NaN as parameter errors

[cverstege]

This example https://gist.github.com/cverstege/1e085a6c6349399a5fa957628675794c yields NaN as parameter errors.
Fitting with iminuit seems to work.
This should be investigated either way.

[cverstege]

The upper bound for b is 0, while iminuit fits it to: b = -0.017 +/- 0.052
When setting the starting value for b to -0.1, the scipy fit converges and is able to calculate the errors.

[cverstege]

The underlying issue might be, that scipy is unable to calculate uncertainties, once the 1sigma interval is outside the parameter limits. iminuit does it anyway.