Séminaire Doctoral / Seminar PhD |
| « The Lion and the Mouse fable for cosmological simulations: why you want a very poor estimator to help you » |
|
Nicolas Chartier |
| The next generations of galaxy surveys, such as the Euclid mission, will analyze the distribution of matter within broad cosmological volumes. Such studies require large N-body simulations catalogues to compute statistics such as the matter power spectrum, the halo mass function, the matter bispectrum, as well as their higher order statistics. Moreover, likelihood-free inference frameworks require forward simulations only, and are a relevant solution when the likelihood is intractable and when we do not want to rely on ad hoc calibrations. Robust mock data is necessary for each cosmological model to be tested, and the computational cost remains the bottleneck for future datesets. For instance, the Quijote Simulations of Villaescusa-Navarro et al (2019), required 35 million CPU core hours. While it seems, at first glance, that you have to choose between unbiased and costly full N-body simulations on the one hand and cheap and approximate methods on the other, we will examine in this presentation a variance reduction technique designed to get the best of both worlds. Results on common cosmological statistics will be presented to demonstrate that you can actually do much better than the standard sample mean when it comes to unbiased estimation. |
|
vendredi 8 mai 2020 - 16:00 Webinaire, Institut d'Astrophysique |
| Page web du séminaire / Seminar's webpage |