Contrarily to sampling approaches, a model is assumed (the parametrised family), implying a bias but also a lower variance. In general VI methods are less accurate that MCMC ones but produce results much faster: these metho…
Contrarily to VI methods described in the next section, MCMC approaches assume no model for the studied probability distribution (the posterior in the Bayesian inference case). As a consequence, these methods have a low bias but a high variance and it implies that results are most of the time more costly to obtain but also more accurate than the one we can get from VI.