Recasts the mini-bucket approximation as exact inference on a simplified model made by splitting nodes, so that approximations can borrow every advance in exact inference.
Mini-bucket is a well-known way to approximate inference. The paper shows it can be understood as exact inference on an approximate model, one produced by splitting nodes in the network. Three things follow. Branch-and-bound searches that use mini-bucket bounds can work in a drastically smaller space. The view suggests new mini-bucket heuristics and explains existing ones. And it lets these approximations benefit directly from advances in exact inference, which extends their reach.
It shows the traffic running the other way: progress in exact reasoning improving the approximate methods most of the field had moved to.
A methods paper on bounds. No product claims attach to it.