Next: PRECONDITIONING THE REGULARIZATION
Up: PRECONDITIONED DATA FITTING
Previous: Preconditioner with a starting
We are tasked with coming up with ``trial solution''--a pretty vague assignment.
Some kind of a scaling, smoothing, or shaping transformation
of some mysterious ``preconditioned space''
should represent the model
we seek.
We begin by investigating how the shaper
alters the gradient.
|
= |
|
introduces
, implicitly defines
|
|
= |
|
consequence of the above |
|
= |
|
gradient is adjoint upon residual |
|
= |
|
residual in terms of
|
|
= |
|
residual in terms of
|
|
= |
|
reordering calculation |
|
= |
|
gradient is adjoint upon residual |
|
= |
|
reordering |
|
= |
|
recalling
|
We may compare the gradient
with and without preconditioning.
|
= |
|
original |
|
= |
|
with preconditioning transformation |
When the first vanishes, the second also vanishes.
When the second vanishes, the first vanishes provided
is a nonsingular matrix.
As our choice of
is quite arbitrary,
it is marvelous the freedom we have to monkey with the gradient.
Remember that
starts off being
.
Compare the
scaled gradient to the analytic solution.
|
= |
|
modified gradient |
|
= |
|
analytic solution |
Mathematically,
we see it would be delightful if
were something like
,
but we rarely have ideas how to arrange it.
We do,
however, have some understanding of the world of images,
and understand where on the image we would like iterations to concentrate first,
and what spatial frequencies are more relevant than others.
If we cannot go all the way, as we cannot in giant imaging problems,
it is important to make the important steps early.
Next: PRECONDITIONING THE REGULARIZATION
Up: PRECONDITIONED DATA FITTING
Previous: Preconditioner with a starting
2015-05-07