Digital Signal Processing Reference
In-Depth Information
The example shows that if we sample from
f
(x(
n
)
j
y(1 :
n
)) a large number
of particles
M
, we will be able to estimate
E
(
h
(X(
n
)) with arbitrary accuracy. In
practice, however, the problem is that we often cannot draw samples directly from
the
a posteriori
PDF
f
(x(
n
)
j
y(1 :
n
)). An attractive alternative is to use the concept
of importance sampling [58]. The idea behind it is based on the use of another function
for drawing particles. This function is called importance sampling function or
proposal distribution, and we denote it by
p
(x(
n
)).
When the particles are drawn from
p
(x(
n
)), the estimate of
E
(
h
(X(
n
))) in (5.11) can
be obtained either by
M
X
M
1
w
(
m
)
(
n
)
h
(x
(
m
)
(
n
))
E
(
h
(X(
n
)))
(5
:
15)
m¼
1
or by
E
(
h
(X(
n
)))
X
M
w
(
m
)
(
n
)
h
(x
(
m
)
(
n
))
(5
:
16)
m¼
1
where
f
(x
(
m
)
(
n
)
j
y(1 :
n
))
p
(x
w
(
m
)
(
n
)
¼
(5
:
17)
(
m
)
(
n
))
and
w
(
m
)
(
n
)
P
i¼
1
w
(
i
)
(
n
)
w
(
m
)
(
n
)
¼
(5
:
18)
where
w
(
m
)
(
n
)
¼ cw
(
m
)
(
n
)
with
c
being some unknown constant. The symbols
w
(
m
)
(
n
) and
w
(
m
)
(
n
) are known as
true and normalized importance weights of the particles x
(
m
)
(
n
), respectively. They are
introduced to correct for the bias that arises due to sampling from a different function
than the one that is being approximated,
f
(x(
n
)
j
y(1 :
n
)). The estimate in (5.15) is
unbiased whereas the one from (5.16) is with a small bias but often with a smaller
mean-squared error than the one in (5.15) [55]. An advantage in using (5.16)
over (5.15) is that we only need to know the ratio
f
(x(
n
)
j
y(1 :
n
))
=p
(x(
n
)) up to a
multiplicative constant and not the exact ratio in order to compute the estimate of
the expectation of
h
(X(
n
)).
How is (5.18) obtained? Suppose that the true weight cannot be found and instead
we can only compute it up to a proportionality constant, that is
(
m
)
(
n
)
j
y(1 :
n
))
p
(x
w
(
m
)
(
n
)
¼ c
f
(x
(
m
)
(
n
))
¼ cw
(
m
)
(
n
)
(5
:
19)
Search WWH ::
Custom Search