EECS 224B: Fundamentals of Wireless Communications

Spring 2006
Problem Set 6: Solutions
1) Exercise 3.24:
1. Assume uniform scattering around the mobile. In this case the fading correlation is independent of the direction of movement, so we can assume that the mobile is moving in the same
direction as where the second antenna would be located. The mobile reaches the location
of the second antenna after traveling a distance d, after a time interval d/v where v is the
speed of movement. The correlation of the fading gains between these 2 locations is given by
R[bd/v W c], where 1/W is the sampling interval of the discrete time model.
The fading gains are zero mean, circularly symmetric jointly complex Gaussian, so their joint
distribution is completely determined by their correlation matrix. Letting h = [h 1 h2 ]T be the
gains at the locations of the 2 antennas at a given time, we have:
h CN (0, Kh )

(1)

where
Kh =

R[0]
R[bd/v W c]
R[bd/v W c]
R[0]

(2)

2. The received signal at the 2 antennas is y[n] = hx[n] + w[n] where w[n] CN (0, N 0 I2 ).
p
Conditioned on h the error probability for BPSK is Q( 2khk2 SNR) and the average error
probability is:

i
h p
Pe = E Q( 2khk2 SNR)

(3)

where the expectation is taken with respect to the distribution of h which was found in part
(1).
3. The problem of directly doing a high SNR approximation in the computation of (3) is that
the 2 components of h are correlated. However, noting that any nonsingular transformation
of y[n] is a sufficient statistic, we can do a transformation that decorrelates the entries of h.
Let Kh = UU where U is unitary and is diagonal with non-negative entries. We
[n] =
can always find this decomposition because K h is positive semidefinite. Then define y
1

U y = U h[n]x[n] + U w[n] = h[n]x[n]

+ w[n].
It follows that h[n]
CN (0, ) and

w[n]
CN (0, N0 I2 ). Now we can use the error probability expression found in part 2) but
with uncorrelated fading gains, which in the case of circularly symmetric complex Gaussian
random variables implies independence, and use a high SNR approximation for the Q function:
Pe

 q

h
i
2
2

= E Q
2khk SNR
E ekhk SNR
h
i h
i
2
2
= E e|h1 | SNR E e|h2 | SNR =

1
1
(1 SNR + 1) (2 SNR + 1)

(4)

where 1 and 2 are the diagonal elements of (the eigenvalues of K h ). These eigenvalues
can be computed explicitly. Assuming R[0] = 1 and letting = R[bd/v W c] we obtain
1 = 1 + || and 2 = 1 ||. If || > 0 then i > 0 (i = 1, 2), and we can further approximate
(4) for high SNR:
Pe

1
1
2 =
1 2 SNR
(1 ||2 )SNR2

(5)

In this case we get a diversity gain of 2 and the correlation between antennas increases the
error probability by the factor 1/(1 || 2 ) as compared to the uncorrelated antenna case. If
on the other hand || = 1 (perfect correlation between antennas) then 2 = 0 and Pe

1
2SNR .

In this case the diversity gain reduces to 1.

As we increase the antenna separation d the correlation || decreases since the correlation
function R[m] is monotonically decreasing, and the probability of error decreases.
2) Exercise 3.28:
L1
The DFT of the fading coefficients {h` }`=0
after extending the sequence adding N L zeros

is:
L1
1 X

h` ej2`n/N
hn =
N `=0

(6)

L1
n }N 1 are circularly
for n = 0, 1, . . . , N 1. Since {h` }`=0
are i.i.d. CN (0, 1/L) it follows that { h
n=0

symmetric jointly complex Gaussian, so their statistics are completely specified by the correlation
function
] =
n+r h
R[r] = E[h
n

L1 L1
2
1 XX
E[hm h` ]ej N [m(n+r)nl]
N
m=0
`=0

L1
1 X j 2 `r
1 sin
e N =
NL
N L sin
`=0

rL
N

r
N

ej

r(L1)
N

(1 [r]) +

1
[r]
N

(7)

valid of |r| (N 1)/2.

The coherence bandwidth is given by W c = W/(2L) and the tone spacing is W/N , so the
coherence bandwidth expressed in number of tones is N/(2L). We expect the gains of the carriers
separated by more than the coherence bandwidth to be approximately independent. To state this
more precisely, we want to show that |R[r]|  R[0] for N/(2L) |r| (N 1)/2, where we have
assumed N to be odd for simplicity, and is some constant, say 10. Note that the carrier gains are
periodic in n so the correlation function is also periodic in r with period N , so we need to consider
values of r only within one period.
To prove this statement we need to upper bound |R[r]| for N/(2L) |r| (N 1)/2. We




| 1 and | sin(x)| 2|x|/ for |x| /2, so sin r
first note that | sin rL
N
N 2|r|/N /L in
the range of interest. Therefore for N/(2L) |r| (N 1)/2,
|R[r]|

1 L
1
1
=
= R[0]
NL
N

(8)

It follows that the correlation decays at least as 1 for tones separated by times the
coherence bandwidth.
3) Exercise 3.30:
We will need to solve first Exercise 2.16:
N
X

h` [m] =

i=1

Let =

1
N

PN

i=1 i (0)

h` [m] = e

j2fc

ai (m/W )ej2fc i (m/W ) sinc(` i (m/W )W )

and i (m/W ) = i (m/W ) . Then,

N
X
i=1

ai (m/W )ej2fc i (m/W ) sinc(` W i (m/W )W )

Often in practice fc fc r/c >> 1

so it is a reasonable assumption to model e j2fc =

ej where U nif orm[0, 2] and is independent of everything else. Note that does not
1

r is the distance between transmit and receive antennas

depend on m so a particular realization of is the same for all components of h. Since e j has
uniformly distributed phase, its distribution does not change if we introduce an arbitrary phase
shift . So ej ej ej .
It follows that

ej h

ej ej

=d ej

=d h

PN

ai (m/W )ej2fc i (m/W ) sinc(` W i (m/W )W )

PN
j2fc i ((m+1)/W ) sinc(`
W i ((m + 1)/W )W )
i=1 ai ((m + 1)/W )e
..
.
PN
j2fc i ((m+n)/W ) sinc(`
a
((m
+
n)/W
)e
W i ((m + n)/W )W )
i=1 i
i=1

PN

j2fc i (m/W ) sinc(`

W i (m/W )W )
i=1 ai (m/W )e
PN
j2f
c i ((m+1)/W )
sinc(` W i ((m + 1)/W )W )
i=1 ai ((m + 1)/W )e

..
.

PN

i=1 ai ((m

+ n)/W )ej2fc i ((m+n)/W ) sinc(` W i ((m + n)/W )W )

Since this is true for all , under the previous assumptions h is circularly symmetric.
We now solve Exercise 3.30:
1. We know that the taps h` [i]s are circularly symmetric. For fixed time i and different values
of ` the taps are independent random variables. This follows because the different taps
correspond to different signal paths which experience independent reflections (attenuations
and phase shifts). The complex gain of the n th carrier at the ith OFDM symbol, for an OFDM
block length of N and L taps is given by:
L1
X
n [i] = 1
h` [i]ej2`n/N
h
N `=0

(9)

The multiplication by the complex exponential does not modify the distribution of h ` [i]. Also,
since the L terms in the sum are independent, the multiplication by the complex exponentials
does not modify the joint distribution of the different terms, and hence does not modify
the distribution of the sum. Since the dependence on n only appears through the complex
exponential, and we can remove the complex exponentials without modifying the distribution
n [i], it follows that the {h
n [i]}N are identically distributed for fixed i.
of h
n=1

2. From the text, we know that the effect of movement of paths from tap to tap is negligible
compared to the variation of the taps due to Doppler spread whenever f c  W . Making
this assumption the movement of paths between taps occurs in a time scale much larger than
the variation due to Doppler shifts, and therefore we have that the processes {h ` [m]]}m are
independent across `. Now we can use an argument similar to that of part (1) but using
vectors.
n [1] . . . h
n [r]]T , n = 1, 2, . . . , N , and h` = [h` [1]h` [2] . . . h` [r]]T , ` = 0, 1, . . . , L 1.
n = [h
Let h
We can write:
L1
X
n = 1
h` ej2n`/N
h
N n=0

(10)

We can use the result of Exercise 2.16 to conclude that the vectors h ` , ` = 0, 1, . . . , L 1
are circularly symmetric. Hence the distribution of each of the terms in the above sum
is not modified by the product with the complex exponential. Also the initial observation
implies that the vectors h` , ` = 0, 1, . . . , L 1 are independent, and therefore the sum (10)
is not modified by the product with the complex exponentials. Since the dependence on n
is only through the complex exponentials, and these can be removed without modifying the
n }N are identically distributed. Since r is arbitrary
distribution of the sum, it follows that { h
n=1
the result follows.
4) Exercise 3.27:

1. The channel matrix is



h0 0
h1 h0

2. The ZF equalizer would only look at y[0] since there is interference on the y[1] component.
But then it makes an error whene ver h 0 is in deep fade. Hence, its diversity is 1. This is less
than the diversity gain of 2 for the ML approach.