INFINITY

The converse also holds as gcd(b, 4ab 1) = 1. Similarly, 4ab 1 divides (a

2
2
b) 4a2 1 if and only if 4ab 1 divides (a b) . So, the original condition is
equivalent to the condition
2
4ab 1 | (a b) .
This condition is symmetric in a and b, so (a, b) is a bad pair if and only if (b, a)
is a bad pair. Thus, we may assume without loss of generality that a > b and that
our bad pair of this type has been chosen with the smallest possible vales of its first
element. Write (a b)2 = m(4ab 1), where m is a positive integer, and treat this
as a quadratic in a:

a2 + (2b 4ma)a + b2 + m = 0.
Since this quadratic has an integer root, its discriminant

2
(2b + 4mb) 4 b2 + m = 4 4mb2 + 4m2 b2 m
must be a perfect square, so 4mb2 + 4m2 b2 m is a perfect square. Let his be the
square of 2mb + t and note that 0 < t < b. Let s = b t. Rearranging again gives:
2

4mb2 + 4m2 b2 m = (2mb + t)

m 4b2 4bt 1 = t2

m 4b2 4b(b s) 1 = (b s)2

m(4bs 1) = (b s)2 .
Therefore, (b, s) is a bad pair with a smaller first element, and we have a contradiction.

Second Solution. (by UNK at IMO 2007) This solution is inspired by the solution of
NZL7 and Atanasovs special prize solution at IMO 1988 in Canberra. We begin
by copying the argument of NZL7. A counter-example (a, b) is called a bad pair.

Consider a bad pair (a, b) so 4ab 1| 4a2 1 . Notice that

b 4a2 1 (4ab

1)a = a b so working modulo 4ab 1 we have b2 4a2 1 (a b)2 . Now, b2 an

2
4ab1 are coprime so 4ab1 divides 4a2 1 if and only if 4ab1 divides (ab)2 .
This condition is symmetric in a and b, so we learn that (a, b) is a bad pair if and
only if (b, a) is a bad pair. Thus, we may assume that a > b and we may as well
choose a to be minimal among all bad pairs where the first component is larger than
the second. Next, we deviate from NZL7s solution. Write (a b)2 = m(4ab 1)
and treat it as a quadratic so a is a root of

x2 + (2b 4mb)x + b2 + m = 0.
The other root must be an integer c since a + c = 2b + 4mb is an integer. Also,
ac = b2 + m > 0 so c is positive. We will show that c < b, and then the pair (b, c)
will violate the minimality of (a, b). It suffices to show that 2b + 4mb < a + b, i.e.,
4mb < a b. Now,
4b(a b)2 = 4mb(4ab 1)
so it suffices to show that 4b(a b) < 4ab 1 or rather 1 < 4b2 which is true.
Delta 4. [IMO 1988/6 FRG] Let a and b be positive integers such that ab + 1 divides
a2 + b2 . Show that

is the square of an integer.

a2 + b 2
ab + 1

INFINITY

Delta 5. (Canada 1998) Let m be a positive integer. Define the sequence {an }n0
by
a0 = 0, a1 = m, an+1 = m2 an an1 .
Prove that an ordered pair (a, b) of non-negative integers, with a b, gives a
solution to the equation
a2 + b2
= m2
ab + 1
if and only if (a, b) is of the form (an , an+1 ) for some n 0.
Delta 6. Let x and y be positive integers such that xy divides x2 + y 2 + 1. Show
that
x2 + y 2 + 1
= 3.
xy
Delta 7. Find all triple (x, y, z) of integers such that
x2 + y 2 + z 2 = 2xyz.
Delta 8. (APMO 1989) Prove that the equation

6 6a2 + 3b2 + c2 = 5n2

has no solutions in integers except a = b = c = n = 0.

INFINITY

1.3. Monotone Multiplicative Functions. In this section, we study when multiplicative functions has the monotonicity.
Example 3. (Canada 1969) Let N = {1, 2, 3, } denote the set of positive integers.
Find all functions f : N N such that for all m, n N: f (2) = 2, f (mn) =
f (m)f (n), f (n + 1) > f (n).
First Solution. We first evaluate f (n) for small n. It follows from f (1 1) =
f (1) f (1) that f (1) = 1. By the multiplicity, we get f (4) = f (2)2 = 4. It follows
from the inequality 2 = f (2) < f (3) < f (4) = 4 that f (3) = 3. Also, we compute
f (6) = f (2)f (3) = 6. Since 4 = f (4) < f (5) < f (6) = 6, we get f (5) = 5. We prove
by induction that f (n) = n for all n N. It holds for n = 1, 2, 3. Now, let n > 2
and suppose that f (k) = k for all k {1, , n}. We show that f (n + 1) = n + 1.
Case 1. n + 1 is composite. One may write n + 1 = ab for some positive integers a and b with 2 a b n. By the inductive hypothesis, we have f (a) = a
and f (b) = b. It follows that f (n + 1) = f (a)f (b) = ab = n + 1.
Case 2. n + 1 is prime. In this case, n + 2 is even. Write n + 2 = 2k for
some positive integer k. Since n 2, we get 2k = n + 2 4 or k 2. Since
k = n+2
n, by the inductive hypothesis, we have f (k) = k. It follows that
2
f (n + 2) = f (2k) = f (2)f (k) = 2k = n + 2. From the inequality
n = f (n) < f (n + 1) < f (n + 2) = n + 2
we conclude that f (n + 1) = n + 1. By induction, f (n) = n holds for all positive
integers n.

Second Solution. As in the previous solution, we get f (1) = 1. We find that

f (2n) = f (2)f (n) = 2f (n)
for all positive integers n. This implies that, for all positive integers k,

f 2k = 2k
Let k N. From the assumption, we obtain the inequality

2k = f 2k < f 2k + 1 < < f 2k+1 1 < f 2k+1 = 2k+1 .

In other words, the increasing sequence of 2k + 1 positive integers

f 2k , f 2k + 1 , , f 2k+1 1 , f 2k+1
lies in the set of 2k + 1 consecutive integers {2k , 2k + 1, , 2k+1 1, 2k+1 }. This
means that f (n) = n for all 2k n 2k+1 . Since this holds for all positive integers
k, we conclude that f (n) = n for all n 2.

The conditions in the problem are too restrictive. Lets throw out the condition
f (2) = 2.
Epsilon 4. Let f : N R+ be a function satisfying the conditions:
(a) f (mn) = f (m)f (n) for all positive integers m and n, and
(b) f (n + 1) f (n) for all positive integers n.
Then, there is a constant R such that f (n) = n for all n N.
We can weaken the assumption that f is completely multiplicative, but we bring
back the condition f (2) = 2.

INFINITY

Epsilon 5. (Putnam 1963/A2) Let f : N N be a strictly increasing function

satisfying that f (2) = 2 and f (mn) = f (m)f (n) for all relatively prime m and n.
Then, f is the identity function on N.
In fact, we can completely drop the constraint f (2) = 2. In 1946, P. Erdos
proved the following result in [PE]:
Theorem 1.4. Let f : N R be a function satisfying the conditions:
(a) f (mn) = f (m) + f (n) for all relatively prime m and n, and
(b) f (n + 1) f (n) for all positive integers n.
Then, there exists a constant R such that f (n) = ln n for all n N.
This implies the following multiplicative result.
Theorem 1.5. Let f : N R+ be a function satisfying the conditions:
(a) f (mn) = f (m)f (n) for all relatively prime m and n, and
(b) f (n + 1) f (n) for all positive integers n.
Then, there is a constant R such that f (n) = n for all n N.
Proof. 2 It is enough to show that the function f is completely multiplicative:
f (mn) = f (m)f (n) for all m and n. We split the proof in three steps.
Step 1. Let a 2 be a positive integer and let a = {x N | gcd(x, a) = 1}.
Then, we find that
f (x + a)
=1
L := inf
xa
f (x)
and

f ak+1 f ak f (a)
for all positive integers k.
Proof of Step 1. Since f is monotone increasing, it is clear that L 1. Now,
we notice that f (k + a) Lf (k) whenever k a . Let m be a positive integer.
We take a sufficiently large integer x0 > ma with gcd (x0 , a) = gcd (x0 , 2) = 1 to
obtain
f (2)f (x0 ) = f (2x0 ) f (x0 + ma) Lf (x0 + (m 1)a) Lm f (x0 )
or
f (2) Lm .
Since m is arbitrary, this and L 1 force to L = 1. Whenever x a , we obtain

f ak+1 f (x)
f ak+1 x
f ak+1 x + ak
=

= f (ax + 1) f ax + a2
k
k
k
f (a )
f (a )
f (a )
or

f ak+1 f (x)
f (a)f (x + a)
f (ak )
or

f ak+1
f (x + a)

.
f (x)
f (a)f (ak )
2We present a slightly modified proof in [EH]. For another short proof, see [MJ].

INFINITY

It follows that

f ak+1
f (x + a)
1 = inf

xa
f (x)
f (a)f (ak )

so that

f ak+1 f ak f (a),

as desired.
Step 2. Similarly, we have
U := sup
xa

and

f (x)
=1
f (x + a)

f ak+1 f ak f (a)

for all positive integers k.

Proof of Step 2. The first result immediately follows from Step 1.
f (x)
1
=
= 1.
sup
xa f (x + a)
inf xa f (x+a)
f (x)

Whenever x a and x > a, we have

f ak+1 f (x)
f ak+1 x
f ak+1 x ak
=

= f (ax 1) f ax a2
f (ak )
f (ak )
f (ak )
or

f ak+1 f (x)
f (a)f (x a).
f (ak )
It therefore follows that

f ak+1
f (x)
f (x a)
1 = sup
= sup

,
f (x)
f (a)f (ak )
xa f (x + a)
xa , x>a
as desired.
Step 3. From the two previous results, whenever a 2, we have

f ak+1 = f ak f (a).
Then, the straightforward induction gives that

k
f ak = f (a)
for all positive integers a and k. Since f is multiplicative, whenever
n = p1 k1 pl kl
gives the standard factorization of n, we obtain

k
k
f (n) = f p1 k1 f pl kl = f (p1 ) 1 f (pl ) l .
We therefore conclude that f is completely multiplicative.

10

INFINITY

1.4. There are Infinitely Many Primes. The purpose of this subsection is to offer
various proofs of Euclids Theorem.
Theorem 1.6. (Euclids Theorem) The number of primes is infinite.
Proof. Assume to the contrary {p1 = 2, p2 = 3, , pn } is the set of all primes.
Consider the positive integer
P = p1 pn + 1.
Since P > 1, P must admit a prime divisor pi for some i {1, , n}. Since both
P and p1 pn are divisible by pi , we find that 1 = P p1 pn is also divisible
by pi , which is a contradiction.

In fact, more is true. We now present four proofs of Eulers Theorem that the
sum of the reciprocals of all prime numbers diverges.
Theorem 1.7. (Eulers Theorem, PEN E24) Let pn denote the nth prime number.
The infinite series

X
1
n=1

pn

diverges.
First Proof. [NZM, pp.21-23] We first prepare a lemma. Let %(n) denote the set of
prime divisors of n. Let Sn (N ) denote the set of positive integers i N satisfying
that %(i) {p1 , , pn }.

Lemma 1.3. We have |Sn (N )| 2n N .

Proof of Lemma. It is because every positive integer i
Sn (N ) has a unique
factorization i = st2 , where s is a divisor of p1 pn and t N . In other words,

i 7 (s, t) is an injective map from Sn (N ) toTn (N ) = { (s, t) | s | p1 pn , t N },

which means that |Sn (N )| |Tn (N )| 2n N .
Now, assume to the contrary that the infinite series p11 + p12 + converges.
Then we can take a sufficiently large positive integer n satisfying that

1 X 1
1
1

=
+
+ .
2 i>n pi
pn+1
pn+2
Take a sufficiently large positive integer N so that N > 4n+1 . By its definition of
Sn (N ), we see that each element i in {1, , N } Sn (N ) is divisible by at least
onej prime
pj for some j > n. Since the number of multiples of pj not exceeding N
k
is

N
pj

, we have
|{1, , N } Sn (N )|

X N
j>n

or
N |Sn (N )|

X N
j>n

or

pj

N
|Sn (N )| .
2

pj

XN
N

p
2
j>n j

INFINITY

11

It follows from this and from the lemma that N2 2n N so that N 4n+1 .
However, it is a contradiction for the choice of N .

Second Proof. We employ an auxiliary inequality without a proof.

Lemma 1.4. The inequality 1 + t et holds for all t R.
2
Let n > 1. Since each positive
integer i n has a unique factorization i = st ,
where s is square free and t n, we obtain

n
X
Y
1
1 X 1

1+
.
2
k p:prime,
p
t
k=1

t n

pn

Together with the estimation

X
X
X
1
1
1
1
1+
=1+

= 2,
t2
t(t 1)
t1
t
t=1
t=2
t=2

we conclude that

n
X
Y
Y
1
1
1
2
1+
2
ep
k
p
p:prime,
p:prime

k=1

or

pn

X 1
ln
p
p:prime

pn

1X1
2
k

!
.

k=1

pn

Since the divergence of the harmonic series 1 +

Comparison Test, the series diverges.

1
2

1
3

+ is well-known, by

Third Proof. [NZM, pp.21-23] We exploit an auxiliary inequality without a proof.

2
1
et+t holds for all t 0, 12 .
Lemma 1.5. The inequality 1t
Let l N. By The Fundamental Theorem of Arithmetic, each positive integer
i pl has a unique factorization i = p1 e1 p1 el for some e1 , , el Z0 . It
follows that

!
pl
l
l
l
1
1
X
X
Y
X 1
Y
Y
1
1
1
pj + p 2
j

=
=

e
i
p1 e1 pl el
pj k
1 p1j
i=1
j=1
j=1
j=1
e1 , ,el Z0

so that

k=0

p
!

l
l
X
X
1
1
1
+ 2 ln
.
pj
pj
i
j=1
i=1

Together with the estimation

l
l
l
X
X
X
X
1
1
1
1
1
1

= lim 1
= 1,
n
p 2
(j + 1)2
(j + 1)j
j
j+1
n+1
j=1 j
j=1
j=1
j=1
we conclude that

p
!
l
l
X
X
1
1
ln
1.
p
i
i=1
j=1 j

12

INFINITY

Since the harmonic series 1 +

get the result.

1
2

1
3

+ diverges, by The Comparison Test, we

Fourth Proof. [DB, p.334] It is a consequence of The Prime Number Theorem.

Let (x) denote the prime counting function. Since The Prime Number Theorem
says that (x) lnxx as x , we can find a constant > 0 satisfying that
(x) > lnxx for all sufficiently large positive real numbers x. This means that

n > lnppnn when n is sufficiently large. Since lnxx > x for all sufficiently large
x > 0, we also have
pn

n>
> pn
ln pn
or
n2 > pn
for all sufficiently large n. We conclude that, when n is sufficiently large,
pn
pn
n>
>
,
ln pn
ln (n2 )
or equivalently,
1

>
.
pn
2n ln n
P
1
Since we have n=2 n ln
n = , The Comparison Test yields the desired result.
We close this subsection with a striking result establish by Viggo Brun.
Theorem 1.8. (Bruns Theorem) The sum of the reciprocals of the twin primes
converges:

X
1
1
1 1
1 1
1
1
B=
+
=
+
+
+
+
+
+ <
p p+2
3 5
5 7
11 13
p, p+2: prime

The constant B = 1.90216 is called Bruns Constant.

INFINITY

13

1.5. Towards $1 Million Prize Inequalities. In this section, we follow [JL]. We

consider two conjectures.
Open Problem 1.1. (J. C. Lagarias) Given a positive integer n, let Hn denote the
n-th harmonic number
n
X
1
1
Hn =
= 1 + +
i
n
i=1
and let (n) denote the sum of positive divisors of n. Prove that that the inequality
(n) Hn + eHn ln Hn
holds for all positive integers n.
Open Problem 1.2. Let denote the prime counting function, that is, (x) counts
the number of primes p with 1 < p x. Let > 0. Prove that that there exists a
positive constant C such that the inequality

Z x
1 +

1
2
(x)
dt

C
x

ln t
2

holds for all real numbers x 2.

These two unseemingly problems are, in fact, equivalent. Furthermore, more
strikingly, they are equivalent to The Riemann Hypothesis from complex analysis.
In 2000, The Clay Mathematics Institute of Cambridge, Massachusetts (CMI) has
named seven prize problems. If you knock them down, you earn at least $1 Million.3
For more info, visit the CMI website at
http://www.claymath.org/millennium

Wir m
ussen wissen. Wir werden wissen.
- D. Hilbert

3However, in that case, be aware that Mafias can knock you down and take money from you :]

14

INFINITY

2. Symmetries
Each problem that I solved became a rule, which served afterwards to solve other problems.
- R. Descartes

2.1. Exploiting Symmetry. We begin with the following example.

Example 4. Let a, b, c be positive real numbers. Prove the inequality
b4 + c4
c 4 + a4
a4 + b 4
+
+
a3 + b3 + c3 .
a+b
b+c
c+a
First Solution. After brute-force computation, i.e, clearing denominators, we reach
a5 b + a5 c + b5 c + b5 a + c5 a + c5 b a3 b2 c + a3 bc2 + b3 c2 a + b3 ca2 + c3 a2 b + c3 ab2 .
Now, we deduce

a5 b + a5 c + b5 c + b5 a + c5 a + c5 b
`

a b5 + c5 + b c5 + a5 + c a5 + b5
`

a b3 c2 + b2 c3 + b c3 a2 + c2 b3 + c c3 a2 + c2 b3

a3 b2 c + a3 bc2 + b3 c2 a + b3 ca2 + c3 a2 b + c3 ab2 .

Here, we used the the auxiliary inequality

x5 + y 5 x3 y 2 + x2 y 3 ,
where x, y 0. Indeed, we obtain the equality

`
`

x5 + y 5 x3 y 2 x2 y 3 = x3 y 3 x2 y 2 .
`
`

It is clear that the final term x3 y 3 x2 y 2 is always non-negative.

Here goes a more economical solution without the brute-force computation.

Second Solution. The trick is to observe that the right hand side admits a nice decomposition:
a3 + b 3
b3 + c3
c 3 + a3
a3 + b3 + c3 =
+
+
.
2
2
2
We then see that the inequality has the symmetric face:
b4 + c4
c 4 + a4
a3 + b 3
b3 + c3
c 3 + a3
a4 + b 4
+
+

+
+
.
a+b
b+c
c+a
2
2
2
Now, the symmetry of this expression gives the right approach. We check that, for x, y > 0,
x4 + y 4
x3 + y 3

.
x+y
2
However, we obtain the identity
`
`

2 x4 + y 4 x3 + y 3 (x + y) = x4 + y 4 x3 y xy 3 = x3 y 3 (x y) .
`

It is clear that the final term x3 y 3 (x y) is always non-negative.

Delta 9. [LL 1967 POL] Prove that, for all a, b, c > 0,
a8 + b8 + c8
1
1
1
+ + .
a3 b3 c3
a
b
c
Delta 10. [LL 1970 AUT] Prove that, for all a, b, c > 0,
bc
ca
ab
a+b+c

+
+
2
b+c
c+a
a+b

INFINITY

15

Delta 11. [SL 1995 UKR] Let n be an integer, n 3. Let a1 , , an be real numbers
such that 2 ai 3 for i = 1, , n. If s = a1 + + an , prove that
a1 2 + a2 2 a3 2
a2 2 + a 3 2 a 4 2
an 2 + a1 2 a2 2
+
+ +
2s 2n.
a1 + a2 + a3
a2 + a3 + a4
an + a1 + a2
Delta 12. [SL 2006 ] Let a1 , , an be positive real numbers. Prove the inequality
X
X
n
ai aj
ai aj
2(a1 + a2 + + an )
ai + aj
1i<jn

1i<jn

Epsilon 6. Let a, b, c be positive real numbers. Prove the inequality

`
`
`

1 + a2 1 + b2 1 + c2 (a + b)(b + c)(c + a).

Show that the equality holds if and only if (a, b, c) = (1, 1, 1).
Epsilon 7. (Poland 2006) Let a, b, c be positive real numbers with ab+bc+ca = abc. Prove
that
b4 + c4
c 4 + a4
a4 + b4
+
+
1.
3
3
3
3
ab(a + b )
bc(b + c )
ca(c3 + a3 )
Epsilon 8. (APMO 1996) Let a, b, c be the lengths of the sides of a triangle. Prove that

a + b c + b + c a + c + a b a + b + c.

16

INFINITY

2.2. Breaking Symmetry. We now learn how to break the symmetry. Lets attack the
following problem.
Example 5. Let a, b, c be non-negative real numbers. Show the inequality
`

4
a4 + b4 + c4 + 3 (abc) 3 2 a2 b2 + b2 c2 + c2 a2 .
There are many ways to prove this inequality. In fact, it can be proved either with
Schurs Inequality or with Popovicius Inequality. Here, we try to give another proof. One
natural starting point is to apply The AM-GM Inequality to obtain the estimations

1
4
4
4
4
4
c4 + 3 (abc) 3 4 c4 (abc) 3 (abc) 3 (abc) 3
= 4abc2
and

a4 + b4 2a2 b2 .
Adding these two inequalities, we obtain
4

a4 + b4 + c4 + 3 (abc) 3 2a2 b2 + 4abc2 .

Hence, it now remains to show that

2a2 b2 + 4abc2 2 a2 b2 + b2 c2 + c2 a2

or equivalently

0 2c2 (a b)2 ,
which is clearly untrue in general. It is reversed! However, we can exploit the above
idea to finsh the proof.
Proof. Using the symmetry of the inequality, we break the symmetry. Since the inequality
is symmetric, we may consider the case a, b c only. Since The AM-GM Inequality implies
4
the inequality c4 + 3 (abc) 3 4abc2 , we obtain the estimation
`

4
a4 + b4 + c4 + 3 (abc) 3 2 a2 b2 + b2 c2 + c2 a2
` 4

a + b4 2a2 b2 + 4abc2 2 b2 c2 + c2 a2
` 2
2
=
a b2 2c2 (a b)2
`

= (a b)2 (a + b)2 2c2 .

Since we have a, b c, the last term is clearly non-negative.

Epsilon 9. Let a, b, c be the lengths of a triangle. Show that

a
b
c
+
+
< 2.
b+c
c+a
a+b
Epsilon 10. (USA 1980) Prove that, for all real numbers a, b, c [0, 1],
a
b
c
+
+
+ (1 a)(1 b)(1 c) 1.
b+c+1
c+a+1
a+b+1
Epsilon 11. [AE, p. 186] Show that, for all a, b, c [0, 1],
a
b
c
+
+
2.
1 + bc
1 + ca
1 + ab
Epsilon 12. [SL 2006 KOR] Let a, b, c be the lengths of the sides of a triangle. Prove the
inequality

b+ca
c+ab
a+bc

3.
b+ c a
c+ a b
a+ b c
`

Epsilon 13. Let f (x, y) = xy x3 + y 3 for x, y 0 with x + y = 2. Prove the inequality

1
1
1
1
= f 1 ,1 +
.
f (x, y) f 1 + , 1
3
3
3
3

INFINITY

17

Epsilon 14. Let a, b 0 with a + b = 1. Prove that

p
p

a2 + b + a + b2 + 1 + ab 3.
Show that the equality holds if and only if (a, b) = (1, 0) or (a, b) = (0, 1).
Epsilon 15. (USA 1981) Let ABC be a triangle. Prove that

3 3
sin 3A + sin 3B + sin 3C
.
2
The above examples say that, in general, symmetric problems does not admit
symmetric solutions. We now introduce an extremely useful inequality when we make
the ordering assmption.
Epsilon 16. (Chebyshevs Inequality) Let x1 , , xn and y1 , yn be two monotone increasing sequences of real numbers:
x1 xn , y1 yn .
Then, we have the estimation
n
X

1
xi yi
n
i=1

n
X

!
xi

i=1

n
X

!
yi

i=1

Corollary 2.1. (The AM-HM Inequality) Let x1 , , xn > 0. Then, we have

x1 + + xn
n
1
n
+
x1
x
1

or

1
1
n2
+

.
x1
xn
x1 + + xn
The equality holds if and only if x1 = = xn .
Proof. Since the inequality is symmetric, we may assume that x1 xn . We have
1
1

.
x1
xn
Chebyshevs Inequality shows that

1
1
1
1
1
x1
+ + x1
(x1 + + xn )
+ +
.
x1
x1
n
x1
x1

Remark 2.1. In Chebyshevs Inequality, we do not require that the variables are positive.
It also implies that if x1 xn and y1 yn , then we have the reverse estimation
! n
!
n
n
X
X
1 X
xi yi
xi
yi .
n i=1
i=1
i=1
Epsilon 17. (United Kingdom 2002) For all a, b, c (0, 1), show that

a
b
c
3 3 abc

+
+

.
1a
1b
1c
1 3 abc
Epsilon 18. [IMO 1995/2 RUS] Let a, b, c be positive numbers such that abc = 1. Prove
that
1
1
1
3
+ 3
+ 3
.
a3 (b + c)
b (c + a)
c (a + b)
2
Epsilon 19. (Iran 1996) Let x, y, z be positive real numbers. Prove that

1
1
1
9
(xy + yz + zx)
+
+
.
(x + y)2
(y + z)2
(z + x)2
4

18

INFINITY

We now present three different proofs of Nesbitts Inequality:

Proposition 2.1. (Nesbitt) For all positive real numbers a, b, c, we have
a
b
c
3
+
+
.
b+c
c+a
a+b
2
Proof 1. We denote L the left hand side. Since the inequality is symmetric in the three
1
1
1
variables, we may assume that a b c. Since b+c
c+a
a+b
, Chebyshevs Inequality
yields that

1
1
1
1
L
(a + b + c)
+
+
3
b+c
c+a
a+b

a+b+c
a+b+c
1 a+b+c
+
+
=
3
b+c
c+a
a+b

a
b
c
= 3 1+
+1+
+1+
b+c
c+a
a+b
1
=
(3 + L),
3
so that L 23 , as desired.
Proof 2. We now break the symmetry by a suitable normalization. Since the inequality is
symmetric in the three variables, we may assume that a b c. After the substitution
x = ac , y = cb , we have x y 1. It becomes
a
c
b
c

+1

a
c

b
c

+1

a
c

1
+

b
c

3
2

or

x
y
3
1
+

.
y+1
x+1
2
x+y
We first apply The AM-GM Inequality to deduce
x+1
y+1
+
2
y+1
x+1
or
y
1
1
x
+
2

.
y+1
x+1
y+1
x+1
It is now enough to show that
1
1
3
1
2

y+1
x+1
2
x+y
or
1
1
1
1

2
y+1
x+1
x+y
or
y1
y1

.
2(1 + y)
(x + 1)(x + y)
However, the last inequality clearly holds for x y 1.
Proof 3. As in the previous proof, we may assume a b 1 = c. We present a proof of
b
1
3
a
+
+
.
b+1
a+1
a+b
2
Let A = a + b and B = ab. What we want to prove is
a2 + b 2 + a + b
1
3
+

(a + 1)(b + 1)
a+b
2
or

A2 2B + A
1
3
+

A+B+1
A
2

INFINITY

or

19

2A3 A2 A + 2 B(7A 2).

Since 7A 2 > 2(a + b 1) > 0 and A2 = (a + b)2 4ab = 4B, its enough to show that
4(2A3 A2 A + 2) A2 (7A 2) A3 2A2 4A + 8 0.
However, its easy to check that A3 2A2 4A + 8 = (A 2)2 (A + 2) 0.

20

INFINITY

2.3. Symmetrizations. We now attack non-symmetrical inequalities by transforming them

into symmetric ones.
Example 6. Let x, y, z be positive real numbers. Show the cyclic inequality
x2
y2
z2
x
y
z
+ 2 + 2 + + .
y2
z
x
y
z
x
First Solution. We break the homogeneity. After the substitution a = xy , b =
becomes
a2 + b2 + c2 a + b + c.
We now obtain
1
1
a2 + b2 + c2 (a + b + c)2 (a + b + c)(abc) 3 = a + b + c.
3

y
,c
z

z
,
x

it

Epsilon 20. (APMO 1991) Let a1 , , an , b1 , , bn be positive real numbers such that
a1 + + an = b1 + + bn . Show that
a1 2
an 2
a1 + + an
.
+ +

a1 + b 1
an + b n
2
Epsilon 21. Let x, y, z be positive real numbers. Show the cyclic inequality
x
y
z
+
+
1.
2x + y
2y + z
2z + x
Epsilon 22. Let x, y, z be positive real numbers with x + y + z = 3. Show the cyclic
inequality
x3
y3
z3
+ 2
+ 2
1.
2
2
2
x + xy + y
y + yz + z
z + zx + x2
Epsilon 23. [SL 1985 CAN] Let x, y, z be positive real numbers. Show the cyclic inequality
x2

x2
y2
z2
+ 2
+ 2
2.
+ yz
y + zx
z + xy

Epsilon 24. [SL 1990 THA] Let a, b, c, d 0 with ab + bc + cd + da = 1. show that

a3
b3
c3
d3
1
+
+
+
.
b+c+d
c+d+a
d+a+b
a+b+c
3
Delta 13. [SL 1998 MNG] Let a1 , , an be positive real numbers such that a1 + +an <
1. Prove that
a1 an (1 a1 an )
1
n+1 .
(a1 + + an ) (1 a1 ) (1 an )
n

Dont just read it; fight it! Ask your own questions, look for your own examples, discover
your own proofs. Is the hypothesis necessary? Is the converse true? What happens in the
classical special case? What about the degenerate cases? Where does the proof use the
hypothesis?
- P. Halmos, I Want to be a Mathematician

INFINITY

21

3. Geometric Inequalities
Geometry is the science of correct reasoning on incorrect figures.
- G. P
olya

3.1. Triangle Inequalities. Many inequalities are simplified by some suitable substitutions.
We begin with a classical inequality in triangle geometry. What is the first4 nontrivial
geometric inequality?
Theorem 3.1. (Chapple 1746, Euler 1765) Let R and r denote the radii of the circumcircle
and incircle of the triangle ABC. Then, we have R 2r and the equality holds if and
only if ABC is equilateral.
Proof. Let BC = a, CA = b, AB = c, s = a+b+c
and S = [ABC].5 We now recall the
2
well-known identities:
abc
S=
, S = rs, S 2 = s(s a)(s b)(s c).
4R
Hence, the inequality R 2r is equivalent to
S
abc
2
4S
s
or
abc 8
or

S2
s

abc 8(s a)(s b)(s c).

We need to prove the following.

Theorem 3.2. (A. Padoa) Let a, b, c be the lengths of a triangle. Then, we have
abc 8(s a)(s b)(s c)
or
abc (b + c a)(c + a b)(a + b c)
Here, the equality holds if and only if a = b = c.
Proof. We exploit The Ravi Substitution. Since a, b, c are the lengths of a triangle, there
are positive reals x, y, z such that a = y + z, b = z + x, c = x + y. (Why?) Then, the
inequality is (y + z)(z + x)(x + y) 8xyz for x, y, z > 0. However, we get
(y + z)(z + x)(x + y) 8xyz = x(y z)2 + y(z x)2 + z(x y)2 0.

Does the above inequality hold for arbitrary positive reals a, b, c? Yes ! Its possible
to prove the inequality without the additional condition that a, b, c are the lengths of a
triangle :
Theorem 3.3. Whenever x, y, z > 0, we have
xyz (y + z x)(z + x y)(x + y z).
Here, the equality holds if and only if x = y = z.
4

The first geometric inequality is the Triangle Inequality: AB + BC AC

In this book, [P ] stands for the area of the polygon P .

22

INFINITY

Proof. Since the inequality is symmetric in the variables, without loss of generality, we
may assume that x y z. Then, we have x + y > z and z + x > y. If y + z > x,
then x, y, z are the lengths of the sides of a triangle. In this case, by the previous
theorem, we get the result. Now, we may assume that y + z x. Then, it is clear that
xyz > 0 (y + z x)(z + x y)(x + y z).

The above inequality holds when some of x, y, z are zeros:

Theorem 3.4. Let x, y, z 0. Then, we have xyz (y + z x)(z + x y)(x + y z).
Proof. Since x, y, z 0, we can find strictly positive sequences {xn }, {yn }, {zn } for which
lim xn = x, lim yn = y, lim zn = z.

The above theorem says that

xn yn zn (yn + zn xn )(zn + xn yn )(xn + yn zn ).
Now, taking the limits to both sides, we get the result.

We now notice that, when x, y, z 0, the equality xyz = (y +z x)(z +xy)(x+y z)

does not guarantee that x = y = z. In fact, for x, y, z 0, the equality xyz = (y + z
x)(z + x y)(x + y z) implies that
x = y = z or x = y, z = 0 or y = z, x = 0 or z = x, y = 0.
(Verify this!) Its straightforward to verify the equality
xyz (y + z x)(z + x y)(x + y z) = x(x y)(x z) + y(y z)(y x) + z(z x)(z y).
Hence, it is a particular case of Schurs Inequality.
Epsilon 25. [IMO 2000/2 USA] Let a, b, c be positive numbers such that abc = 1. Prove
that

1
1
1
a1+
b1+
c1+
1.
b
c
a
Delta 14. Let R and r denote the radii of the circumcircle and incircle of the right triangle
ABC, resepectively. Show that

R (1 + 2)r.
When does the equality hold ?
Delta 15. [LL 1988 ESP] Let ABC be a triangle with inradius r and circumradius R.
Show that
A
B
B
C
C
A
5
r
sin sin + sin sin + sin sin +
.
2
2
2
2
2
2
8
4R
In 1965, W. J. Blundon[WJB] found the best possible inequalities of the form
A(R, r) s2 B(R, r),
where A(x, y) and B(x, y) are real quadratic forms x2 + xy + y 2 .
Delta 16. Let R and r denote the radii of the circumcircle and incircle of the triangle
ABC. Let s be the semiperimeter of ABC. Show that
16Rr 5r2 s2 4R2 + 4Rr + 3r2 .
Delta 17. [WJB2, RS] Let R and r denote the radii of the circumcircle and incircle of the
triangle ABC. Let s be the semiperimeter of ABC. Show that

s 2R + (3 3 4)r.

INFINITY

23

Delta 18. With the usual notation for a triangle, show the inequality6

4R + r 3s.
The Ravi Substitution is useful for inequalities for the lengths a, b, c of a triangle.
After The Ravi Substitution, we can remove the condition that they are the lengths of
the sides of a triangle.
Epsilon 26. [IMO 1983/6 USA] Let a, b, c be the lengths of the sides of a triangle. Prove
that
a2 b(a b) + b2 c(b c) + c2 a(c a) 0.
Delta 19. (Darij Grinberg) Let a, b, c be the lengths of a triangle. Show the inequalities
a3 + b3 + c3 + 3abc 2b2 a 2c2 b 2a2 c 0,
and
3a2 b + 3b2 c + 3c2 a 3abc 2b2 a 2c2 b 2a2 c 0.
Delta 20. [LL 1983 UNK] Show that if the sides a, b, c of a triangle satisfy the equation
`

2 ab2 + bc2 + ca2 = a2 b + b2 c + c2 a + 3abc

then the triangle is equilateral. Show also that the equation can be satisfied by positive real
numbers that are not the sides of a triangle.
Delta 21. [IMO 1991/1 USS] Prove for each triangle ABC the inequality
1
IA IB IC
8
<

,
4
lA lB lC
27
where I is the incenter and lA , lB , lC are the lengths of the angle bisectors of ABC.
We now discuss Weitzenb
ocks Inequality and related theorems.
Epsilon 27. [IMO 1961/2 POL] (Weitzenb
ocks Inequality) Let a, b, c be the lengths of a
triangle with area S. Show that

a2 + b2 + c2 4 3S.
Epsilon 28. (The Hadwiger-Finsler Inequality) For any triangle ABC with sides a, b, c and
area F , the following inequality holds:

a2 + b2 + c2 4 3F + (a b)2 + (b c)2 + (c a)2

or

2ab + 2bc + 2ca (a2 + b2 + c2 ) 4 3F.

Here is a simultaneous generalization of Weitzenb

ocks Inequality and Nesbitts Inequality.
Epsilon 29. (Tsintsifas) Let p, q, r be positive real numbers and let a, b, c denote the sides
of a triangle with area F . Then, we have

p
q 2
r 2
a2 +
b +
c 2 3F.
q+r
r+p
p+q
Epsilon 30. (The Neuberg-Pedoe Inequality) Let a1 , b1 , c1 denote the sides of the triangle
A1 B1 C1 with area F1 . Let a2 , b2 , c2 denote the sides of the triangle A2 B2 C2 with area F2 .
Then, we have
a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 ) 16F1 F2 .
Notice that its a generalization of Weitzenb
ocks Inequality. Carlitz observed that The
Neuberg-Pedoe Inequality can be deduced from Aczels Inequality.
6

It is equivalent to The Hadwiger-Finsler Inequality.

24

INFINITY

Epsilon 31. (Aczels Inequality) If a1 , , an , b1 , , bn > 0 satisfies the inequality

a1 2 a2 2 + + an 2 and b1 2 b2 2 + + bn 2 ,
then the following inequality holds.
q

`
`
a1 b1 (a2 b2 + + an bn ) (a1 2 (a2 2 + + an 2 )) b1 2 b2 2 + + bn 2

INFINITY

25

3.2. Conway Substitution. As we saw earlier, transforming geometric inequalities to algebraic ones (and vice-versa), in order to solve them, may prove to be very useful. Besides
the Ravi Substitution, we remind another technique, known to the authors as the Conway
Substitution Theorem.
Theorem 3.5. (Conway) Let u, v, w be three reals such that the numbers v + w, w + u,
u + v and vw + wu + uv
a triangle XY Z with
Then, there exists
are all nonnegative.
sidelengths x = Y Z = v + w, y = ZX = w + u, z = XY = u + v. This triangle
satisfies y 2 + z 2 x2
= 2u, z 2 + x2 y 2 = 2v, x2 + y 2 z 2 = 2w. The area T of this
1
triangle equals T = 2 vw + wu + uv. If X = \ZXY , Y = \XY Z, Z = \Y ZX are the
u
v
w
angles of this triangle, then cot X = 2T
, cot Y = 2T
and cot Z = 2T
.

Proof.
Since

the numbers v + w, w + u, u + v are nonnegative, their square roots v + w,

w + u, u + v exist,
compu and, ofcourse, are
nonnegative as well.
A straightforward

tation
shows
that
w
+
u
+
u
+
v

v
+
w.
Similarly,
u
+
v
+
v
+
w

w+u

and v + w + w + u u + v. Thus, there exists a triangle XYZ with sidelengths

x = Y Z = v + w, y = ZX = w + u, z = XY = u + v.
It follows that
y 2 + z 2 x2 =

w+u

u+v

v+w

= 2u.

Similarly, z 2 + x2 y 2 = 2v and x2 + y 2 z 2 = 2w. According now to the fact that

x2 + y 2 z 2
,
4T
w
we deduce that so that cot Z = 2T
, and similarly cot X =
well-known trigonometric identity
cot Z =

u
2T

and cot Y =

v
2T

. The

cot Y cot Z + cot Z cot X + cot X cot Y = 1,

now becomes
or

v
w
w
u
u
v

=1
2T 2T
2T 2T
2T 2T
vw + wu + uv = 4T 2 .

or
T =

1 2
1
4T =
vw + wu + uv.
2
2

Note that the positive real numbers m, n, p satisfy the above conditions,
and therefore,

there exists a triangle with sidelengths m = n + p, n = p + m, p = m + n. However,

we will further see that there are such cases when we need the version in which the numbers
m, n, p are not all necessarily nonnegative.
Delta 22. (Turkey 2006) If x, y, z are positive numbers with xy + yz + zx = 1, show that

27
(x + y)(y + z)(z + x) ( x + y + y + z + z + x)2 6 3.
4
We continue with an interesting inequality discussed on the MathLinks Forum.
Proposition 3.1. If x, y, z are three reals such that the numbers y + z, z + x, x + y and
yz + zx + xy are all nonnegative, then
Xp

(z + x) (x + y) x + y + z + 3 yz + zx + xy.

26

INFINITY

Proof. (Darij Grinberg) Applying the Conway substitution theorem to the reals x, y, z,
we see that, since the numbers y + z, z + x, x + y and yz + zx + xy are all nonnegative,

we can conclude
that there exists a triangle ABC with sidelengths a = BC = y + z,

1
b = CA = z + x, c = AB = x + y and area S = 2 yz + zx + xy. Now, we have
Xp
X
X

(z + x) (x + y) =
z+x x+y =
b c = bc + ca + ab,

1
1` 2
2
2
2
x+y+z =
a + b2 + c2 ,
y+z +
z+x +
x+y
=
2
2
and

1

3 yz + zx + xy = 2 3
yz + zx + xy = 2 3 S.
2
Hence, the inequality in question becomes

1` 2
bc + ca + ab
a + b2 + c2 + 2 3 S,
2
which is equivalent with

a2 + b2 + c2 4 3 S + (b c)2 + (c a)2 + (a b)2 .

But this is the well-known refinement of the Weintzenbock Inequality, discovered by Finsler
and Hadwiger in 1937. See [FiHa].

Five years later, Pedoe [DP2] proved a magnificent generalization of the same Weitzenb
ock
Inequality. In Mitrinovic, Pecaric, and Volenecs classic Recent Advances in Geometric
Inequalities, this generalization is referred to as the Neuberg-Pedoe Inequality. See also
[DP1], [DP2], [DP3], [DP5] and [JN].
Proposition 3.2. (Neuberg-Pedoe) Let a, b, c, and x, y, z be the side lengths of two given
triangles ABC, XY Z with areas S, and T , respectively. Then,
`

a2 y 2 + z 2 x2 + b2 z 2 + x2 y 2 + c2 x2 + y 2 z 2 16ST,
with equality if and only if the triangles ABC and XY Z are similar.
Proof. (Darij Grinberg) First note that the inequality is homogeneous in the sidelengths
x, y, z of the triangle XY Z (in fact, these sidelengths occur in the power 2 on the left
hand side, and on the right hand side they occur in the power 2 as well, since the area
of a triangle is quadratically dependant from its sidelengths). Hence, this inequality is
invariant under any similitude transformation executed on triangle XY Z. In other words,
we can move, reflect, rotate and stretch the triangle XY Z as we wish, but the inequality
remains equivalent. But, of course, by applying similitude transformations to triangle
XY Z, we can always achieve a situation when Y = B and Z = C and the point X lies in
the same half-plane with respect to the line BC as the point A. Hence, in order to prove
the Neuberg-Pedoe Inequality for any two triangles ABC and XY Z, it is enough to prove
it for two triangles ABC and XY Z in this special situation.
So, assume that the triangles ABC and XY Z are in this special situation, i. e. that
we have Y = B and Z = C and the point X lies in the same half-plane with respect to
the line BC as the point A. We, thus, have to prove the inequality
`

a2 y 2 + z 2 x2 + b2 z 2 + x2 y 2 + c2 x2 + y 2 z 2 16ST.
Well, by the cosine law in triangle ABX, we have
AX 2 = AB 2 + XB 2 2 AB XB cos \ABX.
Lets figure out now what this equation means. At first, AB = c. Then, since B =
Y , we have XB = XY = z. Finally, we have either \ABX = \ABC \XBC or
\ABX = \XBC \ABC (depending on the arrangement of the points), but in both

INFINITY

27

cases cos \ABX = cos(\ABC \XBC). Since B = Y and C = Z, we can rewrite the
angle \XBC as \XY Z. Thus,
cos \ABX = cos (\ABC \XY Z) = cos \ABC cos \XY Z + sin \ABC sin \XY Z.
By the Cosine Law in triangles ABC and XY Z, we have
cos \ABC =

c 2 + a2 b 2
z 2 + x2 y 2
, and cos \XY Z =
.
2ca
2zx

Also, since
2T
2S
, and sin \XY Z =
,
ca
zx

sin \ABC =
we have that
cos \ABX
=
=

cos \ABC cos \XY Z + sin \ABC sin \XY Z

c 2 + a2 b 2 z 2 + x 2 y 2
2S 2T

.
2ca
2zx
ca zx

This makes the equation

AX 2 = AB 2 + XB 2 2 AB XB cos \ABX
transform into

AX 2 = c2 + z 2 2 c z
which immediately simplifies to
AX 2 = c2 + z 2 2

c2 + a2 b2 z 2 + x2 y 2
2S 2T

2ca
2zx
ca zx

!
`

c2 + a2 b2 z 2 + x2 y 2
4ST
+
,
4ax
ax

and since Y Z = BC,

` 2` 2

a y + z 2 x2 + b2 z 2 + x2 y 2 + c2 x2 + y 2 z 2 16ST
2
AX =
.
2ax
Thus, according to the (obvious) fact that AX 2 0, we conclude that
`

a2 y 2 + z 2 x2 + b2 z 2 + x2 y 2 + c2 x2 + y 2 z 2 16ST,
which proves the Neuberg-Pedoe Inequality. The equality holds if and only if the points
A and X coincide, i. e. if the triangles ABC and XY Z are congruent. Now, of course,
since the triangle XY Z we are dealing with is not the initial triangle XY Z, but just
its image under a similitude transformation, the general equality condition is that the
triangles ABC and XY Z are similar (not necessarily being congruent).

Delta 23. (Bottema [BK]) Let a, b, c, and x, y, z be the side lengths of two given triangles
ABC, XY Z with areas S, and T , respectively. If P is an arbitrary point in the plane of
triangle ABC, then we have the inequality
r
a2 (y 2 + z 2 x2 ) + b2 (z 2 + x2 y 2 ) + c2 (x2 + y 2 z 2 )
+ 8ST .
xAP +yBP +zCP
2
Epsilon 32. If A, B, C, X, Y , Z denote the magnitudes of the corresponding angles of
triangles ABC, and XY Z, respectively, then
cot A cot Y + cot A cot Z + cot B cot Z + cot B cot X + cot C cot X + cot C cot Y 2.

28

INFINITY

Epsilon 33. (Vasile C

artoaje) Let a, b, c, x, y, z be nonnegative reals. Prove the inequality
(ay + az + bz + bx + cx + cy)2 4 (bc + ca + ab) (yz + zx + xy) ,
with equality if and only if a : x = b : y = c : z.
Delta 24. (The Extended Tsintsifas Inequality) Let p, q, r be positive real numbers such
that the terms q + r, r + p, p + q are all positive, and let a, b, c denote the sides of a triangle
with area F . Then, we have

q 2
r 2
p
a2 +
b +
c 2 3F.
q+r
r+p
p+q
Epsilon 34. (Walter Janous, Crux Mathematicorum) If u, v, w, x, y, z are six reals such
that the terms y + z, z + x, x + y, v + w, w + u, u + v, and vw + wu + uv are all nonnegative,
then
p
x
y
z
(v + w) +
(w + u) +
(u + v) 3 (vw + wu + uv).
y+z
z+x
x+y
Note that the Neuberg-Pedoe Inequality is a generalization (actually the better word is
parametrization) of the Weitzenb
ock Inequality. How about deducing Hadwiger-Finslers
Inequality from it? Apparently this is not possible. However, the Conway Substitution
Theorem will change our mind.
Lemma 3.1. Let ABC be a triangle with side lengths a, b, c, and area S, and let u, v,
w be three reals such that the numbers v + w, w + u, u + v and vw + wu + uv are all
nonnegative. Then,

ua2 + vb2 + wc2 4 vw + wu + uv S.

Proof. According
a triangle with
to the Conway
Substitution
Theorem, we can construct

sidelenghts x = v + w, y = w + u, z = u + v and area T = vw + wu + uv/2. Let

this triangle be XY Z. In this case, by the Neuberg-Pedoe Inequality, applied for the
triangles ABC and XY Z, we get that
`

a2 y 2 + z 2 x2 + b2 z 2 + x2 y 2 + c2 x2 + y 2 z 2 16ST.
By the formulas given in the Conway Substitution Theorem, this becomes equivalent with
1
a2 2u + b2 2v + c2 2w 16S
vw + wu + uv
2

which simplifies to ua2 + vb2 + wc2 4 vw + wu + uv S.

Proposition 3.3. (Cosmin Pohoat

a) Let ABC be a triangle with side lengths a, b, c, and
area S and let x, y, z be three positive real numbers. Then,
2

2
x yz 2 y 2 zx 2 z 2 xy 2
a2 + b2 + c2 4 3S +
a +
b +
c .
x+y+z
x
y
z
Proof. Let m = xyz(x + y + z) 2yz(x2 yz), n = xyz(x + y + z) 2zx(y 2 zx), and
p = xyz(x + y + z) 2xy(z 2 xy). The three terms n + p, p + m, and m + n are all
positive, and since
mn + np + pm = 3x2 y 2 z 2 (x + y + z)2 0,
by Lemma 3.1, we get that
X

[xyz(x + y + z) 2yz(x2 yz)]a2 4xyz(x + y + z) 3S.

cyc

INFINITY

This rewrites as

X
cyc

x2 yz 2
(x + y + z) 2
a 4(x + y + z) 3S,
x

and, thus,

2
a + b + c 4 3S +
x+y+z
2

29

x2 yz 2 y 2 zx 2 z 2 xy 2
a +
b +
c .
x
y
z

Obviously, for x = a, y = b, z = c, and following the fact that

1
a3 + b3 + c3 3abc = (a + b + c) (a b)2 + (b c)2 + (c a)2 ,
2
Proposition 3.3 becomes equivalent with the Hadwiger-Finsler Inequality. Note also that
for x = y = z, Proposition 3.3 turns out to be the Weintzenbock Inequality. Therefore,
by using only Conways Substitution Theorem, weve transformed a result which strictly
generalizes the Weintzebock Inequality (the Neuberg-Pedoe Inequality) into one which
generalizes both the Weintzenbock Inequality and, surprisingly or not, the HadwigerFinsler Inequality.

30

INFINITY

3.3. Hadwiger-Finsler Revisited. The Hadwiger-Finsler inequality is known in literature

as a refinement of Weitzenb
ocks Inequality. Due to its great importance and beautiful
aspect, many proofs for this inequality are now known. For example, in [AE] one can find
eleven proofs. Is the Hadwiger-Inequality the best we can do? The answer is indeed no.
Here, we shall enlighten a few of its sharpening.
We begin with an interesting phenomenon. Most of you might know that according
to the formulas ab + bc + ca = s2 + r2 + 4Rr, and a2 + b2 + c2 = 2(s2 r2 4Rr), the
Hadwiger-Finsler Inequality rewrites as

4R + r s 3,
where s is the semiperimeter of the triangle. However, by using this last equivalent form
in a trickier way, we may obtain a slightly sharper result:
Proposition 3.4. (Cezar Lupu, Cosmin Pohoat
a) In any triangle ABC with sidelengths a,
b, c, circumradius R, inradius r, and area S, we have that

a2 + b2 + c2 2S 3 + 2r(4R + r) + (a b)2 + (b c)2 + (c a)2 .

Proof. As announced, we start with

4R + r s 3.

By multiplying with 2 and adding 2r(4R + r) to both terms, we obtain that

16Rr + 4r2 2S 3 + 2r(4R + r).

According now to the fact that ab+bc+ca = s2 +r2 +4Rr, and a2 +b2 +c2 = 2(s2 r2 4Rr),
this rewrites as

2(ab + bc + ca) (a2 + b2 + c2 ) 2S 3 + 2r(4R + r).

Therefore, we obtain

a2 + b2 + c2 2S 3 + 2r(4R + r) + (a b)2 + (b c)2 + (c a)2 .

This might seem strange, but wait until you see how does the geometric version of
Schurs Inequality look like (of course, since we expect to run through another refinement
of the Hadwiger-Finsler Inequality, we obviously refer to the third degree case of Schurs
Inequality).
Proposition 3.5. (Cezar Lupu, Cosmin Pohoat
a [LuPo]) In any triangle ABC with sidelengths a, b, c, circumradius R, inradius r, and area S, we have that
r
4(R 2r)
a2 + b2 + c2 4S 3 +
+ (a b)2 + (b c)2 + (c a)2 .
4R + r
Proof. The third degree case of Schurs Inequality says that for any three nonnegative real
numbers m, n, p, we have that
m3 + n3 + p3 + 3mnp m2 (n + p) + n2 (p + m) + p2 (m + n).
Note that this can be rewritten as
2(np + pm + mn) (m2 + n2 + p2 )
and by plugging in the substitutions x =

1
,
m

y=

1
,
n

9mnp
,
m+n+p

and z = p1 , we obtain that

zx
xy
9xyz
yz
+
+
+
2(x + y + z).
x
y
z
yz + zx + xy

INFINITY

31

So far so good, but lets take this now geometrically. Using the Ravi Substitution (i. e.
1
1
1
x = (b + c a), y = (c + a b), and p = (a + b c),
2
2
2
where a, b, c are the sidelengths of triangle ABC), we get that the above inequality
rewrites as
X (b + c a)(c + a b)
9(b + c a)(c + a b)(a + b c)
P
+
2(a + b + c).
(b + c a)(c + a b)
(a
+
b

c)
cyc
Since ab + bc + ca = s2 + r2 + 4Rr and a2 + b2 + c2 = 2(s2 r2 4Rr), it follows that
X
(b + c a)(c + a b) = 4r(4R + r).
cyc

Thus, according to Herons area formula that

p
S = s(s a)(s b)(s c),
we obtain

X (b + c a)(c + a b)
18sr
+
4s.
(a + b c)
4R + r
This is now equivalent to
X (s a)(s b)
9sr
+
2s,
(s

c)
4R
+r
cyc
and so

(s a)2 (s b)2 +

cyc

9s2 r3
2s2 r2 .
4R + r

By the identity
X

(s a) (s b) =

cyc

!2
(s a)(s b)

2s2 r2 ,

cyc

we have
X

!2
(s a)(s b)

2s2 r2 +

cyc

and since

9s2 r3
2s2 r2 ,
4R + r

(s a)(s b) = r(4R + r),

cyc

we deduce that
r2 (4R + r)2 +
This finally rewrites as

4R + r
s

9s2 r3
4s2 r2 .
4R + r

2
+

9r
4.
4R + r

According again to ab + bc + ca = s2 + r2 + 4Rr and a2 + b2 + c2 = 2(s2 r2 4Rr), we

have

2
2(ab + bc + ca) (a2 + b2 + c2 )
9r
4
.
4S
4R + r
Therefore,
r
4(R 2r)
a2 + b2 + c2 4S 3 +
+ (a b)2 + (b c)2 + (c a)2 .
4R + r

32

INFINITY

Epsilon 35. (Tran Quang Hung) In any triangle ABC with sidelengths a, b, c, circumradius
R, inradius r, and area S, we have
X

A X
B
C
a2 + b2 + c2 4S 3 + (a b)2 + (b c)2 + (c a)2 + 16Rr
cos2
cos cos
.
2
2
2
Delta 25. Let a, b, c be the lengths of a triangle with area S.
(a) (Cosmin Pohoat
a) Prove that

1
a2 + b2 + c2 4S 3 + (|a b| + |b c| + |c a|)2 .
2
(b) Show that, for all positive integers n,

n
4
a2n + b2n + c2n 3
+ (a b)2n + (b c)2n + (c a)2n .
3

INFINITY

33

3.4. Trigonometry Rocks! Trigonometry is an extremely powerful tool in geometry. We

begin with Fagnanos theorem that among all inscribed triangles in a given acute-angled
triangle, the feet of its altitudes are the vertices of the one with the least perimeter. Despite
of its apparent simplicity, the problem proved itself really challenging and attractive to
many mathematicians of the twentieth century. Several proofs are presented at [Fag].
Theorem 3.6. (Fagnanos Theorem) Let ABC be any triangle, with sidelengths a, b, c,
and area S. If XY Z is inscribed in ABC, then
8S 2
.
abc
Equality holds if and only if ABC is acute-angled, and then only if XY Z is its orthic
triangle.
XY + Y Z + ZX

Proof. (Finbarr Holland [FH]) Let XY Z be a triangle inscribed in ABC. Let x = BX,
y = CY , and z = AZ. Then 0 < x < a, 0 < y < b, 0 < z < c. By applying the Cosine
Law in the triangle ZBX, we have
ZX 2

(c z)2 + x2 2x(c z) cos B

(c z)2 + x2 + 2xc z) cos (A + C)

(x cos A + (c z) cos C)2 + (x sin A (c z) sin C)2 .

Hence, we have
ZX |x cos A + (c z) cos C|,
with equality if and only if x sin A = (c z) sin C or ax + cz = c2 , Similarly, we obtain
XY |y cos B + (a x) cos A|,
with equality if and only if ax + by = a2 . And
Y Z |z cos C + (b y) cos B|,
with equality if and only if by + cz = b2 . Thus, we get
XY + Y Z + ZX

|y cos B + (a x) cos A| + |z cos C + (b y) cos B| + |x cos A + (c z) cos C|

|y cos B + (a x) cos A + z cos C + (b y) cos B + x cos A + (c z) cos C|

|a cos A + b cos B + c cos C|

a2 (b2 + c2 a2 ) + b2 (c2 + a2 b2 ) + c2 (a2 + b2 c2 )

2abc
8S 2
=
.
abc
Note that we have equality here if and only if
=

ax + cz = c2 , ax + by = a2 , and by + cz = b2 ,
and moreover the expressions
u = x cos A + (c z) cos C, v = y cos B + (a x) cos A, w = z cos C + (b y) cos B,
are either all negative or all nonnegative. Now it is easy to very that the system of
equations
ax + cz = c2 , ax + by = a2 , by + cz = b2
has an unique solution given by
x = c cos B, y = a cos C, and z = b cos A,
in which case
u = b cos C, v = c cos C, and w = a cos A.

34

INFINITY

Thus, in this case, at most one of u, v, w can be negative. But, if one of u, v, w is zero,
then one of x, y, z must be zero, which is not possible. It follows that
8S 2
,
abc
unless ABC is acute-angled, and XY Z is its orthic triangle. If ABC is acute-angled, then
8S 2
is the perimeter of its orthic triangle, in which case we recover Fagnanos theorem.
abc
XY + Y Z + ZX >

We continue with Morleys miracle. We first prepare two well-known trigonometric

identities.
Epsilon 36. For all R, we have
sin (3) = 4 sin sin

2
+ sin
+ .
3
3

Epsilon 37. For all A, B, C R with A + B + C = 2, we have

cos2 A + cos2 B + cos2 C + 2 cos A cos B cos C = 1.
Theorem 3.7. (Morleys Theorem) The three points of intersections of the adjacent internal
angle trisectors of a triangle forms an equilateral triangle.
Proof. We want to show that the triangle E1 E2 E3 is equilateral.
Let R denote the circumradius of A1 A2 A3 . Setting \Ai = 3i for i = 1, 2, 3, we get
1 + 2 + 3 = 3 . We now apply The Sine Law twice to deduce

sin 2
sin
` 2 2 2R sin (33 ) = 8R sin 2 sin 3 sin
A1 E3 =
A1 A2 =
+ 3 .
sin ( 1 2 )
3
sin 3 + 3
By symmetry, we also have
A1 E2 = 8R sin 3 sin 2 sin

+ 2 .

3
Now, we present two different ways to complete the proof. The first method is more direct
and the second one gives more information.
First Method. One of the most natural approaches to crack this is to compute the lengths
of E1 E2 E3 . We apply The Cosine Law to obtain
E1 E2 2
AE3 2 + AE2 2 2 cos (\E3 A1 E2 ) AE3 AE1

i
h

= 64R2 sin2 2 sin2 3 sin2

+ 3 + sin2
+ 2 2 cos 1 sin
+ 3 sin
+ 2
3
3
3
3
To avoid long computation, here, we employ a trick. In the view of the equality

( 1 ) +
2 +
3 = ,
6
6
we have

cos2 ( 1 )+cos2
2 +cos2
3 +2 cos ( 1 ) cos
2 cos
3 = 1
6
6
6
6
or

cos2 1 + sin2
+ 2 + sin2
+ 3 2 cos 1
+ 2 cos
+ 3 = 1
3
3
3
3
or

sin2
+ 2 + sin2
+ 3 2 cos 1 sin
+ 2 sin
+ 3 = sin2 1 .
3
3
3
3
We therefore find that
E1 E2 2 = 64R2 sin2 1 sin2 2 sin2 3
=

INFINITY

35

so that
E1 E2 = 8R sin 1 sin 2 sin 3
Remarkably, the length of E1 E2 is symmetric in the angles! By symmetry, we therefore
conclude that E1 E2 E3 is an equilateral triangle with the length 8R sin 1 sin 2 sin 3 .
Second Method. We find the angles in the picture explicitly. Look at the triangle E3 A1 E2 .
The equality

1 +
+ 2 +
+ 3 =
3
3
allows us to invite a ghost triangle ABC having the angles

A = 1 , B = + 2 , C = + 3 .
3
3
Observe that two triangles BAC and E3 A1 E2 are similar. Indeed, we have \BAC =
\E3 A1 E2 and
`

8R sin 2 sin 3 sin 3 + 3

sin 3 + 3
sin C
AB
A1 E3
`
=
`
=
=
=
.
A1 E2
sin B
AC
8R sin 3 sin 2 sin 3 + 2
sin 3 + 2
It therefore follows that
(\A1 E3 E2 , \A1 E2 E3 ) =
Similarly, we also have

+ 3 .
3

+ 1 .
3
3
and

+ 1 ,
+ 2 .
(\A3 E2 E1 , \A3 E1 E2 ) =
3
3
An angle computation yields
(\A2 E1 E3 , \A2 E3 E1 ) =

\E1 E2 E3

+ 2 ,

+ 3 ,

2 (\A1 E2 E3 + \E1 E2 A3 + \A3 E2 A1 )

h

i
= 2
+ 3 +
+ 1 + ( 3 1 )
3
3

=
.
3
Similarly, we also have \E2 E3 E1 = 3 = \E3 E1 E2 . It follows that E1 E2 E3 is equilateral.
Furthermore, we apply The Sine Law to reach
sin
` 1 A1 E3
E2 E3 =
sin 3 + 3

sin
` 1 8R sin 2 sin 3 sin
+ 3
=
3
sin 3 + 3
=

8R sin 1 sin 2 sin 3 .

Hence, we find that the triangle E1 E2 E3 has the length 8R sin 1 sin 2 sin 3 .

We pass now to another miracle: the Steiner-Lehmus theorem.

Theorem 3.8. (The Steiner-Lehmus Theorem) If the internal angle-bisectors of two angles
of a triangle are congruent, then the triangle is isosceles.
Proof. [MH] Let BB 0 and CC 0 be the respective internal angle bisectors of angles B and
C in triangle ABC, and let a, b, and c denote the sidelengths of the triangle. We set
\B = 2, \C = 2, u = AB 0 , U = B 0 C, v = AC 0 , V = C 0 B.
We shall see that the assumptions BB 0 = CC 0 and C > B (and hence c > b) lead to the
contradiction that
b
c
b
c
<
and
.
u
v
u
v

36

INFINITY

Geometrically, this means that the line B 0 C 0 intersects both rays BC and CB. On the
one hand, we have
c
u+U
v+V
U
V
a
a
b
=

= <0
u
v
u
v
u
v
c
b
or

b
c
< .
u
v
On the other hand, we use the identity sin 2 = 2 sin cos to obtain
b v

c u

=
=
=
=
=

It thus follows that

b
u

sin B v

sin C u
2 cos sin v

2 cos sin u
cos sin
v

cos
u
sin
cos sin A CC 0

cos BB 0 sin A
cos
.
cos

> vc . We meet a contradiction.

The next inequality is probably the most beautiful modern geometric inequality in
triangle geometry.
Theorem 3.9. (The Erd
os-Mordell Theorem) If from a point P inside a given triangle ABC
perpendiculars P H1 , P H2 , P H3 are drawn to its sides, then
P A + P B + P C 2(P H1 + P H2 + P H3 ).
This was conjectured by Paul Erd
os in 1935, and first proved by Mordell in the same
year. Several proofs of this inequality have been given, using Ptolemys Theorem by Andre
Avez, angular computations with similar triangles by Leon Bankoff, area inequality by V.
Komornik, or using trigonometry by Mordell and Barrow.
Proof. [MB] We transform it to a trigonometric inequality. Let h1 = P H1 , h2 = P H2 and
h3 = P H 3 .
Apply the Since Law and the Cosine Law to obtain
q
h2 2 + h3 2 2h2 h3 cos( A),
P A sin A = H2 H3 =
q
P B sin B = H3 H1 =
h3 2 + h1 2 2h3 h1 cos( B),
q
P C sin C = H1 H2 =
h1 2 + h2 2 2h1 h2 cos( C).
So, we need to prove that
X 1 q
cyclic

sin A

h2 2 + h3 2 2h2 h3 cos( A) 2(h1 + h2 + h3 ).

The main trouble is that the left hand side has too heavy terms with square root expressions. Our strategy is to find a lower bound without square roots. To this end, we express
the terms inside the square root as the sum of two squares.
H2 H3 2

h2 2 + h3 2 2h2 h3 cos( A)

h2 2 + h3 2 2h2 h3 cos(B + C)

h2 2 + h3 2 2h2 h3 (cos B cos C sin B sin C).

INFINITY

37

Using cos2 B + sin2 B = 1 and cos2 C + sin2 C = 1, one finds that

H2 H3 2 = (h2 sin C + h3 sin B)2 + (h2 cos C h3 cos B)2 .
Since (h2 cos C h3 cos B)2 is clearly nonnegative, we get H2 H3 h2 sin C + h3 sin B.
Hence,
p 2
X
X h2 sin C + h3 sin B
h2 + h3 2 2h2 h3 cos( A)

sin A
sin A
cyclic
cyclic

X sin B
sin C
=
+
h1
sin C
sin B
cyclic
r
X
sin B sin C

h1

2
sin C sin B
cyclic

2h1 + 2h2 + 2h3 .

Epsilon 38. [SL 2005 KOR] In an acute triangle ABC, let D, E, F , P , Q, R be the feet
of perpendiculars from A, B, C, A, B, C to BC, CA, AB, EF , F D, DE, respectively.
Prove that
p(ABC)p(P QR) p(DEF )2 ,
where p(T ) denotes the perimeter of triangle T .
Epsilon 39. [IMO 2001/1 KOR] Let ABC be an acute-angled triangle with O as its
circumcenter. Let P on line BC be the foot of the altitude from A. Assume that \BCA
\ABC + 30 . Prove that \CAB + \COP < 90 .
Epsilon 40. [IMO 1961/2 POL] (Weitzenb
ocks Inequality) Let a, b, c be the lengths of a
triangle with area S. Show that

a2 + b2 + c2 4 3S.
Epsilon 41. (The Neuberg-Pedoe Inequality) Let a1 , b1 , c1 denote the sides of the triangle
A1 B1 C1 with area F1 . Let a2 , b2 , c2 denote the sides of the triangle A2 B2 C2 with area F2 .
Then, we have
a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 ) 16F1 F2 .
We close this subsection with Barrows Inequality stronger than The Erd
os-Mordell
Theorem. We need the following trigonometric inequality:
Proposition 3.6. (Wolstenholmes Inequality) Let x, y, z, 1 , 2 , 3 be real numbers with
1 + 2 + 3 = . Then, the following inequality holds:
x2 + y 2 + z 2 2(yz cos 1 + zx cos 2 + xy cos 3 ).
Proof. Using 3 = (1 + 2 ), we have the identity
x2 + y 2 + z 2 2(yz cos 1 + zx cos 2 + xy cos 3 )
=

[ z (x cos 2 + y cos 1 ) ]2 + [ x sin 2 y sin 1 ]2 .

Corollary 3.1. Let p, q, and r be positive real numbers. Let 1 , 2 , and 3 be real numbers
satisfying 1 + 2 + 3 = . Then, the following inequality holds.

1 qr
rp
pq
p cos 1 + q cos 2 + r cos 3
+
+
.
2 p
q
r
q
q
p
qr
and apply the above proposition.

Proof. Take (x, y, z) =

, rp
, pq
p
q
r

38

INFINITY

Delta 26. (Cosmin Pohoat

a) Let a, b, c be the sidelengths of a given triangle ABC with
circumradius R, and let x, y, z be three arbitrary real numbers. Then, we have that
r
r
r
p
yz
zx
xy
R
+
+
xa2 + yb2 + zc2 .
x
y
z
Epsilon 42. (Barrows Inequality) Let P be an interior point of a triangle ABC and let
U , V , W be the points where the bisectors of angles BP C, CP A, AP B cut the sides
BC,CA,AB respectively. Then, we have
P A + P B + P C 2(P U + P V + P W ).
Epsilon 43. [AK] Let x1 , , x4 be positive real numbers. Let 1 , , 4 be real numbers
such that 1 + + 4 = . Then, we have
r
(x1 x2 + x3 x4 )(x1 x3 + x2 x4 )(x1 x4 + x2 x3 )
x1 cos 1 + x2 cos 2 + x3 cos 3 + x4 cos 4
.
x1 x2 x3 x4
Delta 27. [RS] Let R, r, s > 0. Show that a necessary and sufficient condition for the
existence of a triangle with circumradius R, inradius r, and semiperimeter s is
s4 2(2R2 + 10Rr r2 )s2 + r(4R + r)2 0.

INFINITY

39

3.5. Erd
os, Brocard, and Weitzenb
ock. In this section, we touch Brocard geometry. We
begin with a consequence of The Erd
os-Mordell Theorem.
Epsilon 44. [IMO 1991/5 FRA] Let ABC be a triangle and P an interior point in ABC.
Show that at least one of the angles \P AB, \P BC, \P CA is less than or equal to 30 .
As an immediate consequence, one may consider the following symmetric situation:
Proposition 3.7. Let ABC be a triangle. If there exists an interior point P in ABC
satisfying that
\P AB = \P BC = \P CA =
for some positive real number . Then, we have the inequality

.
6

We omit the geometrical proof of the existence and the uniqueness of such point in an
arbitrary triangle.(Prove it!)
Delta 28. Let ABC be a triangle. There exists a unique interior point 1 , which bear the
name the first Brocard point of ABC, such that
\1 AB = \1 BC = \1 CA = 1
for some 1 , the first Brocard angle.
By symmetry, we also include
Delta 29. Let ABC be a triangle. There exists a unique interior point 2 with
\2 BA = \2 CB = \2 AC = 2
for some 2 , the second Brocard angle. The point 2 is called the second Brocard point of
ABC.
Delta 30. If a triangle ABC has an interior point P such that \P AB = \P BC =
\P CA = 30 , then it is equilateral.
Epsilon 45. Any triangle has the same Brocard angles.
As a historical remark, we state that H. Brocard (1845-1922) was not the first one who
discovered the Brocard points. They were also known to A. Crelle (1780-1855), C. Jacobi
(1804-1851), and others some 60 years earlier. However, their results in this area were
soon forgotten [RH]. Our next job is to evaluate the Brocard angle quite explicitly.
Epsilon 46. The Brocard angle of the triangle ABC satisfies
cot = cot A + cot B + cot C.
Proposition 3.8. The Brocard angle of the triangle with sides a, b, c and area S satisfies
cot =

a2 + b 2 + c 2
.
4S

Proof. We have
cot A + cot B + cot C

=
=
=

2bc cos A
2ca cos B
2ab cos C
+
+
2bc sin A
2ca sin B
2ab sin C
b2 + c2 a2
c 2 + a2 b 2
a2 + b 2 c 2
+
+
4S
4S
4S
a2 + b2 + c2
.
4S

We revisit Weitzenb
ocks Inequality. It is a corollary of The Erd
os-Mordell Theorem!

40

INFINITY

Proposition 3.9. [IMO 1961/2 POL] (Weitzenb

ocks Inequality) Let a, b, c be the lengths
of a triangle with area S. Show that

a2 + b2 + c2 4 3S.
Third Proof. Letting denote its Brocard angle, by combining results we proved, we
obtain

a2 + b 2 + c 2
= 3.
= cot cot
4S
6

We present interesting theorems from Brocard geometry.

Delta 31. [RH] Let 1 and 2 denote the Brocard points of a triangle ABC with the
circumcenter O. Let the circumcircle of O1 2 , called the Brocard circle of ABC, meet
the line A1 , B1 , C1 at R, P , Q, respectively, again. The triangle P QR bears the
name the first Brocard triangle of ABC.
(a) O1 = O2 .
(b) Two triangles P QR and ABC are similar.
(c) Two triangles P QR and ABC have the same centroid.
(d) Let U , V , W denote the midpoints of QR, RP , P Q, respectively. Let UH , VH , WH
denote the feet of the perpendiculars from U , V , W respectively. Then, the three lines
U UH , V UH , W WH meet at the nine point circle of triangle ABC.
The story is not over. We establish an inequality which implies the problem [IMO
1991/5 FRA].
Epsilon 47. (The Trigonometric Versions of Cevas Theorem) For an interior point P of a
triangle A1 A2 A3 , we write
\A3 A1 A2 = 1 , \P A1 A2 = 1 , \P A1 A3 = 1 ,
\A1 A2 A3 = 2 , \P A2 A3 = 2 , \P A2 A1 = 2 ,
\A2 A3 A1 = 3 , \P A3 A1 = 3 , \P A3 A2 = 3 .
Then, we find a hidden symmetry:
sin 1 sin 2 sin 3

= 1,
sin 1 sin 2 sin 3
or equivalently,
1
= [cot 1 cot 1 ] [cot 2 cot 2 ] [cot 3 cot 3 ] .
sin 1 sin 2 sin 3
Epsilon 48. Let P be an interior point of a triangle ABC. Show that

cot (\P AB) + cot (\P BC) + cot (\P CA) 3 3.

Proposition 3.10. [IMO 1991/5 FRA] Let ABC be a triangle and P an interior point in
ABC. Show that at least one of the angles \P AB, \P BC, \P CA is less than or equal
to 30 .
Second Solution. The above inequality implies
max{ cot (\P AB) , cot (\P BC) , cot (\P CA) }

3 = cot 30 .

Since the cotangent function is strictly decreasing on (0, ), we get the result.

INFINITY

41

3.6. From Incenter to Centroid. We begin with an inequality regarding the incenter. In
fact, the geometric inequality is equivalent to an algebraic one, Schurs Inequality!
Example 7. (Korea 1998) Let I be the incenter of a triangle ABC. Prove that
IA2 + IB 2 + IC 2

BC 2 + CA2 + AB 2
.
3

Proof. Let BC = a, CA = b, AB = c, and s = a+b+c

. Letting r denote the inradius of
2
4ABC, we have
(s a)(s b)(s c)
.
r2 =
s
By The Pythagoras Theorem, the inequality is equivalent to

1` 2
(s a)2 + r2 + (s b)2 + r2 + (s c)2 + r2
a + b2 + c2 .
3
or

3(s a)(s b)(s c)

1` 2
(s a)2 + (s b)2 + (s c)2 +

a + b2 + c2 .
s
3
After The Ravi Substitution x = s a, y = s b, z = s c, it becomes
x2 + y 2 + z 2 +
or
or

(x + y)2 + (y + z)2 + (z + x)2

3xyz

x+y+z
3

3 x2 + y 2 + z 2 (x + y + z) + 9xyz (x + y + z) (x + y)2 + (y + z)2 + (z + x)2

`

9xyz (x + y + z) 2xy + 2yz + 2zx x2 y 2 z 2

or
9xyz x2 y + x2 z + y 2 z + y 2 x + z 2 x + z 2 y + 6xyz x3 y 3 z 3
or
x3 + y 3 + z 3 + 3xyz x2 (y + z) + y 2 (z + x) + z 2 (x + y).
This is a particular case of Schurs Inequality.

Now, one may ask more questions. Can we replace the incenter by other classical points
in triangle geometry? The answer is yes. We first take the centroid.
Example 8. Let G denote the centroid of the triangle ABC. Then, we have the geometric
identity
BC 2 + CA2 + AB 2
GA2 + GB 2 + GC 2 =
.
3
Proof. Let M denote the midpoint of BC. The Pappus Theorem implies that

2
2
2
GB 2 + GC 2
BC
GA
BC
= GM 2 +
=
+
2
2
2
2
or
GA2 + 2GB 2 + 2GC 2 = BC 2
2
2
Similarly, 2GA GB + 2GC 2 = CA2 and 2GA2 + 2GB 2 2GC 2 = AB 2 . Adding these
three equalities, we get the identity.

Before we take other classical points, we need to rethink this unexpected situation. We
have an equality, instead of an inequality. According to this equality, we find that the
previous inequality can be rewritten as
IA2 + IB 2 + IC 2 GA2 + GB 2 + GC 2 .
Now, it is quite reasonable to make a conjecture which states that, given a triangle ABC,
the minimum value of P A2 + P B 2 + P C 2 is attained when P is the centroid of 4ABC.
This guess is true!

42

INFINITY

Theorem 3.10. Let A1 A2 A3 be a triangle with the centroid G. For any point P , we have
P A1 2 + P A2 2 + P A3 2 GA2 + GB 2 + GC 2 .
Proof. Just toss the picture on the real plane R2 so that
x + x + x y + y + y
1
2
3
1
2
3
P (p, q), A1 (x1 , y1 ) , A2 (x2 , y2 ) , A3 (x3 , y3 ) , G
,
.
3
3
What we need to do is to compute
`
`

3 P A1 2 + P A2 2 + P A3 2 BC 2 + CA2 + AB 2
3
3
2 y + y + y
2
X
X
x1 + x2 + x3
1
2
3
xi +
yi
= 3
(p xi )2 + (q yi )2
3
3
i=1
i=1
=

3
X

(p xi )2

i=1

3
X
x1 + x2 + x3
i=1

xi

+3

3
X

(q yi )2

i=1

3
X
y1 + y2 + y3
i=1

yi

A moments thought shows that the quadratic polynomials are squares.

3
3
2

X
X
x1 + x2 + x3
x1 + x2 + x3 2
xi = 9 p
3
(p xi )2
,
3
3
i=1
i=1

y1 + y2 + y3 2
=9 q
.
3
3
i=1
i=1
`

Hence, the quantity 3 P A1 2 + P A2 2 + P A3 2 BC 2 + CA2 + AB 2 is clearly nonnegative. Furthermore, we notice that the above proof of the geometric inequality discovers a geometric identity:
`
`

P A2 + P B 2 + P C 2 GA2 + GB 2 + GC 2 = 9GP 2 .
3

3
X

(q yi )2

3
X
y1 + y2 + y3

yi

It is clear that the equality in the above inequality holds only when GP = 0 or P = G.

After removing the special condition that P is the incenter, we get a more general
inequality, even without using a heavy machine, like Schurs Inequality. Sometimes, generalizations are more easy! Taking the point P as the circumcenter, we have
Proposition 3.11. Let ABC be a triangle with circumradius R. Then, we have
AB 2 + BC 2 + CA2 9R2 .
Proof. Let O and G denote its circumcenter and centroid, respectively. It reads
`

9GO2 + AB 2 + BC 2 + CA2 = 3 OA2 + OB 2 + OC 2 = 9R2 .

The readers can rediscover many geometric inequalities by taking other classical points
from triangle geometry.(Do it!) Here goes another inequality regarding the incenter.
Example 9. Let I be the incenter of the triangle ABC with BC = a, CA = b and AB = c.
Prove that, for all points X,
aXA2 + bXB 2 + cXC 2 abc.
First Solution. It turns out that the non-negative quantity
aXA2 + bXB 2 + cXC 2 abc
has a geometric meaning. This geometric inequality follows from the following geometric
identity:
aXA2 + bXB 2 + cXC 2 = (a + b + c)XI 2 + abc. 7
7

[SL 1988 SGP]

INFINITY

43

There are many ways to establish this identity. To euler8 it, we toss the picture on the
real plane R2 with the coordinates
A(c cos B, c sin B), B(0, 0), C(a, 0).
, we get I(sb, r). It is well-known
Let r denote the inradius of 4ABC. Setting s = a+b+c
2
that
(s a)(s b)(s c)
r2 =
.
s
Set X(p, q). On the one hand, we obtain
=
=
=

aXA2 + bXB 2 + cXC 2

a (p c cos B)2 + (q c sin B)2 + b p2 + q 2 + c (p a)2 + q 2

(a + b + c)p2 2acp(1 + cos B) + (a + b + c)q 2 2acq sin B + ac2 + a2 c

[4ABC]
a2 + c 2 b 2
2sp2 2acp 1 +
+ 2sq 2 2acq 1
+ ac2 + a2 c
2ac
ac
2

2sp2 p(a + c + b) (a + c b) + 2sq 2 4q[4ABC] + ac2 + a2 c

2sp2 p(2s) (2s 2b) + 2sq 2 4qsr + ac2 + a2 c

2sp2 4s (s b) p + 2sq 2 4rsq + ac2 + a2 c.

On the other hand, we obtain

(a + b + c)XI 2 + abc

2s (p (s b))2 + (q r)2
2

2s p 2(s b)p + (s b)2 + q 2 2qr + r2

2sp2 4s (s b) p + 2s(s b)2 + 2sq 2 4rsq + 2sr2 + abc.

It thus follows that

aXA2 + bXB 2 + cXC 2 (a + b + c)XI 2 abc
=

ac2 + a2 c 2s(s b)2 2sr2 abc

ac(a + c) 2s(s b)2 2(s a)(s b)(s c) abc

ac(a + c b) 2s(s b)2 2(s a)(s b)(s c)

2ac(s b) 2s(s b)2 2(s a)(s b)(s c)

2(s b) [ac s(s b) 2(s a)(s c)] .

However, we compute ac s(s b) 2(s a)(s c) = 2s2 + (a + b + c)s = 0.

Now, throw out the special condition that I is the incenter! Then, the essence appears:
Delta 32. (The Leibniz Theorem) Let 1 , 2 , 3 be real numbers such that 1 +2 +3 6= 0.
We characterize the generalized centroid G = G(1 ,2 ,3 ) by
3

X
XG =
i=1

i
XAi .
1 + 2 + 3

Then G is well-defined in the sense that it doesnt depend on the choice of X. For all
points P , we have
3
X

i P Ai 2 = (1 + 2 + 3 )P G 2 +

i=1

3
X
i=1

i i+1
Ai Ai+1 2 .
1 + 2 + 3

We show that the geometric identity aXA + bXB 2 + cXC 2 = (a + b + c)XI 2 + abc is
a straightforward consequence of The Leibniz Theorem.
8
euler v. (in Mathematics) transform the geometric identity in triangle geometry to trigonometric or algebraic identity.

44

INFINITY

Second Solution. Let BC = a, CA = b, AB = c. With the weights (a, b, c), we have

I = G(a,b,c) . Hence,
aXA2 + bXB 2 + cXC 2

=
=

bc
ca
ab
a2 + +
b2 +
c2
a+b+c
a+b+c
a+b+c
(a + b + c)XI 2 + abc.
(a + b + c)XI 2 +

Epsilon 49. [IMO 1961/2 POL] (Weitzenb

ocks Inequality) Let a, b, c be the lengths of a
triangle with area S. Show that

a2 + b2 + c2 4 3S.
Epsilon 50. (The Neuberg-Pedoe Inequality) Let a1 , b1 , c1 denote the sides of the triangle
A1 B1 C1 with area F1 . Let a2 , b2 , c2 denote the sides of the triangle A2 B2 C2 with area F2 .
Then, we have
a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 ) 16F1 F2 .
Delta 33. [SL 1988 UNK] The triangle ABC is acute-angled. Let L be any line in
the plane of the triangle and let u, v, w be lengths of the perpendiculars from A, B, C
respectively to L. Prove that
u2 tan A + v 2 tan B + w2 tan C 24,
where 4 is the area of the triangle, and determine the lines L for which equality holds.
Delta 34. [KWL] Let G and I be the centroid and incenter of the triangle ABC with
inradius r, semiperimeter s, circumradius R. Show that

1` 2
IG2 =
s + 5r2 16Rr .
9

Inspiration is needed in geometry, just as much as in poetry. - A. Pushkin

INFINITY

45

4. Geometry Revisited

It gives me the same pleasure when someone else proves a good

theorem as when I do it myself.
- E. Landau

4.1. Areal Co-ordinates. In this section we aim to briefly introduce develop the theory of
areal (or barycentric) co-ordinate methods with a view to making them accessible to a
reader as a means for solving problems in plane geometry. Areal co-ordinate methods are
particularly useful and important for solving problems based upon a triangle, because,
unlike Cartesian co-ordinates, they exploit the natural symmetries of the triangle and
many of its key points in a very beautiful and useful way.
4.1.1. Setting up the co-ordinate system. If we are going to solve a problem using areal coordinates, the first thing we must do is choose a triangle ABC, which we call the triangle
of reference, and which plays a similar role to the axes in a cartesian co-ordinate system.
Once this triangle is chosen, we can assign to each point P in the plane a unique triple
(x, y, z) fixed such that x + y + z = 1, which we call the areal co-ordinates of P . The
way these numbers are assigned can be thought of in three different ways, all of which
are useful in different circumstances. We leave the proofs that these three conditions are
equivalent, along with a proof of the uniqueness of areal co-ordinate representation, for
the reader. The first definition we shall see is probably the most intuitive and most useful
for working with. It also explains why they are known as areal co-ordinates.
1st Definition: A point P internal to the triangle ABC has areal co-ordinates

[P BC] [P CA] [P AB]

,
,
.
[ABC] [ABC] [ABC]
If a sign convention is adopted, such that a triangle whose vertices are labelled clockwise
has negative area, this definition applies for all P in the plane.
2nd Definition: If x, y, z are the masses we must place at the vertices A, B, C respectively such that the resulting system has centre of mass P , then (x, y, z) are the areal
co-ordinates of P (hence the alternative name barycentric)
3rd Definition: If we take a system of vectors with arbitrary origin (not on the sides of
triangle ABC) and let a, b, c, p be the position vectors of A, B, C, P respectively, then
p = xa + yb + zc for some triple (x, y, z) such that x + y + z = 1. We define this triple
as the areal co-ordinates of P .
There are some remarks immediately worth making:
The vertices A, B, C of the triangle of reference have co-ordinates (1, 0, 0), (0, 1, 0),
(0, 0, 1) respectively.
All the co-ordinates of a point are positive if and only if the point lies within the
triangle of reference, and if any of the co-ordinates are zero, the point lies on one
of the sides (or extensions of the sides) of ABC.

46

INFINITY

4.1.2. The Equation of a Line. A line is a geometrical object such that any pair of nonparallel lines meet at one and only one point. We would therefore expect the equation of
a line to be linear, such that any pair of simultaneous line equations, together with the
condition x + y + z = 1, can be solved for a unique triple (x, y, z) corresponding to the
areal co-ordinates of the point of intersection of the two lines. Indeed, it follows (using
the equation x + y + z = 1 to eliminate any constant terms) that the general equation of
a line is of the form
lx + my + nz = 0
where l, m, n are constants and not all zero. Clearly there exists a unique line (up to
multiplication by a constant) containing any two given points P (xp , yp , zp ), Q(xq , yq , zq ).
This line can be written explicitly as
(yp zq yq zp )x + (zp xq zq xp )y + (xp yq xq yp )z = 0
This equation is perhaps more neatly expressed in the determinant9 form:
0
1
x xp xq
Det @ y yp yq A = 0.
z z p zq
While the above form is useful, it is often quicker to just spot the line automatically.
For example try to spot the equation of the line BC, containing the points B(0, 1, 0) and
C(0, 0, 1), without using the above equation.
Of particular interest (and simplicity) are Cevian lines, which pass through the vertices
of the triangle of reference. We define a Cevian through A as a line whose equation is
of the form my = nz. Clearly any line containing A must have this form, because setting
y = z = 0, x = 1 any equation with a nonzero x coefficient would not vanish. It is easy to
see that any point on this line therefore has form (x, y, z) = (1 mt nt, nt, mt) where t is
a parameter. In particular, it will intersect the side BC with equation x = 0 at the point
n
m
U (0, m+n
, m+n
). Note that from definition 1 (or 3) of areal co-ordinates, this implies that
the ratio

BU
UC

[ABU ]
[AU C]

m
.
n

4.1.3. Example: Cevas Theorem. We are now in a position to start using areal coordinates to prove useful theorems. In this section we shall state and prove (one direction
of) an important result of Euclidean geometry known as Cevas Theorem. The author
recommends a keen reader only reads the statement of Cevas theorem initially and tries
to prove it for themselves using the ideas introduced above, before reading the proof given.
9
The Determinant of a 3 3 Matrix. Matrix determinants play an important role in areal
co-ordinate methods. We define the determinant of a 3 by 3 square matrix A as
0
1
ax bx cx
@
Det(A) = Det
ay by cy A = ax (by cz bz cy ) + ay (bz cx bxc z) + az (bx cy by cx ).
az bz cz

This can be thought of as (as the above equation suggests) multiplying each element of the first
column by the determinants of 2x2 matrices formed in the 2nd and 3rd columns and the rows not
containing the element of the first column. Alternatively, if you think of the matrix as wrapping
around (so bx is in some sense directly beneath bz in the above matrix) you can simply take the
sum of the products of diagonals running from top-left to bottom-right and subtract from it the
sum of the products of diagonals running from bottom-left to top-right (so think of the above
RHS as (ax by cz + ay bz cx + az bx cy ) (az by cx + ax bz cy + ay bx cz )). In any case, it is worth making
sure you are able to quickly evaluate these determinants if you are to be successful with areal
co-ordinates.

INFINITY

47

Theorem 4.1. (Cevas Theorem) Let ABC be a triangle and let L, M , N be points on
the sides BC, CA, AB respectively. Then the cevians AL, BM , CN are concurrent at a
point P if and only if
BL CM AN

=1
LC M A N B
Proof. Suppose first that the cevians are concurrent at a point P , and let P have areal
co-ordinates (p, q, r).
Then ALhas equation qz = ry (following the discussion of Cevian
q
, r
q+r q+r
we get BL
LC

lines above), so L 0,

, which implies

Taking their product

leave the converse to the reader.

CM
MA

AN
NB

BL
LC

r
.
q

Similarly,

CM
MA

p AN
,
r NB

q
.
p

= 1, proving one direction of the theorem. We

The above proof was very typical of many areal co-ordinate proofs. We only had to
go through the details for one of the three cevians, and then could say similarly and
obtain ratios for the other two by symmetry. This is one of the great advantages of the
areal co-ordinate system in solving problems where such symmetries do exist (particularly
problems symmetric in a triangle ABC: such that relabelling the triangle vertices would
result in the same problem).
4.1.4. Areas and Parallel Lines. One might expect there to be an elegant formula for the
area of a triangle in areal co-ordinates, given they are a system constructed on areas. Indeed, there is. If P QR is an arbitrary triangle with P (xp , yp , zp ) , Q (xq , yq , zq ) , R (xr , yr , zr )
then
0
1
xp xq xr
[P QR]
= Det @ yp yq yr A
[ABC]
zp zq zr
An astute reader might notice that this seems like a plausible formula, because if P, Q, R
are collinear, it tells us that the triangle P QR has area zero, by the line formula already
mentioned. It should be noted that the area comes out as negative if the vertices P QR
are labelled in the opposite direction to ABC.
It is now fairly obvious what the general equation for a line parallel to a given line passing through two points (x1 , y1 , z1 ), (x2 , y2 , z2 ) should be, because the area of the triangle
formed by any point on such a line and these two points must be constant, having a
constant base and constant height. Therefore this line has equation
0
1
x x1 x2
Det @ y y1 y2 A = k = k(x + y + z),
z z1 z2
where k R is a constant.
Delta 35. (United Kingdom 2007) Given a triangle ABC and an arbitrary point P internal
to it, let the line through P parallel to BC meet AC at M , and similarly let the lines
through P parallel to CA,AB meet AB,BC at N ,L respectively. Show that
BL CM AN
1

LC M A N B
8
Delta 36. (Nikolaos Dergiades) Let DEF be the medial triangle of ABC, and P a point
with cevian triangle XY Z (with respect to ABC). Find P such that the lines DX, EY ,
F Z are parallel to the internal bisectors of angles A, B, C, respectively.

48

INFINITY

4.1.5. To infinity and beyond. Before we start looking at some more definite specific useful
tools (like the positions of various interesting points in the triangle), we round off the
general theory with a device that, with practice, greatly simplifies areal manipulations.
Until now we have been acting subject to the constraint that x + y + z = 1. In reality, if we
are just intersecting lines with lines or lines with conics, and not trying to calculate any
ratios, it is legitimate to ignore this constraint and to just consider the points (x, y, z) and
(kx, ky, kz) as being the same point for all k 6= 0. This is because areal co-ordinates are a
special case of a more general class of co-ordinates called projective homogeneous coordinates10, where here the projective line at infinity is taken to be the line x + y + z = 0.
This system only works if one makes all equations homogeneous (of the same degree in
x, y, z), so, for example, x+y = 1 and x2 +y = z are not homogeneous, whereas x+yz = 0
and a2 yz + b2 zx + c2 xy = 0 are homogeneous. We can therefore, once all our line and
conic equations are happily in this form, no longer insist on x + y + z = 1, meaning points
a
b
c
like the incentre ( a+b+c
, a+b+c
, a+b+c
) can just be written (a, b, c). Such represenataions
are called unnormalised areal co-ordinates and usually provide a significant advantage for
the practical purposes of doing manipulations. However, if any ratios or areas are to be
calculated, it is imperative that the co-ordinates are normalised again to make x+y+z = 1.
This process is easy: just apply the map

x
y
z
(x, y, z) 7
,
,
x+y+z x+y+z x+y+z
4.1.6. Significant areal points and formulae in the triangle. We have seen that the vertices
are given by A(1, 0, 0), B(0, 1, 0), C(0, 0, 1), and the sides by x = 0,y = 0,z = 0. In the
section on the equation of a line we examined the equation of a cevian, and this theory
can, together with other knowledge of the triangle, be used to give areal expressions for
familiar points in Euclidean triangle geometry. We invite the reader to prove some of the
facts below as exercises.

Triangle centroid: G(1, 1, 1).11

Centre of the inscribed circle: I(a, b, c).12
Centres of escribed circles: Ia (a, b, c), Ib (a, b, c), Ic (a, b, c).
Symmedian point: K(a2 , b2 , c2 ).
Circumcentre: O(sin 2A, sin 2B, sin 2C).
Orthocentre: H(tan A, tan B, tan C).

The isogonal conjugate of P (x, y, z): P

The isotomic conjugate of P (x, y, z): P

a2 b2 c2
, y, z
x
1 1 1
, ,
x y z

It should be noted that the rather nasty trigonometric forms of O and H mean that
they should be approached using areals with caution, preferably only if the calculations
will be relatively simple.
Delta 37. Let D, E be the feet of the altitudes from A and B respectively, and P, Q
the meets of the angle bisectors AI,BI with BC,CA respectively. Show that D,I,E are
collinear if and only if P ,O,Q are.
10

The author regrets that, in the interests of concision, he is unable to deal with these coordinates in this document, but strongly recommends Christopher Bradleys The Algebra of Geometry, published by Highperception, as a good modern reference also with a more detailed account of areals and a plethora of applications of the methods touched on in this document. Even
better, though only for projectives and lacking in the wealth of fascinating modern examples, is
E.A.Maxwells The methods of plane projective geometry based on the use of general homogeneous
coordinates, recommended to the present author by the author of the first book.
11
The midpoints of the sides BC, CA, AB are given by (0, 1, 1),(1, 0, 1) and (1, 1, 0) respectively.
12
Hint: use the angle bisector theorem.

INFINITY

49

4.1.7. Distances and circles. We finally quickly outline some slightly more advanced theory, which is occasionally quite useful in some problems, We show how to manipulate
conics (with an emphasis on circles) in areal co-ordinates, and how to find the distance
between two points in areal co-ordinates. These are placed in the same section because the
formulae look quite similar and the underlying theory is quite closely related. Derivations
can be found in [Bra1].
Firstly, the general equation of a conic in areal co-ordinates is, since a conic is a general equation of the second degree, and areals are a homogeneous system, given by
px2 + qy 2 + rz 2 + 2dyz + 2ezx + 2f xy = 0
Since multiplication by a nonzero constant gives the same equation, we have five independent degrees of freedom, and so may choose the coefficients uniquely (up to multiplication
by a constant) in such a way as to ensure five given points lie on such a conic.
In Euclidean geometry, the conic we most often have to work with is the circle. The
most important circle in areal co-ordinates is the circumcircle of the reference triangle,
which has the equation (with a, b, c equal to BC, CA, AB respectively)
a2 yz + b2 zx + c2 xy = 0
In fact, sharing two infinite points13 with the above, a general circle is just a variation on
this theme, being of the form
a2 yz + b2 zx + c2 xy + (x + y + z)(ux + vy + wz) = 0
We can, given three points, solve the above equation for u, v, w substituting in the three
desired points to obtain the equation for the unique circle passing through them.
Now, the areal distance formula looks very similar to the circumcircle equation. If we
have a pair of points P (x1 , y1 , z1 ) and Q(x2 , y2 , z2 ), which must be normalised, we may
define the displacement P Q : (x2 x1 , y2 y1 , z2 z1 ) = (u, v, w), and it is this we shall
measure the distance of. So the distance of a displacement P Q(u, v, w), u + v + w = 0 is
given by
P Q2 = a2 vw b2 wu c2 uv
Since u+v+w = 0 this is, despite the negative signs, always positive unless u = v = w = 0.
Delta 38. Use the vector definition of areal co-ordinates to prove the areal distance formula
and the circumcircle formula.
4.1.8. Miscellaneous Exercises. Here we attach a selection of problems compiled by Tim
Hennock, largely from UK IMO activities in 2007 and 2008. None of them are trivial, and
some are quite difficult. Good luck!
Delta 39. (UK Pre-IMO training 2007) Let ABC be a triangle. Let D, E, F be the reflections of A, B, C in BC, AC, AB respectively. Show that D, E, F are collinear if and only
if OH = 2R.
Delta 40. (Balkan MO 2005) Let ABC be an acute-angled triangle whose inscribed circle
touches AB and AC at D and E respectively. Let X and Y be the points of intersection of
the bisectors of the angles \ACB and \ABC with the line DE and let Z be the midpoint
of BC. Prove that the triangle XY Z is equilateral if and only if \A = 60
13

All circles have two (imaginary) points in common on the line at infinity. It follows that if
a conic is a circle, its behaviour at the line at infinity x + y + z = 0 must be the same as that of
the circumcircle, hence the equation given.

50

INFINITY

Delta 41. (United Kingdom 2007) Triangle ABC has circumcentre O and centroid M .
The lines OM and AM are perpendicular. Let AM meet the circumcircle of ABC again
at A0 . Lines CA0 and AB intersect at D and BA0 and AC intersect at E. Prove that the
circumcentre of triangle ADE lies on the circumcircle of ABC.
Delta 42. [IMO 2007/4] In triangle ABC the bisector of \BCA intersects the circumcircle again at R, the perpendicular bisector of BC at P , and the perpendicular bisector
of AC at Q. The midpoint of BC is K and the midpoint of AC is L. Prove that the
triangles RP K and RQL have the same area.
Delta 43. (RMM 2008) Let ABC be an equilateral triangle. P is a variable point internal
to the triangle, and its perpendicular distances to the sides are denoted by a2 , b2 and c2
for positive real numbers a, b and c. Find the locus of points P such that a, b and c can
be the side lengths of a non-degenerate triangle.
Delta 44. [SL 2006] Let ABC be a triangle such that \C < \A < 2 . Let D be on AC
such that BD = BA. The incircle of ABC touches AB at K and AC at L. Let J be the
incentre of triangle BCD. Prove that KL bisects AJ.
Delta 45. (United Kingdom 2007) The excircle of a triangle ABC touches the side AB
and the extensions of the sides BC and CA at points M, N and P , respectively, and the
other excircle touches the side AC and the extensions of the sides AB and BC at points
S, Q and R, respectively. If X is the intersection point of the lines P N and RQ, and Y
the intersection point of RS and M N , prove that the points X, A and Y are collinear.
Delta 46. (Sharygin GMO 2008) Let ABC be a triangle and let the excircle opposite A
be tangent to the side BC at A1 . N is the Nagel point of ABC, and P is the point on
AA1 such that AP = N A1 . Prove that P lies on the incircle of ABC.
Delta 47. (United Kingdom 2007) Let ABC be a triangle with \B 6= \C. The incircle I of
ABC touches the sides BC, CA, AB at the points D, E, F , respectively. Let AD intersect
I at D and P . Let Q be the intersection of the lines EF and the line passing through
P and perpendicular to AD, and let X, Y be intersections of the line AQ and DE, DF ,
respectively. Show that the point A is the midpoint of XY .
Delta 48. (Sharygin GMO 2008) Given a triangle ABC. Point A1 is chosen on the ray
BA so that the segments BA1 and BC are equal. Point A2 is chosen on the ray CA so
that the segments CA2 and BC are equal. Points B1 , B2 and C1 , C2 are chosen similarly.
Prove that the lines A1 A2 , B1 B2 and C1 C2 are parallel.
4.2. Concurrencies around Cevas Theorem. In this section, we shall present some corollaries and applications of Cevas theorem.
Theorem 4.2. Let 4ABC be a given triangle and let A1 , B1 , C1 be three points on lying
on its sides BC, CA and AB, respectively. Then, the three lines AA1 ,BB1 , CC1 concur
if and only if
A0 B B 0 C C 0 A

= 1.
A0 C B 0 A C 0 B
Proof. We shall resume to proving only the direct implication. After reading the following
proof, you will understand why. Denote by P the intersection of the lines AA1 , BB1 , CC1 .
The parallel to BC through P meets CA at Ba and AB at Ca . The parallel to CA through
P meets AB at Cb and BC at Ab . The parallel to AB through P meets BC at Ac and
C1 A
Bc
= P
. On the other hand, we get
CA at Bc . As segments on parallels, we get C
P Ac
1B
Bc P
P B1
P Ac
P A1
=
and
=
.
AB
BB1
AB
AA1
It follows that

Bc P P A c
P B1 P A1
:
=
:
AB
AB
BB1 AA1

INFINITY

so that

51

Bc P
P B1 P A1
=
:
.
P Ac
BB1 AA1

Consequently, we obtain

C1 A
P B1 P A1
=
:
.
C1 B
BB1 AA1
P C1
P B1
A1
P C1
1B
1C
= CC
= P
Similarly, we deduce that A
: BB
and B
: CC
. Now
A1 C
B1 A
AA1
1
1
1

0
0
0
P C 1 P B1
P A1 P C1
P B1 P A1
AB BC CA

=
:

:
= 1,
A0 C B 0 A C 0 B
CC1 BB1
AA1 CC1
BB1 AA1
which proves Cevas theorem.

Corollary 4.1. (The Trigonometric Version of Cevas Theorem) In the configuration described above, the lines AA1 , BB1 , CC1 are concurrent if and only if
sin A1 AB sin C1 CA sin B1 BC

= 1.
sin A1 AC sin C1 CB sin B1 BA
Proof. By the Sine Law, applied in the triangles A1 AB and A1 AC, we have
AB
A1 C
AC
A1 B
=
, and
=
.
sin A1 AB
sin AA1 B
sin A1 AC
sin AA1 C
Hence,
A1 B
AB sin A1 AB
=

.
A1 C
AC sin A1 AC
B1 BC
C1 A
sin C1 CA
AC
1C
= BC
sin
and C
= BC
sin
. Thus, we conclude that
Similarly, B
B1 A
AB
sin B1 BA
C1 CB
1B

=
=

sin A1 AB sin C1 CA sin B1 BC

sin A1 AC sin C1 CB sin B1 BA

A1 B AC
C1 A BC
B1 C AB

A1 C AB
C1 B AC
B1 A BC
1.

We begin now with a result, which most of you might know it as Jacobis theorem.
Proposition 4.1. (Jacobis Theorem) Let ABC be a triangle, and let X, Y , Z be three
points in its plane such that \Y AC = \BAZ, \ZBA = \CBX and \XCB = \ACY .
Then, the lines AX, BY , CZ are concurrent.
Proof. We use directed angles taken modulo 180 . Denote by A, B, C, x, y, z the magnitudes of the angles \CAB, \ABC, \BCA, \Y AC, \ZBA, and \XCB, respectively.
Since the lines AX, BX, CX are (obviously) concurrent (at X), the trigonometric version
of Cevas theorem yields
sin CAX sin ABX sin BCX

= 1.
sin XAB sin XBC sin XCA
We now notice that
\ABX = \ABC + \CBX = B + y, \XBC = \CBX = y,
\BCX = \XCB = z, \XCA = \XCB + \BCA = z + C.
Hence, we get
sin (z)
sin CAX sin (B + y)

= 1.
sin XAB
sin (y)
sin (C + z)
Similarly, we can find
sin (x)
sin ABY sin (C + z)

= 1,
sin Y BC
sin (z)
sin (A + x)

52

INFINITY

sin (y)
sin BCZ sin (A + x)

= 1.
sin ZCA
sin (x)
sin (B + y)
Multiplying all these three equations and canceling the same terms, we get
sin CAX sin ABY sin BCZ

= 1.
sin XAB sin Y BC sin ZCA
According to the trigonometric version of Cevas theorem, the lines AX, BY , CZ are
concurrent.

We will see that Jacobis theorem has many interesting applications. We start with the
well-known Karyia theorem.
Theorem 4.3. (Kariyas Theorem) Let I be the incenter of a given triangle ABC, and let
D, E, F be the points where the incircle of ABC touches the sides BC, CA, AB. Now,
let X, Y , Z be three points on the lines ID, IE, IF such that the directed segments IX,
IY , IZ are congruent. Then, the lines AX, BY , CZ are concurrent.
Proof. (Darij Grinberg) Being the points of tangency of the incircle of triangle ABC with
the sides AB and BC, the points F and D are symmetric to each other with respect to
the angle bisector of the angle ABC, i. e. with respect to the line BI. Thus, the triangles
BF I and BDI are inversely congruent. Now, the points Z and X are corresponding points
in these two inversely congruent triangles, since they lie on the (corresponding) sides IF
and ID of these two triangles and satisfy IZ = IX. Corresponding points in inversely
congruent triangles form oppositely equal angles, i.e. \ZBF = \XBD. In other words,
\ZBA = \CBX. Similarly, we have that \XCB = \ACY and \Y AC = \BAZ. Note
that the points X, Y , Z satisfy the condition from Jacobis theorem, and therefore, we
conclude that the lines AX, BY , CZ are concurrent.

Another such corollary is the Kiepert theorem, which generalizes the existence of the
Fermat points.
Delta 49. (Kieperts Theorem) Let ABC be a triangle, and let BXC, CY A, AZB be
three directly similar isosceles triangles erected on its sides BC, CA, and AB, respectively.
Then, the lines AX, BY , CZ concur at one point.
Delta 50. (Floor van Lamoen) Let A0 , B 0 , C 0 be three points in the plane of a triangle
ABC such that \B 0 AC = \BAC 0 , \C 0 BA = \CBA0 and \A0 CB = \ACB 0 . Let X, Y ,
Z be the feet of the perpendiculars from the points A0 , B 0 , C 0 to the lines BC, CA, AB.
Then, the lines AX, BY , CZ are concurrent.
Delta 51. (Cosmin Pohoat
a) Let ABC be a given triangle in plane. From each of its
vertices we draw two arbitrary isogonals. Then, these six isogonals determine a hexagon
with concurrent diagonals.
Epsilon 51. (USA 2003) Let ABC be a triangle. A circle passing through A and B
intersects the segments AC and BC at D and E, respectively. Lines AB and DE intersect
at F , while lines BD and CF intersect at M . Prove that M F = M C if and only if
M B M D = M C2.
Delta 52. (Romanian jBMO 2007 Team Selection Test) Let ABC be a right triangle with
\A = 90 , and let D be a point lying on the side AC. Denote by E reflection of A into
the line BD, and by F the intersection point of CE with the perpendicular in D to the
line BC. Prove that AF , DE and BC are concurrent.
Delta 53. Denote by AA1 , BB1 , CC1 the altitudes of an acute triangle ABC, where A1 ,
B1 , C1 lie on the sides BC, CA, and AB, respectively. A circle passing through A1 and
B1 touches the arc AB of its circumcircle at C2 . The points A2 , B2 are defined similarly.
1. (Tuymaada Olympiad 2007) Prove that the lines AA2 , BB2 , CC2 are concurrent.
2. (Cosmin Pohoat
a, MathLinks Contest 2008, Round 1) Prove that the lines A1 A2 ,
B1 B2 , C1 C2 are concurrent on the Euler line of ABC.

INFINITY

53

4.3. Tossing onto Complex Plane. Here, we discuss some applications of complex numbers
to geometric inequality. Every complex number corresponds to a unique point in the
complex plane. The standard symbol for the set of all complex numbers is C, and we
also refer to the complex plane as C. We can identify the points in the real plane R2 as
numbers in C. The main tool is the following fundamental inequality.
Theorem 4.4. (Triangle Inequality) If z1 , , zn C, then |z1 | + + |zn | |z1 + + zn |.
Proof. Induction on n.

Theorem 4.5. (Ptolemys Inequality) For any points A, B, C, D in the plane, we have
AB CD + BC DA AC BD.
Proof. Let a, b, c and 0 be complex numbers that correspond to A, B, C, D in the complex
plane C. It then becomes
|a b| |c| + |b c| |a| |a c| |b|.
Applying the Triangle Inequality to the identity (a b)c + (b c)a = (a c)b, we get the
result.

Remark 4.1. Investigate the equality case in Ptolemys Inequality.

Delta 54. [SL 1997 RUS] Let ABCDEF be a convex hexagon such that AB = BC,
CD = DE, EF = F A. Prove that
BC
DE
FA
3
+
+
.
BE
DA
FC
2
When does the equality occur?
Epsilon 52. [TD] Let P be an arbitrary point in the plane of a triangle ABC with the
centroid G. Show the following inequalities
(1) BC P B P C + AB P A P B + CA P C P A BC CA AB,
(2) P A3 BC + P B 3 CA + P C 3 AB 3P G BC CA AB.
Delta 55. Let H denote the orthocenter of an acute triangle ABC. Prove the geometric
identity
BC HB HC + AB HA HB + CA HC HA = BC CA AB.
Epsilon 53. (The Neuberg-Pedoe Inequality) Let a1 , b1 , c1 denote the sides of the triangle
A1 B1 C1 with area F1 . Let a2 , b2 , c2 denote the sides of the triangle A2 B2 C2 with area F2 .
Then, we have
a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 ) 16F1 F2 .
Epsilon 54. [SL 2002 KOR] Let ABC be a triangle for which there exists an interior point
F such that \AF B = \BF C = \CF A. Let the lines BF and CF meet the sides AC
and AB at D and E, respectively. Prove that AB + AC 4DE.

54

INFINITY

4.4. Generalize Ptolemys Theorem! The story begins with three trigonometric proofs of
Ptolemys Theorem.
Theorem 4.6. (Ptolemys Theorem) Let ABCD be a convex quadrilateral. If ABCD is
cyclic, then we have
AB CD + BC DA = AC BD.
First Proof. Set AB = a, BC = b, CD = c, DA = d. One natural approach is to compute
BD = x and AC = y in terms of a, b, c and d. We apply the Cosine Law to obtain
x2 = a2 + d2 2ad cos A
and
x2 = b2 + c2 2bc cos D = b2 + c2 + 2bc cos A.
Equating two equations, we meet
a2 + d2 2ad cos A = b2 + c2 + 2bc cos A
or
cos A =

a2 + d2 b2 c2
.
2(ad + bc)

It follows that

x2 = a2 + d2 2ad cos A = a2 + d2 2ad

a2 + d2 b2 c2
2(ad + bc)

(ac + bd)(ab + cd)

.
ad + bc

Similarly, we also obtain

(ac + bd)(ad + bc)
.
ab + cd
2 2
2
Multiplying these two, we obtain x y = (ac + bd) or xy = ac + bd, as desired.
y2 =

Second Proof. (Hojoo Lee) As in the classical proof via the inversion, we rewrite it as
AB
BC
AC
+
=
.
DA DB
DB DC
DA DC
We now trigonometrize each term. Letting R denote the circumradius of ABCD and
noticing that sin(\ADB) = sin (\DBA + \DAB) in triangle DAB, we obtain
AB
DA DB

=
=
=

Likewise, we have

2R sin(\ADB)
2R sin(\DBA) 2R sin(\DAB)
sin \DBA cos \DAB + cos \DBA sin \DAB
2R sin(\DBA) sin(\DAB)
1
(cot \DAB + cot \DBA) .
2R

BC
1
=
(cot \DBC + cot \DCB)
DB DC
2R

and

AC
1
=
(cot \DAC + cot \DCA) .
DA DC
2R
Hence, the geometric identity in Ptolemys Theorem is equivalent to the cotangent identity
(cot \DAB + cot \DBA) + (cot \DBC + cot \DCB) = (cot \DAC + cot \DCA) .
However, since the convex quadrilateral ABCD admits a circumcircle, it is clear that
\DAB + \DCB = , \DBA = \DCA, \DBC = \DAC
so that
cot \DAB + cot \DCB = 0, cot \DBA = cot \DCA, cot \DBC = cot \DAC.

INFINITY

55

Third Proof. We exploit the Sine Law to convert the geometric identity to the trigonometric identity. Let R denote the circumradius of ABCD. We set
\AOB = 21 , \BOC = 22 , \COD = 23 , \DOA = 24 ,
where O is the center of the circumcircle of ABCD. Its clear that 1 + 2 + 3 + 4 = .
It follows that AB = 2R sin 1 , BC = 2R sin 2 , CD = 2R sin 3 , DA = 2R sin 4 ,
AC = 2R sin (1 + 2 ), AB = 2R sin (2 + 3 ). Our job is to establish
AB CD + BC DA = AC BD
or
(2R sin 1 ) (2R sin 3 ) + (2R sin 2 ) (2R sin 4 ) = (2R sin (1 + 2 )) (2R sin (2 + 3 ))
or equivalently
sin 1 sin 3 + sin 2 sin 4 = sin (1 + 2 ) sin (2 + 3 ) .
We use the well-known identity sin sin =

1
2

[ cos( ) cos( + ) ] to rewrite it as

cos(1 3 ) cos(1 + 3 )
cos(2 4 ) cos(2 + 4 )
+
2
2
cos(1 3 ) cos(1 + 22 + 3 )
2

=
or equivalently

cos(1 + 3 ) + cos(2 4 ) cos(2 + 4 ) = cos(1 + 22 + 3 ).

Since 1 + 2 + 3 + 4 = or since cos(1 + 3 ) + cos(2 + 4 ) = 0, it is equivalent to
cos(2 4 ) = cos(1 + 22 + 3 ).
However, we employ 1 + 2 + 3 = 4 to deduce
cos(1 + 22 + 3 ) = cos(2 + 4 ) = cos(2 4 ).

When the second author of this weblication was a high school student, one day, he
was trying to device a coordinate proof of Ptolemys Theorem. However, we immediately
realize that the direct approach using only the distance formula is hopeless. The geometric
identity reads, in coordinates,
q

(x1 x2 )2 + (y1 y2 )2 (x3 x4 )2 + (y3 y4 )2

q

+
(x2 x3 )2 + (y2 y3 )2 (x4 x1 )2 + (y4 y1 )2
q

=
(x1 x3 )2 + (y1 y3 )2 (x2 x4 )2 + (y2 y4 )2 .
What he realized was that the key point is to find a natural coordinate condition. First,
forget about the destination AB CD + BC DA = AC BD and, instead, find out what
the initial condition that ABCD is cyclic says in coordinates.
Lemma 4.1. Let ABCD be a convex quadrilateral. We toss ABCD on the real plane R2
with the coordinates A (a1 , a2 ) , B (b1 , b2 ) , C (c1 , c2 ) , D (d1 , d2 ). Then, the necessary and
sufficient condition that ABCD is cyclic is that the following equality holds.

a1 2 + a2 2 (a1 b1 + a2 b2 + a1 c1 + a2 c2 b1 c1 b2 c2 )
b1 a2 + a1 c2 + c1 b2 a1 b2 c1 a2 b1 c2
d1 2 + d2 2 (d1 b1 + d2 b2 + d1 a1 + d2 a2 b1 a1 b2 a2 )
b1 d2 + d1 a2 + a1 b2 d1 b2 a1 d2 b1 a2

56

INFINITY

Proof. The quadrilateral ABCD is cyclic if and only if \BAC = \BDC, or equivalently
cot (\BAC) = cot (\BDC). It is equivalent to
cos (\BAC)
cos (\BDC)
=
sin (\BAC)
sin (\BDC)
or

BA2 +AC 2 CB 2
2BAAC
2[ABC]
BAAC

or

BD 2 +DC 2 CB 2
2BDDC
2[DBC]
BDDC

BA2 + AC 2 CB 2
BD 2 + DC 2 CB 2
=
2[ABC]
2[DBC]

or in coordinates,

a1 2 + a2 2 (a1 b1 + a2 b2 + a1 c1 + a2 c2 b1 c1 b2 c2 )
b1 a2 + a1 c2 + c1 b2 a1 b2 c1 a2 b1 c2
d1 2 + d2 2 (d1 b1 + d2 b2 + d1 a1 + d2 a2 b1 a1 b2 a2 )
.
b1 d2 + d1 a2 + a1 b2 d1 b2 a1 d2 b1 a2

The coordinate condition and its proof is natural. However, something weird happens
here. It does not look like being cyclic in the coordinates . Indeed, when ABCD is a
cyclic quadrilateral, we notice that the same quadrilateral BCDA, CDAB, DABC are
also trivially cyclic. It turns out that the coordinate condition indeed admits a certain
symmetry. Now, it is time to consider the destination
AB CD + BC DA = AC BD.
As we see above, the direct application of the distance formula gives us a monster identity
with square roots. What we need is a reformulation without square roots. We recall
that Ptolemys Theorem is trivialized by the inversive geometry. As in the proof via the
inversion, we rewrite it in the symmetric form
AB
BC
AC
+
=
.
DA DB
DB DC
DA DC
Now, we reach the key step. Let R denote the circumcircle of ABCD. The formulas
AB DA DB
BC DB DC
CA DC DA
[DAB] =
, [DBC] =
, [DCA] =
4R
4R
4R
allows us to realize that it is equivalent to the geometric identity.
[DAB]
[DBC]
[DAC]
+
=
DA2 DB 2
DB 2 DC 2
DA2 DC 2
or
DC 2 [DAB] + DA2 [DBC] = DB 2 [DAC].
Summarizing up the result, we have
Lemma 4.2. Let ABCD be a convex and cyclic quadrilateral. Then the following two
geometric identities are equivalent.
(1)

AB CD + BC DA = AC BD.

(2)

DC 2 [DAB] + DA2 [DBC] = DB 2 [DAC].

It is awesome. Why? It is because we can express the second condition in the coordinates without the horrible square root! After a long, very long computation by hand, we
can check that
a1 2 + a2 2 (a1 b1 + a2 b2 + a1 c1 + a2 c2 b1 c1 b2 c2 )
b1 a2 + a1 c2 + c1 b2 a1 b2 c1 a2 b1 c2
d1 2 + d2 2 (d1 b1 + d2 b2 + d1 a1 + d2 a2 b1 a1 b2 a2 )
=
.
b1 d2 + d1 a2 + a1 b2 d1 b2 a1 d2 b1 a2

INFINITY

57

indeed implies the coordinate condition of the reformulation (2). The brute-force computation will be simplified if we take D (d1 , d2 ) = (0, 0). However, it is not the end of
the story. Actually, he found a symmetry in the coordinate computation. It leads him to
rediscover The Feuerbach-Luchterhand Theorem, which generalize Ptolemys Theorem.
Theorem 4.7. (The Feuerbach-Luchterhand Theorem) Let ABCD be a convex and cyclic
quadrilateral. For any point O in the plane, we have
OA2 BC CD DB OB 2 CD DA AB + OC 2 DA AB BD OD2 AB BC CD = 0.
Proof. We toss the picture on the real plane R2 with the coordinates
O(0, 0), A (a1 , a2 ) , B (b1 , b2 ) , C (c1 , c2 ) , D (d1 , d2 ) ,
Letting R denote the circumcircle of ABCD, it can be rewritten as
OA2

[BCD]
[CDA]
[DAB]
[DBC]
OB 2
+ OC 2
OD2
=0
4R
4R
4R
4R

or
OA2 [BCD] OB 2 [CDA] + OC 2 [DAB] OD2 [ABC] = 0.
We can rewrite this in the coordinates without square roots. Now, after long computation,
we can check, by hand, that it is equivalent to the coordinate condition that ABCD is
cyclic:

a1 2 + a2 2 (a1 b1 + a2 b2 + a1 c1 + a2 c2 b1 c1 b2 c2 )
b1 a2 + a1 c2 + c1 b2 a1 b2 c1 a2 b1 c2
d1 2 + d2 2 (d1 b1 + d2 b2 + d1 a1 + d2 a2 b1 a1 b2 a2 )
.
b1 d2 + d1 a2 + a1 b2 d1 b2 a1 d2 b1 a2

The end? Not yet. It turns out that after throwing out the essential condition that the
quadrilateral is cyclic, we can extend the theorem to arbitrary quadrilaterals!
Theorem 4.8. (Hojoo Lee) For an arbitrary point P in the plane of the convex quadrilateral
A1 A2 A3 A4 , we obtain
=

P A1 2 [4A2 A3 A4 ] P A2 2 [4A3 A4 A1 ] + P A3 2 [4A4 A1 A2 ] P A4 2 [4A1 A2 A3 ]

A1 A2 A1 A3 [4A2 A3 A4 ] A4 A2 A4 A3 [4A1 A2 A3 ].

After removing the convexity of A1 A2 A3 A4 , we get the same result regarding the signed
area of triangle.
Outline of Proof. We toss the figure on the real plane R2 and write P (0, 0) and Ai =
(xi , yi ), where 1 i 4. Our task is to check that two matrices
0
1
P A1 2 P A 2 2 P A 3 2 P A 4 2
B x1
x2
x3
x4 C
C
L=B
@ y1
y2
y3
y4 A
1
1
1
1
and

A1 A2 A1 A3
B
x1
R=B
@
y1
1

0
x2
y2
1

0
x3
y3
1

A4 A2 A4 A3
x4
y4
1

1
C
C
A

have the same determinant. We now invite the readers to find a neat proof, of course
without the brute force expansion of the determinants!

58

INFINITY

Is it a re-discovery again? Is this the end of the story? No. There is no end of
generalizations in Mathematics. The lesson we want to deliver here is simple: Even the
brute-force coordinate proofs offer good motivations. There is no bad proof. We now
present some applications of the The Feuerbach-Luchterhand Theorem.
Corollary 4.2. Let ABCD be a rectangle. For any point P , we have
P A2 P B 2 + P C 2 P D2 = 0.
Now, lets see what happens if we apply The Feuerbach-Luchterhand Theorem to a
geometric situation from the triangle geometry. Let ABC be a triangle with the incenter
I and the circumcenter O. Let BC = a, CA = b, AB = c, s = a+b+c
. Let R and r denote
2
the circumradius and inradius, respectively. Let P and Q denote the feet of perpendiculars
from I to the sides CA and CB, respectively. Since \IP C = 90 = \IQC, we find that
IQCP is cyclic.
We then apply The Feuerbach-Luchterhand Theorem to the pair (O, IQCP ) to deduce
the geometric identity
0 = OI 2 QC CP P Q OQ2 CP P I IC + OC 2 P I IQ QP OP 2 IQ QC CI.
What does it mean? We observe that, in the isosceles triangles COA and BOC,
OP 2 = R2 AP P C = R2 (s a)(s c),
OQ2 = R2 BQ QC = R2 (s b)(s c).
Now, it becomes
0

=
+

c 2
OI 2 (s c)2 CI
R (s b)(s c) (s c)r IC
2R

c 2
R2 r2 CI
R (s a)(s c) r(s c) CI
2R

or
0

=
+

c
R2 (s b)(s c) (s c)r
2R

c
R2 r 2
R2 (s a)(s c) r(s c).
2R
OI 2 (s c)2

or

c
Rr2 c 2
=
+ 2R c(s c) (s c)r
2R
2
or

OI 2
Rr2
4R2 r
4Rr(s c) r2 c
=
+

2r
=
R
2r.
R
(s c)2
c(s c)
(s c)2 c
Now, we apply Ptolemys Theorem and The Pythagoras Theorem to deduce

c
c
2r(s c) = CP IQ + P I IQ = P Q CI = CI 2
= (s c)2 + r2
2R
2R
or
`

4Rr(s c) = c (s c)2 + r2
OI 2 (s c)2

or
4Rr(s c) r2 c = (s c)2 c
or

4Rr(s c) r2 c
= 1.
(s c)2 c

It therefore follows that

OI 2 = R2 2rR.
It is the theorem proved first by L. Euler. There are lots of way to establish this. Device
your own proofs! Find other corollaries of The Feuerbach-Luchterhand Theorem. Another
possible generalization of Ptolemys relation is Caseys theorem:

INFINITY

59

Theorem 4.9. (Caseys Theorem) Given four circles Ci , i = 1, 2, 3, 4, let tij be the length
of a common tangent between Ci and Cj . The four circles are tangent to a fifth circle (or
line) if and only if for an appropriate choice of signs, we have that
t12 t34 t13 t42 t14 t23 = 0.
The most common proof for this result is by making use of inversion. See [RJ]. We
shall omit it here. We now work on the Feuerbachs celebrated theorem (actually its first
version).
Theorem 4.10. (Feuerbachs Theorem) The incircle and nine-point circle of a triangle are
tangent to one another.
Why first version? Of course, most of you might know that the nine-point circle is also
tangent to the three excircles of the triangle. Most of the geometry textbooks include this
last remark in the theorems statement as well, but this is mostly for sake of completeness,
since the proof is similar with the incenter case.
Proof. Let the sides BC, CA, AB of triangle ABC have midpoints D, E, F respectively,
and let be the incircle of the triangle. Let a, b, c be the sidelengths of ABC, and let s
be its semiperimeter. We now consider the 4-tuple of circles (D, E, F, ). Here is what
we find:
b
a
c
tDE = , tDF = , tEF = ,
2
2
2

a
b

,
tD = (s b) =
2
2

b
a c

tE = (s c) =
,
2
2

b a
c

tF = (s a) =
.
2
2
We need to check whether, for some combination of +, signs, we have
(c(b a) a(b c) b(a c) = 0.
But this is immediate! According to Caseys theorem there exists a circle what touches
each of D, E, F and . Since the circle passing through D, E, F is the ninepoint circle
of the triangle, it follows that and the nine-point circle are tangent to each other.

We shall see now an interesting particular case of Thebaults theorem.

Proposition 4.2. (IMO Longlist 1991, proposed by India) Circles 1 and 2 are externally
tangent at a point I, and both are enclosed by and tangent to a third circle . One
common tangent to 1 and 2 meets in B and C, while the common tangent at I meets
in A on the same side of BC as I. Then, we have that I is the incenter of triangle ABC.
Proof. Let X, Y be the tangency points of BC with the circles 1 , and 2 , respectively,
and let x, y be the lengths of the tangents from B and C to 1 and 2 . Denote by D
the intersection of AI with the line BC, and let z = AI, u = ID. According to Caseys
theorem, applied for the two 4-tuples of circles (A, 1 , B, C) and (A, 2 , C, B), we
obtain az + bx = c(2u + y) and az + cy = b(2u + x). Subtracting the second equation from
= cb , that is, BD
= AB
,
the first, we obtain that bx cy = u(c b), and therefore x+u
y+u
DC
AC
ac
. Adding the two equations
which implies that AI bisects \BAC, and that BD = b+c
AI
AB
mentioned before, we finally obtain that az = u(b + c), which rewrites as ID
= BD
. This
implies that BI bisects \ABC. Thus, I is the incenter of triangle ABC.

But can we prove Thebaults theorem using Casey? S. Gueron [SG] says yes!

60

INFINITY

Delta 56. (Thebault [VT]) Through the vertex A of a triangle ABC, a straight line AD
is drawn, cutting the side BC at D. I is the incenter of triangle ABC, and let P be the
center of the circle which touches DC, DA, and (internally) the circumcircle of ABC, and
let Q be the center of the circle which touches DB, DA, and (internally) the circumcircle
of ABC. Then, the points P , I, Q are collinear.
Delta 57. (Jean-Pierre Ehrmann and Cosmin Pohoat
a, MathLinks Contest 2008) Let P be
an arbitrary point on the side BC of a given triangle ABC with circumcircle . Let TAb
be the circle tangent to AP , P B, and internally to , and let TAc be the circle tangent to
AP , P C, and internally to . Then, the circles TAb and TAc are congruent if and only if
AP passes through the Nagel point of triangle ABC.
Delta 58. (Lev Emelyanov [LE]) Let P be a point in the interior of a given triangle ABC.
Denote by A1 , B1 , C1 the intersections of AP , BP , CP with the sidelines BC, CA, and
AB, respectively (in other words, the triangle A1 B1 C1 is the cevian triangle of P with
respect to ABC). Construct the three circles (O1 ), (O2 ) and (O3 ) outside the triangle
which are tangent to the sides of ABC at A1 , B1 , C1 , and also tangent to the circumcircle
of ABC. Then, the circle tangent externally to these three circles is also tangent to teh
incircle of triangle ABC.

It is impossible to be a mathematician without being a poet in soul. - S. Kovalevskaya

INFINITY

61

5. Three Terrific Techniques (EAT)

A long time ago an older and well-known number theorist made
some disparaging remarks about Paul Erd
oss work. You admire
Erd
oss contributions to mathematics as much as I do, and I
felt annoyed when the older mathematician flatly and definitively
stated that all of Erd
oss work could be reduced to a few tricks
which Erd
os repeatedly relied on in his proofs. What the number theorist did not realize is that other mathematicians, even
the very best, also rely on a few tricks which they use over and
over. Take Hilbert. The second volume of Hilberts collected
papers contains Hilberts papers in invariant theory. I have made
a point of reading some of these papers with care. It is sad
to note that some of Hilbert0 s beautiful results have been completely forgotten. But on reading the proofs of Hilberts striking
and deep theorems in invariant theory, it was surprising to verify
that Hilberts proofs relied on the same few tricks. Even Hilbert
had only a few tricks!
- G-C Rota, Ten Lessons I Wish I Had Been Taught

5.1. Trigonometric Substitutions. If you are faced with an integral that contains square
root expressions such as
Z p
Z p
Z p
1 x2 dx,
1 + y 2 dy,
z 2 1 dz
then trigonometric substitutions such as x = sin t, y = tan t, z = sec t are very useful. We
will learn that making a suitable trigonometric substitution simplifies the given inequality.
Epsilon 55. (APMO 2004/5) Prove that, for all positive real numbers a, b, c,
(a2 + 2)(b2 + 2)(c2 + 2) 9(ab + bc + ca).
Epsilon 56. (Latvia 2002) Let a, b, c, d be the positive real numbers such that
1
1
1
1
+
+
+
= 1.
1 + a4
1 + b4
1 + c4
1 + d4
Prove that abcd 3.
Epsilon 57. (Korea 1998) Let x, y, z be the positive reals with x + y + z = xyz. Show
that
1
1
1
3

+p
+
.
2
1 + x2
1 + z2
1 + y2
is not concave on R+ , we cannot apply Jensens
`

Inequality directly. However, the function f (tan ) is concave on 0, 2 !

Since the function f (t) = 1

1+t2

Proposition 5.1. In any acute triangle ABC, we have cos A + cos B + cos C 32 .

`
Proof. Since cos x is concave on 0, 2 , its a direct consequence of Jensens Inequality.
`

We note that the function cos x is not concave on (0, ). In fact, its convex on 2 , .
3
One may think that the inequality cos A+cos B +cos C 2 doesnt hold for any triangles.
However, its known that it holds for all triangles.
Proposition 5.2. In any triangle ABC, we have
cos A + cos B + cos C

3
.
2

62

INFINITY

First Proof. It follows from C = A + B that

cos C = cos(A + B) = cos A cos B + sin A sin B
or

3 2(cos A + cos B + cos C) = (sin A sin B)2 + (cos A + cos B 1)2 0.

Second Proof. Let BC = a, CA = b, AB = c. Use The Cosine Law to rewrite the given
inequality in the terms of a, b, c :
b 2 + c 2 a2
c 2 + a2 b 2
a2 + b 2 c 2
3
+
+
.
2bc
2ca
2ab
2
Clearing denominators, this becomes
3abc a(b2 + c2 a2 ) + b(c2 + a2 b2 ) + c(a2 + b2 c2 ),
which is equivalent to abc (b + c a)(c + a b)(a + b c).

We remind that the geometric inequality R 2r is equivalent to the algebraic inequality

abc (b + c a)(c + a b)(a + b c). We now find that, in the proof of the above theorem,
abc (b + c a)(c + a b)(a + b c) is equivalent to the trigonometric inequality
cos A + cos B + cos C 32 . One may ask that
in any triangles ABC, is there a natural relation between cos A + cos B +
cos C and R
, where R and r are the radii of the circumcircle and incircle
r
of ABC?
Theorem 5.1. Let R and r denote the radii of the circumcircle and incircle of the triangle
ABC. Then, we have
r
cos A + cos B + cos C = 1 + .
R
Proof. Use the algebraic identity
a(b2 + c2 a2 ) + b(c2 + a2 b2 ) + c(a2 + b2 c2 ) = 2abc + (b + c a)(c + a b)(a + b c).
We leave the details for the readers.

Delta 59. (China 2004) Let ABC be a triangle with BC = a, CA = b, AB = c. Prove

that, for all x 0,
1
ax cos A + bx cos B + cx cos C (ax + bx + cx ) .
2
Delta 60. (a) Let p, q, r be the positive real numbers such that p2 + q 2 + r2 + 2pqr = 1.
Show that there exists an acute triangle ABC such that p = cos A, q = cos B, r = cos
C.
(b) Let p, q, r 0 with p2 + q 2 + r2 + 2pqr = 1. Show that there are A, B, C 0, 2 with
p = cos A, q = cos B, r = cos C, and A + B + C = .
Epsilon 58. (USA 2001) Let a, b, and c be nonnegative real numbers such that a2 + b2 +
c2 + abc = 4. Prove that 0 ab + bc + ca abc 2.

Life is good for only two things, discovering mathematics and teaching mathematics.
- S. Poisson

INFINITY

63

5.2. Algebraic Substitutions. We know that some inequalities in triangle geometry can
be treated by the Ravi substitution and trigonometric substitutions. We can also transform
the given inequalities into easier ones through some clever algebraic substitutions.
Epsilon 59. [IMO 2001/2 KOR] Let a, b, c be positive real numbers. Prove that

a
b
c
+
+
1.
a2 + 8bc
b2 + 8ca
c2 + 8ab

Epsilon 60. [IMO 1995/2 RUS] Let a, b, c be positive numbers such that abc = 1. Prove
that
1
1
1
3
+ 3
+ 3
.
a3 (b + c)
b (c + a)
c (a + b)
2
Epsilon 61. (Korea 1998) Let x, y, z be the positive reals with x + y + z = xyz. Show
that
1
1
1
3

+p
+
.
2
2
2
1 + x2
1
+
z
1+y
We now prove a classical theorem in various ways.
Proposition 5.3. (Nesbitt) For all positive real numbers a, b, c, we have
b
c
3
a
+
+
.
b+c
c+a
a+b
2
Proof 4. After the substitution x = b + c, y = c + a, z = a + b, it becomes
X y+zx
X y+z
3

or
6,
2x
2
x
cyclic

cyclic

which follows from The AM-GM Inequality as following:

1
X y+z
y
z
z
x
x
y
y z z x x y 6
= + + + + + 6

= 6.
x
x
x
y
y
z
z
x x y y z z
cyclic

Proof 5. We make the substitution

a
b
c
x=
, y=
, z=
.
b+c
c+a
a+b
It follows that
X
X
a
f (x) =
= 1,
a+b+c
cyclic

where f (t) =

t
.
1+t

cyclic

Since f is concave on (0, ), Jensens Inequality shows that

x + y + z
1
1
1 X
f (x) f
f
= =
2
3
3
3
cyclic

Since f is monotone increasing, it implies that

1
x+y+z

2
3
or
X a
3
=x+y+z .
b+c
2
cyclic

Proof 6. As in the previous proof, it suffices to show that

x+y+z
1
T :=
,
3
2
where we have
X x
=1
1+x
cyclic

64

INFINITY

or equivalently,
1 = 2xyz + xy + yz + zx.
We apply The AM-GM Inequality to deduce
1 = 2xyz + xy + yz + zx 2T 3 + 3T 2
It follows that
so that

2T 3 + 3T 2 1 0
(2T 1)(T + 1)2 0

or

1
.
2
Epsilon 62. [IMO 2000/2 USA] Let a, b, c be positive numbers such that abc = 1. Prove
that

1
1
1
a1+
b1+
c1+
1.
b
c
a
T

Epsilon 63. Let a, b, c be positive real numbers satisfying a + b + c = 1. Show that

b
3 3
a
abc
+
+
1+
.
a + bc
b + ca
c + ab
4
Epsilon 64. (Latvia 2002) Let a, b, c, d be the positive real numbers such that
1
1
1
1
+
+
+
= 1.
1 + a4
1 + b4
1 + c4
1 + d4
Prove that abcd 3.
Delta 61. [SL 1993 USA] Prove that
a
b
c
d
2
+
+
+

b + 2c + 3d
c + 2d + 3a
d + 2a + 3b
a + 2b + 3c
3
for all positive real numbers a, b, c, d.
Epsilon 65. [LL 1992 UNK] (Iran 1998) Prove that, for all x, y, z > 1 such that
2,
p

x + y + z x 1 + y 1 + z 1.

1
+ y1
x

+ z1 =

Epsilon 66. (Belarus 1998) Prove that, for all a, b, c > 0,

a
b
c
a+b
b+c
+ +
+
+ 1.
b
c
a
b+c
c+a
Delta 62. [IMO 1969 USS] Under the conditions x1 , x2 > 0, x1 y1 > z1 2 , and x2 y2 > z2 2 ,
prove the inequality
8
1
1
+
.
2
2
x
y

z
x
y
z2 2
1
1
1
2
2
(x1 + x2 ) (y1 + y2 ) (z1 + z2 )
Epsilon 67. [SL 2001 ] Let x1 , , xn be arbitrary real numbers. Prove the inequality.

x1
x2
xn
+
+ +
< n.
1 + x1 2
1 + x1 2 + x2 2
1 + x1 2 + + xn 2
Delta 63. [LL 1987 FRA] Given n real numbers 0 t1 t2 tn < 1, prove that

t2 2
tn n
t1
+
+

+
1 tn 2
< 1.
2
2
2
(1 t1 2 )
(1 t2 3 )
(1 tn n+1 )

INFINITY

65

5.3. Establishing New Bounds. The following examples give a nice description of the
title of this subsection.
Example 10. Let x, y, z be positive real numbers. Show the cyclic inequality
x2
y2
z2
x
y
z
+ 2 + 2 + + .
2
y
z
x
y
z
x
Second Solution. We first use the auxiliary inequality t2 2t 1 to deduce
x2
y
z
y2
z2
x
+
+
2 1 + 2 1 + 2 1.
y2
z2
x2
y
z
x
It now remains to check that
x
y
z
x
y
z
2 1+2 1+2 1 + +
y
z
x
y
z
x
or equivalently
y
z
x
+ + 3.
y
z
x
However, The AM-GM Inequality shows that

1
x
y
z
x y z 3
+ + 3

= 3.
y
z
x
y z x

Proposition 5.4. (Nesbitt) For all positive real numbers a, b, c, we have

a
b
c
3
+
+
.
b+c
c+a
a+b
2

Proof 7. From

a
b+c

1
2

0, we deduce that

a
1

b+c
4
It follows that

X
cyclic

8a
b+c
a
b+c

1
+1

8a b c
.
4(a + b + c)

X 8a b c
a
3

= .
b+c
4(a + b + c)
2
cyclic

Proof 8. We claim that

3

3
3
1
a
3a 2
2 + b2 + c2

3
or
2
a
3a 2 (b + c).
3
3
b+c
2 a2 + b2 + c2
3

The AM-GM inequality gives a2 + b 2 + b 2 3a 2 b and a 2 + c 2 + c 2 3a 2 c . Adding

3
1
3
3
these two inequalities yields 2 a 2 + b 2 + c 2 3a 2 (b + c), as desired. Therefore, we
have
3
X a
3 X
a2
3

.
3
3
3 =
b+c
2
2
a2 + b2 + c2
cyclic

cyclic

Epsilon 68. Let a, b, c be the lengths of a triangle. Show that

b
c
a
+
+
< 2.
b+c
c+a
a+b

66

INFINITY

Some cyclic inequalities can be established by finding some clever bounds. Suppose
that we want to establish that
X
F (x, y, z) C
cyclic

for some given constant C R. Whenever we have a function G such that, for all
x, y, z > 0,
F (x, y, z) G(x, y, z)
and

G(x, y, z) = C,

cyclic

we then deduce that

F (x, y, z)

cyclic

G(x, y, z) = C.

cyclic

For instance, if a function F satisfies the inequality

x
F (x, y, z)
x+y+z
for all x, y, z > 0, then F obeys the inequality
X
F (x, y, z) 1.
cyclic

Epsilon 69. [IMO 2001/2 KOR] Let a, b, c be positive real numbers. Prove that

b
c
a
+
+
1.
a2 + 8bc
b2 + 8ca
c2 + 8ab

Epsilon 70. [IMO 2005/3 KOR] Let x, y, and z be positive numbers such that xyz 1.
Prove that
x5 x2
y5 y2
z5 z2
+ 5
+ 5
0.
5
2
2
2
2
x +y +z
y +z +x
z + x2 + y 2
Epsilon 71. (KMO Weekend Program 2007) Prove that, for all a, b, c, x, y, z > 0,
(a + b + c)(x + y + z)
ax
by
cz
+
+

.
a+x
b+y
c+z
a+b+c+x+y+z
Epsilon 72. (USAMO Summer Program 2002) Let a, b, c be positive real numbers. Prove
that

2
2
2
3
3
3
2a
2b
2c
+
+
3.
b+c
c+a
a+b
Epsilon 73. (APMO 2005) Let a, b, c be positive real numbers with abc = 8. Prove that
p

a2
b2
c2
4
+p
+p

3
3
3
3
3
(1 + a )(1 + b )
(1 + b )(1 + c )
(1 + c3 )(1 + a3 )

Delta 64. [SL 1996 SVN] Let a, b, and c be positive real numbers such that abc = 1. Prove
that
bc
ca
ab
+ 5
+ 5
1.
a5 + b5 + ab
b + c5 + bc
c + a5 + ca
Delta 65. [SL 1971 YUG] Prove the inequality
a2 + a4
a3 + a1
a4 + a2
a1 + a3
+
+
+
4
a1 + a2
a2 + a3
a3 + a4
a4 + a1
where a1 , a2 , a3 , a4 > 0.
There is a simple way to find new bounds for given differentiable functions. We begin
to show that every supporting lines are tangent lines in the following sense.

INFINITY

67

Proposition 5.5. (The Characterization of Supporting Lines) Let f be a real valued function.
Let m, n R. Suppose that
(1) f () = m + n for some R,
(2) f (x) mx + n for all x in some interval (1 , 2 ) including , and
(3) f is differentiable at .
Then, the supporting line y = mx + n of f is the tangent line of f at x = .
Proof. Let us define a function F : (1 , 2 ) R by F (x) = f (x) mx n for all
x (1 , 2 ). Then, F is differentiable at and we obtain F 0 () = f 0 () m. By the
assumption (1) and (2), we see that F has a local minimum at . So, the first derivative
theorem for local extreme values implies that 0 = F 0 () = f 0 ()m so that m = f 0 () and
that n = f ()m = f ()f 0 (). It follows that y = mx+n = f 0 ()(x)+f ().
Proposition 5.6. (Nesbitt) For all positive real numbers a, b, c, we have
a
b
c
3
+
+
.
b+c
c+a
a+b
2
Proof 9. We may normalize to a + b + c = 1. Note that 0 < a, b, c < 1. The problem is
now to prove
X
3
f (a)
2
cyclic

or
f (a) + f (b) + f (c)
f
3

1
,
3

x
where where f (x) = 1x
. The equation of the tangent line of f at x =
9x1
y = 4 . We claim that the inequality

f (x)

1
3

is given by

9x 1
4

holds for all x (0, 1). However, it immediately follows from the equality
f (x)
Now, we conclude that
X
cyclic

(3x 1)2
9x 1
=
.
4
4(1 x)

X 9a 1
a
9 X
3
3
a = .

=
1a
4
4
4
2
cyclic

cyclic

The above argument can be generalized. If a function f has a supporting line at some
point on the graph of f , then f satisfies Jensens Inequality in the following sense.
Theorem 5.2. (Supporting Line Inequality) Let f : [a, b] R be a function. Suppose that
[a, b] and m R satisfy
f (x) m(x ) + f ()
for all x [a, b]. Let 1 , , n > 0 with 1 + + n = 1. Then, the following inequality
holds
1 f (x1 ) + + n f (xn ) f ()
for all x1 , , xn [a, b] such that = 1 x1 + + n xn . In particular, we obtain
s
f (x1 ) + + f (xn )
f
,
n
n
where x1 , , xn [a, b] with x1 + + xn = s for some s [na, nb].

68

INFINITY

Proof.
1 f (x1 ) + + n f (xn )

1 [m(x1 ) + f ()] + + 1 [m(xn ) + f ()]

f ().

We can apply the supporting line inequality to deduce Jensens inequality for differentiable functions.
Lemma 5.1. Let f : (a, b) R be a convex function which is differentiable twice on (a, b).
Let y = l (x) be the tangent line at (a, b). Then, f (x) l (x) for all x (a, b). So,
the convex function f admits the supporting lines.
Proof. Let (a, b). We want to show that the tangent line y = l (x) = f 0 ()(x
) + f () is the supporting line of f at x = such that f (x) l (x) for all x (a, b).
However, by Taylors Theorem, we can find a real number x between and x such that
f (x) = f () + f 0 ()(x ) +

f 00 (x )
(x )2 f () + f 0 ()(x ).
2

Theorem 5.3. (The Weighted Jensens Inequality) Let f : [a, b] R be a continuous

convex function which is differentiable twice on (a, b). Let 1 , , n > 0 with 1 + +
n = 1. For all x1 , , xn [a, b],
1 f (x1 ) + + n f (xn ) f (1 x1 + + n xn ).
First Proof. By the continuity of f , we may assume that x1 , , xn (a, b). Now, let
= 1 x1 + + n xn . Then, (a, b). By the above lemma, f has the tangent line
y = l (x) = f 0 ()(x ) + f () at x = satisfying f (x) l (x) for all x (a, b). Hence,
the supporting line inequality shows that
1 f (x1 ) + + n f (xn ) 1 f () + + n f () = f () = f (1 x1 + + n xn ).

Non-convex functions can be convex locally and have supporting lines at some points.
This means that the supporting line inequality is a powerful tool because we can also
produce Jensen-type inequalities for non-convex functions.
Epsilon 74. (Titu Andreescu, Gabriel Dospinescu) Let x, y, and z be real numbers such
that x, y, z 1 and x + y + z = 1. Prove that
1
1
1
27
+
+

.
1 + x2
1 + y2
1 + z2
10
Epsilon 75. (Japan 1997) Let a, b, and c be positive real numbers. Prove that
(b + c a)2
(c + a b)2
(a + b c)2
3
+
+
.
2
2
2
2
(b + c) + a
(c + a) + b
(a + b)2 + c2
5

Any good idea can be stated in fifty words or less. - S. Ulam

INFINITY

69

6. Homogenizations and Normalizations

Mathematicians do not study objects, but relations between objects.
- H. Poincare

6.1. Homogenizations. Many inequality problems come with constraints such as ab = 1,

xyz = 1, x + y + z = 1. A non-homogeneous symmetric inequality can be transformed
into a homogeneous one. Then we apply two powerful theorems: Schurs Inequality and
Muirheads Theorem. We begin with a simple example.
Example 11. (Hungary, 1996) Let a and b be positive real numbers with a + b = 1. Prove
that
b2
1
a2
+
.
a+1
b+1
3
Solution. Using the condition a+b = 1, we can reduce the given inequality to homogeneous
one:
1
a2
b2

+
3
(a + b)(a + (a + b))
(a + b)(b + (a + b))
or
a2 b + ab2 a3 + b3 ,
which follows from
(a3 + b3 ) (a2 b + ab2 ) = (a b)2 (a + b) 0.
The equality holds if and only if a = b = 12 .

Theorem 6.1. Let a1 , a2 , b1 , b2 be positive real numbers such that a1 + a2 = b1 + b2 and

max(a1 , a2 ) max(b1 , b2 ). Let x and y be nonnegative real numbers. Then, we have
xa1 y a2 + xa2 y a1 xb1 y b2 + xb2 y b1 .
Proof. Without loss of generality, we can assume that a1 a2 , b1 b2 , a1 b1 . If x or
y is zero, then it clearly holds. So, we assume that both x and y are nonzero. It follows
from a1 + a2 = b1 + b2 that a1 a2 = (b1 a2 ) + (b2 a2 ). Its easy to check
=
=
=

xa1 y a2 + xa2 y a1 xb1 y b2 xb2 y b1

xa2 y a2 xa1 a2 + y a1 a2 xb1 a2 y b2 a2 xb2 a2 y b1 a2

xa2 y a2 xb1 a2 y b1 a2
xb2 a2 y b2 a2

1
xb2 y b2 0.
xb1 y b1
a
a
2
2
x y

Remark 6.1. When does the equality hold in the above theorem?
We now introduce two summation notations. Let P(x, y, z) be a three variables function
of x, y, z. Let us define
X
P(x, y, z) = P(x, y, z) + P(y, z, x) + P(z, x, y)
cyclic

and
X
sym

P(x, y, z) = P(x, y, z) + P(x, z, y) + P(y, x, z) + P(y, z, x) + P(z, x, y) + P(z, y, x).

70

INFINITY

Here, we have some examples:

X 3
x y = x3 y + y 3 z + z 3 x,
X

x3 = 2(x3 + y 3 + z 3 ),

sym

cyclic
2

x y = x y + x z + y z + y x + z 2 x + z 2 y,

sym

xyz = 6xyz

sym

Example 12. Let x, y, z be positive real numbers. Show the cyclic inequality
x2
y2
z2
x
y
z
+ 2 + 2 + + .
2
y
z
x
y
z
x
Third Solution. We break the homogeneity. After the substitution a = xy , b = yz , c =
becomes
a2 + b2 + c2 a + b + c.
Using the constraint abc = 1, we now impose the homogeneity to this as follows:

z
,
x

it

a2 + b2 + c2 (abc) 3 (a + b + c) .
After setting a = x3 , b = y 3 , c = z 3 with x, y, z > 0, it then becomes
x6 + y 6 + z 6 x4 yz + xy 4 z + xyz 4 .
We now deduce
X 6
X x6 + y 6
X x4 y 2 + x2 y 4
X 4 y2 + z2
X 4
x =

=
x

x yz.
2
2
2
cyclic

cyclic

cyclic

cyclic

cyclic

Epsilon 76. [IMO 1984/1 FRG] Let x, y, z be nonnegative real numbers such that x+y+z =
1. Prove that
7
0 xy + yz + zx 2xyz
.
27
Epsilon 77. [LL 1992 UNK] (Iran 1998) Prove that, for all x, y, z > 1 such that
2,
p

x + y + z x 1 + y 1 + z 1.

1
+ y1
x

+ z1 =

INFINITY

71

6.2. Schur and Muirhead.

Theorem 6.2. (Schurs Inequality) Let x, y, z be nonnegative real numbers. For any r > 0,
we have
X r
x (x y)(x z) 0.
cyclic

Proof. Since the inequality is symmetric in the three variables, we may assume without
loss of generality that x y z. Then the given inequality may be rewritten as
(x y)[xr (x z) y r (y z)] + z r (x z)(y z) 0,
and every term on the left-hand side is clearly nonnegative.

Remark 6.2. When does the equality hold in Schurs Inequality?

Delta 66. Disprove the following proposition: for all a, b, c, d 0 and r > 0, we have
ar (ab)(ac)(ad)+br (bc)(bd)(ba)+cr (ca)(cc)(ad)+dr (da)(db)(dc) 0.
Delta 67. [LL 1971 HUN] Let a, b, c, d, e be real numbers. Prove the expression
(a b) (a c) (a d) (a e) + (b a) (b c) (b d) (b e)
+

(c a) (c b) (c d) (c e) + (d a) (d b) (d c) (a e)

(e a) (e b) (e c) (e d)

is nonnegative.
The following special case of Schurs Inequality is useful:
X
X 3 X 2
X
X 3
X 2
x(x y)(x z) 0 3xyz +
x
x y
xyz +
x 2
x y.
cyclic

cyclic

sym

sym

sym

sym

Epsilon 78. Let x, y, z be nonnegative real numbers. Then, we have

3
3
3
3xyz + x3 + y 3 + z 3 2 (xy) 2 + (yz) 2 + (zx) 2 .
Epsilon 79. Let t (0, 3]. For all a, b, c 0, we have
X 2
X
2
(3 t) + t(abc) t +
a 2
ab.
cyclic

cyclic

Epsilon 80. (APMO 2004/5) Prove that, for all positive real numbers a, b, c,
(a2 + 2)(b2 + 2)(c2 + 2) 9(ab + bc + ca).
Epsilon 81. [IMO 2000/2 USA] Let a, b, c be positive numbers such that abc = 1. Prove
that

1
1
1
a1+
b1+
c1+
1.
b
c
a
Epsilon 82. (Tournament of Towns 1997) Let a, b, c be positive numbers such that abc = 1.
Prove that
1
1
1
+
+
1.
a+b+1
b+c+1
c+a+1
Delta 68. [TZ, p.142] Prove that for any acute triangle ABC,
cot3 A + cot3 B + cot3 C + 6 cot A cot B cot C cot A + cot B + cot C.
Delta 69. [IN, p.103] Let a, b, c be the lengths of a triangle. Prove that
a2 b + a2 c + b2 c + b2 a + c2 a + c2 b > a3 + b3 + c3 + 2abc.
Delta 70. (Sur
anyis Inequality) Show that, for all x1 , , xn 0,
`

(n 1) (x1 n + xn n ) + nx1 xn (x1 + + xn ) x1 n1 + xn n1 .

72

INFINITY

Epsilon 83. (Muirheads Theorem) Let a1 , a2 , a3 , b1 , b2 , b3 be non-negative real numbers

such that
a1 a2 a3 , b1 b2 b3 , a1 b1 , a1 + a2 b1 + b2 , a1 + a2 + a3 = b1 + b2 + b3 .
(In this case, we say that the vector a = (a1 , a2 , a3 ) majorizes the vector b = (b1 , b2 , b3 )
and write a b.) For all positive real numbers x, y, z, we have
X a a a
X b b b
x 1y 2z 3
x 1y 2z 3.
sym

sym

Remark 6.3. The equality holds if and only if x = y = z. However, if we allow x = 0

or y = 0 or z = 0, then one may easily check that the equality holds (after assuming
a1 , a2 , a3 > 0 and b1 , b2 , b3 > 0) if and only if
x = y = z or x = y, z = 0 or y = z, x = 0 or z = x, y = 0.
We can apply Muirheads Theorem to establish Nesbitts Inequality.
Proposition 6.1. (Nesbitt) For all positive real numbers a, b, c, we have
b
c
3
a
+
+
.
b+c
c+a
a+b
2
Proof 10. Clearing the denominators of the inequality, it becomes
X
2
a(a + b)(a + c) 3(a + b)(b + c)(c + a)
cyclic

or

X
sym

a3

a2 b.

sym

Epsilon 84. [IMO 1995/2 RUS] Let a, b, c be positive numbers such that abc = 1. Prove
that
1
1
1
3
+ 3
+ 3
.
a3 (b + c)
b (c + a)
c (a + b)
2
Epsilon 85. (Iran 1996) Let x, y, z be positive real numbers. Prove that

1
1
1
9
(xy + yz + zx)
+
+
.
(x + y)2
(y + z)2
(z + x)2
4
Epsilon 86. Let x, y, z be nonnegative real numbers with xy + yz + zx = 1. Prove that
1
1
5
1
+
+
.
x+y
y+z
z+x
2
Epsilon 87. [SC] If ma ,mb ,mc are medians and ra ,rb ,rc the exradii of a triangle, prove
that
ra rb
rb rc
rc ra
+
+
3.
ma mb
mb mc
mc ma
We now offer a criterion for the homogeneous symmetric polynomial inequalities with
degree 3. It is a direct consequence of Schurs Inequality and Muirheads Theorem.
Epsilon 88. Let P(u, v, w) R[u, v, w] be a homogeneous symmetric polynomial with degree
3. Then the following two statements are equivalent.
(a) P(1, 1, 1), P(1, 1, 0), P(1, 0, 0) 0.
(b) P(x, y, z) 0 for all x, y, z 0.
Example 13. [IMO 1984/1 FRG] Let x, y, z be nonnegative real numbers such that x +
y + z = 1. Prove that
7
0 xy + yz + zx 2xyz
.
27

INFINITY

73

Solution. Using x + y + z = 1, we convert the given inequality to the equivalent form:

7
0 (xy + yz + zx)(x + y + z) 2xyz
(x + y + z)3 .
27
Let us define L(u, v, w), R(u, v, w) R[u, v, w] by
L(u, v, w) = (uv + vw + wu)(u + v + w) 2uvw,
7
R(u, v, w) =
(u + v + w)3 (uv + vw + wu)(u + v + w) + 2uvw.
27
However, one may easily check that
L(1, 1, 1) = 7, L(1, 1, 0) = 2, L(1, 0, 0) = 0,
7
2
, R(1, 0, 0) =
.
R(1, 1, 1) = 0, R(1, 1, 0) =
27
27

In other words, we dont need to employ Schurs Inequality and Muirheads Theorem
to get a straightforward result.
Delta 71. (M. S. Klamkin) Determine the maximum and minimum values of
x2 + y 2 + z 2 + xyz
where x + y + z = 1, x, y, z 0, and is a given constant.
Delta 72. (W. Janous) Let x, y, z 0 with x + y + z = 1. For fixed real numbers a 0
and b, determine the maximum c = c(a, b) such that
a + bxyz c(xy + yz + zx).
As a corollary of the above criterion, we obtain the following proposition for homogeneous symmetric polynomial inequalities for the triangles :
Theorem 6.3. (K. B. Stolarsky) Let P(u, v, w) be a real symmetric form of degree 3. If
we have
P(1, 1, 1), P(1, 1, 0), P(2, 1, 1) 0,
then P(a, b, c) 0, where a, b, c are the lengths of the sides of a triangle.
Proof. Employ The Ravi Substitution together with the above crieterion. We leave the
details for the readers. For an alternative proof, see [KS].

Delta 73. (China 2007) Let a, b, c be the lengths of a triangle with a+b+c = 3. Determine
the minimum value of
4abc
a2 + b2 + c2 +
.
3
As noted in [KS], applying Stolarskys Crieterion, we obtain various cubic inequalities
in triangle geometry.
Example 14. Let a, b, c be the lengths of the sides of a triangle. Let s be the semiperimeter
of the triangle. Then, the following inequalities holds.
(a) 4(ab + bc + ca) > (a + b`+ c)2 3(ab
+ bc + ca)

36
(b) [DM] a2 + b2 + c2 35
s2 + abc
s
(c) [AP] abc 8(s a)(s b)(s c)
(d) [EC] 8abc (a + b)(b + c)(c + a)
(e) [AP] 8(a3 + b3 + c3 ) 3(a + b)(b + c)(c + a)
(f) [MC] 2(a + b + c)(a2 + b2 + c2 ) 3(a3 + b3 + c3 + 3abc)
(g) 32 abc a2 (s a) + b2 (s b) + c2 (s c) > abc
(h) bc(b + c) + ca(c + a) + ab(a + b) 48(s a)(s b)(s c)
1
1
1
+ sb
+ sc
9s
(i) sa
a
b
c
(j) [AN, MP] 2 > b+c + c+a
+ a+b
32

74

INFINITY
s+b
s+c
(k) 92 > s+a
+ c+a
+ a+b
15
b+c
4
(l) [SR1] 5[ab(a + b) + bc(b + c) + ca(c + a)] 3abc (a + b + c)3

Proof. We only check the left hand side inequality in (j). One may easily check that it is
equivalent to the cubic inequality T (a, b, c) 0, where

a
b
c
T (a, b, c) = 2(a + b)(b + c)(c + a) (a + b)(b + c)(c + a)
+
+
.
b+c
c+a
a+b
Since T (1, 1, 1) = 4, T (1, 1, 0) = 0, and T (2, 1, 1) = 6, the result follows from Stolarskys
Criterion. For alternative proofs, see [BDJMV].

INFINITY

75

6.3. Normalizations. In the previous subsections, we transformed non-homogeneous inequalities into homogeneous ones. On the other hand, homogeneous inequalities also can
be normalized in various ways. We offer two alternative solutions of the problem 8 by
normalizations :
Epsilon 89. [IMO 2001/2 KOR] Let a, b, c be positive real numbers. Prove that
a
b
c

+
+
1.
a2 + 8bc
b2 + 8ca
c2 + 8ab
Epsilon 90. [IMO 1983/6 USA] Let a, b, c be the lengths of the sides of a triangle. Prove
that
a2 b(a b) + b2 c(b c) + c2 a(c a) 0.
Epsilon 91. (KMO Winter Program Test 2001) Prove that, for all a, b, c > 0,
p
p
(a2 b + b2 c + c2 a) (ab2 + bc2 + ca2 ) abc + 3 (a3 + abc) (b3 + abc) (c3 + abc)
Epsilon 92. [IMO 1999/2 POL] Let n be an integer with n 2. (a) Determine the least
constant C such that the inequality
0
14
X
X
xi A
xi xj (x2i + x2j ) C @
1in

1i<jn

holds for all real numbers x1 , , xn 0.

(b) For this constant C, determine when equality holds.
Delta 74. [SL 1991 POL] Let n be a given integer with n 2. Find the maximum value
of
X
xi xj (xi + xj ),
1i<jn

where x1 , , xn 0 and x1 + + xn = 1.
We close this subsection with another proofs of Nesbitts Inequality.
Proposition 6.2. (Nesbitt) For all positive real numbers a, b, c, we have
b
c
3
a
+
+
.
b+c
c+a
a+b
2
Proof 11. We may normalize to a + b + c = 1. Note that 0 < a, b, c < 1. The problem is
now to prove
X a
X
3
x
=
f (a) , where f (x) =
.
b+c
2
1x
cyclic

cyclic

Since f is convex on (0, 1), Jensens Inequality shows that

1 X
a+b+c
1
1
f (a) f
=f
=
or
3
3
3
2
cyclic

f (a)

cyclic

3
.
2

Proof 12. (Cao Minh Quang) Assume that a + b + c = 1. Note that ab + bc + ca

1
(a + b + c)2 = 13 . More strongly, we establish that
3
b
c
9
a
+
+
3 (ab + bc + ca)
b+c
c+a
a+b
2
or

9a(b + c)
9b(c + a)
9c(a + b)
a
b
c
+
+
+
+
+
3.
b+c
4
c+a
4
a+b
4
The AM-GM inequality shows that
r
X
X
X a
9a(b + c)
9a(b + c)
a
+

=
3a = 3.
b+c
4
b+c
4
cyclic

cyclic

cyclic

76

INFINITY

6.4. Cauchy-Schwarz and H

older. We begin with the following famous theorem:
Theorem 6.4. (The Cauchy-Schwarz Inequality) Whenever a1 , , an , b1 , , bn R, we
have
(a1 2 + + an 2 )(b1 2 + + bn 2 ) (a1 b1 + + an bn )2 .
p

First Proof. Let A = a1 2 + + an 2 and B =

b1 2 + + bn 2 . In the case when
A = 0, we get a1 = = an = 0. Thus, the given inequality clearly holds. From now on,
we assume that A, B > 0. Since the inequality is homogeneous, we may normalize to
1 = a1 2 + + an 2 = b1 2 + + bn 2 .
We now need to to show that
|a1 b1 + + an bn | 1.
Indeed, we deduce
|a1 b1 + + an bn | |a1 b1 | + + |an bn |

an 2 + b n 2
a1 2 + b 1 2
+ +
= 1.
2
2

Second Proof. It immediately follows from The Lagrange Identity:

! n
!
!2
n
n
X
X
X 2
X
2
=
(ai bj aj bi )2 .
ai
bi

ai b i
i=1

i=1

i=1

1i<jn

Delta 75. [IMO 2003/5 IRL] Let n be a positive integer and let x1 xn be real
numbers. Prove that
0
12
`

X
X
2 n2 1
@
A
(xi xj )2 .
|xi xj |

3
1i,jn

1i,jn

Show that the equality holds if and only if x1 , , xn is an arithmetic progression.

Delta 76. (Darij Grinberg) Suppose that 0 < a1 an and 0 < b1 bn be real
numbers. Show that
!2
!2
! n
!
!2
n
n
n
n
X
X
X 2
X
1 X
2
ak
bk
>
ak
bk

ak b k
4
k=1

k=1

k=1

k=1

k=1

Delta 77. [LL 1971 AUT] Let a, b, c be positive real numbers, 0 <
for any x, y, z > 0 the following inequality holds:
x
(a + c)2
y
(x + y + z)2 (ax + by + cz)
+ +
4ac
a
b

a b c. Prove that
z
.
c

Delta 78. [LL 1987 AUS] Let a1 , a2 , a3 , b1 , b2 , b3 be positive real numbers. Prove that
(a1 b2 + a1 b3 + a2 b1 + a2 b3 + a3 b1 + a3 b2 )2 4 (a1 a2 + a2 a3 + a3 a1 ) (b1 b2 + b2 b3 + b3 b1 )
and show that the two sides of the inequality are equal if and only if

a1
b1

a2
b2

a3
.
b3

Delta 79. [PF] Let a1 , , an , b1 , , bn R. Suppose that x [0, 1]. Show that
12
! n
! 0 n
n
X
X
X
X
X 2
X
2
ai b i + x
ai b j A .
ai + 2x
ai aj
bi + 2x
bi bj @
i=1

i<j

i=1

i<j

i=1

ij

INFINITY

77

Delta 80. Let a1 , , an , b1 , , bn be positive real numbers. Show that

8
p

>
(1) (a1 + + an )(b1 + + bn ) a1 b1 + + an bn ,
>
>
>
2
2
2
>
<(2) a1 + + an (a1 ++an ) ,
b1
bn
b1 ++bn
2
a
a
a
a
1
>
>(3) b112 + + bnn2 a1 ++an b11 + + bnn ,
>
>
>
:(4) a1 + + an (a1 ++an )2 .
b1

bn

a1 b1 ++an bn

Delta 81. [SL 1993 USA] Prove that

a
b
c
d
2
+
+
+

b + 2c + 3d
c + 2d + 3a
d + 2a + 3b
a + 2b + 3c
3
for all positive real numbers a, b, c, d.
Epsilon 93. (APMO 1991) Let a1 , , an , b1 , , bn be positive real numbers such that
a1 + + an = b1 + + bn . Show that
a1 2
an 2
a1 + + an
.
+ +

a1 + b 1
an + b n
2
Epsilon 94. Let a, b 0 with a + b = 1. Prove that
p
p

a2 + b + a + b2 + 1 + ab 3.
Show that the equality holds if and only if (a, b) = (1, 0) or (a, b) = (0, 1).
Epsilon 95. [LL 1992 UNK] (Iran 1998) Prove that, for all x, y, z > 1 such that
2,
p

x + y + z x 1 + y 1 + z 1.

1
+ y1
x

+ z1 =

We now apply The Cauchy-Schwarz Inequality to prove Nesbitts Inequality.

Proposition 6.3. (Nesbitt) For all positive real numbers a, b, c, we have
a
b
c
3
+
+
.
b+c
c+a
a+b
2
Proof 13. Applying The Cauchy-Schwarz Inequality, we have

1
1
1
((b + c) + (c + a) + (a + b))
+
+
32 .
b+c
c+a
a+b
It follows that
or

a+b+c
a+b+c
9
a+b+c
+
+

b+c
c+a
a+b
2
X a
9
3+
.
b+c
2
cyclic

Proof 14. The Cauchy-Schwarz Inequality yields

X
cyclic

or

0
12
X
a X
a(b + c) @
aA
b+c

X
cyclic

cyclic

cyclic

(a + b + c)2
a
3

.
b+c
2(ab + bc + ca)
2

Epsilon 96. (Gazeta Matematic

a) Prove that, for all a, b, c > 0,
p
p
p
p
p
p
a4 + a2 b2 + b4 + b4 + b2 c2 + c4 + c4 + c2 a2 + a4 a 2a2 + bc+b 2b2 + ca+c 2c2 + ab.
Epsilon 97. (KMO Winter Program Test 2001) Prove that, for all a, b, c > 0,
p
p
(a2 b + b2 c + c2 a) (ab2 + bc2 + ca2 ) abc + 3 (a3 + abc) (b3 + abc) (c3 + abc)

78

INFINITY

Epsilon 98. (Andrei Ciupan) Let a, b, c be positive real numbers such that
1
1
1
+
+
1.
a+b+1
b+c+1
c+a+1
Show that a + b + c ab + bc + ca.
We now illustrate normalization techniques to establish classical theorems. Using the
same idea in the proof of The Cauchy-Schwarz Inequality, we find a natural generalization
:
Theorem 6.5. Let aij (i, j = 1, , n) be positive real numbers. Then, we have
(a11 n + + a1n n ) (an1 n + + ann n ) (a11 a21 an1 + + a1n a2n ann )n .
Proof. The inequality is homogeneous. We make the normalizations:
1

(ai1 n + + ain n ) n = 1
or
ai1 n + + ain n = 1,
for all i = 1, , n. Then, the inequality takes the form
a11 a21 an1 + + a1n a2n ann 1
or
n
X

ai1 ain 1.

i=1

Hence, it suffices to show that, for all i = 1, , n,

ai1 ain

1
n

where ai1 n + + ain n = 1. To finish the proof, it remains to show the following homogeneous inequality.

Theorem 6.6. (The AM-GM Inequality) Let a1 , , an be positive real numbers. Then,
we have

a1 + + an
n a1 an .
n
Proof. Since its homogeneous, we may rescale a1 , , an so that a1 an = 1.
want to show that

14

We

a1 an = 1 = a1 + + an n.
The proof is by induction on n. If n = 1, its trivial. If n = 2, then we get a1 +a2 2 = a1 +

a2 2 a1 a2 = ( a1 a2 )2 0. Now, we assume that it holds for some positive integer

n 2. And let a1 , , an+1 be positive numbers such that a1 an an+1 =1. We may
assume that a1 1 a2 . (Why?) It follows that a1 a2 + 1 a1 a2 = (a1 1)(a2 1) 0
so that a1 a2 + 1 a1 + a2 . Since (a1 a2 )a3 an = 1, by the induction hypothesis, we
have
a1 a2 + a3 + + an+1 n.
It follows that a1 + a2 1 + a3 + + an+1 n.
14

Set xi =

n.

ai

(a1 an ) n

(i = 1, , n). Then, we get x1 xn = 1 and it becomes x1 + +xn

INFINITY

79

We now make simple observation. Let a, b > 0 and m, n N. Take x1 = = xm = a

and xm+1 = = xxm+n = b. Applying the AM-GM inequality to x1 , , xm+n > 0, we
obtain
1
m
n
ma + nb
m
n
(am bn ) m+n or
a+
b a m+n b m+n .
m+n
m+n
m+n
Hence, for all positive rational numbers 1 and 2 with 1 + 2 = 1, we get
1 a + 2 b a 1 b 2 .
We now immediately have
Theorem 6.7. Let 1 , 2 > 0 with 1 + 2 = 1. For all x, y > 0, we have
1 x + 2 y x 1 y 2 .
Proof. We can choose a sequence a1 , a2 , a3 , (0, 1) of rational numbers such that
lim an = 1 .

Set bi = 1 ai , where i N. Then, b1 , b2 , b3 , (0, 1) is a sequence of rational numbers

with
lim bn = 2 .
n

From the previous observation, we have an x + bn y xan y bn . By taking the limits to

both sides, we get the result.

We can extend the above arguments to the n-variables.

Theorem 6.8. (The Weighted AM-GM Inequality) Let 1 , , n > 0 with 1 + +n = 1.
For all x1 , , xn > 0, we have
1 x1 + + n xn x1 1 xn n .
Since we now get the weighted version of The AM-GM Inequality, we establish weighted
version of The Cauchy-Schwarz Inequality.
Epsilon 99. (H
olders Inequality) Let xij (i = 1, , m, j = 1, n) be positive real numbers. Suppose that 1 , , n are positive real numbers satisfying 1 + + n = 1.
Then, we have
!j
!
n
m
m
n
Y
X
X
Y
j
xij

xij
.
j=1

i=1

i=1

j=1

My brain is open. - P. Erd

os

80

INFINITY

7. Convexity and Its Applications

The art of doing mathematics consists in finding that special case
which contains all the germs of generality.
- D. Hilbert

7.1. Jensens Inequality. In the previous section, we deduced the weighted AM-GM inequality from The AM-GM Inequality. We use the same idea to study the following
functional inequalities.
Epsilon 100. Let f : [a, b] R be a continuous function. Then, the followings are
equivalent.
(1) For all n N, the following inequality holds.
1 f (x1 ) + + n f (xn ) f (1 x1 + + n xn )
for all x1 , , xn [a, b] and 1 , , n > 0 with 1 + + n = 1.
(2) For all n N, the following inequality holds.
r1 f (x1 ) + + rn f (xn ) f (r1 x1 + + rn xn )
for all x1 , , xn [a, b] and r1 , , rn Q+ with r1 + + rn = 1.
(3) For all N N, the following inequality holds.
y + + y
f (y1 ) + + f (yN )
1
N
f
N
N
for all y1 , , yN [a, b].
(4) For all k {0, 1, 2, }, the following inequality holds.
y + + y
f (y1 ) + + f (y2k )
1
2k
f
k
2
2k
for all y1 , , y2k [a, b].
`

(5) We have 12 f (x) + 12 f (y) f x+y

for all x, y [a, b].
2
(6) We have f (x) + (1 )f (y) f (x + (1 )y) for all x, y [a, b]
and (0, 1).
Definition 7.1. A real valued function f : [a, b] R is said to be convex if the inequality
f (x) + (1 )f (y) f (x + (1 )y)
holds for all x, y [a, b] and (0, 1).
The above proposition says that
Corollary 7.1. (Jensens Inequality) If f : [a, b] R is a continuous convex function, then
for all x1 , , xn [a, b], we have
x + + x
f (x1 ) + + f (xn )
1
n
f
.
n
n
Delta 82. [SL 1998 AUS] Let r1 , , rn be real numbers greater than or equal to 1. Prove
that
1
1
n
+ +

.
n
r1 + 1
rn + 1
r1 rn + 1
Corollary 7.2. (The Weighted Jensens Inequality) Let f : [a, b] R be a continuous
convex function. Let 1 , , n > 0 with 1 + + n = 1. For all x1 , , xn [a, b],
we have
1 f (x1 ) + + n f (xn ) f (1 x1 + + n xn ).

INFINITY

81

In fact, we can almost drop the continuity of f . As an exercise, show that every convex
function on [a, b] is continuous on (a, b). Hence, every convex function on R is continuous
on R.
Corollary 7.3. (The Convexity Criterion I) If a continuous function f : [a, b] R satisfies
the midpoint convexity
x + y
f (x) + f (y)
f
2
2
for all x, y [a, b], then the function f is convex on [a, b].
Delta 83. (The Convexity Criterion II) Let f : [a, b] R be a continuous function which
are differentiable twice in (a, b). Show that (1) f 00 (x) 0 for all x (a, b) if and only if
(2) f is convex on (a, b).
We now present an inductive proof of The Weighted Jensens Inequality. It turns out
that we can completely drop the continuity of f .
Third Proof. It clearly holds for n = 1, 2. We now assume that it holds for some n N.
Let x1 , , xn , xn+1 [a, b] and 1 , , n+1 > 0 with 1 + + n+1 = 1. Since we
have the equality
1
n
+ +
= 1,
1 n+1
1 n+1
by the induction hypothesis, we obtain

1 f (x1 ) + + n+1 f (xn+1 )

1
n
(1 n+1 )
f (x1 ) + +
f (xn ) + n+1 f (xn+1 )
1 n+1
1 n+1

n
1
x1 + +
xn + n+1 f (xn+1 )
(1 n+1 )f
1 n+1
1 n+1

1
n
f (1 n+1 )
x1 + +
xn + n+1 xn+1
1 n+1
1 n+1
f (1 x1 + + n+1 xn+1 ).

82

INFINITY

7.2. Power Mean Inequality. The notion of convexity is one of the most important concepts in analysis. Jensens Inequality is the most powerful tool in theory of inequalities.
We begin with a convexity proof of The Weighted AM-GM Inequality.
Theorem 7.1. (The Weighted AM-GM Inequality) Let 1 , , n > 0 with 1 + +n = 1.
For all x1 , , xn > 0, we have
1 x1 + + n xn x1 1 xn n .
Proof. It is a straightforward consequence of the concavity of ln x. Indeed, The Weighted
Jensens Inequality shows that
ln(1 x1 + + n xn ) 1 ln(x1 ) + + n ln(xn ) = ln(x1 1 xn n ).

The Power Mean Inequality can be proved by exploiting Jensens inequality in two
ways. We begin with two simple lemmas.
Lemma 7.1. Let a, b, and c be positive real numbers. Let
x

a + bx + cx
f (x) = ln
3

for all x R. Then, we obtain f 0 (0) = ln 3 abc.

Proof. We compute
f 0 (x) =
It follows that
f 0 (0) =

ax ln a + bx ln b + cx ln c
.
a x + bx + cx

ln a + ln b + ln c
3
= ln abc.
3

Lemma 7.2. Let f : R R be a continuous function. Suppose that f is monotone

increasing on (0, ) and monotone increasing on (, 0). Then, the function f is monotone increasing on R.
Proof. We first show that f is monotone increasing on [0, ). By the hypothesis, it
remains to show that f (x) f (0) for all x > 0. For all (0, x), we have f (x) f ().
Since f is continuous at 0, we obtain
f (x) lim f () = f (0).
0+

Similarly, we find that f is monotone increasing on (, 0]. We now show that f is

monotone increasing on R. Let x and y be real numbers with x > y. We want to show
that f (x) f (y). In case 0 6 (x, y), we get the result by the hypothesis. In case x 0 y,
it follows that f (x) f (0) f (y).

Theorem 7.2. (Power Mean inequality for Three Variables) Let a, b, and c be positive real
numbers. We define a function M(a,b,c) : R R by
M(a,b,c) (0) =

abc , M(a,b,c) (r) =

ar + b r + c r
3

Then, M(a,b,c) is a monotone increasing continuous function.

1
r

(r 6= 0).

INFINITY

83

First Proof. Write M (r) = M(a,b,c) (r). We first establish that the function M is continuous. Since M is continuous at r for all r 6= 0, its enough to show that

3
lim M (r) = abc.
r0
x x x
Let f (x) = ln a +b3 +c , where x R. Since f (0) = 0, the above lemma implies that

f (r)
f (r) f (0)
3
= lim
= f 0 (0) = ln abc .
r0
r0
r
r0
Since ex is a continuous function, this means that

f (r)
3
3
lim M (r) = lim e r = eln abc = abc.
lim

r0

r0

Now, we show that the function M is monotone increasing. It will be enough to establish
that M is monotone increasing on (0, ) and monotone increasing on (, 0). We first
show that M is monotone increasing on (0, ). Let x y > 0. We want to show that
x
1
y
1
a + bx + cx x
a + by + cy y

.
3
3
After the substitution u = ay , v = ay , w = az , it becomes
!1
x
x
x
x
u + v + w y1
uy + vy + wy

.
3
3
Since it is homogeneous, we may normalize to u + v + w = 3. We are now required to
show that
G(u) + G(v) + G(w)
1,
3
x
where G(t) = t y , where t > 0. Since xy 1, we find that G is convex. Jensens inequality
shows that
u + v + w
G(u) + G(v) + G(w)
G
= G(1) = 1.
3
3
Similarly, we may deduce that M is monotone increasing on (, 0).

Weve learned that the convexity of f (x) = x ( 1) implies the monotonicity of

the power means. Now, we shall show that the convexity of x ln x also implies The Power
Mean Inequality.
Second Proof of the Monotonicity. Write f (x) = M(a,b,c) (x). We use the increasing function theorem. Its enough to show that f 0 (x) 0 for all x 6= 0. Let x R {0}. We
compute
x

x
x
x
f 0 (x)
d
1
a + bx + cx
1 13 (a ln a + b ln b + c ln c)
=
(ln f (x)) = 2 ln
+
1
f (x)
dx
x
3
x
(ax + bx + cx )
3
or
x

x2 f 0 (x)
a + bx + cx
ax ln ax + bx ln bx + cx ln cx
= ln
+
.
f (x)
3
ax + b x + c x
To establish f 0 (x) 0, we now need to establish that
x

a + bx + cx
ax ln ax + bx ln bx + cx ln cx (ax + bx + cx ) ln
.
3
Let us introduce a function f : (0, ) R by f (t) = t ln t, where t > 0. After the
substitution p = ax , q = ay , r = az , it becomes
p + q + r
.
f (p) + f (q) + f (r) 3f
3
Since f is convex on (0, ), it follows immediately from Jensens Inequality.

84

INFINITY

In particular, we deduce The RMS-AM-GM-HM Inequality for three variables.

Corollary 7.4. For all positive real numbers a, b, and c, we have
r

a2 + b2 + c2
a+b+c
3
3

abc 1
1
3
3
+
+
a
b

1
c

Proof. The Power Mean Inequality implies that

M(a,b,c) (2) M(a,b,c) (1) M(a,b,c) (0) M(a,b,c) (1).

Delta 84. [SL 2004 THA] Let a, b, c > 0 and ab + bc + ca = 1. Prove the inequality
r
r
r
1
3 1
3 1
3 1
+ 6b +
+ 6c +
+ 6a
.
a
b
c
abc
Delta 85. [SL 1998 RUS] Let x, y, and z be positive real numbers such that xyz = 1.
Prove that
y3
z3
3
x3
+
+
.
(1 + y)(1 + z)
(1 + z)(1 + x)
(1 + x)(1 + y)
4
Delta 86. [LL 1992 POL] For positive real numbers a, b, c, define

a+b+c
3
3
A=
, G = abc, H = 1
.
3
+ 1b + 1c
a
Prove that

A
G

1
3 A
+ .
4
4 H

Using the convexity of x ln x or the convexity of x ( 1), we can also establish the
monotonicity of the power means for n positive real numbers.
Theorem 7.3. (The Power Mean Inequality) Let x1 , , xn be positive real numbers. The
power mean of order r is defined by
1
r

x1 + + xn r r
M(x1 , ,xn ) (0) = n x1 xn , M(x1 , ,xn ) (r) =
(r 6= 0).
n
Then, the function M(x1 , ,xn ) : R R is continuous and monotone increasing.
Corollary 7.5. (The Geometric Mean as a Limit) Let x1 , , xn > 0. Then,
r
1

x1 + + xn r r
n
x1 xn = lim
.
r0
n
Theorem 7.4. (The RMS-AM-GM-HM Inequality) For all x1 , , xn > 0, we have
r

x1 2 + + xn 2
x1 + + xn
n

n x1 xn 1
.
n
n
+

+ x1
x
1

Delta 87. [SL 2004 IRL] Let a1 , , an be positive real numbers, n > 1. Denote by gn
their geometric mean, and by A1 , , An the sequence of arithmetic means defined by
a1 + + ak
, k = 1, , n.
Ak =
k
Let Gn be the geometric mean of A1 , , An . Prove the inequality
r
Gn
gn
n+1 n
+
An
Gn
and establish the cases of equality.

INFINITY

85

7.3. Hardy-Littlewood-P
olya Inequality. We first meet a famous inequality established by
the Romanian mathematician T. Popoviciu.
Theorem 7.5. (Popovicius Inequality) Let f : [a, b] R be a convex function. For all
x, y, z [a, b], we have
x + y
y + z
z + x
x + y + z
f (x) + f (y) + f (z) + 3f
2f
+ 2f
+ 2f
.
3
2
2
2
Proof. We break the symmetry. Since the inequality is symmetric, we may assume that
x y z.
Case 1. y

x+y+z
:
3

The key idea is to make the following geometric observation:

z +x x+y h x+ y + zi
,
x,
.
2
2
3
It guarantees the existence of two positive weights 1 , 2 [0, 1] satisfying that
8
x+y+z
z+x
>
< 2 = (1 1 ) x + 1 3 ,
x+y
= (1 2 ) x + 2 x+y+z
,
2
3
>
:
1 + 2 = 32 .
Now, Jensens inequality shows that
x + y
y + z
z + x
f
+f
+f
2
2
2
x + y + z f (y) + f (z)
x + y + z
(1 2 ) f (x) + 2 f
+
+ (1 1 ) f (x) + 1 f
3
2
3
3 x + y + z
1
(f (x) + f (y) + f (z)) + f
.

2
2
3
The proof of the second case uses the same idea.
x+y+z
:
3

We make the following geometric observation:

z + x y + z hx + y + z i
,

,z .
2
2
3
It guarantees the existence of two positive weights 1 , 2 [0, 1] satisfying that
8
x+y+z
z+x
>
< 2 = (1 1 ) z + 1 3 ,
y+z
x+y+z
= (1 2 ) z + 2 3 ,
2
>
:
1 + 2 = 32 .
Case 2. y

Jensens inequality implies that

x + y
y + z
z + x
f
+f
+f
2
2
2
x + y + z
x + y + z
f (x) + f (y)

+ (1 2 ) f (z) + 2 f
+ (1 1 ) f (z) + 1 f
2
3
3
1
3 x + y + z

(f (x) + f (y) + f (z)) + f

.
2
2
3

Epsilon 101. Let x, y, z be nonnegative real numbers. Then, we have

3
3
3
3xyz + x3 + y 3 + z 3 2 (xy) 2 + (yz) 2 + (zx) 2 .
Extending the proof of Popovicius Inequality, we can establish a majorization inequality.
Definition 7.2. We say that a vector x = (x1 , , xn ) Rn majorizes a vector y =
(y1 , , yn ) Rn if we have

86

INFINITY

(1) x1 xn , y1 yn ,
(2) x1 + + xk y1 + + yk for all 1 k n 1,
(3) x1 + + xn = y1 + + yn .
In this case, we write x y.
Theorem 7.6. (The Hardy-Littlewood-P
olya Inequality) Let f : [a, b] R be a convex
function. Suppose that (x1 , , xn ) majorizes (y1 , , yn ), where x1 , , xn , y1 , , yn
[a, b]. Then, we obtain
f (x1 ) + + f (xn ) f (y1 ) + + f (yn ).
Epsilon 102. Let ABC be an acute triangle. Show that
cos A + cos B + cos C 1.
Epsilon 103. Let ABC be a triangle. Show that



A
B
C
tan 2
+ tan 2
+ tan 2
1.
4
4
4
Epsilon 104. Use The Hardy-Littlewood-P
olya Inequality to deduce Popovicius Inequality.
Epsilon 105. [IMO 1999/2 POL] Let n be an integer with n 2. (a) Determine the least
constant C such that the inequality
0
14
X
X
xi xj (x2i + x2j ) C @
xi A
1i<jn

1in

holds for all real numbers x1 , , xn 0.

(b) For this constant C, determine when equality holds.

Its not that Im so smart, its just that I stay with problems longer. - A. Einstein

INFINITY

87

8. Epsilons
God has a transfinite book with all the theorems and their best
proofs
- P. Erd
os

Epsilon 1. [NS] Let a and b be positive integers such that

ak | bk+1
for all positive integers k. Show that b is divisible by a.
Solution. Let p be a prime. Our job is to establish the equality
ordp (b) ordp (a) .
According to the condition that bk+1 is divisibly by ak , we find that the inequality

(k + 1)ordp (b) = ordp bk+1 ordp ak = k ordp (a)
or

ordp (b)
k

ordp (a)
k+1
holds for all positive integers k. Letting k , we have the estimation
ordp (b)
1.
ordp (a)

88

INFINITY

Epsilon 2. [IMO 1972/3 UNK] Let m and n be arbitrary non-negative integers. Prove
that
(2m)!(2n)!
m!n!(m + n)!
is an integer.
Solution. We want to show that L = (2m)!(2n)! is divisible by R = m!n!(m + n)! Let p
be a prime. Our job is to establish the inequality
ordp (L) ordp (R) .
or



X
2m
k=1

pk

2n
pk



X
m
k=1

pk

n
pk

m+n
pk


.

It is an easy job to check the auxiliary inequality

b2xc + b2yc bxc + byc + bx + yc
holds for all real numbers x and y.

INFINITY

89

Epsilon 3. Let n N. Show that Ln := lcm(1, 2, , 2n) is divisible by Rn :=

`2n
n

(2n)!
.
(n!)2

Solution. Let p be a prime. We want to show the inequality

ordp (Kn ) ordp (Ln ) .
Now, we first compute ordp (Kn ).




X
(2n)!
2n
n
ordp (Kn ) = ordp
=
ord
(
(2n)!
)

2ord
(
n!
)
=

2
.
p
p
pk
pk
(n!)2
k=1
The key observation is that both

2n
pk

and

n
pk

vanish for all sufficiently large integer k. Let

N denote the the largest integer N 0 such that pN 2n. The maximality of the
and pnk are smaller
exponent N = ordp (Ln ) guarantees that, whenever k > N , both 2n
pk
2n
n
than 1, so that the term b pk c 2b pk c vanishes. It follows that


N
X
2n
n
ordp (Kn ) =

2
.
pk
pk
k=1

Since b2xc 2 bxc is either 0 or 1 for all x R, this gives the estimation
ordp (Kn )

N
X

1 = N.

k=1

However, since Ln is the least common multiple of 1, , 2n, we see that ordp (Ln ) is the
largest integer N 0 such that pN 2n

90

INFINITY

Epsilon 4. Let f : N R+ be a function satisfying the conditions:

(a) f (mn) = f (m)f (n) for all positive integers m and n, and
(b) f (n + 1) f (n) for all positive integers n.
Then, there is a constant R such that f (n) = n for all n N.
Proof. We have f (1) = 1. Our job is to show that the function
ln f (n)
ln n
is constant when n > 1. Assume to the contrary that
ln f (n)
ln f (m)
>
ln m
ln n
for some positive integers m, n > 1. Writing f (m) = mx and f (n) = ny , we have x > y or
ln n
ln n y
>

ln m
ln m x
A
So, we can pick a positive rational number B , where A, B N, so that
ln n
A
ln n y
>
>
.
ln m
B
ln m x
Hence, mA < nB and mAx > nBy . One the` one hand,` since
f is monotone increasing,
the first inequality mA < nB means that f mA f nB . On the other hand, since
`

`
f mA = f (m)A = mAx and f nB = f (n)B = nBy , the second inequality mAx > nBy
means that

f mA = mAx > nBy = f nB
This is a contradiction.

INFINITY

91

Epsilon 5. (Putnam 1963/A2) Let f : N N be a strictly increasing function satisfying

that f (2) = 2 and f (mn) = f (m)f (n) for all relatively prime m and n. Then, f is the
identity function on N.
Proof. Since f is strictly increasing, we find that f (n + 1) f (n) + 1 for all positive
integers n. It follows that f (n + k) f (n) + k for all positive integers n and k. We now
determine p = f (3). On the one hand, we obtain
f (18) f (15) + 3 f (3)f (5) + 3 f (3)(f (3) + 2) + 3 = p2 + 2p + 3.
On the other hand, we obtain
f (18) = f (2)f (9) 2(f (10)1) = 2f (2)f (5)2 4(f (6)1)2 = 4f (2)f (3)6 = 8p6.
Combining these two, we deduce p2 + 2p + 3 8p 6 or (p 3)2 0. So, we have
f (3) = p = 3.
`

We now prove that f 2l + 1 = 2l + 1 for all positive integers l. Since f (3) = 3, it

`

clearly holds for l = 1. Assuming that f 2l + 1 = 2l + 1 for some positive integer l, we

obtain

f 2l+1 + 2 = f (2)f 2l + 1 = 2 2l + 1 = 2l+1 + 2.

`

Since f is strictly increasing, this means that f 2l + k = 2l + k for all k {1, , 2l + 2}.
` l+1

In particular, we get f 2
+ 1 = 2l+1 + 1, as desired.
Now, we find that f (n) = n for all positive
n. It clearly
` integers

` holds for
n = 1, 2.
Let l be a fixed positive integer. We have f 2l + 1 = 2l + 1 and f 2l+1 + 1 = 2l+1 + 1.
`

Since f is strictly increasing, this means that f 2l + k = 2l + k for all k {1, , 2l + 1}.
Since it holds for all positive integers l, we conclude that f (n) = n for all n 3. This
completes the proof.

92

INFINITY

Epsilon 6. Let a, b, c be positive real numbers. Prove the inequality

`
`
`

1 + a2 1 + b2 1 + c2 (a + b)(b + c)(c + a).

Show that the equality holds if and only if (a, b, c) = (1, 1, 1).
Solution. The inequality has the symmetric face:
`
`
`
`
`
`

1 + a2 1 + b2 1 + b2 1 + c2 1 + c2 1 + a2 (a + b)2 (b + c)2 (c + a)2 .

Now, the symmetry of this expression gives the right approach. We check that, for x, y > 0,
`
`

1 + x2 1 + y 2 (x + y)2
with the equality condition xy = 1. However, it immediately follows from the identity
`
`

1 + x2 1 + y 2 (x + y)2 = (1 xy)2 .
It is easy to check that the equality in the original inequalty occurs only when a = b =
c = 1.

INFINITY

93

Epsilon 7. (Poland 2006) Let a, b, c be positive real numbers with ab+bc+ca = abc. Prove
that
a4 + b4
b4 + c4
c 4 + a4
+
+
1.
3
3
3
3
ab(a + b )
bc(b + c )
ca(c3 + a3 )
Solution. We first notice that the constraint can be written as
1
1
1
+ + = 1.
a
b
c
It is now enough to establish the auxiliary inequality

1 1
1
x4 + y 4

+
xy(x3 + y 3 )
2 x
y
or
` 4

2 x + y 4 x3 + y 3 (x + y) ,
where x, y > 0. However, we obtain
`
`

2 x4 + y 4 x3 + y 3 (x + y) = x4 + y 4 x3 y xy 3 = x3 y 3 (x y) 0.

94

INFINITY

Epsilon 8. (APMO 1996) Let a, b, c be the lengths of the sides of a triangle. Prove that

a + b c + b + c a + c + a b a + b + c.
Proof. The left hand side admits the following decomposition

c+ab+ a+bc
a+bc+ b+ca
b+ca+ c+ab
+
+
.
2
2
2
q

x+ y
We now use the inequality
x+y
to deduce
2
2

c+ab+ a+bc
a,
2

a+bc+ b+ca
b,
2

b+ca+ c+ab
c.
2
Adding these three inequalities, we get the result.

INFINITY

95

Epsilon 9. Let a, b, c be the lengths of a triangle. Show that

b
c
a
+
+
< 2.
b+c
c+a
a+b
Proof. Since the inequality is symmetric in the three variables, we may assume that a
b c. We obtain
a
a
b
b
c

,
< 1.
b+c
a+b c+a
a+b a+b
Adding these three inequalities, we get the result.

96

INFINITY

Epsilon 10. (USA 1980) Prove that, for all positive real numbers a, b, c [0, 1],
a
b
c
+
+
+ (1 a)(1 b)(1 c) 1.
b+c+1
c+a+1
a+b+1
Solution. Since the inequality is symmetric in the three variables, we may assume that
0 a b c 1. Our first step is to bring the estimation
a
b
c
a
b
c
a+b+c
+
+

+
+

.
b+c+1
c+a+1
a+b+1
a+b+1
a+b+1
a+b+1
a+b+1
It now remains to check that
a+b+c
+ (1 a)(1 b)(1 c) 1.
a+b+1
or
1c
(1 a)(1 b)(1 c)
a+b+1
or
(1 a)(1 b)(a + b + 1) 1.
We indeed obtain the estimation
`
`

(1 a)(1 b)(a + b + 1) (1 a)(1 b)(1 + a)(1 + b) = 1 a2 1 b2 1.

INFINITY

97

Epsilon 11. [AE, p. 186] Show that, for all a, b, c [0, 1],
a
b
c
+
+
2.
1 + bc
1 + ca
1 + ab
Proof. Since the inequality is symmetric in the three variables, we may begin with the
assumption 0 a b c 1. We first give term-by-term estimation:
a
a
b
b
c
1

.
1 + bc
1 + ab 1 + ca
1 + ab 1 + ab
1 + ab
Summing up these three, we reach
a
b
c
a+b+1
+
+

.
1 + bc
1 + ca
1 + ab
1 + ab
We now want to show the inequality
a+b+1
2
1 + ab
or
a + b + 1 2 + 2ab
or
a + b 1 + 2ab.
However, it is immediate that 1+2abab = ab+(1a)(1b) is clearly non-negative.

98

INFINITY

Epsilon 12. [SL 2006 KOR] Let a, b, c be the lengths of the sides of a triangle. Prove the
inequality

b+ca
c+ab
a+bc

3.
b+ c a
c+ a b
a+ b c
Solution. Since the inequality is symmetric in the three variables, we may assume that
a b c. We claim that

a+bc

1
a+ b c
and

b+ca
c+ab

2.

b+ c a
c+ a b
It is clear that the denominators are positive. So, the first inequality is equivalent to

a + b a + b c + c.
or
or

a+

2
2
b
a+bc+ c

ab

c(a + b c)

or
ab c(a + b c),
which immediately
follows
from
(a

second inequality.

c)(b c) 0. Now, we prove the

Setting p = a + b and q = a b, we obtain a b = pq and p 2 c. It now becomes

c pq
c + pq

+
2.
cq
c+q
We now apply The Cauchy-Schwartz Inequality to deduce

c pq
c + pq
c pq
c + pq
1
1

+
+
+
cq
c+q
cq
c+q
cq
c+q

2 c c pq 2
2 c
=

c q2
c q2

c2 cpq 2
= 4
(c q 2 )2

c2 2cq 2
(c q 2 )2

c2 2cq 2 + q 4
(c q 2 )2

4.

We find that the equality holds if and only if a = b = c.

INFINITY

99

Epsilon 13. Let f (x, y) = xy x3 + y 3 for x, y 0 with x + y = 2. Prove the inequality

1
1
1
1
f (x, y) f 1 + , 1
= f 1 ,1 +
.
3
3
3
3
First Solution. We write (x, y) = (1 + , 1 ) for some (1, 1). It follows that
`

f (x, y) = (1 + ) (1 ) (1 + )3 + (1 )3
`
`

=
1 2 62 + 2

2
1
8
= 6 2
+
3
3
8

1
1
= f 1 ,1
.
3
3

Second Solution. The AM-GM Inequality gives

`

2
f (x, y) = xy(x + y) (x + y)2 3xy = 2xy(4 4xy)
3

3xy + (4 3xy)
2

2
=

8
.
3

100

INFINITY

Epsilon 14. Let a, b 0 with a + b = 1. Prove that

p
p

a2 + b + a + b2 + 1 + ab 3.
Show that the equality holds if and only if (a, b) = (1, 0) or (a, b) = (0, 1).
First Solution. We may begin with the assumption a 12 b. The AM-GM Inequality
yields

2 + b 1 + (1 + ab) 2 1 + ab
with the equality b = 0. We next show that
p
3 + a 4 a2 a + 1
or

(3 + a)2 16 a2 a + 1

or
Since we have a
have

(15a 7)(1 a) 0.

,
1
,
the
inequality
clearly holds with the equality a = 1. Since we
2

a2 + b = a2 a + 1 = a + (1 a)2 = a + b2

we conclude that

p
p

2 a2 + b + 2 a + b2 + 2 1 + ab 3 + a + (2 + b) = 6.

INFINITY

101

Epsilon 15. (USA 1981) Let ABC be a triangle. Prove that

3 3
sin 3A + sin 3B + sin 3C
.
2
Solution. We observe that the sine function is not cocave on [0, 3] and that it is negative
on (, 2). Since the inequality is symmetric in the three variables, we may assume that
A B C. Observe that A + B + C = and that 3A, 3B, 3C [0, 3]. It is clear that
A 3 C.
We see that either 3B [2, 3) or 3C (0, ) is impossible. In the case when 3B
[, 2), we obtain the estimation

3 3
.
sin 3A + sin 3B + sin 3C 1 + 0 + 1 = 2 <
2
So, we may assume that 3B (0, ). Similarly, in the case when 3C [, 2], we obtain

3 3
sin 3A + sin 3B + sin 3C 1 + 1 + 0 = 2 <
.
2
Hence, we also assume 3C (2, 3). Now, our assmptions become A B < 13 and
2
< C. After the substitution = C 23 , the trigonometric inequality becomes
3

3 3
sin 3A + sin 3B + sin 3
.
2
Since 3A, 3B, 3 (0, ) and since the sine function is concave on [0, ], Jensens Inequality
gives

3A + 3B + 3
3A + 3B + 3C 2
sin 3A+sin 3B+sin 3 3 sin
= 3 sin
= 3 sin
.
3
3
3
`1 1 7
Under the assumption A B C, the equality occurs only when (A, B, C) = 9 , 9 , 9 .

102

INFINITY

Epsilon 16. (Chebyshevs Inequality) Let x1 , , xn and y1 , yn be two monotone increasing sequences of real numbers:
x1 xn , y1 yn .
Then, we have the estimation
n
X

1
xi yi
n
i=1

n
X
i=1

!
xi

n
X

!
yi

i=1

Proof. We observe that two sequences are similarly ordered in the sense that
(xi xj ) (yi yj ) 0
for all 1 i, j n. Now, the given inequality is an immediate consequence of the identity
!
!
n
n
n
1X
1 X
1 X
1 X
xi yi
xi
yi = 2
(xi xj ) (yi yj ) .
n i=1
n i=1
n i=1
n
1i,jn

INFINITY

103

Epsilon 17. (United Kingdom 2002) For all a, b, c (0, 1), show that

a
b
c
3 3 abc

+
+

.
1a
1b
1c
1 3 abc
First Solution. Since the inequality is symmetric in the three variables, we may assume
1
1
1
that a b c. Then, we have 1a
1b
1c
. By Chebyshevs Inequality, The
AM-HM Inequality and The AM-GM Inequality, we obtain

b
c
1
1
1
1
a
+
+

(a + b + c)
+
+
1a
1b
1c
3
1a
1b
1c

1
9

(a + b + c)
3
(1 a) + (1 b) + (1 c)

1
a+b+c
=
3 3 (a + b + c)

1
3 3 abc

3 3 3 3 abc

104

INFINITY

Epsilon 18. [IMO 1995/2 RUS] Let a, b, c be positive numbers such that abc = 1. Prove
that
1
1
1
3
+ 3
+ 3
.
a3 (b + c)
b (c + a)
c (a + b)
2
First Solution. After the substitution a = x1 , b = y1 , c = z1 , we get xyz = 1. The inequality
takes the form
x2
y2
z2
3
+
+
.
y+z
z+x
x+y
2
Since the inequality is symmetric in the three variables, we may assume that x y z.
1
1
1
Observe that x2 y 2 z 2 and y+z
z+x
x+y
. Chebyshevs Inequality and The
AM-HM Inequality offer the estimation

y2
z2
1` 2
1
1
x2
1
+
+

+
+
x + y2 + z2
y+z
z+x
x+y
3
y+z
z+x
x+y

1 2
9

x + y2 + z2
3
(y + z) + (z + x) + (x + y)
=

3 x2 + y 2 + z 2

.
2
x+y+z

Finally, we have x2 + y 2 + z 2 13 (x + y + z)2 (x + y + z) 3 xyz = x + y + z.

INFINITY

105

Epsilon 19. (Iran 1996) Let x, y, z be positive real numbers. Prove that

1
1
1
9
(xy + yz + zx)
+
+
.
(x + y)2
(y + z)2
(z + x)2
4
First Solution. [MEK1] We assume that x y z 0 and y > 0 (not excluding z = 0).
Let F denote the left hand side of the inequality. We define
A = (2x + 2y z)(x z)(y z) + z(x + y)2 ,
B = 14 x(x + y 2z)(11x + 11y + 2z),
C = (x + y)(x + z)(y + z),
D = (x + y + z)(x + y 2z) + x(y z) + y(z x) + (x y)2 ,
E = 14 (x + y)z(x + y + 2z)2 (x + y 2z)2 .
It can be verified that
1
C 2 (4F 9) = (x y)2 [(x + y)(A + B + C) + (x + z)(y + z)D] + E.
2
The right hand side is clearly nonnegative. It becomes an equality only for x = y = z and
for x = y > 0, z = 0.

106

INFINITY

Epsilon 20. (APMO 1991) Let a1 , , an , b1 , , bn be positive real numbers such that
a1 + + an = b1 + + bn . Show that
an 2
a1 + + an
a1 2
+ +

.
a1 + b 1
an + b n
2
First Solution. The key observation is the following identity:
n
n
X
ai 2
1 X ai 2 + b i 2
=
,
ai + b i
2 i=1 ai + bi
i=1
which is equivalent to

n
X
i=1

X bi 2
ai 2
=
,
ai + b i
ai + b i
i=1

which immediately follows from

n
n
n
n
n
n
X
X
X
X
X
X
ai 2
bi 2
ai 2 b i 2

=
=
(ai bi ) =
ai
bi = 0.
ai + b i
ai + bi
ai + b i
i=1
i=1
i=1
i=1
i=1
i=1
Our strategy is to establish the following symmetric inequality
n
a1 + + an + b 1 + + b n
1 X ai 2 + b i 2

.
2 i=1 ai + bi
4
It now remains to check the the auxiliary inequality
a+b
a2 + b 2

,
a+b
2
` 2

where a, b > 0. Indeed, we have 2 a + b2 (a + b)2 = (a b)2 0.

INFINITY

107

Epsilon 21. Let x, y, z be positive real numbers. Show the cyclic inequality
x
y
z
+
+
1.
2x + y
2y + z
2z + x
Solution. We first break the homogeneity. The original inequality can be rewritten as
1
1
1
+
+
1
2 + xy
2 + yz
2 + xz
The key idea is to employ the substitution
y
z
x
a= , b= , c= .
x
y
z
It follows that abc = 1. It now admits the symmetry in the variables:
1
1
1
+
+
1
2+a
2+b
2+c
Clearing denominators, it becomes
(2 + a)(2 + b) + (2 + b)(2 + c) + (2 + c)(2 + a) (2 + a)(2 + b)(2 + c)
or
12 + 4(a + b + c) + ab + bc + ca 8 + 4(a + b + c) + 2(ab + bc + ca) + 1
or
3 ab + bc + ca.
1

Applying The AM-GM Inequality, we obtain ab + bc + ca 3 (abc) 3 = 3.

108

INFINITY

Epsilon 22. Let x, y, z be positive real numbers with x + y + z = 3. Show the cyclic
inequality
x3
y3
z3
+
+
1.
x2 + xy + y 2
y 2 + yz + z 2
z 2 + zx + x2
Proof. We begin with the observation
x2

x3 y 3
y3 z3
z 3 x3
+ 2
+ 2
2
2
+ xy + y
y + yz + z
z + zx + x2
=

(x y) + (y z) + (z x)

or
x3
y3
z3
y3
z3
x3
+ 2
+ 2
= 2
+ 2
+ 2
.
2
2
2
2
2
+ xy + y
y + yz + z
z + zx + x
x + xy + y
y + yz + z
z + zx + x2
Our strategy is to establish the following symmetric inequality
x2

y3 + z3
z 3 + x3
x3 + y 3
+ 2
+ 2
2.
2
2
+ xy + y
y + yz + z
z + zx + x2
It now remains to check the the auxiliary inequality
x2

a3 + b 3
a+b

,
a2 + ab + b2
3
where a, b > 0. Indeed, we obtain the equality
`

3 a3 + b3 (a + b) a2 + ab + b2 = 2(a + b)(a b)2 .

We now conclude that
y3 + z3
z 3 + x3
x+y
y+z
z+x
x3 + y 3
+ 2
+ 2

+
+
= 2.
2
2
2
x + xy + y
y + yz + z
z + zx + x2
3
3
3

INFINITY

109

Epsilon 23. [SL 1985 CAN] Let x, y, z be positive real numbers. Show the cyclic inequality
x2

x2
y2
z2
+ 2
+ 2
2.
+ yz
y + zx
z + xy

First Solution. We first break the homogeneity. The original inequality can be rewritten
as
1
1
1
+
+
xy 2
1 + xyz2
1 + zx
1
+
2
y
z2
The key idea is to employ the substitution
yz
zx
z2
.
, b= 2, c=
2
x
y
xy
It then follows that abc = 1. It now admits the symmetry in the variables:
1
1
1
+
+
2
1+a
1+b
1+c
Since it is symmetric in the three variables, we may break the symmetry. Lets assume
1
a b, c. Since it is obvious that 1+a
< 1, it is enough to check the estimation
a=

1
1
+
1
1+b
1+c
or equivalently

2+b+c
1
1 + b + c + bc

or equivalently
bc 1.
However, it follows from abc = 1 and from a b, c that a 1 and so that bc 1.

110

INFINITY

Epsilon 24. [SL 1990 THA] Let a, b, c, d 0 with ab + bc + cd + da = 1. show that

a3
b3
c3
d3
1
+
+
+
.
b+c+d
c+d+a
d+a+b
a+b+c
3
Solution. Since the constraint ab + bc + cd + da = 1 is not symmetric in the variables, we
cannot consider the case when a b c d only. We first make the observation that
a2 + b 2
b2 + c2
c2 + d 2
d2 + a2
+
+
+
ab + bc + cd + da = 1.
2
2
2
2
Our strategy is to establish the following result. It is symmetric.
Let a, b, c, d 0 with a2 + b2 + c2 + d2 1. Then, we obtain
a2 + b2 + c2 + d2 =

b3
c3
d3
1
a3
+
+
+
.
b+c+d
c+d+a
d+a+b
a+b+c
3
We now exploit the symmetry! Since everything is symmetric in the variables, we may assume that a b c d. Two applications of Chebyshevs Inequality and one application
of The AM-GM Inequality yield

a3
b3
c3
d3
+
+
+
b+c+d
c+d+a
d+a+b
a+b+c

1` 3
1
1
1
1
3
3
3
a +b +c +d
+
+
+
4
b+c+d
c+d+a
d+a+b
a+b+c
2

1` 3
4
a + b3 + c3 + d3
4
(b + c + d) + (c + d + a) + (d + a + b) + (a + b + c)
`

1
42
a2 + b2 + c2 + d2 (a + b + c + d)
2
4
3(a + b + c + d)
1
.
3

INFINITY

111

Epsilon 25. [IMO 2000/2 USA] Let a, b, c be positive numbers such that abc = 1. Prove
that

1
1
1
a1+
b1+
c1+
1.
b
c
a
First Solution. Since abc = 1, we can make the substitution a = xy , b = yz , c = xz for some
positive real numbers x, y, z.15 Then, it becomes a well-known symmetric inequality:

x
z y
xz
y
1+
1+
1+
1
y
y
z
z
x
x
or
xyz (y + z x)(z + x y)(x + y z).

15

For example, take x = 1, y =

1
,
a

z=

1
.
ab

112

INFINITY

Epsilon 26. [IMO 1983/6 USA] Let a, b, c be the lengths of the sides of a triangle. Prove
that
a2 b(a b) + b2 c(b c) + c2 a(c a) 0.
First Solution. After setting a = y + z, b = z + x, c = x + y for x, y, z > 0, it becomes
x3 z + y 3 x + z 3 y x2 yz + xy 2 z + xyz 2
or

x2
y2
z2
+
+
x + y + z.
y
z
x
However, an application of The Cauchy-Schwarz Inequality gives
2

x
y2
z2
+
+
(y + z + x)
(x + y + z)2 .
y
z
x

INFINITY

113

Epsilon 27. [IMO 1961/2 POL] (Weitzenb

ocks Inequality) Let a, b, c be the lengths of a
triangle with area S. Show that

a2 + b2 + c2 4 3S.
First Solution. Write a = y + z, b = z + x, c = x + y for x, y, z > 0. Its equivalent to
((y + z)2 + (z + x)2 + (x + y)2 )2 48(x + y + z)xyz,
which can be obtained as following :
((y + z)2 + (z + x)2 + (x + y)2 )2 16(yz + zx + xy)2 16 3(xy yz + yz zx + xy yz).
Here, we used the well-known inequalities p2 +q 2 2pq and (p+q+r)2 3(pq+qr+rp).

114

INFINITY

Epsilon 28. (Hadwiger-Finsler Inequality) For any triangle ABC with sides a, b, c and area
F , the following inequality holds.

2ab + 2bc + 2ca (a2 + b2 + c2 ) 4 3F.

First Proof. After the substitution a = y + z, b = z + x, c = x + y, where x, y, z > 0, it
becomes
p
xy + yz + zx 3xyz(x + y + z),
which follows from the identity
(xy + yz + zx)2 3xyz(x + y + z) =

(xy yz)2 + (yz zx)2 + (zx xy)2

.
2

Second Proof. We now present a convexity proof. It is easy to deduce

2ab + 2bc + 2ca (a2 + b2 + c2)
A
B
C
+ tan + tan
=
.
2
2
2
4F
`
Since the function tan x is convex on 0, 2 , Jensens Inequality implies that
!
A

+ B2 + C2
2ab + 2bc + 2ca (a2 + b2 + c2)
2
3 tan
= 3.
4F
3
tan

INFINITY

115

Epsilon 29. (Tsintsifas) Let p, q, r be positive real numbers and let a, b, c denote the sides
of a triangle with area F . Then, we have

p
q 2
r 2
a2 +
b +
c 2 3F.
q+r
r+p
p+q
Proof. (V. Pambuccian) By Hadwiger-Finsler Inequality, it suffices to show that
p
q 2
r 2
1
a2 +
b +
c (a + b + c)2 (a2 + b2 + c2 )
q+r
r+p
p+q
2
or

p+q+r
p+q+r
p+q+r
1
2
2
a +
b +
c2 (a + b + c)2
q+r
r+p
p+q
2
or

1
1 2
1 2
((q + r) + (r + p) + (p + q))
a2 +
b +
c
(a + b + c)2 .
q+r
r+p
p+q
However, this is a straightforward consequence of The Cauchy-Schwarz Inequality.

116

INFINITY

Epsilon 30. (The Neuberg-Pedoe Inequality) Let a1 , b1 , c1 denote the sides of the triangle
A1 B1 C1 with area F1 . Let a2 , b2 , c2 denote the sides of the triangle A2 B2 C2 with area F2 .
Then, we have
a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 ) 16F1 F2 .
First Proof. ([LC1], Carlitz) We begin with the following lemma.
Lemma 8.1. We have
a1 2 (a2 2 + b2 2 c2 2 ) + b1 2 (b2 2 + c2 2 a2 2 ) + c1 2 (c2 2 + a2 2 b2 2 ) > 0.
Proof. Observe that its equivalent to
(a1 2 + b1 2 + c1 2 )(a2 2 + b2 2 + c2 2 ) > 2(a1 2 a2 2 + b1 2 b2 2 + c1 2 c2 2 ).
From Herons Formula, we find that, for i = 1, 2,

16Fi 2 = (ai 2 +bi 2 +ci 2 )2 2(ai 4 +bi 4 +ci 4 ) > 0 or ai 2 +bi 2 +ci 2 >

2(ai 4 + bi 4 + ci 4 ) .

The Cauchy-Schwarz Inequality implies that

q
(a1 2 +b1 2 +c1 2 )(a2 2 +b2 2 +c2 2 ) > 2 (a1 4 + b1 4 + c1 4 )(a2 4 + b2 4 + c2 4 ) 2(a1 2 a2 2 +b1 2 b2 2 +c1 2 c2 2 ).

By the lemma, we obtain

L = a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 ) > 0,
Hence, we need to show that
L2 (16F1 2 )(16F2 2 ) 0.
One may easily check the following identity
L2 (16F1 2 )(16F2 2 ) = 4(U V + V W + W U ),
where
U = b1 2 c2 2 b2 2 c1 2 , V = c1 2 a2 2 c2 2 a1 2 and W = a1 2 b2 2 a2 2 b1 2 .
Using the identity
a1 2 U + b1 2 V + c1 2 W = 0 or W =
one may also deduce that
UV + V W + W U =

a1 2
c1 2

c1 2 a1 2 b1 2
V
2a1 2

It follows that
UV + V W + W U =

a1 2
c1 2

a1 2
b1 2
U 2 V,
2
c1
c1

4a1 2 b1 2 (c1 2 a1 2 b1 2 )2 2
V .
4a1 2 c1 2

c 1 2 a1 2 b 1 2
V
2a1 2

16F1 2 2
V 0.
4a1 2 c1 2

Second Proof. ([LC2], Carlitz) We rewrite it in terms of a1 , b1 , c1 , a2 , b2 , c2 :

r

(a1 2 + b1 2 + c1 2 )(a2 2 + b2 2 + c2 2 ) 2(a1 2 a2 2 + b1 2 b2 2 + c1 2 c2 2 )

`

`
2
2
a1 2 + b1 2 + c1 2 2(a1 4 + b1 4 + c1 4 )
a2 2 + b2 2 + c2 2 2(a2 4 + b2 4 + c2 4 ) .

We employ the following substitutions

2 c1 2 ,

y1 = a2 2 + b2 2 + c2 2 , y2 = 2 a2 2 , y3 = 2 b2 2 , y4 = 2 c2 2 .

x1 = a1 2 + b1 2 + c1 2 , x2 =

2 a1 2 , x 3 =

2 b1 2 , x 4 =

INFINITY

117

We now observe
x1 2 > x2 2 + y3 2 + x4 2 and y1 2 > y2 2 + y3 2 + y4 2 .
We now apply Aczels inequality to get the inequality
p
x1 y1 x2 y2 x3 y3 x4 y4 (x1 2 (x2 2 + y3 2 + x4 2 )) (y1 2 (y2 2 + y3 2 + y4 2 )).

118

INFINITY

Epsilon 31. (Aczels Inequality) If a1 , , an , b1 , , bn > 0 satisfies the inequality

a1 2 a2 2 + + an 2 and b1 2 b2 2 + + bn 2 ,
then the following inequality holds.
q

`
`
a1 b1 (a2 b2 + + an bn ) (a1 2 (a2 2 + + an 2 )) b1 2 b2 2 + + bn 2
Proof. [MV] The Cauchy-Schwarz Inequality shows that
q
a1 b1 (a2 2 + + an 2 )(b2 2 + + bn 2 ) a2 b2 + + an bn .
Then, the above inequality is equivalent to
`
`
` 2 ` 2

(a1 b1 (a2 b2 + + an bn ))2 a1 2 a2 2 + + an 2

b1 b2 + + bn 2 .
In case a1 2 (a2 2 + + an 2 ) = 0, its trivial. Hence, we now assume that a1 2 (a2 2 +
+ an 2 ) > 0. The main trick is to think of the following quadratic polynomial
!
!
!
n
n
n
n
X
X
X
X
P(x) = (a1 xb1 )2 (ai xbi )2 = a1 2
ai 2 x2 +2 a1 b1
ai bi x+ b1 2
bi 2 .
i=2

i=2

We now observe that

b1
a1

n
X
i=2

i=2

ai

b1
a1

i=2

2
bi

Since P ab11 0 and since the coefficient of x2 in the quadratic polynomial P is positive,
P should have at least one real root. Therefore, P has nonnegative discriminant. It follows
that
!!2
!
!
n
n
n
X
X
X
2
2
2
2
2 a1 b 1
ai b i
4 a1
ai
b1
bi
0.
i=2

i=2

i=2

INFINITY

119

Epsilon 32. If A, B, C, X, Y , Z denote the magnitudes of the corresponding angles of

triangles ABC, and XY Z, respectively, then
cot A cot Y + cot A cot Z + cot B cot Z + cot B cot X + cot C cot X + cot C cot Y 2.
Proof. By the Cosine Law in triangle ABC, we have cos A = (b2 + c2 a2 )/2bc. On the
other hand, since sin A = 2S/bc, we deduce that
b 2 + c 2 a2
b2 + c2 a2 2S
:
=
.
2bc
bc
4S
2
2
2
Analogously, we have that cot B = (c + a b )/4S, and so,
cot A = cos A : sin A =

b 2 + c 2 a2
c2 + a2 b2
2c2
c2
+
=
=
.
4S
4S
4S
2S
Now since cot Z = (x2 + y 2 z 2 )/4T , it follows that
cot A + cot B =

cot A cot Z + cot B cot Z = (cot A + cot B) cot Z

=
=

c2 x2 + y 2 z 2

2S`
4T

2
c x2 + y 2 z 2
.
8ST

Similarly, we obtain

a2 y 2 + z 2 x2
,
8ST
and
`

b2 z 2 + x2 y 2
cot C cot Y + cot A cot Y =
.
8ST
Hence, we conclude that
cot B cot X + cot C cot X =

cot A cot Y + cot A cot Z + cot B cot Z + cot B cot X + cot C cot X + cot C cot Y
=
=
=

(cot B cot X + cot C cot X) + (

`

`
a2 y 2 + z 2 x2
b2 z 2 + x
+
8ST
` 8ST

`
a2 y 2 + z 2 x2 + b2 z 2 + x2
8ST

From the Neuberg-Pedoe Inequality, we have

`

a2 y 2 + z 2 x2 + b2 z 2 + x2 y 2 + c2 x2 + y 2 z 2 16ST,
and so
cot A cot Y + cot A cot Z + cot B cot Z + cot B cot X + cot C cot X + cot C cot Y 2,
with equality if and only if the triangles ABC and XY Z are similar.

120

INFINITY

Epsilon 33. (Vasile C

artoaje) Let a, b, c, x, y, z be nonnegative reals. Prove the inequality
(ay + az + bz + bx + cx + cy)2 4 (bc + ca + ab) (yz + zx + xy) ,
with equality if and only if a : x = b : y = c : z.
Proof. According to the Conway substitution theorem,
since a, b, c are nonnegative reals,

there exists a triangle ABC with area S = 12 bc + ca + ab and with its angles A, B,
a
b
c
C satisfying cot A = 2S
, cot B = 2S
, cot C = 2S
(note that we cannot denote the
sidelengths of triangle ABC by a, b, c here, since a, b, c already stand for something
different). Similarly, since x, y, z are nonnegative reals, there exists a triangle XY Z with

y
x
area T = 12 yz + zx + xy and with its angles X, Y , Z satisfying cot X = 2T
, cot Y = 2T
,
z
cot Z = 2T
. Now, by Epsilon 32, we have
cot A cot Y + cot A cot Z + cot B cot Z + cot B cot X + cot C cot X + cot C cot Y 2,
which rewrites as
y
a
z
b
z
b
x
c
x
c
y
a

2,
2S 2T
2S 2T
2S 2T
2S 2T
2S 2T
2T 2T
and thus,
ay + az + bz + bx + cx + cy

=
=

2 2S 2T
1
1
22
bc + ca + ab 2
yz + zx + xy
2
2
p
2 (bc + ca + ab) (yz + zx + xy).

Upon squaring, this becomes

(ay + az + bz + bx + cx + cy)2 4 (bc + ca + ab) (yz + zx + xy) .

INFINITY

121

Epsilon 34. (Walter Janous, Crux Mathematicorum) If u, v, w, x, y, z are six reals such
that the terms y + z, z + x, x + y, v + w, w + u, u + v, and vw + wu + uv are all nonnegative,
then
p
x
y
z
(v + w) +
(w + u) +
(u + v) 3 (vw + wu + uv).
y+z
z+x
x+y
Proof. According to the Conway substitution theorem, since the reals v + w, w + u, u + v
and vw
there exists a triangle
ABC with sidelengths
+ wu + uvare all nonnegative,

a = v + w, b = w + u, c = u + v and area S = 12 vw + wu + uv. Applying the

Extended Tsintsifas Inequality to this triangle ABC and to the reals x, y, z satisfying the
condition that the reals y + z, z + x, x + y are all positive, we obtain

x
y
z
a2 +
b2 +
c2 2 3S,
y+z
z+x
x+y
which rewrites as
1
`
2
`
2
`
2
x
y
z

v+w +

w+u +

u+v 2 3
vw + wu + uv,
y+z
z+x
x+y
2
and thus,
p
y
z
x
(v + w) +
(w + u) +
(u + v) 3 (vw + wu + uv).
y+z
z+x
x+y

122

INFINITY

Epsilon 35. (Tran Quang Hung) In any triangle ABC with sidelengths a, b, c, circumradius
R, inradius r, and area S, we have that
X

A X
B
C
a2 + b2 + c2 4S 3 + (a b)2 + (b c)2 + (c a)2 + 16Rr
cos2
cos cos
.
2
2
2
Proof. We know that Hadwiger-Finslers Inequality states that

x2 + y 2 + z 2 4T 3 (x y)2 + (y z)2 + (z x)2 ,

for any triangle XY Z with sidelengths x, y, z, and area T . Let us apply this for the
triangle XY Z = Ia Ib Ic , where X = Ia , Y = Ib , Z = Ic are the excenters of ABC. In this
case, it is well-known that
A
B
C
x = 4R cos , y = 4R cos , z = 4R cos , T = 2sR,
2
2
2
where s is the semiperimeter of triangle ABC. Therefore,

A
B
C
16R2 cos2 + cos2
+ cos2
8s 3
2
2
2
"
2
2
2 #
2 B
2 C
2 A
2
2 A
2 B
2 C
16R
cos
cos
+ cos
cos
+ cos
cos
,
2
2
2
2
2
2
which according to the well-known formulas
r
r
r
s(s a)
s(s b)
s(s c)
A
B
C
cos =
, cos
=
, cos
=
,
2
bc
2
ca
2
ab
easily reduces to
X

A X
B
C
a2 + b2 + c2 4S 3 + (a b)2 + (b c)2 + (c a)2 + 16Rr
cos2
cos cos
.
2
2
2

INFINITY

Epsilon 36. For all R, we have

sin (3) = 4 sin sin

123

2
+ sin
+ .
3
3

Proof. It follows that

sin (3)

=
=
=
=

3 sin 4 sin3
`

sin 3 cos2 sin2

1
1
3
3
4 sin
cos + sin
cos sin
2
2
2
2

2
4 sin sin
+ sin
+ .
3
3

124

INFINITY

Epsilon 37. For all A, B, C R with A + B + C = 2, we have

cos2 A + cos2 B + cos2 C + 2 cos A cos B cos C = 1.
Proof. Our job is to show that the quadratic eqaution
t2 + (2 cos B cos C) t + cos2 B + cos2 C 1 = 0
has a root t = cos A. We find that it admits roots
p
2 cos B cos C 4 cos2 B cos2 C 4 (cos2 B + cos2 C 1)
t =
2
p
= cos B cos C (1 cos2 B) (1 cos2 C)
=

cos B cos C | sin B sin C|.

Since we have
cos B cos C + sin B sin C = cos(B + C) = cos ( A) = cos A,
we find that t = cos A satisfies the quadratic equation, as desired.

INFINITY

125

Epsilon 38. [SL 2005 KOR] In an acute triangle ABC, let D, E, F , P , Q, R be the feet
of perpendiculars from A, B, C, A, B, C to BC, CA, AB, EF , F D, DE, respectively.
Prove that
p(ABC)p(P QR) p(DEF )2 ,
where p(T ) denotes the perimeter of triangle T .
Solution. Lets euler this problem. Let be the circumradius of the triangle ABC. Its
easy to show that BC = 2 sin A and EF = 2 sin A cos A. Since DQ = 2 sin C cos B cos A,
DR = 2 sin B cos C cos A, and \F DE = 2A, the Cosine Law gives us
QR2

=
=

DQ2 + DR2 2DQ DR cos( 2A)

42 cos2 A (sin C cos B)2 + (sin B cos C)2 + 2 sin C cos B sin B cos C cos(2A)

or

p
QR = 2 cos A f (A, B, C),

where
f (A, B, C) = (sin C cos B)2 + (sin B cos C)2 + 2 sin C cos B sin B cos C cos(2A).
So, what we need to attack is the following inequality:
1 0
12
0
10
X
X
X
p
@
2 sin A cos AA
2 sin AA @
2 cos A f (A, B, C)A @
cyclic

or

10

0
@

cyclic

cyclic

sin AA @

12
1 0
X
p
cos A f (A, B, C)A @
sin A cos AA .

cyclic

cyclic

cyclic

Our job is now to find a reasonable lower bound of f (A, B, C). Once again, we express
f (A, B, C) as the sum of two squares. We observe that
f (A, B, C)

(sin C cos B)2 + (sin B cos C)2 + 2 sin C cos B sin B cos C cos(2A)

(sin C cos B + sin B cos C)2 + 2 sin C cos B sin B cos C [1 + cos(2A)]

sin2 (C + B) 2 sin C cos B sin B cos C 2 sin2 A

sin2 A [1 4 sin B sin C cos B cos C] .

So, we shall express

` 1 4 sin B sin C
` cos B cos C as the sum of two squares. The trick is
to replace 1 with sin2 B + cos2 B sin2 C + cos2 C . Indeed, we get
` 2
`

1 4 sin B sin C cos B cos C =

sin B + cos2 B sin2 C + cos2 C 4 sin B sin C cos B cos C
=

(sin B cos C sin C cos B)2 + (cos B cos C sin B sin C)2

sin2 (B C) + cos2 (B + C)

sin2 (B C) + cos2 A.

It therefore follows that

f (A, B, C) = sin2 A sin2 (B C) + cos2 A sin2 A cos2 A

so that

cos A

f (A, B, C)

cyclic

sin A cos2 A.

cyclic

So, we can complete the proof if we establish that

0
10
1 0
12
X
X
X
2
@
sin AA @
sin A cos AA @
sin A cos AA .
cyclic

cyclic

cyclic

126

INFINITY

Indeed, one sees that its a direct consequence of The Cauchy-Schwarz Inequality

(p + q + r)(x + y + z) ( px + qy + rz)2 ,
where p, q, r, x, y and z are positive real numbers.

Remark 8.1. Alternatively, one may obtain another lower bound of f (A, B, C):
f (A, B, C)

(sin C cos B)2 + (sin B cos C)2 + 2 sin C cos B sin B cos C cos(2A)

(sin C cos B sin B cos C)2 + 2 sin C cos B sin B cos C [1 + cos(2A)]
sin(2B) sin(2C)
sin2 (B C) + 2

2 cos2 A
2
2
cos2 A sin(2B) sin(2C).

Then, we can use this to offer a lower bound of the perimeter of triangle P QR:
X
X
p

p(P QR) =
2 cos A f (A, B, C)
2 cos2 A sin 2B sin 2C
cyclic

cyclic

So, one may consider the following inequality:

X

p(ABC)
2 cos2 A sin 2B sin 2C p(DEF )2
cyclic

or

10

@2

sin AA @

cyclic

or

0
@

1 0
12
X

2 cos A sin 2B sin 2C A @2

sin A cos AA .
2

cyclic

cyclic

10
X
cyclic

sin AA @

1
X
cyclic

0
X

cos2 A sin 2B sin 2C A @

cyclic

However, it turned out that this doesnt hold. Disprove this!

12
sin A cos AA .

INFINITY

127

Epsilon 39. [IMO 2001/1 KOR] Let ABC be an acute-angled triangle with O as its
circumcenter. Let P on line BC be the foot of the altitude from A. Assume that \BCA
\ABC + 30 . Prove that \CAB + \COP < 90 .
Solution. The angle inequality \CAB +\COP < 90 can be written as \COP < \P CO.
This can be shown if we establish the length inequality OP > P C. Since the power of
P with respect to the circumcircle of ABC is OP 2 = R2 BP P C, where R is the
circumradius of the triangle ABC, it becomes R2 BP P C > P C 2 or R2 > BC P C.
We euler this. Its an easy job to get BC = 2R sin A and P C = 2R sin B cos C. Hence,
we show the inequality R2 > 2R sin A 2R sin B cos C or sin A sin B cos C < 14 . Since
sin A < 1, it suffices to show that sin A sin B cos C < 41 . Finally, we use the angle condition
\C \B + 30 to obtain the trigonometric inequality
sin(B + C) sin(C B)
1 sin(C B)
1 sin 30
1
sin B cos C =

= .
2
2
2
4

128

INFINITY

Epsilon 40. [IMO 1961/2 POL] (Weitzenb

ocks Inequality) Let a, b, c be the lengths of a
triangle with area S. Show that

a2 + b2 + c2 4 3S.
Second Proof. [AE, p.171] Let ABC be a triangle with sides BC = a, CA = b and AB = c.
After taking the point P on the same side of BC as the vertex A so that 4P BC is
equilateral, we use The Cosine Law to deduce the geometric identity

AP 2 = b2 + c2 2bc cos C
6

2
2
= b + c 2bc cos C
6

= b2 + c2 bc cos C 3bc sin C

b2 + c2 a2
= b2 + c2
2 3K
2
which implies the geometric inequality
b2 + c2
or equivalently

b 2 + c 2 a2
2 3K
2

a2 + b2 + c2 4 3S.

INFINITY

129

Epsilon 41. (The Neuberg-Pedoe Inequality) Let a1 , b1 , c1 denote the sides of the triangle
A1 B1 C1 with area F1 . Let a2 , b2 , c2 denote the sides of the triangle A2 B2 C2 with area F2 .
Then, we have
a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 ) 16F1 F2 .
Third Proof. [DP2] We take the point P on the same side of B1 C1 as the vertex A1 so that
4P B1 C1 4A2 B2 C2 . Now, we use The Cosine Law to deduce the geometric identity
a2 2 A 1 P 2
=

a2 2 b1 2 + b2 2 a1 2 2a1 a2 b1 b2 cos |C1 C2 |

a2 2 b1 2 + b2 2 a1 2 2a1 a2 b1 b2 cos (C1 C2 )

1
1
1
2 2
2
2
= a2 b1 + b2 a1 (2a1 b1 cos C1 ) (2a2 b2 cos C2 ) 8
a1 b1 sin C1
a2 b2 sin C2
2
2
2
`

1` 2
= a2 2 b 1 2 + b 2 2 a1 2
a1 + b1 2 c1 2 a1 2 + b1 2 c1 2 8F1 F2 ,
2
which implies the geometric inequality
`

1` 2
a2 2 b 1 2 + b 2 2 a1 2
a1 + b1 2 c1 2 a1 2 + b1 2 c1 2 8F1 F2
2
or equivalently
=

a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 ) 16F1 F2 .

130

INFINITY

Epsilon 42. (Barrows Inequality) Let P be an interior point of a triangle ABC and let
U , V , W be the points where the bisectors of angles BP C, CP A, AP B cut the sides
BC,CA,AB respectively. Then, we have
P A + P B + P C 2(P U + P V + P W ).
Proof. ([MB] and [AK]) Let d1 = P A, d2 = P B, d3 = P C, l1 = P U , l2 = P V , l3 = P W ,
21 = \BP C, 22 = \CP A, and 23 = \AP B. We need to show that d1 + d2 + d3
2(l1 + l2 + l3 ). Its easy to deduce the following identities
2d2 d3
2d3 d1
2d1 d2
l1 =
cos 1 , l2 =
cos 2 , and l3 =
cos 3 ,
d2 + d3
d3 + d1
d1 + d2
It now follows that

1
l1 + l2 + l3 d2 d3 cos 1 + d3 d1 cos 2 + d1 d2 cos 3 (d1 + d2 + d3 ) .
2

INFINITY

131

Epsilon 43. ([AK], Abi-Khuzam) Let x1 , , x4 be positive real numbers. Let 1 , , 4

be real numbers such that 1 + + 4 = . Then, we have
r
(x1 x2 + x3 x4 )(x1 x3 + x2 x4 )(x1 x4 + x2 x3 )
x1 cos 1 + x2 cos 2 + x3 cos 3 + x4 cos 4
.
x1 x2 x3 x4
q
2
2
2
2
2
4
3 x4
+ x32x+x
q = x1 x2 +x
and = pq . In the view of 1 + 2 +
Proof. Let p = x12x+x
2
1 x2
3 x4
(3 + 4 ) = and 3 + 4 + (1 + 2 ) = , we have
x1 cos 1 + x2 cos 2 + cos(3 + 4 ) p =
and

pq,

q
= pq.

Since cos(3 + 4 ) + cos(1 + 2 ) = 0, adding these two above inequalities yields

r

(x1 x2 + x3 x4 )(x1 x3 + x2 x4 )(x1 x4 + x2 x3 )

.
x1 cos 1 +x2 cos 2 +x3 cos 3 +x4 cos 4 2 pq =
x1 x2 x3 x4

x3 cos 3 + x4 cos 4 + cos(1 + 2 )

132

INFINITY

Epsilon 44. [IMO 1991/5 FRA] Let ABC be a triangle and P an interior point in ABC.
Show that at least one of the angles \P AB, \P BC, \P CA is less than or equal to 30 .
First Proof. Set A1 = A, A2 = B, A3 = C, A4 = A and write \P Ai Ai+1 = i . Let
H1 , H2 , H3 denote the feet of perpendiculars from P to the sides BC, CA, AB, respectively. Now, we assume to the contrary that 1 , 2 , 3 > 6 . Since the angle sum of a
triangle is 180 , it is immediate that 1 , 2 , 3 < 5
. Hence,
6
1
P Hi
= sin i > ,
P Ai+1
2
for all i = 1, 2, 3. We now find that
2 (P H1 + P H2 + P H3 ) > P A2 + P A3 + P A1 ,
which contradicts for The Erd
os-Mordell Theorem.

INFINITY

133

Epsilon 45. Any triangle has the same Brocard angles.

Proof. More strongly, we show that the isogonal conjugate of the first Brocard point
is the second Brocard point. Let 1 , 2 denote the Brocard points of a triangle ABC,
respectively. Let 1 , 2 be the corresponding Brocard angles. Take the isogonal conjugate
point of 1 . Then, by the definition of isogonal conjugate point, we find that
\BA = \2 CB = \2 AC = 1 .
Hence, we see that the interior point is the the second Brocard point of ABC. By the
uniqueness of the second Brocard point of ABC, we see that = 2 and that 1 = 2 .

134

INFINITY

Epsilon 46. The Brocard angle of the triangle ABC satisfies

cot = cot A + cot B + cot C.
Proof. Let denote the first Brocard point of ABC. We only prove it in the case when
ABC is acute. Let AH, P Q denote the altitude from A, Q, respectively. Both angles
\B and \C are acute, the point H lies on the interior side of BC. Let P 6= be the
intersubsection point of the circumcircle of triangle CA with ray B. Since \AP B =
\AP = \AC = = \BC = \P BC, we find that AP is parallel to BC so that
AH = P Q. Since \A is acute or since \P CB = \P BA + \C = \B + \C = 180 \A
is obtuse, we see that the point H lies on the outside of side BC. Since the four points
B, H, C, Q are collinear in this order, we have BQ = BH + HC + CQ. It thus follows that
BQ
BH
HC
CQ
cot =
=
+
+
= cot A + cot B + cot C.
PQ
AH
AH
PQ

INFINITY

135

Epsilon 47. (The Trigonometric Version of Cevas Theorem) For an interior point P of a
triangle A1 A2 A3 , we write
\A3 A1 A2 = 1 , \P A1 A2 = 1 , \P A1 A3 = 1 ,
\A1 A2 A3 = 2 , \P A2 A3 = 2 , \P A2 A1 = 2 ,
\A2 A3 A1 = 3 , \P A3 A1 = 3 , \P A3 A2 = 3 .
Then, we find a hidden symmetry:
sin 1 sin 2 sin 3

=1
sin 1 sin 2 sin 3
or equivalently
1
= [cot 1 cot 1 ] [cot 2 cot 2 ] [cot 3 cot 3 ] .
sin 1 sin 2 sin 3
Proof. Applying The Sine Law, we have
sin 1
P A2 sin 2
P A3 sin 3
P A1
=
,
=
,
=
.
sin 1
P A1 sin 2
P A2 sin 3
P A3
It follows that
P A2 P A3 P A1
sin 1 sin 2 sin 3

= 1.
sin 1 sin 2 sin 3
P A1 P A2 P A3
We now observe that, for i = 1, 2, 3,
sin (i i )
cos i
cos i
sin i
cot i cot i =

=
=
.
sin i
sin i
sin i sin i
sin i sin i
It therefore follows that

=
=
=

[cot 1 cot 1 ] [cot 2 cot 2 ] [cot 3 cot 3 ]

sin 2
sin 3
sin 1

sin 1 sin 1 sin 2 sin 2 sin 3 sin 3

1
sin 1 sin 2 sin 3

sin 1 sin 2 sin 3 sin 1 sin 2 sin 3

1
.
sin 1 sin 2 sin 3

136

INFINITY

Epsilon 48. Let P be an interior point of a triangle ABC. Show that

cot (\P AB) + cot (\P BC) + cot (\P CA) 3 3.

Proof. Set A1 = A, A2 = B, A3 = C, A4 = A and write \Ai = i and \P Ai Ai+1 = i
for i = 1, 2, 3. Our job is to establish the inequality

cot 1 + cot 2 + cot 3 3 3.

We begin with The Trigonometric Version of Cevas Theorem
1
= [cot 1 cot 1 ] [cot 2 cot 2 ] [cot 3 cot 3 ] .
sin 1 sin 2 sin 3
We first apply The AM-GM Inequality and Jensens Inequality to deduce
3

+ + 3 3
sin 1 + sin 2 + sin 3
1
2
3
sin 1 sin 2 sin 3
sin3
=
3
3
2
or

3
2

[cot 1 cot 1 ] [cot 2 cot 2 ] [cot 3 cot 3 ] .

3
Since i (0, i ), the monotonicity of the cotangent function shows that cot i cot i
is positive. Hence, by The AM-GM Inequality, the above inequality guarantees that
p
2

3 [cot 1 cot 1 ] [cot 2 cot 2 ] [cot 3 cot 3 ]

3
[cot 1 cot 1 ] + [cot 2 cot 2 ] + [cot 3 cot 3 ]

3
[cot 1 + cot 2 + cot 3 ] [cot 1 + cot 2 + cot 3 ]
=
3
or

cot 1 + cot 2 + cot 3 cot 1 + cot 2 + cot 3 + 2 3.

Since we know cot 1 + cot 2 + cot 3 3, we get the desired inequality.

INFINITY

137

Epsilon 49. [IMO 1961/2 POL] (Weitzenb

ocks Inequality) Let a, b, c be the lengths of a
triangle with area S. Show that

a2 + b2 + c2 4 3S.
Fourth Proof. ([RW], R. Weitzenb
ock) Let ABC be a triangle with sides a, b, and
` c. To
euler it, we toss the picture on the real plane R2 with the coordinates A(, ), B a2 , 0

`
and C a2 , 0 . Now, we obtain

2
` 2
2 2
3 2 ` 2
a + 2
+ 162 2 0.
a + b2 + c2 4 3S =
2

138

INFINITY

Epsilon 50. (The Neuberg-Pedoe Inequality) Let a1 , b1 , c1 denote the sides of the triangle
A1 B1 C1 with area F1 . Let a2 , b2 , c2 denote the sides of the triangle A2 B2 C2 with area F2 .
Then, we have
a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 ) 16F1 F2 .
Fourth Proof. (By a participant from KMO16 summer program.) We toss 4A1 B1 C1 and
4A2 B2 C2 onto the real plane R2 :
A1 (0, p1 ), B1 (p2 , 0), C1 (p3 , 0), A2 (0, q1 ), B2 (q2 , 0), and C2 (q3 , 0).
It therefore follows from the inequality x2 + y 2 2|xy| that
a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 )
=

(p3 p2 )2 (2q1 2 + 2q1 q2 ) + (p1 2 + p3 2 )(2q2 2 2q2 q3 ) + (p1 2 + p2 2 )(2q3 2 2q2 q3 )

2(p3 p2 )2 q1 2 + 2(q3 q2 )2 p1 2 + 2(p3 q2 p2 q3 )2

2((p3 p2 )q1 )2 + 2((q3 q2 )p1 )2

4|(p3 p2 )q1 | |(q3 q2 )p1 |

16F1 F2 .

16

Korean Mathematical Olympiads

INFINITY

139

Epsilon 51. (USA 2003) Let ABC be a triangle. A circle passing through A and B
intersects the segments AC and BC at D and E, respectively. Lines AB and DE intersect
at F , while lines BD and CF intersect at M . Prove that M F = M C if and only if
M B M D = M C2.
Solution. (Darij Grinberg) By Cevas theorem, applied to the triangle BCF and the concurrent cevians BM , CA and F E (in fact, these cevians concur at the point D), we
have
M F EC AB

= 1.
M C EB AF
MF
AF
EB
AF
EC
AF
EC
Hence, M C = AB EC = AB : EB . Thus, M F = M C holds if and only if AB
= EB
.
EC
AF
But by Thales theorem, AB = EB is equivalent to AE|F C, and obviously we have
AE|F C if and only if \EAC = \ACF . Now, since the points A, B, D and E lie on
one circle, we have that \EAD = \EBD, what rewrites as \EAC = \CBM . On other
hand, we trivially have that \ACF = \DCM . Thus, \EAC = \ACF if and only if
\CBM = \DCM . Now, as it is clear that \CM B = \DM C, we have \CBM = \DCM
if and only if the triangles CM B and DM C are similar. But, the triangles CM B and
B
MC
DM C are similar if and only if M
= M
. This is finally equivalent to M B M D = M C 2 ,
MC
D
and so, by combining all these equivalences, the conclusion follows.

140

INFINITY

Epsilon 52. [TD] Let P be an arbitrary point in the plane of a triangle ABC with the
centroid G. Show the following inequalities
(1) BC P B P C + AB P A P B + CA P C P A BC CA AB and
3
3
3
(2) P A BC + P B CA + P C AB 3P G BC CA AB.
Solution. We only check the first inequality. We regard A, B, C, P as complex numbers
and assume that P corresponds to 0. Were required to prove that
|(B C)BC| + |(A B)AB| + |(C A)CA| |(B C)(C A)(A B)|.
It remains to apply The Triangle Inequality to the algebraic identity
(B C)BC + (A B)AB + (C A)CA = (B C)(C A)(A B).

INFINITY

141

Epsilon 53. (The Neuberg-Pedoe Inequality) Let a1 , b1 , c1 denote the sides of the triangle
A1 B1 C1 with area F1 . Let a2 , b2 , c2 denote the sides of the triangle A2 B2 C2 with area F2 .
Then, we have
a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 ) 16F1 F2 .
Fifth Proof. ([GC], G. Chang) We regard A, B, C, A0 , B 0 , C 0 as complex numbers and assume that C corresponds to 0. Rewriting the both sides in the inequality in terms of
complex numbers, we get
=

a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 )

`
`

2
2
2 |A0 | |B|2 + |A|2 |B 0 | AB + AB A0 B 0 + A0 B

and

`
`

16F1 F2 = AB AB A0 B 0 + A0 B 0 ,
where the sign begin chose to make the right hand positive. According to whether the
triangle ABC and the triangle A0 B 0 C 0 have the same orientation or not, we obtain either

2
a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 ) 16F1 F2 = 2AB 0 A0 B
or

2
a1 2 (b2 2 + c2 2 a2 2 ) + b1 2 (c2 2 + a2 2 b2 2 ) + c1 2 (a2 2 + b2 2 c2 2 ) 16F1 F2 = 2AB 0 A0 B .
This completes the proof.

142

INFINITY

Epsilon 54. [SL 2002 KOR] Let ABC be a triangle for which there exists an interior point
F such that \AF B = \BF C = \CF A. Let the lines BF and CF meet the sides AC
and AB at D and E, respectively. Prove that AB + AC 4DE.
+ i sin 2
. We can toss the
Solution. Let AF = x, BF = y, CF = z and let = cos 2
3
3
pictures on C so that the points F , A, B, C, D, and E are represented by the complex
xz
numbers 0, x, y, z 2 , d, and e. Its an easy exercise to establish that DF = x+z
and
xy
xy
xz
EF = x+y . This means that d = x+z
and e = x+y . Were now required to prove
that

zx
xy 2
+
.
|x y| + |z 2 x| 4
z+x
x+y
3
2
2
Since || = 1 and = 1, we have |z x| = |(z x)| = |z x|. Therefore, we
need to prove

4zx
4xy
|x y| + |z x|

.
z+x
x+y

4zx
4xy
More strongly, we establish that |(x y) + (z x)| z+x
x+y
or |p q|
|r s|, where p = z + x, q = y + x, r =
q s > 0. It follows that

4zx
z+x

and s =

4xy
.
x+y

Its clear that p r > 0 and

|p q|2 |r s|2 = (pq)(p q)(rs)(r s) = (p2 r2 )+(pqrs)+(q 2 s2 ) 0.

Its easy to check that the equality holds if and only if 4ABC is equilateral.

INFINITY

143

Epsilon 55. (APMO 2004/5) Prove that, for all positive real numbers a, b, c,
(a2 + 2)(b2 + 2)(c2 + 2) 9(ab + bc + ca).

First Solution. Choose A, B, C 0, 2 with a = 2 tan A, b = 2 tan B, and c =

2 tan C. Using the trigonometric identity 1 + tan2 = cos12 , one may rewrite it as
4
cos A cos B cos C (cos A sin B sin C + sin A cos B sin C + sin A sin B cos C) .
9
One may easily check the following trigonometric identity
cos(A+B +C) = cos A cos B cos C cos A sin B sin C sin A cos B sin C sin A sin B cos C.
Then, the above trigonometric inequality takes the form
4
cos A cos B cos C (cos A cos B cos C cos(A + B + C)) .
9
Let = A+B+C
. Applying The AM-GM Inequality and Jesens Inequality, we have
3

3
cos A + cos B + cos C
cos A cos B cos C
cos3 .
3
We now need to show that
4
cos3 (cos3 cos 3).
9
Using the trigonometric identity
cos 3 = 4 cos3 3 cos or cos3 cos 3 = 3 cos 3 cos3 ,
it becomes

4
cos4 1 cos2 ,
27
which follows from The AM-GM Inequality

cos2 cos2 `
1 cos2
cos2 `
1

1 cos2

+
+ 1 cos2
= .
2
2
3
2
2
3
One find that the equality holds if and only if tan A = tan B = tan C =
a = b = c = 1.

if and only if

144

INFINITY

Epsilon 56. (Latvia 2002) Let a, b, c, d be the positive real numbers such that
1
1
1
1
+
+
+
= 1.
1 + a4
1 + b4
1 + c4
1 + d4
Prove that abcd 3.
First Solution.` We can write a2 = tan A, b2 = tan B, c2 = tan C, d2 = tan D, where
A, B, C, D 0, 2 . Then, the algebraic identity becomes the following trigonometric
identity :
cos2 A + cos2 B + cos2 C + cos2 D = 1.
Applying The AM-GM Inequality, we obtain
2

sin2 A = 1 cos2 A = cos2 B + cos2 C + cos2 D 3 (cos B cos C cos D) 3 .

Similarly, we obtain
2

sin2 B 3 (cos C cos D cos A) 3 , sin2 C 3 (cos D cos A cos B) 3 , and sin2 D 3 (cos A cos B cos C) 3 .
Multiplying these four inequalities, we get the result!

INFINITY

145

Epsilon 57. (Korea 1998) Let x, y, z be the positive reals with x + y + z = xyz. Show
that
1
1
1
3

+p
+
.
2
2
2
2
1+x
1+z
1+y
First Solution. We give a convexity proof. We can write x = tan A, y = tan B, z = tan C,
`

`
2
where A, B, C 0, 2 . Using the fact that 1 + tan2 = cos1 , we rewrite it in the
terms of A, B, C :
3
cos A + cos B + cos C .
2
x+y
It follows from tan( C) = z = 1xy
= tan(A + B) and from C, A + B (0, )
that C = A + B or A + B + C = . Hence, it suffices to show the following.

146

INFINITY

Epsilon 58. (USA 2001) Let a, b, and c be nonnegative real numbers such that a2 + b2 +
c2 + abc = 4. Prove that 0 ab + bc + ca abc 2.
Solution. Notice that a, b, c > 1 implies that a2 + b2 + c2 + abc > 4. If a 1, then we have
ab + bc + ca abc (1 a)bc 0. We now prove that ab + bc + ca abc 2. Letting
a = 2p, b = 2q, c = 2r, we get p2 + q 2 + r2 + 2pqr = 1. By the above exercise, we can
write
h i
with A + B + C = .
a = 2 cos A, b = 2 cos B, c = 2 cos C for some A, B, C 0,
2
We are required to prove
1
cos A cos B + cos B cos C + cos C cos A 2 cos A cos B cos C .
2
One may assume that A 3 or 1 2 cos A 0. Note that
cos A cos B+cos B cos C+cos C cos A2 cos A cos B cos C = cos A(cos B+cos C)+cos B cos C(12 cos A).
We apply Jensens Inequality to deduce cos B+cos C 23 cos A. Note that 2 cos B cos C =
cos(B C) + cos(B + C) 1 cos A. These imply that

1 cos A
3
cos A(cos B+cos C)+cos B cos C(12 cos A) cos A
cos A +
(12 cos A).
2
2
`
`

However, its easy to verify that cos A 32 cos A + 1cos

(1 2 cos A) = 12 .
2

INFINITY

147

Epsilon 59. [IMO 2001/2 KOR] Let a, b, c be positive real numbers. Prove that
a
b
c

+
+
1.
2
2
2
a + 8bc
b + 8ca
c + 8ab
First Solution. To remove the square roots, we make the following substitution :
a
b
c
x=
, y=
, z=
.
2
2
2
a + 8bc
b + 8ca
c + 8ab
Clearly, x, y, z (0, 1). Our aim is to show that x + y + z 1. We notice that

a2
x2
b2
y2
c2
z2
1
x2
y2
z2
=
,
=
,
=
=
=
.
8bc
1 x2 8ac
1 y 2 8ab
1 z2
512
1 x2
1 y2
1 z2
Hence, we need to show that
x + y + z 1, where 0 < x, y, z < 1 and (1 x2 )(1 y 2 )(1 z 2 ) = 512(xyz)2 .
However, 1 > x + y + z implies that, by The AM-GM Inequality,
(1x2 )(1y 2 )(1z 2 ) > ((x+y+z)2 x2 )((x+y+z)2 y 2 )((x+y+z)2 z 2 ) = (x+x+y+z)(y+z)
1

(x+y+y+z)(z+x)(x+y+z+z)(x+y) 4(x2 yz) 4 2(yz) 2 4(y 2 zx) 4 2(zx) 2 4(z 2 xy) 4 2(xy) 2
= 512(xyz)2 . This is a contradiction !

148

INFINITY

Epsilon 60. [IMO 1995/2 RUS] Let a, b, c be positive numbers such that abc = 1. Prove
that
1
1
1
3
+ 3
+ 3
.
a3 (b + c)
b (c + a)
c (a + b)
2
Second Solution. After the substitution a = x1 , b = y1 , c = z1 , we get xyz = 1. The
inequality takes the form
x2
y2
z2
3
+
+
.
y+z
z+x
x+y
2
It follows from The Cauchy-Schwarz Inequality that
2

y2
z2
x
+
+
[(y + z) + (z + x) + (x + y)]
(x + y + z)2
y+z
z+x
x+y
so that, by The AM-GM Inequality,
1

3(xyz) 3
x2
y2
z2
x+y+z
3
+
+

= .
y+z
z+x
x+y
2
2
2

INFINITY

149

Epsilon 61. (Korea 1998) Let x, y, z be the positive reals with x + y + z = xyz. Show
that
1
1
1
3

+p
+
.
2
2
2
2
1+x
1+z
1+y
Second Solution. The starting point is letting a = x1 , b = y1 , c = z1 . We find that
a + b + c = abc is equivalent to 1 = xy + yz + zx. The inequality becomes
x
z
y
3

+
+p

2
x2 + 1
z2 + 1
y2 + 1
or
x
y
z
3
p
+p
+p

2
x2 + xy + yz + zx
y 2 + xy + yz + zx
z 2 + xy + yz + zx
or
x
y
z
3
p
+p
+p
.
2
(x + y)(x + z)
(y + z)(y + x)
(z + x)(z + y)
By the AM-GM inequality, we have
p

x (x + y)(x + z)
x
1 x[(x + y) + (x + z)]
1
x
x
p
=

=
+
.
(x + y)(x + z)
2 (x + y)(x + z)
2 x+z
x+z
(x + y)(x + z)
In a like manner, we obtain

y
y
y
z
z
z
1
1
p
+
and p
+
.

2 y+z
y+x
2 z+x
z+y
(y + z)(y + x)
(z + x)(z + y)
Adding these three yields the required result.

150

INFINITY

Epsilon 62. [IMO 2000/2 USA] Let a, b, c be positive numbers such that abc = 1. Prove
that

1
1
1
a1+
b1+
c1+
1.
b
c
a
Second Solution. ([IV], Ilan Vardi) Since abc = 1, we may assume that a 1 b. 17 It
follows that

(a 1)(1 b) 18
1
1
1
1
1
1 a 1 +
b1+
c1+
= c+ 2
a+ 1 +
.
b
c
a
c
b
a

x
,
y

y
,
z

z
x

Third Solution. As in the first solution, after the substitution a =

b=
c = for x,
y, z > 0, we can rewrite it as xyz (y + z x)(z + x y)(x + y z). Without loss of
generality, we can assume that z y x. Set y x = p and z x = q with p, q 0. Its
straightforward to verify that
xyz (y + z x)(z + x y)(x + y z) = (p2 pq + q 2 )x + (p3 + q 3 p2 q pq 2 ).
Since p2 pq + q 2 (p q)2 0 and p3 + q 3 p2 q pq 2 = (p q)2 (p + q) 0, we get
the result.

Fourth Solution. (From the IMO 2000 Short List) Using the condition abc = 1, its straightforward to verify the equalities

1
1
1
2=
a1+
+c b1+
,
a
b
c

1
1
1
2=
b1+
+a c1+
,
b
c
a

1
1
1
2=
c1+
+b a1+
.
c
a
c
In particular, they show that at most one of the numbers u = a 1 + 1b , v = b 1 + 1c ,
w = c 1 + a1 is negative. If there is such a number, we have

1
1
1
a1+
b1+
c1+
= uvw < 0 < 1.
b
c
a
And if u, v, w 0, The AM-GM Inequality yields
r
r
r
1
c
1
a
1
b
2 = u + cv 2
uv, 2 = v + aw 2
vw, 2 = w + aw 2
wu.
a
a
b
b
c
c
Thus, uv ac , vw
completes the proof.

17

b
,
a

wu

c
,
b

so (uvw)2

a
c

b
a

c
b

= 1. Since u, v, w 0, this

Why? Note that the inequality is not symmetric in the three variables. Check it!
For a verification of the identity, see [IV].

18

INFINITY

151

Epsilon 63. Let a, b, c be positive real numbers satisfying a + b + c = 1. Show that

a
b
abc
3 3
+
+
1+
.
a + bc
b + ca
c + ab
4
Solution. We want to establish that

ab
1
3 3
1
c
+

1
+
+
.
1 + ca
4
1 + bc
1 + ab
b
a
c
q
q
p ca
ab
,
y
=
,
z
=
. We need to prove that
Set x = bc
a
b
c

1
1
z
3 3
+
+

1
+
,
1 + x2
1 + y2
1 + z2
4
where x, y, z > 0 and xy + yz + zx = 1. Its not hard to show that there exists A, B, C
(0, ) with
A
B
C
x = tan , y = tan , z = tan , and A + B + C = .
2
2
2
The inequality becomes

tan C2
1
1
3 3
+
+

1
+
`
`
`
2
2
2
4
1 + tan A
1 + tan B
1 + tan C
2

or

1+
or

1
3 3
(cos A + cos B + sin C) 1 +
2
4

3 3
cos A + cos B + sin C
.
2

AB
` AB

Note that cos A + cos B = 2 cos

cos
. Since 2 < 2 , this means that
2
2

A+B
C
cos A + cos B 2 cos
= 2 cos
.
2
2
` A+B

It will be enough to show that

2 cos

C
2

+ sin C

3 3
,
2

where C (0, ). This is a one-variable inequality.19 Its left as an exercise for the reader.

19

Differentiate!

Shiing-Shen Chern

152

INFINITY

Epsilon 64. (Latvia 2002) Let a, b, c, d be the positive real numbers such that
1
1
1
1
+
+
+
= 1.
1 + a4
1 + b4
1 + c4
1 + d4
Prove that abcd 3.
Second Solution. (given by Jeong Soo Sim at the KMO Weekend Program 2007) We need
to prove the inequality a4 b4 c4 d4 81. After making the substitution
1
1
1
1
A=
, B=
, C=
, D=
,
1 + a4
1 + b4
1 + c4
1 + d4
we obtain
1A 4
1B 4
1C 4
1D
a4 =
, b =
, c =
, d =
.
A
B
C
D
The constraint becomes A + B + C + D = 1 and the inequality can be written as
1A 1B 1C 1D

81.
A
B
C
D
or
B+C +D C +D+A D+A+B A+B+C

81.
A
B
C
D
or
(B + C + D)(C + D + A)(D + A + B)(A + B + C) 81ABCD.
However, this is an immediate consequence of The AM-GM Inequality:
1

(B+C+D)(C+D+A)(D+A+B)(A+B+C) 3 (BCD) 3 3 (CDA) 3 3 (DAB) 3 3 (ABC) 3 .

INFINITY

153

Epsilon 65. [LL 1992 UNK] (Iran 1998) Prove that, for all x, y, z > 1 such that x1 + y1 + z1 =
2,
p

x + y + z x 1 + y 1 + z 1.

First
Solution. We begin with the algebraic substitution a = x 1, b = y 1, c =

z 1. Then, the condition becomes

1
1
1
+
+
= 2 a2 b2 + b2 c2 + c2 a2 + 2a2 b2 c2 = 1
1 + a2
1 + b2
1 + c2
and the inequality is equivalent to
p
3
a2 + b2 + c2 + 3 a + b + c ab + bc + ca .
2
Let p = bc, q = ca, r = ab. Our job is to prove that p+q+r 23 where p2 +q 2 +r2 +2pqr =
1. Now, we can make the trigonometric substitution

p = cos A, q = cos B, r = cos C for some A, B, C 0,
with A + B + C = .
2
What we need to show is now that cos A + cos B + cos C 32 . It follows from Jensens
Inequality.

154

INFINITY

Epsilon 66. (Belarus 1998) Prove that, for all a, b, c > 0,

a
b
c
a+b
b+c
+ +
+
+ 1.
b
c
a
b+c
c+a
Solution. After writing x =

a
b

and y = cb , we get

c
y a+b
x+1 b+c
1+y
= ,
=
,
=
.
a
x b+c
1+y c+a
y+x
One may rewrite the inequality as
x3 y 2 + x2 + x + y 3 + y 2 x2 y + 2xy + 2xy 2 .
Apply The AM-GM Inequality to obtain
x3 y 2 + x
x3 y 2 + x + y 3 + y 3
x2 y,
2xy 2 , x2 + y 2 2xy.
2
2
Adding these three inequalities, we get the result. The equality holds if and only if
x = y = 1 or a = b = c.

INFINITY

155

Epsilon 67. [SL 2001 ] Let x1 , , xn be arbitrary real numbers. Prove the inequality.

x1
x2
xn
+
+ +
< n.
1 + x1 2
1 + x1 2 + x2 2
1 + x1 2 + + xn 2
First Solution. We only consider the case when x1 , , xn are all nonnegative real numbers.(Why?)20 Let x0 = 1. After the substitution yi = x0 2 + + xi 2 for all i = 0, , n,

we obtain xi = yi yi1 . We need to prove the following inequality

n
X

yi yi1
< n.
y
i
i=0
Since yi yi1 for all i = 1, , n, we have an upper bound of the left hand side:
n
n
n r
X
X
X
yi yi1
yi yi1
1
1

y
y
y
y
i
i yi1
i1
i
i=0
i=0
i=0
We now apply the Cauchy-Schwarz inequality to give an upper bound of the last term:
v
u
s

n r
n
X
u X
1
1
1
1
1
1

tn

= n
.
yi1
yi
yi1
yi
y0
yn
i=0
i=0

Since y0 = 1 and yn > 0, this yields the desired upper bound n.

Second Solution. We may assume that x1 , , xn are all nonnegative real numbers. Let
x0 = 0. We make the following algebraic substitution
xi
1
ti
ti =
, ci =
and si =
2
2
2
x0 + + xi
1 + ti
1 + ti 2
xi
for all i = 0, , n. Its an easy exercise to show that x0 2 ++x 2 = c0 ci si . Since
i

si = 1 ci 2 , the desired inequality becomes

p
p
p

c0 c1 1 c1 2 + c0 c1 c2 1 c2 2 + + c0 c1 cn 1 cn 2 < n.
Since 0 < ci 1 for all i = 1, , n, we have
n
n
n
p
p
X
X
X
p
c0 ci 1 ci 2
c0 ci1 1 ci 2 =
(c0 ci1 )2 (c0 ci1 ci )2 .
i=1

i=1

i=1

Since c0 = 1, by The Cauchy-Schwarz Inequality, we obtain

v
u n
n
X
p
p
u X
(c0 ci1 )2 (c0 ci1 ci )2 tn
[(c0 ci1 )2 (c0 ci1 ci )2 ] = n [1 (c0 cn )2 ].
i=1

i=1

20

x1
1+x1 2

x2
1+x1 2 +x2 2

+ +

xn
1+x1 2 ++xn 2

|x1 |
1+x1 2

|x2 |
1+x1 2 +x2 2

+ +

|xn |
.
1+x1 2 ++xn 2

156

INFINITY

Epsilon 68. Let a, b, c be the lengths of a triangle. Show that

b
c
a
+
+
< 2.
b+c
c+a
a+b
Solution. We dont employ The Ravi Substitution! It follows from the triangle inequality
that
X a
X
a
<
= 2.
1
b+c
(a
+
b + c)
2
cyclic

cyclic

INFINITY

157

Epsilon 69. [IMO 2001/2 KOR] Let a, b, c be positive real numbers. Prove that
a
b
c

+
+
1.
2
2
2
a + 8bc
b + 8ca
c + 8ab
Second Solution. Lets try to find a new lower bound of (x + y + z)2 where x, y, z > 0.
2
There are well-known lower bounds such as 3(xy + yz + zx) and 9(xyz) 3 . Here, we break
the symmetry. We notice that
(x + y + z)2 = x2 + y 2 + z 2 + xy + xy + yz + yz + zx + zx.
We apply The AM-GM Inequality to the right hand side except the term x2 :
1

y 2 + z 2 + xy + xy + yz + yz + zx + zx 8x 2 y 4 z 4 .
It follows that

1
3
3
1
3
3
(x + y + z)2 x2 + 8x 2 y 4 z 4 = x 2 x 2 + 8y 4 z 4 .

We proved the estimation, for x, y, z > 0,

r
x+y+z
It follows that

1
3
3
x 2 x 2 + 8y 4 z 4 .

x4
3
2

3
4

x + 8y z

cyclic
4
3

4
3

3
4

X
cyclic

x
= 1.
x+y+z

After the substitution x = a , y = b , and z = c 3 , it now becomes the inequality

X
a

1.
2 + 8bc
a
cyclic

158

INFINITY

Epsilon 70. [IMO 2005/3 KOR] Let x, y, and z be positive numbers such that xyz 1.
Prove that
x5 x2
y5 y2
z5 z2
+ 5
+ 5
0.
5
2
2
2
2
x +y +z
y +z +x
z + x2 + y 2
First Solution. Its equivalent to the following inequality

x2 x5
y2 y5
z2 z5
+
1
+
+
1
+
+
1
3
x5 + y 2 + z 2
y 5 + z 2 + x2
z 5 + x2 + y 2
or
x2 + y 2 + z 2
x2 + y 2 + z 2
x2 + y 2 + z 2
+
+
3.
x5 + y 2 + z 2
y 5 + z 2 + x2
z 5 + x2 + y 2
With The Cauchy-Schwarz Inequality and the fact that xyz 1, we have
(x5 + y 2 + z 2 )(yz + y 2 + z 2 ) (x2 + y 2 + z 2 )2
or
x2 + y 2 + z 2
yz + y 2 + z 2
2
.
5
2
2
x +y +z
x + y2 + z2
Taking the cyclic sum, we reach
x2 + y 2 + z 2
x2 + y 2 + z 2
x2 + y 2 + z 2
xy + yz + zx
+ 5
+ 5
2+ 2
3.
x5 + y 2 + z 2
y + z 2 + x2
z + x2 + y 2
x + y2 + z2

Second Solution. The main idea is to think of 1 as follows :

x5

x5
y5
z5
x2
y2
z2
+ 5
+ 5
1 5
+ 5
+ 5
.
2
2
2
2
2
2
2
2
2
2
+y +z
y +z +x
z +x +y
x +y +z
y +z +x
z + x2 + y 2

We first show the left-hand. It follows from y 4 + z 4 y 3 z + yz 3 = yz(y 2 + z 2 ) that

x(y 4 + z 4 ) xyz(y 2 + z 2 ) y 2 + z 2 or

x5
x5
x4
5
= 4
.
x5 + y 2 + z 2
x + xy 4 + xz 4
x + y4 + z4

Taking the cyclic sum, we have the required inequality. It remains to show the right-hand.
As in the first solution, The Cauchy-Schwarz Inequality and xyz 1 imply that
(x5 + y 2 + z 2 )(yz + y 2 + z 2 ) (x2 + y 2 + z 2 )2 or

x2 (yz + y 2 + z 2 )
x2
5
.
(x2 + y 2 + z 2 )2
x + y2 + z2

Taking the cyclic sum, we have

X x2 (yz + y 2 + z 2 )
X
x2

.
2
2
2
2
5
(x + y + z )
x + y2 + z2

cyclic

cyclic

Our job is now to establish the following homogeneous inequality

1

X x2 (yz + y 2 + z 2 )
X 2 2 X 2
X 4
X 2
(x2 +y 2 +z 2 )2 2
x y +
x yz
x
x yz.
(x2 + y 2 + z 2 )2

cyclic

cyclic

cyclic

cyclic

cyclic

However, by The AM-GM Inequality, we obtain

X
cyclic

x4 =

X 2 2
X 2 y2 + z2
X 2
X x4 + y 4

x y =
x

x yz.
2
2

cyclic

cyclic

cyclic

cyclic

INFINITY

159

Remark 8.2. Here is an alternative way to reach the right hand side inequality. We claim
that
2x4 + y 4 + z 4 + 4x2 y 2 + 4x2 z 2
x2
5
.
2
2
2
2
4(x + y + z )
x + y2 + z2
We do this by proving
2x4 + y 4 + z 4 + 4x2 y 2 + 4x2 z 2
x2 yz

4(x2 + y 2 + z 2 )2
x4 + y 3 z + yz 3
because xyz 1 implies that
x2 yz
=
x4 + y 3 z + yz 3

x2
x5
xyz

+ y2 + z2

x2
.
x5 + y 2 + z 2

Hence, we need to show the homogeneous inequality

(2x4 + y 4 + z 4 + 4x2 y 2 + 4x2 z 2 )(x4 + y 3 z + yz 3 ) 4x2 yz(x2 + y 2 + z 2 )2 .
However, this is a straightforward consequence of The AM-GM Inequality.
(2x4 + y 4 + z 4 + 4x2 y 2 + 4x2 z 2 )(x4 + y 3 z + yz 3 ) 4x2 yz(x2 + y 2 + z 2 )2
=

(x8 + x4 y 4 + x6 y 2 + x6 y 2 + y 7 z + y 3 z 5 ) + (x8 + x4 z 4 + x6 z 2 + x6 z 2 + yz 7 + y 5 z 3 )

+2(x6 y 2 + x6 z 2 ) 6x4 y 3 z 6x4 yz 3 2x6 yz

p
p
6 6 x8 x4 y 4 x6 y 2 x6 y 2 y 7 z y 3 z 5 + 6 6 x8 x4 z 4 x6 z 2 x6 z 2 yz 7 y 5 z 3
p
+2 x6 y 2 x6 z 2 6x4 y 3 z 6x4 yz 3 2x6 yz

0.

Taking the cyclic sum, we obtain

X
X 2x4 + y 4 + z 4 + 4x2 y 2 + 4x2 z 2
x2

.
1=
2
2
2
2
5
4(x + y + z )
x + y2 + z2
cyclic

cyclic

Third Solution. (by an IMO 2005 contestant Iurie Boreico21 from Moldova) We establish
that
x5 x2
x5 x2
3 2
.
5
2
2
x +y +z
x (x + y 2 + z 2 )
It follows immediately from the identity
(x3 1)2 x2 (y 2 + z 2 )
x5 x2
x5 x2
3 2
= 3 2
.
2
2
2
2
+y +z
x (x + y + z )
x (x + y 2 + z 2 )(x5 + y 2 + z 2 )
Taking the cyclic sum and using xyz 1, we have
X
X 2 1
X ` 2

x5 x2
1
1

x
5
x yz 0.
5
2
2
5
2
2
2
2
x +y +z
x +y +z
x
x +y +z
x5

cyclic

cyclic

cyclic

21

He received the special prize for this solution.

160

INFINITY

Epsilon 71. (KMO Weekend Program 2007) Prove that, for all a, b, c, x, y, z > 0,
(a + b + c)(x + y + z)
ax
by
cz
+
+

.
a+x
b+y
c+z
a+b+c+x+y+z
Solution. (by Sanghoon at the KMO Weekend Program 2007) We need the following
lemma:
Lemma 8.2. For all p, q, 1 , 2 > 0, we have
1 2 p + 2 2 q
pq

.
p+q
(1 + 2 )2
Proof. After expanding, it becomes
`

(p + q) 1 2 p + 2 2 q (1 + 2 )2 pq 0.
However, it can be written as

(1 p 2 q)2 0.

Now, taking (p, q, 1 , 2 ) = (a, x, x + y + z, a + b + c) in the lemma, we get

(x + y + z)2 a + (a + b + c)2 x
ax

.
a+x
(x + y + z + a + b + c)2
Similarly, we obtain

(x + y + z)2 b + (a + b + c)2 y
by

b+y
(x + y + z + a + b + c)2

and

(x + y + z)2 c + (a + b + c)2 z
cz

.
c+z
(x + y + z + a + b + c)2
Adding the above three inequalities, we get
(x + y + z)2 (a + b + c) + (a + b + c)2 (x + y + z)
ax
by
cz
+
+

.
a+x
b+y
c+z
(x + y + z + a + b + c)2
or

(a + b + c)(x + y + z)
ax
by
cz
+
+

,
a+x
b+y
c+z
a+b+c+x+y+z

as desired.

INFINITY

161

Epsilon 72. (USAMO Summer Program 2002) Let a, b, c be positive real numbers. Prove
that
2
2
2

3
3
3
2a
2b
2c
+
+
3.
b+c
c+a
a+b
Proof. Establish the inequality

2a
b+c

2
3

a
a+b+c

162

INFINITY

Epsilon 73. (APMO 2005) Let a, b, c be positive real numbers with abc = 8. Prove that
p

a2
b2
c2
4
+p
+p

3
3
3
3
3
(1 + a )(1 + b )
(1 + b )(1 + c )
(1 + c3 )(1 + a3 )

Proof. Use the auxiliary inequality

1
2
.

2 + x2
1 + x3

INFINITY

163

Epsilon 74. (Titu Andreescu, Gabriel Dospinescu) Let x, y, and z be real numbers such
that x, y, z 1 and x + y + z = 1. Prove that
1
1
27
1
+
+

.
1 + x2
1 + y2
1 + z2
10
Solution. Employ the following inequality
1
27
(t 2) ,
1 + t2
50
where t 1.

164

INFINITY

Epsilon 75. (Japan 1997) Let a, b, and c be positive real numbers. Prove that
(b + c a)2
(c + a b)2
(a + b c)2
3
+
+
.
2
2
2
2
(b + c) + a
(c + a) + b
(a + b)2 + c2
5
Solution. Because of the homogeneity of the inequality, we may normalize to a + b + c = 1.
It takes the form
(1 2a)2
(1 2b)2
(1 2c)2
3
+
+

(1 a)2 + a2
(1 b)2 + b2
(1 c)2 + c2
5
or
1
1
1
27
+ 2
+ 2

.
2a2 2a + 1
2b 2b + 1
2c 2c + 1
5
1
We find that the equation of the tangent line of f (x) = 2x2 2x+1
at x = 13 is given by
y=
and that

f (x)

for all x > 0. It follows that

54
27
x+
25
25

cyclic

54
27
x+
25
25

f (a)

2(3x 1)2 (6x + 1)

0.
25(2x2 2x + 1)

X 54
27
27
a+
=
.
25
25
5

cyclic

INFINITY

165

Epsilon 76. [IMO 1984/1 FRG] Let x, y, z be nonnegative real numbers such that x+y+z =
7
1. Prove that 0 xy + yz + zx 2xyz 27
.
First Solution. Using the constraint x+y+z = 1, we reduce the inequality to homogeneous
one:
7
0 (xy + yz + zx)(x + y + z) 2xyz
(x + y + z)3 .
27
The left hand side inequality is trivial because its equivalent to
X 2
0 xyz +
x y.
sym

The right hand side inequality simplifies to

X 3
X 2
7
x + 15xyz 6
x y 0.
sym

cyclic

In the view of
7

x + 15xyz 6

0
X

x y = @2

sym

cyclic

1
X

x3

3xyz +

x y A + 5 @3xyz +

1
X
cyclic

x yA ,
2

sym

x2 y

sym

cyclic

and

sym

cyclic

its enough to show that

x3

cyclic

x2 y.

sym

The first inequality follows from

X 3 X 2
X 3
X 2
X 3
2
x
x y=
(x + y 3 )
(x y + xy 2 ) =
(x + y 3 x2 y xy 2 ) 0.
cyclic

sym

cyclic

cyclic

cyclic

The second inequality can be rewritten as

X
x(x y)(x z) 0,
cyclic

which is a particular case of Schurs Theorem.

166

INFINITY

Epsilon 77. [LL 1992 UNK] (Iran 1998) Prove that, for all x, y, z > 1 such that
2,
p

x + y + z x 1 + y 1 + z 1.

1
+ y1
x

+ z1 =

Second Solution. After the algebraic substitution a = x1 , b = y1 , c = z1 , we are required to

prove that
r
r
r
r
1
1a
1b
1c
1
1
+ +
+
+
,
a
b
c
a
b
c
where a, b, c (0, 1) and a + b + c = 2. Using the constraint a + b + c = 2, we obtain a
homogeneous inequality
s
s
s
s

a+b+c
a+b+c
a+b+c

b
c
1
1
1
1
2
2
2
(a + b + c)
+ +
+
+

2
a
b
c
a
b
c
or
s
r
r

r
1
1
1
b+ca
c+ab
a+bc
(a + b + c)
+ +
+
+
,

a
b
c
a
b
c
which immediately follows from The Cauchy-Schwarz Inequality
s
r
r

r
1
1
1
b+ca
c+ab
a+bc
[(b + c a) + (c + a b) + (a + b c)]
+ +

+
+
.
a
b
c
a
b
c

INFINITY

167

Epsilon 78. Let x, y, z be nonnegative real numbers. Then, we have

3
3
3
3xyz + x3 + y 3 + z 3 2 (xy) 2 + (yz) 2 + (zx) 2 .
First Solution. By Schurs Inequality and The AM-GM Inequality, we have
X 3
X 2
X
3
3xyz +
x
x y + xy 2
2(xy) 2 .
cyclic

cyclic

cyclic

168

INFINITY

Epsilon 79. Let t (0, 3]. For all a, b, c 0, we have

X 2
X
2
(3 t) + t(abc) t +
a 2
ab.
cyclic
2
3

2
3

cyclic

2
3

Proof. After setting x = a , y = b , z = c , it becomes

X 3
X
3
3
3 t + t(xyz) t +
x 2
(xy) 2 .
cyclic

cyclic

By the previous epsilon, it will be enough to show that

3

3 t + t(xyz) t 3xyz,
which is a straightforward consequence of the weighted AM-GM inequality :

t
3t
3
3
3t
t
3
1 + (xyz) t 1 3 (xyz) t
= 3xyz.
3
3
One may check that the equality holds if and only if a = b = c = 1.
Remark 8.3. In particular, we obtain non-homogeneous inequalities
1
5
+ (abc)4 + a2 + b2 + c2 2(ab + bc + ca),
2
2
2 + (abc)2 + a2 + b2 + c2 2(ab + bc + ca),
1 + 2abc + a2 + b2 + c2 2(ab + bc + ca).

INFINITY

169

Epsilon 80. (APMO 2004/5) Prove that, for all positive real numbers a, b, c,
(a2 + 2)(b2 + 2)(c2 + 2) 9(ab + bc + ca).
Second Solution. After expanding, it becomes
X 2 2
X 2
X
8 + (abc)2 + 2
a b +4
a 9
ab.
cyclic

cyclic

cyclic

From the inequality (ab 1)2 + (bc 1)2 + (ca 1)2 0, we obtain
X 2 2
X
6+2
a b 4
ab.
cyclic

cyclic

Hence, it will be enough to show that

2 + (abc)2 + 4

a2 5

cyclic

ab.

cyclic

Since 3(a2 + b2 + c2 ) 3(ab + bc + ca), it will be enough to show that

X 2
X
2 + (abc)2 +
a 2
ab,
cyclic

which is a particular case of the previous epsilon.

cyclic

170

INFINITY

Epsilon 81. [IMO 2000/2 USA] Let a, b, c be positive numbers such that abc = 1. Prove
that

1
1
1
a1+
b1+
c1+
1.
b
c
a
Second Solution. It is equivalent to the following homogeneous inequality:

(abc)2/3
(abc)2/3
(abc)2/3
a (abc)1/3 +
b (abc)1/3 +
c (abc)1/3 +
abc.
b
c
a
After the substitution a = x3 , b = y 3 , c = z 3 with x, y, z > 0, it becomes

(xyz)2
(xyz)2
(xyz)2
3
3
x3 xyz +
y

xyz
+
z

xyz
+
x3 y 3 z 3 ,
y3
z3
x3
which simplifies to
` 2
`
`

x y y 2 z + z 2 x y 2 z z 2 x + x2 y z 2 x x2 y + y 2 z x3 y 3 z 3
or

3x3 y 3 z 3 +

X
cyclic

or

x6 y 3

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

cyclic

(x2 y)3

cyclic

which is a special case of Schurs Inequality.

x4 y 4 z +

x5 y 2 z 2

cyclic

(x2 y)2 (y 2 z)

sym

INFINITY

171

Epsilon 82. (Tournament of Towns 1997) Let a, b, c be positive numbers such that abc = 1.
Prove that
1
1
1
+
+
1.
a+b+1
b+c+1
c+a+1
Solution. We can rewrite the given inequality as following :
1
1
1
1
+
+

.
a + b + (abc)1/3
b + c + (abc)1/3
c + a + (abc)1/3
(abc)1/3
We make the substitution a = x3 , b = y 3 , c = z 3 with x, y, z > 0. Then, it becomes
1
1
1
1
+ 3
+ 3

x3 + y 3 + xyz
y + z 3 + xyz
z + x3 + xyz
xyz
which is equivalent to
X 3
xyz
(x + y 3 + xyz)(y 3 + z 3 + xyz) (x3 + y 3 + xyz)(y 3 + z 3 + xyz)(z 3 + x3 + xyz)
cyclic

or

x6 y 3

sym

We now obtain

x6 y 3

x5 y 2 z 2 !

sym

sym

x6 y 3 + y 6 x3

cyclic

x5 y 4 + y 5 x4

cyclic

x5 (y 4 + z 4 )

cyclic

x5 (y 2 z 2 + y 2 z 2 )

cyclic

x5 y 2 z 2 .

sym

172

INFINITY

Epsilon 83. (Muirheads Theorem) Let a1 , a2 , a3 , b1 , b2 , b3 be real numbers such that

a1 a2 a3 0, b1 b2 b3 0, a1 b1 , a1 + a2 b1 + b2 , a1 + a2 + a3 = b1 + b2 + b3 .
Let x, y, z be positive real numbers. Then, we have
X a a a
X b b b
x 1y 2z 3
x 1y 2z 3.
sym

sym

Solution. We distinguish two cases.

Case 1. b1 a2 : It follows from a1 a1 +a2 b1 and from a1 b1 that a1 max(a1 +a2
b1 , b1 ) so that max(a1 , a2 ) = a1 max(a1 +a2 b1 , b1 ). From a1 +a2 b1 b1 +a3 b1 = a3
and a1 + a2 b1 b2 b3 , we have max(a1 + a2 b1 , a3 ) max(b2 , b3 ). It follows that
X a a a
X a a a
x 1y 2z 3 =
z 3 (x 1 y 2 + xa2 y a1 )
sym

cyclic

z a3 (xa1 +a2 b1 y b1 + xb1 y a1 +a2 b1 )

cyclic

xb1 (y a1 +a2 b1 z a3 + y a3 z a1 +a2 b1 )

cyclic

xb1 (y b2 z b3 + y b3 z b2 )

cyclic

xb1 y b2 z b3 .

sym

Case 2. b1 a2 : It follows from 3b1 b1 + b2 + b3 = a1 + a2 + a3 b1 + a2 + a3

that b1 a2 + a3 b1 and that a1 a2 b1 a2 + a3 b1 . Therefore, we have
max(a2 , a3 ) max(b1 , a2 + a3 b1 ) and max(a1 , a2 + a3 b1 ) max(b2 , b3 ). It follows
that
X a a a
X a a a
x 1y 2z 3 =
x 1 (y 2 z 3 + y a3 z a2 )
sym

cyclic

xa1 (y b1 z a2 +a3 b1 + y a2 +a3 b1 z b1 )

cyclic

y b1 (xa1 z a2 +a3 b1 + xa2 +a3 b1 z a1 )

cyclic

y b1 (xb2 z b3 + xb3 z b2 )

cyclic

xb1 y b2 z b3 .

sym

INFINITY

173

Epsilon 84. [IMO 1995/2 RUS] Let a, b, c be positive numbers such that abc = 1. Prove
that
1
1
1
3
+ 3
+ 3
.
a3 (b + c)
b (c + a)
c (a + b)
2
Third Solution. Its equivalent to
1
1
1
3
+ 3
+ 3

.
a3 (b + c)
b (c + a)
c (a + b)
2(abc)4/3
Set a = x3 , b = y 3 , c = z 3 with x, y, z > 0. Then, it becomes
X
1
3

.
x9 (y 3 + z 3 )
2x4 y 4 z 4
cyclic

Clearing denominators, this can be rewritten as

X 12 12
X 12 9 3 X 9 9 6
X 11 8 5
x y +2
x y z +
x y z 3
x y z + 6x8 y 8 z 8
sym

or
X
sym

12 12

x y

X
sym

sym

!
11 8 5

x y z

+2

sym

X
sym

12 9 3

x y z

sym

X
sym

!
11 8 5

x y z

9 9 6

x y z

sym

By Muirheads Theorem, every term on the left hand side is nonnegative.

!
8 8 8

x y z

sym

0,

174

INFINITY

Epsilon 85. (Iran 1996) Let x, y, z be positive real numbers. Prove that

1
1
1
9
(xy + yz + zx)
+
+
.
(x + y)2
(y + z)2
(z + x)2
4
Second Solution. Its equivalent to
X 5
X 4
X 4 2
X 3 3
X 3 2
4
x y+2
x yz + 6x2 y 2 z 2
x y 6
x y 2
x y z 0.
sym

sym

cyclic

cyclic

sym

We rewrite this as following

0
1
!
!
X 5
X 4 2
X 5
X 3 3
X 3 X 2
x y
x y +3
x y
x y + 2xyz @3xyz +
x
x y A 0.
sym

sym

sym

sym

cyclic

sym

By Muirheads Theorem and Schurs Inequality, its a sum of three nonnegative terms.

INFINITY

175

Epsilon 86. Let x, y, z be nonnegative real numbers with xy + yz + zx = 1. Prove that

1
1
1
5
+
+
.
x+y
y+z
z+x
2
Solution. Using xy + yz + zx = 1, we homogenize the given inequality as following :

2 2
1
1
1
5
(xy + yz + zx)
+
+

x+y
y+z
z+x
2
or
X 5
X 4
X 3 2
X
X 3 3
4
x y+
x yz + 14
x y z + 38x2 y 2 z 2
x4 y 2 + 3
x y
sym

or
X
sym

x y

X
sym

sym

!
4 2

x y

+3

sym

X
sym

x y

X
sym

sym

!
3 3

x y

+xyz

X
sym

x + 14

sym

!
2

x y + 38xyz

0.

sym

By Muirheads Theorem, we get the result. In the above inequality, without the condition
xy + yz + zx = 1, the equality holds if and only if x = y, z = 0 or y = z, x =
0 or z = x, y = 0. Since xy + yz + zx = 1, the equality occurs when (x, y, z) =
(1, 1, 0), (1, 0, 1), (0, 1, 1).

176

INFINITY

Epsilon 87. [SC] If ma ,mb ,mc are medians and ra ,rb ,rc the exradii of a triangle, prove
that
ra rb
rc ra
rb rc
+
3.
+
ma mb
mb mc
mc ma
Solution. Set 2s = a + b + c. Using the well-known identities
r
s(s b)(s c)
1p 2
ra =
2b + 2c2 a2 , etc.
, ma =
sa
2
we obtain
X rb rc
X
4s(s a)
p
=
.
2
2
mb mc
(2c + 2a b2 )(2a2 + 2b2 c2 )
cyclic
cyclic
Applying the AM-GM inequality, we obtain
X rb rc
X 2(a + b + c)(b + c a)
X
8s(s a)
=
.

2
2
2
2
2
2
mb mc
(2c + 2a b ) + (2a + 2b c )
4a2 + b2 + c2
cyclic

cyclic

cyclic

Thus, it will be enough to show that

X 2(a + b + c)(b + c a)
4a2 + b2 + c2

cyclic

3.

After expanding the above inequality, we see that it becomes

X 6
X 4
X 3 2
X 3 3
X 5
X 4 2
2
a +4
a bc+20
a b c+68
a b +16
a b 276a2 b2 c2 +27
a b .
cyclic

cyclic

sym

cyclic

cyclic

cyclic

We note that this cannot be proven by just applying Muirheads Theorem. Since a, b,
c are the sides of a triangle, we can make The Ravi Substitution a = y + z, b = z + x,
c = x + y, where x, y, z > 0. After some brute-force algebra, we can rewrite the above
inequality as
X 6
X 5
X 4 2
X 3 3
X 4
25
x + 230
x y + 115
x y + 10
x y + 80
x yz
sym

sym

336

sym
3 2

x y z + 124

sym

Now, by Muirheads Theorem, we get the result !

sym

sym

2 2 2

x y z .

sym

INFINITY

177

Epsilon 88. Let P(u, v, w) R[u, v, w] be a homogeneous symmetric polynomial with degree
3. Then the following two statements are equivalent.
(a) P(1, 1, 1), P(1, 1, 0), P(1, 0, 0) 0.
(b) P(x, y, z) 0 for all x, y, z 0.
Proof. [SR1] We only prove that (a) implies (b). Let
X 2
X 3
u +B
u v + Cuvw.
P (u, v, w) = A
sym

cyclic

Letting p = P (1, 1, 1) = 3A + 6B + C, q = P (1, 1, 0) = A + B, and r = P (1, 0, 0) = A, we

have A = r, B = q r, C = p 6q + 3r, and p, q, r 0. For x, y, z 0, we have
X 3
X 2
P (x, y, z) = r
x + (q r)
x y + (p 6q + 3r)xyz
sym

cyclic

0
=

r@

x + 3xyz

x yA + q
2

sym

cyclic

1
2

x y

sym

!
xyz

+ pxyz

sym

0.

Remark 8.4. Here is an alternative way to prove the inequality P (x, y, z) 0.

Case 1. q r : We compute
r
P (x, y, z) =
2

sym

!
xyz

+ (q r)

sym

X
sym

x y

!
xyz

+ pxyz.

sym

Every term on the right hand side is nonnegative.

Case 2. q r : We compute
q
P (x, y, z) =
2

X
sym

!
xyz

0
+ (r q) @

sym

Every term on the right hand side is nonnegative.

1
X
cyclic

x + 3xyz

X
sym

x y A + pxyz.
2

178

INFINITY

Epsilon 89. [IMO 2001/2 KOR] Let a, b, c be positive real numbers. Prove that
a
b
c

+
+
1.
2
2
2
a + 8bc
b + 8ca
c + 8ab
Third Solution. We offer a convexity proof. We make the substitution
a
b
c
x=
, y=
, z=
.
a+b+c
a+b+c
a+b+c
The inequality becomes
xf (x2 + 8yz) + yf (y 2 + 8zx) + zf (z 2 + 8xy) 1,
where f (t) = 1t .22 Since f is convex on R+ and x + y + z = 1, we apply (the weighted)
Jensens Inequality to obtain
xf (x2 + 8yz) + yf (y 2 + 8zx) + zf (z 2 + 8xy) f (x(x2 + 8yz) + y(y 2 + 8zx) + z(z 2 + 8xy)).
Note that f (1) = 1. Since the function f is strictly decreasing, it suffices to show that
1 x(x2 + 8yz) + y(y 2 + 8zx) + z(z 2 + 8xy).
Using x + y + z = 1, we homogenize it as
(x + y + z)3 x(x2 + 8yz) + y(y 2 + 8zx) + z(z 2 + 8xy).
However, it is easily seen from
(x+y+z)3 x(x2 +8yz)y(y 2 +8zx)z(z 2 +8xy) = 3[x(yz)2 +y(zx)2 +z(xy)2 ] 0.

Fourth Solution. We begin with the substitution

bc
ca
ab
x = 2,y = 2 ,z = 2 .
a
b
c
Then, we get xyz = 1 and the inequality becomes
1
1
1

+
+
1
1 + 8y
1 + 8x
1 + 8z
which is equivalent to
X p
p
(1 + 8x)(1 + 8y) (1 + 8x)(1 + 8y)(1 + 8z).
cyclic

After squaring both sides, its equivalent to

X
p
8(x + y + z) + 2 (1 + 8x)(1 + 8y)(1 + 8z)
1 + 8x 510.
cyclic

Recall that xyz = 1. The AM-GM Inequality gives us x + y + z 3,

q
X
X
8
8
8
4
8
(1+8x)(1+8y)(1+8z) 9x 9 9y 9 9z 9 = 729 and
1 + 8x
9x 9 9(xyz) 27 = 9.
cyclic

cyclic

Using these three inequalities, we get the result.

22

Dividing by a + b + c gives the equivalent inequality

P
cyclic

a
a+b+c
a2
(a+b+c)2

8bc
(a+b+c)2

1.

INFINITY

179

Epsilon 90. [IMO 1983/6 USA] Let a, b, c be the lengths of the sides of a triangle. Prove
that
a2 b(a b) + b2 c(b c) + c2 a(c a) 0.
Second Solution. We present a convexity proof. After setting a = y+z, b = z+x, c = x+y
for x, y, z > 0, it becomes
x3 z + y 3 x + z 3 y x2 yz + xy 2 z + xyz 2
or

y2
z2
x2
+
+
x + y + z.
y
z
x
Since its homogeneous, we can restrict our attention to the case x + y + z = 1. Then, it
becomes

y
z
x
yf
+ zf
+ xf
1,
y
z
x
where f (t) = t2 . Since f is convex on R, we apply (the weighted) Jensens Inequality to
obtain

y
z
x
x
y
z
yf
+ zf
+ xf
f y +z +x
= f (1) = 1.
y
z
x
y
z
x

180

INFINITY

Epsilon 91. (KMO Winter Program Test 2001) Prove that, for all a, b, c > 0,
p
p
(a2 b + b2 c + c2 a) (ab2 + bc2 + ca2 ) abc + 3 (a3 + abc) (b3 + abc) (c3 + abc)
First Solution. Dividing by abc, it becomes
s
s

2
2

a
c
a2
b
c
b
c
a
b
+ +
+ +
1+ 3
+1
+1
+1 .
c
a
b
a
b
c
bc
ca
ab
After the substitution x =
the form
p

a
,
b

y = cb , z =

c
,
a

we obtain the constraint xyz = 1. It takes

(x + y + z) (xy + yz + zx) 1 +

s
3

x
z

+1

y
x

+1

z
y

+1 .

From the constraint xyz = 1, we obtain the identity

x
y
z
x + z y + x z + y
+1
+1
+1 =
= (z + x)(x + y)(y + z).
z
x
y
z
x
y
Hence, we are required to prove that
p
p
(x + y + z) (xy + yz + zx) 1 + 3 (x + y)(y + z)(z + x).
Now, we offer two ways to finish the proof.
First Method. Observe that
(x + y + z) (xy + yz + zx) = (x + y)(y + z)(z + x) + xyz = (x + y)(y + z)(z + x) + 1.
p
Letting p = 3 (x + y)(y + z)(z + x), the inequality we want to prove now becomes
p
p3 + 1 1 + p.
Applying The AM-GM Inequality yields
q

p 3 2 xy 2 yz 2 zx = 2.
It follows that

(p3 + 1) (1 + p)2 = p(p + 1)(p 2) 0,

as desired.
Second Method. More strongly, we establish that, for all x, y, z > 0,

p
1 y+z
x+y
z+x

(x + y + z) (xy + yz + zx) 1 +
+
+
.

3
yz
xy
zx
However, an application of The Cauchy-Schwarz Inequality yields
2

2
p
y+z
[x + (y + z)] [yz + x(y + z)]
xyz + x(y + z)2 = 1 +
yz
or
p
y+z
(x + y + z) (xy + yz + zx) 1 + .
yz
Similarly, we also have
p
z+x
(x + y + z) (xy + yz + zx) 1 + .
zx
and
p
x+y
(x + y + z) (xy + yz + zx) 1 + .
xy
Adding these three, we get the desired inequality.

INFINITY

181

Epsilon 92. [IMO 1999/2 POL] Let n be an integer with n 2.

(a) Determine the least constant C such that the inequality
0
14
X
X
` 2
2
xi A
xi xj xi + xj C @
1in

1i<jn

holds for all real numbers x1 , , xn 0.

(b) For this constant C, determine when equality holds.
First Solution. (Marcin E. Kuczma23) For x1 = = xn = 0, it holds for any C 0.
Hence, we consider the case when x1 + + xn > 0. Since the inequality is homogeneous,
we may normalize to x1 + + xn = 1. From the assumption x1 + + xn = 1, we have
X
`

F(x1 , , xn ) =
xi xj x2i + x2j
1i<jn

1i<jn

1in

xi 3 xj +
xi

xi xj 3

1i<jn

xi

j6=i

xi 3 (1 xi )

1in
n
X

xi (xi 2 xi 3 ).

i=1

We claim that C = 18 . It suffices to show that F (x1 , , xn )

Lemma 8.3. 0 x y

1
2

1
8

=F

`1
2

, 12 , 0, , 0 .

implies x2 x3 y 2 y 3 .

Proof. Since x + y 1, we get x + y (x + y)2 x2 + xy + y 2 . Since y x 0, this

implies that y 2 x2 y 3 x3 or y 2 y 3 x2 x3 , as desired.

Case 1.

1
2

x1 x2 xn :
n
X

xi (xi xi )

i=1

Case 2. x1
y x2 , , xn ,

1
2

n
X

xi

i=1

2 3 !
n
1
1
1
1X
xi = .

=
2
2
8 i=1
8

x2 xn : Let x1 = x and y = 1 x = x2 + + xn . Since

F(x1 , , xn ) = x3 y +

n
X

xi (xi 2 xi 3 ) x3 y +

i=2

n
X

xi (y 2 y 3 ) = x3 y + y(y 2 y 3 ).

i=2

Since x3 y + y(y 2 y 3 ) = x3 y + y 3 (1 y) = xy(x2 + y 2 ), it remains to show that

1
xy(x2 + y 2 ) .
8
Using x + y = 1, we homogenize the above inequality as following.
1
xy(x2 + y 2 ) (x + y)4 .
8
However, we immediately find that (x + y)4 8xy(x2 + y 2 ) = (x y)4 0.

23

I slightly modified his solution in [AS].

182

INFINITY

Epsilon 93. (APMO 1991) Let a1 , , an , b1 , , bn be positive real numbers such that
a1 + + an = b1 + + bn . Show that
an 2
a1 + + an
a1 2
+ +

.
a1 + b 1
an + b n
2
Second Solution. By The Cauchy-Schwarz Inequality, we have
! n
!
!2
n
n
X
X
X ai 2
ai + b i

ai
ai + b i
i=1
i=1
i=1
or

n
X
i=1

`Pn 2
n
ai 2
1X
i=1
Pn
Pn
=
ai
ai + b i
2 i=1
i=1 ai +
i=1 bi

INFINITY

183

Epsilon 94. Let a, b 0 with a + b = 1. Prove that

p
p

a2 + b + a + b2 + 1 + ab 3.
Show that the equality holds if and only if (a, b) = (1, 0) or (a, b) = (0, 1).
Second Solution. The Cauchy-Schwarz Inequality shows that
p
p
p

a2 + b + a + b2 + 1 + ab
3 (a2 + b + a + b2 + 1 + ab)
p
=
3 (a2 + ab + b2 + a + b + 1)
p

3 ((a + b)2 + a + b + 1)
=

3.

184

INFINITY

Epsilon 95. [LL 1992 UNK] (Iran 1998) Prove that, for all x, y, z > 1 such that
2,
p

x + y + z x 1 + y 1 + z 1.

1
+ y1
x

+ z1 =

Third Solution. We first note that

x1
y1
z1
+
+
= 1.
x
y
z
Apply The Cauchy-Schwarz Inequality to deduce
s

y1
z1
x1
+
+
x + y + z = (x + y + z)
x 1 + y 1 + z 1.
x
y
z

INFINITY

185

Epsilon 96. (Gazeta Matematic

a) Prove that, for all a, b, c > 0,
p
p
p
p
p
p
a4 + a2 b2 + b4 + b4 + b2 c2 + c4 + c4 + c2 a2 + a4 a 2a2 + bc+b 2b2 + ca+c 2c2 + ab.
Solution. We obtain the chain of equalities and inequalities
s

X p
X
a2 b 2
a2 b 2
a4 + a2 b 2 + b 4 =
a4 +
+ b4 +
2
2
cyclic
cyclic
!
r
r
2 b2
2 b2
a
a
1 X
a4 +
+ b4 +

2
2
2 cyclic
!
r
r
1 X
a2 b 2
a2 c 2
=
a4 +
+ a4 +
2
2
2 cyclic
s

X 4
a2 b 2
a2 c 2
2
a4 +
a4 +

2
2
cyclic
r
X
a2 bc

2
a4 +
2
cyclic
X p
=
2a4 + a2 bc .

(Cauchy Schwarz)

(AM GM)
(Cauchy Schwarz)

cyclic