Vous êtes sur la page 1sur 10

Optimization in Multicarrier (MC-) CDMA Systems

Julius Kusuma and Laurent El-Ghaoui and Kannan Ramchandran Department of Electrical Engineering and Computer Sciences, University of California, Berkeley. CA - 94720. fkusuma, elghaoui, kannanrg@eecs.berkeley.edu

1 Introduction
Diversity is one of the most important aspects of wireless systems design 1]. There are many ways to achieve diversity: time/delay diversity 2], frequency diversity, code diversity, and spatial diversity 3, 4]. The design problem now becomes how much diversity should the communication system has within some degrees of freedom to achieve some performance requirement? One notable technique is multiuser detection 5] where multiple access interference (MAI) resulting from the correlation of the spreading codes of di erent users is explicitly taken into account in the receiver design, rather than just as background noise. Another technique is to use adaptive antenna arrays 6] and power allocation between the antenna elements or beamforming. These techniques are used to enhance the desired signal's SNR and suppress MAI 7]. Another engineering approach is to use Multicarrier CDMA (MC-CDMA) 8]. MC-CDMA systems can be categorized into two di erent categories: a combination of OFDM and CDMA, and MC-DS CDMA. In the former, a spreading sequence is rst serial-to-parallel converted, and each chip is used to modulate a carrier. In the latter scheme, multicarrier signalling is used to modulate M DS-CDMA waveforms. It has been shown that both schemes show similar fading mitigation performance over a frequency-selective channel. However, the latter system requires only M adaptive gain ampli ers in the maximal ratio combiner, which greatly simpli es receiver design 9]. This work will consider only the latter system. An overview of the di erent schemes for combining multicarrier and CDMA is available in 10]. In this approach, di erent users are assigned di erent spreading sequences ak . After spreading, the signal is modulated into M di erent orthogonal frequency bins. While the signals at di erent frequency tones are orthogonal by design, we are still faced with the problem of maximizing performance in the face of correlation between the spreading sequences of di erent users. With this approach, we can assign di erent weights ck to the di erent frequency tones of the di erent users depending on the correlation structure and channel pro le. This can be viewed as a \water- lling" optimization where you put more energy into the \better" channel. Optimization over \cross-coupled" MIMO (Multiple-Input, MultipleOutput) channels have been considered in 11], and in our scenario this cross-coupling is exactly the cross-correlation between the di erent spreading codes ak . MC-CDMA signalling scheme has several desirable attributes: robustness to fading, narrowband interference suppression and robustness to fading, without requiring an explicit RAKE receiver or an interence suppression lter 9]. This scheme also operates at a lower chip rate compared to the single-carrier CDMA system, thus greatly reducing synchronization requirements. In addition, there are fast, e cient (FFT and iFFT) algorithms for the creation of narrowband multicarrier signals 12]. 1

In addition to this, by using optimal power allocation / control the MC-CDMA system can be near-far resistant. The near-far problem arises from the fact that di erent users may have vastly di erent received signal power at the base station, often up to 70 dB of received power di erence. This makes it di cult to guarantee a certain received Signal-to-Interference-and-Noise-Ratio (SINR) for the users far away from the base station. In this work we consider a system where there is a base station receiving signals from multiple mobile users. There are two optimization problems: determining the optimal weighting factors at the receivers wk and the optimal power allocation at the transmitters ck as shown in Fig. 1. The goal of the optimization of the transmission power is two-fold: P 1. Minimize total transmitted power K=0 jck j2 in the network. k 2. Constrained by a certain SINRk requirement of user k. Every user k = 1; : : :; K generates a stream of data symbols bk . The data symbols bk;j are random variables and assumed to be normalized so that E jbk;j j2] = 1. The assumption that these data symbols are independent and uncorrelated is not necessary in this context: each user can apply di erent modulation and coding schemes according to the type of information sent and performance requirements. Therefore each user can change its modulation and coding scheme without a ecting other users. Each user k is also assigned a di erent signature/spreading sequence ak of length N , the spreading factor on each carrier. The chip period Tc is then related to the symbol period T by T = NTc . Therefore each symbol bk;i is multiplied by ak = (ak;i ; ak;i+1; : : :; ak;i+(N ?1)). The same signature sequence is used to modulate each of the M carriers of the kth user. The transmitted signal can then be written as:
8 1 < X bk;bi=N cak;i x(t) = ck;m : i=?1 m=1

2 System Model

M X

9 = (t ? iTc ); ej!m t

(1)

where ck;m is the complex weighting/gain factor for the mth carrier of the kth user to adjust the amplitude and phase of the mth carrier. This structure is shown in Fig 1. The receiver will be a lter bank of M chip correlators, with wk;m complex weighting factors for the M correlator outputs. The M outputs will then be added together to provide an input to the detector. The receiver's optimization objective is to select wk to maximize the signal-to-interference ratio at the output for every user. The receiver structure is shown in Fig 2.

3 Channel Model
We assume that the overall channel is a frequency selective fading channel. But we assume that by choosing the appropriate M and the (t) with a su ciently narrow bandwidth, each carrier is subjected to independent slowly-varying at Rayleigh fading. The received signal is also subjected to AWGN with power spectral density N0 . In order to make this assumption valid we need to consider a few channel parameters in our design 9]: 2

Figure 1: MC-CDMA transmitter model

Figure 2: MC-CDMA receiver model

1. Delay spread Tm : the di erence between the propagation delay of the longest path minus the shortest path. 2. The coherence bandwidth ( f )c : if the coherence bandwidth is larger than the bandwidth of the transmitted signal, the received signal is not distorted appreciably by the multipath e ect. This is the phenomenon called at fading 5]. Otherwise the signal will be distorted in such a way that causes Inter-Symbol Interference (ISI). This is called it frequency-selective fading. Another intuition is that signals which frequencies di er by an amount greater than ( f )c will be subjected to independent fading. The coherence bandwidth is related to the delay spread by ( f )c = 1=Tm. We will choose M such that: 1. There is no frequency-selectivity in the fading of each sub-band, i.e., Tm =(MTc) 1 where Tc is the chip/code period. 2. The di erent sub-bands are subjected to independent fading, i.e., BWm ( f )c . These two conditions can be written as: where is the guard band between adjacent multicarrier signals. We will select N the spreading factor such that the transmitter signals of di erent sub-bands do not overlap. To make our analysis tractable we will assume perfect carrier and symbol synchronization at the receiver. The received signal is then:
8 1 < X r(t) = bk;bi=N cak;i ck;m : m=1 i=?1 k=1

Tm Tc

M (1 + ) Tm Tc

(2)

K M XX

9 = (t ? iTc ); ej!m t k;m + n(t)

(3)

where k;m is the channel fading coe cient subjected to the signal from the mth carrier of the kth user. The assumption of at fading for each sub-band and independent fading across the di erent sub-band has greatly simpli ed our analysis. With the assumption of perfect synchronization and independent fading, we can now focus only on the transmitted power ck and received power at the output of the lter banks denoted dk . The vectors dk , representing the desired signal component, is related to ck by the matrix Ak:

dk = Ak ck

(4)

where Ak is an M M diagonal matrix whose diagonal entries are k;m . The diagonality is due to the fact that the di erent frequency tones are chosen to be orthogonal. This would have been all that we need to consider in our optimization if we were considering only one user with one signature/spreading sequence.However, in a multi-user scenario, we have to take into account the interference from the other users, MAI. To consider the e ect of MAI, we now have to decompose the outputs of the correlators into the desired and undesired components. Let zk be a vector of the outputs of the correlator lter bank of the kth user. Without loss of generality, let us consider only the output corresponding to the rst (0th) symbol:

zk = Tbk; dk} + |{z} + nk | {z


desired part AWGN 4
0

ik;j j 6=k | {z }
MAI

(5)

where ik;j = ik;j;1ik;j;2 : : :ik;j;M ], and ik;j;m denotes the interference term from the j th user to the kth user in the m th frequency tone:

ik;j;m = cj;m

j;m Tc bj;0

N ?1 X i=0

ak;i aj;i

(6) (7) (8)

For convenience we will introduce a compact notation for the desired signal part: sk = Tbk;0dk So now Eqn. 5 can be compactly written as: X zk = |{z} + |{z} + ik;j sk nk j desired part AWGN |6=k } {z MAI

4 Receiver Optimization
The objective of the receiver optimization for the kth user is then to select the weight vector wk to maximize the SINR at the total output for the kth user yk : H yk = wk zk (9) This problem has been considered in 13] for the blind scenario, and the solution is as follows. First we de ne a noise and interference correlation matrix for the kth user by considering only the noise and interference terms from Eqn. 6:

4.1 Objective

Rni;k =
=

where E :] denotes conditional expectation given the channel transfer function k;m . The identity matrix comes from the fact that the noise at the di erent frequency tones will be independent. The second term on the summation in 6 captures the correlation structure between the kth user and all other users. Rni;k is related to the total correlation matrix Rtot;k by the following: Rtot;k = E zk zH ] = E sk sH ] + Rni;k = Rsk + Rni;k (11) k k We have thus decoupled the desired signal component from the noise and interference components. One method to calculate the weighting factors wk is through MaximalhRatio Combining (MRC) 14]. i A MRC sets the gains according to the signal energy calculated by E b(k) zm and the noise in the 0 h (k) i signal calculated by var b0 zm :

2 3 X E 4nk nH + ik;j iH 5 k k;j j 6=k 1 X TN0I + N T 2 dj dH j j 6=k

(10)

p2P NT c k k c m i= h k wk;m = Rni;k ]mm var b zk;m

E b k zk;m
( ) 0 ( ) 0

( )

k;m

(12)

where Rni;k ]mm is the mth diagonal element of Rni;k . However, it is shown in 13] that this is optimal only when the noise and interference across di erent carriers are uncorrelated. For example, this is optimal when only AWGN is present and there is no MAI. In addition to this, the channel coe cients k;m have to be estimated, which is di cult when MAI is also present. Since MAI across di erent carriers / sub-bands are correlated because each user uses the same code sequence across his/her di erent subbands, the MRC will be sub-optimal. This is because this correlation is not taken into calculation in Eqn. 12. From this argument we can deduce that the optimal weight vector wk for user k will be the one that maximizes the following measure:
H H E jwk sj wk R w h i = H sk k SINRk = P H wk Rni;k wk E jwk nk + j6 k ij j
2 = 2

(13)

4.2 Blind method

Equivalently 13], we nd the vector weight that maximizes:


H E jwk zj wH R w i = kH tot;k k SINRk + 1 = h H wk Rni;k wk E jwk nk + Pj6 k ij j
2 = 2

(14)

By summoning some crazy matrix theory from 15], it can be shown that the optimal weight vector that maximizes Eqn. 14, and hence the SNR given in Eqn. 13, is given by the generalized eigenvector associated with the largest eigenvalue of the matrix pencil (Rk ; Rni;k ). We can readily estimate Rz from the outputs of the correlators zk . Now we need to estimate Rni;k . Due to the special structure of the spread-spectrum signal we can remove the desired signal component while maintaining the same statistics on the noise and interference terms 13], i.e., orthogonalize the good part from the bad parts. To do this, for each correlator output we nd a sequence ak that is orthogonal to ak . We can nd this orthogonal ^ sequence by simply changing the sign of every other chip in ak . Then the output of each correlator is put into two correlator branches, one matched to ak and the other to ak . This will orthogonalize the desired ^ signal component from the noise and interference component, and all we need to do now is to subtract the latter from the former as shown in Fig. 3. If we denote the output of the latter as ^k , we can readily z show that: i h (15) z zk Rtot;k = E ^k^H = Rni;k ^ Now the the optimal weight vector wk can be determined. In 7] it is shown that the generalized eigenvector associated with the largest eigenvalue of a matrix pencil (Rk ; RNI;k ) can also be found using a constrained stochastic gradient search algorithm similar to the Least-Mean Squares (LMS) algorithm but without a training sequence. The computational complexity is of the order M per iteration, much smaller than the calculation of the generalized eigenvectors.

4.3 Optimization solution

( t )
* exp ( j m t )

Tc a(1) i

zm

wm

a(1) i

zm

Figure 3: The mth branch of the receiver for weight factor determination

The objective of the transmitter optimization is similar to water-pouring in standard multicarrier systems, which is you want to assign more power to the better subchannels. However, in our multiaccess scenario we also have to consider the e ect of MAI on other users. In addition to this, in our multiaccess scenario we can also design the optimal spreading sequences ak user k. The objectives are as follows: P 1. Minimize total transmitted power K=0 jck j2 in the network. k 2. Constrained by a certain SINRk requirement of user k. There are two optimization variables here: 1. The weighting factors at the receivers wk . 2. The signature sequences ak . This problem has been examined and solved independently by Wong and Lok 13, 16], and Anantharam, et. al. 17]. The power allocation optimization solves the problem given above. In order to begin our analysis we will rst investigate the necessary and su cient condition for solving this problem.

5.1 Objectives

5 Transmitter Optimization

5.2 Power Allocation

in Eqn. 11. Using the matrix inversion lemma we can show that 13]:

5.2.1 Necessary and Su cient Condition We again consider the total correlation matrix Rtot;k and the noise-and-inteference matrix Rni;k as given
T 2dH R?1 dk = k tot;k
1 + T 2 dH R?1 dk N k ni;k

T 2dH R?1 dk k ni;k

(16)

The term on the left side can be called the signal-to-total-power-ratio (STR) for the kth user. The derivation of the proof is shown in 13]. The optimization solution is presented here brie y. Note that 7

STRk is a monotonically increasing function of SNRk , and 0 STRk < N . Therefore we can rewrite the optimization problem as: Minimize subject to where:
k K X k=1 H R?1 d dk tot;k k

jck j

(17) (18)

and k is the target SNR for the kth user. If the target SNR for all users is equal to then the number of users that can be supported is given by: K < M (N + ) (19) A closed form solution to (17) can be very di cult to obtain except for some very special cases: 1. AWGN Channel In an AWGN channel, there are two cases: when K M and when K > M where K is the number of users and M is the number of carriers or subchannels. In the rst case the solution is obvious: assign di erent users to di erent carriers. In the second case, the solution is given by the following formula:

= T 2(1 +k =N ) k

MNT k jdk j = N T MN ? T PK j
2 0 2 2

=1

(20)

2. Fading Channel In a fading channel, we need to incorporate the SNR requirements into a penalty function and formula a Lagrange optimization problem:

L=

K X k=1

jck j +
2

K X k=1

k ?T

dH R? dk k k
1

(21)

where is the Lagrange multiplier. This can be solved by taking the derivative of L w.r.t. ck , but an iterative approach is also presented:
K X k=1

@L ck ? @ c k ? k ? T dH Rk dk k

ck

(22) (23)

13] also considers a distributed iterative search that is not really distributed. It simply approximates the values without taking MAI into account:

ck

ck ?

k?T

dH R? dk )AH R? dk k ni;k k ni;k H ? + k ? T dk Rni;k dk


2 1 1 2 1

(24)
2

(25)

The AWGN case where K M , as expected, makes the system completely FDMA (Frequency-Division Multiple Access): di erent users are assigned only one carrier frequency/subchannel. In the fading channel case, when K M , the system also tends to be FDMA

5.2.2 Simulation Results

5.3 Optimal Sequences

References

1] D. Divsalar and M. K. Simon, \The design of trellis coded MPSK for fading channels," IEEE Transactions on Communications, vol. 36, pp. 1004{1012, September 1988. 2] J. H. Winters, \The diversity gain of transmit diversity in wireless systems with Rayleigh fading," in Proc. Int. Conf. Communications, pp. 1121{1125, May 1994. 3] B. Hochwald, T. Marzetta, T. Richardson, W. Sweldens, and R. Urbanke, \Systematic design of unitary space-time constellations," tech. rep., Bell Laboratories, Lucent Technologies, September 1998. 4] B. Hochwald and W. Sweldens, \Di erential unitary space-time modulation," tech. rep., Bell Laboratories, Lucent Technologies, March 1999. Download available at http://mars.bell-labs.com. 5] S. Verdu, Multiuser Detection. Cambridge, UK: Cambridge Univ Pr., 1998. 6] B. Suard, A. Naguib, G. Xu, and A. Paulraj, \Performance analysis of CDMA mobile communication systems using antenna arrays," in Proc. ICASSP '93, pp. 153{156, April 1993. 7] T. M. Lok, T. F. Wong, J. S. Lehnert, and M. D. Zoltowski, \A linear receiver for direct-sequence spread-spectrum multiple-access systems with antenna arrays and blind adaptation," IEEE Transactions on Information Theory, vol. 44, pp. 659{676, March 1998. 8] T. M. Lok and T. F. Wong, \Transmitter and receiver optimization in multicarrier CDMA systems," IEEE Transactions on Communications, vol. {, pp. {, March 2000. 9] S. Kondo and L. B. Milstein, \Performance of multicarrier DS CDMA systems," IEEE Transactions on Communications, vol. 44, pp. 238{246, February 1996. 10] S. Hara and R. Prasad, \Overview of multicarrier CDMA," IEEE Communications Magazine, pp. 126{133, December 1997. 11] J. Salz, \Digital transmission over cross-coupled linear channels," ATT Technical Journal, vol. 64, pp. 1147{1159, Jul.-Aug. 1985. 12] J. A. C. Bingham, \Multicarrier modulation for data transmission: An idea whose time has come," IEEE Communications Magazine, pp. 5{14, May 1990. 13] T. M. Lok, T. F. Wong, and J. S. Lehnert, \Blind adaptive signal reception for MC-CDMA systems in Rayleigh fading channels," IEEE Transactions on Communications, vol. 47, pp. 464{471, March 1999. 14] J. G. Proakis, Digital Communications, 3rd Ed. New York: McGraw Hill, 1995. 9

15] F. R. Gantmacher, The Theory of Matrices, Volume I. New York: Chelsea, 1959. 16] T. F. Wong and T. M. Lok, \Transmitter adaptation in multicode DS-CDMA systems," IEEE Journal on Selected Areas in Communications, vol. {, pp. {, { 2000. 17] P. Viswanath, V. Anantharam, and D. Tse, \Optimal sequences, power control and capacity maximization of CDMA systems," IEEE Transactions on Information Theory, vol. 45, pp. 1968{1983, September 1999.

10

Vous aimerez peut-être aussi