Digital Signal Processing Reference
In-Depth Information
which is independent of the precoder. Thus,
any unitary precoder is as good as
any other
if we restrict
F
to be
M
M
unitary.
1
×
For any chosen
F
we simply
take the equalizer to be
G
=(
HF
)
#
(15
.
7)
A
#
represents the
to satisfy the zero-forcing condition.
Here the notation
(minimum-norm) left inverse of the matrix
A
(see Eq.
(C.7b) in Appendix
C). The reconstruction error then simplifies to
M−
1
1
σ
h,k
E
mse
=
σ
q
Tr (
H
†
H
)
−
1
=
σ
q
(15
.
8)
k
=0
where
σ
h,k
are the
M
singular values of the channel
H
(all
σ
h,k
>
0 because the
channelisassumedtohaverank
M
).
2
We now present three examples of the choice of unitary
F
.
Since any unitary
F
is as good as any other, all of these are “optimal” solutions.
1.
Lazy precoder
.If
F
=
I
we have
G
=
H
#
.
This transceiver is shown in Fig.
15.1(a). This system has the advantage that all “collaboration” between
the different
s
k
(
n
)'s is done at the receiver. This is suitable for a multiuser
system in multiple access mode (Sec. 4.5).
2.
SVD precoder
. With the channel expressed in SVD form (15.4), if we decide
to choose the unitary precoder
F
=
V
h
,
then the equalizer is
G
=(
HF
)
#
=(
U
h
Σ
h
)
#
=[[
Σ
h
]
−
M
0
]
U
h
,
(15
.
9)
where [
Σ
h
]
M
is the
M
M
diagonal matrix with diagonal elements equal
to
σ
h,k
.
This transceiver is shown in Fig. 15.1(b).
×
3.
Cyclic-prefix precoder
. Recall that in the cyclic-prefix system the channel
is made to look like an
M
×
M
circulant matrix. In this case
H
=
W
−
1
Λ
c
W
.
(15
.
10)
Here
Λ
c
is a diagonal matrix containing the DFT coe
cients of the
cha
n-
nel, and
W
is the DFT matrix. Since
W
†
W
=
M
I
,thematrix
W
†
/
√
M
is
unitary and we can take the unitary precoder and the zero-forcing equalizer
to be
F
=
W
†
G
=
Λ
−
c
W
√
M
,
√
M
.
(15
.
11)
This results in the familiar OFDM system described in Sec. 7.3. See Fig.
15.1(c). Since any u
nit
ary
F
is as good as any other, we can say that
the
IDFT matrix
W
†
/
√
M is also an “optimum unitary precoder” solution.
1
Note that we are only concerned about minimization of MSE. If we are interested in
minimizing error probability, as in Chap. 16, then the situation is different.
2
The second equality in Eq. (15.8) follows because (a) the trace is the sum of eignevalues,
(b) the eigenvalues of
H
†
H
are
σ
h,k
, and (c) the eigenvalues of (
H
†
H
)
−
1
are the reciprocals
of those of
H
†
H
.
Search WWH ::
Custom Search