Authors: David Gillman, Jacob Platnick, Dana Randall

Weighted equitable partitioning of a graph has been of interest lately due to several applications, including redistricting, network algorithms, and image decomposition. Weighting a partition according to the spanning-tree metric has been of mathematical and practical interest because it typically favors partitions with more compact pieces. An appealing algorithm suggested by Charikar et al. is to sample a random spanning tree and remove k-1 edges, producing a random forest. If the components of the forest form a balanced partition, the partition is equitable under an easily computed acceptance probability. Cannon et al. recently showed that spanning trees on grid graphs and grid-like graphs on $n$ vertices are splittable into $k$ equal sized pieces with probability at least $n^{-2k}$, leading to the first rigorous sampling algorithm for a class of graphs. We present complementary results showing that spanning trees on dense random graphs also have inverse polynomial probability of being splittable, giving another class of graphs where equitable partitions can be efficiently sampled exactly. These proofs also guarantee fast almost-uniform sampling for the up-down walk on forests, giving another provably efficient randomized method for generating equitable partitions. Further, we show that problems with the well-studied ReCom algorithm for equitable partitioning are more extensive than previously known, even in special cases that were believed to be more promising. We present a family of graphs where the Markov chain fails to be irreducible when it must keep the components perfectly equitable; yet when the chain is allowed an imbalance of just one vertex between components, the rejection sampling step may take exponential time. This is true even when the graph satisfies desirable properties that have been conjectured to be sufficient for fast sampling.

Read original post