INEQUALITIES INVOLVING THE MEAN AND
THE STANDARD DEVIATION OF NONNEGATIVE
REAL NUMBERS
OSCAR ROJO
Received 22 December 2005; Revised 18 August 2006; Accepted 21 Septe mber 2006
Let m(y)
=
n
j
=1
y
j
/n and s(y) =
m(y
2
) −m
2
(y) be the mean and the standard deviation
of the components of the vector y
= (y
1
, y
2
, , y
n−1
, y
n
), where y
q
= (y
q
1
, y
q
2
, , y
q
n
−1
, y
q
n
)
with q a positive integer. Here, we prove that if y
≥ 0,thenm(y
2
p
)+(1/
√
n −1)s(y
2
p
) ≤
m(y
2
p+1
)+(1/
√
n −1)s(y
2
p+1
)forp = 0,1,2, The equality holds if and only if
the (n
− 1) largest components of y are equal. It follows that (l
2
p
(y))
∞
p=0
, l
2
p
(y) =
(m(y
2
p
)+(1/
√
n −1)s(y
2
p
))
2
−p
, is a strictly increasing sequence converging to y
1
,the
largest component of y,exceptifthe(n
−1) largest components of y are equal. In this
case, l
2
p
(y) = y
1
for all p.
Copyright © 2006 Oscar Rojo. This is an op en access article distributed under the Cre-
ative Commons Attribution License, which permits unrestricted use, distribution, and
reproduction in any medium, provided the original work is properly cited.
1. Introduction
Let
m(x)
=
n
j
=1
x
j
n
, s(x)
=
m
x
2
−
m
2
(x) (1.1)
be the mean and the standard deviation of the components of x
= (x
1
,x
2
, , x
n−1
,x
n
),
where x
q
= (x
q
1
,x
q
2
, , x
q
n
−1
,x
q
n
) for a positive integer q.
The following theorem is due to Wolkowicz and Styan [3, Theorem 2.1.].
Theorem 1.1. Let
x
1
≥ x
2
≥···≥x
n−1
≥ x
n
. (1.2)
Hindawi Publishing Corporation
Journal of Inequalities and Applications
Volume 2006, Article ID 43465, Pages 1–15
DOI 10.1155/JIA/2006/43465
2 Inequalities on the mean and standard deviation
Then
m(x)+
1
√
n −1
s(x)
≤ x
1
, (1.3)
x
1
≤ m(x)+
√
n −1s(x). (1.4)
Equality holds in ( 1.3)ifandonlyifx
1
= x
2
=···=x
n−1
. Equality holds in (1.4)ifandonly
if x
2
= x
3
=···=x
n
.
Let x
1
,x
2
, ,x
n−1
,x
n
be complex numbers such that x
1
is a positive real number and
x
1
≥
x
2
≥···≥
x
n−1
≥
x
n
. (1.5)
Then,
x
p
1
≥
x
2
p
≥···≥
x
n−1
p
≥
x
n
p
(1.6)
for any positive integer p.WeapplyTheorem 1.1 to (1.6)toobtain
m
|
x|
p
+
1
√
n −1
s
|
x|
p
≤
x
p
1
,
x
p
1
≤ m
|
x|
p
+
√
n −1s
|
x|
p
,
(1.7)
where
|x|=(|x
1
|,|x
2
|, , |x
n−1
|,|x
n
|).
Then,
l
p
(x) =
m
|
x|
p
+
1
√
n −1
s
|
x|
p
1/p
(1.8)
isasequenceoflowerboundsforx
1
and
u
p
(x) =
m
|
x|
p
+
√
n −1s
|
x|
p
1/p
(1.9)
isasequenceofupperboundsforx
1
.
We recall that the p-norm and the infinity-norm of a vector x
= (x
1
,x
2
, , x
n
)are
x
p
=
n
i=1
x
i
p
1/p
,1≤ p<∞,
x
∞
= max
i
x
i
.
(1.10)
It is well known that lim
p→∞
x
p
=x
∞
.
Oscar Rojo 3
Then,
l
p
(x) =
⎛
⎜
⎝
x
p
p
n
+
1
n(n −1)
x
2p
2p
−
x
2p
p
n
⎞
⎟
⎠
1/p
,
u
p
(x) =
⎛
⎜
⎝
x
p
p
n
+
n −1
n
x
2p
2p
−
x
2p
p
n
⎞
⎟
⎠
1/p
.
(1.11)
In [2, Theorem 11], we proved that if y
1
≥ y
2
≥ y
3
≥···≥y
n
≥ 0, then
m
y
2
p
+
√
n −1s
y
2
p
≥
m
y
2
p+1
+
√
n −1s
y
2
p+1
(1.12)
for p
= 0,1,2, The equality holds if and only if y
2
= y
3
=···=y
n
. Using this inequal-
ity, we proved in [2, Theorems 14 and 15] that if y
2
= y
3
=···=y
n
,thenu
p
(y) = y
1
for all p,andify
i
<y
j
for some 2 ≤ j<i≤ n,then(u
2
p
(y))
∞
p=0
is a strictly decreasing
sequence converging to y
1
.
The main purpose of this paper is to prove that if y
1
≥ y
2
≥ y
3
≥···≥y
n
≥ 0, then
m
y
2
p
+
1
√
n −1
s
y
2
p
≤
m
y
2
p+1
+
1
√
n −1
s
y
2
p+1
(1.13)
for p
= 0,1,2, The equality holds if and only if y
1
= y
2
=···=y
n−1
. Using this in-
equality, we prove that if y
1
= y
2
=···=y
n−1
,thenu
p
(y) = y
1
for all p,andify
i
<y
j
for
some 1
≤ j<i≤ n −1, then (l
2
p
(y))
∞
p=0
is a strictly increasing sequence converging to y
1
.
2. New inequalities involving m(x) and s(x)
Theorem 2.1. Let x
= (x
1
,x
2
, , x
n−1
,x
n
) be a vector of complex numbers s uch that x
1
is a
positive real number and
x
1
≥
x
2
≥···≥
x
n−1
≥
x
n
. (2.1)
The sequence (l
p
(x))
∞
p=1
converges to x
1
.
Proof. From (1.11),
l
p
(x) ≥
x
p
p
√
n
∀p. (2.2)
Then, 0
≤|l
p
(x) −x
1
|=x
1
−l
p
(x) ≤ x
1
−x
p
/
p
√
n for all p.Sincelim
p→∞
x
p
= x
1
and lim
p→∞
p
√
n=1, it follows that the sequence (l
p
(x)) converges and lim
p→∞
l
p
(x) =x
1
.
We introduce the following notations:
(i) e
=(1,1, ,1),
(ii) Ᏸ
= R
n
−{λe :λ ∈ R},
(iii) Ꮿ
={x =(x
1
,x
2
, , x
n
):0≤ x
k
≤ 1, k = 1,2, ,n},
4 Inequalities on the mean and standard deviation
(iv) Ᏹ
={x =(1,x
2
, , x
n
):0≤ x
n
≤ x
n−1
≤···≤x
2
≤ 1},
(v)
x,y=
n
k
=1
x
k
y
k
for x,y ∈ R
n
,
(vi)
∇g(x) =(∂
1
g(x),∂
2
g(x), ,∂
n
g(x)) denotes the gradient of a differentiable func-
tion g at the point x,where∂
k
g(x) is the partial derivative of g with respect to x
k
,
evaluated at x.
Clearly, if x
∈ Ᏹ,thenx
q
∈ Ᏹ with q a positive integer.
Let v
1
,v
2
, ,v
n
be the points
v
1
= (1,0, ,0),
v
2
= (1,1,0, ,0),
v
3
= (1,1,1,0, ,0),
.
.
.
v
n−2
= (1,1, ,1,0,0),
v
n−1
= (1,1, ,1,1,0),
v
n
= (1,1, ,1,1) =e.
(2.3)
Observe that v
1
,v
2
, , v
n
lie in Ᏹ.Foranyx =(1,x
2
,x
3
, ,x
n−1
,x
n
) ∈Ᏹ,wehave
x
=
1 −x
2
v
1
+
x
2
−x
3
v
2
+
x
3
−x
4
v
3
+ ···+
x
n−2
−x
n−1
v
n−2
+
x
n−1
−x
n
v
n−1
+ x
n
v
n
.
(2.4)
Therefore, Ᏹ is a convex set. We define the function
f (x)
= m(x)+
1
√
n −1
s(x), (2.5)
where x
= (x
1
,x
2
, , x
n
) ∈R
n
.Weobservethat
ns
2
(x) =
n
k=1
x
2
k
−
n
j
=1
x
j
2
n
=
n
k=1
x
k
−m(x)
2
=
x−m(x)e
2
2
.
(2.6)
Then,
f (x)
= m(x)+
1
n(n −1)
x−m(x)e
2
=
n
j
=1
x
j
n
+
1
n(n −1)
n
k=1
x
2
k
−
n
j
=1
x
j
2
n
.
(2.7)
Next, we give properties of f . Some of the proofs are similar to those in [2].
Oscar Rojo 5
Lemma 2.2. The function f has continuous first partial derivatives on Ᏸ,andforx
=
(x
1
,x
2
, , x
n
) ∈Ᏸ and 1 ≤k ≤ n,
∂
k
f (x) =
1
n
+
1
n(n −1)
x
k
−m(x)
f (x) −m(x)
, (2.8)
n
k=1
∂
k
f (x) =1, (2.9)
∇
f (x),x
=
f (x). (2.10)
Proof. From (2.7), it is clear that f is differentiable at every point x
= m(x )e,andfor
1
≤ k ≤n,
∂
k
f (x) =
1
n
+
1
n(n −1)
x
k
−
n
j
=1
x
j
/n
n
i
=1
x
2
i
−
n
j
=1
x
j
2
/n
=
1
n
+
1
n(n −1)
x
k
−m(x)
f (x) −m(x)
,
(2.11)
which is a continuous function on Ᏸ. Then,
n
k
=1
∂
k
f (x) =1. Finally,
∇
f (x),x
=
n
k=1
x
k
∂
k
f (x)
=
n
k
=1
x
k
n
+
1
n(n −1)
n
k
=1
x
2
k
−m(x)
n
k
=1
x
k
f (x) −m(x)
= m(x)+
1
n(n −1)
x −a(x)e
2
= f (x).
(2.12)
This completes the proof.
Lemma 2.3. The function f is convex on Ꮿ.Moreprecisely,forx,y ∈ Ꮿ and t ∈[0,1],
f
(1 −t)x + ty
≤
(1 −t) f (x)+tf(y) (2.13)
with equality if and only if
x
−m(x)e =α
y −m(y)e
(2.14)
for some α
≥ 0.
Proof. Clearly Ꮿ is a convex set. Let x, y
∈ Ꮿ and t ∈[0,1]. Then,
f
(1 −t)x + ty
=
m
(1 −t)x + ty
+
1
n(n −1)
(1 −t)x + ty −m
(1 −t)x + ty
e
2
=(1−t)m(x)+tm(y )+
1
n(n−1)
(1−t)
x−m(x)e
+t
y−m(y)e
2
.
(2.15)
6 Inequalities on the mean and standard deviation
Moreover ,
(1 −t)
x −m(x)e
+ t
y −m(y)e
2
2
= (1 −t)
2
x −m(x)e
2
2
+2(1−t)t
x −m(x)e,y −m(y)e
+ t
2
y −m(y)e
2
2
.
(2.16)
We recall the Cauchy-Schwarz inequality to obtain
x −m(x)e,y −m(y)e
≤
x −m(x)e
2
y −m(y)e
2
(2.17)
with equality if and only if (2.14)holds.Thus,
(1 −t)
x −m(x)e
+ t
y −m(y)e
2
≤ (1−t)
x −m(x)e
2
+ t
y −m(y)e
2
(2.18)
with equality if and only if (2.14)holds.Finally,from(2.15)and(2.18), the lemma fol-
lows.
Lemma 2.4. For x,y ∈ Ᏹ −{e},
f (x)
≥
∇f (y),x
(2.19)
with equality if and only if (2.14) holds for some α>0.
Proof. Ᏹ isaconvexsubsetofᏯ and f isaconvexfunctiononᏱ.Moreover, f isadiffer-
entiable function on Ᏹ
−{e}.Letx,y ∈ Ᏹ −{e}.Forallt ∈[0, 1],
f
tx+(1 −t)y
≤
tf(x)+(1−t) f (y). (2.20)
Thus, for 0 <t
≤ 1,
f
y + t(x −y)
−
f (y)
t
≤ f (x) − f (y). (2.21)
Letting t
→ 0
+
yields
lim
t→0
+
f
y + t(x −y)
−
f (y)
t
=
∇
f (y),x −y
≤
f (x) − f (y). (2.22)
Hence,
f (x)
− f (y) ≥
∇f (y),x
−
∇
f (y),y
. (2.23)
Now, we use the fact that
∇f (y),y=f (y)toconcludethat
f (x)
≥
∇f (y),x
. (2.24)
The equality in all the above inequalities holds if and only if x
−a(x)e =α(y −m(y)e)for
some α
≥ 0.
Oscar Rojo 7
Corollar y 2.5. For x
∈ Ᏹ −{e},
f (x)
≥
∇f
x
2
,x
, (2.25)
where
∇f (x
2
) is the gradient of f with respect to x evaluated at x
2
. The equality in (2.25)
holds if and only if x is one of the following convex combinations:
x
i
(t) =te+(1 −t)v
i
, i =1, 2, ,n −1, some t ∈[0,1). (2.26)
Proof. Let x
= (1,x
2
,x
3
, , x
m
) ∈Ᏹ −{e}.Then,x
2
∈ Ᏹ −{e}. Using Lemma 2.4,weob-
tain
f (x)
≥
∇f
x
2
,x
(2.27)
with equality if and only if
x
−m(x)e =α
x
2
−m
x
2
e
(2.28)
for some α
≥ 0. Thus, we have proved (2.25). In order to complete the proof, we observe
that condition (2.28)isequivalentto
x
−αx
2
= m
x −αx
2
e (2.29)
for some α
≥ 0. Since x
1
= 1, (2.29)isequivalentto
1
−α =x
2
−αx
2
2
= x
3
−αx
2
3
=···=x
n
−αx
2
n
(2.30)
for some α
≥ 0. Hence, (2.28)isequivalentto(2.30).
Suppose that (2.30)istrue.Ifα
= 0, then 1 = x
2
=···=x
n
. This is a contradiction
because x
= e,thusα>0.
If x
2
= 0, then x
3
= x
4
=···=x
n
= 0, and thus x =v
1
.Let0<x
2
< 1. Suppose x
3
<x
2
.
From (2.30),
1
−x
2
= α
1+x
2
1 −x
2
,
x
2
−x
3
= α
x
2
+ x
3
x
2
−x
3
.
(2.31)
From these equations, we obtain x
3
= 1, which is a contradiction. Hence, 0 <x
2
< 1im-
plies x
3
= x
2
.Now,ifx
4
<x
3
,fromx
2
= x
3
and the equations
1
−x
2
= α
1+x
2
1 −x
2
,
x
3
−x
4
= α
x
3
+ x
4
x
3
−x
4
,
(2.32)
we obtain x
4
= 1, which is a contradiction. Hence, x
4
= x
3
if 0 <x
2
< 1. We continue in
this fashion to conclude that x
n
= x
n−1
=···=x
3
= x
2
.Wehaveprovedthatx
1
= 1and
0
≤ x
2
< 1implythatx =(1,t, ,t) =te +(1−t)v
1
for some t ∈ [0,1). Let x
2
= 1.
8 Inequalities on the mean and standard deviation
If x
3
= 0, then x
4
= x
5
=···=x
m
= 0, and thus x =v
2
.Let0<x
3
< 1andx
4
<x
3
.
From (2.30),
1
−x
3
= α
1+x
3
1 −x
3
,
x
3
−x
4
= α
x
3
+ x
4
x
3
−x
4
.
(2.33)
From these equations, we obtain x
4
= 1, which is a contradiction. Hence, 0 <x
3
< 1im-
plies x
4
= x
3
.Now,ifx
5
<x
4
,fromx
3
= x
4
and the equations
1 −x
3
= α
1+x
3
1 −x
3
,
x
4
−x
5
= α
x
4
+ x
5
x
4
−x
5
,
(2.34)
we obtain x
5
= 1, which is a contradiction. Therefore, x
5
= x
4
. We continue in this fashion
to get x
n
= x
n−1
=···=x
3
.Thus,x
1
= x
2
= 1, and 0 ≤x
3
< 1 implies that x = (1,1,t, ,t)
= te+(1 −t)v
2
for some t ∈ [0,1).
For 3
≤ k ≤ n −2, arguing as above, it can be proved that x
1
= x
2
=···=x
k
= 1and
0
≤ x
k+1
< 1 implies that x = (1, ,1,t, ,t) = te+(1 −t)v
k
.Finally,forx
1
= x
2
=···=
x
n−1
= 1and0≤x
n
< 1, we have x =te + v
n−1
.
Conversely, if x is any of the convex combinations in (2.26), then (2.30)holdsby
choosing α
= 1/(1 + t).
Let us define the following optimization problem.
Problem 2.6. Let
F :
R
n
−→ R (2.35)
be given by
F(x)
= f
x
2
−
f (x)
2
. (2.36)
We want to find m in
x∈Ᏹ
F(x). That is, find
minF(x) (2.37)
subject to the constraints
h
1
(x) =x
1
−1 =0,
h
i
(x) =x
i
−x
i−1
≤ 0, 2 ≤ i ≤ n,
h
n+1
(x) =−x
n
≤ 0.
(2.38)
Lemma 2.7. (1) If x
∈ Ᏹ −{e}, then
n
k
=1
∂
k
F(x) ≤0 with equality if and only if x is one of
the convex combinations x
k
(t) in (2.26).
(2) If x
= x
N
(t) with 1 ≤N ≤ n −2, then
∂
1
F(x) =···=∂
N
F(x) > 0, (2.39)
∂
N+1
F(x) =···=∂
n
F(x) < 0. (2.40)
Oscar Rojo 9
Proof. (1) The function F has continuous first partial derivatives on Ᏸ,andforx
∈ Ᏸ
and 1
≤ k ≤n,
∂
k
F(x) =2x
k
∂
k
f (x
2
) −2 f (x)∂
k
f (x). (2.41)
By (2.9),
n
k=1
∂
k
F(x) =2
n
k=1
x
k
∂
k
f
x
2
−
2 f (x)
n
k=1
∂
k
f (x)
= 2
∇
f
x
2
,x
−
2 f (x).
(2.42)
It follows from Corollary 2.5 that
n
k
=1
∂
k
F(x) ≤ 0 with equality if and only if x
i
= te+
(1
−t)v
i
, i = 1, ,n−1.
(2) Let x
= x
N
(t)with1≤ N ≤ n −2fixed.Then,x = te+(1 −t)v
N
,somet ∈ [0,1).
Thus, x
1
= x
2
=···=x
N
= 1, x
N+1
= x
N+2
=···=x
n
= t.FromTheorem 1.1, f (x) < 1.
Moreover ,
f (x)
−m(x) =
1
n(n −1)
N +(n −N)t
2
−
N +(n −N)t
2
n
=
1
n(n −1)
nN + n(n −N)t
2
−N
2
−2N(n −N)t −(n −N)
2
t
2
n
=
1
n
√
n −1
N(n −N)(1−t).
(2.43)
Replacing this result in (2.8), we obtain
∂
1
f (x) =∂
2
f (x) =···=∂
N
f (x)
=
1
n
+
1
n(n −1)
1
−m(x)
f (x) −m(x)
=
1
n
+
1
√
n −1
1
−
N +(n −N)t
/n
N(n −N)(1−t)
=
1
n
+
1
√
n −1n
√
n −N
√
N
> 0.
(2.44)
Similarly,
f
x
2
−
m
x
2
=
1
n
√
n −1
N(n −N)
1 −t
2
,
∂
1
f
x
2
=
∂
2
f
x
2
=···=
∂
N
f
x
2
=
1
n
+
1
n
√
n −1
√
n −N
√
N
> 0.
(2.45)
10 Inequalities on the mean and standard deviation
Therefore,
∂
1
F(x) =∂
2
F(x) =···=∂
N
F(x)
= 2∂
1
f
x
2
−
2 f (x)∂
1
f (x) =2
1 − f (x)
∂
1
f (x) > 0.
(2.46)
We have thu s proved (2.39). We easily see that
∂
N+1
F(x) =∂
N+2
F(x) =···=∂
n
F(x). (2.47)
We have
n
k
=1
∂
k
F(x) =0. Hence,
n
k=N+1
∂
k
F(x) =(n −N)∂
N+1
F(x) =−
N
k=1
∂
k
F(x) < 0. (2.48)
Thus, (2.40)follows.
We recall the following necessary condition for the existence of a minimum in nonlin-
ear programming.
Theorem 2.8 (see [1, Theorem 9.2-4(1)]). Let J : Ω
⊆ V → R be a function defined over
an open, convex subset Ω of a Hilbert space V and let
U
=
v ∈Ω : ϕ
i
(v) ≤0, 1 ≤i ≤m
(2.49)
be a subset of Ω, the constraints ϕ
i
: Ω → R, 1 ≤ i ≤ m, being assumed to be convex. Let
u
∈ U be a point at which the functions ϕ
i
, 1 ≤i ≤m,andJ are differentiable. If the function
J has at u a relative minimum with respect to the set U and if the constraints are qualified,
then there exist numbe rs λ
i
(u), 1 ≤i ≤m, such that the Kuhn-Tucker conditions
∇J(u)+
m
i=1
λ
i
(u)∇ϕ
i
(u) =0,
λ
i
(u) ≥0, 1 ≤i ≤m,
m
i=1
λ
i
(u)ϕ
i
(u) =0
(2.50)
are satisfied.
The convex constraints ϕ
i
in the above necessary condition are said to be qualified if
either all the functions ϕ
i
are affine and the set U is nonempt y, or there exists a point
w
∈ Ω such that for each i, ϕ
i
(w) ≤0 with strict inequality holding if ϕ
i
is not affine.
The solution to Problem 2.6 is given in the following theorem.
Theorem 2.9. One has
min
x∈Ᏹ
F(x) =0 =F(1,1,1, ,1,t) (2.51)
for any t
∈ [0,1].
Oscar Rojo 11
Proof. We observe that Ᏹ is a compact set and F isacontinuousfunctiononᏱ.Then,
there exists x
0
∈ Ᏹ such that F(x
0
) = min
x∈Ᏹ
F(x). The proof is based on the applica-
tion of the necessary condition given in the preceding theorem. In Problem 2.6,wehave
Ω
= V =
R
n
with the inner product x,y=
n
k
=1
x
k
y
k
, ϕ
i
(x) = h
i
(x), 1 ≤i ≤ n +1,U =
Ᏹ and J = F. T he functions h
i
,2≤ i ≤ n + 1, are linear. Therefore, they are convex and
affine. In addition, the function h
1
(x) = x
1
−1isaffine and convex and Ᏹ is nonempty.
Consequently, the functions h
i
,1≤i ≤ n + 1, are qualified. Moreover, these functions and
the objective function F are differentiable at any point in Ᏹ
−{e}. The gradients of the
constraint functions are
∇h
1
(x) =(1,0,0,0, ,0) = e
1
,
∇h
2
(x) =(−1,1,0,0, ,0),
∇h
3
(x) =(0,−1,1,0, ,0),
.
.
.
∇h
n−1
(x) =(0,0, ,0,−1,1,0),
∇h
n
(x) =(0,0, ,0,−1,1),
∇h
n+1
(x) =(0,0, ,0,−1).
(2.52)
Suppose that F has a relative minimum at x
∈ Ᏹ−{e} with respect to the set Ᏹ.Then,
there exist λ
i
(x) ≥ 0(forbrevityλ
i
= λ
i
(x)), 1 ≤ i ≤ n + 1, such that the Kuhn-Tucker
conditions
∇F(x)+
n+1
i=1
λ
i
∇h
i
(x) =0,
n+1
i=1
λ
i
h
i
(x) =0
(2.53)
hold. Hence,
∇F(x)+
λ
1
−λ
2
,λ
2
−λ
3
,λ
3
−λ
4
, , λ
n
−λ
n+1
=
0, (2.54)
λ
2
x
2
−1
+ λ
3
x
3
−x
2
+ ···+ λ
n
x
n
−x
n−1
+ λ
n+1
−
x
n
=
0. (2.55)
From (2.55), as λ
i
≥ 0, 1 ≤ i ≤ n +1,and0≤x
n
≤ x
n−1
≤···≤x
2
≤ 1, we have
λ
k
x
k−1
−x
k
=
0, 2 ≤k ≤ n, λ
n+1
x
n
= 0. (2.56)
Now, from (2.54),
n
k=1
∂
k
F(x)+λ
1
−λ
n+1
= 0. (2.57)
We will conclude that λ
1
= 0 by showing that the cases λ
1
> 0, x
n
> 0andλ
1
> 0, x
n
= 0
yield contradictions.
12 Inequalities on the mean and standard deviation
Suppose λ
1
> 0andx
n
> 0. In this case, λ
n+1
x
n
= 0 implies λ
n+1
= 0. Thus, (2.57)
becomes
n
k=1
∂
k
F(x) =−λ
1
< 0. (2.58)
We apply Lemma 2.7 to conclude that x is not one of the convex combinations in (2.26).
From (2.4),
x
=
1 −x
2
v
1
+
x
2
−x
3
v
2
+
x
3
−x
4
v
3
+ ···+
x
n−2
−x
n−1
v
n−2
+
x
n−1
−x
n
v
n−1
+ x
n
v
n
.
(2.59)
Then, there are at least two indexes i, j such that
1
=···=x
i
>x
i+1
=···=x
j
>x
j+1
. (2.60)
Therefore,
∂
1
F(x) =···=∂
i
F(x),
∂
i+1
F(x) =···=∂
j
F(x).
(2.61)
From (2.56), we get λ
i+1
= 0andλ
j+1
= 0. Now, from (2.54),
∂
i
F(x) =−λ
i
≤ 0,
∂
i+1
F(x) =λ
i+2
≥ 0,
∂
j
F(x) =−λ
j
≤ 0,
∂
n
F(x) =−λ
n
≤ 0.
(2.62)
The above equalities and inequalities together with (2.8)and(2.41)give
1
n
1 − f (x)
+
1
n(n −1)
1 −m
x
2
f
x
2
−
m
x
2
−
1 −m(x)
f (x) −m(x)
≤
0, (2.63)
1
n
1 − f (x)
+
1
n(n −1)
x
2
j
−m
x
2
f
x
2
−
m
x
2
−
x
j
−m(x)
f (x) −m(x)
=
0, (2.64)
1
n
1 − f (x)
+
1
n(n −1)
x
2
n
−m
x
2
f
x
2
−
m
x
2
−
x
n
−m(x)
f (x) −m(x)
≤
0. (2.65)
Subtracting (2.64)from(2.63)and(2.65), we obtain
1
−x
2
j
f
x
2
−
m
x
2
≤
1 −x
j
f
x
2
−
m
x
2
,
x
2
n
−x
2
j
f
x
2
−
m
x
2
≤
x
n
−x
j
f
x
2
−
m
x
2
.
(2.66)
Oscar Rojo 13
Dividing these inequalities by (1
−x
j
)and(x
n
−x
j
), respectively, we get
1+x
j
f
x
2
−
m
x
2
≤
1
f
x
2
−
m
x
2
,
x
n
+ x
j
f
x
2
−
a
x
2
≥
1
f
x
2
−
a
x
2
.
(2.67)
The last two inequalities imply x
n
≥ x
j
, which is contradiction.
Suppose now that λ
1
> 0andx
n
= 0. Let l be the largest index such t hat x
l
> 0. Thus,
x
l+1
= 0. From (2.55),
λ
2
x
2
−1
+ λ
3
x
3
−x
2
+ ···+ λ
l
x
l
−x
l−1
+ λ
l+1
−
x
l
=
0. (2.68)
Then,
λ
k
x
k−1
−x
k
=
0, 2 ≤k ≤ l, λ
l+1
x
l
= 0. (2.69)
Hence, λ
l+1
= 0. If l = n −1, then λ
n
= 0and∂
n
F(x) =λ
n+1
≥ 0. If l ≤ n −2, then ∂
l
F(x) =
−
λ
l
≤ 0. In both situations, we conclude that x is not one of the convex combinations in
(2.26). Therefore, there are at least two indexes i, j such that
1
=···=x
i
>x
i+1
=···=x
j
>x
j+1
. (2.70)
Now, we repeat the argument used above to get that x
l
≥ x
j
, which is a contradiction.
Consequently, λ
1
= 0. From (2.57),
n
k=1
∂
k
F(x) =λ
n+1
≥ 0. (2.71)
We apply now Lemma 2.7 to conclude that x is one of the convex combinations in (2.26).
Let x
= x
N
(t) =te+(1 −t)v
N
,1≤ N ≤ n −2, and t ∈ [0,1). Then, x
1
= x
2
=···=x
N
= 1,
x
N+1
= x
N+2
=···=x
n
= t,andh
N+1
(x) = t −1 < 0. From (2.56), we obtain λ
N+1
= 0.
Thus, from (2.54), ∂
N+1
F(x) =λ
N+2
≥ 0. This contradicts (2.40). Thus, x = x
N
(t)forN =
1,2, ,n−2andt ∈ [0,1). Consequently, x =x
n−1
(t) =(1,1, ,1,t)forsomet ∈ [0,1).
Finally,
F(1,1, ,1,t)
= f
1,1, ,1,t
2
−
f (1,1, ,1,t)
2
= 1 −1 =0 (2.72)
for any t
∈ [0,1]. Hence, min
x∈Ᏹ
F(x) = 0 = F(1,1, ,1,t)foranyt ∈ [0,1]. Thus, the
theorem has been proved.
Theorem 2.10. If y
1
≥ y
2
≥ y
3
≥···≥y
n
≥ 0, then
m
y
2
p
+
1
√
n −1
s
y
2
p
≤
m
y
2
p+1
+
1
√
n −1
s
y
2
p+1
, (2.73)
14 Inequalities on the mean and standard deviation
that is,
n
k
=1
y
2
p
k
n
+
1
n(n −1)
n
k=1
y
2
p+1
k
−
n
k
=1
y
2
p
k
2
n
≤
⎡
⎢
⎢
⎢
⎣
n
k
=1
y
2
p+1
k
n
+
1
n(n −1)
n
k=1
y
2
p+2
k
−
n
k
=1
y
2
p+1
k
2
n
⎤
⎥
⎥
⎥
⎦
1/2
(2.74)
for p
= 0,1,2, The equality holds if and only if y
1
= y
2
=···=y
n−1
.
Proof. If y
1
= 0, then y
2
= y
3
=···= y
n
= 0 and the theorem is immediate. Hence, we
assume that y
1
> 0. Let p be a nonnegative integer and let x
k
= y
k
/y
1
for k = 1,2, ,n.
Clearly, 1
= x
2
p
1
≥ x
2
p
2
≥ x
2
p
3
≥···≥x
2
p
n
≥ 0. From Theorem 2.9,wehave
f
1,x
2
p
2
,x
2
p
3
, , x
2
p
m
2
≤ f
1,x
2
p+1
2
,x
2
p+1
3
, , x
2
p+1
m
, (2.75)
that is,
⎛
⎜
⎜
⎜
⎝
1+
n
k
=2
x
2
p
k
n
+
1
n(n −1)
1+
n
k=2
x
2
p+1
k
−
1+
n
j
=2
x
2
p
j
2
n
⎞
⎟
⎟
⎟
⎠
2
≤
1+
n
k
=2
x
2
p+1
k
n
+
1
n(n −1)
1+
n
k=2
x
2
p+2
k
−
1+
n
j
=2
x
2
p+1
j
2
n
(2.76)
with equality if and only if x
1
= x
2
=···=x
n−1
.Multiplyingbyy
2
p+1
1
, the inequality in
(2.74) i s obtained with equality if and only if y
1
= y
2
=···=y
n−1
. This completes the
proof.
Corollar y 2.11. Let y
1
≥ y
2
≥ y
3
≥···≥y
n
≥ 0. Then (l
2
p
(y))
∞
p=0
,
l
2
p
(y) =
⎛
⎝
y
2
p
2
p
n
+
1
n(n −1)
y
2
p+1
2
p+1
−
y
2
p+1
2
p
n
⎞
⎠
2
−p
=
m
y
2
p
+
1
√
n −1
s
y
2
p
2
−p
,
(2.77)
is an strictly increasing sequence converging to y
1
except if y
1
= y
2
=···=y
n−1
. In this case,
l
2
p
(y) = y
1
for all p.
Oscar Rojo 15
Proof. We know that (l
2
p
(y))
∞
p=0
is a sequence of lower bounds for y
1
.FromTheorem 2.1,
this sequence converges to y
1
. Applying inequality (2.74), we obtain
⎛
⎜
⎜
⎜
⎝
n
k
=1
y
2
p
k
n
+
1
n(n −1)
n
k=1
y
2
p+1
k
−
n
j
=1
y
2
p
j
2
n
⎞
⎟
⎟
⎟
⎠
2
≤
n
k
=1
y
2
p+1
k
n
+
1
n(n −1)
n
k=1
y
2
p+2
k
−
n
j
=1
y
2
p+1
j
2
n
.
(2.78)
Therefore, l
2
p+1
2
p
(y) ≤l
2
p+1
2
p+1
(y), that is, l
2
p
(y) ≤l
2
p+1
(y). The equality in all the above inequal-
ities takes place if and only if λ
1
=y
2
=···=y
n−1
. In this case, l
2
p
(y)=λ
1
for all p.
Acknowledgment
This work is supported by Fondecyt 1040218, Chile.
References
[1] P.G.Ciarlet,Introduction to Numerical Linear Algebra and Optimisation, Cambridge Texts in
Applied Mathematics, Cambridge University Press, Cambridge, 1991.
[2] O. Rojo and H. Rojo, A decreasing sequence of upper bounds on the largest Laplacian eigenvalue of
a graph, Linear Algebra and Its Applications 381 (2004), 97–116.
[3] H. Wolkowicz and G. P. H. Styan, Bounds for eigenvalues using traces, Linear Algebra and Its
Applications 29 (1980), 471–506.
Oscar Rojo: Departamento de Matem
´
aticas, Universidad Cat
´
olica del Norte, Casilla 1280,
Antofagasta, Chile
E-mail address: