1: \begin{abstract}

2:   Among Monte Carlo techniques, the importance sampling requires fine

3:   tuning of a proposal distribution, which is now fluently resolved

4:   through iterative schemes. 

5:   The Adaptive Multiple Importance Sampling (AMIS) of \cite{cornuet:2012}

6:   provides a significant improvement in stability and effective sample size

7:   due to the introduction of a recycling

8:   procedure. % These numerical properties are particularly

9:   % adapted to the Bayesian paradigm in population genetics where the

10:   % modelization involves a large number of parameters.

11:   However, the consistency of the AMIS estimator remains largely open.

12:   In this work we prove the convergence of the AMIS, at a cost of a

13:   slight modification in the learning process.  Contrary to

14:   \cite{douc:2007:a}, results are obtained here in the asymptotic

15:   regime where the number of iterations is going to infinity while the

16:   number of drawings per iteration is a fixed, but growing sequence of

17:   integers. Hence some of the

18:   results shed new light on adaptive population Monte

19:   Carlo algorithms in that last regime. % Finally, first numerical

20:   % experiments exhibit that the modification of the scheme might even

21:   % improve the original AMIS.

22:   % In this

23:   % work, we provide a novel Adaptive Multiple Importance Sampling

24:   % scheme corresponding to a slight modification of \cite{cornuet:2012}

25:   % proposition that preserves the above-mentioned improvements.

26:   % Finally, using limit theorems on triangular arrays of conditionally

27:   % independent random variables, we give a consistency result for the

28:   % final particle system returned by our new scheme.

29: \end{abstract}

30: