> Which, for many things I'd care about, if I knew the underlying distribution that well, I probably don't need to sample You don't have to sample directly. Th

> Which, for many things I'd care about, if I knew the underlying distribution that well, I probably don't need to sample You don't have to sample directly. The entire field of Bayesian variational learning exist to deal with that very problem. Look up Markov chain Monte Carlo, Metropolis algorithm, conjugate priors, reparametrization tricks.