Now we begin to appreciate the strange flavor of L1. We appreciate the idea that solutions are ``intervals''. But it is distressing to realize that we could often have graphical difficulty displaying the results. In practice we might need to settle for ``seeing some examples.'' Perhaps a satisfactory way of generating those examples would be by using random starting values for the fitting.
Suppose we set up the Busch problem with L1. Perhaps we will find the solution is not a unique surface. It might turn out to be a ``mat'' of variable thickness. It would be annoying to try to display the thickness, but perhaps the thickness is related to the uncertainty of the result. That should have value.