A forward-modeling approach to cosmic shear

Abstract

The current cosmological concordance model, ΛCDM, is very successful at describing the statistical properties of the Universe and its evolution with cosmic time at both low and high redshifts. However, two major ingredients of ΛCDM, cold dark matter (CDM) and dark energy (Λ), are only phenomenologically motivated and cosmologists lacks deeper understanding of their origins. Therefore, investigating the physical nature of this dark sector of the ΛCDM model is one of the most pressing issues in modern cosmology and multiple major observational programs aimed at investigating the dark components of ΛCDM are either on the way or already in operation. At low redshifts, three major wide-field surveys, the Kilo-Degree Survey (KiDS), the Dark Energy Survey (DES) and the Hyper Suprime-Cam (HSC) survey, have recently published updated cosmology constraints. They all rely on cosmic shear, the weak gravitational lensing by large-scale structures, as a powerful probe of both the expansion history of the Universe and the growth of structures. While cosmic shear has great potential to shine light on the dark sector of ΛCDM, the effect is challenging to measure and prone to systematic effects. Therefore, Refregier & Amara (2014, DOI: 10.1016/j.dark.2014.01.002) proposed the Monte-Carlo Control Loops (MCCL) framework. This method employs large amounts of forward simulations to quantify the systematic uncertainty of cosmic shear measurements and propagate it through the analysis in a probabilistic way. In this thesis, we develop methods for measuring cosmology with cosmic shear based on the MCCL framework. We first implement and test a forward-modeling approach to measuring the redshift distribution n(z) of typical weak lensing samples. To this end, we devise an empirical model of the intrinsic galaxy population based on redshift-dependent luminosity functions. We then use Approximate Bayesian Computation (ABC) to adjust our simulations to survey data in a Bayesian framework. This yields a family of likely posterior n(z) curves which quantifies the uncertainty of the measurement. Moreover, we develop a method for fast point spread function (PSF) estimation and modeling based on Deep Learning, specifically a convolutional neural network (CNN). Once trained, the computational speed of this algorithm allows it to be used within the MCCL framework to analyze large volumes of synthetic data. Based on the methods described above, we next present the first end-to-end application of the MCCL framework to survey data. In a non-tomographic setup, we constrain cosmology with cosmic shear using the DES Year (Y1) data. The core of our method is the joint measurement of the shear 2-point function and the associated redshift distribution. By simulating the full survey footprint numerous times, we quantify the systematic uncertainty of our analysis and are furthermore able to disentangle statistical and systematic errors. Building on this achievement, we implement a tomographic shear pipeline for the DES Year 3 (Y3) data. We classify galaxies into redshift bins with a machine-learning approach which enables us to measure tomographic shear 2-point functions along with the redshift distributions. The current results with this pipeline offer great prospects for applying the MCCL framework to current and future tomographic weak lensing datasets.