pxcomp deal with missing data

[josephwb]

Not sure why, but for a certain data set pxcomp is looping forever. The stopping criterion expects a small number, but this value becomes -nan and so never completes. This alignment is longish (30K sites) and contains a lot of missing data (7%), so I don't yet know what is responsible...

[josephwb]

Indeed, the ratio that should be getting smaller is instead getting larger. Hmmmmmmmm....

[josephwb]

Ugh. isnan and isinf do not work with fast-math, as it assumes all math involved is finite and NaNs do not exist. Considering whether it is needed (or what alternatives might be available).

[josephwb]

I believe this has to do with missing data, which are not supported. When we count the (valid) characters, the row totals get messed up, so we are not even doing a valid chi-square test anymore. This should be fixed with 7457e45.

[josephwb]

Changing this issue to reflect that missing data are not supported.

Can either: 1) through out columns with missing data or 2) something smarter (divide each amongst code states)? Check that p4/PAUP do in this case.