Closed AnderGray closed 7 months ago
I'll investigate.
Next time please just paste the code here instead of uploading a zip file.
Turns out the issue here is using a sample size where N*target
leads to less samples in the levels >=2
. For example using 205 and 0.1 leads to 189 samples in the subsets because of the way the number of chains and samples per chain are calculated
number_of_seeds = Int64(max(1, ceil(sim.n * sim.target)))
samples_per_seed = Int64(floor(sim.n / number_of_seeds))
here 21 seeds and 9 chains = 189 != 205.
However the sample size passed to the cov estimation is always n
.
This is easily fixed by computing n
for the estimation from the available samples.
Performing convergence study for MC, subset, and subset infinity.
Sometimes (randomly)
estimate_cov
is trying to take the sqrt of a complex number. Script for the study