2009 International ITG Workshop on Smart Antennas – WSA 2009, February 16–18, Berlin, Germany

DFT-based vs. Cooperative MET-based MU-MIMO in the Downlink of Cellular OFDM Systems

L. Thiele, M. Schellmann, T. Wirth and V. Jungnickel F. Boccardi and H. Huang Fraunhofer Institute for Telecommunications Bell Labs, Heinrich-Hertz-Institut Alcatel-lucent Einsteinufer 37, 10587 Berlin, Germany fb,hchuang @alcatel-lucent.com { } thiele, schellmann, thomas.wirth, jungnickel @hhi.fraunhofer.de { }

Abstract—In this work, we compare different pre-coding tech- work, we include multi-user selection diversity in the user niques and their performance in a cellular MIMO OFDM down- grouping process. Multi-user diversity from a large number link. The first target is to determine the achievable additional of users per cell may help to increase the system performance gain by using a near-optium pre-coder instead of DFT-based pre-coding. In the second step, we include multi-user further. From [2] it is well known that this would require a selection diversity, which turned out to be advantageous especially brute force or at least a greedy search over the user space for downlink cooperation. An extended score-based scheduler, to choose appropriate users for the active set . To combat M which is known to asymptotically reach proportional fairness, residual CCI from surrounding cells we employ MMSE equal- is used for the user grouping and resource allocation in the ization, also known as optimum combining (OC) [4], at the cellular downlink. Further, we use minimum mean square error equalization at the terminal side to combat residual cochannel terminal side. interference. II.DOWNLINKSYSTEMMODEL NTRODUCTION I.I The downlink MIMO-OFDM transmission system with NT To enable ubiquitous broadband wireless access in fu- transmit and NR receive antennas per MT is described on each ture cellular systems supporting multiple-input multiple-output subcarrier by (MIMO) technology, transmission must be made robust against y = HCx + n , (1) multi-cell interference. Recently, it was shown that the capac- where H is the N N channel matrix and C the unitary ity scaling law, known from an isolated cell, also holds for R × T N N pre-coding matrix; x denotes the N 1 vector the interference limited case of a multi-cellular radio system T × T T × of transmit symbols; y and n denote the NR 1 vectors [1] with NT = NR transmit and receive antennas. This × work mainly focused on the optimization at the receiver side. of the received signals and of the additive white Gaus- However, cochannel interference (CCI) is still the dominant sian noise (AWGN) samples, respectively, with covariance E nnH = σ2I. source of performance degradation in the cellular network, { } α especially if N > N . Removing CCI may lead to a Assume that a group of cooperating base station (BS) T R C αN additional performance gain. sectors provides a beam set i. The beam set contains T b u 1, ..., αN In this work we compare different pre-coding techniques pre-coding beams i,u with T . In the follow- b u ∈ { } and their performance in a cellular MIMO orthogonal fre- ing we denote i,u as the -th pre-coding vector provided by i ym quency division multiplexing (OFDM) downlink. As a baseline the -th cell cluster. The received downlink signal at the m concept, we consider independent DFT-based pre-coding and MT in the cellular environment is given by multi-user MIMO (MU-MIMO) service in each sector of αNT m m m the system. As a first target, we determine the available y = Hi bi,u xi,u + Hi bi,jxi,j additional beamforming gain by using a near-optium pre- j=1 h jX=u coder instead. Therefore, we employ multi-user eigenmode i,u 6 transmission (MET), known to realize 90% of the dirty paper | {z } ζi,u coding (DPC) capacity in an isolated cell context [2], based NT m| {z } on the dominant eigenmodes of the mobile terminals (MTs) + Hl bl,jxl,j + n , (2) l j=1 in the serving area. Xl∀=i X A second target of this work is to realize limited localized 6 cooperative transmission in a multi-cellular network. Recent zi,u results obtained for a cellular MIMO OFDM downlink with a The desired data stream| xi,u transmitted{z on} the u-th beam sparse user distribution per sector, show potential performance from the i-th cluster is distorted by the intra-cluster and inter- gains for cooperation [3]. In this context it turned out, that cluster interference aggregated in ζi,u and zi,u, respectively. downlink cooperation based on DFT beams is unsuitable and Hm spans the N αN channel matrix for user m formed by i R× T MET should get the precedence. Additionally to preceding the cluster i. Thus, ζi,u denotes the interference generated in 2009 International ITG Workshop on Smart Antennas – WSA 2009, February 16–18, Berlin, Germany

the cooperation area. In the scope of this paper, it is assumed that all αNT beams in the beam set Ci are simultaneously ac- tive, whereby the total available power pi is assumed to be uni- formly distributed over the αN beams. Thus, E x 2 = T | i,j| p /(αN ) holds, and p = αNT E x 2 = αp with p i T i j=1 | i,j|  s s being the transmit power per sector. P  A. Determine serving BS or cooperative BS cluster The general assumption for single-cell operation is that each MT is assigned to the BS sector yielding the highest receive power over the entire frequency band, which is denoted as top-1 signal. Thus, the BS assignment is based on broadband Fig. 1. System concept assuming multiple antennas at the base station for power conditions, and a fast cell handover is assumed. For the purpose unitary fixed DFT-based pre-coded beamforming. SINR feedback downlink cooperation, we extend this scheme by evaluating is provided by the terminal for possible transmission modes using a narrow the top-α strongest signals and grouping the users selecting band feedback channel. the same set of α BSs for joint signal transmission. By subdividing the signal bandwidth into single sub-bands dominant eigenmodes of their channel to the cell cluster confined to a fixed number of consecutive subcarriers, we together with the corresponding eigenvalues. MET supports to define sets of contiguous transmission resources, which are simultaneously transmit up to αNT data streams via unitary denoted as resource blocks (RBs) in the following. Each RB beams, while up to NR beams may be assigned is processed independently, and thus the sectors to form a to a single MT. However, it has been indicated in [2] that group for cooperation can be selected individually per RB. MU-MIMO service for distinct terminals using MET is more III.DOWNLINKPRE-CODING efficient than time-multiplexing multi-stream transmission to a single user. Thus, for our investigations, we let all MTs A. Baseline: DFT-based fixed pre-coding report their dominant eigenmode only (λ = 1), which keeps As a baseline, we consider a system concept from [1], the required amount of feedback per user limited. where all sectors operate independently, while the inter- Determination of the MET-based pre-coding beams in ma- cell interference is accounted for at the multi-antenna MTs trix C is briefly sketched as follows: Consider a fixed set M only. The following evaluation is carried out for each RB of users, which should be served in a RB. Each user m decom- independently (refer to Fig. 1): Each BS provides a fixed m poses its NR αNT channel matrix Hi according to the sin- matrix C consisting of unitary DFT beams. Assuming that × m H gular value decomposition (SVD), yielding Hi = UiΣiVi . the inter-cell interference is completely known at the MTs, The dominant eigenmode is the first column vector of matrix the MTs evaluate the achievable rate per beam and convey Vi, denoted as vi,1. Together with the dominant eigenvalue in this information to their BS. At the BS, the feedback from the Σi, denoted as Σi,1, we obtain the user’s effective eigenmode different MTs is collected, and the DFT beams from matrix C H channel Γm = Σi,1vi,1. This measure needs to be fed back are assigned individually to the MTs. This simple approach has from each MT to the cell cluster. the convenient property that with the fixed beam set C used User orthogonalization at the BS: To obtain the pre-coding for all BSs, the CCI, i.e. ζi,u +zi,u, becomes fully predictable, vector for the m-th user, the BS cluster aggregates the interfer- enabling interference-aware scheduling in a cellular system. ing eigenmodes Γ with n 1, ..., (m 1), (m+1), ..., αN n ∈ { − T } In combination with fair, interference-aware scheduling from the other terminals in the active set , yielding a matrix M policies, it has been shown that users profit from almost of dimension (αNT 1) αNT doubled spectral efficiencies in the MIMO 2 2 system, as − × × ˜ H H H H H compared to the single-input single-output (SISO) setup [1]. Γm = [Γ1 ... Γm 1 Γm+1 ... ΓαNT ] (3) − ˜ B. Downlink cooperation: MET-based pre-coding Performing the SVD of Γm yields There are several concepts for cooperative downlink trans- ˜ ˜ 1 0 H Γm = Um Σ˜ m 0 V˜ v˜ , (4) mission, all imposing different demands on the system archi- m m 0 tecture. As a basic requirement, coherent downlink transmis- where v˜m corresponds to the eigenvector  associated with sion is mandatory. Thus, downlink transmission from all BSs the null space of Γ˜ m. Note, in principle the null space is has to be synchronized with respect to the carrier frequency represented by a matrix of dimensions λ αN . Since we × T and the frame start. A basic concept to achieve this kind of limit each user to be served on its dominant eigenmode only, synchronization has been presented in [5]. i.e. λ = 1, v˜0 is of dimension 1 αN and thus a vector. m × T As a reference design for downlink cooperation in cellular This vector is used for pre-coded transmission to user m, systems, we use an approach based on MET, which is known which ensures that all other users in do not experience any M to achieve a near-optimum pre-coding performance for an interference from this beam under ideal conditions. Note that isolated cell cluster [2]. For MET, the MTs report the λ NR the block-diagonalization constraint αNT m NR(m) ≤ ≥ ∀ ∈M P 2009 International ITG Workshop on Smart Antennas – WSA 2009, February 16–18, Berlin, Germany

is relaxed by the use of dominant eigenmodes, resulting in policy as proposed in [7]). Resource assignment is then done αN instead [2]. for each RB individually by assigning each beam to the user T ≥ |M| The selected pre-coding matrix on a RB and time slot is providing minimum score for that beam. given by For the MET-based pre-coding, each MTs determines the 0 0 0 C = [v˜1 ... v˜m ... v˜αNT ] , (5) effective eigenmode channel Γm per RB and conveys this information to the CPU. At the CPU, an equivalent score-based where tr CCH = p . As the beamforming vectors are i scheduling process is carried out for each of the αN spatial unitary, C implicitly includes the constraint of equal transmit T   layers that are available for transmission in a RB. For the first power per beam and sum power per cell cluster.1 layer, only the effective eigenmode channel Γm per user is IV. LINEAR MMSE RECEIVER considered, whereas for all successive layers the correlation of the effective channels from all possible user pairings is taken Assuming a linear equalizer wu at the MTs, which is m into account. The applied algorithm is closely related to the required to extract the useful signal xi,u from y according to (2), yields a post-equalization signal to interference and noise greedy user selection approach from [2], but has been modified to be applicable with the score-based scheduling technique. ratio (SINR) at the MT for stream xi,u given by Thus, a direct comparison with the baseline system presented H H w hi,uh wu in [1] is enabled. After the proper user selection together with = p u i,u SINRu i H , (6) the corresponding precoding beams in (5) have been found per wu Zuwu RB, the SINR according to (6) can be determined, yielding the where Z is the covariance matrix of the u achievable performance in terms of the spectral efficiency of interfering signals aggregated in ζi,u and zi,u, i.e. H the system. Zu = E (ζi,u + zi,u)(ζi,u + zi,u) , with E[.] being Note that for both precoding concepts, each user is served on the expectationh operator. i a single spatial stream in a RB only, i.e. only pure MU-MIMO In this work, we consider minimum mean square error access is considered without allowing spatial multiplexing (MMSE) equalization for the purpose of inter-cell interference transmission to any of the multi-antenna MTs. This constraint suppression at the receiver side according to is motivated by findings from [2], [8], [9], where the authors emphasize that MU-MIMO transmission mode is preferential p R 1Hmb MMSE i yy− i i,u as long as the number of users K is sufficiently large. wm = , (7) αNT m A. Note on the applied scheduler where Ryy denotes the covariance matrix of y from (2), m m H i.e. Ryy = E y (y ) . This receiver yields a post- The objective of the score-based resource allocation process is to assign each user his best RBs and therein his favoured equalization SINRh on a giveni RB for user m precoding beams, thus guaranteeing a certain amount of fair- pi m H 1 m ness for users during scheduling. Clearly, the process is of SINRm = [Hi bi,u] Zu− Hi bi,u (8) αNT heuristic nature, and hence the global scheduling target of In a practical context, the covariance matrix assigning each user an equal amount of resources is achieved m m H on average only or if the number of available resources Ryy = Zu + Hi bi,u (Hi bi,u) (9) tends to infinity. However, its convenient property for practical may be obtained by using multi-cell channel estimates based applications is its flexible utilization, as the set of resources can on common and dedicated reference signals [6]. Thus, each be defined over arbitrary dimensions (time/frequency/space). terminal is able to combat residual CCI from surrounding cells Thus, fairness can be established on a small time scale, e.g. aggregated in zi,u from (2). even for the scheduling of resources contained within a single OFDM symbol. V. RESOURCE ALLOCATION AND FAIR USER SELECTION Resource allocation is conducted based on a score-based VI.SIMULATION ENVIRONMENT scheduling process at a central processing unit (CPU) con- The performance is investigated in a triple-sectored hexago- trolling the cell cluster of size α. For the baseline system, nal cellular network with 19 BSs in total. The extended spatial each MT firstly evaluates the achievable SINR conditions per channel model (SCME) with urban macro scenario parameters fixed DFT-beam according to (6) based on the current channel is used [10] yielding an user’s geometry for the center cell, conditions in each RB. After conveying this information to refer to Fig. 2 (right), which is equivalent to [11]. The basic the BS2 , the collected per-beam rates from any user are system settings for our simulations are summarized in Table I. ranked over all available RBs according to their quality, and For the evaluation of cooperative transmission strategies, we corresponding scores are assigned (score-based scheduling have to take users from surrounding cells into account. For a realistic performance evaluation of the cellular network, we 1 Note that we are not considering any optimal power allocation scheme employ a wrap-around, which ensures that the interference here. 2As for the baseline system α = 1 holds, the CPU may be located directly scenario follows independent identically distributed (i.i.d.) at the BS. statistics for all users. 2009 International ITG Workshop on Smart Antennas – WSA 2009, February 16–18, Berlin, Germany

TABLE I SIMULATION ASSUMPTIONS. efficiency per sector. The achievable rates are determined from parameter value Shannon’s formula, which represent theoretical limits in a practical system. The receiver is assumed to to have perfect channel model 3GPP SCME type Monte Carlo channel state information (CSI), and feedback given by the scenario urban-macro MTs is considered to be error free and not affected by any additional modifications LOS-NLOS propagation mix delays. traffic model full buffer fc 2 GHz VII.RESULTS frequency reuse 1 signal bandwidth 18 MHz, 100 RBs Beamforming and cooperation gains: Fig. 3(a) depicts the intersite distance 500m spectral efficiency of a system with a sparse user distribution, number of BSs 19 having 3 sectors each in particular K = 2 users per sector. For reference purpose, we NT ; spacing 1,2 ; 4λ include results obtained for a non-cooperationg SISO system, transmit power 46 dBm i.e. without any optimizations at receiver and transmitter side. sectorization triple, with FWHM of 68◦ The baseline is configured as a MIMO 2 2 system. The BS height 32m × spectral efficiency for this setup is given as a blue line. By NT ; spacing 1,2 ; λ/2 MT height 2m comparing this performance with a non-cooperative MET- based pre-coded system (α = 1, green line), we observe a similar throughput. In the practical context, DFT-based pre- coding would be favorable due to complexity issues. How- ever, this DFT-based pre-coding approach lacks the ability of downlink cooperation, which is intended to provide additional performance gains due to intra-cluster interference mitigation of ζi,u. These gains are shown by the results from MET- based downlink BS cooperation using α 2, 3 BS sectors. ∈ { } We observe that MET benefits from additional degrees of freedom due to an increased number of transmit antennas spread over different BS locations, and thus enables additional beamforming gains. The relative gain with respect to the DFT Fig. 2. Left: Triple-sectored cellular setup. The region bounded by the red system amounts to 59% for MET and α = 3. and blue dashed line indicates a cooperative cell cluster jointly serving their terminals on different RBs. Right: User geometries obtained from the center Multi-user selection diversity: In the following, the perfor- cell with id= 1,2,3 and parameters according to Table I. mance of the same system setup, but with K = 10 users per { } sector is given in Fig. 3(b). For all system settings considered As an initial step, we place up to 20 terminals per sector in this work, we clearly observe a significant performance gain inside the inner cell with cell ID= 1,2,3 . Since evaluation in due to multi-user selection diversity. { } this work is limited to linear pre-coded beams, up to αNT In case of DFT-based pre-coding, the system beneficially uses user may be served in the same RB. NT corresponds to the the selection diversity for the purpose of interference-aware available spatial dimensions per sector antenna array. If the scheduling. This scheduling approach can be seen as a simple number of terminals per sector exceeds NT , interference-aware way of interference avoidance applicable in a cellular system. score-based scheduling is applied, which increases the system Therefore, the scheduler is taking channel quality identifier throughput further by exploiting multi-user diversity [1]. (CQI) information into account, which is extracted from the The users in the cooperating cell of a cluster are generated feedback provided by the MTs. This results in an intra-cluster interference ζ = 0. The median spectral efficiency amounts by dropping the users in the center cell constituted from the i,u 6 sectors with ID= 1,2,3 and then shifting the origin of the cell to approx. 5.1 bit/s/Hz in this case. { } topology into the desired direction. In particular, if cells 1,3 Turning the focus to MET-based pre-coding, we observe and 8 (indicated by the blue framed region in Fig. 2 (left)) form that this pre-coder realizes higher benefits from using multi- a cooperation cluster, the users distributions in cell 1 and 3 are user selection diversity, refer to blue and green dashed lines in generated without any shift of the origin, while for the user in Fig. 3(b). The near-optimum pre-coder mainly benefits from cell 8, the center cell is shifted to the north-east direction, as the diversity within the user grouping process, which results in illustrated by the light blue region in Fig. 2 (left). The figure a orthogonal service to different users. Thus, the MET system 3 depicts another clustering given as red framed region. The mitigates the intra-cluster interference , while ensuring user active user set in each cluster is indicated as blue triangles orthogonalization with low costs in receive power reduction. In M and red dots, respectively. contrast to the DFT-based pre-coded system, the scheduler in Performance is evaluated for the sum throughput in a 3Since the MMSE equalizer considers full CCI, ζ + z , optimum α i,u i,u specific cell cluster of size . This value is divided by receiver weights are determined, so that the intra-cluster interference ζi,u the signal bandwidth and α, yielding an effective spectral cannot completely be forced to zero. 2009 International ITG Workshop on Smart Antennas – WSA 2009, February 16–18, Berlin, Germany

1 scheduling, which is a simple way of interference avoidance in 0.95 cellular systems. In contrats to that, the multi-user diversity in the MET system enables to determine most suited user groups 0.75

abscissa) with lowest correlation between different users’ eigenmodes. ≤ Thus, the MET system mitigates the intra-cluster interference, 0.5 while ensuring user orthogonalization with low costs in receive power reduction per data stream thanks to multi-user selection SISO, α=1 0.25 DFT, α=1 diversity. MET, α=1 MET, α=2 ACKNOWLEDGEMENTS P(spectral efficiency 0.05 MET, α=3 0 0 2 4 6 8 10 12 The authors are grateful for financial support from the spectral efficiency [bit/s/Hz] German Ministry of Education and Research (BMBF) in the (a) K = 2 users per sector national collaborative project EASY-C under contract No. 01BU0631. 1 0.95 REFERENCES

0.75 [1] L. Thiele, M. Schellmann, W. Zirwas, and V. Jungnickel, “Capacity

abscissa) scaling of multi-user MIMO with limited feedback in a multi-cell ≤ environment,” in 41st Asilomar Conference on Signals, Systems and 0.5 Computers. Monterey, USA: IEEE, Nov. 2007, invited. [2] F. Boccardi and H. Huang, “A near-optimum technique using linear precoding for the MIMO broadcast channel,” Acoustics, Speech and Signal Processing, 2007. ICASSP 2007. IEEE International Conference 0.25 DFT, α=1 MET, α=1 on, vol. 3, pp. III–17–III–20, April 2007. MET, α=2 [3] L. Thiele, M. Schellmann, T. Wirth, and V. Jungnickel, “Cooperative P(spectral efficiency 0.05 MET, α=3 multi-user MIMO based on reduced feedback in downlink OFDM sys- 0 0 2 4 6 8 10 12 tems,” in 42nd Asilomar Conference on Signals, Systems and Computers. spectral efficiency [bit/s/Hz] Monterey, USA: IEEE, Nov. 2008. [4] J. Winters, “Optimum combining in digital mobile radio with cochannel (b) K = 10 users per sector interference,” IEEE Journal on Selected Areas in Communications, vol. 2, no. 4, pp. 528–539, 1984. Fig. 3. Achievable spectral efficiency (Shannon) for different types of pre- [5] V. Jungnickel, T. Wirth, M. Schellmann, T. Haustein, and W. Zirwas, coding; with and without multi-user selection diversity. “Synchronization of cooperative base stations,” in IEEE International Symposium on Wireless Communication Systems 2008 (ISWCS08), Oct. 2008. the BS takes the correlation of different users’ effective eigen- [6] L. Thiele, M. Schellmann, S. Schiffermuller,¨ and V. Jungnickel, “Multi- cell channel estimation using virtual pilots,” in IEEE 67th Vehicular mode channels Γm into account. Further, Fig. 3(b) depicts the Technology Conference VTC2008-Spring, Singapore, May 2008. achievable spectral efficiency for the cooperative MIMO 2 2 [7] T. Bonald, “A score-based opportunistic scheduler for fading radio × for α 2, 3 . The relative gains compared to the DFT system channels,” in 5th European Wireless Conference, Feb. 2004. ∈ { } [8] L. Thiele, M. Schellmann, T. Wirth, and V. Jungnickel, “Interference- increase to 72% for MET and α = 3, substantiating that the Aware Scheduling in the Synchronous Cellular Multi-Antenna Down- cooperation gains increase in the context of muti-user diversity. link,” Apr. 2009, invited. In this case the median achievable spectral efficiency amounts [9] M. Schellmann, L. Thiele, T. Wirth, T. Haustein, and V. Jungnickel, “Resource Management in MIMO-OFDM systems,” in OFDMA: Fun- to approx. 8.8 bit/s/Hz. damentals and Applications, T. Jiang, L. Song, and Y. Zhang, Eds. CRC Press, Taylor&Francis Group, 2009. VIII.CONCLUSION [10] 3GPP TR 25.996 V7.0.0, “Spatial channel model for multiple input multiple output (MIMO) simulations (release 7),” July 2007. [Online]. In this work we compared different pre-coding techniques Available: http://www.tkk.fi/Units/Radio/scm/ and their achievable performance in a cellular MU-MIMO [11] H. Huang, S. Venkatesan, A. Kogiantis, and N. Sharma, “Increasing OFDM downlink. As a baseline, we considered DFT-based the peak data rate of 3G downlink packet data systems using multiple beamforming as suggested in [1]. To evaluate the additional antennas,” vol. 1, april 2003, pp. 311–315 vol.1. beamforming gains enabled by a near-optimum pre-coder, we use MET-based beamforming under the assumption of ideal knowledge of the users’ effective eigenmode channels. In non- cooperative system with a sparse user distribution, we observe equivalent spectral efficiencies for both, DFT- and MET-based pre-coding. However, the downlink cooperation among α = 3 sectors using MET is able to provide up to 59% performance gain compared to the non-cooperative DFT pre-coded sys- tem. Multi-user selection diversity was shown to enhance the system performance significantly, while MET turned out to benefit more from this diversity than DFT-based pre-coding. The gains for DFT beams are obtained by interference-aware