Beyond Causal Discovery for Astronomy: Learning Meaningful Representations with Independent Component Analysis

Abstract

We present the first steps toward applying causal representation learning to astronomy. Following up on previous work that introduced causal discovery to the field for the first time, here we solve a long standing conundrum by identifying the direction of the causal relation between supermassive black hole (SMBH) mass and their host galaxy properties. This leverages a score-based causal discovery approach with an exact posterior calculation. Causal relations between SMBHs and their host galaxies are further clarified by Independent Component Analysis (ICA). The astrophysical problem we focus on is one of the most important open issues in the field and one that has not seen a definitive resolution in decades. We consider the space of six physical properties of galaxies, subdivided by morphology: elliptical, lenticular, and spiral, plus SMBH mass. We calculate an exact posterior over the space of directed acyclic graphs for these variables based on a flat prior and the Bayesian Gaussian equivalent score. The nature of the causal relation between galaxy properties and SMBH mass is found to vary smoothly with morphology, with galaxy properties determining SMBH mass in ellipticals and vice versa in spirals. This settles a long-standing debate and is compatible with our theoretical understanding of galaxy evolution. ICA reveals a decreasing number of meaningful Independent Components (ICs) from ellipticals and lenticular to spiral. Moreover, we find that only one IC correlates with SMBH mass in spirals while multiple ones do in ellipticals, further confirming our finding that SMBH mass causes galaxy properties in spirals, but the reverse holds in ellipticals.