1: \begin{abstract}

2: 

3:   We present a new approach to Bayesian inference that entirely avoids

4:   Markov chain simulation, by constructing a map that \textit{pushes

5:     forward} the prior measure to the posterior measure. Existence and

6:   uniqueness of a suitable measure-preserving map is established by

7:   formulating the problem in the context of optimal transport theory.

8:   %

9:   We discuss various means of explicitly parameterizing the map and

10:   computing it efficiently through solution of an optimization

11:   problem, exploiting gradient information from the forward model when

12:   possible. The resulting algorithm overcomes many of the

13:   computational bottlenecks associated with Markov chain Monte Carlo.

14:   %

15:   Advantages of a map-based representation of the posterior include

16:   analytical expressions for posterior moments and the ability to

17:   generate arbitrary numbers of independent posterior samples without

18:   additional likelihood evaluations or forward solves. The

19:   optimization approach also provides clear convergence criteria for

20:   posterior approximation and facilitates model selection through

21:   automatic evaluation of the marginal likelihood.

22:   %

23:   We demonstrate the accuracy and efficiency of the approach on

24:   nonlinear inverse problems of varying dimension, involving the

25:   inference of parameters appearing in ordinary and partial

26:   differential equations.

27:   % ALT:

28:   % We demonstrate the accuracy and efficiency of the method on

29:   % inference problems of varying dimension, focusing on point and

30:   % distributed parameters appearing in ordinary and partial

31:   % differential equations.

32: 

33: \end{abstract}

34: