Packet Combining Error Control for CDMA Systems

Slim Souissi and Stephen B. Wicker

School of Electrical and Computer Engineering
Georgia Institute of Technology
Atlanta, Georgia 30332-0250

Ahtract- Two examples of the application of p d e t combining Two methods of packet combining are considered: Av-
to CDMA are presented and analyzed: an uynduonolu DS/CDMA eraged diversity combining (ADC) and Code Combining
system supported by an averwed divcnity combining scheme and a (CC). The comparison of the two schemes is based on ana-
similarsystcmuaingcodecombining. The ADC scheme uses a Viterbi
decoder combined with a CRC code to generate retrwmission r e lytically derived performance bounds. It is shown that the
qu"(., while the CC &e employs a aequcntid decoder operating ADC approach outperforms the CC method.
with the time-out dgoritbm. Multiple received w e t s ue combined
to form a single. mom nliable &st. The error correcting decoder
operata on the combined packet, U opposed to the mcat ruenlly
received individual packet (e.g. u in a type-I hybrid ARQ proto- 11. DESCRIPTIONOF THE SYSTEM
col), substantidy inc-ng the probsbility of acceptana with each The network under consideration consists of numerous sub-
additional transmission. We &ow through numerical andysis that scribers communicating via a single channel with one or
both technigua provide a significant increase in the CDMA system
capacity, rhroughput and reliability. multiple base stations. We assume that during each packet
transmission, the number of users sharing the channel is
fixed, In practice this can be achieved by setting t i m 4 o t
I. I N T R O D U C T ~ O N intervals during which transmissions are allowed. An asyn-
Code division multiple access (CDMA) systems are very chronous direct sequence CDMA system model similar to
attractive for mobile communication applications because that described in [4] is used t o characterize the multiple-
of their efficient use of the channel, their inherent resis- access channel. Gold sequences (21 with good correlation
tance to fading, and their allowance for simultaneous dig- properties are used t o spread the data '.
ital transmissions by a large community of relatively un- As opposed to conventional ARQ implementations where
coordinated users. Unlike FDMA and TDMA, which are the decoder operates on each received codeword without
primarily bandwidth limited, CDMA is only interference regard to earlier received erroneous words, this paper as-
limited. Any increase in the number of simultaneous trans- sumes that previously received copies of each data packet
missions converts directly and proportionally into an in- are available in the decoder buffer. The buffer is assumed
crease in the effective channel noise level a t the input to t o have sufficient length to prevent overflow. In the re-
each receiver. One appropriate way for combatting channel mainder of the section, we describe in more detail the di-
degradation in CDMA systems is the use of forward error rect sequence CDMA system as well as the two methods of
correcting (FEC) techniques. Despite their good through- packet combining.
put performance, FEC schemes have limited reliability per-
formance. Such a handicap is surmounted by properly A . The CDMA sysiem model
combining an FEC scheme with an ARQ scheme into a
We consider the Spread Spectrum Multiple Aceeas sys-
type-I hybrid ARQ protocol. The ARQ portion provides
tem model that was proposed by Pursley [4]. It is one of
very high reliability while the FEC portion, by correcting
the most frequently used models for the analysis of asyn-
the most likely error patterns, reduces the frequency of re-
chronous direct sequence CDMA systems. The resulting
transmission requests [6].
transmitted signal for the kth user is given by
When type-I HARQ protocols are used in very noisy en-
vironments, a large number of unreliable data packets are
rejected. In order to make use of the information provided s&) = dGak(t)dt(t)cos(w,t + eb), (1)
by these noise-corrupted packets, Chase [l] has suggested
that the packets be combined t o form a single, more reliable where & ( t ) and ak(1) represent, respectively, the transmit-
packet. Such a scheme is adaptive in that only the mini- ted sequence of data bits and the spreading code sequence
mum number of packets necessary t o form a reliable packet ( d i ( t ) and a t ( t ) E {-1, l ) ) , is the phase angle for the
is transmitted; the average number of retransmissions is k" user, P is the common signal power, and fc = 2nw,
minimized, allowing greater throughput while maintaining is the common carrier frequency. st.(t) is passed through
the reliability of the accepted data. the channel where it is combined with the other signals in
In this paper, we analyze the performance of adaptive the system. Assuming IC active transmitters, the received
rat,e coding techniques in direct sequence CDMA systems.
'These sequences arc required to have very low cross-correlationfunc-
This work w w supported by National Science Foundation Grant SCR- tions. The autocorrelation function must be very sharp t o gumantee
9016276 proper acquisition

0-7803-1750-5/94/$3.00 8 1994 IEEE

 X(')= (zy', z!), ...,z$)): i f h transmitted copy of the
same data packet.

Y ( *=
) (y:), y;), ...,~2)):
i f h received copy of the same
data packet.

E'') = (zy),$),...,$)): combined packet formed

by L copies of Y ( ; ) .
The transmision/retransmissionprotocol adopted for this
Figure 1: A CDMA system receiver scheme uses two codes: a CRC code C Ofor error detection
and a convolutional code C1 for error correction. A noise-
less feedback channel is available 80 that the transmitte-
is reliably informed of the status of each decoded packet.
When a message I of k bits is ready for transmission it
is first encoded for error detection using C O ,and then en-
coded for error correction using C1. T h e resulting N-bit
codeword X(') is transmitted. T h e receiver attempts to
decode this first copy Y ( l ) . If the packet is recognized to
Here 7 4 t ) is a white Gaussian Process with a tw*sided have a correctable error pattern then it is acknowledged;
spectral density % and rt is the time delay of the k f h user. otherwise, a retransmission request is sent t o the transmit-
B i and r k are two independent, uniformly distributed ran- ter. The transmitter responds to retransmission requests
dom variables. There is no loss in generality in considering by sending another copy ~ ( 2 ) .The receiver then attempts
only the first user's receiver. As shown in figure 1, the to decode the new received packet ~ ( 2 ) If . the packet is
decision variable a t the output of the integrate and dump decoded correctly, then it is acknowledged, the decoded
filter is given by packet is sent t o the data sink, and all copies of the re-

IT
z(T) = r(t)al(t)cos(w,t).
ceived packet are discarded. Otherwise the receiver com-
(3) bines the packets in the buffer along with the most recently
received copy and the decoder attempts t o decode the com-
The makes a decision as to whether +1 or -1 was bined packet. If the combined packet cannot be decoded
sent by examining the sign of z ( T ) can be expressed reliably, then a retransmission is requested and the trans-
mitter sends another copy X ( 3 ) . T h e receiver continues to
[41
+
z(T)= D ( T ) M ( T ) I(T) + request and combine packets until a decoding operation is
(4) successful.
The first term is the desired signal. The second term repre- Assume that L copies, { Y ( i ) } have
, been received and
sents the contribution of the additive white Gaussian noise recognized to be unreliable. These packets are combined
to z ( T ) ;it is a Gaussian random variable with mean zero into a single packet, $'), by taking the average of the soft
and variance equal to N 0 / 4 . The 1 s t term represents the decision values of each repeated packet symbol,
interference from the other (IC - 1) users. The signal to
noise ratio at the input of the first user's receiver, aver-
aged over all possible phase shifts, time delays and data &L) = -
l L $),
(8)
symbols, is defined as i=l

-
SNR= ~ D(TI2 (5) where yj'" = z r ) + TI:), zy) = fl and TI?) is a Gaussian
v a r ( z ( T ) )'
random variable of variance U:. If we assume that the vari-
ance is constant for all i and equal t o u2,then v a r ( e ) ) =
A Gaussian approximation on the probability of bit error
is given by [4] 0 2 / L , ( j = ,,.

PE = Q(&%), (6) C. Code combining with sequential decoding

where the function Q(.) is the Gaussian integral function In sequential decoding, the most PoPu~ar algorithm is
defined as due to Fano [6]. When the channel is discrete and mem-
oryless (DMC), the Fano branch metric at node i is given
Q(+)=
1
/i "
e-('/')y'dy. (7) by
(9)

B. The Averaged Diversity Combiner ( A D C ) where R is the code rate, t i is the i l h channel input symbol
We first introduce the following notation: and r; is the corresponding received symbol. The path

249
metric through node m is simply The number of computations per decoded bit still follows
m the Pareto distribution given in equation (11). The appar-
ent Pareto exponent QL relative to the L-repetition channel
i=O
is given by

Let C denote the number of computations required to de-

--
E Y ( Q L ) - R,
QL
code the first j information symbols. Under the time-out
condition, a packet is assumed t o be in error if the decoding where E L L ) ( a ~is) the Gallager function of the apparent
time exceeds a certain threshold T,,,. Since the decod- channel given by
ing time is proportional to the number of computations,
a packet is declared unreliable if the number of computa- E i L ) ( n r )=
tions C exceeds a certain threshold Cm,,,. The probability
of decoding failure (erasure) of a packet is therefore

P ( F )= P ( C LGas) CPCL, (11)

(16)
where k is the number of information bits per packet, 0 is
a constant less than or equal t o one and a is the Pareto L is the number of symbol repetitions and m and j repra
exponent. The Pareto exponent is given by the implicit sent the channel input and output symbols respectively.
solution of
--
E&) - R,
(12) 111. APPLICATION TO C D M A
P We consider the two methods of packet combining (ADC
E&) is the Gallager function given by and CC) in conjunction with a direct sequence CDMA sys-
tem. For the case of the BSC channel, the channel crossover
probability is given by equation ( 6 ) and for the c a ~ eof the
AWGN channel, the multiple access noise is assumed to be
additive Gaussian with noise variance given by uar(r(T)).
where j and i are the input and output channel symbols,
respectively, and P ( i ) is the probability assignment of in- A . CDMA with Averaged D i v e r s i t y Combining
put letter k. For a system using FEC/ARQ protocols, the The input t o the averaged diversity combiner is the de-
probability of retransmission for a given packet is equal to cision variable* z given by equation (3). Let Y(i) = ( y f ) ,
P(F). y f ) , __., be the it* received copy. Each bit yy’ is
In [3], the method of packet combining was applied to a associated with its individual decision variable 27’. The
system using convolutional encoding and sequential decod-
ing. When L copies of the same data packet are received corresponding set of decision variables Z(’) = ( z y ) ,z t ’ , ..
and declared unreliable, they are combined and the result- , &I) is formed. The averaged diversity combining process
ing packet decoded. At this point, the encoder is viewed consists of taking the average over all the sets of decision
from the receiver as double-staged. The first stage is the variables Z(’) t o form a single set of updated decision vari-
ordinary convolutional encoder and the second stage is an ables
L-repetition encoder. A k-bit information packet I=(il ,iz
_. ik) is viewed as first encoded into an N-bit packet
X=(zl,z2 .. z , ~ and ) then into an LN-bit packet X ( L )=
( z ~ ’ ) , z ~ ’ . . ...; ) ; zj2)..zjL);...; zc),z$)..zE)),based
z ~zj’),
on the L-repetition code. The resulting packet is transmit- The combined packet = (cy’, c‘,“’, ..., is now.E’)
ted over the channel. T h e received packet is formed by the action of the decision device (see figure 1)
on the decision variables 5”’.
Without loss of generality
we consider the first bit of the combined packet. Since the
where yjil denotes the j l h bit of the ith received copy-. The CDMA system is asynchronous, the random variables zli)
cliannel may also be viewed as a DMC “apparent channel” (i = 1 ... L ) are statistically independent. Assuming that
wliicli produces for each input repetition-symbol (zj ,zj... they all have the same mean and variance (i.e. given by
zj) L output letters (y,!’),~,!’) ___
yjL)). In order to apply
the fundamental sequential decoding results to the sequen-
var($)) ) then the variance of is e)
tial decoder with L-repetition-code combining, we begin by
defining the L-repetition Fano metric as
The effective signal-to-noise ratio after L packets are com-
bined is
SNR(L)= L .SNR,
’We omit t h e bit period T ~n the notation z ( T ) .

250
where SNR represents the signal-to-noise ratio without upper and lower bounds developed earlier for the BER at
combining. The averaged diversity combining of L packets the output of the Viterbi decoder, we can bound P(RiM’).
is equivalent to an increase of the effective received SNR
by a factor of L. P(R$YJw)= 1 - (1 - p(‘))N
UP

Throughput Performance: The system is a concatenated < P(R$‘)) 5 P(R&;j) = 1 - (1 -pf,”,’)”.

-
coding system in which an inner channel is formed by a (26)
convolutional encoderlviterbi decoder pair and an outer
Tr,& is then bounded by [3]
cliannel uses CRC error detection. To determine the per-
formance of this system, we must first characterize the bit
stream at the output of the Viterbi decoder (i.e. the in-
p u t of the CRC decoder). Let Ri’’, dM’ and RiM’ de-
1 +
m
ni

i=l j=1
P(RYjo,,,)5 Tr,& 5 1
m
+ E P ( R $ L p ) . (27)
i=l
iiote, respectively, the events “Viterbi-decoded sequence The system throughput is defined as the ratio the num-
contains no errors”, “Viterbi-decoded sequence contains ber of illformation bits accepted by the receiver to the total
errors that are undetectable by the CRC decoder” and
number of bits transmitted, It is given by
“Viterbi-decoded sequence contains errors that are detectable
by the CRC decode;”. Note that this last event is equiv- RcRCrc
alent to the event “a retransmission request is generated”. Bode = -, (28)
Ti.adc
The superscript M denotes the number of combined copies
which form the sequence at the input to the Viterbi de- where R, is the rate of the convolutional code (the asymp-
coder. It is clear that totic rate loss due t o the encoder memory is assumed t o be
negligible) and R,,, is the rate of the CRC code.
+
P ( R i M ) ) P(RiM))+ P ( R y ) )= 1. (20)
B. CC in Conjunciion with CDMA
If we denote by p ( M ) the effective channel BER a t the out-
In order to derive upper and lower bounds on the aver-
put of the Viterbi decoder given the combination of M
age number of transmissions (Tr,cc), we use equation (27)
received copies of a packet, then
in which P(RY’) is replaced by P ( F ( L ) )(probability of de-
P(R!”)) = ( 1 - P ( ~ ) ) ~ , (21) coding failure a t the LIh attempt). P ( F ( L ) )is determined
where N is the number of bits per transmitted sequence. by
Since P ( d L ’ ) is quite small for most CRC codes, this
P ( F ( L ) )k PLkC;;:, (29)
quantity is neglected and P(R&L’)is written as where k is the number of information bits, CmOzis the
maximum number of computations allowed t o decode the
packet in the time-out algorithm and IYL is the effective
Pareto exponent corresponding t o the combination of L
p(’) is now characterized using upper and lower bounds packets. a~ depends on the convolutional code rate, the
that are a function of the convolutional code in use. When numbers of users in the CDMA system and the number
soft decision Viterbi decoding is used, p ( M ) is bounded by of combined packets. In order to evaluate P ( F ( L ) ) ,we
need to determine the explicit formula for computing the
repetition Pareto exponent QL. This can be done by
solving (15). Once T,,,, determined, the data throughput
of the system is given by e,, = R/T,+.

IV. NUMERICAL RESULTS

A DS/CDMA system using packet combining has been nu-
merically analyzed. Spreading is implemented through a
set of Gold codes of period G5. When averaged diversity
combining is used, error detection is provided by a CRC-12
code while error correction is based on a (2,1,3) convolu-
(24) tional code with generator polynomials (15J7). At the
receiver a soft decision Viterbi decoder is employed. The
where (%)(M) is the effective bit energy to noise ratio
system supported by code combining uses the same con-
alter L packets are combined, volutional code. The decoding is based on the time-out
algorithm for which Cma== 4094 computations and the
(25) packet size N is equal t o 2047 bits. For this case, upper
and lower bounds are very tight and are therefore repre-
and T ( S ,1’) is the weight enumerator for the convolutional sented by the same curve.
code i n use. Let Tr,adc be the expected number of transmis- Figure 2 and figure 3 show upper and lower bounds on
sions before a packet is accepted by the receiver. Using t.lie the probability of retransmission request by the receiver for

25 1
different values of the number of combined packets ( L ) . As
L increases we observe a significant decrease in the proba-
bility of packet rejection.
Figure 4 shows the throughput of the CDMA system
(with CC and ADC) as a function of the number of active
users. We note that ADC has a better throughput perfor-
mance than CC. Actually, this is due the fact that the soft
decision Viterbi decoder has better error correction perfor-
mance than the sequential decoder. Furthermore, when er-
ror detection is based on the time-out algorithm, a correct
packet can be declared in error if the maximum number of
computations CmOz is reached, while for the ADC scheme
a packet is first completely decoded and then it is checked
for the presence of errors. Only erroneous data packets
are retransmitted, thus allowing a lower average number of
transmissions and therefore a higher throughput.
Figure 2: CC: Probability of Decoding Failure per Packet
( k 1 0 Users and Cmnz/L = 2).
VI. CONCLUSION
An analysis of the performance of two methods of packet
combining in a DS/CDMA system over the AWGN chan-
nel has been presented. It has been shown that accept-
able levels of throughput and reliability performance are
provided well beyond the point a t which a conventional
CDMA system would have collapsed. The ADC scheme
using a CRC error detecting code has been shown to have
better throughput performance than the CC scheme o g
crating under the time-out algorithm. By employing an
additional error detecting code, ADC restricts retransmis-
sions to the erroneous data while the CC system, limited by
the time-out condition, occasionally requests the retrans-
mission of packets that could have been correctly decoded.
\\
J l l l
EW(U1

Figure 3: ADC: Probability of Packet Error (K=10 Users

REFERENCES and Packet Size = 2047 bits).

(I] D. Chase, “Code Combining- A Maximum-Likelihood Decod-

ing Approach for Combining M Arbitrary Number of Noisy
Packets”, IEEE Tronwctions on Communicaliona, Vol.COM-
33. No.5, pp. 385393, May 1985. 1.0
[2] R. Gold, “Optimal Binary Sequences for Spread Spectrum Mul-
tiplexing”, IEEE fianmclionr on Informalion Theory, Vol. IT-
13, pp. 619-621, October 1967. 0.1
[3] S. Kallcl and D. Haccoun, “Sequential Decoding with ARQ
and Code Combining”, IEEE Tranraclionr on Communica-
lion.. Vol. COM-36, pp. 775-779, July 1988. 10.6
[4] M. B. Pursley, “Perfonnance Evaluation for Phase-Coded
Spread Spectrum Multiple Access Communications- Part I“,
j
o.21 1 1
IEEE Tmnaacliona on Commrnicalionr. Vol. COM-25. pp. 795- 0.4
799, August 1977.
.[SI. s. SoUis.4 and s. B. Wicker. “ADD~icationof Packet Combin-
~ ~.~~~...
ing Techniques to a DSfCDMA’System Using Convolutional
Encoding and Scquentid Decoding,” Infernational Symposium ’
$7:
- , , ,

on Comm~nicalionTheory and Applicaiions”, Ambleside, Ul<

11-16 July 1993.
[S] S.B.Wicker. Ermr Control Sgdems l o r Digilal Communicalion “0 10 20 30 Q IQ
and Storngr. Prentice Hall. To appear in 1994. N U M M M . U-

Figure 4: ADC and CC: Throughput Performance (Eb/No

= 5 dB).

252