A large deviation result is established for the bootstrap empirical di
stribution in a finite sample space, thereby validating both nonparame
tric and parametric bootstrapping in certain phylogenetic inference pr
oblems. The bias previously observed in the bootstrap distribution of
the estimated tree topology is shown to stem from dispersion effects i
n the joint distribution of sample and bootstrap empirical distributio
ns. Both results are examined for maximum likelihood estimation in a t
hree-taxon model having particularly simple geometry. They are also ap
plicable to discrete parameter problems outside of phylogenetic infere
nce.