Biomedical Engineering Reference
In-Depth Information
Following the derivation in Sect. C.3.1, the solution of the above optimization
problem is expressed as the eigenvalue problem,
xy
T
2
ʠ
xy
ʠ
ʱ
=
ˈ
I
ʱ
,
(7.82)
ʠ
xy
=
ʓ
−
1
/
2
ʥ
xy
ʓ
−
1
/
2
T
where
. In the above equation, since the matrix,
ʠ
xy
ʠ
xy
is
xx
yy
a real symmetric matrix, the eigenvector
ʱ
is real-valued. The vector
ʲ
is obtained
using
1
ˈ
I
ʥ
T
ʲ
=
xy
ʱ
,
which is also real-valued.
The matrices
T
ʓ
−
1
xx
ʥ
xy
ʓ
−
1
T
xy
have the same eigenvalues, accord-
ing to Sect. C.8 (Property No. 9). Let us define the eigenvalues of the matrix
ʓ
−
1
ʠ
xy
ʠ
xy
and
yy
ʥ
xx
ʥ
xy
ʓ
−
1
T
2
I
yy
ʥ
xy
as
ʶ
j
(
j
=
1
,...,
d
). The canonical imaginary coherence
ˈ
is derived as
2
I
=
S
max
{
ʓ
−
1
xx
ʥ
xy
ʓ
−
1
T
ˈ
yy
ʥ
xy
}=
ʶ
1
.
(7.83)
Using the same arguments in the preceding section, the imaginary-coherence-based
mutual information is given by
d
1
I
s
(
,
)
=
−
ʶ
j
.
x
y
log
(7.84)
1
j
=
1
The alternative definition of canonical imaginary coherence, based on
I
s
(
x
,
y
)
,is
obtained as
d
ˈ
2
I
=
1
−
exp
[−
I
s
(
x
,
y
)
]=
1
−
1
(
1
−
ʶ
j
),
(7.85)
j
=
which uses all the eigenvalues
ʶ
1
,...,ʶ
d
. When we can assume
ʶ
j
1 (where
j
=
1
,...,
d
), we have
d
ˈ
2
I
≈
1
ʶ
j
.
(7.86)
j
=
The connectivity metric above has been proposed and called the multivariate inter-
action measure (MIM) in [
12
].
The mutual information independent of the sizes of vectors is defined such that,
d
1
d
1
I
s
(
x
,
y
)
=
log
−
ʶ
j
.
(7.87)
1
j
=
1