Regression , Density and Spectral Density Estimation

We consider posterior inference in wavelet based models for non-parametric regression with unequally spaced data, density estimation and spectral density estimation. The common theme in all three applications is the lack of posterior independence for the wavelet coe cients djk . In contrast, most commonly considered applications of wavelet decompositions in Statistics are based on a setup which implies a posteriori independent coe cients, essentially reducing the inference problem to a series of univariate problems. This is generally true for regression with equally spaced data, image reconstruction, density estimation based on smoothing the empirical distribution, time series applications and deconvolution problems. We propose a hierarchical mixture model as prior probability model on the wavelet coe cients. The model includes a level-dependent positive prior probability mass at zero, i.e., for vanishing coe cients. This implements wavelet coe cient thresholding as a formal Bayes rule. For non-zero coe cients we introduce shrinkage by assuming normal priors. Allowing di erent prior variance at each level of detail we obtain level-dependent shrinkage for non-zero coe cients. We implement inference in all three proposed models by a Markov chain Monte Carlo scheme which requires only minor modi cations for the di erent applications. Allowing zero coe cients requires simulation over variable dimension parameter space (Green 1995). We use a pseudo-prior mechanism (Carlin and Chib 1995) to achieve this.