Gibbs and autoregressive Markov processes (Q2467382): Difference between revisions

The authors compare Gibbs sampler processes to autoregressive Markov processes with the same marginals. E.g., for the beta distribution the Gibbs sampler is \(\pi_k\) with \[ \pi_k|u_{k-1}\sim \text{Beta}(a+u_{k-1},b+c-u_{k-1}), \;u_{k-1}|\pi_{k-1}\sim \text{Binomial}(c,\pi_{k-1}), \] and the Markov autoregressive process is \[ \pi_k=w_k\pi_{k-1}+(1-w_k)z_k, \] where \(w_k\sim \text{Beta}(a+b,c)\), \(z_k\sim \text{Beta}(y_k,c-y_k)\), \(y_k\sim \text{Binomial}(c, \text{Beta}(a,b))\). Analogous constructions are considered for gamma distributions and for distributions generated by a Dirichlet process. For gamma autoregressive processes convergence to a Lévy process is demonstrated.