How much is lost in pairwise correlations: the case of phylogenetics

Among the many popular techniques for reconstructing evolutionary trees from molecular sequences, so-called distance-matrix methods are typically the fastest. This speed stems from a straightforward, intuitive approach: repeated merging of the closest clusters of sequences. However, unlike more elaborate techniques such as maximum likelihood, distance-matrix methods only exploit empirical correlations between pairs of sequences. This limited use of the data is often cited as a serious weakness. In this talk, I will discuss surprising theoretical results shedding some light on this question.