1: \begin{abstract}

2: Variational Inference (VI) is a popular alternative to asymptotically exact sampling in Bayesian inference. Its main workhorse is optimization over a reverse Kullback-Leibler divergence (RKL), which typically underestimates the tail of the posterior leading to miscalibration and potential degeneracy. 

3: Importance sampling (IS), on the other hand, is often used to fine-tune and de-bias the estimates of approximate Bayesian inference procedures. 

4: The quality of IS crucially depends on the choice of the proposal distribution. 

5: % Coincidentally, underestimation of the tail in the proposal is also a serious hindrance of its application.

6: Ideally, the proposal distribution has heavier tails than the target, which is rarely achievable by minimizing the RKL.

7: We thus propose a novel combination of optimization and sampling techniques for approximate Bayesian inference by constructing an IS proposal distribution through the minimization of a forward KL (FKL) divergence. 

8: This approach guarantees asymptotic consistency and a fast convergence towards both the optimal IS estimator and the optimal variational approximation.

9: We empirically demonstrate on real data that our method is competitive with variational boosting and MCMC.

10: \end{abstract}

11: