Dependent Dirichlet Processes

A dependent Dirichlet process is a nonparametric Bayesian model in which the base distribution of a standard Dirichlet process varies as a function of covariates — time, space, or any other conditioning variable. Unlike the standard Dirichlet process, which assumes exchangeability and a single stationary mixing distribution, the dependent variant allows the clustering structure to evolve across the covariate domain, producing a collection of related random measures that share statistical strength while maintaining local flexibility.

The construction is typically achieved through stick-breaking representations in which the atom locations or weights depend on the covariate, or through generalizations of the Pólya urn scheme in which the seating probabilities shift according to local context. The dependent Dirichlet process is the nonparametric analog of a hierarchical linear model: it replaces the assumption that all observations come from the same population with the weaker assumption that nearby observations in covariate space come from similar populations.

The dependent Dirichlet process exposes the same design tension that appears in all nonparametric modeling: the trade-off between sharing information across the covariate space and respecting local variation. Too much sharing produces oversmoothing; too little produces noisy, unstable estimates. The covariate-dependent structure is not a convenience — it is a commitment to the principle that similarity in input space should imply similarity in model structure.