Distributed Transmission Based on Inference on Graphs for Cooperative Base Stations

Burak S ekerlisoy
Dep. of Electrical & Electronics Eng. Bilkent University Ankara, Turkey Email: [email protected]

Defne Aktas
Dep. of Electrical & Electronics Eng. Bilkent University Ankara, Turkey Email: [email protected]

AbstractIn this paper, the downlink of a cooperative cellular network is considered. In this network, multiple base stations are cooperatively communicating with multiple mobile stations. The downlink Multi-user MIMO system is visualized as a MIMO broadcast channel. We develop a distributed algorithm using message passing on a factor graph to form the transmit signals from the base stations so that ISI is mitigated and decoding requires a simple modulo operation at the mobile stations. The proposed algorithm is a distributed version of vector perturbation technique which maximizes the sum capacity for MIMO broadcast cahnnels. The factor graph corresponding to this problem is very loopy and using current message passing algorithms result in poor performance. The proposed algorithm uses LLL reduction algorithm as a way to force the distributed algorithm to converge. We test the algorithm under various settings for the number of BSs and MSs. We also test the algorithm using linear array and hexagonal array model for the cellular network.

I. I NTRODUCTION Downlink processing with Cooperation between Base Station is an attractive research area. The coopartive network is used to form the transmit signal vector from base stations to mobile stations. This way the intercellular interference is reduced and higher sum capacity can be achieved. This network with the assumption of innite backhaul capacity can be modeled as a MIMO broadcast channel where one base station with multiple antennas communicates with multiple mobile stations. In [1], it is shown that the capacity for MIMO BC can be achieved by applying dirty-paper coding (DPC) [2] at the transmitter side using sequential encoding. In order to avoid the complexity of DPC, practical schemes are also developed. The channel inversion [3], [4] multiplies the signal to be transmitted with the inverse of the channel matrix. In [5], the channel inversion matrix is regularized. The authors in [6], [7] propose a downlink beamforming scheme which minimizes the worst signal-to-interference-and-noise ratio (SINR). Beside the linear transmission schemes which are suboptimal, nonlinear techniques are also introduced. In [8], a THP [9] based modulo precoding technique is introduced and they show that this technique can achieve performance near capacity at high SNR. In [21] an overview of the theory and currently known techniques for multi-cell MIMO cooperation in wireless networks is presented. The classical BS downlink cooperation scheme assumes central processing unit which does all the processing for the

precoding. All the base stations in the cooperative network are connected to the central unit. This way the cellular network is visualized as a single cell MIMO system with many distributed antennas in various geographical locations. There are several drawbacks of this scheme. First of all, the central unit is a single point of failure. Any malfunction in the central unit will affect the whole cellular network. In addition, this scheme is not suitable for the expansion of the network. As the network expands, the delays of the central processing unit will be higher as a result of the growing complexity. At this point, the idea of distributing the processing among the base stations in the cellular network is more feasible. In [10], the authors investigated the downlink of cooperative network in a distributed manner. The network is represented by a factor graph and by message passings between the nodes of the graph [11], the downlink transmit beamforming algorithm is implemented. The linear precoding method in [10] is suboptimal. We instead use the vector perturbation algorithm [8] which can achieve performance near capacity. The application of this algorithm requires solving integer least squares (ILS) problem. There are algorithms developed to solve the ILS problem. A review of these algorithm can be found in [12], [13], [14]. However, solving ILS distributively on graphs is still an open problem. In [15], factor graphs for ZF, MMSE, VBLAST and ML solutions to LS problem in the context of MIMO decoding are given. In [16], the convergence criterion for the solution of the LS problem on a graph using belief propagation is given. In [17], the convergence of the min-sum algorithm is improved by using prior information found from the ZF or MMSE solution. In [18], [19] the loopy structure of the distributed LS problem is converted to a tree using Gaussian tree approximation. In [15] the VBLAST type factor graph solution is given. In all these studies, the convergence caused by the structure of the channel matrix is an indispensible problem. The proposed distributed method includes a preprocessing using the LLL reduction algorithm [23] and the min-sum algorith running on factor graph. The organization of the paper is as follows : In Section II the cooperative downlink transmission model is presented. Section III describes the proposed transmission algorithm. In Section IV the simulation results are given. We conclude the paper in Section V.

II. C OOPERATIVE D OWNLINK C OMMUNICATION M ODEL We consider a multicell network consisting of n BSs and n MSs. For simplicity we assume that each BS and each MS is equipped with single antenna. MSs in the same cell do not interfere with each other through TDMA or FDMA. Intercell interference is allowed to provide full reuse in the network. We focus on the multicell downlink transmission scheme where each BS transmits data symbols to its intended user.

If the transmitter knows the channel, then one way that comes to the mind is to nd the inverse of the channel matrix and multiply the transmitted symbol with this inverse. However, power limitation at the transmitter is a constraint and the large singular values of the channel inverse prevent us from directly multiplying the transmitted symbol with the inverse of the channel matrix. The proposed way in [5] is to make sure that the transmitted data does not lie along the large singular values of the inverse of the channel matrix. The data is perturbed so that the data vector is approximately orthogonal to the right singular vectors associated with the large singular values of the inverse of the channel matrix. In other words, we form d 1 from the data vector d so that x = H d has norm smaller can still be decoded at MSs. At this point than H1 d and d an idea derived from Tomlinson-Harashima (TH) precoding [9], [22] is used and each element of d is perturbed by a complex number whose real and imaginary parts are integer.

Fig. 1: MIMO Model for Multicell Downlink Transmission Scheme In g. 1 the model used in this paper is depicted. Tha backhaul provides high speed communication link between BSs. Through this link information can be shared and cooperative transmission is enabled. Mathematically the model can be represented as :
n

Fig. 2: Vector Perturbation The vector perturbation algorithm is depicted in g. 2. In this case : = d + l d (3)

yi = hii xi +
k=1,k=i

hik xk + wi

(1) where is a positive real number and l is a complex vector whose real and imaginary parts are integers. is determined according to the constellation which d belongs to. For the minimization of the output transmission power, l is chosen . The problem can be to minimize the norm of x = H1 d expressed as : l = arg min H1 (d + l) 2n
lR 2

where yi is the received signal at MSi and xi is the transmitted signal at BSi. {wi } are independent identically distributed Gaussian noise variables with zero mean and variance 2 . hik represents the channel from BSk to MSi. The vector form for the model can be expressed as: y = Hx + w (2)

We assume that the system in (2) can be converted to a real system so that y R2n is the received signal vector at MSs, x R2n is the transmit signal vector at BSs, H R2n2n is the channel matrix and w R2n is the noise vector. d R2n is the vector of data symbols, where di is the data symbol intended for MSi and it is available at BSi. The main focus of this paper is to form the transmitted signal y using d and H such that MSs can decode easily. If we consider the cooperation between BSs, we can visualize the network formed by n BSs and n MSs as a MIMO broadcast channel. III. D OWNLINK T RANSMISSION BASED ON I NFERENCE ON G RAPH The downlink algorithms has to be designed so that most of the processing is done in the BSs. Assuming that the channel is known to BSs, the transmitted signal is modied so that the MSs decode easily. The method that we considered is the regularized perturbation technique [5] which can achive performance near sum capacity for MIMO Broadcast Channel.

(4)

This problem is an integer lattice least squares problem. Using the following conversion of variables the integer least squares problem is written as : x = arg min z Dx 2n
xR 2

(5)

where z H1 d, D H1 , x l, x l conversions are applied. The sphere decoding algorithm in [12], [13], [14] can be used to solve this problem if a central processor was assumed. For the distributed processing, the problem will be solved on a factor graph by message passing between nodes. The problem in ( 5) can be formulated as :
n n1 n

x = arg min

xR2n

Aii x2 i 2bi xi +
i=1 i=1 j =i+1

Aij xi xj

(6)

th where bi = yT di , Aij = dT column of D. i dj and di is the i

The factor graph corresoponding to the min-sum algorithm includes loops in its sructure. The loopy message passing algorithm fails to converge when the raph is not sparse. Therefore, for the downlink MIMO BC transmission, the performance becomes poor. In [16], it is shown that, in order for the iterative min-sum algorithm to converge on a graph, the diagonal dominancy condition has to be satised. The diagonal dominancy brings a constraint on the D matrix which we cannot always guarantee to meet for this problem. The diagonal dominancy constraint can be expressed as: Fig. 3: Factor Graph Representation of ILS Problem Aii
j =i

|Aij |

(12)

The representation of min-sum problem in (6) is depicted in g.3. The messages passing between the factor and variable nodes are given in (7-9). fi xi (xi ) and gij xi (xi ) represent the messages from factor nodes to variable nodes and xi gij (xi ) represents the messages from the variable nodes to factor nodes. C in (7) and D in (9) is used to normalize the messages. The initialization of the messages is dened as in (10). The iterations of the messages is terminated when the maximum number of iterations is reached or the difference between the current messages and the previous ones is below a tolerance value. variable node to factor node:
i 1

where Aii and Aij are dened in (6). For sparse matrices, the diagonal dominancy constraint is easier to meet. However, D matrix is in our case, related to the inverse of the channel matrix H. Even if the channel matrix is sparse, its inverse may not be sparse. We exploit the LLL basis reduction algorithm [23] in order to make the min-sum algorithm converge. In the proposed method, we rst preprocess the D matrix and convert the problem into another problem which tends to converge. For the problem in (5), LLL reduction algorithm starts with the QR decomposition of D and returns reduced upper triangular matrix R R2n2n , unimodular matrix R R2n2n and the vector z R2n . Using the LLL reduction algorithm we can decompose the D as: D = QRZ1 (13)

xi gij (xi ) =fi xi (xi ) +

gki xi (xi ) +

(7)

gik xi (xi ) + C
k=i+1,k=j

where Q is a unitary matrix. Then we convert the problem in (6) as: 2 x = arg min z QRZ1 x (14) 2n
xR

factor node to variable node: fi xi (xi ) =fi (xi ) gij xi (xi ) = min gji (xi , xj ) + xj gkl (xj ) + D
xj

setting u = Z as in (15). (8) (9)

x and z = QT z we can convert the problem u = arg min z Ru 2n

(15)

where: k =min(i, j ) l =max(i, j ) fi (xi ) =Aii x2 i 2bi xi for i=1,...,n for i=1,...,n-1; j=i+1,...,n gij (xi , xj ) =2Aij xi xj initialization: xi gij (xi ) =0 termination:
i1

The transformed problem in (15) is a reduced form of (5) and it tends to converge. After nding the solution using (15), the transformation x = Zu is applied and the solution of (5) is found. IV. EXPERIMENTAL RESULTS In this section we provide simulation results for our proposed method, Min-Sum with LLL Reduction (MS-LLL) over a hexagonal cell network consisting of 7 cells as in g.4. Pathloss exponent is taken as 3. We assume Rayleigh at fading channel and users are distributed uniformly in their cells. The channel is quasi-static and the channel information is available to all BSs. The system is simulated using QPSK and 16-QAM constellations with Gray mapping. The transmit signal vector is normalized so that the total power is unity. We used 50000 realizations for the channel matrix. We compare our proposed algorithm with the distributed downlink beamformer (DDLB) algorithm proposed in [10] and the centralized sphere encoder (SE) algorithm. The symbol error rate(SER) performance with respect to the received SNR

gij xi (xi ) = 0

(10)

xi gij (xi ) =fi xi (xi ) +

gki xi (xi ) +

(11)

gik xi (xi )
k=i+1

Fig. 4: Hexagonal Celuular Array Model

per MS is experimented for both methods. The number of iterations are also experimented and plotted wrt SNR. The results for SER and iteration counts are seen in g.6, g.6 and g.7 respectively.

Fig. 6: Symbol error rate vs. SNR for 16QAM For the factor nodes in the MS-LLL, the minimization is taken over only one variable and independent of the size of the network, each factor node is connected to only two variable nodes. When calculating a message to a variable node, only the message from the other variable node is used. The complexity of the factor nodes of the DDLB increases with n2 , since every factor node in that graph is connected to every factor node. The factor node calculates the message to a variable node using the messages from the rest n-1 variable nodes and it does this calculation n times.
DDLB MS-LLL Nv n n Nf 2n
n(n+1) 2

Fig. 5: Symbol error rate vs. SNR for QPSK For QPSK constellation, at low SNR the performance of the MS-LLL is close to the SE. The SER of DDLB method has an error oor caused by the convergence of the algorithm. For a terget SER of 102 MS-LLL is 5dB close to the SE. DDLB algorithm cannot achieve SER better than 101 . The same performance comparison between the three methods is also valid for 16QAM constellation. In order to compare the complexities of DDLB and MSLLL, the number of factor nodes and variable nodes is analyzed as seen in TableI. The number of nodes variable nodes increases by n2 for MS-LLL and linear with n for DDLB, where n is the number BSs in the network. However, for these distributed algorithms, the complexity is determined by the complexity per node per iteration. Each node in the graph can be visualized as a processor and algorithm is shared among those processors. So as wee increase the degree of parallelization, we decrease the complexity per processor. So we can conclude that, a system with higher number of processors with low complexity per processor is less complex than the system with less processors with high complexity. In fact, this is why the distributrd algorithm is developed. The complexities per processor is lower for MS-LLL algorithm.

TABLE I: Number of factor and variable nodes

Fig. 7: Iteration Count vs. SNR for QPSK In g.7 the iterations counts for DDLB and MS-LLL are depicted wrt SNR. Especially at high SNR, the number of

iterations are higher for DDLB. This is expected since the algorithm does not converge even at high SNR. This result also brings more complexity for the DDLB method.

Fig. 10: Symbol error rate vs. SNR for QPSK, n=6

Fig. 8: Symbol error rate vs. SNR for QPSK, n=2 The MS-LLL method is also simulated under various settings for the number of BSs and MSs. In g.??, g.?? and g.??, the simulation results are given for n = 2, 4, 6 respectively. For all these simulations, the MS-LLL algorithm performs better than DDLB. For n=2, the performance of MS-LLL is same as SE. This is a result of LLL reduction algorithm. LLL reduction algorithm is more effective for small number of BSs. The symbol error rate for DDLB algorithm has an error oor caused by the convergence problem of the algorithm.

multicell processing, we propose an algorithm which forms the transmit signal vector from BSs to MSs so that intercellular interference is mitigated and decoding at MSs is done easily. We rst consider the vector perturbation method and apply it to the MIMO BS scenario. The problem is then converted to an integer least squares problem. We use min-sum algorithm on factor graph to solve the ILS problem in a distributed manner. The convergence problem of the algorithm is mitigated using LLL reduction algorithm. The performance of the algorithm in a hexagonal cellular model is analyzed and it is shown that it performs better than the distributed downlink beamforming method in terms of symbol error rate averaged over MSs. We also consider the complexity of MS-LLL and compare it with DDLB. The degree of parallelization is higher for MSLLL which reduces the complexity of the complexity per node per iteration. The complexity per node per iteration is also independent from the network size. Future work will include achieveing higher performance for the MS-LLL algorithm and making it perform closer to the sphere encoder. Cellular networks with bigger sizes will also be experimented. Also we will consider the case in which number of BSs and MSs are different and they are equipped with multiple antennas. R EFERENCES
[1] S. Vishwanath et al, Duality, achievable rates, and sum-rate capacity of Gaussian MIMO broadcast channels IEEE Trans. Inf. Theory, vol. 49, no.10, pp. 2658.2668, Oct. 2003. [2] M. Costa, Writing on dirty paper IEEE Trans. Inf. Theory, vol. IT-29, no.3, pp. 439.441, May 1983. [3] T. Haustein, C. von Helmolt, E. Jorswieck, V. Jungnickel, and V. Pohl, Performance of MIMO systems with channel inversion in Proceedings 55th IEEE Vehicular Technology Conference (Spring), vol. 1, (Birmingham, Alabama), pp. 35.39, May 2002. [4] Q. H. Spencer et al, Zero-Forcing Methods for Downlink Spatial Multiplexing in Multiuser MIMO Channels IEEE Trans. Signal Process., vol. 52, no. 2, pp.461-471, Feb. 2004. [5] C. B. Peel, B. M. Hochwald, and A. L. Swindlehurst, A vector-perturbation technique for near-capacity multiantenna multiuser communication-Part I: Channel inversion and regularization, IEEE Trans. Commun., vol. 53, pp. 195202, Jan. 2005.

Fig. 9: Symbol error rate vs. SNR for QPSK, n=4 V. C ONCLUSION BS Cooperation for Downlink MIMO BC Channel is used to achieve higher sum capacity in a cellular network. Using

[6] A. Wiesel, Y. C. Eldar, and S. Shamai, Linear precoding via conic optimization for xed MIMO receivers, to appear in IEEE Trans. Signal Processing. [7] H. Boche and M. Schubert, A general duality theory for uplink and downlink beamforming, in Proc. IEEE Vehicular Technology Conference (VTC-2002 Fall), pp. 87 91, Sept. 2002. [8] C. B. Peel, B. M. Hochwald and A. L. Swindlehurst, A vectorperturbation technique for near-capacity multi-antenna multi-user communication - part II: perturbation, IEEE Trans. Commun., vol. 53, pp. 537544, March 2005. [9] M. Tomlinson, New automatic equaliser employing modulo arithmetic, Electron. Lett., vol. 7, pp. 138139, Mar. 1971. [10] Boon Loong Ng, Jamie S. Evans, Stephen V. Hanly, Defne Aktas Distributed Downlink Beamforming with Cooperative Base Stations IEEE Trans. On Information Theory, vol. 54, No.12, December 2008. [11] F. R. Kschischang, B. J. Frey, and H.-A. Loeliger, Factor graphs and the sum-product algorithm, IEEE Trans. Inf. Theory, vol. 47, no. 2, pp. 498519, Feb. 2001. [12] U. Fincke and M. Pohst, Improved methods for calculating vectors of short lengths in a lattice, including a complexity analysis, Math. Computat., vol. 44, pp. 463471, Apr. 1985. [13] E. Agrell, T. Eriksson, A. Vardy, and K. Zeger, Closest point searches in lattices, IEEE Trans. Inf. Theory, vol. 48, pp. 22012214, Aug. 2002. [14] M. O. Damen, H. El Gamal, and G. Caire, On maximum-likelihood detection and the search for the closest lattice point, IEEE Trans. Inf. Theory, vol. 49, pp. 23892402, Oct. 2003. [15] E. Biglieri, R. Calderbank, A. Constantinides, A. Goldsmith, A. Paulraj, H. V. Poor. MIMO Wireless Communications Chapter 5. Cambridge University Press 2007. [16] Moallemi, C.C.; Van Roy, B.; , Convergence of Min-Sum MessagePassing for Convex Optimization, IEEE Transactions on Information Theory, vol. 56, no.4, pp.2041-2050, April 2010 [17] Goldberger, J. Leshem, A.; , Pseudo Prior Belief Propagation for densely connected discrete graphs, 2010 IEEE Information Theory Workshop (ITW), pp.1-5, 6-8, Jan. 2010 [18] Goldberger, J. Leshem, A.; , MIMO Detection for High-Order QAM Based on a Gaussian Tree Approximation, IEEE Transactions on Information Theory, vol. 57, no.8, pp.4973-4982, Aug. 2011 [19] Goldberger, J. Leshem, A.; , A Gaussian Tree Approximation for Integer Least-Squares Advances in Neural Information Processing Systems 22 (2009), pp. 638-645. [20] R. C. Prim, Shortest connection networks and some generalizations, J. Bell Syst. Tech., pp. 13891401, 1957. [21] Gesbert, D.; Hanly, S.; Huang, H.; Shamai Shitz, S.; Simeone, O.; Wei Yu; , Multi-Cell MIMO Cooperative Networks: A New Look at Interference, IEEE Journal on Selected Areas in Communications, vol. 28, no.9, pp.1380-1408, December 2010 [22] H. Harashima and H. Miyakawa, Matched-transmission technique for channels with intersymbol interference, IEEE Trans. Commun., vol. COM-20, pp. 774780, Aug. 1972. [23] A.K. Lenstra, H.W. Lenstra, and L. Lovasz, Factoring polynomials with rational coefcients, Mathematicsche Annalen, vol. 261, no. 4, pp. 515534, 1982.