Design of High Performance MIMO Receivers for

LTE/LTE-A Uplink
Meilong Jiang
Broadband and Mobile Networks
NEC Labs America, Inc.
Email: [email protected]
Narayan Prasad
Broadband and Mobile Networks
NEC Labs America, Inc.
Email: [email protected]
Xiaodong Wang
Electrical Engineering Department
Columbia University
Email: [email protected]
AbstractIn this paper we design high performance multiple-
input-multiple-output (MIMO) receivers for the DFT-Spread-
OFDM based long term evolution (LTE) cellular uplink. In the
LTE uplink multiple single-antenna users can be scheduled on the
same time-frequency resource block via space division multiple
access. The designed receivers are also applicable to the LTE-
Advanced cellular uplink wherein simultaneous transmission of
multiple streams by a single user will be possible. Two types
of advanced non-linear receivers are considered and optimized,
namely, a receiver based on a two-symbol max-log soft-output
demodulator (two-symbol MLD) and a turbo minimum mean
squared error successive interference cancelation (turbo MMSE-
SIC) receiver. Based on extensive simulations, it is shown that
both the two-symbol MLD and the turbo MMSE-SIC receivers
exhibit superior performance compared to the conventional linear
MMSE (LMMSE) receiver. In general, the turbo MMSE-SIC re-
ceiver is robust to timing offsets and offers the best performance
but also introduces larger latency and higher computational
complexity. Upon employing a proposed new pairing method,
the two-symbol MLD based receiver is also found to yield a good
performance that is robust to timing offsets and which entails a
moderate complexity and latency.
I. INTRODUCTION
DFT-Spread-OFDM (DFT-S-OFDM) based multiple access
technique has been adopted as the uplink access scheme in
emerging cellular systems such as the 3GPP long term evolu-
tion (3GPP-LTE) [1], [2]. Due to the DFT-spreading operation,
the received sufcient statistics can be modeled as the channel
output of a large MIMO system. The large dimension of this
equivalent MIMO model makes it a challenge to design a
MIMO receiver with good performance, low complexity as
well as a low latency. The conventional linear MIMO receiver
such as the linear MMSE receiver has a low complexity and
a low latency but results in a poor performance.
In this paper, we shall investigate two types of advanced
non-linear receivers for the DFT-S-OFDM uplink when multi-
ple singe antenna users are co-scheduled on an identical time-
frequency resource block. This scenario, henceforth referred
to as the DFT-S-OFDM-SDMA uplink, is expected to be im-
portant in the LTE cellular uplink. In particular, we consider a
two-symbol MLD based receiver employing a proposed novel
pairing rule and the turbo MMSE-SIC receiver. Such advanced
MIMO receivers are needed to meet the high throughput
requirements of both LTE and LTE-Advanced cellular systems.
The two symbol MLD based receiver (albeit with a xed
pairing) was initially proposed in [3] and shown to strike a
judicious performance-complexity tradeoff compared to other
non-iterative receivers. However, when there are timing offsets
between the signals received from the co-scheduled users
(UEs), which is quite likely in practise, the receiver was found
to yield a degraded performance. As a remedy, we propose a
new pairing method that results in a robust performance in
the presence of timing offsets. The turbo iterative receiver
using linear ltering and successive interference cancelation
has attracted extensive attention [4], [5]. We present an imple-
mentation friendly turbo-SIC receiver design and evaluate its
performance over the LTE/LTE-A uplink.
The remainder of this paper is organized as follows. Section
II introduces the DFT-S-OFDM-SDMA uplink system model.
The receiver algorithms, including the two-symbol MLD based
one with an improved paring and the turbo MMSE-SIC
are presented in Section III. Section IV presents the BLER
performance results for the proposed receivers. Finally, the
concluding remarks are made in Section V.
Notation: ()

is reserved for complex conjugate; ()

T
for
matrix transpose; ()

for matrix Hermitian (i.e., transpose

and conjugate); ()
1
for matrix inverse. diag (B
1
, . . . , B
M
)
denotes a block diagonal matrix with B
m
as its m-th diagonal
block; E{} is the expectation operator; 0 is a zero matrix; I
denotes the identity matrix; and denotes the Frobenius
norm and C denotes the set of complex numbers.
II. SYSTEM MODEL FOR THE DFT-S-OFDM-SDMA
UPLINK
In this section, we introduce the system model for the DFT-
S-OFDM-SDMA uplink which consists of one base station
with n
R
receive antennas and multiple co-scheduled UEs with
a single antenna each. The same model also applies to a single
user MIMO DFT-S-OFDM uplink. All UEs are scheduled on
the same subset of time-frequency resource blocks for data
transmission. Without loss of generality, we assume that the
UEs are assigned tones 1 through M out of the N total
available subcarriers.
For convenience we consider an SDMA system with two
UEs but the receiver designs given in the sequel can be ex-
tended to more than two UEs. For the m-th tone, the effective
(frequency domain) channel response vector corresponding
the k-th user is h
(k)
m
C
n
R
and the DFT-spread symbol is
Fig. 1. Receiver structures for DFT-S-OFDM-SDMA system
x
(k)
m
C, k = 1, 2. The received signal vector on the m-th
tone is now given by
y
m
= H
m
x
m
+n
m
, (1)
where H
m

= [h
(1)
m
, h
(2)
m
] and x
m

= [x
(1)
m
, x
(2)
m
]
T
. We assume
uncorrelated noise vectors i.e. E{n
m
n

m
} = I.
Let F be the M M DFT matrix with its (k, n)
th
element given by F
k,n
=
1

M
e
j
2(k1)(n1)
M
. With s
(k)
=
[s
(k)
1
, s
(k)
2
, , s
(k)
M
]
T
denoting the vector of (unit average
energy) QAM symbols and x
(k)
= [x
(k)
1
, x
(k)
2
, , x
(k)
M
]
T
=
Fs
(k)
, we can collect the received signals over all the M tones
and obtain
y = [H
(1)
, H
(2)
]
_
x
(1)
x
(2)
_
+n
= [H
(1)
F, H
(2)
F]
_
s
(1)
s
(2)
_
+n, (2)
where y

= [y
T
1
, y
T
2
, , y
T
M
]
T
C
n
R
M
, and H
(k)

=
diag(h
(k)
1
, . . . , h
(k)
M
) C
n
R
MM
.
In what follows we discuss the design of two-symbol MLD
(with a new pairing method) and turbo MMSE-SIC receivers
based on the system model obtained in (2).
III. MIMO RECEIVER DESIGNS FOR DFT-S-OFDM
SYSTEMS
A. Conventional linear MMSE (LMMSE) Receiver
For comparison, we rst derive a linear MMSE receiver.
The linear MMSE estimate of x
(k)
m
, k = 1, 2, based on y
m
in
(1) is given by [6]
x
(k)
m
=
h
(k)
m
_

j=k
h
(j)
m
h
(j)
m
+I
_
1
y
m
1 +h
(k)
m
_

j=k
h
(j)
m
h
(j)
m
+I
_
1
h
(k)
m
,
= h
(k)
m
_
I +H
m
H

m
_
1
y
m
, m = 1, . . . , M. (3)
Dening x
(k)
= [ x
(k)
1
, , x
(k)
M
]
T
and applying the inverse
DFT on x
(k)
, we obtain
s
(k)
= F

x
(k)
= F

D
(k)
Fs
(k)
+n
(k)
, (4)
where D
(k)
= diag
_
d
(k)
1
, . . . , d
(k)
M
_
and d
(k)
m
=
[
_
I +H

m
H
m
_
1
H

m
H
m
]
k,k
, k = 1, 2. It can be veried
that (4) can be simplied as
s
(k)
i
=
(k)
s
(k)
i
+ v
(k)
i
, i = 1, . . . , M, (5)
with
(k)
=
1
M
M

m=1
d
(k)
m
, (6)
where v
(k)
i
contains the residual interference and noise, with
variance
E{|v
(k)
i
|
2
} =
(k)
(1
(k)
). (7)
With (5), we can then calculate the LLR of each bit associated
with the QAM symbol s
(k)
i
.
B. Two-Symbol MLD Receiver with a new pairing method
We now consider the two-symbol MLD based receiver,
that was originally presented in [3]. We now introduce a
new pairing method for the two-symbol MLD scheme, which
ensures a robust performance against timing offsets unlike the
simple xed pairing [7]. The basic idea is to divide the symbol
vectors s
(k)
, k = 1, 2 into M pairs, each with two symbols.
Each pair is demodulated using a two-symbol demodulator and
the pairing rule determines the composition of the M pairs.
All operations up-to equation (4) are same as the con-
ventional LMMSE receiver. We thus obtain s
(1)
= F

x
(1)
and s
(2)
= F

x
(2)
. Let us expand s
(1)
= [ s
(1)
1
, , s
(1)
M
]
T
and s
(2)
= [ s
(2)
1
, , s
(2)
M
]
T
. We rst derive the two-symbol
MLD receiver with xed pairing by forming the pairs s
m
=
[ s
(1)
m
, s
(2)
m
]
T
for 1 m M. The QAM symbols in each
one of the M pairs shall be demodulated using a two-symbol
max-log demodulator. Before that we need to do a noise-
whitening operation on each one of the M pairs. To do this,
we determine C =
1
M

M
m=1
(I + H

m
H
m
)
1
. Note that
the terms (I + H

m
H
m
)
1
, 1 m M are computed in
the LMMSE lter so they need not be re-computed. Next,
we compute the 2 2 matrix Q C
22
using the Cholesky
decomposition
QQ

= (I C)C (8)
and then determine z
m

= Q
1
s
m
, 1 m M. z
m
C
21
permits the expansion
z
m
= Q
1
(I C)
. .
T
s
m
+ n
m
, 1 m M, (9)
with T C
22
, s
m
= [s
(1)
m
, s
(2)
m
]
T
and E[ n
m
n

m
] = I. The
two symbols in s
m
can now be jointly demodulated using the
two-symbol max-log demodulator on z
m
for 1 m M [3].
Notice that the choice of xed pairing in the two-symbol
MLD receiver was made for simplicity but a resulting draw-
back is that the receiver can be sensitive to the timing offset be-
tween the streams transmitted by the two users. An improved
pairing rule is proposed next to achieve robust performance in
the presence of timing offset between the two UEs.
Suppose we form the pairs s
m,q
= [ s
(1)
m
, s
(2)
[m+q]
]
T
for 1
m M and any given q : 0 q M 1 and where
[m + q] = (m + q 1)mod(M) + 1. Then we determine
the matrix X(q) using (10) where where R
k
= I +H
k
H

k
.
Please note that the pairing used in (9) always uses q = 0.
Next, we compute the 2 2 matrix Q(q) C
22
using the
Cholesky decomposition
Q(q)Q(q)

= (I X(q))X(q) (11)
and then determine z
m,q

= Q(q)
1
s
m,q
, 1 m M.
z
m,q
C
21
permits the expansion
z
m,q
= Q(q)
1
(I X(q))
. .
T(q)
s
m,q
+ n
m,q
, 1 m M, (12)
with T(q) C
22
, s
m,q
= [s
(1)
m
, s
(2)
[m+q]
]
T
and
E[ n
m,q
n

m,q
] = I. The two symbols in s
m,q
can now be
jointly demodulated using the two-symbol max-log demodu-
lator on z
m,q
for 1 m M [3].
A way to determine an optimal choice of q is introduced in
Appendix 1.
In order to implement the two-symbol MLD with the
new pairing, we can rst determine the optimal q as de-
scribed in Appendix 1. Next, we can form the pairs s
m, q
=
[ s
(1)
m
, s
(2)
[m+ q]
]
T
for 1 m M. Further, note that all
the terms needed to determine X( q) are available since
1
M

M
k=1
h
(1)
k
R
1
k
h
(2)
k
exp(j2 q(k1)/M) is the ( q+1)-
th element of the vector r computed to determine the optimal
pairing.
C. Turbo MMSE-SIC Receiver
Fig. 2 illustrates the structure of turbo MMSE-SIC receiver
which uses the received signal vector y C
Mn
R
given in
(2), representing the (frequency domain) received observations
over the M tones of interest. The turbo MMSE-SIC receiver
consists of the per-tone LMMSE preltering, stream order-
ing, replica regeneration (symbol and variance estimator) and
interference cancelation blocks.
1) Two Stream MMSE-SIC lter: Let k be the index of
the stream or codeword currently under detection and

k the
interfering stream or codeword index to be subtracted, where

k = k {1, 2}. Each codeword spans one DFT block (M

tones per block) and multiple OFDM symbols. For each DFT
block and OFDM symbol, the received signal y
m
C
n
R
at
m-th tone after canceling the interference data stream at i-th
iteration is given by
y
m
= h
(k)
m
x
(k)
m
+h
(

k)
m
(x
(

k)
m
x
(

k,i)
m
) +n
m
,
= h
(k)
m
x
(k)
m
+
_

k,i)
h
(

k)
m

x
(

k,i)
m
+n
m
. .

n
, m = 1, ..., M, (13)
where the subtracted interfering stream x
(

k,i)
=
[ x
(

k,i)
1
, ..., x
(

k,i)
M
] is reconstructed based on the most
recent available estimates of the interfering symbols,
x
(

k,i)
= F s
(

k,i)
. s
(

k,i)
= [ s
(

k,i)
1
, ..., s
(

k,i)
M
] are the estimated
QAM symbols for the

k-th stream. Let

x
(

k,i)
m
=
(x
(

k)
m
x
(

k,i)
m
)

k,i)
be the normalized residual interference from stream

k at
iteration i, with
(

k,i)
= E[|x
(

k)
m
x
(

k,i)
m
|
2
] denoting the
residual error variance of the interfering codeword, which
turns out to be invariant to m. Thus

x
(

k,i)
m
can be modeled
as a Gaussian variable with zero mean and unit variance. For
the sake of notional simplicity, the iteration index is omitted
in the sequel except where it is necessary.
We can now apply the conventional LMMSE lter on the
received signal model in (13) .
The MMSE estimate of k-th stream x
(k)
m
after canceling the
interfering stream is given by
x
(k)
m
=
h
(k)
m
_

k)
h
(

k)
m
h
(

k)
m
+I
_
1
y
m
1 +h
(k)
m
_

k)
h
(

k)
m
h
(

k)
m
+I
_
1
h
(k)
m
,
= h
(k)
m
_
I +

H
m

m
_
1
y
m
, m = 1, . . . , M. (14)
where

H
m

= [h
(k)
m
,
_

k)
h
(

k)
m
].
Notice that when
(

k)
= 1, the MMSE lter (14) reduces
to the conventional MMSE without SIC given in (3).
With the MMSE estimate of k-th stream, the LLRs of
the k-th stream can be calculated and fed into a soft output
channel decoder to perform turbo decoding. Dening x
(k)
=
[ x
(k)
1
, , x
(k)
M
]
T
and applying the inverse DFT on x
(k)
, we
obtain
s
(k)
= F

x
(k)
= F

D
(k)
Fs
(k)
+ n
(k)
, (15)
where

D
(k)
= diag
_

d
(k)
1
, . . . ,

d
(k)
M
_
and

d
(k)
m
=
_
_
I +

H

H
m
_
1

H
m
_
1,1
.
It can be veried that (15) can be simplied as narrowband
SISO (single input single output) transmit model
s
(k)
i
=
(k)
s
(k)
i
+ v
(k)
i
, i = 1, . . . , M, (16)
with
(k)
=
1
M
M

m=1

d
(k)
m
, (17)
where v
(k)
i
contains the residual interference and noise, with
I X(q) =
1
M
_

M
k=1
h
(1)
k
R
1
k
h
(1)
k

M
k=1
h
(1)
k
R
1
k
h
(2)
k
exp(j2q(k 1)/M)

M
k=1
h
(2)
k
R
1
k
h
(1)
k
exp(j2q(k 1)/M)

M
k=1
h
(2)
k
R
1
k
h
(2)
k
_
, (10)
variance
E{|v
(k)
i
|
2
} =
(k)
(1
(k)
). (18)
Thus the effective SINR for the turbo SIC receiver can be
obtained as
SINR
(k)
eff
=

(k)
(
(

k)
)
1
(k)
(
(

k)
)
. (19)
where
(

k)
denotes the residual error variance of the interfer-
ing stream/codeword.
The LLRs for the decoder input corresponding to the sym-
bols s
(k)
i
of k-th stream can be calculated using the effective
SINR and post-IDFT ltered signal (16) [8]. As shown in Fig.
2, the soft output of the Turbo channel decoder LLE is fed into
a symbol replica and variance estimator to obtain the replica
of k-th stream at i-th iteration x
(k,i)
= [ x
(k,i)
1
, ..., x
(k,i)
M
] and
variance
(k,i)
, which are used to derive the LMMSE lters
for the next detection of codeword

k as given in (13) and (14).
2) Symbol regeneration and variance estimation: Now we
describe the computation of the symbol replica x
(k)
, k
{1, 2} and its residual error variance
(k)
. Assuming {
i
}
are the turbo decoder soft outputs after a specied number
of inner iterations and {b
i
} are the corresponding coded
bits, the a-posteriori i-th bit probability can be obtained as
Pr{b
i
= 0} =
1
(1+e

i )
and Pr(b
i
= 1) =
e

i
(1+e

i )
,
respectively. The rst and second statistical moments of the
transmitted QAM symbols can then be obtained as
s
m
=

siA
s
i
Pr(s
m
= s
i
); (20)
and

s
2
m
=

siA
|s
i
|
2
Pr(s
m
= s
i
); (21)
where A represents the QAM constellation set (QPSK,
16QAM or 64QAM). The symbol probability Pr(s
m
= s
i
)
can be derived as the product of the associated bit probabilities.
Letting s = [ s
1
, , s
M
]
T
, we can compute
x = Fs. (22)
and
= E[|x
(k)
m
x
(k,i)
m
|
2
] = f
m
Tf
m

=
1
M
M

m=1

t
j
(23)
where f
m
is the m-th row of DFT matrix F. The matrix
T = diag{

t
1
, ...,

t
M
} is diagonal with

t
m
= E[|s
m
s
m
|
2
] =
(

s
2
m
| s
m
|
2
).
3) Turbo MMSE-SIC algorithm: The iterative MMSE-SIC
receiver involves the following steps:
Step 1:Stream ordering.
The stream ordering determines the demodulation / de-
coding sequence of the two codewords at each subframe
based on instantaneous channel state information. The
index of the codeword to be decoded rst, denoted by
o, can be chosen as the stream index with a higher
equivalent SISO channel gain out of the conventional
LMMSE pre-ltering.
o = arg max
k=1,2

(k)
(24)
where
(k)
is the equivalent SISO channel gain given in
(16) (with
(

k)
= 1).
Step 2: Decoding the 1st stream in the given order.
LMMSE equalization ((14) with
( o)
= 1) for the stream
index determined in Step 1 (the 1st stream demodulated
in the 1st iteration), followed by a symbol replica and
error variance
(o)
estimation for the decoded stream.
Step 3: Decoding the 2nd stream in the given order.
MMSE-SIC equalization (14) for the stream demodulated
second in the 1st iteration (with soft cancelation of the
other interfering stream), followed by a symbol and error
variance estimation for the 2nd decoded stream.
Step 4: MMSE-SIC equalization (14) for both streams
in the remaining iterations (according to the determined
order with soft cancelation of the other interfering stream
using the latest available symbol and variance estimates)
until the maximum number of outer iterations is reached.
In the LTE uplink channel coding scheme, a 24-bit CRC
is attached to each codeblock. It is worth noting that the
per codeblock CRC attachment can be utilized to devise
rules for early termination of the turbo MMSE-SIC iterations
to reduce processing complexity in practical implementation.
Specically, the channel decoder is designed to output the
soft information of the whole coded sequence as well as the
hard decision of systematic bits at each turbo SIC iteration.
Thus, when the CRC check gets passed for a code block,
the decoding of the current code word is completed and the
symbol estimation and error variance of the current code block
can be saved for the other code words decoding.
IV. PERFORMANCE ANALYSIS
A. Simulation results
We now compare the block error rate (BLER) performance
of the three types of receivers. Specically, the required SNRs
for BLER=0.1 are plotted for different schemes. The simula-
tion parameters are summarized in Table I. The descriptions
of the urban macro channel model can be found in [9]. In the
gure legend, MMSE-n denotes conventional MMSE with n
inner iterations (within the turbo decoder); MLD-n denotes
two-symbol MLD receiver with n inner iterations and which
employs the new pairing unless otherwise specied; TM-mn
Fig. 2. Turbo receiver structure
Parameter Assumption
Bandwidth 10.0 MHz
FFT size 1024
Sub-carrier spacing 15.0 kHz
Sampling frequency 15.36 MHz
Number of occupied sub-carriers 600
Tone mapping method Localized
Number of antennas at Node-B 2 & 4
Number of antennas at UE 1 & 2
Number of UEs 1 & 2
Channel model urban macro
Modulation QPSK,16QAM,64QAM
Channel coding rates 1/3,1/2,2/3,3/4
Channel estimation Ideal
TABLE I
SIMULATION PARAMETERS
denotes the turbo MMSE-SIC with m outer iterations and n
inner iterations.
Fig. 3 illustrates the performance of the two-symbol MLD
(with new pairing and xed pairing) in an SDMA scenario
when there is an offset of 18 time domain samples between
the two UEs over the urban macro channel. Also plotted is
the performance of the two-symbol MLD (with new pairing
and xed pairing) without any offset. Notice that the two-
symbol MLD with new pairing achieves a much more robust
performance compared to its counterpart with the xed paring.
Fig. 4 and Fig. 5 compare the receivers performance with
and without timing offset, respectively. It can be seen that
turbo MMSE-SIC has the best performance at the cost of
higher complexity and larger delay. In addition, the proposed
two-symbol MLD exhibits reasonably good performance espe-
cially for smaller constellation and higher coding rate regimes.
Similar to the two-symbol MLD with new pairing, MMSE
and MMSE-SIC receivers are quite robust with respect to the
timing offset.
V. CONCLUSIONS
In this paper, we have presented the design and performance
analysis of the two-symbol MLD (with an improved symbol
pairing), turbo MMSE-SIC and LMMSE receivers for the
0.3333 0.5 0.6667 0.75
5
10
15
20
25
30
35
40
Coding Rate
S
N
R

p
e
r

r
e
c
e
i
v
e

a
n
t
e
n
n
a
(
d
B
)

@

B
L
E
R
=
0
.
1
Rx2, SCM (urban macro), MLD fixed and new pairing


64QAM, fixed pairing, offset=18
64QAM, fixed pairing, offset=0
64QAM, new pairing, offset=18
64QAM, new pairing, offset=0
16QAM, fixed pairing, offset=18
16QAM, fixed pairing, offset=0
16QAM, new pairing, offset=18
16QAM, new pairing, offset=0
QPSK, fixed pairing, offset=18
QPSK, fixed pairing, offset=0
QPSK, new pairing, offset=18
QPSK, new pairing, offset=0
Fig. 3. Performance comparison of two-symbol MLD with new pairing and
xed pairing; SCM urban macro SDMA; n
R
= 2.
0.3333 0.5 0.6667 0.75
5
10
15
20
25
30
35
40
Coding Rate
S
N
R

(
d
B
)

@

B
L
E
R
=
0
.
1
Rx2, SCM (urban macro), offset=18


MMSE8, QPSK
MLD8, QPSK
TM24, QPSK
TM44, QPSK
MMSE8, 16QAM
MLD8, 16QAM
TM24, 16QAM
TM44, 16QAM
MMSE8, 64QAM
MLD8, 64QAM
TM24, 64QAM
TM44, 64QAM
Fig. 4. Performance of receivers for DFT-S-OFDM-SDMA; SCM (urban
macro SDMA); n
R
= 2; offset=18.
0.3333 0.5 0.6667 0.75
5
10
15
20
25
30
35
40
Coding Rate
S
N
R

(
d
B
)

@

B
L
E
R
=
0
.
1
Rx2, SCM (urban macro), offset=0


MMSE8, QPSK
MLD8, QPSK
TM24, QPSK
TM44, QPSK
MMSE8, 16QAM
MLD8, 16QAM
TM24, 16QAM
TM44, 16QAM
MMSE8, 64QAM
MLD8, 64QAM
TM24, 64QAM
TM44, 64QAM
Fig. 5. Performance of receivers for DFT-S-OFDM-SDMA; SCM (urban
macro SDMA); n
R
= 2; offset=0.
DFT-S-OFDM (MIMO or SDMA) uplink. Both the two-
symbol MLD and turbo MMSE-SIC receivers exhibit supe-
rior performance compared to the LMMSE receiver. Turbo
MMSE-SIC receiver in general offers the best performance
but can also introduce larger latency and higher computational
complexity. The two-symbol MLD receiver yields a good
performance at a moderate complexity and latency. On the
other hand, the conventional MMSE receiver is good enough
when enough receiver diversity is available. All the three
receivers: MMSE, turbo MMSE-SIC and two-symbol MLD
with new pairing, are robust with respect to the timing offsets
in an SDMA scenario.
APPENDIX 1: BEST PAIRING SELECTION
To determine an optimal q (or equivalently an optimal pair
(m, [m+q])) we can use the capacity metric on the model in
(12) and determine a suitable q as
arg max
0qM1
det(I +T(q)

T(q)) = (25)
arg max
0qM1
det(X(q)
1
) = arg min
0qM1
det(X(q)).
Thus, we can equivalently rst determine the vector
r = F[h
(1)
1
R
1
1
h
(2)
1
, , h
(1)
M
R
1
M
h
(2)
M
]
T
(24)
and expanding r as r = [r
1
, , r
M
], we can compute q as
q = arg max
1kM
{|r
k
|} 1. (24)
Note that the terms h
(1)
m
R
1
m
h
(2)
m
= [(I +
H

m
H
m
)
1
H

m
H
m
]
1,2
, 1 m M are already
available after computing the LMMSE lters so determining
the optimal pairing requires one additional DFT.
REFERENCES
[1] H. G. Myung, J. Lim, and D. J. Goodman, Single carrier FDMA for
uplink wireless transmission, IEEE Vehicular Tech. Mag., vol. 1, pp.
3038, Sept. 2006.
[2] Physical layer aspects for evolved Universal Terrestrial Radio Access
(UTRA), ETSI Std. 3GPP TS 25.814, Rev. 7.1.0, 2006.
[3] N. Prasad, S. Wang, and X. Wang, Efcient receiver algorithms for dft-
spread ofdm systems, Trans. Wireless. Comm., vol. 8, no. 6, pp. 3216
3225, 2009.
[4] G. Berardinelli, C. Manchon, L. Deneire, T. Sorensen, P. Mogensen, and
K. Pajukoski, Turbo receivers for single user mimo lte-a uplink, in
Vehicular Technology Conference, 2009. VTC Spring 2009. IEEE 69th,
Apr. 2009, pp. 15.
[5] C. Manchon, L. Deneire, P. Mogensen, and T. Sorensen, On the
design of a mimo-sic receiver for lte downlink, in Vehicular Technology
Conference, 2008. VTC 2008-Fall. IEEE 68th, Sep. 2008, pp. 15.
[6] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.
Cambridge, UK: Cambridge University Press, 2005.
[7] M. Jiang, N. Prasad, and X. Wang, Design of efcient receivers for
dft-s-ofdma systems, in Information Sciences and Systems, 2009. CISS
2009. 43rd Annual Conference on, 18-20 2009, pp. 557 562.
[8] E. Akay and E. Ayanoglu, Low complexity decoding of bicm stbc, in
Vehicular Technology Conference, 2005. VTC 2005-Spring. 2005 IEEE
61st, vol. 1, Apr. 2005, pp. 715718.
[9] Spatial channel model for Multiple Input Multiple Output (MIMO)
simulations, ETSI Std. 3GPP TS 25.996, Rev. 6.1.0, 2003.