1

Adaptive Coded Modulation with Receive Antenna

Diversity and Imperfect Channel Knowledge at
Receiver and Transmitter
Duc V. Duong, Student Member, IEEE, Geir E. ien, Member, IEEE, and Kjell J. Hole, Member, IEEE
AbstractUnifying and generalizing the work in [1] and [2],
we analyze and optimize the performance of an adaptive trellis-
coded modulation system where receive antenna diversity is
implemented by means of maximum ratio combining. As in
[1], the analysis is done in the presence of both estimation
and prediction errors. Rayleigh fading on each subchannel is
considered, with the estimation and prediction being performed
independently on each subchannel. The system optimization
process is done in such a way that the throughput is maximized
under a bit error rate (BER) constraint. The numerical example
employs a Jakes fading spectrum and shows how the power
should be distributed between pilot and data symbols, and
how often pilot symbols should be transmitted for maximum
throughput under an instantaneous (with respect to the predicted
channel) BER constraint.
Index TermsAdaptive coded modulation, PSAM, estimation
and prediction error, optimal pilot spacing, optimal power
allocation, antenna diversity, and maximum ratio combining.
I. INTRODUCTION
A
DAPTIVE coded modulation (ACM) is an efcient tool
to counteract fading in wireless channel communications,
and to improve spectral efciency [3]. In addition, it is well
known that receive diversity stabilizes the channel quality. The
scheme presented in this paper is a generalization of two recent
papers; one by Cai and Giannakis [1], in which an adaptive
modulation system based on uncoded M-QAM constellations
with single transmit and receive antenna is investigated, and
one by ien et al. [2] where an ACM system with maximum
ratio combining (MRC) reception was investigated assuming
perfect receiver channel state information (CSI). We extend
and unify the idea of these papers by analyzing the case when,
as in [2], coding is included and when MRC is implemented
on the receiver side. The system parameters are optimized as
in [1], for maximal spectral efciency under an instantaneous
bit error rate (BER) constraint.
As in [1] and [2], the estimator and predictor in our adaptive
scheme are based on received known pilot symbols which
are sent from the transmitter along with data information.
This technique is known as pilot-symbol-assisted modulation
(PSAM) [4]. The predicted CSI is fed back to the transmitter
Duc V. Duong and Geir E. ien are both with the Department of
Electronics and Telecommunications, Norwegian University of Science and
Technology, O. S. Bragstads pl. 2B, N-7491 Trondheim, Norway (email:
{duong, oien}@iet.ntnu.no).
Kjell J. Hole is with the Department of Informatics, University of Bergen,
Pb. 7800, N-5020 Bergen, Norway (email: [email protected]).
The work is supported by the Research Council of Norway under the project
BEATS (URL: http://www.iet.ntnu.no/projects/beats/).
via a zero-error, non-zero delay feedback channel. Based
on this CSI the transmitter dynamically adapts the rate and
the power to a mode which maximizes the average spectral
efciency (ASE) while maintaining the instantaneous BER
below some pre-dened target value BER
0
. The instantaneous
BER here is with respect to the predicted channel-signal-to-
noise ratio (CSNR). As in [2], [5], 4-dimensional (4-D)thus,
multi-dimensionaltrellis codes designed for additive white
Gaussian noise (AWGN) channels are used as component
codes.
In [6], [7] and [8] multiple-input multiple-output (MIMO)
systems (with single-input multiple output as a special case)
are analyzed where perfect, but delayed CSI (thus outdated)
was assumed. No pilot symbol is transmitted in their system,
and the BER constraint is subject to instantaneous BER and
average BER, respectively. Note also that in [6] 2-D trellis
codes are utilized.
Thoughout the paper we will use the following notations:
superscripts ()
T
, ()

, and ()
H
stand for transpose, complex
conjugate, and Hermitian transpose, respectively. z means
the integer part of z, and E[z] stands for the expectation of z.
The outline of the remainder of this paper is as follows.
We describe our system in Section II. Channel estimation and
prediction is dealt with in Section III. The analysis of BER
performance and ASE performance can be found in Sections
IV and V, respectively. A numerical example is given in
Section VI before the conclusions are drawn in Section VII.
II. SYSTEM MODEL
The system we are considering is illustrated in Fig. 1, where
we have n
R
i.i.d. receive antennas. The adaptive encoder
always chooses to transmit symbols from the constellation of
size M
n
that is best suited to the predicted channel state, out
of a set of constellations of sizes {M
n
}
N
n=1
, corresponding
to the set of spectral efciencies (SEs) {R
n
}
N
n=1
. The SEs
are ordered such that R
n
< R
n+1
and hence M
n
< M
n+1
n. The choice of code is based on the CSI fed back from the
receiver: code n is chosen if the predicted CSNR falls between
the switching thresholds
n
and
n+1
. By letting
0
= 0 and

N+1
= , we have
n
<
n+1
for all n {0, 1, , N}.
The power allocation between data and pilot symbols will
be chosen in an optimal way. The criterion for choosing
the code and the transmit power distribution is to fulll a
target BER performance while maximizing the throughput in
terms of ASE. We include MRC on the receiver side with the
2
.
.
.
.
.
Pilot
insertion
Power
control
Adaptive encoder
and modulator channel
Fading
Predictor
Predictor
Buer
Buer
Zero-error,
delay
feedback channel
Pilot period,
power, and
constellation
selector
Estimator
Estimator
1
n
R
coherent detector,
adaptive decoder,
and demodulator
Combiner,
Fig. 1. The adaptive PSAM system model combined with multi reception of receiving signal.
assumption that the branches are mutually uncorrelated. This
assumption can be fullled if the antennas are spaced at least
half the wavelength from each other [9, p. 330]. However,
due to the difference of the scattering environment around the
base station and the remote station, about quarter wavelength
spacing of the antennas is sufcient to provide enough spatial
decorrelation between the antennas at the mobile terminal
[10]. On the other hand, seperation of 1020 wavelengths is
required at the base station [10]. It is noted that impact of
spatial correlation between different receive antennas on our
system is considered in [11] and [12]. Also, we have extended
the system further to operate on a MIMO channel in [13].
The issue about the needed accuracy of PSAM-based channel
prediction for MIMO channels using transmit-beamforming is
considered in [14].
Similar to in [1], [15], the transmitted data stream is divided
into frameseach of length Lwith a pilot symbol at the
beginning of each frame, and is followed by L1 information
symbols. Both the estimator and predictor are linear and are
made optimal in the maximum a posteriori (MAP) sense
[16]. The difference between the two is that the estimator is
non-causal and uses information from both past and future
symbols, while the predictor is strictly causal and predict
relatively far into the future. Due to this, the order of the
estimator can usually be lower than the order of the predictor.
The received, noisy, and faded pilot symbols of the jth
branch are written in complex baseband as
y
pl;j
(k; l) =
_
E
pl
h
j
(k; l)s(k; l) + n
j
(k; l),
l = 0; j = 1, , n
R
, (1)
and the received data symbols as
y
d;j
(k; l) =
_
E
d
h
j
(k; l)s(k; l) + n
j
(k; l),
l = 1, , L 1; j = 1, , n
R
. (2)
The index k counts the frame and l is the symbol index in
that frame. E
pl
and E
d
is the power per pilot and per data
symbol, respectively, to be optimized later. s(k; 0) is the pilot
symbol and {s(k; l)}
L1
l=1
are data symbols in the kth frame.
For simplicity, we assume that E[|s(k; l)|
2
] = |s(k; 0)|
2
=
1. Furthermore, n
j
(, ) denotes zero mean complex-valued
AWGN with variance N
0
/2 per dimension and dimensions
being uncorrelated. The fading envelope of the jth branch,
h
j
(k; l), is assumed to be a stationary complex Gaussian
random process with zero mean and variance
2
hj
= 1. Each
subchannel is assumed to be slowly varying so that it remains
constant over many channel symbols, and we perform the
estimation and prediction independently on each subchannel.
Note that the slowly fading conditions is in the sense that
the channel does not reveal its ergodic properties over one
data packet. I.e., it is assumed that the data packet length is
signicantly smaller than the coherence time of the channel,
implying low terimal velocity in the system.
The total average power for both pilot and data symbols is
denoted by E. The average power per data symbol is

E
d
=
LE/(L1), and per pilot symbol is E
pl
= (1 )LE. The
constant determines how power is distributed between pilot
and data symbols. Equal data and pilot power is obtained when
= 1 1/L.
III. CHANNEL ESTIMATION AND PREDICTION
The analysis in this section is done assuming that the
channel is wide-sense stationary, and that the correlation
function of the channel is known.
A. Estimation
The optimally estimated channel in the MAP sense for
the complex Gaussian case is a linear combination of the
noisy observations [16, p. 741742]. Following the approach
in [1] we can nd the estimated channel h
e;j
(k; l), with the
corresponding minimum MSE (MMSE) of the estimated jth
branch as

2
e;j
(l) = 1
Ke

=1
|u
H

r
e
|
2
(1 )L
j
(1 )L
j

+ 1
. (3)
In (3), K
e
is the estimator order,
j
= E/N
0
is the expected
CSNR on any receive branch, {u

} denote the eigenvectors of

the covariance matrix of the channel gain R
e
= E[hh
H
] where
h = [h
j
(kK
e
/2; 0), , h
j
(k+(K
e
1)/2; 0)]
T
, {

}
are the corresponding eigenvalues, and r
e
is the covariance
vector; r
e
= E[hh

j
(k; l)].
Due to the orthogonality principle [17, p. 264] h
e;j
(k; l)
and the estimation error
e;j
(k; l) = h
j
(k; l) h
e;j
(k; l) are
uncorrelated. Both are zero mean Gaussian random variables
[18, Chap. 7.2]
B. Prediction
For notational simplicity, we restrict the system feedback
delay
1
to be = DLT
s
, where D is a positive integer and T
s
1
The feedback delay here includes the time it takes to perform prediction,
actual transmission delay on the feedback channel, and the processing time
needed by the transmitter to activate the code to be transmitted.
3
is the duration of a channel symbol. The predictor uses K
p
pilot symbols from the past to predict one sample in the set
{h
j
(k; l)}
L1
l=1
of the kth frame. Similarly to the estimation
case we can nd the predicted channel h
p;j
(k; l), which is
Gaussian distributed with zero mean and variance
2
hp;j
(l) =

2
hj

2
p;j
(l) = 1
2
p;j
(l). The corresponding MMSE of the
prediction error is

2
p;j
(l) = 1
Kp

=1
|u
H

r
p
|
2
(1 )L
j
(1 )L
j

+ 1
. (4)
Note that {u

} now denote the eigenvectors of the covariance

matrix of the channel gain R
p
= E[hh
H
] where h = [h
j
(k
D; 0), , h
j
(kDK
p
+1; 0)]
T
, {

} are the corresponding

eigenvalues, and r
p
= E[hh

j
(k; l)] is the covariance vector.
IV. BER PERFORMANCE ANALYSIS
A. BER Analysis in the Presence of Estimation Errors
To reduce the complexity of the receiver, we use the subop-
timal symbol-by-symbol maximum likelihood (ML) detection
on each subchannel as z(k; l) = y
d;j
(k; l)/(

E
d
h
e;j
(k; l)). As
a result of that detection rule, the CSNR of a single branch
at the time of detection can be found in [1]. The total CSNR
after MRC is the sum of the individual branch CSNRs,
(k; l) =
nR

j=1
E
d
|h
e;j
(k; l)|
2
N
0
+ gE
d

2
e;j
(l)
=
E
d

nR
j=1
|h
e;j
(k; l)|
2
N
0
+ gE
d

2
e
(l)
, (5)
where the last equality is obtained by assuming that the
variance of the estimation error is the same for all branches.
This is reasonable since the branches are assumed uncorrelated
and the same estimation algorithm is used on all the branches.
The constant g = 1 for 4-QAM and g = 1.3 for M-QAM
when M > 4 [1].
ACM systems using 2-D trellis codes was rst introduced
by Goldsmith and Chua in [19]. However, to the best of the
authors knowledge, extending ACM to 4-D trellis codes was
rst done by Hole, Holm, and ien in [20]. In this paper the
4-D trellis codes in [20] are utilized. Tight approximations
for 4-D trellis-coded modulation (TCM) BER performance on
AWGN channels can be found in [2], [5], and [20]. However,
in order to obtain a closed-form and mathematically tractable
solution when MRC is considered we use a somewhat looser
BER approximation expression which, similar to [21], is given
by
BER
_
M
n

_
=
L

=1
a
n
() exp
_

b
n
()
M
n

_
. (6)
Again, M
n
is the number of points in the symbol constellation,
and a
n
() and b
n
() are constellation dependent constants
(given in Tab. I) which can be found by rst simulating the
codes BER performance and then using curve tting with the
least squares method. L is the number of exponential functions
which approximate the simulated BER. The approximation is
illustrated in Fig. 2 for different constellation sizes. At rst
sight, the approximations seem to be quite coarse. However,
it is shown later that the ACM performance results of this
approximation are very close to those obtained when the
TABLE I
THE CODE-DEPENDENT CONSTANTS {an()}
3
=1
AND {bn()}
3
=1
FOR
THE EXAMPLE 4-D TRELLIS CODES.
n an(1) bn(1) an(2) bn(2) an(3) bn(3)
1 233.8034 12.4335 -280.8712 11.4405 51.3394 8.6131
2 210.6415 8.4208 -242.0657 7.9916 34.3732 6.0432
3 246.0565 7.7677 -334.6900 8.1130 89.4924 9.1087
4 99.7887 7.7426 -160.6040 8.3843 61.5091 9.4667
5 78.0083 7.0135 -100.8414 7.4793 23.4319 9.0714
6 86.2181 7.3704 -96.0270 7.6780 10.3583 10.3191
7 87.6912 7.0471 -94.5020 7.2852 7.3344 10.1898
8 89.3099 7.2848 -95.6889 7.4987 6.8972 10.4428
0 5 10 15 20 25 30 35
10
6
10
5
10
4
10
3
10
2
10
1
10
0
M=4
M=8
M=16
M=32
M=64
M=128
M=256
M=512
CSNR [dB]
B
E
R
Fig. 2. BER performance of TCM codes on AWGN channels for different
M-QAM constellations. The solid lines denote the approximations, while the
stars represent the simulated values.
tighter approximations in [2] and [5] are used. This is seen
by comparing the results for the single antenna system in [5]
and what we achieve here for n
R
= 1.
Inserting (5) into (6) the BER becomes
BER
_
M
n

{h
e;j
}
_
=
L

=1
a
n
()
nR

j=1
exp
_
A
n
E
d
|h
e;j
(k; l)|
2
_
(7)
where A
n
= b
n
()/(M
n
(E/
j
+ gE
d

2
e
(l))). This equation
will be useful when deriving the BER in the presence of both
the prediction and estimation errors in the next subsection.
B. BER Analysis in the Presence of both Estimation and
Prediction Errors
We may express the estimated channel as
h
e;j
(k; l) = h
p;j
(k; l) +
p;j
(k; l)
. .
hj(k;l)

e;j
(k; l). (8)
For a single antenna system p
_
|h
e;j
|

h
p;j
_
is Rice distributed
[1]. Now we consider the MRC technique with the branches
assumed independent and identical. Thus, p
_
{|h
e
|}

{h
p
}
_
=

nR
j=1
p
_
|h
e;j
|

h
p;j
_
. Each branch is still Ricean distributed,
4
with the Rice factor K = |(1 )h
p;j
(k; l)|
2
/
2
he;j
,
where = E[
e;j
(k; l)h

p;j
(k; l)]/
2
hp;j
(l) is the correlation
between predicted channel and the estimation error, and

2
he;j
= Var (h
e;j
(k; l)|h
p;j
(k; l)) = E[|
p;j
(k; l)
e;j
(k; l) +
h
p;j
(k; l)|
2
]. It is shown in [1] that typically takes on very
small values, and that the deviations in the BER performance
obtained by using the exact value of instead of setting = 0
is of less than 0.2 dB. We therefore choose to set = 0 in our
analysis. Moreover, in contrary to [1],
e;j
(k; l) is correlated
with
p;j
(k; l). Thus, as a result,
2
he;j
=
2
p;j
(l)
2
e;j
(l).
Hence, the BER conditioned on the combined predicted chan-
nels is given by
BER
_
M
n

{h
p;j
}
_
=
_

0

_

0
. .
nRfold
BER
_
M
n

{|h
e;j
|}
_
p
_
{|h
e;j
|}

{h
p;j
}
_
d|h
e;1
| d|h
e;nR
|
=
L

=1
a
n
()
_

0

_

0
. .
nRfold
nR

j=1
exp
_
A
n
E
d
|h
e;j
|
2
_
p
_
|h
e;j
|

h
p;j
_
d|h
e;1
| d|h
e;nR
|. (9)
With the help of [22, Eq. (6.633-4)], by assuming
2
he;j
=
2
he
j, and letting d
n
= 1/(A
n
E
d

2
he
+ 1), Eq. (9) becomes
BER
_
M
n

{h
p;j
}
_
=
L

=1
a
n
()d
nR
n
exp
_
d
n
A
n
E
d
nR

j=1
|h
p;j
|
2
_
.
(10)
The predicted CSNR on each branch is dened as in [2] by

j
=

E
d
|h
p;j
(k; l)|
2
/N
0
, and the combined predicted CSNR
using the MRC scheme is obtained as [23, Eq. (26)]
=

E
d
N
0
nR

j=1
|h
p;j
(k; l)|
2
. (11)
Solving this wrt.

|h
p;j
|
2
and inserting the result into (10)
gives
BER(M
n
| ) =
L

=1
a
n
()d
nR
n
exp
_

d
n
A
n
EE
d

j

E
d
_
. (12)
The combined predicted CSNR is fed back to the trans-
mitter via the return channel and is used for deciding which
code to use. If
n
< <
n+1
, code n (or constellation of
size M
n
) is activated to transmit. To nd the optimal switching
thresholds {
n
}
N
n=1
in a maximal ASE sense, subject to BER
and power constraints, we set (12) equal to BER
0
and solve
for
n
. In contrast to [1], here we have to use a numerical
approach to obtain the solutions; this will be explained in
Section V. The optimal switching thresholds here will be
dependent on the average expected subchannel CSNR, which
is in contrast to [2], where constant switching thresholds
optimized for perfect CSI were used.
From (11) we can nd the average predicted CSNR as

E
d
N
0
nR

j=1
E[|h
p;j
(k; l)|
2
] = r
j
n
R
, (13)
where r =

E
d
(1
2
p
)/E. The second equality is obtained by
assuming that
2
p;j
=
2
p
b, since the branches are assumed
uncorrelated and the same prediction algorithm is used on
all the branches. The overall predicted CSNR with MRC
of n
R
branches will follow a Gamma distribution [2], i.e.,
G(n
R
, r
j
) =

n
R
1
(nR)(r j)
n
R
exp( /(r
j
)).
C. Overall Average BER Performance Analysis
The average BER for the nth constellation is found by
averaging (12) over the Gamma probability density function
of :
BER(M
n
)=
_
n+1
n
BER(M
n
| )p( )d
=
L

=1
a
n
()
_
d
n

E
d
rd
n
A
n
EE
d
+

E
d
_
nR

_
n
R
,
n
rd
n
A
n
EE
d
+

E
d
r
j

E
d
_

_
n
R
,
n+1
rd
n
A
n
EE
d
+

E
d
r
j

E
d
__
, (14)
where (a, z) = (
_

z
t
a1
e
t
dt)/(a) is the normalized
incomplete gamma function [24, Eq. (11.3)].
The most frequently used estimate of the average BER (over
all codes) is the ratio between the average number of bits in
error, and the number of bits transmitted in total [1], [25],
[26]:
BER =

N
n=1
BER(M
n
)R
n

N
n=1
P
n
R
n
, (15)
where P
n
is the probability that [
n
,
n+1
; P
n
=
_
n+1
n
p( )d = (n
R
,
n
/r
j
) (n
R
,
n+1
/r
j
).
V. OPTIMIZATION OF ASE
The goal of this section is to design an adaptive system
where both and L are optimized in such a way that the
ASE is maximized while BER(M
n
| ) BER
0
.
It is obvious that the variance of the prediction error is
largest when predicting the last symbol in a framei.e., l =
L1. On the other hand, the variance of the estimation error
is almost the same for all l if the order of the estimator is
K
e
20 [1]. Thus, we use the variance of the estimation
error
2
e;j
(L1) and the conservative choice of the prediction
error variance
2
p;j
(L1) when nding the optimal switching
thresholds {
n
}, as well as in the further optimization process.
A. ASE Performance Analysis
The system is experiencing an outage when the predicted
CSNR falls below
1
, since by denition there is no code in
our code set which then guarantees the BER performance. In
that case the system does not send anything but the pilots
in order to perform the channel estimation and prediction
while the data must be buffered at the transmitter. And since
no transmission is allowed when <
1
, we do not use any
5
data power during that time. Therefore the actual transmitted
power per data symbol can be set to
E
d
=

E
d
_

1
p( )d
=

E
d
(n
R
,
1
/r
j
)
. (16)
The SE of the nth constellation used by the 2G-
dimensional
2
trellis code is R
n
= (11/L)(log
2
(M
n
)1/G).
Hence, the ASE is given by
ASE=
N

n=1
R
n
P
n
=
L 1
L
N

n=1
_
log
2
(M
n
)
1
G
_

_
n
R
,

n
r
j
_

_
n
R
,

n+1
r
j
__
, (17)
in which the term 11/L accounts for the fact that every Lth
symbol is a pilotin which no information is transmitted.
When using Nyquist sampling, L must be less than L
max
=
1/(2f
d
T
s
) [4] where f
d
is the maximum Doppler shift.
Thus, for L [2, , L
max
] we have the following optimiza-
tion problem:
max

ASE
subject to 0 < < 1. (18)
In order to solve BER(M
n
| ) = BER
0
wrt. we need the
value of E
d
, which again is a function of
1
(cf. (16)). The
problem of nding {
n
} is reduced to nding the
1
since
{
n
}
N
n=2
is easily found when E
d
is given. Thus, for the
rst constellation M
1
and given an , we let E
d
vary as a
function of the whole range of (note that it is not
1
), i.e.,
E
d
( ). The solution of BER(M
1
| ) = BER
0
for that E
d
( )
yields the
1
which is optimal if is optimal. Once
1
is
known, E
d
is explicitly given by (16), and it can be used to
nd other thresholds. In that way, ASE() is reduced from a
2-dimensional (2-D) function S(, E
d
) in [1, Eq. (23)] to a
1-D function. Thus, the optimization of 1-D function of (18)
is easily done by picking an 0, 1 by numerical search
which maximizes the ASE.
3
After solving (18) for all the possible L values, the maxi-
mum ASE is found by searching over all L in order to nd
the and the L values which simultaneously maximize ASE.
VI. NUMERICAL EXAMPLE
At this point, we consider an example ACM system which
has a set of N = 8 QAM signal constellations of sizes
{M
n
} = {4, 8, 16, 32, 64, 128, 256, 512} to switch between.
Those constellations are used to represent eight 4-D trellis
codes. The carrier frequency is 2 GHz and the length of a
channel symbol is 5 scorresponding to a channel band-
width of 200 kHz using Nyquist sampling. The mobile velocity
is v = 30 m/s and the system delay considered is = 1 ms.
With those parameters, the Doppler frequency is f
d
= 200 Hz
and the normalized delay f
d
= 0.2. We require the system to
2
As in [2] and [20], G is equal to 2 for our example codes in Section VI.
3
For this purpose we have used the function fminbnd in MATLAB.
5 10 15 20 25 30 35
0
20
40
60
80
100
120
140
160
180
200
Expected subchannel CSNR [dB]
L
Optimal power
Equal power
n
R
= 1
n
R
= 2
n
R
= 4
Fig. 3. Optimum pilot period L when the power is equally and optimally
allocated between pilot and data symbols. The number of antennas to combine
is n
R
= 1,2, and 4.
5 10 15 20 25 30 35
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Expected subchannel CSNR [dB]
F
r
a
c
t
i
o
n

o
f

p
o
w
e
r

a
l
l
o
c
a
t
i
o
n

t
o

p
i
l
o
t

s
y
m
b
o
l

(
1

)
Equal power allocation
Optimal power allocation
n
R
= 1
n
R
= 2
n
R
= 4
Fig. 4. Fraction of power allocation to pilot symbols (i.e., 1 ) when the
pilot period L is optimal. The number of receive antennas is n
R
= 1,2, and
4.
tolerate an BER
0
= 10
5
, and as in [1] we choose the order
of the estimator and predictor to be K
e
= 20 and K
p
= 250,
respectively. That choice of K
p
leads to a suboptimal but
satisfactory predictor [27]. Furthermore, we assume that the
expected subchannel CSNR is the same for all the branches
i.e., =
1
=
2
= =
nR
.
Figs. 3 and 4 respectively shows the optimum pilot symbol
period
4
and the optimum fraction of power allocation to the
pilot symbolsfor different numbers of branches to combine.
As can be seen from Fig. 3, when the power distribution
between pilot and data symbols is optimized fewer pilot
symbol is needed. The spacing between two pilot symbols is
also larger when there is more antennas available to combine.
4
Note that since = DLTs = 1 ms, and both D and L must be positive
integers, the value of L must obey L = 200/D. Thus, L has to be in the set
L {2, 4, 5, 8, 10, 20, 25, 40, 50, 100, 200}.
6
5 10 15 20 25 30 35
0
1
2
3
4
5
6
7
8
9
n
R
=1
n
R
=2
n
R
=4
Capacity
Expected subchannel CSNR [dB]
A
v
e
r
a
g
e

s
p
e
c
t
r
a
l

e
f
f
i
c
i
e
n
c
y

[
b
i
t
s
/
s
/
H
z
]
Optimal power, optimal L
Optimal power, L = 10
Equal power, optimal L
Capacity
Fig. 5. Average spectral efciency for both optimal parameters and some
xed parameters together with the channel capacity for Rayleigh fading. n
R
is the number of receive antennas and is set to 1, 2, and 4.
5 10 15 20 25 30 35
10
8
10
7
10
6
10
5
Expected subchannel CSNR [dB]
A
v
e
r
a
g
e

B
E
R
n
R
= 1
n
R
= 2
n
R
= 4
Fig. 6. Average BER for optimal power and optimal L with different number
of receive antennas (n
R
= 1, 2, and 4).
The power allocated to pilot symbols is lower for higher
diversity orders for both the equal power and optimal power
allocation cases (cf. Fig. 4). This is due to the array gain and
to the fact that MRC is optimal in the sense that it maximizes
the output CSNR, thus less pilot power is needed when more
antenna is available. To correspond to the larger pilot period
the optimal power allocation scheme puts more power on the
pilots than when the power is equally allocated.
On the other hand, the pilot power increases again at very
high CSNR for the optimal case. This corresponds to the
steeply increasing pilot period in the same CSNR region.
In order to have a good channel prediction and estimation
which the system can rely onso that the BER requirement
is maintainedmore power should be allocated to the pilots.
Because of the nite number of codes, the ASE reaches a
ceiling when the CSNR grows large. As expected, the ASE is
higher when the transmit power and pilot period are optimized,
and when we have more antennas available to combine. This
is due to the array gain as well as to the fact that the pilot
period is larger for higher n
R
. This is shown clearly in Fig. 5.
As can be seen in Fig. 6, the theoretical average BER
curvesproduced from Eq. (15)are always lower than
BER
0
= 10
5
. This is due to the requirement that the
instantaneous BER is less than or equal to BER
0
.
We observe that our optimum ASE is approximately 0.7
bits/s/Hz higher than for uncoded M-QAM for n
R
= 1cf.
Fig. 5 and [1, Fig. 5] at 20 dB. This corresponds to about 3
dB gain in average CSNR due to the coding.
With the same delay and system parameters, the system in
[2] will not be able to operate acceptably with 1, or even
with 2 receive antennas. Combining 4 antennas, the system
starts functioning at average subchannel CSNR = 21 dB, for
L = 10. To conclude, our system operates satisfactory and
superior to the system in [2] for the whole range of CNSRs
under consideration and for the considered delay. This holds
especially at the high CSNR region where xing of pilot
spacing is clearly a disadvantage since the channel can also
be satisfactorily tracked with larger pilot period.
VII. CONCLUSIONS
We have investigated an ACM system where receive diver-
sity is implemented by means of MRC. The ASE is a result
of a joint optimization of the pilot symbol period, and the
power distribution between pilot and data symbols. The ASE
is considerably increased compared to when the pilot spacing
is xed, and when the power is equally allocated to pilot and
data symbols. The increased ASE is obtained without losing
BER performance; especially in the high CSNR regions. This
gain is due to the fact that the rate of the pilot symbols is
substantially reduced in that region, thus, xing of the pilot
symbol rate is clearly a disadvantage.
The optimal power allocation strategy is the one which
puts more power on pilot symbols; as the response to the
decreased pilot symbol rate. The pilot spacing increases, and
the pilot power decreases for higher diversity orders and for
both optimal and non-optimal power distribution scheme.
ACKNOWLEDGMENT
The authors would like to thank the anonymous reviewers
for their careful reading, their comments, and feedback on the
manuscript, which signicantly improved the quality of the
paper.
REFERENCES
[1] X. Cai and G. B. Giannakis, Adaptive PSAM accounting for channel
estimation and prediction errors, IEEE Transactions on Wireless Com-
munications, vol. 4, no. 1, pp. 246256, Jan 2005.
[2] G. E. ien, H. Holm, and K. J. Hole, Impact of channel prediction
on adaptive coded modulation performance in Rayleigh fading, IEEE
Transactions on Vehicular Technology, vol. 53, no. 3, pp. 758769, May
2004.
[3] A. J. Goldsmith and S.-G. Chua, Variable-rate variable-power MQAM
for fading channels, IEEE Transactions on Communications, vol. 45,
no. 10, pp. 12181230, Oct 1997.
7
[4] J. K. Cavers, An analysis of pilot symbol assisted modulation for
Rayleigh fading channels, IEEE Transactions on Vehicular Technology,
vol. 40, no. 4, pp. 686693, Nov 1991.
[5] D. V. Duong and G. E. ien, Adaptive trellis-coded modulation with
imperfect channel state information at the receiver and transmitter, in
Proc. Nordic Radio Symposium, Oulu, Finland, Aug 2004.
[6] S. Zhou and G. B. Giannakis, Adaptive modulation for multi-antenna
transmissions with channel mean feedback, IEEE Transactions on
Wireless Communications, vol. 3, no. 5, pp. 16261636, Sept 2004.
[7] Y. Ko and C. Tepedelenlioglu, Space-time block coded rate-adaptive
modulation with uncertain SNR feedback, in Asilomar Conf. On
Systems and Computers, 2003.
[8] , Optimal switching thresholds for space-time block coded rate-
adaptive M-QAM, in ICASSP, Montreal, Canada, May 2004.
[9] T. S. Rappaport, Wireless Communications Principles & Practice. Pren-
tice Hall, 1999.
[10] J. H. Winters, Smart antennas for wireless systems, IEEE Personal
Communications, vol. 5, pp. 2327, Feb 1998.
[11] B. Holter and G. E. ien, Impact of spartial correlation on adaptive
coded modulation performance in Rayleigh fading, submitted to IEEE
Trans. on Vehicular Techn., Feb 2005.
[12] D. V. Duong, M.

A. Wingar, and G. E. ien, Link adaptation in
correlated antenna diversity environments, in Proc. Norwegian Signal
Processing Symposium (NORSIG), Stavanger, Norway, Sept 2005.
[13] D. V. Duong, B. Holter, and G. E. ien, Optimal pilot spacing and
power in rate-adaptive MIMO diversity systems with imperfect trans-
mitter CSI, in Proc. IEEE Workshop on Signal Processing Advances in
Wireless Communications (SPAWC), New York City, USA, June 2005.
[14] S. Zhou and G. B. Giannakis, How accurate channel prediction needs
to be for transmit-beamforming with adaptive modulation over Rayleigh
MIMO channels? IEEE Transactions on Wireless Communications,
vol. 3, no. 4, pp. 12851294, July 2004.
[15] X. Tang, M.-S. Alouini, and A. J. Goldsmith, Effect of channel
estimation error on M-QAM BER performance in Rayleigh fading,
IEEE Transactions on Communications, vol. 47, no. 12, pp. 18561864,
Dec 1999.
[16] H. Meyr, M. Moeneclaey, and S. A. Fechtel, Digital Communication
Receivers: Synchronization, Channel Estimation and Signal Processing.
John Wiley & Sons, 1998.
[17] A. Papoulis and S. U. Pillai, Probability, Random Variables and Stochas-
tic Processes, 4th ed. McGraw-Hill Higher Education, 2002.
[18] C. W. Therrien, Discrete Random Signals and Statistical Signal Pro-
cessing. Signal Processing Series. Englewood Cliffs, NJ: Prentice Hall,
1992.
[19] A. J. Goldsmith and S.-G. Chua, Adaptive coded modulation for fading
channels, IEEE Transactions on Communications, vol. 46, no. 5, pp.
595602, May 1998.
[20] K. J. Hole, H. Holm, and G. E. ien, Adaptive multi-dimensional coded
modulation over at fading channels, IEEE Journal on Selected Areas
in Communications, vol. 18, no. 7, pp. 11531158, July 2000.
[21] S. Falahati, A. Svensson, M. Sternad, and H. Mei, Adaptive trellis-
coded modulation over predicted at fading channels, in Proc. IEEE
Vehicular Technology Conference (VTC), Orlando, Florida, Oct 2003.
[22] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series and
Products, 6th ed. Academic Press, 2000.
[23] D. G. Brennan, Linear diversity combining techniques, Proceedings
of the IRE, vol. 47, pp. 10751102, June 1959.
[24] N. M. Temme, Special Functions: An Introduction to the Classical
Functions of Mathematical Physics. New York: John Wiley & Sons
Inc., 1996.
[25] M.-S. Alouini and A. J. Goldsmith, Adaptive modulation over Nak-
agami fading channels, Kluwer J. Wireless Communications, vol. 13,
pp. 119143, May 2000.
[26] S. T. Chung and A. J. Goldsmith, Degrees of freedom in adaptive
modulation: a unied view, IEEE Transactions on Communications,
vol. 49, no. 9, pp. 15611571, Sept 2001.
[27] G. E. ien, R. K. Hansen, D. V. Duong, H. Holm, and K. J. Hole, Bit
error rate analysis of adaptive coded modulation with mismatched and
complexity-limited channel prediction, in Proc. IEEE Nordic Signal
Processing Symposium (NORSIG), Hurtigruten, Norway, Oct 2002.
Duc V. Duong (S03) was born in Qui Nhon,
Vietnam, in 1977. He received the M.Sc. degree
in electrical engineering from the Department of
Telecommunications, Norwegian University of Sci-
ence and Technology (NTNU), Trondheim, Norway,
in 2002.
Currently, he is working towards the Ph.D. degree
at Department of Electronics and Telecommunica-
tions, NTNU. His research interests include the area
of wireless communciations; especially analysis of
fading channel, channel estimation, channel predic-
tion with application to link adaptation.
Geir E. ien (S90M01) was born in Trondheim,
Norway, in 1965. He received the M.Sc. and the
Ph.D. degrees, both from the Norwegian Institute
of Technology (NTH), Trondheim, Norway, in 1989
and 1993, respectively.
From 1994 to 1996, he was an Associate Professor
with Stavanger University College, Stavanger, Nor-
way. In 1996, he joined The Norwegian University
of Science and Technology, Trondheim, as Associate
Professor and in 2001 he was promoted to Full
Professor. His current research interests include the
analysis and design of bandwidth-efcient channel coding and modulation
schemes for fading channels; wireless channel analysis, estimation, and
prediction; analysis and characterization of spatial diversity and multiple-
input multiple-output (MIMO) systems; and orthogonal frequency-division
multiplexing (OFDM)-based wireless systems.
Kjell J. Hole (S89M90) was born in Molde,
Norway, on June 1, 1961. He received the B.Sc.,
M.Sc., and Ph.D. degrees in computer science from
the University of Bergen, Bergen, Norway, in 1984,
1987, and 1991, respectively.
From August 1988 to May 1990, he was a Visiting
Scholar with the Center for Magnetic Recording
Research, the University of California, San Diego.
In 1993, he was with the IBM Almaden Research
Center, San Jose, CA. From 1995 to 2005 he was a
Research Scientist at the University of Bergen and
at the Norwegian University of Science and Technology (NTNU), Trondheim,
with funding from the Norwegian Research Council. In 2005 he became a
Full Professor at University of Bergen. His current research interests are in
the areas of wireless communication and data security.