Leveraging machine learning for optimal cosmological parameter extraction
Simulation-based inference (SBI) offers a transformative approach to cosmological data analysis by moving beyond traditional methods that rely on assumed likelihood forms, approximate covariance matrices, or perturbative models. Instead, SBI leverages simulations together with neural network to directly learn the relationship between cosmological parameters and observables.
Galaxy skew spectra provide an efficient way to compress the information contained in three-point statistics into a form that resembles two-point functions, making them particularly powerful for extracting amplitude-like parameters such as primordial non-Gaussianities predicted by extensions of standard inflation. They are constructed by correlating a single density field with a weighted pairs of galaxies given by theory-motivated functions. As a compressed statistic, skew spectra help overcome key challenges of higher-order analyses, notably the high computational cost of estimators and the difficulty of covariance estimation.
Our application of SBI to skew spectra analysis of BOSS data demonstrates significant improvements in cosmological parameter constraints, achieving up to 30% improvement in precision for key cosmological parameters compared to traditional power spectrum analysis alone.
Related Publications: JH, Moradinezhad et al. (2024), arXiv:2401.15074
JH, Moradinezhad et al. (2022), arXiv:2210.12743