A target distribution is the probability distribution a sampling procedure aims to reproduce. It specifies what the generated samples should represent, even when the procedure reaches it indirectly. section 16.6 section 15.6
Different sampling routes use different information about the target:
- A known quantile function transforms a standard uniform sample directly into a target-distributed sample. section 15.6
- A PDF can support rejection sampling through a helper distribution, although finding a suitable helper and avoiding repeated rejection can be difficult. section 15.6
- The Metropolis algorithm uses information about high-weight regions to construct a Markov process whose states sample the target distribution. section 16.6
Updating uncertainty
A prior distribution captures beliefs before observations; combining it with a likelihood function produces a posterior distribution that updates uncertainty. That posterior can serve as the target for continuous sampling. section 16.6
