On some elementary inequalities

Titu Andreescu Gabriel Dospinescu

University of Texas at Dallas "Louis le Grand" College
Dallas, TX, USA Paris, France

A remarkable inequality was proposed by G. Zbaganu as the sixth USAMO

2000 problem: for any nonnegative real numbers a1 ; a2 ; :::; an ; b1 ; b2 ; :::; bn , the
following inequality holds:
X X
min(ai aj ; bi bj ) min(ai bj ; aj bi ) (1)
1 i;j n 1 i;j n

Ravi Boppana recently found a beautiful solution to this very di¢ cult question,
based on the following lemma.
Lemma 1. For any positive real numbers a1 ; a2 ; :::; an and any real numbers
x1 ; x2 ; :::; xn , the inequality
X
xi xj min(ai ; aj ) 0 (2)
1 i;j n

holds.
The aim of this paper is to present some applications and re…nements of
(2), which will lead to an even stronger form of (1). We will also discuss some
related problems from various mathematical contests.

Two proofs of the lemma

We present a very simple proof of Lemma 1, as well as a proof involving integrals.
The second proof will allow us to prove some other inequalities involving the
function.
The elementary solution starts with the simple observation that, without
loss of generality, we may assume a1 a2 an . Then the inequality
becomes
Xn nX1 n
X
ai x2i + 2 ai xi xj 0:
i=1 i=1 j=i+1
Pn
Let us denote bi = j=i xj . Then xi = bi bi+1 and the above inequality
reduces to
a1 b21 + (a2 a1 )b22 + ::: + (an an 2
1 )bn 0;
which is true, hence Lemma 1 is proved.
Let us see now an even shorter proof of this result. The crucial observation
is the simple identity
Z 1
min(x; y) = [0;x] (t) [0;y] (t) dt (3)
0

Mathematical reflections 1, (2006) 1

where A is the characteristic function of the set A and x; y are positive real
numbers. Of course, proving (3) is just a formality: it su¢ ces to observe that
[0;x] [0;y] equals 1 on [0; min(x; y)] and 0 elsewhere. Using (3), we infer that

X X Z 1
xi xj min(ai ; aj ) = xi [0;ai ] (t) xj [0;aj ] (t) dt =
1 i;j n 1 i;j n 0
Z 1X n
2
= (xi [0;ai ] (t)) dt 0
0 i=1

and Lemma 1 is proved.

The second proof also shows that (2) can be strengthened to
X Pn
( i=1 ai xi )2
xi xj min(ai ; aj ) : (4)
max1 i n ai
1 i;j n

Indeed, all we need is to observe that

Z n
1X Z max1 i n ai n
X
2 2
(xi [0;ai ] (t)) dt = (xi [0;ai ] (t)) dt:
0 i=1 0 i=1

and to apply the Cauchy-Schwarz inequality

Z max1 i n ai Xn R max n ai
Pn
2 ( 0 1 i
( i=1 xi [0;ai ] (t)dt))2
( xi [0;ai ] (t)) dt :
0 i=1
max1 i n ai

Since Z max1 i n Xn n
X
( xi [0;ai ] (t)) dt = ai xi ;
0 i=1 i=1

the conclusion follows.

But Lemma 1 and the relation (3) have other applications besides the US-
AMO inequality! Before presenting the solution to (1), let us focus on some
applications of (2) and (3).

Three related problems

We promised some nice applications of (2) and (3). We will discuss three prob-
lems, two of which are from mathematical contests. The third one is a di¢ cult
inequality due to Don Zagier. We strongly recommend to the readers to …nd
other solutions to these questions, in order to convince themselves that, despite
the following apparently simple solutions, these problems are really di¢ cult.
We begin with an immediate consequence of (3). The way the problem is
formulated does not even remotely suggest using (3).

Mathematical reflections 1, (2006) 2

 Pn 1. Let a1 ; a2 ; : : : ; an > 0 and let x1 ; x2 ; : : : ; xn be real numbers such
Problem
that i=1 ai xi = 0.
Prove that X
xi xj jai aj j 0:
1 i<j n

Show that the equality holds if and only if there exists a partition
PA1 ; A2 ; : : : ; Ak
of the set f1; 2; :::; ng such that for all i with 1 i k we have j2Ai xj = 0
and aj1 = aj2 if j1 ; j2 2 Ai .
(Gabriel Dospinescu, Mathlinks Contest)
Solution. Let us use the formula
ai + aj jai aj j
min(ai ; aj ) =
2
in (3). Since
n
! n
!
X X X
xi xj (ai + aj ) = 2 xi ai xi = 0;
1 i;j n i=1 i=1

the …rst part of the problem follows from (3).

For the second question, observe that in order to have equality, we must
have !2
Z 1 X n
xi [0;ai ] (t) dt = 0
0 i=1
Pn
and therefore f = i=1 xi [0;ai ] is zero at every point of continuity. Now,
let b1 ; b2 ; :::; bk be the distinct numbers that appear among a1 ; a2 ; :::; an and
let Ai = fj j 1 j n; a = bi g. Then A1 ; A2 ; :::; Ak is a partition of the
Pk Pj
set f1; 2; :::; ng and i=1 ( j2A Pi xj ) [0;bi ] = 0, except for a …nite number of
points. We now conclude that j2Ai xj = 0 for all 1 i k.
We will see in the following problem that Arthur Engel’s principle, according
to which any inequality reduces in the end to x2 0, has merits.
Problem 2. Prove that for any real numbers a1 ; a2 ; :::; an ; b1 ; b2 ; :::; bn the fol-
lowing inequality holds
X X
(jai aj j + jbi bj j) jai bj j:
1 i<j n 1 i;j n

(Poland, 1999)
Solution. Of course, by considering M ai and M bj where M is the
greatest number among all 2n numbers, we may assume that ai ; bj 0.
First of all, we note that the inequality is equivalent to
X X X
jai aj j + jbi bj j 2 jai bj j:
1 i;j n 1 i;j n 1 i;j n

Mathematical reflections 1, (2006) 3

 a +a ja aj j
The same useful formula min(ai ; aj ) = i j 2 i shows that it can be rewrit-
ten as
X X X
min(ai ; aj ) + min(bi ; bj ) 2 min(ai ; bj ): (5)
1 i;j n 1 i;j n 1 i;j n

A simple look at (3) shows now that (5) is a sister of the classical inequality
x2 + y 2 2xy. Indeed, (5) can be rewritten in the form
Z 1 Z 1 Z 1
2 2
f (x) dx + g (x) dx 2 f (x)g(x) dx;
0 0 0
Pn Pn
where f = i=1 [0;ai ] and g R= i=1 [0;bi ] .
The last inequality is easy to
1
prove, since it is equivalent to 0 (f (x) g(x))2 dx 0, which is clearly true.

Observe that the same method that was used to …nd a re…nement of (2)
allows here to prove the stronger inequality
X X Pn Pn 2
( i=1 ai i=1 bi )
jai bj j (jai aj j + jbi bj j) + :
max1 i n max(ai ; bi )
1 i;j n 1 i<j n

Of course, Cauchy-Schwarz cannot be absent from a note on inequalities!

Here is a di¢ cult problem which, using (3), boils down to this famous inequality.
Problem 3. Prove that for all a1 ; a2 ; :::; an ; b1 ; b2 ; :::; bn 0 the following in-
equality holds
0 1 0 1 0 12
X X X
@ min(ai ; aj )A @ min(bi ; bj )A @ min(ai ; bj )A :
1 i;j n 1 i;j n 1 i;j n

(Don Zagier)
Solution. It would be interesting to …nd another proof to this inequality,
but it seems that any attempt to prove it without using (3) fails. Anyway, using
(3) the problem becomes very simple. Indeed, the inequality reduces to
Z 1 Z 1 Z 1 2
2 2
f (x) dx g (x) dx f (x)g(x) dx ;
0 0 0

which is nothing else than, yes, the Cauchy-Schwarz inequality.

The reader has already noticed that there are virtually hundreds of inequal-
ities that can be created using this technique. For example, take an inequality
such as the classical Schur’s inequality

x3 + y 3 + z 3 + 3xyz x2 (y + z) + y 2 (z + x) + z 2 (x + y)

Mathematical reflections 1, (2006) 4

and apply it to x = A [0;a] , y = B [0;b] , z =C [0;c] in order to deduce
(after integration) a stronger form

aA3 + bB 3 + cC 3 + 3ABC min(a; b; c)

AB(A + B) min(a; b) + BC(B + C) min(b; c) + CA(C + A) min(c; a):

There are countless examples and we invite the reader to …nd other “new”
inequalities.

Solution to the USAMO problem

Finally, we go back and discuss a solution to the USAMO 2000 inequality. The
reader will be surprised that only one step is enough to solve this problem, once
(2) is assumed. Yes, but as we will see, this step is terribly well hidden.
As usual, for very di¢ cult inequalities, identities play the most important
role. First of all, let us write the inequality in a simpler form. Observe that
continuity reasons show that it su¢ ces to regard all variables positive. Now,
put xi = abii . The inequality becomes
X
ai aj (min(xi ; xj ) min(1; xi xj )) 0 (6)
1 i;j n

One could try at this moment to prove that the matrix ((min(xi ; xj ) min(1; xi xj ))1 i;j n
has all eigenvalues nonnegative, but this is not easy at all (at least, we could not
…nd a proof of this assertion yet). Ravi Boppana’s observation is the following
strange identity.
Lemma 2. For all positive real numbers x; y the following identity holds

jx 1j jy 1j
min(x; y) min(1; xy) = f (x)f (y) min ;
min(x; 1) min(y; 1)

where f (x) = sgn(x 1) min(x; 1).

Impressive, isn’t it? Proving this identity is (as usual) not di¢ cult; the hard
part is to …nd it. It su¢ ces to assume that x y and to see that in each of the
cases x y 1, x 1 y and 1 x y the identity holds. We will treat only
the …rst case, when the identity reduces to y 1 = sgn(x 1) sgn(y 1) (y 1),
which is clearly true. The other two cases are similar.
Using this lemma, the inequality (6) becomes
X jxi 1j jxj 1j
(ai f (xi )) (aj f (xj )) min ; 0
min(xi ; 1) min(xj ; 1)
1 i;j n

which is nothing else than (2) disguised. The solution of the celebrated USAMO
2000 problem ends here.

Mathematical reflections 1, (2006) 5

 Of course, using (4) and Lemma 2, we can easily establish the following
stronger form
X X Pn
( i=1 (ai bi ))2
min(ai bj ; aj bi ) min(ai aj ; bi bj ) + jai bi j
:
1 i;j n 1 i;j n max1 i n min(ai ;bi )

An interesting consequence
We will end this note with an unexpected problem, which is a direct consequence
of (1), but whose statement does not suggest this fact. Unfortunately, it seems
very di¢ cult to …nd a solution di¤erent from the one that reduces the problem
to (1).
Problem 4. Let x1 ; x2 ; :::; xn be positive real numbers satisfying
X X
j1 xi xj j = jxi xj j:
1 i;j n 1 i;j n

Prove that these numbers add up to n.

(Gabriel Dospinescu)
x+y jx yj
Solution. Let us take bi = 1 in (1). The formula min(x; y) = 2
allows us to rewrite the inequality (1) in the form

n n
!2
X X X X
2
2n xi jxi xj j n + xi j1 xi xj j:
i=1 1 i;j n i=1 1 i;j n

Pn
UsingPthe hypothesis, we now conclude that ( i=1 xi n)2 0, which shows
n
that i=1 xi = n.
That is how a simple identity comes handy in solving a variety of di¢ cult
problems. The readers will certainly discover other problems that can be solved
using these ideas and we strongly encourage them to search for di¤erent solu-
tions.

Mathematical reflections 1, (2006) 6