Default priors for density estimation with mixture models

The infinite mixture of normals model has become a popular method for density estimation problems. This paper proposes an alternative hierarchical model that leads to hyperparameters that can be interpreted as the location, scale and smoothness of the density. The priors on other parts of the model have little effect on the density estimates and can be given default choices. Automatic Bayesian density estimation can be implemented by using uninformative priors for location and scale and default priors for the smoothness. The performance of these methods for density estimation are compared to previously proposed default priors for four data sets.