Optimal and scalable methods to approximate the solutions of large-scale Bayesian problems: Theory and application to atmospheric inversions and data assimilation.