1: \begin{abstract}

2: This paper studies posterior concentration behavior of the base 

3: probability measure $G$ of a Dirichlet measure $\DP_{\alpha G}$,

4: given observations associated with $m$ Dirichlet processes sampled

5: from $\DP_{\alpha G}$, as $m$ and the number of observations 

6: $m\times n$ tend to infinity. The base measure itself is

7: endowed with another Dirichlet prior, 

8: a construction known as the hierarchical Dirichlet 

9: processes~\citep{Teh-etal-06}. 

10: %The Dirichlet processes sampled from $\Dcal$

11: %serve as mixing measures for an exchangeable collection of $m$

12: %mixture distributions, posterior concentration behavior of 

13: %which is also investigated. 

14: Convergence rates are established in transportation distances (i.e. 

15: Wasserstein metrics) under various geometrically sparse 

16: conditions on the support of the true base measure. 

17: As a consequence of the theory we demonstrate

18: the benefit of ``borrowing strength'' in the inference of multiple

19: groups of data --- a 

20: heuristic argument commonly used to motivate hierarchical modeling.

21: In certain settings, the gain in efficiency due to the latent hierarchy

22: can be dramatic, improving

23: from a standard nonparametric rate to a parametric rate of 

24: convergence. Tools developed include transportation distances for

25: nonparametric Bayesian hierarchies of random measures, the existence of tests for Dirichlet measures, 

26: and geometric properties of the support of Dirichlet measures.

27: \end{abstract}

28: