Digital Signal Processing Reference
In-Depth Information
where
'
t
(
W
)
¼
log
p
(
x
(
t
)
jW
)
¼
log
p
S
(
Wx
)
þ
log
j
det
Wj
and the density of the transformed random variables is written through the compu-
tation of the Jacobian as
p
(
x
)
¼j
det
Wjp
S
(
Wx
)
(1
:
48)
where
W
is defined in (1.11).
We use the notation that
p
S
(
Wx
)
W
Q
n¼
1
p
S
n
(
w
n
x
), where
w
n
is the
n
th row of
W
,
p
S
n
(
u
n
)
W
p
S
n
(
u
n
r
,
u
n
i
) is the joint pdf of source
n
,
n ¼
1,
...
,
N
, with
u
n
¼ u
n
r
þ ju
n
i
,
and defined
W ¼ A
2
1
, that is, we express the likelihood in terms of the inverse mixing
matrix, which provides a convenient change of parameter. Note that the time index in
x
(
t
) has been omitted in the expressions for simplicity.
To
take
advantage
of Wirtinger
calculus, we write
each
pdf
as
N
N
N
p
S
n
(
u
r
,
u
i
)
¼ g
n
(
u
,
u
) to define
g
(
u
,
u
):
C
C
7!
R
so that we can directly
evaluate
@
log
g
(
u
,
u
)
@W
¼
@
log
g
(
u
,
u
)
x
H
W
c
(
u
,
u
)
x
H
(1
:
49)
@u
where
u ¼ Wx
and we have defined the score function
c
(
u
,
u
) that is written directly
by using the result in Brandwood's theorem given by (1.5)
1
2
@
log
p
S
(
u
r
,
u
i
)
@u
r
þ j
@
log
p
S
(
u
r
,
u
i
)
c
(
u
,
u
)
¼
:
(1
:
50)
@u
i
When writing (1.49) and (1.50), we used a compact vector notation where each
element of the score function is given by
c
n
(
u
,
u
)
¼
@
log
g
n
(
u
n
,
u
n
)
1
2
@
log
p
S
n
(
u
r
,
n
,
u
i
,
n
)
@u
r
,
n
þ j
@
log
p
S
n
(
u
r
,
n
,
u
i
,
n
)
@u
i
,
n
¼
:
@u
n
(1
:
51)
To compute
@
log
j
det
Wj=@W
, we first observe that
@
log
j
det
Wj¼
Trace(
W
1
@W
)
¼
Trace(
@WPP
1
W
1
), and then choose
1
2
I jI
jII
P ¼
to write
@
log
j
det
Wj¼
Trace(
W
1
@W
)
þ
Trace((
W
)
1
@W
)
¼hW
H
,
@WiþhW
T
,
@W
i:
(1
:
52)
Search WWH ::
Custom Search