Haykin_ch09_pp3.

fm Page 551 Friday, January 4, 2013 4:58 PM

9.11 MIMO Capacity for Channel Known at the Receiver 551

In light of (9.100) and the assumption that hik is a standard Gaussian random variable with
zero mean and unit variance, the average SNR at each receiver input of the MIMO channel
is given by
P
 = ------
-
2
w
2 (9.111)
Nt s
= -----------
-
2
w
2
which is, for a prescribed noise variance  w, fixed once the total transmit power P is fixed.
Note also that, first, all the Nt transmitted signals occupy a common channel bandwidth
and, second, the average SNR  is independent of Nr.
The idealized Gaussian model just described of a MIMO wireless communication
system is applicable to indoor local area networks and other wireless environments, where
the extent of user-terminal mobilities is limited.11

9.11 MIMO Capacity for Channel Known at the Receiver

With the basic complex channel model of Figure 9.30 at our disposal, we are now ready to
focus attention on the primary issue of interest: the channel capacity of a MIMO wireless
link. In what follows, two special cases will be considered: the first case, entitled “ergodic
capacity,” assumes that the MIMO channel is weakly (wide-sense) stationary and, therefore,
ergodic. The second case, entitled “outage capacity,” considers a nonergodic MIMO channel
under the assumption of quasi-stationarity from one burst of data transmission to the next.

Ergodic Capacity
According to Shannon’s information capacity law discussed in Chapter 5, the capacity of a
real AWGN channel, subject to the constraint of a fixed transmit power P, is defined by
 P- bits/s
C = B log 2 1 + ------ (9.112)
  
2
w
2
where B is the channel bandwidth and w
is the noise variance measured over the
bandwidth B. Given a time-invariant channel, (9.112) defines the maximum data rate that
can be transmitted over the channel with an arbitrarily small probability of error being
incurred as a result of the transmission. With the channel used K times for the transmission
of K symbols in T seconds, the transmission capacity per unit time is KT times the
formula for C given in (9.112). Recognizing that K = 2BT in accordance with the sampling
theorem discussed in Chapter 6, we may express the information capacity of the AWGN
channel in the equivalent form
1  P
C = --- log 2 1 + ------- bits/(s Hz) (9.113)
2   
2
w
Note that one bit per second per hertz corresponds to one bit per transmission.
Haykin_ch09_pp3.fm Page 552 Friday, January 4, 2013 4:58 PM

552 Chapter 9 Signaling over Fading Channels

With wireless communications as the medium of interest, consider next the case of a
complex flat-fading channel with the receiver having perfect knowledge of the channel
state. The capacity of such a channel is given by
 h P
2
C = ⺕ log 2 1 + ------------ bits/(s Hz) (9.114)
  
2
w
where the expectation is taken over the gain of the channel h 2 and the channel is assumed
to be stationary and ergodic. In recognition of this assumption, C is commonly referred to
as the ergodic capacity of the flat-fading channel and the channel coding is applied across
fading intervals (i.e., over an “ergodic” interval of channel variation with time).
It is important to note that the scaling factor of 12 is missing from the capacity
formula of (9.114). The reason for this omission is that this equation refers to a complex
baseband channel, whereas (9.113) refers to a real channel. The fading channel covered by
(9.114) operates on a complex signal, namely a signal with in-phase and quadrature
components. Therefore, such a complex channel is equivalent to two real channels with
equal capacities and operating in parallel; hence the result presented in (9.114).
Equation (9.114) applies to the simple case of a single-input, single-output (SISO)
flat-fading channel. Generalizing this formula to the case of a multiple-input, multiple-
output MIMO flat-fading channel governed by the Gaussian model described in Figure
9.30, we find that the ergodic capacity of the MIMO channel is given by the following
formula:12
†
 det  R w + HR s H  
C = ⺕ log 2 ----------------------------------------------  bits/(s Hz) (9.115)
 det  R w  
which is subject to the constraint
max tr  R s   P
Rs
where P is the constant transmit power and tr   denotes the trace of the enclosed
matrix. The expectation in (9.115) is over the random channel matrix H, and the
superscript dagger notes Hermitian transposition; Rs and Rw are respectively the
correlation matrices of the transmitted signal vector s and channel noise vector w. A
detailed derivation of (9.115) is presented in Appendix E.
In general, it is difficult to evaluate (9.115) except for a Gaussian model. In particular,
substituting (9.107) and (9.110) into (9.115) and simplifying yields

 2 
  s † 
C = ⺕ log  det  I N + ------- HH   bits/(s Hz) (9.116)
  r w 
2 2
 
Next, invoking the definition of the average SNR  introduced in (9.111), we may rewrite
(9.116) in the equivalent form
  † 
C = ⺕ log  det  I N + ----- HH   bits/(s Hz), for N t  N r (9.117)
 r Nt 
2
 
Haykin_ch09_pp3.fm Page 553 Friday, January 4, 2013 4:58 PM

9.11 MIMO Capacity for Channel Known at the Receiver 553

Equation (9.117), defining the ergodic capacity of a MIMO flat-fading channel, involves
the determinant of an Nr-by-Nr sum matrix (inside the braces) followed by the logarithm to
base 2. It is for this reason that this equation is referred to as the log-det capacity formula
for a Gaussian MIMO channel.
As indicated in (9.117), the log-det capacity formula therein assumes that Nt  Nr for the
†
matrix product HH to be of full rank. The alternative case, Nr  Nt makes the Nt-by-Nt
†
matrix product H H to be of full rank, in which case the log-det capacity formula of the
MIMO link takes the form

  † 
C = ⺕ log  det  I N + ----- H H  bits/(s Hz), Nr  Nt (9.118)
 t Nr 
2
 
where, as before, the expectation is taken over the channel matrix H.
Despite the apparent differences between (9.117) and (9.118), they are equivalent in
that either one of them applies to all {Nr, Nt} antenna configurations. The two formulas
differentiate themselves only when the full-rank issue is of concern.
Clearly, the capacity formula of (9.114), pertaining to a complex, flat-fading link with a
single antenna at both ends of the link, is a special case of the log-det capacity formula.
2
Specifically, for Nt = Nr = 1 (i.e., no spatial diversity),  = P   w , and H = h (with
dependence on discrete-time n suppressed, (9.116) reduces to that of (9.114).
Another insightful result that follows from the log-det capacity formula is that if
Nt = Nr = N, then, as N approaches infinity, the capacity C defined in (9.117) grows
asymptotically (at least) linearly with N; that is,
C
lim ----  constant (9.119)
N N
In words, the asymptotic formula of (9.119) may be stated as follows:

The ergodic capacity of a MIMO flat-fading wireless link with an equal number
of transmit and receive antennas N grows roughly proportionately with N.

What this statement teaches us is that, by increasing computational complexity resulting

from the use of multiple antennas at both the transmit and receive ends of a wireless link,
we are able to increase the spectral efficiency of the link in a far greater manner than is
possible by conventional means (e.g., increasing the transmit SNR). The potential for this
very sizable increase in the spectral efficiency of a MIMO wireless communication system
is attributed to the key parameter
N = min{Nt, Nr}
which defines the number of degrees of freedom provided by the system.

Two Other Special Cases of the Log-Det Formula: Capacities of

Receive and Transmit Diversity Links
Naturally, the log-det capacity formula for the channel capacity of an Nt, Nr wireless link
includes the channel capacities of receive and transmit diversity links as special cases:
Haykin_ch09_pp3.fm Page 554 Friday, January 4, 2013 4:58 PM

554 Chapter 9 Signaling over Fading Channels

1. Diversity-on-receive channel. The log-det capacity formula (9.118) applies to this

case. Specifically, for Nt = 1, the channel matrix H reduces to a column vector and
with it (9.118) reduces to

 Nr 
 2 
C = ⺕ log 2  1 + 

  hi 

bits/(s Hz) (9.120)
i=1 
 
Compared with the channel capacity of (9.114), for an SISO fading channel with
2
 = P   w , the squared channel gain |h|2 is replaced by the sum of squared
magnitudes |hi|2, i = 1, 2, , Nr. Equation (9.120) expresses the ergodic capacity
due to the linear combination of the receive-antenna outputs, which is designed to
maximize the information contained in the Nr received signals about the transmitted
signal. This is simply a restatement of the maximal-ratio combining principle
discussed in Section 9.8.
2. Diversity-on-transmit channel. The log-det capacity formula of (9.117) applies to
this second case. Specifically, for Nr = 1, the channel matrix H reduces to a row
vector, and with it (9.117) reduces to

 N 
 t
C = ⺕ log 2  1 + -----
 hk 
2
bits/(s Hz) (9.121)
 Nt 
 k=1 
where the matrix product HH† is replaced by the sum of squared magnitudes |hk|2,
k = 1, 2, , Nt. Compared with case 1 on receive diversity, the capacity of the
diversity-on-transmit channel is reduced because the total transmit power is being
held constant, independent of the number of Nt transmit antennas.

Outage Capacity
To realize the log-det capacity formula of (9.117), the MIMO channel must be described
by an ergodic process. In practice, however, the MIMO wireless channel is often
nonergodic and the requirement is to operate the channel under delay constraints. The
issue of interest is then summed up as follows:
How much information can be transmitted across a nonergodic channel,
particularly if the channel code is long enough to see just one random
channel matrix?
In the situation described here, the rate of reliable information transmission (i.e., the strict
Shannon-sense capacity) is zero, since for any positive rate there exists a nonzero
probability that the channel would not support such a rate.
To get around this serious difficulty, the notion of outage is introduced into
characterization of the MIMO link. (Outage was discussed previously in the context of
diversity on receive in Section 9.8.) Specifically, we offer the following definition:
The outage probability of a MIMO link is defined as the probability for which
the link is in a state of outage (i.e., failure) for data transmitted across the link at
a certain rate R, measured in bits per second per hertz.
Haykin_ch09_pp3.fm Page 555 Friday, January 4, 2013 4:58 PM

9.11 MIMO Capacity for Channel Known at the Receiver 555

To proceed on this probabilistic basis, it is customary to operate the MIMO link by

transmitting data in the form of bursts or frames and invoke a quasi-stationary model
governed by four points:
1. The burst is long enough to accommodate the transmission of a large number of
symbols, which, in turn, permits the use of an idealized infinite-time horizon basic to
information theory.
2. Yet, the burst is short enough to treat the wireless link as quasi-stationary during
each burst; the slow variation is used to justify the assumption that the receiver has
perfect knowledge of the channel state.
3. The channel matrix is permitted to change, from burst k to the next burst k + 1,
thereby accounting for statistical variations of the link.
4. Different realizations of the transmitted signal vector s are drawn from a white
Gaussian codebook; that is, the correlation matrix of s is defined by (9.107).
Points 1 and 4 pertain to signal transmission, whereas points 2 and 3 pertain to the MIMO
channel itself.
To proceed with the evaluation of outage probability under this model, we first note
that, in light of the log-det capacity formula (9.117), we may view the random variable
  † 
C k = log 2 det  I N + ----- H k H k  bits/(s Hz) for burst k (9.122)
 r Nt 
 
as the expression for a “sample realization” of the MIMO link. In other words, with the
random-channel matrix Hk varying from one burst to the next, Ck will itself vary in a
corresponding way. A consequence of this random behavior is that, occasionally, a sample
drawn from the cumulative distribution function of the MIMO link results in a value for Ck
that is inadequate to support reliable communication over the link. In this kind of situation
the link is said to be in an outage state. Correspondingly, for a given transmission strategy,
we define the outage probability at rate R as
P outage  R  = ⺠  C k  R k  for some burst k (9.123)

Equivalently, we may write

   †  
P outage  R  = ⺠  log 2 det  I N + ----- H k H k   R for some burst k  (9.124)
 r Nt 
   
On this basis, we may offer the following definition:
The outage capacity of the MIMO link is the maximum bit rate that can be
maintained across the link for all bursts of data transmissions (i.e., all possible
channel states) for a prescribed outage probability.
By the very nature of it, the study of outage capacity can only be conducted using Monte
Carlo simulation.

Channel Known at the Transmitter

The log-det capacity formula of (9.117) is based on the premise that the transmitter has no
knowledge of the channel state. Knowledge of the channel state, however, can be made
Haykin_ch09_pp3.fm Page 556 Friday, January 4, 2013 4:58 PM

556 Chapter 9 Signaling over Fading Channels

available to the transmitter by first estimating the channel matrix H at the receiver and
then sending this estimate to the transmitter via a feedback channel. In such a scenario, the
capacity is optimized over the correlation matrix of the transmitted signal vector s, subject
to the power constraint; that is, the trace of this correlation matrix is less than or equal to
the constant transmit power P. Naturally, formulation of the log-det capacity formula of a
MIMO channel for which the channel is known in both the transmitter and receiver is
more challenging than when it is only known to the receiver. For details of this
formulation, the reader is referred to Appendix E.

9.12 Orthogonal Frequency Division Multiplexing

In Chapter 8 we introduced the DMT method as one discrete form of multichannel

modulation for signaling over band-limited channels. Orthogonal frequency division
multiplexing (OFDM)13 is another clearly related form of multifrequency modulation.
OFDM is particularly well suited for high data-rate transmission over delay-dispersive
channels. In its own way, OFDM solves the problem by following the engineering
paradigm of “divide and conquer.” Specifically, a large number of closely spaced
orthogonal subcarriers (tones) is used to support the transmission. Correspondingly, the
incoming data stream is divided into a number of low data-rate substreams, one for each
carrier, with the subchannels so formed operating in parallel. For the modulation process,
a modulation scheme such as QPSK is used.
What we have just briefly described here is essentially the same as the procedure used
in DMT modulation. In other words, the underlying mathematical theory of DMT
described in Chapter 8 applies equally well to OFDM, except for the fact that the signal
constellation encoder does not include the use of loading for bit allocation. In addition,
two other changes have to be made in the implementation of OFDM:
1. In the transmitter, an upconverter is included after the digital-to-analog converter to
appropriately translate the transmitted frequency, so as to facilitate propogation of
the transmitted signal over the radio channel.
2. In the receiver, a downconverter is included before the analog-to-digital converter to
undo the frequency translation that was performed by the upconverter in the
transmitter.
Figure 9.31 shows the block diagram of an OFDM system, the components of which are
configured to accommodate the transmission of a binary data stream at 36 Mbit/s as an
illustrative example. Parts a and b of the figure depict the transmitter and receiver of the
system, respectively. Specifically, pertinent values of data carrier rates as well as sub-
carrier frequencies at the various functional blocks are included in part a of the figure
dealing with the transmitter. One last comment is in order: the front end of the transmitter
and the back end of the receiver are allocated to forward error-correction encoding and
decoding, respectively, for improved reliability of the system. (Error-control coding of the
forward error-correction variety is discussed in Chapter 10.)

The Peak-to-Average Power Ratio Problem

A compelling practical importance of OFDM to wireless communications is attributed to
the computational benefits brought about by the FFT algorithm that plays a key role in its