Decomposable (probabilistic) models are log-linear models generated by acyclic hypergraphs, and a number of nice properties enjoyed by them are known. In many applications the following selection problem naturally arises: given a probability distribution $p$ over a finite set $V$ of $n$ discrete variables and a positive integer $k$, find a decomposable model with tree-width $k$ that best fits $p$. If $\mathcal{H}$ is the generating hypergraph of a decomposable model and $p_{\mathcal{H}}$ is the estimate of $p$ under the model, we can measure the closeness of $p_{\mathcal{H}}$ to $p$ by the information divergence $D(p: p_{\mathcal{H}})$, so that the problem above reads: given $p$ and $k$, find an acyclic, connected hypergraph ${\mathcal{H}}$ of tree-width $k$ such that $D(p: p_{\mathcal{H}})$ is minimum. It is well-known that this problem is $NP$-hard. However, for $k = 1$ it was solved by Chow and Liu in a very efficient way; thus, starting from an optimal Chow--Liu solution, a few forward-selection procedures have been proposed with the aim at finding a `good' solution for an arbitrary $k$. We propose a backward-selection procedure which starts from the (trivial) optimal solution for $k=n-1$, and we show that, in a study case taken from literature, our procedure succeeds in finding an optimal solution for every $k$.
decomposable model, acyclic hypergraph, backward selection, information divergence, $k$-hypertree
05C65, 62-09, 68R10, 68T05