1: \begin{abstract}
2:
3:
4: % !!!! 300 words limit
5: In this paper, we study efficient approximate sampling for probability distributions known up to normalization constants. We specifically focus on a problem class arising in Bayesian inference for large-scale inverse problems in science and engineering applications. The computational challenges we address with the proposed methodology are:
6: (i) the need for repeated evaluations of expensive forward models;
7: (ii) the potential existence of multiple modes; and (iii) the fact that
8: gradient of, or adjoint solver for, the forward model might not be feasible.
9:
10:
11: While existing Bayesian inference methods meet some of these challenges individually,
12: we propose a framework that tackles all three systematically.
13: Our approach builds upon the Fisher-Rao gradient flow in probability space, yielding a dynamical system for probability densities that converges towards the target
14: distribution at a uniform exponential rate. This rapid convergence is advantageous
15: for the computational burden outlined in (i). We apply Gaussian mixture approximations with operator splitting techniques to simulate the flow numerically; the resulting
16: approximation can capture multiple modes thus addressing (ii).
17: Furthermore, we employ the Kalman methodology to facilitate a derivative-free update
18: of these Gaussian components and their respective weights,
19: addressing the issue in (iii).
20:
21: The proposed methodology results in an efficient derivative-free sampler flexible
22: enough to handle multi-modal distributions:
23: Gaussian Mixture Kalman Inversion (GMKI).
24: The effectiveness of GMKI is demonstrated both theoretically and numerically in several experiments with multimodal target distributions, including proof-of-concept
25: and two-dimensional examples, as well as
26: a large-scale application: recovering the Navier-Stokes initial condition from solution data at positive times.
27: \end{abstract}
28: