1: \begin{abstract}

2: Bayesian inverse problems often involve sampling posterior distributions on infinite-dimensional function spaces. Traditional Markov chain Monte Carlo (MCMC) algorithms are characterized by deteriorating mixing times upon mesh-refinement, when the finite-dimensional 

3: approximations become more accurate.

4: Such methods are typically forced to reduce step-sizes as the discretization gets finer, and thus are expensive as a function 

5: of dimension. Recently, a new class of MCMC methods with mesh-independent convergence times has emerged.

6: However, few of them take into account the geometry of the posterior informed by the data. 

7: At the same time, recently developed geometric MCMC algorithms have been 

8: found to be powerful in exploring complicated distributions that deviate 

9: significantly from elliptic Gaussian laws, but are in general computationally

10: intractable for models defined in infinite dimensions. In this work, we 

11: combine geometric methods on a finite-dimensional subspace with 

12: mesh-independent infinite-dimensional approaches. Our objective is to speed 

13: up MCMC mixing times, without significantly increasing the computational 

14: cost per step (for instance, in comparison with the vanilla preconditioned Crank-Nicolson 

15: (pCN) method). This is achieved by using ideas from geometric MCMC to

16: probe the complex structure of an intrinsic finite-dimensional subspace 

17: where most data information concentrates, while retaining robust mixing 

18: times as the dimension grows by using pCN-like methods in the

19: complementary subspace. The resulting algorithms are demonstrated in

20: the context of three challenging inverse problems arising in subsurface flow,

21: heat conduction and incompressible flow control. The algorithms

22: exhibit up to two orders of magnitude improvement in sampling efficiency 

23: when compared with the pCN method.

24: \end{abstract}

25: