Performance of Linear Multi-User MIMO Precoding

in LTE System
C´ assio B. Ribeiro, Klaus Hugl
Nokia Research Center
P.O. Box 407, FIN-00045
Nokia Group, Finland.
[cassio.ribeiro,klaus.hugl]@nokia.com
Marko Lampinen
Nokia Devices
P.O. Box 50, FIN-90571
Oulu, Finland.
marko.lampinen@nokia.com
Markku Kuusela
Nokia Devices
P.O. Box 407, FIN-00045
Nokia Group, Finland.
markku.kuusela@nokia.com
Abstract—For scenarios with a large number of users to
be served in one cell, high capacity gains can be achieved by
transmitting independent data streams to different users sharing
the same time-frequency resources. This technique is known as
Multi-User MIMO (MU-MIMO). In this paper we investigate the
performance of MU-MIMO operation in 3GPP LTE for different
frequency granularities of the precoder at the OFDM transmitter.
We also investigate the impact of channel correlation on the
performance of the receiver when it is unaware of the interferer’s
precoding vector. The performance is evaluated by means of semi-
static system simulations.
I. INTRODUCTION
Multiple Input Multiple Output (MIMO) techniques using
multiple antennas at the transmitter and receiver are essential
technologies for future wireless communications systems due
to the large capacity gains presented by such techniques. In
the 3GPP Long Term Evolution (LTE) system several MIMO
modes of operation are considered in order to achieve high data
rates and high system capacity [1]. The decision on which tech-
nique to use depends on characteristics of the MIMO channel,
like spatial correlation and frequency selectivity, and also on
system parameters, e.g., availability of feedback, number of
users to be served, and type of data traffic.
In particular, for scenarios where there is a large number
of users to be served in one cell, high capacity gains can be
achieved by transmitting multiple streams to different users
sharing the same time-frequency resources. This technique is
known as Multi-User MIMO (MU-MIMO) [2]. In general,
there are non-linear and linear MU-MIMO schemes. The non-
linear schemes have been mainly inspired by [3] whereas the
linear schemes are based on typical linear processing principles
such as beamforming or zero-forcing filtering [4]. Only linear
MU-MIMO precoding is considered for 3GPP LTE [1].
In 3GPP LTE it is assumed that MU-MIMO is applied only
in scenarios with high spatial correlation. If the spatial corre-
lation is not frequency-dependent, then spatial characteristics
of the channel are similar for the whole transmitted bandwidth
of 20MHz, and this fact can be used to simplify signaling and
scheduling.
In this paper we investigate the performance of MU-MIMO
operation in LTE downlink for different frequency granularities
of the precoder at the transmitter and for different values
of spatial correlation at transmitter side. We also investigate
the impact of channel correlation on the performance of the
receiver when it is not aware of the interferer’s signal. Perfor-
mance is evaluated by means of semi-static system simulations
using best effort traffic model, where at each simulation drop
the users are fixed at a random location, but the fast-fading is
changing over time.
This paper is organized as follows: in Section II an overview
of 3GPP LTE is provided. In Section III the signal model
used in the paper is described. In Section IV the precoding
design techniques are described. In Section V the Linear
Minimum Mean Square Error (LMMSE) receiver is presented.
In Section VI channel feedback mechanisms are described.
Section VII describes the scheduling algorithm assumed in this
paper. Finally, Section VIII presents simulation results based
on a semi-static system simulator.
II. 3GPP LONG TERM EVOLUTION
The standardization process for 3GPP Long Term Evolution
(LTE) is on final stages of completion. In this Section we
summarize the main aspects of LTE physical layer in dowlink
direction for Frequency Division Duplex (FDD) type of oper-
ation. Detailed specifications can be found in [5], [1].
The physical layer technology employed in 3GPP LTE
in downlink is Orthogonal Frequency Division Multiple Ac-
cess (OFDMA). Several different transmission bandwidths are
possible, ranging from 1.4 MHz to 20 MHz [6]. The 3GPP
LTE system is specified such that the subcarrier spacing is
constant for all transmission bandwidths, and equal to 15 kHz.
Resource Blocks (RB) are defined as groups of 12 consecutive
subcarriers, which corresponds to 180 kHz bandwidth for each
RB.
In 3GPP LTE, the minimum allocation in time-domain is
a subframe, which has a duration of 1 ms, consisting of
14 OFDM symbols. Hence, one RB consists of 14 OFDM
symbols over 12 subcarriers. It should be noted that not all
subcarriers contain actual data, since some of them are reserved
for reference symbols.
Several MIMO operation modes are specified in LTE down-
link, including Single-User MIMO (SU-MIMO) and Multi-
User MIMO (MU-MIMO). In SU-MIMO operation several
parallel data streams are transmitted simultaneously to one
terminal, or User Equipment (UE) is 3GPP terminology. The
978-1-4244-1653-0/08/$25.00 c 2008 IEEE ISWPC 2008
410
UE is supposed to feedback the preferred number of data
streams depending on the observed channel. The UE can also
feedback an index to a matrix in a codebook, which can be
used by eNodeB as a precoder. The precoder is chosen such
that throughput is maximized. The granularity for computation
and signaling of the precoding index can range from a couple
of RBs to the full bandwidth.
In MU-MIMO operation two or more UEs share the same
time-frequency resources. Several parallel data streams are
transmitted simultaneously, one for each UE. It is assumed
that the UE feeds back a quantized version of the observed
channel, so that eNodeB can schedule in MU-MIMO mode
those terminals with good channel separation (cf Section IV
and Section VII). The transmitted data is precoded such that
each data stream is transmitted to the corresponding UE
with maximum throughput. The precoder must be designed
jointly for all data streams, such that interference between data
streams can be minimized, as seen in Section IV.
III. SIGNAL MODEL
Let us define N
s
as the number of streams transmitted
simultaneously to each user, N
t
and N
r
as the number of trans-
mit and receive antennas, respectively, and N
u
as the number
of users transmitting simultaneously at the same subcarrier. For
each subcarrier, we can write the received symbol estimates for
the l-th user after receiver filtering as
¯ x
l
= W
R,l
H
l
N
u

u=1
W
T,u
x
u
+W
R,l
n, (1)
where the N
t
× N
s
matrix W
T,l
and the N
s
× N
r
matrix
W
R,l
are the transmitter precoding filter and receiver baseband
processing filter for the l-th user, respectively, H
l
is the
MIMO channel matrix for the l-th user, x
l
are the transmitted
symbols, and n is the circularly-complex Gaussian white noise.
The dependency on subcarrier index and time instant is not
explicitly indicated in (1), since the processing is assumed to
be performed on a subcarrier basis for each received OFDM
symbol.
IV. PRECODER COMPUTATION
The singular value decomposition (SVD) of the N
r
× N
t
channel H is given by H = UΣV
H
. Since matrices U and
V are unitary, the SVD decouples the channel into orthogonal
directions. Assuming the receiver is given by the left singular
vectors, U, and that we are interested in the channel direction
that corresponds to the largest singular value, we can write the
equivalent channel seen by the l-th user as
H
eq,l
= σ
l
v
H
l
, (2)
where v
l
denotes the first column of V and σ
l
is the first
element of the main diagonal of Σ. The equivalent channel is
quantized to a codebook before it is fed back to eNodeB [1].
The quantized version of the equivalent channel is denoted by
H
eq,l
and computed as
H
eq,l
= arg max
c
H
eq,l
c
H
, (3)
where c denotes a vector that belongs to the codebook.
Zero-Forcing (ZF) precoding is a potential precoder design
technique for DL MU-MIMO. The main benefits of ZF pre-
coding is that the interference is pre-canceled at the transmitter
side. This implies that eNodeB has most of the computational
complexity in designing the precoder, and each terminal needs
only information regarding its own data streams for reception.
The ZF precoder can be designed using the Moore-Penrose
pseudo-inverse as
W
T
= H
H
eq
_
H
eq
H
H
eq
_
−1
, (4)
where
H
eq
=
_
H
T
eq,1
. . . H
T
eq,N
u
_
T
(5)
W
T
=
_
W
T,1
. . . W
T,N
u
¸
(6)
In practice, the precoder has to be quantized to a codebook
as well, or else dedicated pilots must be used for channel
estimation. Detailed description of codebook definitions for
MIMO operation in LTE can be found in [1].
A special case of the ZF precoder is obtained when the
equivalent channel observed by different users are orthogonal
to each other. In this case the expression for the transmitter in
(4) simplifies to
W
T
= H
H
eq
. (7)
When the scheduler imposes the constraint that only users with
orthogonal channels can be multiplexed, the resulting multi-
plexing scheme is known as unitary precoding. In principle,
unitary precoding is more robust to channel quantization and
variation than ZF precoding. However, the probability that any
two users feedback orthogonal channels decreases with the
number of codewords in the codebook, assuming the codebook
is designed such that all codewords are fed back with non-zero
probability. If the number of codewords is small, then only a
coarse quantization of the channel is possible, which limits
the precoding gain. Hence, with unitary precoding there is a
trade-off between multiplexing and precoding gains.
V. RECEIVER
In principle, there is no need to cancel the interference
of the other user at the receiver, since the ZF precoder is
designed such that the received signal is free from multi-
user interference. However, due to channel quantization and
feedback delay, some MU interference will exist. An LMMSE
receiver can be employed at the receiver to reduce the interfer-
ence and improve system performance, but this requires that
the precoding vectors applied to the streams transmitted to
different users are known. This information could be signaled
in downlink control channel, or else estimated from dedicated
pilots. In both cases, this implies additional overhead. The
LMMSE receiver for the l-th user is given by
W
R
= W
H
T,l
H
H
l
_
H
l
W
T
(H
l
W
T
)
H
+ σ
2
n
I
_
−1
, (8)
where σ
2
n
is the noise variance.
411
In LTE, the working assumption in the ongoing standard-
ization [1] is that only one stream can be transmitted to a
user in MU-MIMO mode, i.e. N
s
= 1. Hence, if the receiver
is not aware of the precoding vectors applied to the streams
transmitted to the other users, the receiver is not able to reject
the interference from the other users, and can only maximize
the received power. Hence the LMMSE receiver in (8) is
equivalent to the Maximum Ratio Combiner (MRC).
VI. CHANNEL FEEDBACK
Feedback from the terminal is crucial in order to design pre-
coders taking into account current channel state. The terminal
is supposed to feed back a Precoding Matrix Indication (PMI)
which is an index in the codebook for the preferred precoder.
It is not practical to feedback one PMI for each subcarrier, and
hence the terminal feeds back one PMI for a given group of
subcarriers.
For each subcarrier the optimum precoding vector is given
by the right singular vector corresponding to the largest singu-
lar value. The vector is then quantized to the codebook using
the metric in equation (3) for all subcarriers in the group.
Moreover, accurate Channel Quality Indication (CQI) is
important for proper link adaptation at eNodeB. Otherwise, low
rate modulation and coding schemes should be used in order
to avoid detection errors, resulting in reduced throughput.
CQI definition for MU-MIMO is still an open issue in LTE,
and in this paper we assume that the terminal reports the Signal
to Interference plus Noise Ratio (SINR) assuming a single-
stream, single-user transmission. The eNodeB then applies a
reduction factor in order to take into account the reduced
power in each stream and other losses due to, e.g. interference.
As seen in Section V, for the single-stream transmission the
LMMSE receiver is equivalent to the MRC receiver, and the
SNR is given by
δ =
|H
eq,l
W
T,l
|
2
σ
2
n
, (9)
where W
T,l
is the preferred (quantized) precoding vector,
H
eq,l
is the equivalent channel given by equation (2), and σ
2
n
is the noise variance.
VII. USER SCHEDULING ALGORITHMS
In SU-MIMO transmission, several parallel data streams
are transmitted to the same terminal, while in MU-MIMO
transmission the streams are transmitted to different users who
share the same time-frequency resources. In the 3GPP LTE
system, it is assumed that UEs are semi-statically allocated in
MU-MIMO mode, implying that it is not allowed for a UE to
be scheduled in one subframe in MU-MIMO and in Single-
User MIMO (SU-MIMO) in the next subframe. Moreover,
it is assumed that only one stream can be transmitted to a
UE operating in MU-MIMO mode (N
s
= 1), as noted in
Section II.
For each resource allocation the scheduler has to decide
between single-stream single-user transmission or MU-MIMO
transmission. Since the transmitted power must remain con-
stant, the power of each stream in MU-MIMO mode is the
UE 1
UE 3 UE 2 UE 2
Frequency
P
o
w
e
r
Figure 1. Distribution of power over different resource blocks. The eNodeB
transmits with maximum power to UE 1 on those RBs where it is not
multiplexed with any other user.
total TX power divided by the number of streams, and hence
MU-MIMO transmission does not necessarily imply higher
data rates. We assume that the scheduler assigns one user for
transmission, and decides on transmitting in MU-MIMO mode
only if the estimated data rate in MU-MIMO mode is higher
than for single-user transmission.
For frequency-dependent (FD) scheduling it is assumed
that the same modulation and coding scheme is used for the
whole allocation, according to 3GPP LTE. The FD scheduling
algorithm can be summarized as:
• Primary user selection and the resource allocation for the
primary users is done by the FD scheduler independently
in time and frequency domains. Well-known schedulers
can be used, such as Round Robin and Proportional Fair
schedulers [7].
• Candidates for MU-MIMO are selected among users that
have not been scheduled as primary users.
– Identify which UEs can be transmitted in MU-MIMO
mode with the primary UE.
– Estimate the rate for single-stream transmission and
MU-MIMO transmission for each candidate UE. De-
cide on single-stream or MU-MIMO allocation as we
will describe in Section VII-B.
• Compute the precoding matrix as in Section IV, assuming
either ZF or unitary precoding.
• Limitation on the maximum number of scheduled users
per TTI due to control signaling restrictions must be taken
into account.
• Users can be allocated in MU-MIMO mode with different
primary users.
The selection of UEs to be scheduled and the allocation of
frequency resources are performed independently in time and
frequency domains. The evaluation of MU-MIMO allocation is
performed independently for each resource block. A terminal
is allocated in MU-MIMO mode for each resource block
depending on the precoding vector and channel condition, and
hence it is not guaranteed that a UE can be allocated in MU-
MIMO mode for all resource blocks it has been allocated
to. Hence, eNodeB does not perform power sharing in those
resource blocks where there is no actual user multiplexing, in
order to guarantee that the total output power is constant over
all subcarriers. Figure 1 illustrates this arrangement.
412
A. Search for MU-MIMO candidates
The MU-MIMO candidates are identified by means of cor-
relation between the reported channel vectors. If the correlation
is below a pre-defined threshold, then the users are marked as
candidates to be scheduled in MU-MIMO mode. For unitary
precoding this threshold should be zero and for ZF precoding
the threshold can be close to unity.
However, even for the ZF case it is not recommended
to accept terminals with similar channels as candidates. The
reason is that the ZF solution can be more sensitive to errors
if the channels are highly correlated. Hence, by setting a
more conservative threshold (i.e. closer to zero) the overall
complexity of the scheduler is simplified. This is due to the
smaller number of terminals that have to be evaluated for each
RB, thus reducing the number of computed precoding matrices.
B. Decision on MU-MIMO allocation
For each candidate set, eNodeB uses fed back information
to estimate the transmission rate for single-user and multi-
user allocation. Let R
S
P
denote the rate of the primary user
scheduled in single-user mode, R
M
P
denote the rate of the
primary user scheduled in multi-user MIMO mode, and R
M
S
denote the rate of the secondary user scheduled in multi-user
MIMO mode.
With these definitions, a set of users is allocated in MU-
MIMO mode if and only if
R
M
P
+ R
M
S
≥ R
S
P
(10)
and
R
M
P

R
min
N
, (11)
where N is the number of scheduled resources allocated to
the user and R
min
is a QoS parameter specifying the minimum
supported data rate. The rates R
S
P
, R
M
P
, and R
M
S
are estimated
from Channel Quality Indication (CQI) fed back by the UE.
The purpose of (11) is to avoid that a weak UE is forced
to transmit in MU-MIMO mode in order to favor transmission
for a much stronger UE which is a secondary user, i.e., R
S
P
and R
M
P
are small, but R
M
S
is large.
VIII. SIMULATION RESULTS
In this Section we provide system simulation results to
evaluate the impact of varying precoding granularity. We also
evaluate the performance loss if the interferer’s precoding
vector and transmission is not known.
System simulations were done for 2x2 antenna configura-
tion, TU Case 1 channel model [8], 10 MHz bandwidth, 20
users per sector, and a regular grid of 19 cells (57 sectors).
Precoding is based on 3-bit SU-MIMO codebook agreed in
3GPP [1]. For unitary precoding, only four precoding vectors
are considered, corresponding to two unitary matrices. The
receiver is as defined in Section V. Frequency domain packet
scheduling as described in Section VII, with scheduling granu-
larity of 5RBs, i.e. the minimum allocation for a terminal is 60
consecutive subcarriers, corresponding to 900 kHz. The total
number of available RBs for 10 MHz bandwidth is equal to
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
12.5
13
13.5
14
14.5
15
15.5
16
TX correlation
A
v
g
.

s
e
c
t
o
r

t
h
r
o
u
g
h
p
u
t

[
M
b
p
s
]


SU−MIMO
UN−512
UN−256
ZF−512
ZF−256
Figure 2. Average sector throughput for R
min
= 256 kbps and R
min
=
512 kbps as a function of TX correlation. Also shown is the performance of
open-loop SU-MIMO transmission.
50 [6]. Proportional-fair scheduling algorithm is used in both
time and frequency domains [7].
In principle, the precoder should be computed indepen-
dently for each subcarrier, as shown in Section IV. However,
such frequency granularity is not of practical use, and the same
precoding vector must be applied to a group of subcarriers.
Two different precoding granularities have been simulated:
5 RBs or 50 RBs. A granularity of 5 RBs implies that the UE
must feedback 10 PMI for each subframe, compared to a single
PMI feedback in the case of 50 RBs granularity. Moreover, the
applied precoding vectors must be transmitted in the downlink
control channel as well, especially if eNodeB is allowed to
utilize different precoding vectors than the ones signaled by
the UE. Such situation can happen if ZF precoding is used, or
if the PMI was received with error.
Figure 2 shows the average sector throughput for R
min
=
256 kbps and R
min
= 512 kbps as a function of TX correlation.
Performance of open-loop SU-MIMO transmission is shown
for comparison. The SU-MIMO scheme simulated in this
article is the Selective Per Antenna Rate Control (S-PARC)
[9]. It can be seen from the figure that performance of open-
loop SU-MIMO degrades with increased TX correlation, as
expected. However, performance of MU-MIMO improves with
TX correlation, since this allows better separation between the
streams transmitted to each user. It is observed that the system
only benefits from the higher utilization of MU-MIMO for
very high spatial correlation. Otherwise, a more conservative
adaptation between single stream and MU-MIMO transmission
results in higher sector throughput.
Figure 3 shows the average sector throughput for unitary
precoding for different precoding granularities in frequency
domain and for different receivers. The results are shown
for a receiver that is aware of the transmission to other
users (LMMSE), and for a receiver that is not aware of
413
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
11.5
12
12.5
13
13.5
14
14.5
15
15.5
TX correlation
A
v
g
.

s
e
c
t
o
r

t
h
r
o
u
g
h
p
u
t

[
M
b
p
s
]


SU−MIMO
ZF−Prec−5RB−MMSE
ZF−Prec−FBW−MMSE
ZF−Prec−5RB−MRC
ZF−Prec−FBW−MRC
Figure 4. Average sector throughput for zero-forcing precoding as a function
of TX correlation.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
12
12.5
13
13.5
14
14.5
15
15.5
16
TX correlation
A
v
g
.

s
e
c
t
o
r

t
h
r
o
u
g
h
p
u
t

[
M
b
p
s
]


SU−MIMO
UN−Prec−5RB−MMSE
UN−Prec−FBW−MMSE
UN−Prec−5RB−MRC
UN−Prec−FBW−MRC
Figure 3. Average sector throughput for unitary precoding as a function
of TX correlation. Also shown is the performance of open-loop SU-MIMO
transmission.
the transmission to other users (MRC). It is shown that the
higher precoding granularity only gives gain for scenarios
with low spatial correlation. For scenarios with high spatial
correlation, which are the key scenarios for application of MU-
MIMO in 3GPP LTE, one precoder for the whole bandwidth
is sufficient. This observation allows for reduced signaling
in downlink control channel, or else a reduced number of
dedicated pilots. Both alternatives imply reduced overhead.
Moreover, there is performance gain if an LMMSE receiver
is used instead of the MRC receiver, especially for low spatial
correlation. The performance difference is smaller for higher
TX correlation, which implies that for high TX correlation
the MRC receiver could be sufficient. This would reduce the
signaling overhead, since it would not be needed to signal in
downlink the precoding vectors of the interfering UEs.
Similar behavior is observed for the average sector through-
put for ZF precoding, as shown in Figure 4. The limited
number of codewords in the codebook limits the beamform-
ing gains that could be achieved with ZF precoding, and
as a consequence it performs worse than unitary precoding.
Moreover, finer granularity of the codebook allows for better
approximation of the precoder itself, as defined in equation
(8).
IX. CONCLUSIONS
In this paper we evaluate the performance of MU-MIMO
transmission in LTE for different precoding granularities and
for different receivers. System simulation results show that
there is no significant performance degradation if only one
precoder is used for the whole bandwidth if the spatial correla-
tion is high. It is also observed that the performance difference
between the LMMSE and MRC receivers is less pronounced
for high spatial correlation. For the given codebook, unitary
precoding overperforms zero-forcing precoding, for all values
of TX correlation.
These observations lead to significant reduction on the
amount of overhead in MU-MIMO operation in 3GPP LTE.
The overhead in uplink and downlink directions is reduced
once a reduced number of precoding indexes have to be fed
back by the UEs. If the MRC receiver is used, even further
reduction of overhead in downlink direction is possible, since
it is not needed to signal the precoding vectors of interfering
UEs, and no extra dedicated pilots are needed.
REFERENCES
[1] 3GPP TR 36.211 V8.1.0, “Evolved universal terrestrial radio access (E-
UTRA); physical channels and modulation,” http://www.3gpp.org, Nov.
2007.
[2] Q. H. Spencer, C. Peel, A. L. Swindlehurst, and M. Haardt, “An
introduction to the multi-user MIMO downlink,” IEEE Comm. Mag.,
vol. 42, no. 10, pp. 60–67, Oct. 2004.
[3] M. Costa, “Writing on dirty paper,” IEEE Trans. on Information Theory,
vol. 29, no. 3, pp. 439–441, May 1983.
[4] Q. H. Spencer, A. L. Swindlehurst, and M. Haardt, “Zero-forcing methods
for downlink spatial multiplexing in multiuser MIMO channels,” IEEE
Transactions on Signal Processing, vol. 52, no. 2, pp. 461–471, Feb. 2004.
[5] 3GPP TR 36.201 V8.1.0, “Evolved universal terrestrial radio access (E-
UTRA); LTE physical layer - general description,” http://www.3gpp.org,
Nov. 2007.
[6] 3GPP TR 36.104 V8.0.0, “Evolved universal terrestrial radio ac-
cess (E-UTRA); base station (BS) radio transmission and reception,”
http://www.3gpp.org, Dec. 2007.
[7] V. K. N. Lau, “Proportional fair space-time scheduling for wireless
communications,” IEEE Trans. on Communications, vol. 53, no. 8, pp.
1353–1360, Aug. 2005.
[8] 3GPP TR 25.943 V7.0.0, “Deployment aspects,” http://www.3gpp.org,
Jun. 2007.
[9] S. Grant, J.-F. Cheng, L. Krasny, K. Molnar, and Y.-P. Wang, “Per-
antenna-rate-control (PARC) in frequency selective fading with SIC-
GRAKE receiver,” in Proc. of IEEE Vehicular Technology Conference
– VTC Fall, Sep 2004, vol. 2, pp. 1458–1462.
414

Sign up to vote on this title
UsefulNot useful