I am using djacobix to differentiate a function which can only take positve inputs. As my program proceeds jacobix the point of interest approaches zero and eventually jacobix tries to evaluate the function with negative arguments.
It seems to me that at some point the way jacobix is using the "eps" parameter changes; for large point of interest x the step size used in (what i assume is) finite differencing is
dx = (plus and minus) eps*x
which remains positive as x approaches zero. However at some point jacobix starts using
dx = (plus and minus) eps
resulting in negative arguments to the target function.
Can someone explain this behaviour so I can suppress it. I'm happy with a cluge that modifies the value of eps to avoid negative values if that is necessary - in which case knowing when jacobix changes behaviour is enough. If it's complicated I will probably write my own routine.