Professional Documents
Culture Documents
Spring 2006
Problem Set 5: Solutions
1.i) Exercise 3.2, Part 1:
Let = SNR. For Rayleigh fading |h[0]| 2 Exp(1) so we have:
Pe
i Z
h p
2
2|h[0]| =
= E Q
Z t2 /(2)
2x
1
2
et /2 ex dtdx
2
1
2
et /2 ex dxdt =
2
i
h
1
2
2
et /2 1 et /(2) dt
2
0
0
0
s
"
#
r
Z r
1
1 + t2 (1+1/)
1
1
1
e 2
dt =
1
2
1 +
2
2
1 + 1/
We can approximate
p
1/(1 + x) = 1 x/2 + o(x) for x 0 1 . Then,
1
11
1
+ o(1/) =
+ o(1/)
11+
Pe =
2
2
4
and
1
4
lim Pe =
Part 2:
We will need the following result:
Z
Q( y)dy =
1
2
et /2 dtdy =
2
Z t2
0
1
2
et /2 dydt =
2
limx0 o(x)/x = 0
Q(
2x)f (x)dx
1
t2
2
et /2 dt =
2
2
Assuming that f () is right continuous at 0, that f (0) > 0 and that f () is bounded (this last
condtion enables us to use the bounded convergence theorem to exchange limit and integral):
lim Pe =
lim
f (0)
2
Z
p
Q( 2x)f (x)dx = lim
Q( y)f
y
2
1
dy
2
f (0)
Q( y)dy =
4
of independent random variables equals the convolution of the corresponding pdfs we can write for
x 0:
f2 (x) =
f3 (x) =
Z
Z
0
x
f2 (t)g3 (x t)dt =
x2
+ o(x2 )
2
..
.
fL (x) =
x
0
L
Y
gl (0)
l=1
where we defined = [
x
0
"
L1
Y
gl (0)
l=1
#
tL2
L2
+ o(t
) gL (x t)dt
(L 2)!
xL1
+ o(xL1 ) = xL1 + o(xL1 )
(L 1)!
QL
l=1 gl (0)]/(L
h p
i Z
2khk2 =
Pe = E Q
Q(
p
2x)fL (x)dx
Multiplying by L , taking limit for and assuming that we can exchange the order
of limits and integrals we have:
lim Pe L =
lim
Q(
Z
p
2x)L fL (x)dx = lim
Q( y)L1 fL
y
2
1
dy
2
lim
2L
2L
y
Z
y L1 1
y L1 1
f
L
2
Q( y) L1
Q( y)
dy =
dy
2
2
2
2
y
0
2
1
2
et /2 y L1 dtdy = L
2
2
2L
t
0
2
1
1
et /2 dt = L
L 2
2 2L
1 (2L)!
=
2L 2L L!2L
2L 1
L
Z t2
0
1
2
et /2 y L1 dydt
2
2
1
t2L et /2 dt
2
L
1 Y
gl (0)
4L
l=1
L (2L 1)!
(2L 1)!
= L
L
4
L!
4 (L 1)!
We observe that this scheme still achieves a diversity gain of L but the error performance
Q
degrades by the factor L/ = L!/( L
i=1 gi [0]) = L! with respect to that of optimal combining.
2. Exercise 3.16:
1. We have
y = Ad + w,
X
=
ai di + w.
i
Since A is orthogonal, all the ai s are orthogonal, thus for detecting d i , we can project along
ai :
ai y = ||ai ||2 di + ai w.
(1)
Since A is orthogonal, the noise ai w is independent of other noise terms (and hence the other
projections).Thus, each of the di s can be decoded separately.
2. If ||am || = ||h||, then we can normalize the equation (1) and get a fading coefficient of ||h||
which implies a full diversity gain for each symbol.
3. We have
h X = d t At ,
which along with orthogonality and the full diversity property of A implies that
h XX h = ||d||2 ||h||2 IL ,
h (XX ||d||2 IL )h = 0,
for every h. Thus, XX must be ||d||2 IL .
3. Exercise 3.19:
1. Let H be the fading matrix for the MIMO channel. Then the channel model can be written
as:
Y = HX + W.
Now, this channel model can be rewritten as a MISO channel with block-length n r N . Let X
and h be
=
X
0
0
h =
0
..
.
0
0 ,
X
with y and w appropriately defined in term of Y and W. Then the probability of pairwise
error can be written as:
s
E Q
SNRh(XA XB )(XA XB ) h
2
2. Since we have reduced the MIMO problem with i.i.d. Rayleigh fading to a MISO problem
with i.i.d. Rayleigh fading, probability of pairwise error can be upper bounded as:
P(XA XB )
4Lnr
,
4L
SNRL det ((XA XB )(XA XB ) )
nr
where the last step follows from the diagonal structure of XA and XB .
3. Thus, the code design criterion of maximizing the minimum determinant remains unchanged.
4. Exercise 3.26:
1. With x[0] = 1,
det(XA XB )(XA XB ) 4L
and
P {XA XB } SNRL
for any pair of codewords that differ in the first component. Hence by the union bound, the
probability of error p0 on the first sy mbol is
p0 < 2(L1) SNRL (
1 L
) .
2SNR
2. To get the same rate using the naive scheme, one has to use 2 L - PAM. The distance between
constellation points is of the orde r of 2 L . Hence , the error probability is of the order of
4 2L
SNR
5
L
Hence, the first scheme uses a factor of 2 (L+1) less energy than the naive scheme for the
same error probability. This coding ga in is exponential in L (linear in L in dB.)
3. Even if we are trying to calculate the error probability for a middle stream, the determinant
is still lower bounded by 4L (consider the first non-zero stream). Then, the probability of
error is upper bounded by: 2(N 1) SNRL .