Digital Signal Processing Reference
In-Depth Information
σ
q
Tr
Σ
f
(
U
f
Σ
h
U
f
)
M
−
1
=
σ
q
Tr
(
U
f
Σ
h
U
f
)
M
−
1
Σ
−
f
=
σ
q
M−
1
B
kk
σ
f,k
=
,
k
=0
where (
A
)
M
denotes the
M
M
leading principal submatrix of
A
as usual.
In the preceding,
B
kk
>
0 are the diagonal elements of the Hermitian positive
definite matrix
×
B
=
(
U
f
Σ
h
U
f
)
M
−
1
.
(12
.
84)
Here it is assumed that
σ
f,k
>
0 (i.e., that
F
has rank
M
). Let us take a closer
look at the matrix
B
and its relation to
U
f
.
Write the unitary matrix
U
f
in
partitioned form:
MP
−
M
U
f
=
.
U
1
U
2
If we replace
U
f
with
MP
−
M
U
f,new
=
U
1
U
b
U
2
(12
.
85)
for some
M
×
M
unitary
U
b
,
then
B
in Eq. (12.84) is replaced with (Problem
12.5)
B
new
=
U
b
BU
b
.
(12
.
86)
So, as far as
B
is concerned, the freedom to choose
U
f
is equivalent to the
freedom to choose the unitary matrix
U
b
.
12.B.3 Conversion to a majorization problem
To proceed further we will use the ideas of majorization and Schur convexity
reviewed in Chap. 21. If we assume the ordering
σ
f,
0
≤
σ
f,
1
≤
...
≤
σ
f,M−
1
(12
.
87)
then it can be shown that
B
00
≤
B
11
≤
...
≤
B
M−
1
,M−
1
.
(12
.
88)
If this is not the case then we can interchange pairs of elements
B
kk
and
B
mm
to reduce the MSE further (as we did in Sec. 12.4.1); we have the freedom to do
that by interchanging the columns of the matrix
U
b
.
So without loss of generality
we assume the above orderings. Now, from Ex. 1 in Sec. 21.4.1 (Chap. 21), we
know that the sum
M−
1
a
k
x
k
(12
.
89)
k
=0
is
Schur-concave
in the region
x
0
≤
x
1
≤
...
≤
x
M−
1
Search WWH ::
Custom Search