Inconsistency of evolutionary tree topology reconstruction methods when substitution rates vary across characters

A fundamental problem in reconstructing the evolutionary history of a set of species is to infer the topology of the evolutionary tree that relates those species. A statistical method for estimating such a topology from character data is called consistent if, given data from more and more characters, the method is sure to converge to the true topology. A number of popular methods are based on modeling the evolution of each character as a Markov process along the evolutionary tree. The standard models further assume that each character has in fact evolved according to the same Markov process. This homogeneity assumption is unrealistic; for example, different types of characters are known to experience substitutions at different rates. Certain distance and maximum likelihood methods for topology estimation have been shown to be consistent under the homogeneity assumption. Here we give examples showing that these methods can fail to be consistent when the homogeneity assumption is relaxed. The examples are very simple, requiring only four taxa, binary characters, and characters that evolve at two different rates.