1: \begin{abstract}

2: 

3: We investigate how ideas from covariance localization in numerical

4: weather prediction can be used in Markov chain Monte Carlo (MCMC) sampling

5: of high-dimensional posterior distributions arising in Bayesian inverse problems.  

6: To localize an inverse problem is to enforce an anticipated ``local'' structure by

7: (\textit{i}) neglecting small off-diagonal elements of the prior precision and covariance matrices;

8: and (\textit{ii}) restricting the influence of observations to their neighborhood.

9: For linear problems we can specify the conditions 

10: under which posterior moments of the localized problem 

11: are close to those of the original problem.

12: We explain physical interpretations of our assumptions

13: about local structure

14: and discuss the notion of high dimensionality in local problems,

15: which is different from the usual notion of high dimensionality in function space MCMC.

16: The Gibbs sampler is a natural choice of MCMC algorithm for localized inverse problems

17: and we demonstrate that its convergence rate is independent of dimension

18: for localized linear problems.

19: Nonlinear problems can also be tackled efficiently by localization

20: and, as a simple illustration of these ideas, 

21: we present a localized Metropolis-within-Gibbs sampler.

22: Several linear and nonlinear numerical examples 

23: illustrate localization in the context of MCMC samplers for inverse problems.

24: 

25: \end{abstract}

26: