y Rutgers University.pdfVIP

On the Neighborhoods of Neighbor Joining TreesMartin Farach Steve FerenciyRutgers UniversityAugust 8, 1996AbstractWe consider the problem of tting an n  n distance matrix D by a tree metric T . Thisproblem is NP-hard for most reasonable distance functions between D and T . Within thebiology community, the so-called Neighbor-Joining (NJ) heuristic [16] has wide acceptance. Weexamine the local optimality of NJ trees.We provide a systematic approach to local optimization for NJ trees. Further, we explorethe pros and cons of the two most popular criteria for tree tting optimization: parsimony andL2. We show that these two exhibit very dierent behavior for local optimization and discusssome of the implications these dierences may have for practitioners. Department of Computer Science, Rutgers University, Piscataway, NJ 08855, USA. (farach@,/farach) Supported by NSF Career Development Award CCR-9501942, an Alfred P. SloanResearch Fellowship and NATO Grant 960215.yDepartment of Computer Science, Rutgers University, Piscataway, NJ 08855, USA. (ferenci@).Supported by NSF grant CCR-9501942. 1 IntroductionThe problem of constructing an evolutionary tree, or phylogeny, for a set of species is one of thecentral problems of computational biology. While many approaches to this problem have beenproposed, distance methods are the only methods to have been shown to be consistent, that is, toconverge on the right answer in the limit, and ecient, that is, to have a convergence rate that iswithin a polynomial of the best possible rate [10].The distance method approach involves measuring pairwise evolutionary distances between thespecies, and then tting the data to a tree metric, which is dened by a weighted tree spanningthe points of the metric, the distance between two points being the sum of the weights of the edgeson the path between them. This problem is called the Numerical Taxonomy problem, and it hasreceived a great deal of attention (see [2, 18, 19] for extensive surv