Information Technology Reference
In-Depth Information
Appendix B
Properties of Differential Entropy
B.1 Shannon's Entropy
H
S
(
X
)=
−
f
(
x
)ln
f
(
x
)
dx
=
−
E
[
lnf
(
X
)]
.
(B.1)
X
A list of important properties of Shannon's entropy [48, 184, 62] is:
1.
H
S
(
X
)
∈
]
−∞
,
ln
X
],where
X
is the support length of
X
. The equal-
ity
H
S
(
X
)=
holds for a uniform distribution in a bounded support.
The minimum value (
X
) corresponds to a sequence of continuous Dirac-
δ
functions (Dirac-
δ
comb).
2. Invariance to translations:
H
S
(
X
+
c
)=
H
S
(
X
) for a constant
c
.
3. Change of scale:
H
S
(
aX
)=
H
S
(
X
)+ln
−∞
|
a
|
for a constant
a
.If
X
is a
random vector,
H
S
(
A
X
)=
H
S
(
X
)+ln
|
det
A
|
.
4. Conditional entropy:
H
S
(
X
|
Y
)=
−
E
[ln
f
(
X
|
Y
)]
≤
H
S
(
X
) with equality
iff
X
and
Y
are independent.
5. Sub-additivity for joint distributions:
H
S
(
X
1
,...,X
n
)=
i
=1
H
S
(
X
i
|
X
1
,
≤
i
=1
H
S
(
X
i
), with equality (additivity) only if the r.v.s are
independent.
6. Bijective transformation
Y
=
ϕ
(
X
):
H
S
(
Y
)=
H
S
(
X
)
...,X
i−
1
)
−
E
X
[ln
|
J
ϕ
(
Y
)
|
],
where
J
ϕ
(
Y
)=
∂ϕ
−
1
(
y
i
)
∂y
k
,
i, k
=1
,...,d
, is the Jacobian of the transfor-
mation. Note that this implies properties 2 and 3, and also the invariance
under an orthonormal transformation
Y
=
A
X
,with
|
A
|
=1.
7. If a random vector
X
with support in
R
n
has covariance matrix
Σ
,then
1
2
ln[(2
πe
)
n
H
S
(
X
)
≤
|
Σ
|
]. The equality holds for the multivariate Gaussian
distribution.
8. Let
X
1
,...,X
n
be independent r.v.s with densities and finite variances.
Then,
n
e
2
H
S
(
X
1
+
...
+
X
n
)
e
2
H
S
(
X
i
)
.
≥
(B.2)
i
=1