Digital Signal Processing Reference
In-Depth Information
M−
1
p
(
s
k
r
)
f
(
r
)
d
r
=1
−
R
k
k
=0
p
(
s
k
r
)
f
(
r
)
d
r
,
=1
−
R
where
R
is the union of the disjoint regions
R
k
.
The third equality is obtained
by using Bayes' rule, that is,
f
(
r
s
k
)
p
(
s
k
)=
p
(
s
k
r
)
f
(
r
). So we have proved
that the average error probability is
p
(
s
k
r
)
f
(
r
)
d
r
.
P
e
=1
−
(5
.
55)
R
It is now clear that this is minimized by maximizing
p
(
s
k
r
)foreach
k
.Sothe
MAP estimate also minimizes average error probabilities
. Since the ML estimate
agrees with the MAP estimate when all
p
(
s
k
)areequal,the
ML estimate also
minimizes error probabilty when all
s
k
have identical probabilities.
5.5.4 The ML estimate in the Gaussian case
The maximum likelihood (ML) method finds an estimate
s
est
from the mea-
surement
r
such that the conditional probability
f
(
r
s
est
) is maximized. Now
assume that we have an additive white Gaussian noise or
AWGN
channel, that
is,
|
r
(
n
)=
s
(
n
)+
q
(
n
)
,
(5
.
56)
where
q
(
n
) is zero-mean white Gaussian noise with variance
σ
q
.
For a fixed
transmitted symbol
s
(
n
)=
s
k
, we see that
r
(
n
) is a Gaussian random variable
with mean
s
k
. Its density function is therefore
f
(
r
s
k
)=
exp
−
,
s
k
)
2
2
σ
q
1
2
πσ
q
(
r
−
(5
.
57)
where the time argument (
n
) has been omitted for simplicity. Maximizing this
quantity is therefore equivalent to minimizing
D
2
(
r, s
k
)=(
r
s
k
)
2
.
−
(5
.
58)
That is, given the received sample
r
(
n
)attime
n
, the best estimate of the symbol
s
(
n
)attime
n
would be that value
s
k
in the constellation which
minimizes the
distance
D
(
r, s
k
)
.
Similarly, suppose we have received a sequence of
K
samples
r
=[
r
(0)
r
(1)
...
r
(
K
−
1) ]
(5
.
59)
and want to estimate the first
K
symbols
s
=[
s
(0)
s
(1)
...
s
(
K
−
1) ]
(5
.
60)
such that the conditional pdf
f
(
r
s
k
)
(5
.
61)
Search WWH ::
Custom Search