THE TEACHING OF MATHEMATICS

2011, Vol. XIV, 2, pp. 97–106

SOME CLASSICAL INEQUALITIES

AND THEIR APPLICATION TO OLYMPIAD PROBLEMS

Zoran Kadelburg

Dedicated to Professor Milosav Marjanović on the occasion of his 80th birthday

Abstract. The technique introduced by M. Marjanović in [10] is used to prove

several classical inequalities. Examples of application are given which can be used for
preparing students for mathematical competitions.
ZDM Subject Classification: H34; AMS Subject Classification: 97H30.
Key words and phrases: Chebyshev’s inequality; Karamata’s inequality; Stef-
fensen’s inequality; Jensen’s inequality.

1. Introduction

M. Marjanović deduced in [10] some refinements of classical inequalities of

Karamata [7] and Steffensen [16], making also shorter proofs (see also [8]). Similar
technique was used in [12] to make a short proof of Chebyshev’s inequality [4].
These proofs are reproduced in Section 2 of this paper.
It is well known that these classical results can be used in solving some of
the most difficult problems in mathematical olympiads (in fact, the authors of
these problems probably used the mentioned results, and made elementary solutions
afterwards). We shall present in Section 3 some examples of this kind, giving
solutions based on classical inequalities.

2. Majorization of sequences and functions and its application

Recall that, for decreasing finite sequences a = (ai )ni=1 and b = (bi )ni=1 of real
numbers, a is said to majorize b, what is denoted by a º b, or b ¹ a, if the terms
of these sequences satisfy the following two conditions:
k
X k
X n
X n
X
1◦ ai > bi , for each k ∈ {1, 2, . . . , n − 1}; 2◦ ai = bi .
i=1 i=1 i=1 i=1

Analogously, for two integrable functions ψ1 , ψ2 : [α, β] → R, ψ1 is said to majorize

ψ2 , what is denoted by ψ1 º ψ2 , or ψ2 ¹ ψ1 , if the following two conditions are
satisfied:
Z x Z x Z β Z β
1◦ ψ1 dt > ψ2 dt for x ∈ [α, β); 2◦ ψ1 dt = ψ2 dt.
α α α α
98 Z. Kadelburg

Example 1. (a) If a = (ai )ni=1 is an arbitrary decreasing sequence of nonneg-

ative numbers, having the sum equal to 1, then
³1 1 1´
(1, 0, . . . , 0) º (a1 , a2 , . . . , an ) º , ,...,
n n n
(for details see Lemma 2 in [12]).
(b) If ψ1 : [α, β] → R is a decreasing integrable function, and
Z x
1
ψ2 (x) = ψ1 dt for x ∈ [α, β],
β−α α
then ψ1 º ψ2 .
(c) The sequences (4, 4, 1) and (5, 2, 2) are incomparable in the sense of the
relation º, i.e., none of the two majorizes the other one. 4
Lemma 1. [15] Let ψ1 , ψ2 : [α, β] → R be two integrable functions, such that
ψ1 º ψ2 , and let ϕ : [α, β] → R be an increasing (integrable) function. Then
Z β Z β
ϕψ1 dx 6 ϕψ2 dx.
α α

Rx
Proof. [8] Put ψ(x) = ψ1 (x) − ψ2 (x) and g(x) = α ψ(t) dt. Then, by the
hypothesis, g(x) > 0 for x ∈ [α, β] and g(α) = g(β) = 0. Using integration by parts
in the Stieltjes integral, we get
Z β Z β ¯β Z β
¯
ϕ(t)ψ(t) dt = ϕ(t) dg(t) = ϕ(t)g(t)¯ − g(t) dϕ(t)
α α α α
Z β
=− g(t) dϕ(t) 6 0.
α

Following [10], one can use this lemma to deduce the following classical in-
equality, which is connected with various names—I. Schur [14], G. H. Hardy, J. I.
Littlewood, G. Polya [3], H. Weyl [18], and J. Karamata [9]. Following articles [5],
[10] and [13], we shall call it Karamata’s inequality.
Theorem 1. Let a = (ai )ni=1 and b = (bi )ni=1 be two (finite) decreasing se-
quences of real numbers from an interval (α, β). If a º b, and if f : (α, β) → R is
a convex function, then the following inequality holds
n
X n
X
(1) f (ai ) > f (bi ).
i=1 i=1

Proof. [10] The given function

R x f , being convex, is continuous and it can be
represented in the form f (x) = α ϕ dt for an increasing function ϕ. Introduce
functions A, B : [α, β] → R by
n
X n
X
A(x) = (min{x, ai } − α), B(x) = (min{x, bi } − α).
i=1 i=1
 Some classical inequalities 99

It is easy to see that A(x) 6 B(x), for x ∈ [α, β] and A(a1 ) = B(a1 ). More-
over, A0 (x) and B 0 (x) exist everywhere except in a finite set of points. Applying
Lemma 1, we conclude that
Z a1 Z a1
(2) ϕ dA(x) > ϕ dB(x).
α α

But,
Z a1 Z an Z an−1 Z a1
ϕ dA(x) = n ϕ dx + (n − 1) ϕ dx + · · · + ϕ dx
α α an a2
= f (a1 ) + f (a2 ) + · · · + f (an ),

and the similar relation holds for the integral on the right-hand side of (2). This
proves Karamata’s inequality.
If the function f is strictly convex, it can be easily checked that the equality
in (1) is obtained if and only if the sequences (ai ) and (bi ) coincide.
By the standard technique (passing from natural to rational and then to real
weights) one can deduce the weighted form of Karamata’s inequality (sometimes
called Fuchs’ inequality, see [2]):
n
X n
X
λi f (ai ) > λi f (bi )
i=1 i=1

Pk Pk
if λi ∈ R+ and (ai ) and (bi ) are decreasing sequences, λi ai > λi bi for k ∈
P
n P
n i=1 i=1
{1, 2, . . . , n − 1} and λi a i = λi bi . An immediate consequence is the classical
i=1 i=1
Jensen’s inequality:

Theorem 2. [6] Let f : [α, β] → R be a convex function, let xi ∈ [α, β],

P
n
i ∈ {1, 2, . . . , n} and let λi ∈ [0, 1] be such that λi = 1. Then
i=1
µX
n ¶ n
X
(3) f λi xi 6 λi f (xi ).
i=1 i=1

If f is strictly convex, equality in (3) holds if and only if x1 = x2 = · · · = xn or all

but one λi ’s are equal to 0.

Lemma 1 can also be used in proving Steffensen’s inequality [16]:

Theorem
R x 3. Let f, g : [0, a] → R, 0 6 g(x) 6 1, f be decreasing on [0, a], and
let F (x) = 0 f dt. Then
Z a µZ a ¶
f g dx 6 F g dx .
0 0
100 Z. Kadelburg

Ra
Proof. [10] If we denote c = g dx, then 0 < c 6 a. Let
0
½
1, x ∈ [0, c],
h(x) =
0, x ∈ (c, a].
Then, it is easy to check that h º g, and so, applying Lemma 1, we obtain Stef-
fensen’s inequality in the form
Z c Z a Z a
f dx = f h dx > f g dx.
0 0 0

In order to deduce Chebyshev’s inequality, we first prove the following lem-

ma [12].

Lemma 2. Let a = (ai )ni=1 , b = (bi )ni=1 and c = (ci )ni=1 be three decreasing
sequences of real numbers, such that a º b. Then the following inequality holds:
n
X n
X
ai ci > bi ci .
i=1 i=1

P
i Pi
Proof. Denote Ai = aj , Bi = bj , for i ∈ {1, 2, . . . , n}, and put A0 =
B0 = 0. Then we have j=1 j=1

n
X n
X n
X n
X
ai ci − bi ci = (ai − bi )ci = (Ai − Ai−1 − Bi + Bi−1 )ci
i=1 i=1 i=1 i=1
n
X Xn
= (Ai − Bi )ci − (Ai−1 − Bi−1 )ci
i=1 i=1
n−1
X n−1
X
= (Ai − Bi )ci − (Ai − Bi )ci+1
i=1 i=0
n−1
X
= (Ai − Bi )(ci − ci+1 ) > 0,
i=1

being Ai − Bi > 0 and ci − ci+1 > 0 for each i ∈ {1, 2, . . . , n − 1}.

Pn P
n
In particular, when a and b are decreasing, a º b and ai = bi = 1, the
i=1 i=1
above inequality holds for convex combinations of points c1 , c2 , . . . , cn .

Theorem 4. If (xi )ni=1 and (yi )ni=1 are decreasing sequences of real numbers,
then the following inequality holds:
µX n ¶µXn ¶ X n
(4) xi yi 6 n xi yi .
i=1 i=1 i=1

Equality in (4) holds if and only if x1 = x2 = · · · = xn or y1 = y2 = · · · = yn .

 Some classical inequalities 101

Proof. [12] Without loss of generality, we can assume that the terms xi and yi
of the given sequences are nonnegative (if, for example, some of xi ’s or yi ’s were
negative, we would apply the procedure that follows to the terms x0i = xi − xn > 0
and yi0 = yi − yn > 0).
P n xi 1
Denote X = xi , ai = and bi = . Then the sequences (ai ) i (bi ) are
i=1 X n
decreasing and, by Example 1, (ai ) º (bi ) holds true. Applying Lemma 2 to the
sequences (ai ), (bi ) and taking ci = yi , we obtain
n
X Xn
xi 1
· yi > · yi ,
X i=1 i=1
n
µ n ¶µ n ¶
P
n P P
i.e., n xi yi > xi yi .
i=1 i=1 i=1

We state also the following version of Chebyshev’s inequality:

Theorem 4’. Let (xi )ni=1 and (yi )ni=1 be decreasing sequences of real numbers,
and let π be an arbitrary permutation of the set {1, 2, . . . , n}. Then the inequality
n
X n
X
(5) xi yπ(i) 6 xi yi
i=1 i=1

holds. If the sequence (xi )ni=1 is strictly decreasing, then equality in (5) holds if and
only if yπ(i) = yi for i ∈ {1, 2, . . . , n}.

3. Examples of olympiad problems

Problem 1. [Asian-Pacific Olympiad, 1996] Let a, b, c be the length of sides

of a triangle. Prove that the inequality
√ √ √ √ √ √
a+b−c+ b+c−a+ c+a−b6 a+ b+ c
holds.
Solution. Suppose, without loss of generality, that
√ a > b > c and apply
Karamata’s inequality to the concave function f (x) = x and sequences
(a + b − c, c + a − b, b + c − a) º (a, b, c). 4

Problem 2. [13] Prove that the inequality

a31 a3 a3
+ 2 + · · · + n > a21 + a22 + · · · + a2n
a2 a3 a1
holds for arbitrary positive numbers a1 , a2 , . . . , an .
Solution. Making the substitution xi = log ai , i ∈ {1, 2, . . . , n}, we obtain an
equivalent inequality
e3x1 −x2 + e3x2 −x3 + · · · + e3xn −x1 > e2x1 + e2x2 + · · · + e2xn .
102 Z. Kadelburg

This is obtained by applying Karamata’s inequality to the (convex) function

f (x) = ex , and the sequences a = (3x1 − x2 , 3x2 − x3 , . . . , 3xn − x1 ) and
b = (2x1 , 2x2 , . . . , 2xn ). It is enough to prove that these sequences, when arranged
to be decreasing, satisfy a º b.
Let indices m1 , . . . , mn and k1 , . . . , kn be chosen so that
{m1 , . . . , mn } = {k1 , . . . , kn } = {1, . . . , n},
(6) 3xm1 − xm1 +1 > 3xm2 − xm2 +1 > · · · > 3xmn − xmn +1 ,
(7) 2xk1 > 2xk2 > · · · > 2xkn .
Then
3xm1 − xm1 +1 > 3xk1 − xk1 +1 > 2xk1
(the first inequality holds because 3xm1 − xm1 +1 is, by the choice of numbers mi ,
the greatest of numbers of the form 3xmi − xmi +1 ; the second one follows by the
choice of numbers ki ). By similar reasons,
(3xm1 − xm1 +1 ) + (3xm2 − xm2 +1 ) > (3xk1 − xk1 +1 ) + (3xk2 − xk2 +1 ) > 2xk1 + 2xk2 ,
and, generally, the sum of the first l terms of sequence (6) is not less than the sum
of the first l terms of sequence (7), for l ∈ {1, . . . , n − 1}. For l = n, obviously, the
equality is obtained, and so all the conditions for applying Karamata’s inequality
are fulfilled. 4
Problem 3. [International Mathematical Olympiad 1999] Let n be a fixed
integer, n > 2.
(a) Determine the minimal constant C such that the inequality
X µ X ¶4
2 2
xi xj (xi + xj ) 6 C xi
16i<j6n 16i6n

is valid for all real numbers x1 , x2 , . . . , xn > 0.

(b) For the constant C found in (a) determine when the equality is obtained.
Solution. As far as the given inequality is homogeneous, we can assume that
x1 + x2 + · · · + xn = 1. In this case the inequality can be written as
x31 (1 − x1 ) + x32 (1 − x2 ) + · · · + x3n (1 − xn ) 6 C.
The function f (x) = x3 (1 − x) is increasing and convex on the segment [0, 1/2]. Let
x1 be the greatest of the given numbers. Then the numbers x2 , x3 , . . . , xn are not
greater than 1/2. If x1 ∈ [0, 1/2] as well, then from (x1 , x2 , . . . , xn ) ¹ ( 12 , 21 , 0, . . . )
using Theorem 1, we obtain that
³1´ ³1´ 1
f (x1 ) + f (x2 ) · · · + f (xn ) 6 f +f + (n − 2)f (0) = .
2 2 8
If, to the contrary, x1 > 1/2, then it is 1 − x1 < 1/2 and we have that
(x2 , x3 , . . . , xn ) ¹ (1 − x1 , 0, . . . , 0). Applying Karamata’s inequality once more,
we obtain that
f (x1 ) + f (x2 ) · · · + f (xn ) 6 f (x1 ) + f (1 − x1 ) + (n − 2)f (0) = f (x1 ) + f (1 − x1 ).
 Some classical inequalities 103

It is easy to prove that the function g(x) = f (x) + f (1 − x) has the maximum on
the segment [0, 1] equal to g(1/2) = 1/8. Thus, in this case also, f (x1 ) + f (x2 ) +
· · · + f (xn ) 6 1/8 follows.
Equality holds, e.g., for x1 = x2 = 1/2, which proves that C = 1/8. 4
Problem 4. [G. Szegö, [17]] Let f : [0, a1 ] → R be a convex function and
a1 > a2 > . . . > a2n+1 > 0. Then the inequality
f (a1 − a2 + a3 − · · · + a2n+1 ) 6 f (a1 ) − f (a2 ) + f (a3 ) − · · · + f (a2n+1 )
holds.
Solution. [5] Put a = a1 − a2 + a3 − · · · + a2n+1 . Then the given inequality
can be rewritten as
f (a1 ) + f (a3 ) + · · · + f (a2n+1 ) > f (a) + f (a2 ) + · · · + f (a2n ).
To apply Karamata’s inequality it is enough to check that (a1 , a3 , . . . , a2n+1 ) º
(a2 , . . . , a2n , a). But this follows directly because a2k−1 > a2k for all k.
For another proof of Szegö’s inequality see [11] or [8]. 4
Problem 5. Let a, b and c be the length of sides of a triangle and let s be its
semiperimeter. Prove that for a positive integer n, the inequality
µ ¶n−2
an bn cn 2
+ + > sn−1
b+c c+a a+b 3
holds.
Solution. [7] Without loss of generality, we can suppose that a 6 b 6 c; then
1 1 1
also 6 6 . Chebyshev’s inequality, applied to the sequences
b+c cµ
+a a+b ¶
n n n 1 1 1
(a , b , c ) and , , , implies that
b+c c+a a+b
µ ¶
an bn cn an + bn + cn 1 1 1
+ + > + + .
b+c c+a a+b 3 a+b b+c c+a
By Cauchy-Schwarz inequality, we have
µ ¶
1 1 1
2(a + b + c) + + > 9,
a+b b+c c+a
and by the mean inequality of order n,
µ ¶n
an + bn + cn a+b+c
> .
3 3
Now,
µ ¶n µ ¶
an bn cn a+b+c 1 1 1
+ + > + +
b+c c+a a+b 3 a+b b+c c+a
µ ¶n−1 µ ¶n−2
1 1 2 2
> · · s ·9= sn−1 . 4
3 2 3 3
104 Z. Kadelburg

Problem 6. [University student’s competition, Ostrawa 2002] Let 0 < x1 6

x2 6 · · · 6 xn (n > 2) and
1 1 1
+ + ··· + = 1.
1 + x1 1 + x2 1 + xn
Prove that
µ ¶
√ √ √ 1 1 1
x1 + x2 + · · · + xn > (n − 1) √ + √ + ··· + √ .
x1 x2 xn
Solution. [7] It is enough to prove that
µ ¶ µ ¶ µ ¶
√ 1 √ 1 √ 1
x1 + √ + x2 + √ + ··· + xn + √
x1 x2 xn
µ ¶
1 1 1
> n √ + √ + ··· + √ ,
x1 x2 xn
or, equivalently,
µ ¶µ ¶
1 + x1 1 + xn 1 1 1
(8) √ + ··· + √ + + ··· +
x1 xn 1 + x1 1 + x2 1 + xn
µ ¶
1 1 1
> n √ + √ + ··· + √ .
x1 x2 xn
√ 1 x+1
Take the function f (x) = x+ √ = √ , x ∈ (0, +∞). It is easy to check that
x x µ1¶
f is increasing on (1, +∞) and that f (x) = f for each x > 0.
x
1
By the assumptions, it follows that only x1 could be less than 1 and 6
1 x1 1 1 + x2
1− = . Hence, x2 > . Now it is clear that (both in the case x1 > 1
1 + x1 1 + x1 x1
and in the case x1 < 1)
µ ¶
1
f (x1 ) = f 6 f (x2 ) 6 · · · 6 f (xn ).
x1
µ ¶n
1 + xk
This means that the sequence √ is increasing. Applying Chebyshev’s
xk k=1
inequality, we obtain that inequality (8) holds.
We leave it to the reader to check that for n > 2 equality holds if and only if
x1 = x2 = · · · = xn ¯= n − 1. If n = 2, then equality holds for
n ³1 ´ o
¯
(x1 , x2 ) ∈ ,t ¯ t > 1 . 4
t
Problem 7. [Serbian Mathematical Olympiad 2007] Let k be a positive in-
teger. Prove that for positive real numbers x, y, z, having the sum equal to 1, the
following inequality holds
xk+2 y k+2 z k+2 1
+ k+1 + k+1 > .
xk+1 k
+y +z k y k
+z +x k z + xk + y k 7
 Some classical inequalities 105

Solution. The inequality will be proved by several applications of Chebyshev’s

inequality. Note first, that the expression on left-hand side is symmetric, so we can
suppose that x > y > z. Let us prove now that

xk+1 + y k + z k 6 y k+1 + z k + xk 6 z k+1 + xk + y k .

It is enough to prove the first inequality, which is equivalent to xk+1 + y k 6 y k+1 +

³ y ´k 1−x
xk , i.e., to 6 (numbers x and y are less than 1). Since y 6 x, it is
x 1−y
y 1−x
enough to prove that 6 , which is equivalent to 0 6 x − x2 − y + y 2 =
x 1−y
(x − y)(1 − x − y) = (x − y)z, and this inequality obviously holds.
Applying Chebyshev’s inequality to the triples
µ ¶
1 1 1
(xk+2 , y k+2 , z k+2 ) and , , ,
xk+1 + y k + z k y k+1 + z k + xk z k+1 + xk + y k
we obtain that the left-hand side of the given inequality is not less than
µ ¶
1 k+2 k+2 k+2 1 1 1
A = (x +y +z ) + k+1 + k+1 .
3 xk+1 + y k + z k y + z k + xk z + xk + y k
A new application of Chebyshev’s inequality, this time to the triples (x, y, z) and
(xk+1 , y k+1 , z k+1 ), gives the inequality

xk+2 + y k+2 + z k+2 = xk+1 · x + y k+1 · y + z k+1 · z

1 1
> (xk+1 + y k+1 + z k+1 )(x + y + z) = (xk+1 + y k+1 + z k+1 ).
3 3
Thus,
1 1 k+1
A> · (x + y k+1 + z k+1 ) ×
3 3 µ ¶
1 1 1
× + k+1 + k+1 =: B.
xk+1 + y k + z k y + z k + xk z + xk + y k
But, Cauchy-Schwarz inequality implies that
µ ¶
1 1 1
+ k+1 + k+1 ×
xk+1 + y k + z k y + z k + xk z + xk + y k
× ((xk+1 + y k + z k ) + (y k+1 + z k + xk ) + (z k+1 + xk + y k )) > 9,

xk+1 + y k+1 + z k+1

and hence B > . It remains only to prove
xk+1 + y k+1
+ z k+1 + 2(xk + y k + z k )
that
3(xk+1 + y k+1 + z k+1 ) > xk + y k + z k .
This follows directly by another application of Chebyshev’s inequality.
Equality holds if and only if x = y = z = 1/3. 4
106 Z. Kadelburg

REFERENCES

[1] D. D- ukić, V. Janković, I. Matić, N. Petrović, The IMO Compendium, Springer, 2006.
[2] L. Fuchs, A new proof of an inequality of Hardy-Littlewood-Pólya, Math. Tidsskr. B. (1947),
53–54.
[3] G. H. Hardy, J. E. Littlewood, G. Pólya, Some simple inequalities satisfied by convex function,
Messenger Math. 58 (1928/29), 145–152.
[4] G. H. Hardy, J. E. Littlewood, G. Pólya, Inequalities, Cambridge University Press, 2nd ed.,
Cambridge 1952.
[5] A. I. Hrabrov, Around Mongolian inequality (in Russian), Matematicheskoe prosveshchenie,
3, 7 (2003), 149–162.
[6] J. L. W. V. Jensen, Sur les fonctions convexes et les inégalités entre les valeurs moyennes,
Acta Math. 30 (1906), 175–193.
[7] Z. Kadelburg, D. D - ukić, M. Lukić, I. Matić, Inequalities (in Serbian), Društvo matematičara
Srbije, Beograd, 2003.
[8] Z. Kadelburg, D. D - ukić, M. Lukić, I. Matić: Inequalities of Karamata, Schur and Muirhead,
and some applications, The Teaching of Mathematics VIII, 1 (2005), 31–45.
[9] J. Karamata, Sur une inégalité relative aux fonctions convexes, Publ. Math. Univ. Belgrade
1 (1932), 145–148.
[10] M. Marjanović, Some inequalities with convex functions, Publ. de l’Inst. Math. (Belgrade) 8
(22) (1968), 66–68.
[11] M. Marjanović, Convex and concave functions and corresponding inequalities (in Serbian),
Nastava Matematike XL, 1–2 (1994), pp. 1–11 and 3–4 (1995), pp. 1–14, Društvo matema-
tičara Srbije, Beograd.
[12] M. M. Marjanović, Z. Kadelburg, A proof of Chebyshev’s inequality, The Teaching of Math-
ematics X, 2 (2007), 107–108.
[13] D. Nomirovskiı̆, Karamata’s Inequality (in Russian), Kvant 4 (2003), 43–45.
[14] I. Schur, Über eine Klasse von Mittelbildungen mit Anwendungen die Determinanten, Theorie
Sitzungsler, Berlin, Math. Gesellschaft 22 (1923), 9–20.
[15] J. E. Steffensen, On a generalization of certain inequalities by Tchebychef and Jensen, Skand.
Aktuarietidskr. (1925), 137–147.
[16] J. E. Steffensen, Bounds of certain trigonometric integrals, Tenth Scand. Math. Congress
1946, 181–186, Copenhagen, Gjellerups, 1947.
[17] G. Szegö, Über eine Verallgemeinerung des Dirichletschen Integrals, Math. Zeitschrift 52
(1950), 676–685.
[18] H. Weyl, Inequalities between the two kinds of eigenvalues of a linear transformation, Proc.
Nat. Acad. Sci. U.S.A. 35 (1949), 408–411.

University of Belgrade, Faculty of Mathematics, Studentski trg 16/IV, 11000 Beograd, Serbia
E-mail: kadelbur@matf.bg.ac.rs