Quoting part of Jack Sullivan's comment,
>question pertains to figures for 50% bootstrap trees. So, what I typically
>do (again, for what it's worth) is pick one of the optimal trees (mp,
>ml...whatever), depict it as a phylogram (for the reasons Doug gave, among
>others), use a scale bar to indicate branch lengths and put estimates of
>nodal support (bootstrap/jacknife values, decay indices) on each node. If a
>node receives <50% bootstrap support, I still show it as resolved unless
>there's a zero-length internal branch in all optimal trees...
Of course one could always do a similar thing, by picking one of the
best trees, and then showing at each node the relative support for
the other possible resolutions of it. For binary trees there are two
alternate resolutions (according to the Nearest Neighbour
Interchange) though if you have a load of polytomies there would be
more ( (2k-3)!!-1 if it's a k-tomy) which would be problematic. If
you choose a binary representative from the original set of trees,
that's not an issue of course.
Also there's a known correlation between bootstrap support of a
branch and the ratio of relative split support to conflict with that
branch (if you like, the amount of the data which supports one branch
as compared with the data which breaks that branch).
I also get the feeling (though I may be wrong) that unless the data
sets are offensively unhelpful, the vast majority of signal which
conflicts with a given split (branch, edge, or whatever) comes from
the edges of the tree which are just one NNI removed from the split
of interest. If *that's* true, then much of the uncertainty in the
original trees can be adequately represented by showing the relative
split support for the NNI alternatives, for each edge in the tree.