Improving model convergence and sampling speed

@junpenglao - is there any literature about the amount by which ADVI underestimates, or the conditions under which underestimation is most severe for mean-field ADVI? Are other approximtions, such as fullrank ADVI also prone to this problem?