1: \begin{abstract}

2:   There is an increasing interest in estimating expectations outside of the classical

3:   inference framework, such as for models expressed as probabilistic programs. Many of

4:   these contexts call for some form of \emph{nested inference} to be applied.  In this

5:   paper, we analyse the behaviour of nested Monte Carlo (NMC) schemes, for which classical

6:   convergence proofs are insufficient. We give conditions under which NMC will converge,

7:   establish a rate of convergence, and provide empirical data that suggests that this rate

8:   is observable in practice. Finally, we prove that general-purpose nested inference

9:   schemes are inherently biased.  Our results serve to warn of the dangers associated with

10:   na\"ive composition of inference and models.

11: %Improvements in approximate inference techniques and computing hardware have been

12: %increasing the scope of models for which one can reasonably hope to estimate expectations.

13: %In particular, probabilistic programming systems are making it increasingly simple to

14: %define problems that fall outside of the classical inference framework.  Many of these

15: %problems fall into a \emph{nested inference} setting, for which classical convergence

16: %proofs are insufficient.  In this paper, we consider the convergence characteristics of

17: %Monte Carlo in this more general setting and provide results that serve as a warning about

18: %the dangers of na\"ive composition of inference algorithms and models.  We show that

19: %general-purpose nested inference schemes are inherently biased, require significantly

20: %stricter assumptions to guarentee converge than standard Monte Carlo, and have convergence

21: %rates that exponentially decrease with the depth of the nesting.

22: \end{abstract}

23: