3226 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 11, NO.
9, SEPTEMBER 2012
Achieving Massive MIMO Spectral Efficiency
with a Not-so-Large Number of Antennas
Hoon Huh, Member, IEEE, Giuseppe Caire, Fellow, IEEE, Haralabos C. Papadopoulos, Member, IEEE,
and Sean A. Ramprashad, Senior Member, IEEE
AbstractTime-Division Duplexing (TDD) allows to estimate
the downlink channels for an arbitrarily large number of base
station antennas from a finite number of orthogonal uplink
pilot signals, by exploiting channel reciprocity. Based on this
observation, a recently proposed Massive MIMO scheme was
shown to achieve unprecedented spectral efficiency in realistic
conditions of distance-dependent pathloss and channel coherence
time and bandwidth.
The main focus and contribution of this paper is an improved
Network-MIMO TDD architecture achieving spectral efficiencies
comparable with Massive MIMO, with one order of magnitude
fewer antennas per active user per cell (roughly, from 500 to
50 antennas). The proposed architecture is based on a family of
Network-MIMO schemes defined by small clusters of cooperating
base stations, zero-forcing multiuser MIMO precoding with
suitable inter-cluster interference mitigation constraints, uplink
pilot signals allocation and frequency reuse across cells. The key
idea consists of partitioning the users into equivalence classes,
optimizing the Network-MIMO scheme for each equivalence
class, and letting a scheduler allocate the channel time-frequency
dimensions to the different classes in order to maximize a suitable
network utility function that captures a desired notion of fairness.
This results in a mixed-mode Network-MIMO architecture,
where different schemes, each of which is optimized for the served
user equivalence class, are multiplexed in time-frequency.
In order to carry out the performance analysis and the
optimization of the proposed architecture in a systematic and
computationally efficient way, we consider the large-system
regime where the number of users, the number of antennas,
and the channel coherence block length go to infinity with fixed
ratios.
Index TermsChannel training, downlink scheduling, fre-
quency reuse, inter-cell cooperation, large-system analysis, linear
precoding, Massive MIMO, pilot contamination, time-division
duplex.
I. I NTRODUCTION
M ULTIUSER MIMO (MU-MIMO) technology is the
subject of extensive theoretical and practical investi-
gation for the next generation wireless cellular systems (e.g.,
LTE-Advanced [1]). Schemes where antennas of different
Base Stations (BSs) are jointly processed by centralized BS
Manuscript received July 22, 2011; revised March 8, 2012; accepted May 8,
2012. The associate editor coordinating the review of this paper and approving
it for publication was R. Nabar.
H. Huh is with the Media Lab., Samsung Electronics Co., Ltd., Suwon,
Gyeonggi-do 443-742, Korea (e-mail: hhuh@samsung.com).
G. Caire is with the Department of Electrical Engineering, University of
Southern California, Los Angeles, CA 90089, USA (e-mail: caire@usc.edu).
H. C. Papadopoulos and S. A. Ramprashad are with Docomo In-
novations, Palo Alto, CA 94304, USA (e-mail: {hpapadopoulos, ram-
prashad}@docomoinnovations.com).
Digital Object Identifier 10.1109/TWC.2012.070912.111383
1536-1276/12$31.00 
controllers are referred to as Network-MIMO architectures
(e.g., [2][8]). It has been recognized that the improvement
obtained from transmit antenna joint processing is limited by
a dimensionality bottleneck [9][11]. In particular, the high-
SNR capacity of a single-user MIMO system with Nt transmit
antennas, Nr receiving antennas, and fading coherence block
length T complex dimensions,1 scales as C(SNR) = M (1
M /T ) log SNR+O(1) where M = min{Nt , Nr , T /2} [13].
Therefore, even by pooling all BSs into a single distributed
macro-transmitter with Nt  1 antennas and all user terminals
into a single distributed macro-receiver with Nr  1 antennas,
the system degrees of freedom 2 are limited by the fading
coherence block length T . The same dimensionality bottleneck
also arises in the high-SNR behavior of MU-MIMO systems
based on explicit training for channel estimation, and can be
interpreted as the effect of the overhead incurred by pilot
signals [14], [15].
For frequency-division duplex (FDD) systems, the training
overhead required to collect channel state information at
the transmitters (CSIT) grows linearly with the number of
cooperating transmit antennas. Such overhead significantly
limits the performance improvement that can be expected by
increasing the number of jointly processed BS antennas in
FDD, as shown in [10], [11].
For Time Division Duplexing (TDD) systems, exploiting
channel reciprocity [16], [17], the CSIT can be obtained from
the uplink pilot signals. In this case, the training overhead
scales linearly with the number of active users3 per cell, but
it is independent of the number of cooperating antennas at
the BSs. As a result, for a fixed number of users scheduled
for transmission, the TDD system performance can be signif-
icantly improved by increasing the number of BS antennas.
Following this idea, Marzetta [17] showed that simple
Linear Single-User BeamForming (LSUBF) and random user
scheduling, without any BS joint processing, yields very high
spectral efficiency in TDD cellular systems, provided that
a sufficiently large number of transmit antennas per active
1 The fading coherence block length T , measured in signal complex
dimensions in the time-frequency domain is proportional to the product
Wc Tc , where Tc (s) denotes the channel coherence interval, and Wc (Hz)
denotes the channel coherence bandwidth [12].
2 The system Degrees of Freedom (DoFs) are defined as the pre-log factor
C(SNR)
of the system capacity C(SNR), i.e., DoFs = limSNR log SNR , and
quantify the number of equivalent parallel single-user Gaussian channels,
in a first-order approximation with respect to log SNR.
3 By active users we denote the users simultaneously served on a given
time-frequency slot in a cell. In practice, there might be many more users,
that are not scheduled on the current slot and therefore are not active.
c 2012 IEEE
HUH et al.: ACHIEVING MASSIVE MIMO SPECTRAL EFFICIENCY WITH A NOT-SO-LARGE NUMBER OF ANTENNAS
user are employed at each BS. This scheme, referred to
hereafter as Massive MIMO, was analyzed in the limit of
infinite number of BS antennas per user per cell. In this
regime, the effects of Gaussian noise and uncorrelated intra-
and inter-cell interference disappear, and the only remaining
impairment is the correlated inter-cell interference due to pilot
contamination [18], i.e., the correlated interference due to re-
using the same pilot signals in other cells (see Section III for
more details).
In this work, we also focus on TDD systems and exploit
reciprocity. The main contribution of this paper is a novel
Network-MIMO architecture that achieves spectral efficiencies
comparable with Massive MIMO, with a much smaller
number of BS antennas per active user (one order of mag-
nitude fewer antennas for approximately the same spectral
efficiency). As in [17], we also analyze the performance of the
proposed system in the limit of a large number of antennas.
However, a different system scaling is considered, for which
the number of antennas per active user per cell is finite. 4
The analysis is obtained by letting the number of users
per cell, the number of antennas per BS, and the channel
coherence block length go to infinity, with fixed ratios [11].
In this regime we find that the LSUBF scheme advocated in
[17] performs very poorly. In contrast, we consider a family of
Network-MIMO schemes based on small clusters of cooperat-
ing base stations, Linear Zero-Forcing BeamForming (LZFBF)
with suitable inter-cluster interference constraints, uplink pilot
signals allocation and frequency reuse across cells. The key
idea consists of partitioning the users into geographically
determined equivalence classes, referred to in the following
as bins. Users in the same bin share the same set of
distances from all BSs, and therefore they have identical fading
channel statistics (determined by the distance-dependent path
loss coefficients). Users with identical fading channel statistics
are referred to, in the following, as statistically equivalent.
The system serves users in the same bin simultaneously,
on the same time-frequency slot, using a Network-MIMO
scheme in the family optimized for the specific bin. Different
bins are scheduled over the time-frequency slots in order to
maximize an appropriately chosen network utility function
reflecting some desired notion of fairness. The proposed
scheme yields very simple system operation, where each time
a given bin is scheduled, the subset of users in the selected bin
is chosen at random or in a deterministic round robin fashion,
without performing any CSIT-based user selection. This allows
a fast turn-around between feedback and transmission as in
[17], which can take place in the same channel coherence
block. The resulting system is a mixed-mode Network-MIMO
architecture, where different Network-MIMO schemes, each
of which is optimized for the specific user bin, are multiplexed
on the time-frequency slots by the scheduler.
Using results and tools from the large-system analysis
developed in [11], [20] and adapted to the present scenario,
we obtain the asymptotic achievable rate for each scheme
in closed form. The performance predicted by the large-
system analysis matches very well with finite-dimensional
4 Analogous results and conclusions, in striking agreement with ours, have
been developed independently and at the same time in [19].
3227
simulations, in agreement with [11], [21] and several well-
known works on single-user MIMO in the large antenna
regime [22], [23]. The large-system analysis developed here is
instrumental to the systematic design and optimization of the
proposed system architecture, since it allows an accurate and
rapid selection of the best Network-MIMO scheme for each
user bin without resorting to cumbersome and time-consuming
Monte Carlo simulation.
It is worthwhile to remark that the ideas of dynamic
clustering of cooperating BSs, and mixed-mode MU-MIMO
downlink have appeared in a large number of previous works
(see for example [24][28]). Giving a fair account of this
vast literature would be impossible within the space limits
of this paper. Nevertheless, we wish to stress here that the
novel contribution of this paper is a systematic approach to
multi-modal system optimization based on simple closed-form
expressions of the spectral efficiency of each Network-MIMO
scheme in the family, and on scheduling across the schemes
(or modes) in order to maximize a desired network utility
function.
The remainder of this paper is organized as follows. In
Section II, we describe the family of proposed Network-
MIMO schemes. Section III presents the uplink training,
MMSE channel estimation and pilot contamination effect for
TDD-based systems. In Section IV, we provide expressions
for the spectral efficiency achievable with the Network-MIMO
schemes under considerations in the large-system limit. Sec-
tion V discusses the downlink scheduling problem under spe-
cific fairness criteria. Numerical results including comparison
with finite dimensional simulation results are provided in
Section VI and concluding remarks are given in Section VII.
II. S YSTEM M ODEL
The TDD cellular architecture for high-data rate downlink
proposed in this work is based on the following elements:
1) A family of Network-MIMO schemes. Each scheme of
the family is defined by the size and shape of clusters of
cooperating BSs, pilot reuse across clusters, frequency
reuse factor, and downlink linear precoding scheme;
2) A partition of the user population into bins of statisti-
cally equivalent users, according to their position in the
cellular coverage area;
3) The determination of the optimal Network-MIMO
scheme in the family for each user bin, creating an asso-
ciation between user bins and Network-MIMO schemes;
4) Scheduling of the user bins in time-frequency in order to
maximize a suitable concave and componentwise non-
decreasing network utility function of the ergodic user
rates, reflecting a desired notion of fairness [21], [29]
[32].
The system operation can be readily described. Each bin is
assigned a fraction of the time-frequency plane resources by
the scheduler. When a bin is scheduled, a subset of users is
chosen (e.g., in a round robin fashion) and is served by the
(bin-optimized) Network-MIMO architecture.
Invoking well-known convergence results [22], [23], we
use the large-system analysis approach for multi-antenna
cellular systems pioneered in [21], [33][35]. In particular,
3228 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 11, NO. 9, SEPTEMBER 2012

we rely on results from [11] and analyze the performance

of the Network-MIMO schemes in the considered family
when scaling the number of users in each bin, the number of
antennas per BS and the small-scale fading coherence block
length to infinity, with fixed ratios. We define a system size
parameter, indicated by N , and let all the above quantities
increase linearly with N . Specifically, we let the product
M N denote the number of BS antennas, LN denote the
channel coherence block length, and U N denote the number
of users per location (a bin is defined as a set of discrete
locations in the cellular coverage, see Section II-A), for given
constants M, L and U .
Fig. 1. Two dimensional hexagonal cell layout with B = 19. The triangle
marks indicate the BS positions (points of bs ). The large hexagons indicate
A. Cellular layout and frequency reuse the Voronoi regions of , and the solid-line large hexagon indicate V. The
small hexagons inside V indicate the cells Vb for b bs V. We have
Base stations, cells and clusters: The system geometry is B = 19 cell centers inside V.
concisely described by using a regular grid of points (lattices
and lattice translates) on the real line R (for 1-dimensional
2
layouts [8], [10]) or on the real plane R (for 2-dimensional
layouts [17]). Consider nested lattices bs u in
R (resp., R2 ). The system coverage region is given by the
Voronoi cell V of centered at the origin.5 BSs are located
at points b bs V. The finer lattice u defines a grid of
discrete user locations, as explained later in this section. We
let B = |bs V| denote the number of BSs in the system
coverage region.
Example 1: Consider the 1-dimensional layout defined by
= B Z and bs = Z. The coverage region is V =
[B/2, B/2) and the BS locations b are given by all integer-
coordinate points in the interval [B/2, B/2).
Example 2: In system studies reported in the standardiza-
tion of 4th generation cellular systems [36], [37] it is custom-
ary to consider a 2-dimensional hexagonal layout formed by Fig. 2. Cluster pattern geometry and user bins in one-dimensional and two
2 dimensional layouts.
19 cells, as shown in Fig. 1. In this case, bs = AZ and
2
= ABZ , with
   
3r 3 0 4 3 Consider the set of BS locations C = {b0 , . . . , bC1 }, with
A= and B = , bj bs V and b0 = 0. We define the clustering pattern
2 1 2 3 4
u(C) as the collection of BS location sets (referred to as
where r denotes the distance between the center of a small clusters in the following)
hexagon and one of its vertices. We have B = det(B) = 19,
and the distance between the closest two points in bs is 3r. u(C) = {{C + c} : c bs V} . (2)
We focus on systems based on single-cell processing (C = 1),
In order to avoid border effects at the edges of the or with joint processing over clusters of small size: C = 2 in
coverage region, all distances and all spatial coordinates are the 1-dimensional case, and C = 3 in the 2-dimensional case,
2
defined modulo : for x R (resp., x R ), let Q (x) as shown in Fig. 2.
denote the lattice quantization of x, i.e., the point of closest Example 3: In the 1-dimensional case of Example 1, with
to x. Then, C = 1 and C = 2, we have (recall that coordinates are defined
[x] mod = x Q (x) modulo , i.e., in this case, modulo B Z)
and the modulo distance between two points x, x in R u({0}) = {{0}, {1}, . . . , {B 1}} .
2
(resp., R ) is given by
and
d (x, x ) = |[x x ] mod |. (1)
u({0, 1}) = {{0, 1}, {1, 2}, . . . , {B 1, 0}} ,
Cell Vb is defined as the Voronoi region of BS b bs V
with respect to the modulo- distance. The collection of cells respectively.
{Vb } forms a partition of V into congruent regions. User location bins: We assume a uniform user spatial
distribution over the coverage region. For the sake of analytical
5 The Voronoi cell of a lattice point x R n simplicity, we discretize the user distribution into a regular
is the set of points
n
y R closer to x than to any other lattice point. grid of user locations with U N users each, corresponding to
HUH et al.: ACHIEVING MASSIVE MIMO SPECTRAL EFFICIENCY WITH A NOT-SO-LARGE NUMBER OF ANTENNAS 3229

 u = u + u0 , where u0 = 0
the points of the lattice translate

is chosen such that u is symmetric with respect to the origin
and no points of  u fall on the cell boundaries. 6 Notice that
for a sufficiently fine quantization grid, any compactly sup-
ported user distribution can be approximated by its discretized
version, and the symmetric construction does not involve any
loss of generality.
Example 4: In the 1-dimensional case of Example 1, let K
1 1
denote an even integer and let u = K Z, and u0 = 2K . Then,

the points of u Vb are symmetrically located with respect Fig. 3. 1-dimensional layout with C = 2 and F = 2.
to each BS of coordinate b = 0, 1, . . . , B 1.
Consider the set of user locations X = {x0 , x1 , . . . , xm1 }
with xi  u . We define the user bin v(X ) as the collection that mU CM . Therefore, the downlink DoFs are always
of user location sets (indicated by groups in the following) limited by the number of BS antennas.7 The number of active
users effectively scheduled on each time-frequency slot is
v(X ) = {{X + c} : c bs V} . (3) denoted by SN , where the coefficient S [0, CM ] is referred
to as the loading factor. In general, S can be optimized
In particular, we choose X to be a symmetric set of points
depending on X and C and on the type of network MIMO
with respect to the positions of the BSs comprising cluster
schemes employed (see Section IV). We restrict to consider
C. The reason for this symmetry is two-fold: on one hand,
schemes that serve an equal number SN/m of active users
a symmetric set generalizes the single location case and yet
per location. By construction, the users in the same bin are
provides a set of statistically equivalent users (same set of
statistically equivalent. Therefore, without loss of generality,
distances from all BSs in the cluster), thus providing a richer
we can assume that a round-robin scheduling picks all subsets
system optimization parameter space. On the other hand,
of size SN out of the whole mU N users in each group with
symmetry yields very simple closed-form expressions in the
the same fraction of time. In this way, the aggregate spectral
large-system analysis, by means of [11, Th. 3]. It should be
efficiency of the group (indicated in the following a group
noticed here that the bin geometry is a design choice, i.e., we
spectral efficiency) is shared evenly among all the users in
do not assume that the bins are symmetric, but we choose
the group.
(by system design choice) to serve users in symmetric sets of
Frequency reuse: The frequency reuse factor of the scheme
locations.
is denoted by F , and can also be optimized for each {X , C}.
Example 5: In the 1-dimensional case of Example 1, we
The system bandwidth is partitioned into F subbands of equal
are interested in the cases X = {x, x} and X = {x, 1 x},
 u [0, 1/2], as shown in Fig. 2. This yields width. For F = 1, all clusters in u(C) transmit on the whole
for some x
system bandwidth. For F > 1, clusters are assigned different
the bins
subbands according to a regular reuse pattern [8], [10]. For the
v({x, x}) = 1-dimensional layout, any integer F dividing B is possible.
{{x, x}, {1 x, 1 + x}, . . . , {B 1 x, B 1 + x}} For the 2-dimensional layout, we consider reuse factors given
by F = i2 + ij + j 2 for non-negative integer i and j [38]. For
and later use, we define D(f ) as the set of clusters transmitting
v({x, 1 x}) = on subband f {0, . . . , F 1}.
Example 6: Fig. 3 shows a 1-dimensional system with
{{x, 1 x}, {1 + x, 2 x}, . . . , {B 1 + x, B x}} ,
frequency reuse F = 2 for the clustering pattern of size
respectively. C = 2 defined by C = {0, 1} and the user bin defined by
Cluster/group association and user group rate: The X = {x, 1x}. Even-numbered clusters operate on subband 0
BSs forming a cluster are jointly coordinated by a cluster and odd-numbered clusters operate on subband 1. An example
controller that collects all relevant channel state information for the 2-dimensional hexagonal layout with F = 3 and C = 1
and computes the beamforming coefficients for the desired is shown in Fig. 1, where cells with the same color operate
MU-MIMO precoding scheme. For given sets {X , C}, the on the same subband.
users in group X + c are served by the cluster C + c, for all
c bs V (see Fig. 2). By construction, each BS belongs to B. Channel statistics and received signal model
C clusters and transmits the superposition of the signals from
all the corresponding cluster controllers. These signals may The average received signal power for a user located at
share the same frequency band, or be defined on orthogonal x V from a BS antenna located at b V is denoted by
subbands, depending on the system frequency reuse factor, g(x, b), a polynomially decreasing function of the distance
as defined later in this section. There are mU N users in d (x, b). The AWGN noise power spectral density is nor-
each group, and CM N jointly coordinated antennas in each malized to 1. For a given clustering pattern u(C) and user
cluster. We assume a sufficiently large number of users such bin v(X ), the fading channel coefficients from the CM N BS
antennas of cluster C + c to an active user k {1, . . . , S/m}
6 If a user location falls at the cell boundary it can be assigned arbitrarily
to one of the nearest neighbor cells. We checked that the effect of different 7 A system with mU < CM is not fully loaded, in the sense that the
assignments on the overall system performance is minimal. infrastructure would support potentially a larger number of users.
3230 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 11, NO. 9, SEPTEMBER 2012
at location x + c : x X on frequency subband f , form a
CMN 1
random vector indicated by hk,c ,c (f ; x) C , with
circularly-symmetric complex Gaussian independent entries
(independent Rayleigh fading). In the considered Network-
MIMO schemes, active users are served with equal power
1/S. Hence, the total transmit average power per cluster is
equal to N . Since each BS simultaneously participates in C
(7)
clusters, the average power per BS is also equal to N . Since
we consider the limit for N , the channel coefficients
are normalized to have variance 1/N , such that the received
signal power is independent of N . This provides the correct
scaling of the elements of the random channel vectors such
that the large-system limits hold [11], [20]. We let the channel
vector covariance matrix scaled by N be given by
 
Gc ,c (x) = N E hk,c ,c (f ; x)hH 
k,c ,c (f ; x) , (4)
where Gc ,c (x) = diag (g(x + c , b + c)IMN : b C).8 No-
tice that Gc ,c (x) is independent of the user index k and on the
subband index f , since the channels are identically distributed
across subbands and co-located users.
Under the standard block-fading assumption [11][13], [17],
[39], the channel vectors are constant on each subband for
blocks of length LN signal dimensions. Without loss of gen-
erality, we assume that these coherence blocks also correspond
to the scheduling slot. Each slot is partitioned into an uplink
training phase, of length LP N and a downlink data phase, of
length LD N . In this section we deal with the data phase, while
the training phase is addressed in Section III. For the sake of
notation simplicity, the slot time index is omitted: since we
focus on ergodic (average) rates, only the per-block marginal
channel statistics matter. The data-bearing signal transmitted
by cluster C + c on subband f is given by
Xc (f ) = Uc (f )VcH (f )
L N SN
where the matrix Uc (f ) C D contains the coded
information-bearing symbols arranged by columns, for the
SN served active users. We assume that users codebooks are
drawn from an i.i.d. Gaussian random coding ensemble with
symbols CN (0, 1/S). Achievable rates are obtained via the
familiar random coding argument [40] with respect to this
CMN SN
input distribution. The matrix Vc (f ) C contains
the beamforming vectors arranged by columns, normalized
to have unit norm. It is immediate to verify that, indeed,
the average transmit power of any cluster C + c, active on
frequency f , is given by
 
transmitting on subband f and using training codebook q.
1
tr E XH c (f )Xc (f )
LD N
1  
{0, 1} and X = {x, 1 x} we may have F = 1 (i.e., each
= tr E Vc (f )UH H
c (f )Uc (f )Vc (f )
LD N
LD N   H
= tr E Vc (f )Vc (f )
SLD N
SN
= = N, (6)
S 9 As done in [17], also our analysis is slightly optimistic since it only
as desired.
8 We use diag(M : a A) to indicate a block-diagonal matrix with
a
diagonal blocks Ma , for some index a taking values in the ordered set A,
and In to indicate the n n identity matrix.
Recalling the (reuse-factor dependent) definition of D(f )
as the set of subclusters transmitting on band f , the received
signal for user k at location x + c : x X is given by

yk,c (f ; x) = Uc (f )VcH (f )hk,c,c (f ; x) + zk,c (f ; x)
c D(f )

where zk,c (f ; x) CN 0, F1 ILD N . Notice that a scheme
using frequency reuse F > 1 transmits with total cluster power
N over a fraction 1/F of the whole system bandwidth. This is
taken into account by letting the noise variance per component
be equal to 1/F , in the signal model (7).
By construction, the encoded data symbols for user k at
location x + c : x X , are the entries of the k-th column of
Uc (f ). The columns k  = k of Uc (f ) form the intra-cluster
(multiuser) interference for user k. All other signals Uc (f ),
with c D(f ), c = c, form the Inter-Cluster Interference
(ICI). As seen in Section IV, intra-cluster interference and ICI
are handled by a combination of beamforming and frequency
reuse.
III. U PLINK T RAINING AND C HANNEL E STIMATION
The CSIT is obtained on a per-slot basis, by letting the
active users send pilot signals over LP N dimensions dedicated
to uplink training.9 We fix {X , C} and focus on the SN active
users in the groups X + c : c D(f ). These users must send
SN orthogonal pilot signals to allow channel estimation at
their corresponding serving clusters C + c : c D(f ).
A. Pilot reuse scheme
(5)
Let LP = QS, where Q 1 is an integer pilot reuse factor
QSN QSN
that can be optimized for each {X , C}. Let C
H
be a scaled unitary matrix, such that = ul QSN IQSN ,
where ul denotes the uplink transmit power per user during
the training phase. The matrix is partitioned into Q disjoint
blocks of SN columns each, denoted by 0 , . . . , Q1 and
referred to as training codebooks. These are assigned to the
groups in a periodic fashion, such that the same training
codebook q is reused every Q-th groups X + c : c D(f ).
For later use, we let q(c) {0, . . . , Q 1} denote the index
of the training codebook allocated to group X + c, and define
P(q, f ) = {c D(f ) : q(c) = q} to be the set of clusters
Example 7: In the 1-dimensional layout with C = 2, C =
cluster transmits on the whole system bandwidth) and Q =
2 (i.e., two mutually orthogonal training codebooks are used
alternately, such that the same set of uplink pilot signals is
reused in every other cluster, as shown in Fig. 4).
accounts for the overhead and degradation due to uplink noisy channel
estimation, while it assumes genie-aided overhead-free dedicated training
to support coherent detection during data-transmission. As shown in [15], the
effect of noisy dedicated training is minor relatively to the CSIT estimation
error.
HUH et al.: ACHIEVING MASSIVE MIMO SPECTRAL EFFICIENCY WITH A NOT-SO-LARGE NUMBER OF ANTENNAS 3231

define
g(x + c , b + c)
c ,c,b (f ; x) = (14)
1 + c ,c,b (f ; x)
c ,c,b (f ; x) = g(x + c , b + c) c ,c,b (f ; x)
g(x + c , b + c)
Fig. 4. Pilot reuse and contamination for C = 2, F = 1, and Q = 2. The = , (15)
dashed lines show the pilot contamination at cluster 0 from a user in cluster 1 + c ,c,b (f ; x)1
2, sharing the same pilot signal.
with

c ,c,b (f ; x) =
B. MMSE channel estimation and pilot contamination g(x + c , b + c)
 . (16)
The uplink signal received by the CM N antennas of cluster (ul QS)1 + c P(q(c),f )\c g(x + c , b + c)
C + c : c D(f ), during the training phase, is given by
 The desired channel estimate at cluster C + c is given by
Yc (f ) = q(c ) HH 
h 
c ,c (f ; X ) + Zc (f ). (8) k,c,c (f ; x), obtained by letting c = c in (10) (16). Notice
c D(f ) that the training phase observation rk,c (f ; x) in (9) contains
the superposition of all the channel vectors hk,c ,c (f ; x) of
Thanks to TDD reciprocity, the uplink channel matrix
CMN SN
the users k at locations x + c : x X sharing the
Hc ,c (f ; X ) C contains the downlink channels same pilot signal, i.e., for all c P(q(c), f ). This is the
hk,c ,c (f ; x) arranged by columns, for all active users k = so-called pilot contamination effect of TDD systems [17],
1, . . . , SN/m at all locations x + c : x X . In (8), [18]. Because of pilot contamination, the MMSE estimate
L CMN
Zc (f ) C P denotes the uplink AWGN with compo- 
h k,c,c (f ; x) is correlated with the channels hk,c ,c (f ; x), for
nents CN (0, 1). The goal of the uplink training phase is to all c P(q(c), f ).
provide to each cluster C +c an estimate of the channel vectors Next, we express the channel vector hk,c ,c (f ; x) for c
hk,c,c (f ; x) for all the active users in the corresponding served P(q(c), f ) in terms of the estimate h 
k,c,c (f ; x) and a com-
group X + c. 
ponent independent of h k,c,c (f ; x). This decomposition is
By projecting Yc (f ) onto the columns of q(c) and di-
instrumental to the proofs of Theorems 1, 2 and 3 of Section
viding by ul QSN , the relevant observation for estimating
IV-B (see [41] for details), and it is the key to understanding
hk,c,c (f ; x) is given by
qualitatively the pilot contamination effect. From (10), and
 since Gc,c (x) is invertible, we have
rk,c (f ; x) = hk,c ,c (f ; x) + nk,c (f ) (9)

c P(q(c),f )   (f ; x) = E h  (f ; x)|r (f ; x)
h k,c ,c k,c ,c k,c

where nk,c (f ) CN (0, (ul QSN )1 ICMN ). For any = Gc ,c (x)G1 (x)h  (f ; x). (17)
c,c k,c,c
c P(q(c), f ), the MMSE estimate of hk,c ,c (f ; x) from
rk,c (f ; x) is obtained as Using (11) into (17), the channel vector hk,c ,c (f ; x) from the
antennas of cluster C + c to the unintended user k at location
  (f ; x) = Gc ,c (x)
h x + c : x X can be written as
k,c ,c
1 
 hk,c ,c (f ; x) = Gc ,c (x)G1
c,c (x)hk,c,c (f ; x) + ek,c ,c (f ; x).
( QS)
ul 1
ICMN + Gc ,c (x) rk,c (f ; x). (18)
c P(q(c),f ) The Joint Gaussianity, the mutual orthogonality of
(10)   (f ; x) and e  (f ; x) and the fact that all covariance
h k,c ,c k,c ,c
matrices are diagonal imply that h 
k,c,c (f ; x) and ek,c ,c (f ; x)
Invoking the well-known MMSE decomposition, we can write
are mutually independent.
  (f ; x) + e  (f ; x),
hk,c ,c (f ; x) = h (11) Pilot contamination: As anticipated before, (18) reveals
k,c ,c k,c ,c
qualitatively the pilot contamination effect. With LSUBF,
where the channel estimate h   (f ; x) and the error vector as in [17], cluster C + c serves user k at location x + c
k,c ,c
with beamforming vector h  
ek,c ,c (f ; x) are zero-mean uncorrelated complex circularly
symmetric jointly Gaussian vectors. After some straightfor-
ward algebra (omitted for brevity), we obtain the covariance
matrices (scaled by N )
k,c,c (f ; x)/hk,c,c (f ; x), which is
strongly correlated with the vector hk,c ,c (f ; x) of the channel
from cluster C + c to the unintended user k at location x + c ,
sharing the same pilot signal. It follows that some constant
amount of interfering power, which does not vanish even
 H  (f ; x)],
  (f ; x)h
c ,c (x) = N E[h (12) for N , is leaked into the spatial direction of this
k,c ,c k,c ,c
user, leading to an interference limited system. Theorem 1 in
and Section IV-B precisely quantifies these effects on the ergodic
c ,c (x) = N E[ek,c ,c (f ; x)eH
k,c ,c (f ; x)], (13) user rates. For the family of LZFBF schemes considered in
this work, the pilot contamination effect is less intuitive, but
where c ,c (x) = diag (c ,c,b (f ; x)IMN : b C) and nevertheless precisely quantified by Theorems 2 and 3 in
c ,c (x) = diag (c ,c,b (f ; x)IMN : b C), and where we Section IV-B.
3232 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 11, NO. 9, SEPTEMBER 2012

IV. MU-MIMO P RECODERS AND ACHIEVABLE R ATES are treated as zero.10 Since these schemes are complicated
In the family of Network-MIMO schemes considered in this to explain in full generality, we shall illustrate two specific
work, the beamforming matrix Vc (f ) is calculated by the c-th examples, the generalization of which may be cumbersome
cluster controller as a function of the estimated channel matrix but is conceptually straightforward.
 (f ; X ). For fixed c, c , we say that a user k at location
H Example 8: Consider Fig. 5(a), illustrating a 1-dimensional
c,c
x + c : x X imposes a Zero-Forcing (ZF) constraint on system with C = {0, 1}, X = {x, 1 x}, F = 1 and Q =
cluster c if the beamforming matrix Vc (f ) must satisfy the 2. The beamforming matrix of the c-th cluster satisfies ZF
set of linear equations constraints for its own active users and for the active users
in the 2 locations at minimum distance in the neighboring
H  
vj,c  (f ; x )hk,c,c (f ; x) = 0,
for all (j, x , c ) = (k, x, c), where vj,c
H 
 (f ; x ) is the column
of Vc (f ) (i.e., the beam) corresponding to user j at location
x + c : x X . The schemes considered in this work employ
different forms of LZFBF, distinguished by the number of
ZF constraints imposed. For given {X , C}, we define the
parameter J 0 as the number of ZF constraints imposed
by each active user in the served bin v(X ).
A. Beamforming
(19) clusters c1 and c+1. These locations collectively use distinct
columns of the training codebooks q = q(c). For example,
the reference cluster c = 0 uses training codebook 0 , and
the nearest locations on the left and on the right of cluster
0 use the first and second half (SN/2 columns each) of the
training codebook 1 , respectively. Hence, the 0-th cluster
controller can estimate all the channels of its own active users,
at locations x, 1 x, and of the users in adjacent locations x
and 1 + x, as shown in the figure. The beamforming matrix
in the case of Fig. 5(a) is obtained as follows. Define
Mc (f ; X ) =
   
 (f ; X )  H 
Next we provide expressions for the cluster precoders for
H  (f ; {1 x})  H (f ; {x}) (22)
different choice of the parameter J. c,c c1,c c+1,c
        
Case J = 0: In this case no ZF constraints are imposed. 2MN SN 2MN SN/2 2MN SN/2
Hence, we have
  to be the matrix of dimension 2M N 2SN of estimated chan-
Vc (f ) = UNorm H  c,c (f ; X ) (20) nels at cluster controller c, where the first block corresponds to
the desired active users and the remaining blocks correspond
where UNorm{} indicates a scaling of the columns of the to users in the neighboring clusters for which a ZF constraint
matrix argument to have unit norm. It is immediate to see is imposed. Then,
that (20) coincides with the LSUBF considered in [17].    SN
Case J = 1: In this case any active user imposes ZF con- Vc (f ) = UNorm M+ c (f ; X ) 1 (23)
straints on its own serving cluster. This yields the conventional where []mn extracts the columns from n to m of the matrix
single-cluster LZFBF, for which argument. This scheme can be generalized to J = Q, where
 
Vc (f ) = UNorm H  + (f ; X ) , (21) each c-th cluster satisfies ZF constraints for its own SN active
c,c
users and for a total of (J 1)SN additional users in the
 1 nearest location of neighboring clusters.
where M+ = M MH M denotes the Moore-Penrose
Example 9: Fig. 5(b) shows the same 1-dimensional system
pseudo-inverse of the full column-rank matrix M. It fol-
of Example 8 with a different beamforming design. In this
lows that vk,c (f ; x) is orthogonal to the estimated channels
  
h j,c,c (f ; x ) for all (j, x ) = (k, x) in the group X +c, i.e., ZF
is used to tackle intra-cluster interference while disregarding
ICI.
Case J > 1: In this case, beyond the ZF constraints
imposed to the serving cluster, each user imposes additional
ZF constraints to J 1 neighboring clusters in order to mitigate
also ICI. This provides an alternative approach to frequency
reuse for combating ICI. Any cluster C + c is subject to ZF
constraints imposed by its own users (i.e., users in group
X + c), as well as by users at some locations x + c : x X
for J 1 neighboring clusters c = c. In order to enable
such constraints, the c-th cluster controller must be able to
estimate the channels of these out-of-cluster users. This can
be done if these users employ training codebooks with indices
q = q(c). In particular, J > 1 can be used only if the
pilot reuse factor Q is larger than 1. In some cases, only
the channel subvectors to the nearest BS in the cluster can
be effectively estimated, since there are other users sharing
the same pilot signal that are received with a stronger path
coefficient. The channel subvectors that cannot be estimated
case, the beamforming matrix of each cluster c satisfies ZF
constraints for its own served users and all the users in the
nearest neighbor clusters. However, some of these users share
the same pilot signals. In this specific example, the reference
cluster c = 0 uses training codebook 0 , and clusters c = 1
the other training codebook 1 . Users at location 1 + x use
the same pilot signals of users at location 1 + x, and users
at location x use the same pilot signals of users at location
2x. Then the 0-th cluster controller assumes that the channel
coefficients for BSs at larger distance are equal to zero. In the
example, for locations 1 + x and x, only the subvector
of dimension M N corresponding to the antennas of BS 0 is
estimated, while the remaining subvector to BS 1 is treated as
zero. Similarly, for locations 1+x and 2x only the subvector
of dimension M N corresponding to the antennas of BS 1 is
10 Although treating channel entries with no estimates as zero may seem
arbitrary, this is not so. First, notice that the expected value of the channel
coefficients is zero, which corresponds to the channel MMSE estimate in the
absence of measurements. Moreover, this is consistent with the conventional
approach in cellular networks, which simply ignore users whose channel
coefficients are not acquired.
HUH et al.: ACHIEVING MASSIVE MIMO SPECTRAL EFFICIENCY WITH A NOT-SO-LARGE NUMBER OF ANTENNAS 3233

considered in this work. These expressions are obtained by

letting N . The proof of the following results is an
application of the results previously obtained by some of the
authors in [11]. They are not conceptually difficult, but require
a good deal of technicalities and therefore are omitted for the
(a) J = Q
sake of brevity. In Appendix A we give a very high-level
overview of the proofs. The interested reader can find the full
details in [41].
Theorem 1: For given sets X , C, and system parameters
F, S and Q, in the limit of N , the following group
spectral efficiency of bin v(X ) is achievable with LSUBF
precoding (J = 0):
(b) J = C(Q 1) + 1
RX ,C (F, C, J = 0) =
Fig. 5. Two cases of a precoding scheme for C = 2, F = 1, and Q = 2,
 CM 
with J = Q (a) and J = C(Q 1) + 1 (b). The dashed blue lines indicate S  S 0,0 (x)
log 1 + 1 CM
, (26)
the channel vectors of out-of-cluster user locations imposing a ZF constraint. mF F + (x) + S (x)
In Figure (b), the light-shaded grey dashed lines indicate the channel vectors xX
assumed zero in the beamforming calculation.
where11
1    c,c,b (x )g(x, c + b)
estimated, while the remaining subvector to BS 0 is treated as (x) = (27)
mC  c,c (x )
zero. x X bC cD(0)

The beamforming matrix corresponding to the scheme of and

Fig. 5(b) is obtained as follows. Define  2
 1 1  g(x, c + b)
Mc (f ; X ) = (x) = c,c,b (x) ,
   
 (f ; X )  H  c,c (x) C g(x, b)
H 
c1,c (f ; X ) L Hc+1,c (f ; X ) R (24) cP(0,0)\0 bC
c,c
         (28)
2MN SN 2MN SN 2MN SN with
to be the matrix of dimension 2M N 3SN of all estimated 1 
c,c (x) = c,c,b (x), (29)
channels at cluster controller c, where indicates element- C
bC
wise product and where
    and c,c ,b (x) given by (15), are coefficients that depend
1MN SN 0MN SN
L= , R= uniquely on the system geometry, frequency and pilot reuse,
0MN SN 1MN SN but are independent of the loading factor S and of the BS
are masking matrices that null out the subvectors of the antenna factor M .
channels that are treated as zero in the beamforming design. Eqn. (26) concisely captures the dependence of the group
Then, Vc (f ) is again given by (23) although in this case ergodic rate on the cluster size C, antenna and loading factors,
Mc (f ; X ) is given by (24) instead of (22). This scheme can M and S, and frequency reuse F . Focusing on the effect
be generalized to J = C(Q 1) + 1, where each c-th cluster of the interference terms, as captured by the denominator of
satisfies ZF constraints for its own SN active users and for a the SINR inside the logarithm, the term 1/F captures the
total of (J 1)SN additional users in the neighboring clusters, effect of noise, the term CM (x)/S captures the effect of
with some channel sub-vectors set to zero. intercell-interference due to user streams for users that trained
on the same pilot (effect of pilot contamination), while (x)
B. Achievable group spectral efficiency captures aggregate effect of intra- and the remaining inter-cell
(N ) interference. As observed in [17], in the regime M , the
Let Rk,c (f ; x) denote the spectral efficiency (in bit/s/Hz) pilot contamination dominates the other interference terms,
of user k at location x + c : x X , served by cluster C + c and thus the system spectral efficiency is uniquely limited by
according to one of the schemes defined above. The group the ICI due to pilot contamination. In particular, the result of
spectral efficiency of bin v(X ) served by the cluster pattern [17] can be recovered as a corollary of Theorem 1, by letting
u(C) is defined as the sum rate of all users in bin v(X ), M in (26) (infinite number of BS antennas per active
averaged over all subbands f = 0, . . . , F 1, and normalized user), and setting C = 1, and Q = 1 as in [17]. We obtain
by the number of cells B and by the system scaling parameter
N . Explicitly, this is given by lim RX ,{0} (F, 1, 0) =
M
 
1 
F 1   SN/m
 (N ) S  g(x, 0)2
RX ,C (F, C, J) = Rk,c (f ; x) log 1 +  2
. (30)
F BN mF cP(0,0)\0 g(x, c)
f =0 cbs V xX k=1 xX
(25)
Next, we provide closed-form expressions for the achievable 11 In (27) and (28) it is assumed, without loss of generality, that cluster
group spectral efficiency for all the Network-MIMO schemes c = 0 uses subband f = 0 and training codebook q = 0.
3234 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 11, NO. 9, SEPTEMBER 2012

As observed in [17], in this regime the system spectral in cases (a) and (b), and
efficiency is uniquely limited by the ICI due to pilot con- 
tamination. (x) = 0,0 (x) + g0,c (x)
The next result yields the achievable group spectral effi- cD(0)\0E(x)

ciency of LZFBF in the case of single-cell processing (i.e., 1  
for C = 1). The distance between a user location x X + 0,c,b(x,c) (x) + g(x, c + b) (38)
C
and a cluster C + c is defined as the minimum of the distance cE(x) bC\b(x,c)
between x and any BS in the cluster, i.e.,
in case (c), and where
d (x, C + c) = min{d (x, b + c) : b C}. (31)   2
1  g(x, c + b)
(x) = c,c,b (x), (39)
We let E(x) denote the set of J 1 clusters c = 0 closest to C g(x, b)
cP(0,0)\0 bC
x X (if J = 1 then E(x) = ). With these definitions, we
have: with
Theorem 2: For given set X , C = 1 (i.e., C = {0}), and 1  1 
system parameters F, S and Q, in the limit of N , the g0,c (x) = g(x, c + b), 0,c (x) = 0,c,b (x),
C C
following group spectral efficiency of bin v(X ) is achievable bC bC
with LZFBF precoding (J 1): (40)
where c ,c,b (f ; x) and c,c ,b (x) given by (14) and (15), re-
RX ,{0} (F, 1, J 1) = spectively, are coefficients that depend uniquely on the system
 
S  MJS
0,0,0 (x) geometry, frequency and pilot reuse, but are independent of
S
log 1 + 1 MJS
(32) the loading factor S and of the BS antenna factor M .
mF F + (x) + S (x)
xX

where V. S CHEDULING AND FAIRNESS

 
(x) = 0,c,0 (x) + g(x, c) Consider a system with K bins, {v(X0 ), . . . , v(XK1 )},
cP(0,0)E(x) cD(0)P(0,0)E(x) defined by sets of symmetric locations chosen to uniformly
(33) discretize the cellular coverage region V.
and  2 The net bin spectral efficiency (in bit/s/Hz) provided to
 g(x, c)
(x) = c,c,0 (x), (34) a bin v(Xk ) by a Network-MIMO scheme with a specific
g(x, 0) combination of parameters is given by the product
cP(0,0)\0

and c,c ,b (x) given by (15), are coefficients that depend
uniquely on the system geometry, frequency and pilot reuse,
but are independent of the loading factor S and of the BS
antenna factor M .
In passing, we notice that the limit of (32) for M
coincides with (30). Therefore, as observed in [17], in the
Massive MIMO regime LZFBF yields no advantage over
the simpler LSUBF.
The case of LZFBF with multicell processing (C > 1) needs
the definition of some more notation. As in Examples 8 and
9, we consider the cases J = 1, J = Q and J = C(Q
1) + 1, referred to as cases (a), (b) and (c), respectively, in
the following. In case (c), we define b(x, c) to be the BS in
cluster c E(x) closest to user location x X , i.e.,
b(x, c) = arg min{d (x, b + c) : b C}.
Theorem 3: For given sets X , C with C > 1, and system
parameters F, S and Q, in the limit of N , the following
group spectral efficiency of bin v(X ) is achievable with
LZFBF precoding (J 1):
RX ,C (F, C, J 1) =
 CMJS 
S  S 0,0 (x)
log 1 + 1 CM
mF F + (x) + S (x)
xX
max{1 QS/L, 0} RXk ,C (F, C, J), (41)
where the first term is the ratio between data-phase and
total slot channel uses, and takes into account the pilot
dimensionality overhead. The second term in (41) is the bin
spectral efficiency of the data phase for the given Network-
MIMO scheme, and is given by one of the three closed-
form expressions in Theorems 1, 2 and 3, depending on the
specific beamforming case. The net bin spectral efficiency
at a bin v(Xk ) can be maximized by choosing among all
Network-MIMO schemes in a family, the one that maximizes
the product (41). The maximization of (41) is subject to the
constraint JS CM , which becomes relevant for J > 0 (i.e.,
for LZFBF precoding) and requires searching over a discrete
parameter space (apart from S, which is continuous). The
(35) simple closed-form expressions given in Theorems 1, 2 and
3 allow for an efficient system optimization, avoiding lengthy
Monte Carlo simulations.
Let R (Xk ) denote the maximum of (41) for given Xk .
Then, a scheduler allocates the different bins on the time-
frequency slots in order to maximize some desired network
utility function of the user rates. With randomized or round-
robin user selection, each user in bin v(Xk ) shares on av-
(36) erage an equal fraction of the product k R (Xk ), where k
is the fraction of the transmission resource (time-frequency

where (x) takes on the expressions slots) allocated to bin v(Xk ). In general, the scheduler allo-
  cates the transmission resource to the user bins in order to
(x) = 0,c (x) + g 0,c (x), (37) maximize a componentwise non-decreasing concave network
cE(x){0} cD(0)\0E(x) utility function of the bin spectral efficiencies, denoted by
HUH et al.: ACHIEVING MASSIVE MIMO SPECTRAL EFFICIENCY WITH A NOT-SO-LARGE NUMBER OF ANTENNAS
G(R0 , . . . , RK1 ) [21], [29][32]. In our setting, the sched-
uler determines the transmission resource allocation {k } by
solving the following convex problem:
maximize G(R0 , . . . , RK1 )
K1
subject to Rk k R (Xk ), k 1, k 0.
k=0
(42)
For example, the popular Proportional Fairness (PF) criterion
[30] corresponds to the choice
K1
G(R0 , . . . , RK1 ) = log Rk ,
k=0
resulting in the bin time-frequency sharing fractions k =
1/K (each bin is given an equal amount of slots). Instead, if
the minimum user rate constitutes a more relevant figure of
merit for the system performance, Max-Min fairness can be
imposed by considering the function
G(R0 , . . . , RK1 ) = min Rk ,
k=0,...,K1
resulting in the bin time-frequency sharing fractions k =
1
R (Xk )
K1 1 . In general, a whole family of scheduling rules
j=0 R (Xj )
including PF and Max-Min as special cases is obtained by
using the so-called -fairness network utility function, as
defined in [29].
VI. N UMERICAL R ESULTS AND D ISCUSSION
In this section we illustrate the performance benefits
provided by the proposed architectures, over the Massive
MIMO scheme of [17]. In particular, we first verify the ac-
curacy of our large-system limit closed-form rate expressions
(as N ) in terms of approximating the finite system
setting. We then use the large-system analysis to show that
significant gains can be provided by the proposed architectures
with respect to the Massive MIMO architecture of [17], in the
relevant regime of a finite number of BS antenna per active
user.
Fig. 6 shows the group spectral efficiency in (41) as a
function of the bin locations within a cell for different schemes
identified by the parameters (F, C, J) and Q. The group
spectral efficiency is obtained by Monte Carlo simulation
(dotted) and is compared against the corresponding values
from the closed-form large-system analysis (solid), for the 1-
dimensional cell layout of Fig. 2 with B = 24 BSs, M = 30
BS antenna factor, L = 40 coherence block dimension factor,
and K = 20 bins in each cluster, where clusters and location
bins are given in Example 3 and 5, with x equally spaced
in [0, 1/2]. The pathloss coefficients are given by g(x, b) =
G0 /(1 + (d (x, b)/) ), with G0 = 106 , = 3.76, and
= 0.05, and reflects (after suitable normalizations) a typical
cellular scenario with 1 km diameter cells in a sub-urban
environment [28]. The (1,1,1) scheme with Q = 1 yields the
best performance for locations near the cell center. However,
at the cell edges, C = 2, J = 3, attain significantly better
performance. This hints that a mixed-mode architecture that
serves users in the cell center using the (1,1,1), Q = 1 scheme,
3235
Fig. 6. Bin spectral efficiency vs. location within a cell obtained from the
large system analysis (solid) and the finite dimension (N = 1) simulation
(dotted) for various (F, C, J). M = 30 and L = 40.
and users at the cell edge using the (1,2,3), Q = 2 scheme,
can achieve significant gains over a fixed-mode architecture,
especially in the case where worst-case users dominate the
system performance, as in the case of Max-Min fairness. This
observation is confirmed by the results of Figs. 9 and 10,
presented at the end of this section.
It is also worthwhile to notice that, as expected, the limit for
N matches very accurately with the finite-dimensional
Monte Carlo simulation even for very small N (the simulated
performance in Fig. 6 corresponds to estimated spectral effi-
ciencies based on simulations for the case N = 1). For this
reason, in the following we present only the results for the
large-system limit, obtained using the closed-form expressions
of Theorems 1, 2 and 3.
In the 2-dimensional case, we considered the layout with
B = 19 hexagonal cells as shown in Fig. 1. For comparison,
we assume the same system model as in [17], with channel
coherence block dimension, the cell radius, and pathloss model
given by LP = 84, 1.6 km, and g(x, b) in the same form as
before, with parameters G0 = 106 , = 0.1 km, and =
3.8, respectively. Log-normal shadowing, considered in [17],
is not considered here (see the comment in Section VII). We
considered schemes with cluster size C = 1 and C = 3,
K = 16 bins with 48 user locations, where the cluster and
bin layout are qualitatively described in Fig. 2. The frequency
reuse factor F and pilot reuse factor Q are selected between
1 and 3. For example, when F = 3 (or Q = 3) and C = 1,
the frequency subbands (or training codebooks) are allocated
to clusters as shown in Fig. 1 where different colors denotes
different subbands (or training codebooks). For C = 3, the
same pattern is used for the clusters formed as shown in Fig. 2.
Fig. 7 illustrates the optimum over the family of Network-
MIMO schemes for (a) M = 20 and (b) M = 100. In both
cases, (1, 1, 1) is optimal in the inner part of the cell, but
schemes with (3, 3, 1) or (3, 1, 1) yield better performance for
locations near the cell boundary. We notice also that the inner
region within which the (1,1,1) scheme is the best increases
with the BS antenna factor M .
Next, we compare the performance of the proposed ar-
3236 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 11, NO. 9, SEPTEMBER 2012
(a) M=20
(b) M=100
Fig. 7. Best scheme at each user location. M = 20 and 100, K = 16, and
L = 84.
chitecture with the one advocated in [17]. Fig. 8 shows the
bin-optimized spectral efficiency normalized by the spectral
efficiency of the (1, 1, 0), Q = 1 scheme (corresponding to
[17]), under two-dimensional layout with M = 50. The
gain of the proposed architecture ranges from about 40%
to 580%, depending on the users location. Fig. 9 shows
the system throughput under Proportional Fairness Scheduling
as a function of M in the two-dimensional layout. Both
the throughput for some fixed Network-MIMO schemes, and
the throughput of the bin-optimized mixed-mode scheme as
described in Section V are shown, and compared with the
reference performance of the (1, 1, 0), Q = 1 scheme. The
cluster scheme includes two cases where the cluster pattern
is fixed to be one of two cases shown in Fig. 2, or where
the cluster pattern is chosen among the two cases of Fig. 2
depending on the user locations. For the sake of comparison
with the results in [17], we considered 20 MHz system
bandwidth and the coherence block size L = 84 as in
(considering the parameters of 3GPP LTE TDD system). As
the figure reveals, the (3, 3, 1) fixed schemes perform very
well for small M < 20 while, as M increases, the (1, 1, 1)
scheme is best. The bin-optimized architecture improves the
throughput further at any value of M . The dotted horizontal
line in Fig. 9 denotes the cell throughput claimed in [17] in
the limit of an infinite number of transmit antennas per user
with the (1, 1, 0), Q = 1 scheme. We notice that this limit
is approached very slowly, and more than 10000 antennas
Fig. 8. Bin-optimized spectral efficiencies normalized by the (1,1,0) spectral
efficiencies. M = 50, K = 16,and L = 84.
Fig. 9. Cluster sum throughput vs. M for various (F, C, J) and for a bin-
optimized architecture under PF scheduling. K = 16 and L = 84. The arrow
indicates that the proposed architecture achieves the same spectral efficiency
as the fixed scheme (1, 1, 0) of [17], with a 10-fold reduction of the number
of BS antennas.
per BS are required (clearly impractical). For finite number
of antennas, the proposed architecture achieves the same
throughput of the scheme in [17] with a 10-fold reduction
in the number of antennas at the base stations (roughly, from
500 to 50 antennas, as indicated by the arrow).
Finally, Fig. 10 depicts the cell-throughput performance for
the same set of system parameters and designs as in Fig. 9, un-
der Max-Min fairness scheduling. We notice that in this case,
unlike the case of PFS, the system performance is severely
limited by the worst-case user locations, which represent the
system bottleneck. In this case, the advocated bin-optimized
architecture achieves significant throughput gains over the
fixed schemes, because of its ability of improving the bin
spectral efficiency of the worst-case user locations.
VII. C ONCLUSIONS
We studied a novel Network-MIMO TDD architecture
that can achieve spectral efficiencies comparable with the
recently proposed Massive MIMO scheme, while requiring
an order of magnitude fewer antennas per active user per
HUH et al.: ACHIEVING MASSIVE MIMO SPECTRAL EFFICIENCY WITH A NOT-SO-LARGE NUMBER OF ANTENNAS 3237

utility function, can be further generalized to include shad-

owing, users with different mobility and fractional frequency
reuse. These generalizations are, however, non-trivial. For
example, in the presence of slow frequency-flat shadowing,
bins are no-longer determined by the users geographic
position. Rather, the set of large-scale channel gains (including
shadowing), should be used to classify the users in equiva-
lence classes. Also, in the presence of users with different
mobility and/or frequency selectivity, users can be partitioned
according to their different channel coherence block length.
The issue of how to optimally and dynamically classify users
into equivalence classes and how to serve each class with
a specifically optimized MU-MIMO scheme represents an
interesting avenue for future work.

A PPENDIX A
Fig. 10. Cluster sum throughput vs. M for various (F, C, J) and for a
bin-optimized architecture under Max-Min scheduling. K = 16 and L = 84. P ROOF OF T HEOREMS 1, 2 AND 3: M AIN I DEAS
We focus on the reference cluster C (i.e., c = 0), with cor-
responding served group of locations X = {x0 , . . . , xm1 }.
cell. The proposed strategy operates by partitioning the users For brevity, the subband index f is omitted and D denotes the
population into geographically determined bins. The time- set of clusters transmitting on the same subband of cluster 0.
frequency scheduling slots are allocated to the bins in order From (7), the (scalar) signal received at some symbol interval
to form independent MU-MIMO transmissions, each of which of the data phase, at the k-th active user receiver at location
is optimized for the corresponding bin. This strategy allows x X , is given by
the uplink training reuse factor, the frequency reuse factor, H
yk,0 (x) = uk,0 (x)vk,0 (x)hk,0,0 (x) (43)
the active user loading factor, the BS cooperative cluster 
H
size and the type of MU-MIMO linear beamforming to be + uj,0 (x)vj,0 (x)hk,0,0 (x)
finely tailored to the particular user bin, rather than for the j =k
 
system worst-case. We considered system optimization over + uj,0 (x )vj,0
H
(x )hk,0,0 (x) (44)
1-dimensional and 2-dimensional cell layouts, based on a x X \x j
family of Network-MIMO schemes ranging from single-cell   
+ uj,c (x )vj,c
H 
 (x )hk,0,c (x) + zk,0 (x),
processing to joint processing over clusters of coordinated
c D\0 x X j
BSs, with linear precoders ranging from conventional linear
(45)
single-user beamforming to zero-forcing beamforming with
additional zero-forcing constraints for neighboring cells. In where uj,c (x ) denotes the code symbol transmitted by cluster
order to carry out the system optimization, we developed c , to user j at location x + c : x X .
closed-form expressions for the achievable spectral efficiency Using the MMSE decomposition (11), we isolate the useful
for each scheme in the family and each bin in the cellular signal term from (43), given by,
layout. Our closed-form analysis is based on the large-system H
uk,0 (x)vk,0 
(x)h (46)
limit, where all system dimensions scale to infinity with fixed k,0,0 (x).
ratios, and make use of recent results (by some of the authors The sum of the residual self-interference term due to the
of this paper) on the analysis of cellular systems with linear channel estimation error with the signals in (44) transmitted
zero-forcing beamforming and channel estimation errors [11]. by cluster 0 to the other users, results in the intra-cluster
The performance predicted by the large-system asymptotic interference term
analysis is shown to match very well with finite-dimensional 
H H
simulations. Our numerical results show that different schemes uk,0 (x)vk,0 (x)ek,0,0 (x) + uj,0 (x)vj,0 (x)hk,0,0 (x)
in the considered family achieve the best spectral efficiency at j =k
 
 H
different user locations. This suggests the need for a location- + uj,0 (x )vj,0 (x )hk,0,0 (x). (47)
adaptive scheme selection to serve efficiently the whole cover- x X \x j
age region. The resulting overall system is therefore a mixed-
mode Network-MIMO architecture, where different schemes, Finally, the ICI term and background noise are given in (45).
each of which is optimized for the corresponding user bin, are From [11, Theorem 4] the rate
multiplexed in the time-frequency plane. (N )
Rk,0 (x) =
As a final remark, it is worthwhile pointing out that the  

E |useful sig.|2 | vk,0 (x), h
approach of partitioning the users in homogeneous sets, and k,0,0 (x)
serving each set according to a specifically optimized scheme, E log 1 +  

E |interf. + noise| | vk,0 (x), hk,0,0 (x)
2
while letting the overall scheduler allocate the transmission
resource to the sets in order to maximize some desired network (48)
3238
is achievable by a Gaussian i.i.d. random coding ensemble,
assuming that the receiver has perfect knowledge of its own
estimated channel and beamforming vector. Both numerator
and denominator of the Signal-to-Interference plus Noise
Ratio (SINR) appearing inside the log in (48) converge to
deterministic limits as N , so that the outer expectation
in (48) becomes irrelevant asymptotically. These limits can be
calculated in closed-form by using the representation of the
channel MMSE estimates as
1 analysis, IEEE Trans. Inf. Theory, vol. 57, no. 12, Dec. 2011.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 11, NO. 9, SEPTEMBER 2012
[18] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, Pilot
contamination and precoding in multi-cell TDD systems, accepted
for publication in IEEE Trans. Wireless Commun., 2011. Available:
http://arxiv.org/abs/0901.1703
[19] J. Hoydis, M. Debbah, and S. T. Brink, Massive MIMO: how many
antennas do we need? July 2011. Available: http://arxiv.org/abs/1107.
1709v1
[20] A. M. Tulino and S. Verdu, Random Matrix Theory and Wireless
Communications. Foundations and Trends in Communications and In-
formation Theory, 2004, vol. 1, no. 1.
[21] H. Huh, S.-H. Moon, Y.-T. Kim, I. Lee, and G. Caire, Multi-cell MIMO
downlink with cell cooperation and fair scheduling: a large-system limit

h 
1/2  
(49)
j,c,c (x ) = c,c (x )aj,c,c (x )
N through correlated channels in the presence of correlated interferers and
where the vectors aj,c,c (x ) are i.i.d. CN (0, ICMN ), with
generic components denoted by {an,b : n = 1, . . . , M N } for
all b C, the pilot contamination decomposition (18) and the
analysis in [11], which is directly applicable to this setting
(see details in [41, Appendix A and B]).
R EFERENCES
[1] 3GPP technical specification group radio access network, Further
advancements for E-UTRA: LTE-Advanced feasibility studies in RAN
WG4, 3GPP TR 36.815, Tech. Rep., Mar. 2010.
[2] G. Boudreau, J. Panicker, N. Guo, R. Chang, N. Wang, and S. Vrzic,
Interference coordination and cancellation for 4G networks, IEEE
Commun. Mag., vol. 47, no. 4, pp. 7481, Apr. 2009.
[3] H. Dahrouj and W. Yu, Coordinated beamforming for the multicell
multi-antenna wireless system, IEEE Trans. Wireless Commun., vol. 9,
no. 5, pp. 17481759, May 2010.
[4] H. Huh, H. C. Papadopoulos, and G. Caire, Multiuser MISO transmitter
optimization for intercell interference mitigation, IEEE Trans. Signal
Process., vol. 58, no. 8, pp. 42724285, Aug. 2010.
[5] G. J. Foschini, K. Karakayali, and R. A. Valenzuela, Coordinating mul-
tiple antenna cellular networks to achieve enormous spectral efficiency,
IEE Proc. Commun., vol. 153, no. 4, pp. 548555, Aug. 2006.
[6] S. Jing, D. N. C. Tse, J. B. Soriaga, J. Hou, J. E. Smee, and R. Padovani,
Downlink macro-diversity in cellular networks, in Proc. 2007 IEEE
Int. Symp. Inf. Theory.
[7] F. Boccardi and H. Huang, Limited downlink network coordination
in cellular networks, in Proc. 2007 IEEE Int. Symp. Personal, Indoor,
Mobile Radio Commun.
[8] G. Caire, S. A. Ramprashad, H. C. Papadopoulos, C. Pepin, and C.-
E. W. Sundberg, Multiuser MIMO downlink with limited inter-cell
cooperation: approximate interference alignment in time, frequency and
space, in Proc. 2008 Allerton Conf. Commun., Control, Comput.
[9] S. A. Ramprashad and G. Caire, Cellular vs. network MIMO: a
comparison including the channel state information overhead, in Proc.
2009 IEEE Int. Symp. Personal, Indoor, Mobile Radio Commun.
[10] S. A. Ramprashad, G. Caire, and H. C. Papadopoulos, Cellular and
network MIMO architectures: MU-MIMO spectral efficiency and costs
of channel state information, in Proc. 2009 IEEE Asilomar Conf.
Signals, Syst., Comput.
[11] H. Huh, A. Tulino, and G. Caire, Network MIMO with linear zero-
forcing beamforming: large system analysis, impact of channel estima-
tion and reduced-complexity scheduling, to appear in IEEE Trans. Inf.
Theory, Dec. 2010. Available: http://arxiv.org/abs/1012.3198
[12] J. G. Proakis, Digital Communications. McGraw-Hill, 2000.
[13] L. Zheng and D. N. C. Tse, Communication on the Grassmann
manifold: a geometric approach to the noncoherent multiple-antenna
channel, IEEE Trans. Inf. Theory, vol. 48, no. 2, pp. 359383, Feb.
2002.
[14] B. Hassibi and B. M. Hochwald, How much training is needed in
multiple-antenna wireless links? IEEE Trans. Inf. Theory, vol. 49, no. 4,
pp. 951963, Apr. 2003.
[15] G. Caire, N. Jindal, M. Kobayashi, and N. Ravindran, Multiuser MIMO
achievable rates with downlink training and channel state feedback,
IEEE Trans. Inf. Theory, vol. 56, no. 6, pp. 28452866, June 2010.
[16] T. L. Marzetta, How much training is required for multiuser MIMO?
in Proc. 2006 IEEE Asilomar Conf. Signals, Syst., Comput.
[17] , Noncooperative cellular wireless with unlimited numbers of base
station antennas, IEEE Trans. Wireless Commun., vol. 9, no. 11, pp.
35903600, Nov. 2010.
[22] A. L. Moustakas, S. H. Simon, and A. M. Sengupta, MIMO capacity
noise: a (not so) large N analysis, IEEE Trans. Inf. Theory, vol. 49,
no. 10, pp. 25452561, Oct. 2003.
[23] K. R. Kumar, G. Caire, and A. Moustakas, Asymptotic performance
of linear receivers in MIMO fading channels, IEEE Trans. Inf. Theory,
vol. 55, no. 10, pp. 43984418, Oct. 2009.
[24] A. Forenza, M. Airy, M. Kountouris, R. Heath Jr, D. Gesbert, and
S. Shakkottai, Performance of the MIMO downlink channel with multi-
mode adaptation and scheduling, in Proc. 2005 IEEE Workshop Signal
Process. Adv. Wireless Commun.), pp. 695699.
[25] A. Papadogiannis, D. Gesbert, and E. Hardouin, A dynamic clustering
approach in wireless networks with multi-cell cooperative processing,
in Proc. 2008 IEEE Int. Conf. Commun., pp. 40334037.
[26] R. Chen, Z. Shen, J. Andrews, and R. Heath, Multimode transmission
for multiuser MIMO systems with block diagonalization, IEEE Trans.
Signal Process., vol. 56, no. 7, pp. 32943302, July 2008.
[27] H. Shirani-Mehr, G. Caire, and M. J. Neely, MIMO downlink schedul-
ing with non-perfect channel state knowledge, IEEE Trans. Commun.,
vol. 58, no. 7, pp. 20552066, July 2010.
[28] H. C. Papadopoulos, G. Caire, and S. A. Ramprashad, Achieving large
spectral efficiencies from MU-MIMO with tens of antennas: location-
adaptive TDD MU-MIMO design and user scheduling, in Proc. 2010
IEEE Asilomar Conf. Signals, Syst., Comput.
[29] J. Mo and J. Walrand, Fair end-to-end window-based congestion
control, IEEE/ACM Trans. Netw., vol. 8, pp. 556567, Oct. 2000.
[30] P. Viswanath, D. N. C. Tse, and R. Laroia, Opportunistic beamforming
using dumb antennas, IEEE Trans. Inf. Theory, vol. 48, no. 6, pp. 1277
1294, June 2002.
[31] P. Bender, P. Black, M. Grob, R. Padovani, N. Sindhushayana, and
A. Viterbi, CDMA/HDR: a bandwidth-efficient high-speed wireless
data service for nomadic users, IEEE Commun. Mag., vol. 38, no. 7,
pp. 7077, July 2000.
[32] S. Parkvall, E. Englund, M. Lundevall, and J. Torsner, Evolving 3G
mobile systems: broadband and broadcast services in WCDMA, IEEE
Commun. Mag., vol. 44, no. 2, pp. 3036, Feb. 2006.
[33] D. Aktas, M. N. Bacha, J. S. Evans, and S. V. Hanly, Scaling results on
the sum capacity of cellular networks with MIMO links, IEEE Trans.
Inf. Theory, vol. 52, no. 7, pp. 32643274, July 2006.
[34] J. Hoydis, M. Kobayashi, and M. Debbah, On the optimal number
of cooperative base stations in network MIMO systems, submitted to
IEEE Trans. Signal Process. Available: http://arxiv.org/abs/1003.0332
[35] R. Zakhour and S. V. Hanly, Base station cooperation on the downlink:
large system analysis, submitted to IEEE J. Sel. Areas Commun., June
2010. Available: http://arxiv.org/abs/1006.3360
[36] 3GPP technical specification group radio access network, Further
advancements for E-UTRA: physical layer aspects, 3GPP TR 36.814,
Tech. Rep., Mar. 2010.
[37] IEEE 802.16 broadband wireless access working group, IEEE 802.16m
evaluation methodology document (EMD), IEEE 802.16m-08/004,
Tech. Rep., Jan. 2009.
[38] T. S. Rappaport, Wireless Communications: Principles & Practice.
Prentice Hall, 2002.
[39] T. L. Marzetta and B. M. Hochwald, Capacity of a mobile multiple-
antenna communication link in Rayleigh flat fading, IEEE Trans. Inf.
Theory, vol. 45, no. 1, pp. 139157, Jan. 1999.
[40] T. M. Cover and J. A. Thomas, Elements of Information Theory. Wiley,
2005.
[41] H. Huh, G. Caire, H. C. Papadopoulos, and S. A. Ramprashad, Achiev-
ing massive MIMO spectral efficiency with a not-so-large number of
antennas, 2011. Available: http://arxiv.org/abs/1107.3862
HUH et al.: ACHIEVING MASSIVE MIMO SPECTRAL EFFICIENCY WITH A NOT-SO-LARGE NUMBER OF ANTENNAS
Hoon Huh (S 09 M 11) received the B.S.
and M.S. degrees from Seoul National University,
Seoul, Korea and the Ph.D. degree from the Uni-
versity of Southern California, Los Angeles, CA,
all in electrical engineering, in 1997, 1999, and
2011, respectively. Since 1999, he has been with
Samsung Electronics, Suwon, Korea, working on the
physical and MAC layer standards and the MODEM
and scheduler algorithms for the 3GPP LTE/LTE-
Advanced, Mobile WiMAX, and CDMA2000 1xEV-
DO systems. His research interests are multiple-
antenna techniques, resource allocation, and cooperative communications
in the areas of communications theory, information theory, and wireless
communications.
Giuseppe Caire (S 92 M 94 SM 03 F 05) in the areas of communications and signal processing, with emphasis on
was born in Torino, Italy, in 1965. He received the
B.Sc. in Electrical Engineering from Politecnico di
Torino (Italy), in 1990, the M.Sc. in Electrical Engi-
neering from Princeton University in 1992 and the
Ph.D. from Politecnico di Torino in 1994. He was a
recipient of the AEI G.Someda Scholarship in 1991,
has been with the European Space Agency (ESTEC,
Noordwijk, The Netherlands) from May 1994 to
February 1995, was a recipient of the COTRAO
Scholarship in 1996 and of a CNR Scholarship in
1997.
He has been visiting Princeton University in Summer 1997 and Sydney
University in Summer 2000. He has been Assistant Professor in Telecommu-
nications at the Politecnico di Torino, Associate Professor at the University of
Parma, Italy, Professor with the Department of Mobile Communications at the
Eurecom Institute, Sophia-Antipolis, France, and he is currently a professor
of Electrical Engineering with the Viterbi School of Engineering, University
of Southern California, Los Angeles, CA.
He served as Associate Editor for the IEEE T RANSACTIONS ON C OM -
MUNICATIONS in 1998-2001 and as Associate Editor for the IEEE T RANS -
ACTIONS ON I NFORMATION T HEORY in 2001-2003. He received the Jack
Neubauer Best System Paper Award from the IEEE Vehicular Technology
Society in 2003, and the IEEE Communications Society & Information
Theory Society Joint Paper Award in 2004 and in 2011. Giuseppe Caire is
Fellow of IEEE since 2005. He has served in the Board of Governors of the
IEEE Information Theory Society from 2004 to 2007, and as President of the
IEEE Information Theory Society in 2011. His main research interests are in
the field of communications theory, information theory, channel and source
coding with particular focus on wireless communications.
3239
Haralabos C. Papadopoulos (S 92 M 98)
was born in Serres, Greece, in 1968. He received
the S.B., S.M., and Ph.D. degrees from the Mas-
sachusetts Institute of Technology, Cambridge, MA,
all in electrical engineering and computer science, in
1990, 1993, and 1998, respectively. Since December
2005, he has been with DOCOMO Innovations, Palo
Alto, CA, working on physical-layer algorithms for
wireless communication systems and architectures.
From 1998 to 2005, he was on the faculty of the
Department of Electrical and Computer Engineering,
University of Maryland, College Park, MD, and held a joint appointment
with the Institute of Systems Research. During his 19931995 summer visits
to AT&T Bell Labs, Murray Hill, NJ, he worked on shared time-division
duplexing systems and digital audio broadcasting. His research interests are
resource-efficient algorithms and architectures for wireless communication
systems.
Dr. Papadopoulos is the recipient of an NSF CAREER Award (2000), the
G. Corcoran Award (2000) given by the University of Maryland, College Park,
and the 1994 F. C. Hennie Award (1994) given by the MIT EECS department.
He is also a coauthor of the VTC Fall 2009 Best Student Paper Award. He is
a member of Eta Kappa Nu and Tau Beta Pi. He is also active in the industry
and an inventor on several issued and pending patents.
Sean A. Ramprashad received a B.S.E. degree
summa cum laude from Princeton University in 1991
and M.S. and Ph.D. degrees from Cornell University
in 1993 and 1996, all in Electrical Engineering.
Dr. Ramprashad was an intern with AT&T Bell
Laboratories during the years 1993 to 1995, a Mem-
ber of Technical Staff at Lucent Technologies Bell
Laboratories from 1996 to 2000, and a Distinguished
Member of Technical Staff at Agere Systems from
2000 to 2003. Up until April 2012, Dr. Ramprashad
was a Principal Engineer and Project Manager at
DoCoMo Innovations where he managed teams focused on multimedia and
wireless communications.
Dr. Ramprashads research interests have covered a range of areas including
video coding, speech and audio processing, detection theory, vector quanti-
zation, multimedia systems, 802.11, MIMO, and wireless communications.
He served on the committees for the 1997 and 2002 IEEE Workshops on
Speech Coding, served as an Associate Editor for the IEEE T RANSACTIONS
ON S PEECH AND AUDIO P ROCESSING , co-supervised a summer intern who
was awarded the best student paper award in Fall VTC 2009, and has
planned and chaired numerous conference sessions. Dr. Ramprashad has been
involved in a variety of product teams focused ITU-T speech coding standards,
LTE standardization, audio and video-conferencing systems, satellite audio
broadcasting, media over 802.11 and Bluetooth, and multi-processor mobile
multimedia platforms.