1: \begin{abstract}  

2: Many problems arising in applications result in 

3: the need to probe a probability distribution for functions.

4: Examples include Bayesian nonparametric statistics

5: and conditioned diffusion processes.

6: Standard MCMC algorithms typically become arbitrarily

7: slow under the mesh refinement dictated 

8: by nonparametric description of the unknown function.

9: We describe an approach to modifying a whole 

10: range of MCMC methods which ensures 

11: that their speed of convergence is

12: robust under mesh refinement. 

13: In the applications of interest the data is often sparse

14: and the prior specification is an essential part

15: of the overall modeling strategy. 

16: The algorithmic approach that we describe

17: is applicable whenever the

18: desired probability measure has density with

19: respect to a Gaussian process or Gaussian random field 

20: prior, and to some useful non-Gaussian priors constructed

21: through random truncation. 

22: Applications are shown in density estimation,

23: data assimilation in fluid mechanics, subsurface

24: geophysics and image registration. The key design

25: principle is to formulate the MCMC method for

26: functions. This leads to algorithms which

27: can be implemented via minor modification of

28: existing algorithms, yet which show enormous

29: speed-up on a wide range of applied problems. 

30: \end{abstract}