Professional Documents
Culture Documents
Linear Precoder Design For MIMO Interference Channels With Finite-Alphabet Signaling
Linear Precoder Design For MIMO Interference Channels With Finite-Alphabet Signaling
Linear Precoder Design For MIMO Interference Channels With Finite-Alphabet Signaling
I. I NTRODUCTION
HE recent decade has witnessed the widespread application of multiple-input multiple-output (MIMO) wireless
communication systems because of their superb spectral efficiency and link reliability [13]. However, potential benefits
of MIMO systems are often hampered by the omni-present
Manuscript received February 15, 2013; revised May 26, 2013. The
associate editor coordinating the review of this paper and approving it for
publication was Tony Q. S. Quek.
Part of the material in this paper will be presented at IEEE GLOBECOM
2013. The work of Y. Wu and X. Gao was supported in part by National
Natural Science Foundation of China under Grants 60925004 and 61222102,
the China High-Tech 863 Plan under Grant 2012AA01A506, and National
Science and Technology Major Project of China under Grants 2011ZX03003001 and 2011ZX03003-003-03. The work of C. Xiao was supported in part by
National Science Foundation under Grants CCF-0915846 and ECCS-1231848.
The work of Z. Ding was supported by NSF Grants 0917251 and 1147930.
This work began while Y. Wu was a visiting Missouri University of Science
and Technology and was completed while C. Xiao was a summer faculty
fellow at the Air Force Research Laboratory in Rome, NY. Approved for
Public Release; Distribution Unlimited: 88ABW-2013-1062.
Acknowledgmentof Support and Disclaimer: (a) Contractor acknowledges
Governments support in the publication of this paper. This material is based
upon work funded by the AF Summer Faculty Fellow Program, under AFRL
Contract No. FA8750-11-2-0218. (b) Any opinions, findings and conclusions
or recommendations expressed in this material are those of the author(s) and
do not necessarily reflect the views of AFRL.
Y. Wu and X. Gao are with the National Mobile Communications Research
Laboratory, Southeast University, Nanjing, 210096, P. R. China (e-mail:
{ypwu, xqgao}@seu.edu.cn).
C. Xiao is with the Department of Electrical and Computer Engineering,
Missouri University of Science and Technology, Rolla, MO 65409, USA (email: xiaoc@mst.edu).
J. D. Matyjas is with the Air Force Research Laboratory/RIT, Rome, NY
13441, USA (e-mail: John.Matyjas@rl.af.mil).
Z. Ding is with the Department of Electrical and Computer Engineering,
University of California, Davis, CA 95616, USA (e-mail: zding@ucdavis.ca).
Digital Object Identifier 10.1109/TCOMM.2013.072213.130132
c 2013 IEEE
0090-6778/13$31.00
WU et al.: LINEAR PRECODER DESIGN FOR MIMO INTERFERENCE CHANNELS WITH FINITE-ALPHABET SIGNALING
3767
algorithm converges within several steps and achieves significant WSR gains over the conventional iterative designs. In
addition to WSR, the coded bit error rate (BER) of the MIMO
interference channel systems is another important performance
evaluation criterion in practice. To further examine the robust
performance of the proposed design, we present iterative
transceiver systems for the MIMO interference channels by
deploying low-density parity-check (LDPC) encoders and the
proposed precoders at the transmitters, and the soft maximum
a posteriori (MAP) multiuser detectors and LDPC decoders at
the receivers. Simulations show that the proposed precoding
design achieves substantial coded BER improvement through
the iterative decoding and detection operations. Furthermore,
recognizing interference signals as finite alphabet inputs rather
than colored Gaussian noise, we designed a novel detector
structure that provides additional coded BER gains.
The following notations are adopted throughout the paper: Column vectors are represented by lower-case bold-face
letters, and matrices are represented using upper-case boldface letters. Superscripts ()T , () , and ()H represent the
matrix/vector transpose, conjugate, and conjugate-transpose
operations, respectively. We let x and XF denote the
Euclidean norm of vector x and the Frobenius norm of matrix
X, respectively. C denotes the complex field. IM denotes an
M M identity matrix (sometimes without using subscript
M ) and EV represents the expectation of random variable
(scalar, vector, or matrix) V .
II. S YSTEM M ODEL AND E XISTING R ESULTS OF MIMO
I NTERFERENCE C HANNEL W ITH G AUSSIAN I NPUTS
We consider a K-user interference channel system in which
each transmitter has dedicated information for its intended receiver and generates co-channel interference to other receivers
as illustrated in Figure 1. Suppose the i-th transmitter has Nti
antennas and the j-th receiver has Nrj antennas for every
i, j = 1, 2, , K. In addition, we assume no collaboration
among transmitters or receivers. Then, the received signal
yj CNrj 1 observed at the j-th receiver can be described
as
yj = Hjj Gj xj +
K
i=1,i=j
Hji Gi xi + nj , j = 1, 2, , K
(1)
where xj CNtj 1 and nj CNrj 1 denote the zeromean transmitted information vector and noise vector for the
j-th user with covariance matrices I and 2 I, respectively.
Also, Gj represents the linear precoding matrix for the j-th
user, and Hji CNrj Nti stands for the channel response
matrix between the i-th transmitter
and the j-th receiver,
which is normalized as tr Hji HH
ji = Nrj . Here we make
the common assumption (e.g., as in [12, 15], among others)
that the channel state information is globally available, i.e.,
each transmitter has access to perfect channel knowledge of
all users. Moreover, we assume the precoding matrix does
not increase the transmission power. Henceforth, we have the
power constraint
H
= tr Gj GH
Pj , j = 1, 2, , K
tr E Gj xj xH
j Gj
j
(2)
...
...
x1,out
...
Receiver 2
...
x2,out
Receiver K
...
...
Transmitter 2
Receiver 1
...
Transmitter 1
...
x2
...
x1
...
3768
xK,out
Fig. 1.
Transmitter K
...
xK
...
.. .
...
Hij
i=1
K
2
H H
Hji Gi Gi Hji .
log2 det I +
i=1,i=j
(3)
Therefore, the optimum precoding matrices for maximizing
the WSR can be expressed as
{G1 , G2 , , GK } =
arg max
G1 ,G2 , ,GK
Rwsum
(4)
M
K
mK =1
M
1
K 1
i=1 Mi
En log2
M
2
m1 =1 m2 =1
(7)
where
H1,j (m1 , m2 , , mK , n)
2
K
Hji Gi (xi,mi xi,ni ) +n
M
M
1 M
2
K
i=1
exp
2
n1 =1 n2 =1
nK =1
where
K
K
H
Rwsum =
j log2 det 2 I +
Hji Gi GH
H
i
ji
j=1
i=1
K
K
H
j log2 det 2 I +
Hji Gi GH
i Hji
j=1
i=1,i=j
(5)
and j denotes the weighting factor of the j-th user, j =
K
1, 2, , K, where
j=1 j = K. Then, by exploiting the
matrix derivative results in [27], along with the complex
matrix differentiation conclusions in [28], we can calculate
the gradient of the WSR with respect to Gk as2
Gk Rwsum =
1
K
K
H
2
H H
log2 e
j Hjk I +
Hji Gi Gi Hji
Hjk Gk
j=1
i=1
1
K
K
2
H
j H H
HjiGiGH
HjkGk,
i Hji
jk I+
j=1,j=k
i=1,i=j
k = 1, 2, , K.
(6)
Given the derived gradient expression, an iterative algorithm was proposed in [15] to solve the optimal precoders
{G1 , G2 , , GK }.
2 We note that the derived gradient form here is a little different from [15,
Eq. (38)]. This is because we rewrite the WSR expression in [15, Eq. (35)]
only for notation simplicity. Mathematically, they are equivalent.
M
1
(8)
M
j1
n1 =1
n
=1
j12
K
.
Hji Gi (xi,mi xi,ni ) +n
M
j+1
M
K
i=1,i=j
exp
nj+1 =1
nK =1
(9)
Proof: See Appendix A.
The results in Proposition 1 are general for an arbitrary
number of users and arbitrary antenna configurations. For two
users, the achievable rate of R1 and R2 can be simplified as
in (10) and (11) at the top of the next page.
In practice, a great amount of wireless systems may work
at low power, especially for the wireless units operated by
batteries. For instance, it was reported in [29, 30] that 40% of
the geographical locations undergo receiver SNR levels below
0 dB. In addition, a key objective in future digital communication systems, the energy-efficient requirement necessitates
the operation in low SNR region. It was indicated in [31]
that the energy efficiency enhances as one operates in low
SNR region, and the minimum bit energy is achieved as SNR
vanishes. Therefore, next we present closed form expressions
for a near-optimal transmit strategy in low SNR region.
Proposition 2: For the interference channel model (1), a
near-optimal transmit precoding design in low SNR region
( 2 ) is given by
Gj = Pj [vmax,j 0] , j = 1, 2, , K
(12)
WU et al.: LINEAR PRECODER DESIGN FOR MIMO INTERFERENCE CHANNELS WITH FINITE-ALPHABET SIGNALING
R1, nite
3769
2
H12 G2 (x2,m2 x2,n2 )+n
= log2 M1 +
En log2
exp
2
m2 =1
n2 =1
2
2
H1i Gi (xi,mi xi,ni )+n
M
M
M
M
1
2
1
2
i=1
1
En log2
exp
M1 M2
2
1
M2
M
2
M
2
m1 =1 m2 =1
M11M2
n1 =1 n2 =1
2
H21 G1 (x1,m1 x1,n1 )+n
exp
2
m1 =1
n1 =1
2
2
H2i Gi (xi,mi xi,ni )+n
M
M
1 M
2
1 M
2
i=1
.
En log2
exp
2
1
M1
M
1
m1 =1 m2 =1
En log2
M
1
(13)
(14)
Rsum
=
K
j=1
log2 Mj .
(11)
n1 =1 n2 =1
(10)
(17)
Rsum, IA =
log Mj .
+ 1 j=1 2
(18)
3770
[34].
a normalization step : Gj := Pj Gj / tr Gj GH
j
(19)
subject to tr GH
j Gj Pj , j = 1, 2, , K.
(20)
(0)
In
general,
the
WSR
objective
function
Rwsum, nite (G1 , G2 , , GK ) in (19) is not concave
with respect to precoding matrices {G1 , G2 , , GK }.
Thus, a set of necessary conditions for this optimization
problem are determined, as given in the following proposition.
Proposition 4: The optimal precoding matrices for the
WSR maximization problem in (19) and (20), satisfies the
following conditions
log2 e
j Gj = 2
K
i=1
i H H
ii T1,i,j
K
i=1,i=j
i H H
ii T2,i,j
j = 1, 2, , K
(21)
j tr GH
j Gj Pj = 0, j = 1, 2, , K
(22)
tr GH
j Gj Pj 0, j = 1, 2, , K
(23)
j 0, j = 1, 2, , K
(24)
where the Nrj Ntj matrices T1,i,j and T2,i,j are given as
in (25) and (26) at the top of the next page, where dj,mj ,nj =
xj,mj xj,nj .
Proof: See Appendix F.
1) Initialize Gj , with tr
H
(0)
(0)
Gj
Gj
= Pj , j =
1, 2, , K. Set n = 0.
(n)
(n)
(n)
(n)
(n)
2) Compute Rwsum, nite (G1 , G2 , , GK ), T1,i,j
(n)
and T2,i,j , i = 1, 2, , K, j = 1, 2, , K.
3) Set j := 1.
(n)
(n)
(n)
=
4) Compute Gj Rwsum, nite G1 , G2 , , GK
(n)
(n)
K
K
log2 e
H
H
i=1 i Hi T1,i,j
i=1,i=j i Hi T2,i,j .
2
5) Set step size t := 1.
6) Evaluate:
(n)
(n)
(n)
c = t Gj Rwsum, nite G1 , G2 , , GK 2F .
If c is sufficiently small, then go to step 11.
7) Compute:
(n)
(n)
(n)
(n)
(n)
Gj =Gj +tGj Rwsum, nite G1 , G2 , , GK .
H
(n)
(n)
(n+1)
8) If tr Gj
=
Gj
> Pj , update Gj
(n)
P G
(n) .
j j ; Otherwise, G(n+1) = G
(n)
j
j
Gj
9) Compute:
(n+1)
(n+1)
(n+1)
(n+1)
(n)
(n)
Rwsum, nite(G1
, , Gj1 ,Gj
,Gj+1 , , GK ),
set t := t.
(n+1)
(n+1)
(n+1)
(n+1)
(n)
10) If Rwsum, nite (G1
, , Gj1 , Gj
, Gj+1 , ,
(n)
(n)
(n)
(n)
(n)
GK ) < Rwsum, nite (G1 , G2 , , GK ) + c, go
to step 6.
11) If j K, j := j + 1, go to step 4.
12) Set n := n + 1, go to step 2 until a stopping criterion
is reached.
It is important to note that Algorithm 1 iterates over
Gj , j = 1, 2, , K in each step increasing the WSR in
(19). Expressions in (7) imply that the WSR under finite
alphabet constraint is upper-bounded. As a result, Algorithm
1 generates increasing sequences which are upper-bounded.
Thus, it is guaranteed to converge. Due to the non-concavity
of the objective function Rwsum, nite (G1 , G2 , , GK ), the
proposed algorithm may only find local maxima. To mitigate
the local convergence, we randomly initialize the precoding
matrices multiple times, and choose the resulting precoders
with the maximum WSR performance to be the final solution
[18, 21, 23].
(n)
to the boundary
conclusion in [34], the best solution is to project G
j
of the feasible set. At the start of the iteration, the initial weighted sumrate is low. Thus, the weighted sum-rate will be increased with the new
(n+1)
precoder Gj
that satisfies the individual power constraint. This has been
confirmed in various scenarios in [21, 23, 25, 34]. After several iteration steps,
the weighted sum-rate might be high and it might not be able to find a feasible
(n+1)
new precoder Gj
that still increases the weighted sum-rate. Hence, in
step 6, we set a condition that when c is sufficiently small, which corresponds
to the sufficiently small step t, the update process stops and Algorithm 1 goes
to step 11 directly.
4 It
WU et al.: LINEAR PRECODER DESIGN FOR MIMO INTERFERENCE CHANNELS WITH FINITE-ALPHABET SIGNALING
t=1
n1 =1 n2 =1 nK =1
En
m1 =1 m2 =1
mK =1
M
1
Mt
M
2
M
K
2
K
t=1 Hit Gt dt,mt ,nt +n
exp
2
M
2
M
1
n1 =1 n2 =1
T2,i,j =
K
M
1
t=1,t=i Mt
m1 =1
M
i1
M
i+1
mi1 =1 mi+1 =1
M
K
mK =1
MK
nK =1
2
K
t=1 Hit Gt dt,mt ,nt +n
exp
2
2
K
Mi+1
Fig. 2.
Encoder
uK
Mi+1
ni1 =1 ni+1 =1
M
K
nK =1
K
t=1,t=i
S/P
GK
LE(u )
1
...
y
K
MIMO IC
LA(u )
1
Soft Detector
LE(u )
K
LA(u )
K
Fig. 3.
j ,nj
LA(s1)
LE(s )
1
MIMO IC
Soft Detector
...
...
G1
xK
Mod
y1
...
...
...
Mod
S/P
...
Mi1
x1
...
u1
.. .
n1 =1
...
sK
2
K
Hit Gt dt,m ,n +n
t t
t=1,t=i
exp
2
.. .
LDPC
M
1
...
bK
...
User K
s1
...
LDPC
Encoder
...
User 1
b1
MK
En
2
n1 =1 ni1 =1 ni+1 =1 n =1 exp
K
Mi1
M
1
K
H
LDPC
Decoder
LA(s )
K
LE(s )
K
(25)
Hard
Decision
(26)
User 1
LD(s )
1
...
K 1
MK
M
2
M
1
LDPC
Decoder
...
...
T1,i,j =
3771
Hard
Decision
b
K
User K
LD(s )
K
K
H
2
Hji Gi GH
i Hji + INt , j = 1, 2, , K. (27)
i=1,i=j
yj = Cj
yj , j = 1, 2, , K.
(28)
log2
sj Sj,i,+1
sj Sj,i,1
(29)
where sj,i denotes the i-th bit of the j-th user, with 1 i
Ntj Mj , and 1 j K. The Ntj Mj 1 vector sj represents
the coded bits for the j-th user. Sj,i,+1 and Sj,i,1 are the
sets of 2Ntj Mj 1 bit vectors with the i-th element being +1
and 1, respectively. sj,[i] demonstrates
the subvector of sj by
omitting the i-th elements. LA sj,[i] is the (Ntj Mj 1) 1
vector with the a priori information of sj,[i] . xj = map (sj )
denotes the modulation from the bit vector sj to symbol vector
xj . For the MIMO interference channel systems (1), we have
the likelihood function
M
M
j1
1
p (yj |xj = map (sj ) ) = (2 )Nr 1K
t=1,t=j Mt m =1
mj1 =1
1
2
K
M
j+1
mj+1 =1
M
K
mK =1
exp
Hjt Gt xt,mt
yj Hjj Gj xj
t=1,t=j
4
3.5
3
Sum Rate (b/s/Hz)
3772
2.5
2
1.5
1
SNR = 5 (dB)
SNR = 0 (dB)
SNR = 5 (dB)
0.5
10
12
Iteration Index
14
16
18
20
(30)
Fig. 4.
Fig. 5.
SNR = 5 (dB)
SNR = 0 (dB)
SNR = 5 (dB)
2
10
12
Iteration Index
14
16
18
20
WU et al.: LINEAR PRECODER DESIGN FOR MIMO INTERFERENCE CHANNELS WITH FINITE-ALPHABET SIGNALING
4.5
7
6
3.5
3
2.5
2
5
4
3
1.5
1
0.5
0
20
3773
10
10
20
2
1
0
20
30
10
SNR (dB)
20
30
Fig. 7.
Sum rate of 2-user MIMO interference channels with QPSK
modulation.
10
10
10
BER
Fig. 6.
Sum rate of 2-user MIMO interference channels with BPSK
modulation.
10
SNR (dB)
10
10
8
10
SNR (dB)
12
14
16
3774
2.5
4.5
4
3.5
1.5
User 1
User 2
Proposed Precoding
Best Rotation
Gaussian Design
0.5
0
10
Fig. 9.
10
SNR (dB)
15
20
25
3
2.5
2
1.5
0.5
0
10
30
User 1
User 2
Proposed Precoding
Best Rotation
Gaussian Design
Fig. 11.
10
SNR (dB)
15
20
25
30
Modulation
10
BPSK
BER
TABLE I
BER R ESULTS U SING D IFFERENT D ETECTION M ETHODS .
10
10
QPSK
4
10
10
10
12
14
SNR (dB)
16
18
20
Fig. 10. BER of 2-user MIMO interference channels with QPSK modulation.
SNR (dB)
Proposed
Detection
2
2.5
3
3.5
9
10
11
12
0.0013
0
0
0
0.0086
0
0
0
Gaussian
Assumption
Detection
0.0114
7.513 104
0
0
0.0780
0.0712
0.0578
0.0476
0.6131 0.1955j
0.7792 + 0.3374j
0.4643 + 0.6778j 0.7344 0.0113j
=
0.4052 0.6845j 0.3806 0.0892j
0.8238 0.4134j 0.5425 + 0.1126j
=
0.1321 + 0.2715j
0.7267 0.4734j
H21 =
H22
H23
0.3382 0.7046j
0.4195 + 0.2793j
WU et al.: LINEAR PRECODER DESIGN FOR MIMO INTERFERENCE CHANNELS WITH FINITE-ALPHABET SIGNALING
3775
K
=
at Rsum
j=1 log2 Mj b/s/Hz, and the Interference
Alignment has 50% sum-rate losses8 as implied in Corollary
2.
5
Sum Rate (b/s/Hz)
V. C ONCLUSION
4
2
Gaussian Input Design
BPSK, Proposed Precoding
BPSK, Gaussian Design
Finite Alphabet Input High SNR Design
Interference Alignment Technique
0
15
10
10
15
SNR (dB)
20
25
30
35
Fig. 12.
Sum rate of 3-user MIMO interference channels with BPSK
modulation.
12
10
ACKNOWLEDGMENT
0
10
10
20
30
40
A PPENDIX A
P ROOF OF P ROPOSITION 1
SNR (dB)
Fig. 13.
Sum rate of 3-user MIMO interference channels with QPSK
modulation.
H31 =
H32 =
H33 =
0.9707 + 0.2271j
0.0265 0.7569j
0.4520 0.2644j
0.2748 0.2878j
0.2200 0.0000j
0.5837 0.1839j
0.0113 + 0.6334j
0.0279 1.0840j
0.2200 0.0000j
0.5837 0.1839j
0.0113 + 0.6334j
0.0279 1.0840j
1
N
( 2 ) rj
exp
2
K
yj Hji Gi xi.m
i
i=1
(33)
8 It is noted that each user transmits one data stream for the IA method.
Although the IA approach can align other users interference into a signal
space and eliminate them, it is still suboptimal. This is because the mutual
information with finite alphabet inputs is bounded. Thus, allocating more
power to the already saturated signals can not further improve the mutual
information. This implies that transmitting only one data stream for each
user will result in a constant sum-rate loss in finite alphabet input scenarios.
Instead, for the proposed method in Algorithm 1 and the high SNR design in
Proposition 3, each user transmits two data streams. It is shown in Figure 12
and Figure 13 that both Algorithm 1 and the proposed high SNR design in
Proposition
3 can effectively combat interference and achieve the saturated
sum-rate K
j=1 log2 Mj b/s/Hz at high SNR.
3776
H (yj ) =
K
i=1
M
1
K 1
log2 Mi +
Mi
i=1
log2
m1 =1 m2 =1
M
2
M
1
n1 =1 n2 =1
H (yj |xj ) =
K
i=1,i=j
Mi +
log2
K 1
i=1
M
1
n1 =1
1
N
( 2 ) rj
Mi
M
1
exp
M
2
m1 =1 m2 =1
Mj+1
Mj1
nj1 =1 nj+1 =1
M
2
MK
nK =1
mK =1
nK =1
'
i=1
yj Hjj Gj xj.mj
exp
i=1,i=j
2
Hji Gi xi.ni
(31)
dyj
)
2
K
yj Hji Gi xi.n
i
(
p yj (xj,mj , x1,n1 , , xj1,nj1 , xj+1,nj+1 , , xK,n
2K
K
=
mK =1
1
'
(34)
1
N
( 2 ) rj
M
K
M
K
M
K
.
(35)
(32)
dyj .
)
A PPENDIX B
P ROOF OF P ROPOSITION 2
First, we apply Jensens inequality [33] to obtain an approximation of (7) as follows
1
Rj,nite log Mj )K
Mi
i=1
*2
K
*
*
*
f1 En exp 1 *
H
G
(x
x
)
+n
*
*
ji i
i,mi
i,ni
*
2 * i=1
*
*2
*
*
K
*
1 *
*
K
H (yj ) =
log2 Mi
i=1
M
K
'
mK =1
K 1
i=1
M
1
Mi
M
2
m1 =1 m2 =1
M1
MK
MK
M1
M2
M2
.
f1 () =
log2
()
m1 =1 m2 =1
K
i=1
log2 Mi
M
K
mK =1
K 1
i=1
M
1
Mi
M
2
m1 =1 m2 =1
mK =1
M
K
'
mK =1
i=1,i=j
mK =1
log2 Mi
K 1
i=1
M
1
Mi
M
2
m1 =1 m2 =1
(38)
K
i=1,i=j
log2 Mi
K 1
i=1 Mi
M
1
Mj+1
nj1 =1 nj+1 =1
MK
()
nK =1
M1
MK
MK
M2
M1
M2
.
()
f3 () =
=
M
K
K
n1 =1
(41)
nK =1
(42)
m1 =1 m2 =1
M
2
m1 =1 m2 =1
mK =1 n1 =1 n2 =1
Mj1
MK
M1
M2
M1
.
f4 () =
m1 =1 m2 =1
H (yj |xj ) =
n1 =1 n2 =1
Mj1
M1
MK
M1
M2
.
log2
f2 () =
m1 =1 m2 =1
mK =1
mK =1n1 =1
(43)
nK =1
Mj+1
nj1 =1 nj+1 =1
MK
().
nK =1
(44)
Recalling the definition of n in (7), the first expectation in
(40) can be evaluated as
*
*2
* K
*
1 *
En exp 2 *
Hji Gi (xi,mi xi,ni ) + n*
*
i=1
* K
*2
*
*
'
= n exp 12 *
Hji Gi (xi,mi xi,ni ) + n*
*
*
i=1
Nr
2
dn 2
exp n
2
= 2Nr exp
21 2
(45)
*
*2
K
*
*
*
Hji Gi (xi,mi xi,ni ) *
*
*
i=1
'
n I1 (n) dn
(46)
WU et al.: LINEAR PRECODER DESIGN FOR MIMO INTERFERENCE CHANNELS WITH FINITE-ALPHABET SIGNALING
where
2 Nr
*
K
*2
* 1
*
2
*
*
exp 2 *
Hji Gi (xi,mi xi,ni ) + n* (47)
*
* 2 i=1
I1 (n) =
denotes a probability density function of an independent and identically distributed Gaussian vector with mean
K
2
Hji Gi (xi,mi xi,ni ) and covariance matrix 2 I.
12
i=1
Therefore, we have
(xi,mi xi,ni )
(48)
K
Mi
i=1
*
*2
K
*
*
1
*
*
f1 exp
H
G
(x
x
)
*
*
ji
i
i,m
i,n
i
i
*
22 * i=1
*
*2
*
*
K
*
*
1
* .
H
G
(x
x
)
f2exp 2 *
ji i
i,mi
i,ni
*
*
2 *
*
i=1,i=j
(49)
It is noted that here we use the Jensens inequality in
two terms in (7) and then subtract them. When SNR is
zero, the approximated rate Rj, nite in (49) is zero, which
corresponds to the exact achievable rate in (7). This indicates that the approximation based on Jensens inequality is
asymptotically accurate when SNR goes to zero. In addition,
since the bounding errors for both terms are to be similar
and therefore subtracting these will have a canceling effect,
which in turn will yield a fairly accurate approximation. This
will be confirmed by numerical results where the obtained
precoding design based on (49) performs nearly close to sumrate achieved by Gaussian inputs in low SNR region, which
implies the obtained precoding design is actually near-optimal
over all the possible solution sets.
In low SNR region where 2 +, utilizing the Taylor
expansion of the exponent function exp(x) = 1 + x + o(x),
(49) can be computed as
j,nite = log Mj ) 1
R
K
Mi
i=1
*K
*2
*
*
1
1
*
*
f1 1
H
G
(x
x
)
*
ji
i
i,m
i,n
i
i * +o
2
*
2 * i=1
2
*
*2
*
*
K
*
*
1
1
Hji Gi (xi,mixi,ni )*
f2 1 2 *
* +o 2 .
2 *
*
*
i=1,i=j
(50)
Next, we exploit the Taylor expansion of function log2 (1
H
GH
i Hji
Mj
K
f4 (tr (HjiGi
i=1,i=j
H
H
(xi,mi xi,ni ) (xi,mi xi,ni ) GH
i Hji
1
+o
. (51)
2
I1 (n) dn = 1.
3777
m=1
xj,m xH
j,m = Mj I,
M
j M
j
m=1 n=1
xj,m xH
j,n = 0,
j = 1, 2, , K.
(52)
To this end, (51) can be further reduced to
2
K
,K
e
log
2
Mi
tr HjiGiGH
HH
Rj,nite= )
2
i
ji
i=1
K
i=1
2
i=1 Mi
,
2
K
K
1
H
+o
Mi
tr HjiGiGH
H
i
ji
i=1
2
i=1,i=j
log2 e
1
H H
=
tr
H
G
G
H
.
(53)
+
o
jj j j
jj
2
2
In (53), the achievable rate of each user is determined by its
own channel gain and precoding matrix. Also, it is straightforward to identify that the precoding design maximizing the
first-order term in (53) is to perform beamforming along the
eigenvector corresponding to the largest eigenvalue of matrix
HH
jj Hjj , which completes the proof.
A PPENDIX C
P ROOF OF C OROLLARY 1
Here we prove that in low SNR region, the optimal precoding under Gaussian input assumption conforms to the precoding structure in Proposition 2. First, we rewrite achievable
rate under Gaussian input assumption in (3) as
K
H H
Rj (u) = log2 det I + u
Hji Gi Gi Hji
i=1
K
H
Hji Gi GH
log2 det I + u
i Hji
i=1,i=j
(54)
where u = 12 . In low SNR region, we expand Rj (u) with
respect to u at u = 0 as
.
(55)
(56)
(57)
3778
A PPENDIX D
P ROOF OF P ROPOSITION 3
2
n
. (70)
lim H1,j (m1 , m2 , , mK , n) = exp 2
2 0
i=1
lim
exp
= 0.
2 0
(58)
=
(59)
c1,2
j,t (m2 , m3 , , mK , n2 , n3 , , nK )
K
i Pi
=
Nt ai,j,mi ,ni ,t .
i
i=2
(60)
Since m1 = n1 , we can always find a t 1, Nrj , satisfies
(
(
P1
(
( 1,1
1,min
(61)
(cj,t (m1 , n1 )(
Nt1
(
(
( 1,2
(
(cj,t (m2 , m3 , , mK , n2 , n3 , , nK )(
K K (a)
i Pi
i Pi
|
|a
i,j,m
,n
,t
i
i
Nt
Nt i,max
i=2
i=2
(62)
where (a) is from the Minkowskis inequality [41]. Then,
combining (61), (62), and the conditions in (16) (i = 1), we
have
(
( (
(
( 1,1
( ( 1,2
(
(cj,t (m1 , n1 )( > (cj,t (m2 , m3 , , mK , n2 , n3 , , nK )( .
(63)
(63) implies that
1,2
c1,1
j,t (m1 , n1 )+cj,t (m2 , m3 , , mK , n2 , n3 , , nK ) = 0
(64)
for arbitrary arrays (m1 , m2 , , mK ) = (n1 , n2 , , nK ).
If m1 = n1 and m2 = n2 , we define
2 P2
2,1
cj,t (m2 , n2 ) =
a2,j,m2 ,n2 ,t
(65)
Nt2
c2,2
j,t (m3 , m4 , , mK , n3 , n4 , , nK )
K
i Pi
=
Nt ai,j,mi ,ni ,t .
i=3
(66)
,
m
,
n
,
n
,
,
n
)
=
0.
c2,1
2
2
3
4
K
3
4
K
j,t
j,t
(68)
= exp n
2
.
(71)
Combining (7), (70), and (71), the achievable rate for the
j-th user is given by
lim Rj, nite = log Mj , j = 1, 2, , K.
2 0
(72)
(72) is the maximum rate that can be achieved for the j-th user
with respect to finite alphabet constraints, which completes the
proof.
It is noted that the optimal precoding design when
2 0 is not
Kunique. For any precoders which fulfill
the condition
i=1 Hji Gi (xi,mi xi,ni ) = 0 for arrays
(m1 , m2 , , mK ) = (n1 , n2 , , nK ) can achieve the rate
in (72).
A PPENDIX E
P ROOF OF C OROLLARY 2
Based on [13, Theorem 3], we can transform the K user
MIMO interference channel in (1) into a KNT user SIMO
interference channel with antenna at each receiver. Let
= KNT (KNT 1). Then, recalling the interference
alignment scheme in [13, Appendix A], the length of desired
signal transmitted over a n = ( + 1)(n + 1) symbol
extension for the j-th user, j = 1, 2, , K, is given by
.
(n + 1) , j = 1, 2, , + 1
(73)
lj =
j = + 2, + 3, , KNT
n ,
At the receiver, each user can decode its desired signal by zero
forcing the aligned interference signals. Thus, the sum-rate of
all the receivers at high SNR is given by
Rsum,
IA (n) = (n + 1)
+1
j=1
log2 qj + n
KN
T
j=+2
log2 qj
(74)
WU et al.: LINEAR PRECODER DESIGN FOR MIMO INTERFERENCE CHANNELS WITH FINITE-ALPHABET SIGNALING
(i 1) NT + 1 j iNT ,
Rsum,IA = sup
n
1
R
(n).
n sum,IA
(76)
(n) and
By plugging in the expressions of n and Rsum,IA
taking the supremum over all n, it yields
Rsum, IA
KN
K
T
=
log2 qj =
log2 Mj . (77)
+ 1 j=1
+ 1 j=1
A PPENDIX F
P ROOF OF P ROPOSITION 4
The Lagrangian cost function for the precoding matrices in
(19)(20) is given by
f (G, ) = Rwsum, nite (G1 , G2 , , GK )
K
j tr Gj GH
+
Pj
j
j=1
(78)
in which j 0, j = 1, 2, , K. Following the similar
approaches in [17], we define the complex gradient operator
th
f
as Gj f = G
element of matrix with the
. The (i, j)
j
.
gradient is defined as Gj f i,j = {Gj }i,j f = Gf
{ j }i,j
Then, the KKT conditions in [33] are as follows
Gj f (G, ) = Gj Rwsum,nite (G1 , G2 , , GK )
+j Gj = 0
(79)
P
=
0
(80)
j tr Gj GH
j
j
tr Gj GH
Pj 0
(81)
j
j 0
(82)
3779
3780
Xiqi Gao received the Ph.D. degree in electrical engineering from Southeast University, Nanjing,
China, in 1997. He joined the Department of Radio
Engineering, Southeast University, in April 1992.
Since May 2001, he has been a professor of information systems and communications. From September
1999 to August 2000, he was a visiting scholar at
Massachusetts Institute of Technology, Cambridge,
and Boston University, Boston, MA. From August
2007 to July 2008, he visited the Darmstadt University of Technology, Darmstadt, Germany, as a Humboldt scholar. His current research interests include broadband multicarrier
communications, MIMO wireless communications, channel estimation and
turbo equalization, and multirate signal processing for wireless communications. He served as an Editor for the IEEE T RANSACTIONS ON W IRELESS
C OMMUNICATIONS from 2007 to 2012. He now serves as an Associate Editor
for the IEEE T RANSACTIONS ON S IGNAL P ROCESSING.
Dr. Gao received the Science and Technology Awards of the State Education
Ministry of China in 1998, 2006 and 2009, the National Technological
Invention Award of China in 2011, and the 2011 IEEE Communications
Society Stephen O. Rice Prize Paper Award in the field of communications
theory.
John D. Matyjas received the A.S. degree in preengineering from Niagara University in 1996 and the
B.S., M.S., and Ph.D. degrees in electrical engineering from the State University of New York at Buffalo
in 1998, 2000, and 2004, respectively. Currently, he
is a Principal Research Engineer employed since
2004 by the Air Force Research Laboratory in
Rome, NY, performing RD in the area of information connectivity. His research interests include
cognitive wireless multiple-access communications
and networking, statistical signal processing and
optimization, and neural networks. Additionally, he serves as Chair of the
Communications Tech Panel, an international C3I Group, for The Technical
Cooperation Program (TTCP). Also, he is the Connectivity and Dissemination
Core Technical Competency (CTC) Lead at the Information Directorate.
Dr. Matyjas is the recipient of the 2012 IEEE R1 Technological Innovation
Award; 2012 AFRL Harry Davis Award for Excellence in Basic Research,
the 2010 IEEE International Conference on Communications Best Paper
Award, and 2009 Mohawk Valley Engineering Executive Council Engineer
of the Year Award. He is a member of the IEEE Communications, Information Theory, Computational Intelligence, and Signal Processing Societies;
Chair of the IEEE Mohawk Valley Chapter Signal Processing Society; and a
member of the Tau Beta Pi and Eta Kappa Nu engineering honor societies.
Zhi Ding (S88M90SM95F03) is the Child
Family Endowed Professor of Engineering and Entrepreneurship at the University of California, Davis.
He also holds a joint appointment as a thousandtalent professorship at Southeast University in Nanjing, China. He received his Ph.D. degree in Electrical Engineering from Cornell University in 1990.
From 1990 to 2000, he was a faculty member of
Auburn University and later, University of Iowa.
Prof. Ding has held visiting positions in Australian
National University, Hong Kong University of Science and Technology, NASA Lewis Research Center and USAF Wright
Laboratory. Prof. Ding has active collaboration with researchers from several
countries including Australia, China, Japan, Canada, Taiwan, Korea, Singapore, and Hong Kong.
Dr. Ding is a Fellow of IEEE and has been an active member of IEEE,
serving on technical programs of several workshops and conferences. He was
associate editor for IEEE T RANSACTIONS ON S IGNAL P ROCESSING from
1994-1997, 2001-2004, and associate editor of IEEE S IGNAL P ROCESSING
L ETTERS 2002-2005. He was a member of technical committee on Statistical
Signal and Array Processing and member of technical committee on Signal
Processing for Communications (1994-2003). Dr. Ding was the Technical
Program Chair of the 2006 IEEE Globecom. He is also an IEEE Distinguished Lecturer (Circuits and Systems Society, 2004-06, Communications
Society, 2008-09). He served as an IEEE T RANSACTIONS ON W IRELESS
C OMMUNICATIONS Steering Committee Member (2007-2009) and its Chair
(2009-2010). Dr. Ding is a coauthor of the book Modern Digital and Analog
Communication Systems, 4th edition, Oxford University Press, 2009.