Digital Signal Processing Reference
In-Depth Information
Plugging this solution back into the LS problem, we can first rewrite
C
−
1
/
2
(
r
−
M
s
s
−
M
n
ˆ
n
)
w
C
−
1
/
2
C
−
1
/
2
M
n
C
−
w
M
n
)
−
1
M
n
C
−
1
=[
−
M
n
(
]
r
−
w
w
w
C
−
1
/
2
C
−
1
/
2
M
n
C
−
w
M
n
)
−
1
M
n
C
−
1
−
[
−
M
n
(
]
M
s
s
w
w
w
P
⊥
C
−
1
/
2
=
(
r
−
M
s
s
)
w
where
C
−
1
/
2
M
n
C
−
w
M
n
)
−
1
M
n
C
−
1
/
2
P
⊥
=
I
−
P
,
P
=
M
n
(
.
w
w
, we can show that
P
is an orthogonal projection. Hence, also
P
⊥
=
Similar to
˘
P
is an orthogonal projection, onto the complement of the range of
C
−
1
/
2
−
I
M
n
,
w
˙
i.e., the weighted range of the noise matrix. If
n
is diagonal, this is equivalent to
“projecting out the diagonal”, thus omitting these entries in the WLS fitting. It is
interesting to note that, in current telescopes, the autocorrelations (main diagonal of
R
) are usually not estimated. That fits in very well with this scheme, as the projection
would project them out anyway!
The resulting “compressed” WLS problem is
C
−
1
/
2
2
ˆ
=
s
(
r
−
)
arg min
M
s
s
s
s
where
C
−
1
/
2
P
⊥
C
−
1
/
2
C
−
1
=
.
:
w
w
s
(This is with some abuse of notation:
C
s
is singular due to the projection, hence not
invertible. However, we will only need
C
−
1
, and will use the above definition for
s
it.) The solution
ˆ
but now it is obtained in two steps: first
ˆ
s
and then, if required,
ˆ
As the expression for the compressed problem is very similar to the original WLS
problem, we obtain similar results: the solution is
M
s
C
−
1
)
−
1
M
s
C
−
1
ˆ
=(
r
M
s
s
s
s
which can also be written as
M
−
1
ds
M
s
C
−
1
M
s
C
−
1
s
=
ˆ
ds
,
ˆ
M
ds
=
M
s
,
ds
=
ˆ
r
.
s
s
M
ds
is the deconvolution matrix, and ˆ
ds
is the WLS dirty image. This time, the
dirty image is really a (vectorized) image, whereas in the previous discussion, the
vector ˆ
d
had an image component and a noise component.
The covariance of the image estimate ˆ
s
is, using
C
−
1
C
w
C
−
1
C
−
1
s
=
,
s
s
M
s
C
−
1
M
s
)
−
1
M
s
C
−
1
C
−
1
s
M
s
C
−
1
M
s
)
−
1
M
−
1
C
s
=(
(
C
w
)
M
s
(
=
ds
.
s
s
s