Capacity of Multi-Antenna Array Systems

Chen-Nee Chuah
supervised by
Professor Joseph Kahn and Professor David Tse
Department of Electrical Engineering and Computer Sciences
University of California
Berkeley, CA 94720
January 25, 2000
Abstract:
This work is inspired by the large capacity increase that may be achieved by using multi-ele-
ment antenna arrays (MEA) at both transmitting and receiving sites. We present a numerical
study of the information-theoretic capacity of indoor wireless systems that employ MEAs,
based on realistic ray-tracing modeling of propagation in an ofce building. The Shannon
capacity for multi-antenna systems assuming that the transmitter knows the channel (C
n
) is
computed, and is compared to the mutual information achieved with equal power allocation
when the transmitter does not know the channel (MI
n
). Results show that for a xed average
received SNR, the gap between C
n
and MI
n
grows as the number of antennas is increased. We
also investiage the effect of the received SNR on C
n
and MI
n
. We derive expressions for the
asymptotic growth of capacity as the number of antennas grows large and compare these
asymptotic results with the capacities computed using simulated channel responses. The
results reported in this paper serve as a preliminary exploration of MEA systems, and we
anticipate obtaining different capacity estimates for different environments or when actual
measurements are made.
This work was supported by the University of California MICRO Program.
CONTENTS
I. Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
II. Channel Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
III. Information-Theoretic Results . . . . . . . . . . . . . . . . . . . . . 4
A. Capacity Without Channel Knowledge at the Transmitter . . . . . . . . . . . . . . . . 4
B. Capacity With Perfect Channel Knowledge at the Transmitter . . . . . . . . . . . . . . 4
C. Asymptotic Behavior of Capacity . . . . . . . . . . . . . . . . . . . . . . . . . 6
1. With Perfect Channel Knowledge at the Transmitter . . . . . . . . . . 7
2. Without Channel Knowledge at the Transmitter . . . . . . . . . . . 8
IV. Numerical Studies Using WiSE. . . . . . . . . . . . . . . . . . . . . 9
A. WiSE System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1. Channel Outage . . . . . . . . . . . . . . . . . . 10
2. Received SNR . . . . . . . . . . . . . . . . . . . 11
B. Simulation Results and Discussions . . . . . . . . . . . . . . . . . . . . . . . . 11
1. Capacity Improvement due to Water Filling . . . . . . . . . . . 12
2. Asymptotic Behavior of Capacities . . . . . . . . . . . . . . 14
V. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
VI. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . 17
VII. References . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
VIII. Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
IX. Figures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Appendix A: Capacity for Channels with Frequency-Selective Fading. . . . . . 29
Capacity of Multi-Antenna Array Systems 1 of 29
I. Introduction
In order to meet the competitive demand for higher and higher bit rates in wireless local-area
networks (LANs), researchers have explored the utilization of multiple-element arrays (MEAs) at
both transmitting and receiving sites. Space diversity has long been known to improve reception of
signals for wireless systems. Numerous studies have considered using multiple receivers to combat
multipath fading of the desired signal, or to suppress interfering signals [1], [2]. Using diversity at
both transmitters and receivers promises a huge capacity gain because of the additional spatial
degrees of freedom it affords. One fundamental question arises: how many bits per second per Hertz
can one transmit from a transmitting MEA to a receiving MEA given a specic propagation environ-
ment?
Foschini and Gans reported information-theoretic MEA capacities in [3], [4] for a narrowband
Rayleigh fading environment, with i. i. d. channel responses between each antenna pair. They assume
that the transmitter does not know the channel, and equal number of antennas, n, are used at both
transmitter and receiver. Foschini and Gans showed that for large n, the capacity increases linearly
with n in this case. Our work seeks to determine whether the capacity will further increase if the
transmitter knows the channel, and if so, is the increase signicant enough to justify the additional
complexity required in signal processing and feedback from the receiver to the transmitter. In addi-
tion, we also investigate whether the linear capacity growth found in [4] can be retained in a more
realistic propagation environment, in which the transmitted and received signals are correlated or
when there is a dominant line-of-sight (LOS) signal.
We only consider a point-to-point link with additive Gaussian noise between the transmitter and
the receiver, assuming there is no co-channel (multi-user) interference. We use the WiSE [5] simula-
tion tool to model the indoor channel from a base station to receivers located in different rooms of an
ofce building. A 5.2-GHz carrier is launched from transmitting antennas, and the amplitudes, delays
and phase angles of rays impinging on the receiving antennas are recorded. Similar studies have been
reported by Foschini and Valenzuela for MEA systems operating at 5.2 GHz [6].
The remainder of this paper is organized as follows. In Section II, we model the channel as a mul-
tiple-input-multiple-output (MIMO) system with at frequency response. Using this mathematical
2 of 29 Channel Model
model in Section III, we present information-theoretic results for the capacity of MEA systems and
analyze its asymptotic behavior as the number of antennas grows large. In Section IV, we present
capacity estimates for the simulated channels and discuss the discrepancies between these results and
the asymptotic capacities predicted by theory. We briey describe how WiSE is used to represent the
indoor propagation environment that our numerical study is based on. We also include details about
placements of transmitting and receiving MEA, arrangement of antennas in an array, and basic
assumptions for the antenna elements. Then conclusions are presented in Section V.
II. Channel Model
The following notation will be used throughout the paper: for vector transpose, for transpose
conjugate, for the identity matrix, E[
.
] for expectation,
*
for convolution, and underline for
vectors.
We consider a single-user, point-to-point communication channel with n transmitting antennas
and n receiving antennas. The transmitted signal is , an vector
whose jth component represents the signal transmitted by the jth antenna. Similarly, the received
signal and received noise are represented by vectors, and , respectively, where
and represent the signal and noise received at the ith antenna. The complex baseband impulse
response between antenna i and antenna j is represented by , for i = 1, 2, ..., n and j = 1, 2, ..., n.
Assuming that the channel is linear and time-invariant, the transmitted and received vectors are
related by:
(1)
We assume the communication bandwidth, W, that we consider throughout this report is narrow
enough that the channel frequency characteristic can be treated as at over frequency. For this to be a
reasonable approximation, the communication bandwidth must be much less than the coherent band-
width. We have determined from WiSE that, for our channels, the maximum delay spread
1
is 24 ns.
The coherence bandwidth is approximately the reciprocal of the multipath delay spread, which is
41.7 MHz. Therefore the frequency response can indeed be considered at for a communication
1.
Delay spread here refers to the difference, maximum delay - minimum delay, of strong rays that arrive at the
receiver with power above a chosen threshold.
'

I
n
n n
X t ( ) x
1
t ( ) x
2
t ( ) x
T
t ( ) , , , [ ]' = n 1
n 1 Y t ( ) Z t ( ) y
i
t ( )
z
i
t ( )
h
ij
t ( )
Y t ( ) h t ( )
*
X t ( ) Z t ( ) + =
Capacity of Multi-Antenna Array Systems 3 of 29
bandwidth much less than 40 MHz, or a symbol rate much less than 80 Mbaud/s, assuming zero
excess bandwidth.
For the remaining analysis and discussions, we use the discrete-time equivalent model and (1) can
be simplied to:
(2)
where and whose entry H
ij
, which is the discrete time sample of h
ij
(t), rep-
resents the fading path gain from antenna j to antenna i.
We further assume that:
The symbol duration is short compared to the coherence time
1
of the channel. Hence, the
channel response H is xed throughout the time scale that we consider.
The total radiated average power (sum over all transmitting antennas) is P
tot,
regardless of the
number of transmitting antennas n.
The noise vector, Z, is an additive white complex Gaussian random vector, which means its
components, for i = 1, 2, ..., n, are i. i. d. circularly symmetric complex Gaussian random
variables with variance where W is the signal bandwidth.
We consider the following two cases:
1. H is known only to the receiver but not the transmitter. Power is distributed equally over all
transmitting antennas in this case.
2. H is known at the transmitter and receiver, so that power can be allocated appropriately to
maximize the achievable rate over the channel.
In this work, we treat H as quasi-static. Since H is xed for the whole duration of communication,
capacity can be computed for each realization of H without time averaging. On the other hand, H
changes if the receiver is moved from one place to the other, i.e. H varies over a much larger time
scale than a specic communication session when the receiver remains xed on the same location.
Therefore, the associated capacity and mutual information for each specic realization of H can be
viewed as random variables that depend on the location of transmitting and receiving MEA.
1.
Coherence time of the channel refers to the duration for which the channel response remains constant.
Y HX Z + =
X Y Z , , C
n
H C
n n

Z
i
E Z
i
2
[ ] N
0
W =
4 of 29 Information-Theoretic Results
III. Information-Theoretic Results
In this section, we state the generalized formula for capacity for different cases. The transmitting
MEA sends the signal vector X, which consists of symbols chosen from some alphabet set. The recep-
tion is subject to additive white Gaussian noise. Channel capacity is dened as the highest rate at
which information can be sent with arbitrarily low probability of error (Chapter 8, [7]).
For the case with one antenna, n = 1, the Shannon capacity is:
bps/Hz. (3)
In the high-SNR regime, each 3-dB increase of P
tot
/N
0
W yields a capacity increase of 1 bps/Hz.
A. Capacity Without Channel Knowledge at the Transmitter
Here, we assume that the transmitter does not know the channel, and equal power is radiated from
each transmitting antenna. The mutual information in this case is denoted by MI
n
(when n transmit-
ting and n receiving antennas are used), and has been derived in [3] as:
bps/Hz, (4)
where H is an matrix that represents the channel as in (2).
B. Capacity With Perfect Channel Knowledge at the Transmitter
In this section, we derive the capacity for MIMO channel assuming the transmitter has perfect
knowledge about the channel. This assumption is reasonable if the channel is changing very slowly,
so that it can be tracked by the transmitter through feedback from the receiver. With this knowledge of
the channel, total transmit power can be allocated in the most efcient way over the different trans-
mitting antennas to achieve the highest possible bit rate, and we refer to this as the optimal power
allocation. Based on the model in Section II and denitions in [7], the MEA capacity with optimal
power allocation, C
n
, can be derived as follows:
bps/Hz, (5)
subject to the average power constraints:
C P
tot
H , ( )
2
1
P
tot
H
2
N
0
W
------------------- +
,

_
log =
MI
n
P
tot
H , ( )
2
det log I
n
P
tot
nN
0
W
----------------
,
_
HH

+ =
n n
C
n
P
tot
H , ( )
max
Q 2
det log I
n
HQH

N
0
W
---------------- + =
Capacity of Multi-Antenna Array Systems 5 of 29
, (6)
where Q is the covariance matrix of X, and P
tot
is the total average transmit
power.
By singular value decomposition, , where and are unitary matrices.
is diagonal and has non-negative entries, which are the square roots of the eigenvalues of . We
can rewrite (2) as
. (7a)
If we let , , and , (7a) becomes
. (7b)
Since is unitary, is just a rotation of in n-dimensional space, therefore has the same distri-
bution as , i.e. is also a white complex Gaussian random vector.
We dene , which is related to Q by:
. (7c)
Since unitary operators preserve the inner product, . Hence, (5) and (6) can be simpli-
ed to:
, (8a)
and
. (8b)
To maximize the mutual information to achieve capacity, we need to choose to be independent
complex Gaussian ([8]), i.e. should be diagonal. The diagonal entries (the variances of ) are
chosen such that the average power constraint (8b) is satised. The total capacity can be expressed as:
tr Q ( ) E X
i
2
[ ]
i 1 =
n

= P
tot

n n Q E XX

[ ] =
H UDV

= U V D R
n n

HH

Y UDV

X Z + =
Y

Y = X

X = Z

Z =
Y

DX

+ =
U

Z Z

Z Z

E X

[ ] =
Q

QV =
tr Q

( ) tr Q ( ) =
C
n
P
tot
H , ( )
max
Q

2
det log I
n
D
2
Q

N
0
W
------------ - + =
tr Q

( ) P
tot

i
Q

ii
X

i
6 of 29 Information-Theoretic Results
, (9)
subject to the same power constraint (8b). This optimization problem can be solved by using
Lagrange multipliers, which gives us the following optimum solutions:
, (10a)
for i = 1, 2,..., n where satises
. (10b)
Therefore,
. (11)
The optimal solution, (10a) and (10b), is analogous to the water-lling solutions for parallel Gaussian
channels discussed by Cover and Thomas in [7]. Intuitively, the above equation suggests that the orig-
inal channel can virtually be decomposed into n parallel independent sub-channels, and we allot
power to the channels with higher SNRs (smaller N
0
W/
i
). Here, is the water level that marks the
height of the power that is poured into the water vessel formed by the function N
0
W/
i
, i = 1, 2, ...,
n. When the available power P
tot
is increased, some of the power will be allocated to the noisier chan-
nels. Each of these sub-channels contributes to the total capacity through . If
, (12)
we say that this sub-channel provides an effective mode of transmission and the corresponding
i
is
called a strong eigenmode. If all the eigenvalues are much larger than 1/, the capacity is a sum of n
terms of a similar magnitude.
C. Asymptotic Behavior of Capacity
We investigate the growth of capacity of MEA system as the number of antennas grows large for
two cases: (a) when the transmitter knows the channel and (b) when it does not. The capacity, MI
n
and
C
n
P
tot
H , ( )
max
Q

ii
2
log 1

i
Q

ii
N
0
W
------------ - +
i 1 =
n

=
Q

ii

N
0
W

i
------------
,
_
+
=

i
----
,
_
+
i

P
tot
=
C
n
P
tot
H , ( )
2
log
i
( )
+
i 1 =
n

=
2
log
i
( )
+

i
1
Capacity of Multi-Antenna Array Systems 7 of 29
C
n
are random variables that depend on the specic realization of H. The channel is described by
equation (2), and we assume that Z is complex, zero mean white Gaussian noise vector with variance
N
0
W, and H has i. i. d. complex entries with variance . All the assumptions in
Section II hold. For simplicity, when n antennas are used, we denote MI
n
(P
tot
, H) and C
n
(P
tot
, H) as
MI
n
and C
n
respectively.
1. With Perfect Channel Knowledge at the Transmitter
For a given H, the capacity is given by the water-lling solution (9):
(13a)
where satises
(13b)
and
i
are the eigenvalues of , which are random variables that depend on the realization of H.
For each n, let F
n
be the empirical distribution of
i
, i.e., the fraction of
i
less than or equal to for
the case with n antennas:
.
Note that the capacity depends on H only through the empricial distribution of
i
, F
n
(). The
asymptotic properties of the random variable C
n
depends on how the distribution F
n
behaves as n
approaches innity. The following theorem is the result of the work by Marcenko, Yin, Silverstein et
al in [9]-[11].
Theorem 1. Dene G
n
():= F
n
(n). Then, almost surely, G
n
converges to a nonrandom distri-
bution G*, which has a density given by:
E H
ij
E H
ij
( )
2

2
=
C
n
P
tot
H , ( )
2
log
i
( ) ( )
+
i 1 =
n

i
----
,
_
+
i

P
tot
=
HH

F
n
( )
1
n
--- i:
i
( ) { } =
g

( )
1

---
1

---
1
4
--- 0 4
0 otherwise.

'

=
8 of 29 Information-Theoretic Results
The scaling by n in the denition of F
n
means that the
i
are growing as order n. After rescaling,
the distribution converges to a deterministic limiting distribution, i.e. for large n, the empirical distri-
bution of the singular values looks similar for almost all realization of H. Using this theorem, we get
the following result:
Proposition 1.1. Let . Almost surely,
,
where
(14a)
and satises
. (14b)
This indicates that C
n
scales like nC*.
2. Without Channel Knowledge at the Transmitter
Let us now consider the case when the transmitter does not have any information about the
channel, and always allocates an equal power P
tot
/n to each of the transmitting antennas. The mutual
information under equal power allocation is
, (15)
where {
i
} are the eigenvalues of . Using Theorem 1, we can prove the following proposition.
Proposition 1.2. Again let . With almost sure convergence,
,
where

2
P
tot
N
0
W
-------------- =
C
n
P
tot
H , ( )
n
---------------------------- C

( )
C

( )
2
log ( ) ( )
+
g

( ) d
0
4

---
,
_
+
g

( ) d
0
4

=
MI
n
2
log 1
P
tot
nN
0
W
----------------
i
+
,
_
i 1 =
n

=
HH

2
P
tot
N
0
W
-------------- =
MI
n
P
tot
H , ( )
n
------------------------------- MI

( )
Capacity of Multi-Antenna Array Systems 9 of 29
. (16)
With the above two propositions, we get
almost surely.
The above ratio quanties the performance gain achieved asymptotically by water-lling, and
depends only on . Using LHopitals rule, it can be shown that at low SNR,
,
while at high SNR,
.
In the next section, we estimate the average capacities C
n
and MI
n
of simulated channel responses
and study how they grow at large n. The simulation results are then compared to the asymptotic
results nC* and nMI* derived above.
IV. Numerical Studies Using WiSE
A. WiSE System Model
We use the experimentally based Wireless System Engineering (WiSE) ray tracing simulator to
generate realistic realizations of the channel matrix H for indoor wireless propagation. C
n
and MI
n
are
then computed for each realization of H. WiSE has been developed by Steve Fortune et al at Bell Lab-
oratories [5] to model radio propagation in buildings. The predicted baseband channel impulse
response is of the form:
, (17)
MI

( )
2
log 1 + ( ) ( )
+
g

( ) d
0
4

=
C
n
MI
n
----------
C

MI

----------
C

MI

----------
P
tot
0
lim 4 =
C

MI

----------
P
tot

lim 1 =
h t ( )
k
e
i
k
t
k
( )
k 0 =
M

=
10 of 29 Numerical Studies Using WiSE
where is the real positive gain, the associated phase shift, and the propagation delay of the
kth ray. M is the total number of rays and is the unit impulse function. We can compute the fre-
quency response as:
, (18)
As long as the communication bandwidth W is narrower than the coherence bandwidth, H(f) can be
considered constant over the band of interest. For our analysis, we evaluated H(f) at the carrier fre-
quency, f
c
.
In our work, we model the rst oor of a two-oor ofce building at Crawford Hill, New Jersey
(see Fig. 1) in WiSE. We consider the case of multiple omnidirectional transmitting and receiving
antennas, with a 5.2-GHz carrier frequency. For all the simulations, we place the transmitting MEA
on the rst oor ceiling near the middle of the ofce building (see Fig. 1). Antennas are arranged in
square grids within horizontal planes at both the transmitting and receiving sites. The separation
between antenna elements is the same at both the transmitting and receiving MEAs. The antenna
spacing is denoted as d and we consider d = 0.5 (2.9 cm). The location of the transmitter is xed.
Receiving antenna-arrays are placed with random rotations on the horizontal plane at 1000 randomly
chosen positions in each of three rooms: L117 (closest to the transmitter), L147 (at intermediate dis-
tance) and L175 (furthest from the transmitter).
H varies for different receiver locations. When we take expectation with respect to different real-
izations of H, we mean taking the ensemble average over the 1000 sample receiver locations. The
capacities with and without water lling, C
n
and MI
n
, are computed for different n. Since the channel
changes as the receiver is moved around, the capacity is treated as a random variable. The values of
C
n
and MI
n
vary for different receive locations. The results will be presented in terms of complemen-
tary cumulative distribution functions (CCDFs). The averages, C
n
and MI
n
are then computed.
1. Channel Outage
As mentioned before, capacity is dependent on the channel realization and hence might change
considerably when the receiver is moved from one place to another. The communication is considered

k

k

k
t ( )
H f ( )
k
e
i
k
e
i2f
k

k 0 =
M

=
Capacity of Multi-Antenna Array Systems 11 of 29
successful only if the capacity is above a certain threshold. When this is not met, we say that a
channel outage has occurred. Outage is a probabilistic event and we are interested in its tail distribu-
tion. The gures of merit that are of greatest interest are the values of C
n
and MI
n
that can be attained
with probability 1 - , denoted as C
n

and MI
n

. For example, P (C
n
< C
n

) = . Here we consider =
0.05 and call this 5 % channel outage. From the CCDFs that we obtain from WiSE, we determine the
capacities at 5 % outage probability, C
n
0.05
and MI
n
0.05
.
2. Received SNR
We assume the fading path gains, H
ij
, for i, j = 1, 2, ..., n, are identically distributed with the same
variance v
2
. We assume that v
2
is the same for all fading gain H
ij
for the receiving antennas that are
co-located in a specic ofce room. However, v
2
varies when the receiving MEAs are moved into
another ofce room, for example, v
2
is different between room L117 and L147. Within each room, we
estimate , from the channel response H generated using WiSE by averaging
over 1000 channel realizations, and over all possible antenna pairs, j to i.
The average received SNR is dened to be
, (19)
where v
2
is estimated as explained above.
We assume that should be high enough for low-error-rate communication. The appropriate
range is 18-22 dB, due to the practical constraints on A/D converters that are available with current
technology. For all our simulations, we assume W to be 10 MHz, and N
0
to be -170 dBm/Hz
1
, giving
a total noise variance N
0
W of -100.8 dBm. P
tot
is a parameter that we can vary in our experiments in
order to vary the received SNR .
B. Simulation Results and Discussions
The larger the separation between the transmitter and the receiver, the smaller the values of
achievable capacities C
n
and MI
n
. Fig. 2 illustrates this phenomenon. The three pairs of curves are the
CCDFs of C
n
and MI
n
in each of the three ofce rooms for the case when n = 4 and the antenna ele-
1.
Typical two sided power spectral density of thermal noise at 300 K (room temperature) for a receiver that is
modeled as a 50 ohm resistance is -170.8dBm/Hz.
v
2
E H
ij
E H
ij
[ ]
2
[ ] =

P
tot

2

N
0
W
------------------ =
12 of 29 Numerical Studies Using WiSE
ments are separated by d = /2. We x the transmit power P
tot
to be 20 dBm, so that P
tot
/N
0
=
120.8 dB for all three ofce rooms. The average received SNR, , depends on v
2
, which is different
for each ofce room. From our empirical data, = 17 dB when receivers are placed in the furthest
room L175, = 38 dB for L147 and = 55.7 dB for L117. At 5 % outage, MI
n
0.05
is 60.7 bps/Hz in
L117, 18.9 bps/Hz in L147 and 4.6 bps/Hz in L175. For n = 4, there are 4 distinct eigenvalues (
i
) for
. The smaller the separation between the transmitter and receiver, the more strong eigenmodes
exists for effective communication, i.e. the larger the fraction of
i
that satisfy the condition
. Similarly, when optimal power allocation is employed when the separation is smaller, the
higher the proportion of the
i
that satisfy condition (12). Therefore there are more strong sub-chan-
nels that contribute to the total capacity in room L117 compared to other rooms. C
n
0.05
is 60.7 bps/Hz
in L117, followed by 19.0 bps/Hz in L147 and 6.3 bps/Hz in L175. With a 10-MHz bandwidth, one
can expect capacities in the range of 607 Mbps in L117. Such huge capacities might be too high to be
fully exploited in a practical implementation with current technology. Nevertheless, the information-
theoretic capacity serves as the ultimate limit for the achievable bit rate.
1. Capacity Improvement due to Water Filling
The advantage of having channel knowledge at the transmitter so that water lling can be
employed is illustrated in Fig. 2 by the horizontal gap between the CCDFs of MI
n
(solid curves) and
C
n
. The gain in capacity is more apparent in L175 where there are fewer strong eigenmodes for effec-
tive transmission, which explains the wider gap between the solid and dashed curve for L175 com-
pared to the gap for L147 or L117. When all
i
are equally strong eigenmodes, equal power allocation
does as well as the water-lling solutions, resulting in little difference between MI
n
and C
n
(as in
L117). On the other hand, in L175 where only a few
i
are strong eigenmodes, water lling solution
allows us to pour more power into the strong sub-channels. Therefore the strong sub-channels will
contribute more to the total capacity C
n
. If we allocate power equally over all antennas, some power
will be wasted in the weak sub-channels. Therefore, the improvement of C
n
over MI
n
is more signi-
cant in this case. The ratio of improvement at 5 % outage probability C
n
0.05
/MI
n
0.05
, depends on the
SNR level, number of antennas n, and location of the receivers. For example, with = 38 dB and n =
4, C
n
0.05
/MI
n
0.05
= 1.7, 1.1 and 1.0 for L175, L147, and L117, respectively.
HH

P
tot

i
N
0
W
-------------- 1
Capacity of Multi-Antenna Array Systems 13 of 29
The difference between C
n
and MI
n
is dependent on the distribution of
i
, the average received
SNR, , as well as the number of antennas n. To separate the effect that the distribution of
i
might
have, we only consider one ofce room, L147, in the remaining simulation results and discuss the
effect of and n on C
n
, MI
n
, C
n
0.05
, MI
n
0.05
, and C
n
0.05
/MI
n
0.05
.
The CCDFs for MI
n
and C
n
in room L147 are plotted for different antenna array sizes: n = 1, 4, 9,
16, 25 and 36 at = 38 dB in Fig. 3. The rightward shift of the curves shows that the capacity
increases with the number of antennas, n, because space diversity provides additional spatial degrees
of freedom for transmission. At 5 % outage, using a single antenna yields MI1
0.05
= 8.63 bps/Hz
while MEA system achieves value up to 19 bps/Hz when n = 4 and 33.8 bps/Hz when n = 9. Note how
the gap between C
n
and MI
n
(dashed and solid lines, respectively) grows with the array size. MI
n
0.05
=
8.63 bps/Hz when n = 1, while C
n
0.05
= 8.86 bps/Hz. This improvement is negligible. However for a
system with n = 4, the difference between C
n
0.05
and MI
n
0.05
is about 1 bps/Hz, which is more than a
5 % improvement. The difference between C
n
0.05
and MI
n
0.05
keeps increasing as the number of
antennas increases, and reaches 11.32 % for the case n = 36. Table 1 summarizes the improvement in
capacity when water lling is employed as compared to the case of using equal power allocation at
the transmitter. The gain is computed for 5 % outage probability.
In Fig. 4(a), C
n
0.05
and MI
n
0.05
are shown by dashed and solid lines, respectively for receivers in
L147 at various values of SNR . When is small, there are few strong eigenmodes. Knowing the
channel allows us to allocate power more efciently to stronger subchannels and therefore achieve
higher capacity as compared to equal power distribution over all subchannels. As increases, the dif-
ference between C
n
and MI
n
decreases.
The corresponding water-lling gain, C
n
0.05
/MI
n
0.05
is plotted in Fig. 5(a). For n = 4, the ratio
decreases from 3 at = -10 dB to 1 at = 50 dB. We observe that we need lower average received
SNR to achieve the same water-lling gain when the number of antennas increases. This is surprising
because the asymptotic results from Section III-C predicts that C
n
and MI
n
will scale as nC* and
nMI* respectively. Therefore, one would expect that the ratio will converge to a constant as shown in
Section III-C, but this is not the case in Fig. 5(a). To investigate this further, C
n
/MI
n
is plotted in Fig.
5(b). At low received SNR, the three solid curves (for n = 4, 9 and 16) are closer to one another in Fig.
14 of 29 Numerical Studies Using WiSE
5(b) compared to the corresponding curves in Fig. 5(a). We conjecture that this is because the rate of
convergence for C
n
0.05
and MI
n
0.05
(at 5 % outage probability) to the asymptotic value is lower than
the rate of convergence for C
n
and MI
n
2. Asymptotic Behavior of Capacities
We also study the behavior of C
n
and MI
n
as n grows large using channel matrces generated using
WiSE. As shown in Fig. 3, the rightward shifts of CCDFs for MI
n
and C
n
as n increases indicates the
capacity gain achieved by using antenna arrays. The theory of Section III-C predicts that as n grows
large, the mutual information and capacity of the MIMO channel will scale like nMI* and nC*,
respectively. Here, we estimate the variance v
2
from the simulated channels to compute nMI* and
nC*. We note that the asymptotic results of Section III-C, nMI* and nC*, assumes i. i. d. channel
responses, while the channels used to estimate MI
n
and C
n
are subjtect to correlation between the
entries of H.
The average capacity C
n
increases when the antenna separation d is increased. The C
n
computed
from simulated channels are plotted in Fig. 6(a) as a function of n for different d = 0.5, 1 and 5 for
high average received SNR, = 40 dB. When the antenna spacing is equal to 0.5, there is a huge gap
between the C
n
and the nC* predicted from the asymptotic theory of Section III-C. C
n
obtained is
22.8 % of nC* for n = 1. The fraction of predicted asymptotic capacity achieved in simulated chan-
nels increases as n increases. In this case, C
n
reaches 54 %, 59 % and 64 % of nC* for n = 4, 9, 16
respectively. The decrease in capacity when d is small is due to the correlation between the fading
gains, H
ij
, which is not captured in the theoretical model used to derive asymptotic results, which
assumes H
ij
to be i. i. d.. As the antenna spacing is increased to 1 and 5, correlation between
antenna pairs is reduced and therefore C
n
more closely approaches nC*, as shown in Fig. 6(a). For d =
5, the simulated channels have average capacity C
n
that reaches 70 %, 84 % and 89 % of the pre-
dicted limit nC* for n = 4, 9 and 16, respectively.
Fig. 6(b) shows an analogous comparison between MI
n
and nMI*. As d is increased from 0.5 to
1 and 5 , MI
n
increases monotonically and more closely approaches nMI* at each value n. For n =
4, 9, 16, MI
n
also reaches 70 %, 84 % and 89 % of nMI*, respectively.
Capacity of Multi-Antenna Array Systems 15 of 29
In the limit of large n, C
n
/n and MI
n
/n converge almost surely to C* and MI* according to our
asymptotic results. We try to study this asymptotic behavior of the capacities computed using WiSE
simulated channels in the case when is large. The empirical probability density functions (PDFs) of
C
n
/n are plotted for n = 4, 9 and 16 with d = 0.5 in Fig. 7(a) and d = 5 in Fig. 7(b). As n increases,
the PDF of C
n
/n becomes narrower with higher peak value, i.e. C
n
/n becomes less random. In the
limit of large n, we expect the PDF of C
n
/n to converge to an impulse function centered at the value
C*. This illustrates the almost-surely convergence of C
n
/n to C*. Note that when d is large (5), the
PDF of C
n
/n grows narrower and taller much faster than when d is small (0.5). This indicates that the
rate of convergence is higher when d is larger, i.e. when the correlation between H
ij
is less. Further
analysis is needed to understand how correlation affects the validity of the aymptotic results in Sec-
tion III-C.
V. Conclusions
There is a potentially huge gain in capacity when multiple antennas are employed for an indoor
wireless transmission system compared to a single antenna system. The achievable rates, C
n
and MI
n
,
depend on the propagation environment, and grow with the increase of average received SNR ,
number of antennas n, and antenna separation d. With perfect channel knowledge at the transmitter,
power can be allocated efciently over different transmitting antennas to achieve better rates, C
n
. The
water lling gain is more signicant when there are fewer strong eigenmodes (71.4 % for n = 36 in
room L175 compared to 0.06 % in L117). The benet of water-lling is also more obvious at low
average received SNR, e.g. when = -10 dB, water lling yields 3.5 times larger capacity at 5 %
outage probability (C
n
0.05
/MI
n
0.05
= 3.5), but at = 50 dB, water lling gain is negligible
( ).
Assuming i. i. d. channels between different antenna pairs, theoretical analysis shows that the
capacity grows linearly asymptotically with the number of antennas n, in the limit of large n. How-
ever, in a more realistic propagation environment, correlation does exist between antenna pairs and
causes a smaller rate of growth in capacity when n grows large. Our simulation results show that for
0.5 antenna spacing, the simulated average capacity C
n
is only 64 % of the predicted value nC* for
n = 16. When the antenna spacing is increased, we see more agreement between C
n
and nC*. Indeed
C
n
0.05
MI
n
0.05
1
16 of 29 Conclusions
with d = 5 , the C
n
achieved at = 40 dB is 89 % of nC* when n = 16. To utilize MEA systems in the
most efcient way, we need to choose appropriate antenna spacing d, transmit power P
tot
, and decide
on the best arrangement of antenna elements in an array as well as the actual locations of transmitting
and receiving MEAs.
Capacity of Multi-Antenna Array Systems 17 of 29
VI. Acknowledgments
I would like to express my sincere appreciation to Professor Joseph Kahn and Professor David
Tse for their guidance and encouragement during this work. The opportunity of working closely with
them have been a stimulating and rewarding experience in my Master program.
I am grateful to Reinaldo Valenzuela, Jerry Foschini, Jonathan Ling and Dmitry Chizhik for
allowing us to use their WiSE simulation tools, and for their valuable advice & suggestions. Discus-
sion with Jack Salz is equally enlightening. In addition, I truly appreciate my EE comrade, Da-Shan
Shiu for patiently sharing with me his broad range of knowledge. His strong faith in me is a valuable
source of motivation. The many heated discussions and arguments we had certainly help sharpen my
critical thinking. I am also thankful to Ma Yi who has tutored me in EE226 during my rst semester at
Berkeley. Surviving the course keeps me in graduate school, and it also helps to build the background
knowledge that I need for further research work.
Special thanks are dedicated to Professor David Goodman, Professor Roy Yates and Professor
Christopher Rose at Rutgers University, who have motivated me to come to graduate school. Their
advice will guide me a long way in life, both in college and after graduation.
I am extremely fortunate to be surrounded by friends who are caring and understanding. They
have become my extended family at Berkeley. I want to take this opportunity to thank my soul mates,
Ching Shang and Fang-Pei Chen, for always being there for me. I cherish every moment of their com-
pany. I would like to thank Jocelyn Nee and Janice Hudgings in particular for their emotional support,
to Gene Cheung, Theen Theen Tan, Kean Hock Yeap, Niny Khor, Amy Aweisbin and Tz-Yin Lin for
their encouragement. I also wish to thank David Lee for helping me with all sorts of computer prob-
lems and thank all the group members for sharing their expertise and humor all these past years.
I wish to thank the U.C. Berkeley EECS department, Tau Beta Pi and Schumberger Coorporation,
whose nancial support made my graduate program possible.
Last but not least, I wont have accomplished this without the love and support of my family
members, who are far away in distance but close at heart. They are always available with love and
patience, through all my past twenty four years. In my heart, my family is the greatest masterpeice on
earth.
18 of 29 References
VII. References
[1] W. Jakes Jr., Microwave Mobile Communications, New Wiley,1974.
[2] J. Winters, Optimum Combining for Indoor Radio Systems with Multiple Users, IEEE Trans.
Commun., vol. com-35, no. 11, pp. 1222-1230, Nov. 1987.
[3] G. J. Foschini and M. J. Gans, On Limits of Wireless Communication in a Fading Environment
When Using Multiple Antennas, accepted for publication in Wireless Personal Communications.
[4] G. J. Foschini and M. J. Gans, Capacity When Using Diversity At Transmit And Receive Sites and
The Rayleigh-Faded Matrix Channel Is Unknown At The Transmitter, WINLAB Workshop on Wire-
less Information Network, March 20-21, New Brunswick, NJ.
[5] S. J. Fortune, D. H. Gay, B. W. Kernighan, O. Landron, R. A. Valenzuela and M. H. Wright, WiSE
design of Indoor Wireless Systems: Practical Computation and Optimization, IEEE Computational
Science and Engineering, Mar. 1995.
[6] G. J. Foschini and R. A. Valenzuela, Initial Estimation of Communication Efciency of Indoor
Wireless Channels, Wireless Network, vol.3, no. 2, pp. 141-54, 1997.
[7] T. M. Cover and J. A. Thomas, Elementary of Information Theory, John Wiley & Sons, New York,
1991.
[8] I. E. Telatar, Capacity of Multi-antenna Gaussian Channels, AT&T Bell Laboratories, Murray Hill,
NJ.
[9] V.A. Marcenko, and L.A. Pastur, Distribution of Eigenvalues for Some Sets of Random Matrices,
USSR-Sb. 1, 457-483, 1967.
[10] Y.Q. Yin, Limiting Spectral Distribution for A Class of Random Matrices, Journal of Multivariate
Analysis, 20: 50-68, 1986.
[11] Jack Silverstein, Strong Convergence of the Empirical Distribution of Eigenvalues of large Dimen-
sional Random Matrices, Journal of Multivariate Analysis, 55(2):331-339, 1995.
Capacity of Multi-Antenna Array Systems 19 of 29
VIII. Tables
number of antennas, n 1 4 9 16 25 36
C
n
0.05
(bps/Hz) 19.14 60.70 126.86 211.51 334.36 459.58
MI
n
0.05
(bps/Hz) 19.14 60.69 126.83 211.40 334.21 459.30
% difference 0.00 0.00 0.02 0.05 0.45 0.06
Table 1a: L117
number of antennas, n 1 4 9 16 25 36
C
n
0.05
(bps/Hz) 8.86 20.02 36.28 56.87 75.36 106.25
MI
n
0.05
(bps/Hz) 8.63 18.93 33.83 52.29 68.23 95.43
% difference 2.65 5.73 7.22 8.75 10.44 11.32
Table 1b: L147
number of antennas, n 1 4 9 16 25 36
C
n
0.05
(bps/Hz) 2.45 6.32 10.01 14.05 20.10 26.23
MI
n
0.05
(bps/Hz) 2.44 4.59 6.65 10.31 12.20 15.30
% difference 0.19 37.78 50.51 36.28 64.80 71.40
Table 1c: L175
Table 1: The percentage difference, (C
n
0.05
- MI
n
0.05
)/MI
n
0.05
x 100 %, for room (a)
L117 (b) L147 and (c) L175. = 38 dB, d = 0.5 .
20 of 29 Figures
IX. Figures
L
1
1
7
L
1
4
7
L
1
7
5
118 meters
1
4

m
e
t
e
r
s
/ 2
Receiving
MEA
Transmitting MEA
x
y
Fig. 1. Floor Plan for the rst oor of Bell Laboratories Building at
Crawford Hill, New Jersey. Receivers with antennas positioned in
square grids are placed randomly at 1000 locations in room L117,
L147 and L175. The transmitting MEA is placed with its adjacent sides
parallel to x-axis and y-axis, respectively. The receiving MEA is placed
with random orientation at each of the sample location.
0 10 20 30 40 50 60 70 80
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Capacity, C
n
& MI
n
(bps/Hz)
P
r
o
b

(
C
n
o
r
M
I
n
>
=

a
b
s
c
i
s
s
a
)
L175
L147
L117
CCDFs for C
n
CCDFs for MI
n
Fig. 2. The CCDFs of C
n
(solid curve) and MI
n
(dashed curve). C
n
is
the capacity with optimum water-lling power allocation, and MI
n
is the
mutual information with at power allocation. P
tot
/N
0
is xed at 120.8
dB and the measured is indicated for the three rooms.
=
17 dB
= 38 dB
55.7 dB =
Capacity of Multi-Antenna Array Systems 21 of 29
Fig. 3. The CCDFs of C
n
(achieved via water pouring) and MI
n
(with
equal power allocation) for L147 with equal number of antennas at
both transmitter and receiver (n= 1, 4, 9, 16, 25 & 36) at received =
20dB.
0 20 40 60 80 100 120 140 160
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Capacity, C
n
& MI
n
(bps/Hz)
P
r
o
b

(
C
n

o
r
M
I
n
>
=

a
b
s
c
i
s
s
a
)
CCDFs for C
n
L147
n = 1 4 9 16 25 36
CCDFs for MI
n
22 of 29 Figures
Fig. 4(a). C
n
0.05
(capacity with optimum water-lling power
allocation) and MI
n
0.05
(mutual information with at power allocation)
at 5 % outage probability for receivers in room L147.
-10 0 10 20 30 40 50
0
9
18
27
36
45
54
63
72
81
90
Average received SNR, (dB)
C
n
0
.
0
5

a
n
d
M
I
n
0
.
0
5
(
b
p
s
/
H
z
)
n =16
n = 9
n = 4
C
n
0.05
MI
n
0.05
Capacity of Multi-Antenna Array Systems 23 of 29
-10 0 10 20 30 40 50
0.5
1
1.5
2
2.5
3
3.5
Average received SNR, (dB)
C
n
0
.
0
5
/
M
I
n
0
.
0
5
n = 4
n = 9
n = 16
C*/MI*
Fig. 5(a). Water-lling gain C
n
0.05
/MI
n
0.05
(solid lines) over varying
average received SNR, , in room L147 for n = 4, 9 and 16. The
reference curve is C*/MI* (dashed line) is predicted by asymptotic
theory in Section III-C.
C
n
0.05
/MI
n
0.05
C*/MI*
24 of 29 Figures
-10 0 10 20 30 40 50
0.5
1
1.5
2
2.5
3
3.5
Average received SNR, (dB)
C
n
/
M
I
n
C
n
/MI
n
C*/MI*
n = 4
n = 9
n = 16
C*/MI*
Fig. 5(b). Water-lling gain C
n
/MI
n
(solid lines) over varying average
received SNR, in room L147 for n = 4, 9 and 16. The reference curve
is C*/MI* (dashed line) is predicted by asymptotic theory in Section III-
C.
Capacity of Multi-Antenna Array Systems 25 of 29
0 2 4 6 8 10 12 14 16
0
20
40
60
80
100
120
140
Number of Antennas, n
C
n
a
n
d
n
C
*

(
b
p
s
/
H
z
)
d

=

0
.
5

d

=

1

d

=

5

n
C
*
= 18.7 dB
nC
*
(asymptotic result)
C
n
(from simulations)
Fig. 6(a). Average C
n
with optimum water-lling power allocation in
room L147 for different antenna spacing as n is increased from 1 to 16.
The reference curve (dashed line) is nC*, which is predicted by
asymptotic theory using variance v
2
computed from WiSE channels.
26 of 29 Figures
Fig. 6(b). Average MI
n
with equal power allocation in room L147 for
different antenna spacing as n is increased from 1 to 16. The reference
curve (dashed line) is nMI
*
.
0 2 4 6 8 10 12 14 16
0
20
40
60
80
100
120
140
Number of Antennas, n
M
I
n
a
n
d
n
M
I
*

(
b
p
s
/
H
z
)
d

=

0
.
5

d

=

1

d

=

5

n
M
I
*
= 18.7 dB
nMI
*
(asymptotic result)
MI
n
(from simulations)
Capacity of Multi-Antenna Array Systems 27 of 29
Fig. 7(a). Empirical probability density function for the normalized
capacity C
n
/n for n = 4, 9 and 16. The reference value is C*, the
constant that C
n
converges to as predicted by the aymptotic theory.
Antenna spacing d = 0.5.
1 2 3 4 5 6 7 8 9
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05
C
n
/n (bps/Hz/antenna)
E
m
p
i
r
i
c
a
l

P
D
F

f
o
r
C
n
/
n
= 18.7 dB
d = 0.5
C* = 7.8
n=4
n=9
n=16
(1)
28 of 29 Figures
Fig. 7(b). Empirical probability density function for the normalized
capacity C
n
/n for n = 4, 9 and 16. The reference value is C*, the
constant that C
n
converges to as predicted by the aymptotic theory.
Antenna spacing d = 5.
1 2 3 4 5 6 7 8 9
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05
C
n
/n (bps/Hz/antenna)
E
m
p
i
r
i
c
a
l

P
D
F

f
o
r
C
n
/
n
= 18.7 dB
d = 5
n = 4
n = 9
n = 16
C* = 7.8
(1)
Capacity of Multi-Antenna Array Systems 29 of 29
Appendix A: Capacity for Channels with Frequency-Selective Fading
We will extend the result to the frequency-selective fading case. Consider where each entry
is a function of frequency. Assume the signal is bandlimited to W Hz. We will approximate the fre-
quency response in piecewise-constant fashion by dividing the total bandwidth into B frequency bins
of width , each of which is narrow enough so that is approximately at within the bin.
Applying SVD separately to each frequency band m = 1, 2, ..., B, we have
.
The capacity per unit bandwidth can be approximated as:
, (20)
where are chosen to meet the power constraints:
, (20a)
. (20b)
Assuming that signals in different frequency bins are orthogonal to one another, the channel
reduces to parallel Gaussian channels that provide BT degrees of freedom in choosing the variances of
the T signals transmitted in B frequency sub-bands.
H f ( )
f H f ( )
H

m ( ) U

m ( )D

m ( )V

m ( ) =
C
n
W
------
C
n
H W P
i
P
tot
, , , ( )
B f
-------------------------------------------- =
1
B
---

i
m ( ) Q

ii
m ( )
N
o
----------------------------------- 1 +
,

_
log
i 1 =
min T R , { }

' ;

m 1 =
B

ii

ii

m ( )
i 1 =
T

m 1 =
B

P
tot

V
ji
m ( )
2
Q

ii

m ( )
i 1 =
T

m 1 =
B

P
j
for j 1 2 3 T , , , , =