Efficient data augmentation techniques for Gaussian state space models
share
We propose a data augmentation scheme for improving the rate of convergence of the EM algorithm in estimating Gaussian state space models. The scheme is based on a linear transformation of the latent states, and two working parameters are introduced for simultaneous rescaling and re-centering. A variable portion of the mean and scale are thus being moved into the missing data. We derive optimal values of the working parameters (which maximize the speed of the EM algorithm) by minimizing the fraction of missing information. We also study the large sample properties of the working parameters and their dependence on the autocorrelation and signal-to-noise ratio. We show that instant convergence is achievable when the mean is the only unknown parameter and this result is extended to Gibbs samplers and variational Bayes algorithms.
READ FULL TEXT