is anybody aware of any results for importance sampling, where we don't have access to the true unnormalized target density, but only to an unbiased estimator of its log density? I'm thinking e.g. about a setup where we may wanna estimate posterior densities on mini batches