Digital Signal Processing Reference
In-Depth Information
Theorem F.2.
Power spectrum of error
. Assume the noise
w
(
n
) is zero-
mean and uncorrelated to the signal
x
(
n
)
,
so that the Wiener filter is as in Eq.
(F.25). Then the error
e
(
n
) has the power spectrum
S
ee
(
e
jω
)=
♠
xx
(
e
jω
)+
S
−
ww
(
e
jω
)
−
1
,
S
−
1
(F
.
26)
assuming that
S
ww
(
e
jω
)and
S
xx
(
e
jω
) are invertible for all
ω.
Thus the error
spectrum is the
inverse of the sum of inverses
of the spectra of the signal and
the noise.
♦
Proof.
Since
x
(
n
)=[
g
∗
(
x
+
w
)](
n
)
,
where
g
(
n
) is the impulse response of
G
(
z
)and
∗
denotes convolution, we see that
e
(
n
)=[
g
∗
x
](
n
)+[
g
∗
w
](
n
)
−
x
(
n
)
.
(F
.
27)
The orthogonality principle dictates that
e
(
n
) be orthogonal to
y
(
m
) for all
n, m.
Since
x
(
n
) is a linear combination of samples of
y
(
.
)
,
it then follows
that
x
(
n
) is orthogonal to
e
(
m
) for all
n, m.
Thus
x
†
(
n
E
[
e
(
n
)
−
k
)] =
0
,
for all
k
, which implies that
S
e
x
(
z
)=
0
.
Thus
S
ee
(
z
)=
S
e
,
(
(
z
)=
S
e
x
(
z
)
−
S
ex
(
z
)=
−
S
ex
(
z
)
.
x
−
x
)
From Eq. (F.27) we therefore obtain
S
ee
(
z
)=
−
S
ex
(
z
)=
−
G
(
z
)
S
xx
(
z
)
−
G
(
z
)
S
wx
(
z
)+
S
xx
(
z
)
.
In view of the assumption that
S
wx
(
z
)=
0
,
it therefore follows that
S
ee
(
z
)=(
I
−
G
(
z
))
S
xx
(
z
)
.
(F
.
28)
Substituting for
G
(
z
) from the Wiener filter expression (F.25) we have
I
−
G
(
z
)=
I
−
S
xx
(
z
)[
S
xx
(
z
)+
S
ww
(
z
)]
−
1
=
S
xx
(
z
)+
S
ww
(
z
)
−
S
xx
(
z
)][
S
xx
(
z
)+
S
ww
(
z
)]
−
1
S
ww
(
z
)[
S
xx
(
z
)+
S
ww
(
z
)]
−
1
.
=
Using this in Eq. (F.28) we have
S
ee
(
z
)=
S
ww
(
z
)[
S
xx
(
z
)+
S
ww
(
z
)]
−
1
S
xx
(
z
)
,
which indeed can be rewritten as Eq. (F.26).
The error covariance martrix is simply the integral of the power spectral matrix,
that is,
2
π
1
2
π
R
ee
(0) =
E
[
e
(
n
)
e
†
(
n
)]
,
=
S
ee
(
e
jω
)
dω,
0
Search WWH ::
Custom Search