This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts

for publication in the WCNC 2008 proceedings.

Non-Unitary Codebook Based Precoding Scheme for

Multi-User MIMO with Limited Feedback
Fang Shu1,2, Li Lihua1,2, Cui Qimei1,2 and Zhang Ping1,2
1. Key Laboratory of Universal Wireless Communications, Ministry of Education
2. Wireless Tech. Innovation Institute
Beijing University of Posts and Telecom., Beijing 100876,China
Email: [email protected]

AbstractIn this paper, we propose a non-unitary codebook the single-user environment, the optimal power allocation that
based precoding scheme for multi-user MIMO (MU-MIMO) maximizes the mutual information of the MIMO channel is
downlink transmission with limited feedback. The proposed MU- achieved through waterfilling over its eigenvalues. For the
MIMO scheme will perform scheduling and precoding just MU-MIMO channel, the problem is more complex. The sum
according to the limited feedback of the signal to interference and capacity in a multi-user broadcast channel is defined to be the
noise ratio (SINR) of each user to enhance the system capacity. A maximum aggregation of all users data rates. It requires
precoder codebook design method is related to Grassmannian maximizing each users data rate and minimizing the co-
line packing criterion. We group vectors from the Grassmannian channel interference (CCI) to other users. It was proven that
codebook into precoding matrices according to the correlation when a transmitter has advance knowledge of the interference,
coefficient of the vectors to suppress the multi-user co-channel
it could design a code to compensate for it [4]. An approach
interference (CCI). It terms as non-unitary precoding and
outperforms the traditional single user MIMO system as well as
that was defined as dirty paper coding (DPC) [4][5] by non-
the unitary codebook based MU-MIMO scheme with limited linear preprocessing was developed. Although DPC can
feedback and low complexity. achieve the sum capacity of multi-user broadcast channel,
deploying DPC in real-time systems is difficult because of the
Keywords-multi-user MIMO, Grassmannian line packing, non- complicated encoding and decoding schemes.
unitary precoding, SINR, limited feedback Many linear precoding MU-MIMO techniques are emerged
recently. Such as the channel inversion method [6], the
I. INTRODUCTION diagonalization (BD) method [7][8], the unitary codebook
Multiple-input multiple-output (MIMO) communication based precoding method [9] and other precoding methods
technique has attracted enormous interest for the next- [10][11]. Channel inversion method is simple but has capacity
generation wireless communications because of its potential for limit. When the spatial correlation increases, the multi-user
high capacity. The traditional MIMO system is focused on the channel capacity decreased rapidly. BD can perfectly cancel
point-to-point single-user MIMO (SU-MIMO) such as the CCI, but has antenna constraint at the base station and the
spatial diversity (STC: Space-Time Coding) and the spatial mobile station [8] and the computation burden for system is
multiplexing (VBLAST: Vertical Bell Lab Layered Space- very heavy when the number of users is large. Both channel
Time) techniques. Spatial diversity technique is employed to inversion method and the BD method are based on the
combat the channel fading, meanwhile robust to various feedback of the MIMO channel matrix, so the feedback is very
channel environments [1], but has limit to the transmission rate. large. The unitary codebook based precoding method is firstly
Spatial multiplexing technique can obtain considerable system proposed in 3GPP LTE (Long Term Evolution) [12]. It can
throughput by transmitting parallel data streams. However, it suppress the CCI efficiently through orthogonal precoding
has limit to the transmission environment that when the spatial vectors and increase the MU-MIMO system capacity with the
channel correlation is high, the performance degrades severely limited feedback. It is much more practical for its simplicity,
[2]. When the base station (BS) and also the mobile station low feedback and has become a hot research topic in LTE.
(MS) are equipped with multiple antennas, the space dimension In this paper, we proposed a non-unitary codebook based
can be exploited besides time dimension and frequency precoding method for MU-MIMO downlink transmission with
dimension for transmission the data streams of multiple users. limited feedback and low complexity. The proposed non-
As a result, the study of point-to-multipoint MIMO system unitary codebook based precoding scheme exploits efficient
termed as the multi-user MIMO (MU-MIMO) has emerged multi-user diversity gain and precoding gain by scheduling and
recently as a hot research topic. Such system has many precoding at the BS. The adoption of Grassmannian precoding
advantages that the multiple antenna diversity gain can enhance [13] aims to provide strong precoding gain for each user. We
the bit error rate performance and the multiple antenna spatial construct the precoding matrices according to the correlation
multiplexing gain can enlarge the channel capacity area. coefficient of the Grassmannian vectors to suppress the CCI.
Telatar generalized the Shannon capacity of the single-input So, it terms as non-unitary precoding since the vector pair in
single-output channel to the MIMO case [3]. He showed that in Grassmannian codebook is not orthogonal. The proposed
This work was supported by the National Natural Science Foundation of
China (60702051), the National High Technology Research and Development
Program of China (2007AA01Z261) and the Huawei Foundation of TDD LTE
Reaserch (YBWL2005025).

1525-3511/08/$25.00 2008 IEEE

 This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2008 proceedings.

scheme shows considerable system capacity improvement than MS, the elements of n k obeys distribution CN (0, N 0 ) and is
the traditional SU-MIMO system and outperforms the unitary spatially and temporarily white. w is the M K precoding
codebook based precoding scheme [9].
matrix, which contains K precoding vectors. vi is the M 1
II. SYSTEM MODEL OF THE CODEBOOK BASED MU-MIMO column vector with unit norm viH v i = 1 and []T denotes the
matrix transposition. s is the transmit vector with K data
We consider a MU-MIMO system with M transmit
antennas at the BS and N receive antennas at each MS as streams. pk is the transmitted signal power of the k-th data
shown in figure 1. There is a codebook known prior by both stream, p0 is the total transmit power.
the BS and the MS. If X users are waiting to be scheduled at
For the MU-MIMO system, the virtual MIMO channel
the BS, the scheduler will determine K ( K M ) users to be
matrix of user k after precoding can be assumed as
transmitted according to the scheduling criterion and the
feedback information. At the same time, the scheduler will  =H w
H (6)
k k
select a precoding matrix w from the codebook to precode for
the scheduled K users before transmission. For traditional MIMO detection, a linear receiver is
designed to detect the transmit data. Zero Forcing (ZF) or
1
2
Minimum Mean Squared Error (MMSE) detection criterions
User1 are commonly employed. In order to obtain good performance,
...

User1 1 N
User1 data
linear MMSE receiver is adopted at MS in this paper.
Scheduler

User2 data User2 2

Precoder

1
2  for the k-th user or the k-th
The linear MMSE receiver G
...
...

...

User2 k
...

N
UserX data UserK M data stream is
1
2
UserK
 = h H (H
G  H  H + KN 0 I )1 (7)
...

k k k k N
N p0
Feedback of Channel Information
  = [H w ]
h k = H k v k =  H (8)
k k k k
Figure 1. The configuration of MU-MIMO system

The MIMO channel matrix H k for the k-th MS is as follow. Where () 1 indicates the inverse of the matrix, () H is the
matrix conjugation transposition and I N is the N N identity
 h1,1 ! h1, M 
  matrix. v k is the precoding vector for the k-th user and
Hk =  # % #  (1)
 hN ,1 ! hN , M 
[]k denotes the k-th column of the matrix.
 
The received signal for k-th user can be indicated as
Where hi , j indicates the channel impulse response coupling K
the j-th antenna at the BS to the i-th antenna at the MS and its yk = pk H k v k s k +  pi H k vi si +n k (9)
amplitude obeys independent and identically Rayleigh- i =1, i k
distribution.
The middle part of (9) is the CCI from other users and the last
The orthogonal frequency division multiplexing (OFDM) part is the noise. Then the detected data stream sk for the k-th
technique has become one of the brilliant techniques in the MS is
next generation of wireless communications. Since OFDM
technique can deal frequency selective fading as flat fading, so  y
sk = G (10)
k k
in this paper, we model the MIMO channel as the flat MIMO
fading. The detected SINR for the k-th user with the linear receiver
is
The received signal at the k-th MS can be denoted as
 H v
pk G
2
y k = H k ws + n k (2) k k k
SINRk = K
(11)
 H v 2+ G


2
T pi G N0
s =  p1 s1 p2 s2 ... pK sK  (3) i =1, i k
k k i k 2

w = [ v1 v2 ... v K ] (4) Where 2

denotes the matrix two-norm.

p
k =1
k = p0 (5) III. THE UNITARY PRECODING CODEBOOK
The unitary precoding codebook contains a set of unitary
Where H k is a N M wireless MIMO channel matrix from precoding matrices [9] as
the BS to the k-th MS, y k is the received vector of the k-th E = {E0 ,..., E L 1} (12)
 This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2008 proceedings.

Where El is the (l+1)-th precoding matrix in the codebook V = arg max (X) (18)
N
E. XU Mt

El = e(0l ) ... e (Ml )1  (13) The codebook V consists of Nt constant transmit

precoding weights, which can be denoted as a P Nt
Where e (ml ) is the (m+1)-th precoding vector of the (l+1)-th Grassmannian codebook.
precoding matrix and M is the transmit antenna number. According to the above criterion, the Grassmannian line
1 T packing codebook with different P and Nt could be obtained.
e m(l ) = u0(lm) ... u((Ml ) 1) m  (14)
M
V. NON-UNITARY CODEBOOK BASED PRECODING FOR MU-
2 n l MIMO WITH LIMITED FEEDBACK
u (l )
nm = exp{ j (m + )} (15)
M L
We construct a codebook W  \w1 ,..., w L ^ of L
The generation of the precoding matrices is based on precoding matrices provided for users. It is known prior at
Fourier transform. The precoding matrices are unitary matrices both the BS and each MS. The generation of a precoding
since Ei EiH = I . According to the above criterion, the unitary matrix w l (l=1L) is processed by selecting M ( M Nt )
precoding codebook with arbitrary antenna number M and different vectors from a M Nt Grassmannian codebook. So
matrix number L could be generated.
w l is a M M dimensional matrix. There will be Q = CNMt
IV. GRASSMANNIAN LINE PACKING CRITERION different precoding matrices in total. Where CNMt denotes the
The Grassmannian line packing problem is the problem of total methods of selecting M vectors from Nt vectors. The
optimally packing one-dimensional subspaces, which results
precoding terms as non-unitary precoding since w l w lH I and
in finding the set or packing of Nt lines in ^ that has
there is viH v i = 1 . We order these precoding matrices
maximum minimum distance between any pair of lines [13].
Where Nt is the packing number or vector number of the according to as follow
Grassmannian codebook. We represent a packing of Nt lines K K

= v H
vj (19)
in ^ by a P N t matrix V =  v1 ... v Nt  whose v2 i =1 j = i +1
i

column space vi is the i -th line in the packing. vi is a P 1 The L precoding matrices are selected prior with small
unit vector with viH v i = 1 and v iH v j 1 ( i j ). It is proved from the totally Q matrices, since small will result in
that maximizing the transmitting precoding gain small CCI.
2
Gvi (H ) = arg max Hv i 2
with a finite codebook vectors For each user, the received SINR not only depends on its
1i N
corresponds to maximizing the minimum distance between own precoding vector, but also on other users precoding
any pair of lines spanned by the codebook vectors. vectors of the same precoding matrix according to (11). If user
k obtains information about the precoding vectors and its
Defining a distance function d ( v1 , v 2 ) by letting the channel matrix H k , the SINR of user k can be computed. Thus,
distance between the two lines generated from unit vectors v1 for a precoding matrix w l , it results in M different SINR for
and v 2 be the sine of the angle 1,2 between the two lines. each user. Each SINR corresponds to precoding by one of the
The distance is vectors in w l . The codebook W  \w1 ,..., w L ^ with L non-
2 unitary matrices results in M L different SINR for each user.
d ( v1 , v 2 ) = sin(1,2 ) = 1 v1H v 2 (16) Then, each user selects one best SINR of a precoding matrix
w l , and L best SINR are selected for each user with the
The minimum distance of a packing is the sine of the
smallest angle between any pair of lines. codebook W  \w1 ,..., w L ^ . And they will feed back L best
SINR and the corresponding vector index of the codebook to
Grassmannian packing criterion is that designing the set of the BS.
codebook vectors {v i }iN=t1 such that the corresponding
We propose a criterion to maximize the system capacity of
codebook V maximizes MU-MIMO according to the feedback SINR and vector index
2 of each user. The aim of the proposed scheme is to select
(V ) = min 1 v kH vl (17) proper users and a precoding matrix w l from the codebook
1 k < l N t
W . The selected users will be precoded by w l to share the
So, a Grassmannian line packing codebook is designed by
solving [13] same time and frequency resources to maximize the system
capacity.
 This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2008 proceedings.

The system capacity of the MU-MIMO depends on the a single user. However, for MU-MIMO, the BS selects K = 2
received SINR of the scheduled users as uses to transmit one data stream for each user simultaneously.
K
CMU ( H k , w l ) =  log 2 (1 + SINRk (H k , w l ) ) (20) 1
SU-MIMO
k =1 Propos ed MU-MIMO with L=1
0.9 Propos ed MU-MIMO with L=2
Propos ed MU-MIMO with L=4

We proposed maximum capacity criterion to select 0.8

Propos ed MU-MIMO with L=8

precoding matrix and the scheduled users as 0.7

{w, k = 1,..., K } = arg max CMU (H k , w l ) (21) 0.6

CDF
wW , k X
0.5

The detail procedure is as follow 0.4

1) Constructing the non-unitary precoding codebook 0.3

W  \w1 ,..., w L ^ according to the above illustration. 0.2

0.1

2) Each MS computes M L SINR according to the non- 0

4 5 6 7 8 9 10

unitary precoding codebook and its channel matrix, then Capacity(bps/Hz)

feeds back L best SINR correspond to L non-unitary

matrices and the L vector indexes to the BS. Figure 2. System Capacity CDF of the SU-MIMO and the proposed non-
unitary MU-MIMO scheme
3) The BS gathers the feedback information from each MS.
From the simulation results, we can see that the proposed
4) According to all feedback SINR values and index scheme can effectively improve the MU-MIMO system
information, the scheduler at the BS will select k=1,K capacity. It provides more than 1 bps/Hz capacity gain than the
users who declare the same precoding matrix w with SU-MIMO system when L = 4 with the feedback of L SINR
different precoding vector index in the precoding matrix values and L vector indexes. For a MU-MIMO system with M
to maximize the system capacity. That is the selected transmit antennas, the vector index feedback just requires
users declare their precoding vectors in a same precoding log 2 M bits for each index, since each precoding matrix
matrix w and its precoding vector with v k = [ w ]k to consists M precoding vectors. With the increasing of the
obtain the maximal system capacity CMU . codebook size L, the system capacity of the proposed MU-
MIMO scheme will increase as well as the feedback.
5) The precoder at the BS will perform precoding to the
selected users with the selected precoding matrix 1
SU-MIMO

w before transmission. 0.9

Unitary precoding with L=8
Non-unitary prec oding with L=8

0.8

VI. SIMULATION RESULTS 0.7

The proposed non-unitary codebook based MU-MIMO

CDF

0.6

scheme, the traditional SU-MIMO scheme and the unitary 0.5

codebook based MU-MIMO scheme were simulated. 0.4

For SU-MIMO system, the scheduler selects the user with 0.3

the maximum Shannon capacity among the users waiting to be 0.2

scheduled, so the maximum SU-MIMO system capacity with 0.1

equal power allocation is [14]
0
4 5 6 7 8 9 10

 n
 p 
Capacity(bps/Hz)

CSU = max  log 2 1 + 0 ik  (22)

k
 i =1  N0 M  Figure 3. System Capacity CDF of the three schemes

Figure 3 shows the CDF of the system capacity for the SU-
Where ik is the i-th eigenvalues of H Hk H k , n=min{ M , N }. MIMO, the unitary codebook based MU-MIMO and the
Figure 2 is the cumulative distribution function (CDF) of proposed non-unitary codebook based MU-MIMO with
the system capacity for SU-MIMO and the proposed non- M = 2 , N = 2 , K = 2 , X = 10 , p0 / N0 = 5dB , linear
unitary codebook based MU-MIMO with different codebook MMSE detection and equal power allocation. The codebook
size L. M = 2 , N = 2 , equal power allocation is applied at size with L=8 is employed for both the two MU-MIMO
BS, linear MMSE detection is applied at each MS, X = 10 schemes. 2 64 Grassmannian codebook is employed for the
users at the base station waiting to be scheduled and proposed scheme.
p0 / N0 = 5dB for both schemes. In order to obtain small The proposed non-unitary codebook based MU-MIMO
for the non-unitary precoding matrices, 2 64 Grassmannian scheme also outperforms the unitary codebook based MU-
codebook is adopted for the proposed scheme. For SU-MIMO, MIMO scheme, especially when L is large. Because the non-
the BS selects just one user to transmit K = 2 data streams for unitary codebook will be close to the Grassmannian space
 This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2008 proceedings.

when L is large, thus the Grassmannian precoding provides detection, equal power allocation are employed for all the
larger precoding gain than the unitary precoding method, three schemes and no channel coding for them. The codebook
while both schemes could suppress the CCI efficiently. size with L=8 is employed for both the two MU-MIMO
schemes. 2 64 Grassmannian codebook is employed for the
8
SU-MIMO
proposed scheme. The result shows that the proposed scheme
Unitary precoding with L=8
Non-unitary precoding with L=8 provides a significant improvement than SU-MIMO and
7.5 outperforms the unitary codebook based MU-MIMO.
Average System Capacity (bps/Hz)

7
VII. CONCLUSIONS
6.5 In this paper, a non-unitary codebook based precoding
scheme for multi-user MIMO downlink transmission with
6 limited feedback is proposed. The proposed scheme constructs
a codebook of non-unitary precoding matrices by grouping
5.5 Grassmannian precoding vectors according to their correlation
coefficient. This method just requires the limited feedback
5
2 4 6 8 10 12 14 16
information to exploit multi-user diversity gain and the
User Number precoding gain as well as suppressing the CCI effectively. It
provides significant improvement to the system capacity than
Figure 4. Average system capacity CDF of the three schemes
the traditional SU-MIMO scheme and also outperforms the
Figure 4 describes the average system capacity of the SU- unitary codebook based MU-MIMO scheme.
MIMO, the unitary codebook based MU-MIMO and the
proposed non-unitary codebook based MU-MIMO scheme REFERENCES
with different number of users to be scheduled at the BS. [1] E.A. Jorswieck, A. Sezgin, Impact of spatial correlation on the
M = 2 , N = 2 , K = 2 , p0 / N0 = 5dB , linear MMSE performance of orthogonal space-time block codes, Communications
detection and equal power allocation are employed for all the Letters, vol.8, no. 1, 2004, pp. 21~23.
three schemes. The codebook size with L=8 is employed for [2] P. Tarasak, Hlaing Minn, V.K. Bhargava, Linear prediction receiver for
differential space-time block codes with spatial correlation,
both the two MU-MIMO schemes. 2 64 Grassmannian Communications Letters, vol. 7, no.11, 2003, pp. 543~545.
codebook is employed for the proposed MU-MIMO scheme. [3] I. E. Telatar, Capacity of multi-antenna Gaussion channels, Eur.
The user number X scales from 2 to 16 for the three schemes. Trans. Telecommunn., vol. 10, no. 6, Nov./Dec. 1999, pp. 585-595.
[4] M. Costa, Writing on dirty paper, IEEE Trans. Information Theory,
When more users X are waiting to be scheduled, the multi- vol. 29, no.3, May 1983, pp. 439-441.
user diversity gain for MU-MIMO system will increase and [5] U. Erez, St. Brink, A close-to-capacity dirty paper coding scheme,
the maximum Shannon capacity of SU-MIMO system will IEEE Transactions on Information Theory, vol. 51, no. 10, Oct. 2004, pp.
3417-3432.
increase too. With the increasing number of users to be [6] T. Haustein, C. von Helmolt, E. Jorswieck, et al, Performance of
scheduled at the BS, the performance of the proposed scheme MIMO systems with channel inversion, IEEE 55th VTC, vol. 1, May
will increase simultaneously. It is clear that the proposed 2002, pp. 35-39.
scheme exploits multi-user diversity gain and the precoding [7] Q. H. spencer, A. L. Swindlehurst, and M. Haardt, Zero-forcing
gain to improve the MU-MIMO system capacity efficiently. methods for downlink spatial multiplexing in multiuser MIMO
channels, IEEE Trans. Signal Processing, vol. 52, no. 2, Feb. 2004,
-1
pp.461-471.
10
[8] L. U. Choi and R. D. Murch, A transmit preprocessing techinique for
multiuser MIMO systems using a decomposition approach, IEEE
Trans. Wireless Commun., vol. 3, no. 1, Jan. 2004, pp.20-24.
[9] Samsung, Downlink MIMO for EUTRA, 3GPP TSG RAN WG1
10
-2 Meeting #43, Korea, Nov. 2005, R1-051353.
[10] Eun Yong Kim, Joohwan Chun, Random beamforming in MIMO
systems exploiting efficient multiuser diversity, VTC, vol.1, 2005, pp.
BER

202-205.
[11] Alireza Tarighat, Mirette Sadek, Ali H. Sayed, A multi user
10
-3
beamforming scheme for downlink MIMO channels based on
maximizing signal-to-leakage ratios, ICASSP, vol. 1, Mar. 2005, pp.
129-132.
SU-MIMO
Unitary precoding with L=8 [12] 3GPP, Multiple Input Multiple Output (MIMO) antennae in UTRA
Non-unitary precoding with L=8
-4
(Release 7), 2007, TR 25.876.
10
0 1 2 3 4 5 6 7 8 [13] David J.Love, Robert W. Heath, Thomas Strohmer, Grassmannian
SNR (dB) beamforming for Multiple-Input Multiple-Output Wireless Systems,
Figure 5. BER performance of three schemes IEEE Transactions on Information Theory, vol. 49, No. 10, Oct. 2003,
pp. 2735-2747.
Figure 5 shows the BER performance of the SU-MIMO [14] N. Chiurtu, B. Rimoldi, Varying the antenna locations to optimize the
scheme, the unitary codebook based MU-MIMO scheme and capacity of multi-antenna Gaussian channels, ICASSP, vol. 5, Jun.
the proposed non-unitary codebook based MU-MIMO scheme. 2000, pp.3121-31.
M = 2 , N = 2 , K = 2 , X = 10 , QPSK, linear MMSE