1: \begin{abstract}

2: This article develops a general-purpose adaptive sampler that approximates the

3: target density by a mixture of multivariate $t$ densities. The adaptive

4: sampler is based on reversible proposal distributions each of which has the

5: mixture of multivariate $t$ densities as its invariant density. The reversible

6: proposals consist of a combination of independent and correlated steps that

7: allow the sampler to traverse the parameter space efficiently as well as

8: allowing the sampler to keep moving and locally exploring the parameter space.

9: We employ a two-chain approach, in which a trial chain is used to adapt the

10: proposal densities used in the main chain. Convergence of the main chain

11: and a strong law of large numbers are proved under reasonable conditions, and

12: without imposing a Diminishing Adaptation condition. The mixtures of

13: multivariate $t$ densities are fitted by an efficient Variational

14: Approximation algorithm in which the number of components is determined

15: automatically. The performance of the sampler is evaluated using simulated and

16: real examples. Our autocorrelated framework is quite general and can handle mixtures

17: other than  multivariate $t$.

18: 

19: \vspace{1.3ex plus 0.7ex minus 0.5ex}\noindent\textbf{\boldmath{Keywords.}}

20:  Ergodic convergence; Markov Chain Monte Carlo; Metropolis-within Gibbs composite sampling;

21: Multivariate

22: $t$ mixtures; Simulated annealing; Variational Approximation.

23: 

24: \end{abstract}