Digital Signal Processing Reference
In-Depth Information
In Appendix 19.C at the end of the chapter we show that when
G
is as in Eq.
(19.69), the preceding expression for
R
ee
can be rewritten in the following form:
R
ee
=
σ
s
(
B
+
I
)
−
1
(
B
+
I
)
†
,
I
+
σ
s
σ
q
F
†
H
†
HF
(19
.
72)
which is more useful for optimization. The goal is now to optimize the matrices
F
and
B
to minimize the trace of Eq. (19.72) subject to the power constraint
and the constraint that
B
be strictly upper triangular. Thus the objective to be
minimized is
E
mse
=
σ
s
Tr
(
B
+
I
)
−
1
(
B
+
I
)
†
,
I
+
σ
s
σ
q
F
†
H
†
HF
(19
.
73)
subject to the power constraint
σ
s
Tr
FF
†
=
p
0
(19
.
74)
and the upper triangular constraint
b
km
=0
,
k
≥
m.
(19
.
75)
19.4.3 Bound on the error
We now find a lower bound on the reconstruction error
E
mse
in terms of quantities
that depend only on
F
and
H
.
For the reader who is eager for a preview, this
bound is shown in Eq. (19.88), where
σ
f,k
are the singular values of the precoder
F
,
and
σ
h,k
are the first
M
dominant singular values of the channel
H
.
Depending
on the available power, this bound can further be refined as in Eq. (19.98), where
K
and
λ
are quantities that depend on power
p
0
.
Finally we show in Sec. 19.4.5
how this bound can be achieved by choice of
F
,
G
,
and
B
.
Now for the details. The trace is the sum of the eigenvalues
λ
k
of
R
ee
,
whereas the determinant is the product of
λ
k
.
Since
R
ee
is positive definite, we
have
λ
k
>
0, and the AM-GM inequality can be applied to obtain
Tr
R
ee
M
(det
R
ee
)
1
/M
.
≥
(19
.
76)
That is,
M
det
(
B
+
I
)
−
1
(
B
+
I
)
†
1
/M
I
+
σ
s
E
mse
σ
s
≥
σ
q
F
†
H
†
HF
=
M
det
−
1
/M
I
+
σ
s
σ
q
F
†
H
†
HF
,
where we have used the fact that
B
+
I
is upper triangular with diagonal elements
equal to unity. So
Mσ
s
det
−
1
/M
I
+
σ
s
σ
q
F
†
H
†
HF
E
mse
≥
.
(19
.
77)
Search WWH ::
Custom Search