1: \begin{abstract}

2:       %

3:       Solving inverse problems without the use of derivatives or adjoints of the

4:       forward model is highly desirable in many applications arising in science and

5:       engineering. In this paper we propose a new version of such a methodology, a

6:       framework for its analysis, and numerical evidence of the practicality of the

7:       method proposed. Our starting point is an ensemble of over-damped Langevin

8:       diffusions which interact through a single preconditioner computed as the

9:       empirical ensemble covariance. We demonstrate that the nonlinear Fokker-Planck

10:       equation arising from the mean-field limit of the associated stochastic

11:       differential equation (SDE) has a novel gradient flow structure, built on the

12:       Wasserstein metric and the covariance matrix of the noisy flow. Using this

13:       structure, we investigate large time properties of the Fokker-Planck equation,

14:       showing that its invariant measure coincides with that of a single Langevin

15:       diffusion, and demonstrating exponential convergence to the invariant measure in

16:       a number of settings. We introduce a new noisy variant on ensemble Kalman

17:       inversion (EKI) algorithms found from the original SDE by replacing exact

18:       gradients with ensemble differences; this defines the ensemble Kalman sampler

19:       (EKS). Numerical results are presented which demonstrate its efficacy as a

20:       derivative-free approximate sampler for the Bayesian posterior arising from

21:       inverse problems.

22:       %

23:     \end{abstract}

24: