You are on page 1of 6

ON THE CAPACITY OF THE MIMO CHANNEL

- A TUTORIAL INTRODUCTION -
Bengt Holter
Norwegian University of Science and Technology
Department of Telecommunications
O.S.Bragstads plass 2B, N-7491 Trondheim, Norway
bholter@tele.ntnu.no
ABSTRACT
Multiple-input multiple-output (MIMO) systems
are today regarded as one of the most promising re-
search areas of wireless communications. This is due
to the fact that a MIMO channel can oer a sig-
nicant capacity gain over a traditional single-input
single-output (SISO) channel. In this paper, a tutor-
ial introduction on the channel capacity of a MIMO
channel will be given.
1. INTRODUCTION
The increase in spectral eciency oered by MIMO
systems is based on the utilization of space (or an-
tenna) diversity at both the transmitter and the re-
ceiver. Due to the utilization of space diversity, MIMO
systems are also referred to as multiple-element an-
tenna systems (MEAs). With a MIMO system, the
data stream from a single user is demultiplexed into
n
T
separate sub-streams. The number n
T
equals the
number of transmit antennas. Each sub-stream is
then encoded into channel symbols. It is common
to impose the same data rate on all transmitters, but
adaptive modulation rate can also be utilized on each
of the sub-streams [1]. The signals are received by n
R
receive antennas.
With this transmission scheme, there is a linear in-
crease in spectral eciency compared to a logarithmic
increase in more traditional systems utilizing receive-
diversity or no diversity. The high spectral ecien-
cies attained by a MIMO system are enabled by the
fact that in a rich scattering environment, the sig-
nals from each individual transmitter appear highly
uncorrelated at each of the receive antennas. When
the signals are conveyed through uncorrelated chan-
nels between the transmitter and receiver, the signals
corresponding to each of the individual transmit an-
tennas have attained dierent spatial signatures. The
receiver can use these dierences in spatial signature
to simultaneously and at the same frequency separate
the signals that originated from dierent transmit an-
tennas.
2. CHANNEL CAPACITY
At the input of a communication system, discrete
source symbols are mapped into a sequence of chan-
nel symbols. The channel symbols are then transmit-
ted/conveyed through a wireless channel that by na-
ture is random. In addition, random noise is added to
the channel symbols. In general, it is possible that two
dierent input sequences may give rise to the same
output sequence, causing dierent input sequences to
be confusable at the output. To avoid this situation,
a non-confusable subset of input sequences must be
chosen so that with a high probability, there is only
one input sequence causing a particular output. It is
then possible to reconstruct all the input sequences at
the output with negligible probability of error.
A measure of how much information that can be
transmitted and received with a negligible probability
of error is called the channel capacity. To determine
this measure of channel potential, assume that a chan-
nel encoder receives a source symbol every T
s
second.
With an optimal source code, the average code length
of all source symbols is equal to the entropy rate of the
source. If S represents the set of all source symbols
and the entropy rate of the source is written as H(S),
the channel encoder will receive on average
H(S)
Ts
in-
formation bits per second.
1
Assume that a channel
codeword leaves the channel encoder every T
c
second.
In order to be able to transmit all the information
from the source, there must be
R =
H(S)T
c
T
s
(1)
1
The entropy rate is a function of the statistical distribution
of the source S. If the source S represent a discrete memoryless
random variable, the entropy rate of the source is equal to the
source entropy, and is dened as H(S) = E[log
b
(p
S
)]. It
is common to use b = 2 and the entropy is then expressed in
information bits per source symbol.
information bits per channel symbol. The number R
is called the information rate of the channel encoder.
The maximum information rate that can be used caus-
ing negligible probability of errors at the output is
called the capacity of the channel. By transmitting
information with rate R, the channel is used every
T
c
seconds. The channel capacity is then measured
in bits per channel use. Assuming that the channel
has bandwidth W, the input and output can be rep-
resented by samples taken T
s
=
1
2W
seconds apart.
With a band-limited channel, the capacity is mea-
sured in information bits per second. It is common
to represent the channel capacity within a unit band-
with of the channel. The channel capacity is then
measured in bits/s/Hz.
It is desirable to design transmission schemes that
exploit the channel capacity as much as possible. Rep-
resenting the input and output of a memoryless wire-
less channel with the random variables X and Y re-
spectively, the channel capacity is dened as [2]
C = max
p(x)
I(X; Y ), (2)
where I(X; Y ) represents the mutual information be-
tween X and Y . Eq.(2) states that the mutual in-
formation is maximized with respect to all possible
transmitter statistical distributions p(x). Mutual in-
formation is a measure of the amount of information
that one random variable contains about another vari-
able. The mutual information between X and Y can
also be written as
I(X; Y ) = H(Y ) H(Y |X), (3)
where H(Y |X) represents the conditional entropy be-
tween the random variables X and Y . The entropy
of a random variable can be described as a measure
of the amount of information required on average to
describe the random variable. It can also described as
a measure of the uncertainty of the random variable.
Due to (3), mutual information can be described as
the reduction in the uncertainty of one random vari-
able due to the knowledge of the other. Note that the
mutual information between X and Y depends on the
properties of the channel (through a channel matrix
H) and the properties of X (through the probability
distribution of X). The channel matrix H used in
the representation of the input/output relations of a
MIMO channel is dened in the next section.
3. SYSTEM MODEL
It is common to represent the input/output relations
of a narrowband, single-user MIMO link by the com-
plex baseband vector notation
y = Hx +n, (4)
where x is the (n
T
1) transmit vector, y is the (n
R

1) receive vector, H is the (n


R
n
T
) channel matrix,
and n is the (n
R
1) additive white Gaussian noise
(AWGN) vector at a given instant in time. Through-
out the paper, it is assumed that the channel matrix
is random and that the receiver has perfect channel
knowledge. It is also assumed that the channel is
memoryless, i.e., for each use of the channel an inde-
pendent realization of H is drawn. This means that
the capacity can be computed as the maximum of the
mutual information as dened in (2). The results are
also valid when H is generated by an ergodic process
because as long as the receiver observes the Hprocess,
only the rst order statistics are needed to determine
the channel capacity [3].
A general entry of the channel matrix is denoted
by {h
ij
}. This represents the complex gain of the
channel between the jth transmitter and the ith re-
ceiver. With a MIMO system consisting of n
T
trans-
mit antennas and n
R
receive antennas, the channel
matrix is written as
H =
_

_
h
11
h
1nT
h
21
h
2nT
.
.
.
.
.
.
.
.
.
h
nR1
h
nRnT
_

_
, (5)
where
h
ij
= + (6)
=
_

2
+
2
e
arctan

(7)
= |h
ij
| e
ij
. (8)
In a rich scattering environment with no line-of-sight
(LOS), the channel gains |h
ij
| are usually Rayleigh
distributed. If and are independent and normal
distributed random variables, then |h
ij
| is a Rayleigh
distributed random variable.
4. SISO CHANNEL CAPACITY
The ergodic (mean) capacity of a random channel
with n
T
= n
R
= 1 and an average transmit power
constraint P
T
can be expressed as [2]
C = E
H
_
max
p(x):PPT
I(X; Y )
_
, (9)
where P is the average power of a single channel code-
word transmitted over the channel and E
H
denotes
the expectation over all channel realizations. Com-
pared to the denition in (2), the capacity of the
channel is now dened as the maximum of the mu-
tual information between the input and the output
over all statistical distributions on the input that sat-
isfy the power constraint. If each channel symbol at
the transmitter is denoted by s, the average power
constraint can be expressed as
P = E
_
|s|
2

P
T
. (10)
Using (9), the ergodic (mean) capacity of a SISO sys-
tem (n
T
= n
R
= 1) with a random complex channel
gain h
11
is given by [4]
C = E
H
_
log
2
_
1 + |h
11
|
2
__
, (11)
where is the average signal-to-noise (SNR) ratio at
the receiver branch. If |h
11
| is Rayleigh, |h
11
|
2
follows
a chi-squared distribution with two degrees of freedom
[5]. Eq.(11) can then be written as [4]
C = E
H
_
log
2
_
1 +
2
2
__
, (12)
where
2
2
is a chi-square distributed random variable
with two degrees of freedom. Figure 1 shows the
0 2 4 6 8 10 12 14 16 18 20
0
1
2
3
4
5
6
SISO capacity
C
a
p
a
c
i
t
y

[
b
i
t
/
s
/
H
z
]
SNR [dB]
Figure 1: Ergodic capacity a Rayleigh fading SISO
channel (dotted line) compared to the Shannon ca-
pacity of a SISO channel (solid line).
Shannon capacity for a gaussian channel (solid line)
and the capacity of a Rayleigh fading channel (dot-
ted line) according to (12). The Rayleigh fading curve
presented in Figure 1 equals the result in [6] (capacity
of a Rayleigh fading channel with optimal power and
rate adaption at the transmitter under the assump-
tion of perfect channel estimation and return channel
free of errors and delay).
5. MIMO CHANNEL CAPACITY
The capacity of a random MIMO channel with power
constraint P
T
can be expressed as
C = E
H
_
max
p(x):tr()PT
I(x; y)
_
, (13)
where = E{xx

}
2
is the covariance matrix of the
transmit signal vector x. The total transmit power
is limited to P
T
, irrespective of the number of trans-
mit antennas. By using (4) and the relationship be-
tween mutual information and entropy, (13) can be
expanded as follows for a given H
I(x; y) = h(y) h(y|x) (14)
= h(y) h(Hx +n|x) (15)
= h(y) h(n|x) (16)
= h(y) h(n), (17)
where h() in this case denotes the dierential entropy
of a continuous random variable. It is assumed that
the transmit vector x and the noise vector n are in-
dependent.
Eq. (17) is maximized when y is gaussian, since
the normal distribution maximizes the entropy for a
given variance [2]. The dierential entropy of a real
gaussian vector y R
n
with zero mean and covari-
ance matrix K is equal to
1
2
log
2
((2e)
n
det K). For
a complex gaussian vector y C
n
, the dierential
entropy is less than or equal to log
2
det(eK), with
equality if and only if y is a circularly symmetric com-
plex Gaussian with E{yy

} = K [3]. Assuming the


optimal gaussian distribution for the transmit vector
x, the covariance matrix of the received complex vec-
tor y is given by
E
_
yy

_
= E
_
(Hx +n) (Hx +n)

_
(18)
= E
_
Hxx

_
+E
_
nn

_
(19)
= HH

+K
n
(20)
= K
d
+K
n
. (21)
The superscript d and n denotes respectively the de-
sired part and the noise part of (21). The maximum
mutual information of a random MIMO channel is
then given by
I = h(y) h(n)
= log
2
_
det
_
e
_
K
d
+K
n
__
log
2
[det (eK
n
)]
= log
2
_
det
_
K
d
+K
n
_
log
2
[det (K
n
)]
= log
2
_
det
_
_
K
d
+K
n
_
(K
n
)
1
__
= log
2
_
det
_
K
d
(K
n
)
1
+I
nR
__
= log
2
_
det
_
HH

(K
n
)
1
+I
nR
__
.
When the transmitter has no knowledge about the
channel, it is optimal to use a uniform power distri-
bution [3]. The transmit covariance matrix is then
given by =
PT
nT
I
nT
. It is also common to assume
2
The superscript

denotes Hermitian transpose
uncorrelated noise in each receiver branch described
by the covariance matrix K
n
=
2
I
nR
. The ergodic
(mean) capacity for a complex AWGN MIMO channel
can then be expressed as [3, 4]
C = E
H
_
log
2
_
det
_
I
nR
+
P
T

2
n
T
HH

___
. (22)
This can also be written as
C = E
H
_
log
2
_
det
_
I
nR
+

n
T
HH

___
, (23)
where =
PT

2
is the average signal-to-noise (SNR)
ratio at each receiver branch. By the law of large
numbers, the term
1
nT
HH

I
nR
as n
T
gets large
and n
R
is xed. Thus the capacity in the limit of large
n
T
is
C = E
H
{n
R
log
2
(1 +)} . (24)
Further analysis of the MIMO channel capacity given
in (23) is possible by diagonalizing the product matrix
HH

either by eigenvalue decomposition or singular


value decomposition. By using eigenvalue decompo-
sition, the matrix product is written as
HH

= EE

, (25)
where E is the eigenvector matrix with orthonormal
columns and is a diagonal matrix with the eigenval-
ues on the main diagonal. Using this notation, (23)
can be written as:
C = E
H
_
log
2
_
det
_
I
nR
+

n
T
EE

___
. (26)
The matrix product HH

can also be described by us-


ing singular value decomposition on the channel ma-
trix H written as
H= UV

, (27)
where U and V are unitary matrices of left and right
singular vectors respectively, and is a diagonal ma-
trix with singular values on the main diagonal.
All elements on the diagonal are zero except for
the rst k elements. The number of non-zero singular
values k equals the rank of the channel matrix. Us-
ing (27) in (23), the MIMO channel capacity can be
written as
C = E
H
_
log
2
_
det
_
I
nR
+

n
T
U

___
. (28)
After diagonalizing the product matrix HH

, the ca-
pacity formulas of the MIMO channel now includes
unitary and diagonal matrices only. It is then easier to
see that the total capacity of a MIMO channel is made
up by the sum of parallel AWGN SISO subchannels.
The number of parallel subchannels is determined by
the rank of the channel matrix. In general, the rank
of the channel matrix is given by
rank(H) = k min{n
T
, n
R
}. (29)
Using (29) together with the fact that the determinant
of a unitary matrix is equal to 1, (26) and (28) can
be expressed respectively as
C = E
H
_
k

i=1
log
2
_
1 +

n
T

i
_
_
(30)
= E
H
_
k

i=1
log
2
_
1 +

n
T

2
i
_
_
. (31)
In (30),
i
are the eigenvalues of the diagonal matrix
and in and (31),
2
i
are the squared singular values
of the diagonal matrix . The maximum capacity of a
MIMO channel is reached in the unrealistic situation
when each of the n
T
transmitted signals is received
by the same set of n
R
antennas without interference.
It can also be described as if each transmitted sig-
nal where received by a separate set receive antennas,
giving a total number of n
T
n
R
receiving antennas.
With optimal combining at the receiver and re-
ceive diversity only (n
T
= 1), the channel capacity
can be expressed as [4]
C = E
H
_
log
2
_
1 +
2
2nR
__
, (32)
where
2
2nR
is a chi-distributed random variable with
2n
R
degrees of freedom. If there are n
T
transmit an-
tennas and optimal combining between n
R
antennas
at the receiver, the capacity can be written as
3
[4]
C = E
H
_
n
T
log
2
_
1 +

n
T

2
2nR
__
. (33)
Eq.(33) represent the upper bound of a Rayleigh fad-
ing MIMO channel. In Figure 2, the Shannon capacity
of a SISO channel is compared to the upper bound of
(33) with n
T
= n
R
= 6. Even though this bound on
the MIMO channel represent a special case, Figure 2
clearly shows the potential of the MIMO technology.
When the channel is known at the transmitter,
the maximum capacity of a MIMO channel can be
achieved by using the water-lling principle [2] on
the transmit covariance matrix. The capacity is then
given by
C = E
H
_
k

i=1
log
2
_
1 +
i

n
T

i
_
_
(34)
= E
H
_
k

i=1
log
2
_
1 +
i

n
T

2
i
_
_
, (35)
3
Assuming the articial case of no interference between the
received n
T
signals.
10 5 0 5 10 15 20 25 30 35 40
0
10
20
30
40
50
60
70
Ergodic capacity of a MIMO fading channel
C
a
p
a
c
i
t
y

[
b
i
t
/
s
/
H
z
]
SNR [dB]
Figure 2: The Shannon capacity of a SISO channel
(dotted line) compared to the ergodic capacity of a
Rayleigh fading MIMO channel (solid line) with n
T
=
n
R
= 6.
where
i
is a scalar, representing the portion of the
available transmit power going into the ith subchan-
nel. The power constraint at the transmitter can be
expressed as

nT
i=1

i
n
T
.
Clearly, with a reduced number of non-zero singu-
lar values in (31) and (35), the capacity of the MIMO
channel will be reduced because of a rank decient
channel matrix. This is the situation when the signals
arriving at the receivers are correlated. Even though
a high channel rank is necessary to obtain high spec-
tral eciency on a MIMO channel, low correlation is
not a guarantee of high capacity [7]. In [8], the ex-
istence of pin-hole channels is demonstrated. Such
channels exhibit low fading correlation between the
antennas at both the receiver and transmitter side,
but the channels still have poor rank properties and
hence low capacity.
6. ANTENNA SELECTION
The MIMO channel capacity has so far been opti-
mized based on the assumption that all transmit and
receive antennas are used at the same time. Recently,
several authors have presented papers on MIMO sys-
tems with either transmit or receive antenna selec-
tion. As seen earlier in this paper, the capacity of the
MIMO channel is reduced with a rank decient chan-
nel matrix. A rank decient channel matrix means
that some columns in the channel matrix are linearly
dependent. When they are linearly dependent, they
can be expressed as a linear combination of the other
columns in the matrix. The information within these
columns is then in some way redundant and is not con-
tributing to the capacity of the channel. The idea of
transmit antenna selection is to improve the capacity
by not using the transmit antennas that correspond
to the linearly dependent columns, but instead redis-
tributing the power among the other antennas. Since
the total number of parallel subchannels in the sums
of eq.(31) and (35) is equal to the rank of the channel
matrix, the optimal choice is to distribute the trans-
mit power on a subset of k transmit antennas that
maximizes the channel capacity. It is shown in [9]
that the optimal choice of k transmit antennas that
maximizes the channel capacity results in a channel
matrix that is full rank. In [10], a computationally ef-
cient, near-optimal search technique for the optimal
subset based on classical waterpouring is described.
In [11], the capacity of MIMO systems with receive
antenna selection is analyzed. With such a reduced-
complexity MIMO scheme, a selection of the L best
antennas of the available n
R
antennas at the reciever
is used. This has the advantage that only L receiver
chains are required compared to n
R
in the full-complexity
scheme. In [11], it is demonstrated through Monte
Carlo simulations that for n
T
= 3 and n
R
= 8, the ca-
pacity of the reduced-complexity scheme is 20bits/s/Hz
compared to 23bits/s/Hz of a full-complexity scheme.
7. OUTAGE CAPACITY
In this paper, the ergodic (mean) capacity has been
used as a measure for the spectral eciency of the
MIMO channel. The capacity under channel ergodic-
ity where in (9) and (13) dened as the average of the
maximal value of the mutual information between the
transmitted and the received signal, where the maxi-
mization was carried out with respect to all possible
transmitter statistical distributions. Another mea-
sure of channel capacity that is frequently used is out-
age capacity. With outage capacity, the channel ca-
pacity is associated to an outage probability. Capac-
ity is treated as a random variable which depends on
the channel instantaneous response and remains con-
stant during the transmission of a nite-length coded
block of information. If the channel capacity falls be-
low the outage capacity, there is no possibility that
the transmitted block of information can be decoded
with no errors, whichever coding scheme is employed.
The probability that the capacity is less than the out-
age capacity denoted by C
outage
is q. This can be
expressed in mathematical terms by
Prob{C C
outage
} = q. (36)
In this case, (36) represents an upper bound due to
fact that there is a nite probability q that the channel
capacity is less than the outage capacity. It can also
be written as a lower bound, representing the case
where there is a nite probability (1 q) that the
channel capacity is higher than C
outage
, i.e.,
Prob{C > C
outage
} = 1 q. (37)
8. SUMMARY
In this paper, a tutorial introduction on the capac-
ity of the MIMO channel has been given. The use
of multiple antennas on both the transmitter and re-
ceiver side of a communication link have shown to
greatly improve the spectral eciency of both xed
and wireless systems. The are many research papers
published on MIMO systems, reecting the percep-
tion that MIMO technology is seen as one of the most
promising research areas of radio communication to-
day.
9. REFERENCES
[1] S. Catreux, P. F. Driessen, L. J. Greenstein, At-
tainable throughput of an interference-limited
multiple-input multiple-output (MIMO) cellular
system, IEEE Transactions on Communica-
tions, 49(8):13071311, aug 2001.
[2] T. M. Cover, J. A. Thomas, Elements of Infor-
mation theory, John Wiley & Sons, Inc., 1991.
[3] I. Telatar, Capacity of multi-antenna gaussian
channels, AT&T Technical Memorandum, jun
1995.
[4] G. J. Foschini, M. J. Gans, On limits of wireless
communications in a fading environment when
using multiple antennas, Wireless Personal
Communications, 6:311335, aug 1998.
[5] J. G. Proakis, Digital Communications,
McGraw-Hill, Inc., 1995.
[6] M-S. Alouini, A. Goldsmith, Capacity of nak-
agami multipath fading channels, Proc. Vehic-
ular Technology Conference, 1:358362, 1997.
[7] D. Chizhik, G. J. Foschini, R. A. Valenzuela,
Capacities of multi-element transmit and re-
ceive antennas: Correlations and keyholes, Elec-
tronic Letters, 36(13):10991100, jun 2000.
[8] D. Gesbert, H. B olcskei, D. Gore, A. Paulraj,
MIMO wireless channels: capacity and per-
formance prediction, Global Telecommunica-
tions Conference (GLOBECOM 00), 2:1083
1088, 2000.
[9] D. A. Gore, R. U. Nabar, A. Paulraj, Select-
ing an optimal set of transmit antennas for a
low rank matrix channel, Proc. ICASSP 2000,
49(8):13071311, aug 2000.
[10] S. Sandhu, R. U. Nabar, D. A. Gore, A. Paulraj,
Near-optimal selection of transmit antennas for
a MIMO channel based on shannon capacity,
Conference Record of the 34th Asilomar Confer-
ence on Signals, Systems and Computers, 1:567
571, 2000.
[11] A. F. Molisch, M. Z. Win, J. H. Winters, A.
Paulraj, Capacity of MIMO systems with an-
tenna selection, In Proc. IEEE International
Conference on Communications (ICC), 2:570
574, 2001.

You might also like