Вы находитесь на странице: 1из 16

2112 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO.

10, OCTOBER 2013

Spectral Efficiency of Distributed MIMO Systems


Dongming Wang, Member, IEEE, Jiangzhou Wang, Senior Member, IEEE, Xiaohu You, Fellow, IEEE, Yan Wang,
Ming Chen, and Xiaoyun Hou

AbstractDistributed multi-input multi-output (D-MIMO) [5], and some practical issues in the design of DAS are also
system is a promising system to greatly improve the spectral studied in [6][8]. In DAS, many remote antenna units (RAUs,
efficiency and power efficiency of the cellular system. The also named as remote radio head in LTE) are distributed over a
performance analysis of the spectral efficiency of D-MIMO
system is a fundamental problem for both theoretical study and large area and connected to a baseband processing unit (BPU)
technique evaluation. In this paper, the theoretical performances by fiber. RAUs are only used to transmit and receive signals.
of spectral efficiency for D-MIMO system and traditional col- At the BPU, the signals associated with different connected
located MIMO (C-MIMO) system are studied and compared. RAUs are processed by using the advanced signal processing
First, a composite D-MIMO channel including path loss, shadow techniques. Since there is no sophisticated signal processing
fading and multipath fading is given. Conditioned on the desired
user position, by using the tight bounds and central limit theory, at RAU, its size and cost can be reduced. A special case of
the analytical approximations of the mean and the cumulative DAS is considered as cooperative base-station [9][11], where
distribution function (CDF) of the mutual information (MI) are all RAUs are base stations.
derived for C-MIMO and D-MIMO channels at both high signal
to noise ratio (SNR) and low SNR. Assuming that the users In [12], a generalized DAS was proposed and was also
are randomly distributed in the cell, the CDFs of the spectral called as distributed MIMO (D-MIMO) system. D-MIMO
efficiency are also given for the C-MIMO and D-MIMO cellular system combines the advantages of point-to-point MIMO and
system, and the closed-form expressions for the mean spectral DAS, and thus has the ability of exploiting both spatial micro-
efficiency, mean outage spectral efficiency are derived. Finally, and macro- diversities [12], [13]. However, unlike collocated
the theoretical comparisons between C-MIMO and D-MIMO for
large number of antennas are given, and simulation results are MIMO (C-MIMO) system, D-MIMO system suffers from
presented which validate the analytical results. different degrees of path losses caused by different access
distances to different distributed antennas. This makes the
Index Termsdistributed antenna system (DAS), distributed
multiple-input multiple output (MIMO), spectral efficiency. spectral efficiency analysis of D-MIMO more challenging. In
[14], [15], the uplink spectral efficiency of C-MIMO cellular
system was studied. Based on Wyners circular cellular array
I. I NTRODUCTION model, the studies for the spectral efficiency of cooperative
NE of the most important features in the fourth genera- base stations were given in [16], [17]. However, the Wyner
O tion (4G) mobile systems is to provide much higher rate
data services to large number of users without corresponding
model is highly idealized and it is not practical for cellular
systems. In real environments, path loss and shadowing fading
increases in transmit power and bandwidth. For traditional are crucial factors to realize transmit (or receive) diversity
centralized antenna systems, due to the inter-cell interference, schemes in the D-MIMO system. Yet, due to the complexity
the spectral and power efficiencies are very low, especially at of the composite Rayleigh-lognormal distribution, only few
the cell edge [1]. Distributed antenna system (DAS), which analytical studies are available for the D-MIMO system. In
has been traditionally used to cover the dead spots in indoor [12], [18], the studies for the D-MIMO system were presented
wireless communications, recently received more attention for by using Monte-Carlo simulation. In [13], analytical results of
its capability of lowering radiated power (and hence other-cell the capacity for DAS with single antenna were presented. In
interference), maximizing coverage, and improving spectral [19], the performance comparisons of microcellular and DAS
efficiency [2][4]. The concept of DAS is developed for the were given. In [20], the capacity of uplink multi-user CAS
current version of Long Term Evolution-Advanced (LTE-A) and DAS was compared. Using the composite channel model
of [21], an accurate approximation for the the outage capacity
Manuscript received October 14, 2012; revised March 20, 2013. This of the D-MIMO channel was given by [22]. However, the
work was supported in part by the Natural Science Foundation of China analytical result is complex and it is difficult to further study
(NSFC) under grant 61221002, the National Basic Research Program of China
(973 Program 2013CB336600), China International Cooperative Project No.
the cell-wide performance of D-MIMO system.
2009DFB10110, and the NSFC under grants 61271205, 61071112, 61101086. The cell averaged spectral efficiency and cell averaged
This paper was presented in part at the IEEE International Conference on
Communications (ICC), Beijing, P. R. China, May 2008. outage spectral efficiency are important parameters to evaluate
D. Wang, X. You, Y. Wang, M. Chen are with the National Mobile the performance of the cellular systems, and they are often ob-
Communications Research Lab., Southeast University, Nanjing, 210096, P.R. tained by using complex system-level simulations in industry.
China (email: {wangdm, xhyu, yanwang, chenming}@seu.edu.cn).
J. Wang is with School of Engineering and Digital Arts, University of Kent, For traditional cellular system, the cell-wide spectral efficiency
United Kingdom (email: j.z.wang@kent.ac.uk). was studied in [23]. Recently, in [24], a tractable approach was
X. Hou is with Institute of Signal Processing and Transmission, Nanjing proposed to analyze the coverage and rate of the traditional
University of Posts and Telecommunications, Nanjing, P. R. China (email:
houxy@njupt.edu.cn). cellular networks. Most recently, based on [21], [22], in [25],
Digital Object Identifier 10.1109/JSAC.2013.131012. Lee et al. studied the cell averaged spectral efficiency of D-
0733-8716/13/$31.00 
c 2013 IEEE
WANG et al.: SPECTRAL EFFICIENCY OF DISTRIBUTED MIMO SYSTEMS 2113

MIMO system. But the cell averaged outage spectral efficiency


is not given.
In this paper, our focus is on the performances of the cell-
wide spectral efficiency of the (M, L, N ) D-MIMO systems,
RAU3 D3 , S / 2
where M is the number of antennas of the mobile terminal,
d3
L is the number of antennas per RAU, and N is total number
of RAUs. We define the mean spectral efficiency (MSE) D3
RAU4 D4 ,5S / 6 RAU2 D2 , S / 6
and mean outage spectral efficiency (MOSE) over the whole
cellular systems. More practical composite fading channels
including shadowing, multipath fading and path loss are RAU1

considered. We first study the mean and variance of the mutual
information (MI) of the D-MIMO channel. By using the BPU
tight bounds and central limit theorem (CLT), approximated
distribution of the MI is given conditioned on the user position.
With the statistics of the MI, and taking into account the
RAU5 D5 , 7S / 6 RAU7 D7 ,11S / 6
effect of random location of the user in a cell, we investigate
performance of the MSE and MOSE for the D-MIMO cellular
system. Finally, we compare the performance of C-MIMO and
RAU6 D6 ,3S / 2
D-MIMO systems for the large number of antennas. The main
D
differences between this paper and [21] are as follows:
We present a CLT for the logarithm of the sum of the
random variables. With this theorem, we show that for
M L, the lower bound of the MI is asymptotic
Fig. 1. Distributed radio system (There are seven RAUs, each of which
Gaussian distribution, and then we derive the mean and is equipped with three antennas, and the mobile user is equipped with two
variance of the MI. We also show that the variance of antennas).
the MI is asymptotically zero for large N .
We study the performance of the MI for low SNRs. With
the CLT, a lognormal approximation for the distribution the squares of k i.i.d Gaussian random variables with zero
of the MI is proposed. mean and variance 1/2. erfc() is the complementary error
We give the theoretical comparisons between C-MIMO function defined by

and D-MIMO for large number of antennas. It is shown 2 t2
when M N L, the MSE of the C-MIMO is larger than erfc (x) = e dt,
x
that of D-MIMO. On the contrary, when M L, and
N is very large, the MSE of D-MIMO is larger than and erfc1 () is the inverse complementary error function.
that of C-MIMO. Moreover, in most of the cases, the
D-MIMO cellular system has larger MOSE than the C- II. S YSTEM M ODEL AND F UNDAMENTALS
MIMO cellular system. The first part of this section describes the distributed cellular
The paper is organized as follows. In Section II, the D- systems and presents the mathematical description of the
MIMO cellular system and the channel model are introduced signal and the channel models. The random location model
as well as some fundamentals related to CLT, which are of a user is described in the second part. The MI of a D-
required in the remainder of the paper. In Section III, the MIMO channel is then presented in the third part. The last
statistics of the MI of the D-MIMO channels are presented. part of this section presents fundamentals to be used to derive
The spectral efficiency over the cell is analyzed for the C- capacities of the C-MIMO and the D-MIMO channels. Two
MIMO and the D-MIMO cellular systems in Section IV, and theorems on independent sequences of nonnegative random
the theoretical comparisons are given in Section V. Finally, variables are given.
concluding remarks are summarized in Section VI.
The notation adopted in this paper conforms to the follow- A. Distributed MIMO System and Channel Model
ing convention. Vectors are column vectors and are denoted
We take Fig. 1 as an example to show a distributed radio
in lower case bold: x. Matrices are upper case bold: A. Ik
system with a generalized cell (GN-Cell). In Fig.1, there are
denotes the identity matrix of size kk. ()T and ()H represent
seven RAUs (i.e. N = 7), and the BPU is located at the
transpose, and Hermitian transpose, respectively. Tr(A) is the
central RAU. The location of RAUs could be optimized as
trace of A and det(A) denotes the determinant of A.
in [26]. We assume a single user scenario for simplicity of
denotes the Kronecker product. The operator E() denotes
analysis, but this assumption holds for most practical multiuser
expectation and V() denotes variance. The squared 2 Frobenius
2    systems, such as time division multiple access, and orthogonal
norm of A is defined as AF = i,j [A]i,j  . N (, 2 ) frequency division multiple access [27], [28]. We assume all
stands for normal distribution with mean and variance 2 . N L antennas have a separate feeder to the BPU where all
The chi-squared distribution with k degrees of freedom, which signal processing is performed. When N = 1, it becomes a
is denoted as 2k , is defined as the distribution of the sum of traditional C-MIMO system.
2114 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO. 10, OCTOBER 2013

Due to the nature of largely separated antennas comprising In practice, the user distribution is subject to a constraint that
the array, the signal model for distributed radio system should the shortest distance of the mobile from the RAU, is D0 with
encompass not only small scale fading but also large scale 0 < D0  D. Thus, the probability density functions (PDFs)
fading and path loss. In this study, a noise-limited GN-Cell of the mobiles polar coordinate (, ) relative to the center
is considered. At the BPU, the overall uplink received signal of the coverage area are approximated as
can be written as 2
p () = , 0D (7)
r = H (d) u + n, (1) D2
1
where r = [r1 , . . . , rN L ]T is the discrete-time received signal p () = , 0   2. (8)
T 2
vector, u = [u1 , . . . , uM ] is the transmitted
 signal vector
satisfying the power constraint E uH u  P , and n is the
C. MI of the D-MIMO Channel
complex additive white Gaussian  noise (AWGN) vector with
covariance matrix of E nnH = IN L . The distance vector The MI in bits per second per Hertz (bps/Hz) of the
d of length N indicates the distances from the user to the corresponding D-MIMO channel is given by
T
N RAUs, d = [d1 , d2 , . . . , dN ] . Also, the channel matrix  
I = log2 det IN L + H (d) H H (d) , (9)
H (d), is parameterized by the distance vector d and consists  H
of N independent subchannel matrices for each RAU, where = E uu . It is assumed that the channel is
unknown at the transmitter and perfectly known at the receiver.
H1 (d1 )
.. As such, one transmission strategy is to transmit multiple
H (d) = . independent complex Gaussian data streams and split the total
HN (dN ) power equally among them. One advantage of using equal-
power allocation transmission is that in many situations it
where the subchannel matrix Hn (dn ) (n = 1, 2 . . . , N ) is an maximizes the worst-case capacity and therefore provides a
L by M channel matrix from the mobile to the nth RAU. The robust transmission [30]. The information rate achieved by
composite fading channel matrix Hn (dn ) is modelled as this equal-power transmission scheme is given by
Hn (dn ) = hsh,n Hw,n , (2)  
P H
I = log2 det IM + H (d) H (d) . (10)
where Hw,n represents the small-scale fast fading of the M
channel, and hsh,n represents the large scale fading and path Substituting (2) and (3) into (10) gives rise to
loss. Hw,n is a matrix of size LM which contains zero mean  
circularly symmetric complex Gaussian random variables of I = log2 det IM + HwH ( IL ) Hw , (11)
M
variance 1/2 per dimension. The large scale fading is assumed
to be [29] where
cP
csn  , (12)
hsh,n = , (3) D
dn
Hw,1
where is the path loss exponent, typically between 3.0 and ..
5.0, c is the median of the mean path gain at a reference Hw  . (13)
distance dn = 1km, and sn is a log-normal shadow fading Hw,N
variable, i.e,
(D/d1 ) s1 0
10 log10 sn N (0, sh
2
), (4)
 .. . (14)
.
ln(sn ) N (0, 2 sh
2
), (5) 0
(D/dN ) sN
where  ln1010 . Furthermore, we assume the elements of In C-MIMO system, is the median of the SNR at the cell
Hw,n are independent of sn . boundary [29] and represents the noise level of the D-MIMO
system at the reference distance D.
B. User Random Location Model The MI, I, with equal-power allocation given in (11) is
a random variable depending on channel condition and user
As the distances between the user and the RAUs impact
position (distances from all the RAUs), and it is necessary
the channel directly, a key problem for the capacity analysis
to study its statistics, in particular its mean E (I) and its
is to investigate their characteristics. For simple analysis, the
CDF. Note that when the fading process is ergodic, a Shannon
GN-Cell shape is approximated by a circle of radius D. The
capacity or ergodic capacity exists and is given by E (I), while
RAUs polar coordinates relative to the center of the coverage
the CDF is defined as the probability that the MI exceeds a
area are denoted as (Dn , n ) (n = 1, 2, . . . , N )(see Fig. 1).
predetermined rate threshold.
The mobile user is assumed to be independent and uniformly
In the following three sections, we first study the statistical
distributed in the GN-Cell, and the mobiles polar coordinate
properties of the MI in the C-MIMO and the D-MIMO chan-
is denoted as (, ). Then, the distance dn from the mobile to
nels conditioned on the user position. Its first two moments
the n-th RAU is calculated as
 are defined as
dn = 2 + Dn2 2Dn cos ( n ). (6) I (d)  Es,Hw (I) , (15)
WANG et al.: SPECTRAL EFFICIENCY OF DISTRIBUTED MIMO SYSTEMS 2115

I2 (d)  Vs,Hw (I) . (16) Proof: Since X is the sum of the squares of 2K i.i.d
Gaussian random variables with zero mean and variance, 1/2,
Then, the averages of the ergodic capacity and outage capacity according to the CLT and Theorem2, as K , both the
over the cell are derived. limiting distributions of XK
and K [ln (X) ln (K)] are
K
Gaussian.
D. Preliminary Theorems and Lemmas Lemma 1 shows that for larger K, 22K can be approximated
Based on the CLT for the triangular arrays of independent by log-normal distribution, and it will be used to approximate
random variables [31], the following theorems for the inde- the distribution of the MI of the D-MIMO channels at low-
pendent sequence of nonnegative random variables are proved, SNR.
and they will be used for deriving the limit distribution of the Lemma 2. If pp matrices Ak , k = 1, 2, . . . , K, are positive
MI of the D-MIMO channel. definite or semi-definite matrices, then
Theorem 1. Let X1 , X2 , . . . be independent sequence of K  p1 K
  1
nonnegative random variable, satisfying det Ak det (Ak ) p ,
k=1 k=1
E (Xk ) = k , E(Xk k )2 = k2 , (17)
Proof: By successively using the Minkowskis inequality
K
 K

2 [32, Theorem 17.9.7],
K = k , qK = k2 , (18) 1 1 1
k=1 k=1 det (A1 + A2 ) p det (A1 ) p + det (A2 ) p ,
and the mean K and the variance qK 2
go to infinity as K the result is easily obtained.
2+
. Suppose that |Xk | are integrable for some positive
and that Lyapounov condition III. T HE S TATISTICAL P ROPERTIES OF MI IN THE
K
 D-MIMO C HANNELS WITH F IXED U SER P OSITION
1 2+
lim E|Xk k | =0 (19) In this section, we will study the statistics of MI of the D-
K q 2+
k=1 K MIMO channel with fixed user position. The capacity analysis
holds. QK is defined as for traditional C-MIMO has been studied by many researchers.
 K
 According to the results of [33][35], the ergodic capacity of
 (11) has closed-form expression for arbitrary M , N and L .
QK = ln Xk . (20)
k=1
However, based on the results of [33][35], it is intractable to
obtain the cell-wide spectral-efficiency. Then, in this section,
If we will use tight bounds and CLT to obtain the simple
qK
lim K = 0, (21) expressions of the mean and variance of the MI for C-MIMO
K 
k and D-MIMO channels.
k=1 In the following of the paper, we will consider two cases:
then as K , M N L and M L. We should note in the current LTE-

K A standard, up to four antennas is supported for the user
k   K

 equipment. Then M N L may be possible in practice.
k=1 d
QK ln k N (0, 1) .
Furthermore, as we will see in the Section IV and Section V,
qK when comparing the performance of C-MIMO and D-MIMO,
k=1

Proof: See Appendix A. we will have different conclusions for the two cases.

Theorem 2. Suppose that X1 , X2 , . . . are independent se- A. Approximation of the CDF of the MI in the High-SNR
quence of nonnegative random variables having the same Regime with M N L
distribution with mean and finite positive variance 2 . For
QK given in (20), as K Theorem 3. For (M, L, N ) D-MIMO system, suppose that
  M is greater than or equal to N L and the distances from the
d 2 user to RAUs are known. At high SNR, the mean and variance
K [QK ln (K)] N 0, 2 .
of the MI in (11) are given by
Proof: According to Theorem 1, one can obtain the result. N
D
ID-MIMO (d) = N L log2 + L log2
M n=1
dn
Comparing CLT and Theorem 2, note thatfor i.i.d.nonneg-

K 
K
1 
NL
ative random variables, both Xk and ln Xk can be + (M i + 1), (22)
k=1 k=1 ln 2 i=1
approximated by Gaussian random variable for large K. For
chi-squared distribution, we have the following lemma.  NL

1 

Lemma 1. Suppose that X 22K. If K , the limiting I2D-MIMO (d) = 2 N L2 2 sh
2
+ (M i + 1) ,
distributions of both XK
and K [ln (X) ln (K)] are (ln 2) i=1
K
normal. (23)
2116 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO. 10, OCTOBER 2013

respectively, where () is Eulers digamma function, which Proof: Theorem 4 can be proved similarly to Theorem 3.
for natural arguments can be expressed as
m1
 1 Theorem 4 shows that at high SNR, the mean of the MI is
(m) = (1) + the same as the i.i.d Rayleigh fading MIMO channels [36].
p=1
p
However, the variance of the MI of the composite MIMO
with (1) 0.577215 the Euler-Mascheroni constant, channel with Rayleigh-lognormal fading is much larger than
 () is the first derivative of the digamma function and it that of the i.i.d Rayleigh fading MIMO channel.
can be expressed as

1
 (x) = 2.
B. Asymptotic Distribution of the Lower Bound of MI with
p=1 (p + x 1) M L
Proof: As is very large, (11) can be approximated as When M L, the MI in (11) can be written as
[36]  N   
   D
I log2 det ( IL ) Hw HwH . (24) ID-MIMO = log2 det IM + H
sn Hw,n Hw,n .
M M n=1 dn
Substituting (14) into (24), one obtains (29)
N   The explicit lower bound of (29) is derived as follows.
D
I N L log2 + L log2 Theorem 5. For (M, L, N ) D-MIMO system, suppose that
M n=1
dn
M is less than or equal to L and the distances from the user
N
    to RAUs are known. The lower bound of the ergodic capacity
+L log2 (sn ) + log2 det Hw HwH . (25) can be written as
n=1   M

By using (5) and applying Theorem 2.11 in [37, Sect.2.1.6], 1 
ID-MIMO (d)  M log2 1 + exp (L i + 1)
the mean and variance of I given by (22) and (23), respec- M M i=1
tively, can be numerically calculated.
   N   
In [38], it is proved that log2 det Hw HwH asymptotically D
. (30)
has a normal distribution, when M or N L is or both are large. n=1
dn
Since log2 sn is i.i.d Gaussian, the MI in (25) is asymptotically
Gaussian. As we will show that even for few antennas, Proof: From Lemma 2, (29) is lower bounded according
Gaussian approximation of ID-MIMO also has satisfactory per- to
 N  
formance. For the D-MIMO channel with M N L, we also  D
see that in the high SNR regime the mean of the MI is not ID-MIMO  M log2 1 + sn
M n=1 dn
related to the variance of log-normal shadow fading. That is, 
the macro-diversity has no contribution to the ergodic capacity.  H  M1
(23) shows that the variance of MI is not a function of d. det Hw,n Hw,n . (31)
Since the C-MIMO channel is a special case of the D-
MIMO channel, the following theorem is drawn. To give a For (1, L, N ) D-MIMO
 system (i.e. M = 1), the equality
fair comparison between the D-MIMO and C-MIMO, it is 
N
xi
holds. Note that ln 1 + ai e is a convex function in
assumed that for the C-MIMO system, all of the N L antennas i=1
are collocated at the center of the cell. xi for ai > 0 [39], and using Jensens inequality, we have a
lower bound on the ergodic capacity, given by (32)(34).
Theorem 4. For (M, N L, 1) C-MIMO system, suppose that
M is greater than or equal to N L and the distance from the Theorem 6. For (M, L, N ) D-MIMO system, suppose that M
user to the base station, denoted as d, is known. Then, from is less than or equal to L and the distances from the user to
(11), the MI of the C-MIMO channel reduces to RAUs are known. At high SNR, the lower bound of the MI in
    (31) is defined as
D
IC-MIMO = log2 det IN L + sHwH Hw . (26) N   
M d  D  H  M1
ID-MIMO, LB  M log2 sn det Hw,n Hw,n
At high SNR, the mean and variance   of MIare
 given by n=1
dn
D  
IC-MIMO (d) = N L log2 + M log2 . (35)
M d M
N L
1  As N is very large, the distribution of ID-MIMO, LB is approxi-
+ (M i + 1) (27) mately Gaussian distributed, and the mean and variance are
ln 2 i=1
given by (36) and (37) respectively, where () denotes the
 NL

1  Gamma function
I2C-MIMO (d) = 2 2 2 2
2 N L sh +  (M i + 1) , 
(ln 2) i=1 (a) = xa1 ex dx.
(28) 0

respectively. Proof: See Appendix B.


WANG et al.: SPECTRAL EFFICIENCY OF DISTRIBUTED MIMO SYSTEMS 2117

We should note that according to Jensens inequality, (36) can be approximated by log-normal random variable which
is an upper bound of (35). Then, the approximation (36) is satisfies
M
neither lower bound nor upper bound of the ergodic capacity. D 1 
For C-MIMO with M L, in the high SNR regime, Wn  E [ln (Wn )] = ln + (L j + 1),
dn M j=1
exchanging N L and M in (27) and (28), we obtain the mean
and variance of MI. Different from C-MIMO, Theorem 6 M
1  
shows that when N is large, the ergodic capacity of the D- 2
W n
 V [ln (Wn )] = 2 sh
2
+ (L j + 1).
MIMO channel at least linearly increases with sh 2
, which M 2 j=1
means that D-MIMO can exploit the macro-diversity. Theorem Further, the distribution of the sum of log-normal variables can
6 also turns out that the variance of the MI shrinks as the be approximated by another log-normal distribution [40]. This
number of distributed antennas increases. In order to see this, can also be proved by using Theorem 1. Then, the lower bound
given M L, using (61), as N is very large, the variance of the MI can also be approximated by a Gaussian random
I2D - MIMO,LB goes to zero. However, for C-MIMO with M L, variable. There are many methods to obtain the approximation
as shown in Theorem 4 the variance of the MI becomes a of the CDF of the sum of the log-normal random variable, and
constant number, even for N L . Theorem 6 shows the the most famous method is Schwartz and Yehs method (SYs
composite fading channel approaching the AWGN channel as method) proposed in [40]. In the next section, we will use this
the number of distributed antennas approaches infinity. This method to approximate the CDF of the data throughput.
form of channel hardening [38] can reduce the deterioration
of shadow fading, and is generally beneficial for voice and D. CDF of the MI in the Low-SNR Regime
other traffic that is sensitive to channel fluctuations and delay. At low SNR, the MI in (11) can be lower bounded by [41,
Theorem 2]
C. Approximation of the CDF of the MI with M L for high  H !
ID-MIMO  log2 1 + Tr Hw ( IL ) Hw
SNR  M 
N  
We will next focus on the more accurate approximations of  D  H 
= log2 1 + sn Tr Hw,n Hw,n
the mean and variance of MI. M n=1 dn
With the definition of Wn in (55), ID-MIMO, LB can be  N   
rewritten as  D 2
 N  = log2 1 + sn Hw,n  . (38)
 M n=1 dn

ID-MIMO, LB = M log2 Wn + M log2 . As is very small, using ln (1 + x) x for small x, (38) can
n=1
M
   be written as
As we know log2 det Hw HwH is N  
 asymptotically
  Gaus-  D
sian for large M or N L [38], then det Hw HwH can be ID-MIMO sn Hw,n 2 . (39)
M ln 2 n=1 dn
approximated by a log-normal random variable. That is, Wn

 
N    
 D 1  H 
ID-MIMO (d)  M Es,Hw log2 1 + exp ln sn + ln det Hw,n Hw,n (32)
M n=1 dn M
 N    
 D 1   H 
 M log2 1 + exp E (ln sn ) + E ln det Hw,n Hw,n (33)
M n=1 dn M
  M
 N   
1   D
= M log2 1 + exp (L i + 1) (34)
M M i=1 n=1
dn

 M  
2 sh
2  1
ID - MIMO,LB (d) = M log2 + + log2 Lj+ +1
M 2 ln 2 j=1 M
M
 N   
  D
log2 (L j + 1) + log2 (36)
j=1 n=1
dn


"
L   N
2 2 Li+ 2
M + 1 (L i + 1) d2
n
i=1 n=1
I2D - MIMO,LB (d) = e sh   1  2 (37)
"
L   2 N
Li+ 1
M +1 d
n
i=1 n=1
2118 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO. 10, OCTOBER 2013

32
Ergodic capacity, simulation
50 Ergodic capacity, simulation 31
Asymptotic approximation
Lower bound /D=0.5
Asymptotic approximation 30
45

Ergodic capacity (bit/s/Hz)


Approximation(SYs method) 29
Ergodic capacity (bit/s/Hz)

40 28

27 /D=0.75
35 26

25
30
24
/D=1
25 23

22
20
21
35 40 45 50 55 60 65 70 75 80
0 0.2 0.4 0.6 0.8 1
/D 2sh (dB)

Fig. 2. Bounds, asymptotic results and simulation results for ergodic capacity Fig. 3. Ergodic capacity versus lognormal shadow fading variance for the
of the (2, 2, 7) D-MIMO channel. Polar angle of the user is = 0 . (2, 2, 7) D-MIMO channel. = 20 dB.

2
Note that Hw,n  22ML can be approximated by a
lognormal random variable. Then, the MI in (39) can also Empirical
Analytical Gaussian approx.(SY)
be approximated by a lognormal random variable. Similarly,
the CDF of the MI can be obtained by using SYs method.
( , )=( 0.5D, 0 )
0.1
E. Outage Capacity
In the high SNR regime, the Gaussian-approximated CDF
PDF

of the MI, I, can now be obtained as ( , )=( 0.05D, 0 )



1 y I (d) 0.05
Pr (I y|d) = 1 erfc # . (40)
2 2
2 (d) I

Equivalently, the approximated maximum rate at the outage


probability, (i.e. Pr(I < I |d) = ), can be given by 0
10 20 30 40 50 60 70
# mutual information (bit/s/Hz)
I (d) = I (d) + 2I2 (d)erfc1 (2 (1 )) . (41)
Fig. 4. Empirical distribution and Gaussian approximation to the PDF of the
capacity for various polar radiuses with = 0 . = 20 dB.
F. Numerical Results
For our numerical results, (2, 2, 7) and (5, 1, 5) D-MIMO
cellular systems are considered. For GN-Cell with 7 RAUs, Fig. 2 shows the ergodic capacity lower bound (30), asymp-
the polar coordinates of the RAUs are totic approximation (36), approximation with SYs method, as
well as Monte Carlo simulated curves for (2, 2, 7) D-MIMO
(D1 , 0) , (D2 , /6) , (D3 , /2) , (D4 , 5/6) , systems with = 20 dB. In the simulation, the polar angle of
(D5 , 7/6) , (D6 , 3/2) , (D7 , 11/6) , the user is fixed with = 0, and the polar radiuses vary form
0.02D to D. It is clearly seen that the lower bound is quite
where tight in the high SNR regime. For example, when /D < 0.15,
 
D1 = 0, D2 = = D7 = 3 3 D/2. the lower bound is extremely accurate. It is also shown that
the approximation with SYs method is very accurate, and the
For GN-Cell with 5 RAUs, the polar coordinates of the RAUs accuracy of (36) is also acceptable. Fig.3 shows that for high
are SNR, the ergodic capacity increases linearly with the variance
(D1 , 0) , (D2 , 0) , (D3 , /2) , (D4 , ) , (D5 , 3/2) . of the lognormal shadow fading, which coincides with the
analytical expression (36).
where Fig. 4 shows the numerical Gaussian approximation to the
 
D1 = 0, D2 = . . . = D5 = 3 3 D/2. MI PDF with mean and variance computed by Schwartz and
Yehs (SY) method, as well as the empirically generated PDF
Unless specified otherwise, path loss exponent = 3.7 and in the high SNR regime. As expected, the numerical estimation
log-normal shadow fading standard deviation sh = 8dB. of the mean and variance by using SYs method are almost the
WANG et al.: SPECTRAL EFFICIENCY OF DISTRIBUTED MIMO SYSTEMS 2119

2.5
45
Outage capacity, simulation Empirical
Approximation (SYs method) 2 Analytical lognormal approximatin
40
Outage capacity (bit/s/Hz)

35 1.5

PDF
30
1

25

0.5
20

0
15 0 1 2 3 4 5
0 0.2 0.4 0.6 0.8 1 mutual information (bit/s/Hz)
/D
(a) Empirical distribution and log-normal approximation to the PDF of the
capacity for a (2, 2, 7) D-MIMO channel at low SNR. Polar coordinate
Fig. 5. Approximation of the outage capacity of the (2, 2, 7) D-MIMO of the user is (, ) = (D, 0 ) and = 25dB.
channels. Polar angle of the user is = 0
1.5

Empirical
65 Analytical lognormal approximatin
Simulated ergodic capacity
60 Analytical ergodic capacity
Outage capacity (bit/s/Hz)

Simulated outage capacity


Analytical outage capacity 1
55
capacity (bit/s/Hz)

50
ergodic capacity
45

0.5
40

35
outage capacity
30

25 0
28 26 24 22 20 18 16 14 12
(dB)
20
0 0.2 0.4 0.6 0.8 1
/D (b) Outage capacity of a (2, 2, 7) D-MIMO channel at low SNR. Polar
coordinate of the user is (, ) = (D, 0 ).

Fig. 6. Approximation of the ergodic and outage capacity of the (5, 1, 5) Fig. 7. Statistics of the MI of the D-MIMO channels for low SNRs.
D-MIMO channels. Polar angle of the user is = 0

IV. S PECTRAL E FFICIENCY OVER THE C ELL


same. Moreover, the Gaussian approximation is quite accurate
for high SNR. A. CDF of the Spectral Efficiency Over the Cell
Fig. 5 illustrates the numerical approximation of the outage
In the previous section, the CDFs of the MI for the C-MIMO
capacity of (2, 2, 7) D-MIMO channel with = 20 dB
and D-MIMO channels at a specific user position are derived.
and outage probability = 0.05. Using SYs method, the
The next step is to average this conditional CDF over users
approximation is very accurate. Fig. 6 shows the ergodic
locations, and thereby obtain the CDF of the spectral efficiency
capacity and outage capacity of (5, 1, 5) D-MIMO channel
over the cell. As shown in Subsection II-B, the polar candidate
with = 20 dB. As expected, the analytical ergodic capacity
of the user and the distances d are given by equation (6), and
(22) and outage capacity (substituting (22) and (23) into (41))
then Pr (I  y|d) can be expressed as the function of and .
are very accurate in the high SNR regime.
Thus, using the user random location model, for the C-MIMO
Fig. 7(a) and Fig. 7(b) show the statistics of the MI of the
or D-MIMO system, the CDF of the spectral efficiency over
D-MIMO channel in the low SNR regime. Fig. 7(a) shows
the cell can be written as
the numerical log-normal approximation to the MI PDF with
 D  2
mean and variance computed by SYs method, as well as
the empirically generated PDF. As expected, the log-normal Pr (I y) = Pr (I  y|, ) 2
dd. (42)
0 0 D
approximation is quite accurate for low SNRs. In Fig. 7(b),
it is clearly seen that using log-normal approximation the For high SNRs, Pr (I y|, ) is the CDF of a Gaussian
analytical outage capacity is quite accurate. random variable.
2120 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO. 10, OCTOBER 2013

B. Mean Spectral Efficiency and Mean Outage Spectral Effi- By substituting (36) into (43), and using (51), the MSE can
ciency Over the Cell be approximated as (52), shown at the bottom of the next page.
The average of the ergodic capacity and outage capacity However, for M L, since the variance of MI is a
over all possible values of (, ), called mean spectral effi- complex function of d, it is difficult to obtain the closed-form
ciency (MSE) and mean outage spectral efficiency (MOSE) expression for MOSE of the D-MIMO system.
are defined as
Y  E, [I (, )] , (43)
C. Numerical and Simulation Results
 
Y  E, I (, ) . (44) The exact analytical values of Pr (I y) require numerical
integration. In the following, we use Simpson quadrature to
In the following, the closed-form expressions of (43) and
approximate them. For the C-MIMO and D-MIMO with M
(44) will be discussed with multiple scenarios.
N L, given and , the mean and variance of MI are computed
1) M N L: Using Theorem 3 and Theorem 4, the
according to Theorem 4 and Theorem 3, respectively. For
following theorems are deduced.
the D-MIMO system with M L, we use SYs method to
Theorem 7. For (M, L, N ) D-MIMO system, suppose that M approximate the mean and variance of MI.
is greater than or equal to N L. At high SNR, the MSE and Our basic Monte Carlo simulation approach is as follows.
MOSE are given by (45) and (46), shown at the bottom of the 1) The position of the desired user is randomly picked
next page. according to (7) and (8) as follows.
Proof: See Appendix C. a) Generate two pseudo-random, pseudo-independent
Theorem 8. For (M, N L, 1) C-MIMO system, suppose that numbers 1 and 2 uniformly distributed in [0, 1].
M is greater than or equal to N L. At high SNR, the MSE b) Deduce the polar coordinate (, ) as
and MOSE are given by (47) and (48), shown at the bottom
= D 1
of the next page.
= 22
Proof: For C-MIMO system, the base station is located
at the center of the cell. Then, the MSE of the cell-wide 2) Calculate the distance dn according to (6).
distribution is given by, 3) Generate the complex path gains to the base-station,
 D
following the channel model given in (3).
Y CMIMO = IC-MIMO ()p () d. (49)
0
4) Compute the MI according to (26). For each generated
polar coordinate (, ), we repeat procedures 3) and
Substituting (27) and (7) into (49) and integrating, (47) is
4) over 2000 sets of lognormal and Rayleigh fading.
obtained. The MOSE can be obtained as well, since the
We then get the the ergodic capacity I (d) and outage
variance of the MI is not related to .
capacity I (d) at the given position.
2) M L: For the D-MIMO system with M L, the
5) Do this experiment for 2000 random generated values of
analytical expression of MSE is hard to obtain. Thus, we resort
distance d, and find the CDF of the spectral efficiency.
to tighter bounds. Define
  N    6) Finally, Average I (d) and I (d) over user positions.

 D
d = Ed log2 . (50) In this subsection, (1, 1, 5), (5, 1, 5) and (2, 2, 7) D-MIMO
n=1
dn systems, and the corresponding counterparts (1, 5, 1), (5, 5, 1)
From (36), in order to get the MSE, we should compute d , and (2, 14, 1) C-MIMO systems are considered.
which can be lower bounded by the following proposition. First, we compute the integral (42) numerically and show
the results in Fig. 8(a) and Fig. 8(b), where the analytical
Proposition 1. The mean defined in (50) can be lower CDF and the simulated CDF are compared for the C-MIMO
bounded by  N 
and D-MIMO with different antenna layouts. The median of
 
Dn2 SNR at the cell extremity is = 20 dB. For the particular
d + log2 exp . (51) case displayed here and all others we examined at high SNR,
2 ln 2 n=1
2D2
n  agreement between analysis and simulation was seen.
 xi Fig. 9(a) and Fig. 9(b) compare the MSE of the C-MIMO
Proof: From [39], ln ai e is a convex function in
i=1 and D-MIMO systems. In the two figures, we see that the
xi for ai > 0. Then, using Jensens inequality, we can lower closed-form approximation (47) of MSE for C-MIMO are
bound d as follows   N  extremely accurate. In Fig. 9(a), we see that the closed-form

d = log2 D + Ed log2 exp ( ln dn ) approximations (52) of MSE for the D-MIMO system are also
n=1 satisfactory. Comparing the (2, 2, 7) and (2, 14, 1) systems we
 N
 see a significant increase in the MSE. When M N L, in

log2 D + log2 exp [E (ln dn )] Fig. 9(b), it is shown that the MSE of the C-MIMO system is
n=1 larger than that of the D-MIMO system.
N  
 Dn2 Fig. 10(a) and Fig. 10(b) compare the MOSE of the C-
= + log2 exp . MIMO and D-MIMO systems. The outage probability is =
2 ln 2 n=1
2D2
0.05. In Fig. 9(a) and Fig. 10(a), comparing the (2, 2, 7) and
(2, 14, 1) systems, it is interesting to observe that the increase
WANG et al.: SPECTRAL EFFICIENCY OF DISTRIBUTED MIMO SYSTEMS 2121

in the MOSE is more significant than the increase in the MSE. For the D-MIMO system, with random antenna layout, as
In Fig. 10(b), when M N L, we see that the analytical N , we have
expressions of MOSE are the same in the high SNR regime
N
for both C-MIMO and D-MIMO. In Fig. 10(b), comparing the 1  2 D2
lim Dn = .
(5, 1, 5) and (5, 5, 1) systems, we see a significant increase in N N 2
n=1
the MOSE due to the macro-diversity gain. Then, compared
with C-MIMO, D-MIMO has larger MOSE. Then,
N L
V. C OMPARISON OF THE C-MIMO AND D-MIMO Y DMIMO Y CMIMO = .
4 ln 2
SYSTEMS FOR LARGE N
Note that the expressions in (43) and (44) show the impact It is quite interesting to see that from Theorem 9 when
of various parameters on the spectral efficiency of the C- M N L C-MIMO has larger averaged spectral efficiency.
MIMO or D-MIMO systems. In a orthogonal multiple-access An intuitive explanation is that in this case at high SNR the
cellular system with a large number of user channels, the total multiplexing gain is N L, and because the averaged logarithm
information rate or outage information rate per cell will lie in of the access distance of D-MIMO is larger than that of C-
MIMO (For C-MIMO the base-station is always located at the
a narrow range about Y or Y . We, thus, regard Y and Y
as useful metrics for comparing the C-MIMO and D-MIMO center of the cell), then C-MIMO has larger averaged spectral
cellular systems. efficiency.
To provide more general results, in the following, the RAUs However, comparing (23) and (28), it can be seen that the
are assumed to be uniformly distributed in the GN-Cell, and variance of the MI of the D-MIMO channel is less than that
N is very large. With no intention to optimize the antenna of the C-MIMO channel. The reason is that the D-MIMO
layout, the MSE of a D-MIMO system with random antenna system benefits from the macro-diversity. In practice, the
layout roughly represents a lower bound [20]. outage probability is much less than 0.5, and in (41) we have
erfc1 (2 (1 )) < 0. Then, the D-MIMO system may have
A. M N L larger MOSE.
Theorem 9. Suppose that M is greater than or equal to Fig.11 shows the difference in the MOSE between the D-
N L. At high SNR, the MSE of (M, N L, 1) C-MIMO system MIMO and C-MIMO systems with various N and L. In the
is larger than that of (M, L, N ) D-MIMO system. numerical results, we set
Proof: In the high SNR region, when M N L, accord- N
1  2 1
ing to Theorem 7 and Theorem 8, we have 2
Dn = .
N D n=1 2
N
L  2
Y DMIMO Y CMIMO = D < 0. We can see that, in all cases except N = 2, the MOSE of
2D2 ln 2 n=1 n
the D-MIMO system is significantly larger than that of C-
This completes the proof. MIMO system. Note that for N = 2, if we let D1 = D2 =

N NL
N L L  2 1 
Y DMIMO = N L log2 + D + (M i + 1) (45)
M 2 ln 2 2D2 ln 2 n=1 n ln 2 i=1
$
% NL
2erfc (2 (1 )) %
1
&N L2 2 2 +

Y DMIMO = Y DMIMO + sh  (M i + 1) (46)
ln 2 i=1

NL
N L 1 
Y CMIMO = N L log2 + + (M i + 1) (47)
M 2 ln 2 ln 2 i=1
$
%
2erfc (2 (1 )) %
1 NL


Y CMIMO = Y CMIMO + & 2 2 2 2
N L sh +  (M i + 1) (48)
ln 2 i=1

 N  
+ 2 sh2  Dn2
Y DMIMO = M log2 + + log2 exp
M 2 ln 2 n=1
2D2
M   M

 1 
+ log2 L j + +1 log2 (L j + 1) (52)
j=1
M j=1
2122 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO. 10, OCTOBER 2013

1 40
Monte Carlo simulation

Capacity averaged over the user position (bit/s/Hz)


0.9 Analytical approx.(CMIMO, Eq.47) DMIMO
35 Analytical approx.(DMIMO, Eq.52) (2,2,7)
0.8
CMIMO CMIMO
(1,5,1) (2,14,1) 30
0.7

0.6 25
Pr(I y)

0.5 CMIMO (2,14,1)


DMIMO DMIMO
(1,1,5) (2,2,7) 20 DMIMO
0.4
(1,1,5)

0.3 15

0.2
10 CMIMO (1,5,1)
Numerical integration
0.1 Monte Carlo simulation

0 5
0 10 20 30 40 50 10 12 14 16 18 20 22 24
y (bit/s/Hz) (dB)

(a) CDF of the spectral efficiency over the cell ( = 20dB), M L (a) MSE, M L

Spectral efficiency averaged over the user position (bit/s/Hz)


65
1
CMIMO (5, 5, 1), Analytical approx.
CMIMO (5, 5, 1), Monte Carlo simulation
0.9 60
DMIMO (5, 1, 5), Analytical approx.
0.8 DMIMO (5, 1, 5), Monte Carlo simulation
55
0.7
50
0.6
Pr(I y)

45
0.5

0.4 40

0.3
35
CMIMO (5, 5, 1), numerical integration
0.2
CMIMO (5, 5, 1), Monte Carlo simulation
30
DMIMO (5, 1, 5), numerical integration
0.1
DMIMO (5, 1, 5), Monte Carlo simulation
25
0 14 16 18 20 22 24 26 28 30 32
0 20 40 60 80 100 120
(dB)
y (bit/s/Hz)

(b) MSE, M N L
(b) CDF of the spectral efficiency over the cell ( = 20dB), M N L
Fig. 9. Comparison of MSE.
Fig. 8. CDF of the spectral efficiency.

 
3 3 D/2, or equivalently,
N
1  2 5.0), the second term is negative, and it denotes the
D = 0.4019, spectral efficiency loss of D-MIMO due to the increasing of
N D2 n=1 n
the mean access distance. When the contribution of the first
term is larger than the second term, the MSE of D-MIMO
we also have Y DMIMO Y CMIMO > 0.
is larger than C-MIMO. Fortunately, in the real environment,
2
sh varies between 6 and 13, and sh 38 can be satisfied
B. M L in the most of cases.
Theorem 10. For (M, L, N ) D-MIMO system, suppose that
M is less than or equal to L, all RAUs are uniformly We should note that when M L even without macro-
distributed in the GN-Cell, and the closest distance D0 that diversity D-MIMO may also has larger averaged spectral
the mobile can be from the RAU is far less than the GN-Cell efficiency with carefully choosing the layout of the RAUs [26].
radius D. At high SNR, as N is very large, for practical An intuitive explanation is that in this case distance diversity
(2 5) and sh 2 2
(sh 38), the MSE of (M, L, N ) of D-MIMO is larger than that of C-MIMO, especially for
D-MIMO system is larger than that of (M, N L, 1) C-MIMO large .
system.
According to Theorem 6, for large N , the variance of MI
Proof: See Appendix D. becomes 0. However, for C-MIMO, the variance of MI is a
From (73), we see that, the first term of the right-hand side is constant over the whole cell. Then, for large N , both MSE
related to the macro-diversity. However, for practical (3.0 and MOSE of D-MIMO may larger than that of C-MIMO.
WANG et al.: SPECTRAL EFFICIENCY OF DISTRIBUTED MIMO SYSTEMS 2123

Outage spectral efficiency averaged over the user position (bit/s/Hz)


30 50

Difference of MOSE between DMIMO and CMIMO


Approximation (SYs method) L = 1, M = N
Monte Carlo simulation 45
DMIMO (2,2,7) L = 2, M = 2N
25
40 L = 3, M = 3N
L = 4, M = 4N
35
20
CMIMO (2,14,1) 30

25
15
20
DMIMO (1,1,5)
10 15

10
5 CMIMO (1,5,1)
5

0
0
10 12 14 16 18 20 22 24 2 3 4 5 6 7 8
(dB) N (number of RAU)

(a) MOSE, M L (a) Fixed L

55 30
Outage capacity averaged over the user position (bit/s/Hz)

CMIMO (5, 5, 1), Analytical approx.

Difference of MOSE between DMIMO and CMIMO


50 CMIMO (5, 5, 1), Monte Carlo simulation N = 2, M = 2L
DMIMO (5, 1, 5), Analytical approx. 25 N = 3, M = 3L
45 DMIMO (5, 1, 5), Monte Carlo simulation N = 4, M = 4L

40 20

35
15
30

25
10
20

15 5

10
0
5
16 18 20 22 24 26 28 30 32 34 1 2 3 4 5 6 7 8
(dB) L (number of receive antenna at RAU)

(b) MOSE, M N L (b) Fixed N

Fig. 10. Comparison of MOSE. Fig. 11. Difference of the MOSE of the D-MIMO and C-MIMO systems:
= 0.05, = 3.7, sh = 8dB. .
VI. C ONCLUSIONS
In this paper, we have presented an analytical framework A PPENDIX A
to investigate the spectral efficiency of the C-MIMO and D- P ROOFS OF T HEOREM 1
MIMO cellular systems in a composite Rayleigh-lognormal Define
channel. We evaluated the mean and variance of the MI

K
for the C-MIMO and D-MIMO channels conditioned on the k   K


user position. We prove that the MI can be approximated by k=1
V  QK ln k , (53)
Gaussian distribution at high SNR and log-normal distribution qK
k=1
at low SNR. The closed-form approximation of the mean and
and
K
1 
variance of the mutual information are presented at any SNR
and for different antenna configurations. We showed that when U (Xi k ). (54)
M N L the MSE of C-MIMO is larger than D-MIMO, and qK i=1
the result is opposite when M L and N is very large. For According to the CLT for the triangular arrays of independent
both of the cases, the MOSE of the D-MIMO cellular system d
random variables [31, Chapter 27], we have U N (0, 1).
is larger than that of the C-MIMO cellular system. Using (20) and (54), (53) can be rewritten as


K
ACKNOWLEDGMENT k
qK U
We thank the reviewers for the careful reviews and for V = k=1 ln 1 + K .
qK 
their suggestions which helped in improving the quality of k
the paper. k=1
2124 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO. 10, OCTOBER 2013

The characteristic function of V can be written as According to Theorem 1, asymptotic normality (as N is very
  large) of ID-MIMO, LB can be verified by confirming that, for
V (w) = E ewV some > 0, the Lyapounov condition

qw K
k
K k=1 
N
2+
E [Wn E (Wn )]
U n=1
=E 1 + .

lim . 2+ = 0 (59)
1

K
N

N
qK k E [Wn E (Wn )]
2
k=1 n=1

When K , and the following equation


.
qw 
K
k

N
2
K k=1 E [Wn E (Wn )]
n=1
U lim =0 (60)
lim V (w) = lim E 1 + N 
N
K K 
K
E (Wn )
1
qK k n=1
k=1
are satisfied. To show that the Lyapounov condition is satisfied
 
= E ewU . in the case, we choose = 2. Then, we have

N
Then V
d
N (0, 1). E [Wn E (Wn )]4
n=1
lim / 02
N 
N
2
A PPENDIX B E [Wn E (Wn )]
n=1
P ROOF OF T HEOREM 6

N
Define 4 d4
n
 H  E [G E (G)] n=1
Gn  sn det Hw,n Hw,n
1
M
, = 2 lim  2 .
E 2 [G E (G)] N 
N
d2
n
    n=1
D  H 1 D
Wn  sn det Hw,n Hw,n M = Gn . (55) Since dn is the distance between the n-th RAU and the user,
dn dn
it is satisfied
Obviously, Gn , n = 1, . . . , N is a sequence of i.i.d nonnega- D0  dn  2D.
tive random variables. Further, defining
    Then,
E Gk  E Gkn , k = 1, 2, . . . ,

N 
N
d4
n
1
N d4
n
we have n=1 1 n=1
lim  2 = Nlim  2 = 0
      H k N 
N N 
N
E Gk = E skn E det Hw,n Hw,n M d2
n
1
N d2
n
n=1 n=1
"
L  k

Li+ M +1 Therefore (59) is satisfied. Similarly,
1 2 2 2 i=1
=e 2 k sh , (56) .
"
L

N
(L i + 1) E [Wn E (Wn )]
2
i=1 n=1
lim
where we use the property of the lognormal random variable N 
N
E (Wn )
  1 2 2 2 n=1
E skn = e 2 k sh , k = 1, 2, . . . , (57) .
 H  # 
N
d2
n
and the the moment-generating function of ln det Hw Hw E [G E (G)]2 n=1
= lim =0 (61)
  -M E (G) N 
N
(L i + + 1) d
n
E e ln det( HwH Hw ) = , (58) n=1
i=1
(L i + 1)
(60) is also satisfied.
for 0. Then, according to Theorem 1, ID-MIMO, LB is can be
From (56), we can see that approximated by a Gaussian random variable as N is very
large, and the mean and variance are given by
E [G E (G)]2 < ,
 N   
 D
and E (ID-MIMO, LB ) = M log2 E (G) , (62)
4 M n=1 dn
E [G E (G)] < .
WANG et al.: SPECTRAL EFFICIENCY OF DISTRIBUTED MIMO SYSTEMS 2125


N
2 A PPENDIX D
E [Wn E (Wn )]
V (ID-MIMO, LB ) = n=1 P ROOF OF T HEOREM 10
 2

N
E (Wn ) When M L, the MI of C-MIMO channel can be rewritten
n=1 as
    N
   
d2
n D
E G2
n=1 ICMIMO = log2 det IM + sHwH Hw .
= 1  2 . (63) M d
E (G)
2
N

dn As N is very large, using the strong law of large numbers
n=1
(SLLN), we have HwH Hw N LIM . Then, following the
Applying (56) to (62) and (63) and simplifying, (36) and (37)
proof of Theorem 8, the MSE of C-MIMO system can be
can be obtained.
expressed as
A PPENDIX C N L M
P ROOF OF T HEOREM 7 Y CMIMO = M log2 + . (68)
M 2 ln 2
At high SNR, substituting (22) into (43), we have
For the D-MIMO system, since D0  dn  2D, then, given
NL
1  the users polar coordinate (, ), we have
Y DMIMO = N L log2 + (M i + 1)
M ln 2 i=1  
N
V d
n < .

+ N L log2 D L E (log2 dn ). (64) When N is very large, using (29) and the SLLN we see that
n=1
N    N   
Focusing on the last term in (64), using (6), we have  D 1 2 2
 D

  sn Hw,n Hw,n Le 2
H sh IM .
E, ln d2n n=1
dn n=1
dn
 D  2
  (69)
= 2
ln 2 2Dn cos + Dn2 dd
0 0 D
 D  2   Substituting (69) into (29), integrating over the user position
2Dn PDF, and using (51), we have
= ln 1 cos dd
0 0 D2 2 + Dn2    
 D L + 2 sh
2
2  2  Y DMIMO M log2 +
+ ln + Dn2 d M 2 ln 2
D

2
 N  
0
.  2 
 D Dn2
2 2Dn + M log2 exp . (70)
= ln 1 + 1 d ln 2 2D2
D2 2 + Dn2 n=1
0
 D Note that
2  2 
+ ln + Dn2 d N    D  
1 
D 2
 D
0 Dn2 x2 2x
2  2   lim exp = exp dx
= ln + Dn2 + 2 Dn2  d ln 2 N N
n=1
2D2 0 2D2 D2
2 
D 2
0
 Dn  D 1 e/2 .
2  2  2  2  =

(71)
= 2
ln 2D n d + 2
ln 2 d ln 2
0 D Dn D
Then, substituting (71) into (70), we have
D2
= 2 ln D 1 + n2 , (65)    
D N L + 2 sh
2
lim Y DMIMO M log2 +
where the third equation follows since [42, Sect.4.224] N M 2ln2
     
1 1 2
1 e/2 .
ln (1 + a cos x)dx = ln + 1a ,2 + M log2

(72)
0 2 2
for a2 1. Then we have Then, the difference between (72) and (68) can be given by
1 Dn2  
E, (log2 dn ) = log2 D + . (66) lim Y DMIMO Y CMIMO
2 ln 2 2D2 ln 2 N
  
Substituting (66) into (64), (45) is obtained. M 2 sh2
2 /2
Similarly, the outage capacity averaged over the user posi- + M log2 1e . (73)
2ln2
tion (, ) is given by
 D  2 For practical (3.0 5.0), if sh
2
38, we have

Y DMIMO = IDMIMO
(, ) 2
dd. (67)  
0 0 D lim Y DMIMO Y CMIMO > 0.
N
Since the variance of the MI is not a function of d, substituting
(22) and (23) into (41), (46) is obtained. This completes the proof.
2126 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO. 10, OCTOBER 2013

R EFERENCES [24] J. G. Andrews, F. Baccelli, and R. K. Ganti, A tractable approach to


coverage and rate in cellular networks, IEEE Trans. Commun., vol. 59,
[1] X. You, D. Wang, P. Zhu, and B. Sheng, Cell edge performance of no. 11, pp. 31223134, Nov. 2011.
cellular mobile systems, IEEE J. Sel. Areas Commun., vol. 29, no. 6, [25] S.-R. Lee, S.-H. Moon, J.-S. Kim, and I. Lee, Capacity analysis of
pp. 11391150, June 2011. distributed antenna systems in a composite fading channel, IEEE Trans.
[2] A. A. M. Saleh, A. J. Rustako, and R. S. Roman, Distributed antennas Wireless Commun., vol. 11, no. 3, pp. 10761086, Mar. 2012.
for indoor radio communications, IEEE Trans. Commun., vol. 35, pp. [26] X. Wang, P. Zhu, and M. Chen, Antenna location design for generalized
12451251, Dec. 1987. distributed antenna systems, IEEE Commun. Lett., vol. 13, no. 5, pp.
[3] M. V. Clark, T. M. W. III, L. J. Greenstein, A. J. Rustako, V. Erceg, 315317, May 2009.
and R. S. Roman, Distributed versus centralized antenna arrays in [27] H. Zhu and J. Wang, Chunk-based resource allocation in OFDMA
broadband wireless networks, in Proc. IEEE Veh. Technology Conf. systems - part I: chunk allocation, IEEE Trans. Commun., vol. 57,
(VTC01), May 2001, pp. 3337. no. 9, pp. 27342744, Sept. 2009.
[4] W. Roh and A. Paulraj, Outage performance of the distributed antenna [28] , Chunk-based resource allocation in OFDMA systems - part II:
systems in a composite fading channel, in Proc. IEEE Veh. Technology joint chunk, power and bit allocation, IEEE Trans. Commun., vol. 60,
Conf. (VTC02), vol. 3, Sept. 2002, pp. 15201524. no. 2, pp. 499509, Feb. 2012.
[5] 3GPP-LTE, Technical specification group radio access network: [29] S. Catreux, P. F. Driessen, and L. J. Greenstein, Data throughputs using
Evolved universal terrestrial radio access (E-UTRA): further advance- multiple-input multiple-output (MIMO) techniques in a noise-limited
ments for E-UTRA physical layer aspects (release 9), 3GPP TR36.814, cellular environment, IEEE Trans. Wireless Commun., vol. 1, no. 2,
Mar. 2010. pp. 226235, Apr. 2002.
[30] D. P. Palomar, J. M. Cioffi, and M. A. Lagunas, Uniform power
[6] F. Adachi, K. Takeda, T. Yamamoto, and R. Matsukawa, Gigabit
allocation in MIMO channels: A game-theoretic approach, IEEE Trans.
distributed antenna network and its related wireless techniques, in
Inf. Theory, vol. 49, no. 10, pp. 25452561, Oct. 2003.
Proc. International Wireless Communications and Mobile Computing
[31] P. Billingsley, Probability and measure, 2nd ed. New York: John Willey
Conference (IWCMC), 2011, pp. 15501556.
& Sons, 1985.
[7] W. Peng and F. Adachi, Multi-user downlink transmit beamforming
[32] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed.
for the broadband single-carrier distributed antenna network, in Proc.
New York: John Willey & Sons, 1991.
IEEE Veh. Technology Conf. (VTC 2011 Fall), Sept. 2011, pp. 15.
[33] M. Chiani, M. Z. Win, and H. Shin, MIMO networks: The effects of
[8] , Study on the capacity of distributed antenna network system by interference, IEEE Trans. Inf. Theory, vol. 56, no. 1, pp. 336349, Jan.
using single carrier frequency domain adaptive antenna array, in Proc. 2010.
International Conference on Communications and Networking in China [34] M. Chiani, M. Z. Win, and A. Zanella, On the capacity of spatially
(CHINACOM), 2011, pp. 655658. correlated MIMO Rayleigh fading channels, IEEE Trans. Inf. Theory,
[9] A. D. Wyner, Shannon-theoretical approach to a Gaussian cellular vol. 49, no. 10, pp. 23632371, Oct. 2003.
multiple-access channel, IEEE Trans. Inf. Theory, vol. 40, pp. 1713 [35] P. J. Smith, S. Roy, and M. Shafi, Capacity of MIMO systems with
1727, Nov. 1994. semicorrelated flat fading, IEEE Trans. Inf. Theory, vol. 49, no. 10, pp.
[10] O. Somekh, B. M. Zaidel, and S. Shamai(Shitz), Sum rate character- 27812788, Oct. 2003.
ization of joint multiple cell-site processing, IEEE Trans. Inf. Theory, [36] A. Grant, Rayleigh fading multi-antenna channels, EURASIP J. Ap-
vol. 53, no. 12, pp. 44734497, Dec. 2007. plied Signal Processsing (Special Issue on Space-Time Coding (Part I)),
[11] Y. Liang, T. Yoo, and A. Goldsmith, Coverage spectral efficiency of pp. 316329, Mar. 2002.
cellular systems with cooperative base stations, in Proc. IEEE Global [37] A. M. Tulino and S. Verdu, Random matrix theory and wireless
Communications Conference (Globecom06), Dec. 2006. communications, Foundations and Trends in Communications and
[12] W. Roh and A. Paulraj, MIMO channel capacity for the distributed Information Theory, vol. 1, no. 1, 2004.
antenna systems, in IEEE Veh. Technology Conf. (VTC02), vol. 3, Sept. [38] B. M. Hochwald, T. L. Marzetta, and V. Tarokh, Multi-antenna channel
2002, pp. 15201524. hardening and its implications for rate feedback and scheduling, IEEE
[13] W. Choi and J. G. Andrews, Downlink performance and capacity of Trans. Inf. Theory, vol. 50, pp. 18931909, Sept. 2004.
distributed antenna systems in a multicell environment, IEEE Trans. [39] Q. T. Zhang, X. W. Cui, and X. M. Li, Very tight capacity bounds
Wireless Commun., vol. 6, pp. 6973, Jan 2007. for MIMO correlated Rayleigh fading channels, IEEE Trans. Wireless
[14] W. Peng and F. Adachi, A framework for uplink capacity evaluation Commun., vol. 4, no. 2, pp. 681688, Mar. 2005.
of cellular MIMO system, in Proc. IEEE International Conference on [40] S. Schwartz and Y. Yeh, On the distribution function and moments
Network Infrastructure and Digital Content (IC-NIDC), Nov. 2009, pp. of power sums with lognormal components, Bell System Technique J.,
875880. vol. 61, no. 7, pp. 14411462, Sept. 1982.
[15] D. Wang and X. You, Data throughputs of noise-limited MIMO cellular [41] O. Oyman, R. U. Nabar, H. Bolcskei, and A. J. Paulraj, Characterizing
systems, IEEE Trans. Signal Process., vol. 56, no. 5, pp. 21652170, the statistical properties of mutual information in MIMO channels,
May 2008. IEEE Trans. Signal Process., vol. 51, no. 11, pp. 27842795, Nov. 2003.
[16] D. Aktas, M. N. Bacha, J. Evans, and S. Hanly, Scaling results on the [42] I. S. Gradshteyn and I. M. Ryshik, Table of integrals, Series, and
sum capacity of cellular networks with MIMO links, IEEE Trans. Inf. Products, 6th ed. New York: Academic, 2000.
Theory, vol. 52, no. 7, pp. 32643274, July 2006.
[17] H. Dai and H. V. Poor, Asymptotic spectral efficiency of multicell
MIMO systems with frequency-flat fading, IEEE Trans. Signal Pro-
cess., vol. 51, no. 11, pp. 29762988, Nov. 2003.
[18] H. Zhuang, L. Dai, L. Xiao, and Y. Yao, Spectral efficiency of Dongming Wang received the B.S. degree from
distributed antenna system with random antenna layout, IEE Electron. Chongqing University of Posts and Telecommunica-
Lett., vol. 39, no. 6, pp. 495496, Mar. 2003. tions in 1999, the M.S. degree from Nanjing Univer-
[19] H. Zhu, Performance comparison between distributed antenna and sity of Posts and Telecommunications in 2002, and
microcellular systems, IEEE J. Sel. Areas Commun., vol. 29, no. 6, the Ph. D. degree from Southeast University in 2006.
pp. 11511163, June 2011. He joined the National Mobile Communications
[20] L. Dai, A comparative study on uplink sum capacity with co-located Research Laboratory at Southeast University, China,
and distributed antennas, IEEE J. Sel. Areas Commun., vol. 29, no. 6, in 2006, where he has been an Associate Professor
pp. 12001213, June 2011. since 2010. He is currently also a visiting scholar at
[21] D. Wang, X. You, J. Wang, Y. Wang, and X. Hou, Spectral efficiency University of California, Davis, CA. He serves as an
of distributed MIMO cellular systems in a composite fading channel, associate editor for the SCIENCE CHINA Informa-
in Proc. IEEE International Conference on Communications (ICC08), tion Sciences. His current research interests include turbo detection, channel
Beijing, China, 2008, pp. 12591264. estimation, distributed antenna systems, and large-scale MIMO systems.
[22] F. Heliot, X. Chu, R. Hoshyar, and R. Tafazolli, An accurate closed-
form approximation of the distributed MIMO outage probability, IEEE
Trans. Wireless Commun., vol. 10, no. 1, pp. 511, Jan. 2011.
[23] M.-S. Alouini and A. J. Goldsmith, Area spectral efficiency of cellular
mobile radio systems, IEEE Trans. Veh. Technol., vol. 48, no. 4, pp.
10471066, Jul. 1999.
WANG et al.: SPECTRAL EFFICIENCY OF DISTRIBUTED MIMO SYSTEMS 2127

Jiangzhou Wang (M91, SM94) received the B.S. Yan Wang received the B.S. degree from Xidian
and M.S. degrees from Xidian University, Xian, University, Xian, China in 1990, M.S. and PhD
China, in 1983 and 1985, respectively, and the degree from the University of Science & Technology
Ph.D. degree (with Greatest Distinction) from the of China (USTC), Hefei, China in 1999 and 2003
University of Ghent, Belgium, in 1990, all in com- respectively, both in electrical engineering. Since
munication engineering. 2004, he has been with the National Mobile Com-
Dr. Wang is currently Chair of Telecommunica- munications Research Laboratory, Southeast Univer-
tions and Head of Broadband and Wireless Commu- sity and now is a professor. From 1990 to 2001, he
nications Research Group at the University of Kent, was a System Engineer at the East China Research
England, United Kingdom. From 1995 to 2005, Institute of Electronic Engineering, Hefei, China,
he was with the University of Hong Kong. From where he was a principal investigator of two digital
1992 to 1995, he was a Senior System Engineer at Rockwell International array systems. His research interests span in statistical/array signal processing,
Corporation, California, USA. From 1990 to 1992, he was a Postdoctoral MIMO, wireless location and array system design. Prof. Wang has published
Fellow in the University of California at San Diego, USA. Dr. Wang has more than 40 journal and conference papers. He was the outstanding individ-
published over 200 papers in international journals and conferences in the ual of the 15th Anniversary of the National 863 Hi-Tech Project of China, in
areas of wireless mobile communications. He has written/edited three books, February 2001 and received two other national awards.
two of which have been translated into Chinese.
Dr. Wang is an Editor for IEEE Transactions on Communications since
1998, a six-time Guest Editor for IEEE Journal on Selected Areas in
Communications (Wideband CDMA, 8/2000 and 1/2001; Advances in Multi-
carrier CDMA, 6/2006; Wireless Video Transmission, 3/2010; and Distributed
Broadband Wireless Communications, 6/2011; Device-to-device Communi- Ming Chen received his B.Sc., M.Sc., and Ph.D.
cations in Cellular Networks, 6/2014), a two-time Guest Editor for IEEE degrees from the Mathematics Department of Nan-
Wireless Communications Magazine (Coordinated and Distributed MIMO, jing University in 1990, 1993, and 1996, respec-
6/2010; Recent Advances in Wireless Technologies for Smart Grid, 6/2012), tively. In July 1996, he joined the National Mobile
Guest Editor for IEEE Communications Magazine (Advances on Cooperative Communications Research Laboratory at Southeast
Wireless Networking, 5/2011 and 4/2012). He was IEEE Globecom2011 University, where he has been a professor since
Workshop Co-chair and IEEE VTC2011-Spring Track Co-chair. He is the 2003. His research interests include digital commu-
Technical Program Chairman of IEEE WCNC2013 in Shanghai and the nication theory and signal processing for wireless
Executive Chairman of IEEE ICC2015 in London. He is a Fellow of IET communication systems with emphasis on multi-
and IEEE Distinguished Lecturer. user detection, adaptive antenna, space-time coding,
and radio resource management.

Xiaohu You received the B.S., M.S. and Ph.D. de-


grees in electrical engineering from Nanjing Institute
of Technology, Nanjing, China, in 1982, 1985, and
1989, respectively. From 1987 to 1989, he was with
Nanjing Institute of Technology as a Lecturer. From
1990 to the present time, he has been with South-
east University, first as an Associate Professor and Xiaoyun Hou received the B.S. degree and the
later as a Professor. His research interests include M.S. degree from Nanjing University of Posts and
mobile communications, adaptive signal processing, Telecommunications in 1999 and 2002, and the
and artificial neural networks with applications to Ph.D. degree from Shanghai Jiaotong University
communications and biomedical engineering. He is in 2005. She joined Nanjing University of Posts
the Chief of the Technical Group of China 3G/B3G Mobile Communication and Telecommunications, China, in 2005, where she
R & D Project. Dr. You received the excellent paper prize from the China has been an Associate Professor since 2009. She
Institute of Communications in 1987 and the Elite Outstanding Young Teacher is currently also a visiting scholar at University of
Awards from Southeast University in 1990, 1991, and 1993. He was also a California, Davis, CA. Her current research inter-
recipient of the 1989 Young Teacher Award of Fok Ying Tung Education ests include channel estimation and physical layer
Foundation, State Education Commission of China. security.

Вам также может понравиться