> You would 
> need to choose the length scale at which variations in the data are ignored. 
> E.g. you need to find some means of determining why the curve should skip over A 
> but try to get close to B.
>   
  

That is exactly the problem. I can look at the data curve and know where
the boundary curve should go but trying to figure out an algorithm that
works in general is non trivial. If I have to do it by hand, that will
severely limit the number of datasets I can analyze.

> You might be able to formulate this as a constrained minimization problem using 
> scipy.optimize.fmin_cobyla(). Take some curve f(x), minimize (f(x) - y) under 
> the constraint (f(x) - y >= 0).
  

Once I have something to optimize, I was thinking that genetic
algorithms might yield the fastest solution because I only need an
approximate solution.

Attached is an example. The green curve is boundary and the blue is data.

Bill