But, I'm seeing -H be too large when more phantom points are retained. It should be similar to the amount of compression needs to reach the typical set. Similarly I saw a scaling factor in ESS and logZ_uncert of k+1. It might be that log_dp_mean should be divided by k+1 to solve both problems. Something about the actual number of effective samples vs how shrinkage is computed. Needs to be investigated.
Describe the bug
But, I'm seeing
-H
be too large when more phantom points are retained. It should be similar to the amount of compression needs to reach the typical set. Similarly I saw a scaling factor in ESS and logZ_uncert ofk+1
. It might be thatlog_dp_mean
should be divided byk+1
to solve both problems. Something about the actual number of effective samples vs how shrinkage is computed. Needs to be investigated.JAXNS version 2.3.x, 2.4.0