Information Technology Reference
In-Depth Information
Lemma 14. Bin
can be interpreted as Binomials of biased
coins. More precisely, they are distributed like the sum of a Binomial and a scalar.
(
2
, )
and Bin
(
3
, )
1
4
Proof. It is easy to check that Bin
(
3
,
odd
)
1
+
2
·
Bin
(
1
,
)
and Bin
(
2
,
odd
)
3
4 )
1
2 )
Bin
(
1
,
1
)
,aswellasBin
(
3
,
even
)
2
·
Bin
(
1
,
and Bin
(
2
,
even
)
2
·
Bin
(
1
,
.
Corollary 2. E has the distribution of a mixture of a sum of a scalar and a sum of
independent Binomials.
Having this corollary, we can then apply a well known result of Hoeffding [ 8 ].
Theorem 5 (Hoeffding). If X p 1 ,
X p 2 ,...,
X p
are independent Bernoulli trials with
,
,...,
parameters p 1
p 2
p respectively, then
i = 1 X p i c ] P [ b Bin ( , p ) c ] , when 0 b p c ≤ ,
P
[
b
1
i = 1 p i .
where p
=
Recall that we are interested in a lower bound on the median of the independent
sum E
t
i =
t
i =
(
, )
1 E i
1 Bin
d i
. We know that E is a mixture of independent
sums of Bin
, which are (rescaled) biased coins. We finish the
proof of Theorem 2 by proving that every particular independent sum of this mixture
has a median that is
(
2
, )
and Bin
(
3
, )
d
2
. Suppose that we have an independent sum,
Ξ
, consist-
2
ing of r
,
z
,
a
,
w
∈{
0
,
1
,
2
,...}
terms from Bin
(
3
,
odd
)
,Bin
(
3
,
even
)
,Bin
(
2
,
even
)
and
Bin
(
2
,
odd
)
respectively. Notice that 3 r
+
3 z
+
2 a
+
2 w
=
d .
d 2
2
Lemma 15. Amedianof
Ξ
is
.
d
1
Proof. Suppose first that z
r . In that case we show that Med
( Ξ )
. Denote
2
1
4 )+
1
2 )+
3
4 )
by
Ψ
the independent sum Bin
(
r
,
Bin
(
a
,
Bin
(
z
,
.Then
Ψ =
j if and only
if
Ξ =
r
+
2 j
+
w . Thus a median of
Ξ
can be estimated through a median of
Ψ
and
d
1
r
+
2 a
+
3 z
1
so a median of
Ξ
is
if and only if a median of
Ψ
is
. We apply
z r + 2 a + 3 z
, =
2
4
1
Hoeffding's result with p
=
r
+
a
+
z and c
=
r
+
a
+
z
,
b
=
r
+
a
+
4
r +
2 a +
3 z
1
. This gives that
4
r
+
2 a
+
3 z
1
r
+
2 a
+
3 z
1
P
[ Ψ
]
P
[
Bin
(
r
+
a
+
z
,
p
)
]
4
4
r
+
2 a
+
3 z
P
[
Bin
(
r
+
a
+
z
,
p
)
] .
4
r
+
2 a
+
3 z
1
2 .
Hence the lemma will follow once we prove that P
[
Bin
(
r
+
a
+
z
,
p
)
]
4
r
+
2 a
+
3 z
1
2
Note that the mean of Bin
(
r
+
a
+
z
,
p
)
equals
.Now,if z
r then p
and
4
1
thus Bin
(
r
+
a
+
z
,
p
)
is stochastically larger than Bin
(
r
+
a
+
z
,
2 )
. This means that
1
2
(
+
+
,
)
(
+
+
,
)
a median of Bin
r
a
z
p
is bigger than or equal to a median of Bin
r
a
z
.
 
Search WWH ::




Custom Search