We propose two generalized block-diagonalization (BD) schemes for multiple-input multiple-output (MIMO) relay broadcasting systems with no channel state information (CSI) at base station. We first introduce a generalized zero forcing (ZF) scheme that reduces the complexity of the traditional BD scheme. Then the optimal power loading matrix for the proposed scheme is analyzed and the closed-form solution is derived. Furthermore, an enhanced scheme is proposed by employing the minimum-mean-squared-error (MMSE) criterion. Simulation results show that the proposed generalized MMSE scheme outperforms the other schemes and the optimal power loading scheme improves the sum-rate performance efficiently.
1. Introduction
One major challenge faced by the wireless multimedia broadcasting systems is to provide high data rates for remote users located in the cell boundaries, which experience very low received signal-to-noise ratios (SNRs). An increasingly attractive and cost effective solution is the use of relay stations (RSs). Relays can be classified as full-duplex and half-duplex. Full-duplex relay is still under investigation due to its highly complex hardware implementation. For practical systems, half-duplex relay is more commonly used but suffers significant rate loss as a result of the two or more transmission phases needed to deliver a message. Multiple-input multiple-output (MIMO) technique is well known to provide significant improvement of the spectral efficiency and link reliability because of its multiplexing and diversity gains. So combining relay and MIMO techniques can utilize both of their advantages to increase the data rates of remote users.
From a general information theoretic perspective, the capacity bounds of MIMO relay channel with a single user have been analyzed in [1]. For practical implementation, [2] has investigated the optimal processing matrix at the relay in an amplify-and-forward (AF) relay MIMO system, also with a single user. When multiple antennas are deployed at base station (BS) and RS, multiple users can be scheduled at the same time for simultaneous transmissions. However, the optimal processing problem becomes more complex because of the multiuser interference (MUI). For MIMO relay broadcasting systems where each user is equipped with a single antenna, the authors in [3] proposed an implementable system architecture by exploiting nonlinear precoding at BS and linear processing at RS, while [4] proposed an iterative algorithm for jointly optimizing the precoding matrix at BS and RS to maximize the system capacity. Recently, the studies in [5] generalized the single-antenna user case to multiantenna, and a block-diagonalization (BD) based precoding scheme was applied to transform the system into multiple parallel single-user channels.
All the above research assumed that BS knows the channel state information (CSI) that is needed to perform the precoding. However, to inform BS of the CSI between RS and users would be rather challenging, especially for high-velocity users. The scenario that only RS has the CSI of the links from BS to RS and from RS to users was considered in [6] and a BD based linear processing scheme is employed at RS.
In this paper, we extend the work in [6] to a generalized form for MIMO relay broadcasting systems with no CSI at BS. The AF relay is considered due to its simplicity and practicality. We first introduce a generalized zero forcing (ZF) scheme to reduce the complexity of the traditional BD based scheme. Then the optimal power loading matrix for the proposed scheme is analyzed and the closed-form solution is derived. Furthermore, by employing the minimum-mean-squared-error (MMSE) criterion, an enhanced scheme is proposed to balance the interference and noise. Simulation results show that the proposed generalized MMSE scheme is superior to the other schemes and the optimal power loading scheme improves the sum-rate performance efficiently.
This paper is organized as follows. The system model and problem formulation are given in Section 2. In Section 3, a brief review of the traditional BD based scheme is presented. The generalized BD schemes are developed in Section 4. Simulation results and conclusions are displayed in Section 5 and Section 6, respectively.
The following notations are used throughout the paper. Boldface capitals and lowercases denote matrices and vectors. XT, XH, X-1, and X+ denote the transpose, conjugate transpose, inverse, and pseudoinverse of X, respectively. I represents the identity matrix.
2. System Model and Problem Formulation
We consider a MIMO relay broadcasting system as shown in Figure 1. In this system, the BS with NT antennas transmits independent data streams to K remote users simultaneously and an RS with NS antennas helps the communication. The kth user is equipped with Nk antennas and the total number of receive antennas is NR=∑k=1KNk. In this paper, we assume that NT=NR for simplicity. For the case NT≠NR, a user scheduling algorithm is needed and the multiuser diversity gain can be exploited. The system configuration can be described as NT×NS×N1,…,NK. We ignore the direct link between BS and each user due to very severe large-scale path loss.
System model of the MIMO relay broadcasting system.
The data transmitted from BS to the kth user is assumed to be an Nk-dimensional symbol vector xk, which is normalized as E[xkxkH]=I. Therefore, the total transmitted data vector at BS can be expressed as x=[x1T,…,xKT]T. The received data vector at RS is given by
(1)yr=PTNTHx+nS,
where PT is the transmit power at BS and H∈CNS×NT is the backward channel matrix. nS is the RS’s complex Gaussian noise vector with independent and identically distributed (i.i.d.) elements of zero mean and variance σS2.
After a linear processing, the transmitted data vector at RS is expressed as
(2)xr=Wyr=PTNTWHx+WnS,
where W∈CNS×NS is the linear processing matrix at RS. The transmit power constraint at RS can be written as
(3)trPTNTWHHHWH+σS2WWH≤PS.
Finally, the received data vector at the kth user is represented as
(4)xk=Gkxr+nR=PTNTGkWHx+GkWnS+nR,
where Gk∈CNk×NS is the forward channel matrix from RS to the kth user and nR is the user’s complex Gaussian noise vector with i.i.d. elements of zero mean and variance σR2. Equation (4) can be rewritten as
(5)xk=PTNTGkWHkxk+PTNTGkW∑j=1,j≠kKHjxj+GkWnS+nR,
where Hk∈CNS×Nk represents the channel matrix from the specific Nk transmit antennas at BS which are dedicated to the kth user to the NS receive antennas at RS.
As we can see from (5), the first term of the right-hand side indicates the desired data of the kth user and the others include the MUI and noise. To fully eliminate the MUI in (5), W should satisfy
(6)GkWHj≠0,forj=k=0,otherwise
or
(7)GWH=G1WH1000⋯000GKWHK,
where G=[G1T⋯GKT]T is the combining forward channel matrix. We define W as
(8)W=WGWH.WG and WH are the linear processing matrices for the forward channel and the backward channel, respectively. It is obvious that (7) is satisfied when we set WG=G+ and WH=H+. However, this scheme simply treats each antenna as a single receiver, thus sacrificing some benefit of the multiantenna at each user.
3. Review of the BD Based Scheme
The authors in [6] proposed a BD based scheme, where WG and WH are designed to place each user’s data stream at the null space of the other users’ channels.
To this end, we exclude the kth user’s channel matrix from G and define
(9)G¯k=G1T⋯Gk-1TGk+1T⋯GKTT∈CN¯Gk×NS,
where N¯Gk=NR-Nk. Assuming that NS≥NR, we have rank(G¯k)=N¯Gk in a rich scattering environment. The singular value decomposition (SVD) of G¯k is
(10)G¯k=UGkΣGkVGk,1VGk,0H,
where UGk∈CN¯Gk×N¯Gk is a unitary matrix and the diagonal matrix ΣGk∈CN¯Gk×NS contains the singular values of G¯k. VGk,1∈CNS×N¯Gk consists of the first N¯Gk nonzero singular vectors and VGk,0∈CNS×(NS-N¯Gk) holds the last NS-N¯Gk zero singular vectors. Thus, VGk,0 forms an orthonormal basis for the null space of G¯k and WG can be expressed as
(11)WG=VG1,0⋯VGK,0.
WH can be calculated in a similar way. We define
(12)H¯k=H1⋯Hk-1Hk+1⋯HK∈CNS×N¯Hk,
where N¯Hk=NT-Nk. Assuming that NS≥NT, we have rank(H¯k)=N¯Hk in a rich scattering environment. The SVD of H¯k is
(13)H¯k=UHk,1UHk,0ΣHkVHkH,
where UHk,0H∈C(NS-N¯Hk)×NS forms an orthonormal basis for the null space of H¯k and WH can be expressed as
(14)WH=UH1,0⋯UHK,0H.
After the determination of WG and WH, the MUI at each user is fully eliminated. Since this BD based scheme uses SVD to calculate the linear processing matrix, we term it as the BD-SVD scheme.
4. Generalized BD Schemes
In this section, we introduce the generalized BD schemes to reduce the complexity and improve the performance of MIMO relay broadcasting systems.
4.1. Generalized ZF Scheme
Although the BD-SVD scheme can fully eliminate the MUI and benefit from the multiantenna gain, the SVD operations bring along considerable computational complexity which makes it difficult to implement in practice. In order to reduce the complexity, we propose a BD based generalized ZF scheme (BD-GZF) as follows.
We first define the pseudoinverse of the forward channel matrix G as
(15)G+=GHGGH-1=G1+⋯GK+.
By performing QR decomposition (QRD) on Gk+, we get
(16)Gk+=QGkRGk,k=1,…,K,
where QGk∈CNS×Nk forms an orthonormal basis for the column space of Gk+ and RGk∈CNk×Nk is an upper triangular matrix [7]. It is observed that GjGk+=GjQGkRGk=0, for j≠k. Since RGk is nonsingular, it follows GjQGk=0. Therefore, QGk forms an orthonormal basis for the null space of G¯k, just like VGk,0 in the BD-SVD scheme. Thus WG can be obtained as
(17)WG=QG1⋯QGK.
WH can be calculated in a similar way. We define
(18)H+=HHH-1HH=H1+⋯HK+H.
By performing QRD operation on Hk+, we get
(19)Hk+=QHkRHk,k=1,…,K,
where QHk∈CNS×Nk forms an orthonormal basis for the null space of H¯k and WH can be obtained as
(20)WH=QH1⋯QHKH.
4.2. Complexity Analysis
We use the number of floating point operations (FLOPs) to measure the computational complexity. According to [8], the numbers of FLOPs required for different matrix operations are summarized as follows:
multiplication of m×n and n×p complex matrices: 8mnp-2mp;
SVD of an m×n (m≤n) complex matrix where only Σ and V are obtained: 32(nm2+2m3);
inversion of an m×m real matrix: 4m3/3;
QRD of an m×n (m≤n) complex matrix: 16(n2m-nm2+1/3m3).
The required numbers of FLOPs for the BD-SVD and BD-GZF schemes are illustrated in Tables 1 and 2, respectively.
Complexity of BD-SVD.
Steps
Operations
FLOPs
1
SVD for calculating WG
32∑k=1K(NSN¯Gk2+2N¯Gk3)
2
SVD for calculating WH
32∑k=1K(NSN¯Hk2+2N¯Hk3)
Complexity of BD-GZF.
Steps
Operations
FLOPs
1
Inversion of G
4/3NR3+16NR2NS-2NR2-2NRNS
2
QRD of WG
16∑k=1K(NS2Nk-NSNk2+1/3Nk3)
3
Inversion of H
4/3NT3+16NT2NS-2NT2-2NTNS
4
QRD of WH
16∑k=1K(NS2Nk-NSNk2+1/3Nk3)
In order to make the complexity comparison more comprehensive and intuitive, we plot the numbers of FLOPs required for the two schemes in Figure 2 as a function of user number. We assume that each user is equipped with Nk=2 antennas and NT=NS=Nk×K.
Complexity comparison with Nk=2 and NT=NS=Nk×K.
It can be seen from Figure 2 that the proposed BD-GZF scheme demands much lower computational complexity than the BD-SVD scheme. One reason is that the QRD operation is much simpler than the SVD operation in the case of same matrix dimension. A more important reason is that the SVD operations in the BD-SVD scheme are implemented K times on matrices with dimensions N¯Gk×NS and NS×N¯Hk, while the QRD operations in the BD-GZF scheme are implemented K times on matrices with dimensions NS×Nk and NS×Nk, which are much lower than the former. It is worth noting that, with the increase of the system dimension, the complexity reduction by the proposed BD-GZF scheme becomes more considerable.
4.3. Optimal Power Loading Scheme
To maximize the sum-rate, we define the power loading matrix P as
(21)P=P1000⋯000PK,
where Pk∈CNk×Nk is the power loading matrix for the kth user. With the power loading matrix being introduced, (8) can be rewritten as
(22)W=WGPWH.
By substituting (17), (20), (21), and (22) into (5), we get
(23)yk=PTNTG~kPkH~kxk+G~kPkQHkHnS+nR,
where G~k=GkQGk and H~k=QHkHHk denote the effective channel matrix for the forward and backward channels, respectively. As we can see, after the linear processing, the MIMO relay broadcasting system becomes K parallel MIMO relay systems. Motivated by the optimal design for the single-user MIMO relay system [2], we perform SVD operations on G~k and H~k as G~k=U~GKΣ~GKV~GKH and H~k=U~HKΣ~HKV~HKH. Then the optimal Pk can be expressed as
(24)Pk=V~GkΛkU~HkH,
where Λk=diag(Λk,1,…,Λk,Nk) is for allocating the power. Thus (23) can be rewritten as
(25)yk=PTNTU~GKΣ~GKΛkΣ~HKV~HKHxk+U~GKΣ~GKΛkU~HkHQHkHnS+nR.
The sum-rate at the kth user is derived as
(26)Rk=12log2detI+PT/NTΣ~HK2Λk2Σ~GK2σS2Λk2Σ~GK2+σR2=12∑j=1Nklog21+PT/NTΣ~Hk,jΛk,jΣ~Gk,jσS2Λk,jΣ~Gk,j+σR2,
where the factor of 1/2 comes from the loss of the half-duplex transmission. Σ~Hk,j and Σ~Gk,j are the jth elements of the diagonal matrices Σ~HK2 and Σ~GK2, respectively.
The transmit power constraint at RS can be written as
(27)∑k=1KtrPTNTΛk2Σ~Hk2+σS2Λk2=∑k=1K∑j=1NkΛk,jPTNTΣ~Hk,j+σS2≤PS.
Therefore, the maximizing problem of the system sum-rate R=∑k=1KRk is formulated as
(28)maxmizeΛk,jRsubjecttoΛk,j≥0,∀k,j;hhhhhhhhh∑k=1K∑j=1NkΛk,j(PTNTΣ~Hk,j+σS2)≤PS.
Problem (28) is a standard convex optimization problem and can be solved with the Lagrange multiplier method. The closed-form solution is given as
(29)Λk,j=σR2ρΣ~Hk,j2+4ρμΣ~Hk,jΣ~Gk,j/ln2σR2-ρΣ~Hk,j-2+2σS2Σ~Gk,j1+ρΣ~Hk,j,
where ρ=PT/(NTσS2) and (x)+=max(0,x). μ is a unique root of ∑k=1K∑j=1NkΛk,j(PTΣ~Hk,j/NT+σS2)-PS=0, which can be solved with a numerical root-finding algorithm, such as the bisection method.
4.4. Enhanced MMSE Scheme
It is well known that the ZF criterion eliminates the MUI without appropriate consideration of the noise and therefore some performance loss will occur. To take the noise term into account and improve the performance, an enhanced scheme termed the BD based generalized MMSE scheme (BD-GMMSE) is proposed as follows.
By applying the MMSE criterion, we replace the pseudoinverse in (15) and (18) with MMSE channel inversion as [9]
(30)Gmmse+=GHGGH+αI-1,Hmmse+=HHH+βI-1HH,
where α=NSσR2/PS and β=NTσS2/PT. Then WG and WH can be obtained in the same manner.
Note that the BD-GZF scheme relies on the pseudoinverse of G and H, thus having the dimension constraints; that is, NS≤NT and NR≤NS. However, the BD-GMMSE scheme can be employed when there are more receive antennas than transmit antennas, similar to the MMSE channel inversion.
Unlike the BD-GZF scheme, the optimal power loading matrix for the BD-GMMSE scheme is not easy to identify since the residual MUI varies according to the power loading. However, the above optimal power loading scheme can still be applied in BD-GMMSE and improve the sum-rate performance efficiently as will be shown later.
5. Simulation Results
In this section, we present simulation results to show the sum-rate performance of the proposed schemes. A system under the configuration of 6×6×2,2,2 is considered. All channels are assumed to be quasistatic flat faded and the elements are complex Gaussian variables with zero mean and unit variance. The average received SNRs per antenna at RS and at users are denoted by SNR1 and SNR2, respectively. OPL in the legend denotes the implementation of the proposed optimal power loading scheme and the bisection method is used to calculate μ. We average the sum-rate over 10000 random channel realizations.
Figure 3 shows the sum-rate performance of different schemes as a function of SNR1 with SNR2 fixed at 10 dB, and vice versa in Figure 4. It can be seen that the proposed BD-GZF scheme represents identical performance as the traditional BD-SVD scheme, while offering much lower computational complexity as analyzed in Section 4. The enhanced BD-GMMSE scheme outperforms ZF based schemes by balancing the interference and noise. The proposed optimal power loading scheme improves the sum-rate performance efficiently for both BD-GZF and BD-GMMSE schemes.
Sum-rate as a function of SNR1 with SNR2 fixed at 10 dB.
Sum-rate as a function of SNR2 with SNR1 fixed at 10 dB.
6. Conclusion
In this paper, two generalized BD schemes for MIMO relay broadcasting systems are proposed. We first extend the traditional BD based scheme to a generalized BD-GZF scheme, which reduces the computational complexity significantly. Then an optimal power loading scheme for the proposed scheme is developed to improve the sum-rate performance. By employing the MMSE criterion, an enhanced scheme is proposed to further improve the performance. Simulation results show that the proposed BD-GMMSE scheme is superior to the other schemes and the optimal power loading scheme improves the sum-rate performance efficiently.
Conflict of Interests
The authors declare that there is no conflict of interests regarding the publication of this paper.
Acknowledgments
This work is supported by the 863 Project (no. 2014AA01A701) and the National Natural Science Foundation of China (no. 61072052).
WangB.ZhangJ.Host-MadsenA.On the capacity of MIMO relay channels2005511294310.1109/tit.2004.839487MR22345702-s2.0-12444347518TangX.HuaY.Optimal design of non-regenerative MIMO wireless relays2007641398140610.1109/TWC.2007.3483362-s2.0-34247359530ChaeC.-B.TangT.HeathJ.ChoS.MIMO relaying with linear processing for multiuser transmission in fixed relay networks200856272773810.1109/tsp.2007.907821MR25124432-s2.0-39649100853XuW.DongX.LuW.-S.Joint precoding optimization for multiuser multi-antenna relaying downlinks using quadratic programming20115951228123510.1109/TCOMM.2011.041111.0907132-s2.0-79956283621LiangL.XuW.DongX.Limited feedback-based multi-antenna relay broadcast channels with block diagonalization20131284092410110.1109/twc.2013.071113.1216962-s2.0-84883245919LiuW.LiC.LiJ.-D.HanzoL.Block diagonalisation-based multiuser multiple input multiple output-aided downlink relaying20126152371237710.1049/iet-com.2011.0553MR3057713SungH.LeeS.-R.LeeI.Generalized channel inversion methods for multiuser MIMO systems200957113489349910.1109/tcomm.2009.11.0704042-s2.0-72949091527GolubG.LoanC. V.1996The Johns Hopkins University PressMR1417720PeelC. B.HochwaldB. M.SwindlehurstA. L.A vector-perturbation technique for near-capacity multiantenna multiuser communication—part I: channel inversion and regularization200553119520210.1109/tcomm.2004.8406382-s2.0-14044264157