1: \begin{abstract} 

2: 	The pairwise influence matrix of Dobrushin has long been used as an analytical tool to bound the rate of convergence of Gibbs sampling.

3: 	In this work, we use Dobrushin influence as the basis of a practical tool to certify and efficiently improve the quality of a discrete Gibbs sampler.  

4: 	Our Dobrushin-optimized Gibbs samplers (DoGS) 

5: 	offer

6: 	customized variable selection orders for a given sampling budget and variable subset of interest,

7: 	 explicit bounds on total variation distance to stationarity,

8: 	and certifiable improvements over the standard systematic and uniform random scan Gibbs samplers.

9: 	In our experiments with joint image segmentation and object recognition, Markov chain Monte Carlo maximum likelihood estimation, and Ising model inference,

10: 	DoGS consistently deliver higher-quality inferences with significantly smaller sampling budgets than standard Gibbs samplers.

11: 	

12: 

13: \end{abstract}

14: