SUCCESSIVE INTERFERENCE

CANCELLATION SCHEMES IN
MULTI-RATE DS/CDMA SYSTEMS
*
Ann-Louise Johansson and Arne Svensson
Chalmers University of Technology
Department of Information Theory
S-412 96 Gothenburg, Sweden
ABSTRACT
In this paper we propose successive interference cancellers (IC) for a variable data
rate environment. Firstly, a successive interference cancellation (IC) scheme for M-
level rectangular QAM in DS/CDMA systems is analysed. Secondly, the perform-
ance under Rayleigh fading is analysed for single modulation systems and different
systems for handling multiple data rates. One alternative for handling multi-rate sys-
tems is to use mixed modulation, which imply that each user chooses modulation for-
mat according to need. Another alternative is to use parallel channels, where each
user transmits over one or several synchronous channels. We show that the succes-
sive IC, for mixed modulation or parallel channel systems, has a performance close
to the single BPSK user bound and, consequently, it gives a considerable increase in
performance and exibility compared to both single modulation systems and mixed
modulation or parallel channel systems employing a conventional detector.
1 INTRODUCTION
In the future users will demand mobile telephone systems to be able to handle many
more services than speech, like e.g. facsimile, Hi-Fi audio and computer data, which
is not possible today. To achieve this we need to use a multiple-access method which
is exible and has the prospect of capacity increases and being able to handle varia-
ble data rates. Recently Code Division Multiple Access (CDMA) has been suggested
to be a multiple-access method able to stand these requirements. The rst company to
propose this technique for a practical system is Qualcomm Inc in the USA, who are
working on a CDMA digital cellular radio communication system.
A Direct-Sequence CDMA (DS/CDMA) system has several unique features. Some
of them are spectrum sharing, rejection of multipath signal components or possibility
to utilize them for recombining [1] and frequency reuse factor of one in a cellular
scenario [2]. These features are highly desirable, though a CDMA system employing
a conventional, matched lter, detector is interference limited, which directly deter-
mines the system capacity. The conventional detector is optimal in a single-user
channel corrupted only by additive white Gaussian noise (AWGN). The presence of a
number of users in the system introduces multiple-access interference (MAI), since
the signature sequences used are not perfectly orthogonal, which leads to an irreduci-
* This work was supported by the Swedish National Board for Industrial and Technical Development
(NUTEK), project number 9303363-2.
ble error probability. In a mobile radio scenario the transmitters move in relation to
the receiver and the energies of the received signals are to be neither equal nor con-
stant. In this situation the conventional detector fails to demodulate weak users, even
when the cross-correlation between the signals is relatively low. This is known as the
near/far problem. One way to combat this problem is to use stringent power control
[2]. Another approach is to use more sophisticated receivers which are near/far resist-
ant. Recently a lot of attention has been directed to the area of multi-user detectors,
which has the prospect of mitigating the near/far problem and by cancelling the MAI
also increasing the total system capacity.
The research in this area was trigged by Verdu [3], who worked on the optimal multi-
user detector. This detector is very complex and that has initiated further research in
the area of sub-optimal, lower complexity, detectors both using parallel detection [4]-
[8] and serial or successive interference cancellation (IC) [9]-[13].
The motivation for our work is to evaluate an efcient detector for a multi-user and
multi-rate DS/CDMA system. A method to handle multiple data rates is to let the
users use different forms of modulation [14]. A user could, depending on the need,
choose between using e.g. BPSK, QPSK or 16-QAM modulation. Therefore, an IC
scheme for M-ary QAM is analysed in this paper, based on the IC scheme for coher-
ent BPSK modulation derived by Patel and Holtzman [9]-[11]. The analysis is then
extended to cover systems where the users employ different QAM formats. Another
approach to handle multiple data rates is to let each user transmit over one or several
parallel channels according to requirements [14]. This can of course also be used in
combination with different modulation formats.
The operation of the IC scheme is made in the following manner: The signal ampli-
tude of each user is estimated from the output of a linear correlator, where we corre-
late the received signal with each users signature sequence. It should be noted that
this will not give a perfect estimate though the scheme will be simple. We then rank
the users in decreasing order of their received powers, which are obtained using these
correlator outputs. The users signals are decoded and cancelled from the received
signal successively starting with the strongest user. In an uplink, where we are inter-
ested in all received signals, we will use a scheme where all the users are decoded
and cancelled. On the other hand, in a downlink we would perform the successive IC
scheme on all the interfering users before decoding the desired signal.
Due to limited space we will only give the main performance results in this paper. All
derivations may be found in [15] for the interested reader. We will consider ideal
coherent demodulation and frequency-nonselective fading. The performance meas-
ure used is average bit error probability. We assume perfect knowledge of the phase
and the time delay. We also assume perfect power ranking in the performance analy-
sis, which imply knowledge of the channel gain.
2 SYSTEM MODEL AND DECODER STRUCTURE
We consider a system model for square lattice QAM, where the received signal, for a
K user system, is
(1)
r t ( )
k
2E
0
T
---------d
k
I
t
k
( ) c
k
I
t
k
( )
c
t
k
+ ( ) cos +
k 1 =
K

k
2E
0
T
---------d
k
Q
t
k
( ) c
k
Q
t
k
( )
c
t
k
+ ( ) sin n t ( ) +
which is the sum of all the transmitted signals embedded in AWGN. is a
sequence of amplitude, rectangular pulses of duration T. T is the inverse of the
symbol rate, assumed to be equal for all users. is the information-bearing sig-
nal amplitude of the quadrature carriers for the users symbol element, which
together generate M equiprobable and independent symbols. is the energy of the
signal with lowest amplitude. is the users signature sequence, which
is used for spreading the signal in the in-phase (I) or the quadrature (Q) branch. It
consists of a sequence of, antipodal, unit amplitude, rectangular pulses of duration
. The period of all the users signature sequences are , so there is one
period per data symbol. We assume without loss of generality that the signature
sequences have power equal to one. is the time delay and is the phase of the
user. In the asynchronous, though symbol-synchronous, case they are i.i.d. uni-
form random variables over and , respectively. Both parameters
are assumed to be known in the analysis. represents the common centre fre-
quency, represents the channel gain and is the AWGN with two-sided
power spectral density of .
The receiver is the corresponding matched lter detector for M-ary QAM, from
which we obtain two decision variables, . is our sufcient statis-
tics for the I and Q components and for symbol zero of the rst user we get the fol-
lowing expression [15]
(2)
where is a noise term of the decision variable including both Gaussian noise
and noise caused by interference. For convenience and reasons explained later we
will omit the second subscript of the noise term. It can be shown [15] with the help of
trigonometric functions that is given by
(3)
where the components are uncorrelated Gaussian random variables and the sums
of the and components are interference due to the resulting K-1 users. The
noise component caused by the user in the I branch is dened by [15]
(4)
where is a unit amplitude, rectangular pulse of length T, ,
,and the delay is assumed to be shorter than .The reason for this
is explained later. is given by a similar expression [15].
d
k
I Q /
t ( )
A
k l ,
I Q /
A
k l ,
I Q /
k
th
l
th
2E
0
c
k
I Q /
t ( ) k
th
T
c
N T T
c
=

k

k
k
th
0 T ) , [ 0 2 ) , [

k
n t ( )
N
0
2
S
1 0 ,
I
and S
1 0 ,
Q
S
1 0 ,
I Q /
S
1 0 ,
I Q /
E
0
2T
------
1
A
1 0 ,
I Q /
N
1
I Q /
+ =
N
1
I Q /
N
1
I Q /
N
1
I
E
0
2T
------ I
k 1 ,
II
A
k
I Q /

k 1 ,

k 1 ,
, ,
,
_
1
2
--- n
1
II
n
1
QI
+
k 2 =
K

=
N
1
Q
E
0
2T
------ I
k 1 ,
QQ
A
k
I Q /

k 1 ,

k 1 ,
, ,
,
_
1
2
--- n
1
QQ
n
1
IQ
+
k 2 =
K

'

n
1
I
k 1 ,
II
I
k 1 ,
QQ
k
th
I
k 1 ,
II
A
k
I Q /

k 1 ,

k 1 ,
, , ( )

k
T
------ A
k l ,
I
p
T
l 0 =
1

t
k 1 ,
lT ( ) c
k
I
t
k 1 ,
( ) c
1
I
t ( ) t d
0
T

,

_

k 1 ,
cos + =

k
T
------ A
k l ,
Q
p
T
l 0 =
1

t
k 1 ,
lT ( ) c
k
Q
t
k 1 ,
( ) c
1
I
t ( ) t d
0
T

,

_

k 1 ,
sin
p
T
t ( )
k i ,

k

i
=

k i ,

k

i
=
k

1
I
k 1 ,
QQ
3 SUCCESSIVE INTERFERENCE CANCELLATION
SCHEME WITH M-ARY QAM
A receiver with IC is shown in Fig. 1. Each detector is a matched lter for M-ary
QAM and the outputs are the decision variables, which are used both for deciding
which user is the strongest and in the cancellation of that users signal. The strongest
user, ideally the one with the largest , is decoded rst and cancelled at baseband
from the received signal. The detector is a coherent detector and we assume optimum
decision boundaries. Subsequently all the users are decoded and cancelled in
decreasing order of their powers. Let us assume we have the means of deciding
which user is the strongest and therefore also the one most likely to be decoded cor-
rectly. A suggestion of such a scheme is given in [11], [15].
Without loss of generality we consider the decision of at time corre-
sponding to symbol element 0. All the symbol elements prior to the zeroth have
already been decoded. We assume also without loss of generality that user 1 has the
strongest signal and, accordingly, it is decoded rst and cancelled from the composite
signal. Assume that we have a situation where the strongest user has a time delay
which is shorter than any other users time delay. Then it would not be enough to can-
cel only symbol element 0 of user 1 from the composite signal to reduce the noise
caused by interference since the decision variable of user 2s zeroth symbol will also
include interference from user 1s rst symbol. To overcome this we decode and can-
cel user 1s rst symbol element too before we continue the IC scheme for the second
strongest user. Therefore, without loss of generality, we assume in the analysis
. In this symbol-synchronous case we decode symbol 0 and since
we assume , the interference from the user in originate
from symbol element 0 in and symbol element 1 in .
We use the decision variables and to estimate user 1s baseband signal for
symbol element 0 and cancel it from the composite signal. This cancellation is not
perfect since the signature sequences are not orthogonal and we do not know the
channel gain of user 1. Furthermore the correlator output contains Gaussian noise.
Considering the general case, the resulting baseband signal for the I channel after the
cancellation will be as follows [15]
Fig. 1. M-ary QAM receiver with Interference Cancellation

k
A
k 0 ,
I Q /

k
T +

1

2

K
> > >

k

1
< k
th

1
[
1
T + ) ,

1
[
k
T + ) ,
k
T +
1
T ) + , [
S
1 0 ,
I
S
1 0 ,
Q
h
th
LPF
LPF

c
t cos

c
t sin
Detector 1
Detector 2
Detector K
Select
Max
&
Decode
+
+
+
+
c
1
I
t
1
( )
1
cos
c
1
Q
t
1
( )
1
cos
c
1
Q
t
1
( )
1
sin
c
1
I
t
1
( )
1
sin ( )

+
+

I
t ( )

Q
t ( )
S
1 0 ,
I
S
1 0 ,
Q
S
K 0 ,
I
S
K 0 ,
Q
S
1 0 ,
I
S
1 0 ,
I
S
1 0 ,
Q
S
1 0 ,
Q
r t ( )
(5)
When h is 1 the term corresponds to the remaining baseband signal after
cancellation of all symbol elements prior to the zeroth, and consequently we will get
after cancelling user 1s zeroth symbol. Let user 2 be the second strongest
user and also the user with the second longest time delay. Now there are only K-2
interfering users left since the strongest user has been cancelled from the composite
signal. We proceed in the same manner for the second strongest user as well as for the
subsequent users. The decision variable before the cancellation is given by [15]
(6)
We have now h-1 cancelled and K-h+1 remaining symbol elements 0. Therefore, the
total noise component for the user in the I channel is
(7)
where the rst sum consists of noise caused by the remaining interfering users, the
second term is white Gaussian noise and the last sum is the resulting noise caused by
imperfect cancellations. The expression for will be similar to with some
changes in indices. The correlation terms and , in (7) above, are dened as
(8)
In these expressions the correlation of is with noise caused by imperfect can-
cellation of symbol element -1 and 0 of the rst user, since . See Fig. 2, where
shaded lines indicate cancelled symbol elements. We are also considering a fre-
quency-nonselective, slowly fading channel which implicates that the interference
power can be regarded as equal for two consecutive symbol elements. Therefore we
do not distinguish between and in (7).
Fig. 2. Cross-correlation between users

h 0 ,
I
t ( )
h 1 0 ,
I
t ( ) S
h 0 ,
I
p
T
t
h
( ) c
h
I
t
h
( )
h
cos =
S
h 0 ,
Q
p
T
t
h
( ) c
h
Q
t
h
( )
h
sin

0 0 ,
I
t ( )

1 0 ,
I
t ( )
h
th
S
h 0 ,
I Q /
E
0
2T
------
h
A
h 0 ,
I Q /
N
h
I Q /
+ =
h
th
N
h
I
E
0
2T
------ I
k h ,
II
A
k
I
A
k
Q

k h ,

k h ,
, , ,
,
_
k h 1 + =
K

1
2
--- n
h
II
n
h
QI
+ =
N
j
I
J
j h ,
II

j h ,

j h ,
, ( )
j 1 =
h 1

N
j
Q
J
j h ,
QI

j h ,

j h ,
, ( ) +
N
h
Q
N
h
I
J
1 2 ,
II
J
1 2 ,
QI
J
1 2 ,
II

1 2 ,

1 2 ,
, ( )
1
T
--- c
1
I
t
1 2 ,
( ) c
2
I
t ( )
1 2 ,
( ) cos dt
0
T

=
J
1 2 ,
QI

1 2 ,

1 2 ,
, ( )
1
T
--- c
1
Q
t
1 2 ,
( ) c
2
I
t ( )
1 2 ,
( ) sin dt
0
T

=
c
2
I
t ( )

2

1
<
N
1 1 ,
I
N
1 0 ,
I
-T 0 T 2T

1
User 1
User 2
-1 0 1

1
T +

2
T +
2

2
T

1
T
4 PERFORMANCE ANALYSIS ON A STATIONARY
CHANNEL
Evaluating the performance of the IC scheme will be done by applying Gaussian
approximation [9]. The noise components, caused by interference from the other
users in the system are therefore modelled as independent Gaussian noise. We have
chosen to use the Gaussian approximation partly since it is commonly used and
partly because we have not found any other practical way to evaluate the perform-
ance. By using a Gaussian approximation an increase in noise and interference vari-
ance immediately leads to an increase in error probability. It is likely that this will
occur also for the true distribution and we believe therefore that relative performance
will not change signicantly. Absolute performance is likely to be too optimistic
though [16]. Additionally, since we do not consider any fading in this section, the
channel gain, , will be constant. We also assume that power control is used for dis-
tance and shadow fading and consequently, in this case, all the s will be equal.
4.1 Performance Analysis of QAM IC Scheme
We start with calculating the variance of the I and Q channel decision variables con-
ditioned on , i.e.
(9)
It is easy to show that all the random variables in are independent and with
zero mean. Consequently we can model as independent Gaussian random vari-
ables with zero mean and variance . When using the Gaussian approximation it
is easy to obtain the probability of error from the theory of single transmission of
QAM signals over a AWGN channel [1]. The symbol error rate (SER) with ideal
coherent detection can then be expressed as ( )( ) where
is the probability of error for the transmission over the I channel, respectively
the Q channel, given by [1]
(10)
where the Q-function is the complementary Gaussian error function. in (10) is
a signal-to-noise ratio for the I and Q channel dened as
(11)
for the user. The expression for the variance in (11) depends on which kind of
sequences we use. For deterministic sequences the variance in (9) yields
(12)
where will have a similar expression with some small changes in indices. The
different terms in (12) are average interference dened in [17]. For random
sequences we obtain a similar expression using the equalities and

h
I Q /
Var N
h
I Q /

h
=
N
h
I Q /
N
h
I Q /

h
I Q /
Pe
h
1 = 1 Pe
h
I
1 Pe
h
Q

Pe
h
I Q /
Pe
h
I Q /
2
M 1
M
------------------
,
_
Q
h
I Q /
,
_
=

h
I Q /

h
I Q /
E
0
2T
------
h

h
I Q /
----------------- =
h
th

h
I
N
0
4T
------
M 1
3
--------------
E
0
12N
3
T
----------------
,
_

k
2
r
k h ,
II
r
k h ,
QI
+ ( )
k h 1 + =
K

1
6N
3
---------
j
I
r
j h ,
II

j
Q
r
j h ,
QI
+
j 1 =
h 1

+ + =

h
Q
r
k h ,

j
I

j
Q
=
[17].
The ranking of the users is completely random since all the s are equal. Therefore
the order of cancellation will change continuously and we obtain the average proba-
bility of symbol error for each user if we take the average of all SER calculated for
each stage of cancellation. To compare the results obtained for QAM users with
BPSK users we need to calculate the bit error rate (BER). We assume a Gray-
encoded version of M-ary QAM and estimate the BER by dividing the SER with
, which is an approximation taking into account only the most likely errors.
5 PERFORMANCE ANALYSIS OF QAM IC SCHEME
UNDER FADING
In the case of fading, we consider a system where the K users are received through
independent, frequency-nonselective slowly fading channels. This model is suitable
in areas with small delay spread and for mobiles with slow speed (small Doppler fre-
quency). These conditions also make estimation of and (needed for decision
boundaries for M-ary QAM and power ranking) feasible. No instantaneous power
control is applied, only average power control which takes care of shadowing and
distance attenuation.
5.1 IC Scheme for QAM under Single-Path Rayleigh
Fading
In this paper we will use the same method, for analysing the single-path Rayleigh
fading channel, as was used in [11] for BPSK modulation. The equations for the
noise variance and error probabilities obtained in the previous section were all condi-
tioned on .
The unconditioned error probability of the I channel for each stage of cancellation is
easily obtained from the conditioned probability of error in (10) as follows
(13)
where is the distributions of the ordered amplitudes. The disordered ampli-
tudes are assumed to be Rayleigh distributed with unit mean square value. That is,
the average received power from all the users is equal, which is a consequence of
assuming perfect power control for shadowing and long term fading. Therefore the
distributions of the ordered , , are obtained by using order statistics [19].
We dene the signal-to noise ratio, , for the user in the I channel in the same
manner as in (11). The only difference is that is replaced by , which is
the expected value of the conditioned variance with respect to . The variance is
approximated as a sum of independent Gaussian random variables with zero mean
given by
(14)
for random sequences. We then use and calculate the unconditioned
SER in the same manner as the conditioned SER. Taking the average of all SERs at
all stages of cancellation yields the average probability of symbol error. This is a
E r
k h ,
[ ] 2N
2
=

k
log
2
M

k

k

k
P

e
h
I
Pe
h
I
f

h
x ( ) dx
0

=
f

h
x ( )

h
f

h
x ( )

h
I
h
th

h
I
E

k

h
I

k
E

h
I
N
0
4T
------
M 1
3
--------------
E
0
3NT
-----------
,
_
E

k
2
k h 1 + =
K

2
3N
------- E

j
I
j 1 =
h 1

+ + =
P

e
h
I
P

e
h
Q
=
proper performance measure, since the order of cancellation will change with fading
and the average of all users will be the same as the time average for each user.
5.2 Numerical Examples
The average BER of different single user systems, 40 BPSK, 20 QPSK and 10 16-
QAM users, with the same throughput under Rayleigh fading is shown in Fig. 3. The
length of the random sequences is 127. The result is compared to the single user
bound under fading for both BPSK and 16-QAM users and corresponding systems
employing a conventional detector. Overall, the performance for all the systems
using IC is considerably better compared to the systems employing the conventional
detector. The results also show that the QPSK and BPSK systems (partly overlap-
ping) perform almost equally well as in the case of transmission of a single BPSK
user and that the performance of the 16-QAM system is inferior but it is close to the
single 16-QAM user bound.
6 PERFORMANCE ANALYSIS OF MIXED
MODULATION SYSTEMS
One way to handle multi-rate systems is to let each user choose a modulation format
in correspondence with required transmitted data rate [14]. In the sequel we will
evaluate the performance of a system where the users employ different forms of
modulation, e.g., a combination of BPSK, QPSK and 16-QAM users.
6.1 DS/CDMA Mixed System Model Description
We consider a system where we have BPSK, QPSK and 16-QAM users.
To make a comparison between different forms of modulation we let the transmitted
Fig. 3. Performance of single modulation systems with and without IC under
Rayleigh fading
K
1
K
2
K
3
E
b
/N
0
(dB)
0 5 10 15 20 25
10
3
10
2
10
1
10
0
A
v
e
r
a
g
e

B
i
t

E
r
r
o
r

P
r
o
b
a
b
i
l
i
t
y
Processing Gain = 127
Random Sequences
IC
Conv
BPSK (40)
QPSK (20)
16QAM (10)
Single BPSK
Single 16QAM
bit energy, , be equal for all users independent of the modulation format used. We
then rewrite the energy as a function of , in the following way
(15)
which is valid for M-ary QAM. For BPSK . If we dene
(QPSK) and (16-QAM) we can express the signal-to-noise ratio, ,
conditioned on , for the QPSK user as follows
(16)
where i, h-1 and m species the number of cancelled BPSK, QPSK and 16-QAM
users, respectively. is the expression for the variance of the BPSK users [9] and
and is the variance for QPSK, respectively 16-QAM, users dened in Sec-
tion 4.1. In this case all users employ random sequences. Rewriting (16) we get
(17)
where we nd the noise caused by interference from QPSK users on the rst row,
from BPSK users on the second row and from 16-QAM users on the last row. The
expression for BPSK users is easily obtained in the same manner [15] and for 16-
QAM users the expression will be similar to the one in (17) with some changes in
indices.
6.2 Performance Analysis on a Stationary Channel
Again we use the Gaussian approximation for the noise term of each type of modula-
tion scheme. The total BER for the whole system is given by
(18)
where , and are the data rates for the BPSK, QPSK and 16-QAM users
respectively and the terms are the individual BERs.
E
b
E
0
E
b
E
0
E
b
3 M log
2 M 1 ( )
------------------------ =
E
0
E
b
= M
1
4 =
M
2
16 =
h
I

k
h
th

h
I
E
b

h
2
2T
-------------
3 M
1
log
2 M
1
1 ( )
--------------------------

i

h
I

m
I
+ +
------------------------------------------------- =

h
I

m
I

h
I

h
M
1
1
3
----------------
N
0
E
b
M
1
log
2
------------------------
2
3N
-------
k
2
k h 1 + =
K
2

+
,

_
2
3N
-------
j
I
( )
2
j 1 =
h 1

+ + =
2 M
1
1 ( )
3 M
1
log
2
--------------------------
1
3N
-------
k
2
j i 1 + =
K
1

1
3N
-------
2 M
1
1 ( )
3 M
1
log
2
--------------------------
j
2
j 1 =
i

+ +
2 M
1
1 ( ) M
2
log
2
3 M
1
log
2
---------------------------------------------
1
3N
-------
k
2
k m 1 + =
K
3

M
1
1 ( ) M
2
log
2
M
2
1 ( ) M
1
log
2
-----------------------------------------
2
3N
-------
j
I
( )
2
j 1 =
m

+
1 2 /
Pb
tot
Pb
i
BPSK
R
1
i

Pb
h
QPSK
R
2
h

Pb
m
QAM
m

R
3
+ +
K
1
R
1
K
2
R
2
K
3
R
3
+ +
----------------------------------------------------------------------------------------------------------------- =
R
1
R
2
R
3
Pb
6.3 Mixed Modulation Systems under Rayleigh Fading
In the case of Rayleigh fading we follow the same procedure using order statistics as
in Section 5.1. The difference is that now we have a mixed system and the amplitudes
are ordered independently of modulation. Using (10), (11), modied as in Section
5.1, and (13) together with (17) we can calculate the performance for each QAM user
of the system. Calculating the performance for BPSK users is done in a similar man-
ner [15].
6.4 Numerical Examples
The average performance of a mixed system with 20 BPSK, 10 QPSK and 5 16-
QAM users with and without IC is shown in Fig. 4. The length of the random
sequences is 127. The amplitudes in each mixture have been distributed among the
users independently of modulation format and the result is an average over 100 mix-
tures. The average BER for each kind of users is also shown in the graph together
with the single user BPSK bound. We can also compare the result with single modu-
lation systems with the same throughput (60 BPSK, 30 QPSK and 15 16-QAM) [15].
We would then see that the average performance for the QPSK and BPSK users is
slightly better in the case of 35 mixed users compared to their corresponding single
modulation systems. On the other hand, the average performance of the 16-QAM
users is better for the single modulation system than for mixed system. This can be
explained by the fact that the corresponding number of users in the mixed system is
lower compared to the BPSK and QPSK single modulation systems but higher for the
16-QAM systems. (This is if we count QAM users as two because of the I- and Q-
branch.) Furthermore we have to contemplate the fact that the average is made over a
nite number of mixtures, because of the very time consuming calculations, which
may slightly affect the result.
Fig. 4. Average BER of a mixed modulation system with and without IC under
Rayleigh fading
E
b
/N
0
(dB)
0 5 10 15 20 25
10
3
10
2
10
1
10
0
A
v
e
r
a
g
e

B
i
t

E
r
r
o
r

P
r
o
b
a
b
i
l
i
t
y
Processing Gain = 127
Random Sequences
IC
Mix Conv
BPSK (60)
QPSK (30)
16QAM (15)
Mix (20/10/5)
Single BPSK
7 PERFORMANCE ANALYSIS OF PARALLEL
CHANNEL SYSTEMS
We have previously discussed the possibility to handle multi-rate systems by means
of employing higher modulation formats. Another alternative is to use parallel chan-
nels for transmission of information [14]. We simply let a user send simultaneously
over as many channels as required for a specic data rate. In this kind of system there
are going to be a larger number of interfering signals, though the synchronous signals
corresponding to each user will have considerably lower cross-correlation than asyn-
chronous signals.
7.1 Parallel Channels
We consider rst a system with purely synchronous transmission. We have parallel
channels with exactly the same channel parameters since they belong to the same
user. In other words, the relative time delay and phase between the channels are equal
to zero. We use the expression for the total noise component given in (7) for M-ary
QAM, in the case when and both are zero. Taking the variance we get [15]
(19)
where is the periodic cross-correlation function [17]. In the case of BPSK mod-
ulation the variance of the total noise component will be as follows
(20)
The variance stated in (19) and (20) are for deterministic sequences. We obtain the
equations for random sequences using the fact [16] that [15].
7.2 Combination of Synchronous and Asynchronous
Transmission
We consider now a system where we have K users and where each user, k, transmits
over channels. Thus, we will have a system where the a total number of informa-
tion-bearing channels, , is the sum of all . Accordingly, we have both synchro-
nous and asynchronous interferers and if we use deterministic sequences the major
part of the interference will come from the asynchronous users. Therefore cancella-
tion of parallel signals is excluded and instead we consider a receiver where each
users parallel signals are decoded and cancelled simultaneously. The complexity of
the receiver will not increase considerably since the decision variables for all signals
have to be generated anyway to determine which signal is the strongest. Futhermore
the channel parameters, and , are equal for parallel channels.
Combining (12) and (19) we write the noise variance of the users signal in a
M-ary QAM system with both asynchronous and synchronous interferers as

k h ,

k h ,

h
I
N
0
4T
------
M 1
3
--------------
E
0
2TN
2
-------------
,
_

i
2

i h ,
II
,
_ 2
0 ( )
i h 1 + =

1
N
2
------
j
I

j h ,
II
,
_ 2
0 ( )
j 1 =
h 1

+ + =

i h ,
II

h
N
0
4T
------
E
b
2TN
2
-------------
i
2

i h ,
2
0 ( )
i h 1 + =

1
N
2
------
j

j h ,
2
0 ( )
j 1 =
h 1

+ + =
E
i h ,
2
0 ( ) N =

k

k

h

h
h
th
g
th
(21)
where is the sum of for the h-1 cancelled users and are the remain-
ing interfering asynchronous signals.
7.3 Synchronous and Asynchronous Transmission under
Rayleigh Fading
When we transmit information on parallel channels under Rayleigh fading the paral-
lel channels will be affected by the same channel parameters. That is the delay, the
phase shift and the amplitude will be the same. Since synchronous channels interfere
less with each other we have not considered successive interference cancellation of
parallel channels; instead they are all detected and cancelled simultaneously. The per-
formance of a system with parallel channels is analysed using the same method with
order statistics as described earlier. The difference in this case is that we order the K
user, each one with parallel channels, and then assign the same pdf and
mean square value to all the channels of user k.
7.4 Numerical results
The performance of systems employing parallel channels under Rayleigh fading are
shown in Fig. 5. Systems with K users and 2, 3 or 4 parallel channels per user are
compared with asynchronous systems with , and users. K is in this
Fig. 5. Average BER of parallel channel systems and asynchronous QPSK systems
(employing Gold sequences) with IC under Rayleigh fading

g h ,
I
M 1
3
--------------
E
0
12N
3
T
----------------
,
_

k
2
r
k h ,
II
r
k h ,
QI
+ ( )
k
h
1 + =

1
6N
3
---------
j
I
r
j h ,
II

j
Q
r
j h ,
QI
+
j 1 =

h 1

+ + =
M 1
3
--------------
E
0
2TN
2
-------------
,
_

i
2

i h ,
II
0 ( ) ( )
2
i 1 =
i g

N
0
4T
------ +

h 1

k

h

k
f

k
x ( )
E

k
2

k
2 K 3 K 4 K
E
b
/N
0
(dB)
0 5 10 15 20 25
10
3
10
2
10
1
10
0
A
v
e
r
a
g
e

B
i
t

E
r
r
o
r

P
r
o
b
a
b
i
l
i
t
y
Processing Gain = 127
Gold Sequences
IC
Conv
QPSK (10, P=2,3,4)
QPSK (20/30/40)
Single BPSK
case equal to 10. The performance of the parallel channel systems employing Gold
sequences is very close to the performance of asynchronous systems (the curves are
almost not distinguishable). This is caused by the fact that the interference from the
strongest interferer will be larger in the case of parallel channels since that user will
send over channels instead of one. This will counterbalance the improvement
caused by the good cross-correlation properties of Gold sequences in the synchro-
nous case.
In Fig. 6 we compare the average performance of 15 QPSK users with two parallel
channels each, with the one of 15 16-QAM users. The result show that QPSK users
employing parallel channels perform better than 16-QAM users.
8 DISCUSSION AND CONCLUSIONS
We have demonstrated the use of M-level rectangular QAM with successive IC in
single modulation systems and compared the average BER between systems with the
same throughput. Two different methods, mixed modulation and parallel channels,
for handling multiple data rates have been analysed. These systems have been stud-
ied in the case of Rayleigh fading and they have been compared with single modula-
tion systems.
The conclusion is that the successive IC scheme has relatively low complexity even
in the case of using M-ary QAM, the performance, even for large systems, is close to
the single user bound and, consequently, it yields considerable increase in perform-
ance compared to conventional matched lter detection. Mixed modulation systems
offer more exibility at the cost of a slight decrease in average performance com-
pared to pure asynchronous QPSK systems. The 16-QAM users in the mixed system,
who are most sensitive to noise, will have the highest average BER. On the other
hand, if we instead have a QPSK system where the users employ parallel channels
the average performance is almost equal for all the users. (It would be equal if all the
users in the system use the same number of parallel channels.) Though, to achieve
Fig. 6. Average BER of a parallel channel QPSK system and an asynchronous 16-
QAM system (employing Gold sequences) with IC under Rayleigh fading.

k
E
b
/N
0
(dB)
0 5 10 15 20 25
10
3
10
2
10
1
10
0
A
v
e
r
a
g
e

B
i
t

E
r
r
o
r

P
r
o
b
a
b
i
l
i
t
y
Processing Gain = 127
Gold Sequences
IC
Conv
QPSK (15, P=2)
QPSK (30)
16QAM (15)
Single BPSK
good performance in a parallel channel system it is important to use signature
sequences with good cross-correlation properties to make use of the advantages with
synchronous channels. A minor drawback with parallel channels is that sooner or
later we run out of signature sequences and in some cases it may be better to add a
16-QAM user, if the user can accept a small decrease in performance, instead of a
QPSK user with two parallel channels. Thus, the greatest system exibility is
obtained by a combination of parallel channels and mixed modulation.
Future work within this project will be to analyse mixed systems where the users
have different powers. We have seen from the results that the BER for 16-QAM users
is higher than for BPSK and QPSK users. So by increasing the for the 16-QAM
users we could decrease the BER to meet the performance of the BPSK and QPSK
users. Additionally we will study multi-stage IC schemes, where we decrease the
interference further and increase the capacity by employing a modied version of the
IC scheme in several stages. Multipath fading channels with RAKE combiner will
also be included in the analysis, as will channel coding.
Two different IC schemes, resulted from work independent from ours, have been pro-
posed recently in [20] and [21]. Though their work have only minor similarities to
ours. In both papers they consider solely BPSK systems and Gaussian channels.
Hence, no fading is taken into consideration. In [20] they evaluate the average BER
for a multistage successive IC scheme under the assumption of unequal power con-
trol. Also in [21] they consider a multistage IC scheme, but in this case parallel detec-
tion is employed.
REFERENCES
[1] J. G. Proakis, Digital Communications, 3rd ed, McGraw-Hill, 1995.
[2] K. S. Gilhousen, I. M. Jacobs, R. Padovani, A.J. Viterbi, L.A. Weaver, and C.
E.Wheatley III, On the capacity of a cellular CDMA system, IEEE Trans. on
Vehicular Technology, vol. 40, pp. 303-311, May. 1991.
[3] S. Verdu, Minimum probability of error for asynchronous multiple access
channels, IEEE Trans. on Information Theory, vol. IT-32, pp. 85-96, Jan.
1986.
[4] R. Lupas and S. Verdu, Linear multiuser detectors for synchronous code-divi-
sion multiple access channel, IEEE Trans. on Information Theory, vol. IT-35,
pp. 123-136, Jan. 1989.
[5] R. Lupas and S. Verdu, Near-far resistance of multiuser detectors in asynchro-
nous channels, IEEE Trans. on Communications, vol. COM-38, pp. 497-507,
April 1990.
[6] M. Varanasi and B. Aazhang, Multistage detection in asynchronous code-divi-
sion multiple access communications, IEEE Trans. on Communications, vol.
COM-38, pp. 509-519, April 1990.
[7] Z. Xie, R. T. Short and C. K. Rushforth, A family of suboptimum detectors for
coherent multiuser communications, IEEE Journal on Sel. Areas in Com, vol.
8, pp 683-690, May 1990.
E
b
[8] A. Duel-Hallen, Decorrelating decision-feedback multiuser detector for syn-
chronous code-division multiple access channel, IEEE Trans. on Communica-
tions, vol. COM-41, pp. 285-290. Feb. 1993.
[9] P. Patel and J. Holtzman, Analysis of a DS/CDMA successive interference
cancellation scheme using correlations, Proceedings, Globecom (Houston,
Texas), Dec. 1993.
[10] P. Patel and J. Holtzman, Analysis of successive interference cancellation in
M-ary orthogonal DS-CDMA system with single path rayleigh fading, Pro-
ceedings, International Zurich Seminar (Zurich, Switzerland), March 1994.
[11] P. Patel and J. Holtzman, Analysis of a simple successive interference cancel-
lation scheme in a DS/CDMA, IEEE Journal on Sel. Areas in Com., vol. 12,
pp. 796-807, June 1994.
[12] Magnus Ewerbring, Bjrn Gudmundson, Gustav Larsson and Paul Teder,
CDMA with interference cancellation: A technique for high capacity wireless
systems, Proceedings ICC93 (Geneva, Switzerland), May 1993.
[13] Magnus Ewerbring, Bjrn Gudmundson, Paul Teder and Per Willars, CDMA-
IC: A proposal for future high capacity digital cellular systems, Proceedings
VTC93 (Secaucus, New Jersey), May 1993.
[14] Tony Ottosson, Arne Svensson, Multi-rate schemes in DS/CDMA systems,
To appear in proceedings VTC95 (Chicago, Illinois), July 1995.
[15] Ann-Louise Johansson, Arne Svensson, Analysis of successive interference
cancellation schemes in multiple data rate DS/CDMA systems, Internal
report, Dept. of Information Theory, Chalmers University of Technology, 1995.
[16] E. A. Geraniotis and B. Ghaffari, Performance of binary and quaternary
direct-sequence spread-spectrum multiple-access systems with random signa-
ture sequences, IEEE Trans. on Communications, vol. COM-39, pp.713-724,
May 1991.
[17] M. B. Pursley, Performance evaluation for phase-coded Spread-Spectrum
multiple-access communication - Part I: System analysis, IEEE Trans. on
Communications, vol. COM-25, pp. 795-799, Aug. 1977.
[18] S. Haykin, Digital Communications, John Wiley & Sons, 1988
[19] A. Papoulis, Probability, Random Variables, and Stochastic Processes, 3rd ed,
McGraw-Hill, 1991.
[20] Ying Li and R. Steel, Serial interference cancellation method for CDMA,
Electronics Letters, vol. 30, pp 1581-1583, Sept. 1994.
[21] A. Kaul and B.D. Woerner, Analytic limits on performance of adaptive multi-
stage interference cancellation for CDMA, Electronics Letters, vol. 30, pp
2093-2095, Dec. 1994.