Digital Signal Processing Reference
In-Depth Information
2
f
@ w
C
@ w
C
@
where T
W
H
2
H
1
H
2
H
1
and
(
)
denotes
[(
)
]
1
. Since
is Hermitian,
we finally obtain the complex Newton's method given in (1.31). The expression for
Dw
is the conjugate of (1.31).
In [80], it has been shown that the Newton algorithm for
N
complex variables
cannot be written in a form similar to the real-valued case. However, as we have
shown, by including the conjugate of
N
variables, it can be written as shown in
(1.31), a form that is equivalent to the Newton method in
R
2
n
. This form is also
given in [110] using the variables
w
R
and
w
C
, which is shown to lead to the
form given in (1.31) using the same notation in [64]. Also, a quasi-Newton update
is given in [117] by setting the matrix
H
1
to a zero matrix, which might not define a
descent direction for every case, as also noted in [64].
C
N
1.3.3 Matrix Optimization in
Complex Matrix Gradient
Gradient of a matrix-valued variable can also
be written similarly using Wirtinger calculus. For a real-differentiable
f
(
W
,
W
):
C
NN
NN
C
7!
R
, we recall the first-order Taylor series expansion given in (1.20)
@f
@W
@f
@W
þ DW
,
Df DW
,
@f
@W
¼
2Re
DW
,
(1
:
33)
where
@f
@W
is an
NN
matrix whose (
m
,
n
)th entry is the partial derivative of
f
with
respect to
w
mn
. As in the vector case, the matrix gradient with respect to the conjugate
@f
@W
defines the direction of the maximum rate of change in
f
with respect to the
variable
W
.
Complex Relative Gradient Updates
We can use the first-order Taylor
series expansion to derive the relative gradient update rule [21] for complex matrix
variables, which is usually directly extended to the complex case without a derivation
[9, 18, 34]. To write the relative gradient rule, we consider an update of the parameter
matrix
W
in the invariant form
G
(
W
)
W
[21]. We then write the first-order Taylor
series expansion for the change of the form
G
(
W
)
W
as
@f
@W
þ G
(
W
)
W
,
@
f
@W
Df G
(
W
)
W
,
@f
@W
W
H
¼
2Re
G
(
W
),
Search WWH ::
Custom Search