Академический Документы
Профессиональный Документы
Культура Документы
I. I NTRODUCTION
-LEVEL quadrature amplitude modulation (M -QAM)
provides high spectral efficiency and is thus widely
employed in many wireless communication applications [1]
[3]. The accurate bit error rate (BER) evaluation of the M QAM receivers in various fading conditions has attracted a lot
of interest in the literature [4][14]. In practice, the channel
estimates cannot be perfect in fading channels, and thus the
adverse effect of imperfect channel estimation (ICE) on the
BER performance must be taken into account in the system
design. In [4], for pilot symbol assisted modulation (PSAM),
an approximate BER for M -QAM was obtained for a Rayleigh
fading channel. In [5], the BER for the minimum meansquared-error (MMSE)-based channel estimator for a nondiversity Rician fading channel was derived. In [6], the exact
BERs for pulse amplitude modulation (PAM) and QAM signals in a maximum ratio combining (MRC) diversity Rayleigh
fading channel and a non-diversity Rician fading channel
with ICE were obtained, respectively. In [9], [10] a symbol
error rate (SER) formula involving a two-fold integral was
Manuscript received April 20, 2005; revised August 17, 2005; accepted
November 11, 2005. The associate editor coordinating the review of this paper
and approving it for publication was S. Zhou. This paper has been presented
in part at IEEE Wireless Commun. and Networking Conf. (WCNC), New
Orleans, March 2005.
Yao Ma and Dongbo Zhang are with the Department of ECpE, Iowa State
University, Ames, IA, 50011 (e-mail: {mayao, dbzhang}@iastate.edu).
Robert Schober is with the the Dept. of Electrical & Computer Engineering,
University of British Columbia, 2356 Main Mall, Vancouver, BC, V6T 1Z4,
Canada (e-mail: rschober@ece.ubc.ca).
Digital Object Identifier 10.1109/TWC.2007.05243.
c 2007 IEEE
1536-1276/07$25.00
MA et al.: EXACT BER FOR M-QAM WITH MRC AND IMPERFECT CHANNEL ESTIMATION IN RICIAN FADING CHANNELS
1, h {I + 1, . . . , 1, 1, . . . , I 1}, and v
{J + 1, . . . , 1, 1, . . . , J 1}. Ea is used to normalize the
data symbol energy to unity, i.e., Ed = E[|d(i)|2 ] = 1. For
3
square M -QAM Ea = 2(M1)
; and for rectangular (I, J)3
QAM Ea = I 2 +J 2 2 [8]. c(i) = [c1 (i), . . . , cL (i)]T is
the channel-coefficient vector for the L branches. In Rician
fading channels, c(i) can be written as c(i) = c (i) +
T
cf (i), where c (i) = E[c(i)] = [c1 (i), . . . , cL (i)] and
cf (i) = [cf1 (i), . . . , cfL (i)]T are the line-of-sight (LOS) and
diffuse components of c(i), respectively. At the lth branch,
cl (i) = |cl (i)|ej2(fdl +fol )iTs +l , where fdl is the Doppler
shift of the LOS component, fol is the frequency offset, l
is a random residual phase, and Ts is the symbol duration.
The Rician factor is defined as Kl = |cl (i)|2 /c2l , where
c2l = E[|cfl (i)|2 ] is the variance of the diffuse fading
component of the lth branch. The normalized temporal channel
c (n) =
correlation coefficient at the lth branch is defined as R
l
2
E[cfl (i)cfl (i n)]/cl . For Clarkes fading spectrum,
c (n) = J0 (2nBf Ts )ej2fol nTs ; and for the Gaussian
R
l
l
c (n) = exp((nBf Ts )2 )ej2fol nTs ,
fading spectrum, R
l
l
where J0 (x) is the zeroth order Bessel function and Bfl is
the Doppler fading bandwidth of the diffuse component.
The additive background noise vector, n(i)
=
[n1 (i), . . . , nL (i)]T , is a zero-mean circularly symmetric
complex Gaussian process, with average power
E[|nl (i)|2 ] = N0 , l = 1, . . . , L. In (1), n(i) may
be temporally and spatially correlated. The average bit
signal-to-noise ratio (SNR) of the lth branch is given by
b,l = Pd E[|cl (i)|2 ]/N0 = Pd (|cl (i)|2 + c2l )/N0 . Without
loss of generality, we assume Pd = 1 throughout this paper
unless otherwise stated.
B. PSAM-based Channel Estimators
For a PSAM-based channel estimator (PSAM-CE), a pilot
symbol is inserted into the data stream every P symbol
intervals. To estimate the channel coefficient vector c(i) for
the desired symbol d(i), F pilot symbols are employed which
may be written as an F 1 vector dPS = [d(i P F1 +
ioff ), . . . , d(i P + ioff ), d(i + ioff ), . . . , d(i + P (F2 1) + ioff)]T ,
where F1 and F2 (with F1 + F2 = F ) are the numbers of pilot
symbols on the left and right side of d(i), respectively, and
ioff (ioff = 1, 2, . . . , P 1) is the offset of the desired symbol
d(i) to the closest pilot symbol on its right side. The received
927
(2)
928
H
PPS RH
c (P ) + n (P )
PPS Rc (0) + n (0)
RyPS =
...
...
..
.
b1 = 1
b2 = 1
b2 = 1
~
B1
~
B2
~
B2
3a
3a
11
10
00
01
b2 = 1
~
B2
~
B3
7a
111
~
B1
~
B3
~
B2
~
B3
~
B3
5a
3a
3a
5a
7a
110
100
101
001
000
010
011
b3 = 1
b3 = 1
b3 = 1
1
R < 0|h = 1}]
R < 0|h = 3} + P {D
[P {D
2
P4,H (2) =
1
R < 2a|h = 3}
R < 2a|h = 3} P {D
[P {D
2
R < 2a|h = 1} + P {D
R > 2a|h = 1}]
+P {D
(4)
I = (D)
R < an |d(i) = an }
= P {D
I < an |d(i) = jan }
= P {D
(5)
(6)
=
=
C. Symbol Detection
(i) to detect d(i),
By using the estimated channel vector c
P4,V
1
[P4,H (1) + P4,H (2)]
2
N4
1
[C4 +
wn PH (n |n )]
4
n=1
(7)
N4
1
[C4 +
wn PV (n |n )]
4
n=1
(8)
P8,H (2) =
P8,H (3) =
1
[PH (0|1) + PH (0|3) + PH (0|5)
4
+PH (0|7)],
(9)
1
[2 + PH (4|5) PH (4|5)
4
+PH (4|7) PH (4|7) + PH (4|3)
+PH (4|1) PH (4|3) PH (4|1)] (10)
1
2 + PH (6|7) PH (2|7) + PH (2|7)
4
PH (6|7) PH (6|5) + PH (2|5)
PH (2|5) + PH (6|5) PH (6|3)
+PH (2|3) PH (2|3) + PH (6|3)
PH (2|1) + PH (6|1) + PH (2|1)
PH (6|1)
(11)
MA et al.: EXACT BER FOR M-QAM WITH MRC AND IMPERFECT CHANNEL ESTIMATION IN RICIAN FADING CHANNELS
=
=
1
[P8,H (1) + P8,H (2) + P8,H (3)]
3
N8
1
[C8 +
wn PH (n |n )]
12
n=1
+
wn PH (n |n , J 2m + 1)] (13)
J m=1 n=1
(14)
1
[(log2 I)PI,H + (log2 J)PJ,V ]
log2 (I J)
n
n
n
wn
(12)
B. M -QAM
2
[CJ
J log2 J
NJ
I
1
+
wn PV (n |I 2m + 1, n )]
I m=1 n=1
TABLE I
C OEFFICIENTS FOR BER CALCULATION FOR 4-PAM AND 8-PAM.
(a) 4-PAM
PJ,V =
929
(15)
1
2
3
1
2
-2
3
-1
3
-2
1
1
4
2
1
-1
5
0
3
1
6
0
1
1
(b) 8-PAM
n
1
2
3
4
5
6
7
8
9
10
11
12
13
14
n
0
0
0
0
4
-4
4
-4
-4
-4
4
4
6
2
n
1
3
5
7
5
5
7
7
3
1
3
1
7
7
wn
1
1
1
1
1
-1
1
-1
1
1
-1
-1
1
-1
n
15
16
17
18
19
20
21
22
23
24
25
26
27
28
n
-2
-6
6
2
-2
-6
6
2
-2
-6
2
6
-2
-6
n
7
7
5
5
5
5
3
3
3
3
1
1
1
1
wn
-1
-1
-1
1
-1
1
-1
1
-1
1
-1
1
1
-1
1
2
PI,H =
CI (m)
I log2 I J m=1
+
NI
n=1
PM =
CM
M log2 M
M N
M
1
+
wn PH (n |n , M 2m + 1) (17)
M m=1 n=1
When M = 16 and 64, the values of {wn , n , n } have
to be selected from Table I (a) and Table I (b), respectively.
Next, we re-formulate the conditional BERs PH (n |h , v )
and PV (n |h , v ) for QAM in terms of the distribution of
some new DVs. PH (n |h ) and PV (n |v ) for PAM can be
obtained as special cases.
We note that in [8] the BER of M -QAM is obtained by
extending the result of I-PAM to QAM. The result in [8] is
valid for the case of perfect channel state information (CSI)
930
By definition, PH
(n |h , v ) = P {DR < n Ea |d(i)
R =
= (h + jv ) E
}. Using the equalities D
a L
L
2
(
l (i)yl (i))/ l=1 |
cl (i)|
and
yl (i)
=
l=1 c
cl (i) Ea (h + jv ) + nl (i), we obtain PH (n |h , v ) =
P {DH (n |d(i)) < 0}, where DH (n |d(i)) is a new DV
defined as
L
DH (n |d(i)) =
cl (i)yl (i)
l=1
L
|
cl (i)|2
Ea n
(18)
l=1
cl (i)yl (i) Ea n
|
cl (i)|2
DV (n |d(i)) =
l=1
l=1
where
= vH (d)QH (n )v(d)
= vH (d)QV (n )v(d)
v(d)
QH (n )
QV (n )
=
c(i)
c(i)d + n(i)
n aIL
0.5IL
0.5IL
0LL
n aIL
0.5jIL
0.5jIL
0LL
1
(d))
exp(
vH (d)[Q1
v (d)]1 v
H (n )s
det(I2L sv (d)QH (n ))
(19)
(d) and v (d) are the mean vector and the covariance
where v
(d) and
matrix of v(d), respectively. The procedure to find v
v (d) for generalized Rician fading and any linear channel
estimators is outlined below.
c (i) + cf (i), where
c (i) and cf (i) are the
Let c(i) =
LOS and diffusive component vectors of c(i), respectively.
The mean and the auto-covariance matrix of v(d), respectively, be given by
(d)
v
v (d)
T (i), Tc (i)d]T
[
c
c,c
c,c d
2
H
c,c d c,c |d| + n (0)
(20)
(21)
PPS WPS
(22)
H
yPS cf (i)]
c,c = E[W
H
= WE[( PPS cfPS + nPS )cH
(23)
f (i)] = Wc,yPS
H
c,c = E[W( PPS cfPS + nPS )( PPS cfPS + nPS ) WH ]
=
WyPS WH
(24)
MA et al.: EXACT BER FOR M-QAM WITH MRC AND IMPERFECT CHANNEL ESTIMATION IN RICIAN FADING CHANNELS
H
where c,yPS = E[cf (i)
yPS
] is the cross-covariance matrix
H
PS
between c(i) and yPS , and yPS = E[
yPS y
] is the autocovariance matrix of yPS . They can be obtained by replacing
Rc (m) in (3) and (4) with c (m), respectively, as shown
by the equations given next and at the top of the next page,
respectively,
c,yPS =
PPS [c (P F1 ioff ), . . . , c (ioff ),
. . . , c (P (F2 1) ioff )]
A special case is discussed below. For the case of
white noise, independent diversity branches, and identical fading correlation functions at all the branches (i.e.,
R
c1 (n) = . . . = RcL (n) = Rc (n)), we have c,yPS =
c (P (F2 1)
PPS [Rc (P F1 ioff ), . . . , Rc (ioff ), . . . , R
2
2
ioff )] diag(c1 , . . . , cL ), and yPS = N0 IF L +
c (P (F 1))
c (P ) . . . R
1
R
c (P )
R
1
...,
PPS
..
.
.
.
.,
c (P (F 1))
R
...
1
diag(c21 , . . . , c2L )
Applying (20) (24) to (19), we can evaluate the MGF of
the new DV. The cdf P {DH (n |d) < 0} can be evaluated by
the inverse Laplace transform of the MGF, as shown by
P {DH (n |d) < 0} =
c+j
DH (n |d) (s)
1
ds
2
js
cj
(25)
+R
P {DH (n |d) < 0} =
2N n=1
2N
where ()
= DH (n |d) (c jc tan(/2))(1 j tan(/2)),
931
L
l=1
an
0.5
0.5
0
where QH,2 (n ) =
and vl (d) =
2
cl
l cl cl d
. Let the two eigenvalues of
l cl cl d c2l |d|2 + N0
vl (d)QH,2 (n ) be denoted by
l (n , d). Since the signals in all L branches are independent, all the poles of
+
L
DH (n |d) (s) are given by {1/
l (n , d), 1/l (n , d)}l=1 .
Using the result in the Appendix, we obtain
1
[(l,R h l,I v )acl cl n ac2l ]
2
1
c {a2 [c2l (1 2l,R )2v + c2l (1 2l,I )2h
2 l
2
2c,l
l,R l,I h v 2n cl cl
l (n , d) =
l=1 m=1,m=l
l (n , d)
l (n , d) m (n , d)
(28)
932
yPS =
H
PPS H
c (P ) + n (P ) . . .
PPS c (0) + n (0)
...
..
.
L1
L + l 1
L
P {DH (n |d) < 0} = [A(n |d)]
l
l=0
(29)
(n ,d)
where A(n |d) = (n,d)
+ ( ,d) . By using (27), we
n
obtain (30), shown at the top of the next page.
Let us rewrite A(n |d) as A(n |d) = A(n |h , v ), where
d = a(h +jv ). Then, we obtain a closed-form average BER
expression for M -QAM with L-fold MRC as (31).
For comparison, the result given in [11, eq. (37)] is in
the form of a two-dimensional (2-D) integral of the 2L-order
parabolic cylinder function, and thus is numerically intensive
to evaluate. In [12, eqs. (9), (12)], an alternative closed-form
BER expression for M -QAM with PSAM was presented. Our
result in (31) and (30) explicitly shows the relation between l
and the BER, and thus provides more insight into the behavior
of PSAM, as will be discussed next in Subsection IV-B.3.
3) MMSE-CE: The MMSE channel estimator for PSAM
has attracted a lot of research interest in the literature, and
it provides a performance benchmark for other linear channel
estimators. For MMSE-CE with i.n.d. branches, we can show
that
H ] /
[Wmmse yPS Wmmse
l =
l,l
cl
H
=
[c,yPS 1
(32)
yPS c,yPS ]l,l /cl
2 2
2 2
|l | cl
|l | cl d
vl (d) =
(33)
|l |2 c2l d c2l |d|2 + N0
l (d, n ) =
1 2 2
1
cl al (h n ) l {c4l a2
2
2
[2h + 2l n2 + (1 2l )2v 22l n h ]
(34)
+c2l N0 }1/2
1 1
[ac (h n )]
2 2
[c2 a2 (2h + 2 n2 + (1 2 )2v
22 n h ) + N0 ]1/2
[1 A(n |d)]l
(35)
)
=
a(
),
which
shows
that an error
and +
n
h
n
cl
l
floor does not occur in this case. Thus, {l }L
l=1 is closely
related to the performance of diversity QAM with PSAM. The
error floor for M -QAM with MMSE-CE can be analytical
evaluated by using (32)(35) with (28) or (29). We note
that the normalized residual MSE (which is proportional to
1|l |2 defined in this paper) of MMSE-CE in a non-diversity
Rayleigh channel has recently been studied in [24, Appendix].
It was shown that for a bandlimited fading process (e.g.,
the Jakes model) with Bf Ts P < 1/2 and infinite PSAM
interpolation order (F ), the normalized residual MSE
decreases proprotionally as the SNR increases, which suggests
that under these conditions an error floor does not occur.
Finally, we underscore that we have verified that for all
the square and rectangular M -QAM constellations obtained
in this paper, our new BER results, when simplified to the
case of perfect CSI, are numerically identical to the known
results given in [3], [8].
V. N UMERICAL E XAMPLES AND D ISCUSSIONS
In this section, we provide some numerical results for the
performance of the M -QAM MRC receiver in general Rician
fading channels taking into account the effects of ICE and
various system and channel parameters. We assume Pd = PPS
unless otherwise stated. To clearly illustrate the effects of
Doppler fading bandwidth and the Rician K factor, we also
assume Bf,l = Bf , fd,l = fd , and Kl = K (for all l) for all
the cases studied.
A. Verification of the BER Analysis
For correlated fading, the popular constant spatial correlation model [3] (with correlation coefficient c ) is assumed.
Forexample, for balanced
branches and L = 3, c (0) =
1 c c
c2 c 1 c . To show the effect of the phase vector
c c 1
= [1 , , L ] of the LOS components, we assume fd Ts =
0 and Bf Ts = 0.02. In our simulation, we assume Rc (n) =
Rc (0)J0 (2nBf Ts ), i.e., the spatial and temporal correlations
can be separated.
MA et al.: EXACT BER FOR M-QAM WITH MRC AND IMPERFECT CHANNEL ESTIMATION IN RICIAN FADING CHANNELS
A(n |d) =
PM
933
M N M
4
1
{C M +
wn [A(n |n , M 2m + 1)]L
log2 M
M m=1 n=1
L1
L+l1
[1 A(n |n , M 2m + 1)]l }
l
(31)
l=0
10
= [0, 0, 0]
10
10
10
L=1
10
L=2
10
10
= [0, /2, ]
MMSECE, simu.
10
4
10
L=3
6
10
10
15
Average bit SNR per branch (dB)
20
0.1
0.2
0.3
0.4
0.5
0.6
Spatial correlation coefficient
0.7
0.8
0.9
Fig. 3.
BER (averaged over all ioff s) vs. average bit SNR per branch
for 16-QAM MRC receivers with CSI and the MMSE-CE, respectively, in
independent and correlated Rician (K = 5 dB) fading channels. = 01L ,
L = {1, 2, 3}, P = 10, F1 = F2 = 5, Bf Ts = 0.02, fd Ts = 0, and
fo = 0.
Fig. 4. BER (averaged over all ioff s) vs. spatial correlation coefficient c
for 16-QAM MRC receivers with CSI, MMSE- and sinc-interpolator-based
(with rectangular and Hamming windows) channel estimators, respectively, in
Rician (K = 5 dB) fading channels. = [0, 0, 0] and [0, /2, ],
b = 10
dB, L = 3, P = 10, F1 = F2 = 5, Bf Ts = 0.02, fd Ts = 0, and fo = 0.
the perfect CSI case, and the SNR loss is within 2 dB for
all SNRs shown in the figure; while in Fig. 6, for a larger
Bf Ts (Bf Ts = 0.03) and a larger P the relevant SNR gap
becomes larger. Fig. 5 shows that the BER of sinc-CE with
Hamming window (sinc-Hamm-CE) is uniformly lower than
that of sinc-CE with rectangular window (sinc-rect-CE). On
the contrary, for the parameters studied in Fig. 6, QAM with
sinc-rect-CE performs significantly better than QAM with
sinc-Hamm-CE. In addition, the SNR gaps between sinc-CE
and MMSE-CE depend on the average SNR, P , F , Bf Ts ,
and other parameters. This observation demonstrates that the
relative performance between sinc-rect-CE and sinc-HammCE, and the performance gap between sinc-CE and MMSECE, crucially depend on the channel and system parameters.
B. Effects of Design Parameters
In the following, we show the effects of some design
parameters, including the pilot symbol insertion interval P ,
number of pilot symbols F , and the ratio PPS /Pd on the
performance of the MRC M -QAM.
We show the effect of P on BER of 16-QAM with MMSECE and sinc-CE in Fig. 7, assuming L = 2 and ioff = P/2
.
Our results show that the BER degrades slowly when P 14.
When P > 14, both the performances of MMSE-CE and sincHamm-CE start to degrade rapidly. Sinc-Hamm-CE may give
934
10
10
Perfect CSI
MMSECE
sincHammCE
MMSECE, simu.
32QAM
1
10
SincCE, Hamm
Win., simu.
10
SincCE, Rect.
Win, simu.
Analysis
16QAM
B T = 0.03, L=2
10
10
10
Rayleigh Fading
10
10
10
K = 5 dB
10
10
10
15
Average bit SNR at the first branch (dB)
20
10
25
Fig. 5. BER (averaged over all ioff s) vs. average bit SNR of the first branch
for 32-QAM MRC receivers with CSI, MMSE- and sinc-interpolator-based
channel estimators, respectively, in i.n.d. Rayleigh and Rician (K = 5 dB)
fading channels. L = 2, P = 10, F1 = F2 = 4, Bf Ts = 0.02, fd Ts =
0.01, and fo = 0. The SNR in the first branch is 2 dB larger than that in the
second branch.
16QAM
Dash lines: Rayleigh
Solid lines: K=5 dB
With channel gain power normalization
10
16QAM
25
10
20
Fig. 7. BER versus the pilot symbol insertion interval P for MRC 16-QAM
with MMSE-CE and sinc-CE, respectively, in i.i.d. Rayleigh and Rician (K =
5 dB) fading channels.
b = 20 dB, L = 2, ioff = P/2, F1 = F2 = 6,
Bf Ts = 0.03, fd Ts = 0.015, fo = 0.
10
10
10
15
Pilot symbol interval P
Perfect CSI
MMSECE
sincHammCE
sincRectCE
F=20, P=10
BfT=0.03, L=3
SNR = 10 dB
10
10
10
10
SNR = 20 dB
5
10
10
10
10
10
10
10
15
Average bit SNR per branch (dB)
3
4
5
Power Ratio between Pilot and Data symbols PPS/Pd
20
Fig. 6. BER (averaged over all ioff s) vs. average bit SNR per branch for 16QAM MRC receivers with CSI, MMSE- and sinc-interpolator-based channel
estimators, respectively, in i.i.d. Rayleigh and Rician (K = 5 dB) fading
channels. L = 2, P = 15, F1 = F2 = 5, Bf Ts = 0.03, fd Ts = 0.015,
and fo = 0.
Fig. 8. BER vs. the power allocation ratio PPS /Pd for MRC 16-QAM
with MMSE-CE and sinc-CE (with power normalization), respectively, in
i.i.d. Rayleigh and Rician (K = 5 dB) fading channels.
b = {10, 20} dB,
L = 3, P = 10, ioff = 5, F1 = F2 = 10, Bf Ts = 0.03, fd Ts = 0.015,
fo = 0.
WyPS WH /.
Fig. 8 shows that PPS /Pd [2, 3] gives optimal BER
performance for MMSE-CE and sinc-CE, and compared to
the case of PPS /Pd = 1 the BERs can be reduced by a
factor of about 1.5 2.5. We note that for non-constant
modulus formats, such as M -QAM, the power normalization
of the channel estimate is necessary (except for the MMSE-CE
where the power normalization has been taken into account in
Wmmse ). Otherwise, the case of PPS /Pd = 1 (e.g., for sinc-CE)
may severely degrade the performance.
We study the effects of F2 and F on the BER of the
M -QAM MRC receivers with MMSE-CE and sinc-CE. We
present the BER vs. F2 for 32-QAM in i.i.d. Rician and
Rayleigh fading channels in Fig. 9 for F = 24. The cases of
F2 = 0 and F2 = 12 correspond to a prediction-based (thus
MA et al.: EXACT BER FOR M-QAM WITH MRC AND IMPERFECT CHANNEL ESTIMATION IN RICIAN FADING CHANNELS
935
10
VI. C ONCLUSIONS
Perfect CSI
MMSECE
sincRectCE
B T = 0.03, P=15
f
L=2, SNR=20 dB
1
32QAM
Dashed lines: Rayleigh Fading
Solid lines: K = 5 dB
10
10
10
10
10
10
15
20
25
F2
Fig. 9. BER versus F2 for MRC 32-QAM with MMSE-CE and sinc-rectCE, respectively, in i.i.d. Rayleigh and Rician (K = 5 dB) fading channels.
10
Perfect CSI
MMSECE
sincHammCE
sincRectCE
Bf T = 0.03, P=15
K = 5 dB, L=2
2
10
64QAM
10
10
10
16QAM
6
10
10
15
20
Number of pilot symbolsF
25
30
35
Fig. 10. BER versus F for MRC 16- and 64-QAM receivers with MMSECE and sinc-CE, respectively, in an i.i.d. Rician fading channel. K = 5 dB,
[1] L. Hanzo, W. Webb, and T. Keller, Single- and Multi-Carrier Quadrature Amplitude Modulation. New York: Wiley, 2000.
[2] G. L. Stuber, Principles of Mobile Communications. Norwell, MA:
Kluwer Academic Publishers, second ed., 2003. Third Printing.
[3] M. K. Simon and M.-S. Alouini, Digital Communications over Fading
Channels: A Unified Approach to Performance Analysis. John Wiley &
Sons, 2000.
[4] J. K. Cavers, An analysis of pilot symbol assisted modulation for
Rayleigh fading channels, IEEE Trans. Vehic. Technol., vol. 40,
pp. 686693, Nov. 1991.
[5] M. G. Shayesteh and A. Aghamohammadi, On the error probability
of linearly modulated signals on frequency-flat Ricean, Rayleigh and
AWGN channels, IEEE Trans. Commun., vol. 43, pp. 14541466,
Feb./Mar./Apr. 1995.
[6] S. K. Wilson and J. M. Cioffi, Probability density functions for analyzing multi-amplitude constellations in Rayleigh and Ricean channels,
IEEE Trans. Commun., vol. 47, pp. 380386, March 1999.
[7] X. Tang, M.-S. Alouini, and A. J. Goldsmith, Effect of channel
estimation error on M-QAM BER performance in Rayleigh fading,
IEEE Trans. Commun., vol. 47, pp. 18561864, Dec. 1999.
[8] K. Cho and D. Yoon, On the general BER expression of one- and twodimensional amplitude modulations, IEEE Trans. Commun., vol. 50,
pp. 10741080, July 2002.
[9] X. Dong and L. Xiao, Two-dimensional signaling in Ricean fading
with imperfect channel estimation, in Proc. IEEE Wireless Commun.
and Networking Conf. (WCNC), pp. 11591164, 21-25 Mar. 2004.
[10] X. Dong and L. Xiao, Symbol error probability of two-dimensional
signaling in Ricean fading with imperfect channel estimation, IEEE
Trans. Vehic. Technol., vol. 54, pp. 538549, Mar. 2005.
[11] L. Cao and N. C. Beaulieu, Exact error-rate analysis of diversity 16QAM with channel estimation error, IEEE Trans. Commun., vol. 52,
pp. 10191029, June 2004.
[12] B. Xia and J. Wang, Effect of channel-estimation error on QAM systems with antenna diversity, IEEE Trans. Commun., vol. 53, pp. 481
488, March 2005.
[13] L. Cao and N. C. Beaulieu, Closed-form BER results for MRC diversity
with channel estimation errors in Ricean fading channels, IEEE Trans.
Wireless Commun., vol. 4, pp. 14401447, July 2005.
[14] Y. Ma, R. Schober, and D. Zhang, Exact BERs for M -QAM with
MRC and channel estimation errors in Rician channels, in Proc. IEEE
Wireless Commun. and Networking Conf. (WCNC), vol. 2, pp. 967972,
Mar. 2005.
[15] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis. Cambridge
University Press, 1991.
[16] S. Haykin, Adaptive Filter Theory. Englewood Cliffs, NJ: Prentice Hall,
3rd ed., 1996.
[17] F. D. Neeser and J. L. Massey, Proper complex random processes with
applications to information theory, IEEE Trans. Inform. Theory, vol. 39,
pp. 12931302, July 1993.
[18] G. L. Turin, The characteristic function of hermitian quadratic forms
in complex normal variables, Biometrika, vol. 47, pp. 199201, 1960.
936
A PPENDIX
D ERIVATION OF E QUATION (27)
For the Rayleigh fading with i.n.d. branches, let us
consider the eigenvalue pair
l (n , d) at the lth branch.
For simplicity,
we
suppress
the
subscript
l below. Let
c2
cc d
and QH,2 (n ) =
v (d) =
2
2
cc d c |d| + N0
an 0.5
, where d = (h + jv )a and = R + jI .
0.5
0
Here, R and I are the real and imaginary parts of ,
respectively.
It follows that v (d)QH,2 (n ) =
n ac2 + 0.5 cc d
0.5c2
.
n acc d + 0.5c2 |d|2 +0.5N0 0.5
cc d
c1 c2
For a matrix of the form
we obtain its two
c3 c4
eigenvalues as [26],
1
(36)
(c1 + c4 ) 4c2 c3 + (c1 c4 )2
=
2
Based on this observation, we obtain
(c1 c4 )2 = n2 a2 c4 c2 c2 [(d)]2 + j2n ac3 c (d),
4c2 c3 = 2n ac3 c d + c2 c2 |d|2 + c2 N0 , and c1 + c4 =
(d)cc n ac2 . Applying these equalities to (36), we
obtain the two eigenvalues of v QH,2 (n ) given in (27).