Evaluation of MCMC methods
Alexandre Bouchard-Côté
How to empirically evaluate the performance of MCMC methods
How to investigate the performance of MCMC methods using empirical methods? E.g. for the purpose of
Comparing two algorithms
Understanding how algorithms scale with number of data points/parameters/parallelization
Most common measure of sampling quality: Effective Sample Size
Intuition: the number of independent samples which would give similar Monte Carlo approximation error as your MCMC output
Based on Central Limit Theorems for Markov chains
To estimate ESS from MCMC output: best method out there is the batch mean method
https://projecteuclid.org/euclid.aos/1266586622
R package
To get the full picture:
Use ESS per unit of time (e.g. time to run the software, or number of target evaluation if comparable across methods)
Use log-log plot to approximate scaling trends (as in
https://www.stat.ubc.ca/~bouchard/courses/stat520-sp2020-21/T9-consistency.html#(7)
)
Other methods:
Looking at convergence of Monte Carlo averages
https://arxiv.org/abs/1703.01717
and
https://arxiv.org/abs/1611.06972
https://arxiv.org/pdf/1712.06006.pdf
In the context of parallel tempering, round trip rates (see e.g.
https://arxiv.org/pdf/1905.02939.pdf
)