Chapter 2

Large scale Variations in Gain

2.1 INTRODUCTION
In this chapter we will study the large scale variations in the channel that are captured in the power gain term
G(d, ) of (1.16). The variation of G(d, ) with location is in general quite complex and nature of the vari-
ation changes from one geographical region to the next. Measurements of G(d, ) can be obtained through
drive tests or through detailed simulation using all the characteristics of the environment, after averaging out
the small scale variations. While the details of G(d, ) may be useful for certain kinds of cellular planning to
avoid holes in coverage, no such detailed model can ever be accurate since the environment could change.
More importantly, these details will be too cumbersome to be useful in design and analysis of wireless com-
munication systems.
We will hence try and capture the essential features of G(d, ) in a useful parametric model with a few pa-
rameters. As we shall see, this process is facilitated by resorting to a stochastic model for G(d, ). In cellular
systems, for example, some of these parameters can be expected to remain constant over regions of the cell,
and they can be estimated from eld trials or simulations and stored at the basestation. Other parameters can
be estimated online based on measurements taken on a pilot signal broadcast by the basestation.
2.2 MEDIAN COMPONENT OF GAIN
To understand the variations in G(d, ) with location, we appeal to the physics of electromagnetic wave prop-
agation. For radiation in free space, Friis equation [Par98] dictates that the received power should decay with
distance from the transmitter as d
2
(note that there is no dependence on ). However in a realistic propa-
gation environment, there are several scatterers that reect the wave, and also possibly refract or defract the
wave before it reaches the receiver. Thus the power gain G(d, ) has a much more complicated variation with
location. The book by Parsons [Par98] gives an excellent introduction to the various modes of propagation in
a typical wireless environment, and their effect on the power gain. However, as noted earlier, a detailed study
of the propagation physics is futile since the resulting description of G(d, ) would be too cumbersome to be
useful. Nevertheless, we use the intuition from physics to argue that G(d, ) should have a mean variation
c _V. V. Veeravalli, 2007 9
that is monotonically decreasing in d and is independent of . This motivates the following model for G(d, ):
G(d, ) = S(d, )

G(d) . (2.1)
where

G(d) is the monotonic component. Note that we can always write G in this fashion if we allow S and

G to be arbitrary. Our goal is to nd the

G that best ts the variation in G. In order to do this, it is more
convenient to rewrite the above equation in dB as
G(d, )[dB] = Z(d, ) +

G(d)[dB] , (2.2)
where Z(d, ) = 10 log S(d, ). Now

G(d)[dB] can indeed be considered as the mean variation of G(d, )[dB],
with Z(d, ) being the variation around the mean. The problem of nding the best

G(d)[dB] is facilitated by
using a parametric model. This is where we resort to propagation physics to argue that

G(d) should be roughly
of the form ad
b
, where b > 0, i.e., that

G(d)[dB] is of the form

G(d)[dB] ABlog d . (2.3)

More generally, we could model

G(d)[dB] as being piece-wise linear in log d. The simple linear model is quite
accurate for outdoor channels, whereas a piece-wise linear model may be more appropriate for indoor channels
[Stu96, Pg. 104]. There is also experimental evidence [?] that suggests that

G(d) falls off exponentially for
large values of d, i.e., that

G(d)[dB] is eventually linear in d. Recent work by Franceschetti et al [?] based on
a wandering photon model provides a theoretical justication for this exponential tail.
To obtain the constants Aand B for a given region we use linear regression on measured values of G(d, )[dB]
obtained via drive tests or via detailed simulation. Then, A and B will correspond to the best cone t to
G(d, )[dB] and Z(d, ) will represent the error (see Fig. 2.1). There have been attempts to generate generic
rough estimates of A and B for various kinds of environments using extensive measurements (see, e.g., the
models of Okumura and Hata in [Stu96, Pg. 95]). Clearly, the most accurate computation of

G(d) is obtained
from drive tests taken in the cell. Also, the more accurate the curve t for the mean, the smaller the variance of
the error term Z(d, ). Of course, no matter how accurate we make our model for

G(d), variance of Z(d, )
cannot be made zero, since G is a function of both d and .
PSfrag replacements
(log d, ) (log d, ) (log d, )
G[dB]
G[dB] +Z
A Blog d
Figure 2.1: The best cone t for G(d, )[dB] and the corresponding error Z(d, )
c _V. V. Veeravalli, 2007 10
2.3 SHADOWFADING
In the previous section, we modeled the mean value of the link gain (in dB) using a few parameters (just A
and B in the simplest case). While the mean link gain is modeled as being independent of , the error term
Z(d, ) in (2.2) is clearly a function of both d and . We now address the modeling of Z(d, ). In general,
it would be impossible to nd a simple parametric deterministic model for the measured values of Z(d, ).
Of course we could always simply interpolate between all the available sample values of Z(d, ), but such
a model will be just as cumbersome as that obtained by interpolating between values of G(d, ). We hence
resort to describing Z(d, ) using a stochastic model. The error term Z(d, ), when modelled as stochastic, is
called shadow fading. Note that since Z(d, ) is a (random) function in two dimensions, it is a random eld.
(See Chapter 8 of [WH85] for an introduction to random elds. Also see Appendix C.) We begin our modeling
exercise by studying the rst order statistics of Z(d, ).
2.3.1 FIRST ORDER STATISTICS OF SHADOWFADING
Since Z(d, ) is the residual error obtained from the best curve t

G(d)[dB] to G(d, )[dB], it is reasonable to
assume that the average value of Z(d, ) is zero. Also, assuming that the nature of the environment is roughly
isotropic within the coverage of the basestation, the variance of Z(d, ) may be modeled as being independent
of location (d, ). Finally, Central Limit Theorem arguments suggest a Gaussian model for the statistics. Thus,
we model Z(d, ) by a ^(0,
2
Z
) random variable, with the same value of
2
Z
for every point in a given region
of the cell. This model for Z(d, ) automatically implies a corresponding model for the multiplicative fading
term in (2.1). Since Z = 10 log S, i.e. S = 10
Z/10
, it is easy to show that the pdf of S is given by
p
S
(s) =
10
s ln(10)
1

2
exp
_

(10 log s)
2
2
2
Z
_
u(s) (2.4)
where u() is the unit step function. The distribution of S is called lognormal since the logarithm of S is
normal (Gaussian), and hence shadow fading is also referred to as lognormal fading.
With the Gaussian model for Z(d, ), G(d, )[dB] is Gaussian with both mean and median equal to

G(d)[dB],
and variance
2
Z
. Hence G(d, ) is also lognormal. Note that
E[G] = E
_
10

G[dB]+Z
10
_
< 10

G[dB]
10
=

G
(2.5)
where the inequality is due to Jensens inequality [WH85]. Thus the mean of G does not equal

G. However, if
we consider the median of G, denoted by
G
, we see that
G
is the solution to the equation
PG
G
=
1
2
=P10 log G 10 log
G
=
1
2
=PG[dB]
G
[dB] =
1
2
(2.6)
which implies that
G
[dB] =

G[dB], since G[dB] has median

G[dB]. We summarize this in the following
result.
c _V. V. Veeravalli, 2007 11
Result 2.1. The median of G equals

G.
This is why

G(d) is sometimes referred to as the median link gain.
Thus far we have described only the rst order statistics of Z, i.e., the marginal distribution of the shadow
fading at a single point (d, ). To completely characterize a random eld, it is necessary to specify the joint
distribution of any set of sample values of the eld. Under a Gaussian eld model, all that needs to be specied
is the correlation between any given pair of sample values of the eld.
2.3.2 CORRELATION MODEL FOR SHADOWFADING
To model the eld correlation, it is more convenient to use an (x, y) coordinate system, with x = d cos and
y = d sin . With a slight abuse of notation, we denote the resulting eld by Z(x, y). A general model for the
eld correlation would assume an arbitrary correlation between samples of the eld, but such a model would
be too cumbersome to be useful. If the scatterers in the cell (or geographical region of interest) are roughly of
the same shape and size, it may be reasonable to assume that the correlation between a pair of sample values
of the eld depends only on the distance between the points and not on their absolute locations.
This leads to the homogeneous, isotropic, random eld model described in Appendix C. Since Z(x, y) is
modelled as Gaussian as well, it is completely characterized by its mean (0 in our model) and correlation
function, R
Z
(), which is dened as
R
Z
() = E[Z(x
1
, y
1
) Z(x
2
, y
2
)] , where =
_
(x
2
x
1
)
2
+ (y
2
y
1
)
2
. (2.7)
What is good model for the correlation function R
Z
()? The simplest possible model for the correlation is a
white model with R
Z
() =
2
Z
(). However, the white model is clearly unrealistic for shadow fading,
since presumably the link gains at points that are close together will be correlated. This leads us to the following
rst-order correlation model.
Exponential Model for R
Z
()
Based on measurements taken in various cities, Gudmundson [Gud91] suggested the following simple expo-
nential correlation model:
R
Z
() =
2
Z
exp
_

[[
D
c
_
, (2.8)
where D
c
is dened to be the correlation distance of the fading process. The value of D
c
can be estimated from
eld trials or detailed simulations. Note that there is no physical justication for this exponential correlation
model it is simply a convenient one parameter function that appears to capture the correlation behaviour
observed in eld trials. Clearly, a more accurate t may be obtained by using higher order models with more
parameters.
2.3.3 SHADOWFADING IN TIME FOR MOBILE TERMINAL
Consider the situation where one of the terminals is mobile and other is xed (base) at the origin, with
(x(t), y(t)) representing the location of the mobile at time t. The power gain of the link sees variations as
c _V. V. Veeravalli, 2007 12
PSfrag replacements
Base
Mobile
(x(t), y(t))
_
x
2
(
t
)
+
y
2
(
t
)
Figure 2.2: Trajectory of mobile terminal.
function of t that are given by
G

(t) =

G
_
_
x
2
(t) +y
2
(t)
_
+Z

(t), (2.9)
where Z

(t) = Z(x(t), y(t)) is the shadow fading process (not eld!) seen by the user. Note that
E[Z

(t)Z

(t +)] = R
Z
_
_
(x(t +) x(t))
2
+ (y(t +) y(t))
2
_
, (2.10)
which means that Z

(t) is, in general, a nonstationary process, even if the eld is homogeneous and isotropic.
If the terminal is moving on a straight line with constant velocity v, Z

(t) a stationary process with autocorre-

lation function
R
Z
() = R
Z
(v)
=
2
Z
exp
_

v[[
D
c
_
(for exponential model).
(2.11)
Discrete time (distance) model
Suppose the process Z

(t) is sampled in time such that the k-th sample location is (x

k
, y
k
). Then the link gain
at the k-th sample is given by
G
k
=

G
_
_
x
2
k
+y
2
k
_
+Z
k
, (2.12)
where Z
k
is a discrete-time Gaussian process with zero mean, variance
2
Z
, and autocorrelation function
R
Z
[k, k +m] = E[Z
k
Z
k+m
]
= R
Z
_
_
(x
k+m
x
k
)
2
+ (y
k+m
y
k
)
2
_
.
(2.13)
c _V. V. Veeravalli, 2007 13
PSfrag replacements
Base
Mobile
(x
k
, y
k
)
(x
1
, y
1
)
(x
n
, y
n
)
_ x
2
k
+
y
2
k
Figure 2.3: Discrete time shadow fading process
If the samples are equally-spaced with distance d
s
on a straight line trajectory, then Z
k
is stationary with
autocorrelation function
R
Z
[m] = R
Z
(md
s
)
=
2
Z
exp
_

[m[d
s
D
c
_
=
2
Z
a
|m|
(for exponential model).
(2.14)
where a = exp(d
s
/D
c
) is the correlation coefcient of the process Z
k
.
Based on the models we have developed for the large scale variations, we can see that we need a minimum
of four parameters (A, B,
Z
and D
c
) to characterize these variations statistically in a region of interest. A
piece-wise linear t for

G(d) we would require more than one pair (A
i
, B
i
), and a more accurate correlation
model may involve more than just the one parameter D
c
.
2.4 SIMULATING LARGE SCALE VARIATIONS
An immediate application of the models we have developed thus far is in the simulation of the large scale
variations in a typical cellular environment. To simulate a typical prole for G(d, ) over the coverage of a
cell, we need to generate the mean

G(d) and a realization of the random eld Z(d, ). Generating

G(d) is
easy, especially for a piece-wise linear model in dB. However, generating a set of samples of the shadow fading
eld Z(d, ) is more involved.
2.4.1 SIMULATING THE SHADOWFADING RANDOMFIELD
Consider simulating samples ofZ(d, ) (or Z(x, y)) over some region of interest. Suppose the sample points
are (x
i
, y
i
), i = 1, 2, . . . , n, and the corresponding eld values are:
Z
i

= Z(x
i
, y
i
) , i = 1, 2, . . . , n . (2.15)
c _V. V. Veeravalli, 2007 14
Dene the vector Z = [Z
1
Z
2
Z
n
]

. Assuming that the eld is homogeneous and isotropic with autocor-

relation function R
Z
(), the covariance matrix of Z has components that are given by:

ij
= E[Z
i
Z
j
] = R
Z
_
_
(x
i
x
j
)
2
+ (y
i
y
j
)
2
_
. (2.16)
The random eld values can then be jointly generated from a vector W of i.i.d. ^(0, 1) random variables as:
Z = CW, (2.17)
where C is such that CC

= . One way to obtain C is through an LU decomposition [HJ85] of .

2.4.2 SIMULATING THE SHADOWFADING PROCESS
In the special case, where we are interested in simulating the discrete-time shadow fading process when the
is mobile moving through the cell with a constant velocity, the problem is considerably easier. To simulate
the discrete time process Z
k
for a stationary model with autocorrelation function R
Z
[m], we simply lter
a sequence W
k
of i.i.d. ^(0, 1) random variables (white Gaussian noise) through a LTI system. The lter
transfer function H(z) must satisfy
H(z)H(z
1
) = S
Z
(z) , (2.18)
where S
Z
(z) is the Z-transform of R
Z
[m]. Such a lter can be found by spectral factorization [WH85].
For the exponential autocorrelation function R
Z
[m] =
2
Z
a
|m|
,
S
Z
(z) =

2
Z
(1 a
2
)
(1 az
1
)(1 az)
, (2.19)
which implies that
H(z) =

Z

1 a
2
1 az
1
. (2.20)
Thus Z
k
can be calculated using the recursive equations:
Z
0
=
Z
W
0
Z
k+1
= aZ
k
+
Z
_
1 a
2
W
k+1
, k = 0, 1, . . .
(2.21)
In this case, the sequence Z
k
is also referred to as a rst-order autoregressive (AR-1) discrete-time stochastic
process [WH85].
2.5 APPLICATIONS OF LARGE SCALE FLUCTUATION MODELS
The rst two applications discussed below use only the rst-order statistics of shadow fading, whereas the last
one uses correlation information as well.
c _V. V. Veeravalli, 2007 15
2.5.1 CELLULAR AREA RELIABILITY
Consider the design of a cellular basestation network. If the transmit power is P
t
[dBm], then the received
power at (d, ) is given by
P
r
(d, ) [dBm] = P
t
[dBm] +G(d, ) [dB]
= P
t
[dBm] +ABlog d +Z(d, )
= A
t
Blog d +Z(d, ) .
(2.22)
where A
t
= P
t
[dBm] +A.
The cells are to be designed in such a way that received signal strength is at a reliable level over the coverage
of the cell. The probability of outage at a point (d, ) is dened to be the probability that P
r
(d, ) [dBm] falls
below a threshold P
th
[dBm] required for reliable performance, i.e.,
P
out
(d, ) = PA
t
Blog d +Z(d, ) < P
th
[dBm] . (2.23)
Since Z is Gaussian with mean zero and variance
Z
, we have
P
out
(d, ) = 1 Q(P
th
[dBm] A
t
+Blog d) , (2.24)
where Q() is the tail of unit Gaussian distribution, i.e.,
Q(x) =
1

2
_

x
e

x
2
2
dx . (2.25)
The area reliability of the cell [Rap96] is dened by
F
area
= 1 average outage probability over the cell
= 1
1
R
2
_
R
0
_

P
out
(, ) d d .
(2.26)
where R is the cell radius.
From (2.24) and (2.26), it follows that
F
area
=
2
R
2
_
R
0
Q(a +b ln ) d, (2.27)
where a and b are given by
a =
P
th
[dBm] A
t

Z
, b =
B

Z
ln(10)
. (2.28)
It is a straightforward exercise in integration by parts to show that
F
area
= Q(a +b ln R) +
exp
_
2
b
2

2a
b
_
R
2
_
1 Q
_
a +b ln R
2
b
__
. (2.29)
A related quantity of interest in characterizing the reliability in a cell is the edge reliability, F
edge
, which is
dened by
F
edge
= 1 P
out
(R) = Q(a +b ln R) . (2.30)
c _V. V. Veeravalli, 2007 16
0 1 2 3 4 5 6 7 8
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
sigma hat
F
u
Fu versus sigma hat
Figure 2.4: Area reliability as a function of edge reliability and normalized shadow fading standard deviation.
It easy to show that F
edge
and F
area
are related in the following manner:
F
area
= F
edge
+e
2
b
2
e

2
b
Q
1
(F
edge
)
_
1 Q
_
Q
1
(F
edge
)
2
b
__
(2.31)
For xed F
edge
, F
area
is independent of a and R, and is a function of b only. This means that once we x the
edge reliability to a desired value, then the area reliability is only a function of the pathloss exponent B/10
and the shadow fading standard deviation
Z
. If we dene the normalized standard deviation by

Z
=
10
Z
B
(2.32)
then F
area
can be expressed in terms of F
edge
and
Z
using (2.31). This relationship is captured in Figure 2.4.
2.5.2 SIR IN CHANNELIZED SYSTEMS WITH SHADOWFADING
In channelized (F/TDMA) cellular communication systems, the signal-to-interference ratio (SIR) seen by a
user in the cell of interest is given by
=
P
r,0

N
I
i=1
P
r,i
, (2.33)
where P
r,0
is the power received from the transmitter of interest, P
r,i
is the power received from the i-th
co-channel interferer, and N
I
is the number of co-channel interferers.
If we consider the downlink (basestation to mobile), then with the received power at location (d, ) relative to
the basestation modeled as in (2.22), it is easy to see that
P
r
(d, ) = Kd
n
S(d, ) , (2.34)
where K = 10
At/10
and n = B/10.
c _V. V. Veeravalli, 2007 17
Assuming that median gain parameters (A and B) and the transmit powers (P
t
) are the same for basestations,
K and n are equal for all basestations. If we compute the SIR of(2.33) after averaging out the small scale
variations from the received powers, then
=
S
0
D
n
0
N
I

i=1
S
i
D
n
i
=
S
0
N
I

i=1
S
i
_
D
i
D
0
_
n
, (2.35)
where S
0
is the (lognormal) shadow fading seen on the link from the intended basestation, and S
i
is the fading
seen from the i-th cochanell interferer. Also, D
0
is the distance to the intended basestation, and D
i
is the
distance to the i-th co-channel interferer.
If we model S
0
, S
1
, S
2
, . . . , S
N
I
as independent random variables, then the sum in the denominator of (2.35)
can be approximated reasonably well by a lognormal
1
. This leads to a lognormal model for (since the ratio
of lognormals is lognormal) whose statistics can be determined explicitely. To do this we use the following
result that is easily established from rst principles.
Result 2.2. Suppose X is a lognormal random variable with mean m
X
and second moment
X
, and suppose
Y = 10 log X has mean m
Y
and variance
2
Y
. (Note that Y is Gaussian.)
m
X
= exp
_
(
Y
)
2
2
_
exp(m
Y
),

X
= exp
_
2(
Y
)
2
_
exp(2m
Y
),
m
Y
= 20 log m
X
5 log
X
,

2
Y
=
1

(10 log
X
20 log m
X
),
(2.36)
where = ln(10)/10.
Example: SIR for single tier crude approximation.
Consider the rst tier of six cochannel interfering cells surrounding the cell of interest in a hexagonal cellular
grid. If only these interferers are included in the SIR calculation, then the worst-case SIR can be approximated
as [Rap96]:

worst-case

S
0
R
n

6
i=1
S
i
D
n
i

_
D
R
_
n
S
0

6
i=1
S
i
(2.37)
where D is the distance between cochannel cells, and R is the cell radius.
Let S
T
=

6
i=1
S
i
. For the lognormal approximation we need to compute m
S
T
and
S
T
, which are given by:
m
S
T
=
6

i=1
m
S
i
,

S
T
= m
2
S
T
+
6

i=1
_

S
i
m
2
S
i

.
(2.38)
1
More accurate approximations can be derived (see Section 3.1 of Stuber [Stu96]).
c _V. V. Veeravalli, 2007 18
Now suppose the S
i
are identically distributed, with Z
i
= 10 log S
i
^(0,
2
). Then it is easy to show
from (2.36) that
m
S
T
= 6 exp
_

2
2
_
,

S
T
= 6
_
exp
_
2
2

2
_
+ 5 exp(
2

2
)

(2.39)
and that Z
T
= 10 log S
T
is Gaussian with
m
Z
T
= 20 log m
S
T
5 log
S
T
,

2
Z
T
=
1

(10 log
S
T
20 log m
S
T
) .
(2.40)
Note that m
Z
T
,= 0. Now let

be the SIR in dB. Then from (2.37), we get

= 10nlog
_
D
R
_
+Z
0
Z
T
(2.41)
This means that

is Gaussian with
m

= 10nlog
_
D
R
_
m
Z
T
,

=
2
+
2
Z
T
.
(2.42)
We can now pose a question such as: what is the probability that the worst-case SIR is less than 18 dB? The
answer is easily seen to be 1 Q((18 m

)/

), which can be computed terms of D, R, n and

2
using the
series of equations derived above.
2.5.3 SIGNAL STRENGTH PREDICTION
For radio resource management functions (such as handoff and power control) in cellular systems, it is often of
interest to be able to predict the large scale variations of the channel based on previous measurements. These
measurements may, for example, be available through monitoring of pilot signals.
Suppose k sample values of the signal strength at locations (d
1
,
1
), . . . , (d
k
,
k
) are known, and using these
values we wish to predict the signal strength at location (d
k+1
,
k+1
). Now, the received power for sample i,
1 i k is given by
P
r,i
= A
t
Blog d
i
+Z
i
. (2.43)
If we assume that A
t
, B and the d
i
s are known, the problem of predicting P
r,k+1
based on P
r,1
, . . . , P
r,k
is
equivalent to that of predicting Z
k+1
based on Z
1
, . . . , Z
k
.
The prediction problem can be posed in a minumum mean squared-error (MMSE) framework. Under the
Gaussian model for the Z
i
s, the MMSE predictor for Z
k+1
is linear and is given by

Z
MMSE
k+1
=
k

i=1
a
i
Z
i
, (2.44)
c _V. V. Veeravalli, 2007 19
where the a
i
s are solutions to the Yule-Walker equations [WH85]:
_

_
R
1,1
R
1,2
. . . R
1,k
R
2,1
R
2,2
. . . R
2,k
.
.
.
.
.
.
.
.
.
.
.
.
R
k,1
R
k,2
. . . R
k,k
_

_
_

_
a
1
a
2
.
.
.
a
k
_

_
=
_

_
R
1,k+1
R
2,k+1
.
.
.
R
k,k+1
_

_
(2.45)
with R
i,j
= E[Z
i
Z
j
], which can be computed based on knowledge of the eld autocorrelation function R
Z
().
The MMSE solution takes on a much simpler form in the special case where the samples Z
k
form an AR-1
process (see Section 2.4.2). In this case it easy to see that

Z
MMSE
k+1
= E[Z
k+1
[ Z
k
, . . . , Z
1
] = aZ
k
(2.46)
and also that

P
MMSE
r,k+1
= aP
r,k
+ (1 a)A
t
Blog
_
d
k+1
d
a
k
_
. (2.47)
2.6 CONCLUDING REMARKS
In this chapter we developed a model for the variations in the power gain in a terrestrial cellular system as a
function of the location of the receiver relative to the transmitter. We saw that is was possible to characterize
the variations stochastically using a few parameters. We note that in addition to variations in power gain, there
may be other large-scale variations in the channel introduced by mobility. In Chapter 3, we will see that the
small scale variations in the channel are a function of the Rice factor, which is strength of the line-of-sight
path relative to the diffuse reected paths connecting the Tx and Rx antennas. The variations in the Rice
factor are sometimes modelled using a lognormal distribution. We will also see in Chapter 3, that the time and
frequency selectivity of the channel depend on the delay-Doppler scattering function of the channel, which is
a description of the distribution of received power as a function of delay and Doppler frequency. The delay-
Doppler scattering function depends of the location of the mobile and the varies over the large scale. The same
is true of the spatial scattering function that is studied in Chapters 5. Modeling variations in the scattering
functions is a difcult and open problem that deserves further attention.
c _V. V. Veeravalli, 2007 20
Chapter 3
Small scale Variations in Gain
3.1 INTRODUCTION
In Chapter 1 we justied the separation of the variations in the channel into two scales, and in Chapter 2 we
studied the large scale variations in the channel. In this chapter, we study the small scale variations, treating
the large scale variations as constant (see Fig 3.1). Our analysis here is restricted to the situation where only
one antenna is present at the transmitter and receiver. Generalizations are considered in Chapter 5.
PSfrag replacements
s(t) y(t)
g
h(t; )
x(t)
Figure 3.1: Small scale variations in the channel (with large scale variations treated as constant)
Recall from Section 1.3 that the small scale variations in the channel are captured in a linear system with
response
h
d,
() =
N

n=1

n
(d, ) e
jn(d,)
(
n
(d, )) (3.1)
where the gains
n
are normalized so that

N
n=1

2
n
= 1. Small movements of the terminals (around relative
location (d, )), or movements of the scatterers, cause the gains
n
, the delays
n
and the phases
n
to
vary with time. Thus the channel corresponding to the small scale variations becomes time-varying, with an
impulse response that is given by
h(t; ) =
N

n=1

n
(t) e
jn(t)
(
n
(t)) (3.2)
For notational convenience, from here foreword, we will skip the number of paths, N, in the summations.
c _V. V. Veeravalli, 2007 21
Treating the large scale variations g(d, ) as roughly constant (see Figure 3.1), we obtain the input-output
relationship:
y(t) =
_

0
h(t; )x(t )d
= g
_

0
h(t; )s(t )d .
(3.3)
3.2 DELAY POWER PROFILE AND DELAY SPREAD
As we remarked earlier in Chapter 1, for movements of the order of a few wavelengths,
n
(t) and
n
(t)
can be assumed to be roughly constant, and the time variations in h(t; ) are mainly due to changes in
n
(t),
i.e.,
h(t; )

n
e
jn(t)
(
n
) (3.4)
The power gain associated with the n-th path is given by
2
n
. One way to describe the physical scattering
environment for the purposes of studying the channel is to plot the power gains in the paths as a function of
the delay. Such a plot, which is shown in Figure 3.2, is called the delay power prole (or simply delay prole)
of the channel. Notice that we have associated an impulse or delta function with each path in Figure 3.2. The
reason is that the power delay prole can be also interpreted as a power gain density as a function of delay
since we have normalized the powers in the paths to add up to unity.
With some possible abuse of notation, we use the index 0 to denote the line of sight (LOS) path, if one exists.
The width of the delay prole (delay spread) is of the order of tens of microseconds for outdoor channels, and
of the order of hundreds of nanoseconds for indoor channels. Note that the paths typically appear in clusters
in the delay prole, with each cluster representing a diffuse reector.
Denition 3.1. The quantity
ds
= max
n
min
n
is called the delay spread of the scattering environment
seen by the terminals.
Without loss of generality, we may assume that the delay corresponding to the rst path arriving at the receiver
is 0. The rst path would of course be the LOS path if one exists. Then min
n
= 0,
ds
= max
n
, and (3.3)
can be rewritten as:
y(t) =
_

ds
0
h(t; )x(t )d
= g
_

ds
0
h(t; )s(t )d
(3.5)
The delay prole is the key to characterizing the frequency selectivity of the channel as we shall see in Sec-
tion ??.
3.3 DOPPLER POWER PROFILE AND DOPPLER SPREAD
As we remarked earlier, the time variations on the small scale are entirely due to variations in the path phases.
To understand the nature of these variations, consider the simple scenario where only one of the terminals is
c _V. V. Veeravalli, 2007 22
PSfrag replacements

0
ds
Power

2
0
(LOS)

2
n

n
Figure 3.2: Typical delay power prole of channel with LOS path having delay 0.
mobile and the scatterers are xed. Suppose the mobile is traveling with velocity v and suppose that the n-th
path has an angle of arrival
n
(t) with respect to the velocity vector as shown in Figure 3.3. The angle of
arrival
n
(t) plays a key role in the phase variations. Note that we may assume that
n
is roughly constant over
the small scale. Then for small
t
,

n
(t +
t
)
n
(t)
2f
c
v
t
cos
n
c
=
2v

t
cos
n
, (3.6)
where
c
is the carrier wavelength and c is the velocity of light. The frequency shift in the n-th path introduced
by the movement of the mobile is hence given by

n
= lim
t0

n
(t +
t
)
n
(t)
2
t
=
v

c
cos
n
(3.7)
This frequency shift is familiar Doppler shift that is commonly observed in the context of sound waves travel-
ling from moving sources.
More generally, both terminals and the scatterers may be mobile. In this case, the phase shift introduced in the
n-path will be due all the mobile units (terminals and scatterers) contributing to the path. Over the small scale
it is reasonable to assume that all mobile units are travelling at constant velocities. Hence the phase difference
term shown in (3.6) gets modied as:

n
(t +
t
)
n
(t)
M

i=1
2v
i

t
cos
n,i
, (3.8)
where M is the number of mobile units, v
i
is the velocity of the i-th mobile unit relative to the previous unit
in the path and
n,i
is the angle of arrival at the i-th unit. The corresponding frequency shift is given by:

n
=
M

i=1
v
i

c
cos
n,i
. (3.9)
c _V. V. Veeravalli, 2007 23
PSfrag replacements
v
vt

n
vt cos
n
Figure 3.3: Doppler Shift.
PSfrag replacements

+
0
Power
Figure 3.4: Doppler power prole.
Note that if the maximum possible velocity for the mobile units is v
max
, then the Doppler shift for any path is
upper bounded by
max
= 2v
max
/
c
.
Just as we plotted the power in the paths as a function of their delay, we may also plot the power in the paths as
a function of the Doppler shift associated with them. An example Doppler power prole (or simply Doppler
prole) of the channel is shown in Figure 3.4. Again, we associate an impulse with each path in the Doppler
prole, with the understanding that the prole can also be interpreted as a power gain density as a function of
Doppler shift. We will revisit this interpretation in Section 3.4.1.
We denote the minimum and maximum Doppler frequency (for a given location of the terminals) by

and

+
, respectively, with

= min
n

n
,
+
= max
n

n
. (3.10)
We then have the following denition of Doppler spread that is analogous to the delay spread of Denition 3.1.
Denition 3.2. The quantity
ds
=
+

is called the Doppler spread of the scattering environment seen

by the terminals.
Unlike in the case of the delay prole, we do not change the axes in the Doppler prole to obtain a prole
that spans the range of Doppler frequencies [0,
ds
]. This is because the actual value of the Doppler shift
is important in the determining how the carrier frequency is affected by the phase variations introduced by
channel.
c _V. V. Veeravalli, 2007 24
PSfrag replacements
s(t) y(t)
x(t)
g
E(t)
Figure 3.5: Channel Model for Flat Fading
The Doppler power prole will play a key role in characterizing the time selectivity of the channel as we shall
see in following section.
3.4 PURELY TIME-SELECTIVE (FREQUENCY-FLAT) FADING
If the bandwidth of transmitted signal s(t) is much smaller than 1/
ds
, then s(t), and hence the input x(t),
does not change much over time intervals of the order of
ds
. Thus (3.5) can be approximated as
y(t) x(t)
_

ds
0
h(t; )d = x(t)

n
e
jn(t)
. (3.11)
This implies that the multipath channel simply scales the transmitted signal without introducing signicant
frequency distortion. The variations with time of this scale factor are referred to as frequency nonselective, or
at, fading.
Note that the distortions introduced by the channel depend on the relationship between the delay spread of
the channel and the bandwidth of the signal. The same channel may be frequency selective or at, depending
on the bandwidth of the input signal. With a delay spread of 10 s corresponding to a typical outdoor urban
environment, an analog AMPS signal (30 kHz) undergoes at fading, whereas an IS-95 CDMA signal (1.25
MHz) undergoes frequency selective fading.
For at fading, the channel of Figure 3.1 simplies to that shown in Figure 3.5, i.e.,
y(t) = E(t) x(t) , (3.12)
where
E(t) =
_

ds
0
h(t; )d =

n
e
jn(t)
. (3.13)
3.4.1 STATISTICAL CHARACTERIZATION OF FLAT FADING
To model E(t) statistically, we rst t a stochastic model to the phases
n
(t). Based on the Doppler model
for the phase drift given in (3.8), we may approximate the evolution of the phase process as

n
(t)
n
+ 2
n
t (3.14)
where
n
is an initial phase.
c _V. V. Veeravalli, 2007 25
We may combine the initial phase with the path gain to dene the complex path gain

n
=
n
e
jn
. (3.15)
Furthermore, we dene the complex path process e
n
(t) as
e
n
(t) =
n
e
j2nt
(3.16)
so that
E(t) =

n
e
n
(t) . (3.17)
The randomness in E(t) is captured in the initial phases
n
, which we model using the following reasonable
assumption.
Assumption 3.1. For each n, the initial phase
n
is a random variable that is uniformly distributed on [, ].
This assumption immediately gives us the result that the path process e
n
(t) is a zero mean process. Further-
more, using (3.17), we obtain the following important result.
Result 3.1. E(t) is a zero-mean process.
The uniform phase assumption actually yields an even more powerful result about the process E(t). In partic-
ular, the pseudocovariance function of e
n
(t) can be computed as follows.

C
en
(t +, t) = E[e
n
(t)e
n
(t +)]

2
n
E[e
j(2n+2n(2t+))
] = 0
(3.18)
where the approximation follows from (3.14). This implies that e
n
(t) is (approximately) a proper complex
random process (see Denition B.2 in Appendix B). Furthermore, using the result that linear combinations of
proper complex processes remain proper (see Result B.5 Appendix B), we obtain:
Result 3.2. The process E(t) is (approximately) a proper complex random process.
It is worth pointing out that standard analyses of the fading processes (see, e.g., [Jak74, Pro95, Stu96]) do not
fully exploit the properness of the process E(t). The statistical properties of the fading process are instead
established by considering the inphase and quadrature components separately. By establishing that E(t) is
proper at the start, we will see that the subsequent analysis is much simplied and cleaner.
Thus far we have established that E(t) is a zero mean proper complex process. In order to characterize the
autocorrelation function of E(t), we need to know how the phases
n
are jointly distributed. This leads us
to the most important assumption of this book.
Assumption 3.2. The initial phases
n
of the different paths are independent random variables.
This assumption together with Assumption 3.1 implies the following results:
E[e
jn
e
j
n

] = [n n

], and E[e
jn
e
j
n

] = 0 (3.19)
c _V. V. Veeravalli, 2007 26
where [] is the discrete Dirac function.
We can now compute the autocorrelation function of the proper complex process E(t) as follows:
R
E
(t +, t) = E[E(t +)E

(t)]
= E
_

n
e
jn+2n(t+)

n
e
j
n
2
n
t
_
=

n

2
n
E
_
e
j2n
_
(by (3.19))
=

n

2
n
e
j2n

= R
E
()
(3.20)
This of course implies that E(t) is wide sense stationary.
Result 3.3. The process E(t) may be modelled as a zero-mean, wide sense stationary, proper complex process
over short time horizons corresponding to movements of the order of a few wavelengths in distance, with
autocorrelation function
R
E
() =

2
n
e
j2n
(3.21)
In-Phase and Quadrature Processes
From (3.13) it is clear that the in-phase and quadrature processes E
I
(t) and E
Q
(t) are given by
E
I
(t) =

n

n
cos
n
(t), and E
Q
(t) =

n

n
sin
n
(t) . (3.22)
It is possible to compute the auto- and cross-correlation functions of E
I
(t) and E
Q
(t) from scratch using (3.22)
as is done, for example, in [Pro95, Stu96]. However, if we exploit the fact that E(t) is a proper process, we
immediately have (see (B.12)):
R
E
I
() = R
E
Q
() =
1
2
ReR
E
()
=
1
2

2
n
cos(2
n
) ,
(3.23)
and that
R
E
Q
E
I
() = R
E
I
E
Q
() =
1
2
ImR
E
()
=
1
2

2
n
sin(2
n
) .
(3.24)
Dopper Power Density and Continuum-of-Paths Model
We mentioned earlier in Section 3.3 that the Doppler prole can also be interpreted as a power (gain) density
as a function of Doppler shift, since the power gains
2
n
have been normalized so that

n

2
n
= 1. This
leads to the following denition.
c _V. V. Veeravalli, 2007 27
Denition 3.3. The Doppler power density () is given by
() =

n

2
n
(
n
) . (3.25)
Clearly, as is implied from the term density,
_

+

() d = 1 (3.26)
We can now express the autocorrelation function of E(t) in terms of (). From (3.21) it is easy to see that
R
E
() =
_

+

()e
j2
d . (3.27)
Furthermore, the auto- and cross-correlation function of the inphase and quadrature components can be written
as
R
E
I
() = R
E
Q
() =
1
2
_

+

() cos(2)d ,
R
E
Q
E
I
() = R
E
I
E
Q
() =
1
2
_

+

() sin(2)d .
(3.28)
We may also approximate the discrete density of (3.25) by a continuous density and think of the paths as form-
ing a continuum. The continuous model serves as a good approximation to diffuse scattering that we will study
below in Section 3.4.2. Another justication for the model is that any practical method to measure/estimate
the Doppler prole of the environment will be limited in resolution due to nite windowing in time, and hence
impulses in the Doppler power density will be smoothed out by the measurement process.
It is also clear from (3.27) that () is the Fourier transform of the R
E
(), i.e., that is the power spectral
density of the fading process E(t):
S
E
() = () . (3.29)
This is why the Doppler power density is also called the Doppler power spectral density.
To understand how the Doppler power density manifests itself in the fading channel consider the simple case
where the passband transmitted signal s(t) is a pure carrier without any modulating message signal, i.e.,
s(t) = cos(2f
c
t). Then s(t) = 1 and y(t) = gE(t) (see Figure 3.5). Now consider the passband process
y(t) corresponding to baseband process y(t). Since y(t) is a proper complex process, it follows from (1.6) that
R
y
() = Re
_
R
y
() e
j2fc
_
= g
2
Re
_
R
E
() e
j2fc
_
, (3.30)
and hence that
S
y
(f) =
g
2
2
[S
E
(f f
c
) +S
E
(f f
c
)]
=
g
2
2
[(f f
c
) + (f f
c
)]
(3.31)
This means that the single tone at the carrier frequency is spread out over the range of frequencies dened by
the Doppler power density.
c _V. V. Veeravalli, 2007 28
Special Case of Single Mobile Terminal
For a general (), the integrals in (3.27) and (3.28)) cannot be obtained in closed form. Let us now consider
a special scenario where only one of the terminals is mobile with velocity v, and the scatterers are xed. This
is a reasonable model for certain suburban cellular environments. In this case the Doppler shift along each path
n is completely determined by the angle
n
of made by the path at the mobile terminal as shown in Figure 3.3.
In particular

n
=
max
cos
n
(3.32)
where
max
= v/
c
. Furthermore,
+
=

=
max
.
Since the path angles the mobile determine the time variations in the channel, it is of interest to study the
distribution of power with the angle. To this end we have the following denition.
Denition 3.4. The Angular power density () is given by
() =

2
n
(
n
) . (3.33)
The Doppler power density and the angular power density are related via change of variables as:
() =
_
_
_
(cos
1
(/
max
)) +(cos
1
(/
max
))
_

2
max

2
for [[ <
max
0 otherwise
(3.34)
Note also that we can rewrite (3.27) in terms of () as:
R
E
() =
_
max
max
()e
j2max cos
d . (3.35)
We can rewrite (3.28) similarly.
Now, consider the situation where the propagation environment is such that power is distributed roughly uni-
formly over angles at the mobile terminal. We may approximate such scattering as continuous with a uniform
anglular power density. This leads to the following denition.
Denition 3.5. The fading is isotropic with respect to the mobile terminal if () = 1/2 for all .
For isotropic fading, (3.35) simplies as
R
E
() =
1
2
_

e
2max cos
d
=
1
2
_

cos(2
max
cos )d = J
0
(2
max
)
(3.36)
where J
0
() is the zeroth order Bessel function of the rst kind [AS64].
J
0
(x) =
1
2
_

cos(xcos )d
=
1

_

0
cos(xcos )d =
2

_
/2
0
cos(xcos )d .
(3.37)
Thus
c _V. V. Veeravalli, 2007 29
Result 3.4. For isotropic fading, R
E
() is real-valued:
R
E
() = 2R
E
I
() = J
0
(2
max
) . (3.38)
and
R
E
Q
E
I
() = 0 (3.39)
That is, the in-phase and quadrature processes are uncorrelated.
It is of interest to determine the accuracy of the Bessel ACF model for a real isotropic fading environment. To
this end consider the scenario where there are N discrete paths that have angles that are uniformly distributed
at the mobile terminal. In this case the R
E
() can be computed numerically using
R
E
() =
1
N
N

n=1
e
j2max cos(2/N)
(3.40)
We see in Figure 3.6 that even for N = 8, 16 we get an ACF that is well approximated by a Bessel function.
0 0.5 1 1.5 2 2.5 3
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
Bessel
N=8
N=16
PSfrag replacements

max
or

c
R
E
I
(

)
o
r
R
E
I
(

)
Figure 3.6: Accuracy of the Bessel approximation.
So far we have been able to characterize the mean and autocorrelation function of the at fading process
E(t), using the key assumptions that the path phases are independent and uniformly distributed. Remarkably
these assumptions give us the important properties that E(t) is zero mean and proper. Our goal now is to
make the statistical model for E(t) more precise, but before we do that we distinguish between the cases
where the multipath does or does not have a specular component. If the multipath is produced only from
reections from (rough) objects in the environment, then this form of scattering is said to be purely diffuse
and can be approximated by a continuum of paths, with no one path dominating the others in strength. By
c _V. V. Veeravalli, 2007 30
specular component is meant a subset of paths whose strengths signicantly dominate the other paths. The
most commonly encountered specular path is the line-of-sight (LOS) path, but specular paths may also be
produced by smooth reectors.
3.4.2 PURELY DIFFUSE SCATTERING - RAYLEIGH FADING
Now, if the number of paths is large and they are roughly of equal amplitudes, we may apply the Central Limit
Theorem (see Result B.3) to conclude that E(t) is a proper complex Gaussian (PCG) random process.
Thus, for purely diffuse scattering, we have the following:
Result 3.5. E(t) is well-modelled as a zero-mean, proper complex Gaussian process.
First order statistics of E(t) for purely diffuse scattering
For xed t, E(t) = E
I
(t) +jE
Q
(t) is a proper complex Gaussian random variable with
E
_
[E(t)[
2
_
=

n

2
n
= 1 . (3.41)
Since E(t) is proper, E
I
(t) and E
Q
(t) are uncorrelated and have the same variance, which equals half the
variance of E(t) (see (B.8)). Since E(t) is also Gaussian, E
I
(t) and E
Q
(t) are independent as well. Thus
Result 3.6. For xed t, E
I
(t) and E
Q
(t) are independent ^(0, 1/2) random variables.
Denition 3.6. The envelope process (t) and the phase process (t) are dened by
(t) = [E(t)[ =
_
E
2
I
(t) +E
2
Q
(t) ,
and
(t) = tan
1
_
E
Q
(t)
E
I
(t)
_
.
We can write y(t) of Figure 3.5 in terms of (t) and (t) as:
y(t) = g(t) e
j(t)
s(t) . (3.42)
This means that for at fading, the channel is seen as a single path with gain g(t) and phase shift (t) that
vary with time. Note that and vary much more rapidly than the gains and phases of the individual paths

n
and
n
. For xed t, using the fact that E
I
(t) and E
Q
(t) are independent ^(0, 1/2) random variables, it is
easy to show that (t) and (t) are independent random variables with (t) having a Rayleigh pdf and (t)
being uniform on [0, 2]. The pdf of (t) is given by
p

(x) = 2x e
x
2
u(x) . (3.43)
It can be shown that
E[] =
_
/4 , (3.44)
c _V. V. Veeravalli, 2007 31
0 1 2 3
0
0.2
0.4
0.6
0.8
1
Figure 3.7: Rayleigh pdf
and that
E[
2
] = E
_
[E(t)[
2
_
= 1 . (3.45)
Note that if we scale the envelope by the long-term gain g to produce

= g, then

is also Rayleigh
distributed with E[(

)
2
] = g
2
, i.e.,
p

(x) =
2x
g
2
exp
_

x
2
g
2
_
u(x) (3.46)
Since the envelope has a Rayleigh pdf, purely diffuse fading is referred to as Rayleigh fading. Since the process
E(t) is zero mean and Gaussian, its complete statistical characterization is captured in the autocorrelation
function R
E
(t +, t). Note that
Autocorrelation Function of Envelope and Squared-Envelope
The signal at the output of the at fading channel is given by
y(t) = gE(t)s(t) = g(t) e
j(t)
s(t) . (3.47)
The channel introduces both phase and envelope uctuations in the signal. However, envelope uctuations
directly affect the signal-to-noise ratio at the receiver, and are hence usually more important than phase uctu-
ations in determining communication system performance
1
. It is hence of interest to determine the autocorre-
lation function of the envelope uctuations (t).
Since E[(t)] =
_
/4 (see (3.44)), we have that the autocorrelation function R

() and the autocovariance

function C

() are related as:

R

() = C

() +

4
. (3.48)
The covariance function C

() can be computed exactly but its form is quite complicated and involves a
hypergeometric function [Cla68]. A good approximation for C

() is [Stu96]:
C

()

16
[R
E
()[
2
. (3.49)
1
This is under the assumption that the channel variations can be tracked at the receiver.
c _V. V. Veeravalli, 2007 32
The statistics of the squared-envelope process
2
(t) are also of interest in applications since the instantaneous
power (and SNR) at time t is proportional to
2
(t). The rst-order pdf of
2
(t) is easily computed to be
exponential as shown below:
p

2(x) =
p

x)
2

x
= e
x
u(x) (3.50)
The auto-covariance function of
2
(t) can be computed exactly, and it has a suprisingly simple form:
Result 3.7.
C

2() = [R
E
()[
2
. (3.51)
Proof. The proof follows easily from the fact for zero-mean, real-valued, jointly Gaussian random variables
X
1
, X
2
, X
3
and X
4
E[X
1
X
2
X
3
X
4
] = E[X
1
X
2
]E[X
3
X
4
] +E[X
1
X
3
]E[X
2
X
4
] +E[X
1
X
4
]E[X
2
X
3
]

Now since E[
2
(t)] = 1, the auto-correlation function of
2
(t) is given by
R

2() = 1 +C

2() = 1 +[R
E
()[
2
. (3.52)
We end our discussion of Rayleigh at fading by plotting a typical sample path of a simulated Rayleigh fading
envelope [E(t)[ in Figure 3.8. It can be seen that the fading can be quite severe with path losses of 10 to 20 dB
over very small time scales. We will see in the next section that the fading is not as severe when there is a line
of sight component in the scattering.
3.4.3 SCATTERING WITH A SPECULAR COMPONENT RICEAN FADING
As mentioned earlier, when the scattering has a specular component, there may be more than one specular
(dominant) path. We rst consider the most commonly encountered scenario where there is only specular
path, namely the line-of-sight (LOS) path. Without loss of generality, suppose that the LOS path is given the
index of 1, with angle of arrival
1
, gain
1
and phase process

1
(t) =
1
+ 2
max
t cos
1
. (3.53)
Then the at fading process E(t) of (3.13) gets modied as
E(t) =
1
e
j
1
(t)
+
N

n=2

n
e
jn(t)
. (3.54)
where gain of the LOS path
1
is considerably larger than the gains of the diffuse components
n

N
n=2
. Now,
since the small-scale variations are normalized to have a net average gain of 1, we have
N

n=1

2
n
= 1 , (3.55)
c _V. V. Veeravalli, 2007 33
0 50 100 150 200 250
20
15
10
5
0
5
10
15
t (ms)
C
h
a
n
n
e
l

G
a
i
n

(
d
B
)
Figure 3.8: Rayleigh fading envelope for a carrier frequency of 900 MHz and mobile speed of 72 km/hr.
which implies that
N

n=2

2
n
= 1
2
1
. (3.56)
This means we can write
N

n=2

n
e
jn(t)
=
_
1
2
1

E(t) (3.57)
where

E(t) is a zero mean, proper complex Gaussian, Rayleigh fading process with variance E
_
[

E(t)[
2

= 1,
i.e.,
R

E
() =
_

()e
j2max cos
d , (3.58)
with () being the angular gain density of the diffuse components.
Remark 3.1. It is important to note that E(t) is also a zero-mean process, but it cannot be modelled as
Gaussian since the specular component
1
e
j
1
(t)
dominates the diffuse components in power, and the Central
Limit Theorem does not apply. However, conditioned on
1
, E(t) is indeed a proper complex Gaussian process
with mean
1
e
j
1
(t)
.
The relative strength of the specular component is captured in a quantity called the Rice factor dened as
follows.
Denition 3.7. The Rice factor is dened by
=
power in the specular component
total power in diffuse components
=

2
1
1
2
1
. (3.59)
c _V. V. Veeravalli, 2007 34
From the denition of it follows that

1
=
_

+ 1
, and 1
2
1
=
1
( + 1)
. (3.60)
Thus we have
E(t) =
_

+ 1
e
j
1
(t)
+
_
1
+ 1

E(t) . (3.61)
First Order Statistics of Ricean Fading
For xed t, the pdf of the envelope (t) can be found by rst computing the joint pdf of (t) and (t),
conditioned on
1
. This is facilitated by the fact that conditioned on
1
, E(t) is a proper complex Gaussian
random variable with mean
1
e
j
1
(t)
. We then have:
Result 3.8. The pdf of (t) conditioned on
1
is the Ricean pdf [Ric48], given by
p
|
1
(x) =
2x
1
2
1
I
0
_
2x
1
1
2
1
_
exp
_

x
2
+
2
1
1
2
1
_
u(x) = p

(x) (3.62)
where I
0
() is the zeroth order modied Bessel function of 1st kind [AS64], i.e.,
I
0
(y) =
1
2
_

exp(y cos )d . (3.63)

Notice that the conditional pdf p
|
1
(x) given in (3.62) is not a function of
1
. This means that although E(t)
depends on
1
, the evelope (t) is independent of
1
. Thus the unconditional pdf of (t) is also given by the
Ricean pdf of (3.62). The Ricean pdf can be rewritten in terms of as:
p

(x) = 2x( + 1) I
0
_
2x
_
( + 1)
_
exp
_
x
2
( + 1)

u(x) . (3.64)
It is easy to see that when = 0, p

(x) of (3.64) reduces to a Rayleigh pdf (see (3.43)).

The pdf of
2
(t) is also easily computed as:
p

2(x) =
p

x)
2

x
= ( + 1) I
0
_
2
_
x( + 1)
_
exp [x( + 1) ] u(x) (3.65)
which reduces to the exponential pdf of (3.50) when = 0.
The presence of a LOS component makes the fading more benign, as in seen in Figure 3.10, where typical
sample paths for Ricean envelopes with Rice factors of = 5, 10 are compared with Rayleigh fading.
Correlation and covariance functions for Ricean fading
From (3.61), exploiting the fact that the process
1
(t) is independent of the process

E(t), it is easy to see that
the autocorrelation function of E(t) is given by:
R
E
() = E[E(t +) E

(t)]
=

+ 1
E
_
e
j[
1
(t+)
1
(t)]
_
+
1
+ 1
R

E
()
=

+ 1
e
j2max cos
1
+
1
+ 1
R

E
()
(3.66)
c _V. V. Veeravalli, 2007 35
0 1 2 3
0
0.5
1
1.5
2
Rayleigh
Ricean =1
Ricean =5
Ricean =10
Figure 3.9: Ricean pdf for various Rice factors.
where R

E
() is given in (3.58). Note that the Ricean fading process is also stationary (over short time horizons
corresponding to movements of the order of a few wavelengths in distance.)
For isotropic diffuse components (see (3.36)), we have
R
E
() =

+ 1
e
j2max cos
1
+
1
+ 1
J
0
(2
max
) . (3.67)
Since E(t) is proper, the auto- and cross-correlation functions of the inphase and quadrature processes E
I
(t)
and E
Q
(t) are of course given in directly in terms of the real and imaginary parts of R
E
() (see (B.12)).
The covariance function C

() is not easy to compute for Ricean fading. However, C

2() can be computed

after some tedious but straightforward steps similar to those used in proving (3.51).
Result 3.9.
C

2() =
_
1
+ 1
_
2 _
[R

E
()[
2
+ 2Re
_
R

E
()e
j2max cos
1
__
, (3.68)
Note that (3.68)reduces to the covariance function for Rayleigh fading when = 0. However, for nonzero ,
C

2() ,= [R
E
()[
2
, but rather
C

2() = [R
E
()[
2
+
_

+ 1
_
2
. (3.69)
Scattering With Multiple Specular Paths
If there is more than one specular path, we can rewrite (3.54) as
E(t) =
Ns

n=1

n
e
jn(t)
+
N

n=Ns+1

n
e
jn(t)
(3.70)
where N
s
is the number of specular paths.
c _V. V. Veeravalli, 2007 36
0 50 100 150 200 250
20
15
10
5
0
5
10
15
t (ms)
C
h
a
n
n
e
l

G
a
i
n

(
d
B
)
Rayleigh
Ricean =5
Ricean =10
Figure 3.10: Ricean fading envelopes for a carrier frequency of 900 MHz and mobile speed of 72 km/hr.
A major difference between the case of one specular path and that with multiple specular paths is that the
unconditional distribution of the envelope is no longer Ricean. However, conditioned on the specular phases

1
, . . . ,
Ns
, the process E(t) is proper complex Gaussian with mean

Ns
n=1

n
e
jn(t)
, and the envelope has
a Ricean distribution with Rice factor
=

Ns
n=1

2
n

N
n=Ns+1

2
n
(3.71)
Furthermore, the correlation and covariance functions described above for LOS Ricean fading get appropriately
modied when there are multiple specular components.
3.4.4 FADING ACF AND COHERENCE DISTANCE/TIME
The fading model we have developed so far is for the (small scale) gain variations in time due to the mobile
traveling with some velocity v. We can equivalently treat the gain variations as a random process in a distance
variable . Assuming that we are at distance
0
from some reference point on the trajectory of the mobile at
time t = 0, the distance at time t is given by
=
0
+vt . (3.72)
We can dene the fading process in the distance variable by
E

() = E((
0
)/v) (3.73)
c _V. V. Veeravalli, 2007 37
and the corresponding auto-correlation function over the distance variable is given by
R
E
() = E
_
E

( + )E

()

= R
E
_

v
_
= R
E
_

maxc
_
= J
0
_
2

c
_
(for isotropic fading)
(3.74)
We now introduce the notion of coherence time (distance) which is a measure of how rapidly the channel
varies.
Denition 3.8. Coherence Distance. The coherence distance
c
is a measure how much the MS can move
with E

() being roughly unchanged. We can dene

c
more precisely in terms of the ACF as (say):

c
= largest such that [R
E
()[ > 0.9R
E
(0) = 0.9
The justication for using the absolute value of R
E
() in the above denition is that the squared-envelope
covariance function is a direct function of [R
E
()[
2
(see (3.69)).
We can similarly dene the coherence time T
c
in terms of the ACF R
E
(), but this is not necessary since we
can dene T
c
in terms of
c
.
Denition 3.9. Coherence Time. The coherence time is given by
T
c
=

c
v
.
For isotropic Rayleigh fading, using (3.36), we can show that:

c
0.1
c
, and T
c

0.1

max
. (3.75)
The coherence time of the channel naturally plays an important role in the estimation of the channel the
longer the coherence time, the easier it is to estimate the channel. For digital communications over fading
channels, the relationship between the coherence time T
c
and the symbol duration T
s
is important in designing
signaling strategies and in performance analysis. In this context, we have the following important denition.
Denition 3.10. The fading process is said to be slow if
T
c
T
s
.
The slow fading assumption holds for most modern wireless communication systems. As an example, consider
a system with f
c
= 1 GHz, and suppose the maximum velocity v
max
= 200 km/hr 55 m/s. (It would be
illegal for anyone to be on a terrestrial wireless terminal at speeds greater than this!) Then f
m
200 Hz. If
the symbol rate is greater than 1500 symbols/s (which is true for most applications of interest), then the fading
is slow.
c _V. V. Veeravalli, 2007 38
Angular spread of scattering and coherence distance/time
The ACF of E(t) is a function of the angular spread (and angular location) of the propagation. If the support
of the angular gain density () is restricted to [
1
,
2
], then it follows from (3.27) that
R

E
() =
_

2

1
() e
j2max cos
d . (3.76)
In this case, R
E
() is not necessarily expressible in closed-form, nor is it necessarily real-valued.
The isotropic scattering environment represents one extreme case for which the fading has the smallest coher-
ence time. The other extreme case is when the angular spread is so small that the propagation is effectively on
a single LOS path arriving at angle . In this case,
E(t) e
j(2maxt cos +)
(3.77)
and the channel produces no envelope uctuations in the signal, and simply shifts the frequency by
max
cos
(assuming that the velocity is constant). This means that there is essentially no fading. Note, however, that
it very unlikely that scattering would be diffuse when the angular spread is close to zero this would only
happen in the rare situation where the LOS path is blocked but only one other (diffuse) scatterer exists in the
environment.
The absolute value of the ACF of E(t) for Rayleigh fading is shown for various angular spreads in Figure 3.11.
From this gure it is clear that [R
E
()[, and hence
c
(or T
c
), can be a strong function of the angular spread.
For example in a environment with an angular spread of 30 degrees, the coherence time is 10 times larger than
in an isotropic scattering environment. This dependence of correlation on angular spread will be even more
signicant in the context of frequency selective fading (see Section 3.5).
Finally, coherence times can be considerably larger for Ricean fading with the same angular spreads, as can be
seen in Figure 3.12. This is because the LOS component stabilizes the channel considerably and makes the
fading more benign.
3.4.5 POWER SPECTRAL DENSITY OF FREQUENCY FLAT FADING
For an arbitrary angular gain density (), it can be shown that the power spectral density (PSD) for the
diffuse component of the at fading process can be obtained in closed form by taking the Fourier transform of
the auto-correlation function given in (3.58):
S

E
() =
_
_
_
(cos
1
(/
max
)) +(cos
1
(/
max
))
_

2
max

2
for [[ <
max
0 otherwise
(3.78)
The PSD of for a Ricean fading process with Rice factor and angle of arrival
1
for the LOS component is
then given by the Fourier transform of the ACF given in (3.66):
S
E
() =

+ 1
(
max
cos
1
) +
1
+ 1
S

E
() . (3.79)
c _V. V. Veeravalli, 2007 39
0 1 2 3 4 5
0
0.2
0.4
0.6
0.8
1
1 deg
20 deg
30 deg
120 deg
360 deg
PSfrag replacements

max
or

c
[
R
E
(

)
[
o
r
[
R
E

)
[
Figure 3.11: Correlation functions for various angular spreads around = 45 deg for Rayleigh fading.
0 1 2 3 4 5
0
0.2
0.4
0.6
0.8
1
1 deg
20 deg
30 deg
120 deg
360 deg
PSfrag replacements

max
or

c
[
R
E
(

)
[
o
r
[
R
E

)
[
Figure 3.12: Correlation functions for various angular spreads around = 45 deg for Ricean fading, with the
LOS path at
0
= 45 deg and Rice factor of = 5.
c _V. V. Veeravalli, 2007 40
PSfrag replacements

max

max
0
S

E
(f)
Figure 3.13: Power spectral density of isotropic at fading.
For isotropic fading, the PSD of the diffuse components can be obtained from the Fourier transform of the
Bessel function J
0
in the form given below.
S

E
I
() = S

E
Q
() =
1
2
S

E
() =
_
_
_
1
2

2
max

2
if [[ <
max
0 otherwise
, (3.80)
A plot of this PSD is shown in Figure 3.13.
If the transmitted passband signal is a pure carrier, i.e., s(t) = 1, and the fading is Rayleigh and isotropic, then
from (3.31) we get that the PSD of the passband version y(t) of y(t) is
S
y
(f) =
_

_
g
2
2
_

2
max
(f f
c
)
2
if [f f
c
[ <
max
g
2
2
_

2
max
(f +f
c
)
2
if [f +f
c
[ <
max
0 otherwise
, (3.81)
See page 75 of the book by Jakes [Jak74] for a plot of the ideal isotropic PSD and a comparison with simula-
tions.
3.4.6 SIMULATION OF FREQUENCY FLAT FADING
Method 1: Direct Approach
Pick the number of paths N.
Pick the path gains, delays, and angles of arrival for a discrete set of paths corresponding to desired
angular spread and angular location. For example, to simulate an isotropic Rayleigh fading environment,
set
n
=
_
1/N, and
n
= 2n/N, n = 1, 2, . . . , N.
Approximate the path phase processes
n
(t) by

n
(t) =
n
2
max
t cos
n
,
where
n
are chosen to be i.i.d. Unif[0, 2].
c _V. V. Veeravalli, 2007 41
Sum up paths, i.e., simulate E(t) by

E(t) =
N

n=1

n
e
j

n(t)
. (3.82)
Method 2: Filtered White Gaussian Noise
Find a discrete-time approximation to E(t) by sampling, with sample period T
0
, i.e.,
E[k] E(kT
0
) . (3.83)
Compute the corresponding discrete time autocorrelation function from the (known) continuous-time
ACF R
E
() as
R
E
[m] = E[E[k +m]E

[k]] = R
E
(mT
0
) . (3.84)
Find the square-root lter using spectral factorization [WH85]. That is, the lter H(z) satises the
H(z)H(z
1
) = S
E
(z) , (3.85)
where S
E
(z) is the Z-transform of R
E
[m].
Pass a proper complex white Gaussian sequence, i.e., a sequence of (^(0, 1) random variables, through
the lter to produce samples of E(t).
A Comment on Jakes Simulation Method
Jakes [Jak74] suggested the following method to simulate isotropic Rayleigh at fading (which is valid only
for N such that N/2 is odd). He forms E
J
I
(t) and E
J
Q
(t) using the equations:
E
J
I
(t) =
2

N
_
cos cos 2
max
t +

2
M

n=1
cos
n
cos 2
n
t
_
,
E
J
Q
(t) =
2

N
_
sin cos 2
max
t +

2
M

n=1
sin
n
cos 2
n
t
_
,
(3.86)
where M =
1
2
(
N
2
1),
n
=
max
cos(2n/N), = 0 and
n
= n/M .
The key point to be noted in the above equations is that an isotropic fading environment corresponding to N
paths is being simulated using only about N/4 oscillators. The direct approach would need to have N terms
in the sum for each of

E
I
(t) and

E
Q
(t). This implies a signicant reduction in complexity if we are trying to
simulate fading using hardware (which is what Jakes did!).
Jakes argues that the time correlation properties of the function E
J
(t) generated by (3.86) approximate the
ensemble correlation functions of an isotropic Rayleigh fading process. That is E
J
(t) is a sample path of an
isotropic Rayleigh fading process.
c _V. V. Veeravalli, 2007 42
PSfrag replacements
s(t) y(t)
g
h(t; )
x(t)
Figure 3.14: Channel model for general (possibly frequency selective) fading
If we denote the time average auto-correlation function of a random process X(t) by

R
X
(t + , t), then it is
easy to show that

R
E
J
I
() =
2
N
_
cos
2
cos(2
max
) + 2
M

n=1
cos
2

n
cos(2
n
)
_

R
E
J
Q
() =
2
N
_
sin
2
cos(2
max
) + 2
M

n=1
sin
2

n
cos(2
n
)
_

R
E
J
Q
E
J
I
() =
2
N
[sin cos cos(2
max
)
+2
M

n=1
sin
n
cos
n
cos(2
n
)
_
=

R
E
J
I
E
J
Q
()
(3.87)
Now if we plug in = 0 and
n
= n/M in the above equation, we can show that

R
E
J () is a good
approximation for J
0
(2
max
) for large M.
However, the time correlation functions given in (3.87) are poor approximations to the ensemble correlation
functions of an isotropic Rayleigh fading process for
max
> 1. Furthermore, the complex process that E
J
(t)
represents is not even proper (assuming ergodicity) since

R
E
J
I
() ,=

R
E
J
Q
() and

R
E
J
Q
E
J
I
() =

R
E
J
I
E
J
Q
().
In conclusion, there is really no sense in using Jakes technique for simulating at fading on a computer. The
direct method described above produces much more accurate results and is much more general. A factor of 2
reduction in complexity can be obtained in the direct method if we wish to simulate an isotropic fading process.
3.5 FREQUENCY SELECTIVE FADING
We now revisit the channel model for small-scale variations under different assumptions than the ones we
made in previous section. Recall that the passband signals s(t) and x(t) are assumed to have bandwidth of W,
and hence s(t) and x(t) have (complex) baseband bandwidth of W/2. If W
1

ds
, then the fading is at and
we have the channel model that we studied in Section 3.4 with h(t; ) being replaced by a multiplication by
the at fading process E(t).
Table 3.1, taken from Paulraj and Papadias [PP97], shows the delay spreads for various environments. Most
modern digital communications systems operate at bandwidths of greater than 100 KHz, which means that the
at fading model is rarely useful in practice.
c _V. V. Veeravalli, 2007 43
Environment Spread
Flat Rural 0.5 s
Urban 5 s
Hilly 20 s
Mall 0.3 s
Indoors 0.1 s
Table 3.1: Typical delay spreads for various environments.
If W >
1

ds
, then the fading is said to be frequency selective, and h(t; ) can no longer be replaced by a
multiplication by the at fading process E(t) as in (3.12). The goal of this section is to study how the channel
model should be modied to account for frequency selectivity.
3.5.1 PURELY DIFFUSE SCATTERING
As we did in the study of frequency at fading, we rst characterize the channel for the situation where the
scattering is purely diffuse, i.e., there is no specular component. Recall that time-varying impulse response is
given by
h(t; ) =

n
e
jn(t)
(
n
) (3.88)
where the phase process
n
(t) evolves as:

n
(t)
n
+ 2
n
t
=
n
+ 2
max
t cos
n
(3.89)
with the phases
n
being i.i.d. Unif [, ] (see Assumption 3.2 and (3.14)). We can rewrite (3.88) as
h(t; ) =

n
e
jn
e
j2nt
(
n
) (3.90)
Before we develop a stochastic model for h(t; ), we introduce two additional equivalent represensentations
of the time-varying channel [Bel63].
Denition 3.11. For a channel with impulse response h(t; ), the time-varying transfer function is given by
H(t; f) =
_
h(t; )e
j2f
d
=

n

n
e
jn
e
j2(ntfn)
(3.91)
and the delay-Doppler spreading function is given by
C(; ) =
_
h(t; )e
j2t
dt .
=

n

n
e
jn
(
n
) (
n
)
(3.92)
c _V. V. Veeravalli, 2007 44
Statistical Characterization of H(t; f)
In characterizing the channel stochastically, it is easiest to deal with the time-varying transfer function H(t; f)
since it is written as a sum of complex exponentials (rather than impulses). Since H is a function of two
variables, it must be modelled as a random eld.
The contribution of the n-th path to H(t, f) can be written as

n
(t; f) =
n
e
jn
e
j2(ntfn)
(3.93)
Analogous to the result we derived earlier in the context of frequency at fading (Result 3.2),
n
(t; f) can
easily be seen to be a zero-mean proper complex eld. Now the Central Limit Theorem (Result B.3) can be
applied to obtain:
Result 3.10. For purely diffuse scattering, H(t; f) is well-modelled as a zero-mean, proper complex Gaussian
random eld.
As a consequence, for xed t, f, the magnitude [H(t; f)[ has a Rayleigh distribution and the phase H(t; f)
is uniformly distributed on [, ].
With the zero-mean proper complex Gaussian model for H(t; f), all that is needed to completely characterize
it stochastically is the autocorrelation function. Mimicking the derivation in (3.20), we have
E[H(t
1
; f
1
)H

(t
2
; f
2
)]

2
n
e
j2n(t
1
t
2
)
e
j2n(f
1
f
2
)
. (3.94)
Thus H(t; f) is approximately stationary in both t and f, i.e., it is approximately a homogeneous random eld.
The homogeneous autocorrelation function can then be dened as
R
H
(; ) = E[H(t +; f +)H

(t; f)]

2
n
e
j2n
e
j2n
.
(3.95)
Using the fact that
n
=
max
cos
n
, we can rewrite (3.95) as
R
H
(; ) =

n

2
n
e
j2max cos n
e
j2n
(3.96)
As in (3.35), we can express the autocorrelation function R
H
(; ) in terms of a power gain density. To this
end, we generalize the angular gain density () to a joint density that describes the allocation of power to
both angle of arrival and delay.
Denition 3.12. The joint angle-delay gain density (; ) is given by
(; ) =

n

2
n
(
n
) (
n
) . (3.97)
The discrete joint density of (3.97) can be approximated by continuous density, by considering the paths to
be forming a continuum. As mentioned earlier, the continuous model is representative of diffuse scattering.
Furthermore, the angular gain density () is the marginal of (; ), i.e.,
() =
_

ds
0
(; )d . (3.98)
c _V. V. Veeravalli, 2007 45
Based on (3.97), we can rewrite (3.96) as
R
H
(; ) =
_

_

ds
0
(; ) e
j2max cos
e
j2
dd (3.99)
GWSSUS Model and Scattering Function
The zero-mean proper complex Gaussian eld model for H(t; f) implies that the time-varying transfer func-
tion, h(t; ), and the delay-Doppler spreading function, C(; ), are both also zero-mean proper complex
Gaussian elds.
The autocorrelation function of h(t; ) is easily shown to be
E[h(t
1
;
1
)h

(t
2
;
2
)] =
_

2
n
e
j2n(t
1
t
2
)
(
1

n
)
_
(
1

2
) (3.100)
Note that h(t; ) is not a homogeneous random eld since it is not stationary in the delay variable . How-
ever, it is indeed stationary in the time variable and uncorrelated in the delay variable. The continuous path
generalization of this model for h(t; ) is referred to as the Gaussian Wide Sense Stationary Uncorrelated
Scattering (GWSSUS) model, and was rst studied in great detail by Bello [Bel63]. For the continuous path
generalization, we can rewrite autocorrelation function in terms of the angle-Delay gain density as
E[h(t +;
1
)h

(t;
2
)] =
__

_

ds
0
(; ) e
j2max cos
(
1
) dd
_
(
1

2
) (3.101)
The autocorrelation function of the delay-Doppler spreading function, C(; ), is given by
E[C(
1
;
1
)C

(
2
;
2
)] =
_

2
n
(
1

n
) (
1

n
)
_
(
1

2
) (
1

2
) . (3.102)
Thus, C(; ) is not stationary in or , but it is uncorrelated in both variables. It is interesting to see to that
the quantity inside the square brackets on the RHS of (3.102) describes the distribution of power as a function
of the Doppler frequency and delay. This leads to the following denition.
Denition 3.13. The delay-Doppler scattering function is given by
(; ) =

2
n
(
n
) (
n
) . (3.103)
Thus
E[C(
1
;
1
)C

(
2
;
2
)] = (
1
;
1
) (
1

2
) (
1

2
) . (3.104)
The scattering function can approximated by a continuous (smooth) function for purely diffuse scattering. The
support of the scattering function is restricted to the rectangular region [
max
,
max
] [0,
ds
].
Obviously, there is a one-one relationship between the scattering function, (; ), and the angle-delay gain
density (, ). A straightforward change of variables argument yields
(, ) =
max
sin (
max
cos , ) . (3.105)
c _V. V. Veeravalli, 2007 46
3.5.2 SCATTERING WITH AN SPECULAR COMPONENT
If there is a specular component (with possibly multiple paths) in addition to the diffuse paths, then we can
write the time-varying transfer function as
H(t; f) =
Ns

n=1

n
e
jn
e
j2(ntnf)
+

H(t; f) (3.106)
where

H(t; f) is a zero-mean, proper complex Gaussian, homogeneous eld of the kind described in the
previous section, and

2
= 1
N

n=Ns+1

2
n
. (3.107)
Thus, H(t; f) is a zero-mean, proper complex, homogeneous eld, but is non-Gaussian. However, conditioned
on
1
, . . . ,
Ns
, H(t; f) is a proper complex, non-homogeneous eld, with non-zero mean. Also, just as in
the case of the at fading process E(t), for xed t, f, the distribution of the envelope [H(t; f)[, conditioned on

1
, . . . ,
Ns
, is Ricean with Rice factor given in (3.71). If there is only one specular path, the unconditional
distribution of [H(t; f)[ is also Ricean.
The time-varying impulse response and delay-Doppler spreading functions get modied in a similar fashion:
h(t; ) =
Ns

n=1

n
e
jn
e
j2nt
(
n
) +

h(t; ) (3.108)
where

h(t; ) is a zero-mean, GWSSUS eld.
C(; ) =
Ns

n=1

n
e
jn
(
n
) (
n
) +

C(; ) (3.109)
where

C(; ) is a zero-mean Gaussian eld that is uncorrelated in and .
The autocorrelation functions of H(t; f), h(t; ), and C(; ) can be derived as straightforward extensions to
those obtained previously for purely diffuse scattering.
c _V. V. Veeravalli, 2007 47
Appendix C
Random Fields
Denition C.1. A random eld V (x, y) is said to be homogeneous (or stationary) if for all n, and all choices
of (x
1
, y
1
), . . . , (x
n
, y
n
), and (v
x
, v
y
), the joint distribution of V (x
1
, y
1
), . . . , V (x
n
, y
n
) is the same as that of
V (x
1
+v
x
, y
1
+v
y
), . . . , V (x
n
+v
x
, y
n
+v
y
). That is, the joint distribution is invariant to translations.
Note that for a homogeneous eld, the joint distribution between the eld at two points (say V (x
1
, y
1
) and
V (x
2
, y
2
)) is a function only of the vector joining them, i.e. (x
2
x
1
, y
2
y
1
).
Denition C.2. A random eld V (x, y) is said to be homogeneous and isotropic if for all n, and all choices of
(x
1
, y
1
), . . . , (x
n
, y
n
), the joint distribution of V (x
1
, y
1
), . . . , V (x
n
, y
n
) is the same as that of
V (T
1
(x
1
, y
1
)), . . . , V (T
n
(x
n
, y
n
)), where
T ((x
1
, y
1
), . . . , (x
n
, y
n
)) = (T
1
(x
1
, y
1
), . . . , T
n
(x
n
, y
n
))
is any transformation on the points (x
1
, y
1
), . . . , (x
n
, y
n
) that preserves distances between any pair of these
points. That is, the joint distribution is invariant to rigid-body motions, i.e., translation followed by rotation
(see Figure C.1).
For a homogeneous and isotropic random eld, the joint distribution between the eld at two points (say
V (x
1
, y
1
) and V (x
2
, y
2
)) is a function only of the distance between them. That is
E[V (x
1
, y
1
) V (x
2
, y
2
)] = R
V
(), where =
_
(x
2
x
1
)
2
+ (y
2
y
1
)
2
. (C.1)
c _V. V. Veeravalli, 2007 107
Bibliography
[Cla68] Clarke R. (July-August 1968) A statistical theory of mobile radio reception. Bell Syst. Tech. J. 47(6):
9571000.
[Gud91] Gudmundson M. (1991) Correlation model for shadow fading in mobile radio systems. Electron.
Lett. 27(23): 21452146.
[HJ85] Horn R. A. and Johnson C. R. (1985) Matrix Analysis. Cambridge, New York.
[AS64] Abramowitz M. and Stegun I. A. (1964) Handbook of Mathematical Functions. Dover, New York.
[NM93] Neeser F. D. and Massey J. L. (July 1993) Proper complex random processes with applications to
information theory. IEEE Trans. Inform. Th. 39(4).
[Par98] Parsons J. D. (1998) The Mobile Radio Propagation Channel. Wiley, Chichester, England.
[PP97] Paulraj A. J. and Papadias C. B. (November 1997) Space-time processing for wireless communica-
tions. IEEE Signal Processing Magazine pages 4983.
[Bel63] P. A. Bello. Characterization of randomly time-variant linear channels. IEEE Trans. Commun. Sys-
tems, pages 360393, December 1963.
[Pro95] Proakis J. G. (1995) Digital Communications. Mc-Graw Hill, New York, 3rd edition.
[Rap96] Rappaport T. S. (1996) Wireless Communications: Principles and Practice. Prentice Hall, Upper
Saddle River, NJ.
[Ric48] Rice S. (January 1948) Statistical properties of a sine wave plus noise. Bell Syst. Tech. J. 27(1):
109157.
[Stu96] Stuber G. (1996) Principles of Mobile Communication. Kluwer Academic, Norwell, MA.
[Jak74] Jakes, W. C., Jr. (1974) Microwave Mobile Communications. Wiley, New York.
[WH85] Wong E. and Hajek B. (1985) Stochastic Processes in Engineering Systems. Springer-Verlag, New
York.
[Kai59] Kailath T. (May 1959) Sampling models for linear time-variant lters. Tech. Rep. 352, MIT Research
Laboratory of Electronics.
[Koz97] Kozek W. (Jan. 1997) On the transfer function calculus for underspread LTV channels. IEEE Trans.
Signal Processing pages 219223.
c _V. V. Veeravalli, 2007 109
[SA99] Sayeed A. M. and Aazhang B. (Jan. 1999) Joint multipath-Doppler diversity in mobile wireless com-
munications. IEEE Trans. Commun. 47(1): 123132.
[Sle76] Slepian D. (Mar. 1976) On bandwidth. Proc. IEEE 64(3): 292300.
[AOS98] Aszt ely D., Ottersten B., and Swindlehurst A. L. (Feb. 1998) Generalised array manifold model for
wireless commuinication channels with local scattering. IEE Proc. Radar, Sonar Navig. 145(1): 5157.
[Bre78] Brewer J. W. (Sep. 1978) Kronecker products and matrix calculus in system theory. IEEE Trans. Circ.
and Syst. 25(9): 772781.
[FMB98] Fuhl J., Molisch A. F., and Bonek E. (Feb. 1998) Unied channel model for mobile radio systems
with smart antennas. In IEE Proc. Radar, Sonar Navig., volume 145, pages 3241.
[Fos96] Foschini G. J. (1996) Layered space-time architecture for wireless communication in a fading envi-
ronment when using multi-element antennas. Bell Labs Tech. J. 1(2): 4159.
[GBGP00] Gesbert D., B olcskei H., Gore D. A., and Paulraj A. J. (July 2000) Outdoor MIMO wireless chan-
nels: Models and performance prediction. submitted to the IEEE Trans. Commun. .
[Jak74] Jakes W. C. (1974) Mobile radio propagation. In Jakes W. C. (ed) Microwave Mobile Communica-
tions, pages 1178. Wiley.
[JD93] Johnson D. H. and Dudgeon D. E. (1993) Array Signal Processing. Prentice Hall.
[LZM01] Luo J., Zeidler J. R., and McLaughlin S. (Jan. 2001) Performance analysis of compact antenna
arrays with MRC in correlated nakagami fading channels. IEEE Trans. Vehic. Tech. 50(1): 267277.
[OSW94] Ozarow L. H., Shamai S., and Wyner A. D. (1994) Information theoretic considerations for cellular
mobile radio. IEEE Trans. Vehic. Tech. VT-43: 359378.
[PP97] Paulraj A. J. and Papadias C. B. (Nov. 1997) Space-time processing for wireless communications.
IEEE Signal Processing Magazine pages 4983.
[RC98] Raleigh G. G. and Ciof J. M. (Mar. 1998) Spatio-temporal coding for wireless communication. IEEE
Trans. Commun. pages 357366.
[SFGK00] Shiu D., Foschini G., Gans M., and Kahn J. (Mar. 2000) Fading correlation and its effect on the
capacity of multielement antenna systems. IEEE Trans. Commun. pages 502513.
[SR01] Svantesson T. and Ranheim A. (May 2001) Mutual coupling effects on the capacity of multielement
antenna systems. In Proc. ICASSP 01.
[SW86] Stark H. and Woods J. W. (1986) Probability, Random Processes, and Estimation Theory for Engi-
neers. Prentice Hall, New Jersey.
[Tel95] Telatar E. (June 1995) Capacity of multi-antenna gaussian channels. AT& T-Bell Labs Internal Tech.
Memo. .
[TSC98] Tarokh V., Seshadri N., and Calderbank A. R. (Mar. 1998) Space-time codes for high data rate
wireless communication: Performance criteria and code construction. IEEE Trans. Inform. Th. pages
744765.
c _V. V. Veeravalli, 2007 110
[VP01] Vian J. and Popovi c Z. (June 2001) Smart lens antenna arrays. In IEEE 2001 International Microwave
Symposium. Phoenix.
c _V. V. Veeravalli, 2007 111