That’s essentially what I built, but you have to figure out where that log density threshold is, so for points along the optimization paths I spawned short MCMC chains. If the log density of those went down, they were in the tail, if they went up, they were past the typical set toward the mode, and if they bounced around, they were in the typical set.
How do you use importance sampling? I thought you needed to know the proposal distribution.
OK.