Digital Signal Processing Reference
In-Depth Information
(
)
−
1
ˆ
†
†
†
2
γ
=
WY
=
h
HH
+
σ
I
Y
,
(4.18)
p
p
p
eq q
n
where W
p
denotes the filter,
h
p
is the
p
th
column of H
eq
, and (.)
†
stands for transpose
complex conjugate. Note that the entries of H
eq
are real values, and hence (.)
†
is equiva-
lent to transposition. From the second iteration, we can calculate soft estimates S
˜
of
the transmitted data using the soft decoder outputs. Using these estimates, we perform
interference cancelling followed by a simple zero forcing (ZF) or MMSE filtering:
ˆ
ˆ
,
ˆ
†
YYHS
=−
γ
=
WY
,
(4.19)
p
p
p
p
pp
1
1
ZF:
W
=
hh
h
,
MMSE:
W
=
h
p
,
(4.20)
( )
p
p
p
†
hh
†
σ
2
pp
pp
n
where S
˜
p
of dimension ((2
Q
- 1) × 1) is S
˜
with its
p
th
entry removed, and H
p
of dimension
(2
M
R
T
× (2
Q
- 1)) is the matrix H with its
p
th
column removed. Notice that, compared
to the exact MMSE filtering proposed in [48], (4.20) are simplified solutions that assume
almost perfect estimation of data symbols and permit a considerable reduction of the
computational complexity. Thanks to iterative processing, the performance loss due to
this simplification would be negligible. In the results that we present later, we will con-
sider the simplified ZF solution.
4.4.3.3.2 Conversion to LLR
For QAM modulation with
B
(an even number) bits per symbol, we can attribute
m
=
B
/2 bits to the real and imaginary parts of each symbol. Let, for instance, the bit
c
i
cor-
respond to the real (imaginary) part of the symbol
s
q
. Let also
a
1,
j
and
a
0,
j
,
j
= 1,
…
,
B
/2
denote the real (imaginary) part of the signal constellation points, corresponding to
c
i
= 1 and
c
i
= 0, respectively. Remember that the signal constellation points have nor-
malized average power. The LLR corresponding to
c
i
is calculated as follows [62]:
m
−
1
∑
2
1
2σ
γα
−
( )
2
ˆ
exp
−
p
1
,
j
2
j
=
1
p
=
…
LLR
i
=
log
,
i
1
,
,
m
,
(4.21)
10
m
−
1
∑
−
2
1
−
( )
2
ˆ
exp
σ
γα
p
p
0
,
j
2
2
j
=
1
where σ
p
2
is the variance of noise plus the residual interference (RI) that intervenes in the
detection of γ
ˆ
p
, and is assumed to be Gaussian. Note that as the detection is performed
on blocks of
Q
complex symbols, or in other words, on blocks of
2Q real symbols
in our
model, the RI comes in fact from (2
Q
- 1) other real symbols in the corresponding chan-
nel use [64]. In LLR calculation, we need the variances σ
p
2
. These variances can be cal-
culated analytically as shown in [46], or estimated at each iteration and for each one of
Search WWH ::
Custom Search