1

Asymptotic Analysis of Linear Precoding

Techniques in Correlated Multi-Antenna
Broadcast Channels
Romain Couillet, Student Member, IEEE, Sebastian Wagner, Student Member, IEEE,

arXiv:0906.3682v1 [cs.IT] 19 Jun 2009

and Merouane Debbah, Senior Member, IEEE

Abstract
In this paper we consider MIMO broadcast channels with separable variance profile modelling antenna correlation
at both transmit and receive sides. For large number of antennas we derive the theoretical sum-rate capacity when the
transmitter performs zero-forcing or regularized zero-forcing precoding. In particular, we apply the results to volumelimited devices where the correlation originates from a dense antenna packing. Throughout this contribution we make
extensive use of recent tools from random matrix theory. Simulations confirm the theoretical claims and also indicate
that in most scenarios the asymptotic derivations applied to a finite number of users give good approximations of the
actual ergodic sum-rate capacity.
Index Terms
Broadcast channel (BC), spatial correlation, channel capacity, random matrix theory, linear precoding, beamforming, large system analysis

I. I NTRODUCTION
We consider multiple-input multiple-output (MIMO) broadcast channels (BC) composed of a transmitter equipped
with M antennas and K single-antenna receivers. It has been shown in [1] that the rate region of the MIMO-BC
with additive Gaussian noise coincides with the rate region achieved by the dirty-paper coding scheme (DPC). Since
the K-dimensional rate regions are impractical to work with, the performance of MIMO downlink transmission
schemes is evaluated as a function of the individual rates; in the present work, we will specifically focus on sumrate capacity, defined as the maximum achievable sum of individual transmission rates. Although DPC achieves the
sum-rate capacity, its inherent complexity renders this scheme unattractive for implementation. Linear precoding

R. Couillet and M. Debbah are with the Alcatel-Lucent Chair on Flexible Radio, SUPELEC,
Gif sur Yvette, 91192, Plateau de Moulon, 3,
Rue Joliot-Curie, France e-mail: {romain.couillet, merouane.debbah}@supelec.fr.
S. Wagner is with the EURECOM, Sophia-Antipolis, 06904, Route des Cretes, B.P. 193, France e-mail: sebastian.wagner@eurecom.fr
Manuscript received April 19, 2005; revised January 11, 2007.

techniques such as channel inversion (CI) or regularized-CI (R-CI) [2], [3] offer a convenient trade-off between
complexity and sum-rate performance. In particular, linear precoding schemes can achieve the full multiplexing gain
of min(K, M ) [4], i.e. the ratio sum-rate capacity over signal-to-noise ratio (SNR) grows linearly as min(K, M )
for high SNR.
The first theoretical results on MIMO channels assumed uncorrelated channels, leading to a linear increase in
sum-rate capacity for every added transmit antenna (provided that M K). However, in practical communication
systems the signals transmitted or received by different antennas tend to be correlated. This correlation mainly
depends on antenna geometry, angular spread and angles of arrival. These parameters can significantly diminish
the theoretical gains predicted for uncorrelated antenna systems. Several contributions analyzed the impact of
correlation in single-user MIMO communications for multiple antennas e.g. [5], [6], [7], [8], [9]. Particularly, in
[7], the general model of Gaussian MIMO channels with general variance profile is treated, when the numbers of
transmit and receive antennas grow large.
In this contribution we investigate the problem of antenna correlation in the MIMO-BC when the channel input
data are linearly precoded. More precisely, we analyze the impact of antenna correlation on the sum-rate capacity
for CI and R-CI precoders when the channel is modelled as Gaussian with separable variance profile. The behavior
of CI and R-CI in uncorrelated MIMO-BC has already been studied in [2], [3] for i.i.d. Gaussian channels. In
particular, the authors in [2] showed that CI achieves linear growth in multiplexing gain. As such, the purpose of
the present work is to extend the previous results in [2], [3] to Gaussian channels with separable variance profile.
We derive analytical expressions for the sum-rate capacities of CI and R-CI when (K, M ) grow large. Furthermore
we consider that the available transmit power is distributed equally among the users. Note that the work in [10]
suggests that asymptotically, equal power allocation is optimal when the channel is i.i.d. Gaussian.
The analytic expressions of the sum-rate capacities for CI and R-CI are derived by applying recent tools from
random matrix theory (RMT). These expressions are independent of the specific channel realizations and only
depend on the channel correlation profile. In particular, we utilize a recent result on the Stieltjes transform of
random matrices of the form X1/2 WQWH X1/2 + Y, with W a Gaussian random matrix [11] and X, Q, Y
deterministic matrices. As a case study we apply our results to a scenario where the correlation stems from a
dense antenna packing at the transmitter. The effect of antenna packing on volume-limited devices has already been
studied in the point-to-point MIMO channel scenario, e.g. [8], [12], [13], but not in the context of MIMO-BC, to
the authors knowledge.
In our case study of a dense antenna packing at the transmitter, we find that the sum-rate capacity is significantly
diminished by the effect of correlation at the transmitter. The sum-rate capacity grows only sub-linearly with
min(K, M ). For a fixed ratio M/K with both K, M large, the achievable sum-rate capacity is going to zero for CI
and is saturating for R-CI. By adapting the ratio M/K (i.e. the number of served users for a given M ) according
to the specific correlation, the sum-rate capacity is shown to grow sub-linearly with min(K, M ) for both CI and
R-CI.
The remainder of this paper is organized as follows: Section II briefly reviews various tools of random matrix

theory which will be used in later derivations. Section III introduces the MIMO-BC system model. In Section IV
we study channel inversion precoding, while Section V treats regularized channel inversion precoding. Section VI
provides simulation results which are shown to corroborate the theoretical derivations. In Section VII we discuss
the limitations of the present work and especially of the separable variance channel model. Finally in Section VIII
we provide our conclusions.
Notation: In the following, boldface lower-case symbols represent vectors, capital boldface characters denote
matrices (IN is the N N identity matrix). The Hermitian transpose is denoted ()H . The operator tr[X] represents
the trace of matrix X. The eigenvalue distribution of an Hermitian random matrix X is X (x). The symbol EX []
denotes expectation over random variable X. The derivative of a function f (x) of a single variable x is denoted
f 0 (x). If not stated otherwise, all logarithms are natural logarithms.
II. R ANDOM M ATRIX T HEORY T OOLS
In this work, we are interested in the behaviour of large random Hermitian matrices, and particularly in the
asymptotic distribution of their eigenvalues. Specifically, the empirical eigenvalue distribution of large Hermitian
matrices converges, in many practical cases, to a definite probability distribution, hereafter called the limit distribution
of the random matrix, when the matrix dimensions grow to infinity.
A tool of particular interest in this work is the Stieltjes transform SX of a large Hermitian nonnegative definite
matrix X, defined on the half complex space C \ R+ = {z C, Re(z) < 0}, as
Z +
1
X ()d
SX (z) =
z
0

(1)

where X is the empirical distribution of X.

Couillet et al. [14] derived a fixed-point expression of the Stieltjes transform for Gaussian matrices with leftand right-side correlations in the following theorem,
Theorem 1: Let the entries of the K M matrix W be i.i.d. Gaussian with zero mean and variance 1/M . Let
X and Q be respectively K K and M M Hermitian nonnegative definite matrices with eigenvalue distributions
X and Q . We impose further that the largest eigenvalues of X and Q are bounded independently of K, M .
Let Y be an K K Hermitian matrix with the same eigenvectors as X and let f be some function mapping the
eigenvalues of X to those of Y. Let z C+ = C \ R+ . Then, for M , K large with K/M = , the Stieltjes
transform SH (z) of H = X1/2 WQWH X1/2 + Y is approximately
1
Z 
Z
q Q (q)dq
SH (z) =
f (x) + x
z
X (x)dx
1 + qTH (z)
where TH is a solution of the fixed-point equation
Z 
Z
TH (z) = x f (x) + x

q Q (q)dq
z
1 + qTH (z)

(2)

1
X (x)dx

(3)

An immediate corollary, when only right-correlation is considered, unfolds naturally as follows,

Corollary 2: [15] Let the entries of the K M matrix W be i.i.d. Gaussian with zero mean and variance
1/M . Let Y be an K K Hermitian non-negative matrix with eigenvalue distribution Y (x). Moreover, let Q

be a M M nonnegative definite matrix with eigenvalue distribution Q (x), such that the eigenvalues of Q are
bounded irrespectively of M . Then, for large K, M , such that K/M = , the Stieltjes transform on C+ of the
matrix
H = WQWH + Y

(4)

verifies approximately


SH (z) = SY z

q
Q (q)dq
1 + qSH (z)


(5)

III. S YSTEM M ODEL

Consider the homogeneous network (i.e. all path-losses are equal) where one transmitter with M antennas
communicates to K single-antenna receivers. This is depicted in Figure 1 in the case of UCA. In addition we assume
that M/K 1. Under these assumptions the received signal vector y CK of a narrow-band communication system
reads
y = Hx + n

(6)

with transmit vector x CM , channel matrix H CKM and noise vector n CN (0, 2 IK ). The transmit signal
vector x is obtained from the symbol vector s CN (0, IK ) by a linear precoding G CM K = [g1 . . . gK ]
x = Gs

(7)

tr(E[xxH ]) tr(GGH ) P

(8)

where G satisfies the total transmit power constraint

where P > 0 is the total transmit power. The SNR at the receiver is = P/ 2 . The symbol received by user k
is given by
yk =

hH
k gk sk

K
X

hH
k gi si + nk

(9)

i=1,i6=k
M
where hH
denotes the kth row of H.
k C

With equal power allocation among the users, the signal-to-interference plus noise ratio (SINR) per user is
k =

2
|hH
k gk |
M
X
j=1
j6=k

2
|hH
k gj |

(10)

M
+

We further assume that the channel H follows the widely used Kronecker model, i.e. Gaussian with separable
variance profile [6]
H = R 1/2 Hw T 1/2

(11)

where T CM M and R CKK are non-negative definite correlation matrices at the transmitter and the
receiver, respectively. The matrix Hw CKM has i.i.d. standard Gaussian entries. Note that the Kronecker model

Fig. 1.

Broadcast channel with uniform circular antenna arrays at the transmitter and the receiver

corresponds to the most appropriate model1 when only R , T and the channel mean power are known to the
system modeler [16]. In the MIMO broadcast channel the distance between the users is supposed to be sufficiently
large compared to the signal wavelength ; hence the users are assumed to be uncorrelated, i.e. R 1/2 = IK .
In the simulations of Section VI we will use correlation matrices that originate from a dense antenna packing
on a volume-limited device. To model the correlation matrix T in this dense antenna framework, we will use the
classical Jakes model, such that, denoting dij the distance between the antennas indexed by i and j,
(T )ij = J0 (kdij )

(12)

where k = 2/ and J0 is the zero-order Bessel function of first kind. In particular we consider the uniform circular
array (UCA) model where the distance dij between antenna i and antenna j is given by


|i j|
dij = 2r sin
M

(13)

in which case T is a circulant matrix.

Back to the initial assumptions, the ergodic sum-rate capacity with equal transmit power allocation is given by
"K
#
X
R = EH
log(1 + k )
(14)
k=1

In the subsequent sections we derive asymptotic sum-rate capacities for large (K, M ) with M/K = 1 fixed,
in case of linear precoding at the transmitter and equal power allocation across the users.
1 in

the sense of the maximum entropy principle.

IV. C HANNEL I NVERSION P RECODING

Channel inversion precoding, also referred to as zero-forcing (ZF) precoding, annihilates all the inter-user
interference by performing an inversion of the channel matrix H at the transmitter. The precoding matrix Gci
is given by
Gci = HH HHH

1

(15)

where the parameter is set to fulfill the transmission power constraint (8) and depends only on the channel
realization H. Thus, we have
2 =

P
tr ((HHH )1 )

(16)

and the system model in (6) becomes

y = s + n

(17)

The signal-to-interference plus noise ratio (SINR) k,M for user k therefore expresses simply as
2
2

k,M =

(18)

which is independent of the selected user. In the subsequent derivations, we will therefore drop the user index k.
Let us denote H0 =

1 H
M

and H0w =

1 Hw .
M

It follows from (16)

P

2 =
1
M tr

H0 H0 H

(19)

1

When M is large with M/K = , the denominator of Equation (19) verifies

Z
1  0 0 H 1
1
1
1
tr H H
H0 H0 H ()d = SH0 H0 H (0)
=
M

(20)

According to Corollary 2, SH0 H0 H (0) is the solution of2

SH0 H0H (0) =

!1

1 + SH0 H0H (0)

As a consequence, for large (K, M )

M =

T ()d

2
SH0 H0 H (0)

(21)

(22)

and the sum-rate is approximately


Rci = K log 1 +

SH0 H0 H (0)

(23)

Following [2] we now look for a value ? of the ratio M/K such that, for a fixed number of transmit antennas
M , the sum-rate Rci is maximized. By differentiating (23) with respect to and setting the derivative to zero we
obtain the implicit solution
"
#
d
(S
0 H0 H (0))
?
H
d
S 0 0 H (0)1 ?
= log (1 + )
1+ HH
SH0 H0 H (0)2
2 it

is important to note here that we slightly misapply Corollary 2 since the result is only proven valid outside for any z > 0.

(24)

with = ? SH0 H0 H (0)1 , and

d
d

(SH0 H0 H (0)) is solution of the implicit equation



Z 2 d (S 0 0 H (0)) 1
HH

d
2
(SH0 H0 H (0)) = SH
0 H0 H (0)

2 T ()d
d
1 + SH0 H0 H (0)

(25)

Remark 1: In the absence of transmit correlation (i.e. T = I) Equation (21) simplifies to SH0 H0 H (0) = /(
1). Substitution into (23) leads to
Rci = K log (1 + ( 1))

(26)

?
= log (1 + ( ? 1))
1 + ( ? 1)

(27)

The implicit equation (24) becomes

The results (26) and (27) correspond to those derived in [2].

Figure 2 depicts ? and the corresponding optimal number of users K ? = M/ ? for M = 16 and different
SNR. We observe that with increasing SNR more and more users should be served to maximize the sum-rate.
When the transmit antennas are uncorrelated it is optimal to serve K ? = M users3 for , corresponding to
a multiplexing gain of M . However, if the transmit correlation increases, it is optimal to serve fewer users. In the
extreme case of fully correlated antennas ((T )ij = 1, i, j) the multiplexing gain reduces to one, i.e. TDMA is
the optimal transmission strategy.
V. R EGULARIZED C HANNEL I NVERSION P RECODING
To compensate for a possible ill-conditioning of the channel matrix H, a regularization term is introduced in
the last factor of Equation (15). In particular we will treat two schemes: i) the regularization term is computed
to satisfy the transmit power constraint (8); we refer to this scheme as constrained regularized CI (CR-CI) and
ii) the regularization term is unconstrained and can be chosen to maximize the SINR; an additional subsequent
normalization is performed to ensure that (8) is met. This second technique is referred to as regularized CI (R-CI)
and has been proposed in [3]. By construction the CR-CI scheme necessarily achieves smaller sum-rates than R-CI
due to the constraint on the regularization term. However, this scheme is introduced because it is a straightforward
extension of CI and the sum-rate expressions derived for CR-CI can easily be extended to R-CI.
A. Constrained Regularized Channel Inversion
Consider the CR-CI precoding matrix

1
Gcrci = HH HHH + M IK

1 H
(a)
= HH H + M IM
H
3 to

(28)
(29)

prove this, take > 0 and assume ? () > 1 + for sufficiently large. Than one can show the LHS of (27) is bounded while the

RHS grows unbounded.

16
Uncorrelated
r = 0.1
r = 0.2
r = 0.5
r=

14
12

10
8
6
4
2
0
5

10

15

20

25

30

35

20

25

30

35

SNR [dB]
(a)
16
Uncorrelated
r = 0.1
r = 0.2
r = 0.5
r=

14
12

K?

10
8
6
4
2
0
5

10

15
SNR [dB]
(b)

Fig. 2.

CI, UCA, optimal ? and optimal number of users K ? for different r/, M =16

where (a) follows from the matrix inversion lemma (MIL). The regularization term M IK contains the factor M
to ensure that, as (K, M ) grow large, both tr(HHH ) and tr(M IK ) grow with the same order of magnitude M K.

The parameter is set so to fulfill the transmission power constraint (8) which leads to


2 
1
H
H
P =
tr H0 H0 H0 H0 + IM
M
Z

0 H 0 ()d
=
( + )2 H H

Z 
1

H0 H H0 ()d
=

( + ) ( + )2

(30)
(31)
(32)

0
= SH0 H H0 () SH
0 H H0 ()

(33)

SH0 H H0 (z) = ST 1/2 H0w H H0w T 1/2 (z)

(34)

with

1
1
S 0
0 H (z)
Hw T Hw
z

when the number of antennas grows large, SH0w T H0w (z)H follows from Theorem 2
!1
Z

T ()d z
SH0w T H0w H (z) =
1 + SH0w T H0w H (z)

(35)

(36)

and
0
S
1/2 H0
T

H H0 1/2
T
w

(z) =

1
1 0
S 0
0 H (z) +
Hw T Hw
z 2

0
where after the differentiation of (36), we obtain the implicit equation for SH
0 H0 H (z),
T
w
w

2
Z

0
SH0w T H0w H (z)
0
2
1 +
SH
T ()d SH
0 H0 H (z) =
0 H0 H (z)
T
T
w
w
w
w
(1 + SH0w T H0w H (z))2

(37)

(38)

The received signal can be written

Let us denote H0


1
H
H
y = H0 H0 H0 + IM
H0 s + n
(39)

1
H
= [h01 , . . . , h0K ] and W = H0 H0 + IM
. Without loss of generality we will only focus

on user k. The received symbol of user k is

H

yk = h0k Wh0k sk +

K
X

h0k Wh0i si + nk

(40)

i=1,i6=k

and the SINR of user k reads k

k =

H
|h0k Wh0k |2
H
0
h0k WUH
k Uk Whk

(41)

0
0
0
0
where we write UH
k = [h1 , . . . , hk1 , hk+1 , . . . , hK ] (in other words, we remove column/user k). In the following
H

we will derive the asymptotic expressions of the signal power |h0k Wh0k |2 and the power of the interference
H

0
h0k WUH
k Uk Whk . At this point we need the following result from Bai and Silverstein.

Lemma 3: [17] Let A be a deterministic N N complex matrix with uniformly bounded spectral radius for all
N . Let x =

1 [x1 , . . . , xN ]T
N

where the {xi } are i.i.d. complex random variables with zero mean, unit variance

and finite eighth moment. Then



1
c
E | xH Ax trA |4 2
N
N

(42)

10

where c is a constant that does not depend on N or A.

Corollary 4: In the conditions of Lemma 3,
1
N
trA 0
N

xH Ax

(43)

almost surely.
1) Asymptotic Signal Power: Applying the MIL yields
H

h0k W =

h0k

UH
k Uk + IM

1

1 + h0k H UH
k Uk + IM

1

h0k

(44)

We apply the following change of variables: h0k = T 1/2 k0k and Uk = Vk T 1/2 where k0k is independent of Vk
and both have i.i.d. entries. Thus, using Corollary 4, we have approximately

1 0
H
k0k VkH Vk + T 1
kk
0H
0
hk Whk =


1
H
1
1 + k0k VkH Vk + T
k0k
h
i


1
1
VkH Vk + T 1
M tr
h
=

1 i
1
1+ M
tr VkH Vk + T 1
=

SVH Vk +T 1 (0)
k

1 + SVH Vk +T 1 (0)

(45)

(46)

(47)

2) Asymptotic Interference Power: Similarly, applying twice the MIL, we have in the large M assumption,

1 H

1 0
H
h0k UH
Uk Uk UH
hk
H
k Uk + IM
k Uk + IM
0
h0k WUH
U
Wh
=
(48)

2
k k
k


1
H
0
H
0
1 + hk Uk Uk + IM
hk

1 0


1
H
kk
VkH Vk VkH Vk + T 1
k0k VkH Vk + T 1
=
(49)

2


1
1 + k0k H VkH Vk + T 1
k0k
h

2 H i
1
Vk Vk
VkH Vk + T 1
M tr
(50)
=
h

1 i2
1
1+ M
tr VkH Vk + T 1
where the numerator of (50) can be expanded as

2 H i
1 h H
Vk Vk
tr Vk Vk + T 1
M

2

i
1 h H
=
tr Vk Vk + T 1
VkH Vk + T 1 T 1
M
i

1 i

2
h H
1 h H
=
tr Vk Vk + T 1

tr Vk Vk + T 1
T 1
M
M 
i

2 


1 h H

1
=
tr Vk Vk + T 1

tr T 1/4 VkH Vk T 1/4 + T 1/2

M
M

(51)
(52)
(53)

11

In the large M limit, we have

H

0
h0k WUH
k Uk Whk =

1
M tr


2 

1 i

tr T 1/4 VkH Vk T 1/4 + T 1/2

VkH Vk + T 1
M

h

1 i2
1
1+ M
tr VkH Vk + T 1

d
SVH Vk +T 1 (0) dx
ST 1/4 VH Vk T 1/4 +T 1/2 (0)
k
k

2
1 + SVH Vk +T 1 (0)

(54)

(55)

At this point, we need the following additional lemma, (Lemma 2.1 in [18])
Lemma 5: Let z R , X N N Hermitian nonnegative definite, R and q CN . Then


1
1 
tr (X zIN )1 (X + qqH zIN )1 
N
N |z|
Denote > 0 some value less than the minimum eigenvalue of T 1 , irrespectively of N . We can write

1 h H
1
i
VkH Vk + T 1
= Vk Vk + T 1 IN + IN
2
2

(56)

(57)

The matrix between brackets is still nonnegative definite, and then, approximately, for N large, from Lemma 5,

1 

1 i
1
1 h H
H
tr Vk Vk + T 1
=
tr VkH Vk + k0k k0k + T 1
(58)
M
M

1 
1
1
0 H 0
=
tr Hw Hw + T
(59)
M
Therefore, asymptotically, the removal of a single column in the large matrix H0w does not affect the normalized
trace in (46). The same reasoning can be applied to the normalized trace in Equation (50) and therefore, for large
N,
SVH Vk +T 1 (z) = SH0w H H0w +T 1 (z)

(60)

ST 1/4 VH Vk T 1/4 +T 1/2 (z) = ST 1/4 H0w H H0w T 1/4 +T 1/2 (z)

(61)

Remark 2: Note that the accuracy of the approximation (58) depends on the product N ||. Therefore, for N
finite, either when T 1 has very small eigenvalues or when is small, the inequality in Lemma 5 is rather loose.
This will lead to some inaccuracies in the simulations presented in Section VI, when N is not large.
Consequently, the final expression for the asymptotic SINR k is user-independent and given by
2
SH
0

k =

0
SH0w H H0w +T 1 (0) S

H H0 + 1
T
w

(0)

1/4 H0 H H0 1/4 + 1/2

T
T
T
w
w

(0) +

2
1 + SH0w H H0w +T 1 (0)

where SH0w H H0w +T 1 (0) is the solution of the fixed-point equation

!
Z

1
I ()d
SH0w H H0w +T 1 (0) = ST 1

1 + SH0w H H0w +T 1
Z
1
=

1 T ()d

1
+
1
+
S
0 H H0 + 1 (0)
H
T

w
w

(62)

(63)
(64)

12

and, from Theorem 2,4

Z 
ST 1/4 HHw Hw T 1/4 +T 1/2 (0) =

1/2

1/2
+
(1 + T (0))

1
T ()d

(65)

with T (0) the unique solution of

Z
T (0) =

1/2


1/2 +

1/2
(1 + T (0))

1
T ()d

(66)

Our interest is however in the derivative of Equation (65), which we compute as

Z
2
1/2 T 0 (0) [1 + T (0)] + 1
0
S
(0)
=
1/4 HH H 1/4 + 1/2

2 T ()d
T
w
T
T
w
1
1/2 + 1/2 ([1 + T (0)])
where T (0) verifies Equation (66) and T 0 (0) is the only solution of the fixed-point equation
Z
1
T 0 (0)[1 + T (0)]2 + 1/2
0
T (0) =

2 T ()d

1
1/2 + 1/2 ([1 + T (0)])

(67)

(68)

The corresponding sum-rate is

Rcrci = K log(1 + )

(69)

Remark 3: In the absence of transmit correlation (i.e. T = I) the solution to Equation (33) can be computed
explicitly and takes the form


1
2P + 1
= 1+
+p

P (P + 1)

(70)

and the SINR (62) simplifies to

= P

s
1

P +1
P

!2

1
P + 1
+
P

(71)

Interestingly (assuming 1) in (70) = 0 for = (P + 1)/P or equivalently for P = 1/( 1) and =

i.e. we obtain CI.
B. Regularized Channel Inversion
Consider the precoding matrix

1
Grci = ()HH HHH + M IK

1 H
= () HH H + M IM
H
4 in

(72)
(73)

this case again, we use Theorem 2 for z = 0. However, with the extra assumption (which is always met in practice) that the

eigenvalues of T 1 are greater than some value > 0, then T 1 = (T 1

SH0 H H0 +( 1 I ) ( 2 ),
T
w
w
2 N

for which the hypothesis of Theorem 2 are fulfilled.

I )
2 N

I
2 N

and SH0

H H0 + 1
T
w

(0) =

13

where () is used to normalize the transmit power (8) and the regularization term is set such that the asymptotic
sum-rate is maximized. Note that, when = 0, we fall back to the CI solution. The received signal of user k
becomes
yk = r

K
H
H
X
h0k Wh0k
h0k Wh0i
r
s
+
k

 si + nk


H
H
1
1
0
2
0
0
2
0
i=1,i6
=
k
P M tr H W H
P M tr H W H

(74)

Hence the SINR of user k reads

H

k =

where

1
M tr

|h0k Wh0k |2
0
h0k H WUH
k Uk Whk +

1
M tr

H0 W2 H0 H

(75)



H
has already been calculated in (33). Therefore we obtain the following expression for
H0 W2 H0

the asymptotic SINR

2
SH
H H + 1 (0)
w
T

k () () =

0
SHHw Hw +T 1 (0) S

1/4 HH H 1/4 + 1/2

T
w
T
T
w

(0) +

2
1 + SHHw Hw +T 1 (0) ()
(76)

0
with () = SH0 H H0 () SH
0 H H0 () and the sum-rate is

Rrci = K sup log(1 + ())

(77)

In contrast to the previous model (28), the parameter is not imposed by the transmit power constraint and
the optimal = ? can therefore be chosen to maximize the SINR. In the uncorrelated case ? = 1/() was
shown to be optimal [3]. However if correlation is present, ? depends on the eigenvalue distribution of T and
a derivation of ? from Equation (76) is more involved and will not be treated in the present article.
Remark 4: In the absence of transmit correlation (i.e. T = I) the SINR (76) takes the form
() =
where =

2
SH
H H ()
w



w
1 SHHw Hw () + [1 + SHHw Hw ()]2

q
2 + 2( 1 + 1) + ( 1 1)2 and SHHw Hw () =

1+1/+
2

(78)

is the Stieltjes transform of the

Marchenko-Pastur law.
VI. S IMULATION AND R ESULTS
In this section we evaluate by simulation how transmit correlation impacts the behavior of the sum-rate of linearly
precoded MIMO-BC when the antenna array at the transmitter is dense. We compare numerical results obtained
by Monte-Carlo simulations with our previously derived asymptotic expressions for finite (K, M ). The sum-rate
capacities are expressed in bit/s/Hz. In particular we consider two cases: i.) we vary the SNR and fix the size
of the antenna array, i.e. T is constant, and ii.) we fix the SNR ( = 10 dB) and vary the number of transmit
antennas M on a limited volume, i.e. the transmit correlation increases with increasing M . In both cases the ratio
= M/K is either fixed to = 2 or chosen to maximize the achievable sum-rate, i.e. with our previous notations,
= ? . Furthermore we apply the UCA antenna pattern in all simulations.

14

First of all we observe that in all simulations the asymptotic results closely match the numerical results even for
small values of (K, M ).
However, in order to maintain a close match between simulation and theory, the number of antennas used for
R-CI simulations is deliberately chosen larger than the number of antennas used for CI simulations.
A. Channel Inversion
Case i.) is illustrated by Figures 3 and 4 for M = 16. In Figure 3 we observe that strong correlations heavily
impact the sum-rate capacity for a constant = 2. By adapting the number of served users, i.e. = ? (24), Figure
4 shows a significant gain in the sum-rate capacity. The corresponding optimal number of users K ? is depicted in
Figure 2. Furthermore, compared to = 2, the sum-rates of the curve r = 0.5 at SNRs of 25 dB and 30 dB
remain almost unchanged when = ? . This means that = 2 is close to optimal in this case, as can be verified
from Figure 2.
The second case is presented in Figures 5 and 6. We observe from Figure 5 that the sum-rate capacity goes
systematically to zero if there is transmit correlation and if is constant. This effect can be reversed, cf. Figure 6,
by choosing = ? (24), i.e. sum-rate capacity grows unbounded.
In both cases the sum-rate capacity can be significantly increased by dynamic adaption of the maximum number
of served users.
B. Regularized Channel Inversion
Figures 7 and 8 present the achievable sum-rates in the first case for CR-CI and R-CI, respectively. We observe
that R-CI outperforms CR-CI over the whole range of SNRs. Only in the medium SNR region (10 dB to 15 dB)
both schemes perform equally. The sum-rate capacity of CR-CI saturates for high SNR, due to the constraint on
the regularization term , whereas for R-CI the sum-rate keeps increasing for high SNR. Hence, unlike R-CI, the
multiplexing gain of CR-CI does not converge to the multiplexing gain also achieved by CI for high SNR.
For R-CI, case ii.) is depicted in Figure 9. We omit the corresponding plot for CR-CI because the performance
is similar at the selected SNR of 10 dB. We observe that the sum-rate capacity saturates for large M due to the
increasing transmit correlation. In contrast to CI, the optimal number of served users K ? , mathematically involved,
is not studied in the present article.
VII. D ISCUSSION
It is important to understand the extent to which the previous results are valid. In particular the Kronecker
model with inner Gaussian i.i.d. matrices is asymptotically inappropriate in practice. Indeed, when the transmit and
receive antenna arrays grow dense, the model implicitly assumes an asymptotically large number of scatterers in the
environment. This is dishonest since one obviously knows that the typical density of scatterers in the environment
is somehow limited. Another dishonest assumption which we used in simulations lies in the isotropic directions of
signal departure and arrival which leads to Jakes model. Again, field measurements suggest that the correlation

15

70
Uncorrelated
r = 0.1
r = 0.2
r = 0.5
r=

60

Rci [bit/s/Hz]

50
40
30
20
10
0
5

10

15

20

25

30

35

SNR [dB]
Fig. 3.

CI, UCA, Ergodic MIMO sum-rate Rci for different r/, = 2, M =16, numerical simulations are indicated by circle marks

figure is usually different from Jakes model but general modelling is more involved when this knowledge is taken
into account.
Concerning the accuracy of the theoretical results compared to Monte Carlo simulations, the reader must be
aware of the susceptibility of the different lemmas and theorems used in this work regarding high SNR regions.
When high SNRs are considered, the finite K, M approximations are only trustworthy when both K and M are
extremely large. For instance, the significant discrepancy between theoretical results and simulations observed in
the high SNR region of Figure 8 is linked to Remark 2: in this region, tends to 0 (see e.g. [3]) and therefore
the approximation taken from Lemma 5 is only valid for very large N . On the contrary, for moderate SNR values,
ergodic sum-rate capacities obtained from simulations and theoretical formulas do in general match closely.
Finally, on an information theoretic viewpoint, the fundamental result to be remembered in this contribution is
the fact that, due to correlation, the spatial dimension is not an infinite source of potential capacity gain. Even small
correlations at the transmission is already detrimental to the affordable transmission rates.

16

70
Uncorrelated
r = 0.1
r = 0.2
r = 0.5
r=

60

Rci [bit/s/Hz]

50
40
30
20
10
0
5

10

15

20

25

30

35

SNR [dB]
Fig. 4.

CI, UCA, Ergodic MIMO sum-rate Rci for different r/, = ? , M =16, numerical simulations are indicated by circle marks

17

40
Uncorrelated
r = 0.1
r = 0.2
r = 0.5
r=

35

Rci [bit/s/Hz]

30
25
20
15
10
5
0

8 10 12 14 16 18 20 22 24 26 28 30 32
M

Fig. 5.

CI, UCA, Ergodic MIMO sum-rate Rci for different r/, = 2, SNR=10 dB, numerical simulations are indicated by circle marks

18

40
Uncorrelated
r = 0.1
r = 0.2
r = 0.5
r=

35

Rci [bit/s/Hz]

30
25
20
15
10
5
0

8 10 12 14 16 18 20 22 24 26 28 30 32
M

Fig. 6. CI, UCA, Ergodic MIMO sum-rate Rci for different r/, = ? , SNR=10 dB, numerical simulations are indicated by circle marks,

in simulations ? is rounded to the nearest integer

19

70
Uncorrelated
r = 0.1
r = 0.2
r = 0.5
r=

60

Rcrci [bit/s/Hz]

50
40
30
20
10
0
5

10

15

20

25

30

35

SNR [dB]
Fig. 7.

CR-CI, UCA, Ergodic MIMO sum-rate Rcrci for different r/, = 2, M =64, numerical simulations are indicated by circle marks

20

70
Uncorrelated
r = 0.1
r = 0.2
r = 0.5
r=

60

Rrci [bit/s/Hz]

50
40
30
20
10
0
5

10

15

20

25

30

35

SNR [dB]
Fig. 8.

R-CI, UCA, Ergodic MIMO sum-rate Rrci for different r/, = 2, M =64, numerical simulations are indicated by circle marks

21

70
Uncorrelated
r = 0.1
r = 0.2
r = 0.5
r=

60

Rrci [bit/s/Hz]

50
40
30
20
10
0

8 10 12 14 16 18 20 22 24 26 28 30 32
M

Fig. 9. R-CI, UCA, Ergodic MIMO sum-rate Rrci for different r/, = 2, SNR=10 dB, numerical simulations are indicated by circle marks

22

VIII. C ONCLUSION
In this work, we provide expressions for the sum-rate capacity of MIMO-BC with correlated antennas for channel
inversion and regularized channel-inversion precoding, when the numbers of antennas and users grow large. The
analytical formulas are particularly appealing since they are based on tractable and numerically solvable fixed-point
equations. The results are applied to the problem of dense antenna packing on volume-limited transmitters. In this
case, the asymptotic formulas are verified and, when applied to small numbers of antennas or users, are shown to
be generally good approximations to the actually achievable sum-rates.
ACKNOWLEDGMENT
The authors would like to thank Jack Silverstein for his valuable help with Theorem 1.
R EFERENCES
[1] H. Weingarten, Y. Steinberg, and S. Shamai, The Capacity Region of the Gaussian Multiple-Input Multiple-Output Broadcast Channel,
Information Theory, IEEE Transactions on, vol. 52, no. 9, pp. 39363964, 2006.
[2] B. Hochwald and S. Vishwanath, Space-Time Multiple Access: Linear Growth in the Sum Rate, in Proceedings of the annual allerton
conference on communication control and computing, vol. 40, no. 1.

The University; 1998, 2002, pp. 387396.

[3] C. Peel, B. Hochwald, and A. Swindlehurst, A Vector-Perturbation Technique for Near-Capacity Multiantenna Multiuser Communication
Part I: Channel Inversion and Regularization, Communications, IEEE Transactions on, vol. 53, no. 1, pp. 195202, 2005.
[4] N. Jindal and A. Goldsmith, Dirty-paper coding versus TDMA for MIMO Broadcast channels, Information Theory, IEEE Transactions
on, vol. 51, no. 5, pp. 17831794, 2005.
[5] D. Shiu, G. Foschini, M. Gans, J. Kahn, Q. Inc, and C. Santa Clara, Fading Correlation and Its Effect on the Capacity of Multielement
Antenna Systems, IEEE Transactions on Communications, vol. 48, no. 3, pp. 502513, 2000.
[6] C. Chuah, D. Tse, J. Kahn, and R. Valenzuela, Capacity scaling in MIMO wireless systems under correlated fading, Information Theory,
IEEE Transactions on, vol. 48, no. 3, pp. 637650, 2002.
[7] A. Tulino, A. Lozano, and S. Verdu, Impact of antenna correlation on the capacity of multiantenna channels, Information Theory, IEEE
Transactions on, vol. 51, no. 7, pp. 24912509, 2005.
[8] S. Wei, D. Goeckel, and R. Janaswamy, On the Asymptotic Capacity of MIMO Systems With Antenna Arrays of Fixed Length, Wireless
Communications, IEEE Transactions on, vol. 4, no. 4, pp. 16081621, 2005.
[9] H. Shin, M. Win, and M. Chiani, Asymptotic statistics of mutual information for doubly correlated MIMO channels, Wireless
Communications, IEEE Transactions on, vol. 7, no. 2, pp. 562573, February 2008.
[10] H. Viswanathan and S. Venkatesan, Asymptotics Of Sum Rate for Dirty Paper Coding and Beamforming In Multiple-Antenna Broadcast
Channels, in Proceedings Of The Annual Allerton Conference On Communication Control and Computing, vol. 41, no. 2. The University;
1998, 2003, pp. 10641073.
[11] A. Tulino and S. Verdu, Random Matrix Theory And Wireless Communications.

Now Publishers Inc, 2004.

[12] D. Gesbert, T. Ekman, and N. Christophersen, Capacity Limits of Dense Palm-Sized MIMO Arrays, GLOBECOM-NEW YORK-, vol. 2,
pp. 11871191, 2002.
[13] T. Pollock, T. Abhayapala, and R. Kennedy, Fundamental Limits of MIMO Capacity for Spatially Constrained Arrays, in Proc. 4th
Australian Communications Theory Workshop, pp. 712.
[14] R. Couillet, M. Debbah, and J. W. Silverstein, A deterministic equivalent for the capacity analysis of correlated multi-user MIMO
channels, submitted to IEEE Trans. on Information Theory, arXiv preprint http://arxiv.org/abs/0906.3667.
[15] J. Silverstein and Z. Bai, On the Empirical Distribution of Eigenvalues of a Class of Large Dimensional Random Matrices, Journal of
Multivariate Analysis, vol. 54, pp. 175175, 1995.
[16] M. Debbah and R. Muller, MIMO channel modeling and the principle of maximum entropy, Information Theory, IEEE Transactions on,
vol. 51, no. 5, pp. 16671690, 2005.

23

[17] Z. Bai and J. Silverstein, No eigenvalues outside the support of the limiting spectral distribution of large dimensional sample covariance
matrices, Annals of Probability, vol. 26, pp. 316345, 1998.
[18] , On the signal-to-interference ratio of CDMA systems in wireless communications, ANNALS OF APPLIED PROBABILITY, vol. 17,
no. 1, p. 81, 2007.