Académique Documents
Professionnel Documents
Culture Documents
DOI 10.1007/s11277-008-9553-7
A. Bentrcia · A. Zerguine
Abstract In this paper, a new linear group-wise parallel interference cancellation (LGPIC)
detector is proposed. Four different group-detection schemes are derived, namely, the linear
group matched filter PIC (LGMF-PIC) detector, the linear group decorrelator PIC (LGDEC-
PIC) detector, the linear group minimum mean square error PIC (LGMMSE-PIC) detector
and the linear group parallel interference cancellation weighted PIC (LGPIC-PIC) detector.
The convergence behavior of the proposed detector is analyzed and conditions of conver-
gence are derived. Finally, extensive simulations regarding the convergence behavior and
the effect of the grouping on the convergence behavior of the proposed LGPIC detector are
conducted.
1 Introduction
Interference cancellation (IC) structures are the most promising multiuser detector structures
to be implemented in future commercial systems [1]. Linear interference cancellation struc-
tures are commonly used to implement the decorrelator/LMMSE detectors [2]. Two types of
IC structures can be distinguished: successive interference cancellation (SIC) structures [3]
and the parallel interference cancellation (PIC) structures [4]. The linear SIC structure exhib-
its low computational complexity, however, it suffers from relatively long detection delay.
One solution to this problem is group-detection where groups of users instead of individual
users are detected in series, which reduces considerably the detection delay of the linear SIC
detector. Such structure is known as the linear group-wise SIC (LGSIC) structure [5]. The
123
24 A. Bentrcia, A. Zerguine
PIC detector on the other hand, is attractive due mainly to its inherent parallelism which, if
properly exploited, reduces considerably the computation time.
On the other hand, as for the LGSIC detector, the convergence behavior of the linear PIC
detector can be greatly increased if group-detection is applied, where groups of users instead
of individual users are detected in parallel. Despite of all these apparent advantages, and up
to our knowledge, no linear group-wise PIC structure is proposed in the literature. In this
paper, a linear group-wise PIC (LGPIC) structure is proposed. The LGPIC detector will be
shown to be equivalent to linear matrix filtering which allows the derivation of analytical
expressions for the Bit Error Rate (BER) and Asymptotic Multi-user Efficiency (AME) for
proposed structure. Depending on the group-detection scheme, four different group detection
schemes are obtained. The proposed LGPIC structure is shown to converge to the decorre-
lator detector if it converges and conditions of convergence are derived. Finally, simulation
results are shown to corroborate well with theory.
In this paper, we consider a case of an uplink channel scenario where K users transmit
simultaneously over a synchronous additive white Gaussian noise (AWGN) channel using
Binary Phase Shift Keying (BPSK). Each user is characterized by its own pseudo-noise code
of length N chips. The received signal is expressed in vector form as:
r = SAb + n (1)
√ √ N ,K
where S = (s1 , s2 , . . . , sk , . . . , s K ) ∈ −1/ N , 1/ N , A = diag(a1 , a2 , . . . , ak ,
. . . , a K ) ∈ R K ,K , b = (b1 , b2 , . . . , bk , . . . , b K )T ∈ {−1, 1} K and n = (n 1 , n 2 , . . . , n n , . . . ,
n N )T ∈ R N .
S is a N × K matrix of the spreading codes, sk is the N × 1 spreading code of the kth
user, A is a K × K matrix of the received amplitudes, b is a K -length vector of received
binary symbols, and finally n is a N -length vector of independently and identically distributed
additive white Gaussian samples with zero-mean and variance σ 2 .
In the following we assume that the K users are partitioned into G groups, where the gth
group consists of Ug users such that: K = U1 +U2 + · · · +Ug + · · · +UG and thus the matrix
S can be partitioned as S = (S1 , S2 , . . . , Sg , . . . , SG ) where Sg = (sg,1 , sg,2 , . . . , sg,u g , . . . ,
√ √ N ,Ug
sg,Ug ) ∈ −1/ N , 1/ N . We define: R = ST S as the cross-correlation matrix of
the spreading codes, Ri, j = SiT S j as the (ith, jth) submatrix of R, and Ag as the gth diagonal
submatrix of the matrix A. We assume that R and Rg,g (for g = 1, 2, . . . , G) are nonsingular
(the spreading codes are assumed to be linearly independent).
The LGPIC detector consists of interference cancellation units arranged in a multistage
structure as shown in Fig. 1. The internal structure of each interference cancellation unit is
illustrated in Fig. 2. The vector of decision variables of the ( p − 1)th stage, gth group y p−1,g
is first despreaded added to the vectors of decision variables of the other G groups to form the
interference due to all users at the ( p − 1)th stage, that is, I p−1 = j=1 S j y p−1, j . The
interference I p is subtracted from the received signal r to obtain a purified received signal
(r − I p ) where all users exhibit less mutual interference. The vector of decision variables
of the pth stage, gth group y p,g is obtained by despreading the purified signal, multiplying
123
A New LGPIC Detector 25
the result by a transformation matrix and finally adding the result to the vector of decision
variables of the previous stage, that is:
y p,g = Fg STg r − I p−1 + y p−1,g (2)
123
26 A. Bentrcia, A. Zerguine
S1 ST1 F1
Ip-1 – +
yp-1,1 yp,1
S2 ST2 F2
yp-1,2 yp,2
Sg STg Fg
yp-1,g yp,g
SG STG FG
yp-1,G yp,G
Fig. 2 The pth stage interference cancellation unit of the LGPIC detector
that the computational complexity of the proposed LGPIC structure using the decorrelator
detector as the group-detection scheme is given by:
⎛
G k G ⎞
⎜ 2 g U g + W N + max τ + max τlk g=1 2U g − 1 +⎟
⎜
PW ⎝ 1≤k≤K 1≤k≤K
⎟
G ⎠
2W N + 2 max τ k + 2 max τlk − 1 U
g=1 g
1≤k≤K 1≤k≤K
k
+ P (W G − 1) W N + max τ + max τlk (3)
1≤k≤K
1≤k≤K
k
+ 2P W N + max τ + max τlk + W G g=1 11Ug + 2 Ug + Ug
3 3 2
1≤k≤K 1≤k≤K
k G 2
+2W W N + max τ + max τlk g=1 Ug
1≤k≤K 1≤k≤K
where τlk is the delay of the lth path (L resolvable paths are considered) of the kth user and
τ k is the relative delay of the kth user.
123
A New LGPIC Detector 27
In this section, we show, using an algebraic approach, that the LGPIC detector is equivalent
to matrix filtering of the received chip-matched signal. This enables the determination of
analytical expressions for the BER and AME of the proposed detector.
The vector of decision variables of the pth stage, gth group y p,g in Eq. 2 can be written
in matrix form as:
y p = y p−1 + FST r − Sy p−1 (4)
where F = diag F1 , F2 , . . . , Fg , . . . , FG . Hence (4) is equivalent to:
bk = 1
123
28 A. Bentrcia, A. Zerguine
Similar to the case of the linear GSIC detector proposed in [5] and depending on the
transformation matrix Fg , different group detection schemes can be obtained, namely: the
LGMF-PIC, LGDEC-PIC, LGPIC-PIC and finally the LGMMSE-PIC detectors.
Note that if the group size is equal to one, we obtain the conventional linear PIC detector.
Before discussing the convergence behavior of the proposed scheme, let us establish the
connection between the LGPIC detector and the Jacobi/block-Jacobi iterative method [7].
three parts, that is: R = D − L− L , where DT is block
The matrix R can be decomposed into T
diagonal matrix, that is D = diag R1,1 , R2,2 , . . . , Rg,g , . . . , RG,G , and L and L are the
remaining lower-left and upper-right block triangular parts of R, respectively. On the other
hand, the block-Jacobi iterative method is given by [7]:
y p = y p−1 + D−1 y − Ry p−1 (13)
−1
By comparing (4) and (13), it easy to notice that if Fg = STg Sg then the LGPIC
detector is in fact a realization of the block-Jacobi iterative method. On the other hand, if
123
A New LGPIC Detector 29
−1
Fg = diag STg Sg then the LGPIC detector is in fact a realization of the Jacobi relaxation
iterative method.
From (6), it easy to show that as the number of stages tends to infinity the vector of decision
variables tend to that of the decorrelator detector, that is,:
p
i−1
lim y p = lim I − FST S FST r
p→∞ p→∞
i=1
−1
= FST S FST r (14)
−1
= ST S F−1 FST r
which is the decorrelator detector, therefore, if the proposed LGPIC detector converges, it
converges to the decorrelator detector.
To determine the conditions of convergence, we use the following theorem [7]:
Let B be a square matrix such that |λmax (B)| < 1, then the iteration b p+1 = Bb p + f
converges for any f and b0 where λmax is the largest eigenvalue of the matrix B.
By rewriting Eq. 4 as:
y p = FST r + I − FST S y p−1 (16)
and using the above theorem, it is easy to show that the iteration matrix B of the proposed
detector is given by:
B = I − FST S (17)
If the above inequality is not always satisfied then the LPIC and LGPIC schemes will diverge.
However, as in the case of the weighted LPIC scheme [8,9], a weighted LGPIC scheme can
be used here as well to ensure convergence.
6 Simulation Results
In this section, the convergence behavior of the proposed LGPIC multiuser detector is sim-
ulated and the results obtained are detailed. Two different scenarios are considered: a syn-
chronous CDMA AWGN channel and an asynchronous CDMA multipath Rayleigh fading
channel. The simulation parameters are depicted in Table 1.
Figure 3 depicts the average BER (average of all users) versus the number of linear
group-wise PIC stages. Four different detection schemes are considered. For the LGPIC-PIC
detector, a 2-stage PIC detector is used. It is easy to notice that the LGDEC-PIC converges
faster than the other group-detection schemes. As can be seen from this figure, the LGDEC-
PIC needs only 7 stages whereas the LGPIC-PIC detector needs 8 stages, the LGMMSE-PIC
123
30 A. Bentrcia, A. Zerguine
-1.58
10
-1.59
10
-1.6
10
2 4 6 8 10 12 14 16 18 20
Number of LGPIC stages
detector needs 10 stages and the LGMF-PIC detector needs 11 stages; however, the linear
LGMF-PIC and the LGMMSE-PIC detectors achieve the lowest average BER level among
all detection schemes. Moreover, it is important to notice that lower average BER levels are
achieved prior to convergence, this is more noticeable for highly loaded systems and it has
also been reported in [3].
The effect of grouping is analyzed and depicted in Figs. 4–7. It can be seen that while
convergence speed of the LGDEC-PIC, the LGMMSE-PIC and the LGPIC-PIC detector
increases with decreasing number of groups, the convergence speed of the LGMF-PIC detec-
tor is independent of grouping and is constant for any grouping. This is because the LGMF-
PIC detector is equivalent to the conventional LPIC detector and hence the grouping in this
case is G = K . However, the average BER difference between different groupings is small
and of theoretical importance only.
In Fig. 8, the convergence behavior of different LGPIC detection schemes is evaluated in
an asynchronous CDMA multipath fading channel. Ten users are divided into two equally
sized groups. In addition, a two-stage PIC detector is used for the group-detection in the
LGPIC-PIC detector. The simulation parameters are depicted in Table 1.
This Figure shows that while the LGMF-PIC, which is equivalent to the LPIC detec-
tor, and the LGPIC-PIC detectors are divergent, the LGDEC-PIC and the LGMMSE-PIC
123
A New LGPIC Detector 31
-1.56
10
-1.59
10
-1.6
10
2 4 6 8 10 12 14 16 18 20
Number of LGDEC-PIC stages
-1.56
10
-1.59
10
-1.6
10
2 4 6 8 10 12 14 16 18 20
Number of LGMMSE-PIC stages
detectors are convergent and they need few stages to converge to the decorrelator detector’s
performance. This indicates that while the LPIC detector is divergent, its counterpart LGPIC,
particularly the LGDEC-PIC and the LGMMSE-PIC detectors is convergent. This shows that
group-detection can be used to stabilize a divergent LPIC detector. Hence, it is in fact an alter-
native to the conventional way of stabilizing a divergent LPIC detector where a weighting
factor can be used for this propose.
Finally, Fig. 9 shows the convergence behavior of both the LGDEC-PIC and LGDEC-
SIC detector for two different groupings, namely for G = 2 and G = 10. It is obvious from
this figure that the LGDEC-SIC detector converges relatively faster than the LGDEC-PIC
123
32 A. Bentrcia, A. Zerguine
-1.56
10
-1.59
10
-1.6
10
2 4 6 8 10 12 14 16 18 20
Number of LGPIC-PIC stages
-1.56
10
-1.59
10
-1.6
10
2 4 6 8 10 12 14 16 18 20
Number of LGMF-PIC stages
detector, i.e., for G = 2, the LGDEC-SIC detector needs only 5 stages while LGDEC-PIC
detector needs around 7 stages to converge to the decorrelator’s performance. This is because
the LGDEC-SIC detector uses the most updated estimates of the decision variables at the
expense of more detection delay.
7 Conclusion
In this work, a new linear group-wise PIC multiuser detector is developed. The proposed
scheme exhibits inherent parallelism compared to the group-wise SIC detector, which
123
A New LGPIC Detector 33
0
10
Matched filter detector
Decorrelator detector
LMMSE detector
LGDEC-PIC detector
LGPIC-PIC detector
LGMMSE-PIC detector
LGMF-PIC detector
Average BER
-1
10
-2
10
5 10 15 20 25
Number of LGPIC stages
Fig. 8 Convergence of different LGPIC detection schemes in an asynchronous CDMA multipath Rayleigh
fading channel
-1.58
10
-1.59
10
-1.6
10
2 4 6 8 10 12 14 16 18 20
Number of LGSIC/LGPIC stages
Fig. 9 Convergence behavior of both the LGDEC-PIC and the LGDEC-SIC detectors for G = 2 and G = 10
reduces the computation time. Depending on the group-detection scheme, four different
group-detection PIC structures were derived. First, we used a matrix algebraic approach to
describe the proposed structure. This approach enabled us to derive analytical expressions
for both the BER and AME for the proposed detector. Second, we proved that the proposed
structure converges to the decorrelator detector if it converges. Moreover, we showed that
the proposed structure, like the conventional linear PIC detector, is not always convergent
and that a weighting factor can be used to stabilize the scheme. Finally, extensive simulations
were performed to uncover different the convergence behavior of the proposed structure.
123
34 A. Bentrcia, A. Zerguine
References
1. Kourtis, S., McAndrew, P., & Tottle, P. (1998). W-CDMA: Aspects of implementation. IEE Colloquium
on UMTS – The R&D Challenges 23 Nov 1998 (pp. 10/1–10/8).
2. Grant, A., & Schlegel, C. (2001). Convergence of linear interference cancellation multiuser receivers.
IEEE Transaction on Communication, 49(10), 1824–1834.
3. Rasmussen, L. K., Lim, T. J., & Johansson, A. (2000). A matrix-algebraic approach to successive inter-
ference cancellation in CDMA. IEEE Transaction on Communication, 48(1), 145–151.
4. Guo, D., Rasmussen, L. K., Sun, S., & Lim, T. J. (2000). A matrix-algebraic approach to linear parallel
interference cancellation in CDMA. IEEE Transaction on Communication, 48(1), 152–161.
5. Johansson, A. L., & Rasmussen, L. K. (1998). Linear group-wise successive interference cancellation
in CDMA. IEEE 5th Proceedings of the International Symposium on Spread Spectrum Techniques and
Applications, 1(2–4), 121–126.
6. Verdu, S. (1998). Multi-user Detection. Cambridge University Press.
7. Saad, Y. (2003). Iterative Methods for Sparse Linear Systems, (2nd ed.). publisher: SIAM
8. Rasmussen, L. K., & Oppermann, I. J. (2001). Convergence behaviour of linear parallel cancellation in
CDMA. IEEE Global Telecommunication Conference, San Antonio, TX, December 2001. (pp. 3148–
3152).
9. Rasmussen, L. K., & Oppermann, I. J. (2003). Ping-pong effects in linear parallel interference cancellation
for CDMA. IEEE Transactions on Wireless Communication, 2, 357–363
10. ETSI TR 101 112, UMTS 30.03, V3.2.0, Annex B, Sections 1.2.3, 1.3, 1.4. 1998.
Author Biographies
123