Optimal low-rank approximations for linear Gaussian inverse problems on Hilbert spaces, Part II: posterior mean approximation

Abstract: In this work, we construct optimal low-rank approximations for the Gaussian posterior distribution in linear Gaussian inverse problems. The parameter space is a separable Hilbert space of possibly infinite dimension, and the data space is assumed to be finite-dimensional. We consider various types of approximation families for the posterior. We first consider approximate posteriors in which the means vary among a class of either structure-preserving or structure-ignoring low-rank transformations of the data, and in which the posterior covariance is kept fixed. We give necessary and sufficient conditions for these approximating posteriors to be equivalent to the exact posterior, for all possible realisations of the data simultaneously. For such approximations, we measure approximation error with the Kullback-Leibler, R\'enyi and Amari $\alpha$-divergences for $\alpha\in(0,1)$, and with the Hellinger distance, all averaged over the data distribution. With these losses, we find the optimal approximations and formulate an equivalent condition for their uniqueness, extending the work in finite dimensions of Spantini et al. (SIAM J. Sci. Comput. 2015). We then consider joint approximation of the mean and covariance, by also varying the posterior covariance over the low-rank updates considered in Part I of this work. For the reverse Kullback-Leibler divergence, we show that the separate optimal approximations of the mean and of the covariance can be combined to yield an optimal joint approximation of the mean and covariance. In addition, we interpret the joint approximation with the optimal structure-ignoring approximate mean in terms of an optimal projector in parameter space.