1: \begin{abstract}

2: Markov chain Monte Carlo (MCMC) samplers are numerical methods for drawing samples from a given target probability distribution. We discuss one particular MCMC sampler, 

3: the MALA-within-Gibbs sampler, from the theoretical and practical perspectives.

4: We first show that the acceptance ratio and step size of this sampler are independent of 

5: the overall problem dimension

6: when (\emph{i}) the target distribution has sparse conditional structure,

7: and (\emph{ii}) this structure is reflected in the partial updating strategy of MALA-within-Gibbs.

8: If, in addition, the target density is block-wise log-concave, 

9: then the sampler's convergence rate is independent of dimension. 

10: From a practical perspective, we expect that MALA-within-Gibbs is useful for solving high-dimensional

11: Bayesian inference problems where the posterior exhibits sparse conditional structure at least approximately. 

12: In this context, a partitioning of the state that correctly reflects

13: the sparse conditional structure must be found, and we illustrate this

14: process in two numerical examples.

15: %

16: We also discuss trade-offs between the block size used for partial

17: updating and computational requirements that may increase with the

18: number of blocks. 

19: %\todo{ymm: I thought the earlier version dwelled on

20:  % this issue too much and propose a more compact comment here.}

21: %

22: % From a practical perspective, 

23: % dimension independence of the acceptance ratio or convergence rate

24: % does not imply that the computational cost of the sampler is dimension independent.

25: % We explain that this leads to a trade-off between the block size,

26: % used for the partial updating, and the computational requirements that may increase with the number of blocks. 

27: \end{abstract}

28: