Implementation of "Unbiased Markov-Chain Monte Carlo estimators with couplings". Project done as part of the Bayesian Machine Learning course by Rémi Bardenet and Julyan Arbel.
It includes a coupled Metropolis-Hasting algorithm, the maximal coupling and its generalization, the modified Thorisson's Algorithm.
The R companion-code of the paper is available at https://github.com/pierrejacob/unbiasedmcmc.
This package requires jax
and numpy
. You can use pip install -e .
to install the package.
See the experiments
folder for some examples of how to use the package.
- Pierre E. Jacob, John O'Leary, Yves F. Atchadé, "Unbiased Markov-Chain Monte Carlo estimators with couplings", 2020
- Yves F. Atchadé, Pierre E. Jacob, "Unbiased Markov Chain Monte Carlo: what, why and how"