1: \begin{abstract}

2: Given i.i.d.\ data from an unknown distribution, we consider the

3: problem of predicting future items. An adaptive way to estimate the

4: probability density is to recursively subdivide the domain to an

5: appropriate data-dependent granularity. In Bayesian inference one

6: assigns a data-independent prior probability to ``subdivide'', which

7: leads to a prior over infinite(ly many) trees. We derive an exact,

8: fast, and simple inference algorithm for such a prior, for the data

9: evidence, the predictive distribution, the effective model

10: dimension, moments, and other quantities. We prove asymptotic

11: convergence and consistency results, and illustrate the behavior of

12: our model on some prototypical functions.

13: \def\contentsname{\centering\normalsize Contents}

14: {\parskip=-2.7ex\tableofcontents}

15: \end{abstract}

16: