Biomedical Engineering Reference
In-Depth Information
The mutual information between
x
and
y
is defined as
I(
x
,
y
)
=
H(
x
)
+
H(
y
)
−
H(
x
,
y
).
(C.46)
When
x
and
y
are independent, we have
I(
x
,
y
)
=
0 according to Eq. (
C.43
). Let us
T
. Assuming the Gaussian processes for
x
and
y
, the mutual
information is expressed as
x
T
y
T
define
z
as
z
=[
,
]
I(
x
,
y
)
=
H(
x
)
+
H(
y
)
−
H(
x
,
y
)
2
log
ʣ
yy
−
1
2
log
1
1
2
log
=
|
ʣ
xx
| +
|
ʣ
zz
|
.
(C.47)
Here,
|
ʣ
zz
|
is rewritten as
=
ʣ
yy
ʣ
xx
−
ʣ
xy
ʣ
−
1
xy
,
ʣ
xx
ʣ
xy
ʣ
T
|
ʣ
zz
| =
yy
ʣ
(C.48)
T
xy
ʣ
yy
where the determinant identity in Eq. (
C.94
) is used. Substituting Eq. (
C.48
)into
(
C.47
), we have
1
2
log
|
ʣ
xx
|
ʣ
xx
−
ʣ
xy
ʣ
−
1
xy
I(
x
,
y
)
=
T
yy
ʣ
1
2
log
1
xy
=
.
(C.49)
−
ʣ
−
1
xx
ʣ
xy
ʣ
−
1
I
yy
ʣ
ʣ
−
1
xx
ʣ
xy
ʣ
−
1
xy
as
ʣ
ʳ
j
where
j
=
,...,
Let us define the eigenvalues of
1
d
and
yy
.
6
Using these eigenvalues, we can derive
d
=
min
{
p
,
q
}
1
2
log
1
xy
I(
x
,
y
)
=
−
ʣ
−
1
xx
ʣ
xy
ʣ
−
1
T
I
yy
ʣ
d
1
2
log
1
j
=
1
(
1
2
1
=
=
log
−
ʳ
j
.
(C.50)
1
1
−
ʳ
j
)
j
=
1
2
Note that
ʳ
1
is equal to the canonical squared correlation
ˁ
c
, according to the argu-
ments in Sect.
C.3.1
.
When the random vectors
x
and
y
are complex-valued, the mutual information for
the complex random vectors is expressed as
log
ʣ
yy
−
I(
x
,
y
)
=
H(
x
)
+
H(
y
)
−
H(
x
,
y
)
=
log
|
ʣ
xx
| +
log
|
ʣ
zz
|
.
(C.51)
6
Here, remember that
p
and
q
are the sizes of the column vector
x
and
y