Capacity Evaluation of Various Multiuser MIMO Schemes in Downlink Cellular Environments

The 17th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’06) CAPACITY EVALUATION OF VARIOUS MULTIUSER MIMO SCHEMES IN DOWNLINK CELLULAR ENVIRONMENTS Jingon Joung, Eun Yong Kim, Sung Hoon Lim, Yong-Up Jang, Won-Yong Shin, Sae-Young Chung, Joohwan Chun, and Yong H. Lee ∗ Department of Electrical Engineering and Computer Science Korea Advanced Institute of Science and Technology (KAIST) 373-1, Guseong-dong, Yuseong-gu, Daejeon, 305-701, Republic of Korea TEL: +82-42-869-3437, FAX: +82-42-869-4030 E-mail: yohlee@ee.kaist.ac.kr A BSTRACT Presented in this paper is a study of the capacity evaluation of various multiuser MIMO schemes in cellular environments. The throughputs per user of the generalized zero-forcing with rank adaptation and vector perturbation schemes are compared with the capacity bound of the Gaussian MIMO broadcast channel, obtained by dirty paper coding under proportional fairness scheduling. The average cell throughputs of these schemes are also compared. From these comparisons, this study provides vital information for applying multiuser MIMO schemes in multicell environments. I. I NTRODUCTION Recently, the capacity bounds of Gaussian multiple-input multiple-output (MIMO) broadcast channels have been studied and shown to be achieved by dirty paper coding (DPC) [1]–[7], and several practical progresses using source-channel coding in the dirty paper channel have been made in this area [6], [7]. However, there are still open problems to achieve rates closer to capacity with practical transceivers, whereas many suboptimal schemes have been proposed to simplify the transceiver. In [4], the capacity bound obtained from the DPC and the throughput of linear processing, such as zero forcing (ZF)-based orthogonal space-division multiplexing (OSDM) [8]–[10] and time-division multiple-access (TDMA), have been compared for downlink cellular systems. In this paper, the downlink throughputs of a more general linear OSDM scheme with rank adaptation [11] and nonlinear vector perturbation [12], [13] are evaluated and compared numerically with the capacity bound in the multicell environment. To apply a point-to-point MIMO technique in a point-tomultipoint MIMO system, the proportional fairness (PF) scheduler in [14] is considered. First, the cumulative density functions (CDFs) of the throughput per user are shown, and then the average cell throughputs are compared. The throughput gaps from the capacity bound for suboptimal techniques are shown using computer simulations. All of these results provide useful insights into the design and application of multiuser MIMO techniques in cellular environments. The organization of this paper is as follows. Section II describes the multiuser MIMO system model in multicell envi∗ This work was supported in part by the University Information Technology Research Center Program of the government of Korea. c 1-4244-0330-8/06/$20.00°2006 IEEE ronments. Various multiuser MIMO schemes are then presented in Section III. Section IV shows the computer simulation results to demonstrate the performance evaluation of the optimal and suboptimal methods. Finally, Section V concludes the paper. II. M ULTIUSER MIMO S YSTEM M ODEL The multiuser downlink MIMO system in cellular environments, where a common base station with NT transmit antennas transmits different signals to multiple mobile users, is shown in Fig. 1. Let K be the number of users in the cell and NR,k be the number of receive antennas at the k-th user. To ensure that the users receive their data without coordination, an appropriate preprocessing of the data should be carried out at the transmitter. Let F denote the preprocessing function and xk be the Lk -by-1 data vector for the k-th user, where Lk is the number of spatial modes supported to the k-th user. Then, the transmitted signal is represented by: s = F (x1 , x2 , . . . , xK ) , where s is an NT -by-1 vector. In multicell environments, the received signal at the k-th user, an NR,k -by-1 vector, is represented by yk = Hk s + nk , where Hk is an NR,k -by-NT matrix that denotes the channel matrix between the base station and the k-th user, and nk denotes the additive noise due to both the thermal noise and the interferences from neighboring cells. It is assumed that the entries of Hk are independent and identically distributed (i.i.d.) random variables whose means are zero and variances are determined by the path loss and shadowing factor of the k-th user. Each user decodes the data vector as follows: x̂k = Gk (yk ) ∀k, where Gk is the postprocessing function for the k-th user. We assume that the channel matrices for different users are independent and all the channel matrices are quasi-static, flat fading, and perfectly known at the base station. For simplicity, it is also assumed that every user has the same number of receive antennas, i.e. NR,k = NR ∀k. Throughout this paper, A† , tr(A), and |A| denote the conjugate transpose, trace operation, and determinant of matrix A, respectively. The 17th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’06) BS N1 1 H1 1 x1 LM NR,1 n1 ⊕ ⊕ MS 1 G1 M 1 M L1 x̂1 1 1 H2 2 x 2 LM F 2 s M M1 x K LM HK MS 2 ⊕ NR,2 NT K n2 N1 G2 M ⊕ 1 M L2 x̂ 2 N N1 NR,K MS K nK ⊕ ⊕ 1 M GK M LK x̂ K Figure 1: System model of a multiuser MIMO system III. T ECHNIQUES FOR M ULTIUSER MIMO D OWNLINK In this section, we first present the capacity bound of the multiuser MIMO downlink systems, i.e. the DPC achievable rate regions. Next, we present two suboptimal schemes: the generalized zero-forcing (GZF) and the vector perturbation (VP) schemes. Finally, we explain the PF scheduler with rank adaptation that is also under consideration. A. Capacity Bound (DPC) With the DPC it is possible to precancel interferences that are known noncausally at the transmitter, resulting the same capacity as if there is no interference. When the DPC is applied to a multiuser MIMO downlink, it can be used to precancel other users’ coded signals with an appropriate ordering [3]. The DPC achievable rate region is given by:  CDP C = Co  [ π,Σi  Rπ  , (1) S where Co (A) is the convex hull operation of the set A , is the union operation, Rπ = (Rπ(1) , ..., Rπ(K) ), π is the user permutation vector, and Rπ(i) = ¯ ¯ ³P ´ ¯ ¯ † ¯INr + Hπ(i) j≥i Σπ(j) Hπ(i) ¯ ¯, ³P ´ log ¯¯ ¯ † Σ H ¯INr + Hπ(i) π(j) j>i π(i) ¯ i = 1, ..., K. Here, IN is the N -dimensional indentity matrix, Σk is the covariance matrix of the user k, the union operation is performed over all possible permutations (π(1), π(2), ..., π(K)), and all possible non-negative covariance matrices that are constrained to tr(Σ1 + Σ2 + ...ΣK ) ≤ P . In this paper, we evaluate the achievable rates using the PF scheduler, which needs to find the optimal point on the boundary of the rate region (1) that maximizes the weighted sum rates for a given weight vector. Finding the optimal operating point with (1) is formidable since it is neither convex nor concave. Thus, this problem is transformed into a dual multiple access channel (MAC) optimization problem that is concave of the covariance matrices; then, the following result is obtained [2]: [ CDP C (P, H) = CM AC (P, H† ), P: PK i=1 Pi =P which shows that the capacity of the Gaussian broadcast channel is the same as the union of the MAC capacity regions over all individual power constraints P = (P1 , ..., P2 ) that sum to P . Here, the MAC capacity region for a given power constraint and channel instance, CM AC (P, H† ), is given by the union of the rate regions over all possible non-negative covariance matrices, as follows: ( S P R : i∈S Ri CM AC (P, H† ) = Qi ≥0 tr(Qi )≤Pi ∀i ) ¯ ¯ P ¯ ¯ † ≤ log ¯INt + i∈S Hi Qi Hi ¯ ∀S ⊆ {1, ..., K} , where Qi is the covariance matrix for the i-th user and R = (R1 , R2 , ..., RK ). We locate the operating point on the boundary of the optimal region for a given weight vector, µ, by maximizing the weighed sum rates using the algorithm in [4], which is calculated using the results of the dual MAC and standard convex optimization procedures shown in (2), where R̄ is a vector of user rates of the set R and µ is a weight vector given by PF. The numerical methods for the optimization procedure in (2) are shown in detail in [4]. B. Suboptimal Schemes 1) Generalized Zero-Forcing (GZF) For the downlink of multiuser MIMO systems, one of the suboptimal schemes is the linear OSDM. The OSDM enables the users to receive their own data with zero co-channel interference. Recently, [10] proposed an iterative algorithm that finds the preprocessing and postprocessing linear operators for the OSDM, which maximize the effective channel gains. In this paper, however, to avoid the burden from iterations, we use the conventional ZF block-diagonalization [9], in which the multiuser MIMO channels are decomposed into multiple, singleuser MIMO channels. For the case of NR,k > Lk , the dominant Lk left singular vectors are considered as the effective left The 17th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’06) max R∈RDP C µ · R̄ = max Qi : PK K−1 X i=1 tr(Qi )=P i=1 ¯ ¯ ¯ ¯ K i ¯ ¯ ¯ ¯ X X ¯ ¯ ¯ ¯ † † Hπ(l) Qπ(l) Hπ(l) ¯ (µπ(i) − µπ(i+1) ) log ¯I + Hπ(l) Qπ(l) Hπ(l) ¯ + (µπ(K) ) log ¯I + ¯ ¯ ¯ ¯ singular vectors of Hk for the k-th user in order to preserve the maximum spatial diversity. Then, we apply the singular value decomposition approach [8] to each single-user MIMO channel to achieve the maximum single-user MIMO capacity. This scheme is termed GZF. In the aspect of complexity, GZF has a merit in that it needs only linear processing at both the transmitter and receivers. Let F denote the preprocessing matrix at the base station and Gk denote the postprocessing matrix at the k-th user. We note that F can be separated as [F1 F2 · · · FK ], where Fk is the preprocessing matrix for the k-th user, and the transmitted signal can be represented by: s= K X Fk xk . k=1 To obtain the optimal Fk and Gk , let Wk be a matrix with orthogonal columns such that: Wk ∈ null ¡£ † H1 U1 · · · H†k−1 Uk−1 H†k+1 Uk+1 and h i† Heff = H†1U1 · · · H†K UK [W1 · · · WK ] , °2 ° K ° °X ° ′ ° F (x + τ l ) l = arg min ° ° , k k k ° l′ ° k=1 where τ is a scalar of a positive real value, lk is the integer vector, and k·k denotes the vector 2-norm. Using the same pre- and postprocessing matrices with GZF as in (3), the postprocessed data vector for VP is given by: x̂k = Gk Hk Fk (xk + τ lk ) + Gk nk . (5) Taking the modulo-τ operations for (5), the estimate of x with VP is obtained and is the same as (4). where the effective channel of the k-th user is given by: Heff,k = U†k Hk Wk . Let the columns of Ueff,k (Veff,k ) denote the left (right) singular vectors of Heff,k , such as: † Heff,k = Ueff,k Deff,k Veff,k , where Deff,k denotes a diagonal matrix that consists of singular values of Heff,k . To achieve the maximum capacity, Fk and Gk are Fk , Wk Veff,k Ek , and Gk , U†eff,k U†k , (3) where the power-loading matrix Ek is an Lk -by-Lk diagonal matrix, whose diagonal elements are determined by the wellknown waterfilling algorithm [8]. Then, the postprocessed data vector is represented by: = G k yk = Gk Hk Fk xk + Gk nk . 2) Vector Perturbation (VP) In our comparison, VP, a one-dimensional dirty paper approach, is considered as a simple nonlinear technique. By perturbing the data vector at the transmitter and taking modulo operation at the receivers, VP achieves an excellent performance gain over the linear processing schemes, especially at high signal-to-noise ratios (SNRs) [12]. Recently, VP has been extended to systems with multiple receive antennas and multiple spatial modes, and an optimum VP for minimizing the mean square error has been proposed [13]. Once the precoding and decoding matrices for a ZF block diagonalization are determined, VP chooses the best perturbation vector to minimize the power in the transmitted signal when added to the data vector. In this paper, the data vector is perturbed based on the GZF. The perturbation vector is constrained to an integer vector so as not to affect the zero co-channel interference condition [12]. The transmitted signal can be represented by: K X Fk (xk + τ lk ), s= k=1 ¤† ¢ · · · H†K UK , where null(A) denotes the null space of A and Uk consists of the dominant Lk left singular vectors of Hk . If [W1 · · · WK ] is multiplied to the transmitted signal as the preprocessing matrix, the effective block-diagonal channel matrix becomes: x̂k (2) l=1 l=1 (4) 3) Scheduling with Rank Adaptation When the total number of receive antennas is more than NT , both GZF and VP need an extra scheduling algorithm, while the DPC provides the optimal user selection implicitly. Moreover, if the number of receive antennas for a user can be more than one, the rank adaptation technique [11] should be applied to improve the sum rates. In our comparison, the PF scheduler is employed under the consideration of fairness among the users. The set of spatial modes for users, (L1 , L2 , . . . LK ), are determined by bruteforce searching. IV. C OMPARISON AND D ISCUSSION The CDFs of throughput per user and the average cell throughputs for the DPC, GZF, VP, and TDMA are evaluated. MIMO channels are obtained by generating independent Gaussian random variables with a zero mean, and the results shown below are the averages over 350 independent trials under the PF The 17th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’06) Table 1: Simulation Parameters Values Parameter {2, 1}, {4, 2} Inter-cell 1, 2, 4, 8 Tx antenna pattern PF Path loss model Synchronous Path loss exponent 47 dBm Shadowing STD −94 dBm BS correlation 1000 m Min. separation (d0 ) Rayleigh Max. achievable SINR 1 1 0.9 0.9 0.8 0.8 0.7 0.7 CDF of throughput CDF of average SINR Parameter # of antennas {NT ,NR,k } # of users (K) Scheduling Sync. protocol Tx power Rx noise level Radius of cell Channel model 0.6 0.5 0.4 0.6 0.5 0.3 0.2 0.2 0.1 0.1 −6 −4 −2 0 2 4 6 8 Average SINR (dB) 10 12 14 16 0 18 K=1,2,4,8 0.4 0.3 0 −10 −8 Values 18 (2nd-tier) Omni-direction 3GPP2/TSG-C.R1002 3.5 8.9 dB 0.5 100 m 17.8 dB DPC VP GZF TDM 0 1 2 3 4 Throughput (bits/sec/Hz) 5 6 7 (a) Figure 2: SINR CDF of multicell environments in Table 1. 1 0.9 0.8 0.7 CDF of throughput scheduler for each average SINR point. Every average SINR was assigned independently 8, 000 times to all users according to the distribution in Fig. 2, which was generated under the multicell environment outlined in Table 1. The mobile station locations are constrained to be distanced from base station farther than minimum distance (d0 ) 100 m and the maximum achievable SINR in the receiver is limited to 20 dB , as shown in Table 1. BS correlation is defined as the correlation factor among the inter-cells’ BSs and its value is 0.5. Fig. 3(a) and 3(b) show the CDFs of the throughput per user for {NT , NR } = {2, 1} and {4, 2}, respectively. Here, when NT = 4 and K = 4, 10% of the high end users for each scheme in an inter-cell achieve 4.65, 4.19, 3.92, and 2.86 bit/sec/Hz using the DPC, VP, GZF, and TDMA, respectively. For the low average SINR users, the gains of the GZF, VP, and DPC over TDMA seem to be negligible. This is because the DPC, GZF, and VP all use the solution of allocating all of the power to the proportionally highest rate user in this regime, which is the same as TDMA with PF. In contrast, in the high SINR regime, the DPC, GZF, and VP have a significant gain over TDMA since the multiplexing gain of TDMA is bounded by min{NT , NR }, whereas those of the multiuser MIMO schemes are bounded by min{NT , KNR }. There is an additional gain 0.6 0.5 K=1,2,4,8 0.4 0.3 0.2 DPC VP GZF TDM 0.1 0 0 2 4 6 8 10 Throughput (bits/sec/Hz) 12 14 (b) Figure 3: CDF of throughput of various multiuser MIMO schemes when (a) {NT , NR } = {2, 1}, (b) {NT , NR } = {4, 2}. for the DPC over GZF and VP due to the optimal cancelling of inter-user interferences. The 17th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’06) VP, which also have potential to be improved further, are good candidates for next generation communications. However, all of the previous schemes need perfect channel state information at the transmitter for implementation. Examining the robustness of these schemes against channel uncertainty and the exact complexity comparison of each system remains as work to be undertaken. 12 DPC VP GZF TDM Average cell throughput (bits/sec/Hz) 11 10 9 8 7 6 R EFERENCES {N ,N }={4,2} 5 T R [1] M. Costa, “Writing on dirty paper,” IEEE Trans. Inform. Theory, vol. 29, pp. 439–441, May 1983. 4 3 {N ,N }={2,1} T 2 [2] S. Vishwanath, N. Jindal, and A. Goldsmith, “On the capacity of multiple input multiple output broadcast channels,” in Proc. Int. Conf. Communications, Apr. 2002, pp. 1444–1450. R 1 0 1 2 4 Number of users (K) 8 Figure 4: Average cell throughput versus the number of users when {NT , NR } is {2, 1} and {4, 2}. Fig. 4 shows the average cell throughputs for {NT , NR } = {2, 1} and {4, 2} versus the number of users. Obviously, we can see that the average cell throughputs are the same for each scheme when there is only one user in the cell since it is a pointto-point optimization problem without interference cancellation for all schemes. By increasing the number of users, the average throughputs for both {NT , NR } = {2, 1} and {4, 2} increase, but the rate of increase is much larger for the {4, 2} case with two users, because for more than two users, it will fully utilize the multiplexing gain of 2. For the region where multiplexing gain is fully utilized for both antenna configurations (e.g. users ≥ 2), the rate of increase is slightly larger for {NT , NR } = {4, 2} compared with {NT , NR } = {2, 1}, which results from the array gain of having more receiver antennas. In general, for KNR ≤ NT , the multiplexing gain will be the dominating gain, and this gain is proportional to KNR ; whereas for the KNR > NT case, the dominating gain is a relatively smaller multiuser diversity gain. Following this tendency, the cell average throughput for K = 8 and NT = 2(4) shows a gap of 12(16)%, 14(20)%, and 35(45)% from that of the DPC under the PF scheduler for VP, GZF, and TDMA, respectively. Compared to orthogonal schemes such as TDMA and FDMA, the GZF and VP have a significant throughput improvement of approximately 32(45)% and 34(53)%, respectively, with some additional complexity. Despite the gains that the GZF and VP provide over TDMA, there is still a noticeable gap from the optimal bound. V. C ONCLUSION The CDFs of the throughput per user and the average cell throughputs for various multiuser MIMO schemes are evaluated. The effects of multiplexing, multiuser diversity, and array gain are observed. Also, the performance gaps between the optimal and various suboptimal schemes are shown. Here, it is surmised that the suboptimal schemes such as the GZF and [3] A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, “Capacity limits of MIMO channels,” IEEE J. Select. Areas Commun., vol. 21, pp. 683–702, June 2003. [4] H. Viswanathan, S. Venkatesan, and H. Huang, “Downlink capacity evaluation of celluar networks with known-interference cancellation,” IEEE J. Sel. Areas Commun., vol. 21, pp. 802–811, June 2003. [5] H. Weingarten, Y. Steinberg, and S. Shamai (Shitz), “The capacity region of the Gaussian MIMO broadcast channel,” in Proc. IEEE Int. Symp. Information Theory (ISIT), June 2004, pp. 174. [6] U. Erez and S. ten Brink, “A close-to-capacity dirty paper coding scheme,” IEEE Trans. Inform. Theory, vol. 51, pp. 3417–3432, Oct. 2005. [7] Y. Sun, A. D. Liveris, V. Stanković, and Z. Xiong, “Near-capacity dirty-paper code designs based on TCQ and IRA codes,” in Proc. IEEE Int. Symp. Information Theory (ISIT), Sept. 2005, pp. 184– 188. [8] G. G. Raleigh and J. M. Cioffi, “Spatio-temporal coding for wireless communication,” IEEE Trans. Commun., vol. 46, pp. 357–366, Mar. 1998. [9] Q. H. Spencer, A. L. Swindlehurst, and M. Haardt, “Zero-forcing methods for downlink spatial multiplexing in multiuser MIMO channels,” IEEE Trans. Signal Processing, vol. 52, pp. 461–471, Feb. 2004. [10] Z. G. Pan, K. K. Wong, and T. S. Ng, “Generalized multiuser orthogonal space division multiplexing,” IEEE Trans. Wireless Commun., vol. 3, pp. 1969–1973, Nov. 2004. [11] Y.-U. Jang, H. M. Kwon, and Y. H. Lee, “Adaptive mode selection for multiuser MIMO downlink systems,” in Proc. IEEE Vehicular Technology Conf. (VTC’06 Spring), Melbourne, Canada, May 2006. [12] B. M. Hochwald, C. B. Peel, and A. Lee Swindlehurst, “A vector-perturbation technique for near-capacity multiantenna multiuser communication-Part II: perturbation,” IEEE Trans. Commun., vol. 53, pp. 537–544, Mar. 2005. [13] E. Y. Kim and J. Chun, “Optimal vector perturbation minimizing system-wide mean square error for the downlink of multiuser MIMO systems,” in preparation. [14] E. F. Chaponniere, P. Black, J. M. Holtzman, and D. Tse, “Transmitter directed, multiple receiver system using path diverstiy to equitably maximize the throughput,” U.S. Patent No. 6449490, Sept. 10, 2002.