “A blind source separation technique using second- tion, and parameter estimation of biologi- order statistics,” IEEE Trans. Signal Processing, vol. 45, no. 2, cal systems applied to noninvasive diagnos- pp. 434–444, 1997. tic techniques. At present, his research inter- [16] E. Salerno, C. Baccigalupi, L. Bedini, et al., “Independent est is in the field of digital signal processing, image reconstruction, component analysis approach to detect the cosmic microwave and neural networks applied to image processing. He is a coauthor background radiation from satellite measurements,” Tech. of more than 80 scientific papers. From 1971 to 1989, he was an Rep. B4-04, IEI-CNR, Pisa, Italy, 2000. Associate Professor of system theory at the Computer Science De- [17] G. Patanchon, H. Snoussi, J.-F. Cardoso, and J. Delabrouille, “Component separation for Cosmic Microwave Background partment, University of Pisa, Italy. data: a blind approach based on spectral diversity,” in Proc. D. Herranz received the B.S. degree in 1995 3rd Workshop on Physics in Signal and Image Processing (PSIP ’03), pp. 17–20, Grenoble, France, January 2003. and the M.S. degree in 1995 from the Uni- [18] G. De Zotti, L. Toffolatti, F. Argueso,¨ et al., “The planck versidad Complutense de Madrid, Madrid, surveyor mission: astrophysical prospects,” in 3K Cosmology, Spain, and the Ph.D. degree in astrophysics Proc. EC-TMR Conference, vol. 476, pp. 204–204, American from Universidad de Cantabria, Santander, Institute of Physics, Rome, Italy, October 1999. Spain, in 2002. He was a CMBNET Post- [19] P. Vielva, E. Mart´ınez-Gonzalez,´ L. Cayon,J.M.Diego,J.L.´ doctoral Fellow at the Istituto di Scienza Sanz, and L. Toffolatti, “Predicted Planck extragalactic point- e Tecnologie dell’Informazione “A. Faedo” source catalogue,” Monthly Notices of the Royal Astronomical (CNR), Pisa, Italy, from 2002 to 2004. He Society, vol. 326, no. 1, pp. 181–191, 2001. is currently at the Instituto de Fisica de [20] L. Tenorio, A. H. Jaffe, S. Hanany, and C. H. Lineweaver, “Ap- Cantabria, Santander, Spain, under an MEC Juan de la Cierva con- plications of wavelets to the analysis of cosmic microwave tract. His research interests are in the areas of cosmic microwave background maps,” Monthly Notices of the Royal Astronomi- background astronomy and extragalactic point source statistics cal Society, vol. 310, no. 3, pp. 823–834, 1999. as well as the application of statistical signal processing to astro- [21] L. Cayon,J.L.Sanz,R.B.Barreiro,etal.,“Isotropicwavelets:a´ nomical data, including blind source separation, linear and non- powerful tool to extract point sources from cosmic microwave linear data filtering, and statistical modeling of heavy-tailed pro- background maps,” Monthly Notices of the Royal Astronomical cesses. Society, vol. 315, no. 4, pp. 757–761, 2000. [22] A. K. Barros and A. Cichocki, “Extraction of specific signals E. Salerno graduated in electronic engineering from the Univer- with temporal structure,” Neural Computation,vol.13,no.9, sity of Pisa, Italy, in 1985. In September 1987, he joined the Ital- pp. 1995–2004, 2001. ian National Research Council (CNR) as a permanent researcher. [23] L. Bedini, S. Bottini, C. Baccigalupi, et al., “A semi-blind He is now with the Institute of Information Science and Technolo- approach for statistical source separation in astrophysical gies (ISTI), Signals and Images Laboratory, Pisa. His scientific in- maps,” Tech. Rep. ISTI-2003-TR-35, ISTI-CNR, Pisa, Italy, terests are in applied inverse problems, image reconstruction and 2003. restoration, nondestructive evaluation, and blind signal separation. [24] E. Moulines, J.-F. Cardoso, and E. Gassiat, “Maximum like- He has been assuming various responsibilities in research programs lihood for blind separation and deconvolution of noisy sig- in nondestructive testing, robotics, numerical models for image re- nals using mixture models,” in Proc. IEEE Int. Conf. Acoustics, construction and computer vision, and neural network techniques Speech, Signal Processing (ICASSP ’97), vol. 5, pp. 3617–3620, Munich, Germany, April 1997. in astrophysical imagery. Dr Salerno is an Associate Investigator [25] H. Attias, “Independent factor analysis,” Neural Computation, with the Planck-LFI Consortium, and a Member of the Italian vol. 11, no. 4, pp. 803–851, 1999. Society for Information and Communication Technology (AICT- [26] A. J. Banday and A. W. Wolfendale, “Galactic dust emission AEIT). and the cosmic microwave background,” Monthly Notices of the Royal Astronomical Society, vol. 252, pp. 462–472, 1991. C. Baccigalupi is currently an Assistant Pro- [27] G. Giardino, A. J. Banday, K. M. Gorski,´ K. Bennett, J. L. fessor at SISSA/ISAS. He is a member of the Jonas, and J. Tauber, “Towards a model of full-sky Galactic Planck and EBEx cosmic microwave back- synchrotron intensity and linear polarisation: a re-analysis of ground (CMB) polarization experiments. the Parkes data,” Astronomy & Astrophysics, vol. 387, no. 1, In Planck, he is leading the working group pp. 82–97, 2002. on component separation, and in EBEx, [28] C. L. Bennett, R. S. Hill, G. Hinshaw, et al., “First-year he is responsible for the control of the Wilkinson Microwave Anisotropy Probe (WMAP) observa- foreground polarized contamination to the tions: Foreground emission,” Astrophysical Journal Supple- CMB radiation. He is the author of about 40 ment Series, vol. 148, no. 1, pp. 97–117, 2003. papers on refereed international scientific 2412 EURASIP Journal on Applied Signal Processing reviews, on topics ranging from the theory of gravity to CMB data analysis. He is teaching linear cosmological perturbations and CMB anisotropies courses for the Astroparticle Ph.D. course at SISSA. He is involved in long-term international projects. The most important are the Long Term Space Astrophysics funded by NASA for the du- ration of five years, on component separation on COBE, WMAP, and future CMB experiments, and a one-year Mercator Professor- ship to be carried out in the University of Heidelberg in the aca- demic year 2005/2006.
E. E. Kuruoglu˘ was born in Ankara, Turkey, in 1969. He obtained his B.S. and M.S. de- grees both in electrical and electronics engi- neering from Bilkent University in 1991 and 1993, respectively. He completed his gradu- ate studies with M.Phil. and Ph.D. degrees in information engineering from the Cam- bridge University, in the Signal Processing Laboratory, in 1995 and 1998, respectively. During this period, he received the British Council Scholarship, Cambridge Overseas Trust Scholarship, and the Lundgren Award. Upon graduation from Cambridge, he joined the Xerox Research Center in Cambridge as a permanent member of the Collaborative Multimedia Systems Group. After two years in Xerox, he won an ERCIM Fellowship which he spent in INRIA- Sophia Antipolis, France, and IEI CNR, Pisa, Italy. In January 2002, he joined ISTI-CNR, Pisa, as a permanent member. His research interests are in statistical signal processing, human-computer in- teraction, and information and coding theory with applications in image processing, astronomy, telecommunications, intelligent user interfaces, and bioinformatics. He is currently in the Editorial Board of Digital Signal Processing and an Associate Editor for the IEEE Transactions on Signal Processing. He was the Guest Editor for a special issue on signal processing with heavy-tailed distribu- tions published in signal processing, December 2002. He is the Spe- cial Sessions Chair for EURASIP European Signal Processing Con- ference, EUSIPCO 2005, and is the Tutorials Chair for EUSIPCO 2006. In 2005, he has been elected to become a Member of the IEEE Technical Committee on Signal Processing Theory and Methods. He has more than 50 publications and holds 5 US, European, and Japanese patents.
A. Tonazzini graduated cum laude in math- ematics from the University of Pisa, Italy, in 1981. In 1984, she joined the Istituto di Scienza e Tecnologie dell’Informazione of the Italian National Research Council (CNR) in Pisa, where she is currently a re- searcher at the Signals and Images Labora- tory. She cooperated in special programs for basic and applied research on image pro- cessing and computer vision, and is a coau- thor of over 60 scientific papers. Her present interest is in inverse problems theory, image restoration and reconstruction, document analysis and recognition, independent component analysis, and neural networks and learning. EURASIP Journal on Applied Signal Processing 2005:15, 2413–2425 c 2005 Hindawi Publishing Corporation
Adapted Method for Separating Kinetic SZ Signal from Primary CMB Fluctuations
Olivier Forni IAS-CNRS, Universit´e Paris Sud, Batimentˆ 121, 91405 Orsay Cedex, France Email: [email protected]
Nabila Aghanim IAS-CNRS, Universit´e Paris Sud, Batimentˆ 121, 91405 Orsay Cedex, France Email: [email protected] Division of Theoretical Astronomy, National Astronomical Observatory of Japan, Osawa 2-21-1, Mitaka, Tokyo 181-8588, Japan
Received 30 May 2004; Revised 11 December 2004
In this first attempt to extract a map of the kinetic Sunyaev-Zel’dovich (KSZ) temperature fluctuations from the cosmic microwave background (CMB) anisotropies, we use a method which is based on simple and minimal assumptions. We first focus on the intrinsic limitations of the method due to the cosmological signal itself. We demonstrate using simulated maps that the KSZ reconstructed maps are in quite good agreement with the original input signal with a correlation coefficient between original and reconstructed maps of 0.78 on average, and an error on the standard deviation of the reconstructed KSZ map of only 5% on average. To achieve these results, our method is based on the fact that some first-step component separation provides us with (i) a map of Compton parameters for the thermal Sunyaev-Zel’dovich (TSZ) effect of galaxy clusters, and (ii) a map of temperature fluctuations which is the sum of primary CMB and KSZ signals. Our method takes benefit from the spatial correlation between KSZandTSZeffects which are both due to the same galaxy clusters. This correlation allows us to use the TSZ map as a spatial template in order to mask, in the CMB + KSZ map, the pixels where the clusters must have imprinted an SZ fluctuation. In practice, a series of TSZ thresholds is defined and for each threshold, we estimate the corresponding KSZ signal by interpolating the CMB fluctuations on the masked pixels. The series of estimated KSZ maps is finally used to reconstruct the KSZ map through the minimisation of a criterion taking into account two statistical properties of the KSZ signal (KSZ dominates over primary anisotropies at small scales, KSZ fluctuations are non-Gaussian distributed). We show that the results are quite sensitive to the effect of beam convolution, especially for large beams, and to the corruption by instrumental noise. Keywords and phrases: cosmic microwave background, data analysis.
1. INTRODUCTION DASI [4], and Archeops [5], which achieved a firm detection The cosmic microwave background (CMB) temperature of the so-called “first peak” in the CMB anisotropy angular anisotropies field encloses so-called primary anisotropies, di- power spectrum at the degree scale. This detection was re- rectly related to the initial density fluctuations at early stages cently confirmed by the WMAP satellite [6].Aseriesofsmall of the universe, and so-called secondary anisotropies gener- scale CMB experiments (e.g., VSA [7], CBI [8], ACBAR [9]) ated after matter and radiation decoupled. The secondary showed rather conclusive evidence for a second and a third anisotropies arise from the interaction of the CMB pho- peak. The positions, heights, and widths of these features in tons with gravitational potential wells or with ionised gas the angular power spectrum already give us a good idea of along their way towards us. More “local” contributions to the cosmological model. the CMB signal are due to foreground emissions from our It is clear however that such constraints necessitate the galaxy and from distant galaxies. One of the major goals of “cleanest” possible cosmological signal, or in other words observational cosmology is to use the CMB anisotropies to they need the best possible monitoring of contaminating sig- probe the cosmological model mainly through cosmological nals such as secondary anisotropies or foreground emissions. parameter estimation. This is already performed by a num- This is the objective of the component separation for CMB ob- ber of groups using ground-based and balloon-borne exper- servations. In most cases, the different contributing signals to iments such as TOCO [1], BOOMERanG [2], MAXIMA [3], CMB anisotropies exhibit both different frequency (ν)and 2414 EURASIP Journal on Applied Signal Processing spectral (in Fourier or spherical harmonic domains) depen- 8 dences. As a consequence CMB experiments often observe at several frequencies to be able to separate the different as- 6 trophysical signals. Numerous methods were adapted and 4 developed for the CMB problem like the Wiener filtering, maximum entropy, independent component analysis, and so 2 forth [10, 11, 12, 13, 14]. All gave very satisfactory results and showed clearly that one can extract the CMB primary signal 0
from the observed mixture. Obviously the success of the sep- I (in arbitrary units) −2 aration methods greatly depend on how different from each ∆ other the signals, in the observed mixture, are. Signals shar- −4 ing with the primary anisotropies the same frequency de- pendence and/or the same power spectrum will be badly de- −6 tected or even undetected by the separation methods men- 0 5 10 15 20 x tioned above. We present here a new method optimised to extract, from Figure 1: Frequency dependences of the intensity variations due to the primary signal, temperature fluctuations which have the the TSZ effect (solid line) and KSZ effect (dashed line) as a func- x = hν/kT h same frequency dependence and are associated with a ma- tion of the dimensionless frequency CMB ( is the Planck k T = . jor secondary effect, namely, the kinetic Sunyaev-Zel’dovich constant, is the Boltzmann constant, and CMB 2 728). K is the (KSZ) effect (for details see [15]). Our method is based on mean temperature of the CMB. a two-step strategy in which we derive the best estimate of the CMB signal on masked pixels by interpolation, and then We test and develop our method on a set of numerical we deduce the best estimate of the KSZ map by minimisa- simulations that will be described in Section 2.InSection 3, tion. We show that we retrieve the KSZ signal, in the best we detail the methodology adopted to separate KSZ from pri- possible way, in terms of its amplitude, its distribution and mary CMB signal. We focus on the way our method is intrin- its power spectrum provided we use (i) a well-chosen spa- sically limited by pure cosmological signals primary CMB tial template for the masked pixels, and (ii) adapted signal and SZ effect. We perform some sensitivity tests in Section 4 processing techniques for both interpolating and minimis- and explore the effects of beam convolution and instrumen- ing. For the first point, we use the TSZ maps as a template tal white noise on our results. Finally, we discuss our results since both TSZ and KSZ are associated with the same ob- in Section 5. jects (clusters of galaxies). For the second point, namely, the signal processing techniques, there are several issues to ad- dress in order to optimise the extraction of the KSZ signal 2. THE ASTROPHYSICAL SIGNALS fromtheprimaryCMB.Wehavethustomakesureateach Among all secondary anisotropies, the dominant contribu- step that we use adapted techniques. At the interpolation on tion to CMB signal comes from the Sunyaev-Zel’dovich (SZ) the masked pixels defined by the template, we can obviously effect [20, 21] which represents the inverse Compton scatter- use several methods like for example constrained 2D realisa- ing of CMB photons by free electrons in ionised and hot in- tions of the underlying CMB (sensitive to our knowledge of tracluster gas. This so-called thermal SZ (TSZ) effect, whose the CMB through the confidence intervals on the cosmologi- amplitude is given by the Compton parameter y, is the inte- cal parameters), textures [16] (sensitive to the morphological gral of the pressure along the line of sight. The inverse Comp- description of the processes), or simply cubic B-spline meth- ton effect moves the CMB photons from the lower to the ods. The latter, which we use in the present study, is a classi- higher frequencies of the spectrum. This results in a peculiar cal and robust method giving reliable results especially when spectral signature with a brightness decrement at long wave- we set proper continuity and boundary conditions. In order lengths and an increment at short wavelengths (Figure 1, to separate the primary CMB signal from KSZ, several meth- solid line). When the galaxy cluster moves with respect to the ods can be used such as principal component analysis or least CMB rest frame, with a peculiar radial velocity vr , a Doppler square minimisation. Here, we choose to minimise over sta- shift called the kinetic SZ (KSZ) effect generates a tempera- tistical criteria. At the minimisation step,itisthusimportant ture anisotropies with the same spectral signature as the pri- to use the tools which emphasize the different statistical char- mary CMB fluctuations, at first order (Figure 1, dashed line). acteristics of the signal (power, non-Gaussian signatures). The importance of the SZ effect for cosmology has been Among the numerous possible tools used to exhibit non- recognized very early (see reviews by [22, 23]). It is a pow- Gaussian signatures (higher-order moments in real space, bi- erful tool to detect high-redshift galaxy clusters since it is and tri-spectrum (e.g., [17]), Minkowski functionals (e.g., redshift independent. In combination with X-ray observa- [18]), higher criticism (e.g., [19]), the multiscale transforms tions, it can be used to determine the Hubble constant and seem to be the most satisfactory and will thus be used in the probe the intracluster gas distribution. Moreover, the KSZ following. In the same spirit, we use the most sensitive non- effect may be one of the best ways of measuring the cluster Gaussian estimator among the coefficients in a biorthogonal peculiar velocities by combining thermal and kinetic effects wavelet transform, namely, the diagonal coefficients. [21]. Formally, the KSZ can be distinguished from the TSZ Separating Kinetic SZ from CMB 2415 effect due to their different frequency dependences. The KSZ From this simple statement we build a two-step compo- intensity reaches its maximum at ∼218 GHz, just where the nent separation strategy in which (i) we first derive the best TSZ intensity is zero (Figure 1). In practice, very few mea- estimate of the CMB map by interpolation, and (ii) conse- surements of the peculiar velocities were attempted [24, 25]. quently deduce the best estimate of the KSZ map by minimi- With usual component separation techniques it has been sation. shown that the TSZ signal can be extracted from the CMB rather easily (its frequency dependence is quite different from 3.1. Estimating primary CMB anisotropies a black-body spectrum) while the KSZ component remains The basic idea in order to estimate the primary CMB indistinguishable from the primary CMB anisotropies due anisotropies is to use the TSZ map as a mask to select in to same frequency distribution. In early works, [26] used an the δT map the pixels where only primary anisotropies are optimal filtering (Wiener), with a spatial filter derived from present. These pixels contain no TSZ fluctuations in the y X-ray observations of galaxy clusters, that minimises confu- map. The rest of the pixels in the δT map are masked pix- sion with CMB, and [27] used a matched filter optimised els. We then interpolate the δT signal on these masked pixels on simulated data and independent of the underlying CMB with the constraint that pixels where the signal is associated model. with only primary CMB, keep their values after the interpo- We simulate (512 × 512 pixels) primary CMB (∆T/ lation. We therefore obtain an estimated primary CMB map. T RMS = σ = . × −5 σ = . × −5 y )CMB CMB 1 9 10 ), TSZ (mean y 1 17 10 , It is worth noting that map is an observable quantity that is ∆T/T RMS =−. × −5 rather easy to obtain from multifrequency observations due i.e., mean ( )TSZ 2 34 10 at 2 mm) and KSZ ∆T/T RMS = σ = . × −6 to its spectral signature. This is what makes it useful for the (mean ( )KSZ KSZ 1 85 10 ) maps with a pixel- size of 1.5 arcmin. A precise description of the SZ simulations mask definition. is given in [28]. The CMB signal is a Gaussian field whose Formally, the KSZ map can then be estimated simply by ff δ power spectrum is computed from an inflationary flat, low computing the di erence between the original unmasked T matter density, model. The KSZ effect induces temperature map and the primary CMB map obtained from the interpo- KSZ lation. fluctuations given by δT = (∆T/T)KSZ =−(vr /c)τ,withc τ and the velocity of light and the cluster Thomson optical 3.1.1. Interpolation of the masked pixels depth. The primary CMB and the KSZ anisotropies having the same spectral shape (at first order), we construct maps of The reconstruction of the KSZ map depends on the perfor- mances of the interpolation. We use the method described thermodynamic temperature fluctuations, δT , by adding the in [31] and consider the problem of the minimisation of a two signals δT = (∆T/T)KSZ +(∆T/T)CMB. We are thus left with two simulated datasets of pure cosmological signals, one general criterion written as consisting of the temperature fluctuation maps (CMB+KSZ) E u = w k l f k l − u k l 2 and the other consisting of the Compton parameter maps, y, ( ) ( , ) ( , ) ( , ) (k,l)∈Z2 for the TSZ effect. For our study, we restrict the analysis to 15 2 2 (1) simulated maps which span a representative range of ampli- + λ dx ∗ u(k, l) + dy ∗ u(k, l) , tudes for all signals. (k,l)∈Z2 Note that in “real life,” it is the multifrequency CMB ex- where f is an input image, u is the desired solution, w ≥ 0isa periments that can provide us, after classical component sep- map of space-varying weights, and dx and dy are the horizon- aration, with two sets of maps. One contains both CMB and tal and vertical gradient operators, respectively. The second KSZ temperature fluctuations, as they are indistinguishable, space-invariant term in (1) is a membrane spline regulariser; and the second consists of Compton parameter maps associ- λ ff the amount of smoothness is controlled by the parameter . ated with the TSZ e ect. Taking the partial derivative of (1)withrespecttou,wefind that u is the solution of the differential equation 3. METHOD FOR SEPARATING KSZ f = Wu λLu = Au FROM CMB SIGNAL w + ,(2)
From the two available types of maps (y maps for TSZ and δT where W is the diagonal weight matrix, fw = Wf the maps for CMB + KSZ). Our goal is to obtain the best possi- weighted data vector, L is the discrete Laplacian operator, and ble estimate of the KSZ buried in the dominant CMB signal. A = W + λL a symmetric definite matrix. The inversion of We benefit for this from the fact that TSZ and KSZ features (2) is achieved using a multigrid technique [32]. Typically, are spatially correlated (e.g., [29, 30]). The spatial correla- we need two V-cycles with two iterations in the smoothing tion simply means that both effects are due to galaxy clus- Gauss-Seidel part of the algorithm to reach a residual of the ters. Therefore, where TSZ signal is present, so are KSZ fluc- order of 10−6. tuations regardless of their signs or amplitudes. Conversely, The interpolation of the primary CMB map can be where the TSZ fluctuations are absent, so are the KSZ fluctu- achieved by setting the weights to zero where the data are ations and the signal at that position in the δT map is associ- missing, that is in the masked pixels, and to one elsewhere ated with the CMB primary anisotropies only. Note however and by resolving (2). The value of λ then determines the thatclustersatrestwithrespecttoCMB(vr = 0) will have tightness of the fit at the known data points (unmasked pix- no KSZ signal. els), while the surface u is interpolated such that the value of 2416 EURASIP Journal on Applied Signal Processing the Laplacian of u is zero elsewhere. In the present work, we δT map) and the original KSZ map. When the threshold is impose a low value for λ so that the recovered values at the low, we take into account a large fraction of clusters, but the known data points are equal to the original values. This cri- interpolated surfaces are large and the quality of the interpo- terion can be relaxed to take into account corruption of the lation suffers from that. Moreover, the characteristic scale of data by additive white noise [31]. In this case, the optimum the interpolated surfaces becomes of the order of the CMB regularisation parameter λ can be defined as fluctuations leading to “confusion effects.” ffi σ2 Sinceitisdi cult to choose one single optimal TSZ λ = ,(3)threshold, we retrieve a set of interpolated CMB maps cor- E f · Lf − σ2 ( ) 4 responding to a set of TSZ threshold values. The later are where σ2 is the variance of the noise and E( f · Lf)denotes defined as follows: we compute the cumulative distribution y an estimate of the correlation between the noisy image f and function of the TSZ values in the map and we search for its Laplacian Lf. In the case of nonwhite noise, the optimal the values corresponding to 15%–95% of the total number regularisation parameter λ may be determined from the data of pixels (with a step of 5%). This gives us a set of 17 thresh- y using cross-validation methods [33], or from a given mea- old values. All pixels in the TSZ map that have parameters surement model of the signal + noise [34]. above the thresholds are identified as missing data points in δ Theperformancesoftheinterpolationareimprovedif the simulated T map, that is, the mask. the values of the Laplacian of u at the missing data points 3.1.3. Results are nonzero. Moreover, the values are set such that the first and second derivatives of the interpolated signal are contin- For each of the 15 simulated maps of our datasets, we ob- δ uous throughout the interval. These continuity conditions tain 17 TSZ thresholds, and thus 17 masked T maps. We characterise the cubic B-spline functions which are known interpolate the missing data points to recover the primary for their simplicity and their performances in terms of sig- CMB signal in the masked regions. We evaluate 17 associ- nal reconstruction [35, 36]. In practice, these conditions im- ated KSZ maps by subtracting the interpolated primary CMB δ ply that the source term fw in (2) is modified to impose maps from the total T map. nonzero values at the points where the weights are set to zero We compute for each of the 17 KSZ estimated maps ffi (i.e., masked pixels). An equivalent way to solve (2) with the the correlation coe cient between the original input KSZ ffi above-mentioned conditions is to replace the Laplacian op- map and the 17 estimated KSZ maps. The correlation coe - erator L by the quadratic operator L2. cients are plotted as a function of the standard deviation of Obviously other interpolation methods can be proposed the estimated KSZ map for each of the 17 threshold values and used to estimate the CMB data in the masked pixels. We (Figure 2). The diamonds and the dashed line represent the could for example improve the interpolation by using tex- case where the interpolation is such that the Laplacian val- tures [16]. The latter account for the morphological proper- ues are set to zero, and the triangles and the solid line are for ties of the signal. Such method is limited by our knowledge the case in which the Laplacian values are nonzero. Figure 2a ffi of these characteristics. We could also think of using con- shows our best recovery case in terms of correlation coe - strained 2D realisations of the CMB to obtain the values in cient. Figure 2b is for our worst case. ffi the mask. This method is simple; however, it suffers from the From Figure 2, we see that the correlation coe cient be- precision to which the CMB power spectrum is estimated, or tween the original and the estimated KSZ maps is higher in other words the precision on the cosmological parameters when the Laplacian values are nonzero than when they are set used for the realisation. to zero. This is especially true for the maps with low standard deviations. The improvement due to the biharmonic opera- 3.1.2. Defining the masked pixels tor is of the order of 20% in our worst case (Figure 2b). We We now define the mask, that is, how we select the missing will therefore use, in the following, the L2 operator as it gives data points. Besides the pixels that actually contain no galaxy better interpolation. In addition, we see that the correlation clusters, that is, no SZ contributions, we fix a threshold value coefficient increases when the standard deviation of the es- for the TSZ amplitude below which TSZ signal is considered timated KSZ map increases (i.e., TSZ threshold decreases). too small to be detected. The corresponding pixels in the δT The correlation coefficient reaches a maximum value and maps are then associated only with primary CMB signal. On then decreases for the highest KSZ standard deviations (i.e., the contrary, above this threshold, pixels in the δT map are the lowest TSZ thresholds). considered to be the missing data points (masked pixels) that we want to interpolate. The number and location of the miss- 3.2. Reconstructing the KSZ map ing data depend on the threshold. The choice of this thresh- From the previous step, we obtained a set of 17 estimated old has thus important consequences on the quality of the in- KSZ maps. Now, we search for a method that gives us either terpolation. When the threshold is high, the number of miss- the reconstructed KSZ map which is the closest to the orig- ing data is small and the interpolated surface is good but the inal KSZ signal or the combination of 17 KSZ maps giving selection retains only the clusters with the highest TSZ and the best estimate of the original KSZ map. We compare the misses the majority of clusters. In this case, we expect a low reconstructed maps with the original KSZ maps. This allows correlation coefficient between the retrieved KSZ map (ob- us to calibrate our method and thus provides us with the in- tained by subtracting the interpolated CMB map from the trinsic limitations of the reconstructing methods. Separating Kinetic SZ from CMB 2417
1 However, in this case, the standard deviations of the recon- Best case structed maps are lower than the original KSZ signal by al- most 25% on average. Furthermore, the results of the SVD 0.8 least square minimisation depend on the set of estimated cient maps that are used which is clearly undesirable. ffi The two previous attempts being not quite satisfactory, . 0 6 we thus need as much map-independent results as possible. We must identify a criterion, to minimise on, which should ffi . ideally give the largest possible correlation coe cient and the Correlation coe 0 4 reconstructed KSZ maps with the closest possible standard deviations to those of the original KSZ signal. Moreover, a 0.2 good minimisation criterion would characterise the KSZ sig- 123456nal only, excluding the primary CMB signatures. We have −6 Standard deviation ∆T/T ×10 identified two properties of the KSZ fluctuations that fulfill this definition. (a) (i) The KSZ signal dominates primary CMB at high wave 1 numbers (small angular scales). ff Worst case (ii) The KSZ e ect is a highly non-Gaussian process con- trary to primary CMB which is a Gaussian process. . 0 8 The KSZ effect is due to galaxy clusters whose typical sizes cient
ffi are a few to a few tens of arcmin. As a result, SZ anisotropies produced either by KSZ or TSZ intervene at small angu- 0.6 lar scales where they show a maximum amplitude (Figure 3, dashed line). At those scales, primary CMB anisotropies are . severely damped and the angular power spectrum decreases Correlation coe 0 4 sharply (Figure 3, solid line). Therefore, at small scales, both the power and the statistical properties of the total δT signal 0.2 should be those of the dominant signal, that is, KSZ effect. 0123456In order to focus on the KSZ signal and also to enhance the × −6 Standard deviation ∆T/T 10 signal-to-noise ratio, we perform multiscale wavelet decom- δ (b) position of the T map. The above-mentioned properties re- main true in the wavelet domain as it was first recognised by Figure 2: The correlation coefficient between the original KSZ map [37] and applied by [38]. Thus, the statistical properties of and the series of 17 estimated KSZ maps as a function of the stan- the wavelet coefficients at the lowest decomposition scale (3 dard deviations of the estimated KSZ maps. (a) The best case and arcmin) reflect the properties of the SZ effect only. (b) the worst case. Triangles and solid line stand for the interpola- We use the decimated biorthogonal wavelet transform tion with the biharmonic operator; diamonds and dashed line are which decomposes a signal s as follows: for the interpolation with the Laplacian. The interpolation with the biharmonic operator gives better results especially for the KSZ maps J with low standard deviation. The vertical lines mark the standard s(l) = cJ,kφJ,l(k)+ ψj,l(k)wj,k (4) −6 −6 deviation of the original KSZ maps (2.6 ×10 and 1.2 ×10 ). The k k j=1 standard deviation of the primary CMB is 1.9 ×10−5. − j − j − j − j with φj,l(x) = 2 φ(2 x − l)andψj,l(x) = 2 ψ(2 x − l), where φ and ψ are, respectively, the scaling and the wavelet 3.2.1. Method functions. J is the number of resolutions used in the de- We test the decorrelation by principal component analysis composition, wj the wavelet coefficients (or details) at scale (PCA). The PCA gives us a reconstructed KSZ signal which j,andcJ a smooth version of s (j = 1 corresponds to the is rather close to the original. The correlation coefficient, av- finest scale, highest frequencies). The two-dimensional al- eraged over the 15 maps, between reconstructed and orig- gorithm gives three wavelet subimages at each decompo- inal KSZ reaches 0.73. However, the standard deviation is sition scale. Within this choice, the wavelet analysis pro- on average smaller by almost 50% than the original. This vides us with the wavelet coefficients associated with diag- is not satisfactory. We can also search for a linear combi- onal, vertical, and horizontal details of the analysed map. nation of the 17 estimated KSZ maps that is the closest to Using this tool, we have demonstrated [39, 40] that the ex- the original KSZ in the sense of least squares. This minimi- cess kurtosis of the wavelet coefficients in a biorthogonal sation is done using a standard singular value decomposition decomposition allows us to discriminate between a Gaus- (SVD). The average correlation coefficient (over the 15 sim- sian primary CMB signal and a non-Gaussian process like ulated input maps) between the original KSZ map and the SZ effect better than with an orthogonal wavelet decompo- reconstructed map is 0.8, slightly higher than the PCA result. sition. Moreover, we have shown that diagonal details are 2418 EURASIP Journal on Applied Signal Processing
×10−6 10−8 3
l − C 10 9
+1) 2.5 l T/T (
l −10 10 ∆
10−11 2
10−12 1.5 Standard deviation 10−13 Angular power spectrum
10−14 1 1000 02468101214 Multipole Map number
Figure 3: Angular power spectrum of the primary CMB Figure 4: Standard deviations of our set of 15 KSZ original simu- anisotropies (solid line) and of the KSZ fluctuations from galaxy lated maps (triangles) as compared with standard deviations of the clusters (dashed line). The plots are for one statistical realisation of 15 reconstructed KSZ maps (squares). The reconstruction is based both processes. (The multipole is equivalent to a wave number in on the minimisation of the statistical criterion. the spherical harmonic decomposition of the sky.)
Table 1: The statistical properties of the first scale (3 arcmin) diag- Consequently, we can confidently minimise on the statis- onal wavelet coefficients distribution for the δT map (KSZ + CMB), tical properties of power and non-Gaussianity at the small- the KSZ map, and the primary CMB alone. The two cases stand for est decomposition scale. In practice, we choose the follow- our best case (first pair) and the worst case (second pair). We note ing criterion minimised over the 17 estimated KSZ maps (for that the three moments are almost identical and characterise well each of the 15 maps of our dataset): the KSZ fluctuations; they are very different from the CMB fluctua- tions properties. M w − M w 2 M w − M w 2 ζ = 2 0 2( ) 4 0 4( ) Min M2 w + M2 w , Standard deviation Skewness Excess kurtosis 2 0 4 0 KSZ + CMB 6.45 ×10−7 0.10 8.71 (5) KSZ 6.45 ×10−7 0.10 8.72 where w0 is the distribution of diagonal wavelet coefficients −7 KSZ + CMB 2.05 ×10 0.22 8.97 for the known δT map(KSZ+CMB)andw is the distri- KSZ 2.09 ×10−7 0.23 9.15 bution of diagonal wavelet coefficients for the desired solu- − M M CMB 1.60 ×10 8 −0.02 0.45 tion map (the reconstructed map). 2 and 4 are respec- tively the second and the fourth moments of the wavelet co- efficients. This criterion takes into account both the energy content of the coefficients, through second moment, and the the most sensitive to non-Gaussian signatures (recently con- non-Gaussian character, through fourth moment. We have firmed and explained in [41]). We therefore choose to use chosen the fourth moment because it is the one for which the the diagonal details in a biorthogonal wavelet decomposi- KSZ signal is the most sensitive to non-Gaussianity. Clearly, tion at the smallest decomposition scale to obtain the best we might also include the third moment of the wavelet coef- results. ficients to the criterion. This would be needed in particular In Table 1, we compare, using the 9/7 biorthogonal filter if we were dealing with a “skewed” signal (e.g., weak lensing bank [42] for the worst and best cases, the statistical prop- signal). Taking the fourth moment in the minimisation crite- erties of the diagonal details of KSZ maps and CMB + KSZ rion allows us in turn to focus on the reconstruction of KSZ maps at the first decomposition scale (3 arcmin). We also give maps excluding any skewed signal that might contribute at the values for the primary CMB maps. As expected, we note small scales. that the wavelet coefficients for KSZ and KSZ + CMB share In addition to the conditions of power and non-Gaussian the same statistical properties and are quite different from character, we make use in the minimisation process, of a those of the primary CMB alone. This confirms that KSZ sig- nice property of the wavelet transform, which is that it pre- nal dominates over primary CMB in wavelet domain (same serves the spatial information. Thus instead of minimising standard deviation means same power, c.f. Figure 3 in real over all wavelet coefficients of the data map (w0 in (5)), we space), and that non-Gaussian signatures in the KSZ + CMB can minimise only over those corresponding to clusters. This maps are associated with the KSZ effect (same skewness and enhances the non-Gaussian character and reduces the influ- excess kurtosis) alone at the smallest decomposition scale (3 ence of other possible non-Gaussian processes that could af- arcmin). fect the anisotropy map δT . Separating Kinetic SZ from CMB 2419
106 106 105 105 4 −6 4 −6 10 σreal = 2.684 × 10 10 σreal = 1.177 × 10 σ = . × −6 σ = . × −6 103 est 2 594 10 103 est 1 118 10 102 102 Number count 101 Number count 101 100 100 −4 −20 2 4 −3 −2 −10123 ∆T/T ×10−5 ∆T/T ×10−5
(a) (a)
10−10 10−11
10−11 10−12 l l C C 10−12 10−13 +1) −16 +1) −16 l Preal = 4.384 × 10 l Preal = 2.116 × 10 ( ( l P = . × −16 l P = . × −16 10−13 est 4 107 10 10−14 est 1 274 10
10−14 10−15 10 100 1000 10000 10 100 1000 10000 Multipole l Multipole l
(b) (b)
10 10
ratio 1 ratio 1 l l C C
0.1 0.1 10 100 1000 10000 10 100 1000 10000 Multipole l Multipole l
(c) (c)
Figure 5: (a)-(b) Histogram and power spectrum of the original Figure 6: The same as Figure 5. This is our worst reconstruction KSZ map (solid line) and of the reconstructed KSZ map (dashed case which corresponds to the original KSZ map with the lowest line). The reconstruction was obtained by minimising a statistical standard deviation. Note the low correlation coefficient 0.62. criterion. (c) The ratio of original to reconstructed power spectrum. Note the correlation coefficient between original and reconstructed ∼ P KSZ maps of 0.9 and the agreement between total power real and The quality of the KSZ map reconstruction can be observed P est. in Figures 5 and 6 which display, for our best and worst cases respectively, the histograms of the temperature fluctuations 3.2.2. Results and the power spectra of both original (solid line) and re- constructed (dashed line) KSZ maps as well as the ratio of In Figure 4, we present the standard deviations of the 15 these two power spectra. Note that the ratio is close to one original simulated KSZ maps (triangles) and of the 15 over a large range of multipoles (wave number in the spheri- reconstructed KSZ maps (squares) obtained by the above- cal harmonic decomposition) even in the domain where the mentioned minimisation technique. The agreement is good primary CMB dominates the KSZ signal (see Figure 3). We even for the lowest standard deviations with an error only also notice the correlation coefficient between original and of the order of ∼5%. This is much smaller than what was reconstructed KSZ maps which reaches ∼0.9 in our best case obtained from the PCA method (∼50%) or from the least and 0.62 in our worst case. The comparison between stan- square minimisation (∼25%). Furthermore, the mean value dard deviations of original and reconstructed maps σreal and (over the 15 original maps) of the correlation coefficient be- σest also gives a global indication on how well the method tween the original and the reconstructed KSZ maps is 0.78. works. 2420 EURASIP Journal on Applied Signal Processing
×10−5 ×10−6 2 6 1 4 2
T/T 0 T/T
∆ ∆ 0 − 1 Line 64 −2 Line 64 −2 −4 0 100 200 300 400 500 0 100 200 300 400 500 Pixel number Pixel number
×10−5 ×10−6 1.5 8 1 6 0.5 4 2
T/T 0 T/T ∆ . ∆ 0 −0 5 −2 Line 192 Line 192 −1 −4 −1.5 −6 0 100 200 300 400 500 0 100 200 300 400 500 Pixel number Pixel number
×10−5 ×10−6 2 10 1 5
T/T 0 T/T ∆ ∆ 0 −1 Line 320 Line 320 −2 −5 0 100 200 300 400 500 0 100 200 300 400 500 Pixel number Pixel number
×10−5 ×10−6 1.5 6 1 4 . 0 5 2 0 T/T − . T/T 0
∆ 0 5 ∆ −1 −2 Line 448 Line 448 −1.5 −4 −2 −6 0 100 200 300 400 500 0 100 200 300 400 500 Pixel number Pixel number
Figure 7: Cuts across the best reconstructed KSZ map (dashed line) Figure 8: Same as Figure 7 for the worst reconstructed KSZ map and its original counterpart (solid line). The cuts have the same po- and its original counterpart. sition in both maps.
The method allows us to obtain such results because we structions of the KSZ original maps, both in terms of corre- are able to estimate correctly the amplitude of the KSZ sig- lation coefficient, power spectrum and pixel distribution. We nal for most clusters together with their angular separation, now investigate some of the effects that can affect our results. as well as the amplitude of the background (primary CMB). 4.1. Amplitude of the input KSZ signal This is nicely exhibited by the superposition of the cuts across reconstructed (dashed line) and original (solid line) KSZ The previous results were obtained in a specific model which maps, for the best and worst cases (Figures 7 and 8,resp.). predicts the amplitude of the KSZ signal and thus its ratio The method partially fails to find broad KSZ features due to to primary CMB anisotropies. Obviously the KSZ amplitude their important level of confusion with primary CMB fluctu- can vary for many physical reasons (number of clusters, dis- ations. Moreover, since the minimisation process is an overall tribution of velocities, etc.). It is thus important to test what procedure, relatively large features (i.e., of the order of 10−5 is the performances of our separation method are in response ff in absolute ∆T/T) are occasionally poorly recovered. to di erent mixing ratios. For illustration, we take one KSZ map and add it to the same primary CMB map. The stan- dard deviation of the KSZ signal is reduced while the CMB 4. SENSITIVITY TESTS standard deviation is kept the same (i.e., we reduce the KSZ contribution to the 5δT map). We reduce the standard devi- We have shown in the previous section that statistical min- √ i imisation with a well-chosen criterion gives very good recon- ation following a geometrical progression σi = σ0 2 with Separating Kinetic SZ from CMB 2421
Table 2: Standard deviations of the KSZ maps and correlation coefficients between original and reconstructed KSZ maps for the same KSZ map with standard deviations ranging from 2.5 ×10−7 to 2.0 ×10−6. Two wavelet bases are tested.
9/7 filter 6/10 filter Original σ Estimated σ Correlation coefficient Estimated σ Correlation coefficient 2.5 ×10−7 2.21 ×10−7 0.48 2.68 ×10−7 0.45 3.53 ×10−7 3.36 ×10−7 0.54 4.02 ×10−7 0.52 5.0 ×10−7 5.52 ×10−7 0.56 5.46 ×10−7 0.58 7.07 ×10−7 8.02 ×10−7 0.59 6.60 ×10−7 0.67 1.0 ×10−6 9.74 ×10−7 0.68 9.16 ×10−7 0.71 1.41 ×10−6 1.35 ×10−6 0.74 1.20 ×10−6 0.77 2.0 ×10−6 1.94 ×10−6 0.78 1.77 ×10−6 0.81
−7 i = 0, 6 and σ0 = 2.5 ×10 . The highest standard devia- To illustrate the effectofalargerbeam,wehaveconvolved −6 tion is then σmax = 2.0 ×10 which is a typical value for our observed maps (y and δT maps) by a Gaussian-shaped our dataset. At the same time, we test the sensitivity of our beam (for simplicity) with a size of 3 arcmin. We find that method to the wavelet base by comparing results obtained the reconstructed KSZ map is not satisfying neither in terms using two different biorthogonal wavelet bases, the 9/7 tap of the correlation coefficients (mean coefficient of 0.59), nor filter and the 6/10 tap filter [43]. in terms of the average amplitudes (standard deviations of The results for this test are displayed in Table 2.Wefirst the 15 reconstructed maps are typically 40% smaller than the notice that results do not depend much on the wavelet ba- original), nor in terms of the power spectrum. We show in sis. As expected, the quality of the reconstruction (in terms Figure 10 a cut across a reconstructed KSZ map (which is not of correlation coefficient) increases with the standard devi- our best case) and its original counterpart. We note that only ation of the original KSZ map from 0.5 to ∼0.8. The small- the largest amplitude features are reconstructed but with am- est coefficients are obtained for very-low-standard deviations plitudes which are lower than the original. As expected, we (< 10−6). It is worth noting that decreasing the KSZ ampli- find that the results get worse for larger beam sizes. tudebyafactor2(i.e.,afactor4inpower)stillgivesareason- One way to improve our results in the case of large beam ably good correlation coefficient. At the same time, the stan- experiments might be to use a minimisation criterion in (5) dard deviation of the reconstructed KSZ map is very close to based on other wavelet decomposition scales which should the original even when the input KSZ signal is decreased by be less affected by the beam dilution. For example, the second one order of magnitude in terms of standard deviation. This smallest decomposition scale could be used in the case of a 3 is illustrated in Figure 9 by the reconstructed power spec- arc-minute beam. At that scale, the non-Gaussian character trum of the KSZ map. of the KSZ signal is indeed still preserved (see [39]), how- ever the power is no more dominated by KSZ but rather by 4.2. Beam convolution the CMB primary signal. More adapted criteria should then Our separation method is based on two steps; the first is be investigated, but they will likely require more “a priori” the interpolation and the second is the minimisation. Obvi- knowledge of both KSZ and primary CMB signals. ously, when the sky is observed by an instrument, the δT map suffers from beam dilution, which means that the signal is 4.3. Noise damped at the typical scale of the beam size. The same is true We illustrate possible effects of noise on our separation for the y map for which the damping can be even more severe method by adding to the observed δT and y maps a white since the signal is mainly at small scales. As a consequence, noise at the pixel size whose RMS amplitude in terms of tem- the definition of the mask based on the TSZ template and perature fluctuation is 2 × 10−6. This corresponds to a noise used in the interpolation is also affected by beam dilution. level of about 6 µK which is the typical noise of most future We expect that this reduces the quality of the interpolation SZ experiments. We note that the RMS noise level is of the and in turn that of the 17 estimated KSZ maps. Moreover, the order of the mean standard deviation of the original input minimisation criterion is based on two properties of the KSZ KSZ signal. It is twice as large as the standard deviation of signal (non-Gaussian character and excess of power) as com- some KSZ maps. This not only modifies the amplitude of pared to the primary CMB, which are mostly true at small the fluctuations in the δT at a given position in the map, but scales. When the δT map is convolved by the beam instru- also significantly modifies the position of the maxima and ment, the contribution from KSZ signal is reduced affecting shape of the fluctuations associated with the KSZ signal. As a also the statistical minimisation criterion. consequence, the spatial correlation between TSZ and KSZ is All these effects depend on the size of the beam. The decreased and the reconstructed KSZ signal is different from smallest the beam is, the less affected the recovered signal the input map (see Figure 11). The correlation coefficients is. For a beam-size of 1.5 arc-minute (like that of some between the original and reconstructed KSZ maps are obvi- planned SZ experiments), there should be no effect on our ously very low in this case with values ranging between 0.24 results since our minimum resolution is 1.5arc-minute. and 0.54. 2422 EURASIP Journal on Applied Signal Processing
106 4 ) 105 6 − 4 −7 10 σreal = 2.500 × 10 2 σ = . × −7 103 est 2 208 10 102 0 Number count 101 0 10 − −3 −2 −10123 2 ∆T/T ×10−6 −4 −12
10 Amplitude of KSZ fluctuations (10 −6 10−13 0 100 200 300 l
C Pixel number 10−14 +1) −18 l Preal = 9.459 × 10 ( l P = . × −18 10−15 est 3 156 10 Reconstructed Input 10−16 10 100 1000 10000 Multipole l Figure 10: Cuts across a typical reconstructed KSZ map (dashed line) and its original counterpart (solid line). The cuts have the same position in both maps. (No noise, convolution beam = 3 arcmin.) 10
is that no a priori criteria are needed to obtain the KSZ map.
ratio 1 However, the resulting maps are of low quality in terms of l
C standard deviation. More sophisticated methods such as the independent component analysis [12, 44, 45]canbeusedbut the results obtained need to be rescaled using external con- 0.1 10 100 1000 10000 straints. Multipole l In our study, we choose to use a reconstruction method based on a minimisation technique. We propose a minimisa- tion criterion taking into account statistical properties of the Figure 9: Same as Figure 5. The standard deviation of the origi- −7 KSZ signal: (i) KSZ dominates over primary anisotropies at nal KSZ map is very low (σreal = 2.5 ×10 ). Note the excess of near-zero values in the histogram of the estimated map (logarith- small angular scales, and (ii) KSZ fluctuations follow a non- micscale).Notealsotheverylowcorrelationcoefficient 0.48. This Gaussian distribution. We use the excess kurtosis of the di- is for the worst case. agonal wavelet coefficients to characterise the non-Gaussian signatures of the KSZ effect. The minimisation method gives reconstructed KSZ maps that are in quite good agreement As noted in Section 3.1.1, the white noise can in princi- with the original signal with an average correlation coeffi- ple be accounted for at the interpolation stage in the regu- cient between original and reconstructed KSZ maps of 0.78, larisation parameter. Such possibilities should have to be ad- and an error of 5% on the standard deviation of recon- dressed. structed KSZ maps. The KSZ reconstruction through min- imisation depends on the minimisation criteria and there- fore on our knowledge of the signals. The available CMB data 5. DISCUSSION seem to agree on the fact that primary CMB anisotropies are We presented a method for separating the KSZ signal from Gaussian distributed at least at small scales [46, 47, 48]; see primary CMB anisotropies based on two steps: (1) interpo- [49] for large scales. The KSZ effect is dominant at small lation and (2) reconstruction. In our case this corresponds to scales since it is associated with galaxy clusters. We have the interpolation of a correlated noise (the CMB). The KSZ tested our results against the relative amplitude of KSZ to reconstruction is based on a set of KSZ estimated maps ob- primary signal. We find satisfactory results even when KSZ tained with a choice of TSZ thresholds (from the cumulative is twice as small (in RMS) as predicted. distribution of the pixels in the TSZ template map), more so- The results above are for the case where only the two sig- phisticated methods optimising the series of TSZ thresholds nals CMB and SZ are taken into account, which allows us can be proposed. Using the set of KSZ estimated maps, we to investigate the intrinsic limitations of the method. Addi- can investigate several methods to reconstruct the final KSZ tional astrophysical contributions should be partly treated maps. We tested a decorrelation-based approach using the in a first-step component separation (from which we ob- PCA. The decorrelation is a blind method whose advantage tain the observed y and δT maps). For example if some Separating Kinetic SZ from CMB 2423
20 ACKNOWLEDGMENT ) 6 − 10 The authors wish to thank the Editor C. Baccigalupi for his incitement and anonymous referees for their comments on a previous version. 0
−10 REFERENCES [1] A. D. Miller, R. Caldwell, M. J. Devlin, et al., “A measurement −20 of the angular power spectrum of the CMB from l = 100 to 400,” Astrophysical Journal, vol. 524, pp. L1–L4, 1999. −30 [2] P. de Bernardis, P. A. R. Ade, J. J. Bock, et al., “A flat universe from high-resolution maps of the cosmic microwave back- Amplitude of KSZ fluctuations (10 −40 ground radiation,” Nature, vol. 404, no. 6781, pp. 955–959, 0 100 200 300 2000. Pixel number [3] S . Hanany, P. A. R. Ade, A. Balbi, et al., “MAXIMA-1: a mea- surement of the cosmic microwave background anisotropy on ◦ Reconstructed angular scales of 10 −5 ,” Astrophysical Journal, vol. 545, no. 1, Input pp. L5–L9, 2000. [4] N. W. Halverson, E. M. Leitch, C. Pryke, et al., “Degree an- Figure 11: Cuts across a typical reconstructed KSZ map (dashed gular scale interferometer first results: a measurement of the line) and its original counterpart (solid line). The cuts have the same cosmic microwave background angular power spectrum,” As- position in both maps. (Noise = 2 × 10−6, no convolution.) trophysical Journal, vol. 568, no. 1, pp. 38–45, 2002. [5] A. Benoit, P. A. R. Ade, A. Amblard, et al., “The cosmic mi- crowave background anisotropy power spectrum measured by Archeops,” Astronomy and Astrophysics, vol. 399, no. 3, contribution from the TSZ signal remains in the δT map, it pp. L19–L23, 2003. will act as a correlated noise. We can account for it at the [6]C.L.Bennett,M.Halpern,G.Hinshaw,etal.,“Firstyear interpolation stage with the additional constraint that the Wilkinson Microwave Anisotropy Probe (WMAP) Observa- skewness should be zero (which is the case for the primary tions: preliminary maps and basic results,” Astrophysical Jour- CMB anisotropies), or in the minimisation procedure using nal Supplement Series, vol. 148, no. 1, pp. 1–27, 2003. [7] P. F. Scott, P. Carreira, K. Cleary, et al., “First results from a generalised criterion including the skewness as well as the the very small array—III. The cosmic microwave background excess kurtosis. In the present study, we have tested for the power spectrum,” Monthly Notices of the Royal Astronomical presence of an instrumental white noise at the pixel scale Society, vol. 341, no. 4, pp. 1076–1083, 2003. with 6 µK RMS amplitude. We find that such noise level af- [8] T. J. Pearson, B. S. Mason, A. C. S. Readhead, et al., “The fects the KSZ map reconstruction making it difficult to re- anisotropy of the microwave background to l = 3500: mosaic cover the KSZ signal buried in CMB. In theory, instrumental observations with the cosmic background imager,” Astrophys- ical Journal, vol. 591, pp. 556–574, 2003. noise can be taken into account in the interpolation step by λ [9] M. C. Runyan, P. A. R. Ade, J. J. Bock, et al., “First results from relaxing the parameter . Another way to deal with noise is the arcminute cosmology bolometer array receiver,” New As- to minimise not on the non-Gaussian character of the KSZ, tronomy Reviews, vol. 47, no. 11-12, pp. 915–923, 2003. but rather on the statistical properties of the remainder (i.e., [10]M.P.Hobson,A.W.Jones,A.N.Lasenby,andF.R.Bouchet, CMB+noise+other components) at scales where CMB dom- “Foreground separation methods for satellite observations of inates. We should then obtain an estimate of all the compo- the cosmic microwave background,” Monthly Notices of the nents except KSZ that can then be subtracted from the total Royal Astronomical Society, vol. 300, no. 1, pp. 1–29, 1998. [11] F. R. Bouchet and R. Gispert, “Foregrounds and CMB signal. These methods will need to be investigated in the fu- experiments—I. Semi-analytical estimates of contamination,” ture. New Astronomy, vol. 4, no. 6, pp. 443–479, 1999. Another key element of our separation method is the use [12] C. Baccigalupi, L. Bedini, C. Burigana, et al., “Neural net- of a spatial template. The choice of a spatial template is an works and the separation of cosmic microwave background important issue since it is used to define the mask and hence and astrophysical signals in sky maps,” Monthly Notices of the the interpolated regions. The template should then be the Royal Astronomical Society, vol. 318, no. 3, pp. 769–780, 2000. closest possible to the signal. In our case, the optimal choice [13] J. Delabrouille, J.-F. Cardoso, and G. Patanchon, “Multidetec- tor multicomponent spectral matching and applications for is the TSZ signal itself as it allows us to evaluate the temper- cosmic microwave background data analysis,” Monthly Notices ature fluctuations associated with KSZ in the map without of the Royal Astronomical Society, vol. 346, no. 4, pp. 1089– resorting to the knowledge or the measurement of cluster pa- 1102, 2003. rameters. However, the beam dilution caused by observation [14] M. P.Hobson and C. McLachlan, “A Bayesian approach to dis- suppresses the signal at small scales and can significantly af- crete object detection in astronomical data sets,” Monthly No- fect the results especially for large beam sizes. (Note that our tices of the Royal Astronomical Society, vol. 338, no. 3, pp. 765– . 784, 2003. previous results are equivalent to a 1 5 arc-minute beamsize.) [15] O. Forni and N. Aghanim, “Separating the kinetic Sunyaev- One way around the problem is to resort to multiscale min- Zel’dovich effect from primary cosmic microwave back- imisation criteria at the reconstruction step; we will investi- ground fluctuations,” Astronomy and Astrophysics, vol. 420, gate this question in the future. no. 1, pp. 49–60, 2004. 2424 EURASIP Journal on Applied Signal Processing
[16] D. J. Heeger and J. R. Bergen, “Pyramid-based texture anal- [35] M. Unser, A. Aldroubi, and M. Eden, “B-spline signal process- ysis/synthesis,” in Proc. 22nd Annual Conference on Computer ing: Part II—efficient design and applications,” IEEE Trans. Graphics (SIGGRAPH ’95), pp. 229–238, Los Angeles, Calif, Signal Processing, vol. 41, no. 2, pp. 834–848, 1993. USA , August 1995. [36] P. Thevenaz,´ T. Blu, and M. Unser, “Interpolation revisited,” [17]M.Kunz,A.J.Banday,P.G.Castro,P.G.Ferreira,andK.M. IEEE Trans. Med. Imag., vol. 19, no. 7, pp. 739–758, 2000. Gorski, “The trispectrum of the 4 year COBE DMR data,” As- [37] P. G. Ferreira, J. Magueijo, and J. Silk, “Cumulants as trophysical Journal, vol. 563, no. 2, pp. L99–L102, 2001. non-Gaussian qualifiers,” Physical Review D,vol.56,no.8, [18] S. F. Shandarin, “Testing non-gaussianity in cosmic mi- pp. 4592–4603, 1997. crowave background maps by morphological statistics,” [38] J. Pando, D. Valls-Gabaud, and L.-Z. Fang, “Evidence for Monthly Notices of the Royal Astronomical Society, vol. 331, scale-scale correlations in the cosmic microwave background no. 4, pp. 865–874, 2002. radiation,” Physical Review Letters, vol. 81, no. 21, pp. 4568– [19] D. Donoho and J. Jin, “Higher criticism for detecting sparse 4571, 1998. heterogeneous mixtures,” Annals of Statistics, vol. 32, no. 3, [39] N. Aghanim and O. Forni, “Searching for the non-Gaussian pp. 962–994, 2004. signature of the CMB secondary anisotropies,” Astronomy and [20] R. A. Sunyaev and Y. B. Zel’dovich, “The observations of Astrophysics, vol. 347, pp. 409–418, 1999. relic radiation as a test of the nature of X-Ray radiation from [40] O. Forni and N. Aghanim, “Searching for non-gaussianity: the clusters of galaxies,” Comments on Astrophysics and Space statistical tests,” Astronomy and Astrophysics Supplement Se- Physics, vol. 4, pp. 173–178, 1972. ries, vol. 137, pp. 553–567, 1999. [21] R. A. Sunyaev and Y. B. Zel’dovich, “Microwave background [41] J.-L. Starck, N. Aghanim, and O. Forni, “Detection and radiation as a probe of the contemporary structure and his- discrimination of cosmological non-Gaussian signatures by tory of the universe,” Annual Review of Astronomy and Astro- multi-scale methods,” Astronomy and Astrophysics, vol. 416, physics, vol. 18, pp. 537–560, 1980. pp. 9–17, 2004. [22] Y. Rephaeli, “Comptonization of the cosmic microwave back- [42] A. Cohen, I. Daubechies, and J. C. Feauveau, “Biorthogonal ground: the Sunyaev-Zel’dovich effect,” Annual Review of As- bases of compactly supported wavelets,” Tech. Rep., AT&T tronomy and Astrophysics, vol. 33, pp. 541–579, 1995. ff Bell Lab, Murray Hill, NJ, USA, 1990, Page: TM 11217- [23] M. Birkinshaw, “The Sunyaev-Zel’dovich e ect,” Physics Re- 900529-07. ports, vol. 310, pp. 97–195, 1999. [43] J. D. Villasenor, B. Belzer, and J. Liao, “Wavelet filter evalu- [24] J. M. Lamarre, M. Giard, E. Pointecouteau, et al., “First mea- ff ation for image compression,” IEEE Trans. Image Processing, surement of the submillimeter Sunyaev-Zel’dovich e ect,” As- vol. 4, no. 8, pp. 1053–1060, 1995. trophysical Journal, vol. 507, pp. L5–L8, 1998. [44] J.-F. Cardoso and A. Souloumiac, “Blind beamforming for [25] B. A. Benson, S. E. Church, P. A. R. Ade, et al., “Peculiar veloc- non Gaussian signals,” IEE Proceedings-F, vol. 140, no. 6, ity limits from measurements of the spectrum of the Sunyaev- pp. 362–370, 1993. Zel’dovich effect in six clusters of galaxies,” Astrophysical Jour- nal, vol. 592, no. 2, pp. 674–691, 2003. [45] A. Hyvarinen,¨ “Fast and robust fixed-point algorithms for independent component analysis,” IEEE Trans. Neural Net- [26] M. G. Haehnelt and M. Tegmark, “Using the kinematic works, vol. 10, no. 3, pp. 626–634, 1999. Sunyaev-Zel’dovich effect to determine the peculiar velocities of clusters of galaxies,” Monthly Notices of the Royal Astronom- [46] L. Cayon, E. Martinez-Gonzalez, F. Argueso, A. J. Banday, and ical Society, vol. 279, pp. 545–556, 1996. K. M. Gorski, “COBE-DMR constraints on the nonlinear cou- [27] N. Aghanim, A. De Luca, F. R. Bouchet, R. Gispert, and J. pling parameter: a wavelet based method,” Monthly Notices of L. Puget, “Cosmology with Sunyaev-Zel’dovich observations the Royal Astronomical Society, vol. 339, no. 4, pp. 1189–1194, from space,” Astronomy and Astrophysics, vol. 325, pp. 9–18, 2003. 1997. [47] E. Komatsu, A. Kogut, M. R. Nolta, et al., “First-year [28] N. Aghanim, K. M. Gorski, and J.-L. Pujet, “How accu- Wilkinson Microwave Anisotropy Probe (WMAP) observa- rately can the SZ effect measure peculiar cluster velocities and tions: tests of gaussianity,” Astrophysical Journal Supplement, bulk flows?” Astronomy and Astrophysics, vol. 374, pp. 1–12, vol. 148, pp. 119–134, 2003. 2001. [48] M. G. Santos, A. Cooray, Z. Haiman, L. Knox, and C.-P. Ma, [29] A. Diaferio, R. A. Sunyaev, and A. Nusser, “Large-Scale mo- “Small-scale cosmic microwave background temperature and tions in superclusters: their imprint in the cosmic microwave polarization anisotropies due to patchy reionization,” Astro- background,” Astrophysical Journal, vol. 533, no. 2, pp. L71– physical Journal, vol. 598, pp. 756–766, 2003. L74, 2000. [49] P. Vielva, E. Martinez-Gonzalez, R. B. Barreiro, J. L. Sanz, and [30] M. Sorel, N. Aghanim, and O. Forni, “Blind statistical indica- L. Cayon, “Detection of non-Gaussianity in the WMAP 1-year tors of the kinetic Sunyaev-Zel’dovich anisotropies,” Astron- data using spherical wavelets,” Astrophysical Journal, vol. 609, omy and Astrophysics, vol. 395, no. 3, pp. 747–751, 2002. no. 1, pp. 22–34, 2004. [31] M. Unser, “Multigrid adaptive image processing,” in Proc. IEEE International Conference on Image Processing (ICIP ’95), vol. 1, pp. 49–52, Washington, DC, USA, October 1995. Olivier Forni is a planetologist at the Institut d’Astrophysique Spa- [32] P. Wesseling, An Introduction to Multigrid Methods,JohnWi- tiale in Orsay (France). His main research activities deal with the ley & Sons, Chichester, UK, 1992. evolution of the planets and satellites of the solar system. Recently [33] G. Wahba, “Practical approximate solutions to linear operator he has been working on the statistical properties of the cosmic mi- equations when the data are noisy,” SIAM Journal on Numeri- crowave background (CMB) and of the secondary anisotropies by cal Analysis, vol. 14, no. 4, pp. 651–667, 1977. means of multiscale transform analysis. He also got involved in [34] S. J. Reeves, “Optimal space-varying regularization in iterative component separation techniques in order to improve the detec- image restoration,” IEEE Trans. Image Processing, vol. 3, no. 3, tion of low-power signatures and to analyse hyperspectral infrared pp. 319–324, 1994. data on Mars. Separating Kinetic SZ from CMB 2425
Nabila Aghanim is a cosmologist at the Institut d’Astrophysique Spatiale in Orsay (France). Her main research interests are cosmic microwave background (CMB) and large-scale structure. She has been working during the last ten years on the statistical charac- terisation of CMB temperature anisotropies through power spec- trum analyses and higher-order moments of wavelet coefficients. She naturally got invloved and interested in signal processing tech- niques in order to improve the detection of low signal-to-noise ra- tios such as those associated with secondary anisotropies and sepa- rate them from the primary signal. EURASIP Journal on Applied Signal Processing 2005:15, 2426–2436 c 2005 Hindawi Publishing Corporation
Detection of Point Sources on Two-Dimensional Images Based on Peaks
M. Lopez-Caniego´ Instituto de F´ısica de Cantabria, CSIC-Universidad de Cantabria, and Departamento de F´ısica Moderna, Universidad de Cantabria, avenida de los Castros s/n, 39005 Santander, Spain Email: [email protected] D. Herranz Istituto di Scienze e Tecnologie dell’Informazione “A. Faedo,” CNR, via Moruzzi 1, 56124 Pisa, Italy Email: [email protected]
J. L. Sanz Instituto de F´ısica de Cantabria, CSIC-Universidad de Cantabria, avenida de los Castros s/n, 39905 Santander, Spain Email: [email protected]
R. B. Barreiro Instituto de F´ısica de Cantabria, CSIC-Universidad de Cantabria, avenida de los Castros s/n, 39905 Santander, Spain Email: [email protected]
Received 8 June 2004; Revised 7 February 2005
This paper considers the detection of point sources in two-dimensional astronomical images. The detection scheme we propose is based on peak statistics. We discuss the example of the detection of far galaxies in cosmic microwave background experiments throughout the paper, although the method we present is totally general and can be used in many other fields of data analysis. We consider sources with a Gaussian profile—that is, a fair approximation of the profile of a point source convolved with the detector beam in microwave experiments—on a background modeled by a homogeneous and isotropic Gaussian random field characterized by a scale-free power spectrum. Point sources are enhanced with respect to the background by means of linear filters. After filtering, we identify local maxima and apply our detection scheme, a Neyman-Pearson detector that defines our region of acceptance based on the a priori pdf of the sources and the ratio of number densities. We study the different performances of some linear filters that have been used in this context in the literature: the Mexican hat wavelet, the matched filter, and the scale-adaptive filter. We consider as well an extension to two dimensions of the biparametric scale-adaptive filter (BSAF). The BSAF depends on two parameters which are determined by maximizing the number density of real detections while fixing the number density of spurious detections. For our detection criterion the BSAF outperforms the other filters in the interesting case of white noise. Keywords and phrases: analytical methods, data analysis methods, image processing techniques.
1. INTRODUCTION The CMB is the remnant of the radiation that filled the universe immediately after the big bang. This weak radiation A very challenging aspect of data analysis in astronomy is the can provide us with answers to one of the most important detection of pointlike sources embedded in one- and two- set of questions asked in modern science—how the universe dimensional images. Some common examples are the sep- did begin, how it evolved to the state we observe today, and aration of individual stars in crowded optical images, the how it will continue to evolve in the future. Unfortunately, we identification of emission and absorption lines in noisy one- do not measure the CMB alone but a mixture of it with in- dimensional spectra, and the detection of faint extragalactic strumental noise and other astrophysical radiations that are objects at microwave frequencies. This latter case, for exam- usually referred to as foregrounds. ple, is one of the most critical issues for the new generation of Some foregrounds are due to our own galaxy, for exam- experiments that observe the cosmic microwave background ple, the thermal emission due to dust grains in the galactic (CMB). plane or the synchrotron emission by relativistic electrons Detection of Point Sources 2427 moving along the galactic magnetic field. These foregrounds ff appear as di use emission in the sky, and their spectral be- 5 haviors (the way the emission scales from one wavelength of observation to another) are reasonably well known. An- 200 other foreground with a well-known spectral behavior is the Sunyaev-Zel’dovich effect, which is due to the hot gas con- tained in galaxy clusters that distorts the energy distribution 100 of CMB photons. Foreground emissions carry information 0 about the galaxy structure, composition, and physical pa- (degrees) y rameters as well as about the number, distribution, and evo- lution of galaxy clusters that map the distribution of mat- 0 ter in the universe. Therefore, the study of the different fore- grounds has great scientific relevance by itself. In order to properly study the CMB and the different foregrounds, it −5 is mandatory to separate the signals (components) that are −100 mixed in the observations. This can be done by observing −5 0 5 the sky at a number of frequencies at least as big as the num- x (degrees) ber of components and then applying some statistical compo- nent separation method in order to recover the different as- Figure 1: Residual map of a 12.8 × 12.8 square degrees sky patch trophysical signals. Several component separation techniques at 30 GHz after the application of a maximum entropy compo- have been suggested, including blind (Baccigalupi et al. [1], nent separation. The residual map is obtained by subtracting from the 30 GHz map the different components (CMB and foregrounds) Maino et al. [2], Delabrouille et al. [3]), semi-blind (Bedini et given by the maximum entropy algorithm. Bright point sources ap- al. [4]) and nonblind (Hobson et al. [5], Bouchet and Gispert pear as spots in the images whereas faint point sources are masked [6], Stolyarov et al. [7], Barreiro et al. [8]) approaches. by the residual noise. Another important foreground is due to the emission of far galaxies. Since the typical angular size of the galaxies in the sky is a few arcseconds and the angular resolution of the lar power spectrum and hampering the statistical study (e.g., microwave detectors is typically greater than a few arcmin- the study of Gaussianity) of CMB and other foregrounds at utes,1 galaxies appear as points to the detector, which is un- such scales. Moreover, while there are good galaxy surveys at able to resolve their inner structure. Therefore, they are usu- radio and infrared frequencies, the microwave window of the ally referred to as extragalactic point sources (EPS) in the CMB electromagnetic spectrum is a practically unknown zone for jargon. Note that, however, they do not appear as points in extragalactic astronomy. Therefore, it is important to have the images but as the convolution of a pointlike impulse with detection techniques that are able to detect EPS with fluxes the angular response of the detector (beam). The instruments as low as possible. (radiometers and bolometers) that are used in CMB experi- One possibility is to consider the EPS emission at each ments have angular responses that are approximately Gaus- frequency as an additional noise to be considered in the equa- sian and therefore EPS appear as small Gaussian (or nearly tions of a statistical component separation method. Once the Gaussian) spots in the images.2 algorithm has separated the different components, the resid- The problem with EPS is that galaxies are a very hetero- ual that is obtained by subtracting the output foregrounds geneous bundle of objects, from the radio galaxies that emit from the original data should contain the EPS plus the in- most of their radiation in the low-frequency part of the elec- strumental noise and some amount of foreground residuals tromagnetic spectrum to the dusty galaxies that emit mainly that remain due to a nonperfect separation. As an example, in the infrared (Toffolatti et al. [9], Guiderdoni et al. [10], Figure 1 shows the residual at 30 GHz after applying a max- Tucci et al. [11]). This makes it impossible to consider all of imum entropy component separation algorithm (Hobson et them as a single foreground to be separated from the other by al. [12]) to a 12.8×12.8 square degrees simulated sky patch as means of multiwavelength observations and statistical com- would be observed by the Planck satellite. The brightest point ponent separation techniques. EPS constitute an important sources can be clearly observed over the residual noise. How- contaminant in CMB studies at small angular scales (Toffo- ever, fainter point sources are still masked by a residual noise latti et al. [9]), affecting the determination of the CMB angu- that is approximately Gaussian and must be detected some- how. Besides, the situation is more complex because the pres- ence of bright EPS in the data affects the performance of the 1For example, the upcoming ESA’s Planck satellite will have angular res- component separation algorithms so the recovered compo- olutions ranging from 5 arcminutes (for the 217–857 GHz channels) to 33 nents are contaminated by point sources in a way that is dif- arcminutes (for the 30 GHz channel). ficult to control. Therefore, any satisfactory method should 2 It is also common to speak of compact sources, describing a source that is detect and extract at least the bright sources before the com- comparable to the size of the beam being used. Non-pointlike sources (such as large galaxy clusters with arcminute angular scales) will have more com- ponent separation. Then, after separation some additional plicated responses when convolved with a beam, but if the source profile is low intensity EPS could be detected from the residual maps known, it is always possible to apply the methods presented in this work. such as the one in Figure 1. 2428 EURASIP Journal on Applied Signal Processing
Several techniques based on linear filters have been pro- In Section 4 we briefly review some of the linear filters pro- posed in the literature for the detection of point sources in posed in the literature. In Section 5 we describe a probability CMB data. Linear filtering techniques are suitable for this distribution of sources that is of interest and compare the problem because they can isolate structures with a given performance of the filters, regarding our choice of detector. characteristic scale, as is the case of pointlike sources, while Finally, in Section 6 we summarize our results. canceling the contribution of diffuse foregrounds. Among the methods proposed in the literature, we emphasize the Mexican hat wavelet (MHW, Cayon´ et al. [13], Vielva et al. 2. PEAK STATISTICS [14, 15, 16]), the classic matched filter (MF, Tegmark and de In this section we will study the statistics of peaks for a two- Oliveira-Costa [17]), the adaptive top hat filter (Chiang et al. dimensional Gaussian background in both the absence and [18]), and the scale-adaptive filter (SAF, Sanz et al. [19], Her- presence of a source. We will focus on three quantities that ranz et al. [20]). Moreover, linear filters can be used in com- define the properties of the peaks: the intensity of the field, bination with statistical component separation techniques in the curvature, and the shear at the position of the peak. The order to produce a more accurate separation of the different first quantity gives the amplitude of the peak. The curvature foregrounds (Vielva et al. [15]). and the shear give information about the spatial structure of The goal of filtering is to enhance the contrast between the peak and are related to its sharpness and eccentricity, re- the source to be detected and the background that masks it. spectively. For example, if we filter the image in Figure 1, assuming that the background can be characterized by a white noise, with 2.1. Background the well-known matched filter (see Section 4.1) at the scale of the 30 GHz detector beam (FWHM = 33 arcminutes) the We consider a two-dimensional (2D) background repre- signal-to-noise ratio of the sources increases by more than sented by a Gaussian random field ξ(x )withaveragevalue 25%. Therefore, a source whose signal-to-noise ratio was ∼ 3 ξ(x )=0andpowerspectrumP(q), before filtering becomes a source with signal-to-noise ratio ∼ 4 and will be easier to detect. ∗ 2 After filtering, a detection rule is applied to the data in ξ(Q)ξ (Q ) = P(q)δD(Q − Q ), q ≡|Q|,(1) order to decide whether the source is present or not. The usual detection approach in astronomy is thresholding:for 3 2 where ξ(Q) is the Fourier transform of ξ(x ) and δD is the any given candidate (e.g., a local peak in the data), a posi- Dirac distribution in 2D. tive detection is considered if the candidate has a signal-to- We are interested in the distribution of maxima of the noise ratio greater than a certain threshold (in many astro- background with respect to the three variables already men- nomical applications, a typical value of this threshold is 5σ). tioned: intensity, curvature, and shear. We define the normal- This naive approach works fine for bright sources, but weak ized field intensity ν, the normalized curvature κ, and the nor- sources can be easily missed. malized shear as More sophisticated detection schemes can use additional information in order to improve the detection. If the detec- ξ λ λ λ − λ tion is performed by means of the study of the statistics of ν ≡ κ ≡ 1 + 2 ≡ 1 2 σ , σ , σ ,(2) maxima in the images, such information includes not only 0 2 2 2 the amplitude of the maxima but also spatial information re- lated to the source profile, for example, the derivatives of the where ν ∈ (−∞, ∞), κ ∈ [0, ∞), ∈ [0, κ/2), λ1 and λ2 are intensity. In our approach we will consider the amplitude, the the eigenvalues of the negative Hessian matrix, and the σn are curvature, and the shear of the sources (the last two quanti- defined as ties are given by the properties of the beam in the case of point sources) to discriminate between maxima of the back- ∞ 2 1 1+2n ground and real sources. Moreover, in some cases a priori σn ≡ dq q P(q). (3) 2π information on the distribution of intensity of the sources 0 is known. We will therefore use a Neyman-Pearson detec- σ tor that uses the three above-mentioned elements of infor- The moment 0 is equal to the dispersion of the field. mation (amplitude, curvature, and shear) of the maxima as The distribution of maxima of the background in one well as the a priori probability distribution of the sources. dimension (1D) with respect to the intensity and curvature This technique has been successfully tested in images of one- (the shear is not defined in 1D) was studied by Rice [23]. dimensional fields (Lopez-Caniego´ et al. [21, 22]). In this If we generalize it to 2D, including the shear, the expected work we will generalize it to two dimensions. The overview of this work is as follows. In Section 2 we 3 describe the statistics of the peaks for a two-dimensional Throughout this paper we will use the following notation for the Fourier transform: the same symbol will be used for the real space and the Gaussian background in the absence and presence of a Fourier space versions of a given function. The argument of the function will source. In Section 3 we introduce the detection problem, specify in each case which is the space we are referring to. For instance, f (q) define the region of acceptance, and derive our detector. will be the Fourier transform of the function f (x). Detection of Point Sources 2429 number density of maxima per intervals (x,x + dx ), (ν, ν + 3. THE DETECTION PROBLEM dν κ κ dκ d ), ( , + ), and ( , + )isgivenby Equations (4)and(8) can be used to decide whether a source √ ispresentornotinadataset.Thetoolthatallowsustode- 8 3nb − / ν2− 2− κ−ρν 2/ −ρ2 nb(ν, κ, ) = κ2 − 42 e (1 2) 4 ( ) 2(1 ), cide whether a point source is present or not in the data is π 1 − ρ2 called a detector. In this section we will describe the Neyman- (4) Pearson detector (NPD). We will study its performance in where nb is the expected total number density of maxima terms of two quantities: the number of true detections and (i.e., number of maxima per unit area dx ), the number of false (spurious) detections that emerge from the detection process. Our approach fixes the number density 1 nb ≡ √ ,(5)of spurious detections and determines the number density of π θ2 4 3 m true detections in each case. ρ θ and and m are defined as 3.1. The region of acceptance √ σ σ2 θ √ σ We consider a peak in the 2D dataset characterized by the 1 1 m 0 θm ≡ 2 , ρ ≡ = , θc ≡ 2 . (6) normalized amplitude, curvature, and shear (ν, κ, ). The σ2 σ0σ2 θc σ1 number density of background maxima nb(ν, κ, ) represents In the previous equations θc and θm are the coherence scale the null hypothesis H0 that the peak is due to the background of the field and maxima, respectively. The formula in (4)can in the absence of source. Conversely, the local number den- be derived from previous works (Bond and Efstathiou [24], sity of maxima n(ν, κ, ) represents the alternative hypoth- Barreiro et al. [25]). esis, that the peak is due to the source added to the back- ground. The local number density of maxima n(ν, κ, )can 2.2. Background plus point source be calculated as To the previous 2D background we add a source with a ∞ known spatial profile τ(x ) and an amplitude A, so that the n(ν, κ, ) = dνs p νs n ν, κ, |νs . (10) 0 intensity due to the source at a given position x0 is ξs(x ) = Aτ x − x ( 0). For simplicity, we will consider a spherical Gaus- In the last equation we have used the a priori probability sian profile given by p(νs) that gives the amplitude distribution of the sources. R ν κ ν κ 2 We can associate to any region ∗( , , ) in the ( , , ) x ∗ ∗ τ(x) = exp − , x ≡|x|,(7)parameterspacetwonumberdensitiesnb and n , 2R2 R ∗ where is the Gaussian width (in the case of point sources nb = dν dκd nb(ν, κ, ), R R∗ convolved with a Gaussian beam, is the beam width). We (11) 4 could easily consider other functional profiles without any n∗ = dν dκd n(ν, κ, ), loss of generality. The expected number density of maxima R∗ per intervals (x,x+dx ), (ν, ν+dν), (κ, κ+dκ), and (, +d), ∗ given a source of amplitude A in such spatial interval, is where nb is the expected number density of spurious sources, R ν κ that is, due to the background, in the region ∗( , , ), n∗ n ν, κ, |νs whereas is the number density of maxima expected in the √ ν κ same region of the ( , , ) space in the presence of a local 8 3nb − / ν−ν 2− 2− κ− κ −ρ ν−ν 2/ −ρ2 = κ2 −42 e (1 2)( s) 4 ( 2 s ( s)) 2(1 ), source. The region R∗ will be called the region of acceptance. π −ρ2 1 In order to define the region of acceptance R∗ that gives (8) the highest number density of detections n∗ for a given ∗ number density of spurious detections nb ,weconsidera where νs = A/σ0 is the normalized amplitude of the source, Neyman-Pearson detector (NPD) using number densities in- κs =−Aτ /σ2 is the normalized curvature of the source, and steadofprobabilities τ is the second derivative of the source profile τ with respect to x at the position of the source. Note that in (8)weare n(ν, κ, ) L(ν, κ, ) ≡ ≥ L∗, (12) taking into account that the shear of the source is zero since nb(ν, κ, ) we are considering a spherical profile. It is useful to define a L quantity ys that is related to the curvature of the source: where ∗ is a constant. The proof follows the same approach as for the standard Neyman-Pearson detector. If L ≥ L∗ we 2 θmτ νs ys decide that the signal is present, whereas if L where ϕ∗ is a constant and ϕ(ν, κ)isgivenby the background is statistically homogeneous and isotropic, we will consider spherically symmetric filters, 1 − ρys ys − ρ ϕ(ν, κ) ≡ aν + bκ, a ≡ , b ≡ . (14) 1 − ρ2 1 − ρ2 1 |x − b| Ψ(x; R, b) ≡ ψ . (18) R2 R We remark that the detector is independent of the shear . This is due to the fact that we are considering a source with a spherical profile with shear s = 0. If the profile is not spher- If we filter our background with Ψ(x; R, b), the filtered field ical, the detector may depend on the shear. is 3.2. Spurious sources and real detections w(R, b) = dxξ(x )Ψ(x; R, b). (19) Given a region of acceptance R∗, we can calculate the num- ber density of spurious sources and the number density of detections as given by (11): The filter is normalized such that the amplitude of the source is the same after filtering: √ n ∞ n∗ = √3 b dκ κ2 − e−κ2 e−κ2/2 M b π 1+ erfc( ), 2 0 dxτ(x )Ψ(x; R,0) = 1. (20) (15) ϕ∗ − ysκ M ≡ ; a − ρ2 2 1 For the filtered field, (3)becomes √ ∞ ∗ 3nb n = √ dνs p νs ∞ π 2 1+2n 2 2 0 σn ≡ 2π dq q P(q)ψ (q). (21) (16) ∞ 0 × dκ κ2 − 1+e−κ2 e−(1/2)(κ−νs ys)2 erfc(Q), 0 Thevaluesofρ, θm, θc, and all the derived quantities change ρys − 1 accordingly. The curvature of the filtered source κs can be Q ≡ M + νs . (17) 2 1 − ρ2 obtained through (9), taking into account that for the filtered source, Our approach is to fix the number density of spurious de- ∞ tections and then to determine the region of acceptance that 3 −τψ = π dq q τ(q)ψ(q). (22) gives the maximum number of true detections. This can be 0 ∗ done by inverting (15)toobtainϕ∗ = ϕ∗(nb /nb; ρ, ys). Once ϕ∗ is known, we can calculate the number density of detec- Note that the function ψ(q) will depend as well on the scal- tions using (16). ing R. As an application of the previous ideas, we study the detection of point sources characterized by a Gaussian pro- file τ(x) = exp(−x2/2R2), x =|x|, and Fourier transform 4. THE FILTERS τ(q) = R2 exp(−(qR)2/2). This is the case we find in CMB Detection can, in principle, be performed on the raw data, experiments, where the profile of the point source is given by but in most cases it is convenient to transform first the data the instrumental beam that can be approximated by a Gaus- in order to enhance the contrast between the distributions sian. nb(ν, κ, )andn(ν, κ, ). Hopefully, such an enhancement This profile introduces in a natural way the scale of the R will help the detector to give better results (namely, a higher source , the scale at which we filter. However, previous number of true detections). In this paper we will focus on the works in 1D using the MHW, MF, SAF, and BSAF have shown αR use of linear filters as a means to transform the data in such a that the use of a modified scale can significantly improve way. Filters are suitable for this task because background fluc- the number of detections (Cayon´ et al. [13], Vielva et al. tuations that have variation scales different from the source [14, 15], Lopez-Caniego´ et al. [21, 22]). Therefore, we gener- scale can be easily filtered out while preserving the sources. alize the functional form of these filters to 2D and allow for α Different filters will improve detection in different ways: this this additional degree of freedom . paper compares the performance of several filters. The fil- ter that gives the highest number density of detections, for 4.1. The matched filter a fixed number density of spurious sources, will be the pre- We introduce a circularly symmetric filter Ψ(x; R, b). The fil- ferred filter among the considered filters. tered field is given by (19). Now, we express the conditions We consider a filter Ψ(x; R, b), where R and b define a to obtain a filter for the detection of the source s(x) = Aτ(x) scaling and a translation respectively. Since the sources we are at the origin taking into account the fact that the source is considering are spherically symmetric and we assume that characterized by a single scale R0. We assume the following Detection of Point Sources 2431 conditions: where (1) w(R0,0)=s(0) ≡ A, that is, w(R0, 0) is an unbiased α2 1 1 estimator of the amplitude of the source; N(α) = , (29) ∆m πΓ(m) γ +(2t/m)∆ (2) the variance of w(R, b) has a minimum at the scale R0, that is, it is an efficient estimator. and where m and ∆ are defined as in (24), t is defined as in Then, the 2D filter satisfying these conditions is the so-called (27). The parameters of the filtered background and source matched filter.Asmentionedbefore,wewillallowthefilter are scale to be modified by a factor α.Ifα = 1 we have the well- known standard matched filter use in the literature. For a m H1 2 H1 ρ(α)=ρ= , θm(α) = αR , source with a Gaussian profile, a scale-free power spectrum 1+m H H 1+m H P q ∝ q−γ 2 3 3 ( ) , and allowing the filter scale to vary through the α parameter, the modified matched filter is m H2 γ + c(1 + m)∆ ys(α) = ∆ , 1+m H3 γ + cm∆ γ −(1/2)z2 (30) ψMF(q) = N(α)z e , z ≡ qαR, (23) where c = 2t/m and where 2 2 H1 = γ +2γc(1 + m)+c (1 + m)(2 + m), 2+γ α2 1 1 2α2 m ≡ , N(α) = , ∆ = , (24) H = γ2 γcm c2m m 2 ∆m π Γ(m) (1 + α2) 2 +2 + (1 + ), (31) 2 2 H3 = γ +2γc(2 + m)+c (2 + m)(3 + m). and Γ is the standard Gamma function. The parameters of the filtered background and source are The corresponding threshold as compared to the stan- dard matched filter (α = 1) is m 2 t− m ρ α =ρ= θ α =αR y α =ρ∆. ν α 2∆ (γ + cm∆) ( ) m, m( ) m, s( ) = . 1+ 1+ ν H (32) (25) MF(α=1) 2 The corresponding threshold as compared to the stan- 4.3. The Mexican hat wavelet dard matched filter (α = 1) is The MHW is defined to be proportional to the Laplacian of ν(α) the Gaussian function in 2D real space = αt−2∆m, (26) ν α= MF( 1) 2 −(1/2)x2 ψMHW(x) ∝ 1 − x e , x ≡|x|. (33) where Thus, in Fourier space we get the modified Mexican hat 2 − γ wavelet introducing the α parameter as follows: t ≡ . (27) 2 2 −(1/2)z2 ψMHW(q) = N(α)z e , z ≡ qαR, We remark that for the standard matched filter the cur- ff 1 α 2 (34) vature does not a ect the region of acceptance and the linear N(α) = . detector ϕ(ν, κ) is reduced to ϕ = ν. π ∆ 4.2. The scale-adaptive filter Thus, the filtered background and source parameters are The scale-adaptive filter (or optimal pseudo-filter) has been proposed by Sanz et al. [19]. The filter is obtained by impos- 2+t 2 ρ(α) = ρ = , θm(α) = αR , ing an additional condition to the conditions that define the 3+t 3+t MF: (35) 2 w R R ys(α) = ∆, (3) ( , 0) has a maximum at ( 0, 0). (2 + t)(3 + t) −γ Considering a scale-free power spectrum, P(q) ∝ q ,a m ∆ t αR where and are defined as in (24)and is defined as in modified scale , and a Gaussian profile for the source, the (27). The corresponding threshold is functional form of the filter in 2D is t ν α αt−2∆2 γ −(1/2)z2 2 2 ( ) = . ψSAF(q) = N(α)z e γ + z , z ≡ qαR, (28) (36) m νMF(α=1) Γ(m)Γ(2 + t) 2432 EURASIP Journal on Applied Signal Processing Table 1: Number density of detections n∗ for the BSAF and the standard MF (α = 1) with optimal values of c and α for different values of n∗ R ff ≡ − n∗ /n∗ b and . RD means relative di erence in number densities in percentage: RD 100( 1+ BSAF MF). R n∗ αcn∗ n∗ b BSAF MF RD(%) 0.005 0.5 −0.44 0.0507 0.0484 4.7 1.5 0.01 0.5 −0.46 0.0709 0.0620 14.3 0.005 0.4 −0.54 0.0396 0.0335 18.2 2 0.01 0.4 −0.54 0.0567 0.0406 39.6 2.5 0.005 0.3 −0.64 0.0320 0.0245 30.6 4.4. The biparametric scale-adaptive filter 5. ANALYTICAL RESULTS Lopez-Caniego´ et al. [21] have shown that removing condi- In this section we will compare the performance of the dif- tion (3) defining the SAF and introducing instead the condi- ferent filters previously introduced. We use as an example the tion interesting case of white noise as background. This is a fair approximation to the case presented in Figure 1, where the (3) w(R , b) has a maximum at (R ,0) 0 0 sources are embedded in a background that is a combination leads to the new filter of instrumental noise (approximately Gaussian) and a small contribution of residual foregrounds that have not been per- τ(q) fectly separated. For this example, we will consider sources ψ(q) ∝ 1+c(qR)2 , (37) P(q) with intensities distributed uniformly between zero and an upper cutoff. c The comparison of the filters is performed as follows. We where is an arbitrary constant related to the curvature of fix the number density of spurious detections, the same for the maximum. For the case of a scale-free spectrum, and al- all the filters. Then, for any given filter we calculate the quan- lowing for a modified scale αR, the filter is given by the pa- tities σn, ρ,andys. Using (15) it is possible to calculate the rameterized equation value of ϕ∗ that defines the region of acceptance. Then we calculate the number density of real detections using (16). γ −(1/2)z2 2 ψBSAF(q) = N(α)z e 1+cz , z ≡ qαR, The filter that leads to the highest number density of detec- tions will be the preferred one. We do this for different values α2 (38) N α = 1 1 1 of α in order to test how the variation of the filtering scale af- ( ) m , ∆ π Γ(m) 1+cm∆ fects the number of detections. where m and ∆ aredefinedasin(24). We remark that c = 0 5.1. A priori probability distribution c ≡ t/mγ t leads to the MF, and if 2 ,with defined as in (27), As mentioned before, we will test a pdf of source intensities the BSAF becomes the SAF. The parameters of the filtered that is uniform in the interval 0 ≤ A ≤ Ac.Intermsofnor- background and source are malized intensities, we have the pdf 1 m D1 2 D1 p ν = ν ∈ ν . ρ α =ρ = θm α = αR s , s 0, c (42) ( ) , ( ) , νc 1+m D2 D3 1+m D3 ff m D2 1+c(1 + m)∆ We will consider a cuto in the amplitude of the sources such ys(α) = ∆ , ν = 1+m D 1+cm∆ that c 2 after filtering with the standard MF, that is, we will 3 ffi (39) focus on the case of faint sources that would be very di cult to detect if no filtering was applied. Note that while the value νc is different for each filter (because each filter leads to a where different dispersion σ0 of the filtered field), the distribution A 2 in source intensities is the same for all the cases. D1 = 1+2c(1 + m)+c (1 + m)(2 + m), 2 5.2. Results for white noise D2 = 1+2cm + c m(1 + m), (40) We want to find the optimal filter in the sense of the max- D = 1+2c(2 + m)+c2(2 + m)(3 + m). 3 imum number of detections. For the sources, we use a uniform distribution with amplitudes in the interval A ∈ The equivalent threshold is given by [0, 2]σ0,whereσ0 is the dispersion of the linearly filtered map with the standard MF. We focus on the interesting case of t− m ν(α) α 2∆ (γ + cm∆) white noise (γ = 0) and explore different values of n∗ and R. = . (41) b νMF(α=1) D3 The results are summarised in Table 1. Detection of Point Sources 2433 0.06 0.08 0.06 n∗ 0.04 n∗ 0.04 0.02 0.02 0.4 0.6 0.8 1 1.2 1.4 . . . . . 0 4 0 6 0 8 1 1 2 1 4 α α MF MF SAF SAF BSAF BSAF Figure 2: The expected number density of detections n∗ as a func- Figure 3: The expected number density of detections n∗ as a func- tion of α for γ = 0fortheBSAF(c has been obtained by maximizing tion of α for γ = 0fortheBSAF(c has been obtained by maximising the number of detections for each value of α), MF, and SAF filters. the number of detections for each value of α), MF, and SAF filters. ∗ R = n∗ = . We consider the case R = 1.5, nb = 0.01. We consider the case 2, b 0 01. We study the performance of the different filters as a follow what would be found in a real image. Therefore, we function of α. This allows us to test how the variation of the present the results only for those values of α such that αR is natural scale of the filters helps the detection. In the case of at least ∼ 1. the BSAF, which has an additional free parameter, c in (38), for each value of α we determine numerically the value of c 6. CONCLUSIONS that gives the highest number of detections. Then the BSAF ∗ with such c parameter (i.e., a function of α, nb ,andR)is Severaltechniqueshavebeenintroducedintheliterature compared with the other filters. to detect point sources in two-dimensional images. Exam- In Figure 2, we plot the expected number density of de- ples of point sources in astronomy are far galaxies as de- tections n∗ for different values of α, R = 1.5 pixels, and tected by CMB experiments. An approach that has been thor- ∗ nb = 0.01. Note that for the 2D case the MHW and SAF are oughly used in the literature for this case consists in linear the same filter for γ = 0, and we have only included the latter filtering the data and applying detectors based on thresh- in our figures. In this case, the curve for the BSAF always goes olding. Such approach uses only information on the ampli- above the other filters. The maximum number of detections tude of the sources: the potentially useful information con- is found for small values of α. In this region, the improve- tained in the local spatial structure of the peaks is not used ment of the BSAF with respect to the standard matched filter at all. In our work, we design a detector based on peak is of order 15%. statistics that uses the information contained in the am- In Figure 3, we show the results for R = 2. We have in- plitude, curvature, and shear of the maxima. These quan- creased the beam width as compared to the previous example tities describe the local properties of the maxima and are and left unchanged the number density of false detections. used to distinguish statistically between peaks due to back- The BSAF outperforms all the other filters, and for small val- ground fluctuations and peaks due to the presence of a ues of α the improvement is of order 40%. Note that in source. this figure the MF takes values α ∈ [0, 1]. For greater values We derive a Neyman-Pearson detector (NPD) that con- ∗ of α,withR = 2andnb = 0.01, we cannot solve for ϕ∗ in the siders number densities of peaks which leads to a sufficient implicit equation (15) and cannot calculate n∗. detector that, in the case of the spherically symmetric sources We remark that filtering at scales much smaller than the that we consider, is linear in the amplitude and curvature of scale of the pixel does not make sense. This is due to the fact the sources. For this particular case, then, the information of that we are not including the effect of the pixel in our the- the shear of the peaks is not relevant. In other cases, however, oretical calculations and, thus, the results would not exactly it could be useful. 2434 EURASIP Journal on Applied Signal Processing It is a common practice in astronomy to linear filter the The criterion for detection can be written as images in order to enhance very faint point sources and help ∞ the detection. The best filter would be the one that makes it Ł(ν, κ) ≡ dνs p νs L ν, κ|νs ≥ L∗,(A.2) easier to distinguish between peaks coming from the back- 0 ground alone and those due to the presence of a source, ac- where L∗ is a constant. L is a function of ϕ, cording to the information used by the detector. In the case of 1 − ρys ys − ρ simple thresholding, which considers only the amplitude of ϕ(ν, κ) ≡ aν + bκ, a = , b = . (A.3) the peaks, the answer to the question of which is the best filter 1 − ρ2 1 − ρ2 (in the previous sense) is well known: the standard matched By differentiating L with respect to ϕ we find that filter. But in the case of the Neyman-Pearson detector, which ∞ considers other things apart from mere amplitudes, this is no ∂L ϕν − / ν2 ρν − κ 2 = dνs p νs νse s (1 2)( s +( s 2 s) ) ≥ 0, (A.4) longer true. ∂ϕ 0 We have compared three commonly used filters in the lit- erature in order to assess which one of them performs bet- and therefore setting a threshold in L is equivalent to setting ter when detecting sources with our scheme. In addition, we a threshold in ϕ: have designed a filter such that it optimizes the number of ν κ ≥ L ⇐⇒ ϕ ν κ ≥ ϕ true detections for a fixed number of spurious sources. The Ł( , ) ∗ ( , ) ∗,(A.5) optimization of the number of true detections is performed where ϕ(ν, κ)isgivenby(A.3)andϕ∗ is a constant. by using the a priori pdf of the amplitudes of the sources. This filter depends on two free parameters and it is therefore called biparametric scale-adaptive filter (BSAF). By construc- ACKNOWLEDGMENTS tion, the functional form of the BSAF includes the standard The authors thank Patricio Vielva for useful discussions. MF as a particular case and its performance in terms of num- Lopez-Caniego´ thanks the Ministerio de Ciencia y Tec- beroftruedetectionsforafixednumberofspuriousdetec- nolog´ıa (MCYT) for a predoctoral FPI fellowship. Barreiro tions must be at least as good as the standard MF’s one. thanks the MCYT and the Universidad de Cantabria for Following the work done in the 1D case, we generalize the aRamon´ y Cajal contract. Herranz acknowledges support functional form of the filters to 2D and introduce an extra from the European Community’s Human Potential Pro- degree of freedom α thatwillallowustofilteratdifferent gramme under contract HPRN-CT-2000-00124, CMBNET, scales αR,whereR is the scale of the source. This significantly and from an ISTI fellowship since September 2004. We ac- improves the results. knowledge partial support from the Spanish MCYT project We have considered an interesting case, a uniform distri- ESP2002-04141-C03-01 and from the EU Research Train- bution of weak sources with amplitudes A ∈ [0, 2]σ0,where ing Network “Cosmic Microwave Background in Europe for σ0 is the dispersion of the field filtered with the standard Theory and Data Analysis.” matched filter, embedded in white noise (γ = 0). We have tested different values of the source size R and of the number REFERENCES n∗ density of spurious detections b that we fix. We find that [1] C. Baccigalupi, L. Bedini, C. Burigana, et al., “Neural net- the BSAF improves the number density of detections up to works and separation of cosmic microwave background and 40% with respect to the standard MF (α = 1) for certain astrophysical signals in sky maps,” Monthly Notices of the Royal cases. Note that since the Neyman-Pearson detector for the Astronomical Society, vol. 318, no. 3, pp. 769–780, 2000. standard MF (α = 1) defaults to the classic thresholding de- [2] D. Maino, A. Farusi, C. Baccigalupi, et al., “All-sky astrophys- tector that is commonly used in astronomy, the results of this ical component separation with fast independent component analysis (FASTICA),” Monthly Notices of the Royal Astronomi- work imply that it is possible, under certain circumstances, to cal Society, vol. 334, no. 1, pp. 53–68, 2002. detect more point sources than in the classical approach. [3] J. Delabrouille, J. F. Cardoso, and G. Patanchon, “Multidetec- The generalization of these ideas to other source profiles tor multicomponent spectral matching and applications for and non-Gaussian backgrounds is relevant and will be dis- cosmic microwave background data analysis,” Monthly Notices cussed in a future work. of the Royal Astronomical Society, vol. 346, no. 4, pp. 1089– 1102, 2003. [4] L. Bedini, D. Herranz, E. Salerno, C. Baccigalupi, E. E. APPENDIX Kuruoglu,˘ and A. Tonazzini, “Separation of correlated as- trophysical sources using multiple-lag covariance matrices,” We will show in this appendix that ϕ(ν, κ) ≥ ϕ∗ given in (13) EURASIP Journal on Applied Signal Proccessing, vol. 2005, is a sufficient linear detector, that is, the detector is linear in no. 15, pp. 2400–2412, 2005, Special issue on applications of ν κ signal processing in astrophysics and cosmology. the threshold and the curvature and the data it uses is a [5] M. P. Hobson, A. W. Jones, A. N. Lasenby, and F. R. Bouchet, sufficient statistic to decide if a peak is a source (independent “Foreground separation methods for satellite observations of of the a priori probability P(νs)). The ratio L(ν, κ, |νs) ≡ the cosmic microwave background,” Monthly Notices of the n(ν, κ, |νs)/nb(ν, κ, ) can be explicitly written as Royal Astronomical Society, vol. 300, no. 1, pp. 1–29, 1998. [6] F. R. Bouchet and R. Gispert, “Foregrounds and CMB experi- ments I. Semi-analytical estimates of contamination,” New As- ϕν − / ν2 ρν − κ 2 L ν, κ, |νs = e s (1 2)( s +( s 2 s) ). (A.1) tronomy, vol. 4, no. 6, pp. 443–479, 1999. Detection of Point Sources 2435 [7] V. Stolyarov, M. P. Hobson, M. A. J. Ashdown, and A. the Neyman-Pearson detector,” Monthly Notices of the Royal N. Lasenby, “All-sky component separation for the Planck Astronomical Society, vol. 359, pp. 993–1006, 2005. mission,” Monthly Notice of the Royal Astronomical Society, [23] S. O. Rice, “Mathematical analyses of random noise,” in Se- vol. 336, no. 1, pp. 97–111, 2002. lected Papers on Noise and Stochastic Processes, N. Wax, Ed., [8]R.B.Barreiro,M.P.Hobson,A.J.Banday,etal.,“Foreground Dover Publications, New York, NY, USA, 1954. separation using a flexible maximum-entropy algorithm: an [24] J. R. Bond and G. Efstathiou, “The statistics of cosmic back- application to COBE data,” Monthly Notices of the Royal As- ground radiation fluctuations,” Monthly Notices of the Royal tronomical Society, vol. 351, no. 2, pp. 515–540, 2004. Astronomical Society, vol. 226, pp. 655–687, 1987. [9] L. Toffolatti, F. Argueso,¨ G. De Zotti, et al., “Extragalactic [25]R.B.Barreiro,J.L.Sanz,E.Mart´ınez-Gonzalez,´ L. Cayon,´ and source counts and contributions to the anisotropies of the J. Silk, “Peaks in the cosmic microwave background: flat ver- cosmic microwave background: predictions for the Planck sus open models,” The Astrophysical Journal, vol. 478, part 1, Surveyor mission,” Monthly Notices of the Royal Astronomical no. 1, pp. 1–6, 1997. Society, vol. 297, no. 1, pp. 117–127, 1998. [10] B. Guiderdoni, E. Hivon, F. R. Bouchet, and B. Maffei, “Semi- analytic modelling of galaxy evolution in the IR/submm M. Lopez-Caniego´ received his M.S. degree range,” Monthly Notices of the Royal Astronomical Society, in physics from the Universidad Autonoma vol. 295, no. 4, pp. 877–898, 1998. de Madrid, Madrid, Spain, in 2000, after [11] M. Tucci, E. Mart´ınez-Gonzalez,´ L. Toffolatti, J. Gonzalez-´ completing as part of his degree one year Nuevo, and G. De Zotti, “Predictions on the high-frequency at the University of Frankfurt, Germany. polarization properties of extragalactic radio sources and im- He was a Research Fellow at Bell Labs, Lu- plications for polarization measurements of the cosmic mi- cent Technologies, New Jersey, USA, dur- crowave background,” Monthly Notices of the Royal Astronom- ing 2001, before he continued his postgrad- ical Society, vol. 349, no. 4, pp. 1267–1277, 2004. uate studies at the Universidad Autonoma ff ff [12]M.P.Hobson,R.B.Barreiro,L.To olatti, et al., “The e ect de Madrid in 2002. At the end of this year of point sources on satellite observations of the cosmic mi- he joined a research group at the Instituto de Fisica de Cantabria crowave background,” Monthly Notices of the Royal Astronom- (CSIC-UC), where he is currently pursuing his Ph.D. in the field of ical Society, vol. 306, no. 1, pp. 232–246, 1999. image processing under an MCYT FPI predoctoral fellowship. His [13] L. Cayon,J.L.Sanz,R.B.Barreiro,etal.,“Isotropicwavelets:a´ research interests are in the areas of gravitational lensing and cos- powerful tool to extract point sources from cosmic microwave background maps,” Monthly Notices of the Royal Astronomical mic microwave background astronomy and the development and Society, vol. 315, no. 4, pp. 757–761, 2000. application of statistical signal processing techniques to astronom- [14] P. Vielva, E. Mart´ınez-Gonzalez,´ L. Cayon,J.M.Diego,J.L.´ ical data, in particular, linear and nonlinear filtering for the detec- Sanz, and L. Toffolatti, “Predicted Planck extragalactic point- tion and extraction of extragalactic point sources. source catalogue,” Monthly Notices of the Royal Astronomical Society, vol. 326, no. 1, pp. 181–191, 2001. D. Herranz received the B.S. degree in 1995 [15] P. Vielva, R. B. Barreiro, M. P. Hobson, et al., “Combining and the M.S. degree in 1995 from the Uni- maximum-entropy and the Mexican hat wavelet to recon- versidad Complutense de Madrid, Madrid, struct the microwave sky,” Monthly Notices of the Royal As- Spain, and the Ph.D. degree in astrophysics tronomical Society, vol. 328, no. 1, pp. 1–16, 2001. from Universidad de Cantabria, Santander, [16] P. Vielva, E. Mart´ınez-Gonzalez,´ J. E. Gallegos, L. Toffolatti, Spain, in 2002. He was a CMBNET Post- and J. L. Sanz, “Point source detection using the Spheri- doctoral Fellow at the Istituto di Scienza cal Mexican Hat Wavelet on simulated all-sky Planck maps,” e Tecnologie dell’Informazione “A. Faedo” Monthly Notices of the Royal Astronomical Society, vol. 344, (CNR), Pisa, Italy, from 2002 to 2004. He no. 1, pp. 89–104, 2003. is currently at the Instituto de Fisica de [17] M. Tegmark and A. de Oliveira-Costa, “Removing point Cantabria, Santander, Spain, under an MEC Juan de la Cierva con- sources from cosmic microwave background maps,” The As- tract. His research interests are in the areas of cosmic microwave trophysical Journal Letters, vol. 500, no. 2, pp. L83–L86, 1998. background astronomy and extragalactic point source statistics as [18]L.Y.Chiang,H.E.Jørgensen,I.P.Naselsky,P.D.Naselsky,I. well as the application of statistical signal processing to astronom- D. Novikov, and P. R. Christensen, “An adaptive filter for the ical data, including blind source separation, linear and nonlinear PLANCK compact source catalogue construction,” Monthly data filtering, and statistical modeling of heavy-tailed processes. Notices of the Royal Astronomical Society, vol. 335, no. 4, pp. 1054–1060, 2002. J. L. Sanz received the M.S. degree in [19] J. L. Sanz, D. Herranz, and E. Mart´ınez-Gonzalez,´ “Optimal 1971 from the Universidad Complutense detection of sources on a homogeneous and isotropic back- de Madrid, Spain, and the Ph.D. degree ground,” The Astrophysical Journal, vol. 552, no. 2, pp. 484– in physical sciences from Universidad Au- 492, 2001. [20] D. Herranz, J. L. Sanz, R. B. Barreiro, and E. Mart´ınez- tonoma de Madrid, Spain, in 1976. He was a Gonzalez,´ “Scale-adaptive filters for the detection/separation Postdoctoral Fellow at the Queen Mary Col- of compact sources,” The Astrophysical Journal, vol. 580, no. 1, lege, London, UK, during 1978. He is cur- pp. 610–625, 2002. rently Professor of astrophysics at the In- [21] M. Lopez-Caniego,D.Herranz,R.B.Barreiro,andJ.L.Sanz,´ stituto de Fisica de Cantabria, Santander, “A Bayesian approach to filter design: detection of compact Spain. His research interests are in the areas sources,” in Computational Imaging II, vol. 5299 of Proceedings of cosmic microwave background astronomy (extragalactic point of SPIE, San Jose, Calif, USA, January 2004. sources, component separation, and non-Gaussian studies) as well [22] M. Lopez-Caniego,´ J. L. Sanz, D. Herranz, and R. B. Barreiro, as the application of statistical signal processing and image analysis “Filter design for the detection of compact sources based on to astronomical data (linear and nonlinear data filtering, fusion). 2436 EURASIP Journal on Applied Signal Processing R. B. Barreiro obtained her B.S. degree in 1995 from the Universidad de Santiago de Compostela, Spain, completing also as part of her degree one year at the Uni- versity of Sheffield, UK. She completed her Ph.D. in astrophysics in the Universi- dad de Cantabria in 1999. After her Ph.D., she worked as a Research Associate at the Cavendish Laboratory of the University of Cambridge, UK, until the end of 2001. She is currently at the Instituto de Fisica de Cantabria (CSIC-UC), Spain, under a Ramon´ y Cajal contract. Her research interests are mainly focused in the field of the cosmic microwave background, including statistical data analysis, in particular the study of the Gaussianity of the CMB, and the development of component sepa- ration techniques for both diffuse emissions and compact sources. EURASIP Journal on Applied Signal Processing 2005:15, 2437–2454 c 2005 Hindawi Publishing Corporation Blind Component Separation in Wavelet Space: Application to CMB Analysis Y. Moudden DAPNIA/SEDI-SAP, CEA/Saclay, 91191 Gif-sur-Yvette, France Email: [email protected] J.-F. Cardoso CNRS, Ecole´ National Superieure´ des T´el´ecommunications, 46 rue Barrault, 75634 Paris, France Email: [email protected] J.-L. Starck DAPNIA/SEDI-SAP, CEA/Saclay, 91191 Gif-sur-Yvette, France Email: [email protected] J. Delabrouille CNRS/PCC, Coll`ege de France, 11 place Marcelin Berthelot, 75231 Paris, France Email: [email protected] Received 30 June 2004; Revised 22 November 2004 It is a recurrent issue in astronomical data analysis that observations are incomplete maps with missing patches or intentionally masked parts. In addition, many astrophysical emissions are nonstationary processes over the sky. All these effects impair data processing techniques which work in the Fourier domain. Spectral matching ICA (SMICA) is a source separation method based on spectral matching in Fourier space designed for the separation of diffuse astrophysical emissions in cosmic microwave background observations. This paper proposes an extension of SMICA to the wavelet domain and demonstrates the effectiveness of wavelet- based statistics for dealing with gaps in the data. Keywords and phrases: blind source separation, cosmic microwave background, wavelets, data analysis, missing data. 1. INTRODUCTION constrain these models as well as to measure the cosmologi- cal parameters describing the matter content, the geometry, The detection of cosmic microwave background (CMB) and the evolution of our universe [6]. anisotropies on the sky has been over the past three decades a Accessing this information, however, requires disentan- subject of intense activity in the cosmology community. The gling in the data the contributions of several distinct astro- CMB, discovered in 1965 by Penzias and Wilson, is a relic ra- physical sources, all of which emit radiation in the frequency diation emitted some 13 billion years ago, when the universe range used for CMB observations [7]. This problem of com- was about 370 000 years old. Small fluctuations of this emis- ponent separation, in the field of CMB studies, has thus been sion, tracing the seeds of the primordial inhomogeneities the object of many dedicated studies in the past. which gave rise to present large scale structures as galaxies To first order, the total sky emission can be modeled as and clusters of galaxies, were first discovered in the observa- a linear superposition of a few independent processes. The tions made by COBE [1] and further investigated by a num- observation of the sky in direction (θ, ϕ)withdetectord is ber of experiments among which Archeops [2], boomerang then a noisy linear mixture of Nc components: [3], maxima [4], and WMAP [5]. The precise measurement of these fluctuations is of ut- Nc x ϑ ϕ = A s ϑ ϕ n ϑ ϕ most importance to cosmology. Their statistical properties d( , ) dj j ( , )+ d( , ), (1) (spatial power spectrum, Gaussianity) strongly depend on j=1 the cosmological scenarios describing the properties and where sj is the emission template for the jth astrophysi- evolution of our universe as a whole, and thus permit to cal process, herein referred to as a source or a component. 2438 EURASIP Journal on Applied Signal Processing The coefficients Adj reflect emission laws while nd accounts Blind component separation (and in particular estima- for noise. When Nd detectors provide independent observa- tion of the mixing matrix), as discussed by Cardoso [17], can tions, this equation can be put in vector-matrix form: be achieved in several different ways. The first of these ex- ploits non-Gaussianity of all, but possibly one, components. The component separation method of Baccigalupi [11]and X(ϑ, ϕ) = AS(ϑ, ϕ)+N(ϑ, ϕ), (2) Maino [12] belongs to this set of techniques. In CMB data analysis, however, the main component of interest (the CMB where X and N are vectors of length Nd, S is a vector of length itself) has a Gaussian distribution and the observed mixtures Nc,andA is the Nd × Nc mixing matrix. suffer from additive Gaussian noise, so that better perfor- Given the observations of such a set of independent de- mance can be expected from methods based on Gaussian tectors, component separation consists in recovering esti- models. A second set of techniques exploits spectral diver- mates of the maps of the sources sj (ϑ, ϕ). Explicit component sity and works in the Fourier domain. It has the advantage separation has been investigated first in CMB applications by that detector–dependent beams can be handled easily, since [7, 8, 9]. In these applications, recovering component maps is the convolution with a point spread function in direct space the primary target, and all the parameters of the model (mix- becomes a simple product in Fourier space. SMICA follows ing matrix Adj, noise levels, statistics of the components, in- this approach in the context of noisy observations. Finally, a cluding the spatial power spectra) are assumed to be known third set of methods exploits nonstationarity. It is adapted to and are used to invert the linear system. situations where components are strongly nonstationary in Recent research has addressed the case of an imperfectly real space. known mixing matrix. It is then necessary to estimate it (or It is natural to investigate the possible benefits of ex- at least some of its entries) directly from the data. For in- ploiting both nonstationarity and spectral diversity for blind stance, Tegmark et al. assume power law emission spectra for component separation using wavelets. Indeed wavelets are all components except CMB and SZ, and fit spectral indices powerful tools in revealing the spectral content of nonsta- to the observations [10]. tionary data. Although blind source separation in the wavelet More recently, blind source separation or independent domain has been previously examined, the setting here is component analysis (ICA) methods have been implemented different. We should mention, for instance, the separation specifically for CMB studies. The work of Baccigalupi et method in [18] which is based on the non-Gaussianity of the al. [11], further extended by Maino et al. [12], imple- source signals but after a sparsifying wavelet transform and ments a blind source separation method exploiting the non- the Bayesian approach in [19] which adopts a similar point Gaussianity of the sources for their separation, which permits of view although with a richer source model accounting for to recover the mixing matrix A and the maps of the sources. correlations in the wavelet representation. Accounting for spatially varying instrumental noise in the The paper is organized as follows. In Section 2,wefirst observation model is investigated by Kuruoglu et al. in [13], recall the principle of spectral matching ICA. Then, after as well as the possible inclusion of prior information about a brief reminder of some properties of the atrous` wavelet the distributions of the components using a generic Gaussian transform, we discuss in Section 3 the extension of SMICA to mixture model. componentseparationinwaveletspaceinordertodealwith Snoussi et al. [14] propose a Bayesian approach in the nonstationary data. Considering the problem of incomplete Fourier domain assuming known spectra for the compo- data as a model case of practical significance for the compar- nents as well as possibly non-Gaussian priors for the Fourier ison of SMICA and its extension wSMICA, numerical exper- coefficients of the components. A fully blind, maximum like- iments and results are reported in Section 4. lihood approach is developed in [15, 16], with the new point of view that spatial power spectra are actually the main un- known parameters of interest for CMB observations. A key 2. SMICA benefit is that parameter estimation can then be based on a Spectral matching ICA, or SMICA for short, is a blind set of band-averaged spectral covariance matrices, consider- source separation technique which, unlike most standard ably compressing the data size. ICA methods, is able to recover Gaussian sources in noisy Working in the frequency domain offers several benefits contexts. It operates in the spectral domain and is based on but the nonlocality of the Fourier transform creates some dif- spectral diversity:itisabletoseparatesourcesprovidedthey ficulties. In particular, one may wish to avoid the averaging have different power spectra. This section gives a brief ac- induced by the nonlocality of the Fourier transform when count of SMICA. More details can be found in [16]; first ap- dealing with strongly nonstationary components or noise. In plications to CMB analysis are in [16, 20]. addition, in many experiments, only an incomplete sky cov- erage is available. Either the instrument observes only a frac- 2.1. Model and cost function tion of the sky or some regions of the sky must be masked due to localized strong astrophysical sources of contamination: For a second-order stationary Nd-dimensional process, we compact radio sources or galaxies, strong emitting regions in denote by RX (ν) the Nd×Nd spectral covariance matrix at fre- the galactic plane. These effects can be mitigated in a simple quency ν, that is, the (i, i)th entry of RX (ν) is the power spec- manner thanks to the localization properties of wavelets. trum of the ith coordinate of X while the offdiagonal entries Component Separation in Wavelet Space for CMB Analysis 2439