You are on page 1of 5

Spreading Sequences for Zero-Forcing DS-CDMA Multiuser Detectors

Harald Elders-Boll, Axel Busboom, and Hans Dieter Schotten Institut fiir Elektrische Nachrichtentechnik, Aachen University of Technology (RWTH) Melatener Stralje 23, D-52056 Aachen, Germany Tel/Fax: +49 [241] 807678 / 8888196, E-Mail: elders@ient.rwth-aachen.de
ABSTRACT
In the past, different multiuser detectors for asynchronous code-division multiple-access communications have been proposed, many of them may be characterized as zero-forcing detectors, e.g., the decorrelating detector. We will show that linear interference cancellation schemes are asymptotically zero-forcing which means that they are equivalentto the decorrelating detector if the number of stages approaches infinity. These detectors have been found to be superior to the conventional matched filter detector. However, the design of spreading sequences optimized especially for these receivers has not been considered up to now. Usually, spreading sequences are designed to have a low peak correlation parameter. Pursley [l] has shown that the average interference parameter is an important design parameter since it is related to the average signal-to-interference ratio of the conventional receiver. In this paper, we consider the construction of spreading sequences for zero-forcing multiuser detectors that are optimal in sense of the performance and the near-far resistance. It is shown that sequences with a low AIP are near-optimal. This, again, stresses the importance of the AIP for the design of spreading sequences for CDMA systems employing any kind of receiver. Numerical examples indicate that by using optimized sequences the average signal-to-noiseratio (SNR) can be improved by about 1-2 dB for lengths of interest in applications. In this paper, we consider spreading sequences that are near-optimal for zero-forcing multiuser detectors in sense of the bit-error performance and the near-far resistance. The paper is organized as follows: In the next section, the correlation parameters are defined. In Section 111, we briefly review the main results of [3] that are of interest throughoutthe paper. In Section IV, we shown that any linear interferencecancellation scheme is asymptotically equivalent to the decorrelating detector. In Section V, we investigate the properties of optimal spreading codes for zero-forcing multiuser detectors. Numerical examples are presented in Section VI and the conclusions are given in Section VII.

11. CORRELATION PARAMETERS


Let s k (n) s j (n)be two sequences of length N . Then, the and discrete-time aperiodic cross-correlation function is defined as
f

N--1-v

n=O

For a family of K sequences, the total interferenceparameter (TIP) for user i is

I. INTRODUCTION
We consider asynchronous direct-sequence code-division multiple-access (DS-CDMA) communications. Since the complexity of the optimum multiuser detector is exponential in the number of users [2], a number of suboptimal, linear multiuser detectors have been proposed, many of them may be characterized as zero-forcing detectors, e.g., the decorrelating detector proposed by Lupas and Verd6 [3] and the zeroforcing decision-feedback detector proposed by Duel-Hallen [4]. In this paper, we will show that linear interference cancellation schemes, e.g., the linear multi-stage serial canceller of Johansson and Svensson [5], are asymptotically zero-forcing, which means that they are equivalent to the decorrelating detector if the number of stages approaches infinity. The performance in terms of the bit-error-rate and the near-far resistance of these detectors has been found to be superior to the conventional matched filter detector. However, the design of spreading codes optimized especially for these receivers has not been considered up to now. Usually, spreading codes are designed to have a low peak correlation parameter. Pursley [l] has shown that the average interference parameter is an important design parameter since it is related to the average signal-to-interferenceratio of the conventional receiver.
k#i

where the interference parameters are defined by


f N-1

and their averages are defined by

Usually, the average interference parameter (AIP) is used as a measure for the average system performance of the conventional detector: AIP = 2/40)

+ p(1).

(5)

111. DECORRELATINGDETECTOR
In this section, we review the main results of [3, 41 that are of interest throughout the paper. We consider a DS-CDMA

0-7803-3871-5/97/$10.00 0 1997 IEEE

53

system with K users. Let the receiver input signal be:

A. Limiting Case M

+ CO

k=lm=-M

Taking the Fourier-transform of the discrete-time vector sequence given by equation (1 1) and letting M go to infinity, one obtains

where n(t)is white Gaussian noise with power spectral denis is sity cr2, a k ( m ) E &1 the binary data signal the amplitude, S k (t)is the spreading signd and 7 k is the time delay of the k-th user. The data symbols are supposed to be equally probable and mutually independent. The spreading signals S k ( t )are given by

',d m

Y ( f ) = S(f)[WAI(f) +",
where the hermitian matrix S(f) is given by

(13)

s ( j )= ~ ~ ( 1 + R(O) ~ ( 1 ) e22.rrf) + e--z2xf

(14)

N-1

and [WA](f) is the Fourier-transform of the sequence wa(m) = { [ J m a l ( m )..., J m w ( m ) ] } . Then, , the decorrelator becomes a K-input K-output linear timeinvariant filter with transfer function

where rect(t) is defined as rect(t) = 1 for 0 5 t 5 T, and rect(t) = 0 otherwise. T, is the duration of a chip and T = NT, is the code period. We suppose that the spreading signals have unit energy, e.g., that the spreading sequences S k ( n )are uniform sequences with ISk ( n ) = 1. I The received signal is first fed into a bank of K filters matched to the users spreading sequences and sampled at time instances T k mT:

The assumption that the inverse exists is well justified. It should be noted that S (f)is positive definite in this case. The signal before the sign-decisionis given by

X ( f ) = G(f)Y(f) = [WAl(f)+ Ne(f),

(16)

Yk(mT

+7 k ) =

mTf Ti-Ti,

~ ( t ) s k ( - mT t

-7k)dt.

(8)

mT+n

where Ne( f ) is the Fourier-transform of the filtered Gaussian noise vector sequence with its covariance matrix sequence given by N(1) = E[n,(m)n,(m Z)] = c y S-l(f) +
0

For simplicity and without loss of generality, let us assume an ordering on the time delays T k such that 05 T I 5 7 2 . . .5 TK < T. It is convenient to introduce the following matrix notation. Let y (m)be the output sequence of the bank of matched filters, a(m) the data sequence. Both are sequences of Kdimensional column vectors. Define the K x K signal crosscorrelation matrices R(m) whose entries are given by
R k J (m) =

df. (17)

1,SE
00

It can be seen from equation (16) that the decorrelator eliminates the multiuser interference by "inverting" the channel matrix. Therefore, the decorrelator is a zero-forcing detector. As the zero-forcing equalizer for single-user intersymbol interference channels, the decorrelator enhances noise. The noise power of the k-th user's noise component is
N k

(t - T k ) S J (t - mT - 7 j ) d t .

(9)

=E[n,(m)n,(m)]kk =d

Then, R(1) is an upper triangular matrix with zero diagonal and

l1

S-'(f)kk

df

> F2.

(18)

Therefore, the probability of error for the m-th bit of user k is given by

R(m)=OV\ml > 1, R(m)=R(-m)H, R(0)=R(O)H. (IO)


Moreover, let W(m) = d i a g ( [ d m , ..., With this notation and a(-A4 - 1) = a(-&' 1) = 0 the matched filter output sequence can be expressed as
y(m)

d-1). +

with the asymptotic efficiency


qk=C

= R(-l)W(m+l)a(m+l)

+ R(O)W(m)a(m)
(11)

2N-1
k

R(l)W(m-l)a(m-1)

+ n(m),

(20)

where n(m)is the matched filter output noise vector, with its autocorrelation matrix given by

E[n(m)nH(m+ j ) ] = c2R(j).

(12)

'The assumption of a binary data signal is no resmcnon conceming our analysis of the decorrelating detector. It only simplifies the detection device used in the receiver. With a binary data signal a simple sign-decision device can be used.

and the error function Q ( x ) = 1 e$ Jza dy. The asymp6 totic efficiency was defined in [3] as the ratio of the the actual energy and the effective energy in the limit cr + 0, where the effective energy is defined as the energy that a singleuser system would require to achieve the same bit-error-rate in presence of the same Gaussian background noise. Thus, the asymptotic efficiency is a measure for the performance degradation due to multiple-access interference.

54

IV. LINEAR INTERFERENCE CANCELLATION


A filter with transferfuncton given by equation (15) can be realized in various ways [3, 41. Moreover, there are numerous methods known for matrix inversion. Iterative methods can be computationally very efficient if the iteration converges fast, or the inverse has only to be approximated with relatively low accuracy. The latter is the case in CDMA multiuser detection.. A linear interference cancellation scheme is one where no hard-decisions, e.g., sign-decisions,are made in one of the stages, but there is only one decision device after the interference canceller. In this section, we will show that a linear interference cancellation scheme is asymptotically equivalent to the decorrelating detector since it can be described as an iterative method to invert the channel matrix. In [6] this was shown for the special case of multi-stage serial interference cancellation for synchronous CDMA. Here, we consider asynchronous CDMA and various types of cancellation schemes. We start by trying to find a solution for the system of linear equations given by (13) for the noiseless case:

The convergence of the iteration scheme can be speed up by a scheme that is serial in-between the users because the parallel treatment of the users in the Jacobi iteration does not use the most recently available information. In i-th stage of the interference cancellation for user k, the estimates of the signals of the users 1 , . . . ,k - 1 of that stage can be used instead of the estimates of the previous stage. This is the idea of the GauB-Seidel iteration. Thus, instead of (25) we now have
K
Z" L

(m) = Y k (m) -

Rkl(o)zf )'(

f Rkl @)zf '(

- 1)

l=k+l
k-1

1=1

Rkl(-1)%f+l(m+1) &()f'm. (26) lOZ+()

In matrix notation, we obtain in the time-domain


xz+'(m)

(27) Y(m)-Riy(-l)xi(m+l)- Ru(o>xi(m)- Ru(l)xi(m-l) -RL(-l)xi+'(m+l)- R ~ ( O ) X ~ + ' ( , ) - R ~ ( ~ ) X ~ + ~ ( ~ - ~ ) ,

Let us consider the following decomposition of the channel matrix S(f ):

with the Fourier-transform given by

xi+l(f)=Y(f)-Su(f)xi(f)-SL(f)Xi+l(f). (28)
Since S(f)is positive definite and can be written as S(f) = S u ( f ) I SF(f),the GauB-Seidel iteration converges to Xm(f) = S(f)-'Y(f) [7]. The GauB-Seidel iteration defines a linear multi-stage serial interference cancellation )scheme. Serial cancellers have been proposed e.g., in [lo, 51. ~ From the above, it is obvious that a similar calculation for any other sensible linear multi-stage interference cancellation scheme would reveal its asymptotic equivalence with the decorrelating detector. It should be noted that in principle the iteration defined in (23) or (28) works on infinitely long sequences X i This would prevent a realizable implemen(f). tation. However, in each step of the iteration (25) or (26) the availability of future estimates Z (m U ) of the previous (or : the actual) stage is assumed only for U = 1,2, depending on the ordering of the users. Thus, the iteration can be realized by introducing an appropriate amount of delay between the single stages 2. The cancellation can be done at the correlator input or at its output. The former creates the interfering signals and subtracts them from the total received signal ~ ( t ) . The latter removes the contributions of the interference to the correlator output. Of course both methods are mathematically equivalent 193.

where S L ( ~is lower triangular with zero diagonal, So(f) ) is a diagonal matrix and S u ( f ) is upper triangular with zero diagonal. Note that So(f) = I because the code sequences are normalized to have unit energy, and S L ( ~ = s ~ ( ) since S(f) is hermitian. The simplest iterative scheme is the Jacobi iteration. The transition from X i ( f ) to Xi+' is given by (f)

+ +

Xi+l(f) Y(f) - ( S L ( f ) + S U ( f ) ) X i ( f ) =
If llS~(f) Su(f)II< 1 for any induced matrix norm the Jacobi iteration convergesto XW(f) [WA](f)= S ( f ) - ' Y (f) = for any Xo(f)[7]. Thus, it is asymptotically equivalent to the decorrelating detector. However, the iteration is not carried out in the frequency domain but in the time domain. The inverse Fourier-transform of (23) using (14) yields

xisl(m) = y ( m ) - R ( - l ) x i ( m + l ) (24) - (R(0) - I)xi(m) - R ( l ) x i ( m - 1 ) .


Thus, for the k-th component of x i i l ( m ) corresponding to the k-th user we have
k-1

V. OPTIMUM SEQUENCES
In this section, we derive spreading sequences that are nearoptimum for any multiuser detector equivalent to the decorrelating detector given by (15). The derivation can be viewed as the multi-dimensional analogy of the derivation of the spectral properties of good binary aperiodic invertible sequences in [ 1 I]. The noise power of the users given by (18) and their
'An ordering on the time delays ?-k such that T > 7 2 . . ?-K 2 0 leads 1 to a realization of the GauRSeidel iteration with minimum delay. However, the fastest convergence can be achieved by an ordering of the users according to their signal strength.

1=1

K
l#k

l=k+l

It can easily be seen that (25) describes the linear version of the interference cancellation scheme proposed in [8,9]. Since the signals of all users are treated in parallel in each stage, (25) describes a linear multi-stage parallel interference canceller.

.>

55

asymptotic efficiency given by (20), are respectively the arithmetic and geometric mean of the diagonal elements of N. Our aim is to minimize the average noise power, that is to find sequences which minimize the noise enhancement factor E defined as
E

Computing the trace of the identity

D-' ( f ) 1 (D(f ) -I)2D-1 (f)f 2 I - D(f )


gives tr(D-'(f))
= K+tr((D(f)-I)2D-1(f))

(37)

-E$ = l 11 tr(S-l(f)) df > 1 K


1
K

(29)

k=l

K tr( (D(f)- I)2)- (38)


Xmin(f)

'

where tr(A) denotes the trace of matrix A. Since S(f)is a positive definite hermitian matrix there exists an eigenvector decomposition

where X m j n ( f ) is the minimal eigenvalue of D ( f ) . Finally, using (29) and (32) we obtain an upper bound for E given by
fl
1

where T(f) is an unitary matrix and D(f) = diag([XI ( f ) ,...,XK(f)]) is the diagonal matrix of the realvalued eigenvalues of S(f).Because the trace of a matrix is invariant with respect to a similarity transform given by (30), we can write

A direct minimization of E by selecting optimal sequences according to (29) or (31) would be very difficult since it would involve the computation of either the inverse or the eigenvalues of S(f).Therefore, we try to find an optimization critenon that does not require such computational burden and is directly related to the elements of S(f). Since we have unit energy sequences, tr(S(f)) =tr(D(f)) = K . Therefore, the average noise power Nd takes on its minimum value Nd = 1 if and only if D(f) = I. This would be the case if we had uncorrelated sequences and consequently no interference which is impossible for asynchronous CDMA systems since it would mean that we had orthogonal sequences irrespective of the users' time delays. However, near optimal sequences can be expected to have a reasonably small meansquare deviation

It will be shown by numerical examples in the next section that the left part of this bound is usually quite tight. The right part is only valid if K6 < 1. Nevertheless it shows that the minimization of 6 is sufficient to obtain at least near-optimal sequences with respect to the minimization of E. From (36) it can be seen that 6 is the variance of the multiple-access interference of the conventional detector. Until now we considered only one fixed set of time delays. However, the time delays itself are random processes. Thus Z, the average of (29) with respect to the users' time delays, is a good measure for the average system performance. Taking the average of the variance of the multiple-access interference with respect to the users' time delays yields the average of the TIP

where I I . I

is the Frobenius norm given by

Therefore, spreading sequences which have a low AIP are at least near-optimal with respect to the minimization of $. It should be noted that the computation of the ALP is relatively simple because only the autocorrelation values of the spreading sequences are needed. Since the direct minimization of F is computationallytoo expensive to be a suitable criterion for the construction of the spreading sequences, the minimization of the AIP is the best way to design spreading sequences for zero-forcing multiuser detectors. This is also valid for minimum mean square error (MMSE) multiuser detectors because the MMSE detector reduces to the zero-forcing detector as g2 -+ 0. This result stresses the importance of the AIP for the design of spreading sequences for DS-CDMA systems employing any kind of receiver.

VI. NUMERICAL EXAMPLES


In this section, we compare the noise enhancement factors of different spreading sequences. The AIE' of sequences designed with respect to the periodic peak correlation parameter cannot differ much from 2N2 [12, 131. Therefore, 3 of these sequences is given by (35)
-

Using (30), we can write

since the Frobenius norm is invariant under an orthogonal transform. Then, using the structure of S ( f ) we get

2K-1 6%-3 N -

(41)

In Table 1, the noise enhancement factors of different spreading sequences of length N = 63 are compared. The wellknown Gold and small Kasami families are compared with families of the same size K which have been optimized with

56

K
AIP/2N2 6

Kasami 8 0.909 0.101

AIP-opt. 8 0.125 0.014

Gold 31 0.983 0.468

AIP-opt. 31 0.494 0.235

REFERENCES
[ 11 M. Pursley, Performance evaluation for phase-coded spread-spectrum multiple-access communication - part I: System analysis, IEEE Trans. Commun., vol. COM25, pp. 795-799,1977. [2] S . Verdii, Minimum Probability of Error for Asynchronous Gaussian Multiple Access Channels, IEEE Transactions on Information Theory, vol. 32, no. 1, pp. 85-96, Jan. 1986. [3] R. Lupas and S. Verdli, Near-Far Resistance of Multiuser Detectors in Asynchronous Channels, IEEE Trans. Commun., vol. 38, no. 4, pp. 496-508, Apr. 1990. [4] A. Duel-Hallen, A Family of Multiuser DecisionFeedback Detectors for Asynchronous Code-Division Multiple-Access Channels, IEEE Trans. Commun., vol. 43, no. 21314, pp. 421-434, Feb. 1995. [5] A. Johansson and A. Svensson, Multi-stage Serial Interference Cancellation in multi-rate DS-CDMA, in Proc. IEEE PIMRC, Toronto, Canada, 1995, pp. 965969. [6] K. Jamal and E. Dahlman, Multi-stage Serial Interference Cancellation for DS-CDMA, in Roc. IEEE VTC, Atlanta, USA, 1996, pp. 671-675. [7] G. Golub and C. van Loan, Matrix Computations, London: The Johns Hopkins University Press, 2nd edn., 1989. [SI M. Varanasi and B.Aanzhang, Multistage Dectection in Asynchronous Code-Division Multiple-Access Communications, IEEE Trans. Commun., vol. 38, no. 4, pp. 509-519, Apr. 1990. [9] Y, Yoon, R. Kohno, and H. Imai, A Spread-Spectrum Multiaccess System with Cochannel Interference Cancellation for Multipath Fading Channels, IEEE J. Select. Areas in Commun., vol. l l , no. 7, pp. 1067-1075, Sep. 1993. [IO] P. Pate1 and J. Holtzman, Analysis of a Simple Successive Interference Cancellation Scheme in a DS/CDMA system, IEEE J. Select. Areas in Commun., vol. 12, no. 5, pp. 1713-1724, Jun. 1994. [ 111 J. Ruprecht and M. Rupf, On the Search for Good Aperiodic Binary Invertible Sequences, IEEE Transactions on Information Theory, vol. 42, no. 5, pp. 1604-1612, Sep. 1996. [121 K. Ktirkkainen, Mean-square cross-correlation as a performance measure for spreading code families, in IEEE ISSSTA, 1992, pp. 147-150. [13] H. Schotten, B. Liesenfeld, and H. Elders-Boll, Large Families of Code-Sequences for Direct-Sequence CDMA Systems, AEU, vol. 49, no. 3, pp. 143-150, 1995.

.,
47)

(I

0.029

0.013

11 0.074 I

0.063

Table 1: Mean and standard deviation of the noise enhancement E and the asymptotic efficiency 7 of families of spreading sequences.

respect to the AIP 3. From the original Gold family 31 sequences were picked at random to obtain a smaller family. The values of Z, 7, o(E), ( ~ ( 7for mean and standard devia) tion of E and 7 were obtained from a statistic of 100000 randomly chosen delay constellations. The values obtained from the bound (39) are also depicted. Because the delays were multiples of the chip duration, (40) has to be replaced by
-

s=-- K-1
N

AIP
2N2

Thus, if ( K - 1)/N is small, the difference in E and 77 will be small, too. E.g., the SNR of the AIP-optimized sequences is only 0.4 dB better than the SNR of the Kasami sequences (cp. Table 1, families with K = 8). However, if ( K - l ) / N is large the difference is considerable, e.g., the SNR of the AIP-optimized sequences is 1.55 dB better than the SNR of the Gold sequences (cp. Table 1, families with K = 31).

VII. CONCLUSIONS
We have shown that linear multi-stage interference cancellation schemes are asymptotically zero-forcing which means that they are equivalent to the decorrelating detector if the number of stages approaches infinity. Particularly, a parallel linear interference cancellation scheme is equivalent to a Jacobi iteration, whereas a serial linear canceller corresponds to a GauB-Seidel iteration. We also considered the problem of constructing spreading sequences for zero-forcing multiuser detectors which minimize the average noise power at the detector output. It was found that the minimization of the average noise power is too expensive to be a suitable criterion for the design of the spreading sequences. It was shown that spreading sequences with a low A I P are at least near-optimal. Therefore, the minimization of the AIP is the best way to design spreadingsequences for zero-forcingmultiuser detectors. The fact that this is also valid for MMSE detectors, stresses the importance of the AIP for the design of spreading sequences for DS-CDMA systems employing any kind of receiver. Numerical examples indicate that by using optimized sequences the average SNR can be improved by about 1-2 dB
for lengths of interest in applications.
3These sequences may exhibit very high autocomelation sidelobes.

57

You might also like