1: \begin{abstract}

2: We study the asymptotic behaviour of the posterior distribution in a

3: broad class of statistical

4: models where the ``true'' solution occurs on the boundary of the

5: parameter space. We show that

6: in this case Bayesian inference is consistent, and that the posterior

7: distribution has not only

8: Gaussian components as in the case of regular models (the

9: Bernstein--von Mises theorem) but also

10: has Gamma distribution components whose form depends on the behaviour

11: of the prior distribution

12: near the boundary and have a faster rate of convergence. We also

13: demonstrate a remarkable property of Bayesian inference, that for some models,

14: there appears to be no bound on efficiency of estimating the unknown

15: parameter if it is on the boundary of the parameter space.

16: We illustrate the results on a problem from emission tomography.

17: \end{abstract}