Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
n=0
b
kn
(i)e
j
2n[m(i+1)
G
]
N
, (1)
where iT
S
m
t
< (i + 1)T
S
. b
kn
(i) is the modulation
signal for the k-th antenna at the n-th (0 n < N) subcarrier
of the i-th (i
i
i i
f
) symbol, where i
i
and i
f
denote the
rst and last OFDM symbol numbers in a burst, respectively.
G
(= T
G
/
t
) should be a positive integer.
b) Received Signal: r(t) is an L
R
-by-1 vector having a
signal received by the l
r
-th (1 l
r
L
R
) antenna as the l
r
-th
element, and is given by
r(t) =
L
T
k=1
M
f
m=M
i
h
k
(t m
t
)s
k
(m) + i(t) + n(t), (2)
where M
i
= i
i
(N+
G
) and M
f
= i
f
(N+
G
)1. The L
R
-
by-1 vectors h
k
(t), i(t) and n(t) represent the channel impulse
response, the interfering signal and additive noise components,
respectively.
B. Estimation Error and Metric
An error signal (l
b
, m) of the l
b
-th (1 l
b
L
B
) branch
metric generator at discrete time m
t
is expressed as [3]
(l
b
, m) = y
r
(l
b
, m) y
s
(l
b
, m), (3)
where y
r
(l
b
, m) is the output of STF, and y
s
(l
b
, m) is the
replica signal of y
r
(l
b
, m).
y
r
(l
b
, m) can be approximated with nite FTFs taps as [3]
y
r
(l
b
, m)
= w
H
l
b
x(m), (4)
x
H
(m) = [r
H
(m
t
M) r
H
(m
t
M + )
r
H
(m
t
+ M)], (5)
w
H
l
b
= [w
H
l
b
,M
w
H
l
b
,(M1)
w
H
l
b
,M
], (6)
where
H
denotes Hermitian transposition, and is a sampling
period given by =
t
/ where is a positive integer. M
is a nonnegative integer and (2M +1) is equal to the number
of FTFs taps. w
l
b
,p
with M p M is an L
R
-by-1
coefcient vector, which is multiplied by r(m
t
+ p).
Similarly, the replica signal can be approximated with the
maximum delay D
0
t
as
y
s
(l
b
, m)
= C
H
l
b
s(m), (7)
C
H
l
b
= [C
H
1l
b
C
H
2l
b
C
H
L
T
l
b
], (8)
C
H
kl
b
= [c
l
b
(k, 0) c
l
b
(k, 1) c
l
b
(k, D
0
)], (9)
s
H
(m) = [s
H
1
(m) s
H
2
(m) s
H
L
T
(m)], (10)
2115
s
H
k
(m) = [ s
k
(m) s
k
(m 1) s
k
(m D
0
)], (11)
where the asterisk denotes complex conjugate. C
l
b
is an
L
T
(D
0
+ 1)-by-1 vector representing the equivalent channel
impulse response. s
k
(m) is a replica of s
k
(m) given by
s
k
(m) =
N1
n=0
b
c
kn
(i)e
j
2n[m(i+1)
G
]
N
, (12)
where b
c
kn
(i) is a candidate of b
kn
(i). Note that b
c
kn
(i) =
b
kn
(i) during preambles.
The ML detection performed in the frequency domain uses
the following branch metric that is obtained from the discrete
Fourier transform of (l
b
, m) [3]:
f
(i, n) =
L
B
l
b
=1
Y
l
b
(i, n)
L
T
k=1
H
l
b
(k, n)b
c
kn
(i)
2
, (13)
where Y
l
b
(i, n) is the discrete Fourier transform of y
r
(l
b
, m).
H
l
b
(k, n) is the channel frequency response and is given by
H
l
b
(k, n) =
D
0
d=0
c
l
b
(k, d)e
j
2nd
N
. (14)
The detector searches for {b
c
kn
(i)| 1 k L
T
} that minimize
f
(i, n) in order to detect {b
kn
(i)| 1 k L
T
} .
C. Parameter Estimation
From (4) and (7), (3) can be rewritten as
(l
b
, m) = w
H
l
b
,ext
x
ext
(m), (15)
w
H
l
b
,ext
= [w
H
l
b
C
H
l
b
], (16)
x
H
ext
(m) = [x
H
(m) s
H
(m)], (17)
where w
l
b
,ext
and x
ext
(m) are L-by-1 vectors, and L = (2M+
1)L
R
+ L
T
(D
0
+ 1). Owing to the prewhitening, (l
b
, m)
should satisfy the following equation:
(l
b
, m)(
l
b
, m) = w
H
l
b
,ext
x
ext
(m)x
H
ext
(m)
_
w
l
b
,ext
=
l
b
l
b
|(l
b
, m)|
2
, (18)
where denotes the statistical averaging and
ij
is Kro-
neckers delta. Thus, w
l
b
,ext
can be set parallel to an eigen-
vector of
x
ext
(m)x
H
ext
(m)
_
, and EVD is applied to the
autocorrelation matrix for the parameter estimation.
After obtaining the L eigenvalues {
l
| 1 l L}
and L corresponding eigenvectors {p
l
| 1 l L} of
x
ext
(m)x
H
ext
(m)
_
, {w
l
b
,ext
| 1 l
b
L
B
} are estimated as
w
l
b
,ext
= p
l
b
/
_
b
, (19)
so that |(l
b
, m)|
2
can be normalized to one. Here l
b
(1
l
b
L
B
) is a positive integer and the selection of the
eigenvectors p
l
b
is discussed below.
TABLE I
CAPACITY BASED EIGENVECTOR SELECTION
G = [H
1
(n)
T
H
2
(n)
T
H
L
(n)
T
]
T
for l = 1, ,L
T
k
l
= arg max
lqL
H
q
(n)
2
r
k
l
,k
l
= H
k
l
(n)
H
k
l
(n) H
k
l
(n)/r
k
l
,k
l
for q = l, , L, q = k
l
r
q,k
l
= H
q
(n)H
H
k
l
(n)
H
q
(n) H
q
(n) r
q,k
l
H
k
l
(n)
end
exchange the row l and k
l
in G
end
D. Capacity Based Eigenvector Selection
L
B
is set to L
T
, and L
T
eigenvectors are selected under a
channel capacity criterion. First, {Y
l
b
(i, n)| 1 l
b
L
B
} are
expressed from (13) in a vector form as
Y(i, n) = H(n)b(i, n) + e(i, n), (20)
Y(i, n) = [Y
1
(i, n) Y
2
(i, n) Y
L
B
(i, n)]
T
, (21)
H(n) = [H
T
1
(n) H
T
2
(n) H
T
L
B
(n)]
T
, (22)
H
l
b
(n) = [H
l
b
(1, n) H
l
b
(2, n) H
l
b
(L
T
, n)], (23)
b(i, n) = [b
c
1n
(i) b
c
2n
(i) b
c
L
T
n
(i)]
T
, (24)
where
T
denotes transposition. Each row of the L
B
-by-L
T
matrix H(n) corresponds to the frequency channel response of
the STF output in each branch metric generator. The L
B
-by-1
vector e(i, n) represents the residual interference and noise
after the prewhitening. The covariance matrix of e(i, n) is
equal to
1
N
I where I is the L
B
-by-L
B
identity matrix, because
of the normalization of (19).
Next a hypothetical channel capacity
C is dened as
C = log
2
det[I + NH(n)H
H
(n)]. (25)
Then L
T
eigenvectors to maximize
C are searched among total
L eigenvectors [6]. Although an exhaustive search can solve
this problem, it requires prohibitive complexity. Thus, the
proposed method applies an algorithm that searches L
T
eigen-
vectors to maximize det[H(n)H
H
(n)] instead of
C. The larger
det[H(n)H
H
(n)], the larger
C. Note that det[H(n)H
H
(n)] =
| det[H(n)]|
2
when L
B
is set to L
T
. The optimization problem
becomes equivalent to nding H(n) that has the maximum
determinant. Such an optimization problem can be solved by
the algorithm shown in Table I, where G has the channel
frequency responses of all L eigenvectors.
E. Recursive EVD
To reduce the complexity of parameter update of Pre-FFT
STF, the recursive EVD method is applied [7]. First, the
autocorrelation matrix of x
ext
(m) is expressed in a recursive
form as
R
x
(m) = R
x
(m 1) + x
ext
(m)x
H
ext
(m). (26)
where (0 < 1) is a forgetting factor. Provided that
the L eigenvalues {
l
| 1 l L} and L corresponding
eigenvectors {q
l
| 1 l L} of R
x
(m 1) are obtained and
2116
that
1
2
L
, the eigenvalues {
l
| 1 l L} of
R
x
(m) can be simultaneously searched in parallel as
l
<
l
<
l1
, (27)
where
0
=
1
+ x
H
ext
(m)x
ext
(m). The search interval is
restricted to I
l
= (
l
,
l1
) = (, u). Then the Newtons
method is applied into an iterative search to identify the l-th
eigenvalue
l
[7]
(+1)
l,
=
()
l
_
()
l
_
()
l
_
, (28)
(+1)
l
=
_
(+1)
l,
If
(+1)
l,
I
l
()
l
+u
2
If
(+1)
l,
> u,
()
l
replaces
()
l
+
2
If
(+1)
l,
< ,
()
l
replaces u
(29)
where is the number of iterations, and
() = 1 +
L
l=1
q
H
l
x
ext
(m)
l
, (30)
() =
L
l=1
q
H
l
x
ext
(m)
2
. (31)
The iterative search stops and
l
=
()
l
when
(+1)
l
()
l
<
(+1)
l
, where is the convergence threshold.
After updating the eigenvalues, the corresponding eigenvec-
tors {p
l
| 1 l L} of R
x
(m) are parallelly obtained as
p
l
=
L
=1
q
H
l
x
ext
(m)
l
l
q
l
, (32)
p
l
=
p
l
p
l
. (33)
IV. COMPUTER SIMULATION
A. Simulation Conditions
In order to evaluate the performances of the proposed
scheme, computer simulations under the IEEE 802.11a stan-
dard were conducted in a fading channel. The simulation
parameters are summarized in Table II.
A 17-path Rayleigh fading model with the maximum
Doppler frequency f
D
was used, where the average power
of each path was subject to the exponential distribution. An
average power ratio of the rst propagation path to the last
one was set to 10 dB. It was also assumed that D
0
=
G
. A
linear array was assumed as the receive antenna. The distance
between the antenna elements was half of the wavelength. The
difference among average incident angles of the users was set
to 60 Deg, and the incident angles were assumed to be subject
to Gaussian distribution with the standard deviation of 4 Deg.
B. Effect of Branch Number
Fig. 4 shows BER performance of Pre-FFT STF using the
eigenvector selection under the SINR criterion with L
B
as
a parameter, when there is no interference (N
u
= 1) and
TABLE II
SIMULATION CONDITIONS
Modulation scheme QPSK
Packet format 14 Symbols
Preamble: 4, Data: 10
Number of trans. antennas: L
T
2
Number of rec. antennas: L
R
3
Number of interfering users 0, 1
Number of active subcarriers 52
Pilot: 4, Data: 48
Subcarrier interval:
f
312.5 kHz
Symbol duration: T
S
(= T
G
+T
F
) 4.0 s
Guard interval: T
G
0.8 s
FFT points: N 64
Number of FTFs taps: 2M + 1 3
Sampling rate: 2N
f
40 MHz
Carrier frequency: f
c
5.0 GHz
Convergence threshold: 10
5
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
0 5 10 15 20 25 30
A
v
e
r
a
g
e
B
E
R
Average E
b
/N
0
(dB)
CMLD
Pre-FFT STF
= 2
L
= 5
L
= 4
L
=
L
Nu = 1
f
D
= 0Hz
Fig. 4. BER performance with L
B
as a parameter
f
D
= 0Hz. For comparison, the result of the conventional
MLD (CMLD) that cannot suppress the interference was also
plotted as a lower bound. It can be seen that the performance
of Pre-FFT STF improves as L
B
increases, which brings more
diversity effect to the detector. However, the improvement
is saturated with L
B
= 4L
T
, and the remaining branch
metric generators make little contribution to the improvement.
Thus, the search of the proposed eigenvector selection is
limited to the 4L
T
eigenvectors associated with 4L
T
largest
SINR instead of the total L eigenvectors, which reduces the
complexity of the search without performance degradation.
C. Effect of E
b
/N
0
Fig. 5 shows BER performances of the proposed and con-
ventional schemes versus E
b
/N
0
in a cochannel interference
environment with N
u
= 2. Note the proposed scheme applies
the capacity based eigenvector selection to Pre-FFT STF with
L
B
= L
T
. It is superior to Pre-FFT STF with L
B
= L
T
, and
the performance difference increases with average E
b
/N
0
. In
high E
b
/N
0
region, the proposed scheme nearly achieves BER
performance of Pre-FFT STF with the optimal L
B
because
det[H(n)H
H
(n)] is a good approximation of
C. Conversely,
Post-FFT STF cannot converge to this supreme BER because
2117
0 5 10 15 20 25 30
A
v
e
r
a
g
e
B
E
R
Average E
b
/N
0
(dB)
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
CIR = 0 dB
Nu = 2
f
D
= 0Hz
Average
Pre-FFT STF with the optimal L
Post-FFT STF
L
=
L
Pre-FFT STF,
Proposed scheme
Fig. 5. BER performance versus E
b
/N
0
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
-20 -10 0 10 20 30
A
v
e
r
a
g
e
B
E
R
CIR (dB)
Average15 dB E
b
/N
0
Average30 dB E
b
/N
0
Nu = 2
f
D
= 0Hz
Average
Pre-FFT STF with the optimal L
Post-FFT STF
L
=
L
Pre-FFT STF,
Proposed scheme
Fig. 6. Effect of CIR on BER performances
the interference and limited number of preamble symbols
cause inaccurate parameter estimation.
D. CIR Resistance
Fig. 6 shows BER performances of several schemes versus
CIR. Difference in BER between the proposed scheme and
Pre-FFT STF with L
B
= L
T
or Post-FFT STF increases with
CIR. When average E
b
/N
0
is not high, the performance of the
proposed scheme is a little worse than that of Pre-FFT STF
with the optimal L
B
.
E. Effect of the Doppler Frequency
BER performance of the proposed scheme versus the
Doppler frequency is shown in Fig. 7. There are two update
schemes: one uses the nonrecursive EVD and the other uses
the recursive EVD. The Doppler frequency has signicantly
negative effect on a nonupdate scheme that performs the pa-
rameter estimation only during the preambles, while the update
schemes can track the fading channel sufciently even when
f
D
is very high. It can be also seen that BER performances
of these two update schemes are almost the same.
Table III shows the computational complexity of the two
update schemes. Here the nonrecursive EVD employs the
cyclic Jacobi algorithm, and
M
and n
s
are the maximum
number of iterations and the number of sweeps, respectively.
A
v
e
r
a
g
e
B
E
R
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
0 2.0x10
3
4.0x10
3
6.0x10
3
8.0x10
3
1.0x10
2
f
D
T
S Normalized Doppler Frequency
Nu =2
Nu =1
CIR = 0 dB
= 0.99
= 10
-5
Average
Nonupdate scheme
(Update schemes)
: Nonupdate
: Recursive EVD
: Nonrecursive EVD
Fig. 7. BER performances versus the Doppler frequency
TABLE III
COMPUTATIONAL COMPLEXITY
Operation Recursive EVD Nonrecursive EVD
Mul. N(2L
2
+L
M
) 4L
2
(L 1)n
s
+
3NL
4
(L + 1)
Add. N(2L
2
2L +
L
M
2
) 2L
2
(L 1)n
s
+NL(
L
2
+
1
2
)
Generally, the accurate result of EVD can be obtained in about
4 to 10 sweeps for 4 L 2000. The number of complex
multiplications of the recursive update is about 20 percents of
that of the nonrecursive update when n
s
= 4, and about 9
percents when n
s
= 10. Similarly, the recursive scheme can
reduce the number of complex additions to about 36 percents
when n
s
= 4, and to 15 percents when n
s
= 10.
V. CONCLUSION
In this paper, a capacity based eigenvector selection method
was proposed for Pre-FFT STF in order to alleviate the
performance degradation due to the correlated channels, and a
recursive EVD method was applied to the parameter estimation
of Pre-FFT STF. Computer simulations demonstrated that the
proposed scheme can improve the BER performance and track
fast fading channels with reduced complexity.
REFERENCES
[1] J. Chuang and N. Sollenberger, Beyond 3G: Wideband wireless data
access based on OFDM and dynamic packet assignment, IEEE Comm.
Mag., pp. 78-87, July 2000.
[2] G. J. Foshini, Layered space-time architecture for wireless communica-
tion in a fading environment when using multi-element antennas, Bell
Labs. Tech. J., vol. 1, no. 2, pp. 41-59, 1996.
[3] K. Fukawa, Y. Yamada, H. Suzuki, and S. Suyama, A MIMO-OFDM
maximum likelihood reception scheme in cochannel interference envi-
ronments, IEEE Vehic.Tech.Conf., vol. 4, pp. 1849-1853, May 2004.
[4] J. Ki, K. B. Letaief, and Z. Cao, Co-channel interference cancellation
for space-time coded OFDM systems, IEEE Trans. Wireless Comm.,
vol. 2, no. 1, pp. 41-49, Jan. 2003.
[5] S. T. Park and C. G. Kang, Complexity-reduced iterative MAP receiver
for interference suppression in OFDM-based spatial multiplexing sys-
tems, IEEE Trans. Vehic. Tech., vol. 53, no. 5, pp. 1316-1326, Sept.
2004.
[6] M. Gharavi-Alkhansari, A.B. Gershman, Fast antenna subset selection
in MIMO systems, IEEE Trans. Acoust. Speech Sig. Proc., vol.2, pp.
339-347, Feb. 2004.
[7] K. B. Yu, Recursive updating the eigenvalue decomposition of a
covariance matrix, IEEE Trans. Acoust. Speech Sig. Proc., vol. 39, pp.
1136-1145, May 1991.
2118