Maximizing System Capacity in a Sub-Symbol based DS-CDMA

Multiuser Detector
Z. A. Uzmi, M. J. Narasimha
{uzmi, sim}@nova.Stanford.EDU
Electrical Engineering Department
Stanford University, Stanford, CA. USA

Abstract- A sub-symbol based multiuser detector for DS- The first linear ZF multiuser detection scheme was
CDMA systems was proposed in [1]. This paper addresses the proposed in [7] for a synchronous CDMA system.
optimization of such detectors to maximize the system capacity. Unfortunately, commercial cellular systems do not have
It is shown that the maximum number of simultaneous users synchronous uplinks since users transmit from random
supported in a DS-CDMA system using sub-symbol scheme can locations at random instants of time. Even if the users time
be increased by the proper selection of certain thresholds
their transmission such that the line-of-sight arrivals are
delineated in [1], and by imposing moderate synchronization
requirements on the transmitters. It is also shown that, in a aligned at the base station, the multipath rays render the
moderately synchronized environment, performing system asynchronous. Thus it is necessary to consider
decorrelation with only one sub-symbol is enough for asynchronous multiuser detection schemes in a practical
accommodating a fairly large number of users, if the system situation. For asynchronous systems, the MUD proposed in
signal to noise ratio (SNR) is medium to high. It has been [8] is difficult to implement as its complexity depends on the
verified through simulations that the maximum number of product of the number of users and the number of symbols in
simultaneous users depends on the degree of synchronism and is the whole duration of transmission. Several linear MUD
approximately given by the number of chips in the longest sub- schemes for asynchronous systems have since been proposed
symbol.
and analyzed. A survey of such schemes can be found in [9],
I. INTRODUCTION [10] and the references therein. All these schemes are based
CDMA is the preferred air interface technology for third on full symbol correlation detection at the base station.
generation wireless systems that are currently undergoing the More recently, techniques that consider sub-symbols for
standardization process. It has been shown previously that the partial correlations have been proposed [1]. This paper
capacity of CDMA systems is limited by multiple access focuses on fine tuning the sub-symbol technique and
interference (MAI) and not by thermal noise [2]. Currently maximizing the number of simultaneous users for the
deployed commercial CDMA systems regard MAI as an combining algorithms presented in [11]. We assume that
additive noise and perform detection of each user moderate synchronism can be achieved by issuing downlink
independently. This limits the maximum number of users that commands to the users to adjust their transmission instants.
can be simultaneously supported due to the deterioration of Section 2 briefly describes the system model considered in
the bit error rate (BER). A number of researchers have the sub-symbol scheme. In Section 3, we discuss the
addressed multiuser detection schemes for CDMA systems in algorithms for sub-symbol combining and promote the idea
the last two decades to improve the system capacity. The that for a large number of users, one long sub-symbol is
main reason that these schemes are not used in practical better than using several small sub-symbols. Simulation and
systems is because of their stringent processing requirements. analytic results and guidelines on choosing the optimum
Multiuser detection schemes are specified as optional [3] thresholds for the algorithms used in the sub-symbol scheme
even in the proposals for the third generation wireless are presented in Section 4.
systems.
The complexity of the optimal multiuser detector (MUD) II. SYSTEM MODEL AND SUB-SYMBOL SCHEME
[4] increases exponentially with the number of users. Hence, The baseband model of a CDMA uplink with single-user
several low-complexity sub-optimal schemes have been matched-filter (SUMF) receiver is depicted in Fig. 1 for N u
proposed and evaluated in the literature [5], [6]. The MUD users. The user data signals before spreading are denoted as
schemes researched so far can be broadly classified into d1 (t ), d 2 (t ), K, d N u (t ) . These are multiplied by the spreading
linear schemes and interference cancellation (IC) methods.
codes K1 (t ), K 2 (t ), K, K N (t ) before transmission. The signal
Linear schemes require more computation but are robust u

against near-far resistance. Therefore, we focus on these received at the base station is the sum of individual signals
schemes in this paper. In particular, we present results for from each user modified by the respective channels, and the
schemes that are based on zero forcing (ZF) linear multiuser additive white gaussian noise. Succinctly, the received signal
detection. ZF MUD schemes have lower complexity, as they is given by:
do not require channel gain estimates.

0-7803-6454-6/00/$10.00 (C) 2000

 Nu III. NUMBER OF SUB-SYMBOLS
r (t ) = å K j (t − τ j ) A j (t − τ j )d j (t − τ j ) + n(t ) The decision to perform decorrelation in a particular sub-
j =1
symbol is based on two thresholds defined in [1]. These two
where τ 1 ,τ 2 , K,τ N u
are the delays incurred by individual thresholds are central to the idea of sub-symbol diversity
1
users in the flat wireless channels with single tap gains combining and are defined below:
A1 (t ), A2 (t ), K, AN (t ) . It is assumed that the variance of (a) Chip Threshold (γc) which indicates that the correlation
u
on a particular sub-symbol should only be performed if the
AWGN n(t ) is σ n2 . The single-user matched-filter (SUMF) size (in number of chips) of the sub-symbol exceeds this
receiver performs a full-symbol correlation for each user threshold. This can be used to eliminate cases where the
independently with the output for the ith user given by: number of users is greater than or equal to the number of
chips in a sub-symbol.
yi (t ) = ò r (t ) K i (t − τ j )
symbol period
(b) Condition Number Threshold (γn) which indicates that
the correlation for a particular sub-symbol should not be
To appreciate the use of ZF MUD in a sub-symbol scheme,
performed if the correlation matrix during that sub-symbol is
which is inherently synchronous, the above expression for a
synchronous system becomes: badly scaled. This is used to avoid the cases where the
Nu correlation matrix accidentally becomes ill-conditioned even
yi (t ) = å Rij (t ) A j (t )d j (t ) + ni (t ) though the number of chips over which correlation is
j =1 performed is large.
where Rij (t ) is the crosscorrelation between the sequences The algorithms used for sub-symbol combining are either
simple combining algorithms or condition-number based
K i (t ) and K j (t ) , and the additive noise ni (t ) has a variance combining algorithms. In each algorithm, the sub-symbols
Rii (t )σ n2 . It is important to note that additive noise in yi (t ) are combined based on a diversity principle for each symbol
from each user.
is uncorrelated with the additive noise in y j (t ) for i ≠ j .
In [1] it is assumed that user transmission is uncontrolled
The vector signal formed from the demodulated signal for in the sense that the starting instants of symbols received at
individual user can be represented as: the base station are uniformly distributed over the whole
H H H
y (t ) = R (t ) A(t )d (t ) + n (t ) symbol period. For a small number of users, the sub-symbol
H intervals are long enough to allow easy decorrelation. But as
It is apparent that slicing performed on R −1 (t ) y (t ) would the number of users increase, the expected size of the sub-
result in almost error free detection, save for the additive symbols becomes smaller, and it becomes harder to perform
noise. This is precisely what the ZF Linear MUD does. It is decorrelations within sub-symbols. When the number of
easy to see that this operation results in elimination of MAI at chips in a sub-symbol falls below the number of users, it is
the cost of increased power of additive noise. Simulation impossible to decorrelate all the users successfully in that
results for the ZF linear MUD have been reported in [12] and sub-symbol since the decorrelation matrix always has one or
a detailed analysis can be found in [7] and [9]. more zero eigenvalues. In particular, if Nc is the number of
If the system becomes asynchronous due to either chips in a given sub-symbol, the maximum number of
multipath or random transmission instants from randomly linearly independent users (see [7] for details) in that sub-
located users, the size of the decorrelation matrix increases symbol is also Nc. For the correlation matrix, a regular
making it difficult to implement the MUD [8]. Techniques inverse does not exist and a generalized inverse (or pseudo
like isolation bit insertion [13] and sliding window inverse) has to be used, while setting γn to infinity. Towards
decorrelator [9] have been proposed to limit the size of this end, it is important to distinguish between two types of
decorrelation matrix. sub-symbols for a given user, a clean sub-symbol indicating
In the scheme presented in [1], the symbols for each user that this particular user has been successfully decorrelated in
are broken into sub-symbols at the boundaries defined by the this particular sub-symbol and a corrupted sub-symbol,
start and end of symbols from other users. This makes an which corresponds to one of the zero eigenvalues of the
asynchronous system look like a synchronous one with correlation matrix.
variable length symbols. The decorrelation operation is then When the correlation matrix is not of full rank, the
carried out in each sub-symbol interval. This yields multiple maximum number of clean sub-symbols will be equal to the
estimates for the same symbol, one estimate for each of the rank of the correlation matrix. Although one ‘long’ sub-
sub-symbols in which decorrelation is performed. These symbol in which decorrelation can be performed is good
estimates are combined according to the algorithm described enough to bring the bit error rate (BER) down by a significant
in the next section. amount (see select-best algorithm in [1]), combining many
sub-symbols brings down the SNR gain. Combining the
corrupted sub-symbols with the clean sub-symbols, however,
1
For frequency-selective channel, the received signal is modified
deteriorates the BER, and it would be better not to combine
appropriately. those sub-symbols at all. Simulations for a sub-symbol

0-7803-6454-6/00/$10.00 (C) 2000

scheme in which users transmission instants are uncontrolled σ X2 = Nσ X2
f 1
indicate that combining all sub-symbols greater than a
N
relatively low γc (less than 5% of the spreading gain) would 1
result in the allowable number of users only up to 8-9% of the
Zf =
N
å
i =1
Xi
spreading gain. This is also indicated in Fig. 2 for a spreading σ Z2 = σ n2 / N
gain of 255 with Kasami sequences [14]. As more and more f

users are packed, sub-symbols become smaller and smaller For the sub-symbol case, let M be the number of chips in
and for a given user many of the sub-symbols are corrupted the best symbol, as selected by the select-best algorithm. If
which significantly affect the BER performance. only one sub-symbol is used, then M is equal to the expected
In this paper, we consider a moderately controlled number of chips in the longest sub-symbol within a symbol.
environment in which the users are instructed1 to adjust their M can be analytically computed or it can be found from
transmission instants such that their symbol arrival times at simulations. As an example, for 12 asynchronous users with
the base station fall within some fraction of the symbol uniformly distributed random transmission instants, M turns
period. This kind of moderate synchronization is easier to out to be approximately 55 for the case N=255. The general
implement as compared to a fully synchronized system. The procedure for computing M when user transmission instants
thresholds for sub-symbol decorrelation then play an are moderately synchronized will be outlined later. Similar to
important role in determining the maximum number of users the above computations, for the sub-symbol of length M, we
that can be simultaneously accommodated in the system. The can write:
chip threshold (γc) is set high enough such that only one sub-
M
S s = å1 = M ; X s = X1 + X 2 +L+ X M
symbol is used for demodulation. This has an effect on i =1
processing gain, but if the users can be decorrelated 1 M
successfully, it still yields a significant BER improvement σ X2 = Mσ X2 ;
s 1 å X i ; σ Z2s = σ n2 / M
Zs =
M i =1
over the single-user matched-filter receiver for medium to
high signal to noise ratios. The sub-symbol combining system Thus the SNR penalty (deterioration in SNR with respect
thus becomes the one that only uses selection sub-symbol to ideal ZF MUD) for the select-best algorithm in
diversity since we consider only one dominant sub-symbol approximately given by N/M. This penalty can be reduced if
per symbol. Thus all the sub-symbol diversity-combining fewer users are present and more sub-symbols can be
algorithms converge to the “select best sub-symbol” combined. If the system is moderately synchronized and the
combining algorithm. The only difference is that, in this case, users are allowed to have their transmission instants lie
it is known which sub-symbol would be selected for within the first P chips of a reference symbol, then the
decorrelation. expected value of M is given by:
{( )}
P −1
The SNR penalty for such an algorithm with respect to an E [M ] = ( N − P ) + å i α i − α i +1 + Pα P
ideal case can be easily computed. Assume that S denotes the i =0
total signal, X denotes the total noise and Z represents the
α − α P+1
noise per chip in the decorrelation interval. Also let N be the = ( N − P) +
spreading factor (chips per symbol). The sampling is assumed 1−α
to be one sample per chip and it is also assumed that fading where α is defined as:
Nu
imposes a fixed penalty in each receiver. These assumptions æ P −1ö
can be justified for comparison of different receivers since α ≡ç ÷
è P ø
they are used in similar environments. The subscript “f” is
For practical reasons, P is a large fraction of N (typically
used to refer to the case of full-symbol synchronous and the
50% to 80%) in which case the expected value of M for a
subscript “s” is used to denote the case of sub-symbol
large number of users is approximately given by N-P. The
synchronous receiver. Constant fading penalty assumption
exact expected value is listed in the following table for
allows comparing non-faded systems. For full-symbol
N=255.
synchronous system, the decorrelation is performed over one
full symbol and we can write:
N
P = 64 P = 128 P = 192
S f = å1 = N Nu = 10 users 196.8627 139.2559 81.6535
i =1 Nu = 30 users 192.6558 130.7696 68.8964
X f = X1 + X 2 + L + X N Nu = 50 users 191.8349 129.0826 66.3517
where the Xi’s are i.i.d. Normal(0, σ n2 ) random variables, and Nu = 70 users 191.4972 128.3669 65.2661
indicate the chip noise. Hence, Nu = 90 users 191.3199 127.9750 64.6668

1
Sending commands on the downlink channel can accomplish this.

0-7803-6454-6/00/$10.00 (C) 2000

 IV. OPTIMIZING THE THRESHOLDS moderately synchronized system with a spreading gain of 255
For the decorrelation process in each sub-symbol, either is given in the following table:
the regular inverse (which we will denote by ℑ) can be used
or the generalized inverse (denoted by ℘) can be used. We P = 128 chips, γc = 25% of #users
will consider the effect of thresholds in each case. The Pinv (℘) Inv (ℑ) Inv (ℑ)
condition number threshold γn is important only if the ℑ (γn = 200) (γn = 500)
operation is performed and does not have much significance 116-118 97 110
for the ℘ operation. As long as the number of chips in a sub- P = 192 chips, γc = 100% of #users
symbol is less than the number of users in that sub-symbol, ℑ Pinv (℘) Inv (ℑ) Inv (ℑ)
operation can not be performed. Selecting γc as less than the (γn = 200) (γn = 500)
number of users would result the performance of ℑ operation 62-63 49-52 51-54
to become independent of γc. As the number of users
increases, it may not be possible to perform the ℑ operation The maximum number of supported users shown in the
at all and ℘ is the only recourse. In such a case, a small value above table are determined by the criteria that the sub-symbol
of γc results in more and more sub-symbols being considered detector results in a bit error rate (BER) that is at least an
for decorrelation, which increases the probability of order of magnitude better than the single-user matched-filter
contamination of clean sub-symbols by the corrupted sub- receiver at a signal to noise ratio of 30dB.
symbols. Thus increasing the number of users brings two
benefits: First, more users are accommodated and secondly V. CONCLUSIONS
there is less chance for a corrupted sub-symbol being We presented results on the maximum number of
considered for combining with closely packed users. simultaneous users that can be supported in a DS-CDMA
Therefore, if users transmission instants can be bound within system under the sub-symbol scheme when moderate
a small region, a value of γc exceeding the length of that synchronization can be enforced among the users. Under such
region is desirable. Otherwise, with fewer users, it is better to conditions, it is shown by simulations that only one sub-
use the ℑ operation. symbol is enough for accommodating a fairly large number
If the ℘ operation is to be used, then it should be made of users. It turns out that by using the appropriate values for
sure that in each sub-symbol, there are enough chips to the chip and condition number thresholds, the "select best"
decorrelate the many users. If most of the users in a sub- sub-symbol scheme can support a maximum number of users
symbol remain dependent, they will deteriorate the that is approximately given by the maximum number of
performance achieved by the rest of the sub-symbols, and it is consecutive chips in any sub-symbol. For example, if the
advantageous to isolate that sub-symbol. Obviously, the ideal longest sub-symbol is 63 chips long, irrespective of the
situation is to be able to pick the sub-symbols for only those spreading factor, the maximum number of users supported is
users that have been decorrelated successfully. If it is not approximately 63. In this case, for each user, the sub-symbol
possible to selectively pick users from within a sub-symbol, based MUD yields a BER that is at least an order of
the chip threshold γc should be selected such that there is a magnitude better than a single-user matched-filter detector.
good chance for most of the users to be decorrelated. Thus as
the number of users increases, a larger value of γc is needed,
upto a certain limit, after which it should be set to a constant REFERENCES
value such that we have at least one sub-symbol to carry out [1] Z. A. Uzmi and S. A. Mujtaba, “A Novel Scheme for Linear Multiuser
Detection in Asynchronous DS-CDMA Systems with Frequency
the decorrelation process.
Selective Fading,” Proceedings of the IEEE Globecom, pp. 981-985,
An alternate measure for performance adjustment is the Dec. 1999.
rank threshold γr, which indicates that the decorrelation [2] G. L. Turin, “The Effects of Multipath and Fading on the Performance
operation is performed in a sub-symbol if the corresponding of Direct-Sequence CDMA Systems,” IEEE JSAC, vol. SAC-2, no. 4,
pp. 597-603, July 1984.
correlation matrix has a rank that exceeds the rank threshold. [3] Concept Group Alpha-WCDMA, Evaluation Document, Part-1:
For small number of users, γr should be made equal to the System Description; Performance Evaluation, Tdoc SMG 905/97.
number of active users or even slightly higher in order to [4] S. Verdú, “Minimum Probability of Error for Asynchronous Gaussian
eliminate the risk involved in considering a corrupted sub- Multiple-Access Channels,” IEEE Transactions on Information Theory,
vol. IT-32, no. 1, pp. 85-96, Jan. 1986.
symbol for later combining. Under this condition, γn can be [5] S. Moshavi, “Multi-User Detection for DS-CDMA Communications,”
used in conjunction with γr to great advantage. For a very IEEE Communications Magazine, pp. 124-136, Oct. 1996.
large number of users, we have only one sub-symbol [6] A. Duel-Hallen, J. Holtzman, and Z. Zvonar, “Multiuser Detection for
CDMA Systems,” IEEE Personal Communications, pp. 46-58, Apr.
available that has to be decorrelated and a value of γr slightly 1995.
smaller than the number of users would be the right choice. [7] R. Lupas and S. Verdú, “Linear Multiuser Detectors for Synchronous
Under the above rules, the maximum number of users in a Code-Division Multiple-Access Channels,” IEEE Transactions on
Information Theory, vol. 35, no. 1, pp. 123-136, Jan. 1989.

0-7803-6454-6/00/$10.00 (C) 2000

[8] R. Lupas and S. Verdú, “Near-Far Resistance of Multiuser Detectors in Beamforming and Linear Multiuser Detection,” Proceedings of
Asynchronous Channels,” IEEE Transactions on Communications, vol. ASILOMAR Conference on Signals, Systems, and Computers, January
38, no. 4, pp. 496-508, Apr. 1990. 2000.
[9] S. Verdú, Multiuser Detection, Cambridge University Press, 1998. [13] Fu-Chun Zheng and S. K. Barton, “Near-Far Resistant Detection of
[10] T. Ojanperä and R. Prasad, Wideband CDMA for Third Generation CDMA Signals via Isolation Bit Insertion,” IEEE Transactions on
Mobile Communications, Artech House, 1998. Communications, vol. 43, no. 2/3/4, pp. 1313-1317, Feb./Mar./Apr.
[11] Z. A. Uzmi, S. A. Mujtaba, and M. J. Narasimha, “Simplified Multiuser 1995.
Detection for Asynchronous DS-CDMA Channels,” To be Submitted, [14] D. V. Sarwate and M. B. Pursley, “Crosscorrelation Properties of
IEEE Tranactions on Communications. Pseudorandom and Related Sequences,"”Proceedings of the IEEE, vol.
[12] Z. A. Uzmi and S. A. Mujtaba, “Performance Analysis of a 68, no. 5, pp. 593-619, May 1990.
Convolutionally-Encoded Synchronous CDMA System with Adaptive

K1 A1 K1
d̂1
τ1
d1 y1
ò (⋅)
symbol period

K2 A2 K2
n
τ2
d2 d̂ 2
ò (⋅)
y2

å
symbol period

K Nu AN u K Nu
d Nu d̂ Nu
τN u
ò (⋅)
symbol period
y Nu

Transmitter Channel SUMF Receiver

Fig. 1: Baseband DS-CDMA System Model with SUMF Receiver.

0
10

-1
10

-2
10
Bit Error Rate

10 users, SUMF
14 users, SUMF
-3 18 users, SUMF
10
22 users, SUMF
26 users, SUMF
10 users, Sub-Symbol
-4 14 users, Sub-Symbol
10
18 users, Sub-Symbol
22 users, Sub-Symbol
26 users, Sub-Symbol
-5
10
-3 3 9 15 21 27 33
E b /No in dB

Fig. 2: Maximum Number of Users in Sub-Symbol Schemes with γc=1.25Nu and γn=∞ for
completely asynchronous system.

0-7803-6454-6/00/$10.00 (C) 2000