[Insight-users] Changing the size of the parameters array during registration?

[motes motes]

I have made a deformable image registration process that after a
number of iterations pauses the optimizer and based on some criterion
resizes the parameters array (always to a smaller size).

The the new parameters array is thrown into the image registration
process and the optimization is continued (the remaining parameters
are copied to the new parameters array and the optimizer is updated -
number of optimizer scales).  The jacobian is resized accordingly in
the transform and so are the number of deformable nodes.

Now I am trying to analyse the effect of resizing the parameters array
from the optimizer's point of view.  When the optimizer continues to
work on the new and smaller parameters array (which is initialized
with the  "surviving" coefficients from the previous parameters array)
the optimizer search space is actually reduced and the convergence
should in theory be speed up. But maybe someone has some thought on
this kind of "dynamic" search space reduction and its consequences?