variational inference (VI) provides fast approximations of a Bayesian
posterior in part because it formulates posterior approximation as an
optimization problem: to find the closest distribution to the exact posterior
over some family of distributions. For practical reasons, the family