Random sampling of partitions and contingency tables: Two practical examples of the Burnside process

Abstract: The Burnside process is a general algorithm for sampling a uniformly chosen orbit of a finite group $G$ acting on a finite set $\mathcal{X}$. For example, if $\mathcal{X} = G$ and $G$ acts on itself by conjugation ($s^t = t<sup>{-1}st$), then the orbits are conjugacy classes. When $G$ is the symmetric group $S_n$, the conjugacy classes are indexed by partitions of $n$, so the Burnside process gives a way to sample partitions. If $\mathcal{X}=S_n$ and $G$ is a product of symmetric groups, then the orbits are labeled by contingency tables: non-negative integer arrays with given row and column sums. Actually carrying out the Burnside process requires new combinatorics and group theory. This is worked out and illustrated for these two examples. For partitions, we also developed a new Markov chain called the reflected Burnside process which greatly improves the mixing of the Burnside process.