Abstract—In this paper, we propose simple and efficient provide a multitude of services, especially multimedia and high
algorithms for the code timing acquisition in the direct-sequence bit rate packet data.
code-division multiple-access communication system. The es- CDMA is based on spread-spectrum technique, where
sential assumption is that a preamble or an unmodulated pilot
channel is available for the desired user. Then the correlation the idea is to spread the narrow-band information signal
R
matrix ( ) of the sampled data, where is suitably chosen into a common wide frequency band before transmission. In
time lag, contains the timing information only of desired user, direct-sequence (DS) CDMA, the spreading is performed by
while the contributions of uncorrelated interferers and noise wide-band noise-like signal, which simultaneously identifies
are suppressed out. Hence, compared to the conventional ap- each user in the system. This pseudonoise is also called the
proach, more interference suppression is achieved. Coarse delay
estimates are then obtained by matched filter (MF) or multiple code or the chip sequence of a particular user.
signal classification-type approaches. In the latter case, only L The final objective in the reception of a DS-CDMA system is
eigenvectors are computed, where L is the number of resolvable to estimate the symbols which carry the data, but a prerequisite
paths. If only one path exists, an additional procedure is proposed task is to get the local code generator synchronized to that of
to both approaches, by which the estimation accuracy is greatly received signal. This means estimation of the propagation delay,
improved with negligible increase in computation. More precisely,
the chip timing offset due to chip-asynchronous sampling can be which gives the required knowledge to the receiver about the
determined by solving a system of two second-order polynomials phase of the spreading code. In addition, the strengths of possible
for each chip interval. Therefore, only at most 2C hypotheses are multipaths and carrier phase must be estimated. In this paper,
needed, where C is the processing gain. we consider only code timing estimation, because it tends to be
All the proposed methods are computationally quite simple, con- the most challenging task, and because the other parameters can
taining mainly MF-operations, or at most computation of only few
eigenvectors. Mean acquisition time analysis is carried out semi- be estimated given a reliable delay estimate [2], [3].
analytically. Numerical experiments speaks for the possibility of Conventional CDMA systems rely on single-user techniques,
achieving significant performance gains compared to conventional such as matched filter (MF) [4], in both delay estimation and
acquisition, especially in the presence of strong multiple-access in- detection. Although simple, they are inadequate if the code or-
terference, making them attractive options to be attached for the thogonality conditions are perturbed. This happens even in syn-
next generation mobile receivers.
chronous system with orthogonal codes due to the existence of
Index Terms—Code acquisition, code-division multiple access, multipaths with different delays. Moreover, if the desired signal
differential correlation. is much weaker than the interfering signals, which is commonly
known as a near–far problem, single-user techniques can totally
I. INTRODUCTION collapse. In the uplink (e.g., mobile to base) communication,
near–far problem can be mitigated by power control, but if used
assumed to remain roughly constant in the observation interval. 4) Fading process of each path is stationary. Therefore,
denotes noise, and the chip sequence length (i.e., pro- the correlation coefficient equals
cessing gain) is . From now on, we assume
for simplicity.
First, the received data is sampled by chip-matched filtering,
and the equispaced samples are collected into -vectors where , and .
5) Noise is a random zero mean variable, and independent
(2) with the data and fading. Then
, where and are those of (11).
They have a well-known form (e.g., [8]) Notice that assumption 1) can also be replaced by assuming
that there exists a common unmodulated pilot channel. Notice
(3) also that assumptions 3) and 4), without zero-mean assumption,
are equivalent to the common uncorrelated scatterer (US) model
where denotes noise vector, and [4], and wide-sense stationary (WSS) model [18], respectively.
It should be noted, however, that assumption 3) is not necessary
(4) for the purposes of the paper, but it is made to achieve a simple
(5) and elegant representation of . The case where 3) do not
apply is discussed in Remark 1.
Here the “early” and “late” parts of the code vectors are According to (11), and assumption 5), the differential corre-
lation matrix now is
(6)
(7)
Notice that (4) and (5) and (6) and (7) are due to the chip and
(12)
symbol asynchronous sampling, respectively. To get a more
compact representation of the data, let us define a where the dimensional matrix is defined as
dimensional code matrix
(13)
(8)
According to (10), the matrix contains subma-
and a -vector trices of dimension . Denote these by , i.e.,
(9) (14)
Stacking all the vectors into a -vector Denote the element of a matrix. From (9) we directly
have
(10)
the sample vector (3) can be rewritten as
(11)
Here depends on the codes and the delays, and depends
on the channel coefficients and the symbols.
more. The reason why is not a good choice for the time the zero submatrices were always due to the uncorrelatedness
lag is discussed in Remark 2. and the zero-mean properties of the interfering user’s symbols.
Case 1) . Now all the symbols are equal to Therefore, is free of contributions of interfering users even
1, and depends only on the channel without assumption 3), although the form of is a little bit
coefficients. With we have different. This is because in “Case 1”, both subcases “ ”,
due to assumption 3). On the other hand, with , and “ ” cause nonzero submatrices. Assumption 3) is valid,
we have e.g., for Rayleigh-fading channels [4], but do not apply, e.g., for
Rician fading [4] and fixed multipath channels. In the former
(17) case, not all the paths have a zero mean. However, we still have
for , since only one path has a nonzero
due to assumption 4). mean. In the latter case, all the paths are correlated and have a
Case 2) . Now depends on nonzero mean.
the channel coefficients and the symbols of Remark 2: It is important to notice that the situation
user . Since these symbols are zero mean, is not as desired as the case symbols. The reason is as
for all . follows: due to symbol asynchronous sampling, the sample vec-
Case 3) . In this case, we always have tors usually contain information about two successive sym-
due to assumption 2). bols [see (3)]. Therefore, with the time lag symbol, two
Hence, the matrix contains only nonzero submatrices successive sample vectors also have some degree of correlation
of , and they correspond to the situations where with respect to the interfering users. To see this in details, let
and . Due to the definition (10), these submatrices are , and consider . If now with
situated in the main diagonal of . Therefore, (12) reduces , which belongs to “Case 3” above, we would have a
to nonzero submatrix equal to
.. (21)
.
If is now rewritten in the form of (19), would clearly
have a contribution which corresponds to the interfering user
(18) . For this reason, we only consider cases where
from which we see that depends only on first rows of symbols.
the codematrix . Since these rows correspond to the desired Remark 3: It is worth mentioning that the differential cor-
user only [see (8)], the contribution of interference is sup- relation matrix with contains two kinds of infor-
pressed from . To express in detail the form of , we mation about the channel: the average power of the paths, and
make a mild assumption that , the rapidity of their fading process. The latter is related to the
and denote it . This means that the fading process has correlation coefficient , which is closer to (respectively,
almost the same degree of correlation whether the time dif- 1) the faster (respectively, slower) is the fading process. There-
ference is or symbols. Then, (17) reduces to fore, provided equal energy paths, the slowly varying paths are
, and consequently (see details in Ap- the best represented in . This is a nice property, because
pendix A) such paths are more valuable to be acquired due to their easier
tracking.
.. In practice is estimated by
.
(22)
(29)
(30)
(31)
where
(32)
(33)
i.e.
(36)
(35)
RISTANIEMI AND JOUTSENSALO: CODE TIMING ACQUISITION FOR DS-CDMA IN FADING CHANNELS 905
IV. ANALYSIS
A. Performance Measures
In this paper, we consider delay estimation only in a parallel
manner. This is because the goal is to compare the performance
of the methods without considering any threshold setting. Max-
imum selection is used to determine the delay, i.e., the strongest
peak from the delay spectrum is selected. The definition of ac-
quisition is the case where any multipath component is found.
The performance measure, probability of acquisition, is defined
M
Fig. 4. Typical fading process. The lines indicates the real and imaginary parts
as of the time-varying path attenuation factor for = 500 symbols.
# of acquisitions
(37) The third performance measure is root mean-square error
# of trials
(RMSE). It measures the accuracy of the delay estimation
The other performance measure is mean acquisition time, according to
which measures the average time spent in the acquisition
process. This issue is profoundly considered in [20]. The
acquisition scheme is illustrated in Fig. 3. A preamble of (41)
symbols is sent, which is processed in the receiver’s acquisition
mode. In case of misacquisition, a request for the new preamble where is the true delay, and is the number of realizations.
is sent, which takes a time of symbols. If the delay estimations are unbiased, the RMSE values can be
Suppose a while that it takes no time to request a new pre- directly compared to the Cramer–Rao bound (CRB). CRB is the
amble, i.e., . Then the average time needed for acquisi- lower bound for the variance of any unbiased estimator. The an-
tion is alytical forms of CRBs in different type of channels are derived,
e.g., in [9], [11], and [16].
B. Numerical Experiments
We compare the methods in the downlink environment with
the Rayleigh fading channel. The data rate is assumed to be
kbit/s, carrier frequency GHz, and mobile speed
(38) km/h. This results in normalized Doppler shifts at most
. According to Jakes’ doppler
In practice, however, , since each misacquisition pays power spectrum [18], and related autocorrelation function of the
an additional price of symbols by which a new preamble is channel, the coherent integration interval of symbols
requsted. This additional time is on average is reasonable, and is used from now on unless otherwise stated.
The typical fading process at the symbol level is seen in Fig. 4.
The methods for delay estimation are: DC-MF (24),
DC-MUSIC (29), ADC-MF and ADC-MUSIC (Table I),
MUSIC [9], CMOE [17], and conventional noncoherent and
coherent MF [4]. Any parameters such as time-varying path
strengths, and codes of the interfering users are not assumed
to be known. Only the code of the desired user is known. In
addition, a preamble is available for the desired user
. Gold codes of length are used. Signal-to-noise
(39) ratio (SNR) in the chip-MF output is always 10 dB. Both
frequency-selective and flat fading channels are considered. In
The mean acquisition time is thus frequency-selective fading channels the number of resolvable
paths is either or . The number of users is varied
(40) from to . Therefore, the total number of sources
906 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 49, NO. 5, MAY 2001
Fig. 5. Probability of acquisition as a function of the number of users in an Fig. 7. Probability of acquisition as a function of MAI in an equal energy
K= 10 users. The
M
equal energy two-path Rayleigh fading channel. All the interfering users are two-path Rayleigh fading channel. The system included
10 dB stronger. The length of the preamble is = 400, and the average SNR length of the preamble isM = 400, and the average SNR is 10 dB.
is 10 dB.
Fig. 9. Probability of acquisition as a function of the preamble length in an Fig. 11. Relative number of acquired paths in L = 3 path Rayleigh fading
equal energy two-path Rayleigh fading channel. The system included K = 10 channel as a function of the preamble length. The system included K = 2
users, and all the interfering users are 10 dB stronger. The average SNR is 10 dB. users, and the random codes were of length C = 7. Average SNR was 10 dB.
TABLE II
NUMBER OF CASES A CERTAIN PATH IS ACQUIRED IN A K = 2 USER SYSTEM
WITH RANDOM CODES OF LENGTH C = 7. PREAMBLE LENGTH IS M = 500
SYMBOLS, AND AVERAGE SNR IS 10 dB. FRACTIONAL PARTS ; i = 1; 2; 3;
ARE UNIFORMLY DISTRIBUTED OVER (0; 1)
are available [17]. This is because the estimation for the inverse
of the autocorrelation matrix thus becomes more accurate.
However, in this setup the strong MAI is too much for CMOE,
even if as many as symbols are available. The fact
Fig. 10. Mean acquisition time as a function of the preamble length in an equal
energy two-path Rayleigh fading channel. The system included K = 10 users, that the proposed methods are asymptotically invariant with re-
and all the interfering users are 10 dB stronger. The average SNR is 10 dB. spect to the interfering users and noise can be predicted from
the figure, in which the probability of acquisition seems to tend
fading channel. The length of the preamble is , and to 1. The figure also tells the fact that the performance of the
the average SNR is 10 dB. proposed methods depends essentially on the length of the pre-
Fig. 7 shows the achieved probabilities of acquisition as a amble, because the goal is to “average interference out.” For this
function of MAI per interfering user. One can see that real dif- reason, it is understandable that the methods fail if the preamble
ferences are noticed after MAI of 8 dB, after which DC-MF is is too short. However, this is also the problem for the reference
able to retain the performance the best. With MAI of 12 dB per methods, which may suffer even more.
interfering user, DC-MF still gives a probability of acquisition In Fig. 10 the mean acquisition times in symbols are plotted.
equal to 0.915, while coherent MF drops down to 0.608. Fig. 8 Due to reliable estimations, the acquisition time for DC-MF is
shows the corresponding mean acquisition times. usually no more than the length of the preamble.
3) Setup C: Varying Number of Symbols: In this experiment 4) Setup D: Search of Multiple Paths: In this experiment we
we study the effect of increasing the length of the preamble. demonstrate the ability of DC-MUSIC to find the most paths.
The system includes users and all the interfering users users are given random codes of length in
are 10 dB stronger. The achieved probabilities of acquisition are path Rayleigh fading channel. The paths have 5 dB differences
seen in Fig. 9. Again, noncoherent MF fails in the near–far sce- in power. Fig. 11 shows the average amount of acquired paths
nario, which is predictable, as well as MUSIC due to the high as a function of the preamble length. It is seen that DC-MUSIC
system load. CMOE usually performs better as more symbols always found the most paths. Quite remarkably, DC-MUSIC
908 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 49, NO. 5, MAY 2001
• DC-MUSIC performs well when the processing gain is from which the first equation in (30) follows. The derivation of
small; the second one is entirely similar and is omitted.
• an additional procedure can be attached to both methods,
improving the estimation accuracy by a negligible in- APPENDIX C
crease in computation.
These features were verified by numerical experiments in down- A. Derivation of (31)
link environment, making the proposed methods attractive op- To shorten notations, define and
tions to be attached for the next generation mobile receivers.
. Taking square roots on both
Future research in the context of the paper include, e.g., more
sides of (30), we have
efficient interference suppression mechanisms, which are vital
especially when short preambles are used. Also the possibility (48)
of using the methods for tracking purposes via decision feed-
(49)
back is our ongoing subject.
which is a system of equations with two unknowns, and
APPENDIX A . Dividing (48) by (49) we obtain
A. Derivation of (19)
(50)
First, from (8) we have
By simple algebraic manipulations it leads to
..
.
(51)
[13] Z.-S. Liu, J. Li, and S. L. Miller, “An efficient code-timing estimator for Tapani Ristaniemi (M’00) was born in Kauhava,
receiver diversity DS-CDMA systems,” IEEE Trans. Commun., vol. 46, Finland, in July 1971. He received the M.Sc. degree
pp. 826–835, June 1998. in pure mathematics, the Ph.Lic. degree in applied
mathematics, and the Ph.D. degree in telecommuni-
[14] M. H. Zarrabizadeh and E. S. Sousa, “A differentially coherent PN code
cations from the University of Jyväskylä, Jyväskylä,
acquisition receiver for CDMA systems,” IEEE Trans. Commun., vol. Finland, in 1995, 1997, and 2000, respectively.
45, pp. 1456–1465, Nov. 1997. He is Senior Research Scientist at the Department
[15] C.-D. Chung, “Differentially coherent detection technique for direct-se- of Mathematical Information Technology, University
quence code acquisition in a Rayleigh fading channel,” IEEE Trans. of Jyväskylä, Jyväskylä, Finland. His research in-
Commun., vol. 43, pp. 1116–1126, Feb./Mar./Apr. 1995. terests include signal processing and radio resource
[16] S. Parkvall, “Near-Far Resistant DS-CDMA Systems: Parameter Esti- management for wireless communication systems.
mation and Data Detection,” Ph.D. dissertation, Royal Inst. of Technol.,
Oct. 1996.
[17] M. Latva-Aho, “Advanced Receivers for Wideband CDMA Systems,” Jyrki Joutsensalo (M’94) was born in Kiukainen,
Doctoral dissertation, Acta. Univ. Ouluensis, C 125, Sept. 1998. Finland, in July 1966. He received the diploma en-
[18] J. D. Parsons, The Mobile Radio Propagation Channel. London, U.K.: gineer, licentiate of technology, and doctor of tech-
Pentech, 1992. nology degrees from the Helsinki University of Tech-
[19] J. Joutsensalo and T. Ristaniemi, “Single-user synchronization in fading nology, Espoo, Finland, in 1992, 1993, and 1994, re-
channel,” Proc. IEEE Signal Processing Workshop on Signal Processing spectively.
Currently, he is Professor of Telecommunications
Advances in Wireless Communications, pp. 133–137, May 9–12, 1999.
at the Department of Mathematical Information
[20] A. Polydoros and C. Weber, “A unified approach to serial search spread- Technology, University of Jyvaskyla. His current
spectrum code acquisition—Part II: A matched filter receiver,” IEEE research interests include signal processing for
Trans. Commun., vol. COM-32, pp. 550–560, May 1984. wireless radio communications and data networks.