Coming back to Figure cap:Evaluation-going-below, there was a serious flaw that is associated with the function's insufficiently constrained form. It was possible for the model to be improved by 'cheating' and concealing parts of the data.
There was a period where in order to avoid data breaking and bumps
shrivelling, a height was being forced to remain the same at the pinnacle
of the bump. Several different schemes were attempted. At times, the
position for insertion was projected from the mean and later it was
derived from the model discrepancy. The results were not of very high
quality at that stage and Figure shows
one such case. When run over a long period of time, a sharp tip can
be observed which is due to points being forced to lie in a single
fixed position. On other occasions, parts of the image are still being
hidden. This basically results in gradual darkening of images subjected
to registration.