We got a hint from a colleague:
at wikipedia you can find the following note:(Advantages / 9. ...)
"This type of fit, with the response variable appearing on both sides of the function, should only be used to obtain starting values for the nonlinear fit. The statistical properties of fits like this are not well understood."
This is exactly our approach so far.
Further search on the net shows that Wikipedia is a citation from NIST... see bottom left "Starting Values for ..."
In maths forums you can also find hints, but nothing really reliable.
We should check that and maybe improve our method.
We got a hint from a colleague:
We should check that and maybe improve our method.