Information Technology Reference
In-Depth Information
This last result is of a monotonicity theorem first proved by A. Stam in
1959 [222]. It is also presented and discussed in [50]. Sometimes this (and
other) results are expressed in terms of the
entropy power
defined as
N
(
X
)=
1
2
πe
e
2
H
S
(
X
)
. We now present a Corolary of this theorem justifying larger
Shannon entropy of Gaussian smoothed distributions (see Chap. 3).
Corollary B.1.
Let
X
and
Y
be independent continuous random variables
and
f
Z
=
f
X
⊗
f
Y
.If
Y
is Gaussian distributed with variance
h
2
,then
H
S
(
Z
)=
H
S
(
X
+
Y
)
≥
H
S
(
X
)
.
(B.3)
Proof.
From the monotonicity theorem we have:
N
(
X
+
Y
)
≥
N
(
X
)+
N
(
Y
)
.
(B.4)
2
πe
e
2ln(
h
√
2
πe
)
=
2
πe
e
2
H
(
Y
)
=
1
1
Since
Y
is a Gaussian r.v. we also have
N
(
Y
)=
h
2
>
0, and the above result follows.
B.2 Rényi's Entropy
α
ln
1
1
f
α
(
x
)
dx
=
[
f
α−
1
(
x
)]
,α
H
R
α
(
X
)=
α
ln
E
≥
0
,α
=1
1
−
1
−
X
(B.5)
A list of important properties of Rényi's entropy [47, 62] is:
1.
H
R
α
can be positive or negative, but
H
R
2
is non-negative with minimum
value (0) corresponding to a Dirac-
δ
comb.
2. Invariance to translations:
H
R
α
(
X
+
c
)=
H
R
α
(
X
) for a constant
c
.
3. Change of scale:
H
R
α
(
aX
)=
H
R
α
(
X
)
α
) for a constant
a
.
4. With the conditional entropy defined similarly as for the Shannon entropy
(Property 4 of Sect. B.1), the inequality
H
R
α
(
X
−
ln
|
a
|
/
(1
−
|
Y
)
≤
H
R
α
(
X
) only holds
for
α
≤
1 and
f
(
x
)
≤
1 in the whole support.
5.
H
R
α
(
X
1
,...,X
n
)=
i
=1
H
R
α
(
X
i
) for independent random variables.
6. For
α
=2and univariate distributions with finite variance
σ
2
,themaxi-
mizer of Rényi's quadratic entropy is
Γ
(2)
1
with support
x
2
5
σ
2
Γ
(5
/
2)
1
√
5
πσ
−
√
5
σ
. The general formulas of the maximizing density of the
α
-Rényi
entropy are given in [47]. Note that the Rényi entropy of the univariate
|
x
|≤
normal distribution [162] is
H
R
α
(
g
(
x
;
μ, σ
)) = ln(
√
2
πσ
)
1
−
2
ln
α/
(1
−
α
);
therefore,
H
R
2
(
g
(
x
;
μ, σ
)) = ln(2
σ
√
π
).
7.
H
R
α
(
X
)
α→
1
−−−→
H
S
(
X
).
Note that a result similar to the one in Corolary B.1 for Rényi's quadratic
entropy is a trivial consequence of properties 1 and 5.