Fullsky weaklensing simulation with 70 billion particles
R. Teyssier^{1,2}  S. Pires^{1,3}  S. Prunet^{2}  D. Aubert^{4}  C. Pichon^{1,2}  A. Amara^{1}  K. Benabed^{2}  S. Colombi^{2}  A. Refregier^{1}  J.L. Starck^{1,3}
1  Service d'Astrophysique, CEA Saclay, Bâtiment 709, 91191 GifsurYvette Cedex, France
2 
Institut d'Astrophysique de Paris, 98bis boulevard Arago, 75014 Paris, France
3 
Service d'Électronique, de Detection et d'Informatique, CEA Saclay, Bâtiment 141, 91191 GifsurYvette Cedex, France
4 
Observatoire Astronomique, Université de Strasbourg, UMR 7550, 11 rue de l'Université, 67000 Strasbourg, France
Received 23 July 2008 / Accepted 13 January 2009
Abstract
We have performed a 70 billion darkmatter particles Nbody
simulation in a 2 h^{1} Gpc periodic box, using the concordance,
cosmological model as favored by the latest WMAP3 results. We have
computed a fullsky convergence map with a resolution of
arcmin^{2}, spanning 4 orders of magnitude in
angular dynamical range. Using various highorder statistics on a
realistic cut sky, we have characterized the transition from the
linear to the nonlinear regime at
and shown that
realistic galactic masking affects highorder moments only below
.
Each domain (Gaussian and nonGaussian) spans
2 decades in angular scale. This map is therefore an ideal tool for
testing mapmaking algorithms on the sphere. As a first step in
addressing the full map reconstruction problem, we have benchmarked
in this paper two denoising methods: 1) Wiener filtering applied to
the Spherical Harmonics decomposition of the map and 2) a new
method, called MRLens, based on the modification of the Maximum
Entropy Method on a Wavelet decomposition. While the latter is
optimal on large spatial scales, where the signal is Gaussian,
MRLens outperforms the Wiener method on small spatial scales, where
the signal is highly nonGaussian. The simulated fullsky
convergence map is freely available to the community to help the
development of new mapmaking algorithms dedicated to the next
generation of weaklensing surveys.
Key words: methods: Nbody simulations  methods: data analysis  cosmology: largescale structure of Universe
1 Introduction
Weak gravitational lensing, or ``cosmic shear'', provides a unique tool for mapping the matter density distribution in the Universe (for reviews, see Refregier 2003; Hoekstra 2003; Munshi et al. 2006). Current weaklensing surveys cover altogether about 100 square degrees and have been used to measure the amplitude of the matter power spectrum and other cosmological parameters (see; Fu et al. 2008, and references therein). A number of new instruments are being planned to carry out these surveys over wider sky areas (PanSTARRS, DES, SNAP and LSST)^{} or even over the full extragalactic sky (DUNE^{}). These widefield surveys will yield cosmicshear measurements on both large scales, where gravitational dynamics is in the linear regime, and small scales, where the dynamics is highly nonlinear. The comparison of these measurements with theoretical predictions of the density field evolution will place strong constraints on cosmological parameters, including dark energy parameters (e.g. Hu & Tegmark 1999; Huterer 2002; Amara & Refregier 2006; Albrecht & Bernstein 2007). On small scales, the highly nonlinear nature of the density field ensures that predictions based on analytic calculations are prohibitively difficult and requires the use of numerical simulations. Nbody simulations have thus been used to simulate weaklensing maps across small patches of the sky, using the flat sky approximation (e.g. Jain et al. 2000; Hamana et al. 2001; White & Vale 2004). The simulation of fullsky maps in preparation for future surveys involve a wide range of both mass and length scales and is challenging for current Nbody simulations. The range of scales involved also requires the development of efficient algorithms for deriving a mass map from true noisy data sets. These algorithms need to be wellsuited to both the largescale signal, which is essentially a Gaussian random field, and those on smallscales, where it is highly nonGaussian and exhibits localized features.
In this paper, we used a high resolution Nbody simulation to construct a fullsky weaklensing map and test a new mapreconstruction method based on a multiresolution technique. For this purpose, we use the Horizon simulation, a 70 billion particle Nbody simulation, featuring more than 140 billion cell in the AMR grid of the RAMSES code (Teyssier 2002). The simulation covers a sufficiently large volume ( Gpc) to compute a fullsky convergence map, while resolving MilkyWay size halos with more than 100 particles, and exploring small scales into the nonlinear regime (see Sect. 2). This unprecedented computational effort allows us, for the first time, to close the gap between scales close to the cosmological horizon and scales deep inside virialized darkmatter haloes. A similar effort at lower resolution was presented by Fosalba et al. (2008).
The darkmatter distribution in the simulation was integrated in a light cone to a redshift of 1, around an observer located at the centre of the simulation box (see Sect. 3). This light cone was then used to calculate the corresponding fullsky lensing convergence field, which we mapped using the Healpix^{} pixelisation scheme (Górski et al. 2005) with a pixel resolution of arcmin^{2}( ), and added ``instrumental'' noise for a typical allsky survey with 40 galaxies per arcmin^{2}, as expected for example for the DUNE mission (Réfrégier et al. 2006). Using an Undecimated Isotropic Wavelet Decomposition of this realistic simulated weaklensing map on the sphere, we analyzed the statistics of each wavelet plane using second, third and fourth order moments estimator (Sect. 4). We then applied, in Sect. 5, a multiresolution algorithm to filter a fictitious simulated data set based on an extension of the wavelet filtering technique of Starck et al. (2006b). We characterised the quality of the reconstruction using the power spectrum of the error map and compare this to the result of standard Wiener filtering on the sphere. Our results, summarised in Sect. 6, illustrate the virtue of high resolution simulations such as the one reported here to prepare for future weaklensing surveys and to design new mapmaking techniques.
Figure 1: Fullsky simulated convergence map derived from the Horizon Simulation. Its resolution of 200 million pixels has been downgraded to fit the page. The various inserts display a zoom sequence into smaller and smaller areas of the sky. The pixel size is 0.74 arcmin^{2}. 

Open with DEXTER 
2 The horizon Nbody simulation
This large Nbody simulation was carried out using the RAMSES code (Teyssier 2002) for two months on the 6144 Itanium2 processors of the CEA supercomputer BULL Novascale 3045 hosted in France by CCRT^{}. RAMSES is a parallel hydro and Nbody code based on the Adaptive Mesh Refinement (AMR) techniques. Using a parallel version of the grafic package (Bertschinger 2001), we generated the initial displacement field on a 4096^{3} grid for the cosmological parameters from the WMAP 3rd year results (Spergel et al. 2007), namely , , , n=0.958, H_{0}=73 km s^{1} Mpc^{1}and . We used the Eisenstein & Hu (1999) transfer function, which includes baryon oscillations. The box size was set to 2 Gpc/h, which corresponds roughly to the comoving distance to an object at . We used 68.7 billion particles to simuate the darkmatter density field, yielding a particle mass of and 130 particles per Milky Way halo. This large particle distribution was split across 6144 individual files, one for each processor, according to the RAMSES code domain decomposition strategy (Prunet et al. 2008). Starting with a base (or coarse) grid of 4096^{3} grid points, AMR cells are recursively refined if the number of particles in the cell exceeds 40. In this way, the number of particles per cell varied between 5 and 40, so that the particle shot noise remained at an acceptable level. At the end of the simulation, we had reached 6 levels of refinement with a total of 140 billion AMR cells. This corresponds to a formal resolution of 262 144^{3} or 7.6 h^{1}kpc comoving spatial resolution. Parallel computing is perfomed using the MPI messagepassing library, with a domain decomposition based on the PeanoHilbert spacefilling curve. The work and memory load was adjusted dynamically by reshuffling particles and grid points from each processor to its neighbors. The simulation required 737 main (or coarse) time steps and more than 10^{4} fine time steps for completion.
3 Light cone and convergence map
Born's unperturbedtrajectory assumption for all neighboring light rays is a good approximation in the linear regime of structure formation, but is inaccurate in the nonlinear regime. Consequently, distortion effects of lensing beyond the first order cannot be simulated reliably. As shown by Van Waerbeke et al. (2001), the Born approximation also introduces a relative error in the skewness of the signal of aproximately 10% on large scales where the convergence is Gaussian, and about 1% on small scales in the nonlinear regime. We therefore implemented a multiplelens raytracing method that can be applied more generally than Born's approximation.
We constructed a light cone by recording, at each main time step, the positions of particles within the boundaries of a photon plane: this plane moved at the speed of light towards an observer, who was located at the centre of the box. Our method was developed from the one presented by Hamana et al. (2001). This method produced 348 slices in the light cone, spanning the redshift range [0, 1]. Due to the large size of the simulated volume, the effect of periodic replications of the computational box are minimized. Each slice was then transformed into a fullsky Healpix map ( nside=4096) of the average overdensity using a simple ``Nearest Grid Point'' (NGP) mass projection scheme. The density slices thus represented our physical model of the lens screens used in the raytracing procedure. We note that there is no unique procedure for generating a bandlimited harmonic representation of each slice of particles. We choose to use an NGP interpolation because it is a good compromise between filtering and aliasing, and remains localised in configuration space. More sophisticated interpolation schemes have been developed in the context of either 3D particle distributions (Colombi et al. 2009) or 2D continuous fields (Basak et al. 2008), which, however, remain impractical in significantly large simulations.
After an interpolation kernel has been chosen, all fields (lensing potentials and displacement fields) are computed from the NGP interpolation mass slices at each redshift using a spherical harmonic decomposition. The resampling of the displacement fields outside the pixel centres (as required in a multilens method) is completed using a local linearinterpolation scheme (using covariant, second derivatives of the potential); this last interpolation has the same spectral behavior (and thus the same aliasing contamination) as the NGPinterpolated mass slices, and we therefore do not need to use a higherorder resampling scheme (since the calculation of the potential requires two sets of integration over the mass distribution, while the interpolation of the displacement field corresponds to a secondorder derivative). We provide more details in Appendix A (see Jain et al. 2000; Hamana et al. 2001; and White & Vale 2004, for alternative approaches). We assumed that the background galaxies are within a single source plane located at redshift . The final convergence map was computed using our multiplelens raytracing scheme, for which spherical geometry precludes the use of small angle approximations (as in Das & Bode 2008) especially in the neighborhood of the poles; full rotation matrices for each light ray must therefore be computed from the displacement fields at each redshift.
The resulting fullsky Healpix map with a pixel size of arcmin is shown in Fig. 1, with small inserts to highlight the large dynamical range achieved^{}. The particle shot noise corresponding to our 70 billion particle run has a small impact on the map. As shown in Fig. 4, the particle shot noise is well below the expected instrumental noise, and even sufficiently low to be ignored in the spectral analysis of the signal.
4 Highorder moments and realistic sky cut
Figure 2: Map of the cutsky used in Sect. 4 to compute highordermoments. 

Open with DEXTER 
Figure 3: Moments of the convergence as a function of the average multipole moment on each wavelet scale. The variance, skewness, and kurtosis are shown as black, blue, and red lines, respectively. Solid lines with error bars corresponds ro a fullsky analysis, while dotted lines correspond to our cutsky analysis. 

Open with DEXTER 
Figure 4: Angular power spectrum of the simulated convergence map (black solid line), compared to a fit based on the Smith et al. (2001) analytical model with error bars corresponding to our noise model (pink area). Also shown is the prediction from linear theory (pink dashed curve). The noise power spectrum is plotted as the dashed black line. The green solid line is the power spectrum of the error map obtained with the Wiener filter method, while the blue solid line are that for the MRLens method. 

Open with DEXTER 
In Fig. 1, the signal appears as a typical Gaussian random field on large scales, similar to the Cosmic Microwave Background map seen by the WMAP satellite (Spergel et al. 2007). On small scales, the signal is clearly dominated by clumpy structures (dark matter halos) and is therefore highly nonGaussian. To characterize this quantitatively, we performed a wavelet decomposition of our map using the Undecimated Isotropic Wavelet Transform on the sphere (Starck et al. 2006a), and, for each wavelet scale, we have computed its second, third and fourthorder moment. We used 11 scales with central multipole values of , 4500, 2250, 1125, 562, 282, 141, 71, 35, 18. For each of these maps, we computed the variance , the normalized skewness , and the normalized kurtosis . Results are plotted in Fig. 3 as solid lines of various colors. Error bars were estimated approximately by computing each moment on the 12 Healpix base pixels independently and evaluating the variance in the 12 results. A more appropriate strategy would have been to perform several, independent, 70 billion particle runs, which is currently impossible for us to do. We can see that the variance in the signal steadily increases for higher and higher multipoles, and saturates at a fraction of 10^{4}, corresponding to the value predicted from nonlinear gravitational clustering for . The variance for each wavelet plane can be considered to be a band power estimate of the angular power spectrum, as can be verified using Fig. 4. In the same figure, we have also plotted for comparison the linear power spectrum, to highlight the scale below which nonlinear clustering contributes significantly, i.e., for or equivalently , as first pointed out by Jain & Seljak (1997). Skewness and kurtosis are more direct estimators of the signal nonGaussianity. Departures from Gaussianity occur around , where both statistics cross unity. Due to the large dynamical range of the Horizon simulation, we computed a map spanning two decades in angular scales in the linear, Gaussian regime and two additional decades in angular scales into the nonlinear, nonGaussian regime.
It is clear from Fig. 3 that at small , the skewness and the kurtosis of the map are strongly affected by cosmic variance. The statistics of the convergence field cannot be measured in practice over the whole sky because of sky cuts imposed by the presence of saturated stars and by absorption in the Galactic plane. We estimated the impact of this sky cut on the accuracy of our multiresolution statistical analysis. We computed the expected number of bright stars that would saturate CCD cameras typically employed in widefield survey (Bmagnitude < 20). We then removed from our analysis each pixel contaminated by at least 3 bright stars, based on a random Poissonian realization of bright stars in our Galaxy (according to the model presented in Bahcall & Soneira 1980, Appendix B). We obtain a mask with 40% of the sky removed, corresponding roughly to a cut around the Galactic plane (see Fig. 2). The resulting statistics are overplotted as dotted lines in Fig. 3. The transition scale, for which the departure from Gaussianity is significant, can still be estimated reliabily around . We concluded that the cosmic variance of the cut sky affects highorder moments only below .
Figure 5: Reconstruction of convergence maps with our 2 filtering techniques. The top panels show the square map corresponding to first zoom sequence of Fig. 1. The bottom panels are subset of the corresponding top images with linear size 45'. From left to right, we show the original signal, the noisy image, the Wienerfiltered image and the the MRLensfiltered image. 

Open with DEXTER 
5 Mapmaking using multiresolution filtering
The fullsky simulated convergence maps described above can be used to analyze and compare denoising (or mapmaking) methods on the sphere. For this purpose, we considered a purely white instrumental noise, typical of the next generation allsky surveys, and a root mean square per pixel of area given by for background galaxies per arcmin^{2}. Recovering the most accurate convergence map from noisy data will be an important step in future surveys. This reconstructed map can be used to construct a mass selected halo catalog, measure its statistical properties and constrain cosmological parameters, and be compared directly with other cluster catalogues compiled with other techniques (Xray, galaxy counts or SZ). We restrict ourselves to the fullsky denoising of a convergence map already reconstructed from the shear derived from galaxy ellipticities. In the present work, we do not address filtering in the presence of a cutsky, such as the one shown in Fig. 2. Promising methods based on ``impainting'' have been developed in the CMB context (Abrial et al. 2008), and also weaklensing applications (Pires et al. 2008); these replace missing data with an artificial signal and allow us to optimize the results we obtained with filtering methods for a fullsky analysis.
A straightforward filtering method is the Wiener filtering scheme, which is optimal for Gaussian random fields, and is expected to operate here effectively on large scales. Defining as the power spectrum of the input signal (see Fig. 4) and the power spectrum of the noise, this method involves convolving the noisy map by the Wiener filter defined as . The results of the Wiener filtering approach are shown in Fig. 5. Comparing with the input signal map, we conclude that, although the agreement is satisfactory on large scales, the dense clumps clearly visible in the image are poorly recovered because they have been convolved too significantly.
A dedicated weaklensing waveletrestoration method, called MRLens, has been developed (Starck et al. 2006b). It can be considered to be an extension of the Maximum Entropy Method (MEM) that provides different types of information. In MRLens, the entropy constraint is not applied to the pixels of the solution, but rather its wavelet coefficients. This allows us to take into account more efficiently the multiscale behavior of the information. MRLens was, however, designed for weaklensing maps of smaller surface area on the sky, for which the nonGaussian signal is stronger. MRLens was extended here to the sphere by considering independently each of the 12 Healpix base pixels covering the sphere as 12 independent Cartesian maps, on which we applied the MRLens algorithm of Starck et al. (2006b). Fullsky denoising performed with MRLens is shown in Fig. 5. It performs more efficiently than the Wiener methods on small scales, with dense clumps more accurately estimated, but less efficient than the Wiener method on large scale when recovering low frequency waves in the map. We also computed the angular power spectrum of the error map (see Fig. 4) in both cases (Wiener and MRLens). We can see that Wiener filtering outperforms MRLens on large scales. Interestingly, the MRLens errors decrease significantly above the transition scale we identified in the last section around (see Fig. 4).
To compare both methods more quantitatively, we computed the skewness and kurtosis of both reconstructed maps. Results are shown in Fig. 6. We note that using mapmaking algorithms to recover the skewness and kurtosis of the true signal is not at all the optimal strategy: maximum likelihood estimators are more appropriate. We used highorder statistics here only to compare the relative merits of each method. It is striking to observe in Fig. 6 that the Wiener reconstructed map strongly underestimate the skewness and the kurtosis at small scale. This confirms quantitatively what was already visible in the maps (Fig. 5), namely that the Wiener method strongly suppresses high peaks in the map, affecting the tail of the probability distribution function. On the other hand, the MRLens reconstructed map has a significantly higher skewness and kurtosis than the original map: this waveletbased method is only efficient in recovering high peaks in the signal, affecting the reconstructed probability density function in the opposite direction.
Figure 6: Skewness (blue lines) and kurtosis (red lines) for the original convergence map (solid lines with error bars), compared to the same highorder statistics for the Wiener reconstructed map (dotted lines) and the MRLens reconstructed map (dashed lines). 

Open with DEXTER 
Figure 7: Histogram of the residual maps for Wiener and MRLensfiltering. 

Open with DEXTER 
We now use the probability density function (PDF) of the residual maps to compare each method (see Fig. 7). We confirm our visual impression from Fig. 5 that MRLens performs more efficiently than the Wiener method in recovering the high convergence, nonlinear features in the map. The positive high residual tail is reduced significantly by MRLens, as well as the dozen of strongly outlying pixels in the Wiener filterer map around (see Fig. 7). MRLens, however, performs poorly for small values of the convergence ( ), for which the PDF is well approximated by a Gaussian, an optimal situation for Wiener filtering.
The present analysis, based on using both the power spectrum of the residual maps and the highorder moments of the recovered map, strongly suggests that new methods should be developed using an hybrid, multiresolution formulation; for instance, using spherical harmonics on large scales, while utilizing wavelets coefficients on small scales. The methodology of this combined approach could be based on the idea of Combined Filtering introduced by Starck et al. (2006a).
6 Conclusion
Using the 70 billion particles of the Horizon Nbody Simulation, we have computed for the first time a realistic fullsky convergence map with a pixel resolution of arcmin^{2}. We have analyzed the resulting map using multiresolution statistics (variance, skewness, and kurtosis) and angular powerspectrum analysis. We have shown that this simulated map spans 4 decades of useful signal in angular scale, with 2 decades within the linear, Gaussian regime and 2 decades well into the nonlinear, nonGaussian regime. We have shown that, when considering a realistic sky cut, we can reliabily estimate highorder moments of the map above . Using even higher resolution maps, angular scales smaller than could be explored in future works, although the mass ditribution on these scales might be affected by baryons physics (Jing et al. 2006), so that the present map might already cover all cosmologically relevant scales.
As a first step towards a realistic mapmaking procedure, we have tested two denoising schemes on a simplified fictitious dataset derived from the fullsky map, namely Wiener filtering and the MRLens method (Starck et al. 2006b). We have shown quantitatively that Wiener filtering is the most effective method on large scales, although some signal is lost on small scales. MRLens performs more effectively on small scales and recovers the dense clumps associated with dark matter halos, but deals less accurately with low frequency waves in the map. Hence, this work demonstrates the need for hybrid multiresolution approach, e.g., by combining spherical harmonics and wavelet coefficients. The present analysis will be extended in future work to mapmaking algorithms dealing directly with galaxy shears. The simulated convergence map may prove to be an effective tool for the design of new mapmaking methods and for the preparation of the next generation weaklensing surveys^{}.
Acknowledgements
We would like to thank Julien Devriendt, Pierre Ocvirk, Arthur Petitpierre and Philippe Lachamp for their unvaluable help during the course of this project. The Horizon Simulation presented here was supported by the ``Centre de Calcul Recherche et Technologie'' (CEA, France) as a ``Grand Challenge project''. This work was supported by the Horizon Project. Some of the results in this paper have been derived using the HEALPix (Górski et al. 2005) package.
Appendix A: Computing the convergence maps from simulations
We first recall how to compute the convergence in the Born approximation, and then present our new raytracing scheme.
A.1 Born approximation
We start by the formula relating the convergence to the density
contrast:
which is valid for sources at a single redshift z_{s}, and is the dimensionless, comoving, radial coordinate ( ). We now rewrite this formula in a form that is more suited to integration over redshift slices in a simulation:
where
is a slicerelated weight, and the integral over the density contrast reads
where
is the comoving surface of the spherical pixel. Interpreting all together, we obtain the following formula for the convergence map (omitting the term that corresponds to a constant term):
This is the equation used to derive the convergence map in the Born approximation.
A.2 Raytracig using multiple planes
We discuss here the formulae needed for the multiplane
computations, where we consider the lensing by a number of thin lenses
located at
.
We define
and
with
To follow the light rays, we are interested in computing the angular displacement field for each ray i due to a slice at z_{b}. We then define
where the gradient and Laplacien are computed using angular covariant derivatives on the (unit) sphere, and is the current direction of light ray i when it is incident on the slice b. Now, we start from light rays that are backpropagated from the observer at z=0towards the source (here at z=1). We denote by the location of the Healpix centres, which corresponds to the initial directions of the backpropagated rays emanating from the observer. The tangent vectors to each light ray will be modified by the deflection field at each lens plane, defined by Eq. (A.3). Then, computing the displacement of the rays at slice b reads
We then update the direction of the rays according to the following rotation, :
where (light rays emanate from the observer, thus in a direction perpendicular to the first slice), and is the vector normal to slice b at the intersection of lightray i on slice b. Equation (A.4) can be simplified by noting that is expressed naturally in the local basis of the tangent plane at position :
After calculating the new value of , one needs to compute the intersection of the light rays with the next shell. We call the Cartesian position of the intersection of light ray i with slice b, then the next intersection will be given by the solution for of the system:
assuming that remains strictly unitary, and R_{b} is the comoving radius of slice b. Once is known, it is easy to compute the new positions. The contributions to are then calculated following Eq. (A.1), but where the slice contributions are interpolated at the displaced positions:
We note that may fall into different pixels as a function of the slice b.
References
 Abrial, P., Moudden, Y., Starck, J.L., et al. 2008, Stat. Method., 5, 289 In the text
 Albrecht, A., & Bernstein, G. 2007, Phys. Rev. D, 75, 103003 In the text
 Amara, A., & Refregier, A. 2006, ArXiv Astrophysics eprints In the text
 Bahcall, J. N., & Soneira, R. M. 1980, ApJS, 44, 73 In the text
 Basak, S., Prunet, S., & Benabed, K. 2008, Phys. Rev. D, [arXiv:0811.1677] In the text
 Benjamin, J., Heymans, C., Semboloni, E., et al. 2007, ArXiv Astrophysics eprints
 Bertschinger, E. 2001, ApJS, 137, 1 In the text
 Colombi, S., Jaffe, A., Novikov, D., & Pichon, C. 2009, MNRAS, 393, 511 In the text
 Das, S., & Bode, P. 2008, ApJ, 682, 1 In the text
 Eisenstein, D. J., & Hu, W. 1999, ApJ, 511, 5 In the text
 Fosalba, P., Gaztanaga, E., Castander, F., & Manera, M. 2008, MNRAS, 391, 435 In the text
 Fu, L., Semboloni, E., Hoekstra, H., et al. 2008, A&A, 479, 9 In the text
 Górski, K. M., Hivon, E., Banday, A. J., et al. 2005, ApJ, 622, 759 In the text
 Hamana, T., Colombi, S., & Suto, Y. 2001, A&A, 367, 18 In the text
 Hoekstra, H. 2003, ArXiv Astrophysics eprints In the text
 Hu, W., & Tegmark, M. 1999, ApJ, 514, L65 In the text
 Huterer, D. 2002, Phys. Rev. D, 65, 063001 In the text
 Jain, B., & Seljak, U. 1997, ApJ, 484, 560 In the text
 Jain, B., Seljak, U., & White, S. 2000, ApJ, 530, 547 In the text
 Jing, Y. P., Zhang, P., Lin, W. P., Gao, L., & Springel, V. 2006, ApJ, 640, L119 In the text
 Munshi, D., Valageas, P., Van Waerbeke, L., & Heavens, A. 2006, ArXiv Astrophysics eprints In the text
 Pires, S., Starck, J., Amara, A., et al. 2008, ArXiv eprints In the text
 Prunet, S., Pichon, C., Aubert, D., et al. 2008, ApJS, 178, 179 In the text
 Refregier, A. 2003, ARA&A, 41, 645 In the text
 Réfrégier, A., Boulade, O., Mellier, Y., et al. 2006, in Space Telescopes and Instrumentation I: Optical, Infrared, and Millimeter, ed. J. C. Mather, H. A. MacEwen, & M. W. M. de Graauw, Proc. SPIE, 6265, 62651Y In the text
 Spergel, D. N., Bean, R., Doré, O., et al. 2007, ApJS, 170, 377 In the text
 Starck, J.L., Moudden, Y., Abrial, P., & Nguyen, M. 2006a, A&A, 446, 1191 In the text
 Starck, J.L., Pires, S., & Réfrégier, A. 2006b, A&A, 451, 1139 In the text
 Teyssier, R. 2002, A&A, 385, 337 In the text
 Van Waerbeke, L., Hamana, T., Scoccimarro, R., Colombi, S., & Bernardeau, F. 2001, MNRAS, 322, 918 In the text
 White, M., & Vale, C. 2004, Astropart. Phys., 22, 19 In the text
Footnotes
 ... LSST)^{}
 PanSTARRS: http://panstarrs.ifa.hawaii.edu, DES: https://www.darkenergysurvey.org, SNAP: http://snap.lbl.gov and LSST: http://www.lsst.org
 ... (DUNE^{}
 DUNE: http://www.dunemission.net
 ... Healpix^{}
 HeaPix: http://healpix.jpl.nasa.gov
 ... CCRT^{}
 Centre de Calcul Recherche et Technologie.
 ... achieved^{}
 Higher resolution images are available at http://www.projethorizon.fr
 ... surveys^{}
 The convergence map is freely available for download at http://www.projethorizon.fr
All Figures
Figure 1: Fullsky simulated convergence map derived from the Horizon Simulation. Its resolution of 200 million pixels has been downgraded to fit the page. The various inserts display a zoom sequence into smaller and smaller areas of the sky. The pixel size is 0.74 arcmin^{2}. 

Open with DEXTER  
In the text 
Figure 2: Map of the cutsky used in Sect. 4 to compute highordermoments. 

Open with DEXTER  
In the text 
Figure 3: Moments of the convergence as a function of the average multipole moment on each wavelet scale. The variance, skewness, and kurtosis are shown as black, blue, and red lines, respectively. Solid lines with error bars corresponds ro a fullsky analysis, while dotted lines correspond to our cutsky analysis. 

Open with DEXTER  
In the text 
Figure 4: Angular power spectrum of the simulated convergence map (black solid line), compared to a fit based on the Smith et al. (2001) analytical model with error bars corresponding to our noise model (pink area). Also shown is the prediction from linear theory (pink dashed curve). The noise power spectrum is plotted as the dashed black line. The green solid line is the power spectrum of the error map obtained with the Wiener filter method, while the blue solid line are that for the MRLens method. 

Open with DEXTER  
In the text 
Figure 5: Reconstruction of convergence maps with our 2 filtering techniques. The top panels show the square map corresponding to first zoom sequence of Fig. 1. The bottom panels are subset of the corresponding top images with linear size 45'. From left to right, we show the original signal, the noisy image, the Wienerfiltered image and the the MRLensfiltered image. 

Open with DEXTER  
In the text 
Figure 6: Skewness (blue lines) and kurtosis (red lines) for the original convergence map (solid lines with error bars), compared to the same highorder statistics for the Wiener reconstructed map (dotted lines) and the MRLens reconstructed map (dashed lines). 

Open with DEXTER  
In the text 
Figure 7: Histogram of the residual maps for Wiener and MRLensfiltering. 

Open with DEXTER  
In the text 
Copyright ESO 2009