Strange difference in Minuit between two versions of root

Hello,
I’m wondering if anybody else has seen anything like this. I’m using the same code code root 3.10/02 on cygwin and and root 3.05/04 (compiled with KCC?) on linux. I’m running 100,000 fits using random numbers and getting different answers on the two different machines.

I’m using Minuit in Root to minimize a chi^2 and maximize a likelihood. They are mathematically equivalent and should give the same answer. On linux, they do (see chisq vs likelii numbers below). On cygwin, they do not. The other difference that I see is that when I fit on cygwin, I occasionally get Hesse Errors like:
MINUIT WARNING IN HESSE
============== MATRIX FORCED POS-DEF BY ADDING NaN TO DIAGONAL.
MINUIT WARNING IN HESSE
============== Negative diagonal element 1 in Error Matrix
MINUIT WARNING IN HESSE
============== 4285.19 added to diagonal of error matrix
MINUIT WARNING IN MIGRAD
============== Negative diagonal element 1 in Error Matrix
MINUIT WARNING IN MIGRAD
============== 1.00367 added to diagonal of error matrix

I never get these errors on linux.

 The first thing you're going to say is: "O.k. Charles, this is obvious, the two versions of root are throwing different random numbers."  But their not.  The Likeli numbers agree for both platforms and on the linux platform, the likeli numbers agree with the chi^2 numbers.

 Has anybody else seen anything like this?  Does this look like a cygwin bug, root 3.10.2 bug, user bug?  I realize I haven't given enough information to completely debug this, but I'm hoping that somebody has seen this already.

 Charles

p.s. Here are the outputs. The only thing to note is that I get the SAME EXACT numbers between the chi^2 and the likelihood fits on linux and the likehood numbers on cygwin, but the numbers are slightly different for the chi^2 fit on cygwin. The 100,000 trials are effectively averages, so I believe the reason they are close is that they are the same for almost all trials, but different for the trials where I’m getting Minuit errors. Note that for the single trial, all four (2 cygwin and 2 linux) agree perfectly.

**** Linux ****
100,000 Trials:
chisq: max 6.76 5.44339 8.24868 : 6.76 +1.48868 -1.31661
likeli: max 6.76 5.44339 8.24868 : 6.76 +1.48868 -1.31661

1 Trial:
chi2 6.99188 +2.39904 -2.05268
likeli 6.99188 +2.39904 -2.05268

**** Cygwin ****
100,000 Trials:
chisq: max 6.71667 5.45044 8.25705 : 6.71667 +1.54039 -1.26623
likeli: max 6.76 5.44339 8.24868 : 6.76 +1.48868 -1.31661

1 Trial:
chi2 6.99188 +2.39904 -2.05268
likeli 6.99188 +2.39904 -2.05268