Professional Documents
Culture Documents
Article
Majorization and Karamata Inequality
MathLinks - www.mathlinks.ro
Pham Kim Hung, Stanford University, US
www.VNMATH.com
2
Note
This is an excerpt from the second volume of Secrets In Inequalities, by Pham
Kim Hung. The author thanks sincerely Darij Grinberg for some of his materials
about Symmetric Majorization Theorem, posted on Mathlinks Forum. Please dont
use this excerpt for any commercial purpose.
The Author always appriciates every Contribution to this content- Majorization and
Karamata Inequality.
Best Regard,
Pham Kim Hung
www.VNMATH.com
Chapter 1
Theory of Majorization
The theory of majorization and convex functions is an important and dicult part
of inequalities, with many nice and powerful applications. will discuss in this article
is Karamata inequality; however, its necessary to review rst some basic properties
of majorization.
Denition 1. Given two sequences (a) = (a
1
, a
2
, ..., a
n
) and (b) = (b
1
, b
2
, ..., b
n
)
(where a
i
, b
i
R i {1, 2, ..., n}). We say that the sequence (a) majorizes the
sequence (b), and write (a) (b), if the following conditions are fullled
a
1
a
2
... a
n
;
b
1
b
2
... b
n
;
a
1
+ a
2
+ ... + a
n
= b
1
+ b
2
+ ... + b
n
;
a
1
+ a
2
+ ... + a
k
b
1
+ b
2
+ ... + b
k
k {1, 2, ...n1} .
Denition 2. For an arbitrary sequence (a) = (a
1
, a
2
, ..., a
n
), we denote (a
),
a permutation of elements of (a) which are arranged in increasing order: (a
) =
(a
i1
, a
i2
, ..., a
in
) with a
i1
a
i2
... a
in
and {i
1
, i
2
, ..., i
n
} = {1, 2, ..., n}.
Here are some basic properties of sequences.
Proposition 1. Let a
1
, a
2
, ..., a
n
be real numbers and a =
1
n
(a
1
+a
2
+... +a
n
), then
(a
1
, a
2
, ..., a
n
)
y
i
y
j
i < j, then
(x
1
, x
2
, ..., x
n
) (y
1
, y
2
, ..., y
n
).
Proof. To prove this assertion, we will use induction. Because
x
i
x
1
y
i
y
1
for all
i {1, 2, ..., n}, we get that
x
1
+ x
2
+ ... + x
n
x
1
y
1
+ y
2
+ ... + y
n
y
1
x
1
y
1
.
Consider two sequences (x
1
+ x
2
, x
3
, ..., x
n
) and (y
1
+ y
2
, y
3
, ..., y
n
). By the inductive
hypothesis, we get
(x
1
+ x
2
, x
3
, ..., x
n
) (y
1
+ y
2
, y
3
, ..., y
n
).
Combining this with the result that x
1
y
1
, we have the conclusion immediately.
) (b
) (b
i=1
|a
i
x| kx +
k
i=1
a
i
,
and similarly,
n
i=k+1
|a
i
x| =
n
i=k+1
|x a
i
| (n k)x
n
i=k+1
a
i
.
Combining the two results and noticing that
k
i=1
a
i
k
i=1
b
i
and
n
i=1
a
i
=
n
i=1
b
i
, we get
n
i=1
|a
i
x| (n 2k)x +
k
i=1
a
i
i=k+1
a
i
= 2
k
i=1
a
i
i=1
a
i
+ (n 2k)x 2
k
i=1
b
i
i=1
b
i
+ (n 2k)x =
n
i=1
|b
i
x|.
This last inequality asserts our desired result.
(ii). Sucient condition. Suppose that the inequality
|a
1
x| +|a
2
x| + ... + |a
n
x| |b
1
x| + |b
2
x| + ... + |b
n
x| ()
has been already true for every real number x. We have to prove that (a
) (b
).
Without loss of generality, we may assume that a
1
a
2
... a
n
and b
1
b
2
... b
n
. Because () is true for all x R, if we choose x max{a
i
, b
i
}
n
i=1
then
n
i=1
|a
i
x| = nx
n
i=1
a
i
;
n
i=1
|b
i
x| = nx
n
i=1
b
i
;
a
1
+ a
2
+ ... + a
n
b
1
+ b
2
+ ... + b
n
.
www.VNMATH.com
6
Similarly, if we choose x min{a
i
, b
i
}
n
i=1
, then
n
i=1
|a
i
x| = nx +
n
i=1
a
i
;
n
i=1
|b
i
x| = nx +
n
i=1
b
i
;
a
1
+ a
2
+ ... + a
n
b
1
+ b
2
+ ... + b
n
.
From these results, we get that a
1
+a
2
+...+a
n
= b
1
+b
2
+...+b
n
. Now suppose that x is
a real number in [a
k
, a
k+1
], then we need to prove that a
1
+a
2
+...+a
k
b
1
+b
2
+...+b
k
.
Indeed, we can eliminate the absolute value signs on the left-hand expression of ()
as follows
|a
1
x| +|a
2
x| + ... +|a
k
x| = a
1
+ a
2
+ ... + a
k
kx ;
|a
k+1
x| +|a
k+2
x| + ... +|a
n
x| = (n k)x a
k+1
a
k+2
... a
n
;
i=1
|a
i
x| = (n 2k)x + 2
k
i=1
a
i
i=1
a
i
.
Considering the right-hand side expression of (), we have
n
i=1
|b
i
x| =
k
i=1
|b
i
x| +
n
i=k+1
|x b
i
|
kx +
k
i=1
b
i
+ (n k)x
n
i=k+1
|b
i
| = (n 2k)x + 2
k
i=1
|b
i
|
n
i=1
|b
i
|.
From these relations and (), we conclude that
(n 2k)x + 2
k
i=1
a
i
i=1
a
i
(n 2k)x + 2
k
i=1
|b
i
|
n
i=1
|b
i
|
a
1
+ a
2
+ ... + a
k
b
1
+ b
2
+ ... + b
k
,
which is exactly the desired result. The proof is completed.
The Symmetric Majorization Criterion asserts that when we examine the ma-
jorization of two sequences, its enough to examine only one conditional inequality
which includes a real variable x. This is important because if we use the normal
method, there may too many cases to check.
The essential importance of majorization lies in the Karamata inequality, which
will be discussed right now.
www.VNMATH.com
Chapter 2
Karamata Inequality
Karamata inequality is a strong application of convex functions to inequalities. As we
have already known, the function f is called convex on I if and only if af(x)+bf(y)
f(ax + by) for all x, y I and for all a, b [0, 1]. Moreover, we also have that f is
convex if f
) (b
) (b
(y) x, y I.
From this result, we also have f(a
i
)f(b
i
) (a
i
b
i
)f
(b
i
) i {1, 2, ..., n}. Therefore
n
i=1
f(a
i
)
n
i=1
f(b
i
) =
n
i=1
(f (a
i
) f (b
i
))
n
i=1
(a
i
b
i
)f
(b
i
)
= (a
1
b
1
)(f
(b
1
) f
(b
2
)) + (a
1
+ a
2
b
1
b
2
)(f
(b
2
) f
(b
3
)) + ...+
+
_
n1
i=1
a
i
n1
i=1
b
i
_
(f
(b
n1
) f
(b
n
)) +
_
n
i=1
a
i
i=1
b
i
_
f
(b
n
) 0
because for all k {1, 2, ..., n} we have f
(b
k
) f
(b
k+1
) and
k
i=1
a
i
k
i=1
b
i
.
7
www.VNMATH.com
8
Comment. 1. If f is a non-decreasing function, it is certain that the last condition
n
i=1
a
i
=
n
i=1
b
i
can be replaced by the stronger one
n
i=1
a
i
n
i=1
b
i
.
2. A similar result for concave functions is that
If (a) (b) are number arrays and f is a concave function twice dierentiable
then
f(a
1
) + f(a
2
) + ... + f(a
n
) f(b
1
) + f(b
2
) + ... + f(b
n
).
3. If f is convex (that means f(a) + f(b) f(a + b) , 0, + = 1)
but not twice dierentiable (f
The following examples should give you a sense of how this inequality can be used.
Example 2.1. If f is a convex function then
f(a) + f(b) + f(c) + f
_
a + b + c
3
_
4
3
_
f
_
a + b
2
_
+ f
_
b + c
2
_
+ f
_
c + a
2
__
.
(Popoviciu-Titu Andreescu inequality)
Solution. WLOG, suppose that a b c. Consider the following number sequences
(x) = (a, a, a, b, t, t, t, b, b, c, c, c) ; (y) = (, , , , , , , , , , , ) ;
where
t =
a + b + c
3
, =
a + b
2
, =
a + c
2
, =
b + c
2
.
Clearly, we have that (y) is a monotonic sequence. Moreover
a , 3a + b 4, 3a + b + t 4+ , 3a + b + 3t 4 + 3,
3a + 2b + 3t 4 + 4, 3a + 3b + 3t 4 + 4 + ,
3a + 3b + 3t + c 4+ 4 + 2, 3a + 3b + 3t + 3c 4 + 4 + 4.
Thus (x
) (y
(x, y, z) and
the conclusion follows from Karamata inequality.
) = (2 lna
k1
lna
k1+1
, 2 lna
k2
ln a
k2+1
, ..., 2 lna
kn
ln a
kn+1
).
Because the number sequence (b) = (ln a
k1
, lna
k2
, ..., lna
kn
) is decreasing, we must
have (c
1 + e
x
, we get
1 + a
1
+
1 + a
2
+ ... +
1 + a
n
1 +
a
2
1
a
2
+
1 +
a
2
2
a
3
+ ... +
1 +
a
2
n
a
1
.
www.VNMATH.com
10
2. By Cauchy-Schwarz inequality, we can solve this problem according to the fol-
lowing estimation
_
1 +
a
2
1
a
2
_
(1 + a
2
) (1 + a
1
)
2
.
a
1
a
2
+ ... + a
n
+ ... +
a
n
a
1
+ ... + a
n1
.
Solution. For each i {1, 2, ..., n}, we denote
y
i
=
a
i
a
1
+ a
2
+ ... + a
n
, x
i
=
a
2
i
a
2
1
+ a
2
2
+ ... + a
2
n
then x
1
+ x
2
+ ... + x
n
= y
1
+ y
2
+ ... + y
n
= 1. We need to prove that
n
i=1
x
i
1 x
i
i=1
y
i
1 y
i
.
WLOG, assume that a
1
a
2
... a
n
, then certainly x
1
x
2
... x
n
and
y
1
y
2
... y
n
. Moreover, for all i j, we also have
x
i
x
j
=
a
2
i
a
2
j
a
i
a
j
=
y
i
y
j
.
By property 4, we deduce that (x
1
, x
2
, ..., x
n
) (y
1
, y
2
, ..., y
n
). Furthermore,
f(x) =
x
1 x
is a convex function, so by Karamata inequality, the nal result follows immediately.
) = (x
1
, x
2
, ..., x
n
)
is a permutation of elements of (x) which are rearranged in the decreasing order, then
y
1
+ y
2
+ ... + y
2k
x
1
+ x
2
+ ... + x
2k
,
and therefore (y) (x
).
cyc
e
4a1
+ 2e
a1+b1+c1+d1
sym
e
2a1+2b1
.
Because f(x) = e
x
is convex, it suces to prove that (a
) majorizes (b
) with
(a) = (4a
1
, 4b
1
, 4c
1
, 4d
1
, a
1
+ b
1
+ c
1
+ d
1
, a
1
+ b
1
+ c
1
+ d
1
) ;
(b) = (2a
1
+ 2b
1
, 2b
1
+ 2c
1
, 2c
1
+ 2d
1
, 2d
1
+ 2a
1
, 2a
1
+ 2c
1
, 2b
1
+ 2d
1
) ;
By the symmetric majorization criterion, we need to prove that for all x
1
R then
2|a
1
+ b
1
+ c
1
+ d
1
4x
1
| +
cyc
|4a
1
4x
1
|
sym
|2a
1
+ 2b
1
4x
1
|.
Letting now x = a
1
x
1
, y = b
1
x
1
, z = c
1
x
1
, t = d
1
x
1
, we obtain an equivalent
form as
2
cyc
|x| + |
cyc
x|
sym
|x + y|,
www.VNMATH.com
12
which is exactly the lemma shown above. We are done.
) (b
) with
(a) = (2x
1
, 2x
1
, ..., 2x
1
. .
n1
, 2x
2
, 2x
2
, ..., 2x
2
. .
n1
, ..., 2x
n
, 2x
n
, ..., 2x
n
. .
n1
, 2x, 2x, ..., 2x
. .
n
) ;
(b) = (x
1
+ x
1
, x
1
+ x
2
, x
1
+x
3
, ..., x
1
+x
n
, x
2
+x
1
, x
2
+x
2
, ..., x
2
+x
n
, ..., x
n
+x
n
) ;
and x =
1
n
(x
1
+ x
2
+ ... + x
n
). By the Symmetric Majorization Criterion, it suces
to prove that
(n 2)
n
i=1
|x
i
| +|
n
i=1
x
i
|
n
i=j
|x
i
+ x
j
|.
Denote A = {i
x
i
0}, B = {i
x
i
< 0} and suppose that |A| = m, |B| = k = nm.
We will prove an equivalent form as follows: if x
i
0 i {1, 2, ..., n} then
(n 2)
iA,B
x
i
+ |
iA
x
i
jB
x
j
|
(i,j)A,B
(x
i
+ x
j
) +
iA,jB
|x
i
x
j
|.
Because k + m = n, we can rewrite the inequality above into
(k 1)
iA
x
i
+ (m1)
jB
x
j
+|
iA
x
i
jB
x
j
|
iA,jB
|x
i
x
j
| ()
Without loss of generality, we may assume that
iA
x
i
jB
x
j
. For each i A, let
|A
i
| = {j B|x
i
x
j
} and r
i
= |A
i
|. For each j B, let |B
j
| = {i A|x
j
x
i
} and
s
j
= |B
j
|. Thus the left-hand side expression in () can be rewritten as
iA
(k 2r
i
)x
i
+
jB
(m2s
j
)x
j
.
Therefore () becomes
iA
(2r
i
1)x
i
+
jB
(2s
j
1)x
j
+ |
iA
x
i
jB
x
j
| 0
iA
r
i
x
i
+
jB
(s
j
1)x
j
0.
www.VNMATH.com
13
Notice that if s
j
1 for all j {1, 2, ..., n} then we have the desired result immedi-
ately. Otherwise, assume that there exists a number s
l
= 0, then
max
iAB
x
i
B r
i
1 i {1, 2, ..., m}.
Thus
iA
r
i
x
i
+
jB
(s
j
1)x
j
iA
x
i
jB
x
j
0.
This problem is completely solved. The equality holds for a
1
= a
2
= ... = a
n
and
a
1
= a
2
= ... = a
n1
, a
n
= 0 up to permutation.
a
1
+
1
n1
a
2
+ ... +
1
n1
a
n
_
.
Solution. The inequality can be rewritten in the form
n
i=1
a
i
+ n(n 2)
n
_
n
i=1
a
i
(n 1)
n
i=1
n1
j=i
a
j
.
First we will prove the following result (that helps us prove the previous inequality
immediately): if x
1
, x
2
, ..., x
n
are real numbers then (
) (
) with
() = (x
1
, x
2
, ..., x
n
, x, x, ..., x) ;
() = (y
1
, y
1
, ..., y
1
, y
2
, y
2
, ..., y
2
, ..., y
n
, y
n
, ..., y
n
) ;
where x =
1
n
(x
1
+x
2
+... +x
n
), () includes n(n 2) numbers x, () includes n 1
numbers y
k
(k {1, 2, ..., n}), and each number b
k
is determined from b
k
=
nx x
i
n 1
.
Indeed, by the symmetric majorization criterion, we only need to prove that
|x
1
| + |x
2
| + ... +|x
n
| + (n 2)|S| |S x
1
| +|S x
2
| + ... +|S x
n
| ()
where S = x
1
+ x
2
+ ... + x
n
= nx. In case n = 3, this becomes a well-known result
|x| +|y| + |z| + |x + y + z| |x + y| +|y + z| + |z + x|.
In the general case, assume that x
1
x
2
... x
n
. If x
i
S i {1, 2, ..., n} then
RHS =
n
i=1
(x
i
S) = (n 1)S (n 1)|S|
n
i=1
|x
i
| + (n 2)|S| = LHS.
www.VNMATH.com
14
and the conclusion follows. Case x
i
S i {1, 2, ..., n} is proved similarly. We
consider the nal case. There exists an integer k (1 k n 1) such that x
k
S
x
k+1
. In this case, we can prove () simply as follows
RHS =
k
i=1
(x
i
S) +
n
i=k+1
(S x
i
) =
k
i=1
x
i
i=k+1
x
k+1
+ (n 2k)S,
i=1
|x
i
| + (n 2k)|S|
n
i=1
|x
i
| + (n 2)|S| = LHS,
which is also the desired result. The problem is completely solved.
i=1
|x
i
| + n|S|
n
i,j=1
|x
i
+ (n 1)x
j
| (1)
in which S = x
1
+x
2
+... +x
n
. Indeed, let z
i
= |x
i
| i {1, 2, ..., n} and A = {i
1
i n, i N, x
i
0}, B = {i
1 i n, i N, x
i
< 0}. WLOG, we may assume that
A = {1, 2, ..., k} and B = {k + 1, k + 2, ..., n}, then |A| = k, |B| = n k = m and
z
i
0 for all i A B. The inequality above becomes
n(n 1)
_
_
iA
z
i
+
jB
z
j
_
_
+ n
iA
z
i
jB
z
j
i,i
A
|z
i
+(n1)z
i
|+
j,j
B
|(n1)z
j
+z
j
|+
iA,jB
_
|z
i
(n1)z
j
|+|(n1)z
i
z
j
|
_
Because n = k + m, the previous inequality is equivalent to
n(m1)
iA
z
i
+ n(k 1)
jB
z
j
+ n
iA
z
i
jB
z
j
iA,jB
|z
i
(n 1)z
j
| +
iA,jB
|(n 1)z
i
z
j
| ()
For each i A we denote
B
i
= { j B
(n 1)z
i
z
j
} ; B
i
= {j B
z
i
(n 1)z
j
} ;
www.VNMATH.com
15
For each j B we denote
A
j
= { i A
(n 1)z
j
z
i
} ; A
j
= {i A
z
j
(n 1)z
i
} ;
We have of course B
i
B
i
B and A
i
A
i
A. After giving up the absolute value
signs, the right-hand side expression of () is indeed equal to
iA
(mn 2|B
i
| 2(n 1)|B
i
|) z
i
+
jB
_
kn 2|A
j
| 2(n 1)|A
j
|
_
z
j
.
WLOG, we may assume that
iA
z
i
jB
z
j
. The inequality above becomes
iA
(|B
i
| + (n 1)|B
i
|) z
i
+
jB
_
|A
j
| + (n 1)|A
j
| n
_
z
j
0.
Notice that if for all j B, we have |A
j
| 1, then the conclusion follows immediately
(because A
j
A
j
, then |A
j
| 1 and |A
j
| + (n 1)|A
j
| n 0 j B). If not,
we may assume that there exists a certain number r B for which |A
r
| = 0, and
therefore |A
r
| = 0. Because |A
r
| = 0, it follows that (n 1)z
r
z
i
for all i A. This
implies that |B
i
| |B
i
| 1 for all i A, therefore |B
i
| + (n 1)|B
i
| n and we
conclude that
iA
(|B
i
| + (n 1)|B
i
|) z
i
+
jB
_
|A
j
| + (n 1)|A
j
| n
_
z
j
n
iA
z
i
n
jB
z
j
0.
Therefore (1) has been successfully proved and therefore Suranjis inequality follows
immediately from Karamata inequality and the Symmetric Majorization Criterion.
) in which (x) = (x
1
, x
2
, ..., x
n
), (y) = (y
1
, y
2
, ..., y
n
) and for each i {1, 2, ..., n}
x
i
=
a
i
+ a
i+1
2
, y
i
=
a
i
+ a
i+1
+ a
i+2
3
(with the common notation a
n+1
= a
1
and a
n+2
= a
2
). According to the Symmetric
Majorization Criterion, it suces to prove the following inequality
3
_
n
i=1
|z
i
+ z
i+1
|
_
2
_
n
i=1
|z
i
+ z
i+1
+ z
i+2
|
_
()
www.VNMATH.com
16
for all real numbers z
1
z
2
... z
n
and z
n+1
, z
n+2
stand for z
1
, z
2
respectively.
Notice that () is obviously true if z
i
0 for all i = 1, 2, ..., n. Otherwise, assume
that z
1
z
2
... z
k
0 > z
k+1
... z
n
. We realize rst that its enough to
consider () for 8 numbers (instead of n numbers). Now consider it for 8 numbers
z
1
, z
2
, ..., z
8
. For each number i {1, 2, ..., 8}, we denote c
i
= |z
i
|, then c
i
0. To
prove this problem, we will prove rst the most dicult case z
1
z
2
z
3
z
4
0
z
5
z
6
z
7
z
8
. Giving up the absolute value signs, the problem becomes
3(c
1
+ 2c
2
+ 2c
3
+ c
4
+ c
5
+ 2c
6
+ 2c
7
+ c
8
+ |c
4
c
5
| +|c
8
c
1
|)
2(c
1
+2c
2
+2c
3
+c
4
+|c
3
+c
4
c
5
|+|c
4
c
5
c
6
|+c
5
+2c
6
+2c
7
+c
8
+|c
7
+c
8
c
1
|+|c
8
c
1
c
2
|)
c
1
+ 2c
2
+ 2c
3
+ c
4
+ c
5
+ 2c
6
+ 2c
7
+ c
8
+ 3|c
4
c
5
| + 3|c
8
c
1
|
2|c
3
+ c
4
c
5
| + 2|c
4
c
5
c
6
| + 2|c
7
+ c
8
c
1
| + 2|c
8
c
1
c
2
|
Clearly, this inequality is obtained by adding the following results
2|c
4
c
5
| + 2c
3
2|c
3
+ c
4
+ c
5
|
2|c
8
c
1
| + 2c
7
2|c
7
+ c
8
c
1
|
|c
4
c
5
| + c
4
+ c
5
+ 2c
6
2|c
4
c
5
c
6
|
|c
8
c
1
| + c
8
+ c
1
+ 2c
2
2|c
8
c
1
c
2
|
For other cases when there exist exactly three (or ve); two (or six); only one (or seven)
non-negative numbers in {z
1
, z
2
, ..., z
8
}, the problem is proved completely similarly
(indeed, notice that, for example, if z
1
z
2
z
3
0 z
4
z
5
z
6
z
7
z
8
then
we only need to consider the similar but simpler inequality of seven numbers after
eliminating z
6
). Therefore () is proved and the conclusion follows immediately.
Using Karamata inequality together with the theory of majorization like we have
just done it is an original method for algebraic inequalities. By this method, a purely
algebraic problem can be transformed to a linear inequality with absolute signs, which
is essentially an arithmetic problem, and which can have many original solutions.
www.VNMATH.com