You are on page 1of 5

Cellular Interference Alignment with Imperfect

Channel Knowledge
Roland Tresch, Maxime Guillaud
FTW (Telecommunications Research Center Vienna)
Vienna, Austria
e-mail: {tresch,guillaud}@ftw.at
AbstractInterference alignment is evaluated as a technique
to mitigate inter-cell interference in the downlink of a cellular
network using OFDMA. The sum mutual information achieved
by interference alignment together with a zero-forcing receiver
is considered, and upper and lower bounds are derived for
the case of imperfect channel knowledge. The sum mutual
information achieved by interference alignment when the base
stations share their information about the channels is shown
to compare favorably to the achievable sum-rate of methods
where the base stations do not cooperate, even under moderately
accurate knowledge of the channel state.
I. INTRODUCTION
Interference alignment was rst considered in [1] as a
coding technique for the two-user Multiple-Input Multiple-
Output (MIMO) X channel, where it was shown to achieve
multiplexing gains strictly higher that that of the embedded
MIMO interference channel (IC), multiple-access channel
(MAC) and broadcast channel (BC) taken separately. While
requiring extensive channel knowledge, this coding technique
is based only on linear precoding and zero-forcing at the
receivers. The degrees of freedom region for the X channel was
analyzed in [2] for an arbitrary number of antennas per user.
This transmission technique was later generalized to the K-
user interference channel [3], where it was shown to achieve
almost surely a sum-rate multiplexing gain of
K
2
per time,
frequency and antenna dimension. In comparison, independent
operation of K isolated point-to-point links would incur a
sum-rate multiplexing gain of K per dimension. This indicates
that interference alignment allows virtually interference-free
communications at the cost of each user exploiting only half
of its available degrees of freedom. Thanks to the alignment
of all interference signals in the same subspace from the point
of view of all receivers simultaneously, interference can be
removed simply through zero-forcing ltering.
In [3], an explicit formulation of the precoding vectors
achieving interference alignment is presented for single-
antenna nodes with time-varying channels. In the multiple-
antenna case, no such closed-form solution is known, although
achievability results on multiplexing gain are available. An
iterative algorithm was introduced in [4] to nd the precoding
matrices achieving interference alignment.
The goal of this contribution is to evaluate the applicability
of interference alignment methods to cases representative of
upcoming cellular networks. In current generation of cellular
systems, the density of the base stations and frequency reuse
factors are such that the system is in most cases interference
limited. Therefore, interference alignment has the potential to
signicantly improve the sum-throughput, without requiring
the signicant complexity increase of e.g. dirty-paper coding.
In general, interference alignment requires the knowledge of
all channels in the system. However, since it does not rely
on sharing the user data among base stations, the backhaul
capacity required for its implementation remains limited.
This article is organized as follows: the system model, the
cellular interference alignment method, and the considered
performance metric are introduced in Section II. The sum-
mutual information for the case of imperfect channel state
information is introduced in Section III, and lower and upper
bounds are proposed in Section IV. Section V presents simula-
tion results and compares the proposed interference alignment
scheme to other strategies.
II. CELLULAR INTERFERENCE ALIGNMENT
Let us consider the downlink of a system based on Orthog-
onal Frequency Division Multiple-Access (OFDMA), such as
LTE-Advanced [5], the successor of the Universal Mobile
Telecommunications System (UMTS) Long-Term Evolution
(LTE). Thanks to the frequency orthogonality introduced by
the OFDMA physical layer, we can apply interference align-
ment independently inside each OFDM subband. Therefore,
without loss of generality, we will consider only one frequency
band, with K base stations, each serving only one user,
over a frequency-at MIMO channel. The system therefore
incorporates K cells, with in each cell one base station with
N
T
transmit antennas and one user terminal with N
R
receive
antennas. Here, we will not consider the resource allocation
(scheduling) problem, and assume that the resources have
already been allocated and the channel to each user is given.
We focus our attention on the user in cell i, which receives
interference from other cells in addition to its intended signal.
The baseband model for one resource element of the consid-
ered user, e.g., subcarrier, is given by
y
i
(t) = H
ii
(t)x
i
(t) +

j=i
H
ij
(t)x
j
(t) +n
i
(t), (1)
where x
i
(t) and x
j
(t) are N
T
dimensional vectors repre-
senting the transmitted signals from base-station i and j
978-1-4244-3437-4/09/$25.00 2009 IEEE
respectively. H
ii
(t) and H
ij
(t) are N
R
N
T
matrices rep-
resenting the MIMO channels (assumed frequency-at) of the
link within the cell and the interfering link between cell i
and j, respectively, experienced by user i at time t. n
i
(t)
accounts for the thermal noise generated in the radio frequency
front-end of the receiver and interference from sources other
than the other base-stations. Here, the noise is modeled as
Gaussian i.i.d. with E
n
i
[n
i
(t)n
i
(t)
H
] = I
NR
. In the remainder
of this paper, we will not write the time index t explicitly for
notational convenience.
Interference alignment (IA) is achieved with multiplexing
gains (d
1
, . . . d
K
) (each d
i
corresponds to the multiplexing
gain achieved inside one cell, i.e. d
i
streams per transmitter
are spatially pre-coded in cell i) iff we obtain N
T
d
i
truncated
unitary matrices (precoding matrices) V
i
and N
R
d
i
trun-
cated unitary matrices (zero-forcing interference suppression
matrices) U
i
such that, for i = 1, . . . , K,
U
H
i
H
ij
V
j
= 0, j = i
rank
_
U
H
i
H
ii
V
i
_
= d
i
. (2)
Fig. 1 depicts IA for K = 3 transmitter-receiver pairs. The
iterative algorithm [4, Algorithm 1], which is based on the
minimization of an interference leakage metric (zero leakage
is equivalent to the system of equations in (2)), was used to
align interference.
s
1
s
2
V
2
V
3
s
3
H
11
H
21
H
33
H
23
n
1
n
2
n
3
U
3
U
2
U
1
y
1
y
2
y
3
V
1
d1 x NT
d2 x NT
d3 x NT
NR x d1
NR x d2
NR x d3
H
H
H
Fig. 1. IA incorporating three transmitter-receiver pairs.
The received signal of user i after interference suppression
yields
1
y
i
=U
H
i
H
ii
V
i
s
i
+

j=i
U
H
i
H
ij
V
j
s
j
+U
H
i
n
i
=U
H
i
H
ii
V
i
s
i
+ n
i
, (3)
with s
i
the d
i
dimensional signal vector of base-station
i and s
j
the d
j
dimensional signal vector of interfering
base-station j, respectively. We assume that Tr(Q
x
i
x
i
) =
Tr(V
i
Q
s
i
s
i
V
H
i
) = Tr(Q
s
i
s
i
) =
i
for i = 1 . . . K, with
Q
s
i
s
i
= E
s
i
[s
i
s
H
i
] the signal covariance . n
i
= U
H
i
n
i
is the
1
Throughout this paper v denotes the projection of a signal v onto a
subspace of the received signal space.
H
ij
V
j
H
ik
V
k
U
i
H
ii
V
i
H
H
ii
V
i
(a) perfect CSI.
H
ij
V
j
~
U
i
H
ii
V
i
~
H
~
H
ii
V
i
~
H
ik
V
k
~
U
i
H
ij
V
j
~ ~
U
i
H
ik
V
k
~ ~
H
H
(b) imperfect

U and

V.
Fig. 2. Zero-forcing interference suppression at receiver i in the presence
of interfering users j and k.
effective noise vector, which is spatially white, with covariance
I
di
due to the fact that U
i
is a truncated unitary matrix.
The second equality in (3) comes from the IA condition in
(2). Fig. 2(a) depicts the interference suppression at receiver
i, assuming that there are two other transmitters j and k
present. Here N
R
= 2 and d
i
= 1. Using the inter-cell
IA algorithm with perfect knowledge of the MIMO channels
leads to the following situation in the N
R
dimensional receive
signal space: the interference from transmitter j and k aligns
perfectly in a one-dimensional subspace. Projecting the receive
signal y
i
onto the orthogonal subspace of the interference,
using U
i
, suppresses all inter-cell interference. On the other
hand, the energy of the signal part that lies in the interference
subspace is lost.
Thus, assuming perfect knowledge of H
2
at transmitter and
receiver i leads to the following average achievable rate for
cell i [4]
R
i
=I(s
i
; y
i
|H) = E
H
_
log det
_
I
di
+

H
ii
Q
s
i
s
i

H
H
ii
_
(4)
where

H
ii
= U
H
i
H
ii
V
i
is the effective channel.
III. NOISY CHANNEL STATE INFORMATION
We focus now on the impact of noisy channel state informa-
tion (CSI) on the performance of IA. Let

H
ii
and

H
ij
denote
the MIMO channels estimated by user i and let us assume
that the IA is done based on these noisy estimates. In general,

H = H, due to both estimation noise and time variation of


the channel. Let E =

HH denote the channel measurement
error, and assume that it is a complex Gaussian circularly-
symmetric random matrix with i.i.d. components of variance

2
E
.
The receive signal of user i after interference suppression
with

U and

V computed from

H, i.e.,

U
H
i

H
ij

V
j
= 0 j = i
yields
y
i
=

U
H
i
H
ii

V
i
s
i
+

j=i

U
H
i
H
ij

V
j
s
j
+ n
i
. (5)
Note that we still project y (by the product with

U
H
i
) onto
the orthogonal subspace of the assumed interference, and that
an interference term remains (non-zero interference leakage),
due to the fact that the IA condition is fullled for the
2
H, is short hand notation for the set of all MIMO channels, i.e.,
H
ij
(i, j).
noisy estimated channels. Fig. 2(b) depicts the interference
suppression at receiver i with

U and

V that are based on the
noisy estimates

H. Now the interference from transmitter j
and k spans the full space of receiver i and it is not possible
to suppress it completely. The assumed interference

H
ij

V
j
,

H
ik

V
k
would align as indicated by the dotted arrows in Fig.
2(b).
The equivalent channel (5), assumed known at receiver i
leads to the following average achievable rate for user i
R
i
= I(s
i
; y
i
|H) =E
H,E
_
log det
_
I
di
+

H
ii
Q
s
i
s
i

H
H
ii

j=i

H
ij
Q
s
j
s
j

H
H
ij
+I
di

, (6)
with the transmit signal covariance Q
s
j
s
j
= E
s
j
[s
j
s
H
j
]
and

H
ij
must be redened as

H
ij
=

U
H
i
H
ij

V
j
(i, j).
Furthermore we used the fact that all s
i
are independent.
IV. BOUNDS OF AVERAGE RATE WITH IMPERFECT CSI
In this section we focus on the case that the channels are not
re-estimated for data transmission, i.e., the equivalent channel
(5) is not perfectly known at the receiver. We want to derive
bounds on the performance of IA for this setting, without
focusing on optimizing the signal transmit covariances Q
s
i
s
i
.
Inserting channel estimates used for the IA in (5) yields
y
i
=

U
H
i
(

H
ii
E
ii
)

V
i
s
i
+

j=i

U
H
i
(

H
ij
E
ii
)

V
j
s
j
+ n
i
=

U
H
i
(

H
ii
E
ii
)

V
i
s
i

j=i

U
H
i
E
ij

V
j
s
j
+ n
i
, (7)
where the second equality comes from the IA condition

U
H
i

H
ij

V
j
= 0 j = i. Let us examine in further detail the
structure of the interference terms in (7). Since

U
H
i

U
i
= I
di
and

V
H
j

V
j
= I
dj
,

U
H
i
E
ij

V
j
=

E
ij
has d
i
d
j
Gaussian i.i.d.
coefcients with variance
2
E
. We denote

E
ij
as the d
i
d
j
effective error matrix.
Extending the above derivation for the signal term

U
H
i
E
ii

V
i
turns (7) into
y
i
=(

U
H
i

H
ii

V
i
. .

Hii
+

E
ii
)s
i
+

j=i

E
ij
s
j
+ n
i
, (8)
where

E
ii
is the d
i
d
i
effective error matrix of the in-cell
channel and

H
ii
denotes the d
i
d
i
effective estimated channel
matrix.
A. Lower Bound of Average Mutual Information
We focus on the mutual information I(s
i
; y
i
|

H) given
the estimated channels

H
ii
and

H
ij
j = i at transmitter
and receiver i. Expanding I(s
i
; y
i
|

H =

H) into differential
entropies yields
I(s
i
; y
i
|

H =

H) = h(s
i
|

H =

H) h(s
i
| y
i
,

H =

H). (9)
We choose s
i
given

H to be Gaussian, even though the Gaus-
sian distribution may not be the one that maximizes the mutual
information for the specic interference and noise distribution.
Thus, we x the value of h(s
i
|

H =

H) = log det(eQ
s
i
s
i
)
[6] and nd an upper bound on h(s
i
| y
i
,

H =

H) following [7].
The conditional distribution p(s
i
| y
i
,

H) is not Gaussian. Using
the fact that adding a constant does not change differential
entropy, and that the entropy of a random variable with
given variance is upper-bounded by the entropy of a Gaussian
random variable with the same variance, we can derive an
upper bound
h(s
i
| y
i
,

H =

H) log det
_
eQ
gg
_
, (10)
where g = s
i
A y
i
given

H =

H, A is any given matrix
and Q
gg
= E
s
i
,s
j
,

Eii,

Eij, n
i
[(s
i
A y
i
)(s
i
A y
i
)
H
|

H =

H].
Since (10) holds for any A, we may pick A such that A y
i
is the linear minimum mean-square error (LMMSE) estimate
of s
i
in terms of y
i
given

H =

H, in order to tighten the
bound. Using the formula for the residual error of LMMSE
estimation yields [8]
h(s
i
| y
i
,

H =

H) log det
_
e
_
Q
s
i
s
i
Q
H
y
i
s
i
Q
1
y
i
y
i
Q
y
i
s
i
__
,
(11)
with
Q
y
i
s
i
=E
s
i
,s
j
,

Eii,

Eij, n
i
[ y
i
s
H
i
|

H =

H] =

H
ii
Q
s
i
s
i
,
Q
y
i
y
i
=E
s
i
,s
j
,

Eii,

Eij, n
i
[ y
i
y
H
i
|

H =

H]
=

H
ii
Q
s
i
s
i

H
H
ii
+
2
E
Tr
_
Q
s
i
s
i
_
I
di
+

j=i

2
E
Tr
_
Q
s
j
s
j
_
I
di
+I
di
. (12)
Hence a lower bound of the mutual information given

H =

H
can be derived by inserting (11) into (9) which leads to
I(s
i
; y
i
|

H =

H) log det
_
Q
s
i
s
i
_
Q
s
i
s
i
Q
H
y
i
s
i
Q
1
y
i
y
i
Q
y
i
s
i
_
1
_
. (13)
If we assume that Q
s
i
s
i
= (
i
/d
i
)I
di
i, we can further
simplify (13) using the Woodbury identity, and derive the
bound on the average mutual information as follows
I(s
i
; y
i
|

H) E
H
_
log det
_
1

2
E
d
i

i
+ 1
_

H
ii

H
H
ii
+

2
E

i
+

j=i

2
E

j
+ 1

I
di

. (14)
B. Upper Bound of Average Mutual Information
Expanding I(s
i
; y
i
|

H =

H) into differential entropies
yields
I(s
i
; y
i
|

H =

H) = h( y
i
|

H =

H) h( y
i
|s
i
,

H =

H). (15)
We upper bound the rst term of the right-hand side (RHS) of
(15) by the differential entropy of a Gaussian random variable
with the same variance as y
i
given

H =

H [9], i.e.,
h( y
i
|

H =

H) log det
_
eQ
y
i
y
i
_
, (16)
with Q
y
i
y
i
= E
s
i
,s
j
,

Eii,

Eij, n
i
[ y
i
y
H
i
|

H =

H]. Using the
fact that conditioning reduces entropy we lower bound
h( y
i
|s
i
,

H) with h( y
i
|s
1
, . . . , s
K
,

H). Since we have shown
that

E
ij
(i, j) is Gaussian, y
i
given s
1
, . . . , s
K
,

H is a
Gaussian random variable y

i
with mean

H
ii
s
i
and covariance,
Q
y

i
y

i
= E
Eii,

Eij, n
i
[ y
i
y
H
i
|s
1
, . . . , s
K
,

H]
=

2
E

j=1
s
j

+ 1

I
di
, (17)
where we used the fact that all

E
ij
(i, j) and n
i
are indepen-
dent. Thus, using (16) and the denition of conditional entropy
yields
I(s
i
; y
i
|

H) E
H
_
log det
_
eQ
y
i
y
i
__
E
s
1
,...,s
K
_
log det
_
eQ
y

i
y

i
__
, (18)
where the second expectation can be further simplied to
E
s
1
,...,s
K
_
log det
_
eQ
y

i
y

i
__
= d
i
log
2
(e)
+ d
i

E
s
1
,...,s
K

log

2
E

j=1
s
j

+ 1

.
We can further derive a closed form expression for the second
expectation of the RHS of (18) for the special case that all

j
and all d
j
are equal. We then can sum up the chi-square
random variables s
j

2
(d
j
/
j
) with 2d
j
degrees of freedom
in the RHS of (17), i.e., =

j
s
j

2
(d
j
/
j
) which has a
chi-square distribution with 2n =

j
2d
j
degrees of freedom.
This yields
E
s
1
,...,s
K
_
log det
_
eQ
y

i
y

i
__
= d
i
log
2
(e) + d
i
E

_
log
_

2
E

i
d
i
+ 1
__
= d
i
log
2
(e) + d
i
exp
_
d
i

2
E

i
_
log
2
(e)
n

p=1
E
p
_
d
i

2
E

i
_
,
(19)
where the last equality is from [10], with E
p
() the exponential
integral function of order p, i.e.,
E
p
(z) =
_

1
e
zx
x
p
dx Re{z} > 0. (20)
V. SIMULATION RESULTS
In this section we compare the average sum-rate per cell of
inter-cell IA to those achieved by other strategies that exploit
CSI within each cell independently.
For all simulations, the number of cells in the system is
K = 5, i.e. each mobile receives interference from 4 cells
in addition to the signal from its serving cell. Each base
station is assumed to be equipped with N
T
= 4 antennas,
while each receiver has N
R
= 2 antennas. The degrees of
freedom allocation for IA is d
i
= [1, 1, 1, 1, 1], i.e., one
stream is spatially pre-coded at each base-station. With those
assumptions, the iterative IA algorithm [4] always converges
to zero interference leakage in the simulations. The transmit
power for each cell is set to Tr(Q
s
i
s
i
) =
i
= 18 dB. Note
that the assumptions about the number of signicant interferers
and their power level is in line with the results of [11].
We compare the following strategies with different levels of
CSI for inter-cell IA, Q
s
i
s
i
is set to
i
/d
i
I
di
:
Inter-cell IA with perfect CSI, as in eq. (4), denoted by
IA U
H
HV,
Inter-cell IA with imperfect

U and

V but the receivers
can track the equivalent channel perfectly (for exam-
ple through re-estimation), as in eq.(6), denoted by IA

U
H
H

V,
Lower and upper bounds (eqs. (14) and (18)) introduced
in Section IV for inter-cell IA with imperfect

U and

V
and noisy CSI, denoted by Lower Bound IA

U
H

V and
Upper Bound IA

U
H

V respectively,
and the following strategies that exploit CSI independently in
each cell:
Spatial multiplexing with waterlling (WF) power allo-
cation within each cell, with perfect CSI, denoted by
waterlling HV
WF
. The achievable rate is
R
i
= I(s
i
; y
i
|H) =E
_
log det
_
I
NR
+H
ii
Q
x
i
x
i
H
H
ii

j=i
H
ij
Q
x
j
x
j
H
H
ij
+I
NR

,
(21)
with Q
x
i
x
i
= V
WF,i
Q
s
i
s
i
V
H
WF,i
i and Q
s
i
s
i
is diagonal
containing the waterlling power values. In the case
of imperfect

V
WF
, the receiver can track the channel
perfectly (i.e. Q
x
i
x
i
=

V
WF,i
Q
s
i
s
i

V
H
WF,i
i in (21),
denoted by waterlling

V
WF
).
Maximum eigenmode (EM) transmission within each cell
individually. With perfect CSI, Q
x
i
x
i
= v
EM,i

i
v
H
EM,i
i
in (21), where v
EM,i
is the vector associated to the
strongest eigenmode of the channel. With imperfect CSI,
v
EM
is affected by the channel uncertainty, i.e. Q
x
i
x
i
=
v
EM,i

i
v
H
EM,i
i in (21) (denoted respecitvely by max
Eigenmode Hv
EM
and H v
EM
).
The achieved sum-rates are averaged over the realizations of
the channel. All channels are assumed Gaussian i.i.d., with
unit variance elements for the direct links H
ii
, and variance
for the elements of the interfering channels H
ij
, i = j.
Therefore, refers to the average path-loss of the interfering
links. In Figures 3 and 4, the average sum-rate per cell is
plotted versus the relative channel measurement error denoted
by
2
H
/
2
E
, respectively for = 1 and = 0.5.
For both values of , the IA with perfect CSI clearly outper-
forms the other methods: inter-cell IA converts the interference
network into ve parallel links without interference from other
transmit-receive pairs, and the loss in multiplexing gain is
clearly compensated by the absence of interference. On the
average, the signal power loss of IA caused by the interference
suppression projections (zero-forcing) at the receivers has been
measured in the simulation to around 3dB. IA with imperfect

U and

V performs better than waterlling and maximum EM
transmission if the relative channel estimation error is above
16 dB. An improvement can be achieved by perfectly tracking
the downlink channel at each receiver.
0 5 10 15 20 25 30
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
6.5

H
2
/
E
2
[dB]
a
v
e
r
a
g
e

s
u
m

r
a
t
e

p
e
r

c
e
l
l

(
b
p
s
/
H
z
)
Interference Alignment U HV
H
Interference Alignment U H
H
~ ~
Upper Bound IA U HV
H
~ ~ ~
V
Lower Bound IA U HV
H
~ ~ ~
Waterlling HV
WF
Waterlling HV
WF
~
max Eigenmode v
EM
H
~
max Eigenmode v
EM
H
Fig. 3. Average sum-rate per cell for = 1.
Maximum EM transmission transmits one stream with
full transmit power over the strongest EM of the in-cell
MIMO channels and therefore potentially creates a lot of
interference for the neighboring cell-edge users. The effect
of this interference leakage on the average sum-rate per
cell decreases for smaller values of . The waterlling
strategy transmits two streams with optimal power values
for a maximized sum-rate within each cell. Since those two
streams create inter-cell interference the performance is worse
than maximum EM transmission. The effect of outdated

V
WF
and v
EM
further decreases the sum-rate for
2
H
/
2
E
< 15 dB.
VI. CONCLUSION
Interference alignment was evaluated in the context of
cellular networks. The achieved sum mutual information, as
well as upper and lower bounds, were derived for imperfect
channel knowledge. Performance compares favorably to what
0 5 10 15 20 25 30
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
6.5

H
2
/
E
2
[dB]
a
v
e
r
a
g
e

s
u
m

r
a
t
e

p
e
r

c
e
l
l

(
b
p
s
/
H
z
)
Interference Alignment U HV
H
Interference Alignment U H
H
~ ~
V
Upper Bound IA U HV
H
~ ~ ~
Lower Bound IA U HV
H
~ ~ ~
max Eigenmode v
EM
H
~
max Eigenmode v
EM
H
Waterlling HV
WF
Waterlling HV
WF
~
Fig. 4. Average sum-rate per cell for = 0.5.
can be achieved by other methods where the base stations do
not cooperate, even under moderately accurate channel state
knowledge.
ACKNOWLEDGMENT
This work was supported by the STREP project IST-026905
(MASCOT) within the sixth framework programme of the
European Commission, and by the Comet competence center
program of the Austrian government through the I0 project of
the Vienna Telecommunications Research Center (ftw.).
REFERENCES
[1] M. Maddah-Ali, A. Motahari, and A. Khandani, Signaling over
MIMO multi-base systems: Combination of multi-access and broadcast
schemes, in Proc. IEEE Int. Symp. Information Theory (ISIT), Seattle,
WA, USA, July 2006, pp. 21042108.
[2] S. A. Jafar and S. Shamai, Degrees of Freedom Region of the MIMO X
Channel, IEEE Trans. Information Theory, vol. 54, no. 1, pp. 151170,
Jan. 2008.
[3] V. R. Cadambe and S. A. Jafar, Interference Alignment and Degrees of
Freedom of the K-User Interference Channel, IEEE Trans. Information
Theory, vol. 54, no. 8, pp. 34253441, Aug. 2008.
[4] K. Gomadam, V. R. Cadambe, and S. A. Jafar, Approaching
the Capacity of Wireless Networks through Distributed Interference
Alignment, IEEE Trans. Information Theory, 2008, submitted.
[Online]. Available: http://arxiv.org/abs/0803.3816
[5] 3GPP, Technical Specication Group Radio Access Network; Require-
ments for Further Advancements for E-UTRA (LTE-Advanced), 3GPP,
Tech. Rep. 36.913 V8.0.0, June 2008.
[6] I. E. Telatar, Capacity of multi-antenna gaussian channels, European
Transactions on Telecommunications, vol. 10, no. 6, pp. 585595, 1999.
[7] M. M edard, The Effect upon Channel Capacity in Wireless Commu-
nications of Perfect and Imperfect Knowledge of the Channel, IEEE
Trans. Information Theory, vol. 46, no. 3, pp. 933946, May 2000.
[8] S. M. Kay, Fundamentals of Statistical Signal Processing, Vol. 1:
Estimation Theory. Prentice Hall, 1993.
[9] T. Yoo and A. Goldsmith, Capacity and Power Allocation for Fading
MIMO Channels With Channel Estimation Error, IEEE Trans. Infor-
mation Theory, vol. 52, no. 5, pp. 22032214, May 2006.
[10] H. Shin and J. H. Lee, Capacity of Multiple-Antenna Fading Channels:
Spatial Fading Correlation, Double Scattering, and Keyhole, IEEE
Trans. Information Theory, vol. 49, no. 10, pp. 26362647, Oct. 2003.
[11] 3GPP, Technical Specication Group Radio Access Network; Feasibil-
ity study on interference cancellation for UTRA FDD User Equipment
(UE), 3GPP, Tech. Rep. 25.963 V.7.0.0, Apr. 2007.

You might also like