Tests fail on recent 0.4dev

[sglyon]

The tests are failing for me on recent (5-13-15) 0.4dev.

I have traced the problem down to the call to pinv on this line. For the exact same input matrix, pinv gives very different answers on 0.4dev than 0.3.8.

I have opened JuliaLang/LinearAlgebra.jl#214 in the main Julia repository. Maybe whoever made the changes to pinv over there will know what is going on.

[sglyon]

So that issue over there was closed and a reference to JuliaLang/julia#8859 was given.

It seems that they intentionally changed the behavior of pinv when moving from 0.3.8 to 0.4 to make routines using pinv more stable.

It's too bad that we get very different answers for the J_stat based on the version of Julia. That makes it quite difficult to test that function. This brings me back to JuliaLang/julia#17 though... where are the hard-coded stats and p-values coming from?

Any ideas for how to resolve testing issues on the different Julia versions?

[gragusa]

I think we could do like python and create a pinv2 that keeps doing the same things across version. But this would only solve the consistency issue. The question i have about this is: which of the two approaches (the pinv of 0.3.8 or the pinv of 0.4) works better?

If the latest pinv is better we could write version dependent test.....

[sglyon]

See the comments on the Julia issue I opened. There isn't a general answer for "which pinv is better", rather they said it goes case by case.

I didn't do enough digging to determine which one we want